Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Book

Download as pdf or txt
Download as pdf or txt
You are on page 1of 362

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/327076435

Introduction to Mathematical Economics

Book · August 2018

CITATIONS READS
0 938

1 author:

Wilson Mixon
Berry College
79 PUBLICATIONS   175 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Classical Mechanics with Maxima CAS View project

Undergrad math econ text w/CAS and LaTeX files View project

All content following this page was uploaded by Wilson Mixon on 17 August 2018.

The user has requested enhancement of the downloaded file.


Introduction to Mathematical Economics

Wilson Mixon

May 21, 2018


i

This book is dedicated to the Maxima development team.


Preface

The title page of this book is a bit misleading. I (Wilson Mixon) am not
the author of most of the book’s material. I am grateful to have received
the permission of Professors Anthony L. Ostrosky, Jr. and James V. Koch
to use their textbook [16] as the basis for this project. I have edited their
material slightly and have added some material. Mainly, I have incorporated
material from the Maxima open-source computer algebra system. Also, I
have extended the original discussion of sets and lists.
I hope that this project reflects well on the work of Professors Ostrosky and
Koch. I accept responsibility for any shortcomings that I have introduced in
this rendering of their work.
It is fitting that the bulk of this preface should be in the words of the original
authors:

. . . Economists are called on with increasing frequency to ap-


ply their logic and tools to a variety of societal problems: pollu-
tion; depletion of natural resources; crime; urban sprawl; taxa-
tion; food production; the depreciating dollar–the list swells daily.
The ability of the economist to speak to these problems reflects a
well-developed body of theory, modes of analysis that emphasize
logic, and sophisticated quantitative tools. . . .
Mathematics has played a central role in enabling economists
to rigorously state their theorems, with emphasis on logical infer-
ence, and in enabling them to . . . [test] the empirical validity of
their theories. The primary aim of this book is to show fledgling
students how, where, when, and why they can appropriately uti-
lize mathematics in economics and business. The student who
masters the mathematical tools presented in this book will not

ii
iii

only be able to read and apply [much] of the “language” that


modern economic theory uses, but also he or she will learn (and
perhaps for the first time understand) a great deal of economic
theory. If the readers of this book are similar to the students
at Illinois State university [where the authors taught],then one
can expect that this combination of mathematics and economics
will turn on lights, open doors, and yield deeper understand-
ings. Many are the “math econ” students who have suddenly ex-
claimed, “Now I see what my Econ 100 instructor really meant!”
We do not claim to present all the many applications of math-
ematics to economics and business in this book. This book is
a well-defined one-semester introductory approach to the use of
mathematics in economics and business. . . .
[Even so, an] outstanding feature of this book is the plentiful
use of examples and applications. Each chapter contains a large
section that is entirely devoted to applications of the mathemat-
ical tools; one entire chapter is devoted to specific applications
of matrix algebra. Several examples, such as Stigler’s famous
“diet problem,” are used on a number of occasions in order to
demonstrate the power of applied mathematics. . . .
The organization of the book is based on the view that a thor-
ough review of basic precalculus mathematics and algebra is the
correct place to start. The differential calculus, with its many
applications, is then introduced. Maximization and minimiza-
tion techniques are plentifully’ used and illustrated. The integral
calculus is the next. major topic; two chapters are devoted to its
exposition and application. Finally we cover matrix algebra, and
devote an entire chapter to linear programming and input-output
analysis in a matrix-algebra context. The overall organization of
the book stresses a building-block approach, whereby each newly
introduced topic depends on the topics previously covered.

Much of the material in this text is produced by Maxima, the open-source


computer algebra system. The text is, however, self-contained. Some of the
references to Maxima might affect the flow of the text, but the content can
be accessed without the use of Maxima.
This text serves as an introduction to some of the mathematics that pertains
to economics. It is not a full introduction to Maxima or to the wxMax-
iv

ima user interface. To incorporate Maxima into your study of this material,
look at these two sites: http:// maxima.sourceforge.net/ and http:// an-
drejv.github.io/wxmaxima/. Both sites have links to documentation.1
Another site that readers of this text should visit is http://statmath.wu.ac.at/
leydold/maxima/ [9]. This site contains the text Introduction to Maxima for
Economics, which has a quite complete introduction to both Maxima and
wxMaxima. It also provides an briefer development of much of the material
in this text. Finally, it contains some more advanced material, especially the
treatment of ordinary differential equations.
I am coauthor (with Michael Hammock) of a textbook that develops micro-
economic theory more fully than the confines of the present text allow. That
text, Microeconomic Theory and Computation[7], also provides more detail
on the use of Maxima than is provided here.

1
You probably will not want to download wxMaxima. Windows and MacOS users
should download an executable file that will install wxMaxima. Linux users can access
Maxima and wxMaxima from their repositories.
Acknowledgements

This effort is dedicated to the Maxima team, which maintains and continually
improves this remarkable piece of software. Among the members of that
team, I thank Robert Dodier and Andrej Vodopivec for encouraging my
early efforts to produce material to illustrate Maxima’s power and usefulness.
A third team member, Gunter Königsmann, continues to provide help and
insight in the use of the wxMaxima graphical user interface which greatly
facilitates the use of Maxima, especially by newcomers.
I thank Michael Hammock for working with me on the previously-mentioned
text and for maintaining a website that contains a large amount of material
on the use of Maxima in economic analysis (www.wxmaximaecon.com).
Finally, and above all, I thank my wife Barbara. A suggestion that she
made a few years ago pointed me toward Maxima. Since then, I’ve repaid
her by bending her ears about the most recent tidbit that I’ve discovered by
working on various projects and by providing material for her to proofread.
Remarkably, she continues to encourage me and even offers to proofread new
material. Who would believe such a report?

v
Contents

1 The Role and Power of Mathematics 1


1.1 Stigler’s Diet Problem . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Analysis Using a Computer Algebra System . . . . . . . . . . 5
1.3 The Diet Problem in Maxima . . . . . . . . . . . . . . . . . . 7
1.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Questions and Problems . . . . . . . . . . . . . . . . . . . . . 9

2 Variables, Sets, Lists, and Relations 10


2.1 Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Equations, Roots, and Constants . . . . . . . . . . . . . . . . 14
2.3 The Real Number System . . . . . . . . . . . . . . . . . . . . 16
2.4 Sets and Set Theory . . . . . . . . . . . . . . . . . . . . . . . 18
2.5 Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.6 Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.7 Questions and Problems . . . . . . . . . . . . . . . . . . . . . 33

3 Rectangular Coordinates and Functions 34


3.1 Rectangular Coordinates . . . . . . . . . . . . . . . . . . . . . 35
3.2 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.3 Summation and Multiplication . . . . . . . . . . . . . . . . . . 64
3.4 Questions and Problems . . . . . . . . . . . . . . . . . . . . . 67

vi
CONTENTS vii

4 Limits, Continuity, and Differentiability 70


4.1 Limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2 Extensions of the Limit Concept . . . . . . . . . . . . . . . . . 76
4.3 Continunity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.4 The Derivative of a Function . . . . . . . . . . . . . . . . . . . 87

5 Differentiation: Univariate Functions 95


5.1 Rules for Differentiation . . . . . . . . . . . . . . . . . . . . . 95
5.2 Higher-Order Derivatives . . . . . . . . . . . . . . . . . . . . . 106
5.3 Economic Applications of Derivatives . . . . . . . . . . . . . . 109

6 Differentiation II 120
6.1 Partial Differentiation . . . . . . . . . . . . . . . . . . . . . . 120
6.2 Rules of Differentiation . . . . . . . . . . . . . . . . . . . . . . 121
6.3 Higher-order Partial Derivatives . . . . . . . . . . . . . . . . . 132
6.4 Applications of Partial Derivatives . . . . . . . . . . . . . . . . 134
6.5 Questions and Problems . . . . . . . . . . . . . . . . . . . . . 156

7 Optimization 158
7.1 Extreme Value(s): Functions of One Variable . . . . . . . . . 159
7.2 Inflection Points and Concavity . . . . . . . . . . . . . . . . . 161
7.3 Maxima and Minima II . . . . . . . . . . . . . . . . . . . . . . 169
7.4 Maxima and Minima Subject to
Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
7.5 Economic Applications . . . . . . . . . . . . . . . . . . . . . . 184
7.6 Questions and Problems . . . . . . . . . . . . . . . . . . . . . 202
CONTENTS viii

8 Integral Calculus 206


8.1 The Definite Integral . . . . . . . . . . . . . . . . . . . . . . . 207
8.2 Rules and Properties Relating to the Integral . . . . . . . . . 209
8.3 Applications of the Indefinite Integral . . . . . . . . . . . . . . 218
8.4 The Definite Integral . . . . . . . . . . . . . . . . . . . . . . . 220
8.5 Economic Applications . . . . . . . . . . . . . . . . . . . . . . 242
8.6 Questions and Problems . . . . . . . . . . . . . . . . . . . . . 257

9 Matrix Algebra 259


9.1 Matrices and Vectors . . . . . . . . . . . . . . . . . . . . . . . 260
9.2 Matrix Operations . . . . . . . . . . . . . . . . . . . . . . . . 265
9.3 Special Types of Matrices . . . . . . . . . . . . . . . . . . . . 276
9.4 Determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
9.5 The Inverse of a Matrix . . . . . . . . . . . . . . . . . . . . . 292
9.6 Solving Simultaneous Linear Equations . . . . . . . . . . . . . 297
9.7 Maxima and Minima . . . . . . . . . . . . . . . . . . . . . . . 300
9.8 Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
9.9 Questions and Problems . . . . . . . . . . . . . . . . . . . . . 310

10 Linear Programming 312


10.1 Linear Programming . . . . . . . . . . . . . . . . . . . . . . . 312
10.2 Input-Output Analysis . . . . . . . . . . . . . . . . . . . . . . 329
10.3 Questions and Problems . . . . . . . . . . . . . . . . . . . . . 336

A Additional Review Questions 339


Chapter 1

The Role and Power of


Mathematics

Mathematics is a rigorous and well-defined study of the structures, configu-


rations, and interrelationships that characterize the world in which human
beings live. Mathematics provides an exacting language that articulates the
essential characteristics of a wide range of situations so that the key aspects
of those situations can be dispassionately examined.
Modern mathematics is “economical” in the best sense of the word. It clearly
states the barebones assumptions that underpin a relationship. In addition,
it highlights the logical processes that characterize the relationship. Finally,
it states any conclusions that are implied by the relationship in a clear and
concise form.
Economics is filled with topics that are amenable to mathematical analysis.
Relationships can be specified to relate production cost to output, output to
inputs, wages to worker productivity, and so forth. Further, analysts often
wish to establish the general nature of conditions required to minimize the
cost of achieving a certain objective, or to maximize the output of a particular
productive process. Sometimes, we seek specific values as well as knowledge
of the requisite conditions. In still other circumstances, the analyst may seek
evidence regarding how much change will occur in sales when the firm alters
the amount of advertising it is undertaking, knowing that the statistics-based
evidence will be somewhat imprecise.
Mathematical analysis can apply to abstract concepts as well as to concrete

1
CHAPTER 1. THE ROLE AND POWER OF MATHEMATICS 2

ones. An important member of this category is utility, which is central to


economic analysis. Utility is a highly abstract concept, but salient aspects
can be represented mathematically. The relevant point is that mathematics
is capable of dealing with a wide range of relationships that confront analysts
and decision-makers.
The versatility of mathematics is apparent. Equally important is its power.
With the help of mathematics, analysts can address problems that can be
given only a cursory glance with a strictly verbal analysis. Indeed, we can
make certain statements mathematically that with verbal language either
cannot be made at all or that must be made only in an awkward fashion.
The next section addresses a revealing historical example.

1.1 Stigler’s Diet Problem


In 1945, George Stigler [19] addressed the so-called “diet problem.” He
sought the least expensive combination of foods available to consumers that
would enable them to satisfy the recommended daily dietary allowances es-
tablished by the Food and Nutrition Board of the National Academy of
Sciences. That is, he set out to determine the cheapest way to obtain the
nutrients that individuals need to sustain life.
We sketch Stigler’s approach here, in order to illustrate salient aspects of
mathematical analysis. Chapter XX returns to this problem when we con-
sider linear and nonlinear programming techniques. The equation below
shows the function that is to be minimized. The total daily cost of a subsis-
tence diet, C, is the sum of the amounts spent on each of eighty goods, where
the amount spent on Good j is P j · Xj. The price is P j and the quantity is
Xj.1

C = P 1 · X1 + P 2 · X2 + · · · + P n · Xn
1
For reasons that will become clear later, we do not use subscripts. Much of our
work will involve commands that are written in text, so that subscripts are hard to enter.
Multiplication is indicated with centered dots (·) so that P 2 is a variable name. In contrast,
P · 2 is a product, with the variable name being P . We will not be entirely consistent in
our usage. In some settings, subscripts provide for easier interpretation. The context will
typically make the usage clear.
CHAPTER 1. THE ROLE AND POWER OF MATHEMATICS 3

In Stigler’s specification, (up to) eighty types of food were to be combined


to define the least-cost diet. Suppose that item 1 is peanut butter and the
price of peanut butter is $0.04 per gram. Then the cost of peanut butter in
this diet is $.04 X1, where X1 is the number of grams of peanut butter in a
daily diet. Again, the sum off all such terms is C, the cost of the diet.
As noted, however, the diet must satisfy a set of constraints that dietitians
had established. Nine such constraints were identified and incorporated into
Stigler’s analysis. The following expression shows what one of the nine equa-
tions in this model might look like: a11·X1+a12·X2+· · ·+a1n·Xn ≥ 3000,
where 3000 is the number of calories required.
This equation is the first constraint–one that identifies minimum caloric re-
quirement. The first of the two numbers attached to each coefficient a indi-
cates the constraint number (1 here). The second identifies the food number.
The other eight constraints look much like this one. For the first term in (1.2),
therefore, X1 is peanut butter, which provides about 7 calories per gram, so
a11 ≈ 7.
Stigler’s problem, addressed in the early 1940s, involved repeatedly evaluat-
ing combinations of eighty foods to ensure that they satisfied the nine con-
straints and then determining the cost. Not surprisingly, Stigler concluded
(p. 310) that his approach to solving the problem was “ . . . experimental
because there does not appear to be a direct method of finding the minimum
of a linear function subject to linear conditions.” That is, Stigler was forced
to find a solution by hit-or-miss methods.
Before discussing the tentative nature of this conclusion, we briefly summa-
rize the findings. Stigler determined that the established nutritional requires
could be satisfied for about $60 per year (in 1944 dollars, about $800 per
year in 2014 dollars). The major food items in the diet were these: wheat
flour, cabbage, spinach, pancake flour, and pork liver.
Stigler’s research is instructive at a number of levels. The direct implications
of the results, taking them at face value, are important. This is a classic
economic minimization problem with real world consequences. A second ob-
servation is that advances in applied mathematical analysis greatly facilitated
the analysis that Stigler pioneered. Finally, changes in our understanding of
nutrition require reexamination of the analysis.

• The first and obvious implication of the analysis is that the diet is
CHAPTER 1. THE ROLE AND POWER OF MATHEMATICS 4

not very palatable. Not many individuals would find a diet limited to
Stigler’s set of ingredients to be very tasty.

• Second, the meaning of the seemingly direct phrase “the cost of food”
is not as clear-cut as one might hope. Spending on food in the United
States is around $3500 per capita. We buy much more than subsistence,
and a casual reference to “the cost of food” refers to something quite
different from the cost of a subsistence diet. We eat more, sometimes
too much more, than subsistence requires and, more importantly, we
select foods that have attractions other than mere subsistence.

• Third, linear programming was invented in 1939 in the Soviet Union


and later in the United States. Its existence was unknown to Stigler,
however, because of secrecy surrounding World War II. By 1947, how-
ever, the technique was published and had become widely used. Thus,
within the decade in which Stigler’s work was published, more efficient
solutions became available. Even low-cost personal computers today
provide us the ability to solve systems like Stigler’s analytically, rather
than by approximation. The same is true for more complex systems
that consist of nonlinear relationships. Indeed quite sophisticated soft-
ware is routinely included in programs like Microsoft Excel and the
open-source LibreOffice Calc. Also, it is part of computer algebra sys-
tems, which the next section discusses.
The end result is that advances in computing power coupled with new
mathematical methods, have made solutions to problems like the one
that Stigler pioneered routine. These advances now guide decisions in
manufacturing, in communications, and in transportation.

• Fourth, we have learned more about the characteristics of foods since


Stigler attacked his diet problem. For example, nutritionists now are
aware that liver, though a good source of thiamine and manganese, pro-
tein, vitamin A, and numerous other nutrients, is very high in choles-
terol. Changing knowledge of nutrition is one reason that analysts have
updated Stigler’s findings.

• Fifth, the prices of foods relative to those of other goods and services
has declined and relative prices have changed among types of foods.
CHAPTER 1. THE ROLE AND POWER OF MATHEMATICS 5

• Finally, we now have more foods to choose from for our new cost-
minimizing diet.

In reexamination, Bassi [1] estimated a per-capita cost of about $160 in 1975


dollars (about $670 dollars at the 2014 general price level). Bassi’s mix looks
a lot like Stigler’s, except that red beans play a larger role and beef kidneys
have replaced pork liver as the predominant meat source (again pointing out
the distinction between a subsistence diet and a palatable diet).

1.2 Analysis Using a Computer Algebra Sys-


tem
One product of the confluence of mathematical advances and increased com-
puting power is the Computer Algebra System (CAS). A CAS is a math-
ematical toolkit that can be used as a simple (or advanced) calculator. It
can also be used to find solutions to symbolic mathematical expressions and,
if these expressions cannot be solved, to produce simulations that provide
insights into the expressions’ implications.
Both proprietary CAS programs (ones that must be purchased) and open
source programs (one need not pay to use them) exist. The two most widely-
used proprietary CAS programs are Mathematica and Maple. The most
widely-used open source program, and the one used in the remainder of this
book, is Maxima. An analyst who has gained skill in using any one of these
three programs can quickly transfer that skill to the use of either of the other
two. See citeMeglicki.

1.2.1 The wxMaxima User Interface


The Maxima CAS offers a selection of user interfaces, the most popular one
being wxMaxima. The remainder of this section consists of a quick overview
of wxMaxima, based on a screen shot of a small notebook. We will not actu-
ally begin to learn how to use wxMaxima here, but instead simply preview a
few of its possibilities. The appendix to this text provides more detail. Also,
the website that accompanies this book (http://www.wxmaximaecon.com/)
CHAPTER 1. THE ROLE AND POWER OF MATHEMATICS 6

Figure 1.1: wxMaxima Interface

provides links for downloading Maxima, which includes wxMaxima and for
getting started in wxMaxima. Also, see [7], Chapters 1 and 2.
Figure 1.1 shows three input/output cells. The input is entered as text.
Commands are ended by semicolons (if resulting output is to be printed) or
dollar signs (if resulting output is to be suppressed). Once a set of commands
is to be executed, ctrl-enter generates the output.
The first cell shows a simple way to assign a name to an expression (the two
commands in the first cell). The second cell graphs the expression(s), and the
third cell shows how to find a root of an expression, in this case “Revenue
= Cost.” At this point, just observe the general nature of the workbook
CHAPTER 1. THE ROLE AND POWER OF MATHEMATICS 7

without concern regarding particulars of the commands.


Briefly consider the third cell, which involves the determination of one of
the two quantities at which revenue and cost are equal. This is a “break-
even” quantity, the minimum quantity that the firm must sell in order to
cover costs. The solution is “numerical,” meaning that we just looked for a
numerical value, not an algebraic solution. Later, will we return to examples
like this one, but we also use Maxima to examine the nature of solutions,
including determining the quantity that yields maximum profit. This figure
is the first of many that follow. Most of the cells show the commands that
generate a set of results and those results.
The last part of Figure 1.1 is a horizontal line. This is a “cursor”–wxMaxima
interprets any keyboard entry on such a line as the beginning of a new set of
one or more commands.

1.3 The Diet Problem in Maxima


We close this brief introduction to Maxima by looking back to the Stigler
diet problem. For the sake of clarity and simplification, Stigler reported a
subset of his larger model in order to focus on the logic of his approach. Zhou
provides an extensive development of Stigler’s analysis of that subset, and
we use Zhou’s notation.2 The cell below shows that we wish to minimize z,
which is the cost function, subject to a set of the eight constraints that apply
(actually 13 constraints because we add five nonnegativity constraints) in
this simplified representation of Stigler’s analysis. Each constraint in a linear
expression. The result is a set of five values for the five foods (x1, x2, x3, x4
and x5) that enter into this analysis and the value z = 0.109. Chapter xx
returns to this issue and interprets the values more fully..3
This introduction to wxMaxima is necessarily cursory. Even so, we have
achieved the following: assigning names to expressions, graphing those ex-
2
Wenxiao Zhou, “A Discussion on ‘The Stigler Diet Problem’ by Applying the Simplex
Method & GAMS,”22 April 2013, http: //www. unc.edu/ marzuola/ Math547 S13/
Math547 S13 Projects/W Zhou Section003 SimplexMethod.pdf.
3
This presentation of the input/output cell, where it appears as a single figure, will
not be used beyond this point. Rather, the input will appear after the identifier (%i), and
output will appear after the identifier (%o) except when the output is a graph. Graphs
will follow the input and will be numbered sequentially through each chapter.
CHAPTER 1. THE ROLE AND POWER OF MATHEMATICS 8

pressions, using a numerical method to determine a critical value (analytical


methods follow in subsequent material), and solving a linear programming
problem.

1.4 Summary
Mathematics underlies the analysis of many of the issues that business deci-
sion makers, public policymakers, and economists address. This chapter dis-
cusses the application of mathematical analysis as it applies to an important
issue, the cost of a subsistence diet. This example shows how mathematics
can be applied, and it shows that the mathematical tools are becoming in-
creasingly sophisticated. Coupled with the power of modern computers, this
increased sophistication has broadened the purview of applied mathematical
analysis.
One important advance of the past few decades has been the development
of computer algebra systems. These systems extend analysts abilities by
solving complex problems, by allowing for the examination of a range of
scenarios, and by producing simulations of systems that defy formal solution.
The remainder of this book displays some of these features by applying the
Maxima open-source computer algebra system.
The development and use of mathematical tools to solve business and eco-
nomic problems has expanded very rapidly in recent years. A course covering
the materials presented in this book is now required of business adminis-
tration, accounting, marketing, and economics majors in most colleges and
universities. It behooves the student who wishes to be well prepared and
efficient to master the mathematics that appear in the following chapters,
not only because it probably will be necessary in order to graduate, but
also because mathematics will prove to be very useful in later employment.
Mathematical economics can open new doors to those who take the time to
master its essentials.
CHAPTER 1. THE ROLE AND POWER OF MATHEMATICS 9

1.5 Questions and Problems


1. Mathematics can be used in very useful ways in analyzing many busi-
ness and economics problems. This fact has led some enthusiasts to
contend that “If you can’t measure it, it isn’t worth knowing about,”
or “If you can’t measure it, it doesn’t exist.” Write a critique of these
statements.

2. Many historians claim that they do not use models in writing history
and in arriving at conclusions about historical phenomena. Is it possible
to analyze something without having an underlying model? Will hard
work produce insights and generalizations if you do not have a model?
Explain.

3. Those who use mathematical tools in the analysis of business and eco-
nomics problems frequently contend that it is possible to say things
with mathematics that could not be said verbally. Is this true? Can
the reverse be true?

4. A frequent criticism of business and economics models is that they do


not fit the real world with precision. The real world nearly always
seems to be somewhat different from the world outlined in the model.
Is this a valid criticism? Can we construct models that precisely relate
to a particular situation, or to all situations?

5. Business and economic models that employ mathematics are occasion-


ally criticized on the grounds that they employ unrealistic assump-
tions. For example, economists assume that individuals maximize util-
ity. Some criticize this assumption as being unrealistic. Are realistic
assumptions necessary when one is using mathematical models?

6. Related to Question 5: For a number of reasons, we should suspect


that Stigler’s conclusions are wrong. Does that mean that the analysis
should be ignored? Explain.
Chapter 2

Variables, Sets, Lists, and


Relations

Chapter 1 used Stigler’s diet problem to demonstrate that the appropriate


use of mathematics can provide valuable insights into an important issue.
With mathematics we can isolate and examine the crucial forces operating
in an increasingly complex world.
The world’s complexity is what drives analysts to use mathematics. At the
same time, this complexity bedevils the analyst. The variety of possible inter-
actions in any complex systems means that no decision-maker can consider
all of the factors that might influence that decision. Even so, leaving out
pertinent information can lead to serious error. It is humanly impossible for
any individual to provide a complete description of the richness, complexity,
and variety that characterize the world.
Realistic, successful analysis of a problem that faces a decision-maker requires
that the analyst isolate the key aspects of reality in that problem. That is,
the analyst must abstract and simplify, always taking care to retain those
factors that are deemed crucial to the situation at hand. For example, the
availability of iron ore is a crucial factor in the production of steel, whereas
the religious affiliations of steel workers is probably an irrelevant factor.1
1
We say “probably” because in some contexts this might not be true. If production
is a team activity is a team effort and if religious composition affects team performance
either by promoting cohesion or by providing insights from various viewpoints, then it
should be taken into account. Witness communes like the Amana community.

10
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 11

Skillful analysis of a problem results in a theory. A theory is an abstract set


of relationships from which we can derive meaningful propositions.
A good theory delineates the crucial forces that are at work in a situation,
the circumstances under which those forces are related, the nature of those
relationships, and the probable result of the interaction of those forces. Here
is a theory of your success: Students who are more intelligent, and who study
more, will generally earn higher grades than others not as well equipped or
prepared. Such a theory forthrightly states two of the most important factors
that determine student grades, and also indicates the relationship between
these two factors and grades. This theory is probabilistic, in that it identifies
tendencies but does not insist that, for any two individual, the one that
is both a bit more intelligent and a bit more diligent will inevitably earn
better grades. Furthermore, as stated here this theory says nothing about
the interrelationships and tradeoffs between intelligence and hours of study
in terms of expected grades.
The language and the component parts of theories require further examina-
tion, for it is our ability to construct and use abstract theoretical relationships
that determines our ability to make intelligent choices and decisions. This
statement is itself a theory. Some might argue that intuition is all that mat-
ters: “Good decision-makers are born, not trained.” The theory stated here
does ignore intuition. It does not, however, deny a role for intuition. Sup-
pose that the analysis defines the general nature of some relationships but
does not specify the precise value of some of its parameters. Intuition can be
combined with formal analysis at this point.
Analyzing economic and business problems often involves combining several
theories into a model. We can express an economic model as a series of
mathematical equations, but we need not do so. Many models are developed
verbally, although such models often suffer from lack of precision. A model
identifies the factors and influences that are important in a situation, and
delineates the relationships among those factors and influences. A model is
best thought of as a systematic presentation of interrelated theories.

2.1 Variables
The remainder of this text focuses on mathematical models that relate to
economic activities. The analysis is phrased in terms of expression of how
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 12

variables relate to each other. Most of these variables, like income or a price,
can be quantified. Others, like utility, can be ordered but not quantified in
a meaningful way. A variable is a quantity that can assume different values
at different points of observation.
The magnitude of a variable can assume various values. For example, the
gross national product (GDP) of the United States could be 100, 1000, 1500,
or, indeed, any positive magnitude. Because a variable’s magnitude can as-
sume various different values, a variable must be represented by a general
symbol. Hence the price of a pizza might be represented by the symbol p,
while the tax rate might be represented by the symbol r. Chapter 1 repre-
sented the magnitudes of the 80 different foods as X1, X2,. . . , X80. The
letters at the end of the alphabet, such as u, w, x, y, and z, commonly sym-
bolize the magnitudes of variables. This, however, is a matter of convention
rather than of necessity. In Maxima, we often use a string of letters to name
a variable. Thus cons might be the name assigned to consumption.
Variables can be either cardinal or ordinal. The values assigned to cardinal
numbers have meaning: the difference between $2 and $5 is $3. Ordinal
variables define order alone: we may judge one person to be friendlier or
happier than another but we cannot assign a specific value to the difference
in the levels of friendliness or happiness.
We may classify cardinal variables as being either continuous or discrete in
terms of the magnitudes that are permissible for those variables. A contin-
uous variable is one that can assume any value within a given interval of
values. Annual income is a continuous variable. A discrete variable is one
that can assume at most a limited number of values within a given interval
of values. The number of siblings in your family is a discrete variable.
Consider some examples of continuous and discrete variables. The examples
reveal that the distinction, while quite real, can sometimes be ignored with
impunity. When this distinction can be safely ignored depends, of course, on
theoretical considerations.

• Gross Domestic Product (GDP) is a continuous variable in that it can


assume any positive magnitude. Hence 1,543.324 billion dollars is one
possible magnitude. In practice, however, the actual reports and com-
putations of GDP are typically restricted to values that are truncated
to a specific number of decimal places. The value 18036.649 . . . become
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 13

$18036.65, a discrete value when GDP is reported in billions of dollars


and round to the second decimal place.
• Flipping a coin many times and recording the outcome (1 = heads,
0 = tails) generates values of a discrete value; the value x=22/3, for
example, cannot be observed. Despite this fact, applying the normal
distribution, which describes a continuous variable, can still provide
meaningful analysis of the probabilities associated with the number of
heads per toss when the number of tosses becomes large, as you learned
in statistics.
• Sometimes it is not clear whether the relevant variable is discrete or
continuous. Consider the number of cars in a household. The number
is an integer, but how often cars are traded is continuous. At the
market level, the relevant variable can be treated as continuous, for
two reasons. First, if a small fraction of households change the number
of cars per household, the result will be a very small change in the
number of new cars purchased. The same is true of a small change
in the age at which households trade for a new car. Technically, the
number of new cars sold per year is an integer (though the average
number sold per month or per day is not). Even so, when the total
number sold is in the millions, the relevant variable can be treated as
continuous without compromising the analysis.

Variables may also be classified as to whether or not they are endogenous


or exogenous in nature. An endogenous variable is one whose value is to be
determined inside the model being used. An exogenous variable is one whose
value is taken as given; its value is determined by forces that are outside the
model.
The magnitude of endogenous variables is explicitly examined and deter-
mined by the model. A supply-and-demand model determines the price of
goods and services. Price is endogenous variable in such a case. If, however,
government mandates a price for the good, then price becomes an exogenous
variable in the model.
Recall the simple model of income determination from your macroeconomics
principles class. The Y = C + I + G equation (where Y is a measure of
national income, C is private consumption expenditures, I is business invest-
ment expenditures, and G is governmental purchase of goods and services)
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 14

determines national income. In the simplest models, C is considered to be


endogenous, while I and G are considered to be exogenous. That is, the
value of C is determined inside the model, whereas the values of I and G
are taken as given. (The wxMaxima workbook that accompanies this section
illustrates this example of a model.)

2.2 Equations, Roots, and Constants


Like the simple income determination model above, mathematical models are
usually expressed in the form of equations. An equation is a statement that
asserts the equality or equivalence of two (or more) mathematical expressions.
Each equation must contain at least one variable. For example, the expression
2 · X = 10, is one statement about the variable X. Of all possible values, 5
is the only one for which this statement to be true.
The previous example, 5 is a critical root, or a solution value. Critical root(s)
or solution value(s) is (are) the value(s) of the variable(s) of an equation that
cause(s) the equation to hold true.
Many examples of critical roots (solution values) occur in the field of eco-
nomics. The equilibrium price that clears the market, the magnitudes of
inputs and outputs that maximize profits, and the dollar value of the con-
sumption of private individuals that leads to an equilibrium level of GDP
are all examples of critical roots.
Some equations are characterized by mathematical terms that never change
in value. You are undoubtedly aware that the value of π (the Greek letter
pi) is a constant that is equal to 3.14159.... The value of π never changes.
Another example of a constant is e, the base of the system of natural loga-
rithms, which is equal to 2.71828 . . . . A numerical constant is a magnitude
that is fixed and does not change in value. When a constant is joined to a
variable, that constant is often referred to as the coefficient of that variable.

Many equations include parameters that act as numerical constants in a


limited fashion. A parametric constant or parameter acts as a constant only
within the context of a particular equation or problem, but may assume a
different constant value in other equations or problems.
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 15

An example sharpens the difference between numerical constants and para-


metric constants. Assume that the number attending a St. Louis Cardinals
baseball game is given by the equation Q = 50, 000 − b · P . The number
50,000 is the stadium’s seating capacity and is a constant, at least until
major construction occurs. The number attending will generally be values
within a given interval of values that cannot exceed 50,000. The number
in attendance is less than this 50,000 and depends on the price, P , and the
change in Q per one-unit change in price, −b. The value of the parameter b
depends on many things that are exogenous to this simple theory: the day
of the week, whether the game is critical to a pennant race, the identity of
the pitcher, and the identity of the visiting team among others.
It is customary to use letters at the beginning of the Latin alphabet (for ex-
ample, a, b, c, and d) or of the Greek alphabet (α, β, γ, and δ) to symbolize
parameters in a particular equation. As an example, or first use of Maxima,
consider a simplified Keynesian national-income model to illustrate the vari-
ous terms and definitions that we have developed in this subsection. The cell
below defines the four equations and shows the condition that must be met
for equilibrium to occur. The first three are simple theories of behavior, so
they are behavioral equations. The last of these four equations, entered into
the solve( ) command, is an equilibrium condition. The result is a variable
named Y eq, the equilibrium value of Y .
The four commands below state the components of the model and determine
the condition for equilibrium as follows. Notice some aspects of this entry.
First, we use a fixed font to indicate Maxima commands, which must
consist entirely of text characters. Second, the first three commands end with
dollar signs, so that they produce no printed output (though the variables
name on the left of the colon in each is assigned to the expression on the
right of the colon and these assignments remain in Maxima’s memory). The
fourth command ends with a semicolon, so that the result of executing this
command is printed.

(%i) C:a+b*Y$ I:I0$ G:G0$ solnY: solve(Y=C+I+G, Y);


(%o) [Y = − I0+G0+a
b−1
]

2
The result of executing this command is assigned the name solnY. The
2
Any allowable name would do; some names like value, are reserved and cannot be
applied, though Value could be—Maxima is case-sensitive.
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 16

output is shown as the single item in a list. The textttsolve command always
produces a list. Maxima follows mathematical conventions, so the output
is not always as you might expect to see it. In the example above, we can
restate the expression as follows: Y = (a + I0 + G0)/(1 − b), so that the
multiplier 1/(1 − b) becomes apparent.
Suppose that we have values for the parameters a and b and for the exogenous
values I0 and G0. The following input/output combination shows a set of
values and the implied values of Y and C. The first command below assigns
the name Yeq to the equilibrium income. The second command substitutes
a set of parameter values into Yeq and assigns the name Yeq0 to the result.
The third command substitutes the parameters and the equilibrium output
level into the consumption function, yielding the equilibrium consumption
level. The final command provides a check, to confirm that total spending
sums to the equilibrium output level.3

(%i) Yeq: rhs(solnY[1]);


Yeq0 : subst([a = 150, b=0.75, I0=25, G0=20], Yeq);
C0:subst([a=150,b=0.75,Y=Yeq0],C); C0+25 +20;

(%o) − I0+G0+a
b−1
(%o) 780.0 (%o) 735.0 (%o) 780.0

The above four-equation model has two endogenous variables (Y and C) and
two exogenous variables (I and G), the values of which are assumed to be
determined outside this model. The consumption function illustrates the use
of two parameters (a = 150 and b = 0.75). The values of the exogenous
variables are also numerical constants. The final command states and solves
the equilibrium condition.

2.3 The Real Number System


As we have seen, variables, constants, and parameters usually take on nu-
meric values. We can classify numeric values in terms of their position on
the real number line.
3
When we refer to variables in general, we use standard mathematical notation, so
that the variables appear like this: C = a + b · Y ). When the variables are names assigned
in Maxima input, we use the fixed font, like this: C = a + b*Y.
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 17

Consider the positive integers (1, 2, 3,. . . ), the negative integers (- l, -2, -
3, . . . ), and zero. All these values may be found on the real number line
portrayed below. A real number line has the following characteristics: (1)
The origin (location of zero) on the real number line is arbitrarily chosen.
(2) The units of measurement on the real number line are arbitrarily chosen.
(3) A positive or negative direction along the real number line is indicated
by the sign of the number; this sign reflects the location of a particular
point relative to the origin. (4) The ordering relation among the numbers
on the real number line is that, if x < y, then the point x lies to the left of
point y on the real number line. The number line in Figure 2.1, generated
by Maxima, shows three integers, the values −π and π, the constant e, the
fraction 2/3, and the square root of 5. Confirm that these values are in the
correct sequence.

