Trick

Quantitative Methods
for the Management Sciences

45-760
Course Notes
Instructors:
Gerard Cornuejols
Michael Trick
Graduate School of Industrial Administration
Carnegie Mellon University, Pittsburgh, PA 15213 USA
Fall 1998
Contents
0.1 Introduction : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 4
0.2 Grading Components : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 5
0.3 Schedule : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 6
0.4 Mathematical Methods in Business : : : : : : : : : : : : : : : : : : : : : : : : : : : : 6
0.4.1 History : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 6
0.4.2 Mathematical Modeling Today : : : : : : : : : : : : : : : : : : : : : : : : : : 7
1 Basic Linear Algebra 9
1.1 Linear Equations : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 9
1.2 Operations on Vectors and Matrices : : : : : : : : : : : : : : : : : : : : : : : : : : : 12
1.3 Linear Combinations : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 15
1.4 Inverse of a Square Matrix : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 20
1.5 Determinants : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 23
1.6 Positive Denite Matrices : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 23
2 Unconstrained Optimization: Functions of One Variable 25
2.1 Derivatives : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 25
2.2 Maximum and Minimum : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 26
2.3 Binary Search : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 30
2.4 Convexity : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 33
3 Unconstrained Optimization: Functions of Several Variables 37
3.1 Gradient : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 37
3.2 Maximum and Minimum : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 39
3.3 Global Optima : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 41
4 Constrained Optimization 43
4.1 Equality Constraints (Lagrangians) : : : : : : : : : : : : : : : : : : : : : : : : : : : : 43
4.1.1 Geometric Interpretation : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 45
4.1.2 Economic Interpretation : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 45
4.2 Equality and Inequality Constraints : : : : : : : : : : : : : : : : : : : : : : : : : : : 49
4.3 Exercises : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 53
5 Modeling with Linear Programming 57
5.1 Introductory Example : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 57
5.1.1 Modeling : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 57
5.1.2 Solving the Model : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 58
5.1.3 Using Solver : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 61
1
2 CONTENTS
5.2 Second Example: Market Allocations : : : : : : : : : : : : : : : : : : : : : : : : : : : 62
5.2.1 Modeling : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 63
5.2.2 Solving the Model : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 63
5.2.3 Using Solver : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 64
5.3 Common Features : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 64
5.4 Linear Programming Models : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 65
5.4.1 Diet Problem : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 65
5.4.2 Workforce Planning : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 66
5.4.3 Financial Portfolio : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 67
5.4.4 Production over Time : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 69
5.4.5 Investing over Time : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 70
5.4.6 Problems : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 72
6 Case Study 79
7 The Simplex Method 87
7.1 Linear Programs in Standard Form : : : : : : : : : : : : : : : : : : : : : : : : : : : : 87
7.2 Solution of Linear Programs by the Simplex Method : : : : : : : : : : : : : : : : : : 88
7.3 Alternate Optimal Solutions, Degeneracy, Unboudedness, Infeasibility : : : : : : : : 92
8 Sensitivity Analysis for Linear Programming 97
8.1 Tableau Sensitivity Analysis : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 97
8.1.1 Cost Changes : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 97
8.1.2 Right Hand Side Changes : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 99
8.1.3 New Variable : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 100
8.2 Solver Output : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 104
8.2.1 Tucker Automobiles : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 104
8.2.2 Carla's Maps : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 106
9 Decision Theory 113
9.1 Decisions Under Risk : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 114
9.2 Decision Trees : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 115
10 Game Theory 125
10.1 Two-Person Games : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 125
10.1.1 Pure Strategies : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 125
10.1.2 Mixed Strategies : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 129
10.2 n-Person Games : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 132
11 Network Optimization 135
11.1 Introduction : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 135
11.2 Terminology : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 135
11.3 Examples : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 136
11.3.1 Shortest Paths : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 136
11.3.2 Maximum Flow : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 137
11.3.3 Transportation Problem : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 138
11.3.4 Assignment Problem : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 141
11.4 Unifying Model: Minimum Cost Network Flows : : : : : : : : : : : : : : : : : : : : : 142
CONTENTS 3
11.5 Generalized Networks : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 144

11.6 Conclusions : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 145
12 Data Envelopment Analysis 147
12.1 Numerical Example : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 147
12.2 Graphical Example : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 148
12.3 Using Linear Programming : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 149
12.4 Applications : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 155
12.5 Strengths and Limitations of DEA : : : : : : : : : : : : : : : : : : : : : : : : : : : : 155
12.6 References : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 156
4 CONTENTS
Course Syllabus
0.1 Introduction
Objectives of the Course:
to provide a formal quantitative approach to problem solving;
to give an intuition for managerial situations where a quantitative approach is appropriate;
to introduce some widely used quantitative models;
to introduce software for solving such models.
To accomplish these objectives we provide:
background in linear algebra and nonlinear optimization;
description of the linear programming model, decision theory, etc.;
explanation of the mathematical ideas behind these widely used models;
hands-on experience with software such as SOLVER;
examples of business applications;
a case study that shows how models are used in practice.
Instructors
Gerard Cornuejols, GSIA 232A, phone 268-2284, fax 268-7357
email: [email protected]
secretary: Barbara Carlson 268-1342, GSIA 232.
Michael Trick, GSIA 257C, phone 268-3697, fax 268-7057
email: [email protected]
Required Material for the Course
The course notes are self contained. The software for this course is installed in your notebooks.
Instruction on using it will be provided in class.
0.2. GRADING COMPONENTS 5
0.2 Grading Components

The course grade will be based on homework, a case study, and two exams.
Homework
The only way to learn mathematical material is by working problems.
Homework is due at the beginning of class on the due date (see schedule). No late homework
will be accepted.
You are encouraged to work exercises that are not assigned. Some can be found in the course
notes. More can be found in the textbook by Winston, Operations Research: Applications
and Algorithms, Duxbury Press (1994). Copies are on reserve in the Hunt library.
You should do the homework without consulting other students. If you need help, you may
consult the instructor or a teaching assistant. For nonassigned exercises, it is a good idea to
work them out with other students.
Every attempt will be made to return graded homeworks promptly. Solutions to the homework
will be distributed in class after the due date.
The following rules apply to all homework:
{ Use only 11 by 8.5 in. paper.
{ Start each problem on a new page.
{ Assemble the homework in proper order and staple it together.
{ PRINT your name on the top right hand corner.
Case Study
The case study has two parts. The rst part will require formulating a linear program. The
second part requires solving the linear program and analyzing the results for managerial purposes.
Exams
There will be one midterm and one nal exam. The precise dates of the exams will be
announced by the GSIA administration.
The questions in the exams will be similar to the homework problems and the problems dis-
cussed in class both in form and content. The lectures may occasionally range over peripheral
topics, but the content of the exercises is a reliable guide to which topics can appear on the
exam.
All exams will be open-notes.
Except for provable unforeseen circumstances, makeup exams will not be given. A student is
required to inform the instructor as soon as possible in such a case. In case of an unavoidable
con ict, it is the student's responsibility to schedule a makeup exam before the exam date
after discussing with the instructor.
Grading Weights
Midterm exam = 25%, homework = 15%, case study = 25%, nal exam = 35%.
6 CONTENTS
0.3 Schedule
It is advisable to read through the material in the course notes prior to when it is discussed in class.
Date Topic Chapter Work Due

Aug 25 Linear Algebra Chapter 1
Aug 27 Functions of One Variable Chapter 2
Sept 1 Functions of Several Variables Chapter 3 HW 1 due
Sept 3 Optimization with Constraints Chapter 4
Sept 8 LP modeling / Solver Chapter 5 HW 2 due
Sept 10 LP modeling Chapter 5
Sept ?? MIDTERM EXAM TBA
Sept 15 The Simplex Method Chapter 7 HW 3 due
Sept 17 Sensitivity Analysis Chapter 8
Sept 22 Case Study Part I Chapter 6 Case Part I due
Sept 24 Decision Theory Chapter 9
Sept 29 Game Theory Chapter 10 HW 4 due
Oct 1 Network Models Chapter 11
Oct 6 Data Envelopment Analysis Chapter 12 Case Part II due
Oct 8 Problem Session HW 5 due
Oct ?? FINAL EXAM TBA
0.4 Mathematical Methods in Business

0.4.1 History
Mathematical methods have long played an important role in management and economics but
have become increasingly important in the last few decades. \Business mathematics," such as the
computation of compound interest, appeared in ancient Mesopotamia but became prevalent in the
mercantile economy of the Renaissance. Mathematics later contributed to the engineering advances
that made the Industrial Revolution possible. Mathematics crept into economics during the 18th
and 19th centuries and became rmly established in the 20th.
It is only in the last few decades that management per se has received the sort of rigorous
study that permits the application of mathematics. In the early part of this century \scientic
management," called \Taylorism" after its founder F. Taylor, was fashionable. This movement
made some worthwhile contributions, such as time-and-motion studies, but it was coupled with
dreadful labor relations practices. The symbol of the movement was the eciency expert policing
the shop oor with stopwatch and clipboard.
After Taylorism waned, interest shifted more to making ecient use of labor and other factors
than making employees work harder. Shortly before the Second World War a team of scientists
solved the operational problems of coordinating Britain's radar stations and thereby created \op-
erational research," called \operations research" in the U.S. During the war this practice of putting
scientists and mathematicians to work on operational problems was surprisingly successful in both
the British and American military, partly because members of the professional class who had not
previously \dirtied their hands" had their rst opportunity to apply their training to operational
0.4. MATHEMATICAL METHODS IN BUSINESS 7
problems. Their success attracted attention, and operations research spread to industry during the
fties.
After the war, G. Dantzig and others developed linear programming at about the same time
that computers became available to solve linear programming models. Linear programming proved
a remarkably useful tool for the ecient allocation of resources, and this gave a great boost to
operations research. In the early postwar years J. von Neumann and O. Morgenstern invented
game theory (closely related to linear programming), and H. W. Kuhn and A. W. Tucker broke
ground in nonlinear programming. W. E. Deming and others developed statistical techniques for
quality control, and the statistical methods rst designed for psychometrics in the early 20th century
became the mathematical basis for econometrics. Meanwhile business schools began teaching many
of these techniques along with microeconomics and other quantitative elds. As a result of all this,
mathematical modeling has played an increasingly important role in management and economics.
In fact one can make a case that overreliance on a mathematical approach has sometimes led
to neglect of other approaches. After its rst few years, operations research focused almost exclu-
sively on mathematical modeling. The reigning orthodoxy in much of economics is the \neoclas-
sical" paradigm, which is heavily mathematical. There is little communication between people in
quantitative elds and those in such \soft" elds as organizational science and general management.
In fact both approaches appear to be in a crisis. Economics and operations research have
achieved much, but neither can adequately understand phenomena that involve human beings.
The soft sciences take on the dicult, ill-structured, human-oriented problems, but it is unclear
even what counts as a satisfactory theory in these areas.
One can hope that this double crisis will lead in the coming years to a \paradigm shift" that
will supplant the \hard" and \soft" approaches with an integrated science with the rigor of one
and the breadth of the other. There are faint signs of such a shift, but they are not suciently
crystallized to show up in a business school curriculum.
Despite the limitations of the quantitative methods discussed in this course, in many specic
areas they are quite useful. In fact the potential of most mathematical techniques, particularly those
of operations research, is only beginning to be realized in practice. Aordable personal computers
and user-friendly software are motivating managers to learn about mathematical models and use
them. Also the line between managerial and technical sta is blurring, so that managers now run
models on their notebook computers that systems analysts once ran on mainframes. This again
leads to more widespread use.
0.4.2 Mathematical Modeling Today
The applications of mathematical methods in management and economics today are so manifold
that it is dicult to nd a single person who is aware of their full scope. The following list can
only be incomplete.
Economists use linear and nonlinear programming, the theory of variational inequalities,
optimal control theory, dynamic programming, game theory, probability choice models, utility
theory, regression and factor analysis and other techniques to study equilibrium, optimal
investment, competition, consumer behavior, and a host of other phenomena.
People in operations management use statistical sampling and estimation theory, linear and
integer programming, network programming, dynamic programming and optimal control the-
ory, queuing theory, simulation, articial intelligence techniques, and combinatorial optimiza-
tion methods to solve problems in quality control, allocation of resources, logistics, project
8 CONTENTS
scheduling, labor and machine scheduling, job shop scheduling and assembly line balancing,
and facility layout and location. The introduction of exible manufacturing systems, robots
and other automated devices has posed a whole new array of unsolved mathematical problems.
People in nance use linear, nonlinear and integer programming, optimal control theory
and dynamic programming, Markov decision theory, regression and time series to determine
optimal resource allocation, multiperiod investments, capital budgeting, and investment and
loan portfolio design, and to try to forecast market behavior.
People in marketing use regression and factor analysis, time series, game theory, Markov
decision theory, location theory, mathematical programming, probability choice models and
utility theory to study consumer preferences, determine optimal location in product space,
allocate advertising resources, design distribution systems, forecast market behavior, and
study competitive strategy.
People in information systems and decision support systems use articial intelligence tech-
niques, propositional and quantied logic, Bayesian methods, probabilistic logic, data struc-
tures and other computer science techniques, mathematical programming, and statistical de-
cision theory to design expert and other knowledge-based systems, develop ecient inference
and retrieval methods, and evaluate the economic and organizational eects of information
systems.
The peculiar nature of mathematics unfortunately raises two obstacles to learning it. One is that
students often believe that mathematics can be learned simply by studying a book. Mathematics
is learned by working through problems. A second obstacle is that students often believe that they
can learn to work problems by studying a book. A book can get one started, but learning to work
problems is like learning to play basketball or play the piano|it requires practice, practice, practice.
Mathematical skills are very athletic in this sense. (The phrase \quant jock" is appropriate.) That
is why this course assigns a lot of homework problems.
Acknowledgements: We would like to thank John Hooker, Bill Hrusa, Rick Green and Anuj
Mehrotra for their inputs.
Chapter 1
Basic Linear Algebra

Linear equations form the basis of linear programming. If you have a good understanding of the
Gauss-Jordan method for solving linear equations, then you can also understand the solution of
linear programs. In addition, this chapter introduces matrix notation and concepts that will be
used in Chapters 3 and 4 (optimizing functions of several variables).
1.1 Linear Equations

The Gauss-Jordan elimination procedure is a systematic method for solving systems of linear equa-
tions. It works one variable at a time, eliminating it in all rows but one, and then moves on to the
next variable. We illustrate the procedure on three examples.
Example 1.1.1 (Solving linear equations)
x1 + 2x2 + x3 = 4
2x1 , x2 + 3x3 = 3
x1 + x2 , x3 = 3
In the rst step of the procedure, we use the rst equation to eliminate x1 from the other two.
Specically, in order to eliminate x1 from the second equation, we multiply the rst equation by 2
and subtract the result from the second equation. Similarly, to eliminate x1 from the third equation,
we subtract the rst equation from the third. Such steps are called elementary row operations. We
keep the rst equation and the modied second and third equations. The resulting equations are
x1 + 2x2 + x3 = 4
, 5x2 + x3 = ,5
, x2 , 2x3 = ,1
Note that only one equation was used to eliminate x1 in all the others. This guarantees that the
new system of equations has exactly the same solution(s) as the original one. In the second step of
the procedure, we divide the second equation by -5 to make the coecient of x2 equal to 1. Then,
we use this equation to eliminate x2 from equations 1 and 3. This yields the following new system
of equations.
x1 + 57 x3 = 2
x2 , 15 x3 = 1
, 115 x3 = 0
9
10 CHAPTER 1. BASIC LINEAR ALGEBRA
Once again, only one equation was used to eliminate x2 in all the others and that guarantees
that the new system has the same solution(s) as the original one. Finally, in the last step of the
procedure, we use equation 3 to eliminate x3 in equations 1 and 2.
x1 = 2
x2 = 1
x3 = 0
So, there is a unique solution. Note that, throughout the procedure, we were careful to keep three
equations that have the same solution(s) as the original three equations. Why is it useful? Because,
linear systems of equations do not always have a unique solution and it is important to identify
such situations.
Another example:
x1 + 2x2 + x3 = 4
x1 + x2 + 2x3 = 1
2x1 + 3x2 + 3x3 = 2
First we eliminate x1 from equations 2 and 3.
x1 + 2x2 + x3 = 4
, x2 + x3 = ,3
, x2 + x3 = ,6
Then we eliminate x2 from equations 1 and 3.
x1 + 3x3 = ,2
x2 , x3 = 3
0 = ,3
Equation 3 shows that the linear system has no solution.
A third example:
x1 + 2x2 + x3 = 4
x1 + x2 + 2x3 = 1
2x1 + 3x2 + 3x3 = 5
Doing the same as above, we end up with
x1 + 3x3 = ,2
x2 , x3 = 3
0 = 0
Now equation 3 is an obvious equality. It can be discarded to obtain
x1 = ,2 , 3x3
x2 = 3 + x3
The situation where we can express some of the variables (here x1 and x2 ) in terms of the remaining
variables (here x3 ) is important. These variables are said to be basic and nonbasic respectively.
Any choice of the nonbasic variable x3 yields a solution of the linear system. Therefore the system
has innitely many solutions.
1.1. LINEAR EQUATIONS 11
It is generally true that a system of m linear equations in n variables has

either:
(a) no solution,
(b) a unique solution,
(c) innitely many solutions.
The Gauss-Jordan elimination procedure solves the system of linear equa-
tions using two elementary row operations:
modify some equation by multiplying it by a nonzero scalar (a scalar
is an actual real number, such as 21 or -2; it cannot be one of the
variables in the problem),
modify some equation by adding to it a scalar multiple of another
equation.
The resulting system of m linear equations has the same solution(s) as
the original system. If an equation 0 = 0 is produced, it is discarded
and the procedure is continued. If an equation 0 = a is produced where
a is a nonzero scalar, the procedure is stopped: in this case, the system
has no solution. At each step of the procedure, a new variable is made
basic: it has coecient 1 in one of the equations and 0 in all the others.
The procedure stops when each equation has a basic variable associated
with it. Say p equations remain (remember that some of the original
m equations may have been discarded). When n = p, the system has
a unique solution. When n > p, then p variables are basic and the
remaining n , p are nonbasic. In this case, the system has innitely
many solutions.
Exercise 1 Solve the following systems of linear equations using the Gauss-Jordan elimination
procedure and state whether case (a), (b) or (c) holds.
(1)
3x1 , 4x3 = 2
x1 + x2 + x3 = 4
2x2 + x3 = 3
(2)
2x1 + 2x2 , x3 = 1
4x1 + x3 = 2
x1 , x2 + x3 = 2
(3)
x1 , x2 + x3 = 1
,2x1 + 2x2 , 2x3 = ,2
,x1 + x2 , x3 = ,1
Exercise 2 Indicate whether the following linear system of equations has 0, 1 or innitely many
solutions.
x1 + 2x2 + 4x3 + x4 + 3x5 = 2
2x1 + x2 + x3 + 3x4 + x5 = 1
3x2 + 7x3 , x4 + 5x5 = 6
1.2 Operations on Vectors and Matrices

It is useful to formalize the operations on vectors and matrices that form the basis of linear algebra.
For our purpose, the most useful denitions are the following.
A matrix is a rectangular array of numbers written in the form
0 a11 a12 : : : a1n 1
A=B
@ ..
.
..
.
.. C
. A:
am1 am2 : : : amn
The matrix A has dimensions m n if it has m rows and n columns. When m = 1, the matrix is
called a row vector; when n = 1, the matrix is called a column vector. A vector can be represented
either by a row vector or a column vector.
Equality of two matrices of same dimensions:
0 a11 : : : a1n 1 0 b11 : : : b1n 1
Let A = B@ ... .. C B .
. A and B = @ ..
.. C
. A.
am1 : : : amn bm1 : : : bmn
Then A = B means that aij = bij for all i; j .
Multiplication of a matrix A by a scalar k:
0 ka11 ka12 : : : ka1n 1
kA = B
@ ... ..
.
.. C
. A:
kam1 kam2 : : : kamn
Addition of two matrices of same dimensions:

0 a11 : : : a1n 1 0 b11 : : : b1n 1
Let A = B
@ ... .. C B .
. A and B = @ ..
.. C
. A.
am1 : : : amn
0 a11 + b11 : : : a1n + bb1mn1 1: : : bmn
Then A + B = B@ ..
.
..
.
CA :
am1 + bm1 : : : amn + bmn
Note that A + B is not dened when A and B have dierent dimensions.
! !
Exercise 3 Compute 3 20 35 14 , 2 01 46 16 .
1.2. OPERATIONS ON VECTORS AND MATRICES 13
Multiplication
0 a of a: :matrix 1of dimensions m n by a matrix of dimensions n p:
11 : a 1n
B
B .. .. C 0 b11 : : : b1j : : : b1n 1
B . . C
C
Let A = BB
B ai1 : : : ain CCC and B = B@ ... ..
.
.. C
. A.
B
@ . .
. .
.
. AC bm1 : : : bmj : : : bmn
am1 : : : amn
Then AB is a matrix of dimensions m p computed as follows.
0 a11b11 + : : : + a1nbn1 ::: a11b1p + : : : + a1n bnp 1
AB = B
@ ..
. ai1 b1j + : : : + ain bnj ..
.
CA :
am1 b11 + : : : + amn bn1 ::: am1 b1p + : : :amn bnp
As an example, let us multiply the matrices
! 0 1
11 12
A = 25 31 40 and B = B
@ 14 15 CA :
17 18
! 0 132 141 1
11 + 3 14 + 4 17 2 12 + 3 15 + 4 18 = B
The result is AB = 52 @ C
A:
11 + 1 14 + 0 17 5 12 + 1 15 + 0 18 69 75
Note that AB is dened only when the number of columns of A equals the number of rows of B .
An important remark: even when both AB and BA are dened, the results are usually dierent.
A property of matrix multiplication is the following:
(AB )C = A(BC ):
That is, if you have three matrices A, B , C to multiply and the product is legal (the number of
columns of A equals the number of rows of B and the number of columns of B equals the number
of rows of C ), then you have two possibilities: you can rst compute AB and multiply the result
by C , or you can rst compute BC and multiply A by the result.
Exercise 4 Consider the following matrices
0 1
1
x= 1 1 1 y= 3 2 1 z=B
@ 1 CA :
1
When possible, compute the following quantities:
(a) xz
(b) zx
(c) yzx
(d) xzy
(e) (x + y )z
(f) (xz ) + (yz ).
Remark: A system of linear equations can be written conveniently using matrix notation. Namely,
a11x1 + : : : + a1nxn = b1
::: ::: :::
am1 x1 + : : : + amnxn = bm
can be written as 0 a11 : : : a1n 1 0 x1 1 0 b1 1
B
@ ..
.
.. C B . C B . C
. A @ .. A = @ .. A
am1 : : : amn xn bm
or as 0 a11 1 0 a1n 1 0 b1 1
B
@ .. C B
. A x1 + : : : + @
.. C B
. A xn = @
.. C
. A:
am1 amn bm
So a matrix equation Ax = b where A is a given m n matrix, b is a given m-column vector
and x is an unknown n-column vector, is a linear system of m equations in n variables. Similarly, a
vector equation a1 x1 + : : : + an xn = b where a1 ; : : :; an ; b are given m-column vectors and x1 ; : : :; xn
are n unknown real numbers, is also a system of m equations in n variables.
Exercise 5 (a) Solve the matrix equation
0 10 1 0 1
B@ 2 ,1 3 CA B@ xx12 CA = B@ 43 CA :
1 2 1
1 1 ,1 x3 3
(b) Solve the vector equation
0 1 0 1 0 1 0 1
1 2 1 4
@ 2 CA x1 + B@ ,1 CA x2 + B@ 3 CA x3 = B@ 3 CA :
B
1 1 ,1 3
[Hint: Use Example 1.1.1]
The following standard denitions will be useful:

A square matrix is a matrix with same number of rows as columns.
The 0identity matrix I is 1a square matrix with 1's on the main diagonal and 0's elsewhere, i.e.
BB 10 01 :: :: :: 00 CC
I=B @ : : : : : : : : : : : : CA.
0 0 ::: 1
The zero matrix contains only 0's. In particular, the zero vector has all components equal to 0. A
nonzero vector is one that contains at least one nonzero component.
1.3. LINEAR COMBINATIONS 15
0 1
0 a11 a12 : : : a1n 1 Ba11 : : : am1
CA is the matrix AT = BBa12 : : : am2 CC
The transpose of matrix A = B
@ ..
.
..
.
..
. B@ .. .. C
CA.
am1 am2 : : : amn . .
a1n : : : amn
A square matrix A is symmetric if its elements satisfy aij = aji for all i and j . So a symmetric
matrix satises AT = A.
1.3 Linear Combinations

Suppose that the vector (1; 3) represents the contents of a \Regular" can of mixed nuts (1 lb cashews
and 3 lb peanuts) while (1; 1) represents a \Deluxe" can (1 lb cashews and 1 lb peanuts). Can you
obtain a mixture of 2 lb cashews and 3 lb peanuts by combining the two mixtures in appropriate
amounts? The answer is to use x1 cans of Regular and x2 cans of Deluxe in order to satisfy the
equality ! ! !
1 1
x1 3 + x2 1 = 3 : 2
This is none other than a system of two linear equations:

x1 + x2 = 2
3x1 + x2 = 3
The solution of these equations (obtained by the Gauss-Jordan procedure) is x1 = 1=2, x2 = 3=2.
So the desired combination is to mix 1/2 can of Regular nuts with 3/2 cans of Deluxe nuts. Thus
if some recipe calls for the mixture (2; 3), you can substitute 1/2 can of Regular mix and 3/2 can
of Deluxe mix.
Suppose now that you want to obtain 1 lb cashews and no peanuts. This poses the equations,
! ! !
1 1
x1 3 + x2 1 = 0 : 1
The solution is (x1 ; x2) = (,1=2; 3=2). Thus you can obtain a pound of pure cashews by buying
3/2 cans of Deluxe mix and removing enough nuts to form 1/2 can Regular mix, which can be sold.
In this case it is physically possible to use a negative amount of some ingredient, but in other cases
it may be impossible, as when one ! is mixing ! paint. !
1 1
A vector of the form x1 3 + x2 1 is called a linear combination of the vectors 3 and 1
! ! !
1 . In particular we just saw that the vector 2 is a linear combination of the vectors 1
1 ! ! 3 3
and 11 . And so is 10 .
The question arises: can one obtain any mixture !whatever by taking the appropriate
! combina-!
b1 1
tion of Regular and Deluxe cans? Is any vector b a linear combination of 3 and 1 ? 1
2
To answer the question, solve the equations in general. You want a mixture of b1 lb cashews
and b2 lb peanuts and set ! ! !
1 1
x1 3 + x2 1 = b ; b 1
2
or equivalently,
x1 + x2 = b1
3x1 + x2 = b2:
These equations have a unique solution, no matter what are the values of b1 and b2, namely,
x1 = b2 ,2 b1 ; x2 = 3b1 2, b2 :
No matter what vector (b1; b2) you want, you can get it as a linear combination of (1; 3) and (1; 1).
The vectors (1; 3) and (1; 1) are said to be linearly independent.
Not all pairs of vectors can yield an arbitrary mixture (b1; b2). For instance, no linear combi-
nation of (1; 1) and (2; 2) yields (2,3). In other words, the equations
! ! !
x1 11 + x2 22 = 23
have no solution.
The reason is that (2; 2) is already a multiple of (1; 1), that is (2; 2) is a linear combination of
(1; 1). The vectors (1; 1) and (2; 2) are said to be linearly dependent.
If (2,2) represents, for instance, a large Deluxe can of nuts and (1,1) a small one, it is clear that
you cannot obtain any mixture you want by combining large and small Deluxe cans. In fact, once
you have small cans, the large cans contribute nothing at all to the mixtures you can generate,
since you can always substitute two small cans for a large one.
A more interesting example is (1,2,0), (1,0,1) and (2,2,1). The third vector is clearly a linear
combination of the other two (it equals their sum). Altogether, these three vectors are said to be
linearly dependent. For instance, these three vectors might represent mixtures of nuts as follows:
Brand A B C
cashews 1 1 2
peanuts 2 0 2
almonds 0 1 1
Then once you have brands A and B, brand C adds nothing whatever to the variety of mixtures
you can concoct. This is because you can already obtain a can of brand C from brands A and B
anyway. In other words, if a recipe calls for brand C, you can always substitute a mixture of 1 can
brand A and 1 can brand B.
Suppose you want to obtain the mixture (1; 2; 1) by combining Brands A, B, and C. The
equations are
x1 + x2 + 2x3 = 1
2x1 + 2x3 = 2
x2 + x3 = 1
If you try to solve these equations, you will nd that there is no solution. So the vector (1; 2; 1) is
not a linear combination of (1; 2; 0), (1; 0; 1) and (2; 2; 1).
The concepts of linear combination, linear dependence and linear independence introduced in
the above examples can be dened more formally, for any number of n-component vectors. This is
done as follows.
A vector b having n components is a linear combination of the k

vectors v1 ; : : :; vk , each having n components, if it is possible to nd
k real numbers x1; : : :; xk satisfying:
x1v1 + : : :xk vk = b (1.1)
To nd the numbers x1 ; : : :; xk , view (1.1) as a system of linear
equations and solve by the Gauss-Jordan method.
A set of vectors (all having n components) is linearly dependent if
at least one vector is a linear combination of the others. Otherwise
they are linearly independent.
Given n linearly independent vectors v1; : : :; vn, each having n com-
ponents, any desired vector b with n components can be obtained
as a linear combination of them:
x1 v1 + : : :xn vn = b (1.2)
The desired weights x1 ; : : :; xn are computed by solving (1.2) with
the Gauss-Jordan method: there is a unique solution whenever the
vectors v1 ; : : :; vn are linearly independent.
Exercise 6 A can of Brand A mixed nuts has 1 lb cashews, 1 lb almonds, 2 lb peanuts. Brand B
has 1 lb cashews and 3 lb peanuts. Brand C has 1 lb almonds and 2 lb peanuts. Show how much
of each brand to buy/sell so as to obtain a mixture containing 5 lb of each type of nut.
0 1
1
Exercise 7 Determine whether the vector B@ 2 CA is a linear combination of
,2
0 10 1
1 2
(a) B
@ 0 CA, B
@0C A
2 1
0 10 10 1
2 1 ,1
(b) B
@ 1 CA, B@ 2 CA, B@ 1 C
A.
,1 ,1 ,2
In linear programming, there are typically many more variables than equations. So, this case
warrants looking at another example.
Example 1.3.1 (Basic Solutions.) A craft shop makes deluxe and regular belts. Each deluxe
belt requires a strip of leather and 2 hours of labor. Each regular belt requires a leather strip and
1 hour of labor. 40 leather strips and 60 hours of labor are available. How many belts of either
kind can be made?
This is really a mixing problem. Rather than peanuts and cashews, the mixture will contain
leather and labor. The items to be mixed are four activities: manufacturing a deluxe belt, manu-
facturing a regular belt, leaving a leftover leather strip in inventory, and leaving an hour of labor
idle (or for other work). Just as each Regular can of mixed nuts contributes 1 pound of cashews
and 3 pounds of peanuts to the mixture, each regular belt will consume 1 leather strip and 2 hours
of labor. The aim is to combine the four activities in the right proportion so that 40 strips and 60
hours are accounted for:
! ! ! ! !
1 1 1
x1 2 + x2 1 + s1 0 + s2 1 = 60 : 0 40
So,
x1 = number of deluxe belts made
x2 = number of regular belts made
s1 = number of leather strips left over
s2 = number of labor hours left over
In tableau form, the equations are:
x1 x2 s1 s2 RHS
1 1 1 0 40 (1.3)
2 1 0 1 60
Since there are only two equations, you can only solve for two variables. Let's solve for x1 and x2,
using Gauss-Jordan elimination. After the rst iteration you get,
x1 x2 s1 s2 RHS
1 1 1 0 40 (1.4)
0 ,1 ,2 1 ,20
A second iteration yields the solution,
x1 x2 s1 s2 RHS
1 0 ,1 1 20 (1.5)
0 1 2 ,1 20
This tableau represents the equations
x1 , s1 + s2 = 20
x2 + 2s1 , s2 = 20
or,
x1 = 20 + s1 , s2 (1.6)
x2 = 20 , 2s1 + s2
You can't say how many deluxe belts x1 and regular belts x2 the plant will make until you specify
how much leather s1 and labor s2 will be left over. But (1.6) is a formula for computing how many
belts of either type must be made, for any given s1 and s2 . So the equations have many solutions
(innitely many).
For example, if you want to have nothing left over (s1 = s2 = 0), you will make 20 of each. If
you want to have 5 strips and 5 hours left over, you will make x1 = 20 deluxe and x2 = 15 regular
belts.
The variables x1 ; x2 you solved for are called basic variables. The other variables are nonbasic.
You have control of the nonbasic variables. Once you give them values, the values of the basic
variables follow. A solution in which you make all the nonbasic variables zero is called a basic
solution.
Can you have basic variables other than x1; x2? Sure. Any pair of variables can be basic,
provided the corresponding columns in (1.3) are linearly independent (otherwise, you can't solve
for the basic variables).
Equations (1.3), for instance, are already solved in (1.3) for basic variables s1 ; s2. Here the
two basic activities are having leftover leather and having leftover labor. The basic solution is
(s1 ; s2) = (40; 60). This means that if you decide to produce no belts (x1 = x2 = 0), you must have
40 leftover leather strips and 60 leftover labor hours.
The intermediate step (1.4) solves the equations with basic variables x1 and s2 . Here the basic
solution is unrealizable. If you decide to participate only in the basic activities (making deluxe
belts and having leftover labor), you must make 40 belts and have ,20 leftover hours (i.e., use 20
more than you have), which you can't do within your current resources.
Exercise 8 Consider the system of equations
x1 x2 x3 x4 RHS
1 1 2 4 100
3 1 1 2 200
where the rst four columns on the left represent a Regular mixture of nuts, a Deluxe mixture, a
small can of Premium mixture, and a large can of Premium.
a) Solve the system with x1 and x2 basic.
b) You want 100 cans of mixed nuts, each of which contains 1 lb cashews and 2 lb peanuts (i.e.,
you want the right-hand side of the above equation). How can you get them by blending 10
small cans of Premium with proper amounts of the Regular and Deluxe blends? Hint. Set
x3 = 10 and x4 = 0 in the expression for the solution values of x1, x2 found in (a).
c) How can you obtain a small can of Premium mix by combining (and decombining) the Regular
and Deluxe blends?
d) How can you obtain one can of Regular mix by combining large and small cans of Premium?
If you cannot do it, why not? Hint. It has to do with linear dependence.
Exercise 9 A can of paint A has 1 quart red, 1 quart yellow. A can of paint B has 1 quart red, 1
quart blue. A can of paint C has 1 quart yellow, 1 quart blue.
a) How much of each paint must be mixed to obtain a mixture of 1 quart red, 1 quart yellow and
1 quart blue?
b) How much of each paint must be mixed to obtain one quart of pure red? What do you conclude
about the physical feasibility of such a mixture?
Exercise 10 An eletronics plant wants to make stereos and CB's. Each stereo requires 1 power
supply and 3 speakers. Each CB requires 1 power supply and 1 speaker. 100 power supplies and
200 speakers are in stock. How many stereos and CB's can it make if it wants to use all the power
supplies and all but 10 of the speakers? Hint. Use the following equations.
x1 x2 s1 s2 RHS
1 1 1 0 100
3 1 0 1 200
Exercise 11 A construction foreman needs cement mix containing 8 cubic yards (yd3) cement, 12
yd3 sand and 16 yd3 water. On the site are several mixer trucks containing mix A (1 yd3 cement,
3 yd3 sand, 3 yd3 water), several containing mix B (2 yd3 cement, 2 yd3 sand, 3 yd3 water) and
several containing mix C (2 yd3 cement, 2 yd3 sand, 5 yd3 water).
How many truckloads of each mix should the foreman combine to obtain the desired blend?
1.4 Inverse of a Square Matrix

If A and B are square matrices such that AB = I (the identity matrix), then B is called the inverse
of A and is denoted by A,1 . A square matrix A has either no inverse or a unique inverse A,1 .
In the rst case, it is said to be singular and in the second case nonsingular. Interestingly, linear
independence of vectors plays a role here: a matrix is singular if its columns form a set of linearly
dependent vectors; and it is nonsingular if its columns are linearly independent. Another property
is the following: if B is the inverse of A, then A is the inverse of B .
0 10 1
1 3 ,2 , 3 ,3 2
Exercise 12 (a) Compute the matrix product B@ 0 ,5 4 CA B@ 8 7 ,4 CA :
2 ,3 3 10 9 ,5
0 1
,3 ,3 2
(b) What is the inverse of B
@ 8 7 ,4 CA?
10 9 ,5
0 1
0 0 0
(c) Show that the matrix A = B @ 0 1 1 CA is singular.
1 0 1
0 1
b 11 b12 b 13
[Hint: Assume that the inverse of A is B = B
@ b21 b22 b23 CA and perform the matrix product
b31 b32 b33
AB .0Then show1 that no choice of bij can make this product equal to the identity matrix
1 0 0
I=B
@ 0 1 0 CA.]
0 0 1
An important property of nonsingular square matrices is the following. Consider the system of
linear
0 equations 1 0 1
a11 : : : a1n 0 x 1 b1
BB
a21 : : : a2n C 1
CC B .. C B B b2 C
C
BB .. C @ . A = B .. C
B ; simply written as Ax = b.
. C
..
@ . . A x
n
@ A
an1 : : : ann bn
When A is a square nonsingular matrix, this linear system has a unique solution, which can be
obtained as follows. Multiply the matrix equation Ax = b by A,1 on the left:
A,1 Ax = A,1 b:
This yields Ix = A,1 b and so, the unique solution to the system of linear equations is
x = A,1 b:
1.4. INVERSE OF A SQUARE MATRIX 21
Exercise 13 Solve the system

,3x1 , 3x2 + 2x3 = 1
8x1 + 7x2 , 4x3 = ,1
10x1 + 9x2 , 5x3 = 1
using the result of Exercise 12(b).
Finding the Inverse
0 a11 of a Square Matrix
: : : a1n 1 0 b11 : : : b1n 1
Given A = B
@ ... .. C B
. A, we must nd B = @
..
.
.. C
. A such that AB = I (the
an1 : : : ann bn1 : : : bnn 0 1
0 b11 1 B 1 C
B0C
@ ... CA = BB .. CC (this vector
identity matrix). Therefore, the rst column of B must satisfy A B
bn1 @.A
0
is the 1st column of I ).0Similarly,
1 for the other columns of B. For example, the j th column of B
0
B.C
0 bij 1 BBB .. CCC
satises A B
@ ..
.
CA = BBB 01 CCC (the j th column of I ). So in order to get the inverse of an n n
BB 0 CC
bnj BB . CC
@ .. A
0
matrix, we must solve n linear systems. However, the same steps of the Gauss-Jordan elimination
procedure are needed for all of these systems. So!we solve them all at once, using the matrix form.
Example: Find the inverse of A = ,34 ,23 .
We need to solve the following matrix equation
! !
3 ,2 1 0
,4 3 B = 0 1
We divide the rst row by 3 to introduce a 1 in the top left corner.
! !
1 , 32 B = 1 0
3
,4 3 0 1
Then we add four times the rst row to the second row to introduce a 0 in the rst column.
! !
1 , 23 B=
1
34 0
0 13 3 1
Multiply the second row by 3. ! !
1 , 23 B = 1
3 0
0 1 4 3
Add 23 the second row to the rst. (All this is classical Gauss-Jordan elimination.)
! !
1 0 B= 3 2
0 1 4 3
As IB = B , we get !
B = 34 23 :
It is important to note that, in addition to the two elementary row operations introduced earlier
in the context of the Gauss-Jordan elimination procedure, a third elementary row operation may
sometimes be needed here, namely permuting
0 two
1 rows.
0 1 0
Example: Find the inverse of A = B@ 2 1 0 CA.
0 0 1
0 1 0 1
0 1 0 1 0 0
B@ 2 1 0 CA B = B@ 0 1 0 CA
0 0 1 0 0 1
Because the top left entry of A is 0, we need to permute rows 1 and 2 rst.
0 1 0 1
2 1 0 0 1 0
B@ 0 1 0 CA B = B@ 1 0 0 CA
0 0 1 0 0 1
Now we divide the rst row by 2.
0 1 1 0 1 1
B@ 01 12 00 CA B = B@ 01 20 00 CA
0 0 1 0 0 1
Next we add , 21 the second row to the rst.
0 1 0 1 1 1
B@ 0 1 0 CA B = B@ , 21 20 00 CA
1 0 0
0 0 1 0 0 1
and we are done, since the matrix in front of B is the identity.
Exercise 14 Find the inverse of the matrix
0 1
1 4 2
A=B @ 0 1 0 CA
,1 ,4 2
.
Exercise 15 Find the inverse of the matrix
0 1
0 1 ,2
A=B @ 4 0 5 CA
3 0 4
.
1.5. DETERMINANTS 23
1.5 Determinants
To each square matrix, we associate a number, called its determinant, dened as follows:
If A = (a11), ! then det(A) = a11,
If A = aa11 aa12 , then det(A) = a11 a22 + a12a21 .
21 22
For a square matrix A of dimensions n n, the determinant can be obtained as follows. First,
dene A1j as the matrix of dimensions (n , 1) (n , 1) obtained from A by deleting row 1 and
column j . Then
det(A) = a11det(A11 ) , a12det(A12 ) + a13 det(A13 ) , : : :a1n det(A1n ):
Note that, in this formula, the signs alternate between + and ,.
0 1
a 11 a 12 a 13
For example, if A = B
@ a21 a22 a23 C A, then
a31 a32 a33
det(A) = a11(a22a33 , a23 a32) , a12 (a21a33 , a23a31 ) + a13(a21a32 , a22 a31):
Determinants have several interesting properties. For example, the following statements are
equivalent for a square matrix A:
det(A) = 0,
A has no inverse, i.e. A is singular,
the columns of A form a set of linearly dependent vectors,
the rows of A form a set of linearly dependent vectors.
For our purpose, however, determinants will be needed mainly in our discussion of classical opti-
mization, in conjunction with the material from the following section.
0 1
4 2 3
Exercise 16 Compute the determinant of A = B@ 1 0 1 CA.
0 1 2
1.6 Positive Denite Matrices

When we study functions of several variables (see Chapter 3!), we will need the following matrix
notions.
A square matrix A is positive denite if xT Ax > 0 for all nonzero column vectors x. It is
negative denite if xT Ax < 0 for all nonzero x. It is positive semidenite if xT Ax 0 and negative
semidenite if xT Ax 0 for all x. These denitions are hard to check directly and you might as
well forget them for all practical purposes.
More useful in practice are the following properties, which hold when the matrix A is symmetric
(that will be the case of interest to us), and which are easier to check.
The ith principal minor of A is the matrix Ai formed by the rst i rows and columns of A. So,
the rst principal !minor of A is the matrix A1 = (a11), the second principal minor is the matrix
A2 = aa11 aa12 , and so on.
21 22
The matrix A is positive denite if all its principal minors A1, A2; : : :An have strictly positive
determinants.
If these determinants are nonzero and alternate in signs, starting with det(A1) < 0, then the
matrix A is negative denite.
If the determinants are all nonnegative, then the matrix is positive semidenite,
If the determinant alternate in signs, starting with det(A1) 0, then the matrix is negative
semidenite.
!
To x ideas, consider a 2 2 symmetic matrix A = aa11 aa12 .
12 22
It is positive denite if:
(i) det(A1) = a11 > 0
(ii) det(A2) = a11a22 , a12a12 > 0
and negative denite if:
(i) det(A1) = a11 < 0
(ii) det(A2) = a11a22 , a12a12 > 0.
It is positive semidenite if:
(i) det(A1) = a11 0
(ii) det(A2) = a11a22 , a12a12 0
and negative semidenite if:
(i) det(A1) = a11 0
(ii) det(A2) = a11a22 , a12a12 0.
Exercise 17 Check whether the following matrices are positive denite, negative denite, positive
semidenite, negative semidenite or none of the above.
!
(a) A = 2 1
1 4
!
(b) A = ,24 ,48
!
(c) A = ,22 ,24
!
2 4
(d) A = 4 3 .
Chapter 2
Unconstrained Optimization:
Functions of One Variable
This chapter reviews basic results about functions of one variable, including the notions of deriva-
tive, extremum and convexity. It also describes a numerical method for nding x such that f (x) = 0,
known as \binary search".
2.1 Derivatives
It is important to know the following dierentiation formulas:
Function Derivative
f (x) + g (x) f 0 (x) + g0(x)
af (x) af 0 (x)
f (x)g (x) f 0 (x)g (x) + f (x)g 0(x)
f (x)=g (x) [g (x)f 0(x) , f (x)g 0(x)]=g 2(x)
f (g (x)) f 0 (g (x))g 0(x)
xa axa,1
ex ex
ln x (x > 0) where ln x = loge x 1=x
Exercise 18 Compute the derivatives of the following functions:

(a) x3 , 3x + 1
(b) ln(1 + x2 )
(c) ln(1 + x)2
(d) (3 , 2x)ex2
25
26 CHAPTER 2. UNCONSTRAINED OPTIMIZATION: FUNCTIONS OF ONE VARIABLE
For a function f of one variable x, recall that the derivative f 0 (x) is equal to the slope of a
tangent line at point x. So, if the function has a positive derivative at point x, then the function is
increasing, and if it has a negative derivative, it is decreasing. Since the function and its tangent
line are close around point x, the following formula can be used when is small.
f (x + ) f (x) + f 0 (x)
Example 2.1.1 Suppose that the demand for gasoline at price x (in dollars) is
ae,0:2x
gallons, where a is a constant. Suppose the price of gas rises by 10 cents. By how much does
demand fall?
Let f (x) = ae,0:2x denote the demand for gas at price x. The rate of change is given by the
derivative
f 0 (x) = ,0:2f (x):
Since = 0:10, we get
f (x + 0:10) f (x) + 0:10(,0:2f (x)) = 0:98f (x):
So demand drops by 2%. The factor relating change in demand to change in price is known as
\price elasticity of demand" in economics (You will learn more about this in 45-749 Managerial
Economics and in marketing courses such as 45-720 Marketing Management and 45-834 Pricing).
Here f 0 (x) = ,0:2f (x), so price elasticity of demand is -0.2.
Exercise 19 Suppose that, if x dollars are spent on advertising for a product during a given year,
f (x) = k(1 , e,cx ) customers will purchase the product (k > 0 and c > 0 are two constants).
(a) As x grows large, the number of customers purchasing the product approaches a limit. Find
this limit. Can you give an interpretation for the constant k?
(b) The sales response from a dollar of advertising is f (x + 1) , f (x). Using the formula for
small changes, show that the sales response from a dollar of advertising is proportional to the
number of potential customers who are not purchasing the product at present.
2.2 Maximum and Minimum

Let f be a function of one variable dened for all x in some domain D. A global maximum of f is
a point x0 in D such that f (x) f (x0) for all x in D.
For a constant > 0, the neighborhood N (x0) of a point x0 is the set of all points x such
that x0 , < x < x0 + . A point x0 is a local maximum of f if there exists > 0 such that
f (x) f (x0) for all x in N (x0 ) where f (x) is dened.
Similarly one can dene local and global minima. In Figure 2.1, the function f is dened for x
in domain [a; e]. Points b and d are local maxima and b is a global maximum, whereas a, c, and e
are local minima and e is a global minimum.
Finding extrema
Extrema, whether they are local or global, can occur in three places:
2.2. MAXIMUM AND MINIMUM 27
f(x)
a b c d e x
Figure 2.1: local maxima and minima
1. at the boundary of the domain,

2. at a point without a derivative, or
3. at a point x0 with f 0(x0 ) = 0.
The last case is particularly important. So we discuss it further. Let f be dierentiable in a
neighborhood of x0 . If x0 is a local extremum of f , then f 0 (x0) = 0. Conversely, if f 0 (x0) = 0,
three possibilities may arise: x0 is a local maximum, x0 is a local minimum, or neither!!! To decide
between these three possibilities, one may use the second derivative. Let f be twice dierentiable
in a neighborhood of x0 .
If f 0(x0) = 0 and f 00(x0) > 0 then x0 is a local minimum.
If f 0(x0) = 0 and f 00(x0) < 0 then x0 is a local maximum.
If f 0(x0) = 0 and f 00(x0) = 0 then x0 may or may not be a local extremum.
Figure 2.2 illustrates these three possibilities.
Example 2.2.1 Suppose an oil cartel can price gasoline as it wishes. How will it maximize revenue,
given that demand for gasoline at price x is
f (x) = ae,0:2x
The revenue at price x is

g (x) = xf (x):
We compute the derivative of g and set it to 0.
g0(x) = f (x) + x(,0:2f (x)) = (1 , 0:2x)f (x)
f(x)
Figure 2.2: f 00(x0) > 0, f 00 (x0) < 0 and one of the possibilities with f 00(x0 ) = 0
Since f (x) > 0 for all x, setting g 0(x) = 0 implies 1 , 0:2x = 0. So x = 5. This is the only possible
local optimum. To determine whether it is a maximum or a minimum, we compute the second
derivative.
g00(x) = 2(,0:2f (x)) + x(,0:2)2f (x) = ,0:2(2 , 0:2x)f (x):
Putting in x = 5 shows g 00(x) < 0, so this is a maximum: the oil cartel maximizes its revenue by
pricing gasoline at $5 per gallon.
Example 2.2.2 Economic Order Quantity We want to keep paper towels on hand at all times.
The cost of stockpiling towels is $ h per case per week, including the cost of space, the cost of tying
up capital in inventory, etc. The time and paperwork involved in ordering towels costs $ K per
order (this is in addition to the cost of the towels themselves, which is not relevant here because
we are supposing that there are no price discount for large orders.) Towels are used at a rate of d
cases per week. What quantity Q should we order at a time, to minimize cost?
We must rst calculate the holding cost (cost of stockpiling). If each order period begins with
Q cases and ends with zero cases, and if usage is more or less constant, then the average stock level
is Q=2. This means that the average holding cost is hQ=2.
Since each order of Q cases lasts Q=d weeks, the average ordering cost per week is Q=d K = Kd :
Q
Thus the average total cost per week is
hQ + Kd :
2 Q
To nd Q that minimizes cost, we set to zero the derivative with respect to Q.
h , Kd = 0:
2 Q2
This implies that the optimal order quantity is
s
Q = 2Kd
h :
This is the classical economic order quantity (EOQ) model for inventory management. (You will
learn when the EOQ model is appropriate and when to use other inventory models in 45-765).
Exercise 20 Find the local maxima and minima, if any, of the following functions.
(a) x3 , 3x + 1
(b) x3 , 3x2 + 3x , 1
(c) (3 , 2x)ex2
Exercise 21 We want to design a cone-shaped paper drinking cup of volume Vpthat requires a
minimum of paper to make. That is, we want to minimize its surface area r r2 + h2 , where
r is the radius of the open end and h the height. Since V = 3 r2h, we can solve for h to get
h = 3V=(r2). Thus the area in terms of r is
2 V 2 !1=2 2 4 V 2 !1=2
A(r) = r r + 9 2r4 = r + 9 r2 :
a) Compute the derivative A0(r).

b) Find the radius r at which A(r) is minimized (the solution you get is in fact a minimum).
c) What is the ratio h=r when r has the optimal value?
Exercise 22 An air courier company operates a single depot in a large city and wants to decen-
tralize by establishing several depots, each covering the same delivery radius r. The total area to
be covered is A, and the area covered by each depot is r2. Thus if n depots are installed, we must
have A = nr2 , so that n = A=(r2). Expenses and revenues break down as follows.
Trucking expenses. The number of trucks required at each depot is r3, so that the total
number of trucks required is nr3 = (A=r2)r3 = Ar. Each truck costs $K per day, so
that total expense for trucks is $KAr.
Overhead. Each depot incurs an overhead (xed) expense of $F , for a total overhead of
nF = FA=(r2).
Revenue from customers. Since the total area is xed, the total revenue from customers is
independent of the delivery radius.
Since the total revenue is xed, we can simply minimize costs:
min KAr + FA
r2 :
Your problem. Find the delivery radius r that minimizes cost. Your solution should indicate
that the optimal delivery radius is proportional to the cube root (the 1/3 power) of the xed cost.
Thus if the xed cost increases eightfold (and other costs were constant), the optimal delivery
radius would only double. This is fortunate, since it means that the rent and other xed costs must
rise substantially (relative to other costs) before they justify relocating depots.
2.3 Binary Search

Binary search is a very simple idea for solving numerically f (x) = 0.
It requires:
(i) the function f to be continuous,
(ii) values (a; b) such that f (a) > 0 and f (b) < 0.
Then a desired value x0 such that f (x0) = 0 can be found between the values a and b.
The method consists in computing f (c) for c = a+2 b .
If f (c) > 0, the procedure is repeated with the values (a; b) replaced by the values (c; b).
If f (c) < 0, the procedure is repeated with the values (a; b) replaced by (a; c).
If f (c) = 0, the procedure terminates as c is the value we are looking for.
This method is easy to program and converges rapidly (20 iterations will give you x0 with 5
signicant gures).
Example 2.3.1 Binary search can be used to compute the internal rate of return of an investment.
The internal rate of return of an investment is the interest rate r that makes the present value of
the cash ows from the investment equal to the cost of the investment. Mathematically, r is the
interest rate that satises the equation
F1 + F2 + F3 + : : : + FN , C = 0
1 + r (1 + r)2 (1 + r)3 (1 + r)N
where
Ft = cash ow in year t
N = number of years
C = cost of the investment
For a bond, the internal rate of return r is known as the yield to maturity or simply yield. Let
us consider the case of a noncallable bond, i.e. a bond that the issuer cannot retire prior to its
stated maturity date. For such a bond, the cash ows consist of periodic interest payments to the
maturity date and the par value paid at maturity. As an example, consider a 4-year noncallable
bond with a 10% coupon rate paid annually and a par value of $1000. Such a bond has the following
cash ows:
t Years from now Ft
1 $ 100
2 100
3 100
4 1100
Suppose this bond is now selling for $900. Compute the yield of this bond.
2.3. BINARY SEARCH 31
The yield r of the bond is given by the equation

100 100 100 1100
1 + r + (1 + r)2 + (1 + r)3 + (1 + r)4 , 900 = 0
Let us denote by f (r) the left-hand-side of this equation. We nd r such that f (r) = 0 using binary
search.
We start by nding values (a; b) such that f (a) > 0 and f (b) < 0. In this case, we expect r to
be between 0 and 1. Since f (0) = 500 and f (1) = ,743:75, we have our starting values.
Next, we let c = 0:5 (the midpoint) and calculate f (c). Since f (0:5) = ,541:975, we replace
our range with a = 0 and b = 0:5 and repeat. When we continue, we get the following table of
values:
Iter. a c b f (a) f (c) f (b)
1 0 0.5 1 500 -541.975 -743.75
2 0 0.25 0.5 500 -254.24 -541.975
3 0 0.125 0.25 500 24.85902 -254.24
4 0.125 0.1875 0.25 24.85902 -131.989 -254.24
5 0.125 0.15625 0.1875 24.85902 -58.5833 -131.989
6 0.125 0.140625 0.15625 24.85902 -18.2181 -58.5833
7 0.125 0.132813 0.140625 24.85902 2.967767 -18.2181
8 0.132813 0.136719 0.140625 2.967767 -7.71156 -18.2181
9 0.132813 0.134766 0.136719 2.967767 -2.39372 -7.71156
10 0.132813 0.133789 0.134766 2.967767 0.281543 -2.39372
11 0.133789 0.134277 0.134766 0.281543 -1.05745 -2.39372
12 0.133789 0.134033 0.134277 0.281543 -0.3883 -1.05745
So the yield of the bond is r = 13.4%.
Of course, this routine sort of calculation is perfectly set up for calculation by computer. In
particular, we can use Microsoft's Excel (or any spreadsheet program) to do these calculations. On
the course home page is a spreadsheet that implements binary search for this problem. We will
show this spreadsheet in class. Cells B3:G3 down to B20:G20 contain the body of the table. You
can generate the same table by following these steps:
In cell B3, type 0 [comment: this is initial value of a.]
In cell D3, type 1 [comment: this is the initial value of b.]
In cell C3, type =(B3+D3)/2 [comment: this is the middle point c.]
In cell E3, type =100/(1+B3) +100/(1+B3)^2 +100/(1+B3)^3 +1100/(1+B3)^4 -900
[comment: this is f (a).]
Use the Copy Command to copy cell E3 into cells F3 and G3 [comment: cell F3 contains
the same function as cell E3, except that B3 is replaced by C3. So cell F3 contains f (c). Similarly,
G3 contains f (b).]
In cell B4, type =IF(F3> 0,C3,B3) [comment: If F3> 0, cell B4 contains C3 and otherwise,
cell B4 contains B3.]
In cell D4, type =IF(F3> 0,D3,C3)
Use the Copy Command to copy cell B4 down column B
Similarly, copy C3 down column C, D4 down column D, E3 down column E, F3 down column
F, and G3 down column G.
Now you should have the full table! Read down column C to nd the value of r that makes
f (r) = 0.
Excel also has an addin called Solver that can search for any particular value for a function.
We will also review this system in class (we will use Solver extensively in this course).
Exercise 23 (a) Show the cash ows for the three bonds below, each of which pays interest
annually:
Bond Coupon Rate % Par Value Years to Maturity Price
X 14 $ 1000 5 $ 900
Y 16 2000 7 1900
Z 17 1000 4 950
(b) Calculate the yield to maturity for the three bonds.
Golden Section Search
Golden section search is similar in spirit to binary search. It is used to compute the maximum
of a function f (x) dened on an interval [a; b], when the method presented in Section 2.2 is not
applicable for one reason or another.
It assumes that
(i) f is continuous
(ii) f has a unique local maximum in the interval [a; b].
The golden search method consists in computing f (c) and f(d) for a < d < c < b.
If f (c) > f (d), the procedure is repeated with the interval (a; b) replaced by (d; b).
If f (c) < f (d), the procedure is repeated with the interval (a; b) replaced by (a; c).
Note: The name \golden section" comes from a certain choice of pc and d that yields fast conver-
gence, namely c = a + r(b , a) and d = b + r(a , b), where r = 52,1 = :618034 : : :. This is the
golden ratio, already known to the ancient greeks.
Example 2.3.2 Find the maximum of the function x5 , 10x2 + 2x in the interval [0; 1].
In this case, we begin with a = 0 and b = 1. Using golden section search, that gives d = 0:382
and c = 0:618. The function values are f (a) = 0, f (d) = ,0:687, f (c) = ,2:493, and f (b) = ,7.
Since f (c) < f (d), our new range is a = 0, b = :618. Recalculating from the new range gives
d = :236, c = :382 (note that our current c was our previous d: it is this reuse of calculated values
that gives golden section search its speed). We repeat this process to get the following table:
2.4. CONVEXITY 33
Iter. a d c b f (a) f (d) f (c) f (b)
1 0 0.382 0.618 1 0 -0.6869 -2.4934 -7
2 0 0.2361 0.382 0.618 0 -0.0844 -0.6869 -2.4934
3 0 0.1459 0.2361 0.382 0 0.079 -0.0844 -0.6869
4 0 0.0902 0.1459 0.2361 0 0.099 0.079 -0.0844
5 0 0.0557 0.0902 0.1459 0 0.0804 0.099 0.079
6 0.0557 0.0902 0.1115 0.1459 0.0804 0.099 0.0987 0.079
7 0.0557 0.077 0.0902 0.1115 0.0804 0.0947 0.099 0.0987
8 0.077 0.0902 0.0983 0.1115 0.0947 0.099 0.1 0.0987
9 0.0902 0.0983 0.1033 0.1115 0.099 0.1 0.0999 0.0987
10 0.0902 0.0952 0.0983 0.1033 0.099 0.0998 0.1 0.0999
11 0.0952 0.0983 0.1002 0.1033 0.0998 0.1 0.1 0.0999
12 0.0983 0.1002 0.1014 0.1033 0.1 0.1 0.1 0.0999
13 0.0983 0.0995 0.1002 0.1014 0.1 0.1 0.1 0.1
14 0.0995 0.1002 0.1007 0.1014 0.1 0.1 0.1 0.1
15 0.0995 0.0999 0.1002 0.1007 0.1 0.1 0.1 0.1
16 0.0995 0.0998 0.0999 0.1002 0.1 0.1 0.1 0.1
17 0.0998 0.0999 0.1 0.1002 0.1 0.1 0.1 0.1
18 0.0999 0.1 0.1001 0.1002 0.1 0.1 0.1 0.1
19 0.0999 0.1 0.1 0.1001 0.1 0.1 0.1 0.1
20 0.0999 0.1 0.1 0.1 0.1 0.1 0.1 0.1
21 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1
Again we can use Excel to implement golden section search or use Solver to maximize this
function directly.
2.4 Convexity
Finding global extrema and checking that we have actually found one is harder than nding local
extrema, in general. There is one nice case: that of convex and concave functions. A convex
function is one where the line segment connecting two points (x; f (x)) and (y; f (y )) lies above the
function. Mathematically, a function f is convex if, for all x; y and all 0 < < 1,
f (x + (1 , )y ) f (x) + (1 , )f (y ):
See Figure 2.3. The function f is concave if ,f is convex.
There is an easy way to check for convexity when f is twice dierentiable: the function f is
convex on some domain [a; b] if (and only if) f 00(x) 0 for all x in the domain. Similarly, f is
concave on some domain [a; b] if (and only if) f 00(x) 0 for all x in the domain.
If f (x) is convex, then any local minimum is also a global minimum.
If f (x) is concave, then any local maximum is also a global maximum.
Exercise 24 Find whether the following functions are concave, convex or neither.
(a) x4 , 4x3 + 6x2 + 3x + 1
(b) ,ex2
Exercise 25 Find a local extremum of the function f (x) = xe,x . Indicate whether it is a local
maximum, a local minimum or neither. Is it a global optimum on the domain [-2,2]?
f(x)
convex function
x
x y
f(x)
concave function
x
x y
Figure 2.3: Convex function and concave function
2.4. CONVEXITY 35
Exercise 26 Consider the linear demand function D(P ) = K , aP , where P ( K=a) is the price
and K and a are positive constants. Thus demand for a free good is K , and demand drops linearly
until the price reaches K=a, at which point demand is zero.
a) Find the price P that maximizes revenue PD(P ) by setting the derivative of revenue equal to
zero.
b) Use the second derivative to show that you found a local maximum.
c) Use the second derivative to show that the revenue function is concave. It follows that your
local maximum is a global maximum.
Exercise 27 Oil is to be shipped in barrels having an undetermined height h and radius r. Each
barrel must have a xed volume V . We want to pick r and h so as to minimize the surface area
of each barrel, given by A = 2r2 + 2rh, and thereby minimize the cost of making the barrels.
Since V = r2h, we can solve for h in terms of r to get h = V=r2, and we can substitute this in
the formula for A to get a formula for area solely in terms of r, namely,
A(r) = 2r2 + 2V=r:
a) Compute A0(r), A00(r) and show that the function A is convex on the set of positive reals.
b) Find the radius at which A(r) has its global minimum. Explain how you know the minimum is
global.
c) What is the ratio h=r when r has the optimal value you just derived?
Chapter 3
Unconstrained Optimization:
Functions of Several Variables
Many of the concepts for functions of one variable can be extended to functions of several variables.
For example, the gradient extends the notion of derivative. In this chapter, we review the notion
of gradient, the formula for small changes, how to nd extrema and the notion of convexity.
3.1 Gradient
Given a function f of n variables x1; x2; : : :; xn , we dene the partial derivative relative to variable
@f , to be the derivative of f with respect to x treating all variables except x as
xi , written as @x i i i
constant.
Example 3.1.1 Compute the partial derivatives of f (x1; x2) = (x1 , 2)2 + 2(x2 , 1)2.
@f (x ; x ) = 2(x , 2); @f
The answer is: @x 1 1 2 1 @x2 (x1; x2) = 4(x2 , 1).
Let x denote the vector (x1; x2; : : :; xn ). With this notation, f (x) = f (x1; x20 @f (x) =
; : : :; xn), @x
@f (x) 1
i
BB @x@f1 (x) CC
@f B @x2. CC. The
@xi (x1; x2; : : :; xn), etc. The gradient of f at x, written rf (x), is the vector B B@ .. CA
@f
@xn (x)
gradient vector rf (x) gives the direction of steepest ascent of the function f at point x. The
gradient acts like the derivative in that small changes around a given point x can be estimated
using the gradient.
f (x + ) f (x) + rf (x)
where = (1; : : :; n ) denotes the vector of changes.
Example 3.1.2 If f (x1; x2) = x21 , 3x1x2 + x22, then f (1; 1) = ,1. What about f (1:01; 1:01)?
@f (x ; x ) = 2x , 3x and @f (x ; x ) =
In this case, x = (1; 1) and = (0:01; 0:01). Since @x 1 2 1 2 @x2 1 2
,3x1 + 2x2, we get
1
37
38CHAPTER 3. UNCONSTRAINED OPTIMIZATION: FUNCTIONS OF SEVERAL VARIABLES
!
rf (1; 1) = ,,11
. !
, 1
So f (1:01; 1:01) = f ((1; 1)+(0:01; 0:01)) f (1; 1)+(0:01; 0:01)rf (1; 1) = ,1+(0:01; 0:01) ,1 =
,1:02.
Example 3.1.3 Suppose that we want to put away a shing pole in a closet having dimensions 3
by 5 by 1 feet. If the ends of the pole are placed at opposite corners, there is room for a pole of
length,
f (x1; x2; x3) = f (3; 5; 1) = px21 + x22 + x23 = 5:9ft:
It turns out that the actual dimensions of the closet are 3 + 1 , 5 + 2 and 1 + 3 feet, where
1 , 2 and 3 are small correction terms. What is the change in pole length, taking into account
these corrections?
By the formula for small changes, the change in pole length is
f (3 + 1; 5 + 2 ; 1 + 3 ) , f (3; 5; 1) (1; 2; 3)rf (3; 5; 1):
So, we need to compute the partial derivatives of f . For i = 1; 2; 3
@ f (x ; x ; x ) = xi :
@xi 1 2 3 px1 + x22 + x23
2
Now we get
0 1
0:51
(1; 2; 3)rf (3; 5; 1) = (1; 2; 3) B
@ 0:85 CA = 0:511 + 0:852 + 0:173:
0:17
Exercise 28 Consider the function f (x1; x2) = x1 ln x2.
(a) Compute the gradient of f .
(b) Give the value of the function f and give its gradient at the point (3; 1).
(c) Use the formula for small changes to obtain an approximate value of the function at the point
(2:99; 1:05).
Exercise 29 Consider a conical drinking cup with height h and radius r at the open end. The
volume of the cup is V (r; h) = 3 r2h.
a) Suppose the cone is now 5 cm high with radius 2 cm. Compute its volume.
b) Compute the partial derivatives @V=@r and @V=@h at the current height and radius.
c) By about what fraction (i.e., percentage) would the volume change if the cone were lengthened
10%? (Use the partial derivatives.)
d) If the radius were increased 5%?
e) If both were done simultaneously?
Hessian matrix
Second partials @x@i @xf j (x) are obtained from f (x) by taking the derivative relative to xi (this
2
yields the rst partial @x@f (x) ) and then by taking the derivative of @f (x) relative to x . So we
i @xi j
@ 2f @ 2f
can compute @x1@x1 (x), @x1 @x2 (x) and so on. These values are arranged into the Hessian matrix
0 @2f @ 2f @2f 1
BB @x12@x1 (x) @x12@x2 (x) @x12@xn (x) CC
@f @f @f
H (x) = BBB @x2 @x1 (x) @x2 @x2 (x) @x2 @xn (x) CC
.. .. ... .. CA
@ . . .
@2f @ 2f @2f
@xn @x1 (x) @xn @x2 (x) @xn@xn (x)
@2f @2f
The Hessian matrix is a symmetric matrix, that is @xi @xj (x) = @xj @xi (x).
Example 3.1.1 (continued): Find the Hessian matrix of f (x1; x2) = (x1 , 2)2 + 2(x2 , 1)2:
!
The answer is H (x) = 2 0 :
0 4
3.2 Maximum and Minimum