Figure 2.1: Real Number Line, Segment

The gap between any two whole, integer values found on the real number
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 18

line may be partially filled with rational numbers. A rational number like
2/3 results from the division of one integer by another, provided that the
denominator is not equal to zero. The number 2/3, expressed as the quotient
of the integers 2 and 3, is more commonly known as a fraction. Any integer
may be expressed as the quotient of some two integers. Therefore every
integer is a rational number. For example, 5 = 10/2 = 5/1.
The remaining gaps on the real number line are filled by irrational numbers.
Irrational numbers cannot be expressed as the quotient of two integers.√ An
√. . . The square root of 5 ( 5) is
example is the value of π, which is 3.14159.
another example of an irrational number, 5 = 2.236067 . . ..
To summarize: A rational number is the quotient of two integers, the denom-
inator not being equal to zero; an irrational number cannot be expressed as
the quotient of two integers; and a fraction is a rational number that is not
an integer.
The rational and irrational numbers together form the real number system.
The one-to-one relationship between the real number system and the real
number line means that we may use the terms “real number” and “point”
interchangeably. A real number is a point
√ on the real number line. We omit
complex numbers, which involve i = −1. Complex numbers can occur in
some analysis of dynamic systems, but take us beyond the purview of this
text

2.4 Sets and Set Theory


We could paraphrase the preceding paragraph as follows: The set of rational
numbers can be combined with the set of irrational numbers to form a larger
set, the set of real numbers. In this paraphrase, “set” is used casually. The
concept of the set, however, can be defined much more precisely. The theory
based on that concept, developed in the latter part of the nineteenth century,
forms the foundation of much of modern mathematics. A thorough treatment
of set theory would require an enormous amount of work. However, you
can acquire a working knowledge of the basic concepts of set theory with
considerably less effort.
We begin with a formal definition. A set is a collection of distinct, well-
defined objects. Here “well-defined” means that any object either is an in-
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 19

stance of the set or it is not, and a condition exists for determining which
of these is true. A set may be defined by either the “roster method” or the
“set-builder method.” Consider a simple set: A = {1, 2, 3, 4, 5}. This is
an application of the roster (or enumeration) method: the elements of
the set are listed. Such a set necessarily has a finite number of elements.
Notice the notation: A set’s name is typically a capital letter, and the rule
for identifying its elements is enclosed in curly brackets.
In the set-builder (or definition) approach, these brackets contain a rule
for identifying the elements Consider B = {x|0 < x < 100}. Read this as
“x such that x’s value is between 0 and 100 and does not include 0 or 100.”
Infinitely many real numbers qualify, so a set that is constructed with the
set-builder method can (but need not) be infinite.
Maxima uses the roster method, so the sets that it manipulates are finite,
though they can be very large. The three commands below build sets A,
B, and C. The resulting output appears below the commands. Compare
the third command with the third output line. Maxima has removed the
elements that repeat. It does this because a set consists entirely of distinct
elements; repetitions are not allowed. Also observe that Maxima writes the
elements of set B in alphabetic order, not in the order in which they were
entered. An important aspect of a set is the sequence does not matter.

(%i) A:1, 2, 3, 4, 5, 6, 7, 8, 9;
B:red, orange, yellow, green, blue, indigo, violet;
C:1,2,3,4,5,6,7,8,9,8,7,6,5,4,3,2,1;
(%o) {1, 2, 3, 4, 5, 6, 7, 8, 9}
{blue, green, indigo, orange, red, violet, yellow}
{1, 2, 3, 4, 5, 6, 7, 8, 9}

Consider one more case, one in which we might be tempted to think that the
set contains a single element. We enter these named expressions: X: a/c +
b/c, Y: a/c + b/c and Z: (a + b)/c as Maxima commands. The first two
expressions, assigned the names X and Y are equivalent to each other and
have the same form. The third expression, Y, has the same value but not
the same form. Therefore, it is a distinct element of the set X,Y,Z. Either
of these two equivalent commands generates the relevant set: {X,Y,Z} or
set(X,Y,Z).4 The result is { b+a
c
, bc + ac }.
4
We have distributed the commands through this paragraph, rather than placing them
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 20

An element that is in a set is said to be a member of that set. Membership


is typically signified with the symbol ∈. Read A ∈ G as “set A is an element
of set G,” or as “set A is a member of set G,” or as “set A belongs to set
G.” The notation ∈ / indicates that the element does not belong to the set.
The command elementp(A,x) instructs Maxima to determine whether x
belongs to A. In the example below, Maxima informs us that 5 is an element
of A and that 11 is not an element of this set, where A:{1,2,3,4,5,6,7,8,9}
and the two test commands are these: elementp(5,A) yields true, and
elementp(11,A) yields false.
An important special case is the null set, sometimes called the empty set.
This set, which contains no elements, is denoted by the symbol ∅. The names
of NBA basketball players whose height is under 5 feet is the null set, as is
the set of years in which the real GDP of the United States grew at an annual
rate greater than 50 percent.

2.4.1 Set Algebra


Sets can relate to each other in various ways. Consider the following relation-
ships: equality, subsets, union, intersection, universality, and complementar-
ity.

Equality

Two sets S1, and S2 are said to be equal or identical if and only if S1, and
S2 have exactly the same elements. The next exhibit considers four sets, S,
A, B, and C. Sets A, B, and C are subsets of S, but only C contains all of
S’s elements. Therefore, C = S, while A 6= S and B 6= S. Note that we have
entered some repetitions that do not appear in the first output line, which
identifies set S. Each element of a set must be unique.

(%i) [S: {1,2,3,4,5,6,7,8,9,9,8,7},


A:odds(S), B:evens(S), C:S];
(%o) [{1, 2, 3, 4, 5, 6, 7, 8, 9}, {1, 3, 5, 7, 9},
{2, 4, 6, 8}, {1, 2, 3, 4, 5, 6, 7, 8, 9}]
together and following them with output. When this option seems to provide a better flow,
we will use it. See the accompanying workbook for the input/output cell.
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 21

The foregoing material requires defining a function named evens() and an-
other named odds(). The definition of these functions, which provides a way
for Maxima to emulate the set-builder approach to set building, is sketched
in the workbook that accompanies this chapter. For a more complete treate-
ment, see [13].
Next we confirm that Maxima can discover the relationships between sets
that we have asserted. The following commands are entered: [is(S=S),
is(S=C), is(A=S), is(B=S), is(B=C)]. Maxima treats each of these is(
... ) statements as a condition to be evaluated. The resulting output is
the answers that we expect: [true, true,false, false, false].

Subsets

Set A is said to be a subset of set S if and only if every element of A also


belongs to S. The notation is this: A ⊂ S is read “Set A is a subset of set
S.” In the example above, A ⊂ S, B ⊂ S, and C ⊂ S. We confirm these
assertions with the following commands (both the commands and the output
are entered as lists): [subsetp(A,S), subsetp(B,S), subsetp(C,S)]. As
expected, the result is [true, true, true]: all of the named sets are subsets
of S.

Union

A new set may be formed by the union of two sets. Let S1 and S2 be any
two arbitrary sets. The union of S1 and S2 consists of the elements that are
in S1, in S2, or in both S1 and S2. The notation is S1 ∪ S2, which is read
“S1 union S2.” In the example above, A ∪ B = C. In the next example,
Maxima determines the union of these two sets: all integers from 1 through
10 and the even integers from 2 through 20.
The commands below create a set that consists of the even integers between
1 and 20, set S1. The second set, S2, consists of the squares of the integers
between 1 and 10. The third set consists of the union of S1 and S2.

(%i) S1:setify(makelist(i,i,1,10));
S2:setify(makelist(i^2, i, 1,10)); union(S1,S2);
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 22

(%o) {1, 2, 3, 4, 5, 6, 7, 8, 9, 10}


(%o) {1, 4, 9, 16, 25, 36, 49, 64, 81, 100}
(%o) {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 16, 25, 36, 49, 64, 81, 100}

Intersection

Another way to form a new set is via the intersection of two or more sets.
The intersection of two sets (the definition easily extends to more than two)
S1 and S2 consists of the elements that are in both S1 and S2. The notation
is S1 ∩ S2 is read “S1 intersection S2.” Formally, S1 ∩ S2 is equivalent to
{x|x ∈ S1 and x ∈ S2}. The intersection of S1 and S2 above is the set {2, 4,
6, 8, 10 }, which executing the command intersection(S1,S2) confirms.
The intersection of sets that share no common elements is the null set; such
sets are said to be disjoint. The command intersection({red, yellow,
green}, {up, over, out}) generates this output: {}, which is Maxima’s
notation for the null set.

Set Difference

The union operator defines elements that are members of any of two or more
sets. The intersection operator defines elements that two or more sets share
in common. A third, related operator is the set difference operator, which
defines elements that are in the first set but not in the second set. Order
matters. A formal definition is this: Given any two arbitrary sets S1 and
S2, the set difference of S1 and S2 consists of the set of all elements that
belong to S1 but not to S2. Formally, S1 − S2 = {x|x ∈ S1 and x ∈ / S2}.
Consider these three sets: X : {1, 2, 3}, Y : {3, 4, 5}, and Z : {1, 2, 5, 6, 7}.
The command setdifference(X,Y) produces the set that consists of the
elements that are in X but not in Y : {1,2}. In contrast, setdifference(Y,
X) produces the set that consists of elements of Y that are not in X: {4, 5}.
Other examples based on these sets appear in the workbook that accompanies
this chapter.
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 23

Multiple Sets

Unlike union and intersection, set difference is a binary operation: it com-


pares two sets. Both union and intersection can be applied to more than two
sets. Refer to the three sets above. The union of the three is X ∪ Y ∪ Z =
1, 2, 3, 4, 5, 6, 7. The intersection of these sets is X ∩ Y ∩ Z = {}, the empty
set.

Universal Sets

A universal set includes all elements that are allowed by definition. If the
set is potential results of flipping a coin, then the universal set is {heads,
tails} (assuming the coin never lands on its edge). Thus, a universal sets is
a complete listing of all elements or outcomes that can be associated with a
particular action or situation.
If the contents of a set A is known and if the universal set is given, then we
can deduce the contents of a second set that complements the first set. The
complement to set A is A0 = U −A, where the prime indicates. Alternatively,
the complementary set A0 is {x|x ∈ U and x ∈ / A}.

2.4.2 Set Geometry: Venn Diagrams


The algebraic relationships between sets that can be illustrated visually by
means of Venn diagrams, as shown in Figure 2.2 below. This exhibit shows
the six relationships that involve any two sets and the universe of which they
are a part. The construct of Venn diagrams can extend to any number of
sets. The relationships are these:
a) Both A and B are subsets of the universe U , and B is a subset of A.
B ⊂ A; A ⊂ U ; B ⊂ U .
(b) The union of A and B contains all elements of A and all elements of B:
A ∪ B.
(c) The intersection of A and B contains all element of both A and B. In
this case A and B have no common elements; they are disjoint: A ∩ B = ∅.
(d) The intersection of A and B, when A and B contain some common
elements: A ∩ B.
(e) The set difference A − B shows all elements that are in A but not in B.
(f) The complement to A, A0 , shows all elements of U that are not in A.
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 24

Figure 2.2: Venn Diagrams

Exercise 2 - 1

1. Using set notation, specify each of the following.


(a) The set of all integers greater than -5 but less than 5.
(b) The set of all prime numbers from 0 to 25.
(c) The set of all real numbers greater than 0.
(d) The set of all even numbers that are also ?prime? numbers, in the
sense that they cannot be divided by any integer to obtain another
integer.

2. Let S = {1, 2, 3}, T = {3, 4, 5}, V = {3, 2, 1}, and the universal set
U = {, 2, 3, 4, 5}. Which of the following statements are correct? If a
statement is incorrect, correct it.
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 25

(a) S = T (b) S = V (c) 3 ∈ S


(d) 4 ∈ V (e) S ⊂ V (f)T ⊂ S
(g) V 6⊂ T (h) S ∪ T 6= U (i) S ∩ T = U
(j) V ∩ T = ∅ (k) S ∪ V = S (l) U − S = T
(m) V 0 = T (n) U − S = U − V (o) S ∪ V ∪ T = U
3. Let A = {1, 2, 3}, B = {2, 3, 4, 5}, and C = {1, 3, 5}. Verify that the
following assertions are correct for these sets:
(a) A ∩ (B ∪ C) = (A ∩ B) ∪ (A ∩ C), and
(b) A ∪ (B ∩ C) = (A ∪ B) ∩ (A ∪ C).
4. Use Venn diagrams to verify 3(a) and 3(b).
5. If A ⊂ B and C ⊂ D, does this mean that (A∪B) ⊂ (C ∪D)? Explain.
6. If A ⊂ B and C ⊂ D, does this mean that (A∩B) ⊂ (C ∩D)? Explain.
7. Use Venn diagrams to show when the following expressions are correct.
(a) A ∪ B ∪ C = C
(b)A ∩ B ∩ C = C and
(c) A ∩ B ∩ C = ∅
8. In Figure 2.3, the universe is U . The sets S1, S2, and S3 are as
indicated. Area VI is the difference between U and areas I, II, III, IV,
and IV. Identify each of the following.
(a) The two sets that are disjoint.
(b) The area(s) corresponding to S1 ∪ S2.
(b) The area(s) corresponding to S1 ∩ S2.
(c) The area(s) corresponding to S10 .
(d) The area(s)–if any–corresponding to S1 ∩ S2 ∩ S3.
(f) The complement to the area defined in (d).
(f) The area(s)–if any–corresponding to S1 ∪ S2 ∪ S3.
(g) The complement to the area defined in (f).
(h) S1 − S3.
(i) S3 − S1.

2.4.3 Set Theory: The Formal Algebra


We can formally translate the Venn-diagram analysis of the previous section
into a series of laws that define the algebra of sets. These laws lack the intu-
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 26

Figure 2.3: Three Sets

itive appeal of the Venn diagrams. However, the fact that they are algebraic
rather than graphical in character is advantageous in extended applications
of set theory. Throughout, we assume the existence of three sets–A, B, and
C–that are subsets of the universal set U . You might recognize the similar-
ity of many of these laws to those that provide the foundation for standard
algebra.

• Commutative Laws
(a) A ∪ B = B ∪ A
(b) A ∩ B = B ∩ A

• Associative Laws
(a) A ∪ (B ∪ C) = (A ∪ B) ∪ C
(b) A ∩ (B ∩ C) = (A ∩ B) ∩ C

• Distributative Laws
(a) A ∩ (B ∪ C) = (A ∩ B) ∪ (A ∩ C)
(b) A ∪ (B ∩ C) = (A ∪ B)) ∩ (A ∪ C)
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 27

• Idempotent Laws
(a) A ∪ A = A
(b) A ∩ A = A

• Identity Laws
(a) A ∪ ∅ = A
(b) A ∪ U = U
(c) A ∩ U = A
(d) A ∩ ∅ = ∅

• Complement Laws (a) A ∪ A0 = U


(b) A ∩ A0 = ∅
(c) (A0 )0 = A
(d) U 0 = ∅
(e) ∅0 = U

• DeMorgan’s Laws
(a) (A ∪ B)0 = A0 ∩ B 0
(b) (A ∩ B)0 = A0 ∪ B 0

2.4.4 Ordered and Unordered Pairs


In set theory, the two-element set {x, y} is equal to the two-element set {y, x}.
That is, {x, y} = {y, x}. The pair (x, y) is therefore said to be an unordered
pair, in that the ordering is irrelevant. Contrast this to an ordered pair, in
which (x, y) 6= (y, x) and the ordering of elements x and y is crucial. More
formally, given two elements x and y, a pair (x, y) is said to be an ordered
pair if (x, y) 6= (y, x) unless x = y.
As an example in which ordering matters greatly, consider the ordered pairs
consisting of the wins followed by the losses of an athletic team. For example,
the ordered pair (2, 12) would represent the 2-win, 12-loss record, which is
quite different than a very successful (12 − 2) record. Ordered pairs are
enclosed in parentheses–(2, 12) is not the same as (12, 2)–and unordered pairs
appear within curly brackets–{2,12} is the same as {12,2}.
We can extend the concept of ordered elements in order to distinguish be-
tween ordered and unordered triples, quadruples, and so forth. Consider a
list of the last four U. S. Presidents of the twentieth century.They are Carter,
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 28

Reagan, Bush, and Clinton. As members of a set, the following would be true
{Reagan, Bush, Carter, and Clinton} = {Clinton, Bush, Reagan, Carter}.

2.5 Lists
A historian would not consider the two orderings in the set of presidents to
be the same. They would insist on a list of the presidents: [Reagan, Bush,
and Clinton] would be a chronological list; some other ranking might result
in a different list. A list is an ordered n-tuple of elements, and any of these
elements may consist of text strings, numbers, mathematical expressions,
sets, or other lists. Lists are the basic building blocks for computer algebra
systems like Maxima. The command pList: [Carter, Reagan, BushI,
Clinton,BushII,Obama] produces a list of presidents and assigns it the name
plist. With this information in Maxima’s memory, the command plist[3]
produces the output BushI.5
Lists can be used to assign names to expressions. These expressions can
involve computation or they can consist of strings. Also, a list can contain
another list or a set. The following set of input and resulting output shows
a four-item list. Each item in the list is bound to a member of a set of four
names. The command that creates the lists ends with a $, so that printing
is suppressed. The individual items are then recalled and printed. The four
commands in the second input line result in the four lines of output, one for
each of the named items in the list.

(%i) [a,b,c,d]:["Some text", {x[1],x[2]}, [p,q], log(10.0)]$


a; b; c; d;
(%o) Some text (%o) {x1 , x2 } (%o) [p, q] (%o) 2.30258509299

2.5.1 Creating Lists with Commands


Often rather than entering the items in a list by hand, it is safer and easier
to create the list using the makelist() command. This command, which we
used earlier, requires an expression in terms of a counter variable (any name
5
We repeat that, unlike sets, lists allow multiple copies and that order matters.
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 29

will do), then the counter variable itself, and finally a start and end value
(additional arguments can be inserted; see the Maxima Manual ). Each of
the three commands below creates a list, to which a name has been assigned.
The names can be used to recall the list or items in the list.

(%i) xList: makelist(i,i,0,10);


sqrtList: makelist(sqrt(x),x,0,10);
halfList: makelist(x/2, x, 0, 10);
(%o) [0, 1, 2,√3, 4, √5, 6, √ 7, 8,√ 9, 10] √ √
3
(%o) [ 0, 1, 2, 3, 2, 5, 6, 7, 2 2 , 3, 10 ]
(%o) [0, 12 , 1, 32 , 2, 52 , 3, 72 , 4, 92 , 5]

Maxima contains many built-in operations like sqrt(), sin(), and log()
that can be applied directly to the items in a list, as below. Also, some
operations like dividing all items by the same constant or raising them to
the same power can be applied directly to a list. The next three commands
show a quick way to create the items in sqrtList and halfList, along with
a way to square each item in xList.

(%i) sqrt(xList);
√ √ xList/2;
√ √ √xList^2; √
3
(%o) [0, 1, 2, 3, 2, 5, 6, 7, 2 2 , 3, 10]
(%o) [0, 21 , 1, 32 , 2, 52 , 3, 72 , 4, 92 , 5]
(%o) [0, 1, 4, 9, 16, 25, 36, 49, 64, 81, 100]

Some operations cannot be applied to a list this way. Maxima’s map() com-
mand can be applied in such cases. This command can also be used instead
of some of those that we have already seen. The next input/output group
shows how to apply some basic operations to two lists of equal length (be
aware that you must avoid illegal operations like dividing by zero). Note
the use of quotation marks to indicate the binary operation that is being
conducted.

(%i) A: [2,4,6,8]$ B:[1,5,7,9] map("+",A,B); $


map("-",A,B);map("*",A,B);map("/",A,B); map("^",A,B);

(%o) [3, 9, 13, 17] (%o) [1, −1, −1, −1]


(%o) [2, 20, 42, 72] (%o) [2, 45 , 67 , 89 ]
(%o) [2, 1024, 279936, 134217728]
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 30

The first command below takes the factorial of each term in xList. The
map() command requires the following: a specification of the operation to
be applied, in quotation marks, and the list or lists to which the operation
applies.
Operations or functions can be mapped onto a single function. The next
input/output group shows two ways to return the factorial of the values
in xList. The first approach directly applies the factorial command, !. The
second approach is to create a named function fact(x) and then to map that
function onto the values. Be aware of the slight differences in the syntax.

(%i) map("!",xList);
fact(x):=factorial(x)$ map(fact, xList);
(%o) [1, 1, 2, 6, 24, 120, 720, 5040, 40320, 362880, 3628800]
(%o) [1, 1, 2, 6, 24, 120, 720, 5040, 40320, 362880, 3628800]

To illustrate the second approach in more detail, we create a function with


the command below and than map the function onto xList.

(%i) f(x):=a*x^b$ exprList: map(f, xList);


(%o) [0, a, a 2b , a 3b , a 4b , a 5b , a 6b , a 7b , a 8b , a 9b , a 10b ]

The subst command can be used to determine the values of the expressions
above, given specified values of the parameters a and b.

(%i) valueList: subst([a=5,b=2], exprList);


(%o) [0, 5, 20, 45, 80, 125, 180, 245, 320, 405, 500]

We assigned the list a name, valueList, so individual items can be extracted.


To extract the fourth item in exprList and its counterpart in valueList, we
use the command [exprList[4], valueList[4]]. Note the use of brackets
to indicate the item number. Placing a set of commands inside a list causes
Maxima to place the output into a list: [a 3b , 45].
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 31

2.5.2 Describing a List


Specific information about a list is often useful. For example, if the list is
long, we might wish to know the number of items that it contains. We might
wish to know the sum of the items in the list, or their mean value, or their
standard deviation. The commands below result in output that shows that
xList contains 11 items, the sum of which
√ is 55, and the mean of which is
5. The (sample) standard deviation is 11.

(%i) xLength: length(xList);


xSum:sum(xList[i],i,1, xLength); xMean:(xSum/xLength);
sqrt(sum((xList[i]-xMean)^2,i,1,xLength)/(xLength-1));

(%o) 11 (%o) 55 (%o) 5 (%o) 11

To determine the minimum or maximum value, we must use the apply com-
mand, as below. The workbook provides details regarding this operation.

(%i) [apply(min,xList), apply(max,xList)];


(%o) [0, 10]

Maxima offers a module, descriptive, that computes these values. The


purpose of doing this by applying the formulas directly is to illustrate the
manipulation of lists. Note that the names for the minimum and maximum
values are smin and smax. If you work much with large data lists, then
descriptive will be of value.

(%i) load(descriptive)$
mean(xList); std1(xList); smin(xList); smax(xList);

(%o) 5 (%o) 11 (%o) 0 (%o)10

2.5.3 A Matrix as a List of Lists


In general, a matrix is a list of lists, all of which must have the same length.
Thus a matrix is rectangular. Chapter 9 details the uses of matrices to
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 32

conduct mathematical analysis. Here, we limit their use to creating tables.


Consider the simple example above, in which we listed the minimum and
maximum values of x. With a matrix, we could add a list of names to these
values, as shown below.

(%i) matrix(["Minimum x", "Maximum x"],


[apply(min,xList), apply(max,xList)]);
 
Minimum x Maximum x
(%o)
0 10

2.6 Relations
Any ordered pair (triple, quadruple, and so forth) of values constitutes a
relation. Given the ordered n-tuple (x, y, z, . . .) a relation among the variables
exists whenever every set of values for any of n − 1 of the variables implies
one or more values for the remaining variable. For example, suppose that
x + 2 · y − 1.5 · z 2 = 0. Then specifying values for any two of the variables
implies one or more values for the third.
Using the command expr:x + 2*y - 1.5*z^2 = 0 we enter this expression.
We then assign values to two of the variables and solve for the third variable.
Assigning values to x and y can result in more than one z value. Assigning
values to x and z implies a single y value. Likewise, assigning values to y
and z implies a single x value.
To support these assertions, we use three commands that solve the expression
for one of the two variables in terms of the other two. The subst commands
below determine the implications of given values of y and z for x, of x and z
for y, and of x and y for z. For x and y, singe values result; but for z, a list
of two values is reported.

(%i) expr: x + 2*y - 1.5*z^2 = 0;


subst( [y=4, z = 3], solve(expr, x) );
subst( [x=8, z=4], solve(expr, y) );
subst( [x=1, y=3], solve(expr, z) );
(%o) −1.5 z2 + 2 y + x = 0√ (%o)

= √11
[x √ 2
]]
2 7 2 7
(%o) [y = 8] (%o) [z = − √3 , z = √3 ]
CHAPTER 2. VARIABLES, SETS, LISTS, AND RELATIONS 33

2.7 Questions and Problems


1. Some historians claim that they do not use models in writing history
and in arriving at conclusions about historical phenomena. Is it possible
to analyze something without having an underlying model? Will hard
work produce insights and generalizations if you do not have a model?
Explain.

2. Indicate whether the sets of numbers below can be properly described as


being any or all of the following: integers, fractions, rational numbers,
real numbers, irrational numbers.
(a) {-5, -1, 2, 4}
(b) {4/3,1/2,-3/8,11/12}
p p p
(c) { (2), (3), π, (11)}.

3. Refer to the three sets in (2). Suppose that we replace {} with [],
indicating that the three quadruplets are list, not sets. How would this
change affect the interpretation of the values?

4. Apply the Maxima command sort p to p


these lists:
p (a) [-5, -1, 2, 4],
(b) [4/3,1/2,-3/8,11/12], and (c) [ (2), (3),π, (11]). What do you
perceive is Maxima’s default direction of sorting? If you’re curious
about reversing this order, see the sort command in the manual. (In
wxMaxima, click on the word sort in any command and hit the F1
key.)
Chapter 3

Rectangular Coordinates and


Functions

This chapter builds on material from Chapter 2, which shows that variables
may be related via mathematical expressions. This chapter focuses on a
subset of those expressions, functions. It examines three types of functional
relationships. The first, explicit functions, exist when the value of some
variable is determined by an explicit relationship between that variable and
the value(s) of one or more other variables. The second, implicit functions,
exist when a set of two are more variables must jointly satisfy the condition
that a mathematical expression imposes. We saw an example of an implicit
function at the end of Chapter 2. Finally, two or more variables’ values
can be bound by the fact that each of these variables is functionally related
to another set of one or more variables, which are called parameters. Our
variables are said to be related via parametric equations (sometimes called
freedom equations).
Much of our illustrative analysis involves just two variables. In some cases,
the important relationship actually involves just two variables. In other
cases, the relationship can involve more than two variables, but the method
of approach can be outlined with the two-variable case and then extended.
Because of the importance of the two-variable case, this section begins with
an extension of the number line that Chapter 1 developed, showing two
variables’ values simultaneously with the use of rectangular coordinates. Oc-
casionally, we extend the analysis to include a third dimension. Furthermore,
the reasoning involved in creating rectangular coordinates in a plane can be

34
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 35

extended to any dimension, a topic that we address in Chapter 7 and again


in Chapter 9, which treats linear algebra.

3.1 Rectangular Coordinates


The concepts of the real number line and ordered pairs enable us to the
rectangular (or Cartesian) coordinate system. Suppose that we have two real
number lines that are perpendicular to each other. The two dotted lines in the
next figure are suitable examples. The intersection of these two real number
lines is designated the origin of our coordinate system. The horizontal line
called the x axis) and the vertical line (called the y axis) together form the
coordinate axes.
Before treating the nature of the rectangular coordinate system more com-
pletely, we address a few details regarding the commands and the resulting
output. The draw2d command generates Figure 3.1, a two-dimensional fig-
ure with rectangular coordinates. The “wx” prefix instructs wxMaxima to
place the resulting graph in the workbook. Setting the xaxis and yxaxis
options to true produces the two dotted lines that define the quadrants of
the space. We name the axes (using xlabel and ylabel) and turn off the
listing of values (using xtics and ytics). Then we set the ranges for X and
Y . Finally, we create four labels that describe points in the quadrants. Each
label command contains some text and the coordinates to which the labels
attach.

(%i) wxdraw2d( title="Four Quadrants",


xaxis=true, yaxis=true, xlabel="X", ylabel="Y",
xtics=false, ytics=false,xrange=[-6,6],yrange=[-6,6],
label(["*I (+,+)",5,5]),label(["*II, (-,+)",-5,5]),
label(["*III (-,-)",-5,-5 ]),
label(["*IV (+,-)",5,-5]))$

Both of the coordinate axes have the basic properties of any real number
line. Points to the right of the origin on the, x axis, and upward on the y
axis indicate positive values; points to the left of the origin on the x axis and
downward on the y axis represent negative values.
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 36

Figure 3.1: Rectangular Coordinates: The Four Quadrants

Each ordered pair of real numbers is represented by a unique point on the


plane that is formed by the x and y axes. Examples appear in Figure 3.1
as asterisks. Given an ordered pair (a, b), the x coordinate, or abscissa of
the variable x on the x axis, is always the first element in the ordered pair.
The second element of the ordered pair is the y coordinate, or ordinate of
the variable y on the y axis. This means that the point given by the ordered
pair (a, b) is not the same as the point given by the ordered pair (b, a) unless
a = b.
The x and y coordinates that comprise an ordered pair indicate the location
of the point given by that ordered pair. As Figure 3.1 demonstrates, when
the elements of ordered pair (a, b) are both positive, then the point that
corresponds to this ordered pair lies in the first quadrant, that is, the area
of the coordinate system that lies to the right of the y axis and above the x
axis.
When the signs of the ordered pair (a, b) are (-, +), then the point in question
lies in the second quadrant, to the left of the y axis and above the x axis.
When the signs of the ordered pair (a, b) are both negative, then the point
lies in the third quadrant, to the left of the y axis and below the x axis.
Finally, when the signs of the ordered pair (a, b) are (+, -), the point lies
in the fourth quadrant, to the right of the y axis and below the x axis. By
convention, quadrant numbers are indicated by Roman numerals.
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 37

The two real number lines that form the coordinate system need not have the
same units of measurement. That is, the units in which they are measured
need not be the same. Indeed, often they cannot be the same. One axis might
measure profit and the other axis number of units sold, so the units for one
is dollars per time period and the other is some measure of physical units
per time period (the two time periods need not be the same). Or one axis
might measure price in terms of dollars per unit and the other axis quantity
in terms of physical units. This latter example describes the coordinate axes
that we use to conduct supply-and-demand analysis.

3.2 Functions
Consider these two equations: y = x2 and y 2 = x. Both are equations, but
for only one of the two is y a function of x. When y = x2 , each value of x
implies a single value of y. Such is not the case when y 2 = x, for in that case
a particular x value can be related to either of two y values. For example,
when x = 4, y can be either -2 or +2, for squaring either yields a value of
4 and satisfies the equation. We summarize this introduction with a formal
definition: A function is a relation (a set of ordered pairs) such that no two
ordered pairs have the same first element.
A function is denoted by y = f (x), which is read “y is a function of x” (not
“y equals f times x”). Other functional notations that are frequently used
include g(x), h(x), and F (x). To create a functional expression in Maxima,
use notation like this: f(x, a) := (a*x^2). The next cell shows this ex-
pression when a’s value is not specified, again when each of two values of a
is specified, and finally when both a and x are specified. All five commands
are placed into a list, so the output also appears in a list. Be aware that the
values entered must be entered in the order specified in parentheses, x first
and then a.

(%i) [f(x,a) := a*x^2, f(x,3), f(5,3), f(x,-3), f(5,-3)];

(%o) [f (x, a) := a x2 , 3 x2 , 75, − 3 x2 , − 75]

A function is just a special case of a relation. A function exists when no


two ordered pairs have the same value for x, but different values for y. The
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 38

definition of a function implies that it must be a relation. The reverse,


however, is not true. Not every relation is a function, since in a relation
there may be several ordered pairs that exhibit the same value of x, but
different values of y. Functions are therefore a subset of relations.
A different letter must be used to denote each function used in a particular
problem. For example, if quantity demanded qd and quantity supplied qs are
different functions of price p, then the functional notation used must reflect
the fact that two different functions exist. Hence one might write qd = f (p)
and qs = g(p).This indicates that the two functions are not equivalent, that
is, that f (p) 6= g(p) in general. Of course for one p value, the one that
corresponds to market equilibrium, f (p) = g(p).
Given the functional relationship y = f (x), the value of variable y depends
on the value of variable x. Once x takes on a particular value, y’s value is
determined. Variable y therefore depends on x, and is referred to as the de-
pendent variable, whereas variable x is the independent variable. A function
like y = f (x) is often referred to as an explicit function: y’s value is explicitly
related to x’s value in a way that f (x) defines. Later in this section, we con-
sider two other ways that relationships between x and y might be specified.

3.2.1 Domain and Range


Often only certain values are permissible for both the independent and de-
pendent variables. Most relationships in business and economics involve real
numbers, not imaginary values. Some, like units of output, cannot assume
negative values; others, like profits, can assume negative values. Let output
be defined by f (x) and profits by g(x). A formal statement of the limits that
we observed is that the domain of f (x) consists of 0 and the positive real
numbers. The profit function g(x) is limited to the real numbers, its domain.
In general, given the y = f (x), the set of all values that x may assume is the
domain of the function.
Likewise, sometimes only certain values are permissible for the dependent
variable. Let y = x2 .Then, regardless of what real number value the inde-
pendent variable x assumes, the value of y cannot be negative. The range
of variable y is prescribed and limited. In general, given y = f (x), the set
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 39

of all values that variable y may assume is referred to as the range of the
function.1

3.2.2 Implicit Functions and Parametric Equations


As noted, for a function of the form y = x2 , y is an explicit function of x.
The dependent variable y has a value that is uniquely defined by a value of
the independent variable x. We also noted that y 2 = x is not such a function.
We can, however, express the latter as y 2 –x = 0 and assign that expression
the name F (x, y). This is an implicit function. It expresses a relationship
between x and y but does not assign roles of dependence to either.
Yet another possibility is that x and y are determined by equations that
involve a third variable, which we label t. The value of t can be treated as a
parameter in the two equations and by letting t vary over some range, we can
deduce the behavior of x and y and, therefore, how they relate to each other.
We need not, and perhaps cannot, deduce an equation that relates x and y.
An importanat economic example is this: a large set of production functions
allow for both output and cost to be stated in terms of the employment level
of a single input but does not generate an expression of output in terms of
cost or vice versa.
The graph below shows how draw handles an explicit function and an implicit
function.2 For the first function, y = x2 , the domain is the entire real number
line and the range is the nonnegative portion of the real number line. The
second graph shows the implicit function y 2 –x = 0. For this implicit function,
x can√ take on only nonnegative real numbers. This function implies that
y = x, which involves imaginary numbers for x < 0. The values of y can
range over the entire real number line.

(%i) first:gr2d(title="y = x^2",explicit(x^2, x, -5,5))$


second:gr2d(title = "y^2 = x",
1
In the draw command in Maxima, the options xrange and yrange define the ranges
over which values are to be graphed. Thus, xrange must be within the function’s domain
and yrange must be within the function’s range, as those terms are defined in the text.
2
The two gr2d commands create scenarios that are assigned names. The result of each
is that an object is stored in Maxima’s memory. Executing wxdraw with the object names
inside the parentheses causes Maxima to graph the two named scenarios. As an exercise,
replace gr2d with wxdraw2d and graph each of these expressions separately.
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 40

implicit(y^2=x,x,0,5,y,-sqrt(5),sqrt(5)))$
wxdraw(first,second, columns=2)$

Figure 3.2: Explicit and Implicit Functions


CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 41

Exercise Set 2.1

1. For each of the following, determine the range of the dependent variable
y and the domain of the√independentp variable x.
(a) y = 8 + x (b) y = x (c) y = (4 − x2 )
(d) y = x21−1 (e) y = 8−x 1

2. Write out a few (a, b) that are consistent with the following expressions.
Indicate which expressions can be cast as explicit functions with y as
the dependent variable.
(a) y = px2 (b) y = x4 (c) y 2 = x (d) y 3 = x (e) y 4 = x
(f) y = (x) (g) y = 1/x (h) y = π · x2 (i) x2 + y 2 = 4
2 2
(j) y + x = 1 (k) y + x = 9 (l) x = 3 (m) y = 1/2

3. Use the wxdraw2d command to graph the expressions in (h) and (k)
from the list above.

4. The total revenue, which is defined as T R = P ·Q of a firm per day, is a


function of its daily sales Q. Assume that the firm’s output capacity is
10 to units per day. What p are the domain and range of TR if the price
is defined as P = 1200/ (Q)? (This is the inverse demand curve.)