Optima can occur in three places:
1. at the boundary of the domain,
2. at a nondierentiable point, or
3. at a point x with rf (x ) = 0.
We will identify the rst type of point with Kuhn{Tucker conditions (see next chapter). The
second type is found only by ad hoc methods. The third type of point can be found by solving the
gradient equations.
In the remainder of this chapter, we discuss the important case where rf (x ) = 0. To identify
if a point x with zero gradient is a local maximum or local minimum, check the Hessian.
If H (x) is positive denite then x is a local minimum.
If H (x) is negative denite, then x is a local maximum.
Remember (Section 1.6) that these properties can be checked by computing the determinants of
the principal minors.
Example 3.2.1 Find the local extrema of f (x1; x2) = x31 + x32 , 3x1x2.
This function is everywhere dierentiable, so extrema can only occur at points x such that
rf (x) = 0.
3x 21 , 3x2 !
rf (x) = 3x2 , 3x1
2
This equals 0 i (x1 ; x2) = (0; 0) or (1; 1). The Hessian is
!
H (x) = 6x1 ,3
,3 6x2
So,
!
H (0; 0) = ,03 ,03
Let H1 denote the rst principal minor of H (0; 0) and let H2 denote its second principal minor
(see Section 1.6). Then det(H1) = 0 and det(H2) = ,9. Therefore H (0; 0) is neither positive nor
negative denite.
!
H (1; 1) = ,63 ,63
Its rst principal minor has det(H1) = 6 > 0 and its second principal minor has det(H2) =
36 , 9 = 25 > 0. Therefore H (1; 1) is positive denite, which implies that (1; 1) is a local minimum.
Example 3.2.2 Jane and Jim invested $20,000 in the design and development of a new product.
They can manufacture it for $2 per unit. For the next step, they hired marketing consultants XYZ.
In a nutshell, XYZ's conclusions are the following: if Jane and Jim spend $a on advertizing and
sell the product at price p (per unit), they will sell
p
2; 000 + 4 a , 20p units.
Using this gure, express the prot that Jane and Jim will make as a function of a and p. What
price and level of advertising will maximize their prots?
p
The revenue from sales is (2; 000 + 4pa , 20p)p.
The production costs are (2; 000 + 4 a , 20p)2, the development cost is $20,000 and the cost
of advertizing is a.
Therefore, Jane and Jim's prot is
p
f (p; a) = (2; 000 + 4 a , 20p)(p , 2) , a , 20; 000
To nd the maximum prot, we compute the partial derivatives of f and set them to 0:
@f (p; a) = 2; 040 + 4pa , 40p = 0:
@p
@f (p; a) = 2(p , 2)=pa , 1 = 0:
@a
Solving this system of two equations yields
p = 63:25 a = 15; 006:25
We verify that this is a maximum by computing the Hessian.
3.3. GLOBAL OPTIMA 41
pa !
, 40 2 =
H (x) = 2=pa ,(p , 2)=apa
p
det(H1) = ,40 < 0 and det(H2 ) = 40(p , 2)=a a , 4=a > 0 at the point p = 63:25, a = 15; 006:25.
So, indeed, this solution maximizes prot.
Example 3.2.3 Find the local extrema of f (x1; x2; x3) = x21 + (x1 + x2)2 + (x1 + x3)2.
@f
@x1 (x) = 2x1 + 2(x1 + x2 ) + 2(x1 + x3 )
@f
@x2 (x) = 2(x1 + x2 )
@f
@x3 (x) = 2(x1 + x3 )
Setting these partial derivatives to 0 yields the unique solution x1 = x2 = x3 = 0. The Hessian
matrix is
0 1
6 2 2
H (0; 0; 0) = B
@ 2 2 0 CA
2 0 2
The determinants of the principal minors are det(H1) = 6 > 0, det(H2) = 12 , 4 = 8 > 0 and
det(H3) = 24 , 8 , 8 = 8 > 0. So H (0; 0; 0) is positive denite and the solution x1 = x2 = x3 = 0
is a minimum.
Exercise 30 Find maxima or minima of the following functions when possible.
(a) f (x1; x2; x3) = ,x21 , 3x22 , 10x23 + 4x1 + 24x2 + 20x3
(b) f (x1; x2; x3) = x1 x2 + x2 x3 + x3 x1 , 2x1 , 2x2 , 2x3
Exercise 31 Consider the function of three variables given by
f (x1; x2; x3) = x21 , x1 , x1x2 + x22 , x2 + x43 , 4x3 :
(a) Compute the gradient rf (x1; x2; x3).
(b) Compute the Hessian matrix H (x1; x2; x3).
(c) Use the gradient to nd a local extremum of f .
Hint: if x33 = 1, then x3 = 1.
(d) Compute the three principal minors of the Hessian matrix and use them to identify this
extremum as a local minimum or a local maximum.
3.3 Global Optima

Finding global maxima and minima is harder. There is one case that is of interest.
We say that a domain is convex if every line drawn between two points in the domain lies within
the domain.
We say that a function f is convex if the line connecting any two points lies above the function.
That is, for all x; y in the domain and 0 < < 1, we have f (x + (1 , )y ) f (x) + (1 , )f (y ),
as before (see Chapter 2).
If a function is convex on a convex domain, then any local minimum is a global minimum.
If a function is concave on a convex domain, then any local maximum is a global maximum.
To check that a function is convex on a domain, check that its Hessian matrix H (x) is positive
semidenite for every point x in the domain. To check that a function is concave, check that its
Hessian is negative semidenite for every point in the domain.
Example 3.3.1 Show that the function f (x1; x2; x3) = x41 + (x1 + x2)2 + (x1 + x3)2 is convex over
<3.
@f 3
@x1 (x) = 4x1 + 2(x1 + x2 ) + 2(x1 + x3 )
@f
@x2 (x) = 2(x1 + x2 )
@f
@x3 (x) = 2(x1 + x3 )
0 2 1
12x1 + 4 2 2
H (x1; x2; x3) = B
@ 2 2 0CA
2 0 2
The determinants of the principal minors are det(H1) = 12x21 + 4 > 0, det(H2 ) = 12x21 0 and
det(H3) = 48x21 0. So H (x1; x2; x3) is positive semidenite for all (x1; x2; x3) in <3 . This implies
that f is convex over <3.
Exercise 32 For each of the following, determine whether the function is convex, concave, or
neither over <2.
(a) f (x) = x1 x2 , x21 , x22
(b) f (x) = 10x1 + 20x2
(c) f (x) = x41 + x1x2
(d) f (x) = ,x21 , x1 x2 , 2x22
Exercise 33 Let the following function be dened for all points (x; y) in the plane.
f (x; y) = 2xy , x4 , x2 , y2:
(a) Write the gradient of the function f .
(b) Write the Hessian matrix of f .
(c) Is the function f convex, concave or neither?
(d) Use the gradient to nd a local extremum of f .
(e) Identify this extremum as a minimum, a maximum or neither.
Chapter 4
Constrained Optimization
4.1 Equality Constraints (Lagrangians)
Suppose we have a problem:
Maximize 5 , (x1 , 2)2 , 2(x2 , 1)2
subject to
x1 + 4x2 = 3
If we ignore the constraint, we get the solution x1 = 2; x2 = 1, which is too large for the
constraint. Let us penalize ourselves for making the constraint too big. We end up with a
function
L(x1; x2; ) = 5 , (x1 , 2)2 , 2(x2 , 1)2 + (3 , x1 , 4x2)
This function is called the Lagrangian of the problem. The main idea is to adjust so that we
use exactly the right amount of the resource.
= 0 leads to (2; 1).
= 1 leads to (3=2; 0) which uses too little of the resource.
= 2=3 gives (5=3; 1=3) and the constraint is satised exactly.
We now explore this idea more formally. Given a nonlinear program (P) with equality con-
straints:
Minimize (or maximize) f (x)
subject to
g1(x) = b1
g2(x) = b2
..
.
gm (x) = bm
a solution can be found using the Lagrangian:
X
m
L(x; ) = f (x) + i (bi , gi (x))
i=1
P
(Note: this can also be written f (x) , mi=1 i (gi(x) , bi )).
Each i gives the price associated with constraint i.
43
44 CHAPTER 4. CONSTRAINED OPTIMIZATION
The reason L is of interest is the following:
Assume x = (x1 ; x2; : : :; xn) maximizes or minimizes f (x) subject to the
constraints gi(x) = bi , for i = 1; 2; : : :; m. Then either
(i) the vectors rg1(x); rg2(x); : : :; rgm(x) are linearly dependent,
or
(ii) there exists a vector = (1; 2; : : :; m) such that rL(x; ) = 0.
I.e.
@L (x ; ) = @L (x; ) = = @L (x; ) = 0
@x1 @x2 @xn
and @ @L (x; ) = = @L (x ; ) = 0
1 @m
Of course, Case (i) above cannot occur when there is only one constraint. The following example
shows how it might occur.
Example 4.1.1
Minimize x1 + x2 + x23
subject to
x1 = 1
x21 + x22 = 1:
It is easy to check directly that the minimum is acheived at (x1; x2; x3) = (1; 0; 0). The associ-
ated Lagrangian is
L(x1 ; x2; x3; 1; 2) = x1 + x2 + x23 + 1(1 , x1 ) + 2(1 , x21 , x22):
Observe that
@L (1; 0; 0; ; ) = 1 for all ; ;
@x2 1 2 1 2
and consequently @x@L does not vanish at the optimal solution. The reason for this is the following.
Let g1(x1; x2; x3) = x1 and g2 (x1; x2; x3) = x21 + x22 denote the left hand sides of the constraints.
2
Then rg1(1; 0; 0) = (1; 0; 0) and rg2(1; 0; 0) = (2; 0; 0) are linearly dependent vectors. So Case (i)
occurs here!
Nevertheless, Case (i) will not concern us in this course. When solving optimization problems
with equality constraints, we will only look for solutions x that satisfy Case (ii).
Note that the equation
@L
@i (x ; ) = 0
is nothing more than
bi , gi (x) = 0
or gi (x ) = bi:
In other words, taking the partials with respect to does nothing more than return the original
constraints.
4.1. EQUALITY CONSTRAINTS (LAGRANGIANS) 45
Once we have found candidate solutions x , it is not always easy to gure out whether they
correspond to a minimum, a maximum or neither. The following situation is one when we can
conclude. If f (x) is concave and all of the gi (x) are linear, then any feasible x with a corresponding
making rL(x; ) = 0 maximizes f (x) subject to the constraints. Similarly, if f (x) is convex
and each gi(x) is linear, then any x with a making rL(x ; ) = 0 minimizes f (x) subject to
the constraints.
Example 4.1.2
Minimize 2x21 + x22
subject to
x1 + x2 = 1
L(x1; x2; ) = 2x21 + x22 + 1(1 , x1 , x2)
@L
@x1 (x1 ; x2; ) = 4x1 , 1 = 0
@L (x ; x; ) = 2x , = 0
@x2 1 2 2 1
@L (x; x; ) = 1 , x , x = 0
@ 1 2 1 2
Now, the rst two equations imply 2x1 = x2 . Substituting into the nal equation gives the
solution x1 = 1=3, x2 = 2=3 and = 4=3, with function value 2/3. !
Since f (x1 ; x2) is convex (its Hessian matrix H (x1; x2) = 40 02 is positive denite) and
g(x1; x2) = x1 + x2 is a linear function, the above solution minimizes f (x1; x2) subject to the
constraint.
4.1.1 Geometric Interpretation

There is a geometric interpretation of the conditions an optimal solution must satisfy. If we graph
Example 4.1.2, we get a picture like that in Figure 4.1.
Now, examine the gradients of f and g at the optimum point. They must point in the same
direction, though they may have dierent lengths. This implies:
rf (x) = rg(x)
which, along with the feasibility of x, is exactly the condition rL(x; ) = 0 of Case (ii).
4.1.2 Economic Interpretation

The values i have an important economic interpretation: If the right hand side bi of Constraint i
is increased by , then the optimum objective value increases by approximately i .
In particular, consider the problem
Maximize p(x)
subject to
g (x) = b,
1 f = lambda g
x*
2/3
f(x)=2/3
g(x)=1
0 1/3 1
Figure 4.1: Geometric interpretation
where p(x) is a prot to maximize and b is a limited amount of resource. Then, the optimum
Lagrange multiplier is the marginal value of the resource. Equivalently, if b were increased by ,
prot would increase by . This is an important result to remember. It will be used repeatedly
in your Managerial Economics course.
Similarly, if
Minimize c(x)
subject to
d(x) = b,
represents the minimum cost c(x) of meeting some demand b, the optimum Lagrange multiplier
is the marginal cost of meeting the demand.
In Example 4.1.2
Minimize 2x21 + x22
subject to
x1 + x2 = 1,
if we change the right hand side from 1 to 1:05 (i.e. = 0:05), then the optimum objective function
value goes from 23 to roughly
2 + 4 (0:05) = 2:2 :
3 3 3
4.1. EQUALITY CONSTRAINTS (LAGRANGIANS) 47
If instead the right hand side became 0:98, our estimate of the optimum objective function value
would be
2 + 4 (,0:02) = 1:92
3 3 3
Example 4.1.3 Suppose we have a renery that must ship nished goods to some storage tanks.
Suppose further that there are two pipelines, A and B, to do the shipping. The cost of shipping x
units on A is ax2 ; the cost of shipping y units on B is by 2 , where a > 0 and b > 0 are given. How
can we ship Q units while minimizing cost? What happens to the cost if Q increases by r%?
Minimize ax2 + by 2
Subject to
x+y = Q
L(x; y; ) = ax2 + by2 + (Q , x , y )

@L (x; y; ) = 2ax , = 0
@x
@L (x; y ; ) = 2by , = 0
@y
@L (x; y ; ) = Q , x , y = 0
@
The rst two constraints give x = ab y , which leads to
x = abQ+b
; y = aaQ +b
; = 2aabQ
! +b
and cost of abQ 2a 0 is positive denite since a > 0 and
a+b . The Hessian matrix H (x1; x2) =
2
0 2b
b > 0. So this solution minimizes cost, given a; b; Q.
If Q increases by r%, then2 the RHS of the constraint increases by = rQ and the minimum
cost increases by = 2abrQ
a+b . That is, the minimum cost increases by 2r%.
Example 4.1.4 How should one divide his/her savings between three mutual funds with expected
returns 10%, 10% and 15% repectively, so as to minimize risk while achieving an expected return of
12%. We measure risk as the variance of the return on the investment (you will learn more about
measuring risk in 45-733): when a fraction x of the savings is invested in Fund 1, y in Fund 2 and
z in Fund 3, where x + y + z = 1, the variance of the return has been calculated to be
400x2 + 800y 2 + 200xy + 1600z 2 + 400yz:
So your problem is
min 400x2 + 800y 2 + 200xy + 1600z 2 + 400yz
s:t: x + y + 1:5z = 1:2
x + y + z = 1
Using the Lagrangian method, the following optimal solution was obtained
x = 0:5 y = 0:1 z = 0:4 1 = 1800 2 = ,1380
where 1 is the Lagrange multiplier associated with the rst constraint and 2 with the second
constraint. The corresponding objective function value (i.e. the variance on the return) is 390.
If an expected return of 12.5% was desired (instead of 12%), what would be (approximately) the
correcponding variance of the return?
Since = 0:05, the variance would increase by
1 = 0:05 1800 = 90:
So the answer is 390+90=480.
Exercise 34 Record'm Records needs to produce 100 gold records at one or more of its three
studios. The cost of producing x records at studio 1 is 10x; the cost of producing y records at
studio 2 is 2y 2 ; the cost of producing z records at studio 3 is z 2 + 8z .
(a) Formulate the nonlinear program of producing the 100 records at minimum cost.
(b) What is the Lagrangian associated with your formulation in (a)?
(c) Solve this Lagrangian. What is the optimal production plan?
(d) What is the marginal cost of producing one extra gold record?
(e) Union regulations require that exactly 60 hours of work be done at studios 2 and 3 combined.
Each gold record requires 4 hours at studio 2 and 2 hours at studio 3. Formulate the problem of
nding the optimal production plan, give the Lagrangian, and give the set of equations that must
be solved to nd the optimal production plan. It is not necessary to actually solve the equations.
Exercise 35 (a) Solve the problem
max 2x + y
subject to 4x2 + y 2 = 8
(b) Estimate the change in the optimal objective function value when the right hand side increases
by 5%, i.e. the right hand side increases from 8 to 8.4.
Exercise 36
(a) Solve the following constrained optimization problem using the method of Lagrange multipliers.
max ln x + 2 ln y + 3 ln z
subject to x + y + z = 60
(b) Estimate the change in the optimal objective function value if the right hand side of the con-
straint increases from 60 to 65.
4.2. EQUALITY AND INEQUALITY CONSTRAINTS 49
4.2 Equality and Inequality Constraints

How do we handle both equality and inequality constraints in (P)? Let (P) be:
Maximize f (x)
Subject to
g1(x) = b1
..
.
gm (x) = bm
h1(x) d1
..
.
hp (x) dp
If you have a program with constraints, convert it into by multiplying by ,1. Also convert
a minimization to a maximization.
The Lagrangian is
X
m X
p
L(x; ; ) = f (x) + i(bi , gi (x)) + j (dj , hj (x))
i=1 j =1
The fundamental result is the following:
Assume x = (x1; x2; : : :; xn) maximizes f (x) subject to the constraints
gi (x) = bi, for i = 1; 2; : : :; m and hj (x) dj , for j = 1; 2; : : :; p. Then
either
(i) the vectors rg1(x); : : :; rgm(x), rh1 (x); : : :; rhp(x) are lin-
early dependent, or
(ii) there exists vectors = (1; : : :; m) and = (1 ; : : :; p ) such
that
X m Xp
rf (x ) , i rgi(x ) , j rhj (x) = 0

i=1 j =1
j (hj (x) , dj ) = 0 (Complementarity)
j 0
In this course, we will not concern ourselves with Case (i). We will only look for candidate
solutions x for which we can nd and satisfying the equations in Case (ii) above.
In general, to solve these equations, you begin with complementarity and note that either j
must be zero or hj (x ) , dj = 0. Based on the various possibilities, you come up with one or more
candidate solutions. If there is an optimal solution, then one of your candidates will be it.
The above conditions are called the Kuhn{Tucker (or Karush{Kuhn{Tucker) conditions. Why
do they make sense?
For x optimal, some of the inequalities will be tight and some not. Those not tight can be
ignored (and will have corresponding price j = 0). Those that are tight can be treated as equalities
which leads to the previous Lagrangian stu. So
j (hj (x) , dj ) = 0 (Complementarity)

forces either the price j to be 0 or the constraint to be tight.
Example 4.2.1
Maximize x3 , 3x
Subject to
x2
The Lagrangian is
L = x3 , 3x + (2 , x)
So we need
3x2 , 3 , = 0
x2
(2 , x) = 0
0
Typically, at this point we must break the analysis into cases depending on the complementarity
conditions.
If = 0 then 3x2 , 3 = 0 so x = 1 or x = ,1. Both are feasible. f (1) = ,2, f (,1) = 2.
If x = 2 then = 9 which again is feasible. Since f (2) = 2, we have two solutions: x = ,1 and
x = 2.
Example 4.2.2 Minimize (x , 2)2 + 2(y , 1)2
Subject to
x + 4y 3
xy
First we convert to standard form, to get
Maximize ,(x , 2)2 , 2(y , 1)2
Subject to
x + 4y 3
,x + y 0
L(x; y; 1; 2) = ,(x , 2)2 , 2(y , 1)2 + 1(3 , x , 4y ) + 2(0 + x , y )

which gives optimality conditions
,2(x , 2) , 1 + 2 = 0
,4(y , 1) , 41 , 2 = 0
1(3 , x , 4y) = 0
4.2. EQUALITY AND INEQUALITY CONSTRAINTS 51
2(x , y) = 0
x + 4y 3
,x + y 0
1 ; 2 0
Since there are two complementarity conditions, there are four cases to check:
1 = 0; 2 = 0: gives x = 2, y = 1 which is not feasible.
1 = 0; x , y = 0: gives x = 4=3; y = 4=3; 2 = ,4=3 which is not feasible.
2 = 0; 3 , x , 4y = 0 gives x = 5=3; y = 1=3; 1 = 2=3 which is O.K.
3 , x , 4y = 0; x , y = 0 gives x = 3=5; y = 3=5; 1 = 22=25; 2 = ,48=25 which is not feasible.
Since it is clear that there is an optimal solution, x = 5=3; y = 1=3 is it!
Economic Interpretation
The economic interpretation is essentially the same as the equality case. If the right hand side
of a constraint is changed by a small amount , then the optimal objective changes by , where
is the optimal Lagrange multiplier corresponding to that constraint. Note that if the constraint
is not tight then the objective does not change (since then = 0).
Handling Nonnegativity
A special type of constraint is nonnegativity. If you have a constraint xk 0, you can write
this as ,xk 0 and use the above result. This constraint would get a Lagrange multiplier of its
own, and would be treated just like every other constraint.
An alternative is to treat nonnegativity implicitly. If xk must be nonnegative:
1. Change the equality associated with its partial to a less than or equal to zero:
@f (x) , X @gi (x) , X @hj (x) 0
@xk i @x j @x
i k j k
2. Add a new complementarity constraint:

0 1
X X
@ @f (x) , i @gi(x) , j @hj (x) A xk = 0
@xk i @xk j @xk
3. Don't forget that xk 0 for x to be feasible.
Suciency of conditions
The Karush{Kuhn{Tucker conditions give us candidate optimal solutions x . When are these
conditions sucient for optimality? That is, given x with and satisfying the KKT conditions,
when can we be certain that it is an optimal solution?
The most general condition available is:
1. f (x) is concave, and
2. the feasible region forms a convex region.
While it is straightforward to determine if the objective is concave by computing its Hessian
matrix, it is not so easy to tell if the feasible region is convex. A useful condition is as follows:
The feasible region is convex if all of the gi(x) are linear and all of the hj (x) are convex. If this
condition is satised, then any point that satises the KKT conditions gives a point that maximizes
f (x) subject to the constraints.
Example 4.2.3 Suppose we can buy a chemical for $10 per ounce. There are only 17.25 oz avail-
able. We can transform this chemical into two products: A and B. Transforming to A costs $3 per
oz, while transforming to B costs $5 per oz. If x1 oz of A are produced, the price we command for
A is $30 , x1 ; if x2 oz of B are produced, the price we get for B is $50 , x2 . How much chemical
should we buy, and what should we transform it to?
There are many ways to model this. Let's let x3 be the amount of chemical we purchase. Here
is one model:
Maximize x1(30 , x1 ) + x2 (50 , 2x2 ) , 3x1 , 5x2 , 10x3
Subject to
x1 + x2 , x3 0
x3 17:25
The KKT conditions are the above feasibility constraints along with:
30 , 2x1 , 3 , 1 = 0
50 , 4x2 , 5 , 1 = 0
,10 + 1 , 2 = 0
1 (,x1 , x2 + x3) = 0
2 (17:25 , x3) = 0
1 ; 2 0
There are four cases to check:
1 = 0; 2 = 0. This gives us ,10 = 0 in the third constraint, so is infeasible.
1 = 0; x3 = 17:25. This gives 2 = ,10 so is infeasible.
,x1 , x2 + x3 = 0, 2 = 0. This gives 1 = 10; x1 = 8:5; x2 = 8:75; x3 = 17:25, which is feasible.
Since the objective is concave and the constraints are linear, this must be an optimal solution. So
there is no point in going through the last case (,x1 , x2 + x3 = 0, x3 = 17:25). We are done with
x1 = 8:5; x2 = 8:75; and x3 = 17:25.
What is the value of being able to purchase 1 more unit of chemical?
4.3. EXERCISES 53
This question is equivalent to increasing the right hand side of the constraint x3 17:25 by 1
unit. Since the corresponding lagrangian value is 0, there is no value in increasing the right hand
side.
Review of Optimality Conditions.

The following reviews what we have learned so far:
Single Variable (Unconstrained)
Solve f 0 (x) = 0 to get candidate x.
If f 00 (x) > 0 then x is a local min.
f 00(x) < 0 then x is a local max.
If f (x) is convex then a local min is a global min.
f (x) is concave then a local max is a global max.
Multiple Variable (Unconstrained)
Solve rf (x) = 0 to get candidate x .
If H (x) is positive denite then x is a local min.
H (x) is negative denite x is a local max.
If f (x) is convex then a local min is a global min.
f (x) is concave then a local max is a global max.
P
Multiple Variable (Equality constrained) Form Lagrangian L(x; ) = f (x) + i i (bi , gi (x))
Solve rL(x; ) = 0 to get candidate x (and ).
Best x is optimum if optimum exists.
Multiple Variable (Equality and Inequality constrained)
Put into standard form (maximize and
P constraints) P
Form Lagrangian L(x; ) = f (x) + i i(bi , gi (x)) + j j (dj , hj (x))
Solve
rf (x) , Pi irgi(x) , Pj j rhj (x) = 0
gi (x) = bi
hj (x) dj
j (dj , hj (x)) = 0
j 0
to get candidates x (and , ).
Best x is optimum if optimum exists.
Any x is optimum if f (x) concave, gi (x) convex, hj (x) linear.
4.3 Exercises
Exercise 37 Solve the following constrained optimization problem using the method of Lagrange
multipliers.
max 2 ln x1 + 3 ln x2 + 3 ln x3
s.t. x1 + 2x2 + 2x3 = 10
Exercise 38 Find the two points on the ellipse given by x21 +4x22 = 4 that are at minimum distance
of the point (1; 0). Formulate the problem as a minimization problem and solve it by solving the
Lagrangian equations. [Hint: To minimize the distance d between two points, one can also minimize
d2. The formula for the distance between points (x1; x2) and (y1; y2) is d2 = (x1 , y1 )2 +(x2 , y2 )2.]
Exercise 39 Solve using Lagrange multipliers.
a) min x21 + x22 + x23 subject to x1 + x2 + x3 = b, where b is given.
b) max px1 + px2 + px3 subject to x1 + x2 + x3 = b, where b is given.
c) min c1x21 + c2x22 + c3x23 subject to x1 + x2 + x3 = b, where c1 > 0, c2 > 0, c3 > 0 and b are given.
d) min x21 + x22 + x23 subject to x1 + x2 = b1 and x2 + x3 = b2, where b1 and b2 are given.
Exercise 40 Let a, b and c be given positive scalars. What is the change in the optimum value
of the following constrained optimization problem when the right hand side of the constraint is
5 a.
increased by 5%, i.e. a is changed to a + 100
max by , x4
s.t. x2 + cy = a
Give your answer in terms of a, b and c.
Exercise 41 You want to invest in two mutual funds so as to maximize your expected earnings
while limiting the variance of your earnings to a given gure s2 . The expected yield rates of Mutual
Funds 1 and 2 are r1 and r2 respectively, and the variance of earnings for the portfolio (x1 ; x2) is
2x21 + x1x2 + 2x22. Thus the problem is
max r1x1 + r2x2
s.t. 2 x21 + x1x2 + 2 x22 = s2
(a) Use the method of Lagrange multipliers to compute the optimal investments x1 and x2 in
Mutual Funds 1 and 2 respectively. Your expressions for x1 and x2 should not contain the
Lagrange multiplier .
(b) Suppose both mutual funds have the same yield r. How much should you invest in each?
Exercise 42 You want to minimize the surface area of a cone-shaped drinking cup having xed
volume V0. Solve the problem as a constrainedpoptimization problem. To simplify the algebra,
minimize the square of the area. The area is r r2 + h2 . The problem is,
min 2r4 + 2r2 h2
s.t. 13 r2h = V0 :
Solve the problem using Lagrange multipliers.
Hint. You can assume that r 6= 0 in the optimal solution.
4.3. EXERCISES 55
Exercise 43 A company manufactures two types of products: a standard product, say A, and a
more sophisticated product B . If management charges a price of pA for one unit of product A and
a price of pB for one unit of product B , the company can sell qA units of A and qB units of B ,
where
qA = 400 , 2pA + pB ; qB = 200 + pA , pB :
Manufacturing one unit of product A requires 2 hours of labor and one unit of raw material. For
one unit of B , 3 hours of labor and 2 units of raw material are needed. At present, 1000 hours of
labor and 200 units of raw material are available. Substituting the expressions for qA and qB , the
problem of maximizing the company's total revenue can be formulated as:
max 400pA + 200pB , 2p2A , p2B + 2pA pB
s.t. ,pA , pB ,400
,pB ,600
(a) Use the Khun-Tucker conditions to nd the company's optimal pricing policy.
(b) What is the maximum the company would be willing to pay for
{ another hour of labor,
{ another unit of raw material?
Chapter 5
Modeling with Linear Programming

5.1 Introductory Example
SilComputers makes quarterly decisions about their product mix. While their full product line
includes hundreds of products, we will consider a simpler problem with just two products: notebook
computers and desktop computers. SilComputers would like to know how many of each product to
produce in order to maximize prot for the quarter.
There are a number of limits on what SilComputers can produce. The major constraints are as
follows:
1. Each computer (either notebook or desktop) requires a Processing Chip. Due to tightness in
the market, our supplier has allocated 10,000 such chips to us.
2. Each computer requires memory. Memory comes in 16MB chip sets. A notebook computer
has 16MB memory installed (so needs 1 chip set) while a desktop computer has 32MB (so
requires 2 chip sets). We received a great deal on chip sets, so have a stock of 15,000 chip
sets to use over the next quarter.
3. Each computer requires assembly time. Due to tight tolerances, a notebook computer takes
more time to assemble: 4 minutes versus 3 minutes for a desktop. There are 25,000 minutes
of assembly time available in the next quarter.
Given current market conditions, material cost, and our production system, each notebook
computer produced generates $750 prot, and each desktop produces $1000 prot.
There are many questions SilComputer might ask. The most obvious are such things as \How
many of each type computer should SilComputer produce in the next quarter?" \What is the
maximum prot SilComputer can make?" Less obvious, but perhaps of more managerial interest
are \How much should SilComputer be willing to pay for an extra memory chip set?" \What is
the eect of losing 1,000 minutes of assembly time due to an unexpected machine failure?" \How
much prot would we need to make on a 32MB notebook computer to justify its production?"
Linear programming gives us a mechanism for answering all of these questions quickly and
easily. There are three steps in applying linear programming: modeling, solving, and interpreting.
5.1.1 Modeling
We begin by modeling this problem. Modeling a problem using linear programming involves writing
it in the language of linear programming. There are rules about what you can and cannot do within
57
58 CHAPTER 5. MODELING WITH LINEAR PROGRAMMING
linear programming. These rules are in place to make certain that the remaining steps of the process
(solving and interpreting) can be successful.
Key to a linear program are the decision variables, objective, and constraints.
Decision Variables. The decision variables represent (unknown) decisions to be made. This
is in contrast to problem data, which are values that are either given or can be simply calculated
from what is given. For this problem, the decision variables are the number of notebooks to
produce and the number of desktops to produce. We will represent these unknown values by x1
and x2 respectively. To make the numbers more manageable, we will let x1 be the number of
1000 notebooks produced (so x1 = 5 means a decision to produce 5000 notebooks) and x2 be the
number of 1000 desktops. Note that a value like the quarterly prot is not (in this model) a decision
variable: it is an outcome of decisions x1 and x2 .
Objective. Every linear program has an objective. This objective is to be either minimized
or maximized. This objective has to be linear in the decision variables, which means it must be
the sum of constants times decision variables. 3x1 , 10x2 is a linear function. x1 x2 is not a linear
function. In this case, our objective is to maximize the function 750x1 + 1000x2 (what units is this
in?).
Constraints. Every linear program also has constraints limiting feasible decisions. Here we
have four types of constraints: Processing Chips, Memory Sets, Assembly, and Nonnegativity.
In order to satisfy the limit on the number of chips available, it is necessary that x1 + x2 10.
If this were not the case (say x1 = x2 = 6), the decisions would not be implementable (12,000 chips
would be required, though we only have 10,000). Linear programming cannot handle arbitrary
restrictions: once again, the restrictions have to be linear. This means that a linear function of
the decision variables must be related to a constant, where related can mean less than or equal to,
greater than or equal to, or equal to. So 3x1 , 2x2 10 is a linear constraint, as is ,x1 + x3 = 6.
x1 x2 10 is not a linear constraint, nor is x1 + 3x2 < 3. Our constraint for Processing Chips
x1 + x2 10 is a linear constraint.
The constraint for memory chip sets is x1 + 2x2 15, a linear constraint.
Our constraint on assembly can be written 4x1 + 3x2 25, again a linear constraint.
Finally, we do not want to consider decisions like x1 = ,5, where production is negative. We
add the linear constraints x1 0, x2 0 to enforce nonnegativity of production.
Final Model. This gives us the complete model of this problem:
Maximize 750x1 + 1000x2
Subject to
x1 + x2 10
x1 + 2x2 15
4x1 + 3x2 25
x1 0
x2 0
Formulating a problem as a linear program means going through the above process to clearly
dene the decision variables, objective, and constraints.
5.1.2 Solving the Model
Models are useful in their own right: they allow for a formal denition of a problem and can be
useful in thinking about a problem. Linear programming models are particularly useful, however,
because it is easy to have a computer solve for the optimal decision values.
5.1. INTRODUCTORY EXAMPLE 59
For a model with only two variables, it is possible to solve the problem without a computer by
drawing the feasible region and determining how the objective is optimized on that region. We go
through that process here. The purpose of this exercise is to give you intuition and understanding
of linear programming models and their solution. In any real application, you would use a computer
to solve even two variable problems (we outline how to use Excel's Solver routine to nd solutions
in the next section).
We can represent a model with two variables by labeling the axes of a graph with each of the
variables. The entire graph then represents possible decisions. Constraints are represented by lines
on the graph, with the feasible region lying on one side of the line. The following gure illustrates
this with the constraint x1 + x2 10.
x2
10
6
x1 + x2 10
4
2 4 6 8 10 x1
Figure 5.1: Single Constraint

We can continue this process and add in all of the constraints. Since every constraint must be
satised, the resulting feasible region is the intersection of the feasible region for each constraint.
This is shown in the following gure.
Note that just graphing the model gives us information we did not have before. It seems that
the Chip constraint (x1 + x2 10) plays little role in this model. This constraint is dominated by
other constraints.
Now, how can we nd the optimal solution? We can include the objective function on this
diagram by drawing iso-prot lines: lines along which the prot is the same. Since our goal is to
maximize the prot, we can push the isoprot line out until moving it any further would result in
no feasible point (see diagram, z represents prot). Clearly the optimal prot occurs at point X .
What are the variable values at point X . Note that X is the intersection of the constraints:
x2
10
6
x1 + x2 10
4
2
4x1 + 3x2 25 x1 + 2x2 15
0 2 4 6 8 10 x1
Figure 5.2: All Constraints
x2
10
8
X
6
x1 + x2 10
4 z=6000
x1 + 2x2 15
2 4x1 + 3x2 25
z=3000
0 2 4 6 8 10 x1
Figure 5.3: Finding Optimal Prot

5.1. INTRODUCTORY EXAMPLE 61
x1 + 2x2 = 15
4x1 + 3x2 = 25
We discussed in the previous section how to solve such equations: the solution here is x1 = 1
and x2 = 7. The optimal decision is to produce 1,000 notebooks and 7,000 desktops, for a prot of
$7,750,000.
5.1.3 Using Solver