5. A supply curve is a functional relationship between quantity supplied


and price. Graph the following supply schedule. Be careful when you
label your axes.

(%i) Qlist:["Q, units per week :"


,1000,2000,4000,7000,11000];
Plist: ["P, $ per week : ", 5,6,7,8,9];
matrix(Qlist,Plist); 
Q, units per week : 1000 2000 4000 7000 11000
(%o)
P, $ per week : 5 6 7 8 9

3.2.3 Composite Functions


Given two functions f (x) and g(z), it may be possible to create a new function
by means of an operation known as composition. Given y = f (x) and x =
g(z), the composite function or composition of f (x) and g(z) is the function
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 42

h(z) for which y = f (g(z)) = h(z). The functional relationship y = f (g(x))


is read as, “the function f of a function g, of z,” and indicates that variable y
is a function of variable x, which itself is a function of variable z. To evaluate
the function h(z),one must first compute g(z)and then evaluate f (x) at the
point g(z).The function h(z) is defined only at those points z for which g(z)
is in the domain of f (x).

EXAMPLE: Given that y = f (x) =√ x and x = g(z) = z + 1, the composite
function is given by y = f (g(z)) = z + 1.
Note that the function g(z) is defined for all real values of z, whereas the
function f (x) is defined only for those values such that x > 0. This means
that the composite function f (g(z)) can be satisfied only when (z + 1) ≥ 0.
(The symbol ≥ means, “greater than or equal to,” while the symbol ≤ means
“less than or equal to.”)
WARNING: Be careful in reading the composite function notation: f (g(z)) is
√ f (x) · g(z). Therefore, the composite
read “f of g of z.” It√is not the product
function f (g(x)) is z + 1 ; it is not x · (z + 1).
The input/output cell below illustrates the fact that Maxima understands
composite functions.

(%i) [f(x):=sqrt(x),
√ g(z):=z+1, f(g(z))];

(%o) [f (x) := x, g (z) := z + 1, z + 1]

Using the function notation shown above, f(x) := an expression and greatly
facilitate the evaluation of a function for a number of values. Suppose that
we wish to evaluate for values of x from 51 through 58, or for a single value
of x. The first command below creates a list of x values and binds that list to
a name. The second command creates the functional expression. The third
command applies the function to the list of x values and assigns the result
to the name yList. The the matrix command creates a table of values. The
final command applies the function to a single value, x = 81. The results
appear as a table and as a single value, 5 · 9 = 81.

(%i) xList:makelist(50 + 1.0*i,i,1,)$ f(x):=5*sqrt(x)$


yList: f(xList)$ matrix(xList,yList); f(81); 
51.0 52.0 53.0 54.0 55.0 56.0 57.0 58.0
(%o)
35.707 36.055 36.4 36.742 37.08 37.416 37.749 38.078
(%o) 45
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 43

Exercise Set 2.2 Evaluate these expressions by hand and again with Max-
ima.

1. Given that f (x) = 100 + 7x find (a) f (0), (b) f (5), and (c) f (−10).

2. Given that f (x) = 10 − 4x find (a) f (1), (b) f (10) (c) f (a + h).

3. Given that f (x) = x2 + 4x − 6 find (a) f (0), (b) f (10), (c) f (−2).

4. Given that f (x) = 1/x2 find (a) f (2), (b) f (−4), (c) f (x + h).

5. Given that f (x) = 2x find (a) f (0), (b) f (3), (c) f (−3).

6. Given that f (x) = x2 − 2x + 2 show that f (−2) 6= −f (2).

7. Given that f (x) = x2 show that f (x + h) − f (x) = h(2x + h).

8. Given that f (x) = (1 + x)/(1 − x) show that f (1/x) = −f (x) and


f (−1/x) = 1/f (x).

9. Given that f (x) = x?(x − 1) show that f (x + 1) = f (−x).

10. Given that f (x) = x2 + 8x − 3 and g(z) = 2 find f (g(z)).

11. Given that f (x) = 4x − x2 and g(z) = 1/x find f (g(z)) and g(f (x)).

3.2.4 Functional Forms


This section discusses specific functional forms that we use subsequently. For
now, we limit the discussion to functions of two variables, typically of the
form y = f (x).

Polynomial Functions

A polynomial is a very general functional form that can represent many


relationships. A polynomial function y = f (x) is defined for all real values
of x by an equation of the form y = a0 · x0 + a1 · x1 + a2 · x2 + · · · + an · xn .
Because x0 = 1 and x1 = x, we can rephrase this expression in this more
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 44

familiar form: y = a0 + a1 · x + a2 · x2 + · · · + an · xn . Often, this more compact


representation is useful:
n
X
y = f (x) = ai · x i .
i=0

The symbols, a0 , a1 , . . . , an are coefficients of the polynomial.


It can be useful to refer to polynomials in terms of their degree. The degree
of a polynomial function is the highest nonnegative integer power of any
independent variable in the polynomial. Consider these examples:

• y = 5 · x + 3 is a first-degree polynomial.
• y = 5 · x2 is a second-degree polynomial.
• y = 5 · x + 5 · x2 is also a second-degree polynomial.
• y = 5 · x + 5 · x2 + 5 · x5 is a fifth-degree polynomial.

Exponents and Bases

In considering polynomial functions, we introduced the concept of the degree,


or highest power, of a polynomial. The power to which a variable (or a
number) is raised is referred to as the exponent of that variable or number.
The expression x · x · x · x · · · · xn represents the product of n x’s, all of which
have the same value. The variable x is referred to as the base of the term
xn , and the letter n is called the exponent of variable x.
The cell below states some of the laws that govern the use of exponents and
gives an example of each of these laws in operation. Using the operations
that are reported in the output, use Maxima to determine the results of the
following: (xm )n , xm /xn , (x/y)n , (x · y)1/n , x1/n /y 1/n , xm/n .3
3
We repeat a note made earlier: We use the matrix command throughout the book
to create tables. Chapter xx address the applications of matrix to mathematical analysis.
Notice that a matrix is made up of a number of lists, all of which have equal length. The
lists can consist of values, strings of text, or (as above) combinations of the two.

Note the present of radcan in the third example. Copy this cell, remove the radcan
—don’t forget to remove both parentheses—and see what difference occurs. The term
radcan stands for radical canonical.
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 45

(%i) matrix(["Operation", "General Result", "Example"],


["x^0",x^0, concat("8^0 = ",8^0)],
["x^m*x^n",x^m*x^n,x^5*x^3] ,
["(x*y)^n",radcan((x*y)^n), (x*y)^3] ,
["x^(-n)",x^(-n),x^-2],["x^(1/n)",x^(1/n),x^(1/2)]);

Specific Polynomial Functions

Figure 3.3 shows an example of four specific types of polynomial functions, all
of which will be used in examples in this book and all of which are frequently
used to analyze and illustrate economic and business issues. The general
expressions for the four are these: constant, y = x0 ; linear, y = a + b · x;
quadratic, y = a + b · x + c · x2 ; and cubic, y = a + b · x + c · x2 + d · x3 .

(%i) wxdraw2d(xaxis=true,yaxis=true, yrange=[-15,15],


key="A constant function", explicit(3, x, -5, 5),
color = black, key="A linear function",
explicit(2 + 0.75*x, x, -5, 5), color=red,
key="A quadratic function",
explicit(-2 - 1.5*x + 0.25*x^
2, x, -5, 5),
color = gray40, key="A cubic function",
explicit(-5 + 0.5*x - 0.1*x^
2 + 0.05*x^
3, x, -5, 5))$
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 46

Figure 3.3: Four Polynomial Functions

We have already seen examples of constant and linear functions, both in the
simple Keynesian model. We will encounter all of these functional forms as
we encounter various models, such as those of cost curves, that require the
additional flexibility that terms with larger exponents provide.

Determining a Polynomial’s Coefficients from Points

If we can observe any two points on a straight line, then we can compute the
line’s slope and intercept using techniques that you learned in high-school
algebra. Likewise, if we can observe any three points, then we can compute
the coefficients of a quadratic equation. Finally, any four points determine
the coefficients of a cubic equation.4 The output below shows an example
of using Maxima to determine the coefficients of a quadratic equation and
plotting that equation.

(%i) [x0,y0]:[10,5]$ [x1,y1]:[20,50]$ [x2,y2]:[30,45]$


solve([y0=a+b*x0 + c*x0^2, y1=a+b*x1+ c*x1^2,
y2=a+b*x2+c*x2^2],[a,b,c]);
expression: subst(%[1], y = a+b*x+c*x^2 ) ;
4
Generally, if the number of points equals the degree of the polynomial plus 1, then
the polynomial’s coefficients can be computed.
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 47

wxdraw2d( xaxis=true,title="A Quadratic Equation",


explicit(rhs(expression),x,0,40), point type=
filled diamant,point size = 2, color=red,
points([ [x0,y0],[x1,y1],[x2,y2] ]));
2
(%o) [[a = −90, b = 12, c = − 14 ]] (%o) y = − x4 + 12 x − 90

Figure 3.4: Fitting a Cubic Polynomial to Three Points

The values of the (x, y) pairs were selected arbitrarily. For each x value, this
must be true: y = a + b · x + c · x2 , so the first list in the solve command
is a list of three points at which that statement must be true if a, b, and c
are indeed the proper coefficients. The second list is the list of unknowns
for which we seek values. The solve output consists of a list with another
list embedded. The expression that we name expression is obtained by
extracting the inside list, by using the [1] after the %. The % symbol refers
to the output that results from the most recent command. Thus we are
inserting this list: [a = −90, b = 12, c = −1/4] into the expression for the
quadratic equation. As exercises, repeat this process in order to define the
coefficients of a linear polynomial and a cubic polynomial and to graph the
implied polynomials.

Determining the Coefficients from Points

If we can observe any two points on a straight line, then we can compute the
line’s slope and intercept using techniques that you learned in algebra. Like-
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 48

wise, if we can observe any three points, then we can compute the coefficients
of a quadratic equation. Finally, any four points determine the coefficients of
a cubic equation. The output below shows an example of using Maxima to
determine the coefficients of a quadratic equation and to plot that equation.

(%i) [x0,y0]:[10,5]$ [x1,y1]:[20,50]$ [x2,y2]:[30, 45]$


solve([y0=a+b*x0 + c*x0^2, y1=a+b*x1+ c*x1^2,
y2=a+b*x2+c*x2^2],[a,b,c]); expression: subst(%[1],
y = a+b*x+c*x^2 );
wxdraw2d( xaxis=true,title="A Quadratic Equation",
explicit(rhs(expression),x,0,40),
point type=filled diamant,point size = 2, color=red,
points([ [x0,y0],[x1,y1],[x2,y2] ])); $
2
(%o) [[a = −90, b = 12, c = − 14 ]] (%o) y = − x4 + 12 x − 90

Figure 3.5: Fitting a Polynomial Through Three Points

The values of the (x, y) pairs were selected arbitrarily. For each x value, the
following must be true: y = a + b · x + c · x2 . Therefore, the first list in
the solve command is a list of three points at which that statement must
be true if a, b, and c are indeed the proper coefficients. The second list
contains unknowns for which we seek values, a, b, and c. The solve output
consists of a list with another list embedded. The expression that we name
expression is obtained by extracting the inside list, by using the [1] after
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 49

the %. The % symbol refers to the output that results from the most recent
command. Thus we are inserting this list: [a = -90, b = 12, c = -1/4]
into the expression for the quadratic equation. As exercises, repeat this
process in order to define the coefficients of a linear polynomial and a cubic
polynomial and to graph the polynomials.
Sometimes, we know the value at a point and we know (or have a good
estimate of) the slope at that point and we wish to use that information to
determine the linear function that passes through that point. The cell below
shows an example, where a line with a slope of -2.25 passes through the point
(20, 100). The first command solves the relevant equation and assigns the
solution the name soln. The second command substitutes a = 145 into the
expression y = a−2.25·x to yield the expression for the linear equation. The
wxdraw2d command draws the line over a range of x values, yielding Figure
3.6. The drop line shows the designated point through which the line passes.

(%i) [soln: solve(100 = a -2.25*20,a),


expr:subst(soln, a - 2.25*x)];
wxdraw2d( xaxis=true, xlabel="x",ylabel="y",
key=string(expr),explicit(expr,x,0,70), color=black,
line width=1, line type = dots, points joined=true,
points([[0,100],[20,100],[20,0]]) )$
(%o) [[a = 145], 145 − 2.25 x]

Figure 3.6: A Fitted Line


CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 50

More on Quadratic Functions

The quadratic form is general enough to apply to many economics and busi-
ness issues, so examining it in more detail is warranted. We begin with the
quadratic formula, which is shown below.

(%i) solve(a√ + b*x + c*x^2,√ x);


2 2
(%o) [x = − b −42c
a c+b
, x = b −4
2c
a c−b
]

This formula provides a rule to find the roots of a quadratic equation. That
is, is shows the values of x for which y = 0.5 We see that two such values seem
to occur. We say “seem to” because our attention is limited to real numbers.
If b2 − 4 · a · c < 0, then no real solutions occur. Also, if only positive x values
make economic sense, then the number of solutions can be 0, 1, or 2. The
examples below show three quadratic equations. The first, named p1, has no
real solutions; the second, p2, has two real solutions but only one for x > 0;
and the third, p3, has two real solutions for positive values of x. numbers.
The term b2 − 4 · a · c is called the discriminant: it discriminates between
equations with no real solutions and those with at least one real solution.

(%i) [p1, p2, p3]: [50 - 5*x + 0.5*x^2, 10 + x/3 - x^2/3,


10 - 10*x + x^2];
wxdraw2d( xaxis=true, xlabel="x", ylabel = "y",
key = "No real solution", explicit(p1, x, -5, 10),
color=red, key = "One positive solution",
explicit(p2, x, -6, 10), color = black,
key = "Two positive solutions",
explicit(p3,x,-5,10))$
2
(%o) [0.5 x2 − 5 x + 50, − x3 + x3 + 10, x2 − 10 x + 10]

An Economic Application: Total and Marginal Cost Curves

A firm’s cost curve is often represented by a cubic equation like the one
graphed below. If T C = 50 + 20 · x − 2 · x2 + 0.25 · x3 , then its marginal cost
5
Factoring a quadratic or completing the square are two other ways to find its root.
Every quadratic equation can, however, be solved by using the quadratic formula.
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 51

Figure 3.7: Three Quadratic Equations

curve is M C = 20 − 4 − x + 0.75 · x2 . Figure 3.8 shows this illustration of


how the curves relate to the quantity produced, x. Average cost is simply
T C/x.

(%i) total:gr2d(xlabel="x",ylabel="$", yrange=[0,350],


ytics=100,key="Total Cost",
explicit(50+20*x-2*x^2+0.25*x^3,x,0,10 ) )$
perunit: gr2d( xlabel="x" ,ylabel="$ per unit",
yrange=[0,60], ytics=20, key="Marginal Cost",
explicit(20 - 4*x + 0.75*x^2,x,0,10 ),
color=red, key="Average Cost",
explicit((50+20*x-2*x^2+0.25*x^3)/x,x,0,10))$
wxdraw(total, perunit, dimensions = [480,480])$
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 52

Figure 3.8: Cost Curves

Exercise Set 2.5


Graph the following functions and determine their critical values. Solving
first can guide the drawing of the graph in that it indicates the range of x
values for which f (x) is to be graphed.

1. x2 − x − 6 = 0

2. x2 − 25 = 0

3. x2 + 6 · x + 8 = 0

4. 3 · x2 + 7 · x − 3 = 0

5. x2 − 4 · x + 4 = 0

6. x2 + x − 12 = 0
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 53

7. x2 − 5 · x + 3 = 0

8. x3 − 5 · x2 − x + 5 = 0

Exponential and Logarithmic Functions

Polynomials are useful in conducting analysis, but often other functional


forms serve better. Two related classes of functions, exponential functions
and logarithmic functions, often serve to represent the relationships that
represent economic phenomena.
First, consider exponential functions. Earlier in this chapter, we introduced
the concept of a power function, which we defined as a√variable raised to a
constant power, for example, y = x2 or y = x( 1/2) = x . We extend our
use of exponents to the case in which the exponent itself may be a variable,
for example, y = 3x or y = a · 31/x . In such cases, the base (which is 3)
is fixed and the exponent contains the variable. In general, an exponential
function has the form, y = a · bx where b is the fixed base such that b > 0,
b 6= 1, and x is an independent variable that is any real number.
The first restriction on b’s value, is to preclude roots that are imaginary
numbers. The second restriction, reflects that fact that 1, raised to any
power, is still 1. Figure 3.9 shows the values of y for a range of x values given
three values of b: 0.5, 0.75, 1.25, and 1.5. You should experiment with other
values. Why do all of the curves pass through (0,1)? Observe that all of
these functions’ values are in the first and second quadrants: for all x values.

(%i) wxdraw2d(user preamble="set key left", xlabel="x",


ylabel="y = b^x", key = "b = 0.5",
explicit(0.5^x, x, -2, 5), color=red,key =
"b = 0.75", explicit(0.75^x, x, -2, 5), color=black,
key = "b = 1.25", explicit(1.25^x, x, -2, 5), color=
orange, key= "b = 1.5", explicit(1.5^x,x,-2,4),
color=black,line width=1,line type=dots,
key="(0, 1)", points joined=true,
points([[-2,1],[0,1],[0,0]]), dimensions=[480,360])$
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 54

Figure 3.9: Four Exponential Functions

An especially important value of b is the “Euler number,” 2.718 . . . , the base


of the natural logarithms system. Consider the expression y = (1 + 1/x)x .
The increments to 1, which are 1/x, become increasingly small as x increases.
At the same time, however, the size of the exponent increases in inverse
proportion to the change in the increment size.1 The table shows that f (x)
converges on the value of e as the increment size becomes small and the
exponent becomes large). This process involves the concept of a limit, which
Chapter 4 develops more formally.

(%i) (x):=(1 + 1/x)^x$ xList: [1,2,3,10,20, 100]$


iList:1/xList$ yList:float(map(f,xList))$
matrix(cons("Value of x: ",xList),
cons("Increment size: ", iList),
cons("Value of y: ", yList));
 
Value of x : 1 2 3 10 20 100
1 1 1 1 1
(%o) Increment size : 1 2 3 10 20 100

Value of y : 2.0 2.25 2.3703 2.5937 2.6532 2.7048

The logistic function is an important case of an exponential function. This


function has been used to model growth of forests and wildlife, and it has
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 55

been applied to learning curves and to the fraction of a population that has
adopted a new product or technology. The function is f (t) = K/(1+e1+a+b·x ).
The value K defines the upper limit. The parameters a and b (b < 0)
determine the function’s curvature and height properties. Euler’s number is
e, and t is time. The cell below defines a function for which K = 0.8, a = 0,
and b = −0.2, shows some values of that function, and graphs the function.
The map command is used to map the function onto the list of t values.

(%i) f(t):= 0.8/(1 + exp(-0.2*t))$


tList: makelist(t-10, t, 0, 40,5)$
yList: map(f, tList)$
matrix( cons("t",tList), cons("y",yList) );
wxdraw2d( xlabel="t",ylabel="y", yrange=[0,1],
yaxis=true, key="A Logistic Curve",
explicit(f(t),t,-10,30 ) )$

Figure 3.10: A Logisitic Curve

For this set of values, the population has been growing and has reached one-
half its maximum size when observations begin (t = 0). When y is relatively
small, the per-period growth rate is high and, at first, increasing. Then the
growth rate decelerates, asymptotically approach zero. In this illustration y
is only very slightly less than K, its maximum value, by year 30.
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 56

We can generalize the simple exponential function by attaching a coefficient


a to the base b and another coefficient c to the variable x. Now our ex-
pression is y = a · bc·x . The graphs below show that a affects the function’s
height; specifically a is the function y-intercept. The coefficient c affects the
curvature.

(%i) f(x,a,b,c) := a*b^(c*x)$


avals :gr2d(title="a*b^c*x; b=3, c=1", yaxis=true,
xlabel="x",ylabel="y", yrange=[0,6], xtics=2,
key="a = 1",explicit( f(x,1,3,1),x,-2,2), color=red,
key="a=2",explicit(f(x,2,3,1),x,-2,2) )$
cvals : gr2d(title="a*b^c*x; a=1, b=3", yaxis=true,
xlabel="x",ylabel="y", yrange=[0,6], xtics=2,
key="c = 1",explicit(f(x,1,3,1),x,-2,2), color=red,
key = "c = -1/2",explicit(f(x,1,3,-1/2),x,-2,2) )$
wxdraw(avals, cvals, columns=2)$

Figure 3.11: Four Exponential Functions

Exercise Set 2.6

1. Sketch on the same set of axes the graphs of the exponential functions
y = bx for these values of b: 4, 8, and 12.
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 57

2. Sketch on the same set of axes the graphs of the exponential functions
y = a · bx , with a = 5 and c = 1, for these values of b: 0.3, 1, and 7.
3
3. Draw the graph of this logistic curve y = 1+e1−2·x
. Graph the function
for −2 ≤ x ≤ 5.

Logarithmic functions are closely related to exponential functions. An expo-


nential function of the form y = f (x) = bx is a “one-to-one” function in the
sense that for each value of variable x there is one and only one value of vari-
able y, and vice versa. Any function that exhibits a one-to-one relationship
also has associated with it an inverse function. We develop inverse functions
systematically later in this chapter. Logarithmic functions can be defined in
terms of exponential functions, as follows: Given x = by where b > 0 and
b 6= l, we refer to y = logb x as the logarithmic function of x to the base b.
This definition implies that the logarithm of a number is the exponent to
which a base must be raised in order to yield the original number. In general,
it is true that x = bb ⇔ logb x = y. To repeat, logarithm, base, and exponent
relate as follows: the logarithm of some variable x to the base b is the power
to which we must raise the base b in order to yield the value x.
EXAMPLES: The following common logarithms are used frequently.

1. 100 = 1 ⇔ log10 1 = 0

2. 101 = 10 ⇔ log10 10 = 1

3. 102 = 100 ⇔ log10 100 = 2

4. 103 = 1000 ⇔ log10 1000 = 3

5. 10−3 = 0.001 ⇔ log10 0.001 = −3

6. 10−2 = 0.01 ⇔ log10 0.01 = −2

7. 10−1 = 0.1 ⇔ log10 0.1 = −1

The allowable bases for logarithms are positive, so raising a positive number
to any power yields a positive number. This fact prohibits logarithms of
negative numbers. We can extract the logarithm of quite small positive
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 58

values like 0.0001, for which the base-ten logarithm is -3, as Example 5 above
shows.6
The logarithmic function permits any positive real number as the base b.
Even so, almost all analysis is conducted with b = 10 or b = e ≈ 2.718. Ten
is the base for a system called “common logarithms,” and e is the base for a
system called “natural logarithms” or “Napierian logarithms” (Napier is the
mathematician who developed this system).
A widely-used convention in science texts is to use the notation logx to refer
to common logarithms and lnx to refer to natural logarithms. In contrast
mathematics textbooks use logxto refer to natural logarithms and log10 x to
refer to refer to common logarithms. We use the mathematics convention
with one modification, the use of parentheses, log(x) rather than logx. This
modification accommodates the way the Maxima works: it can take loga-
rithms of expressions as well as numbers, so log is a command for which
the argument must be entered into parentheses. The result of the commands
log(100), log(100.0) and log(100.0)/log(10.0) appear below.

(%i) matrix(["log(100)","log(100.0)","log(10.0)",
"log(100.0)/log(10.0)"],[ log(100),log(100.0),
log(10.0),log(100.0)/log(10.0)]);
 
log(100) log(100.0) log(10.0) log(100.0)/log(10.0)
(%o)
log (100) 4.6051 2.3025 2.0

The command log(100) results in what appears to be the command itself.


Actually, Maxima has evaluated this value and returned its exact represen-
tation, which is kept in Maxima’s memory. Entering 100.0, rather than 100
causes Maxima to report a floating-point representation of the natural log-
arithm of 100.0. Finally, we see that log(100.0)/log(10.0) is the common
logarithm of 100.0.
If you require common logarithms, it is easy to define a function that provides
these values, as the next cell shows. We name the function that returns some
values of log10 (x). The float command ensures that the results are reported
as floating-point values. To show how this function works, we apply it to the
values in Example 1 above, all of which appear in xList.
6
Actually, Maxima does evaluate the logarithms of negative values, but the results are
imaginary numbers.
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 59

(%i) log10(x) := float(log(x)/log(10))$ matrix(


xList:[.001, .01, .1,1,10,100,1000], log10(xList));
 
0.001 0.01 0.1 1 10 100 1000
(%o)
−2.9999 −1.9999 −0.999 0.0 1.0 1.9999 2.9999

The cell below shows some important laws that govern the behavior of loga-
rithms. The command logexpand:super forces Maxima to expand some of
the expressions rather than just evaluating them and reporting them as they
were originally expressed. The last column applies for natural logarithms,
reflecting the general relationship that whenever x = by , then logb x = y.
(%i) logexpand:super$ exprList:[a*b,a/b,1/b,a^b,1/a,exp(a)]$
logList:log(exprList)$ matrix(exprList,logList);
a 1 1
 b a

ab b b
a a
e
(%o)
log (b) + log (a) log (a) − log (b) −log (b) log (a) b −log (a) a
The relationship between the graph of an exponential function such as y = bx
and its inverse function y = logb x can be illustrated graphically. The graph
below shows this symmetry, using the exponential function y = ex and its
inverse function y = log(x).

(%i) wxdraw2d(xaxis=true, yaxis=true, yrange=[-5,5],


user preamble="set key bottom",xlabel="x",ylabel="y",
key="y=x",explicit(x,x, -5, 5) , color = red,
key="y=e^x", explicit(exp(x), x, -5, 5),
line width=3, color=orange, key="y = log(x)",
explicit(log(x),x, - 5, 5))$
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 60

Figure 3.12: Exponential and Logarithmic Functions

Exercise Set 3.7

1. Express the following logarithmic functions in terms of exponents, and


the exponential functions in terms of logarithms. (a) 62 = 36
(b) 104 = 10000 (c) log10 (0.0001) = −4 (d) log2 (8) = 3
(e) 43 = 64 (f) log3 (27) = 3 (g) log2 (x) = 9 (h) 81/3 = 2
(i) 5−2 = 1/25 (j) log2 (64) = x

2. Determine the range of values of x for which the following functions are
defined. (a) y = log(x + 8) (b) y = 1og(8 − x) (c) y = log(x2 − 4)
(d) y = log(25 − x2 ) (e) y = log(x − 9) (f) y = log(x3 + 8)

3. Using the laws concerning the use of logarithms express each of the
following as a single logarithm: log(x) + log(y) + log(z).

4. Use the relationship log(ex ) = x = elog(x) , where log(x) is the natural


5 5
logarithm, to simplify the following. (a) elog(x ) (b) elog(5·x )
2 2 2 5
(c) log(ex ) (d)log( xe2 ) (e) e2·log(5·x) (f) e−log(x ) (g) log(4 · x3 ) · ex
2 5
(h) e3·log(x )+x

5. Pareto’s law of the distribution of incomes says that the fraction of


individuals N from a given population whose incomes exceed x dollars
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 61

is given by N = a/xb . Pareto suggested that the value of b was approx-


imately 1.50. If a = 40, 000, what is the number of individuals whose
incomes exceed x = $60, 000?

6. Given the general exponential form y = f (x) = a · bc·x , graph (and


compare) the shapes of the curves as follows:
(a) Select values for a and c such that a > 0, c > 0. Now graph the
function for four different b values: 0, 0.5,1, 1.5. Place all inside the
same draw-generated graph. Label each curve and select a different
color for each. Set the line width to 2. Note that b = 0 is not allowed
as the base for an exponential function. The graph will reveal why.
(b) Select values such that 0 < b < 1, and c > 0. Draw two curves:
for a < 0 and for a > 0. Place both inside the same draw-generated
graph. Label each curve and select a different color for each. Set the
line width to 2.
(c) Select values such that a > 0 and b > 0. Draw three curves: for
c < 0, for c = 0, and for c > 0. Follow the instructions in (a).
[Create this expression in wxMaxima: f(x,a,b,c) := a*b^(c*x). For
the graphs, let x range from -2 to 2 in all cases. Insert yrange=[-0.5,2]
and line width=2.The last two insertions produce the output in a rel-
atively bold relief. Label the x and y axes.]

Inverse Functions

Previous sections contain references a function that is the inverse of another


function. We now refine our notion of the character of an inverse function.
Consider a function y = f (x) that is strictly increasing or decreasing for all
values of x from some point a to another point b. A function increases (de-
creases) throughout this range of values is said to a monotonically increasing
(decreasing) function. A function f (x) is said to be monotonically increas-
ing (decreasing) if f (xj ) > f (xi ) (f (xj ) < f (xi )) whenever xj > xj . The
graph below shows two linear functions. Both are monotonic over the range
depicted. Indeed these function are monotonic over the entire real number
line.

(%i) [expr1, expr2]:[ 4*x - 5, 25 - 4*x]$


increase: gr2d(xaxis=true,yrange=[-10,30],xlabel="x",
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 62

ylabel="y", key=string(expr1),explicit(expr1,x,0,8))$
decrease: gr2d( xaxis=true,yrange=[-10,30],xlabel="x",
ylabel="y",key=string(expr2),explicit(expr2,x,0,8))$
wxdraw(increase,decrease,columns=2)$

Figure 3.13: Two Monotonic Functions

Each of the functions illustrated above is a “one-to-one mapping.” That is,


for every value of x in the interval (a, b), there is one and only one value of y in
the interval (c, d) such that y = f (x). Hence a change in the value of x always
yields a unique and different value of y, and vice versa. The mathematical
notation for this relationship is the following: x = f −1 (y) = g(y). This is
read, “x is an inverse function of y.” This new function, x = f −1 (y) = g(y)
is called an inverse function. Its domain is the interval (c, d), which is the
range of f (x). The range of g(y) is the interval (a, b), the domain of x for
f (x).
The cell below shows the graphical relationship between a function f and its
inverse function. We can obtain the graph of one from the graph of the other.
First, however, we have Maxima derive the inverse function and recognize
that this function’s domain is limited to the positive real numbers. We define
the expression that constitutes this function and assign it the name expr.
Then we solve y = f (x) for x. The solution involves log(y) and, therefore,
the inverse function’s range (which must be the original function’s domain)
is limited to the positive real numbers.
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 63

(%i) [expr: y=125*(2^x), soln: solve( expr,x)[1]];


y
log( 125 )
(%o) [y = 125 2x , x = log(2) ]

Before we graph the function and its inverse, we determine the range of the
inverse function. Doing so helps guide the graphing of the two functions.

(%i) [y0, y1]: [subst(x=0,expr), subst(x=5, expr)];


(%o) [y = 125, y = 4000]

Figure 3.14 shows that the two functions, except for scale, are mirror images
of each other. We have added reference lines that take into account the
difference in scale for the two variables. In the first, the reference line is
y = 4000 · x, and in second, the line is the equivalent, x = y/4000. We have
extended the axes slightly from the values shown above in order to emphasize
the similarity of these two functions.

(%i) initial: gr2d(user preamble="set key left",


xlabel="x", ylabel="y", ytics=1000, key = "f(x)",
explicit(rhs(expr),x,0,6) , color=black,line width=1,
key="",explicit(1000*x,x,0,6) )$
inverse: gr2d(user preamble="set key left",
xlabel="y", xtics=4000, yrange=[0,6], ylabel="x",
key="g(y)",implicit(soln, y,rhs(y0),8000,x,0,6),
color=black,line width=1,key="",
explicit(y/1000,y,0,rhs(y1)*1.6) )$
wxdraw(initial,inverse, columns=2)$

For emphasis, we repeat that f −1 is not the same as (f )−1 , which is the
inverse of the expression on the right-hand side of the initial function. The
difference between the two is shown below, as created with Maxima’s print
command.
y

1 1 log 125
The inverse of f (x) = = . The inverse function is .
y 125 2x log (2)
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 64

Figure 3.14: A Function and Its Inverse

3.3 Summation and Multiplication


In many business and economic problems, one is required to sum or multiply
a large number of terms (numbers, variables, parameters, coefficients, and
so forth). A shorthand notation that makes use of the uppercase Greek
letter sigma, Σ, enables us to write lengthy sums in a more compact form.
The uppercase Greek letter pi, Π, serves the same role for multiplication.
The statements below were created in Maxima, using the print, sum, and
product commands. Note that Maxima, like all software, has its conventions
about output. In this case, it reports sums and products in reverse order.

(%i) print("Interpret", ’sum(a[i],i,1,10), " as " ,


sum(a[i],i,1,10))$
print("Interpret ",’product(a[i],i,1,10), " as ",
product(a[i],i,1,10))$
P10
Interpret i=1 ai as a10 +a9 +a8 +a7 +a6 +a5 +a4 +a3 +a2 +a1

Q10
Interpret i=1 ai as a1 a2 a3 a4 a5 a6 a7 a8 a9 a10

The first table below shows the results of the following instructions (see the
workbook for details):
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 65

• Use makelist to create a list of the numbers 1 through 10.

• Make a second list of the squared values of these numbers.

• Sum the second through sixth values in the list of squared values.

• Sum all ten squared values.

• Sum k times each of the ten squared values. The illustrates the homo-
geneity property of summation.

• Create a list of first differences, beginning with the second squared


value.

• Sum the list of nine first differences, confirming that this sum equals
the last value less the first value in the original list of squared values.
This result is an example of the telescoping property of addition. For
any list [a1 , a2 , · · · , an ], it must be true that (a2 − a1 ) + (a3 − a2 ) + · · · −
(an − an−1) = an − a1 .

The next table shows the effects of the same sets of commands, except that
product replaces sum, and that ratios of adjacent values replace differences
between adjacent values.
Observe that the homogeneity property of multiplication implies that multi-
plying each of a set of values by k and then multiplying the result yields a
result that is k n as large as the products of the initial values.
Observe also that a variant of the telescoping property also applies. The
product of the ratios created as in the table below equals the terminal value
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 66

in the list of original values. Why this must be so is easily determined by


reference to the second-to-last line: striking out values from left to right
leaves only the numerator of the last item, which is the final value of the
original list.

Exercise Set 2.8

1. For each of the following functions, sketch the graphs of f and f −1 and
determine the range and domain for both the √ function and its inverse.
(a) y = 5 + 4 · x (b) y = 3 − 2 · x (c) y = 2 · x + 8
x
(d) y = x+1 (e) y = x (f) y = 2 · x − 3

2. Evaluate
P44 each of the P
following sums, bothPby hand and using Maxima.
(a) j=40 j (b) 4k=1 2k−1 (c) 5i=0 (−1)i
P4 k−1 P8
(d) k=1 k+1 (e) i=1 (3 · i + 5)

3. Rewrite each of the following in summation notation.


(a) 1 + x + x2 + x3 + x4 (b) 1 + 2 + 4 + 8 + 16 + 32
a a2 a3 an
(c) 1 + x + x2 + x3 + · · · + xn

4. Evaluate each of the following products, both by hand and using Max-
ima. Q
(a) 42 (b) 5j=2 xj (c) 5i=5 i2
Q Q
j=40 j
(d) 4i=1 (3 · i + a)
Q
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 67

3.4 Questions and Problems


1. Solve the equation x2 + 5 · x + 6 = 0 for x using the quadratic formula.

2. Refer to the equation y = 4 · x − 5.


(a) Find the values taken on by the dependent variable y when inde-
pendent variable x takes on integer values between 1 and 3. Do this
twice, first by hand and then with M axima. Create a named list of
x values and then create a functional expression for y and map that
function onto the list.
(b) Graph the function y = 4 · x–5, either by hand or using Maxima.

3. Graph the equation y = x2 − 2 · x + 1. What kind of functional form is


this? What is the range of this function? What is its domain?

4. Given the following price and quantity demanded data for pizza.

Quantity per day 50 60 70 80 90


Price, $ per pizza 15 14 13 12 11

(a) Find the equation of the demand function for pizza such that Q =
f (P ).
(b) What are the slope and the ordinate (intercept) of the demand
function? Beware: Q is the “y” variable given the way the function is
expressed.