Rather than the somewhat tedious and error-prone graphical method (which is limited to linear
programs with 2 variables), special computer programs can be used to nd solutions to linear
programming models. The most widespread program is undoubted Solver, included in all recent
versions of the Excel spreadsheet program. Solver, while not a state of the art code (which can
cost upwards of $15,000 per copy) is a reasonably robust, easy-to-use tool for linear programming.
Solver uses standard spreadsheets together with an interface to dene variables, objective, and
constraints to dene a linear program.
It is dicult to describe in words how to create a Solver spreadsheet, so we will do one or two
in class. Here is a brief outline and some hints and shortcuts:
We will start with a spreadsheet that has all of the data entered in some reasonably neat way.
We will create the model in a separate part of the spreadsheet. We will have one cell for each
variable. Solver will eventually put the optimal values in each cell.
We will have a single cell to represent the objective. We will enter a formula that rep-
resents the objective. This formula must be a linear formula, so it must be of the form:
cell1*cell1'+cell2*cell2'+: : : , where cell1, cell2 and so on contain constant values and cell1',
cell2' and so are the variable cells.
Helpful Hint: Excel has a function sumproduct() that is designed for linear programs.
sumproduct(a1..a10,b1..b10) is identical to a1*b1+a2*b2+a3*b3+: : : +a10*b10. This func-
tion will save much time and aggravation. All that is needed is that the length of the rst
range is the same as the length of the second range (so one can be horizontal and the other
vertical).
Helpful Hint: It is possible to assign names to cells and ranges (under the Insert-Name menu).
Rather than use a1..a10 as the variables, you can name that range var (for example) and
then use var wherever a1..a10 would have been used.
We then have a cell to represent the left hand side of each constraint (again a linear function)
and another cell to represent the right hand side (a constant).
We then select Solver under the Tools menu. This gives a form to ll out to dene the linear
program.
In the \Set Cell" box, select the objective cell. Choose Maximize or Minimize.
In the \By Changing Cells", put in the range containing the variable cells.
We next add the constraints. Press the \Add..." button to add constraints. The dialog box
has three parts for the left hand side, the type of constraint, and the right hand side. Put the
cell references for a constraint in the form, choose the right type, and press \Add". Continue
until all constraints are added. On the nal constraint, press \OK".
We need to explicitly include nonnegativity constraints.
Helpful Hint: It is possible to include ranges of constraints, as long as they all have the same
type. c1..e1 <= c3..e3 means c1 <= c3, d1 <= d3, e1 <= e3. a1..a10 >= 0 means each
individual cell must be greater than or equal to 0.
Push the options button and toggle the \Assume Linear Model" in the resulting dialog box.
This tells Excel to call a linear rather than a nonlinear pgrogramming routine so as to solve
the problem more eciently. This also gives you sensitivity ranges, which are not available
for nonlinear models.
Push the Solve button. In the resulting dialog box, select \Answer" and \Sensitivity". This
will put the answer and sensitivity analysis in two new sheets. Ask Excel to \Keep Solver
values", and your worksheet will be updated so that the optimal values are in the variable
cells.
5.2 Second Example: Market Allocations

MegaMarketing is planning a concentrated one week advertising campaign for their new CutsEv-
erything SuperKnife. The ads have been designed and produced and now they wish to determine
how much money to spend in each advertising outlet. In reality, they have hundreds of possible
outlets to choose from. We will illustrate their problem with two outlets: Prime-time TV, and
newsmagazines.
The problem of optimally spending advertising dollars can be formulated in many ways. For
instance, given a xed budget, the goal might be to maximize the number of target customers
reached (a target customer is a customer with a reasonable chance of purchasing the product).
An alternative approach, which we adopt here, is to dene targets for reaching each market
segment and to minimize the money spent to reach those targets. For this product, the target
segments are Teenage Boys, Auent Women (ages 40-49), and Retired Men. Each minute of
primetime TV and page of newsmagazine advertisement reaches the following number of people (in
millions):
Outlet Boys Women Men Cost
TV 5 1 3 600
Mag 2 6 3 500
Target 24 18 24
Again, MegaMarketing is interested in straightforward answers like how many units of each
outlet to purchase to meet the segment goals. They are also interested in such questions as \How
much will it cost to reach an extra million retired men?", \One radio spot reaches 1 million boys, 1
million women, and 1 million men: how much are we willing to pay for such a spot?", and similar
questions.
5.2. SECOND EXAMPLE: MARKET ALLOCATIONS 63
5.2.1 Modeling
The process of modeling using linear programming is fairly consistent: dene the decision variables,
objective, and constraints.
Decision Variables. In this case, the decision variables are the number of units of each outlet
to purchase. Let x1 be the number of TV spots, and x2 the number of magazine pages.
Objective. Our objective is to minimize the amount we spend: 600x1 + 500x2.
Constraints. We have one constraint for each market segment: we must be certain that we
reach sucient people of each segment. For boys, this corresponds to the constraint:
5x1 + 2x2 24
Similar constraints for the other segments gives us the full formulation as:
Minimize 600x1 + 500x2
Subject to
5x1 + 2x2 24
x1 + 6x2 18
3x1 + 3x2 24
x1 0
x2 0
5.2.2 Solving the Model
Since this model has only two variables, we can solve the problem in two ways: graphically and
using SOLVER (with SOLVER being much easier). We start by solving this graphically.
The rst step is to graph the feasible region, as given in the following gure:
5x1 + 2x2 24
10
x2
8
4
3x1 + 3x2 24
x1 + 6x2 18
2
0 2 4 6 8 10 12
x1
Figure 5.4: Marketing example

The next step is to put some iso-cost lines on the diagram: lines representing points that have
the same cost. We mark the optimal point with an X .
5x1 + 2x2 24
10
x2
z = 6000000
8
6
X
4
3x1 + 3x2 24
x1 + 6x2 18
2
0 2 4 6 8 10 12
x1
Figure 5.5: Marketing example with Isocost Line

X is the intersection of the constraints:
5x1 + 2x2 = 24
3x1 + 3x2 = 24
The optimal solution is x1 = 2:67 and x2 = 5:33
5.2.3 Using Solver

The SOLVER model for this problem is on the course web pages.
5.3 Common Features

Hidden in our models of these problems are a number of assumptions. The usefulness of a model
is directly related to how close reality matches up with these assumptions.
The rst assumption is due to the linear form of our functions. Since the objective is linear, the
contribution to the objective of any decision variable is proportional to the value of the decision
variable. Producing twice as much of a product produces twice as much prot; buying twice as
many pages of ads costs twice as much. This is the Proportionality Assumption.
Furthermore, the contribution of a variable to the objective is independent of the values of
the other variables. One notebook computer is worth $750, independent of how many desktop
computers we produce. This is the Additivity Assumption.
5.4. LINEAR PROGRAMMING MODELS 65
Similarly, since each constraint is linear, the contribution of each variable to the left hand side
of each constraint is proportional to the value of the variable and independent of the values of any
other variable.
These assumptions are quite restrictive. We will see, however, that clever modeling can handle
situations that may appear to violate these assumptions.
The next assumption is the Divisibility Assumption: it is possible to take any fraction of any
variable. Rethinking the Marketing example, what does it mean to purchase 2.67 television ads? It
may be that the divisibility assumption is violated in this example. Or, it may be that the units are
such that 2.67 \ads" actually corresponds to 2666.7 minutes of ads, in which case we can \round
o" our solution to 2667 minutes with little doubt that we are getting an optimal or nearly optimal
solution. Similarly, a fractional production quantity may be worisome if we are producing a small
number of battleships or be innocuous if we are producting millions of paperclips. If the Divisibility
Assumption is important and does not hold, then a technique called integer programming rather
than linear programming is required. This technique takes orders of magnitude more time to nd
solutions but may be necessary to create realistic solutions. You will learn more about this in
45-761.
The nal assumption is the Certainty Assumption: linear programming allows for no uncertainty
about the numbers. An ad will reach the given number of people; the number of assembly hours
we give will certainly be available.
It is very rare that a problem will meet all of the assumptions exactly. That does not negate
the usefulness of a model. A model can still give useful managerial insight even if reality diers
slightly from the rigorous requirements of the model. For instance, the knowledge that our chip
inventory is more than sucient holds in our rst model even if the proportionality assumptions
are not satised completely.
5.4 Linear Programming Models

Linear programming models are found in almost every eld of business (and beyond!). The next
sections go through a number of problems, showing how to model them by the appropriate choice
of decision variables, objective, and constraints. In all cases, we will describe the problem and give
a model. Solver worksheets for each of these is available on the web page.
5.4.1 Diet Problem

Problem Denition
What is the perfect diet? An ideal diet would meet or exceed basic nutritional requirements, be
inexpensive, have variety, and be \pleasing to the palate". How can we nd such a diet?
Suppose the only foods in the world are as follows:
Energy Protein Calcuim Price Limit
Food Serving Size (kcal) (g) (mg) (cents/serving) (servings/day)
Oatmeal 28g 110 4 2 3 4
Chicken 100g 205 32 12 24 3
Eggs 2 Large 160 13 54 13 2
Whole Milk 237cc 160 8 285 9 8
Cherry Pie 170g 420 4 22 20 2
Pork&Beans 260g 260 14 80 19 2
After consulting with nutitionists, we decree that a satisfactory diet has at least 2000 kcal of
energy, 55 g of protein, and 800 mg of calcium (vitamins and iron are supplied by pills). While
some of us would be happy to subsist on 10 servings of pork and beans, we have decided to impose
variety by having a limit on the number of servings/day for each of our six foods. What is the least
cost satisfactory diet?
Problem Modeling
First we decide on decision variables. Let us label the foods 1, 2, : : : 6, and let xi represent the
number of servings of food i in the diet.
Our objective is to minimize cost, which can be written
3x1 + 24x2 + 13x3 + 9x4 + 20x5 + 19x6
We have constraints for energy, protein, calcuim, and for each serving/day limit. This gives the
formulation:
Minimize 3x1 + 24x2 + 13x3 + 9x4 + 20x5 + 19x6
Subject to
110x1 + 205x2 + 160x3 + 160x4 + 420x5 + 260x6 2000
4x1 + 32x2 + 13x3 + 8x4 + 4x5 + 14x6 55
2x1 + 12x2 + 54x3 + 285x4 + 22x5 + 80x6 800
x1 4
x2 3
x3 2
x4 8
x5 2
x6 2
xi 0 (for all i)
Discussion
The creation of optimal diets was one of the rst uses of linear programming. Some diculties
with linear programming include diculties in formulating \palatability" requirements, and issues
of divisibility (no one wants to eat half a green bean). These linear programming models do give
an idea on how much these palatability requirements are costing.
5.4.2 Workforce Planning

Problem Denition.
Consider a restaurant that is open seven days a week. Based on past experience, the number of
workers needed on a particular day is given as follows:
Day Mon Tue Wed Thu Fri Sat Sun
Number 14 13 15 16 19 18 11
Every worker works ve consecutive days, and then takes two days o, repeating this pattern
indenitely. How can we minimize the number of workers that sta the restaurant?
Model.
A natural (and wrong!) rst attempt at this problem is to let xi be the number of people working
on day i. Note that such a variable denition does not match up with what we need to nd. It
does us no good to know that 15 people work Monday, 13 people Tuesday, and so on because it
does not tell us how many workers are needed. Some workers will work both Monday and Tuesday,
some only one day, some neither of those days. Instead, let the days be numbers 1 through 7 and
let xi be the number of workers who begin their ve day shift on day i. Our objective is clearly:
x1 + x2 + x3 + x4 + x5 + x6 + x7
Consider the constraint for Monday's stang level of 14. Who works on Mondays? Clearly
those who start their shift on Monday (x1 ). Those who start on Tuesday (x2 ) do not work on
Monday, nor do those who start on Wednesday (x3). Those who start on Thursday (x4 ) do work
on Monday, as do those who start on Friday, Saturday, and Sunday. This gives the constraint:
x1 + x4 + x5 + x6 + x7 14
Similar arguments give a total formulation of:
Minimize
Px
i i
Subject to
x1 + x4 + x5 + x6 + x7 14
x1 + x2 + x5 + x6 + x7 13
x1 + x2 + x3 + x6 + x7 15
x1 + x2 + x3 + x4 + x7 16
x1 + x2 + x3 + x4 + x5 19
x2 + x3 + x4 + x5 + x6 18
x3 + x4 + x5 + x6 + x7 11
xi 0 (for all i)
Discussion.
Workforce modeling is a well developed area. Note that our model has only one type of shift, but
the model is easily extended to other types of shifts, with diering shift costs.
5.4.3 Financial Portfolio

Problem Denition.
In your nance courses, you will learn a number of techniques for creating \optimal" portfolios. The
optimality of a portfolio depends heavily on the model used for dening risk and other aspects of
nancial instruments. Here is a particularly simple model that is amenable to linear programming
techniques.
Consider a mortgage team with $100,000,000 to nance various investments. There are ve
categories of loans, each with an associated return and risk (1-10, 1 best):
Loan/investment Return (%) Risk
First Mortgages 9 3
Second Mortgages 12 6
Personal Loans 15 8
Commercial Loans 8 2
Government Securities 6 1
Any uninvested money goes into a savings account with no risk and 3% return. The goal for
the mortgage team is to allocate the money to the categories so as to:
(a) Maximize the average return per dollar
(b) Have an average risk of no more than 5 (all averages and fractions taken over the invested
money (not over the saving account)).
(c) Invest at least 20% in commercial loans
(d) The amount in second mortgages and personal loans combined should be no higher than
the amount in rst mortgages.
Model
Let the investments be numbered 1: : : 5, and let xi be the amount invested in investment i. Let xs
be the amount in the savings account. The objective is to maximize
9x1 + 12x2 + 15x3 + 8x4 + 6x5 + 3xs
subject to
x1 + x2 + x3 + x4 + x5 + xs = 100; 000; 000:
Now, let's look at the average risk. Since we want to take the average over only the invested
amount, a direct translation of this constraint is
3x1 + 6x2 + 8x3 + 2x4 + x5 5
x1 + x2 + x3 + x4 + x5
This constraint is not linear, but we can cross multiply, and simplify to get the equivalent linear
constraint:
,2x1 + x2 + 3x3 , 3x4 , 4x5 0
Similarly we need
x4 0:2(x1 + x2 + x3 + x4 + x5 )
or
,0:2x1 , 0:2x2 , 0:2x3 + 0:8x4 , 0:2x5 0
The nal constraint is
x2 + x3 , x1 0
Together with nonnegativity, this gives the entire formulation.
Discussion
Optimal portfolios do not just happen: they must be calculated, and there is a constant interplay
between models and solvability. Linear programming models provide great modeling power with
a great limit: the handling of risk must be done in a linear fashion (like our Risk factors here).
Other models you will see in nance will look at the co-variance of returns between investments, a
fundamentally nonlinear eect. This can give rise to nonlinear models like those that try to minimize
variance subject to return requirements. It is very dicult to embed idiosyncratic constraints (like
(c) and (d) here) in such models.
5.4.4 Production over Time
Problem Description
Many problems are multiperiod, where a series of decisions have to be made over time. For instance,
in our very rst SilComputer example, production decisions for computers have to be made with
the future in mind. In this example, it would be easy to work with both desktop and notebook
computers but let's work just with notebook computers.
SilComputer needs to meet the demand of its largest corporate and educational customers
for notebook computers over the next four quarters (before its current model becomes obsolete).
SilComputer currently has 5,000 notebook computers in inventory. Expected demand over the
next four quarters for its notebook is 7,000; 15,000; 10,000; and 8,000. SilComputer has sucient
capacity and material to produce up to 10,000 computers in each quarter at a cost of $2000 per
notebook. By using overtime, up to an additional 2,500 computers can be produced at a cost of
$2200 each. Computers produced in a quarter can be used either to meet that quarter's demand,
or be held in inventory for use later. Each computer in inventory is charged $100 to re ect carying
costs.
How should SilComputer meet its demand for notebooks at minimum cost?
Modeling.
The decisions seem to be how many of each computer to produce in each period at regular time,
how many to produce at overtime, and how much inventory to carry in each period. Let's denote
our time periods t = 1; 2; 3; 4. Let xt be the number of notebooks produced in period t at regular
time and yt be the number of notebooks produced in period t at overtime. Finally, let it be the
inventory at the end of period t.
Look at the rst quarter: how are these variables restricted and related?
Clearly we need x1 10000 and y1 2500. Now, anything that starts as inventory or is
produced in the period must either be used to meet demand or ends up as inventory at the end of
period 1. This means:
5000 + x1 + y1 = 7000 + i1
For period 2, in addition to the upper bounds, we get the constraint
i1 + x2 + y2 = 15000 + i2
For period 3, we get
i2 + x3 + y3 = 10000 + i3
and for period 4 (assuming no inventory at the end):
i3 + x4 + y4 = 8000
Our objective charges 2000 for each x, 2200 for each y , and 100 for each i:
2000x1 + 2000x2 + 2000x3 + 2000x4 + 2200y1 + 2200y2 + 2200y3 + 2200y4 + 100i1 + 100i2 + 100i3
Discussion.
The constraints tracking where production goes are known as conservation constraints, and model
the requirement that production never disappears. Whenever you have multiple types of variables
in a formulation, be certain to think about whether there are necessary constraints that link them.
5.4.5 Investing over Time
We are going to manage an investment portfolio over a 6-year time horizon. We begin with $1000,
and at various times we can invest in one or more of the following:
Savings account X, annual yield 5%.
Security Y, 2-year maturity, total yield 12% if bought now, 11% thereafter.
Security Z, 3-year maturity, total yield 18%.
Security W, 4-year maturity, total yield 24%.
To keep things simple we will assume that each security can be bought in any denomination.
(This assumption can be relaxed if one uses integer or dynamic programming .) We can make
savings deposits or withdrawals anytime. We can buy Security Y any year but year 3. We can buy
Security Z anytime after the rst year. Security W, now available, is a one-time opportunity.
We let xt be the amount of money invested in the savings account X at the beginning of year
t, and similarly for yt, zt, and wt . We will put any money not tied up in securities into the savings
account. The situation is summed up in the gure.
The problem is really a kind of inventory problem. In a given year, the amount of money
carried forward from the previous year (in savings), plus the yield from securities that mature that
year, equals the amount invested in new securities plus the amount of money left over for next
year. There is one twist: inventory grows while in storage. (In real inventory problems, stock often
decreases over time due to spoilage, etc., and this can be re ected in negative \interest" rates.)
Let v be the nal yield when all securities are cashed in at the end of the sixth year. Then if the
objective is to maximize nal yield, the LP is,
max v
s.t. x1 + y1 + w1 = 1000
x2 + y2 + z2 = 1:05x1
x3 + z3 = 1:05x2 + 1:12y1
x4 + y4 + z4 = 1:05x3 + 1:11y2
x5 + y5 = 1:05x4 + 1:18z2 + 1:24w1
x6 = 1:05x5 + 1:11y4 + 1:18z3
v = 1:05x6 + 1:11y5 + 1:18z4;
xt; yt; zt; wt 0
1 2 3 4 5 6 7
x1 x2 x3 x4 x5 x6
5%
y1 y5
12% 11%
y2 y4
11% 11%
z2
18%
z3
18%
z4
18%
w1
24%
Figure 5.6: Possible investments over a 6-year horizon
The nal assets come to $1387.6, an average annual yield of 5.61%
5.4.6 Problems
Modeling is an art, and it takes practice. The following examples show the variety of problems that
can be attacked by linear programming, and give you the opportunity to try your hand at some
problems.
Exercise 44 A plant produces two types of refrigerators, A and B. There are two production lines,
one dedicated to producing refrigerators of Type A, the other to producing refrigerators of Type
B. The capacity of the production line for A is 60 units per day, the capacity of the production
line for B is 50 units per day. A requires 20 minutes of labor whereas B requires 40 minutes of
labor. Presently, there is a maximum of 40 hours of labor per day which can be assigned to either
production line. Prot contributions are $20 per refrigerator of Type A produced and $30 per Type
B produced. What should the daily production be? Solve graphically and by Solver.
Exercise 45 Albert, Bill, Charles, David, and Edward have gotten into a bind. After a series of
nancial transactions, they have ended up each owing some of the others huge amounts of money.
In fact, near as the lawyers can make out, the debts are as follows
Debtor Creditor Amount
($millions)
A E 10
A C 3
B A 5
C B 6
C D 4
D A 4
E C 7
E D 3
The question is, who is bankrupt? We will say that a person i is bankrupt if there is no possible
transfer of funds among the people such that i completely pays o his obligations, and the transfer
of funds satises the following condition: for every two persons j and k, the amount paid by person
j to person k is no greater than the debt of j to k. For instance, Albert is bankrupt since he
owes 10, and is only owed 9. Formulate the problem of determining whether Bill is bankrupt as
a linear program. Then modify your formulation to determine if each of the others is bankrupt.
This example may look contrived, but it is inspired by a solution to the debts involved in a crash of
Kuwait's al-Mankh stock market.
Exercise 46 Due to an unexpected glut of orders, Blaster Steel has decided to hire temporary
workers for a ve day period to clear out the orders. Each temporary worker can work either a two
day shift or a three day shift for this period (shifts must be consecutive days). At least 10 workers
are needed on days 1, 3, 5, and at least 15 workers are needed on days 2 and 4. A worker on a two
day shift gets paid $125/day, while those on a three day shift gets paid $100/day.
(a) Formulate the problem of hiring temporary workers to minimize cost while meeting the
demand for workers.
(b) Due to a limited number of training personnel, no more than 10 workers can start their shift
on any day. Update your formulation in (a) to take this into account.
(c) Union regulations require that at least half of all money spent on workers go to those who
work three day shifts. Update your formulation in (a) to handle this requirement.
(d) There are four people who are willing to work a shift consisting of days 1, 2, and 5, for a
payment of $110/day. Update your formulation in (a) to handle this possibility.
Exercise 47 You have $1000 to invest in Securities 1 and 2; uninvested funds are deposited in a
savings account with annual yield of 5%. You will sell the securities when they reach maturity.
Security 1 matures after 2 years with a total yield of 12%. Security 2 matures after 3 years with
a total yield of 19%. Your planning horizon is 7 years, and you want to maximize total assets
at the end of the seventh year. Suppose either security may be purchased in arbitrarily small
denominations.
a) Let xit be the amount invested in Security i at the beginning of year t; the savings account can
be considered Security 0. Write the appropriate LP model.
b) Solve the problem by computer and indicate your optimal portfolio in each year 1,...,7.
Exercise 48 Chemco produces two chemicals: A and B. These chemicals are produced via two
manufacturing processes. Process 1 requires 2 hours of labor and 1 lb. of raw material to produce
2 oz. of A and 1 oz. of B . Process 2 requires 3 hours of labor and 2 lb. of raw material to produce
3 oz. of A and 2 oz. of B . Sixty hours of labor and 40 lb. of raw material are available. Chemical
A sells for $16 per oz. and B sells for $14 per oz. Formulate a linear program that maximizes
Chemco's revenue.
Hint: Dene the amounts of Process 1 and Process 2 used, as your decision variables.
Exercise 49 Gotham City National Bank is open Monday-Friday from 9 am to 5 pm. From past
experience, the bank knows that it needs the following number of tellers.
Time Period Tellers Required
9 , 10 4
10 , 11 3
11 , noon 4
noon , 1 6
1,2 5
2,3 6
3,4 8
4,5 8
The bank hires two types of tellers. Full-time tellers work 9-5 ve days a week, except for 1
hour o for lunch, either between noon and 1 pm or between 1 pm and 2 pm. Full-time tellers are
paid (including fringe benets) $25/hour (this includes payment for lunch hour). The bank may
also hire up to 3 part-time tellers. Each part-time teller must work exactly 4 consecutive hours
each day. A part-time teller is paid $20/hour and receives no fringe benets. Formulate a linear
program to meet the teller requirements at minimum cost.
Exercise 50 Sales forecasts for the next four months are (in thousand of units):
October 10
November 16
December 10
January 12
September's production was set at 12,000 units. Varying production rate incurs some cost:
production can be increased from one month to the next at a cost of $2 per unit and it can be
decreased at a cost of $0.50 per unit. In addition, inventory left at the end of a month can be
stored at a cost of $1 per unit per month. Given current demand, there will be no inventory at the
end of September. No inventory is desired at the end of January. Formulate a linear program that
minimizes the total cost (varying production rate + inventory costs) of meeting the above demand.
Exercise 51 An oil company blends gasoline from three ingredients: butane, heavy naphta and
catalytic reformate. The characteristics of the ingredients as well as minimum requirements for
regular gasoline are given below:
Catalytic Heavy
Butane Reformate Naphta Gasoline
Octane 120 100 74 89
Vapor Pressure 60 2:5 4:0 11
Volatility 105 3 12 17
The cost (per gallon) of butane is $0.58, it is $1.55 for catalytic reformate and $0.85 for heavy
naphta. How many gallons of the three ingredients should be blended in order to produce 12,000
gallons of gasoline at minimum cost?
Exercise 52 An electric utility has six power plants on the drawing board. The anticipated useful
life of these plants is 30 years for the coal-red plants (plants 1, 2 and 3) and 40 years for the
fuel-red plants (plants 4, 5 and 6). Plants 1, 2 and 4 will be on line in year 5. Plants 3 and 5 in
year 15. Plant 6 in year 25. The cost of installing generating capacity at Plant i, discounted to the
present, is ci per megawatt, for i = 1; 2; : : :; 6. Projected power demand in year t is Dt megawatts,
for t = 10; 20; 30; 40: Due to environmental regulations, the fraction of generating capacity at
coal-red plants in year t, relative to total generating capacity in year t, can be at most rt for
t = 10; 20; 30; 40.
The generating capacity of each plant has yet to be determined. Write a linear program that as-
signs capacities to the six plants so as to minimize total present cost of installing generating capacity,
while meeting demand and satisfying the environmemtal constraints in years t = 10; 20; 30; 40:
Exercise 53 (Wagner) You must decide how many tons x1 of pure steel and how many tons x2 of
scrap metal to use in manufacturing an alloy casting. Pure steel costs $300 per ton and scrap $600
per ton (larger because the impurities must be skimmed o). The customer wants at least 5 tons
but will accept a larger order, and material loss in melting and casting is negligible. You have 4
tons of steel and 7 tons of scrap to work with, and the weight ratio of scrap to pure cannot exceed
7/8 in the alloy. You are allotted 18 hours melting and casting time in the mill; pure steel requires
3 hrs per ton and scrap 2 hrs per ton.
a) Write a linear programming model for the problem assuming that the objective is to minimize
the total steel and scrap costs.
b) Graph the problem obtained in (a).
c) Solve the problem in (a) using Solver and indicate the optimal values of x1, x2. What is the
weight ratio of scrap to pure steel?
Exercise 54 (Wagner) An airline must decide how many new ight attendants to hire and train
over the next six months. The sta requirements in person- ight-hours are respectively 8000, 9000,
7000, 10,000, 9000, and 11,000 in the months January through June. A new ight attendant is
trained for one month before being assigned to a regular ight and therefore must be hired a month
before he or she is needed. Each trainee requires 100 hours of supervision by experienced ight
attendants during the training month, so that 100 fewer hours are available for ight service by
regular ight attendants.
Each experienced ight attendant can work up to 150 hours a month, and the airline has 60
regular ight attendants available at the beginning of January. If the maximum time available from
experienced ight attendants exceeds a month's ying and training requirements, they work fewer
hours, and none are laid o. At the end of each month, 10% of the experienced ight attendants
quit their jobs.
An experienced ight attendant costs the company $1700 and a trainee $900 a month in salary
and benets.
a) Formulate the problem as a linear programming model. Let xt be the number of ight attendants
that begin training in month t. Hint. Let yt be the number of experienced ight attendants
available in month t; also let x0 = 0, y1 = 60. The problem is really an inventory problem,
with two kinds of stock: trainees and experienced employees. The \holding costs" are the
salaries, and the demands are the number of ight hours needed. \Spoilage" is attrition of
experienced sta.
b) Solve the model with Solver and indicate the solution value for the xt's and yt's.
Exercise 55 (Wagner) A lumber company operates a sawmill that converts timber to lumber or
plywood. A marketable mix of 1000 board feet of lumber products requires 1000 board feet of
spruce and 4000 board feet of Douglas r. Producing 1000 square feet of plywood requires 2000
board feet of spruce and 4000 board feet of r. The company's timberland yields 32,000 board feet
of spruce and 72,000 board feet of r each season.
Sales commitments require that at least 5000 board feet of lumber and 12,000 board feet of
plywood be produced during the season. The prot contributions are $45 per 1000 board feet of
lumber and $60 per 1000 square feet of plywood.
a) Express the problem as a linear programming model.
b) Graph the problem and indicate the optimal solution on the graph.
Exercise 56 (Wagner) An electronics rm manufactures radio models A, B and C, which have
prot contributions of 16, 30 and 50 respectively. Minimum production requirements are 20, 120
and 60 for the three models, respectively.
A dozen units of Model A requires 3 hours for manufacturing of components, 4 for assembling,
and 1 for packaging. The corresponding gures for a dozen units of Model B are 3.5, 5, and 1.5,
and for Model C are 5, 8, and 3. During the forthcoming week the company has available 120
plant-hours for manufacturing components, 160 for assembly, and 48 for packaging. Formulate this
production planning problem as a linear programming problem.
Exercise 57 (Wagner) Process 1 in an oil renery takes a unit feed of 1 bbl (barrel) of crude oil
A and 3 bbl of crude B to make 50 gallons (gal) of gasoline X and 20 gal of gasoline Y. From a
unit feed of 4 bbl crude A and 2 bbl crude B, process 2 makes 30 gal of X and 80 gal of Y. Let xi
be the number of units of feed type i; e.g., x1 = 1 indicates that process 1 uses 1 bbl crude A and
3 bbl crude B.
There are 120 bbl of crude A available and 180 bbl of crude B. Sales commitments call for at
least 2800 gal of gasoline X and 2200 gal of Y. The unit prots of process 1 and 2 are p1 and p2,
respectively. Formulate an LP model.
Exercise 58 (Wagner) An air cargo rm has 8 aircraft of type 1, 15 of type 2 and 11 of type 3
available for today's ights. A type 1 craft can carry 45 tons, type 2, 7 tons and type 3, 5 tons. 20
tons of cargo are to be own to city A and 28 tons to city B. Each plane makes at most one ight
a day.
The costs of ying a plane from the terminal to each city are as follows.
Type 1 Type 2 Type 3
City A 23 15 1.4
City B 58 20 3.8
Let xi be the number of type i planes sent to A and yi the number to B. Formulate an LP for this
routing problem.
Exercise 59 (Wagner) A manufacturing rm produces widgets and distributes them to ve whole-
salers at a xed delivered price of $2.50 per unit. Sales forecasts indicate that monthly deliveries
will be 2700, 2700, 9000, 4500 and 3600 widgets to wholesalers 1-5 respectively. The monthly
production capacities are 4500, 9000 and 11,250 at plants 1, 2 and 3, respectively. The direct costs
of producing each widget are $2 at plant 1, $1 at plant 2 and $1.80 at plant 3.
The transport cost of shipping a widget from a plant to a wholesaler is given below.
Wholesaler 1 2 3 4 5
Plant 1 .05 .07 .11 .15 .16
Plant 2 .08 .06 .10 .12 .15
Plant 3 .10 .09 .09 .10 .16
Formulate an LP model for this production and distribution problem.
Exercise 60 (Wagner) Ft. Loudoun and Watts Bar are two large hydroelectric dams, the former
upstream of the latter. The level of Watts Bar Lake must be kept within limits for recreational
purposes, and the problem is to plan releases from Ft. Loudoun to do so. In reality this problem
is solved simultaneously for numerous dams covering an entire watershed, but we focus on a single
reservoir. There are also sophisticated models for predicting runo into the reservoirs, but we
will suppose that runo is negligible. Thus any water entering Watts Bar Lake must be released
through Ft. Loudoun Dam.
The planning period is 20 months. During month t, let xt be the average water level of Watts
Bar Lake before augmentation by water from Ft. Loudoun; x1 = 25. Let yt be the number of
feet added to the average level in month t from Ft. Loudoun. Lt and Ut are the lower and upper
bounds on the lake level in month t (more restrictive in summer). To model seepage, evaporation
and hydroelectric release through Watts Bar Dam we suppose that Watts Bar Lake begins month
t +1 at a level equal to .75 times the average level of the previous month (including the augmentation
from Ft. Loudoun). The cost of water from Ft. Loudoun Lake is ct for every foot added to the
level of Watts Bar. Formulate the appropriate LP model. (In reality the model is a huge nonlinear
program.)
Exercise 61 Red Dwarf Toasters needs to produce 1000 of their new \Talking Toaster". There
are three ways this toaster can be produced: manually, semi-automatically, and robotically. Man-
ual assembly requires 1 minute of skilled labor, 40 minutes of unskilled labor, and 3 minutes of
assembly room time. The corresponding values for semiautomatic assembly are 4, 30, and 2; while
those for robotic assembly are 8, 20, and 4. There are 4500 minutes of skilled labor, 36,000 min-
utes of unskilled labor, and 2700 minutes of assembly room time available for this product. The
total cost for producing manually is $7/toaster; semiautomatically is $8/toaster; and robotically is
$8.50/toaster.
(a) Formulate the problem of producing 1000 toasters at minimum cost meeting the resource
requirements. Clearly dene your variables, objective and constraints.
(b) Our union contract states that the amount of skilled labor time used is at least 10% of
the total labor (unskilled plus skilled) time used. Update your formulation in (a) to handle this
requirement.
(c) Any unused assembly oor time can be rented out at a prot of $0.50/minute. Update your
formulation to include this possibility.
Answers to Exercise 61:
(a) Let x1 be the number of toasters produced manually, x2 be the number produced semiauto-
matically, and x3 be the number produced robotically.
The objective is to Minimize 7x1 + 8x2 + 8:5x3.
The constraints are:
x1 + x2 + x3 = 1000 (produce enough toasters)
x1 + 4x2 + 8x3 4500 (skilled labor used less than or equal to amount available).
40x1 + 30x2 + 20x3 36000 (unskilled labor constraint)
3x1 + 2x2 + 4x3 2700 (assembly time constraint)
x1 0; x2 0; x3 0 (nonnegativity of production)
(b) Add a constraint x1 + 4x2 + 8x3 :1(41x1 + 34x2 + 28x3)
(c) Add a variable sa to represent the assembly time slack. Add +0:5sa to the objective. Change
the assembly time constraint to
3x1 + 2x2 + 4x3 + sa = 2700 (assembly time constraint)
sa 0
Chapter 6
Case Study
The Global Oil Company
You may complete the project individually or in groups. In the latter case, which is encouraged,
members of the group submit a joint report and all receive the same grade. The ideal size of a
group is three. Groups of up to four are allowed. There should be no collaboration among groups
and/or students working individually.
The Global Oil Company is an international producer, rener, transporter and distributor of oil,
gasoline and petrochemicals. Global Oil is a holding company with subsidiary operating companies
that are wholly or partially owned. A major problem for Global Oil is to coordinate the actions
of these various subsidiaries into an overall corporate plan, while at the same time maintaining a
reasonable amount of operating autonomy for the subsidiary companies.
To deal with this dilemma, the logistics department at Global Oil Headquarters develops an
annual corporate-wide plan which details the pattern of shipments among the various subsidiaries.
The plan is not rigid but provides general guidelines and the plan is revised periodically to re ect
changing conditions. Within the framework of this plan, the operating companies can make their
own decisions and plans. This corporate-wide plan is presently done on a trial and error basis. There
are two problems with this approach. First, the management of the subsidiaries often complains
that the plan does not re ect properly the operating conditions under which the subsidiary operates.
The plan sometimes calls for operations or distribution plans that are impossible to accomplish.
And secondly, corporate management is concerned that the plan does not optimize for the total
company.
The technique of linear programming seems a possible approach to aid in the annual planning
process, that will be able to answer at least in part, the two objections above. In addition the
building of such a model will make it possible to make changes in plans quickly when the need
arises. Before embarking on the development of a world-wide model, Global Oil asks you to build
a model of the Far Eastern operations for the coming year.
Far Eastern Operations
The details of the 1998 planning model for the Far Eastern Operations are now described.
There are two sources of crude oil, Saudi Arabia and Borneo. The Saudi crude is relatively
heavier (24 API), and the Far Eastern sector could obtain as much as 60,000 barrels per day at a
79
80 CHAPTER 6. CASE STUDY
cost of $18.50 per barrel during 1998. A second source of crude is from the Brunei elds in Borneo.
This is a light crude oil (36 API). Under the terms of an agreement with the Netherlands Petroleum
Company in Borneo, a xed quantity of 40,000 b/d of Brunei crude, at a cost of $19.90 per barrel
is to be supplied during 1998.
There are two subsidiaries that have rening operations. The rst is in Australia, operating a
renery in Sydney with a capacity of 50,000 b/d throughput. The company markets its products
throughout Australia, as well as having a surplus of rened products available for shipment to other
subsidiaries.
The second subsidiary is in Japan, which operates a 30,000 b/d capacity renery. Marketing
operations are conducted in Japan, and excess production is available for shipment to other Far
Eastern subsidiaries.
In addition, there are two marketing subsidiaries without rening capacity of their own. One
of these is in New Zealand and the other is in the Philippines. Their needs can be supplied by
shipments from Australia, Japan, or the Global Oil subsidiary in the United States. The latter is
not a regular part of the Far Eastern Operations, but may be used as a source of rened products.
Finally, the company has a eet of tankers that move the crude oil and rened products among
the subsidiaries.
Renery Operations
The operation of a renery is a complex process. The characteristics of the crudes available,
the desired output, the specic technology of the renery, etc., make it dicult to use a simple
model to describe the process. In fact, management at both Australia and Japan have complex
linear programming models involving approximately 300 variables and 100 constraints for making
detailed decisions on a daily or weekly basis.
For annual planning purposes the renery model is greatly simplied. The two crudes (Saudi
and Brunei) are input. Two general products are output { (a) gasoline products and (b) other
products such as distillate, fuel oil, etc. In addition, although the renery has processing exibility
that permits a wide range of yields, for planning purposes it was decided to include only the values
at highest and lowest conversion rates (process intensity). Each renery could use any combination
of the two extreme intensities. These yields are shown in Table 6.1.
The incremental costs of operating the renery depend somewhat upon the type of crude and
process intensity. These costs are shown in Table 6.1. Also shown are the incremental transportation
costs from either Borneo or Saudi Arabia.
Marketing Operations
Marketing is conducted in two home areas (Australia and Japan) as well as in the Philippines
and New Zealand. Demand for gasoline and distillate in all areas has been estimated for 1998.
Area 1998 Demand (thous. of b/d)
Gasoline Distillate
Australia 9.0 21.0
Japan 3.0 12.0
Philippines 5.0 8.0
New Zealand 5.4 8.7
TOTAL 22.4 49.7
81
Table 6.1: Renery Operations