5. The gross domestic product (GDP) of Nuevo Laredo, a relatively less


developed country, was valued at 100 billion pesetas in the year 2015.
The leaders of the country feel that a realistic growth target is for GDP
to grow to 220 within seven years (the year 2022). Consider GDP as
a function of time t. Which of the following functional relationships
most accurately portrays a growth path that is consistent with the
leaders’ belief? (a) GDP = 100 + 2 · t; (b) GN P = 100 + 20 · t; or
(c) GN P = 100 · e0.1125·t . Use wxdraw to graph the three expressions.
Let t = 0 for the year 2015. Specification (c) is consistent with a
compounded annual growth rate of 11.25%. Given your knowledge of
economic growth rates (or on the results of a quick Google search), do
you think any of these growth paths is likely to occur?
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 68

6. The demand function for carpenters in Portland is given by Q =


270, 000 · W −2 , where Q = quantity of carpenters employed, measured
in worker hours per time period, and W = wage rate of carpenters,
measured in dollars per hour. How many carpenter hours will result
if W = $30 per hour? What kind of functional form is this demand
function? (FYI, -2 is the elasticity of demand for labor with respect to
W .)

7. Profit maximization requires that a firm equate marginal cost (MC )


with marginal revenue (MR). The marginal cost curve for the produc-
tion of mathematical economics textbooks is given by MC = 8+0.01·Q,
where Q is the number of textbooks produced. The marginal revenue
realized from the sale of textbooks is given by MR = 104 − 0.05 · Q.
(a) Find the profit-maximizing output and sales of the firm. (b) If the
demand curve is Q = 2080 − 10 · P , where P is the product price, what
price must this firm charge to sell the quantity that you determined?
(c) Determine the inverse demand curve P = g(Q) and use Maxima to
graph this function along with MC and MR.

8. The Numerical Control Company (NCC) produces control units that


attach to machine tools. These tools automatically control a tool’s
operation more precisely than is possible with manual control. NCC
has developed a new model, the DX3. NCC knows the identity of firms
to which it sells and it is confident that eventually 10,000 units will be
purchased. Cost estimates indicate that if 7,000 are adopted within five
years, the DX3 will be profitable. Adoption follows the path defined
by the logistic function p(t) = k/(1 + e−0.2·t ). Here p(t) is the percent
of maximum adoptions that will have occurred by time t, k = 100 is
the maximum percent, and t is the number of years.
(a) If NCC is right about the maximum value, will this project be
profitable?
(b) In approximately how many years will p(t) = 99 percent?

9. Given the production function Q = f (L, K), where Q = output of


the firm in units, and L and K are the number of units of labor and
capital, respectively, that the firm chooses to√hire. The
√ specific form of
the production function of the firm is Q = 3· L+5· K. The marginal
production of labor and the marginal product of√capital are respectively
equal to M P L = 1.5/sqrtL and M P K = 2.5/ K. (Chapter 4 shows
CHAPTER 3. RECTANGULAR COORDINATES AND FUNCTIONS 69

why.) The wage that must be paid to a unit of labor is symbolized by


w and is equal to $1 per unit. The price of capital per unit, r, is $2.
The input prices and the product price are beyond the firm’s control,
and they are not affected by the firm’s decisions regarding employment
or output levels. Thus, the firm treats these values as constants.
The price of the output is symbolized by P and is $4 per unit. Profit
maximization requires that the firm choose each of the input levels so
that the input price equals the product price × the marginal product
of that input. The price of this price-taking firm’s product times the
marginal product of an input is that input’s value marginal product
(VMP). The firm must satisfy w = V M P L and r = V M P K, where
V M P L = P · M P L and V M P K = P · M P K. Determine:
(a) the profit-maximizing magnitudes of L, K, and Q, and
(b) How much profit will the firm will earn if it selects the values
determined in (a).

10. Suppose that the average annual earnings for members of a profession
is represented by this function log(y) = f (x) = 12.5 + 0.07 · x − .001 · x2 ,
where y is average earnings and x is years of experience.
(a) Graph this function, with y (not log(y) on the vertical axis.
(b) Assume that the average age at which members of this profession
begin to practice is 25. What factor(s) might account for the downturn
in earnings at x = 35?
Chapter 4

Limits, Continuity, and


Differentiability

We now explore calculus-based analysis. The calculus consists of two parts,


differential calculus and integral calculus. Differential calculus, as the name
suggests relates to examining differences. Its concern is with how y changes
in response to a change in x. Integral calculus, as its name might suggest,
involves integrating—putting parts back together. If we have a function that
defines y’s changes in terms of x’s changes, then integration can help use to
move from the changes (differences) into the value of x (level) at a given value
of x. The operations of differentiation and integration are the inverse of each
other, just as addition and subtraction, and multiplication and division, are
the inverse of each other.
We use the methods of differential calculus when we know the form of the
function that states the level of a variable in terms of the level of one or
more other variables. Often, the important question is how much a change
(changes) in the independent variable(s) will affect the dependent variable’s
value. This analysis involves taking derivatives. For equations with one
independent variable, the derivative is actually the slope of the tangent to
the graph of that function, as we see later in this chapter.1
1
In at least one inportant respect, this representation of the use of differential calculus
is incomplete. We often illustrate economic models that involve specific expressions. The
comments here apply to those cases. In addition, however, differential calculus allows us
to develop general principles, so that insights can be applied even when the functional
form is not known.

70
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 71

We use the methods of integral calculus when we know the form of the
function that describes changes and we wish to know the form of the function
that describes the level of the variable in question. An important use of
integral calculus is related to the fact that we can use an integral to find the
area under a particular curve or function.
This chapter and the next three address the differential calculus and its
applications. Chapter 8 introduces integral calculus. Before we learn how
to take derivatives and to use them, we develop the concepts of limits and
continuity. The concept of limits is critical to understanding how to interpret
derivatives. Whether or not a function is continuous determines whether
differential calculus methods can be applied.

4.1 Limits
We have already seen one example of a limit. When we examined the logistic
function, we saw that a population grows toward a limiting value. Thinking
of that example gives us a basis for a formal definition of the term. Begin
with a function y = f (x). If, as x approaches some value, x0 , and as a result
f (x) approaches some number A, then A is said to be the limit of f (x) as x
approaches x0 . The standard notation for the preceding statement is this:

lim f (x) = A.
x→x0

The remainder of this section considers a series of examples. These examples


illustrate the concept of the limit. They also introduces some important
aspects of limits, aspects that we will encounter in subsequent analysis.
Example: Consider y = x2 , with x0 = 3. The table below shows values of
x that approach 3 from below and then values that approach 3 from above.
These lists were constructed by subtracting values in diffList from 3 and
by adding those value to 3, respectively. Then corresponding lists of y values
were constructed by inserting smallList and largeList into f (x). For ease
of reading, all values are placed into matrices.

(%i) f(x):=x^2$
diffList: [3, 2, 1, 0.1, 0.01, 0.001, 0.001, 0.0001]$
smallList: 3 - diffList$ largeList: 3 + diffList$
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 72

belowList: f(3-diffList)$ aboveList: f(3 + diffList)$


matrix( cons("x",smallList), cons("y", belowList));
matrix( cons("x",largeList), cons("y", aboveList));
 
x 0 1 2 2.9 2.99 2.999 2.999 2.9999
(%o)
y 0 1 4 8.41 8.9401 8.994 8.994 8.9994
 
x 6 5 4 3.1 3.01 3.001 3.001 3.0001
(%o)
y 36 25 16 9.61 9.06 9.006 9.006 9.0006

From either direction, as x approaches 3, f (x) approaches 9. Maxima’s limit


command can be used to confirm this result:

(%i) limit(f(x),x,3); (%o) 9

Example. As a second example, consider a series for which the important


limit is positive infinity. Suppose that a value grows according to the function
y = (1 + 1/n)n . This value doubles in one period if n = 1. That is the
growth rate is 100% per period. More generally, n is the number of times
that the growth process compounds per period. The question we address is
the limiting value of the expression as n becomes very large (and the term
r/n becomes very small). That is, we look at what happens when the number
of times growth is compounded each period becomes larger. Assuming that
the period is one year, the cell below shows compounding annually (n = 1),
semiannually (n = 2), and so forth up to hourly (n = 8765.81. The results
show that by the time that compounding occurs daily (n = 365 for most
years), the value is close to e = 2.718 . . . .2 The final entry for n, 8765.81,
indicates hourly compounding.
In this cell, the first output line shows values of n. The second line shows
the size of the function at year’s end, given n. The third line shows the value
that is function approaches as n approaches infinity.

(%i) g(n) := (1 + 1/n)^n$


nList: float([1,2, 4,12,52,365,8765.81 ]);
yList: map(g,nList);
[limit(g(n),n,inf), float(limit(g(n),n,inf)) ];
2
Here, e is Euler’s number, which we encountered in Chapter 3.
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 73

(%o) [1.0, 2.0, 4.0, 12.0, 52.0, 365.0, 8765.8]


(%o) [2.0, 2.25, 2.4414, 2.613, 2.6925, 2.7145, 2.7181]
(%o) [e, 2.7182]

The last line above shows the result of applying the float command with
n approaching infinity. The first result shows the exact value; the second
shows its floating-point representation.
Example. Consider a generalization of the example above. Suppose that a
sum of money accumulates for t years at an annual rate of r. The sum plus
interest is compounded n times per year. The next cell shows the formula for
the value after t years and it shows the limit of that formula as the number
of compounding periods per year approaches infinity. This result shows that
Maxima’s limit command is not restricted to finding numerical values. It
can also find limits of expressions, where the limit is another expression.

(%i) [ V: S*(1+r/n)^(n*t), limit(V,n,inf)];


n t
(%o) [ nr + 1 S, er t S]

Example. Next consider the expression function y = (x2 + 1)/(x–1). This


function’s domain is limited, in that y is undefined for x = 1. It might be
useful, however, to determine y’s value as x approaches 1 from above and
from below.3 We show the results of limit when the limit is approached for
above (plus in the command below) and from below (minus in the second
command line below. The two commands are in a list. The graph shows
the behavior of this function as x approaches 1 in each direction. Note the
limited yrange values. As an exercise, execute the commands in this cell
after having removed the yrange option.

(%i) [limit((x^2+1)/(x-1),x,1,plus),
limit((x^2+1)/(x-1),x,1,minus)];
wxdraw2d(yrange=[-100,100],xlabel="x", ylabel="y",
key="(x^2 + 1)/(x - 1)",
explicit( (x^2+1)/(x-1), x, 0, 3 ) )$
(%o) [∞, −∞]
3
We leave as an exercise the determination of y values for x values that are ever closer
to 1. Follow the development of the first example.
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 74

Figure 4.1: No Unique Limit

Example. Another case in which a limit’s value depends on the direction


of approach is a step function. Consider y = |x|/x. The limit of z as x
approaches 0 is either 1 or -1, depending on the direction, as the output
shows and as Figure 4.2 illustrates. This example offers in important lesson:
The limit is not necessarily the same as the value at x0 . For x = x0 , the
function value y is not defined.

(%i) [z :abs(x)/x, limit(z,x,0,plus), limit(z,x,0,minus)];


(yrange=[-2,2], xlabel="x",ylabel = "|x|/x",
explicit(z,x, -2,2) )$
(%o) [ |x|
x
, 1, − 1]

The definition of a limit would be quite difficult to apply in many situations.


Fortunately, a number of theorems that involve combinations of relatively
simple functions remove the requirement that we derive the limit from def-
inition in each application. Using those theorems can greatly simplify the
evaluation of limits for seemingly difficult functions. Furthermore, Maxima
can directly apply these theorems and supply the limits to a large array of
functions. Suppose that a relationship between y and x can be decomposed
into two functions, f (x) and g(x). Furthermore, suppose that the following
are the limits for these functions as x → x0 : f (x) → A and g(x) → B. Then
the following are true:
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 75

Figure 4.2: A Step Function

• If the expression is f (x) + g(x), then its limit is A + B.


• If the expression is f (x) − g(x), then its limit is A − B.
• If the expression is f (x) · g(x), then its limit is A · B.
• If the expression is f (x)/g(x), then its limit is A/B.
• If the expression is K · f (x) where K is a constant, then its limit is
A + B.

The next display shows two functions of x, It evaluates each function as x


approaches 100.0. It then evaluates the limits of sum, difference, product,
and ratio of the two functions as x approaches 100.0. Finally, it evaluates the
product of the two functions and the constant K. The function f (x) returns
the common (base 10) logarithm for 100.0.4 The function g(x) is, of course,
a quadratic function of x.

(%i) [f(x):=log(x)/log(10.0), g(x):=0.05*x^2-2*x+20];


nameList: ["A", "B", "sum", "difference", "product",
4
The common logarithm of 100.0 is 2.0; the value 1.9999 results from rounding error
in the float command. Maxima retains the exact value in its memory and applies that
value to any calculations that involve f (100.0), such as the five limits that are reported in
the table.
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 76

"ratio","times A"]$
limitsList: [limit(f(x),x,100.0),limit(g(x),x,100.0)
limit(f(x)+g(x),x,100.0),
limit(f(x)-g(x),x,100.0),limit(f(x)*g(x),x,100.0),
limit(f(x)/g(x),x,100.0),
limit(K*f(x)*g(x),x,100.0)]$
matrix(nameList, limitsList);
 
A B sum difference product ratio times A
(%o)
1.9999 320.0 322.0 −318.0 640.0 0.00625 640.0 K

EXERCISE 3.1
In each of Exercises 1 through 15, evaluate the limits. Try to determine the
limits yourself before having Maxima provide the value. To keep the exercises
compact we use limx→A f (x) rather than the equivalent

lim f (x).
x→A

1. limx→2 2 · (x − 1) 2. limx→4 (x2 + 4 · x)


x2 −8
3. limx→−1 (3 · x3 − 2 · x2 + 10) 4. limx→0 ( x2 +2·x+1 )
2
5. limx→2 ( xx2 −4
+4
)
6. limx→0 ((x − 4) · (x3 + 4 · x − 3) · (4 · x + 7)
2
2−3·x3
7. limx→3 (x2 − 5)2 8. limx→−1 x2·x+1
9. limx→0 ( x+4
x−4
) 10. limx→1 10−x
2
11. limx→0 8 x
12. limx→1 (2 − x4 )
3 2 +c·x−24)
13.[check] limx→3 (x −3·x(x−6)
14.[check] limx→A f (x)
2 +2·ex2
15. limx→0 x ex3

4.2 Extensions of the Limit Concept


The preceding section shows how a limit can be evaluated. The examples
show three especially important cases: one is which the limit of the function
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 77

is undefined (approaches either infinity or negative infinity) at some value


of x; one at which the limit at a specified value of x differs, depending on
whether the x value is approached from above or from below, and one in
which the limit is found when x itself approaches either positive or negative
infinity. The implication of the first case, where the limit approaches positive
or negative infinity, is direct: the function of x is not defined at that value.
This section addresses the second.

4.2.1 Right-hand and Left-hand Limits


The definition of a limit states that as variable x approaches some finite
number, the value of function f(x) approaches some finite number A. That
is, limx→0 f (x) = A. The limit of a function from above may be the same as
the limit taken from below. As have already seen, however, that need not be
the case. Thus the left-hand limit may be one value, A, and the right-hand
limit may be another value, B.
Only when both the left-hand limit and the right-hand limit exist, and they
are equal to each other are we able to state that a (single) limit of a function
exists. Therefore the function f (x) = |x|/x does not have a uniquely defined
limit at x = 0. Therefore, we must ensure that both limits are defined and
that they are equal. Fortunately, Maxima checks for this. The next exhibit
shows the results of a command to find the limit of the function f (x)|x|/x
that does not specify direction and two commands that do specify direction.
Without this specification, Maxima returns “und” which indicates that the
value of the limit is undefined.

(%i) [limit(abs(x)/x,x,0), limit( abs(x)/x,x,0,plus),


limit( abs(x)/x, x, 0, minus) ];
(%o) [und, 1, −1]

4.2.2 Infinite Limits


Our definition of a limit stipulates that both x and f (x) approach finite nu-
merical values (x0 and A, respectively) in the limit. It is possible, nonetheless,
that either we wish to consider cases in which x0 becomes either arbitrarily
small or arbitrarily large. In such cases, f (x) can either approach a finite
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 78

constant, or it can approach a value A that has no limits. First, consider


the simple function f (x) = A/x. Its infinite limits are defined in the next
exhibit.

(%i) f(x,A):= A/x;


[limit(f(x,A),x,inf),limit(f(x,A),x,minf),
limit(f(x,A),x,0,plus), limit(f(x,A),x,0,minus)];
[limit(f(x,2),x,inf),limit(f(x,2),x,minf),
limit(f(x,2),x,0,plus), limit(f(x,2),x,0,minus)];
[limit(f(x,-2),x,inf),limit(f(x,-2),x,minf),
limit(f(x,-2),x,0,plus), limit(f(x,-2),x,0,minus)];

(%o) f (x, B) := A
x
(%o) [0, 0, infinity, infinity]
(%o) [0, 0, ∞, −∞] (%o) [0, 0, −∞, ∞]

This exhibit shows how the function is defined. The commands specify three
sets of limits. In the first set A’s sign is not specified; in the second, A > 0,
and in the third A < 0. In each case we seek limits as x becomes a very
large positive number (inf) and a very large negative number (minf). Also,
we seek limits as x approaches zero from above and from below. In all cases,
1/x becomes quite small (approaches 0) as x becomes either very large or
very small. Thus the limit of 1/x is zero as x approaches either infinity or
negative infinity.
The behavior as x approaches zero depends on the sign of A. If that sign
is not specified, Maxima provides the ambiguous response infinity which
does not specify a sign. The second and third output lists show the source of
this ambiguity: whether the limiting value of 1/x is positive or negative for
a specified limiting value of x depends on A’s sign. Figure 4.3 adds insight
into the behavior of this function.

(%i) positiveA: gr2d(xaxis=true, yrange=[-10,10],


xlabel="x",ylabel="1/x", key="B = 2",
explicit(f(x,2),x,-20,20))$
negativeA: gr2d( xaxis=true, yrange=[-10,10],
xlabel="x", ylabel="1/x", key="B = -2",
explicit(f(x,-2),x,-20,20))$
wxdraw(positiveA, negativeA);
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 79

Figure 4.3: Graphs of B/x

In all monotonic functions, as x becomes either quite large or quite small,


f (x) can also grow without limit, either toward
√ infinity or toward negative
infinity. Consider the behavior of f (x) = ± x and g(x) = ±log(x). For
both, negative values of x are not in the domain, but x can grow without
limit. As x grows without limit, so do the values of these two functions, as
the next exhibit shows.

(%i) fList: [sqrt(x), -sqrt(x), log(x), -log(x)];


limit(fList,x,inf);
√ √
(%o) [ x, − x, log (x) , −log (x)] (%o)[∞, −∞, ∞, −∞]

A problem can arise if a function consists of a ratio of two expressions, both of


which grow without limit. In many cases, the problem appears more imposing
than it is in reality. Consider f (x) = (4 · x2 + 2 · x − 3)/(x3 − 6 · x + 2).
Both the numerator and the denominator grow without limit as x → inf,
so the behavior of f (x) is not obvious. One can see, however, that the
highest-powered term is in the denominator. Dividing both denominator and
numerator by x3 yields an expression that has a limiting value of 0/1 = 0. A
legitimate question is whether Maxima recognizes this relationship. As the
next exhibit shows, it does.

(%i) limit( (4*x^2+2*x-3)/(x^3-6*x +2),x,inf); (%o) 0


CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 80

Maxima can handle much more complex relationships. Let f (x) = x/ log(x).
Both the numerator and the denominator grow without limit, and no simple
factoring can separate the growth of the two. The next exhibit shows some
values of this function’s components and of the function. These suggest that
f (x) grows without limit. Maxima is able to evaluate this limit and confirms
what the listed values suggest. (The first item in each list identifies the values
that are displayed.)

(%i) xList: ["x", 2.0, 10.0, 100.0, 1000.0];


sqrtxList: sqrt(xList); logxList: log(xList);
ratioList:sqrtxList/logxList;
limit(sqrt(x)/log(x),x, inf);
(%o) [x,
√ 2.0, 10.0, 100.0, 1000.0]
(%o) [ x, 1.4142, 3.1622, 10.0, 31.622]
(%o) [log√
(x) , 0.693, 2.3025, 4.6051, 6.9077]
x
(%o) [ log(x) , 2.0402, 1.3733, 2.1714, 4.5778]
(%o) ∞

4.2.3 An Economic Application: The Cobb-Douglas


Function and the Constant-Elasticity of Substi-
tution Function
.
We saw one economic implication of a limit above when we established that
compound growth can, in the limit, be represented as an exponential function
of time. The limit can also be used to show the relationship between two
functions that are frequently used to illustrate economic principles and as the
basis for empirical work.5 The two are the Cobb-Douglas function xα ∗ y 1−α ,
and the Constant-Elasticity of Substitution (CES) function ( xab + 1−a yb
)−1/b .
For the CES function the elasticity of substitution between the two inputs x
and y is s = 1/(b + 1). For the Cobb-Douglas function s = 1. If s = 1 in the
CES function, however, b = 0, and the CES function is undefined. Even so,
we can evaluate the CES function as b approaches 0. Doing so reveals the
relationship between these two functions.6
5
This relatively advanced material can be omitted without loss of continuity.
6
The tlimit command is used instead of limit. This command creates a Taylor series
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 81

(%i) [CES:(a/x^b + (1-a)/y^b)^(-1/b), tlimit(CES,b,0)];

1
(%o) [  1 , xa y1−a ]
1−a b
+ ab
yb x

The first output entry shows the CES function. The second entry shows the
limit of this function as b → 0. The result is the Cobb-Douglas function.

4.3 Continunity
The first step in determining and interpreting the derivative of a function
is to have a clear understanding of the concept of a limit. The second step
involves using concept of continuity of a function. Very roughly speaking, a
continuous function on a particular interval is one whose graph can be drawn
without lifting one’s pencil or pen from the paper in that interval. Figure
4.4 below shows two continuous functions and two discontinuous functions.7

(%i) [f1:50 + 10*x - 0.1*x^2, f2:150+2*abs(50-x),


f3: if x <=50 then 20+3*x else 50+ x,
f4: x/(x-40)]$
continuous:gr2d(title="Continuous",explicit(f1,x,0,100),
color=black,explicit(f2, x, 0, 100) )$
discontinuous: gr2d(title="Discontinuous",
yrange=[-100,200], explicit(f3, x, 0, 50),
explicit(f3,x,50.001,100),
color=black,explicit(f4,x,0,100) )$
wxdraw(continuous, discontinuous, columns=2)$

Despite the kink at x = 50, the v-shaped function in the first panel is contin-
uous. Both functions in the second panel exhibit discontinuities. Following a
representation of the function and then evaluates the limit of that series. We use this
because on some installations of Maxima, applying the limit command to this exprssion
resulted in an overflow error.
7
The second of the two explicit commands used to generate the first discontinuous
graph contains a “fudge” term in that the second segment begins at 50.001 and not 50.
The reason for this is that draw would connect the points if 50 were both the endpoint of
the first segment and the beginning point of the second segment.
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 82

Figure 4.4: Continuous and Discontinuous Functions

more formal definition of continuity, we catalog the types of discontinuities,


one at x = 40 and the other at x = 50. Look at the expressions for these two
functions and determine why the discontinuities exist.
The following definition applies: A function y = f (x) is said to be continuous
at x = x0 if the following three requirements are fulfilled: (1) f (x0 ) exists
such that point x0 is in the domain of the function, (2) limx→x0 f (x) exists,
and (3) limx→x0 f (x) = f (x0 ). Failure to meet these conditions can fall into
one of four categories: removable, jump, infinite, and essential.8

4.3.1 Removable discontinuity


A removable discontinuity typically occurs when the domain of the indepen-
dent variable does not include a point (or points). Such discontinuities can,
in principle, be removed by reference to the limit of the expression. Consider
 x
1
y= .
1 + 1/x2
8
This category list is used in the MIT Open Course “Single Variable Calculus”:
http://ocw.mit.edu/courses/mathematics/18-01sc-single-variable-calculus-fall-
2010/index.htm. Actually the name applied to what we call essential is “other
(ugly).”
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 83

The domain of x does not include x = 0. Accordingly, when we attempt to


evaluate y(0), the result is an error message.
(%i) f(x):= (1/ (1 + 1/x^2))^x$ f(0);
(%o) expt: undefined: 0 to a negative exponent.
#0: f(x=0) – an error.
To debug this try: debugmode(true);
The next cell shows that y = f (x) does have a defined limit. Whether we
approach x = 0 from above or below, y = 1. We can add this point to the
graph to complete the represent of f (x) over the selected range of x values.
In this case, the limit is “on the line,” but this need not be the case. The
endpoints on the two sections of the curve in Figure 4.5 are chosen so as to
represent the function but to leave space for the point.
(%i) y0: limit(f(x),x,0);
wxdraw2d(explicit(f(x),x,-2,-0.005),
explicit(f(x),x,0.005,2), point type=circle,
points( [ [0,y0] ]) )$
(%o) 1

Figure 4.5: A Removable Discontinuity

The next exhibit shows the new, augmented function for which the domain
spans the real number line. Selected values of this function, named faug(x),
appear as the output. The discontinuity has been removed.
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 84

(%i) faug(x) := if x = 0 then 1 else f(x)$


[faug(-1/2),
√ faug(0),faug(1), faug(2)];
1
(%o) [ 5, 1, 2 , 16
25
]

An example much like the one above is the one that applies to periodic
compounding: We have already evaluated the limit of this function when
n → ∞, finding that the result is y0 · er·t . Use Maxima to confirm that the
limit as n → 0 is y0. Also confirm that the function is not defined at n = 0
(of course, n < 0 makes no economic sense). The fact that n = 0 is not part
of the domain actually makes economic sense: if n = 0, compounding never
occurs.

4.3.2 Jump discontinuity


A jump discontinuity occurs when the right-hand and left-hand limits exist
but the two are not equal. One such discontinuity appears in Figure 4.4.
The function f (x) = |x|/x, which we considered when examining limits, is
another case.

Infinite discontinuity

The next exhibit shows two functions that have infinite discontinuities. As
x approaches a critical value (0 here) the functions’ values become either
very large or very small. The limits themselves may differ or they may be
the same (that is ∞ or −∞). Do not conclude that the fact that the limit
of 1/x2 is the same from either direction implies continuity. Of the three
conditions that must be met for a function to be continuous 1/x2 fails the
first two: 0 is not in the domain, and the limit is no finite value.

(%i) kill(f,g)$ f(x):=1/x$ g(x):=1/x^2$


["f(x) limits:",limit(f(x),x,0),limit(f(x),x,0,plus),
limit(f(x),x,0,minus)];["g(x) limit:",limit(g(x),x,0)];
different:gr2d(yrange=[-100,100],xaxis=true,
yaxis=true,title="f(x)=1/x",explicit(f(x),x,-1,1))$
same:gr2d(yrange=[0,100],xaxis=true,yaxis=true,
title="g(x)=1/x^2", explicit(g(x),x, -1,1))$
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 85

Figure 4.6: Infinite Discontinuities

wxdraw(different,same, columns=2)$
(%o) [f (x)limits :, infinity, ∞, −∞] (%o) [g(x)limit :, ∞]

The first output line above reminds us that Maxima cannot determine the
limit of 1/x as x → 0 unless the direction is provided. The second output
line shows the same is not true of 1/x2 : g(x) → ∞ regardless of the direction
from which zero is approached.

Other (Ugly) Discontinuity

Some functions have no limits as they approach a critical value. A commonly-


used representative for these is limx→x0 sin(1/x). As Figure 4.7 shows, this
function fluctuates increasing wildly between -1 and 1 as x → 0.
A large number of functions exhibit this behavior. Fortunately, many of the
ones that do are functions that depict oscillations of physical systems as time
goes to infinity, and are not likely to apply to phenomena typically analyzed
with economic models. Even so, the possibility of such behavior should be
kept in mind.

(%i) wxdraw2d(yrange=[-1,1.5], xtics=-.1,0,.1,


key="Approaching 0 from below",
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 86

explicit(sin(1/x),x,-.1,0),color=black,
key="Approaching 0 from above",
explicit(sin(1/x),x,0,.1) )$

Figure 4.7: An ”ugly” discontinuity


CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 87

Exercise 3-3
Determine the values of x for which the following functions are continuous.
If a discontinuity exists, determine the type of the discontinuity, and correct
any removable discontinuities.

1. f (x) = 3 · x3 + 2 · x2 + x + 1 2. f (x) = x/(x + 1)


3. f (x = (x3 − 27)/(x2 − 9) 4. f (x) = (x2 − 3 · x + 2)/(x − 2)
(x2 + x − 2)/(x − 1)2
5. f (x) = √ 6. f (x) = (x2 − 4 · x − 21)/(x − 7)
7. f (x) = 4 − x2 8. f (x) =
2
9. f (x) = xx2 +5·x+6
+4·x+4
10. f (x) = 8/(x − 4)
11. f (x) = 5/(1 − 2x ) 12. f (x) = ex
1
13. f (x) = x·(x−4) 14. f (x) = (x − 6)/6
x+4
15. f (x) = x2 +2·x−8 16. f (x) = 1/(x2 + 1)
(x = 4)/(3 · x2 − 27)
17. f (x) = √ 18. f (x) = (x2 − 16)/(x + 4)
x
19. f (x) = x 20. f (x) = 1/(2e − 2)

4.4 The Derivative of a Function


The concepts of a limit and of a continuous function constitute the foundation
for our study of the derivative of a function. Consider the function y = f (x).
We now focus on the effect that an incremental change in x–which we denote
as ∆x– has on y. The magnitude of any change in y, denoted ∆y, that does
occur depends not only on the magnitude of ∆x, but also on the specific
form of f (x).
Consider first the simple linear form. In Chapter 2, a linear function was
graphed as a line with a constant slope. We found that the slope of a straight
line that passed through any two points (x1 , yl ) and (x2 , y2 ), or in functional
notation [x1 , f (x1 )] and [x2 , f (x2 )], was given by the quotient

∆y y2 − y1 f (x2 ) − f (x1 )
m= = = .
∆x x2 − x1 x2 − x1

This definition of the slope m is hereafter referred to as the difference quo-


tient. The exhibit below, for y = 10 + 2 · x, reports an initial value of x,
x1 = 5, the function, and the initial value of y, y1 =. Then it produces a
table, the first row of which is a list of increasingly small ∆x values. The
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 88

second row shows the implied x value; the third row shows the implied y
values. The last two rows show the implied changes in y, ∆y, and the quo-
tients, ∆y/∆x. For this linear function all values on the last row equal the
coefficient of x.

(%i) x1: 5, f(x):=10 + 2*x, y1:f(x1)]$


DeltaxList: [2.00, 1.00, 0.50, 0.25, 0.10, 0.01]$
xList: x1 + DeltaxList$ yList: f(xList)$
DeltayList:yList-y0$ dydxList:DeltayList/DeltaxList$
matrix(cons("x changes",DeltaxList),
cons("x values",xList), cons("y values ", yList),
cons("ychanges", DeltayList),
cons("difference quotients",dydxList) ); 
x changes 2.0 1.0 0.5 0.25 0.1 0.01

 x values 7.0 6.0 5.5 5.25 5.1 5.01 
(%o)  y values 24.0 22.0 21.0 20.5 20.2 20.02 

 y changes 4.0 2.0 1.0 0.5 0.199 0.0199
difference quotients 2.0 2.0 2.0 2.0 1.9999 1.9999

We cannot always use linear functions in decision and choice problems. When
the underlying function is nonlinear, we must determine the slope of the
function at a particular point of interest, since the slope of the function
differs at different points on that function. Consider the nonlinear function
y = f (x) = x2 . We commence by selecting a particular point on the graph of
that function, namely (3, 9) = (x1 , y1 ). Then, as in the preceding example,
we generate a list of values of a variable x2 by specifying changes for which
the absolute values decrease as we move toward the middle column of the
table. The table shows that the difference quotient moves toward 6 as we
move closer to the point (3, 9). That quotient is not defined at 9 because
that value implies ∆x = 0.
The difference quotient method helps us find the slope of a line between two
points. It measures an average rate of change rather than the slope of the
function at a specific point. Given two points (x1 , y1 ) and (x2 , y2 ), ∆y/∆x
measures the average rate of change in y that occurs over the interval per
unit change in x. Commands are omitted because they are virtually the same
as those for the preceding exhibit.
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 89
 
x changes x values y values y changes difference quotients

 −2 1 1 −8 4 


 −1 2 4 −5 5 


 −0.5 2.5 6.25 −2.75 5.5 


 −0.2 2.8 7.8399 −1.16 5.8 


 −0.1 2.9 8.41 −0.589 5.8999 


 −0.01 2.99 8.9401 −0.0598 5.9899 


 0 3 9 0 −−− 


 0.01 3.01 9.06 0.06 6.0099 


 0.1 3.1 9.61 0.61 6.1 


 0.2 3.2 10.24 1.24 6.2 


 0.5 3.5 12.25 3.25 6.5 

 1 4 16 7 7 
2 5 25 16 8
Both the linear example and the quadratic example illustrate that as the
change in variable x, ∆x, becomes increasingly small, approaching 0 in the
limit, the difference quotient approaches some finite value as a limit. In the
case of y = x2 , as x → 3, Deltay/Deltax → 6. This is true whether we
approach x = 3 from above or from below. When this is the case, this limit
is called the derivative of y = f (x) with respect to x. This derivative is
denoted dy/dx and is defined as follows: Given the function y = f (x), the
derivative of y with respect to x, is
dy ∆y f (x + ∆x) − f (x)
= lim = lim
dx ∆x→0 ∆x ∆x→0 ∆x
provided that the limit exists.
This definition of a derivative still measures a rate of change; however, the
rate of change is an infinitesimally small change in variable x. For that reason,
a derivative may be thought of intuitively as being taken at a particular
point on a curve. We apply the definition of a derivative to the functions
y = 10 + 2 · x and y = x2 .
For the linear function
dt f (x + ∆x) − f (x) (10 + 2 · (x + ∆x) − (10 + 2 · x)
= lim = lim =
dx ∆x→0 ∆x ∆x→0 ∆x
(10 + 2 · x + 2 · ∆x) − (10 + 2 · x) (2 · ∆x)
lim = lim = 2.
∆x→0 ∆x ∆x→0 ∆x
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 90

This confirms that the slope of a linear function is the same for all values of
the independent variable.
For the second function, y = x2

dt f (x + ∆x) − f (x) (x + ∆x)2 − x2


= lim = lim =
dx ∆x→0 ∆x ∆x→0 ∆x
x2 + 2 · x · ∆x + (∆x)2 − x2
lim =
∆x→0 ∆x
2 · x + ∆x
lim = 2 · x.
∆x→0 ∆x
The function’s rate of change is directly related to x’s value. For x = 3, for
example, dy/dx = 2 · 3 = 6.
As these examples indicate, the derivative of a function is, in general, a new
function derived from the original function. If the original function was a
function of variable x and only of x, then the derivative is also a function
of x (but of no other variable). In the first, linear case the derivative is a
degenerate function of x: dy/dx = 2 · x0 .
We briefly introduce Maxima’s command to determine a derivative. We
will put it to considerable use as we proceed. The command for the type of
expression that we have encounter so far is simple. It has just two arguments.
This are the expression itself and the identity of the independent variable:
diff(expression, x). The next exhibit applies diff to the two expressions
that have analyzed so far.

(%i) [diff(10 + 2*x, x), diff(x^2, x)];


(%o) [2, 2 x]

Exercise 3-4
For the following functions, find the derivative, dy/dx = df (x)/dx, by evalu-
ating the limit of the difference quotient. Then use Maxima to confirm your
results.
1. f (x) = x3 2. f (x) = x2 + 3 · x + 4 3. f (x) = 4 · x − 1
4. f (x) = 4 − 3 · x 5. f (x) = a · b · x 6. f (x) = b · x2
7. f (x) = 144 − 32 · x
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 91

4.4.1 Geometric Interpretation of the Derivative


The formal definition of a derivative can be illustrated geometrically. The
geometric interpretation of a derivative also leads us directly to the idea of a
derivative as the slope of a tangent line at a point on a curve. We begin with
the geometric interpretation of a difference quotient. Consider the function
y = f (x) = x2 , whose graph is illustrated below. A list of four ∆x values,
two negative and two positive, is used to create a list of four x values. Those
values, in turn, are used to create a list of four y values. The four chords
below connect (3, 9) and the four (3 + ∆x, f (3 + ∆x)) points. The slopes
of these four chords are the difference quotients. Visual examination shows
that as ∆x approaches 0, the slope of the resulting chord becomes closer to
the slope of f (x) at x = 3.

(%i) limit(((x+h)^2-x^2)/h,h,0); f(x):= x^2$ x0:3$


deltaList:[-4, -3, 3, 4]$ xList:x0+deltaList$
yList: f(xList)$
wxdraw2d( explicit(f(x),x,-2,8), line width=1,
color=black,points joined=true,
points([[3,f(3)] ,[xList[1],yList[1]]]),
points([ [3,f(3)] ,[xList[2], yList[2] ] ]),
points([ [3,f(3)] ,[xList[3], yList[3]]]),
points([ [3,f(3)] ,[xList[4], yList[4]]]),
color=orange, line width=2,
key="Tangent line",
explicit(-9 + 6*x,x,0,7),dimensions=[480,480])$
(%o) 2x

The slope of the line that is tangent to f (x) at x = 3 is the limiting value of
chords like these two, when ∆x → 0.
The limit command confirms that the limiting ratio of ∆y to ∆x is 2 · x.
This command, using h to denote ∆x, shows that the limiting slope of the
chords is indeed 2x. The direction from which ∆x → 0 does not affect the
outcome and, therefore, does not need to be specified in the limit command.
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 92

Figure 4.8: Derivative as Tangent

4.4.2 Continuity and Differentiability


Not all functions can be differentiated. For a function y = f (x) to differen-
tiable at a point, it must be continuous at that same point (f (x) must exist).
Thus continuity at a point is a necessary, but not sufficient, condition for a
derivative to exist at that point. Recall our recoverable discontinuity for
!x
1
y= 2
1 + x1

Refer to Figure 4.5 to recall this function. The value x = 0 is not part of
this function’s domain. Maxima can return a derivative for this function,
the messy term below. As it turns out, however, this derivative cannot be
evaluated for x = 0. Trying to evaluate dy/dx at x = 0 yields the same error
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 93

message that resulted in trying to evaluate f(0). As an exercise, confirm that


dy/dx does exist for other values of x.

(%i) f(x):= (1/(1 + (1/x)^2))^x;


diff(f(x),x); subst(x = 0, %);
 x
1
(%o) f (x) := 1 2
1+( x )
2 1
−log ( 2 +1)
( 1 +1) x2 x
(%o) x2 x
( x12 +1)
(%o) expt: undefined: 0 to a negative exponent.
-- an error. To debug this try: debugmode(true);

To repeat, continuity is necessary for the evaluation of a derivative at a


point on a function. It is not, however, sufficient. Suppose that f (x) = |x|.
This function is continuous at x = 0 : f (0) exists, lim f (x) exists, and
∆x→0
lim f (x) = f (0) = 0. Even so, no unique value. The derivative of f (x) is
∆x→0
dy/dx = |x|/x. As from the left the term |x|/x approaches -1; as from the
right, |x|/x approaches 1. At x = 0, it is not defined. In the graph below,
any line that passes through the point (0,0) with a slope between -1 and
1 would be “tangent” to this function. Thus the derivative of f (x) has no
unique value at x = 0.
CHAPTER 4. LIMITS, CONTINUITY, AND DIFFERENTIABILITY 94

Figure 4.9: The absolute value function


Chapter 5

Differentiation: Univariate
Functions

Marginal analysis is the backbone of much of modern economic theory and


its application. Marginal analysis examines the effects on other variables of
small changes in a particular exogenous variable or a set of such values. For
example, it can be used to analyze the effect of a price change on the quantity
demanded for a good, other things equal. It can also be used to address the
effect of a simultaneous change in price and income, again other things equal.
Furthermore, once the analytical framework is defined, the implications of
changes in the “other things” can be addressed.
We can express the rate of change in one variable in response to small changes
in another variable as the first derivative of the function involved. Chapter
4 introduced the concept of a derivative and its geometric interpretation.
This chapter derives and utilizes rules that will assist us in differentiating
functions of one variable. Mastery of the technique of differentiation does
not just enable us to speak to specific problems. This mastery also enhances
our understanding of a wide range of concepts such as marginal cost, marginal
revenue, marginal utility, elasticity, and population growth.

5.1 Rules for Differentiation


Finding the derivative of a function would be tedious if it required us to
compute the derivative as the limit of a difference equation, following its

95
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 96

Figure 5.1: Some functions and their derivatives

definition in Chapter 4, in every case. Fortunately a set of time-saving rules


enables us to find the derivative without referring to the difference quotient.
Furthermore, computer algebra systems know these rules. Figure 5.1 shows
some of the most important rules. We discuss them briefly here and more
fully below.
The first and second functions above are closely related in that both are
subsets of a larger class of functions, polynomials. For a monomial, a · xb
(a polynomial with a single term), the derivative is dy/dx = b · a · xb−1 , as
indicated in the first row in Exhibit 1. A polynomial is the sum of any number
of such terms, and its derivative is the sum of terms like term’s derivative.
The second line illustrates this with a fifth-degree polynomial. Note that
the constant term a0 vanishes. This result illustrates an important general
aspect of sums and differences: The derivatives of sums (differences) of f (x)
and g(x)are sums (differences) of dfdx (x)
and dg(x)
dx
.
The third line shows the derivative for a simple logarithmic function. The
fourth line does the same for a simple exponential function. The fifth and
sixth lines require attention. These relate to the product and quotient of two
functions. For the product f (x) · g(x), taking the derivative requires taking
the derivative of g(x)and multiplying this derivative by f (x), and then taking
the derivative of f (x) and multiplying that derivative by g(x). The quotient
rule is quite similar except for the squared term in the denominator.
Exercise 5.1. Find dy/dx for the following. Solve these expressions by ap-
plying the rules in Exhibit 1 and then with Maxima.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 97

1. y = 1/2 2. y = 1000 3. y = ex 4. y = π
5. y = xn+1 6. y = x2 3/2
7. y = x p 8. y = −x2
9. y = −x−0.5 10. y = x2 · log(x) x
11. y = e / (x)

5.1.1 Polynomials
A quadratic equation illustrates the process of taking a derivative of a poly-
nomial and of interpreting that derivative. Recall that the derivative is the
slope of the function. The expression below says that the derivative is nega-
tive for x < 10 and positive for x > 10. Thus, reaches a minimum value at
x = 10, as the graph confirms.
The Maxima command to determine the expression for the derivative and to
print that expression is this:

print("For ", f(x) := 0.5*x^2 - 10*x + 100,


" the derivative is df(x)/dx =", diff(f(x), x) )$

The resulting output provides the expression for dy/dx:

For f(x) := 0.5x2 − 10x + 100 the derivative is df (x)/dx = 1.0x − 10.

The commands to plot f (x) and its derivative are these:

original: gr2d( xlabel="x",ylabel="f(x)",


explicit(f(x),x,0,20))$
derivative: gr2d( xlabel="x",ylabel="dy/dx",xaxis=true,
explicit(diff(f(x), x), x, 0, 20))$
wxdraw(original,derivative), wxplot size=[480,480];

The graphs in Figure 5.2 show the values of the quadratic function and
its derivative. The two are drawn separately because the y-axis units are
different.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 98

Figure 5.2: Graphing a quadratic function and its derivative


CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 99

5.1.2 Logarithmic Functions


Recall that logarithms are defined only for positive values of a variable. Also
recall, that the logarithm of a variable is a monotonic function of that vari-
able. Specifically, it is an ever-increasing function. Therefore, we should
expect the derivative to be positive for all values of the variable. The com-
mand diff(100*log(50*x), x) yields the result 100 x
. Remember that the
Maxima command log(x) refers to the natural logarithm, not the common
logarirthm.
The result is consistent with our expectation. The result tells us two more
things about the derivative: unlike the original function, the derivative is
monotonic but ever-decreasing, and its value is independent of the coefficient
of x. The second result reflects the fact that and the derivative of the constant
log(50) is zero. As before, the graphs in Figure 5.3 confirm that both the
initial function and its derivative are monotonic.1

5.1.3 Exponential Functions


Exponential functions can exhibit either growth or decay. In either case,
both the function and its derivative are monotonic. For a function showing
growth (positive exponent), the derivative is positive, monotonic, and accel-
erating, as Figure 5.4 shows. For a function that shows decay, the derivative
is negative but increasing toward zero. The rate of approach declines as x
increases. (Again, commands are not reported.)

5.1.4 Product and Quotient Rules


p
Consider the product and then the ratio of these two terms: (x) and e0.05·x .
The next two exhibits show the functions, the derivatives of the component
parts, and the derivative of the product or ratio. In each case, the third line
contains two equivalent statements. The first is the direct result of applying
the rule, and the second is the result of stating that expression in canonical
form. Deriving the expression on the third line of each output from the
material on the first two lines is left as an exercise.
1
The commands to generate this figure are much the same as the ones used previous
and are, therefore, not shown.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 100

Figure 5.3: A logarithmic function and its derivative

These commands

y:sqrt(x)*exp(.05*x);[diff(exp(.05*x),x),diff(sqrt(x),x)];
[dydx: diff(y,x), radcan(dydx)];

produces this output √


x %e0.05x
1
[0.05%e0.05x , √ ]
2 x
x
√ %e0.05x (x + 10) %e 20
[0.05 x %e0.05x + √ , √ ].
2 x 20 x

To analyze the ratio, use these commands


CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 101

Figure 5.4: Two exponential functions and their derivatives

z:sqrt(x)/exp(.05*x);[diff(exp(.05*x),x),diff(sqrt(x),x)];
[dzdx: diff(z,x), radcan(dzdx)];

to generate these results √


x %e−0.05x
1
[0.05%e0.05x , √ ]
2 x
x
%e−0.05x √ −0.05x (x − 10) %e− 20
[ √ − 0.05 x %e ,− √ ].
2 x 20 x

Figure 5.5 graphically depicts the behavior of these two expressions and of
their derivatives over the indicated ranges. For the product, the function is
monotonic and the derivative is always positive. For small values, the func-
tion y grows at a decreasing rate, but for larger values the exponential term’s
rapid growth dominates the damping effect of the relatively slow growth of
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 102

Figure 5.5: Graphs of y and z and of their derivatives

p
(x). The ratio of these two functions of x is not monotonic: at first, the
growth of the square root term dominates. Eventually, however, the expo-
nential term, now in the denominator, overwhelms the growth in the square
root term. The behavior of z is reflected in its derivative, which begins with
positive values but falls to zero when , and remains negative thereafter. For
large values of x, z is asymptotically approaching 0 as is its derivative (from
below).

5.1.5 The Chain Rule


Some of the illustrations above are cases that can be stated like this: y = f (u)
and u = g(x). The chain rule for deriving dx dy
is d(fdx(x)) d(f (u)) d(g(u))
√ = du · dx . The n
example below shows a case of this in which f (u) = u and g(x) √ = a+b·x .
n
The expression for which the derivative is to be determined is b x + a. The
dy
commands below determine dx twice, one by using the product defined above
and again by instructing Maxima to evaluate the expression directly. The
first list of commands below defines the two functions and the derivatives.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 103

dy
The second list of commands calculates dx
in the two ways indicated above.2

[y:sqrt(u),dydu:diff(sqrt(u),u),u:a+b*x^n,dudx:diff(u,x)];
[dydu*dudx, diff(’’y,x)];

The output consists of two lists. The first list shows y = f (u) and its
derivative and then u = g(x) and its derivative. The second list contains
dy
dx
first as a product of two derivatives and then as produced by Maxima.
Confirm the equivalence of the two expressions.
√ 1
[ u, √ , b xn + a, bn xn−1 ]
2 u

bn xn−1 bn xn−1
[ √ , √ ]
2 u 2 b xn + a

The chain rule can be extended to any number of functions.

5.1.6 Trigonometric Functions


Business and economic problems involving trigonometric functions are not as
frequent as problems that involve the functions that we have developed thus
far. We present a few of the rules for differentiating trigonometric functions
primarily to point out that Maxima can produce derivatives for functions
like those in the table below.
2
The “quote-quote” operator, ’’, instructs Maxima to evaluate y rather than treating
it as a noun. Confirm that removing this operator causes the derivative to be evaluated
as zero.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 104

5.1.7 Inverse Functions


Assume that the function y = f (x) is either strictly increasing or strictly
decreasing and continuous on an interval (a, b). If the function y = f (x) is
such that permissible values of x always uniquely determine specific values of
y, then the function y = f (x) has an inverse function of the form x = f 1 (y) =
g(y). That is, the function y = f (x) serves to define a new function g whose
value at each point y is the number x such that y = f . This means that
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 105

not only does a given value of x yield a unique value of y because y = f (x),
but also that a given value of y yields a unique value of x because x = g(y).
There is a one-to-one correspondence between y and x.
Consider the example that appears in the display below. For positive values
of x, f is monotonic. The commands below do the following. The commands
df
in the first line specify f and determine the expression for dx . The commands
in the second line define the inverse function g. The third line of commands
dg
determines the expression for dy and substitutes f where appropriate to
1
confirm that dg/dy = df /dx .

Exercise 4.3
Find dy/dx for the following. Remember that log(x) refers to the natural
logarithm, unless another base is specified. Find the derivatives by hand and
check your work with Maxima.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 106

1. y = u2 + 3 · u + 7 when u = x2 − 7 2. y = u2 when u = 1/x2


2
3. y = u3 + 4 when u = x2 + 2 · x 4. y = u1/2 when u = xx2 −3
+4
3 2 2
√ + 4 whenu = v + 2 · v andv = x
5. y = u 6. y = (x2 + 4)2
7. y = 3 · x 8. y = (2 · x3 + 1)3
x2 −1 3
9. y = (4 · x2 + x4 − 1)5 10. y = ( 2·x 2 +1 )
x 5
11. y = (x2 + 3)4 · (2 · x3 − 5)3 12. y = x+1 )
13. y = loga (x2 + 1) 14. y = log1 (x2 + 1)2
15. y = logb (2 · x3 + 3 · x) 16. y = logx
17. y = log(x2 ) 18. y = log(x2 + 1)
19. y = log(3 · x2 ) 20. y = log(x2 + x − 1)3
2
21. y = log( x1+x
2 −1 )
2
22. y = logx
x
23. y = [log(x + 2)] · (x3 + 3)
2
24. y = ax · xa
2
25. y = a1+x 26. y = 3x−2
27. y = log(1/x) 28. y = x2 · log(x)
2
29. y = a1+x 30. y = 3x−2
31. y = ex 32. y = ax · ex

5.2 Higher-Order Derivatives


The function that we obtain when taking the derivative of some function,
f (x), is also a function of x. This function may also have a derivative. This
process can be continued indefinitely. Some functions have a non-zero deriva-
tives up to a finite order; others have a limitless number of such derivatives.
The next exhibit shows one of each. A cubic function has three non-zero
derivatives; an exponential growth function has a limitless number. We look
at the first three derivatives for each.
The exhibit below shows the commands required to determine the relevant
derivatives and to create a matrix that displays these derivatives, along with
the resulting output matrix. The first line of commands defines the two
functions, y as a function of x and z as a function of t (time). The next two
lines create the first-order through fourth-order derivatives for y. Note the
syntax of the diff( ) command: insert the expression, then the independent
variable, and finally the order of the derivative. Specifying the order is
optional for first-order derivatives. The following two lines do the same for
z. Finally, the last three lines create the output matrix output.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 107

The “0th order” derivative is the original function. The fourth-order deriva-
tive of y is zero, as is any higher-order derivative. All derivatives of z are
positive if r > 0. Use the rules developed above to confirm that the entries
in each row are the derivatives of the expressions in the preceding row.
d y n
th
We use the following notation to indicate order: dx n is the n -order derivative

of y with respect to x. Because Maxima’s commands are entered as text, we


d2 y
adjust by using names like the ones in the preceding exhibit: diffy2 is dx 2,
3
and so forth. In later references, we simplify “first-order derivative” to “first
derivative” and use similar references to higher-order derivatives. We seldom
have occasion to refer to fourth derivatives or higher.
The exhibit above shows a cubic function and three derivatives (four, actually,
but the fourth derivative is 0). Figure 5.6 shows an example. It graphs the
function and each of the three derivatives. The expression for the function
and each derivative appears above the graph.
Suppose that this cubic function defines the amount of output a firm pro-
duces each week as a function of worker hours (maybe in thousands). In-
creasing employment up to just over seven units increases output. Beyond
that employment level, total output falls. Now look just below this graph
at the graph of its first derivative. That function defines the rate of output
change as employment increases. At first the derivative is both positive and
3
The names are arbitrary. Any legal name can be entered.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 108

Figure 5.6: A cubic function and three derivatives

increasing. For x greater than 2.5 or so, the derivative remains positive but
decreases. Look again at the original function and observed that for small
values of x, f (x) is accelerating, rising at an increasing rate. For larger values
of x, f (x) decelerates and finally decreases in value. An exercise: determine
the value at which the first derivative equals zero.
The second derivative’s graph (upper right panel) shows that the first deriva-
tive changes sign at x = 8/3. The second derivative’s derivative (the third
derivative of f , is a constant. The derivative of a constant is zero, so the
fourth and all higher-order derivatives for this function are zero.
These graphs are in separate panels rather than in a single graph for an
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 109

important reason. The units of the functions are quite different. Suppose that
the original function does represent production. Then the original function is
expressed in physical units per time period. The first derivative is in physical
units per time period per worker hour. The second derivative is in physical
units per time period per worker hour per worker hour, and so forth. The
units for higher-order derivatives can become quite unwieldy. Fortunately,
as previously noted, we seldom go beyond the third derivative.
Exercise 4.4
dy d2 y d3 y
Determine the following for each of these functions: ,
dx dx2
, and dx3
.
1. y = 8 · x3 2. y = 8 · x1/2 3. y = x4 + x2 + 1
4. y = ex 5. y = log(x)

5.3 Economic Applications of Derivatives


As noted, the concept of the derivative is central to economic analysis. This
section introduces the application of derivatives to a set of microeconomic
topics, to a macroeconomic topic, and to issues involving growth. Many of
these topics will be revisited in later chapters, where we extend the analysis
that appears here.

5.3.1 Production, Cost, Revenue, Profits, and Elastic-


ity
This section looks at production in the short run and relates production
to cost. It then looks at total cost curves and per-unit cost curves. Next,
demand for the firm’s product and revenue are introduced. Then we look
at profit maximization. Finally, the concept of elasticity is introduced and
demand elasticity is related to profit maximization.

The Production Function

Many important general insights relating to production can be garnered by


referring to a specific function that represents production in a relatively ab-
stract manner. Suppose for now that output q is determined by a function
q = f (L), where L (labor) is a variable input. Economics often refer to
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 110

this function as the total product of labor (TPL) function. Two important
per-unit functions can be derived from this function. The output per unit of
labor, or average product of labor is AP L = q/L = f (L)/L. The change in
output per one (small) unit change in labor is the marginal product of labor,
M P L = dLdq
= dfdL
(L)
.
The law of diminishing returns states that, beyond some employment level,
dq
M P L begins to decrease. That is dL decreases. Chapter 6 returns to this
analysis, extending it to cases in which more than one input is treated as
variable.
One of the simplest of such production functions is the Cobb Douglas func-
tion: q = A · La · K 1−a , where q is the amount that the firm produces each
period, L is the number of worker units employed each period, and K is the
number of units of capital employed each month. The coefficient A reflects
the level of technology and is constant at any point in time. For our pur-
poses, units do not matter and are ignored.4 In input/output exhibit below
shows the total product, average product, and marginal product functions
for a Cobb-Douglas production function. A little manipulation reveals that 1−a
the AP L and M P L functions can be stated as follows: AP L = A · K L
1−a
and M P L = a · A · K L
, so that M P L = a · AP L. This simple relation-
ship between AP L and M P L is a characteristic of this function and does
not generalize to all production functions.

The graphs in Figure 5.7 show the total and per-unit functions. One aspect
of this function that might cause concern is that M P L decreases from the
beginning and is always below AP L. In principles classes you probably saw a
different configuration, with M P L rising at first, until diminishing marginal
product sets in. The failure of the Cobb-Douglas to generate this result is a
4
“Our purposes” are purely illustrative. For applications to specific issues, of course,
knowing the units is crucial.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 111

failure but probably not a serious one. Most production occurs in the range
where diminishing marginal product is present; in fact, in the range for which
AP L > M P L > 0. For the Cobb-Douglas case, AP L > M P L > 0 for all
values of L.5

Figure 5.7: TPL, APL, and MPL for Cobb-Douglas production function

5
In the Maxima commands, the L range does not begin at 0 for the per-unit functions.
The values of AP L and M P L are not defined for L = 0.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 112

Production and Cost

Costs for Cobb-Douglas Production. For the Cobb-Douglas function we


can derive cost curves that express total cost (and both average and marginal
cost) explicitly in terms of the quantity produced.6
To derive a total cost curve requires three steps. First, the command solve(f=q,L)
relates quantity to output, with this result:
1 1
q a K 1− a
[L = 1 ].
Aa
Now, define the total cost function with the command TC: subst%, w*L +
r*K), which uses the output from the previous command to substitute for
L in the expression T C = w · L + r · K, where w and r are unit costs of
labor (L) and capital (K). The resulting expression is this rather daunting
expression:
1 1
q a w K 1− a
1 + rK.
Aa
Finally, substitute values for K, w, r, and the production function param-
eters, using the command subst([K=1000,r=1,w=1,A=1,a=.7], TC).7 The
resulting expression 0.051794q 1.4285 + 1000, is quite easy to interpret. The
fixed cost is 1000 per period, and variable cost increases by 1.4285 percent
per one-percent change in q (reflecting the uniformly decreasing marginal
returns that we noted earlier).
Figure 5.8 shows the per-unit cost curves for this production function (see the
workbook for more detail). Both the average variable cost and the marginal
cost increase monotonically.
A Cubic Cost Function. In actual production, increasing marginal re-
turns may occur over some range before diminishing marginal returns set
in. Therefore, both average variable cost and marginal cost curves may be
U-shaped. A cubic total cost function is the simplest one that can illustrate
6
For some other production functions, such explicit cost functions might not exist.
Often, however, cost curves can be drawn using a parametric approach, with one of the
inputs serving as the parameter.
7
Any values would do for K, r, w, and A. The value of a must be between 0 and 1.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 113

Figure 5.8: Per-unit cost curves for a Cobb-Douglas production function

such cost curves. We use such a curve for the following analysis.8 The cost
functions are as follows:
Total Cost Average Cost
0.05q 3 − 1.7q 2 + 25q + 150 0.05q 2 − 1.7q + 150
q
+ 25

Average Variable Cost Marginal Cost


0.05q 2 − 1.7q + 25 0.15q 2 − 3.4q + 25

Figure 5.9 shows the curves that these expressions generate. Both T C and
T V C increase throughout the range of production. For small output rates,
the rate of cost increase diminishes, but after q ≈ 10 the rate increases. The
M C curve reflects this aspect of the T C curve, decreasing for small q and
increasing for larger q. Both AC and AV C exhibit the U-shape that economic
theory suggests one should observe. The two converge as q increases, with
the vertical distance between them equally the value of AF C. The AF C
curve contains no useful analytical information and will be omitted from
subsequent analysis.
Figure 5.9 reveals three quantities that can be of interest: the quantity at
which M C’s sign changes, the quantity at which M C = AV C (and AV C
8
We could begin with a cubic production function and establish the relationship be-
tween output and cost, but the process is rather involved. See Hammock and Mixon
[7].
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 114

Figure 5.9: Total and Per-unit cost curves, cubic cost function

achieves is minimum value), and the quantity at which M C = AC (and AC


achieves its minimum value. We see below that these values are approxi-
mately 11.333, 17, and 20.551 respectively. In the following analysis, require
that Maxima return decimal representations (via the command numer:true).
The first input line defines the slope of M C twice, once by taking the first
derivative of M C and then by taking the second derivative of T C. As must be
the case, the two results are identical. The commands are these: [MCslope:
diff(MC,q), diff(TC,q,2)]. The result is this list of two identical items:
[0.3q − 3.4, 0.3q − 3.4].
The second input line contains three applications of solve( ), to determine
the three values indicated above. The result of solve(MC = AVC...) is
of note, because Maxima produces a list of values, only one of which is
real. The commands list is as follows: soln1:solve(MCslope,q); soln2:
solve(MC = AVC,q); soln3:solve(MC=AC,q);. The first command in
the list yields [q = 11.33], the quantity at which the marginal cost curve
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 115

reaches its minimum value. The second command yields [q = 0, q = 17]. The
first solution reflects a peculiarity of the cubic functional form and has no
economic import. The second shows the quantity at which AV C reaches its
minimum value.
The final input item in the list of commands, solve(MC=AC,q, determines the
quantity at which the marginal cost curve intersects the average cost curve:
q = 20.55].9

Demand and Revenue

The general expression for the demand curve is q = f (p), where f shows the
quantity demanded at each price. Often, it is easier to work with the inverse
function, p = g(q). An important reason for using the inverse demand curve
is that total revenue is T R = p·q. The illustration here treats a case in which
a homogeneous product is sold in a single market and the same price applies
to each unit. More complex demand curves can be specified with suitable
extensions of this simple representation.
The illustration in this section uses two types of demand curves, linear and
constant-elasticity (the concept of elasticity is the one you encountered in
your principles course and which the next section revisits). The demand
curves are of this form: q = a + b · p and q = A · pe . In the first specification,
a is the horizontal intercept (the quantity demanded when p = 0 and b is the
demand curve’s slope (b < 0). In the second specification, A is a constant
(the quantity when p = 1) and e is the price elasticity of demand (e < 0).
For this specification, no horizontal intercept exists.
The three lines of input below show expressions for two demand curves,
named fL and fCE. The second input line provides Maxima with information
it requires in determining the inverse functions. The third input line produces
expressions for the inverse demand curves.

[fL: a + b*p, fCE: A*p^e];


declare(e, noninteger)$ assume(q>0, A>0)$
[solve(fL=q,p), solve(fCE=q, p)];
9
As the accompanying workbook shows, this command produces three solutions, only
one of which is real and, therefore, of consequence for this analysis.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 116

The resulting output consists of two lists. The first, [bp + a, pe A], is the
original expressions for the quantity demanded. The second, [[p = q−a
b
], [p =
1
qe
1 ]], is the pair of inverse demand curves.
Ae
The inverse of the linear demand curve can be written as p = (a/b)+(1/b)·q,
so that (a/b) is the vertical intercept (graphs appear below) and (1/b) is the
inverse demand curve’s slope. For the constant-elasticity demand curve, the
inverse is p = (1/A( 1/e)) · q ( 1/e) . The following input shows the inverse
demand curves and the associated total revenue curves and marginal revenue
curves. We use functional notation, pL(q,a,b) for example, to facilitate
graphing. The first two input lines specify the inverse demand curves. The
second pair of lines specifies total revenue functions. The “quote-quote”
operator (’’) is used to force Maxima to evaluate the products in the second
pair of commands and the derivatives in the third pair.

pL(q, a, b) := a/b + q/b$


pCE(q,A,e) := q^(1/e)/A^(1/e)$
TRL(q, a,b):=’’(expand(q*pL(q,a,b)))$
MRL(q,a,b) := ’’(diff(TRL(q,a,b),q))$
TRCE(q, A,e):=’’(q*pCE(q,A,e))$
MRCE(q,A,e) :=’’(diff(TRCE(q,A,e),q))$

The resulting expressions are these:


 
Linear, L Constant Elasticity, CE
1
 q a qe 
 P rice b
− b 1

 A1
e 
 q2 aq q E +1
.
 T otal Revenue
b
− b 1

 AE 1 
1
(e )
 +1 q e

2q a
M arginal Revenue b
− b 1
Ae

We specify a set of parameters: a = 60 and b = −2 for the linear demand


curve, and A = 1200ande = −1.5 for the constant-elasticity demand curve.
Figure 5.10 shows the graphical representation of these two inverse demand
curves and the associated marginal revenue curves.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 117

Figure 5.10: Inverse demand curves and marginal revenue curves

Profits

Recall from principles of microeconomics that producing the quantity at


which M R = M C yields maximum profit for a firm. We combine the cost
information above with the CE demand curve to determine how this firm’s
profit level relates to quantity and, in doing so, we confirm that the equality
of M R and M C does correspond to maximum profit (or minimum loss).
Figure 5.11 shows the total and per-unit curves that relate to the cost func-
tion above, combined with the constant-elasticity demand curve.10 The top
10
See the accompanying workbook for the commands that generate these graphs.
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 118

Figure 5.11: Revenue, costs, and profit

panel, which shows total values, reveals that the firm earns losses at all out-
put levels because total revenue lies below total cost. The firm does, however,
cover its variable costs so that producing at a rate between 0 and just un-
der 25 units results in a lower loss than producing nothing and losing the
per-period fixed cost, $150. Minimum loss appear to occur at approximately
q = 15. We determine the actual value below.
The per-unit curves tell the same story. The inverse demand curve (p) is
below average cost but above average variable cost. Maximum profit (mini-
mum loss here) occurs where the marginal revenue and marginal cost curves
intersect, apparently around q = 14. The combination of a polynomial cost
CHAPTER 5. DIFFERENTIATION: UNIVARIATE FUNCTIONS 119

function and constant-elasticity demand curve does not yield an analytical


solution, so we cannot directly solve M R = M C to determine the profit-
maximzing quantity. Rather, we use a numerical root-finding option in
this command: xOptimal: find root(MC-MRCE(q,1200,-1.5),q,5,15),
which yields a value of q = 13.701. Using the command pCE(xOptimal,1200,-1.5)
to plug the optimal quantity into the demand curve shows that the loss-
minimizing price is $19.72. Finally, the command PCE(xOptimal,1200,-1.5)
yields the result that the profit level is -$31.78 per time period.
Chapter 6

Differentiation: Multivariate
Functions

The material up to this point is developed mostly in terms of functions of a


single variable. This chapter extends the analysis to functions that involve
several variables. It focuses on differentiation, but in doing so it addressed
pertinent aspects of functions of more than one variable. Fortunately, most
of what we have learned up to this point applied with relatively little modi-
fication.

6.1 Partial Differentiation


This chapter addresses functions of two or more independent variables. Con-
sider z = f (x, y), where z is the dependent variable and x and y are indepen-
dent variables. Given any permissible values of x and y, we can determine a
value for z. For example, if z = x · y, and x = 2 and y = 4, then z = 8.
This chapter deals with functions of the form y = f (x1, x2, ..., xn), where
the second part of each independent variable’s name distinguishes it from
the other independent variables. That is x1 is a variable, not a value of a
variable named x.1
1
Recall that we use this approach to naming to accommodate the fact that computer
algebra systems typically require that data entry involves text characters. Often in later
material, we use x1 , rather than x1, where the subscript adds clarity.

120
CHAPTER 6. DIFFERENTIATION II 121

The execution of partial differentiation, also called taking a partial derivative,


is quite similar to its single-variable counterpart, and we proceed in the same
way. Suppose that we wish to examine the effect of a change in the value of
x2 with all other independent variables keeping their initial variables. Then
differentiation with respect to x2 alone involves taking the limit of a difference
ratio as follows:
∂y ∆y f (x1, x2 + ∆x2, · · · , xn) − f (x1, x2, · · · , xn)
≡ lim = lim .
∂x2 ∆x2→∞ ∆x2 ∆x2→∞ ∆x2

The derivative taken in this expression measures the rate of change of y with
respect to x2 and is referred to as the partial derivative of y with respect to x2.
The partial derivative indicates the effect of a change in a single independent
variable on the dependent variable, with all other independent variables in
the function held constant while this particular derivative is taken.
The process of partial differentiation is denoted by the variant form of the
lower-case Greek letter delta (δ), namely ∂. Suppose that we have a function
of the form y = f (x1, x2). We represent the partial derivative of y with
∂y
respect to x1 by ∂x1 . Notation regarding partial derivatives is not completely
∂y ∂y
standard. Any of the following can be used to mean the same as ∂x1 : ∂x1 ,
f1 , fx1 , y1 , or yx1 . The context usually makes the meaning clear.

6.2 Rules of Differentiation


Fortunately the rules of differentiation that we derived previously can also
be used to find partial derivatives when we deal with functions of several
independent variables. The major alteration in our differentiation procedure
is that all independent variables not explicitly involved in the differentia-
tion are treated as constants and are differentiated accordingly. Maxima’s
command reflects this similarity: the same diff() command applies to both
simple derivatives and partial derivatives.
Look at two examples. The first is u = x2 + 4 · x · y + y 2 . The commands
u_x : diff(u(x,y), x) and u_y: diff(u(x,y), y) generate this result:

[4y + 2x, 2y + 4x].