Australia Japan
Capacity (b/d of input) 50,000 30,000
Saudi Crude
Transportation Cost ($/b) 0.65 0.70
High Process Intensity ($/b) 1.19 1.26
Yield of Gasoline 0.31 0.30
Yield of Distillate 0.61 0.62
Low Process Intensity ($/b) 0.89 0.88
Brunei Crude
Transportation Cost ($/b) 0.15 0.25
High Process Intensity ($/b) 0.93 0.91
Low Process Intensity ($/b) 0.61 0.55
Variable costs of supplying gasoline or distillate to New Zealand and the Philippines are:
Variable costs of shipment of
gasoline/distillate in $/b
From: To: New Zealand Philippines
Australia .20 .30
Japan .25 .40
Tanker Operations
Tankers are used to bring crude from Saudi Arabia and Borneo to Australia and Japan and
to transport rened products from Australia and Japan to the Philippines and New Zealand. The
variable costs of these operations are included above.
However, there is a limited capacity of tankers available. The eet has a capacity of 6.5 equiv-
alent (standard sized) tankers.
The amount of capacity needed to deliver one barrel from one destination to another depends
upon the distance traveled, port time, and other factors. The table below lists the fraction of one
standard sized tanker needed to deliver 1,000 b/d over the indicated routes.
Tanker Usage Factors
(Fraction of Standard Sized Tanker
Needed to Deliver 1,000 b/d)
Between and Australia Japan
Saudi Arabia .12 .11
Borneo .05 .05
Philippines .02 .01
New Zealand .01 .06
It is also possible to charter independent tankers. The rate for this is $5,400 per day for a
standard sized tanker.
United States Supply
United States operations on the West Coast expect a surplus of 12,000 b/d of distillate during
1998. The cost of distillate at the loading port of Los Angeles is $20.70 per barrel. There is no
excess gasoline capacity. The estimated variable shipping costs and tanker requirements of distillate
shipments from the United States are:
Variable costs Tanker
of shipments requirements
New Zealand 1.40 .18
Philippines 1.10 .15
83
Questions
Part I: Formulate a linear program which can be used to generate a comprehensive plan for the whole
Far Eastern operations. Clearly dene every variable used in your formulation.
Each group should be prepared to make a presentation of their linear programming model in
class, on September 22. No written report is due on this date.
Part II: Written report due on October 6. Start your report with an executive summary (at most
two pages) containing the most important results and recommendations. Attach supporting
material in Appendices.
1. Solve your linear program using SOLVER.
How many barrels of crude should Global Oil purchase from Saudi Arabia for its Far
Eastern operations? How much crude should be rened in Australia? How much in
Japan? Provide tables showing the quantities of gasoline and distillate shipped from
each of the two reneries and from the U.S. to each of the four market areas.
2. Use sensitivity analysis to answer the following questions.
{ What is the marginal value of increasing supply from Brunei elds? Can this
marginal value be used to estimate the total savings for Far Eastern operations
if 41,000 b/d are supplied from Brunei elds instead of 40,000? Explain.
{ What is the marginal value of increasing the tanker eet? Can this marginal value
be used if we want to increase the size to 7 tankers (from the current size of 6.5)?
Explain.
{ What is the additional cost to Far Eastern operations if demand for gasoline in the
Philippines increases to 5,200 b/d? What is the minimum price of gasoline in the
Philippines that would make it protable for Global Oil to consider such an increase
in distribution?
{ By how much should the production costs be reduced at the renery in Japan when
operating at high process intensity in order to make it cost eective to use Saudi
crude?
{ Global Oil is planning a three day shutdown of its Australian renery for mainte-
nance purposes, in the coming year. It has storage facilities and at least two weeks
of inventories, so a renery shutdown for a few days will not disrupt distribution.
What would be the cost of a planned shutdown of the renery in Australia for three
days per year? Same question for the renery in Japan.
{ Currently, it is not economical to ship US distillate to the Philippines. What is the
cost of US distillate at which Global Oil should consider starting such shipments?
3. Several opportunities present themselves to the Global Oil company (see the attached
memos). Consider combinations of these options and prepare a recommendation. Doc-
ument your report.
Memo to: Global Oil Headquarters
From: Australian Aliate
Re: Supplements to Annual Plan
Since submitting data for annual planning purposes, two additional opportunities have arisen.
We would like to include these in the plans.
A. Bid on Gasoline Contract with Australian Government
The government of Australia will submit to bid a contract for 1.5 thousand b/d of gasoline.
We expect we could win this bid at a price of $26.40 per barrel. Estimated costs per barrel
as follows:
Variable Costs
(crude, rening, transportation) $24.90
Allocated Overhead .80
Total $25.70
At these costs, the contract would have a prot contribution of $0.70 per barrel. We would
like permission to bid on the contract. We hope that the linear programming wizards who
are working for your logistics department will not contradict us!
B. Expansion of Australian Renery
For the past two years, the Australian renery has been operating at full capacity. We request
authorization for capital expenditures to increase the renery capacity to 55 thousand b/d.
There are several reasons for the need for this expansion:
1. Australia can supply the current requirements in New Zealand and the Philippines more
cheaply than can Japan.
2. The proposed bid on Australian government gasoline contract (above).
3. We understand the New Zealand aliate is considering increasing its requirements by
4.5 thousand b/d. (See below.)
The cost of this expansion is 4.0 million dollars. To recover this investment1, we need an
annual savings of $702,000.
This assumes a cost of capital rate of 20%. Depreciation tax eects are included. With these considerations, the
1
$702,000 savings per year is equivalent to the $4 million investment

85

From: New Zealand Aliate
Re: Supplement to Annual Plan
Negotiations have been begun with the NOZO Oil Company in New Zealand. This company
is a distributor, with sales of 1.6 thousand b/d of gasoline and 3.2 thousand b/d of distillate. If
negotiations are successfully completed, these requirements would be added to current requirements
for New Zealand, making total requirements of:
Gasoline: 7.0 thousand b/d
Distillate: 11.9 thousand b/d
The anticipated revenue (after subtracting variable marketing costs) for this acquisition are
$30.20 per barrel for gasoline and $24.60 per barrel for distillate. The purchase cost of NOZO oil
is expected to be about $21.0 million. On an annual basis, this would require $3.5 million per year
incremental prot to justify the purchase.

From: Tanker Aliate
We have been made aware that expansions in requirements are being considered in New Zealand
and Australia. We are currently operating the tanker eet at capacity. Additional requirements
will increase the transport requirements both for crude and rened products. This will necessitate
spot chartering unless additional tanker capacity is added.
We can lease on a long time basis additional tankers at a rate of $4.8 thousand dollars per day
per 1 unit tanker equivalent. We propose a lease of 0.5 equivalent tanker units giving us a total
capacity of 7.0 equivalent units. The cost of this would be $2.4 thousand per day or $876,000 per
year. If you prefer other arrangements, we are willing to discuss them.

From: Borneo Oce
We have just been oered the opportunity to increase our contract with our supplier of Brunei
crude. They are willing to supply us an additional 5 thousand b/d at a cost of $20.65 per barrel.
Should we accept the oer? If not, is there a counter oer that we should propose?
Chapter 7
The Simplex Method

In this chapter, you will learn how to solve linear programs. This will give you insights into
what SOLVER and other commercial linear programming software packages actually do. Such an
understanding can be useful in several ways. For example, you will be able to identify when a
problem has alternate optimal solutions (SOLVER never tells you this: it always give you only one
optimal solution). You will also learn about degeneracy in linear programming and how this could
lead to a very large number of iterations when trying to solve the problem.
7.1 Linear Programs in Standard Form

Before we start discussing the simplex method, we point out that every linear program can be
converted into \standard" form
Max c1x1 + c2x2 + : : : + cn xn

subject to a11x1 + a12x2 + : : : + a1n xn = b1
::: ::: :::
am1 x1 + am2x2 + : : : + amn xn = bm
x1 0; : : :xn 0
where the objective is maximized, the constraints are equalities and the variables are all nonnega-
tive.
This is done as follows:
If the problem is min z, convert it to max ,z.
If a constraint is ai1x1 +ai2x2 +: : :+ain xn bi, convert it into an equality constraint by adding
a nonnegative slack variable si . The resulting constraint is ai1 x1 + ai2 x2 + : : : + ain xn + si = bi,
where si 0.
If a constraint is ai1x1 + ai2x2 + : : : + ain xn bi, convert it into an equality constraint by
subtracting a nonnegative surplus variable si . The resulting constraint is ai1x1 + ai2 x2 + : : : +
ain xn , si = bi , where si 0.
If some variable xj is unrestricted in sign, replace it everywhere in the formulation by x0j , x00j ,
where x0j 0 and x00j 0.
87
88 CHAPTER 7. THE SIMPLEX METHOD
Example 7.1.1 Transform the following linear program into standard form.
Min ,2x1 +3x2
x1 ,3x2 +2x3 3
,x1 +2x2 2
x1 urs; x2 0; x3 0
Let us rst turn the objective into a max and the constraints into equalities.
Max 2x1 ,3x2
x1 ,3x2 +2x3 +s1 = 3
,x1 +2x2 ,s2 = 2
x1 urs; x2 0; x3 0 s1 0 s2 0
The last step is to convert the unrestricted variable x1 into two nonnegative variables: x1 =
x01 , x001 .
Max 2x01 ,2x001 ,3x2
x01 ,x001 ,3x2 +2x3 +s1 = 3
,x0 1 00
+x1 +2x2 ,s2 = 2
x01 0; x001 0; x2 0; x3 0 s1 0 s2 0
7.2 Solution of Linear Programs by the Simplex Method
For simplicity, in this course we solve \by hand" only the case where the constraints are of the
form and the right-hand-sides are nonnegative. We will explain the steps of the simplex method
while we progress through an example.
Example 7.2.1 Solve the linear program
max x1 +x2
2x1 +x2 4
x1 +2x2 3
x1 0; x2 0
First, we convert the problem into standard form by adding slack variables x3 0 and x4 0.
max x1 +x2
2x1 +x2 +x3 = 4
x1 +2x2 +x4 = 3
x1 0; x2 0 x3 0; x4 0
Let z denote the objective function value. Here, z = x1 + x2 or, equivalently,
z , x1 , x2 = 0:
Putting this equation together with the constraints, we get the following system of linear equa-
tions.
z ,x1 ,x2 = 0 Row 0
2x1 +x2 +x3 = 4 Row 1 (7.1)
x1 +2x2 +x4 = 3 Row 2
7.2. SOLUTION OF LINEAR PROGRAMS BY THE SIMPLEX METHOD 89
Our goal is to maximize z , while satisfying these equations and, in addition, x1 0, x2 0,

x3 0, x4 0.
Note that the equations are already in the form that we expect at the last step of the Gauss-
Jordan procedure. Namely, the equations are solved in terms of the nonbasic variables x1 , x2.
The variables (other than the special variable z ) which appear in only one equation are the basic
variables. Here the basic variables are x3 and x4 . A basic solution is obtained from the system of
equations by setting the nonbasic variables to zero. Here this yields
x1 = x2 = 0 x3 = 4 x4 = 3 z = 0:
Is this an optimal solution or can we increase z ? (Our goal)
By looking at Row 0 above, we see that we can increase z by increasing x1 or x2 . This is
because these variables have a negative coecient in Row 0. If all coecients in Row 0 had been
nonnegative, we could have concluded that the current basic solution is optimum, since there would
be no way to increase z (remember that all variables xi must remain 0). We have just discovered
the rst rule of the simplex method.
Rule 1 If all variables have a nonnegative coecient in Row 0, the current basic solution is
optimal.
Otherwise, pick a variable xj with a negative coecient in Row 0.
The variable chosen by Rule 1 is called the entering variable. Here let us choose, say, x1 as our
entering variable. It really does not matter which variable we choose as long as it has a negative
coecient in Row 0. The idea is to pivot in order to make the nonbasic variable x1 become a basic
variable. In the process, some basic variable will become nonbasic (the leaving variable). This
change of basis is done using the Gauss-Jordan procedure. What is needed next is to choose the
pivot element. It will be found using Rule 2 of the simplex method. In order to better understand
the rationale behing this second rule, let us try both possible pivots and see why only one is
acceptable.
First, try the pivot element in Row 1.
z ,x1 ,x2 = 0 Row 0
2x1 +x2 +x3 = 4 Row 1
x1 +2x2 +x4 = 3 Row 2
This yields
z , 12 x2 + 13 x3 = 2 Row 0
x1 + 21 x2 + 12 x3 = 2 Row 1
3 1
2 x2 , 2 x3 +x4 = 1 Row 2
with basic solution x2 = x3 = 0 x1 = 2 x4 = 1 z = 2:
Now, try the pivot element in Row 2.
z ,x1 ,x2 = 0 Row 0
2x1 +x2 +x3 = 4 Row 1
x1 +2x2 +x4 = 3 Row 2
This yields
z +x2 +x4 = 3 Row 0
,3x2 +x3 ,2x4 = ,2 Row 1
x1 +2x2 +x4 = 3 Row 2
with basic solution x2 = x4 = 0 x1 = 3 x3 = ,2 z = 3:
Which pivot should we choose? The rst one, of course, since the second yields an infeasible
basic solution! Indeed, remember that we must keep all variables 0. With the second pivot, we
get x3 = ,2 which is infeasible. How could we have known this ahead of time, before actually
performing the pivots? The answer is, by comparing the ratios Entering Right Hand Side
Variable Coecient in
Rows 1 and 2 of (7.1). Here we get 42 in Row 1 and 31 in Row 2. If you pivot in a row with minimum
ratio, you will end up with a feasible basic solution (i.e. you will not introduce negative entries in
the Right Hand Side), whereas if you pivot in a row with a ratio which is not minimum you will
always end up with an infeasible basic solution. Just simple algebra! A negative pivot element
would not be good either, for the same reason. We have just discovered the second rule of the
simplex method.
Rule 2 For each Row i, i 1, where there is a strictly positive \entering variable coecient",
compute the ratio of the Right Hand Side to the \entering variable coecient". Choose the pivot
row as being the one with MINIMUM ratio.
Once you have idendied the pivot element by Rule 2, you perform a Gauss-Jordan pivot. This
gives you a new basic solution. Is it an optimal solution? This question is addressed by Rule 1, so
we have closed the loop. The simplex method iterates between Rules 1, 2 and pivoting until Rule
1 guarantees that the current basic solution is optimal. That's all there is to the simplex method.
After our rst pivot, we obtained the following system of equations.
z , 21 x2 + 13 x3 = 2 Row 0
x1 + 21 x2 + 12 x3 = 2 Row 1
3 x2 , 1 x3 +x4 = 1 Row 2
2 2
with basic solution x2 = x3 = 0 x1 = 2 x4 = 1 z = 2:
Is this solution optimal? No, Rule 1 tells us to choose x2 as entering variable. Where should
we pivot? Rule 2 tells us to pivot in Row 2, since the ratios are 1=22 = 4 for Row 1, and 31=2 = 23 for
Row 2, and the minimum occurs in Row 2. So we pivot on 23 x2 in the above system of equations.
This yields
z + 13 x3 + 31 x4 = 73 Row 0
x1 + 32 x3 , 31 x4 = 53 Row 1
x2 , 31 x3 + 32 x4 = 23 Row 2
with basic solution x3 = x4 = 0 x1 = 35 x2 = 23 z = 73 :
Now Rule 1 tells us that this basic solution is optimal, since there are no more negative entries
in Row 0.
All the above computations can be represented very compactly in tableau form.
7.2. SOLUTION OF LINEAR PROGRAMS BY THE SIMPLEX METHOD 91
z x1 x2 x3 x4 RHS Basic solution
1 ,1 ,1 0 0 0 basic x3 = 4 x4 = 3
0 2 1 1 0 4 nonbasic x1 = x2 = 0
0 1 2 0 1 3 z=0
1 0 , 12 12 0 2 basic x1 = 2 x4 = 1
0 1 12 12 0 2 nonbasic x2 = x3 = 0
0 0 32 , 12 1 1 z=2
1 0 0 13 13 7
35 basic x1 = 53 x2 = 23
0 1 0 23 , 13 32 nonbasic x3 = x4 = 0
0 0 1 , 13 23 3 z = 37
Since the above example has only two variables, it is interesting to interpret the steps of the
simplex method graphically. See Figure 7.1. The simplex method starts in the corner point (x1 =
0; x2 = 0) with z = 0. Then it discovers that z can increase by increasing, say, x1. Since we keep
x2 = 0, this means we move along the x1 axis. How far can we go? Only until we hit a constraint:
if we went any further, the solution would become infeasible. That's exactly what Rule 2 of the
simplex method does: the minimum ratio rule identies the rst constraint that will be encountered.
And when the constraint is reached, its slack x3 becomes zero. So, after the rst pivot, we are at
the point (x1 = 2; x2 = 0). Rule 1 discovers that z can be increased by increasing x2 while keeping
x3 = 0. This means that we move along the boundary of the feasible region 2x1 + x2 = 4 until we
reach another constraint! After pivoting, we reach the optimal point (x1 = 53 ; x2 = 23 ).
x
2
z=7/3
x
z=0 z=2 1
Figure 7.1: Graphical Interpretation

Exercise 62 Solve the following linear program by the simplex method.
max 4x1 +x2 ,x3
x1 +3x3 6
3x1 +x2 +3x3 9
x1 0; x2 0 x3 0
7.3 Alternate Optimal Solutions, Degeneracy, Unboudedness, In-
feasibility
Alternate Optimal Solutions
Let us solve a small variation of the earlier example, with the same constraints but a slightly
dierent objective:
max x1 + 21 x2
2x1 +x2 4
x1 +2x2 3
x1 0; x2 0
As before, we add slacks x3 and x4, and we solve by the simplex method, using tableau repre-
sentation.
1 ,1 , 12 0 0 0 basic x3 = 4 x4 = 3
0 2 1 1 0 4 nonbasic x1 = x2 = 0
0 1 2 0 1 3 z=0
1 0 0 2 0 2 1 basic x1 = 2 x4 = 1
0 1 12 12 0 2 nonbasic x2 = x3 = 0
0 0 32 , 12 1 1 z=2
Now Rule 1 shows that this is an optimal solution. Interestingly, the coecient of the nonbasic
variable x2 in Row 0 happens to be equal to 0. Going back to the rationale that allowed us to
derive Rule 1, we observe that, if we increase x2 (from its current value of 0), this will not eect the
value of z . Increasing x2 produces changes in the other variables, of course, through the equations
in Rows 1 and 2. In fact, we can use Rule 2 and pivot to get a dierent basic solution with the
same objective value z = 2.
1 0 0 21 0 2 basic x1 = 53 x2 = 32
0 1 0 32 , 13 5
32 nonbasic x3 = x4 = 0
0 0 1 , 13 2
3 3 z=2
Note that the coecient of the nonbasic variable x4 in Row 0 is equal to 0. Using x4 as entering
variable and pivoting, we would recover the previous solution!
7.3. ALTERNATE OPTIMAL SOLUTIONS, DEGENERACY, UNBOUDEDNESS, INFEASIBILITY93
Degeneracy
Example 7.3.1
max 2x1 +x2
3x1 +x2 6
x1 ,x2 2
x2 3
x1 0; x2 0
Let us solve this problem using the {by now familiar{ simplex method. In the initial tableau,
we can choose x1 as the entering variable (Rule 1) and Row 2 as the pivot row (the minimum ratio
in Rule 2 is a tie, and ties are broken arbitrarily). We pivot and this yields the second tableau
below.
z x1 x2 x3 x4 x5 RHS Basic solution
1 ,2 ,1 0 0 0 0 basic x3 = 6 x4 = 2 x5 = 3
0 3 1 1 0 0 6 nonbasic x1 = x2 = 0
0 1 ,1 0 1 0 2 z=0
0 0 1 0 0 1 3
1 0 ,3 0 2 0 4 basic x1 = 2 x3 = 0 x5 = 3
0 0 4 1 ,3 0 0 nonbasic x2 = x4 = 0
0 1 ,1 0 1 0 2 z=4
0 0 1 0 0 1 3
Note that this basic solution has a basic variable (namely x3) which is equal to zero. When this
occurs, we say that the basic solution is degenerate. Should this be of concern? Let us continue the
steps of the simplex method. Rule 1 indicates that x2 is the entering variable. Now let us apply
Rule 2. The ratios to consider are 04 in Row 1 and 13 in Row 3. The minimum ratio occurs in Row
1, so let us perform the corresponding pivot.
1 0 0 4 ,4 0 4 3 1 basic x1 = 2 x2 = 0 x5 = 3
0 0 1 41 , 34 0 0 nonbasic x3 = x4 = 0
0 1 0 41 14 0 2 z=4
0 0 0 , 14 34 1 3
We get exactly the same solution! The only dierence is that we have interchanged the names
of a nonbasic variable with that of a degenerate basic variable (x2 and x3 ). Rule 1 tells us the
solution is not optimal, so let us continue the steps of the simplex method. Variable x4 is the
entering variable and the last row wins the minimum ratio test. After pivoting, we get the tableau:
1 0 0 3 0 3 5 2 1 basic x1 = 1 x2 = 3 x4 = 4
0 0 1 0 0 1 3 nonbasic x3 = x5 = 0
0 1 0 31 0 , 13 1 z=5
0 0 0 , 13 1 43 4
By Rule 1, this is the optimal solution. So, after all, degeneracy did not prevent the simplex
method to nd the optimal solution in this example. It just slowed things down a little. Unfor-
tunately, on other examples, degeneracy may lead to cycling, i.e. a sequence of pivots that goes
through the same tableaus and repeats itself indenitely. In theory, cycling can be avoided by
choosing the entering variable with smallest index in Rule 1, among all those with a negative coef-
cient in Row 0, and by breaking ties in the minimum ratio test by choosing the leaving variable
with smallest index (this is known as Bland's rule). This rule, although it guaranties that cycling
will never occur, turns out to be somewhat inecient. Actually, in commercial codes, no eort is
made to avoid cycling. This may come as a surprise, since degeneracy is a frequent occurence. But
there are two reasons for this:
Although degeneracy is frequent, cycling is extremely rare.
The precision of computer arithmetic takes care of cycling by itself: round o errors accumu-
late and eventually gets the method out of cycling.
Our example of degeneracy is a 2-variable problem, so you might want to draw the constraint
set in the plane and interpret degeneracy graphically.
Unbounded Optimum
Example 7.3.2
max 2x1 +x2
,x 1 +x2 1
x1 ,2x2 2
x1 0; x2 0
Solving by the simplex method, we get:
1 ,2 ,1 0 0 0 basic x3 = 1 x4 = 2
0 ,1 1 1 0 1 nonbasic x1 = x2 = 0
0 1 ,2 0 1 2 z=0
1 0 ,5 0 2 4 basic x1 = 2 x3 = 3
0 0 ,1 1 1 3 nonbasic x2 = x4 = 0
0 1 ,2 0 1 2 z=4
At this stage, Rule 1 chooses x2 as the entering variable, but there is no ratio to compute, since
there is no positive entry in the column of x2 . As we start increasing x2 , the value of z increases
(from Row 0) and the values of the basic variables increase as well (from Rows 1 and 2). There is
nothing to stop them going o to innity. So the problem is unbounded.
Interpret the steps of the simplex method graphically for this example.
Infeasible Linear Programs
It is easy to construct constraints that have no solution. The simplex method is able to identify
such cases. We do not discuss it here. The interested reader is referred to Winston, for example.
Properties of Linear Programs
There are three possible outcomes for a linear program: it is infeasible, it has an unbounded
optimum or it has an optimal solution.
7.3. ALTERNATE OPTIMAL SOLUTIONS, DEGENERACY, UNBOUDEDNESS, INFEASIBILITY95
If there is an optimal solution, there is a basic optimal solution. Remember that the number of
basic variables in a basic solution is equal to the number of constraints of the problem, say m. So,
even if the total number of variables, say n, is greater than m, at most m of these variables can
have a positive value in an optimal basic solution.
Exercise 63 The following tableaus were obtained in the course of solving linear programs with 2
nonnegative variables x1 and x2 and 2 inequality constraints (the objective function z is maximized).
Slack variables s1 and s2 were added. In each case, indicate whether the linear program
(i) is unbounded
(ii) has a unique optimum solution
(iii) has an alternate optimum solution
(iv) is degenerate (in this case, indicate whether any of the above holds).
z x1 x2 s1 s2 RHS
(a) 10 0 3 2 0
1 ,2 ,1 0
20
4
0 0 ,1 0 1 2
z x1 x2 s1 s2 RHS
(b) 10 00 ,10 01 ,22 20
5
0 1 ,2 0 3 6
z x1 x2 s1 s2 RHS
(c) 10 23 0 0 1
1 0 ,2
8
4
0 ,2 0 1 1 0
z x1 x2 s1 s2 RHS
(d) 10 0 0 2 0
0 ,1 1 1
5
4
0 1 1 ,1 0 4
Exercise 64 Suppose the following tableau was obtained in the course of solving a linear program
with nonnegative variables x1; x2; x3 and two inequalities. The objective function is maximized and
slack variables s1 and s2 were added.
z x1 x2 x3 s1 s2 RHS
1 0 a b 0 4 82
0 0 ,2 2 1 3 c
0 1 ,1 3 0 ,5 3
Give conditions on a, b and c that are required for the following statements to be true:
(i) The current basic solution is a feasible basic solution.
Assume that the condition found in (i) holds in the rest of the exercise.
(ii) The current basic solution is optimal.
(iii) The linear program is unbounded (for this question, assume that b > 0).
(iv) The currrent basic solution is optimal and there are alternate optimal solutions (for this
question, assume a > 0).
Exercise 65 A plant can manufacture ve products P1, P2, P3, P4 and P5. The plant consists
of two work areas: the job shop area A1 and the assembly area A2 . The time required to process
one unit of product Pj in work area Ai is pij (in hours), for i = 1; 2 and j = 1; : : :; 5. The weekly
capacity of work area Ai is Ci (in hours). The company can sell all it produces of product Pj at a
prot of sj , for i = 1; : : :; 5.
The plant manager thought of writing a linear program to maximize prots, but never actually
did for the following reason: From past experience, he observed that the plant operates best when
at most two products are manufactured at a time. He believes that if he uses linear programming,
the optimal solution will consist of producing all ve products. Do you agree with him? Explain,
based on your knowledge of linear programming.
Exercise 66 Consider the linear program
Maximize 5x1 + 3x2 + x3
Subject to
x1 + x2 + x3 6
5x1 + 3x2 + 6x3 15
x1 ; x2; x3 0
and an associated tableau
z x1 x2 x3 s1 s2 RHS
1 0 0 5 0 1 15
0 0 .4 -0.2 1 -.2 3
0 1 .6 1.2 0 .2 3
(a) What basic solution does this tableau represent? Is this solution optimal? Why or why not?
(b) Does this tableau represent a unique optimum. If not, nd an alternative optimal solution.
Answer:
(a) The solution is x1 = 3, x2 ; x3 = 0, objective 15. It is optimal since cost row is all at least 0.
(b) It is not unique (since x2 has reduced cost 0 but is not basic). Alternative found by pivoting
in x2 (question could have asked for details on such a pivot) for solution x2 = 5, x1 ; x3 = 0,
objective 15.
Chapter 8
Sensitivity Analysis for Linear

Programming
Finding the optimal solution to a linear programming model is important, but it is not the only
information available. There is a tremendous amount of sensitivity information, or information
about what happens when data values are changed.
Recall that in order to formulate a problem as a linear program, we had to invoke a certainty
assumption: we had to know what value the data took on, and we made decisions based on that
data. Often this assumption is somewhat dubious: the data might be unknown, or guessed at,
or otherwise inaccurate. How can we determine the eect on the optimal decisions if the values
change? Clearly some numbers in the data are more important than others. Can we nd the
\important" numbers? Can we determine the eect of misestimation?
Linear programming oers extensive capability for addressing these questions. We begin by
showing how data changes show up in the optimal table. We then give two examples of how to
interpret Solver's extensive output.
8.1 Tableau Sensitivity Analysis

Suppose we solve a linear program \by hand" ending up with an optimal table (or tableau to use
the technical term). We know what an optimal tableau looks like: it has all non{negative values
in Row 0 (which we will often refer to as the cost row), all non{negative right{hand{side values,
and a basis (identity matrix) embedded. To determine the eect of a change in the data, we will
try to determine how that change eected the nal tableau, and try to reform the nal tableau
accordingly.
8.1.1 Cost Changes
The rst change we will consider is changing a cost value by in the original problem. We are
given the original problem and an optimal tableau. If we had done exactly the same calculations
beginning with the modied problem, we would have had the same nal tableau except that the
corresponding cost entry would be lower (this is because we never do anything except add or
subtract scalar multiples of Rows 1 through m to other rows; we never add or subtract Row 0 to
other rows). For example, take the problem
97
98 CHAPTER 8. SENSITIVITY ANALYSIS FOR LINEAR PROGRAMMING
Max 3x+2y
Subject to
x+y <= 4
2x+y <= 6
x,y >= 0
The optimal tableau to this problem (after adding s1 and s2 as slacks to place in standard form)
is:
z x y s1 s2 RHS
1 0 0 1 1 10
0 0 1 2 ,1 2
0 1 0 ,1 1 2
Suppose the cost for x is changed to 3+ in the original formulation, from its previous value 3.
After doing the same operations as before, that is the same pivots, we would end up with the
tableau:
z x y s1 s2 RHS
1 , 0 1 1 10
0 0 1 2 ,1 2
0 1 0 ,1 1 2
Now this is not the optimal tableau: it does not have a correct basis (look at the column of x).
But we can make it correct in form while keeping the same basic variables by adding times the
last row to the cost row. This gives the tableau:
z x y s1 s2 RHS
1 0 0 1, 1+ 10 + 2
0 0 1 2 ,1 2
0 1 0 ,1 1 2
Note that this tableau has the same basic variables and the same variable values (except for z )
that our previous solution had. Does this represent an optimal solution? It does only if the cost
row is all non{negative. This is true only if
1, 0
1+ 0
which holds for ,1 1. For any in that range, our previous basis (and variable values) is
optimal. The objective changes to 10 + 2.
In the previous example, we changed the cost of a basic variable. Let's go through another
example. This example will show what happens when the cost of a nonbasic variable changes.
Max 3x+2y + 2.5w
Subject to
x+y +2w <= 4
2x+y +2w <= 6
x,y,w >= 0
8.1. TABLEAU SENSITIVITY ANALYSIS 99
Here, the optimal tableau is :

z x y w s1 s2 RHS
1 0 0 1:5 1 1 10
0 0 1 2 2 ,1 2
0 1 0 0 ,1 1 2
Now suppose we change the cost on w from 2.5 to 2:5 + in the formulation. Doing the same
calculations as before will result in the tableau:
z x y w s1 s2 RHS
1 0 0 1:5 , 1 1 10
0 0 1 2 2 ,1 2
0 1 0 0 ,1 1 2
In this case, we already have a valid tableau. This will represent an optimal solution if 1:5 ,
0, so 1:5. As long as the objective coecient of w is no more than 2.5+1.5=4 in the original
formulation, our solution of x = 2; y = 2 will remain optimal.
The value in the cost row in the simplex tableau is called the reduced cost. It is zero for a basic
variable and, in an optimal tableau, it is non{negative for all other variables (for a maximization
problem).
Summary: Changing objective function values in the original formulation will result in a changed
cost row in the nal tableau. It might be necessary to add a multiple of a row to the cost row to keep
the form of the basis. The resulting analysis depends only on keeping the cost row non{negative.
8.1.2 Right Hand Side Changes
For these types of changes, we concentrate on maximization problems with all constraints. Other
cases are handled similarly.
Take the following problem:
Max 4x + 5y
Subject to (8.1)
2x + 3y <= 12
x + y <= 5
x; y >= 0
The optimal tableau, after adding slacks s1 and s2 is
z x y s1 s2 RHS
1 0 0 1 2 22
0 0 1 1 ,2 2
0 1 0 ,1 3 3
Now suppose instead of 12 units in the rst constraint, we only had 11. This is equivalent to
forcing s1 to take on value 1. Writing the constraints in the optimal tableau long{hand, we get
z + s1 + 2s2 = 22
y + s1 , 2s2 = 2
x , s1 + 3s2 = 3
If we force s1 to 1 and keep s2 at zero (as a nonbasic variable should be), the new solution
would be z = 21, y = 1, x = 4. Since all variables are nonnegative, this is the optimal solution.
In general, changing the amount of the right{hand{side from 12 to 12+ in the rst constraint
changes the tableau to:
z x y s1 s2 RHS
1 0 0 1 2 22 +
0 0 1 1 ,2 2 +
0 1 0 ,1 3 3 ,
This represents an optimal tableau as long as the righthand side is all non{negative. In other
words, we need between ,2 and 3 in order for the basis not to change. For any in that range,
the optimal objective will be 22 + . For example, with equals 2, the new objective is 24 with
y = 4 and x = 1.
Similarly, if we change the right{hand{side of the second constraint from 5 to 5 + in the
original formulation, we get an objective of 22 + 2 in the nal tableau, as long as ,1 1.
Perhaps the most important concept in sensitivity analysis is the shadow price i of a constraint:
If the RHS of Constraint i changes by in the original formulation, the optimal objective value
changes by i . The shadow price i can be found in the optimal tableau. It is the reduced cost
of the slack variable si . So it is found in the cost row (Row 0) in the column corresponding the
slack for Constraint i. In this case, 1 = 1 (found in Row 0 in the column of s1 ) and 2 = 2 (found
in Row 0 in the column of s2 ). The value i is really the marginal value of the resource associated
with Constraint i. For example, the optimal objective value (currently 22) would increase by 2
if we could increase the RHS of the second constraint by = 1. In other words, the marginal
value of that resource is 2, i.e. we are willing to pay up to 2 to increase the right hand side of
the second constraint by 1 unit. You may have noticed the similarity of interpretation between
shadow prices in linear programming and Lagrange multipliers in constrained optimization. Is this
just a coincidence? Of course not. This parallel should not be too surprising since, after all, linear
programming is a special case of constrained optimization. To derive this equivalence (between
shadow prices and optimal Lagrange multipiers), one could write the KKT conditions for the linear
program...but we will skip this in this course!
In summary, changing the right{hand{side of a constraint is identical to setting the correspond-
ing slack variable to some value. This gives us the shadow price (which equals the reduced cost for
the corresponding slack) and the ranges.
8.1.3 New Variable