CHAPTER 6. DIFFERENTIATION II 122

We use u_x and u_y as names for these partial derivatives, ∂u/∂x = 4y + 2x
and ∂u/∂y = 2y + 4x. It is apparent in this case that the value of both of
the partial derivatives depend on the values of both x and y.2
For the next example, v = 2 · y/x + 4 · x/y, the partial derivatives are also
functions of both variables, as the table below shows. The table below shows
the original function and the two partial derivatives. Both ∂v/∂x and ∂v/∂y
are rather involved functions of both variables, x and y. When we look at
higher-order partial derivatives, we will discover a way to determine more
about how values of the partial derivatives change as x and y change.
 
v(x,y) v_x v_y
2y
x
+ 4x
y
4
y
− x2y2 x2 − 4x
y2

We can generalize the results of the previous examples. Suppose we have


a function y = f (x1, x2 . . . , xn), where the x’s are independent variables
that are independent of one another. It is in principle possible to compute
a partial derivative of y with respect to each of the x’s. A total of n such
partial derivatives exist.3
Exercise Set 6.1.
In each case compute all first-order partial derivatives. Then confirm your
results using Maxima.
4 4 2 2
p+y −4·x ·y
1. f (x, y) = x 2. f (x, y) = x · y + x/y, for y 6= 0
3. f (x, y) = x2 + y 2 4. f (x, y) = (x + y)/(x − y) for x 6= y
2
5. f (x, y) = ex +x·y 6. f (x, y, z) = x · y · z
7.f (x, y) = x/y 2 + y/x2 8. f (x, y) = loge (x · y + y 2 )

9. f (x, y) = x · cos(y) + y · cos(x) 10. f (x, y) = x · y
2 /2+3·y 2
11. f (x, y) = xx·y+y·z

6.2.1 Geometric Interpretation of the Partial Deriva-


tive
We learned in Chapter 4 that when y = f (x), the derivative of y with re-
spect to x can be given a geometric interpretation: dy/dx is the slope of
2
This need not be so: consider, for example z = a · x + b · y 2 . Then ∂z/∂x = a, a
constant, and ∂z/∂y = 2 · y , a function of y alone.
3
The workbook that accompanies this chapter illustrates this point with a function of
four variables.
CHAPTER 6. DIFFERENTIATION II 123


Figure 6.1: Graph of x·y

the curve of that function at a point in a plane. Similarly, a partial deriva-


tive has a geometric interpretation. Consider the function of two variables
√ √ √
f (x, y) = x · y. The partial
√ derivatives are ∂f (x, y)/∂x = (1/2) · y/ x

and ∂f (x, y)/∂y = (1/2) · x/ y . Thus, both partial derivatives are (1)
positive and (2) directly related to the value of the other variable. Figure
6.1 shows this surface for a range of x and y values.
The surface in Figure 6.1 confirms our expectation. Holding y at given value
defines a line on the surface that shows z values for the range of x values.
Likewise, holding x at a given value defines a line that shows z values for the
range of y values. As y increases, the level of z for each x value increases; also
the slope of the line that shows z relative to x becomes steeper. Likewise,
as x increases, z increases at each y value and the slope increases at each
value of y. Compare the slopes of the lines labeled “y = 1” and “y = 4” that
indicate levels at which y is held constant.
CHAPTER 6. DIFFERENTIATION II 124

6.2.2 Differentials
One purpose of analysis is to explain, at least in a qualitative sense, how
changes in one or more independent variables affect the value of a dependent
variable. To approach this question in an intuitive fashion, we begin with a
tautology: ∆y = (∆y/∆x) · ∆x. As it stands this says that the change in y
equals the change in y per unit change in x, multiplied by the change in x.
This, as it stands, tells us nothing. Suppose, however, that we replace with
its limiting value, dy/dx. Now, we can say that ∆y u (dy/dx) · ∆x. This
expression says that, at a given value of x, the line tangent to the function
approximates the effect of a change in x on y.
One step remains in defining the differential. As ∆x becomes a small num-
ber, the product (dy/dx) · ∆x becomes closer to ∆y. Thus, for sufficiently
small change dx, the expression becomes very nearly accurate. Consider an
example. Let y = x3 . Change x from 2 to 2.01 and determine both ∆y and
dy. At x = 2, y = 8, and dy/dx = 3 · x2 . At x = 2.01, y = 2.013 = 8.1206,
so ∆y = 8.1206 − 8 = 0.1206. The predicted change using the differential is
dy = 3 · x2 · 0.01 = 0.12.
We extend the concept of a differential to include functions of two or more
independent variables. We begin by noting that a partial derivative measures
the rate of change of the dependent variable with respect to an infinitesimal
change in one of the independent variables, all other independent variables
held constant. A total differential of a function, however, is a linear ap-
proximation of the rate of change of the dependent variable when all of the
independent variables change by an infinitesimal amount.
The total differential is the sum of the changes in the dependent variable
caused by simultaneous infinitesimal changes in all the independent variables.
Suppose that y = f (x1, x2, · · · , xn). Then the total differential of y is this:
dy = f1 · dx1 + f2 · dx2 + · · · + fn · dxn, where dx1, dx2, . . . , dxn indicate small
changes in the n independent variables. These changes occur independently
of each other.
Extend the example above by making z = x3 + y 2 . The total differential
for z is dz = 3 · x2 · dx + 2 · y · dy. Begin with x = 2 and y = 3, so that
z = 8 + 9 = 17 and dz = 12 · dx + 6 · dy . Let ∆x = 0.01 and ∆y = 0.01.
Then the approximation for dy is 12 · 0.01 − 6 · 0.01 = −0.18, so that the
approximate value of y is now 17.18. The actual value is 17.181.
CHAPTER 6. DIFFERENTIATION II 125

6.2.3 Maxima Notation


To see how Maxima treats a differential, we revisit the preceding paragraph.
The command z: x^3 + y^2 creates the expression to be evaluated. To
find that initially z = 17, use this command: subst([x=2,y=3],z). Like-
wise, to see that z’s value changes to approximately 17.181, use the command
subst([x=2.01,y=3.01],z).
Now use the command diff(z) to define the total differential of this expres-
sion:
2y del(y) + 3x2 del(x).
We can substitute the values of x and y and evaluate the differential with the
command subst([del(x)=.01, del(y)=.01,x=2,y=3],%), getting 0.18 as
the result. Note that Maxima treats del(x) and del(y) as variables.
Exercise Set 6.2
Find the total differentials for these functions by hand and with Maxima.
1. z = x2 /2 + x3 /3 2. z = x2 + x + y + y 2
3. z = loge (x + y) 4. z = x · x · y
2
5. w = ex + y 2 + z 2 6. w = z 2 · (2 · x + 3 · y)

6.2.4 Total Derivatives of Composite Functions


The previous section developed the concept of a total differential. This sec-
tion develops the concept of a total derivative. The slight difference in termi-
nology represents a substantial difference in the type of function with which
we are dealing. Our discussion of partial derivatives and total differentials
relied on the assumption that the independent variables were independent of
one another. Often, however, such an assumption cannot be fulfilled. This
section considers three general cases in which the assumption is violated. It
also considers the case in which no dependent variable can be identified but
in which two are more variables are bound by an composite function.
Case 1: Interrelated Independent Variables
Given the function z = f (x, y), where it also is true that y = g(x). Hence z =
f (x, g(x)), which is a composite function in which x is the only independent
variable. An important underlying assumption of a partial derivative is no
longer fulfilled when we encounter a function such as z = f (x, g(x)). Variable
CHAPTER 6. DIFFERENTIATION II 126

y = g(x) cannot remain constant when x varies. That is, x and y are not
independent of each other. In a case like this, dz = ∂f /∂x · dx + ∂f /∂y · dy
is the total differential of z. We can divide both sides of this expression by
dx to obtain dz/dx = ∂f /∂x + ∂f /∂y · dy/dx. The term dz/dx is the total
derivative of z with respect to x.
This total derivative has two parts. The first part, ∂f /∂x, measures the
change in z brought about by changes in x, all other variables held constant.
The second part, ∂f /∂y · dy/dx measures the change in z brought about by
change in variable x that works through intermediate variable y. The first
part of this expression is often called the direct effect, while the second part
is called the indirect effect. The indirect effect takes account of the fact that
changes in variable x affect variable y, which in turn affects variable z. In
general if z = f (x, y1, y, . . . , yn, then
dz ∂z ∂z dy1 ∂z ∂dy2 ∂z dyn
= + · + · + ··· + · .
dx ∂x ∂y1 dx ∂y2 dx ∂yn dx

Consider three examples, the first of which we also develop with Maxima.
Let z = f (x, y) = x2 + 2 · x · y + y 2 . The total derivative is dz/dx = ∂z/∂x +
(∂z/∂y)·(dy/dx). In our example, y = g(x) = e3·x , so that dy/dx = 3·e3x . We
confirm that the rule for evaluating dz/dx yields the same result as inserting
g(x) into f (x, y) and evaluating the result. We proceed with three sets of
commands, which generate the table below.
y 2 + 2xy + x2 2y + 2x 2y + 2x
6x 3x 2 6x 3x 3x
%e + 2x %e + x 6%e + 6x %e + 2%e + 2x
3%e (2y + 2x) + 2y + 2x 6%e4x + 6x %e3x + 2%ex + 2x
3x

The table above shows the result of using Maxima to determine this total
derivative. The first input line defines z without specifying how x and y are
related, and it takes the two partial derivatives. Maxima responds as if x
and y are independent of each other. The commands are these:
[z : x^2 + 2*x*y + y^2, dzdx01: diff(z,x), dzdy:diff(z,y)]. The
name dzdx0 is assigned to the derivative with respect to x because, we will
determine this derivative two more times in the next two lines of input.
The second input line, [z:subst(y=exp(3*x),z),dzdx02:diff(z,x)], adds
the information that y is a function of x and takes the total differential of z,
which is the second item in the second line of output.
CHAPTER 6. DIFFERENTIATION II 127

The third input line, [dzdx03: dzdx01 + dzdy*diff(%e^(3*x),x),


dy
expand(subst(y=%e^x,dzdx03))], confirms that dy = dx · dx. The value of
dy
dx
is determined by the command diff(%e^(3*x),x) a value that is used in
the final step. The subst command informs Maxima that y = e3·x .4
For a relatively simple example like this one, using Maxima provides little
advantage. In a more complex example, however, Maxima can both simplify
the process of extracting the total derivative and reduce the probability of
errors.
2 2
The second example is the function z = ex −y , with x = 2 · y 3 . For this
2 2 2 2
function, dy dz ∂z
= ∂y ∂z
+ ∂x · dx
dy
= ex −y · (−2 · y) + ex −y · (2 · x) · (3 · 2 · x2 ) =
2 2
12 · x · y 2 · ex −y .
Finally, let w = f (x, y, z) = 2 · x2 + 3 · y 3 + 4 · z 4 , where y = ex and
dy dz dy
z = loge x. In this case, dx = e and dx = 1/x, so dw
dx
= fx + fy · dx dz
+ fz · dx ,
dw 2 x 3 1
or dx = 4 · x + 9 · y · e + 16 · z · x ).

Case 2. “Independent” Variable(s) Affected by One or More Ex-


ternal Variables
Chapter 5 introduced explicit expressions in which x and y are both functions
of a third variable the “parametric” expressions. The case considered here
is similar. Suppose that z is a function of two (or more) such variables.
Let z = f (x, y), where x = g(t) and y = h(t). This is another variant of
a composite function, where z = f (g(t), h(t)). The value of z can change
when the value of either x or y (or both) change independently of t (due
to a parameter shift, for example) or when a changed value of t causes the
change(s) in x and y. We consider two examples.
2
First, suppose that f (x, y) = x2 · y 3 , where z = t2 and y = 3 · t3 . In this case
dz
dt
= fx · dx
dt
+ fy · dy
dt
= (2 · a · x · y 3 ) · t + (3 · x2 · y 2 ) · 9 · t2 . This expression
dz
simplifies to dt = 2 · a · x · y 3 · t + 27 · (x · y · t)2 .
For the second example, we refer to Maxima. The expression to be evaluated
is w = g(x) = ex·y·z . The command w: exp(x*y*z) assigns the name
w to this expression. With this information alone, the command diff(w)
4
This command is embedded in the expand command to make the items in the second
and third lines of output easier to compare. It is not required.
CHAPTER 6. DIFFERENTIATION II 128

produces this output:


xy %exyz del(z) + xz %exyz del(y) + yz %exyz del(x).

Suppose, however, that x = s2 + t2 , y = s · t, and z = t. In this case, none
of the three variables can change while the others remain constant. The
command depends([x,y],[s,t],z,t) tells Maxima that x and y depend
on both s and t, and that z depends on t alone. It does not specify the nature
of the dependencies.
Now the command diff(w) produces this output:
       
xyz d d d
%e xy z +x y z+ x yz del(t)+
dt dt dt

     
d d
x y z+ x yz %exyz del(s),
ds ds
which shows that five derivatives must be evaluated in order to produce dw
in terms of s and t. The rather long subst command
subst([diff(x,s)=diff(s^2+t^2,s), diff(x,t)= diff(s^2 + t^2,t),
diff(y,s)= diff(s*t,s), diff(y,t)= diff(s*t,t),
diff(z,t) =diff(sqrt(t),t) ], diff(w)) provides the necessary infor-
mation for evaluation of dw. Now the command diff(w) generates this
result:
 
xy
2tyz + sxz + √ %exyz del(t) + (2syz + txz) %exyz del(s).
2 t
We are not quite finished, because x, y, and z are in the expression. One
more set of substitutions, subst([x=s^2+t^2,y=s*t,z=sqrt(t)],%) gives
us the result that we seek:

3s t (t2 + s2 )
  3
%es t (t +s ) del(t)+
5 2 2 2
2s t +
2
2
 3  3
t t + s + 2s t %es t (t +s ) del(s).
2 2
2 2 2 23 2

2

The coefficient of del(t) is ∂w/∂t, and the coefficient of del(s)–on the second
line– is ∂w/∂s.5
5
The workbook shows that placing the definitions of s and t directly into w and
executing diff(w) yields the same result.
CHAPTER 6. DIFFERENTIATION II 129

Exercise 6.3
Evaluate the derivatives below, first by following the approach used above
and then with Maxima.

1. For z = x2 · y + x · y 2 , when y = x3 , find dz/dx.


2
p
2. For z = ex ·y · x2 , when y = (x), find dz/dx.

3. For z = loge x, when y = 8 · x + 8, find dz/dx.

4. For z = x/y + x · y, when x = 2 · t and y = t2 , find dz/dt.



5. For z = x + 8 · y, when x = 2 · t + 4 and y = t3 + t, find dz/dt.

6. For w = x2 · y 2 + x · y · z + y 2 · z 2 , when x = 2 · t, y = 2 · t + 2, and z = t2 ,


find dw/dt.
p
7. For z = x2 + y 2 , when x = x−tF and y = x+t, find ∂z/∂s and ∂z/∂t.

8. For z = (x2 +y 2 )3 , when x = s+t and y = 25−t, find ∂z/∂s and ∂z/∂t.

9. For z = loge (x · x · z), when x = s2 · t, y = s · t2 , and z = s · t,


find ∂z/∂s and ∂z/∂t.

Case 3. Implicit Functions


Many of the functions used in the study of business and economics are implicit
functions of the form F (x, y) = 0. Examples include isoquants, indifference
curves, budget constraints, and iso-cost curves. Also, demand and supply
curves can be expressed as implicit functions. Sometimes, as in the case
of a demand curve, the implicit function can be rephrased as an explicit
function, but this need not be so. For example, an indifference curve that
slopes upward over some range (at the margin, on good has become a “bad”)
cannot be rephrased this way. We encounter some examples below. Standard
notation is to express an implicit function of and as follows: F (x, y) = 0, with
upper-case F replacing the lower-case f .
Reasoning that we have developed implies that dF = ∂F ∂x
· dx + ∂F
∂y
· dy.
F (x, y) = 0, a constant, so dF must equal zero. Therefore, dy/dx must equal
−Fx /Fy , given that Fy 6= 0. This important result is the Implicit Function
Theorem. As an example, suppose that F (x, y) = x2 +y 2 −16, which can also
CHAPTER 6. DIFFERENTIATION II 130

Figure 6.2: An implicit function

be written as x2 +y 2 = 16. The table below shows, Fx = 2·x and Fy = 2·y, so


dy/dx = −x/y. Thus the function slopes downward when x and y have the
same sign and upward when they have opposite signs (notation: in Maxima
we use Fx and Fy to indicate Fx and Fy ). The derivative dy/dx is not defined
when y = 0. Figure 6.2 confirms the observations made above.6

table form([["F", "Fx", "Fy","-Fx/Fy"],


Fx (x,y) := ’’(diff(F(x,y),x)),
Fy (x,y) := ’’(diff(F(x,y),y)), -Fx(x,y)/Fy(x,y)] ])$

F Fx Fy −Fx/Fy
2 2
F (x, y) := x + y − 16 Fx (x, y) := 2x Fy (x, y) := 2y − xy

As a second example, suppose that F (x, y) = ex + ey − 20. Following the


same process as before, we can establish that dy/dx = −ex /ey = −ex−y .
Both ex and ey are positive, so dy/dx < 0 for all values of both variables.
Neither variable’s value can exceed loge 20 (why?). As x approaches that
limit, y must become very small, and vice versa. Figure 6.3 illustrates these
conclusions.
6
If the two variables have the same units, then this is the formula for a circle with
radius = 4 and with the center at zero. Maxima does not assume that units are the same.
If you want the output to look circular, add this option: proportional axes=xy.
CHAPTER 6. DIFFERENTIATION II 131

Figure 6.3: Graph of ex + ey − 20

Finally, consider F (x, y)y 3 + x · y − 12. Following the steps in the preceding
examples reveals that dy/dx = −y/((3 · y 2 + x)), as reported in the following
table..
F (x, y) Fx Fy dy/dx
y + xy − 12 y 3y + x − 3y2y+x
3 2

This expression is much more difficult to evaluate than then preceding two.
Graphical analysis helps us to understand it better. Figure 6.4 shows that
the relationship between x and y is not monotonic. For the values of x
and y that yield the upper curve, the derivative is monotonic (dy/dx > 0),
but for the values that generate the lower curve the derivative is no longer
monotonic. If, however, our attention is limited to positive x values, then
the implications are simpler: Both the initial function and its derivative are
monotonic: dy/dx is negative but it approaches zero as x increases.
Partial derivatives can be defined for implicit functions of more than two
variables. If F (x, y, z) = 0, then ∂z/∂x = −Fx /Fz . The other two partial
derivatives can be defined in like fashion. The next display shows an implicit
function of three variables. It also shows Fx and Fy , along with ∂y/∂x. As
the entry in the fourth column shows, for this function ∂y/∂x depends on
the values of all three variables.
F (x, y, z) Fx Fy ∂y/∂x
√ y y
y z 2 + xy
2 √
2 xy
2y z + 2√xxy −
2
√  
2 xy 2y z 2 + 2√xxy
CHAPTER 6. DIFFERENTIATION II 132

Figure 6.4: Graph of x3 + x · y − 12

6.3 Higher-order Partial Derivatives


Given the differentiable function z = f (x,y), we have seen that the process
of partial differentiation produces two new functions, namely ∂z/∂x = fx =
∂f (x, y)/∂x and ∂z/∂y = fy = ∂f (x, y)/∂y. Each of these two functions is
itself a function of variables x and y, so we can differentiate them once again
in order to obtain the rate of change of the partial derivative with respect to
either x or y.
The partial derivative of a partial derivative is referred to as a higher-order
partial derivative. Standard notation for the case in which we take a “second-
order partial derivative” (the partial derivative of a partial derivative) is as
follows:

∂ 2z ∂ 2f
 
∂ ∂z
zxx = = = = fxx
∂x ∂x ∂x2 ∂x2
∂ 2z ∂y 2 f
 
∂ ∂z
zyy = = 2 = = fyy
∂y ∂y ∂y ∂y 2
∂ 2z ∂ 2f
 
∂ ∂z
zxy = = = = fxy
∂y ∂x ∂y∂x ∂y∂x
CHAPTER 6. DIFFERENTIATION II 133

∂ 2z ∂ 2f
 
∂ ∂z
zyx = = = = fyx
∂x ∂y ∂x∂y ∂x∂y

We refer to zxy and zyx as cross (or mixed) partial derivatives. They result
when one differentiates function z first with respect to one variable and then
with respect to the other variable. For example, let z = 5 · x2 · y. Then we
find zxy in two steps. First we differentiate z with respect to x, and obtain
10 · x · y. Then we differentiate 10 · x · y with respect to y, and obtain 10 · x,
which is zxy .
In many cases, zxy = zyx . When this is the case, it makes no difference
whether we first differentiate the function with respect to x, then with respect
to y, or vice versa. We obtain the same result. In general, zx y = zy x when
Young’s Theorem applies.

Young’s Theorem: If a function f (x, y), its two first-order par-


tial derivatives, and both cross partial derivatives are continuous,
then fxy = fyx .

This theorem may be generalized to functions of n variables. It enables us


to disregard the ordering of our differentiation when we find cross partial
derivatives, provided that the continuity property holds as outlined.
The
p table below illustrates these derivatives with the function f (x, y) =
(x · y + x · y 2 + x2 · y). The first line shows the original function. The
next two lines show the first partial derivatives. The fourth and fifth lines
show the second partial derivatives. The last two lines show the cross partial
derivatives, which equal each other, in accordance with Young’s Theorem.

Name Expression

z xy + x y 2 + x2 y
y
zx √
2 xy
+ y 2 + 2xy
zy √x + 2xy + x2
2 xy
y2
zxx 2y − 3
4(xy) 2
x2
zyy 2x − 3
4(xy) 2
xy
zxy √1 − 3 + 2y + 2x
2 xy 4(xy) 2
xy
zyx √1 − 3 + 2y + 2x
2 xy 4(xy) 2
CHAPTER 6. DIFFERENTIATION II 134

Exercise 6.4
Use implicit differentiation to show ∂z/∂x and ∂z/∂y for these expressions.
1. 2 · x2 + 3 · y 2 + 4 · z 2 = 24 2. loge x · y · z = 10
3. a · x + b · y + c · z = e 4. ex + ey + ez = 1000
5. 3 · x2 − 4 · y − z 2 + x2 · y · z 2 = 20 6. −x2 − 4 · y 2 + 2 · z 3 = 60
7. loge x + loge y + loge z = ex 8. x + y − loge z = 0
2 3
9. (x + 8 · y · z) · (x + 5) = 8
10. x3 + y 3 + z + x · y + 2 · x · y 2 + 3 · y · z 3 = 0

Given the following expressions, determine the indicated higher-order par-


tial derivatives.
11. For z = x2 + 2 · x · y + y 2 , find zxx and zyx
12. For z = 4 · x2 · y 2 , find zxx and zyy
13. For z = e( x2 + y 2 ) + 4 · x3 · y 2 , find zxx and zyy
14. For z = x/y − y/x , find zxx and zyy
15. For z = 2 · x2 + y 2 − 4 · x − 8 · x · y 2 , show that zxy = zyz
16. For z = (x − y)/(x + y), show that zxy = zyx
17. For z = loge (x2 + y 2 ), show that zxx + zyy = 0
18. For z = e( x2 + y 2 ) + 4 · x3 · y, show that zxyy = zyxy = zyyx

6.4 Applications of Partial Derivatives


The remainder of this chapter consists of two applications of partial deriva-
tives to economic analysis. The first application is to partial elasticities of
demand. The second relates to an optimal combination of advertising media.

6.4.1 Partial Elasticities


As we have seen, the elasticity is the percentage change of one variable per
one-percent change in some other variable. As such, this measure of response
is valuable precisely because it is unit-free. We need not know the units of
either variable in order to make a meaningful statement about how responsive
one variable’s value is to the other variable’s value. Furthermore, when the
elasticity in question is the price elasticity of demand, we have established an
important relationship between elasticity of demand and marginal revenue,
CHAPTER 6. DIFFERENTIATION II 135

that M R = P · (1 + 1/Ep), where Ep is the price elasticity of demand, a


negative number.
This section extends the analysis of elasticity to include cross-price elastic-
ities and income elasticity. Regarding the latter, we establish an important
feature: depending on whether this elasticity is greater than 1 or less than 1,
consumers will spend a larger or smaller share of income on the good if their
incomes increase.
Consider the general expression for a demand curve, qa = f (pa, pb, m), where
qa is the quantity of Good A, pa is the price of Good A, pb is the price
of Good B, and m is money income. We define the elasticities as follows:
Eaa = ∂qa/∂pa · pa/qa, Eab = ∂qa/∂pb · pb/qa , and Eam = ∂qa/∂m · m/qa.
The first of this set is the “own-price” elasticity of demand, the second is the
cross-price elasticity, and the third is the income elasticity. If Eaa < −1, the
demand for this good is price elastic; if 0 < Eaa < −1, the demand is price
inelastic.
If Eab > 0, then Good B is a substitute for Good A; otherwise the two
goods or complements (or unrelated, if Eab = 0). Finally, if Eam > 1, then
the budget share of Good A moves in the same direction as income (such
goods are sometimes called superior goods). If 0 < Eam < 1, the quantity
purchased of Good A is directly related to money income (such goods are
normal goods; superior goods are also normal goods). If Eam < 0, Good A
is an inferior good.
We illustrate these relationships with two demand curves. One is linear
(constant slope) and the other exhibits constant elasticity. First, consider a
linear case in which the quantity of Good x (qx) is a function of the prices of
Goods x and y (px and py), money income (m), and the number of potential
buyers: qx = b0 + b1 · px + b2 · p + b3 · m + b4 · n.
A detailed interpretation of this expression would require a great deal of
information: physical units of x and y, the time period length, the monetary
unit in which prices and income are stated, and the relevant measure of
the number of potential buyers. Suppose, for example, qx is the number of
riders on a regional transportation system, measured in thousands of riders
per week, px is the fare per ride, py is the per-mile cost of operating a
private automobile, m is average per-family income in the region, and n is
the number of people living within a specified distance of a station.
CHAPTER 6. DIFFERENTIATION II 136

The exhibit below shows a stylized linear demand function for a transit sys-
tem’s services. The list named paramsEst is a hypothetical set of estimated
values for the parameters. The first output line shows the general expression
for this demand function. The second output line shows the expression given
the estimated parameters. Also, it shows the number of rides per period
(3950.0) that is estimated, given the values of the independent variables.7

The bottom three lines show the implied elasticities, along with the calcula-
tions required to generate them.
Supposing that qx is stated in 1000s of riders per period, interpret the values
in the fourth input line as follows: px, the per-ride fare is $4; py, $0.50, is the
per-mile cost of operating a private automobile; m, $40,000, is annual per-
household income in the relevant area; and n, 400,000, is the number of people
living within a defined distance from the transit line. From the estimated
linear relationship we can determine that the demand curve slopes downward,
that transit rides and automobile rides are substitutes, that transit rides are
an inferior good, and that adding 1 person to the area generates 2 more rides
per period.
Expressing these values in terms of elasticities offers a number of advantages.
One is that doing so makes comparison with similar analyses done by others is
easier. Such comparisons can indicate whether the study has been conducted
in a proper fashion (best available set of measures for variables and the
proper set of variables in the model, for example). Also, elasticities provide
some relevant information more directly than the original coefficients. In
particular, −1 < Exx < 0 implies that the transit authority could increase
7
The values used here are stylized, but they are based on research by one of the authors
of the original edition of this text. See Ostrosky and Kuhn, p. 160.
CHAPTER 6. DIFFERENTIATION II 137

revenues by raising the fare. That Exy = 0.038 indicates a weak relationship
between automobile operation cost and ridership.
Next consider a constant-elasticity demand function, qx = A · pxExx · py Exy ·
mExm (other variables could be included).8 At all values of the independent
variables, the elasticities are the same. This constancy of the elasticities
implies that the slopes change as variables’ values change, as Figure 6.5
demonstrates.
To determine the slopes, take the first partial derivatives of the qx function.
To use Maxima, apply this list of commands:
[diff(qx(px,py,m,A,Exx,Exy,Exm),px),
diff(qx(px,py,m,A,Exx,Exy,Exm),py),
diff(qx(px,py,m,A,Exx,Exy,Exm),m)].
The resulting expressions are in the following table.
dqx/dpx = Exx mExm px Exx −1 py Exy A
 
dqx/dpy = Exy mExm px Exx py Exy−1 A 
dqx/dm = Exm mExm−1 px Exx py Exy A
A little manipulation of the results above reveals that the slopes are these:
dqx/dpx = Exx·qx/px, dqx/dpy = Exy ·qx/py, and dqx/dm = Exm·qx/m.
The next table shows qx values for selected combinations of prices and in-
come, given these parameters: A = 0.012, Exx = −1.5, Exy = 0.75, and
Exm = 1.2. Comparing the quantity in the second row with that in the
first row allows computation of the arc own-price elasticity. Likewise, the
values in the third and fourth rows provide the values for calculating the
arc cross-price elasticity and the arc income elasticity. The accompanying
workbook shows the computations using the commonly-used midpoint for-
mula. The arc elasticity values are, respectively, -1.475 (compared to the
point elasticity of -1.5), 0.751 (compared to the point elasticity of 0.75), and
1.19 (compared to the point elasticity of 1.2).
prices and income quantity
py = 4, m = 100, px = 2 3.014263717811496
py = 4, m = 100, px = 3 1.640757346405055
py = 5, m = 100, px = 2 3.563393273053673
py = 4, m = 150, px = 2 4.903325869367984
8
A is a scaling factor. Mathematically, it is the value of qx when all independent
variables equal 1. Given that physical, temporal, and monetary units can be selected at
will, A’s value can be selected for convenience.
CHAPTER 6. DIFFERENTIATION II 138

Figure 6.5: Views of a constant-elasticity demand function


CHAPTER 6. DIFFERENTIATION II 139

6.4.2 Selection of Advertising Media


For a second example, consider a hypothetical firm Eddie’s Electronics Em-
porium (EEE with a fixed budget of B0 dollars that it is willing to spend
on advertising in either of two media, television (T V ) or newspapers (N ).
The price of each unit of television advertising and the price of each unit
of newspaper advertising are pt and pn. The firm can make any select any
combination of television and newspaper spots that do not violate the budget
constraint, given by B0 ≥ pt · T V + pn · N .
Assume that the firm chooses to spend all the B0 dollars at its disposal. Then
B0 = pt · T V + pn · N . This constraint can be phrased in terms of the number
of newspaper ads, given the number of TV ads: N = B0/pn–pt/pn · T V , so
the slope is dN/dT V = −pt/pn.
The effect of advertising is governed by a function S = F (T V, N ), where
S is per-period sales. We expect that adding to either type of advertising
increases sales: fT V = ∂S/∂T V > 0 and fN = ∂S/∂N > 0. Furthermore, we
expect diminishing returns to either of the advertising media: ∂ 2 S/∂T V 2 < 0
and ∂ 2 S/∂N 2 < 0.
The total differential of the advertising sales function is dS = ∂S/∂T V ·dT V +
∂S/∂N · dN . Define an “ isosales curve” as the locus of all combinations of
television and newspaper advertising that yield a specified constant level of
sales. That is, an isosales curve is given by the equation S0 = f (T V, N ).
Along any isosales curve, the change in sales is 0, and we can therefore
rewrite the differential as 0 = ∂S/∂T V · dT V + ∂S/∂N · dN , so the slope of
the isosales curve is –dN/dT V = fT V /fn . We establish below that the firm’s
optimal allocation of its advertising budget occurs when the combination of
TV and N satisfied this condition: pt/pn = fT V /fn .
EEE budgets $10,000 per period to advertise its sales and repair services.
Each TV spot costs $1000, and a spot in the local newspaper costs $100.
The function that relates sales to advertising is S = 10000 · T V 0 .2 · N 0 .6.
The display below shows expressions for the sales function and the advertising
budget (first output list). It also shows the results of setting pt/pn = fT V /fn .
The first result is that, for this pair of expressions, T V and N are used in
a fixed ratio. The second result is that T V = 5/2, (implying that N =
30 · 5/2 = 75.
CHAPTER 6. DIFFERENTIATION II 140

Figure 6.6: Budget line, isosales curves, and optimal mix

Figure 6.6 shows the highest sales level that EEE can achieve with B =
10000. It also shows the budget line, along with three isosales lines. One
of the isosales lines shows various combinations of T V and N that would
yield S = $160, 190. Only one of these combinations can be achieved with
B = $10000. The second, lower isosales line is for S = $12000. Either
of the intersections of this line with the budget line would consistent with
B = $10000, but would be inefficient. The third isosales line, for S = $20000,
cannot be attained with this budget.
CHAPTER 6. DIFFERENTIATION II 141

At the northwest point of intersection of the isosales line S = $200000 ,


fT V /fN > pt/pn, so fT V /pt > fN /pn. This implies that the increase in sales
from increasing TV expenditures by a small amount exceeds the decrease in
sales from reducing newspaper spending by a small amount. A rightward
movement along the budget line increases sales, until the tangency point is
reached. At that combination, the marginal sales per dollar is the same for
both media.
Consider one aspect of the sales function, that it is homogeneous. In general,
if g(x, y) is homogeneous of degree k, then this is true: g(k · x, k · y) = k n ·
g(x, y), where k is a positive value and n measures the degree of homogeneity.
Apply this to the sales function: (k · N )0 .2 · (k · T V )0 .6 = k 0 .8 · N 0 .2 · T V 0 .6 =
k 0 .8 · S. Therefore, our sales function is homogeneous, degree 0.8. We will
encounter homogeneity in later analysis of production. Look at Figure 6.6:
The three isosales lines all have the same slope along any ray (constant
N/T V ). This is an important aspect of homogeneous functions.
Suppose that we double T V and N . As a result, as the next table shows,
sales increase to $278,900, the second entry. The third entry shows that
20.8 · $160190 = $278900. Here, k = 2. If we start from column 2 and let
k = 1/2 , then the fourth entry equals the second entry times (1/2)0.8 .
 
S(5/2, 75) S(5, 150) S(5/2, 75) · 20.8 S(5, 150) · (1/2)0.8
1.6019 · 105 2.789 · 105 2.789 · 105 1.6019 · 105

Before leaving this illustrative example, we consider some managerial aspects


of this analysis. To say that sales are related to a promotional bundle in a
specific way is not to say that the firm knows this relationship. It does not.
Therefore, we are not predicting that EEE will employ precisely this com-
bination. Rather, this analysis offer a normative (prescriptive) framework.
EEE will gain the maximum possible benefit from spending on promotion if
it happens to employ this combination. In order to gain from is promotional
spending, EEE should be considering the marginal gain from each advertising
medium. If ∂S/∂T V > ∂S/∂N , then EEE will gain by moving some funds
from newspaper ads to TV ads; the reverse is true if the inequality sign is
reversed.
EEE’s ignorance is one reason that the precise values of T V and N are
unlikely to be employed. A second reason is lumpiness. While T V = 5/2
might not be impossible (5 ads every two weeks, for example), divisibility is
CHAPTER 6. DIFFERENTIATION II 142

limited. Even so, the prescriptive statement in the preceding paragraph is


still correct, even though the equality of marginal returns per dollar spend
probably cannot be exactly equalized.
This analysis could be extended to determine the conditions that must hold
for EEE to have budgeted the profit-maximizing amount of money to pro-
motion.

6.4.3 Taxation in Competitive Markets


One of the first applications of the model of competitive markets is a demon-
stration that the incidence of either an excise (per-unit) tax or an ad valorem
(per-dollar) tax depends on demand an supply conditions, not on the nominal
incidence of the tax. The accompanying workbook illustrates this result with
linear and constant-elasticity demand and supply curves. Here, we establish
the conditions that determine how much of an excise tax falls on buyers and
how much on sellers.9
Using the inverse demand and supply curves facilitates this analysis. Let
pd = pd(x) and ps = ps(x) where pd and ps indicate the heights of the
demand and supply curve at each output rate x. We assume that both
curves are monotonic and that d(pd)/dx < 0 and d(ps)/dx ≥ 0. We define
an implicit function F (x, t) = pd(x) − ps(x) − t = 0. That is, the per-unit
tax is a “wedge” between the price that consumers pay and the price that
sellers receive. The Implicit Function Rule implies that10
dx Ft 1
=− = .
dt Fx pdx − psx

Our interest is in d(pd)/dt and d(ps)/dt. Multiplying dx/dt by d(pd)/dx =


pdx yields an expression for dp/dt,

d(pd) pdx
= .
dt pdx − psx
In terms of economic impacts, this expression is the change in the price that
buyers pay, given the quantity that is determined above. We could equally
9
We follow Bishop[2].
10
The notation pdx and psx refers to d(pd)/dx and d(ps)/dx.
CHAPTER 6. DIFFERENTIATION II 143

well have multiplied by d(ps)/dx) with the resulting expression being

d(ps) psx
= .
dt pdx − psx
This is the change in the price that sellers receive.
The expressions above indicate the signs of the prices in the presence of a tax.
Commonly, these relationships are expressed in terms of elasticities. For an
price elasticity, E, the value is E = (dx/dp) · (p/x) so that dp/dx = p(/E · x).
Making these substitutions for the supply curve and the demand curve leads
to these relationships (copied from wxMaxima):

Eps Epd
, .
Eps − Epd Eps − Epd
The first term is the effect of the tax on the price that buyers pay, and the
second shows the effect on the price that sellers receive. The ratio of these
two, Eps/Epd, the ratio of the fraction of the tax that is passed to buyers
to the fraction that is passed to sellers.
The difference between the price paid and the price received is 1 · t. Confirm
that subtracting the effect of the tax on the price sellers receive from the
effect on the price that buyers pay yields 1. Also, divide the first term by
the second to confirm that the ratio of the two effects is Eps/Epd.11

6.4.4 Production Theory


The illustration above that addresses the effects of advertising on sales is
mathematically much like production theory. This section develops more
carefully some salient aspects of production theory. It begins with concept
of a production function. A production function defines the maximum output
that a firm can obtain from any given set of inputs that it uses. Assume a
production function of the form Q = F (L, K), where Q = output, L = units
of labor, and K = units of capital. The marginal product of labor (M P L)
measures the change in output that results when a very small change is
11
The accompanying workbook treats taxation in more detail. It considers two specific
functional forms for the demand and supply curves and illustrates the points that this
section sketches.
CHAPTER 6. DIFFERENTIATION II 144

made in the amount of labor being used, the amount of capital being held
constant: M P L = ∂Q/∂L = fL . Likewise, the marginal product of capital
is M P K = ∂Q/∂K = fK .

Optimization

The preceding illustration depicts a firm as having a fixed budget and seeking
to generate maximum sales based on an optimal use of two advertising media.
More generally, we can represent firms as setting out to gain the largest
output for a given cost by selecting the ideal mix of inputs, given a set cost.
Not surprisingly, if the firm uses two inputs, K and L, for which the unit
costs are r and w, then the firm will attain the result if it uses the mix that
spends the allocated cost C = w · L + r · K on a combination such that
fK /fL = r/w.
Suppose, however, that the firm wishes to select a quantity and then find the
lowest-cost input mix consistent with that quantity. This problem is called
the dual to the one above. Figure 6.7 shows an isoquant (for Q = 200 units
in this illustration; see the accompanying workbook) and three isocost lines.
Line T C1 represents a cost that is inconsistent with producing the specified
output level. Line T C2 represents a cost at which either of two points (where
the isoquant intersects T C2) is consistent with the required output level but
at an unnecessarily high cost.
Finally, the isocost line T C0 allows the output level to be produced, given
that labor and capital are combined as indicated by the point of tangency of
T C0 and the isoquant (at approximately L = 25 and K = 15. As in the case
of maximizing output subject to a cost constraint, minimizing cost subject
to an output constraint requires that the ratio of marginal products equal
the ratio of input prices. The negative of the slope of the isoquant is called
the marginal rate of technical substitution (mrts).

Homogeneous and Homothetic Functions

As we have seen a function is homogeneous if changing all inputs in the same


proportion, k , so that their ratio remains the same, causes output to increase
by k n , where n is the degree of homogeneity. Functions with the special
condition that n = 1 are said to exhibit linear homogeneity. (Beware: The
CHAPTER 6. DIFFERENTIATION II 145

Figure 6.7: An isoquant and three isocost lines


CHAPTER 6. DIFFERENTIATION II 146

functions themselves need not be linear.) The following function illustrates


important aspects of homogeneity: Q = A · La · K b . This function is often
called the Cobb-Douglas production function, and we denote it as such.12 For
this function, A · (k · L)a · (k · K)b = A · k a+b · Q. Thus, changing L and K
by a factor k causes Q to change by k a+b .
Figure 6.8 shows three sets of isoquants. Each set depicts the production
of the following quantities: 5, 10, and 15. The titles indicate the degree of
returns to scale. The first function is Q = L.3 · K .2 , which implies n < 1. In
production, this situation is called decreasing returns to scale. The second
production function is Q = L.6 · K .4 , so that n = 1. This production function
is said to exhibit constant returns to scale. Finally, the third function is
Q = L.75 · K .5 , so that n = 1.25 and the function exhibits increasing returns
to scale. In all cases a/b = 3/2.
In all cases, a ray along which the ratio of K and L is the same on each of
three isoquants is the same is added. Each ray’s slope is 2. That is K = 2∗L.
The ray is just for comparison, and any common slope will do. In all cases,
the isoquants are “parallel displacements” of each other: that is, along any
ray like the one shown, the mrts is the same on all isoquants. When n = 0.6,
the isoquants become increasingly farther apart, given constant increments
to output.
Compare the length of line segments along the ray between the first two
isoquants with the length between the second isoquants. With decreasing
returns to scale, the second segment is longer. With constant returns to
scale, the two segments are of equal length. With increasing returns to scale,
the second segment is shorter. A less obvious point is that the isoquant
slopes are the same at all nine points of intersection.
Homogeneous functions are a subset of another important class of functions,
those that are homothetic. A function is homothetic if the mrts for any two
inputs is the same for all values of the inputs as long as they are used in the
same proportion. Such a function need not be homogeneous. The exhibit
below shows a function that is not homogeneous but is homothetic. This
exhibit show that this function is not homogeneous, because no term of the
form k n · Q can be extracted.13
12
Strictly the Cobb-Douglas function requires that a+b = 1. We extend the terminology
to include the slightly more flexible expression.
13
The accompanying workbook generalizes this function and goes into some detail about
CHAPTER 6. DIFFERENTIATION II 147

Figure 6.8: Varying returns to scale

For small values of L and K, Q increases as L and K employment increases


in a fixed proportion. Eventually, however, the second term dominates and
output decreases. Thus, the function is not homogeneous. The second out-
put line confirms this feature of the function. As noted in the preceding
paragraph, the expression cannot be factored so that the output that results
from replacing the original and values with a multiple of those values leads
to an expression that is proportional to the original value.
This function is, however, homothetic. As the fourth column below shows,
the slope of the isoquant equals (3/2) · (K/L). That is, the isoquant’s slope
depends only on the ratio of inputs and is not affected by scale.
 
MP L MP K M PL/M P K =  mrts
3 4 3 2
 6K 54 L 35 −48K 52 32L 35 −4K 25 L 65 K 5 6K 5 L 5 −48K 5
3K

− 2 3 − 2  3 2 6 2L
5L 5 5K 5 L5 32L 5 −4K 5 L 5

the implication of a non-homogeneous function like this one.


CHAPTER 6. DIFFERENTIATION II 148

A production function need not be homothetic. Consider the function

Q (L, K) := Lb log (K)

where 0 < b < 1. We change both inputs by a factor k, using this com-
mand: Q(k*L, k*K). The resulting expression is k b (log (K) + log (k)) Lb .
We cannot derive an expression of the form k n · Q)
The table shows the marginal products of labor and capital and the marginal
rate of technical substitution. We cannot derive an expression in the form
mrts = m · (K/L), so this function is not homothetic.
 
MP L MPK mrts
Lb bK log (K)
b log (K) Lb−1 K L

6.4.5 Homogeneity and Euler’s Theorem


Euler’s theorem states that for a homogeneous function f (x1, x2, . . . , xn),
the relationship between the partial derivatives and the function’s value is as
follows: x1 · fx1 + x2 · fx2 + · · · + fxn = n · f (x1, x2, . . . , xn). This theorem
provides important insights regarding economic issues. For discussion, use
Q = f (L, K) as the relevant production function. First, consider the question
of how to account for output. In general, f (L, K) is such that we cannot
uniquely assign part of the output to each input.
Suppose, however, that a linear homogeneous function provides a good ap-
proximation to reality. Then the following is true: L · M P L + K · M P K = Q.
This provides an accounting framework for thinking about how to attribute
output to the economy’s resources. It does not say anything about incomes.
Suppose, however, that all firms are price-takers in both input and output
markets. Then each firm maximizes profits by hiring labor and capital in
amounts such that each resource’s value marginal product (V M P ) equals
the unit cost of employing the resource. Recall that V M P L = p · M P L
and V M P K = p · M P K, where p is the price for which the firm sells out-
put. At the margin (given the optimal input combination), w = V M P L,
so M P L = w/p, the “real” wage. Likewise, M P K = r/p. These equalities
imply that L · w/p + K · r/p = Q so that L · w + K · r = p · Q. Unlike
the expression in the preceding paragraph, this is a positive statement. It
says that given the conditions above the value of the total output, p · Q, will
CHAPTER 6. DIFFERENTIATION II 149

be distributed by market forces in a specific way, a way that is called the


marginal productivity theory of income distribution.
This theory of distribution might provide useful insights into the general
working of an economy that is largely based on free markets. For various
reasons, outcomes at firm or industry level can deviate from this result.
Suppose that production functions of the firms in an industry exhibit linear
homogeneity, but that each firm faces a downward-sloping demand curve for
its product.
In this case, a firm’s marginal revenue is mr = p · (1 + 1/E), where p is the
price that the firm charges and E, a negative number, is the elasticity of
demand at price. This firm will employ L and K in quantities such that the
marginal revenue product, not the value of the marginal product, equals the
unit cost of the input. That is, employment will occur at levels such that
p·(1+1/E)·M P L = w and p·(1+1/E)·M P K = r. This result implies that
L · w + K · r = p · (1 + 1/E) · Q. Suppose that E = −5.14 Then the amount
paid to L and K is four-fifths of the total product cost. The remainder is
profit.
Second, suppose that a firm is large enough relative to one or more of its input
markets that its employment decision affects the market input price. The
firm has what is loosely called “monopsony power.” In this case the firm’s
marginal input cost ismic = w · (1 + 1/Esupply), where is the price elasticity
of supply to the firm and w is the input price (for a type of labor). Such a
firm will employ this type of labor in an amount such that V M P L = mic.
The variable mic is the marginal input cost. It is, roughly, the amount by
which total cost of the relevant input changes per one-unit change in the
amount of the input the firm employs. In this case, as in the preceding one,
the payments to L and K do not sum to the value of the product, so the
firm receives a profit.
Third, production functions need not homogeneous of degree 1. Hammock
and Mixon show that in a competitive industry all firms except the firm (or
firms) at the margin operate in the area of decreasing returns to scale–the
degree of homogeneity is less than 1. Suppose that n = 0.9 is the degree of
homogeneity. In that case w ·L+r ·K = 0.9·p·Q. The remaining ten percent
14
E is the price elasticity of demand for the firm’s product, not for the industry product.
Unless the industry is a perfect monopoly, E will be a larger negative number than the
elasticity of demand for the industry product.
CHAPTER 6. DIFFERENTIATION II 150

of the revenue is the profit that accrues to the owner of the inframarginal
firm or to the owner of some specialized factor (like location) that makes the
firm’s cost lower that the cost incurred by the marginal firm(s).
Finally, of course, production functions need not be homogeneous at all.
Euler’s theorem applies only to homogeneous functions. Having said this,
however, be aware that being descriptively inaccurate is not the same as
being useless. If production at an aggregate level is well approximated by
a linear homogenous production function, then the marginal productivity
theory of distribution might provide the best model with which to start
thinking about this important issue.

6.4.6 Two More Elasticities


We have encountered four elasticity measures: price elasticity of demand,
cross-price elasticity of demand, income elasticity of demand, and price elas-
ticity of supply (of an input). This section introduces two elasticity measures
that are often used in developing and applying production theory. The first
is the elasticity of substitution, which relates to the ease with which firms
can change their input mix in response to changes in relative input prices.
Second, a set of and output elasticity values indicate how responsive output
is to changes in the individual inputs, holding constant the employment of
the other inputs.

The Elasticity of Substitution

The elasticity of substitution, which we denote Esub, can be thought of as a


measure of an isoquant’s slope–more accurately, as a measure of its curvature.
Figure 6.9 shows two “unit isoquants,” isoquants for which the quantity
produced is one unit. The family of production functions that generated
these is homogeneous, so all other isoquants are parallel to these. These
are from two different production functions. In one case, Esub = 2 and the
isoquant exhibits moderate curvature. For the other, Esub = 0.2 and the
curvature is quite pronounced.
To be more precise, we define the elasticity of substitution as the ratio of
the percentage change in the K/L to the percentage change in the mrts,
the slope of the isoquant. For the first case in Figure 6.9, the percentage
CHAPTER 6. DIFFERENTIATION II 151

Figure 6.9: Two unit isoquants

change in K/L is twice the percentage change mrts. For the second case,
the percentage change in K/L is only 0.2 that of the change in mrts.
The elasticity of substitution takes on special importance when we recall that
optimizing firms select the input combinations that equate their mrts with
the ratio of the relative input prices (more generally, their relative marginal
input costs–see the preceding section). Thus, we can redefine the elasticity
of substitution as
d K
 w
L
Esub = w
· Kr .
d r L
This is the percentage change that the firm makes in its input ratio per
one-percent change in the ratio of the input prices. Given that w · L and
r · K are the costs of the inputs, consider the implications of Esub’s value on
how total cost is distributed. If Esub = 1, then a one-percent change in the
ratio of input costs is exactly offset by a one-percent change in the opposite
direction, so that the share of cost that goes to each input type remains the
same.
If, however, Esub 6= 1, then a change affects income distribution between the
inputs. Specifically, if Esub > 1, then a change in w/r causes labor’s share
to move in the opposite direction. Likewise, if Esub < 1 , then a change w/r
causes labor’s share to move in the same direction that w/r changes. Among
other implications, the value of a firm’s Esub could affect a union’s ability
CHAPTER 6. DIFFERENTIATION II 152

to negotiate for higher wages.


The Cobb-Douglas production function, introduced above, exhibits Esub = 1
for all values of L and K. Indeed, one reason that the Cobb-Douglas produc-
tion function developed is that Douglas, the economist, sought a function for
which labor’s share remained constant.15 His collaborator, Cobb (a mathe-
matician), pointed Douglas to the function that bears their names. For this
function, Q = A · La · K 1−a . Therefore, M P L = a · A · La · K 1−a = a · (Q/L).
Likewise, M P K = (1 − a) · (Q/L). Together, these imply that
w MP L a K
= = · .
r MP K 1−a L
We leave proof that, therefore, Esub = 1 as an exercise.
In part, because of the implication of for income shares, economists have de-
veloped expressions for production function that allow to take values other
than 1. The most widely-used function is the Constant Elasticity of Substi-
tution (CES) function:
 −1/b
a 1−a
A· + .
Lb Kb
The table below shows this function, the two marginal product functions,
and the ratio of these marginal product functions. This ratio, which is the
mrts, can be restated as
 1+b
1 K
mrts = · .
1−a L
 A

CES f unction, Q = 1
 ( Lab + 1−a
Kb
)b 
 −b−1 a
 1
1−a − b −1


 MP L = aA L Lb
+ Kb 

1
−b−1 a 1−a − b −1 

MP K = (1 − a) A K + Kb

 Lb 
a K b+1 L−b−1
mrts = 1−a

To determine that the elasticity of substitution is is quite direct. In the


exhibit below, the expression mrtsCES restates mrts in terms of a variable
15
Cobb’s reading of labor history indicated to him that this constancy had persisted,
at least approximately, for a long period of time.
CHAPTER 6. DIFFERENTIATION II 153

named α, where α = K/L. It shows that


α
mrts = · k 1+b ,
1−α
so that mrts is a monotonic function of α. The second output line shows
that
d mrts α
· = 1 + b.
dα mrts
The inverse function rule implies , therefore, that

d(K/L) mrts 1
Esub = · = .
d mrts K/L 1+b

Refer to Figure 6.9 to recall the implications of the elasticity for the shape
of isoquants for the CES production function. Also, note that as b → 0, the
elasticity of substitution approaches 1, the Cobb-Douglas value.

Output Elasticity

Output elasticities measure the proportionate response in total output that


is elicited by proportionate changes in the quantity of one input, all other
inputs being held constant. Denote a general production function as Q =
f (X1, X2, . . . , Xn). The output elasticity with respect to the ith input is
defined as (∂Q/∂Xi) · (Xi/Q). It is useful to rephrase this expression as
(∂Q/∂Xi)/(Q/Xi), which is the ratio of the marginal product of this input
to its average product.
The Maxima output below this paragraph shows the result of this computa-
tion to the CES production function. The results are not in a form that is
CHAPTER 6. DIFFERENTIATION II 154
 b
a 1−a A

easily interpreted. Note, however, that Lb
+ Kb
= Q
can be deduced
from the production function.
a 1−a
a 1−a
, a 1−a

Lb Lb
+ Kb
Kb L b + K b

Therefore, we can phrase the two entries as


 b
a a Q
= b
(L · A/Q)b A L

and
 b
1−a 1−a Q
b
= .
(K · A/Q) Ab K

As b approaches zero (Esub → 1), these two output elasticity values approach
a and 1 − a as the CES production function approaches the Cobb-Douglas
function, a conclusion that we approach below.

L’Hopital’s Rule, CES, and Cobb-Douglas

Entering the command limit(Q(L,K,A,a,b, b, 0) into Maxima should re-


turn a result that can be phrased as the Cobb-Douglas function, but this
aspect of Maxima is not fully reliable (in the author’s experience). We can,
however, apply L’Hopital’s Rule and let Maxima do some of the work for us.
L’Hopital’s Rule: Let one of the following be true for two functions, f (x) and
g(x) both functions have either 0, ∞, or −∞ as a limit as both approach a
command value. a. That is,

lim f (x)/g(x) = 0/0,


x→a

lim f (x)/g(x) = ∞/∞,


x→a
or
lim f (x)/g(x) = −∞/ − ∞.
x→a

The value of a can be any real number, ∞ or −∞.


CHAPTER 6. DIFFERENTIATION II 155

In any of these cases,

lim f (x)/g(x) = lim f 0 (x)/g 0 (x).


x→a x→a

The terms f 0 (x) and g 0 (x) are the first derivatives of the two functions.
Rather than working with the CES function, we consider its logarithm, which
Maxima provides:
log Lab + 1−a

K b
log (A) − .
b
The second term is a ratio of two terms, both of which approach ∞ as b → ∞.
We can see that if g(b) = b, then g 0 (b) = 1 for all values of b, so for all b, the
limit of g 0 (b) = 1. Therefore we can focus on f 0 (b). The derivative of
 
a 1−a
−log +
Lb Kb

is
a · K b · log (L) + (1 − a) · log (K) · Lb
− .
(a − 1) · Lb − a · K b
As b → 0, this expression approaches

a · log (L) − a · log (K) + log (K) .

We know that the limiting value of log(A), a constant, is just log(A)). There-
fore the limiting value of

log Lab + 1−a



K b
log (A) −
b
is
log(A) + a · log (L) − a · log (K) + log (K) .
Taking the anti-logarithm of this expression (reversing the effect of having
taken the logarithm above) yields (after a couple of manipulations) this re-
sult: A · K 1−a · La , the Cobb-Douglas production function. Therefore, this
function is the limiting case of the CES function when b = 0, which corre-
sponds to Esub = 1.
CHAPTER 6. DIFFERENTIATION II 156

6.5 Questions and Problems


1. (i) For each function below, determine ∂z/∂x, ∂z/∂y, (∂ 2 z)/(∂x2 ),
(∂ 2 z)/(∂y 2 ), and (∂ 2 z)/(∂x∂y), if they exist. Compare your answers
with those produced by Maxima.
a. z = 6 · x2 · y b. z = 0.3 · x4 /y 2 c. z = 0.3 · xx /y x
x+y x+1 2
d. z = y e. z = e ·y f. z = loge 5 · x4 · y
g. z = loge xy h. z = A · xa · y 1−a
(ii) Assuming that x and y might be inputs and z might be output,
determine which of these functions, if any, are plausible candidates to
represent production.
2. One of the earliest applications of regression analysis to economic be-
havior, by Richard Stone, estimated that the demand for beer in Great
Britain pre-World War II was approximately Q = 180 · Y −0.02 · P −1.04 ·
R0.94 , where Q is quantity, Y is aggregate real income, P is the mean
retail price of beer, and R is the mean retail price of other goods and
services.
a. Determine the estimated income, own-price, and cross-price elastic-
ities of demand.
b. If the price of beer rises in a given year, does this estimated demand
curve predict that total spending on beer (which equals total rev-
enue of beer sellers) will rise or fall? If the average price of a
pint was £0.1 during the sample period, what is the estimated
marginal revenue from beer sales?
c. Given this demand specification, one can deteremine marginal rev-
enue but not total revenue based on the information given above.
Explain.
3. Apex Electronics produces a generic hand calculator that is sold under
various store brands. The price at which it can sell is $2.50 per unit.
AE’s total cost is T C = $(500 + Q/4 + Q2 /5000). Determine the
quantity that AE should sell and its profit level.
4. With continuous compounding the present value PV of $1 to be received
t years from now is given by P V = $1 · e−r·t . Determine ∂P V /∂t and
∂P V /∂r. Evaluate both expressions given that the initial values are
t = 15 and r = 0.04.
CHAPTER 6. DIFFERENTIATION II 157

5. Suppose that the “production function” that relates performance on


an accounting examination (G) to hours spend studying (H) and intel-
ligence (I, however it might be measured) is
125
G= I
√ .
− 100 − H
%e + %e +1
a. Use Maxima to draw this function over the ranges 0 ≤ H ≤ 20 and
95 ≤ I ≤ 125.
b. Determine the expressions for ∂G/∂H and ∂G/∂I. Confirm that
both “inputs” exhibit diminishing marginal returns.
c. Determine the values of G and the two “marginal products” when
H = 15 and I = 120.
d. Confirm that GHI = GIH and that this result is a quite small pos-
itive value. Explain what a positive value means in this setting.
Does this seem plausible to you? Explain.
Chapter 7

Optimization: Maximization,
Minimization, and Constraints

Economic analysis assumes that actors have objective functions. These func-
tions include the utility function that represents a consumer’s preferences and
the profit function that represents the outcome of a firm’s actions. Further-
more, this analysis often begins with the assumption that the actors attempt
to maximize (or, for some functions, minimize) the value of these functions.
The actions of the decision-maker are nearly always constrained by limita-
tions such as the amount of money or time (or both), or some minimum
acceptable level of performance or output. When such constraints apply, the
actor generally cannot achieve the maximum value of the objective function.
Rather, the actor is assumed to optimize. That is, the actor finds the set
of options that yield the maximum (or minimum) attainable value of the
function, subject to the constraint.
We have inserted an important and somewhat controversial behavioral as-
sumption here, that the actor is an optimizing agent. This is the standard
approach for neoclassical economics. This approach can be thought of as
either positive or normative. That is, we can interpret the results of the
model as behavior in which we should expect actors to engage (positive).
Alternatively, we may be interested in investigating the conditions that are
required for optimization and using those to prescribe behavior for someone
who is seeking to optimize in a particular setting (normative).
This chapter initially demonstrates how one may find the maximum or min-

158
CHAPTER 7. OPTIMIZATION 159

imum value(s) of a differentiable function. This finding is then generalized


to the case in which one wishes to find the maximum or minimum value(s)
of a differentiable function that subject to some constraint(s). This general-
ization provides an apparatus for dealing with a wide range of problems in
practical and theoretical situations.

7.1 Extreme Value(s): Functions of One Vari-


able
We begin by examining functions of a single variable. We specify conditions
under which such a function’s value is increasing, decreasing, or remaining
constant. Also, for a function that is changing in a given direction, we
examine changes in the rate of increase or decrease.
We have examined a number of functions that either increase or decrease
monotonically. Figure 7.1 shows two pairs of curves, representing the fol-
lowing functions over a range of x values: y = 4 + 2 · x and y = 4 − x, for
−5 ≤ x ≤ 5 and y = e0.5·x and y = e−0.5·x for −3 ≤ x ≤ 3. In all four cases,
dy/dx has the same sign for all x values. For each linear function,dy/dx is
constant, and the values of y can increase or decrease without limit. For the
exponential functions, dy/dx depends on x, and these functions can extend
indefinitely in one direction only, with y = 0 as their lower limit.
Figure 7.2 shows two functions and reports their expressions. Both of these
parabolic functions are non-monotonic. The first function achieves a maxi-
mum value at x = 5. For x < 5, dy/dx > 0; for x > 5, dy/dx < 0, and
for x = 5, dy/dx = 0. The second function achieves its minimum value at
x = 5. Now the first two statements regarding dy/dx are reversed, but at
x = 5, dy/dx = 0, as before. When a function has a single (local) extreme
value, establishing the value of x at which dy/dx = 0 reveals the value of x
that generates an extreme value of y. Whether that extreme is a maximum
value or a minimum value is indicated by the behavior of dy/dx for smaller
and larger values of x. We return to this point below.
Of course, many functions have more than one extreme value. The first of
the two graphs in Figure 7.3 is that of a cubic polynomial. This function
has a single local maximum value, when x is about 1. It has a single local
minimum value, when x is about 7. For this function, y increases without
CHAPTER 7. OPTIMIZATION 160

Figure 7.1: Two pairs of monotonic functions

Figure 7.2: Two parabolas


CHAPTER 7. OPTIMIZATION 161

Figure 7.3: Multiple extreme values

bound as x → ∞, and it decreases without bound as x → −∞. Therefore,


it has no global maximum or minimum value.
The second graph shows a function that combines a sine function with a
simple polynomial. This function has numerous local extreme values. Like
the cubic function, it has no global maximum: it increases without limit as
x becomes either very large or very small. It does, however, have a single
global minimum, at x u 5. In addition, it has a number of local extreme
values, both maxima and minima. The wxMaxima workbook for this chapter
provides more detail.

7.2 Inflection Points and Concavity


The first derivative, dy/dx, measures the slope (rate of change) of a curve
at a given point. It reflects whether the original function is increasing or
decreasing at that point. The second derivative, d2 y/dx2 , measures the rate
of change of the slope of the function f (x). The second derivative reflects
whether the function f (x) is increasing at an increasing (decreasing) rate or
decreasing at an increasing (decreasing) rate.1
1
A commonly used physical interpretation of the first and second derivatives relates
to a moving automobile. The first derivative of distance with respect to time measures
CHAPTER 7. OPTIMIZATION 162

We define concavity as follows: Take as given that dy/dx and d2 y/dx2 exist
for all x in some interval. Then, if d2 y/dx2 > 0 the curve (x) is said to
be concave upward at x. If d2 y/dx2 < 0, then the curve f (x) is said to be
concave downward at x.
Figure 7.4 shows the two graphs from Figure 7.2 and adds three tangent line
segments to each of the two graphs. The first graph illustrates the case in
which the curve of the function is concave downward. At x = 1, the first
point of tangency on the graph, the first derivative is positive and the second
derivative is negative. This means that the function is increasing, but at
a decreasing rate. At x = 9, dy/dx < 0 and d2 y/x2 < 0, indicating that
the function is decreasing at an increasing rate, that is, that the slope of
the function is becoming increasingly negative. Between these two values, at
x = 5, dy/dx = 0, and f (x) achieves its maximum value.2

Figure 7.4: Points of tangency

The second graph depicts the case where the curve of the function is concave
upward. At x = 1, dy/dx < 0, but d2 y/dx2 > 0. This implies that the
velocity (“speed”), while the second derivative of distance with respect to time measures
acceleration. An economic example relates to the general price level (say GDP Deflator
value): The inflation rate is the first derivative, and the annual rate at which inflation is
changing is the second derivative
2
For the first function, d2 y/dx2 = −2 for all x values. For the second function, the
corresponding value is 2.
CHAPTER 7. OPTIMIZATION 163

function is decreasing at a decreasing rate; that is, that the slope of the
function is becoming less negative. When x = 9, dy/dx > 0 and d2 y/dx2 > 0.
The function is now increasing at an increasing rate.
When the concavity of the function changes from downward to upward or
from upward to downward at a value of x, this point on the function is
called an inflection point or point of inflection. Figure 7.5 illustrates the
two different types of points of inflection. In the upper portion of the first
column, the point of inflection occurs where the concavity of the function
changes from downward to upward, at x = 20. The point of inflection in
the upper portion of the second column occurs where the concavity of the
function changes from upward to downward, again at x = 20.3

Figure 7.5: Inflection points

Points of inflection have definite implications for the first and second deriva-
tives. We can see in the lower portion of the first column that the point of
inflection is the minimum value of dy/dx when concavity is changing from
downward to upward. Analogously, the point of inflection in the lower por-
tion of the second column is the maximum value of dz/dx in the case in
x3 3
3
The two functions are 60 − x2 + 25 · x and − x30 + 2x2 − 15 · x.
CHAPTER 7. OPTIMIZATION 164

which the concavity of the function is changing from upward to downward.


Whether the concavity of the function is changing from downward to upward
or upward to downward, the second derivative of the function equals zero.
Hence, when x = 20, both d2 y/dx2 and d2 z/dx2 in Figure 7.5 equal zero. This
reflects the facts that function y is changing from increasing at a decreasing
rate to increasing at an increasing rate, and function z is changing from
increasing at an increasing rate to increasing at a decreasing rate.
We can make a more general statement about points where the concavity of
a function is changing. Given a differentiable function y = f (x), for which
dy/dx and d2 y/dx2 are also continuous, the following must hold. If the value
of the function d2 y/dx2 is changing from a negative value to a positive value,
or from a positive value to a negative value, then there must be one point at
which the value of the function d2 y/dx2 = 0.
A warning is in order. Given the function y = f (x) that has a point of
inflection at x = x0, then it must be true that d2 y/dx2 = 0 when x = x0.
The converse is not true. The fact that d2 y/dx2 = 0 when x = x0 is does
not guarantee the existence of a point of inflection when x = x0. Therefore
d2 y/dx2 = 0 is a necessary condition rather than a sufficient condition for
identifying a point of inflection. Consider a simple example: y = 4 · x. Then,
for all values of x, dy/dx = 4 and d2 y/dx2 = 0. The graph of y = 4 · x is a
straight line, however, so it contains no points of inflection.
Consider y = x4 . For this expression, dy/dx = 4 · x3 and d2 y/dx2 = 12 · x2 .
Solving 12·x2 = 0 yields x = 0. This does not imply, however, that a point of
inflection occurs at x = 0. A point of inflection exists only if the concavity of
the function changes from downward to upward or from upward to downward.
We can determine whether this is the case by evaluating d2 y/dx2 for x < 0
and for x > 0. When x < 0, 12 · x2 > 0. Likewise,when x > 0, 12 · x2 > 0.
Hence d2 y/dx2 > 0 both for values of x that are less than zero and for values
of x that are greater than zero. Therefore, the concavity of the function is not
changing. For a point of inflection to exist in this case, the sign of d2 y/dx2
must change when we go from values of x less than 0 to values greater than
0. Hence there is no point of inflection at x = 0.
CHAPTER 7. OPTIMIZATION 165

7.2.1 Finding Maxima and Minima


The value(s) of x for which f (x) attains a maximum or a minimum are re-
ferred to as extreme values. It is necessary, as suggested earlier, to distinguish
between absolute (or global) and relative (or local) extreme values. We begin
with four definitions. First, we define an absolute (global) maximum: Let
y = f (x) be a real-valued function defined on a set S of real numbers. Then
the function f (x) has an absolute maximum at x = x0 if f (x0) ≥ f (x) for all
x in S. The definition of an absolute minimum is similar: Let y = f (x) be a
real-valued function defined on a set S of real numbers. Then the function
f (x) has an absolute minimum at x = x0 if f (x0) ≤ f (x) for all x in S.
The definitions of a relative (local) maximum and a local minimum are also
similar. Given y = f (x) as above, the function f (x) has a relative (local)
maximum at x = x0 if f (x0) ≥ f (x) for all values of x in a neighborhood.
Finally, the function f (x) has a relative (local) minimum at x = x0 if f (x0) ≤
f (x) for all values of x in a neighborhood. A neighborhood of x0 is an interval
containing x0. Formally, a δ neighborhood of x0 is the interval (x0−δ, x0+δ).
Figure 7.6 shows three graphs that represent functions over a range of values
(not specified). The first of the three has a relative maximum value near the
center of the range of x values, which is also the absolute minimum value
within this range of values (but not necessarily a global maximum–the graph
does not indicate y’s behavior outside this range of values), and the same
absolute maximum value at either end of the range of x values. The second
function exhibits a relative momimum value near the center of the range of
x values, which is also the absolute maximum value in this range, and the
same absolute minimum value at either end of the range of x values. Again,
we cannot conclude how y behaves outside of this range from the graph.
The third graph exhibits the following behavior: as x increases in value: the
absolute minimum, a relative (local) maximum, a relative minimum, and the
absolute maximum for this range of x values. As before, the graph does not
allow us to infer that we have found a global maximum for all x values.
Two tests enable us to identify extreme points. One, the first derivative test,
is only a necessary condition for an extreme point. Even when satisfied, the
first derivative test does not guarantee that an extreme point exists. The
other, the second derivative test, when it accompanies a first derivative test,
is a sufficient condition for an extreme point. When this test is satisfied, an
CHAPTER 7. OPTIMIZATION 166

Figure 7.6: Local extreme values

extreme point exists. These tests do not say whether the extreme value is
absolute or relative. They can say that, at least, a relative extreme value has
been determined.

7.2.2 The First Derivative Test


The first derivative test consists of a set of three steps. First, given that the
first derivative of a function f (x) is df (x)/dx = 0 when x = x0, solve the
equation df (x0)/dx = 0 for its critical root(s). Next, examine each critical
root separately. If, within a given interval, df (x)/dx changes its sign at a
value x = x0, then an extreme point on f (x) has been identified.
The third step is to establish which of the following occurs. If df (x)/dx > 0
for x < x0 and df (x)/dx < 0 for x > x0, then a relative (local) maximum
exists at x = x0. Alternatively, if df (x)/dx < 0 for x < x0 and df (x)/dx >
0f orx > x0, then a relative (local) minimum exists at x = x0. Finally, if
df (x)/dx > 0 for x 6= x0 or if df (x)/dx < 0 for x 6= x0, then no relative
extreme point exists at x = x0.
CHAPTER 7. OPTIMIZATION 167

Consider this simple example: y = x2 , for which dy/dx = 2 · x. This ex-


pression has a single root, x0 = 0. For x > x0, dy/dx > 0, and for x < x0,
dy/dx < 0. We can, therefore, conclude that this function reaches a local
minimum value at x = x0 = 0.
A second example involves the two cubic polynomials defined in the table
below. The first polynomial’s derivative has two roots. To determine that
the first results in a local minimum, we determine that the derivative’s value
at x = 8.7197 is f(8.7197) = -0.898 and that f(8.9197) = 0.890. Thus f(x) is
decreasing for x < 8.8197 and increasing for x > 8.8197. We leave confirming
that the second root, x = 31.181 results in maximum value for f(x) as an
exercise.

 
Function Derivative Solution(s)
− 2·x3 + 8 · x2 − 110 · x − 2·x2 + 16 · x − 110 [x = 8.8197, x = 31.18]
15 5
x3 2 x2
30
− 2 · x + 40 · x 10
− 4 · x + 40 [x = 20.0]

The second polynomial’s derivative has a root at x = 20, but dg(x)/dx > 0
for x < 20 and also for x > 0. For example, at x = 19.8, dg(x)/dx = 0.001;
also, at x = 20.1, dg(x)/dx = 0.001. Thus, x = 20 corresponds to an
inflection point for g(x) = x3 /30–2 · x2 + 40 · x. Figure 7.7 confirms the
results that we have derived using information in the table above.
Exercise 7.1
For each expression below, find any extreme points that exist and determine
whether each such point is a relative maximum, a relative minimum, or a
point of inflection. Confirm your results with Maxima graphs.
1. y = x2 –4 · x + 16 2. y = x3 –6 · x2 + 9 · x 3.y = x · ex
4. y = x · (x − 1)2 5. y = (x − 1)3 + 8 6. y = x + 1/x
7. y = x 2 · x) − 2 · x 8. y = x /3–x + x + 1 9. y = x3
( 3 2

For each of the following, find the absolution maximum and/or minimum in
the designated intervals. Graph each function.
10. y = x2 , where −8 ≤ x ≤ 16
12. y = (25 − 3 · x)0.5 where 0 ≤ x ≤ 3
13. y = (x − 8)2 , where −2 ≤ x ≤ 4
14. y = 150 − 0.8 · x, where 0 ≤ x ≤ 10
14. (a) When an automobile travels s miles per hour, the cost per mile (in
s2 s
dollars) of operating the automobile is O = 5000 − 50 + 189
200
. At what speed
CHAPTER 7. OPTIMIZATION 168

Figure 7.7: Two extreme points and an inflection point

is the cost per hour minimized? (b) Suppose that the cost of the driver’s
time is added to this equation, and that this cost is 30/s. Determine the new
cost-minimizing speed.4
15. The demand curve for a firm’s product is q = 8–p, where q is the number
of units sold and p is the price per unit. (a) What price should the firm charge
if it chooses to maximize its total sales revenue (p·q)? (b) This demand curve
implies that the total revenue function is T R = p · q = 8 · q − q 2 . Confirm that
the marginal revenue function is mr = d(T R)/dq = 8 − 2 · q. (c) Suppose
that marginal cost is constant at mc = 1. What quantity maximizes the
firm’s profit? What price must the firm charge in order to sell this quantity.

7.2.3 The Second Derivative Test


Earlier in this chapter, we used the second derivative to indicate whether a
given function was concave upward or concave downward. We now use the
second derivative to determine whether the critical roots found by the first
derivative test actually relative generate maxima or minima. Application of
this test is a two-step process.
4
At the time of writing (June 2016), self-driving trucks are being considered. One
advantage of this type of truck is that it could save fuel costs by traveling at lower speeds
than are economical when drivers’ wages must be taken into account.
CHAPTER 7. OPTIMIZATION 169

First, given that the first derivative of a function y = f (x) exists, solve the
equation dy/dx = 0 for its critical roots. (This step is identical to the first
step of the first derivative test.) Next, if the second derivative d2 y/dx2 also
exists, then one of the following three conditions must hold:
(a) If d2 y/dx2 < 0, then the function f (x) has a relative maximum at x = x0.
(b) If d2 y/dx2 > 0, then the function f (x) has a relative minimum at x = x0.
(c) If d2 y/dx2 = 0, then the second derivative test fails. We must return to
the first derivative test to ascertain whether a relative maximum or minimum
exists.
Figure 7.7 illustrates the second derivative test graphically. In the first panel,
when x = 8.8197, d2 y/dx2 = 8.9443 (condition b), so this point corresponds
to a local maximum value. When x = 31.18, d2 y/dx2 = −8.9443 (condition
a), so this point corresponds to a local minimum value. In the right-hand
graph, at x = 20, d2 y/dx2 = 0, (condition c), so the second derivative test
cannot detect whether or not a relative maximum or minimum exists at this
point. Hence we cannot be certain what we have, based solely on these two
tests.
Exercise 7.2
Find the extreme values of the following functions, and determine by use
of the second derivative test whether they are maxima or minima. Confirm
your results with Maxima graphs.
1. y = x2 –8 · x + 10 2. y = x · (6 − x)2 3. y = x2 + 8
4. y = x4 − 2 · x2 + 6 5. y = x · e( − x) 6. y = x + 1/x
3 2
7. y = x /3 + x /2 + 12 · x 8. y = x/(x + 1) 9. = 1/(x + 4)

7.3 Maxima and Minima: Functions of Two


Independent Variables
The previous section dealt with the finding of extreme points for functions
of a single independent variable. We now extend this discussion to include
functions of two independent variables. We defer a discussion of how to
identify extreme points in functions of more than two independent variables
until our work with matrix algebra.
As above, we use the term absolute extreme point synonymously with global
extreme point. This underlines the fact that an absolute extreme point is
CHAPTER 7. OPTIMIZATION 170

global in nature with respect to the function in question. We also occasionally


refer to a relative extreme point as a local extreme point. This emphasizes
the fact that a given function may have several extreme points, one for each
locality or neighborhood of the function. Only one of these local extreme
points can be a global extreme point, however (unless some of the points
happen to have the same extreme value).
As with the case of a single independent variable, we define our terms. First,
a global maximum: Let z = f (x, y) be a real-valued function defined on a set
S of ordered pairs of real numbers. When x = x0 and y = y0, the function
f (x, y) has an absolute (global) maximum if (x0, y0) ≥ f (x, y)for all (x, y)
in S.
Likewise, for a global minimum: Let z = f (x, y) be a real-valued function
defined on a set S of ordered pairs of real numbers. When x = x0 and y = y0,
the function f (x, y) has an absolute (global) minimum if f (x0, y0) ≤ f (x, y)
for all (x, y) in S.
Relative maxima and minima are defined in like fashion. Let z = f (xy) be
a real-valued function defined on a set S of ordered pairs of real numbers.
When x = x0 and y = y0, the function f (x, y) has a relative (local) maximum
if f (x0, y0) ≥ f (x, y) for all (x.y) in the immediate vicinity or neighborhood
of x0, yo) in S.5
Figure 7.8 illustrates two functions. Each has two independent variables. The
function on the left reaches a relative maximum at a point (x0, y0) within the
range shown. Likewise, the function on the right reaches a relative minimum
at a point (x0, y0) in the range shown. The gray planes are at the maximum
and minimum values of z. The blue curves show the effect of setting x = x0
to determine z values for the y values in the indicated range. The black curve
reverses the roles of x and y. The intersection of these two curves is also the
point of tangency to the plane.
The slope of the black curve is ∂z/∂x, and the slope of the blue curve is
∂z/∂y. At extreme values of these functions, the two partial derivatives
equal zero. In general for a function f (x, y), which has continuous first
partial derivatives, the following is true: The function reaches a relative
extreme value when ∂z/∂x = ∂z/∂y = 0.
5
A neighborhood of (x0, y0) is a circular area around (x0, y0. Formally. δ neighbor-
hood of (x0, y0) is the disk (x − x0)2 (y − y0)2 ≤ δ 2 .
CHAPTER 7. OPTIMIZATION 171

Figure 7.8: Local maximum and minimum

Figure 7.8 shows that the condition ∂z/∂x = ∂z/∂y = 0 can indicate that
either a local maximum or a local minimum has been achieved. We em-
phasize may have, because two other possibilities exist. The function could
have reached an inflection point, just as we saw could happen with a single
independent variable. Furthermore, it could reach a saddle point.
To envision a saddle point, imagine yourself at a point on a three dimensional
shape. When you look in one direction you appear to be at the top of a hill
(where the slope is zero). Now, turn 90 degrees. You are still at a point where
the slope is zero, but the shape running from the back to the front of the
saddle is now a valley and you are at the minimum. Figure 7.9 illustrates
the case of a saddle point. Standing at the point indicated by +, when you
face in the y direction, you are atop the surface (at least locally). When you
face in the x direction you are (at least locally) at the minimum point on the
surface.
Figure 7.10 shows a more complicated picture. The function here is z = x3 ·y 2 ,
for which ∂z/∂x = 3 · x2 · y 2 and ∂z/∂y = 2 · x3 · y. When either x = 0 or
y = 0, both partial derivatives equal zero, but no local extreme values are
apparent. In particular, observe that moving from negative to positive values
CHAPTER 7. OPTIMIZATION 172

Figure 7.9: A saddle point

of x reveals a line of inflection points. At each of these points, ∂z/∂x = 0 but


on either side of the line ∂z/∂x > 0.That is ∂z/∂x does not change signs.