The shadow prices can be used to determine the eect of a new variable (like a new product in a
production linear program). Suppose that, in formulation (8.1), a new variable w has coecient 4
in the rst constraint and 3 in the second. What objective coecient must it have to be considered
for adding to the basis?
If we look at making w positive, then this is equivalent to decreasing the right hand side of
the rst constraint by 4w and the right hand side of the second constraint by 3w in the original
formulation. We obtain the same eect by making s1 = 4w and s2 = 3w. The overall eect of this
is to decrease the objective by 1(4w) + 2(3w) = 1(4w) + 2(3w) = 10w. The objective value must
be sucient to oset this, so the objective coecient must be more than 10 (exactly 10 would lead
to an alternative optimal solution with no change in objective).
Example 8.1.1 A factory can produce four products denoted by P1, P2, P3 and P4. Each product
must be processed in each of two workshops. The processing times (in hours per unit produced)
are given in the following table.
P1 P2 P3 P4
Workshop 1 3 4 8 6
Workshop 2 6 2 5 8
400 hours of labour are available in each workshop. The prot margins are 4, 6, 10, and 9
dollars per unit of P1 , P2 , P3 and P4 produced, respectively. Everything that is produced can be
sold. Thus, maximizing prots, the following linear program can be used.
MAX 4 X1 + 6 X2 + 10 X3 + 9 X4
SUBJECT TO
3 X1 + 4 X2 + 8 X3 + 6 X4 <= 400 Row 1
6 X1 + 2 X2 + 5 X3 + 8 X4 <= 400 Row 2
X1, X2, X3, X4 >= 0
Introducing slack variables s1 and s2 in Rows 1 and 2, respectively, and applying the simplex
method, we get the nal tableau:
z x1 x2 x3 x4 s1 s2 RHS
1 0:5 0 2 0 1:5 0 600
0 0:75 1 2 1:5 0:25 0 100
0 4:5 0 1 5 ,0:5 1 200
(a) How many units of P1, P2 , P3 and P4 should be produced in order to maximize prots?
(b) Assume that 20 units of P3 have been produced by mistake. What is the resulting decrease
in prot?
(c) In what range can the prot margin per unit of P1 vary without changing the optimal basis?
(d) In what range can the prot margin per unit of P2 vary without changing the optimal basis?
(e) What is the marginal value of increasing the production capacity of Workshop 1?
(f) In what range can the capacity of Workshop 1 vary without changing the optimal basis?
(g) Management is considering the production of a new product P5 that would require 2 hours
in Workshop 1 and ten hours in Workshop 2. What is the minimum prot margin needed on
this new product to make it worth producing?
Answers:
(a) From the nal tableau, we read that x2 = 100 is basic and x1 = x3 = x4 = 0 are nonbasic.
So 100 units of P2 should be produced and none of P1, P3 and P4 . The resuting prot is $ 600 and
that is the maximum possible, given the constraints.
(b) The reduced cost for x3 is 2 (found in Row 0 of the nal tableau). Thus, the eect on prot
of producing x3 units of P3 is ,2x3 . If 20 units of P3 have been produced by mistake, then the prot
will be 2 20 = $40 lower than the maximum stated in (a).
(c) Let 4 + be the prot margin on P1 . The reduced cost remains nonnegative in the nal
tableau if 0:5 , 0. That is 0:5. Therefore, as long as the prot margin on P1 is less than
4:5, the optimal basis remains unchanged.
(d) Let 6 + be the prot margin on P2 . Since x2 is basic, we need to restore a correct basis.
This is done by adding times Row 1 to Row 0. This eects the reduced costs of the nonbasic
variables, namely x1, x3 , x4 and s1 . All these reduced costs must be nonnegative. This implies:
0:5 + 0:75 0
2 + 2 0
0 + 1:5 0
1:5 + 0:25 0.
Combining all these inequalities, we get 0. So, as long as the prot margin on P2 is 6 or
greater, the optimal basis remains unchanged.
(e) The marginal value of increasing capacity in Workshop 1 is 1 = 1:5.
(f) Let 400 + be the capacity of Workshop 1. The resulting RHS in the nal tableau will be:
100 + 0:25 in Row 1, and
200 , 0:5 in Row 2.
The optimal basis remains unchanged as long as these two quantities are nonnegative. Namely,
,400 400. So, the optimal basis remains unchanged as long as the capacity of Workshop 1
is in the range 0 to 800.
(g) The eect on the optimum prot of producing x5 units of P5 would be 1(2x5) + 2(10x5) =
1:5(2x5) + 0(10x5) = 3x5. If the prot margin on P5 is sucient to oset this, then P5 should be
produced. That is, we should produce P5 if its prot margin is at least 3.
Exercise 67 A paper mill converts pulpwood to low, medium and high grade newsprint. The
pulpwood requirements for each newsprint, availability of each pulpwood, and selling price (per
ton) are shown below:
Low Medium High Available
grade grade grade (tons)
Virginia pine 2 2 1 180
White pine 1 2 3 120
Loblolly pine 1 1 2 160
Price $900 $1000 $1200
The associated linear program is
MAX 900 X1 + 1000 X2 + 1200 X3

SUBJECT TO
2 X1 + 2 X2 + X3 + S1 = 180
X1 + 2 X2 + 3 X3 + S2 = 120
X1 + X2 + 2 X3 + S3 = 160
X1, X2, X3, S1, S2, S3 >= 0
with the optimal tableau

z x1 x2 x3 s1 s2 s3 RHS
1 0 200 0 300 300 0 90; 000
0 1 0:8 0 0:6 ,0:2 0 84
0 0 0:4 1 ,0:2 0:4 0 12
0 0 ,0:6 0 ,0:2 ,0:6 1 52
(a) In what range can the price of low grade paper vary without changing the optimal basis?
(b) What is the new optimal solution if the price of low grade paper changes to $800?
(c) At what price should medium grade paper be sold to make it protable to produce?
(d) In what range can the availability of Virginia pine vary without changing the optimal basis?
(e) If 10 additional tons of Virginia pine are obtained, by how much will the optimal prot
increase?
(f) If the pulpwood resources are increased as in (e), what is the new optimal solution (i.e. the
optimal production levels of low, medium and high grade newsprint)?
(g) What would the plant manager be willing to pay for an additional ton of Loblolly pine?
Exercise 68 Snacks 'R Us is deciding what to produce in the upcoming month. Due to past
purchases, they have 500 lb of walnuts, 1000 lb of peanuts, and 500 lb of chocolate. They currently
sell three mixes: Trail Mix, which consists of 1 lb of each material; Nutty Crunch, which has two
pounds of peanuts and 1 lb of walnut; and Choc-o-plenty, which has 2 pounds of chocolate and 1 lb
of peanuts. They can sell an unlimited amount of these mixes, except they can sell no more than
100 units of Choc-o-plenty. The income on the three mixes is $2, $3, and $4, respectively. The
problem of maximizing total income subject to these constraints is
Maximize 2x1 + 3x2 + 4x3
Subject to
x1 + x2 500
x1 + 2x2 + x3 1000
x1 + 2x3 500
x3 100
xi 0 for all i
The optimal tableau for this problem is (after adding slacks for the four constraints):
z x1 x2 x3 s1 s2 s3 s4 RHS
1 0 0 0 1 1 0 3 1800
0 1 0 0 2 ,1 0 1 100
0 0 1 0 ,1 1 0 ,1 400
0 0 0 0 ,2 1 1 ,3 200
0 0 0 1 0 0 0 1 100
Using the above tableau, determine the following:
(a) What is the solution represented by this tableau (both production quantities and total
income)? How do we know it is the optimal tableau?
Production quantities:
Total income:
Optimal? Reason:
(b) Suppose the income for Trail Mix (x1 ) is only an estimate. For what range of values would
the basis given by the above tableau still be optimal? What would be the solution (both production
quantities and total income), if the income was only $2.75?
Range: Lower: Upper:
Solution for $2.75: Total Income Production:
(c) How much should Snacks 'R Us willing to spend to procure an extra pound of peanuts?
How about a pound of walnuts? How about an extra pound of chocolate?
Peanuts: Walnuts: Chocolate:
(d) For what range of peanut{stock values is the basis given by the above tableau still optimal?
What would be the solution (both production quantities and total objective) if there were only 900
lbs of peanuts available?
Range: Lower: Upper:
Solution for 900 lbs: Total Income Production:
(e) A new product, Extra-Walnutty Trail Booster, consists of 1 lb of each of peanuts and
chocolate, and 2 lb of walnuts. What income do we need to make on this product in order to
consider producing it?
Income: Reason:
8.2 Solver Output

A large amount (but not all) of the previous analysis is available from the output of Solver. To
access the information, simply ask for the sensitivity report after optimizing. Rather than simply
give rules for reading the report, here are two reports, each with a set of questions that can be
answered from the output.
8.2.1 Tucker Automobiles
Tucker Inc. needs to produce 1000 Tucker automobiles. The company has four production plants.
Due to diering workforces, technological advances, and so on, the plants dier in the cost of
producing each car. They also use a dierent amount of labor and raw material at each. This is
summarized in the following table:
Plant Cost ('000) Labor Material
1 15 2 3
2 10 3 4
3 9 4 5
4 7 5 6
The labor contract signed requires at least 400 cars to be produced at plant 3; there are 3300
hours of labor and 4000 units of material that can be allocated to the four plants.
This leads to the following formulation:
MIN 15 X1 + 10 X2 + 9 X3 + 7 X4
SUBJECT TO
2 X1 + 3 X2 + 4 X3 + 5 X4 <= 3300
8.2. SOLVER OUTPUT 105
3 X1 + 4 X2 + 5 X3 + 6 X4 <= 4000
X3 >= 400
X1 + X2 + X3 + X4 = 1000
X1, X2, X3, X4 >= 0
Attached is the Solver output for this problem.
Adjustable Cells
Final Reduced Objective Allowable Allowable
Cell Name Value Cost Coefficient Increase Decrease
$B$11 X1 400 0 15 1E+30 3.5
$B$12 X2 200 0 10 2 1E+30
$B$13 X3 400 0 9 1E+30 4
$B$14 X4 0 7 7 1E+30 7
Constraints
Final Shadow Constraint Allowable Allowable
Cell Name Value Price R.H. Side Increase Decrease
$B$18 Labor 3000 0 3300 1E+30 300
$B$19 Material 4000 -5 4000 300 200
$B$20 Plant3Min 400 4 400 100 400
$B$21 Total 1000 30 1000 66.6667 100
Here are some questions to be answered:

1. What are the current production quantities? What is the current cost of production?
2. How much will it cost to produce one more vehicle? How much will we save by producing
one less?
3. How would our solution change if it cost only $8,000 to produce at plant 2? For what ranges
of costs is our solution (except for the objective value) valid for plant 2?
4. How much are we willing to pay for a labor hour?
5. How much is our union contract costing us? What would be the value of reducing the 400
car limit down to 200 cars? To 0 cars? What would be the cost of increasing it by 100 cars?
by 200 cars?
6. How much is our raw material worth (to get one more unit)? How many units are we willing
to buy at that price? What will happen if we want more?
7. A new plant is being designed that will use only one unit of workers and 4 units of raw
material. What is the maximum cost it can have in order for us to consider using it?
8. By how much can the costs at plant 1 increase before we would not produce there?
8.2.2 Carla's Maps
Carla Lee, a current MBA student, decides to spend her summer designing and marketing bicycling
maps of Western Pennsylvania. She has designed 4 maps, corresponding to four quadrants around
Pittsburgh. The maps dier in size, colors used, and complexity of the topographical relief (the
maps are actually 3-dimensional, showing hills and valleys). She has retained a printer to produce
the maps. Each map must be printed, cut, and folded. The time (in minutes) to do this for the
four types of maps is:
Print Cut Fold
A 1 2 3
B 2 4 2
C 3 1 5
D 3 3 3
Avail 15000 20000 20000
The printer has a limited amount of time in his schedule, as noted in the table.
The prot per map, based on the projected selling price minus printers cost and other variable
cost, comes out to approximately $1 for A and B and $2 for C and D. In order to have a suciently
nice display, at least 1000 of each type must be produced.
This gives the formulation:
MAX A + B + 2 C + 2 D
SUBJECT TO
A + 2 B + 3 C + 3 D <= 15000
2 A + 4 B + C + 3 D <= 20000
3 A + 2 B + 5 C + 3 D <= 20000
A >= 1000
B >= 1000
C >= 1000
D >= 1000
Attached is the Solver output.
Adjustable Cells
$B$11 XA 1500 0 1 1 0.333333
$B$12 XB 1000 0 1 0.333333 1E+30
$B$13 XC 1000 0 2 0.333333 1E+30
$B$14 XD 2833.333 0 2 1 0.5
Constraints
$B$17 Print 15000 0.5 15000 100 366.6667
$B$18 Cut 16500 0 20000 1E+30 3500
$B$19 Fold 20000 0.166667 20000 7000 1000
$B$20 MinA 1500 0 1000 500 1E+30
$B$21 MinB 1000 -0.333333 1000 1750 1000
$B$22 MinC 1000 -0.333333 1000 500 1000
$B$23 MinD 2833.333 0 1000 1833.333 1E+30
Here are some questions to answer:
1. What are the production quantities and projected prot?
2. How much is Carla willing to pay for extra printing time? cutting time? folding time? For
each, how many extra hours are we willing to buy at that price?
3. Suppose we reduced the 1000 limit on one item to 900. Which map should be decreased, and
how much more would Carla make?
4. A fth map is being thought about. It would take 2 minutes to print, 2 minutes to cut, and 3
minutes to fold. What is the least amount of prot necessary in order to consider producing
this map? What is the eect of requiring 1000 of these also?
5. The marketing analysis on D is still incomplete, though it is known that the prot of $2 per
item is within $.25 of the correct value. It will cost $500 to complete the analysis. Should
Carla continue with the analysis?
Exercise 69 Recall the Red Dwarf toaster example (Exercise 61).
From Answer report: Final Cost: 7833.33
Adjustable Cells
$B$1 Manual 633.3333333 0 7 0.5 1E+30
$B$2 SemiAut 333.3333333 0 8 1E+30 0.25
$B$3 Robotic 33.33333333 0 8.5 0.5 2.5
Constraints
$B$5 Total 1000 10.83333333 1000 170 100
$B$6 Skilled 2233.333333 0 4500 1E+30 2266.666667
$B$7 Unskill 36000 -0.083333333 36000 1000 6800
$B$8 Assembly 2700 -0.166666667 2700 500 100
(a) What is the optimal allocation of production? What is the average cost/toaster of produc-
tion.
(b) By how much can the cost of robots increase before we will change that production plan.
(c) How much is Red Dwarf willing to pay for more assembly room time? How many units is
Red Dwarf willing to purchase at that price?
(d) How much will we save if we decide to produce only 950 toasters?
(e) A new production process is available that uses only 2 minutes of skilled labor, 10 minutes
of unskilled labor, and an undetermined amount of assembly oor time. Its production cost is
determined to be $10. What is the maximum assembly oor time that the process can take before
it is deemed too expensive to use?
Answer:
(a) 633.3 should be produced manually, 333.3 should be produced semiautomatically, and 33.3
produced robotically, for an average cost of $7.383/toaster.
(b) It can increase by $0.50.
(c) Value is $0.16/minute, willing to purchase 500 at that price.
(d) Objective will go down by 50(10.833).
(e) Cost of $10 versus marginal cost of $10.833, leave 0.83. Unskilled labor costs $0.0833/unit.
Therefore, if the new process takes any time at all, it will be deemed too expensive.
Exercise 70 Kennytrail Amusement park is trying to divide its new 50 acre park into one of three
categories: rides, food, and shops. Each acre used for rides generates $150/hour prot; each acre
used for food generates $200/hour prot. Shops generate $300/hour prot. There are a number of
restrictions on how the space can be divided.
1. Only 10 acres of land is suitable for shops.
2. Zoning regulations require at least 1000 trees in the park. A food acre has 30 trees; a ride
acre has 20 trees; while a shop acre has no trees.
3. No more than 200 people can work in the park. It takes 3 people to work an acre of rides, 6
to work an acre of food, and 5 to work an acre of shops.
The resulting linear program and Solver output is attached:
MAX 150 RIDE + 200 FOOD + 300 SHOP
SUBJECT TO
RIDE + FOOD + SHOP <= 50
SHOP <= 10
20 RIDE + 30 FOOD >= 1000
3 RIDE + 6 FOOD + 5 SHOP <= 200
-------------------------------------------------------------------
Answer report:
Target Cell (Max)

Cell Name Original Value Final Value
$B$5 Cost 0 9062.5
Adjustable Cells
Cell Name Original Value Final Value
$B$1 Ride 0 31.25
$B$2 Food 0 12.5
$B$3 Shop 0 6.25
Constraints
Cell Name Cell Value Formula Status Slack
$B$6 Land 50 $B$6<=$D$6 Binding 0
$B$7 ShopLim 6.25 $B$7<=$D$7 Not Binding 3.75
$B$8 Trees 1000 $B$8>=$D$8 Binding 0
$B$9 Workers 200 $B$9<=$D$9 Binding 0
-----------------------------------------------------------------------------
Sensitivity report:
Adjustable Cells
$B$1 Ride 31.25 0 150 83.33333333 76.6666667
$B$2 Food 12.5 0 200 115 125
$B$3 Shop 6.25 0 300 1E+30 116.666667
Constraints
$B$6 Land 50 143.75 50 10 16.66666667
$B$7 ShopLim 6.25 0 10 1E+30 3.75
$B$8 Trees 1000 -4.375 1000 166.66667 100
$B$9 Workers 200 31.25 200 30 50
For each of the following changes, either nd the answer or state that the information is not
available from the Solver output. In the latter case, state why not.
(a) What is the optimal allocation of the space? What is the prot/hour of the park?
Optimal Allocation: Prot/hour:
(b) Suppose Food only made a prot of $180/hour. What would be the optimal allocation of
the park, and what would be the prot/hour of the park?
Optimal Allocation: Prot/hour:
(c) City Council wants to increase our tree requirement to 1020. How much will that cost us
(in $/hour). What if they increased the tree requirement to 1200?
Increase to 1020: Increase to 1200:
(d) A construction rm is willing to convert 5 acres of land to make it suitable for shops. How
much should Kennytrail be willing to pay for this conversion (in $/hour).
Maximum payment:
(e) Kennytrail is considering putting in a waterslide. Each acre of waterslide can have 2 trees
and requires 4 workers. What prot/hour will the waterslide have to generate for them to consider
adding it?
Minimum Prot: Reason:
(f) An adjacent parcel of land has become available. It is ve acres in size. The owner wants
to share in our prots. How much $/hour is Kennytrail willing to pay?
Maximum payment:
Exercise 71 Attached is the sensitivity report for the Diet Problem (see Section 5.4.1 in Chap-
ter 5).
Adjustable Cells
$B$14 oatmeal 4 -3.1875 3 3.1875 1E+30
$B$15 chicken 0 12.46875 24 1E+30 12.46875
$B$16 eggs 0 4 13 1E+30 4
$B$17 milk 4.5 0 9 2.6923 1.38095
$B$18 pie 2 -3.625 20 3.625 1E+30
$B$19 pork&beans 0 4.375 19 1E+30 4.375
Constraints
$B$23 Energy 2000 0.05625 2000 560 100
$B$24 Protein 60 0 55 5 1E+30
$B$25 Calcium 1334.5 0 800 534.5 1E+30
Answer each of the following questions independently of the others.

1. What does the optimal diet consist of?
2. If the cost of oatmeal doubled to 6 cents/serving, should it be removed from the diet?
3. If the cost of chicken went down to half its current price, should it be added to the diet?
4. At what price would eggs start entering the diet?
5. In what range can the price of milk vary (rounding to the nearest tenth of a cent) while the
current diet still remaining optimal?
6. During midterms, you need a daily diet with energy content increased from 2000 kcal to 2200
kcal. What is the resulting additional cost?
7. Your doctor recommends that you increase the calcium requirement in your diet from 800 mg
to 1200 mg. What is the eect on total cost?
8. Potatoes cost 12 cents/serving and have energy content of 300 kcal per serving, but no protein
nor calcium content. Should they be part of the diet?
Exercise 72 Attached is the sensitivity report for the Workforce Planning Problem (see Sec-
tion 5.4.2 in Chapter 5).
Adjustable Cells
$B$14 Shift1 4 0 1 0.5 1
$B$15 Shift2 7 0 1 0 0.333333
$B$16 Shift3 1 0 1 0.5 0
$B$17 Shift4 4 0 1 0.5 0
$B$18 Shift5 3 0 1 0 0.333333
$B$19 Shift6 3 0 1 0.5 1
$B$20 Shift7 0 0.333333 1 1E+30 0.333333
Constraints
$B$24 Monday 14 0.333333 14 1.5 6
$B$25 Tuesday 17 0 13 4 1E+30
$B$26 Wednesday 15 0.333333 15 6 3
$B$27 Thursday 16 0 16 3 4
$B$28 Friday 19 0.333333 19 4.5 3
$B$29 Saturday 18 0.333333 18 1.5 6
$B$30 Sunday 11 0 11 4 1
Answer each of the following questions independently of the others.

1. What is the current total number of workers needed to sta the restaurant?
2. Due to a special oer, demand on thurdays increases. As a result, 18 workers are needed
instead of 16. What is the eect on the total number of workers needed to sta the restaurant?
3. Assume that demand on mondays decreases: 11 workers are needed instead of 14. What is
the eect on the total number of workers needed to sta the restaurant?
4. Currently, 15 workers are needed on wednesdays. In what range can this number vary without
changing the optimal basis?
5. Currently, every worker in the restaurant is paid $1000 per month. So the objective function
in the formulation can be viewed as total wage expenses (in thousand dollars). Workers have
complained that Shift 4 is the least desirable shift. Management is considering increasing the
wages of workers on Shift 4 to $1100. Would this change the optimal solution? What would
be the eect on total wage expenses?
6. Shift 1, on the other hand, is very desirable (sundays o while on duty fridays and saturdays,
which are the best days for tips). Management is considering reducing the wages of workers
on Shift 1 to $ 900 per month. Would this change the optimal solution? What would be the
eect on total wage expenses?
7. Management is considering introducing a new shift with the days o on tuesdays and sundays.
Because these days are not consecutive, the wages will be $ 1200 per month. Will this increase
or reduce the total wage expenses?
Chapter 9
Decision Theory
Decision theory treats decisions against nature. This refers to a situation where the result (return)
from a decision depends on action of another player (nature). For example, if the decision is to
carry an umbrella or not, the return (get wet or not) depends on what action nature takes. It is
important to note that, in this model, the returns accrue only to the decision maker. Nature does
not care what the outcome is. This condition distinguishes decision theory from game theory. In
game theory, both players have an interest in the outcome.
The fundamental piece of data for decision theory problems is a payo table:
decision state of nature
1 2 ::: m
d1 r11 r12 ::: r1m
d2 r21 r22 ::: r2m
::: ::: ::: ::: :::
dn rn1 rn2 ::: rnm
The entries rij are the payos for each possible combination of decision and state of nature.
The decision process is the following.
The decision maker selects one of the possible decisions d1; d2; : : :; dn. Say di.
After this decision is made, a state of nature occurs. Say state j .
The return received by the decision maker is rij .
The question faced by the decision maker is: which decision to select? The decision will depend
on the decision maker's belief concerning what nature will do, that is, which state of nature will
occur. If we believe state j will occur, we select the decision di associated with the largest number
rij in column j of the payo table. Dierent assumptions about nature's behavior lead to dierent
procedures for selecting \the best" decision.
If we know which state of nature will occur, we simply select the decision that yields the largest
return for the known state of nature. In practice, there may be innitely many possible decisions.
If these possible decisions are represented by a vector d and the return by the real-valued function
r(d), the decision problem can then be formulated as
max r(d) subject to feasibility constraints on d:
113
114 CHAPTER 9. DECISION THEORY
The linear programming model studied earlier in this course ts in this category. In this case,
r(d) is a linear function and the feasibility constraints are linear as well. Many other deterministic
models of operations research also fall in this category.
In the remainder of this chapter, we consider decisions under risk.
9.1 Decisions Under Risk

We make the assumption that there is more than one state of nature and that the decision maker
knows the probability with which each state of nature will occur. Let pj be the probability that
state j will occur. If the decision maker makes decision di , her expected return ERi is
ERi = ri1p1 + ri2p2 + : : : + rimpm:
She will make the decision di that maximizes ERi, namely
ERi = maximum over all i of ERi:
Example 9.1.1 Let us consider the example of the newsboy problem: a newsboy buys papers from
the delivery truck at the beginning of the day. During the day, he sells papers. Leftover papers at
the end of the day are worthless. Assume that each paper costs 15 cents and sells for 50 cents and
that the following probability distribution is known.
p0 = Prob f demand = 0 g = 2=10

How many papers should the newsboy buy from the delivery truck?
To solve this exercise, we rst construct the payo table. Here rij is the reward acheived when
i papers are bought and a demand j occurs.
0 1 2 3
0 0 0 0 0
1 ,15 35 35 35
2 ,30 20 70 70
3 ,45 5 50 105
Next, we compute the expected returns for each possible decision:
ER0 = 0(2=10) + 0(4=10) + 0(3=10) + 0(1=10) = 0

ER1 = ,15(2=10) + 35(4=10) + 35(3=10) + 35(1=10) = 25
ER2 = ,40(2=10) + 10(4=10) + 60(3=10) + 60(1=10) = 30
ER3 = ,45(2=10) + 5(4=10) + 50(3=10) + 105(1=10) = 18:50
9.2. DECISION TREES 115
The maximum occurs when the newsboy buys 2 papers from the delivery truck. His expected
return is then 30 cents.
The fact that the newsboy must make his buying decision before demand is realized has a
considerable impact on his revenues. If he could rst see the demand being realized each day and
then buy the corresponding number of newspapers for that day, his expected return would increase
by an amount known as the expected value of perfect information. Millions of dollars are spent
every year on market research projects, geological tests etc, to determine what state of nature will
occur in a wide variety of applications. The expected value of perfect information indicates the
expected gain from any such endeavor and thus places an upper bound on the amount that should
be spent in gathering information.
Let us compute the expected value of perfect information EV PI for the above newsboy example.
If demand were known before the buying decision is made, the newsboy's expected return would
be
0(2=10) + 35(4=10) + 70(3=10) + 105(1=10) = 45:5
So, EV PI = 45:5 , 30 = 15:5
It should be pointed out that the criterion of maximizing expected return can sometimes produce
unacceptable results. This is because it ignores downside risk. Most people are risk averse, which
means they would feel that the loss of x dollars is more painful than the benet obtained from
the gain of the same amount. Decision theory deals with this problem by introducing a function
that measures the \attractiveness" of money. This function is called the utility function. You are
referred to 45-749 (managerial economics) for more on this notion. Instead of working with a payo
table containing the dollar amounts rij , one would instead work with a payo table containing the
utilities, say uij . The optimal decision di is that which maximizes the expected utility
EUi = ui1 p1 + ui2 p2 + : : : + uim pm
over all i.
9.2 Decision Trees

Example 9.2.1 Company ABC has developed a new line of products. Top management is at-
tempting to decide on the appropriate marketing and production strategy. Three strategies are being
considered, which we will simply refer to as A (aggressive), B (basic) and C (cautious). The market
conditions under study are denoted by S (strong) or W (weak). Management's best estimate of the
net prots (in millions of dollars) in each case are given in the following payo table.
S W
A 30 ,8
B 20 7
C 5 15
Management's best estimates of the probabilities of a strong or a weak market are 0.45 and 0.55
respectively. Which strategy should be chosen?
Using the approach introduced earlier, we can compute the expected return for each decision
and select the best one, just as we did for the newsboy problem.
ERA = 30(0:45) , 8(0:55) = 9:10

ERB = 20(0:45) + 7(0:55) = 12:85
ERC = 5(0:45) + 15(0:55) = 10:50
The optimal decision is to select B.
A convenient way to represent this problem is through the use of decision trees, as in Figure 9.1.
A square node will represent a point at which a decision must been made, and each line leading
from a square will represent a possible decision. A circular node will represent situations where the
outcome is uncertain, and each line leading from a circle will represent a possible outcome.
P(S)=.45 30
P(W)=.55
W
-8
A P(S)=.45 20
S
B
P(W)=.55
C W
7
P(S)=.45 5
S
P(W)=.55
W
15
Figure 9.1: Decision Tree for the ABC Company

Using a decision tree to nd the optimal decision is called solving the tree. To solve a decision
tree, one works backwards. This is called folding back the tree. First, the terminal branches are
folded back by calculating an expected value for each terminal node. See Figure 9.2.
Management now faces the simple problem of choosing the alternative that yields the highest
expected terminal value. So, a decision tree provides another, more graphic, way of viewing the
same problem. Exactly the same information is utilized, and the same calculations are made.
ER = 9.10
A
B ER = 12.85
B
C
ER = 10.50
C
Figure 9.2: Reduced Decision Tree for the ABC Company
Sensitivity Analysis
The expected return of strategy A is
ERA = 30P (S ) , 8P (W )
or, equivalently,
ERA = 30P (S ) , 8(1 , P (W )) = ,8 + 38P (S ):
Thus, this expected return is a linear function of the probability that market conditions will be
strong. Similarly
ERB = 20P (S ) + 7(1 , P (S )) = 7 + 13P (S )
ERC = 5P (S ) + 15(1 , P (S )) = 15 , 10P (S )
We can plot these three linear functions on the same set of axes (see Figure 9.3).
This diagram shows that Company ABC should select the basic strategy (strategy B) as long
as the probability of a strong market demand is between P (S ) = 0:348 and P (S ) = 0:6. This is
reassuring, since the optimal decision in this case is not very sensitive to an accurate estimation of
P (S ). However, if P (S ) falls below 0.348, it becomes optimal to choose the cautious strategy C,
whereas if P (S ) is above 0.6, the agressive strategy A becomes optimal.
Sequential Decisions
Example 9.2.2 Although the basic strategy B is appealing, ABC's management has the option of
asking the marketing research group to perform a market research study. Within a month, this
group can report on whether the study was encouraging (E) or discouraging (D). In the past, such
studies have tended to be in the right direction: When market ended up being strong, such studies
were encouraging 60% of the time and they were discouraging 40% of the time. Whereas, when
market ended up being weak, these studies were discouraging 70% of the time and encouraging 30%
of the time. Such a study would cost $500,000. Should management request the market research
study or not?
20 ERA
15 ERC ERB
10
-5
-10
P(S)=0 0.348 0.45 0.6 P(S)=1
Figure 9.3: Expected Return as a function of P(S)
Let us construct the decision tree for this sequential decision problem. See Figure 9.4. It is
important to note that the tree is created in the chronological order in which information becomes
available. Here, the sequence of events is
Test decision
Test result (if any)
Make decision
Market condition.
The leftmost node correspond to the decision to test or not to test. Moving along the \Test"
branch, the next node to the right is circular, since it corresponds to an uncertain event. There are
two possible results. Either the test is encouraging (E), or it is discouraging (D). The probabilities
of these two outcomes are P (E ) and P (D) respectively. How does one compute these probabilities?
We need some fundamental results about probabilities. Refer to 45-733 for additional material.
The information we are given is conditional. Given S, the probability of E is 60% and the probability
of D is 40%. Similarly, we are told that, given W, the probability of E is 30% and the probability
of D is 70%. We denote these conditional probabilities as follows
P (E jS ) = 0:6 P (E jW ) = 0:3 P (DjS ) = 0:4 P (DjW ) = 0:7
In addition, we know P (S ) = 0:45 and P (W ) = 0:55. This is all the information we need to
compute P (E ) and P (D). Indeed, for events S1; S2; : : :; Sn that partition the space of possible
P(S|E) 29.5
S
P(W|E)
A W -8.5
B
C P(S|E) 19.5
S
P(E) W P(W|E)
6.5
E S P(S|E) 4.5
W P(W|E)
D 14.5
Test S P(S|D) 29.5

P(D)
W P(W|D)
A -8.5
B
C S P(S|D) 19.5
P(W|D)
W 6.5
P(S|D) 4.5
S
W P(W|D)
No Test 14.5
P(S) 30
S
A W P(W)
B -8
C S P(S) 20
P(W)
W
7
S P(S) 5
P(W)
W 15
Figure 9.4: Test versus No-Test Decision Tree

outcomes and an event T , one has
P (T ) = P (T jS1)P (S1) + P (T jS2)P (S2) + : : : + P (T jSn)P (Sn):
Here, this gives
P (E ) = P (E jS )P (S ) + P (E jW )P (W )
= (0:6)(0:45)+ (0:3)(0:55)
= 0:435
and
P (D) = P (DjS )P (S ) + P (DjW )P (W )
= (0:4)(0:45) + (0:7)(0:55)
= 0:565
As we continue to move to the right of the decision tree, the next nodes are square, corresponding
to the three marketing and production strategies. Still further to the right are circular nodes
corresponding to the uncertain market conditions: either weak or strong. The probability of these
two events is now conditional on the outcome of earlier uncertain events, namely the result of the
market reseach study, when such a study was performed. This means that we need to compute the
following conditional probabilities: P (S jE ); P (W jE ); P (S jD) and P (W jD). These quantities are
computed using the formula
P (RjT ) = P (TPjR(T)P) (R) ;
which is valid for any two events R and T . Here, we get
P (S jE ) = P (EPjS(E)P) (S ) = (0:06)(0:45)
:435 = 0:621
Similarly,
P (S jE ) = 0:379 P (S jD) = 0:318 P (W jD) = 0:682
Now, we are ready to solve the decision tree. As earlier, this is done by folding back. See
Figures 9.5, 9.6 and 9.7. You fold back a circular node by calculating the expected returns. You
fold back a square node by selecting the decision that yields the highest expected return. The
expected return when the market research study is performed is 12.96 million dollars, which is
greater than the expected return when no study is performed. So the study should be undertaken.
As a nal note, let us compare the expected value of the study (denoted by EV SI , which stands
for expected value of sample information) to the expected value of perfect information EV PI .
EVSI is computed without incorporating the cost of the study. So
EV SI = (12:96 + 0:5) , 12:85 = 0:61
whereas
EV PI = (30)(0:45) + (15)(0:55) , 12:85 = 21:75 , 12:85 = 8:90
We see that the market research study is not very eective. If it were, the value of EV SI would
be much closer to EV PI . Yet, its value is greater than its cost, so it is worth performing.
15.10
A
B
C
14.57
P(E)
E
8.29
Test P(D) 3.58