7.3.1 Second Order (Sufficient) Condition


In the case of functions of one independent variable of the form y = f (x),
f 0 (x) = 0 was a necessary condition for an extreme point. Also f ”(x) < 0
and f ”(x) > 0 were sufficient conditions for the existence of a maximum and
minimum point, respectively.
The second-order (sufficient) condition for the existence of an extreme point
in the case of a function of two independent variables is analogous, though
more extensive. A second partial derivative such as fxx considers the shape
of a surface only in reference to the plane XZ. Similarly fyy considers the
shape of a surface only in reference to the plane YZ.
Neither second partial derivative considers the shape of any cross section of
the surface. For example, fxx ignores the YZ plane as well as the XY plane.
This means that we cannot rely on the sign of the second partial derivative
CHAPTER 7. OPTIMIZATION 173

Figure 7.10: Figure with inflection points

to identify extreme points, as we did in the case of functions of only one


independent variable. For example, fy = 0 and fyy < 0 is not a guarantee of
a maximum point because we have not also considered fxx and fxy .
Applying the second-order test consists of the following:
1. Given that the first partial derivatives of z = f (x, y) exist, set them equal
to zero. That is, find the equations for ∂z/∂x = 0 and ∂z/∂y = 0. Solve
these two equations for their critical roots. We label these values x = x0 and
y = y0.
2. Determine whether fxx and fyy exist at (x0, y0). If both exist, then one
of the following conditions must hold at this point:
(a) When fxx · ff yy − (fxy )2 > 0, fxx < 0, and fyy < 0, we have a relative
maximum.
(b) When fxx · ff yy − (fxy )2 > 0, fxx > 0, and fyy > 0, we have a relative
minimum.
(c) When fxx · ff yy − (fxy )2 < 0, we have a saddle point.
(d) When fxx ·ff yy −(fxy )2 = 0, the second-order test fails. A relative extreme
point may exist. The second-order test, however, cannot indicate whether
or not that is the case. One must examine the original function z = f (x, y)
CHAPTER 7. OPTIMIZATION 174

in the neighborhood of x = 0, y = y0 in order to determine whether a local


extreme point appears to exist.
Summary of conditions for unconstrained extreme value of z =
f (x, y)
First-order condition
fx = 0, fy = 0
Second-order condition
(a) Maximum fxx · fyy − (fxy )2 > 0 and fxx , fyy < 0
(b) Minimum fxx · fyy − (fxy )2 > 0 and fxx , fyy > 0
(c) Saddle point fxx · fyy − (fxy )2 < 0
(d) Test fails fxx · fyy − (fxy )2 = 0
Three examples follow.
Example 1. Find the extreme value of z = f (x, y) = 8 − x2 − y 2 , if it exists.
fx = −2 · x and fy = −2 · y have a single solution, x = 0 and y = 0.
Therefore, the function may have an extreme value at (0,0).
Furthermore, fxx = −2, fyy = −2 and fxy = 0.
Thus, the second order condition is fxx · fyy − (fxy )2 = 4 − 0 > 0, fxx < 0,
and fyy < 0, indicating a maximum value at (0,0).
Example 2. Find the extreme value of z = f (x, y) = x3 + y 3 − 3 · x · y, if
it exists.
fx = 3 · x2 − 3 · y = 0 and fy = 3 · y 2 − 3 · x have two solutions, one when
x = 0 and y = 0 and another when x = 1, and y = 1.
Therefore, the function may have extreme values at (0,0) and (1,1).
fxx = 6 · x, fyy = 6 · y and fxy = −3.
First, consider the (0,0) case:
The second order condition is fxx · fyy − (fxy )2 = (6 · x) · (6 · y) − (−3)2 =
0 − 9 < 0, indicating a saddle point at (0,0).
Now, consider the (1,1) case:
The second order condition is fxx · fyy − (fxy )2 = (6 · x) · (6 · y) − (−3)2 =
36 − 9 > 0, indicating a local extreme value at (1,1). Because fxx = 3 · x and
fyy = 3 · y, z reaches a minimum point at (1,1).
Figure 7.11 shows this function over the relevant range.
The blue lines show z values for x when y = 0 and for y when x = 0. The
blue lines intersect at the saddle point.
The black lines show z values for x when y = 0 and for y when x = 0. The
black lines intersect at the local minimum value.
CHAPTER 7. OPTIMIZATION 175

Figure 7.11: Example 2 figure


CHAPTER 7. OPTIMIZATION 176

Example 3. Find the extreme value of z = f (x, y) = x2 + y 2 , if it exists.


fx = 2 · x and fy = 2 · y have a single solution, x = 0 and y = 0.
Therefore, the function may have an extreme value at (0,0).
Furthermore, fxx = 2, fyy = 2 and fxy = 0.
Thus, the second order condition is fxx · fyy − (fxy )2 = 4 − 0 > 0, fxx > 0,
and fyy > 0, indicating a minimum value at (0,0).
Exercises 7.3
Find the extreme values of the following functions. If possible, determine by
use of the second derivative test whether each is a relative maximum or a
relative minimum.
1. z = z 2 + (y − 4)2 2. z = x2 − x · y + y 2 − 2 · x + y
2 2
3. z = x − 2 · x · y + y 4. z = x2 + y 2 − 2 · x − 2 · y − x · y + 4
5. z = x3 − 3 · x + y 3 − 12 · y + 6 6. z = x2 + y 2 + x · y + 5 · x + 4 · y
7. z = x2 + 2 · y 2 − 4 · x + 8 · y

7.4 Maxima and Minima Subject to


Constraints
Rare is the decision-maker who makes decisions without reference to con-
straints. Business people and consumers alike have limited budgets, re-
sources. and time. As a consequence, many of the most realistic maximiza-
tion and minimization problems in business and economics involve finding
an extreme point subject to one or more constraints.
For example, the task of a salesperson may be to maximize the sales in a
territory subject to a budget that limits the salesperson’s ability to travel
and service that territory. An academic administrator may wish to construct
a schedule of courses that maximizes the usage of classrooms for certain key
time periods during the day. However, the administrator must do so without
violating constraints on how many classes can be offered, how many classes
can be offered in a single time slot, and so forth. The number of decision-
making problems that involve constrained maximization or minimization is
as large and diverse as the world itself.
A constraint acts as a prohibiting, limiting agent in an optimization problem.
That is, the constraint reduces the feasible or workable area of the objective
function Figure 7.12 show two views of an objective function (blue surface)
CHAPTER 7. OPTIMIZATION 177

Figure 7.12: Two views of a constraint

and a constraint (green plane). The objective is to move as high up the hill
as the constraint allows. The graph does not provide enough information to
determine the highest feasible value of z.6
In general, a constraint must result in an extreme point whose value is less
than or equal to the extreme value obtained when the same objective func-
tion is maximized in the absence of the constraint. Similarly, imposing a
constraint on a minimization problem must result in an extreme point whose
value is greater than or equal to the value obtained when the same objective
function is minimized in the absence of the constraint.
We generally try to solve a constrained optimization problem by one of two
methods. The first involves substituting the constraint into the objective
function, then proceeding as if one were maximizing or minimizing an un-
constrained function. This method seems straightforward. Unfortunately,
it often it becomes complicated and quite troublesome when the objective
function and constraint(s) are something other than very simple functions.
6
We determine below that (x = 7/5, y = 16/5) is the combination that yields the
largest value of z, z = 19/5.
CHAPTER 7. OPTIMIZATION 178

Hence the most popular method of maximizing or minimizing in the face of


a constraint is by the use of Lagrange multipliers.

7.4.1 Lagrange Multipliers


Assume an objective function in the form z = f (x, y)that is to be maximized
or minimized subject to the constraint given by g(x, y) = 0. We now term
a new objective function that contains both the original objective function
and the constraint:

L= L(x, y, λ) = f (x, y)+ λ · g(x, y)


The The The The
Lagrangian New Original Constraint
Expression Objective Objective
Function Function

The Greek letter λ (lambda) is a newly created unknown variable that has
the property at being able to apply to the constrained objective function
precisely the same first-order condition applied when an extremum is found
in the absence of a constraint. The new variable λ has an important inter-
pretation: it equals the change in the objective function per unit change in
the constraint.
First-order Conditions
We determine the optimal values of x and y in three steps: First we differ-
entiate the new objective function in with respect to x, y, and λ. Then we
set these partial derivatives equal to 0. Finally, we solve this system of three
equations for the three unknown values.
Thus, we create a system of these three equations:

Lx = fx − λ · gx = 0,

Ly = fy − λ · gy = 0,
and
Lλ = fλ − λ · gλ = 0.
CHAPTER 7. OPTIMIZATION 179

We can solve these equation for the critical roots of the function L(x, y, λ).
Note that the last of the three first-order conditions is actually nothing more
than the constraint that must be satisfied when the extreme point is found.
Example 1. We now apply this approach to the function and constraint
that generate Figure 7.12. The twoexpressions, stated as Maxima output,
are f (x, y) := − (x − 5)2 + (y − 5)2 + 20 and g(x, y) := x + y/2 − 3.7
The four commands below create the Lagrangian expression and determine
the first partial derivatives.

L : f(x,y)- %lambda*g(x,y);
Lx: diff(L,x);
Ly: diff(L,y);
Llambda: diff(L,%lambda);

The resulting output is the following four expressions.


y 
−λ + x − 3 − (y − 5)2 − (x − 5)2 + 20
2
−2 (x − 5) − λ
λ
−2 (y − 5) −
2
y
− −x+3
2
The command soln: solve([Lx,Ly,Llambda],[x,y,%lambda])[1]; in-
structs Maxima to solve the three partial derivatives and to return a list of
values. That list is assigned the name soln. The result is this list of values:
7 16 36
[x = , y = , λ = ].
5 5 5

Inserting these values into f (x, y) shows the maximum value of z given this
constraint. Use the command subst(soln,f(x,y)); to generate the result
19/5, the maximum attainable value of z.8
7
When an expression is entered into Maxima, it is treated as being equal to zero unless
another value is expressly entered.
8
The accompanying workbook relaxes the constraint by a small amount and shows
that the value of λ closely approximates the resulting change in z’s value.
CHAPTER 7. OPTIMIZATION 180

Example 2. Maximize the utility function u = x · y subject to the budget


constraint given by m = px · x + py · y. As above, we create the Lagrangian
expression. We use these commands: [u : x*y, L : u + %lambda*(m -
px*x - py*y)]. The commands are in brackets to create this output list:
[xy, λ (−pyy − px x + m) + xy].
Commands to create a list of first-order conditions result in this output:
[y − λpx , x − λpy, − pyy − px x + m].
You should derive these conditions as an exercise. Next the command soln:
solve([Lx,Ly,Llambda],[x,y,%lambda])[1] creates the following list of
solutions and assigns it the name soln:
m m m
[x = ,y = ,λ = ].
2px 2py 2px py

The first two entries are the (uncompensated) demand curves for x and y.
Note that this consumer’s income would be equally divided between the two
goods: px · x = m/2 = py · y. The Lagrangian multiplier λ can be interpreted
as the marginal utility of income. Given this function, this value is constant
for a given set of prices.9
Now, suppose that m = 100, px = 2, and py = 5. The command subst([m =
100, px=2, py = 5], soln) performs the (in this case simple) calculations
to yield [x = 25, y = 10, λ = 5]. We see that the consumer does spend
m/2 = 50 on each good. Also, λ is a positive constant so the consumer gains
5 “units” of utility per one-unit increase in m, no matter what the consumer’s
income level might be.
Often a minus sign (-) is used in front of the constraint in a Lagrangian
expression. Making this change does not affect the critical roots of the in-
dependent variables in the original objective function. There is an intuitive
explanation or why the sign of the constraint term is of no consequence. The
value of the constraint term, when the objective function is being maximized
or minimized, as appropriate, is equal to zero. Whether we add or subtract
zero is of no consequence.
9
Beware of two possible errors in interpreting this specific result. First, this constancy
is a characteristic of this class of utility functions and should not be treated as a gen-
eral result. Second, numerical values have no meaning when utility: the measures are
subjective.
CHAPTER 7. OPTIMIZATION 181

Second-order (sufficient) test

The method of Lagrange identifies only those values of the independent vari-
ables that satisfy first-order or necessary conditions for an extreme point.
These values may or may not actually represent an extreme point. A second-
order test is necessary to provide further information on this matter. The
second-order test is as follows.
1. Given: Lx = Ly = 0 at x = x0, y = y0 . Given also: Lxx , Lyy , and Lxy
exist at x = x0, y = y0.
2. Then, one of the following conditions must hold:
(a) If Lxx Lyy − (Lxy )2 > 0, and both Lxx and Lyy are negative, then we have
a relative maximum at x = x0, y = y0.
(b) If Lxx Lyy − (Lxy )2 > 0, and both Lxx and Lyy are positive, then we have
a relative minimum at x = x0, y = y0.
(c) If Lxx Lyy − (Lxy )2 < 0, then the second-order test fails and is incapable of
indicating whether or not a relative extreme point exists. A relative extreme
point may exist. One must analyze the function z = f (x, y) in the neighbor-
hood of x = x0, y = y0 in order to ascertain whether a local extreme point
exists at x = x0, y = y0.
The analysis of such complicated cases is one of the areas in which a computer
algebra system becomes especially useful. Many points in the neighborhood
can typically be evaluated quickly, providing insights into the function’s be-
havior in what might be a critical region.
The second-order test outlined above is quite similar to the second-order test
described for the case when an unconstrained extreme point is being sought.
There is, however, an important difference, Assume that Lxx Lyy −(Lxy )2 ≤ 0.
In the unconstrained case, a saddle point exists when fxx fyy −(fxy )2 < 0, and
an extreme point may exist when fxx fyy − (fxy )2 <= 0. In the constrained
case, however, we can say nothing about the existence of a saddle point when
Lxx Lyy − (Lxy )2 < 0. An extreme point may exist when Lxx Lyy − (Lxy )2 = 0
as well as when Lxx Lyy − (Lxy )2 < 0 in the constrained case.
Summary of Conditions for Constrained Extremum: z = f (x, y) subject to
g(x, y) = 0.
CHAPTER 7. OPTIMIZATION 182

First-order condition
Lx = 0, Ly = Lλ = 0
Second-order condition
(a) Maximum Lxx · Lyy − (Lxy )2 > 0 and Lxx , Lyy < 0
(b) Minimum Lxx · Lyy − (Lxy )2 > 0 and Lxx , Lyy > 0
(c) Test fails Lxx · Lyy − (Lxy )2 ≤ 0
Example 1. Find the extremum of z = x2 + y 2 − 4 · x − 4 · y + 7 subject to
x + y = 4.
Form the Lagrangian function L = x2 + y 2 + 2 · x + 2 · y + 4 + λ · (x + y − 4).
Derive the first-order conditions:
Lx = 2 · x − 4 − λ = 0,
Ly = 2 · y − 4 − λ = 0, and
Lλ = x + y − 4 = 0.
Both Lx and Ly equal λ, so 2 · x − 4 = 2 · y − 4, implying that x = y when
the constraint is satisfied. This implies that x = y = 2.
The second-order expressions are Lxx = 2, Lyy = 2, and Lxy = 0. Therefore
Lxx · Lyy − (Lxy )2 = 4 − 0 > 0 which ensures that an extreme value exists.
Furthermore, Lxx > 0 and Lyy > 0 indicate that a minimum value has been
found.
The value of λ is 0, implying that relaxing the constraint would move us
no closer to the unconstrained maximum. As it happens, this constraint
is irrelevant: The constrained optimum is the local minimum. Repeat this
example, setting x + y = 4 and confirm that λ 6= 0; also, interpret the new
value.10
Example 2. Find the extremum of z = x · log(y) subject to x + y = 4. The
original expression and the Lagrangian expression are these:
√ √
[x y, x y − λ (y + x − 4)].

The associated first-order conditions involve setting the following derivatives


equal to zero:
√ x
[ y − λ, √ − λ, − y − x + 9].
2 y
10
The workbook for this chapter shows z and the plane z = −1 confirming that z is
tangent to the plane at this input combination.
CHAPTER 7. OPTIMIZATION 183

The solution to this system of equations for x, y, and λ are



[x = 6, y = 3, λ = 3].

The terms that relate to the second-order test–fxx , fyy , and fxy are
x 1
[0, − 3 , √ ].
4y 2 2 y

The second-order test relates to the sign of


 2
1 1
0·− √ − √ ,
2 3 2 3
which is −1/12. Therefore, we cannot be certain that we have found an
extreme point.
We can use Maxima’s ability to carry out simulations to provide evidence
regarding the nature of this solution. The commands xList:makelist(8/3
+ i/6,i, -4,4); yList: 4 - xList;
zList: float( xList* sqrt(yList)); can be used to generate this table
of values.
13 7 5 8 17 19 10
x 2 6 3 2 3 6
3 6 3

11 5 3 4 7 5 2
y 2 6 3 2 3 6
1 6 3

z 2.828 2.933 3.012 3.061 3.079 3.06 3.0 2.89 2.721

The center entry (8/3, 4/3) is the solution that we found above. It is also
the largest value of z in this neighborhood of (x, y) combinations.
Exercise 7.4
Solve the following constrained optimization problems by the method of La-
grange multipliers. Use Maxima to confirm your computations.
1. z = 2 · x2 + y 2 subject to x + y = 1
2. z = x2 − 2 · x · y + y 2 subject to x + y = 2
3. z = x2 + 4 · y 2 + 24 subject to x − 4 · y = 10
4. z = 4 · x2 + x · y + 3 · y 2 subject to x + 2 · y = 21
5. z = 6 · x2 − x · y + 5 · y 2 subject to 2 · x + y = 24
6. z = 3 · x2 + y 2 − 2 · x · y − 8 subject to x + y = 1
CHAPTER 7. OPTIMIZATION 184

7.5 Economic Applications


We now apply the tools that we have developed in this chapter to specific
types of problems in business and economics.

7.5.1 Profit Maximization: Price-Taking Firms


A fundamental problem in business and economics is that of determining the
conditions that must hold for a firm to maximize its profits. We focus on a
stylized firm that produces a specific, well-defined product (either a good or a
service). In this section, we look at a firm that purchases its inputs and sells
its output in perfectly competitive markets. That is, the firm’s purchases
of inputs, and its sales of the output it produces, are sufficiently small and
that its output level does not appreciably affect the prices of either inputs or
outputs. We refer to such firms as price takers.
Let p = f (x) be the firm’s inverse demand function. Assume that p, the price
of the firm’s product, is a constant, unaffected by x, the firm’s output rate.
The firm’s total sales revenue T R of the firm is T R = p · x, or T R = x · f (x).
Therefore, we can write T R = g(x).
The firm’s total cost T C function is T C = h(x, k), where k is per-period fixed
cost, that part of cost that is not affected by the output rate. Thus, total
cost is the sum of fixed cost variable variable costs, those that do depend on
x’s value.
The profit of the firm, π, is defined as total revenue minus total cost, and
equals T R − T C. We find the conditions for a profit-maximizing level of
output by satisfying the first-order and second-order conditions for an un-
constrained maximum, namely, dπ/dx = 0 and d2 π/dx2 < 0. Differentiating
T R − T C with respect to x and setting this derivative equal to zero yields
dπ/dx = dT R/dx − dT C/dx = 0 which implies that the firm achieves max-
imum profit if it produces the quantity at which dT R/dx = dT C/dx. In
terms that you have seen before, the firm achieves maximum profit by pro-
ducing the quantity at which marginal revenue (dT R/dx) equals marginal
cost (dT C/dx).
For the price-taking firm p = dT R/dx = marginal revenue. Therefore, such
a firm maximizes its profits by selecting the output rate at which M C =
dT C/dq = p.
CHAPTER 7. OPTIMIZATION 185

Figure 7.13 depicts the situation of a firm in a perfectly competitive market


(i. e., a price-taking firm) The total cost and total revenue functions shown
in the top panel intersect twice, initially at output level x1 and subsequently
at output level x3. At these two levels of output, total profit is equal to
zero. The intervals (0, x1) and (x3, ∞) represent negative profit, whereas
the interval (x1, x3) represents positive profit. The total profit curve in the
same panel reflects these considerations.11
The first-order condition for profit maximization, dπ/dx = 0, implies that
the slope of the total revenue curve must be equal to the slope of a tangent to
the total cost curve. In the top panel of Figure 7.13, this equality occurs at
x = x1.12 The lower panel confirms that the profit maximization condition
can also be phrased in terms of the equality of marginal revenue and marginal
cost.13
The lower panel shows the quantity at which the firm’s average cost is mini-
mized. For the price-taking firm, this is the quantity at which profit per unit
is maximized. One might be tempted to think that this quantity is one to be
produced. It is not. As the graph shows, for each unit between xminac (for
x minimum ac)and x2, the price exceeds the marginal cost. That means that
each of these units adds to the profits that would be earned at x = xminac.
The lesson is general: Focus on the marginal values.

7.5.2 Profit Maximization: Price-Searching Firms


Many firms produce output that is a less than perfect substiute for that of
other firms. These firms must simultaneously determine the price to charge
and the quantity to sell. That is, they must search for the point on their
11
The units on the axes are stated in some detail. They emphasize that output has two
aspects, a physical unit and a time period. Also, note that the monetary units on the y
axes differ for the two graph. The first shows revenue, cost and profit (or loss) per time
period. The second shows per-unit revenue cost and revenue per unit produced.
12
Depending on the curvature of the T C and the price level, it is possible that a
second tangency point can occur for x < x1. If so, this tangency point is consistent with
minimizing profits or, equivalently, with maximizing losses.
13
From a managerial point of view, the marginal revenue, marginal cost comparison
might be more useful. Firms are likely to have some data on how changes in output
affect their cost levels. Knowing the values of total cost of a product is more problematic,
especially for firms that produce more than one product.
CHAPTER 7. OPTIMIZATION 186

Figure 7.13: Cost, revenue, and profits


CHAPTER 7. OPTIMIZATION 187

demand curve that yields maximum profit.14


Let the imperfectly competitive firm’s demand function for its product x (in
inverse form) be p = f (x), where fx < 0. The firm’s total cost function
TC is the same as above. Hence the firm’s profit function is given by π =
f (x) · x − T C. The first-order condition for profit, as before, is that the firm
produce a quantity such that dπ/dx = dT R/dx − dT C/dx = 0. That is,
the firm, like its price-taking counterpart, must select the quantity at which
marginal revenue equals marginal cost.
Also, for this quantity to correspond to maximum, not minimum profit, the
condition d2 π/dx2 < 0. This is equivalent to requiring that d2 T R/dx2 −
d2 T C/dx2 < 0 or d2 T R/dx2 < d2 T C/dx2 must be satisfied. That is, the
slope of the marginal revenue curve be less than the slope of the total cost
curve. The marginal cost curve must cut the total cost curve from below.
Figure 7.14 illustrates profit maximization for the imperfectly competitive
(price searching) firm. The firm’s total revenue function is no longer a ray
through the origin, because price decreases as output increases. The lower
panel shows this fact more expressly. The black curve shows the price that
this firm can charge for each possible output rate per period of time. It also
shows the implied marginal revenue, which is less than the price. Aside from
the divergence of marginal revenue from price (average revenue), the analysis
proceeds as before.
Example. That the Co-op Bookstore considers that to maximize profit on
its sales of Adam Smith’s The Wealth of Nations is quite appropriate. The
(inverse) demand for the book is given by p = 0.1665 · x2 − 0.175 · x + 50
(dollars). The implied marginal revenue is M R = −0.4995 · x2 − 0.35 ·
x + 50 The per-unit cost is a constant, $25 per book. So T C = 25 · x
and marginal cost is 25. The accompanying workbook shows the demand,
marginal revenue, and marginal cost curves.
The first derivative of the profit function, π = −0.1665 · x3 − 0.175 · x2 + 25 · x
is dπ/dx = −0.4995 · x2 − 0.35 · x + 25. Setting this expression equal to zero
yields two values, one of which is negative. The positive value is x = 6.7329
units per week, which we name x1. If the firm is thinking of per-unit sales
14
If the firm could know its demand curve, then its task would be simply to set the
profit-maximizing price (or, equivalently, to sell the profit-maximizing quantity). In fact,
firms do not know the demand curve for their products, so “price searching” is more
descriptive that “price setting.”
CHAPTER 7. OPTIMIZATION 188

Figure 7.14: Demand, cost, and profits Revenue, cost, and profits
CHAPTER 7. OPTIMIZATION 189

over a long time, then fraction units make sense. If it is thinking of this
week, only a discrete number of books can be sold.
Fortunately, Maxima lets us have it both ways. The following three com-
mands determine price, output, and profit if the bookstore sells 6, 6.7329, or
7 books per week:

subst(x=x1,[x,p,TR-TC]);
subst(x=floor(x1),[x,p,TR-TC]);
subst(x=ceiling(x1),[c,p,TR-TC]);.

The results are these.


Quantity Price Profit
6.7329 41.273 109.57
6 42.956 107.73
7 40.616 109.31

The second derivative of the profit function is −0.999 · x − 0.35. For any
of the three x values above, this expression is negative. We have, therefore,
determined a profit-maximizing quantity.

7.5.3 Production: Marginal and Average Products


Consider a production function of the form Q = f (L, K), where Q =output,
and L and K are the labor and capital inputs, respectively. Assume a short-
run situation such that the amount of capital is fixed at K = K0.15 Thus
Q = f (L), given K = K0. The functional form f reflects current technology.

The average product of labor AP L is given by


Q f (K0, L) f (L)
AP L = = = .
L L L

The marginal product of labor MPL is given by


d(f (L))
MP L = = fL (L).
dL
15
Hammock and Mixon, Chapter 6, treats production in more general terms.
CHAPTER 7. OPTIMIZATION 190

As Figure 7.15 illustrates, the M P L curve cuts the AP L curve at the APL
curve’s highest point. That is, M P L = AP L at the value of L for which AP L
is at a maximum. We can demonstrate this mathematically by showing the
conditions under which AP L is at a maximum. We differentiate AP L = Q/L
with respect to L using these to Maxima commands: depends(Q,L) and
d
Q
diff(Q/L, L).16 The result is dLL − LQ2 . Setting this derivative equal to
zero and dividing through by L, which has a positive value, implies that
M P L = AP L is a necessary condition for AP L to achieve an extreme value.
To confirm that this value is a maximum requires evaluating the second
derivative of AP L(L). The derivative supplied by Maxima is this:
d2 d

dL 2 Q 2 dL
Q 2Q
− 2
+ 3.
L L L
The second term is 2 · M P L/L2 and the third term is 2 · Q · 1 = 2 · AP L/L2 .
L L2
Because the first-order condition requires that AP L = M P L, these sum to
zero. Therefore, the sign of the first term determines the sign of second
derivative of AP L(L). The numerator of that term, the first derivative of
M P L(L), is negative for all L. The test confirms what the graph shows: at
L = L1, the average product of labor reaches its maximum value.17

7.5.4 Production and Cost


This section derives the firm’s cost curve from the production function that
the previous section develops. This derivation illustrates the nature of the
dual relationship between production and cost. That is, a firm’s cost curve
is derived from its production function, so that employment decisions and
costs are uniquely related.
Mathematically, however, matters are not quite that simple. The production
function that governs our hypothetical firm is Q = 30 · L + 5 · L2 − 0.2] · L3 , a
cubic function. For such a function Q is not monotonically related to L. If,
16
Of course, for this simple application of the quotient rule, Maxima’s use is not re-
quired.
17
The accompanying workbook shows a slightly more flexible, though probably less
plausible, production function. For it, a quite small value of L results in minimum AP L,
and a larger value results in a maximum. We shall see that, even if the AP L-minimizing
value of L exists, it is likely not relevant to economic analysis.)
CHAPTER 7. OPTIMIZATION 191

Figure 7.15: Total and per-unit output


CHAPTER 7. OPTIMIZATION 192

however, maximizing behavior is added, the difficulty vanishes, because no


maximizing employer would hire so much L that the M P L becomes negative.
As a result of this observation, we can map employment of L to both cost and
output. Therefore, Q and T C are bound by a parametric relationship.18 For
the cubic production used here, the maximum value of output is Q = 1003.6
units, with L = 19.262 units of labor. We restrict all graphs to this range.
To begin, we define the parametric relationships between L and cost, TC(L) :=
400 + 500L for total cost, MC(L) := −0.6L2500 +10L+30
for marginal cost, and
500L+400
AC(L) := −0.2L3 +5L2 +30L for average cost. We also define a parametric rela-
tionship for average variable. Be aware that these are emphatically not the
usual total cost, marginal cost, and average cost functions; these relate costs
to output and not to employment.
Figure 7.16 shows total and per-unit cost curves. These curves show cost as
a function of quantity. They do so via the application of commands like this
one: parametric(Q(L),AC(L),L,.1,Lmax), which appear inside the draw2d
command. This command places Q values from the production function on
the x axis and T C values from the cost (as a function of L) function on the
y axis. Other commands do the same for the per-unit cost curves. Observe
that the resulting cost curves behave much as the stylized curves that we
graphed earlier, using an ad hoc cubic cost function.
Figure 7.16 also contains a price line, which is appropriate for a price-taking
firm. We know that profit-maximization is attained by producing Q such
that M C = p. The graph shows that this equality holds for two values.
Earlier analysis points out the the first of this pair is consistent with loss
maximization. Thus, a quantity Q ≈ 950 seems to be the profit-maximizing
quantity.
We do not have an explicit expression to relate price and Q. Some economic
theory, however, solves our problem. Remember that a price-taking firm
maximizes its profit by employing its variable input at a level such that price·
∂Q/∂L equals the wage rate. In our example, we can phrase this requirement
as MPL(L)*12- 500. We apply the find root command to determine this
value which is Q = 917.53, with an employment level L = 15.404. Confirm
that this firm earns a profit of approximately $2908.
18
We limit the analysis to the short run, where the employment of a single input varies.
More advanced treatments allow multiple variable inputs. See Hammock and Mixon,
Chapter 7, and “Key to Textbooks.”
CHAPTER 7. OPTIMIZATION 193

Figure 7.16: Total and per-unit cost curves


CHAPTER 7. OPTIMIZATION 194

We can now see why production by a price-taking firm occurs within the
range of decreasing marginal product. In this example, production occurs
where M P L = 41.666 and AP L = 59.562. More generally, the following re-
lationships hold: M C = w/M P L and AV C = w/AP L. Profit-maximization
requires that M C = p, so M P L = w/p > 0, where w/p is the “real wage”
that the firm pays. The price level must exceed the minimum value of AV C,
so M C > AV C, implying that M P L < AP L.19

7.5.5 Taxation
Chapter 6 examines the effects of imposing a tax on either the sale or the
purchase of a single good or service when both buyers and sellers are price
takers. It shows that whether the tax is nominally imposed on buyers or
on sellers is beside the point. Either way, the effects on the price paid by
the buyers and the price received by the sellers is determined by the relative
values of the price elasticity of demand and the price elasticity of supply.
This section examines the case in which the seller in question is a price-
searcher. More specifically, we examine the polar case of the pure monopolist.
In the case of pure competition (price takes on both sides of the market),
the analysis proceeds by solving a system of two equations (the demand
and supply curves) and then imposing and tracing the effects of a change.
(This approach is another example of comparative statics analysis.) When
examining the case of price takers’ markets, we did not look at the behavior
of any single firm or buyer; the effects of their behavior are summarized in
the supply curve and the demand curve respectively. Therefore, we did not
have to apply any optimization conditions.
With a single price-searching seller, however, the firm’s optimizing behavior
is central to the analysis. As before, we follow the analytical approach that
Bishop [2] provides. Begin by defining the objective function: the function
π = T R(x) − T C(x) − t · x is to be maximized.
The first-order condition for maximizing π is T Rx −T Cx −t = 0. The second-
19
If the firm is a price-searcher, then employment can occur at a lower L, where M P L
might exceed AP L. The accompanying workbook shows how this can happen. Likewise,
if the firm faces an upward-sloping supply curve of labor, it might employ an amount of
labor such that M P L > AP L.
CHAPTER 7. OPTIMIZATION 195

order condition is that πxx = T Rxx − T Cxx < 0.20 That is, the marginal
revenue’s slope must be less than that of the marginal cost. In most cases,
we expect M R curves to slope downward (T Rxx < 0) and marginal cost
curves to be either horizontal or upward sloping (T Cxx ≥ 0).
In order to determine the nature of dp/dt, we first apply the implicit function
theorem to the first-order condition and determine dx/dt. First, dπx /dx =
T Rxx − T Cxx . And, of course, dπt = 1. Therefore,
dx 1
= .
dt T Rxx − T Cxx
Multiplying this term by dp/dx yields the result that we seek:

dp px
= ,
dt T Rxx − T Cxx
where px is the slope of the inverse demand curve. Both the numerator and
the denominator are negative, so dp/dt > 0.
This formulation allows direct analysis of the case in which the demand curve
and the marginal cost curve are linear. If p = α + a · x, then M R = α + 2 · x.
Likewise, if AV C = b + c · x then M C = b + 2 · x. The slopes of M R and M C
are 2 · a and 2 · c. The table below summarizes these aspects and implications
of the linear price and average variable cost functions. The results imply that
dp/dt = a/(2 · (a − c)). Recalling that a > 0 and c ≥ 0, these values imply
that dp/dt < 1, a result that is similar to the case of price-taker markets.
We can say a bit more. Suppose that c = 0.Then dp/dt = 1/2. For c > 0,
dp/dt < 1/2. As with price-taker markets, the relative values of b and c
determine the effect of the tax on the prices paid and received.
 
p p Slope TR MR M R Slope
ax + α
 a a x2 + αx 2ax + α 2a 

 === === === === === 
 
 AV C AV C slope TV C MC M C Slope
cx + b c c x2 + bx 2cx + b 2c
20
Also, the local maximum profit must be larger than at any other output rate. In
particular, it must exceed π(0), the profit (positive or negative