A
B
C
10.63
11.32
No Test
9.10
A
B
C
12.85
10.50
Figure 9.5: Solving the Tree

15.10
P(E)
Test P(D)
11.32
No Test
12.85

12.96
Test
No Test
12.85
Example 9.2.3 An art dealer has a client who will buy the masterpiece Rain Delay for $50,000.
The dealer can buy the painting now for $40,000 (making a prot of $10,000). Alternatively, he
can wait one day, when the price will go down to $30,000. The dealer can also wait another day
when the price will be $25,000. If the dealer does not buy by that day, then the painting will no
longer be available. On each day, there is a 2/3 chance that the painting will be sold elsewhere and
will no longer be available.
(a) Draw a decision tree representing the dealers decision making process.
(b) Solve the tree. What is the dealers expected prot? When should he buy the painting?
(c) What is the Expected Value of Perfect Information (value the dealer would place on knowing
when the item will be sold)?
Answer:
(a)
10 20
Buy Buy Buy 25
1/3 1/3
Avail Avail
Don’t Don’t 0
Don’t
Sold 2/3
Sold 2/3 0
Figure 9.8: Solution

(b) The value is 10, for an expected prot of $10,000. He should buy the painting immediately.
(c) With probability 2/3, the painting will be sold on the rst day, so should be bought immedi-
ately. With probability 1/3(2/3) it will be sold on the second day, so should be bought after one day.
Finally, with probability 1/3(1/3) it will not be sold on the rst two days, so should be bought after
two days. The value of this is 2/3(10)+1/3(2/3)20+1/3(1/3)25 = 13.89. The EVPI is therefore
$3,889.
Exercise 73 The Scrub Professional Cleaning Service receives preliminary sales contracts from
two sources: its own agent and building managers. Historically, 38 of the contracts have come from
the Scrub agent and 58 from building managers. Unfortunately, not all preliminary contracts result
in actual sales contracts. Actually, only 12 of those preliminary contracts received from building
managers result in a sale, whereas 34 of those received from the Scrub agent result in a sale. The
net return to Scrub from a sale is $6400. The cost of processing and following up on a preliminary
contract that does not result in a sale is $320. What is the expected return associated with a
preliminary sales contract?
Exercise 74 Walter's Dog and Pony show is scheduled to appear in Cedar Rapids on July 4. The
prots obtained are heavily dependent on the weather. In particular, if the weather is rainy, the
show loses $28,000 and if sunny, the show makes a prot of $12,000. (We assume that all days are
either rainy or sunny.) Walter can decide to cancel the show, but if he does, he forfeits a $1,000
deposit he put down when he accepted the date. The historical record shows that on July 4, it has
rained 14 of the time for the last 100 years.
(a) What decision should Walter make to maximize his expected net dollar return?
(b) What is the expected value of perfect information?
Walter has the option to purchase a forecast from Victor's Weather Wonder. Victor's accuracy
varies. On those occasions when it has rained, Victor has been correct (i.e. predicted rain) 90%
of the time. On the other hand, when it has been sunny, he has been right (i.e. he predicted sun)
only 80% of the time.
(c) If Walter had the forecast, what strategy should he follow to maximize his expected net
dollar return? (d) How much should Walter be willing to pay to have the forecast?
Exercise 75 Wildcat Oil is considering spending $100,000 to drill at a particular spot. The result
of such a drilling is either a \Dry Well" (of no value), a \Wet Well" (providing $150,000 in revenues)
or a \Gusher" (providing $250,000 in revenues). The probabilities for these three possibilities are
.5, .3 and .2 respectively.
(a) Draw a decision tree for the problem of deciding whether to drill or not.
(b) Solve the decision tree assuming the goal is to maximize the expected net revenue. Should
the company drill?
The following two problems should be done independently.
(c) SureFire Consultants is able to determine for certain the type of well before drilling. They
oer to tell Wildcat the type for $50,000. Should Wildcat accept their oer?
(d) Close-enough Consultants oer to use their specialized seismic hammer. This hammer
returns either encouraging or discouraging results. In the past, when applied to a Gusher, the
hammer always returned encouraging results. When applied to a Wet Well, it was encouraging 75%
of the time and discouraging 25% of the time. When applied to a Dry Well, it was encouraging
one-third of the time and discouraging two-thirds of the time. What is the maximum Wildcat
should pay Close-enough Consultants for use of their hammer?
Chapter 10
Game Theory
In previous chapters, we have considered situations where a single decision maker chooses an optimal
decision without reference to its eect on other decision makers. In many business situations,
however, the outcome for a given rm depends not just on what strategy it chooses, but also on
what strategies its competitors choose. Game theory is an attempt to formalize such competitive
situations. We present some of the main ideas in the context of Two-Person games and then discuss
brie y n-Person Games.
10.1 Two-Person Games

Each player has several strategies. If the rst player chooses Strategy i while the second player
chooses Strategy j , then Player 1 gains aij while Player 2 gains bij . This outcome is represented by
(aij ; bij . 2-person games where the players' interests are completely opposed are called zero-sum or
constant-sum games: one player's gain is the other player's loss. Games where the players' interests
are not completely opposed are called variable-sum games. Such games arise in business on an
everyday basis, and solving them is not an easy task. Certain 2-person games admit pure strategies
whereas others require mixed strategies. A pure strategy is one where, each time the players play
the game, they choose the same strategy. A mixed strategy is one where the players introduce a
random element in their choice of a strategy, thus leaving the opponent guessing.
10.1.1 Pure Strategies
We start with a constant-sum game: for every possible outcome of the game, the utility u1 of Player
1 plus the utility u2 of Player 2, adds to a constant. For example, if two rms are competing for
market shares, then u1 + u2 = 100%.
Example 10.1.1 (Battle of the Networks)
Two television networks are battling for viewer shares. Viewer share is important because,
the higher it is, the more money the network can make from selling advertising time during that
program. Consider the following situation: the networks make their programming decisions inde-
pendently and simultaneously. Each network can show either sports or a sitcom. Network 1 has a
programming advantage in sitcoms and Network 2 has one in sports: If both networks show sitcoms,
then Network 1 gets a 56% viewer share. If both networks show sports, then Network 2 gets a 54%
viewer share. If Network 1 shows a sitcom and Network 2 shows sports, then Network 1 gets a 51%
viewer share and Network 2 gets 49%. Finally, if Network 1 shows sports and Network 2 shows a
sitcom, then each gets a 50% viewer share.
125
126 CHAPTER 10. GAME THEORY
The possible outcomes are best represented in a table.
Network 2
Sitcom Sports
Sitcom (56%; 44%) (51%; 49%)
Network 1
Sports (50%; 50%) (46%; 54%)
We can solve Battle of the Networks as follows: from Network 1's point of view, it is better
to show a sitcom, whether Network 2 shows a sitcom or sports. The strategy \Show a Sitcom"
is said to dominate the strategy \Show Sports" for Network 1. So Network 1 will show a sitcom.
Similarly, Network 2 is better o showing sports whether Network 1 shows a sitcom or sports. In
other words, Network 2 also has a dominating strategy. So Network 2 shows sports. The resulting
outcome, namely 51% viewer share to Network 1 and 49% to Network 2, is an equilibrium, since
neither of the two players in this game can unilaterally improve their outcome. If Network 1 were
to switch from sitcom to sports, its viewer share would drop 5%, from 51% to 46%. If Network 2
were to switch from sports to sitcom, its viewer share would also drop, from 49% to 44%. Each
network is getting the best viewer share it can, given the competition it is up against.
In a general 2-person game, Strategy i for Player 1 dominates Strategy k for Player 1 if aij akj
for every j . Similarly, Strategy j for Player 2 dominates Strategy ` for Player 2 if bij bi` for every
i.
In a general 2-person game, Strategy i for Player 1 and Strategy j for Player 2 is an equilibrium
if the corresponding outcome (aij ; bij ) has the following property: aij is the largest element akj in
column j and bij is the largest element bi` in row i. When such a pair i; j exists, a pure strategy
for each player provides an optimal solution to the game. When no such a pair exists, the players
must play mixed strategies to maximize their gains (see Section 10.1.2).
There is an easy way to relate constant-sum games and zero-sum games. Let u1 and u2 be the
payos of players 1 and 2, respectively, in a constant-sum game: u1 + u2 = k. Now consider a new
set of payos of the form
v1 = u1 , u2
v2 = u2 , u1
Clearly, v1 + v2 = 0, so we now have a zero-sum game. Using the relation u1 + u2 = k, we can
rewrite the payos v1 and v2 as:
v1 = 2u1 , k
v2 = 2u2 , k
These are positive, linear transformations of utility, and so, according to the expected utility
theorem (see 45-749), they have no eect on decisions. For the Battle of the Networks example,
the new table becomes
10.1. TWO-PERSON GAMES 127
Network 2
Sitcom Sports
Sitcom (12%; ,12%) (2%; ,2%)
Network 1
Sports (0%; 0%) (,8%; 8%)
It represents each network's advantage in viewer share over the other network. Whether you
reason in terms of total viewer shares or in terms of advantage in viewer shares, the solution is the
same.
Example 10.1.2 Competitive Advantage
In a technologically advanced economy like that of the United states, rms constantly encounter
the following situation. A new technological advance becomes available. If one rm adopts the new
technology, it gains an advantage over its competitors, a competitive advantage. If all rms adopt
the new technology, then the advantage vanishes. This is represented in the next table, where a
measures the size of the competitive advantage.
Firm 2
Adopt Stay put
Adopt (0; 0) (a; ,a)
Firm 1
Stay put (,a; a) (0; 0)
Each rm has two strategies, either Stay put, or Adopt the new technology. Firm 1 has an
incentive to adopt the new technology: in the event Firm 2 stays put, then Firm 1 gets the
competitive advantage a, and in the event Firm 2 adopts the new technology, then Firm 1 erases
its competitive disadvantage ,a. So, whatever Firm 2's decision is, Firm 1 is better o adopting the
new technology. This is Firm 1's dominant strategy. Of course, the situation for Firm 2 is identical.
So the equilibrium of Competitive Advantage is for both rms to adopt the new technology. As a
result, both rms get a payo of 0. This may seem like a paradox, since the payos would have
been the same if both rms had stayed put. But, in Competitive Advantage, neither rm can
aord to stay put. The rms in this game are driven to adopt any technology that comes along.
Take, for example, the hospital industry. Magnetic Resonance Imaging (MRI) is a new technology
that enhances conventional X rays. It allows doctors to see body damage in ways that were not
previously possible. Once MRI became available, any hospital that installed an MRI unit gained
a competitive advantage over other hospitals in the area. From a public policy standpoint, it may
not make much sense for every hospital to have its own MRI unit. These units are expensive to
buy and to operate. They can eat up millions of dollars. Often, one MRI unit could handle the
trac of several hospitals. But an individual hospital would be at a competitive disadvantage if
it did not have its own MRI. As long as hospitals play Competitive Advantage, they are going to
adopt every new technology that comes along.
The two-person games we have encountered so far have had unique pure strategy equilibria.
However, a two-person zero-sum game may have multiple equilibria. For example, consider the
game:
Player 2
A B C
A (0; 0) (1; ,1) (0; 0)
Player 1 B (,1; 1) (0; 0) (,1; 1)
C (0; 0) (1; ,1) (0; 0)
Each player can play indierently strategy A or C and so there are four pure strategy equilibria,
corresponding to the four corners of the above table. Note that these equilibria all have the same
payo. This is not a coincidence. It can be shown that this is always the case. Every equilibrium
of a 2-person zero-sum game has the same payo. For pure strategy equilibria, there is a simple
proof: Suppose two equilibria had payos (a; ,a) and (b; ,b) where a 6= b. If these two solutions
lied on the same row or column, we would get an immediate contradiction to the denition of an
equilibrium. Let (a; ,a) lie in row i and column k and let (b; ,b) lie in row j and column `. The
subtable corresponding to these two rows and columns looks as follows
Player 2
k `
i (a; ,a) (c; ,c)
Player 1
j (d; ,d) (b; ,b)
Since (a; ,a) is an equilibrium, we must have a d (from Player 1) and ,a ,c (from Player
2). Similarly, since (b; ,b) is an equilibrium, we must have b c and ,b ,d. Putting these
inequalities together, we get
adbca
This implies that a = b, which completes the proof.
Example 10.1.3 Cigarette Advertising on Television
On January 1, 1971, cigarette advertizing was banned on american television. It came as
something of a surprise to the industry, and a welcome one at that, that prots rose by $91 million
in 1971. Why did this happen? Consider the strategic interaction in the industry before and after
the agreement went into eect. Although there were four large tobacco companies involved, let us
consider the strategic interaction between only two of them, for simplicity. In 1970, the payo
matrix was:
Company 2
No TV ads TV Ads
No TV Ads (50; 50) (20; 60)
Company 1
TV Ads (60; 20) (27; 27)
It is clear from this table that advertising on television is a powerful marketing tool. If Com-
pany 1 switches from not advertising to advertising on television, its prots go up 20%, everything
else being equal, when Company 2 does not advertise on TV, and they go up 35% when Company 2
advertises on TV; the same is true if the roles are reversed. In other words, advertising on television
is a dominant strategy for each of the companies. Hence the equilibrium is when both companies
advertise on television, with a payo of $27 million. We say that an outcome (aij ; bij ) is ecient if
there is no other outcome (akl ; bkl) that pays both players at least as much, and one or both players
strictly more. Namely, the pair i; j is ecient if there is no pair k; l satisfying akl aij and bkl bij
and akl + bkl > aij + bij . Obviously, the equilibrium in the cigarette advertizing game is not ecient.
There is one ecient outcome, however, in this game: that is when neither company advertises
on TV. Then both companies enjoy a payo of $50 million. The ban of cigarette advertizing on
television in 1971 only left the strategy \No TV Ads" and forced the industry into the ecient
outcome! In large part, this is why prots went up in 1971.
It often happens in variable-sum games that the solution is not ecient. One of the biggest
dierences between constant-sum and variable-sum games is that solutions to the former are always
ecient, whereas solutions to the latter rarely are.
10.1.2 Mixed Strategies

Example 10.1.4 Market Niche
Two rms are competing for a single market niche. If one rm occupies the market niche, it
gets a return of 100. If both rms occupy the market niche, each loses 50. If a rm stays out of
the market, it breaks even. The payo table is:
Firm B
Enter Stay out
Enter (,50; ,50) (100; 0)
Firm A
Stay out (0; 100) (0; 0)
This game has two pure strategy equilibria, namely one of the two rms enters the market niche
and the other stays out. But, unlike the games we have encountered thus far, neither player has a
dominant strategy. When a player has no dominant strategy, she should consider playing a mixed
strategy. In a mixed strategy, each of the various pure strategies is played with some probability,
say p1 for Strategy 1, p2 for Strategy 2, etc with p1 + p2 + : : : = 1. What would be the best
mixed strategies for Firms A and B? Denote by p1 the probability that Firm A enters the market
niche. Therefore p2 = 1 , p1 is the probability that Firm A stays out. Similarly, Firm B enters the
niche with probability q1 and stays out with probability q2 = 1 , q1 . The key insight to a mixed
strategy equilibrium is the following. Every pure strategy that is played as part of a mixed strategy
equilibrium has the same expected value. If one pure strategy is expected to pay less than another,
then it should not be played at all. The pure strategies that are not excluded should be expected
to pay the same. We now apply this principle. The expected value of the \Enter" strategy for Firm
A, when Firm B plays its mixed strategy, is
EVEnter = ,50q1 + 100q2:
The expected value of the \Stay out" strategy for Firm A is EVStay out = 0. Setting EVEnter =
EVStay out we get
,50q1 + 100q2 = 0:
Using q1 + q2 = 1, we obtain
q1 = 23 q2 = 31 :
Similarly,
p1 = 23 p2 = 31 :
As you can see, the payos of this mixed strategy equilibrium, namely (0; 0), are inecient. One
of these rms could make a lot of money by entering the market niche, if it was sure that the other
would not enter the same niche. This assurance is precisely what is missing. Each rm has exactly
the same right to enter the market niche. The only way for both rms to exercise this right is to play
the inecient, but symmetrical, mixed strategy equilibrium. In many industrial markets, there is
only room for a few rms { a situation known as natural oligopoly. Chance plays a major role in
the identity of the rms that ultimately enter such markets. If too many rms enter, there are
losses all around and eventually some rms must exit. From the mixed strategy equilibrium, we can
actually predict how often two rms enter a market niche when there is only room for one: with the
above data, the probability of entry by either rm is 2/3, so the probability that both rms enter
is (2=3)2 = 4=9. That is a little over 44% of the time! This is the source of the ineciency. The
ecient solution has total payo of 100, but is not symmetrical. The fair solution pays each player
the same but is inecient. These two principles, eciency and fairness, cannot be reconciled in a
game like Market Niche. Once rms recognize this, they can try to nd mecanisms to reach the
ecient solution. For example, they may consider side payments. Or rms might simply attempt
to scare o competitors by announcing their intention of moving into the market niche before they
actually do so.
An important case where mixed strategies at equilibrium are always ecient is the case of
constant-sum games. Interestingly, the optimal mixed strategies can be computed using linear
programming, one linear program for each of the two players. We illustrate these ideas on an
example.
Example 10.1.5 Stone, Paper, Scissors
Each of two players simultaneously utters one of the three words stone, paper, or scissors. If
both players utter the same word, the game is a draw. Otherwise, one player pays $1 to the other
according to the following: Scissors defeats (cuts) paper, paper defeats (covers) stone, and stone
defeats (breaks) scissors. Find the optimal strategies for this game.
The payo matrix is as follows.

Player 2
Stone Paper Scissors
Stone (0; 0) (,1; 1) (1; ,1)
Player 1 Paper (1; ,1) (0; 0) (,1; 1)
Scissors (,1; 1) (1; ,1) (0; 0)
Suppose Player 1 chooses the mixed strategy (x1; x2; x3), where
x1 = Probability that Player 1 chooses Stone
x2 = Probability that Player 1 chooses Paper
x3 = Probability that Player 1 chooses Scissors
The expected payo of Player 1 is x2 , x3 when Player 2 plays Stone, x3 , x1 when Player 2 plays
Paper and x1 , x2 when Player 2 plays Scissors. The best Player 2 can do is to achieve the minimum
of these three values, namely v = min(x2 , x3 ; x3 , x1 ; x1 , x2 ). Player 1, on the other hand, should
try to make the quantity v as large as possible. So, Player 1 should use probabilities x1 , x2 and
x3 that maximize v = min(x2 , x3; x3 , x1; x1 , x2). This can be written as the following linear
program.
max v
v x2 , x3
v x3 , x1
v x1 , x2
x1 + x2 + x3 = 1
x1; x2; x3 0:
Similarly, we obtain Player 2's optimal mixed strategy (y1 ; y2; y3), where
y1 = Probability that Player 2 chooses Stone
y2 = Probability that Player 2 chooses Paper
y3 = Probability that Player 2 chooses Scissors
by solving the linear program
max z
z y2 , y3
z y3 , y1
z y1 , y2
y1 + y2 + y3 = 1
y1 ; y2 ; y3 0:
For the Stone, Paper, Scissors game, the optimal solutions are x1 = x2 = x3 = 1=3 and
y1 = y2 = y3 = 1=3 with v = z = 0.
In general, for any 2-person zero-sum game, let v and w be the optima of the linear programs
for Players 1 and 2 respectively. Because of the special form of these linear programs, it can be
shown that z = ,v . So, the payo at equilibrium is (v ; ,v ). The fact that, for each player, the
equilibrium solutions are the optimal solutions of a linear has another consequence: All equilibrium
solutions have the same payo.
10.2 n-Person Games

When there are more than two players, fundamental dierences arise. For example, the theorem
about 2-person zero-sum games which says that all equilibrium solutions have the same payo
cannot be generalized to three players. Once there are more than two players, a zero-sum game can
have equilibrium solutions that do not have the same payos. Here is an example with three players:
each player has two strategies, say A and B. Whatever Players 1 and 2 play, the payo vector is
(1; ,21 ; ,21 ) when Player 3 chooses Strategy A and it is ( ,21 ; 1; ,21 ) when Player 3 chooses Strategy
B. We will call the player that receives the payo of 1 the winner and the players that receive a
payo of ,21 , the loosers. In this game, although Player 3 can never be the winner, he determines
which of Players 1 or 2 will be the winner. When this situation occurs, we say that Player 3 is
a spoiler. For example, take the 1992 presidential election: 1=Bush, 2=Clinton, 3=Perot. Perot
could not win, but he could play the spoiler in this campaign. In a two-way race, noone can play
a spoiler; in a three-way race, there is room for a spoiler.
Example 10.2.1 Competitive Advantage with Three Firms
As before, each rm has the choice between staying put or adopting the new technology. If no rm
adopts the new technology, then there is no competitive advantage and the payo vector is (0; 0; 0).
If exactly one rm adopts the new technology, then that rm gets the competitive advantage a, while
each rm at a competitive disadvantage looses a2 . Thus if only Firm 1 adopts the new technology,
then the payo vector is (a; ,2a ; ,2a ). You can think of this as Firm 1 taking market share from
both Firm 2 and Firm 3. If exactly two rms adopt the new technology, then these two rms split
the competitive advantage, each gaining a2 , and the rm at a disadvantage looses a. Finally, if all
rms adopt the new technology, there is no competitive advantage and the payo vector is (0; 0; 0).
We can represent the payos in a table as follows.
Firm 3 Adopt Firm 3 Stay put
Firm 2 Firm 2
Adopt Stay put Adopt Stay put
Adopt (0; 0; 0) ( a2 ; ,a; a2 ) Adopt ( a2 ; a2 ; ,a) (a; ,2a ; ,2a )
Firm 1 Firm 1
Stay put (,a; a2 ; a2 ) ( ,2a ; ,2a ; a) Stay put ( ,2a ; a; ,2a ) (0; 0; 0)
Each rm has a dominant strategy, which is to adopt the new technology. So the unique
equilibrium occurs when all three rms play the pure strategy: Adopt the new technology. This is
10.2. N-PERSON GAMES 133
presicely what happened with two rms. No rm can be left behind in the race to adopt the new
technology. This is just as true for n players as it is for two or three. So, here, the insight from the
2-person game version of Competitive Advantage generalizes nicely to n-person games.
An important aspect of n-person games is that subsets of players can form coalitions. A useful
concept in this case is the core of the game. Another important aspect of game theory is that
of fairness: the Shapley value provides an answer. Game theory also considers games with a
sequential structure, games with imperfect information, market games (you learned about Cournot
and Bertrand competition in 45-749). To learn more about these and other aspects of game theory,
you are referred to the excellent book \Games for Business and Economics" by Roy Gardner (1995),
which was used as a basis for this chapter.
Exercise 76 Here is a game that arises in resource economics.
(a) There are two shing spots, one good and the other excellent. The good spot has 6 sh whereas
the excellent spot has 10 sh. Two shermen know how good each of the spots is. Each sherman
must choose one of the two spots to go shing. If they end up going to the same spot, they divide
the catch equally. If they go to dierent spots, the sherman at the good spot gets 6 sh, the one
at the excellent spot gets 10. Find three equilibria for this game.
(b) In such shing games, con icts often break over the shing grounds. How might an assymetrical
equilibrium, such as the one provided by the 200-mile shing limit among nations, lead to a more
ecient outcome?
Exercise 77 Find each player's optimal strategy and the value of the two-person zero-sum game
with the following payo table.
Player 2
S1 S2 S3 S4
T1 (4; ,4) (5; ,5) (1; ,1) (4; ,4)
Player 1 T2 (2; ,2) (1; ,1) (6; ,6) (3; ,3)
T3 (1; ,1) (0; 0) (0; 0) (2; ,2)

Chapter 11
Network Optimization
11.1 Introduction
Network optimization is a special type of linear programming model. Network models have three
main advantages over linear programming:
1. They can be solved very quickly. Problems whose linear program would have 1000 rows and
30,000 columns can be solved in a matter of seconds. This allows network models to be used
in many applications (such as real-time decision making) for which linear programming would
be inappropriate.
2. They have naturally integer solutions. By recognizing that a problem can be formulated as
a network program, it is possible to solve special types of integer programs without resorting
to the ineective and time consuming integer programming algorithms.
3. They are intuitive. Network models provide a language for talking about problems that is
much more intuitive than the \variables, objective, and constraints" language of linear and
integer programming.
Of course these advantages come with a drawback: network models cannot formulate the wide
range of models that linear and integer programs can. However, they occur often enough that they
form an important tool for real decision making.
11.2 Terminology
A network or graph consists of points, and lines connecting pairs of points. The points are called
nodes or vertices. The lines are called arcs. The arcs may have a direction on them, in which case
they are called directed arcs. If an arc has no direction, it is often called an edge. If all the arcs in a
network are directed, the network is a directed network. If all the arcs are undirected, the network
is an undirected network.
Two nodes may be connected by a series of arcs. A path is a sequence of distinct arcs (no nodes
repeated) connecting the nodes. A directed path from node i to node j is a sequence of arcs, each of
whose direction (if any) is towards j . An undirected path may have directed arcs pointed in either
direction.
A path that begins and ends at the same node is a cycle and may be either directed or undirected.
A network is connected if there exists an undirected path between any pair of nodes. A connected
network without any cycle is called a tree, mainly because it looks like one.
135
136 CHAPTER 11. NETWORK OPTIMIZATION
11.3 Examples
There are many examples of using network ows in practice. Here are a few:
11.3.1 Shortest Paths
Consider a phone network. At any given time, a message may take a certain amount of time to
traverse each line (due to congestion eects, switching delays, and so on). This time can vary greatly
minute by minute and telecommunication companies spend a lot of time and money tracking these
delays and communicating these delays throughout the system. Assuming a centralized switcher
knows these delays, there remains the problem of routing a call so as to minimize the delays. So, in
gure 11.1, what is the least delay path from LA to Boston? How can we nd that path quickly?
2
8
1 7 1
5
3
LA BOS
7 2
3
Figure 11.1: Phone Network

This is an example of a particular type of network model, called the shortest path problem. In
such a problem, you have a network with costs on the edges and two special nodes: a start node
and a nish node. The goal is to nd a path from the start node to the nish node whose total
weight is minimized.
Here is another problem that might not appear to be a shortest path problem, but really is:
At a small but growing airport, the local airline company is purchasing a new tractor for a
tractor-trailer train to bring luggage to and from the airplanes. A new mechanized luggage system
will be installed in 3 years, so the tractor will not be needed after that. However, because it will
receive heavy use, and maintenance costs are high, it may still be more economical to replace the
tractor after 1 or 2 years. The following table gives the total net discounted cost associated with
purchasing a tractor in year i and trading it in in year j (where year 0 is now):
j
1 2 3
0 8 18 31
i 1 | 10 21
2 | | 12
The problem is to determine at what times the tractor should be replaced (if ever) to minimize the
total costs for tractors. How can this be formulated as a shortest path problem?
11.3. EXAMPLES 137
11.3.2 Maximum Flow

Another type of model again has a number on each arc, but now the number corresponds to
a capacity. This limits the ow on the arc to be no more than that value. For instance, in a
distribution system, the capacity might be the limit on the amount of material (in tons, say) that
can go over a particular distribution channel. We would then be concerned with the capcity of
the network: how much can be sent from a source node to the destination node? Using the same
network as above, treating the numbers as capacities, how much can be sent from LA to Boston?
2
8
1 7 1
5
3
LA BOS
7 2
3
Figure 11.2: Distribution Network

Associated with the maximum ow is a bottleneck: a set of arcs whose total capacity equals
the maximum ow, and whose removal leaves no path from source to destination in the network.
It is actually a nontrivial result to show that the maximum ow equals the size of the minimum
bottleneck. It is also an interesting task to nd the bottleneck in the above example.
Maximum ow models occur in applications where cost is not an issue, and the goal is to
maximize the number of items handled (in a broad sense). Here is a similar problem that can be
formulated as a maximum ow problem (this was part of a conversation with a Chicago consulting
rm on how to improve their internal operations):
Over the next three months, there are four projects to be completed. Project A requires 8
person-months of work, and must be done by the end of the second month. Project B requires 3
person-months and must be done by end of the rst month. Project C can start at the beginning of
the second month, and requires 7 person-months of work, no more than 2 of which are in the second
month. Project D requires 6 person-months, no more than 3 of which occur in any month. There
are 9 consultants available in the rst and second month, and 6 in the third. Is it possible to meet
the project deadlines? Surprisingly, this is a maximum ow problem. One nice (and surprising)
aspect of the result is that this shows that no consultant need split a month's work between jobs.
Exercise 78 The Smith's, Jones', Johnson's, and Brown's are attending a picnic. There are four
members of each family. Four cars are available to transport them: two cars that can hold four
people, and two that can hold three. No more than 2 members of any family can be in any car.
The goal is to maximize the number of people who can attend the picnic. Formulate this as a
maximum ow problem. Give the nodes and arcs of the network and state which is the starting
node and which is the destination. For each arc, give its capacity.
2
2 4
2
4 4
4
4
3
4 end
start 3
2
2
Figure 11.3: Solution of the exercise.
11.3.3 Transportation Problem

Consider the following snow removal problem: there are a number of districts in a city. After a
snowfall, the snow in each area must be moved out of the district into a convenient location. In
Montreal (from where this example is taken), these locations are large grates (leading to the sewer
system), a couple large pits, and a couple entry points to the river. Each of these destinations has
a capacity. The goal is to minimize the distance traveled to handle all of the snow.
This problem is an example of a transportation problem. In such a problem, there are a set
of nodes called sources, and a set of nodes called destinations. All arcs go from a source to a
destination. There is a per{unit cost on each arc. Each source has a supply of material, and each
destination has a demand. We assume that the total supply equals the total demand (possibly
adding a fake source or destination as needed). For the snow removal problem, the network might
look like that in gure 11.4.
1000 2
3
10 3000
3
2000 2
2
AREAS 4 2000
7
2500 1
9 7
2000
1500
5
Figure 11.4: Snow Transportation Network

Transportation problems are often used in, surprise, transportation planning. For instance,
11.3. EXAMPLES 139
in an application where goods are at a warehouse, one problem might be to assign customers to
a warehouse so as to meet their demands. In such a case, the warehouses are the sources, the
customers are the destinations, and the costs represent the per{unit transportation costs.
Here is another example of this:
One of the main products of P&T Company is canned peas. The peas are prepared at three
canneries (near Bellingham, Washington; Eugene, Oregon; and Albert Lea, Minnesota) and are then
shipped by truck to four distributing warehouses in Sacremento, California; Salt Lake City, Utah;
Rapid City, South Dakota; and Albuquerque, New Mexico. Because shipping costs are a major
expense, management has begun a study to reduce them. For the upcoming season, an estimate
has been made of what the output will be from each cannery, and how much each warehouse will
require to satisfy its customers. The shipping costs from each cannery to each warehouse has also
been determined. This is summarized in the next table.
Shipping cost Warehouse

per truckload 1 2 3 4 Output
1 464 513 654 867 75
Cannery 2 352 416 690 791 125
3 995 682 388 685 100
Requirement 80 65 70 85
You should nd it an easy exercise to model this as a linear program. If we let xij be the
number of truckloads shipped from cannery i to warehouse j , the problem is to
minimize 464x11 + 513x12 + 654x13 + 867x14 + 352x21 + : : : + 685x34
subject to
x11 + x12 + x13 + x14 = 75
x21 + x22 + x23 + x24 = 125
x31 + x32 + x33 + x34 = 100
x11 + x21 + x31 = 80
x12 + x22 + x32 = 65
x13 + x23 + x33 = 70
x14 + x24 + x34 = 85
xij 0 for all i and j .
This is an example of the transportation model. As has been pointed out, this problem
has a lot of nice structure. All the coecients are 1 and every variable appears in exactly two
constraints. It is this structure that lets the simplex algorithm be specialized into an extremely
ecient algorithm.
What denes a transportation model? In general, the transportation model is concerned with
distributing (literally or guratively) a commodity from a group of supply centers, called sources
to a group of receiving centers, called destinations to minimize total cost.
In general, source i has a supply of si units, and destination j has a demand for dj units. The
cost of distributing items from a source to a destination is proportional to the number of units.
This data can be conveniently represented in a table like that for the sample problem.
We will generally assume that the total supply equals the total demand. If this is not true
for a particular problem, dummy sources or destinations can be added to make it true. The text
refers to such a problem as a balanced transportation problem. These dummy centers may have zero
distribution costs, or costs may be assigned to represent unmet supply or demand.
For example, suppose that cannery 3 makes only 75 truckloads. The total supply is now 25
units too little. A dummy supply node can be added with supply 25 to balance the problem, and
the cost from the dummy to each warehouse can be added to represent the cost of not meeting the
warehouse's demand.
The transportation problem has a couple of nice properties:
Feasibility. As long as the supply equals demand, there exists a feasible solution to the
problem.
Integrality. If the supplies and demands are integer, every basic solution (including optimal
ones) have integer values. Therefore, it is not necessary to resort to integer programming to nd
integer solutions. Linear programming suces. Note that this does not mean that each destination
will be supplied by exactly one source.
In the pea shipping example, a basic solution might be to ship 20 truckloads from cannery 1
to warehouse 2 and the remaining 55 to warehouse 4, 80 from cannery 2 to warehouse 1 and 45 to
warehouse 2 and, nally, 70 truckloads from cannery 3 to warehouse 3 and 30 to warehouse 4. Even
though the linear programming formulation of the pea shipping example has seven constraints other
than nonnegativity, a basic solution has only six basic variables! This is because the constraints
are linearly dependent: the sum of the rst three is identical to the sum of the last four. As a
consequence, the feasible region dened by the constraints would remain the same if we only kept
six of them. In general, a basic solution to the transportation model will have a number of basic
variables equal to the number of sources plus the number of destinations minus one.
Exercise 79 Formulate the following production problem as a transportation model. The demands
for a given item are 150, 250, 200 units for the next three months. The demand may be satised
by
excess production in an earlier month held in stock for later consumption,
production in the current month,
excess production in a later month backordered for preceding months.
The variable production cost per unit in any month is $6.00. A unit produced for later consumption
will incur a storage cost at the rate of $1 per unit per month. On the other hand, backordered
items incur a penalty cost of $3.00 per unit per month. The production capacity in each of the
next three months is 200 units. The objective is to devise a minimum cost production plan.
Exercise 80 Avertz RentaCar needs to redeploy its automobiles to correct imbalances in the
system. Currently Avertz has too many cars in New York (with 10 cars excess) and Chicago (12
cars excess). Pittsburgh would like up to 6 cars, Los Angeles up to 14 cars, and Miami up to 7 cars
(note that more cars are demanded than are available). The cost of transporting a car from one
city to another is given by:
Pittsburgh Los Angeles Miami
New York 50 250 100
Chicago 25 200 125
(a) Formulate this problem as a transportation problem. Clearly give the nodes and arcs. For
each node, give the supply or demand at the node. For each arc, give the cost. Assume that unmet
demand at a city has no cost but that no city can end up with excess cars.
11.3. EXAMPLES 141
(b) It turns out that unmet demand costs $50/car in Pittsburgh, $75/car in LA, and $100/car
in Miami. Update your transportation formulation in (a) to account for this change.
11.3.4 Assignment Problem
A special case of the transportation problem is the assignment problem which occurs when each
supply is 1 and each demand is 1. In this case, the integrality implies that every supplier will be
assigned one destination and every destination will have one supplier. The costs give the charge
for assigning a supplier and destination to each other.
Example 11.3.1 A company has three new machines of dierent types. There are four dierent
plants that could receive the machines. Each plant can only receive one machine, and each machine
can only be given to one plant. The expected prot that can be made by each plant if assigned a
machine is as follows:
Plant
1 2 3 4
1 13 16 12 11
Machine 2 15 0 13 20
3 5 7 10 6
This is a transportation problem with all supplies and demands equal to 1, so it is an assignment
problem.
Note that a balanced problem must have the same number of supplies and demands, so we must
add a dummy machine (corresponding to receiving no machine) and assign a zero cost for assigning
the dummy machine to a plant.
Exercise 81 Albert, Bob, Carl, and David are stranded on a desert island with Elaine, Francine,
Gert, and Holly. The \compatibility measures" in the next table indicate the happiness each couple
would experience if they spent all their time together. If a couple spends only a partial amount of
time together, then the happiness is proportional to the fraction of time spent. So if Albert and
Elaine spend half their time together, they earn happiness of 7/2.
E F G H
A 7 5 8 2
B 7 8 9 4
C 3 5 7 9
D 5 5 6 7
(a) Let xij be the fraction of time that man i spends with woman j . Formulate a linear program
that maximizes the total happiness of the island (assume that no man spends time with any other
man, no woman spends time with any other woman, and no one spends time alone).
(b) Explain why exactly four xij will be 1 and the rest will be 0 in an optimal solution to this
linear program. This result is called the Marriage Theorem for obvious reasons.
(c) Do you think that this theorem will hold if we allow men to spend time with men and women
to spend time with women?
11.4 Unifying Model: Minimum Cost Network Flows

All of the above models are special types of network ow problems: they each have a specialized
algorithm that can nd solutions hundreds of times faster than plain linear programming.
They can all also be seen as examples of a much broader model, the minimum cost network ow
model. This model represents the broadest class of problem that can be solved much faster than
linear programming while still retaining such nice properties as integrality of solution and appeal
of concept.
Like the maximum ow problem, it considers ows in networks with capacities. Like the shortest
path problem, it considers a cost for ow through an arc. Like the transportation problem, it allows
multiple sources and destinations. In fact, all of these problems can be seen as special cases of the
minimum cost ow problem.
Consider a directed network with n nodes. The decision variables are xij , the ow through arc
(i; j ). The given information includes:
cij : cost per unit of ow from i to j (may be negative),
uij : capacity (or upper bound) on ow from i to j ,
bi: net ow generated at i.
This last value has a sign convention:
bi > 0 if i is a supply node,
bi < 0 if i is a demand node,
bi = 0 if i is a transshipment node.
The objective is to minimize the total cost of sending the supply through the network to satisfy
the demand.
Note that for this model, it is not necessary that every arc exists. We will use the convention
that summations are only taken over arcs that exist. The linear programming formulation for this
problem is: PP x
Minimize Pi j cijP ij
Subject to j xij , j xji = bi for all nodes i,
0 xij uij for all arcs (i; j ).
P
Again, we will assume that the network is balanced, so i bi = 0, since dummies can be added
as needed. We also still have a nice integrality property. If all the bi and uij are integral, then the
resulting solution to the linear program is also integral.
Minimum cost network ows are solved by a variation of the simplex algorithm and can be
solved more than 100 times faster than equivalently sized linear programs. From a modeling point
of view, it is most important to know the sort of things that can and cannot be modeled in a single
network ow problem:
Can do
11.4. UNIFYING MODEL: MINIMUM COST NETWORK FLOWS 143
1. Lower bounds on arcs. If a variable xij has a lower bound of lij , upper bound of uij ,
and cost of cij , change the problem as follows:
Replace the upper bound with uij , lij ,
Replace the supply at i with bi , lij ,
Replace the supply at j with bi + lij ,
Now you have a minimum cost ow problem. Add cij lij to the objective after solving
and lij to the ow on arc (i; j ) to obtain a solution of the original problem.
2. Upper bounds on ow through a node. Replace the node i with nodes i0 and i00. Create
an arc from i0 to i00 with the appropriate capacity, and cost 0. Replace every arc (j; i)
with one from j to i0 and every arc (i; j ) with one from i00 to j . Lower bounds can also
be handled this way.
3. Convex, piecewise linear costs on arc ows (for minimization). This is handled by in-
troducing multiple arcs between the nodes, one for each portion of the piecewise linear
function. The convexity will assure that costs are handled correctly in an optimal solu-
tion.
Can't do
1. Fixed cost to use a node.
2. Fixed cost to use an arc.
3. \Proportionality of ow." That is, if one unit enters node i, then you insist that .5 units
go to node j and .5 to node k.
4. Gains and losses of ow along arcs, as in power distribution.
Note that although these cannot be done in a single network, it may be possible to use the
solutions to multiple networks to give you an answer. For instance, if there is only one arc
with a xed cost, you can solve both with and without the arc to see if it is advantageous to
pay the xed cost.
Exercise 82 Here is an example of a problem that doesn't look like a network ow problem, but
it really is:
A company must meet the following demands for cash at the beginning of each of the next six
months:
Month 1 2 3 4 5 6
Needs $200 $100 $50 $80 $160 $140
At the beginning of month 1, the company has $150 in cash and $200 worth of bond 1, $100
worth of bond 2 and $400 worth of bond 3. Of course, the company will have to sell some bonds to
meet demands, but a penalty will be charged for any bonds sold before the end of month 6. The
penalties for selling $1 worth of each bond are shown in the table below.
Month of sale
1 2 3 4 5 6
1 $0.07 $0.06 $0.06 $0.04 $0.03 $0.03
Bond 2 $0.17 $0.17 $0.17 $0.11 $0 $0
3 $0.33 $0.33 $0.33 $0.33 $0.33 $0
(a) Assuming that all bills must be paid on time, formulate a balanced transportation problem
that can be used to minimize the cost of meeting the cash demands for the next six months.
(b) Assume that payment of bills can be made after they are due, but a penalty of $0.02 per month
is assessed for each dollar of cash demands that is postponed for one month. Assuming all
bills must be paid by the end of month 6, develop a transshipment model that can be used
to minimize the cost of paying the next six months' bills.
[ Hint: Transshipment points are needed, in the form Ct = cash available at beginning of
month t after bonds for month t have been sold, but before month t demand is met. Shipments
into Ct occur from bond sales and Ct,1. Shipments out of Ct occur to Ct+1 and demands for
months 1; 2; : : :; t.]
Exercise 83 Oil R' Us has oil elds in San Diego and Los Angeles. The San Diego eld can
produce up to 500,000 barrels per day (bpd); Los Angeles can produce up to 400,000 bpd. Oil
is sent to a renery, either in Dallas or Houston. It costs $700 to rene 100,000 bpd in Dallas
and $900 to rene 100,000 bpd in Houston. The rened oil is then shipped to either Chicago or
New York. Chicago requires exactly 400,000 bpd and New York requires exactly 300,000 bpd. The
shipping costs (per 100,000 bpd) are given as follows:
Dallas Houston New York Chicago
LA $300 $110 | |
San Diego $420 $100 | |
Dallas | | $450 $550
Houston | | $470 $530
(a) Formulate this problem as a minimum cost ow problem. Clearly give the network. Give
the cost and capacity on each arc and the net supply requirement on each node.
(b) Suppose Dallas could process no more than 300,000 bpd. How would you modify your
formulation?
Exercise 84 A company produces a single product at two plants, A and B. A customer demands
12 units of the product this month, and 20 units of the product next month. Plant A can produce
7 items per month. It costs $12 to produce each item at A this month and $18 to produce at A
next month. Plant B can produce 14 items per month. It costs $14 to produce each item at B this
month and $20 to produce each item at B next month. Production from this month can be held in
inventory to satisfy next month's demand at a cost of $3/unit.
(a) Formulate the problem of determining the least cost way of meeting demand as a balanced
transportation problem. Give the costs and node supplies, and describe the interpretation of each
node and arc.
(b) Suppose no more than six items can be held in inventory between this month and next
month. Formulate the resulting problem as a minimum cost ow problem. Draw the network, give
the interpretation for each node and arc, and give the supply/demand at each node and the cost
and capacity for each arc.
11.5 Generalized Networks

The nal model I would like to familiarize you with is called the generalized network model. In this
model, there may be gains or loses as ow goes along an arc. Each arc has a multiplier to represent
11.6. CONCLUSIONS 145
these gains and loses. For instance, if 3 units of ow enter an arc with a multiplier of .5, then 1.5
unit of ow exit the arc. Such a model can still be represented as a linear program, and there are
specialized techniques that can solve such models much faster than linear programming (but a bit
slower than regular network ow problems). The optimal solution need not be integer however.
The standard example of a generalized network is in power generation: as electricty moves over
wires, there is some unavoidable loss along the way. This loss is represented as a mutliplier. Here
is another example on how generalized networks might be used:
Consider the world's currency market. Given two currencies, say the Yen and the USDollar,
there is an exchange rate between them (currently about 110 Yen to the Dollar). It is axiomatic of
a arbitrage-free market that there is no method of converting, say, a Dollar to Yen then to Deutsch
Marks, to Francs, then Pounds, and to Dollars so that you end up with more than a dollar. How
would you recognize when there is an arbitrage possibility?
For this model, we use a node to represent each currency. An arc between currency i and j
gives the exchange rate from i to j . We begin with a single dollar and wish to send ow through
the network and maximize the amount we generate back at the USDollar node. This can be drawn
as in gure 11.5.
Mark
3.3823
.6659 .013493
21.933
Franc .045593 Yen
.1966 111.52
5.0852 .008966
USD Max flow out
Figure 11.5: Financial Exchange Network

These are actual trades made on November 10, 1996. It is not obvious, but the Dollar-Yen-
Mark-Dollar conversion actually makes $0.002. How would you formulate a linear program to
recognize this?
11.6 Conclusions
One class can only provide a bare introduction to this area. The points we would like you to take
away are:
1) Networks are an important subclass of linear programs that are intuitive, easy to solve, and
have nice integrality properties.
2) Problems that might not look like networks might be networks.
3) Networks provide a useful way to think about problems even if there are additional constraints
or variables that preclude use of networks for modeling the whole problem.
In practice, you would generally save using the fastest network codes until the nal implemen-
tation phase. Until then, linear programming codes will tend to be suciently fast to prove the
concepts.
Chapter 12
Data Envelopment Analysis

Data Envelopment Analysis (DEA) is an increasingly popular management tool. This write-up is
an introduction to Data Envelopment Analysis (DEA) for people unfamiliar with the technique.
For a more in-depth discussion of DEA, the interested reader is referred to Seiford and Thrall [1990]
or the seminal work by Charnes, Cooper, and Rhodes [1978].
DEA is commonly used to evaluate the eciency of a number of producers. A typical statistical
approach is characterized as a central tendency approach and it evaluates producers relative to an
average producer In contrast, DEA compares each producer with only the "best" producers. By
the way, in the DEA literature, a producer is usually referred to as a decision making unit or DMU.
DEA is not always the right tool for a problem but is appropriate in certain cases. (See Strengths
and Limitations of DEA.)
In DEA, there are a number of producers. The production process for each producer is to take
a set of inputs and produce a set of outputs. Each producer has a varying level of inputs and gives
a varying level of outputs. For instance, consider a set of banks. Each bank has a certain number
of tellers, a certain square footage of space, and a certain number of managers (the inputs). There
are a number of measures of the output of a bank, including number of checks cashed, number of
loan applications processed, and so on (the outputs). DEA attempts to determine which of the
banks are most ecient, and to point out specic ineciencies of the other banks.
A fundamental assumption behind this method is that if a given producer, A, is capable of
producing Y(A) units of output with X(A) inputs, then other producers should also be able to
do the same if they were to operate eciently. Similarly, if producer B is capable of producing
Y(B) units of output with X(B) inputs, then other producers should also be capable of the same
production schedule. Producers A, B, and others can then be combined to form a composite
producer with composite inputs and composite outputs. Since this composite producer does not
necessarily exist, it is typically called a virtual producer.
The heart of the analysis lies in nding the "best" virtual producer for each real producer. If
the virtual producer is better than the original producer by either making more output with the
same input or making the same output with less input then the original producer is inecient. The
subtleties of DEA are introduced in the various ways that producers A and B can be scaled up or
down and combined.
12.1 Numerical Example

To illustrate how DEA works, let's take an example of three banks. Each bank has exactly 10
tellers (the only input), and we measure a bank based on two outputs: Checks cashed and Loan
147
148 CHAPTER 12. DATA ENVELOPMENT ANALYSIS
applications. The data for these banks is as follows:
Bank A: 10 tellers, 1000 checks, 20 loan applications
Bank B: 10 tellers, 400 checks, 50 loan applications
Bank C: 10 tellers, 200 checks, 150 loan applications
Now, the key to DEA is to determine whether we can create a virtual bank that is better than
one or more of the real banks. Any such dominated bank will be an inecient bank.
Consider trying to create a virtual bank that is better than Bank A. Such a bank would use no
more inputs than A (10 tellers), and produce at least as much output (1000 checks and 20 loans).
Clearly, no combination of banks B and C can possibly do that. Bank A is therefore deemed to be
ecient. Bank C is in the same situation.
However, consider bank B. If we take half of Bank A and combine it with half of Bank C, then
we create a bank that processes 600 checks and 85 loan applications with just 10 tellers. This
dominates B (we would much rather have the virtual bank we created than bank B). Bank B is
therefore inecient.
Another way to see this is that we can scale down the inputs to B (the tellers) and still have at
least as much output. If we assume (and we do), that inputs are linearly scalable, then we estimate
that we can get by with 6.3 tellers. We do that by taking .34 times bank A plus .29 times bank B.
The result uses 6.3 tellers and produces at least as much as bank B does. We say that bank B's
eciency rating is .63. Banks A and C have an eciency rating of 1.
12.2 Graphical Example

The single input two-output or two input-one output problems are easy to analyze graphically.
The previous numerical example is now solved graphically. (An assumption of constant returns to
scale is made and explained in detail later.) The analysis of the eciency for bank B looks like the
following:
Loans
150 C
V
B
50
A
20
O 200 400 1000 Checks

12.3. USING LINEAR PROGRAMMING 149
If it is assumed that convex combinations of banks are allowed, then the line segment connecting
banks A and C shows the possibilities of virtual outputs that can be formed from these two banks.
Similar segments can be drawn between A and B along with B and C. Since the segment AC lies
beyond the segments AB and BC, this means that a convex combination of A and C will create
the most outputs for a given set of inputs.
This line is called the eciency frontier. The eciency frontier denes the maximum combina-
tions of outputs that can be produced for a given set of inputs.
Since bank B lies below the eciency frontier, it is inecient. Its eciency can be determined
by comparing it to a virtual bank formed from bank A and bank C. The virtual player, called V,
is approximately 54% of bank A and 46% of bank C. (This can be determined by an application
of the lever law. Pull out a ruler and measure the lengths of AV, CV, and AC. The percentage of
bank C is then AV/AC and the percentage of bank A is CV/AC.)
The eciency of bank B is then calculated by nding the fraction of inputs that bank V would
need to produce as many outputs as bank B. This is easily calculated by looking at the line from
the origin, O, to V. The eciency of player B is OB/OV which is approximately 63%. This gure
also shows that banks A and C are ecient since they lie on the eciency frontier. In other words,
any virtual bank formed for analyzing banks A and C will lie on banks A and C respectively.
Therefore since the eciency is calculated as the ratio of OA/OV or OA/OV, banks A and C will
have eciency scores equal to 1.0.
The graphical method is useful in this simple two dimensional example but gets much harder in
higher dimensions. The normal method of evaluating the eciency of bank B is by using an linear
programming formulation of DEA.
Since this problem uses a constant input value of 10 for all of the banks, it avoids the com-
plications caused by allowing dierent returns to scale. Returns to scale refers to increasing or
decreasing eciency based on size. For example, a manufacturer can achieve certain economies of
scale by producing a thousand circuit boards at a time rather than one at a time - it might be only
100 times as hard as producing one at a time. This is an example of increasing returns to scale
(IRS.)
On the other hand, the manufacturer might nd it more than a trillion times as dicult to
produce a trillion circuit boards at a time though because of storage problems and limits on the
worldwide copper supply. This range of production illustrates decreasing returns to scale (DRS.)
Combining the two extreme ranges would necessitate variable returns to scale (VRS.)
Constant Returns to Scale (CRS) means that the producers are able to linearly scale the inputs
and outputs without increasing or decreasing eciency. This is a signicant assumption. The
assumption of CRS may be valid over limited ranges but its use must be justied. As an aside,
CRS tends to lower the eciency scores while VRS tends to raise eciency scores.
12.3 Using Linear Programming

Data Envelopment Analysis, is a linear programming procedure for a frontier analysis of inputs and
outputs. DEA assigns a score of 1 to a unit only when comparisons with other relevant units do not
provide evidence of ineciency in the use of any input or output. DEA assigns an eciency score
less than one to (relatively) inecient units. A score less than one means that a linear combination
of other units from the sample could produce the same vector of outputs using a smaller vector of
inputs. The score re ects the radial distance from the estimated production frontier to the DMU
under consideration.
There are a number of equivalent formulations for DEA. The most direct formulation of the
exposition I gave above is as follows:
Let Xi be the vector of inputs into DMU i. Let Yi be the corresponding vector of outputs. Let
X0 be the inputs into a DMU for which we want to determine its eciency and Y0 be the outputs.
So the X 's and the Y 's are the data. The measure of eciency for DMU0 is given by the following
linear program:
Min
s:t: P X X
P i Yi Y 0
i i 0
0
where i is the weight given to DMU i in its eorts to dominate DMU 0 and is the eciency
of DMU 0. So the 's and are the variables. Since DMU 0 appears on the left hand side of
the equations as well, the optimal cannot possibly be more than 1. When we solve this linear
program, we get a number of things:
1. The eciency of DMU 0 (), with = 1 meaning that the unit is ecient.
2. The unit's \comparables" (those DMU with nonzero ).
3.
P
The \goal" inputs (the dierence between X0 and iXi )
4.
P
Alternatively, we can keep inputs xed and get goal outputs ( 1 i Yi )
DEA assumes that the inputs and outputs have been correctly identied. Usually, as the number
of inputs and outputs increase, more DMUs tend to get an eciency rating of 1 as they become
too specialized to be evaluated with respect to other units. On the other hand, if there are too few
inputs and outputs, more DMUs tend to be comparable. In any study, it is important to focus on
correctly specifying inputs and outputs.
Example 12.3.1 Consider analyzing the eciencies of 3 DMUs where 2 inputs and 3 outputs are
used. The data is as follows:
DMU Inputs Outputs
1 5 14 9 4 16
2 8 15 5 7 10
3 7 12 4 9 13
The linear programs for evaluating the 3 DMUs are given by:
LP for evaluating DMU 1:
min THETA
st
5L1+8L2+7L3 - 5THETA <= 0
14L1+15L2+12L3 - 14THETA <= 0
9L1+5L2+4L3 >= 9
4L1+7L2+9L3 >= 4
16L1+10L2+13L3 >= 16
L1, L2, L3 >= 0

min THETA
st
5L1+8L2+7L3 - 8THETA <= 0
14L1+15L2+12L3 - 15THETA <= 0
9L1+5L2+4L3 >= 5
4L1+7L2+9L3 >= 7
16L1+10L2+13L3 >= 10
L1, L2, L3 >= 0

min THETA
st
5L1+8L2+7L3 - 7THETA <= 0
14L1+15L2+12L3 - 12THETA <= 0
9L1+5L2+4L3 >= 4
4L1+7L2+9L3 >= 9
16L1+10L2+13L3 >= 13
L1, L2, L3 >= 0
The solution to each of these is as follows:

DMU 1.
Adjustable Cells
Cell Name Value Cost Coefficien Increase Decrease
$B$10 theta 1 0 1 1E+30 1
$B$11 L1 1 0 0 0.92857142 0.619047619
$B$12 L2 0 0.24285714 0 1E+30 0.242857143
$B$13 L3 0 0 0 0.36710963 0.412698413
Constraints
$B$16 IN1 -0.103473 0 0 1E+30 0
$B$17 IN2 -0.289724 -0.07142857 0 0 1E+30
$B$18 OUT1 9 0.085714286 9 0 0
$B$19 OUT2 4 0.057142857 4 0 0
$B$20 OUT3 16 0 16 0 1E+30
DMU 2.
Adjustable Cells
$B$10 theta 0.773333333 0 1 1E+30 1
$B$11 L1 0.261538462 0 0 0.866666667 0.577777778
$B$12 L2 0 0.22666667 0 1E+30 0.226666667
$B$13 L3 0.661538462 0 0 0.342635659 0.385185185
Constraints
$B$16 IN1 -0.24820512 0 0 1E+30 0.248205128
$B$17 IN2 -0.452651 -0.0666667 0 0.46538461 1E+30
$B$18 OUT1 5 0.08 5 10.75 0.655826558
$B$19 OUT2 7 0.0533333 7 1.05676855 3.41509434
$B$20 OUT3 12.78461538 0 10 2.78461538 1E+30
DMU 3.
Adjustable Cells
$B$10 theta 1 0 1 1E+30 1
$B$11 L1 0 0 0 1.08333333 0.722222222
$B$12 L2 0 0.283333333 0 1E+30 0.283333333
$B$13 L3 1 0 0 0.42829457 0.481481481
Constraints
$B$16 IN1 -0.559375 0 0 1E+30 0
$B$17 IN2 -0.741096 -0.08333333 0 0 1E+30
$B$18 OUT1 4 0.1 4 16.25 0
$B$19 OUT2 9 0.066666667 9 0 0
$B$20 OUT3 13 0 13 0 1E+30
Note that DMUs 1 and 3 are overall ecient and DMU 2 is inecient with an eciency rating
of 0.773333.
Hence the ecient levels of inputs and outputs for DMU 2 are given by:
Ecient levels of Inputs:
" # " # " #
5 7
0:261538 14 + 0:661538 12 = 11:6 5 : 938
Ecient levels of Outputs:

2 3 2 3 2 3
9 4 5
0:261538 64 4 75 + 0:661538 64 9 75 = 64 7 75
16 13 12:785
Note that the outputs are at least as much as the outputs currently produced by DMU 2 and
inputs are at most as big as the 0.773333 times the inputs of DMU 2. This can be used in two
dierent ways: The inecient DMU should target to cut down inputs to equal at most the ecient
levels. Alternatively, an equivalent statement can be made by nding a set of ecient levels of
inputs and outputs by dividing the levels obtained by the eciency of DMU 2. This focus can then
be used to set targets primarily for outputs rather than reduction of inputs.
Alternate Formulation
There is another, probably more common formulation, that provides the same information. We
can think of DEA as providing a price on each of the inputs and a value for each of the outputs.
The eciency of a DMU is simply the ratio of the inputs to the outputs, and is constrained to be
no more than 1. The prices and values have nothing to do with real prices and values: they are an
articial construct. The goal is to nd a set of prices and values that puts the target DMU in the
best possible light. The goal, then is to
Max uT Y0
v T X0
s:t: uT Yj 1; j = 0; :::; n;
v T Xj
uT 0;
vT 0:
Here, the variables are the u's and the v 's. They are vectors of prices and values respectively.
This fractional program can be equivalently stated as the following linear programming problem
(where Y and X are matrices with columns Yj and Xj respectively).
Max uT Y0
s:t: vT X0 = 1;
uT Y , v T X 0;
uT 0;
vT 0:
We denote this linear program by (D). Let us compare it with the one introduced earlier, which
we denote by (P):
Min
s:t: P X X
P iY i Y 0
i i 0
0:
To x ideas, let us write out explicitely these two formulations for DMU 2, say, in our example.
Formulation (P) for DMU 2:
min THETA
st
-5 L1 - 8 L2 - 7 L3 + 8 THETA >= 0
-14L1 -15 L2 -12 L3 + 15THETA >= 0
9 L1 + 5 L2 + 4 L3 >= 5
4 L1 + 7 L2 + 9 L3 >= 7
16L1 +10 L2 +13 L3 >= 10
L1>=0, L2>=0, L3>=0
Formulation (D) for DMU 2:

max 5 U1 + 7 U2 + 10 U3
st
- 5 V1 - 14V2 + 9 U1 + 4 U2 + 16 U3 <= 0
- 8 V1 - 15V2 + 5 U1 + 7 U2 + 12 U3 <= 0
- 7 V1 - 12V2 + 4 U1 + 9 U2 + 13 U3 <= 0
8 V1 + 15V2 = 1
V1>=0, V2>=0, U1>=0, U2>=0, U3>=0
Formulations (P) and (D) are dual linear programs! These two formulations actually give the
same information. You can read the solution to one from the shadow prices of the other. We will
not discuss linear programming duality in this course. You can learn about it in some of the OR
electives.
Exercise 85 Consider the following baseball players:
Name At Bat Hits HomeRuns
Martin 135 41 6
Polcovich 82 25 1
Johnson 187 40 4
(You need know nothing about baseball for this question). In order to determine the eciency
of each of the players, At Bats is dened as an input while Hits and Home Runs are outputs.
Consider the following linear program and its solution:
MIN THETA
SUBJECT TO
- 187 THETA + 135 L1 + 82 L2 + 187 L3 <= 0
41 L1 + 25 L2 + 40 L3 >= 40
6 L1 + L2 + 4 L3 >= 4
L1, L2, L3 >= 0
12.4. APPLICATIONS 155
Adjustable Cells
Final Reduced Objective
Cell Name Value Cost Coefficient
$B$10 THETA 0.703135 0 1
$B$11 L1 0.550459 0 0
$B$12 L2 0.697248 0 0
$B$13 L3 0 0.296865 0
Constraints
Final Shadow Constraint
Cell Name Value Price R.H. Side
$B$16 AT BATS 0 -0.005348 0
$B$17 HITS 0 0.017515 40
$B$18 HOME RUNS 0 0.000638 4
(a) For which player is this a DEA analysis? Is this player ecient? What is the eciency
rating of this player? Give the \virtual producer" that proves that eciency rating (you should
give the At bats, Hits, and Home Runs for this virtual producer).
(b) Formulate the linear program for Jonhson using the alternate formulation and solve using
Solver. Compare the \Final Value" and \Shadow Price" columns from your Solver output with the
solution given above.
12.4 Applications
The simple bank example described earlier may not convey the full view on the usefulness of DEA.
It is most useful when a comparison is sought against "best practices" where the analyst doesn't
want the frequency of poorly run operations to aect the analysis. DEA has been applied in
many situations such as: health care (hospitals, doctors), education (schools, universities), banks,
manufacturing, benchmarking, management evaluation, fast food restaurants, and retail stores.
The analyzed data sets vary in size. Some analysts work on problems with as few as 15 or 20
DMUs while others are tackling problems with over 10,000 DMUs.
12.5 Strengths and Limitations of DEA

As the earlier list of applications suggests, DEA can be a powerful tool when used wisely. A few of
the characteristics that make it powerful are:
DEA can handle multiple input and multiple output models.
It doesn't require an assumption of a functional form relating inputs to outputs.
DMUs are directly compared against a peer or combination of peers.
Inputs and outputs can have very dierent units. For example, X1 could be in units of lives
saved and X2 could be in units of dollars without requiring an a priori tradeo between the
two.
The same characteristics that make DEA a powerful tool can also create problems. An analyst
should keep these limitations in mind when choosing whether or not to use DEA.
Since DEA is an extreme point technique, noise (even symmetrical noise with zero mean)
such as measurement error can cause signicant problems.
DEA is good at estimating "relative" eciency of a DMU but it converges very slowly to
"absolute" eciency. In other words, it can tell you how well you are doing compared to your
peers but not compared to a "theoretical maximum."
Since DEA is a nonparametric technique, statistical hypothesis tests are dicult and are the
focus of ongoing research.
Since a standard formulation of DEA creates a separate linear program for each DMU, large
problems can be computationally intensive.
12.6 References
DEA has become a popular subject since it was rst described in 1978. There have been hundreds
of papers and technical reports published along with a few books. Technical articles about DEA
have been published in a wide variety of places making it hard to nd a good starting point. Here
are a few suggestions as to starting points in the literature.
1. Charnes, A., W.W. Cooper, and E. Rhodes. "Measuring the eciency of decision making
units." European Journal of Operations Research (1978): 429-44.
2. Banker, R.D., A. Charnes, and W.W. Cooper. "Some models for estimating technical and
scale ineciencies in data envelopment analysis." Management Science 30 (1984): 1078-92.
3. Dyson, R.G. and E. Thanassoulis. "Reducing weight exibility in data envelopment analysis."
Journal of the Operational Research Society 39 (1988): 563-76.
4. Seiford, L.M. and R.M. Thrall. "Recent developments in DEA: the mathematical program-
ming approach to frontier analysis." Journal of Econometrics 46 (1990): 7-38.
5. Ali, A.I., W.D. Cook, and L.M. Seiford. "Strict vs. weak ordinal relations for multipliers in
data envelopment analysis." Management Science 37 (1991): 733-8.
6. Andersen, P. and N.C. Petersen. "A procedure for ranking ecient units in data envelopment
analysis." Management Science 39 (1993): 1261-4.
7. Banker, R.D. "Maximum likelihood, consistency and data envelopment analysis: a statistical
foundation." Management Science 39 (1993): 1265-73.
The rst paper was the original paper describing DEA and results in the abbreviation CCR
for the basic constant returns-to-scale model. The Seiford and Thrall paper is a good overview of
the literature. The other papers all introduce important new concepts. This list of references is
certainly incomplete.
A good source covering the eld of productivity analysis is The Measurement of Productive
Eciency edited by Fried, Lovell, and Schmidt, 1993, from Oxford University Press. There is also
12.6. REFERENCES 157
a recent book from Kluwer Publishers, Data Envelopment Analysis: Theory, Methodology, and
Applications by Charnes, Cooper, Lewin, and Seiford.
To stay more current on the topics, some of the most important DEA articles appear in Man-
agement Science, The Journal of Productivity Analysis, The Journal of the Operational Research
Society, and The European Journal of Operational Research. The latter just published a special
issue, "Productivity Analysis: Parametric and Non-Parametric Approaches" edited by Lewin and
Lovell which has several important papers.

Trick

Uploaded by

Trick

Uploaded by

Quantitative Methods

for the Management Sciences

11.5 Generalized Networks : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 144

0.2 Grading Components

Date Topic Chapter Work Due

0.4 Mathematical Methods in Business

Basic Linear Algebra

1.1 Linear Equations

It is generally true that a system of m linear equations in n variables has

1.2 Operations on Vectors and Matrices

Addition of two matrices of same dimensions:

The following standard de nitions will be useful:

1.3 Linear Combinations

This is none other than a system of two linear equations:

 A vector b having n components is a linear combination of the k

1.4 Inverse of a Square Matrix

Exercise 13 Solve the system

1.6 Positive De nite Matrices

Exercise 18 Compute the derivatives of the following functions:

2.2 Maximum and Minimum

1. at the boundary of the domain,

The revenue at price x is

a) Compute the derivative A0(r).

2.3 Binary Search

The yield r of the bond is given by the equation

Golden Section Search

(ii) f has a unique local maximum in the interval [a; b].

e) If both were done simultaneously?

3.2 Maximum and Minimum

3.3 Global Optima

4.1.1 Geometric Interpretation

4.1.2 Economic Interpretation

Figure 4.1: Geometric interpretation

L(x; y; ) = ax2 + by2 + (Q , x , y )

4.2 Equality and Inequality Constraints

j (hj (x) , dj ) = 0 (Complementarity)

j (hj (x) , dj ) = 0 (Complementarity)

L(x; y; 1; 2) = ,(x , 2)2 , 2(y , 1)2 + 1(3 , x , 4y ) + 2(0 + x , y )

2. Add a new complementarity constraint:

Review of Optimality Conditions.

Modeling with Linear Programming

Figure 5.1: Single Constraint

Figure 5.2: All Constraints

Figure 5.3: Finding Optimal Pro t

5.1.3 Using Solver

5.2 Second Example: Market Allocations

Figure 5.4: Marketing example

Figure 5.5: Marketing example with Isocost Line

5.2.3 Using Solver

5.3 Common Features

5.4 Linear Programming Models

5.4.1 Diet Problem

5.4.2 Workforce Planning

5.4.3 Financial Portfolio

Table 6.1: Re nery Operations

$702,000 savings per year is equivalent to the $4 million investment

Memo to: Global Oil Headquarters

Memo to: Global Oil Headquarters

Memo to: Global Oil Headquarters

The Simplex Method

7.1 Linear Programs in Standard Form

Max c1x1 + c2x2 + : : : + cn xn

Our goal is to maximize z , while satisfying these equations and, in addition, x1  0, x2  0,

Figure 7.1: Graphical Interpretation

Sensitivity Analysis for Linear

8.1 Tableau Sensitivity Analysis

Here, the optimal tableau is :

8.1.3 New Variable

MAX 900 X1 + 1000 X2 + 1200 X3

with the optimal tableau

8.2 Solver Output

Attached is the Solver output for this problem.

Here are some questions to be answered:

4. How much are we willing to pay for a labor hour?

Attached is the Solver output.

Here are some questions to answer:

The following standard denitions will be useful:

A vector b having n components is a linear combination of the k

1.6 Positive Denite Matrices

L(x; y; ) = ax2 + by2 + (Q , x , y )

j (hj (x) , dj ) = 0 (Complementarity)

j (hj (x) , dj ) = 0 (Complementarity)

L(x; y; 1; 2) = ,(x , 2)2 , 2(y , 1)2 + 1(3 , x , 4y ) + 2(0 + x , y )

Figure 5.3: Finding Optimal Prot

Table 6.1: Renery Operations

Our goal is to maximize z , while satisfying these equations and, in addition, x1 0, x2 0,

1. What are the production quantities and projected prot?

LP for evaluating DMU 2:

LP for evaluating DMU 3:

Ecient levels of Outputs: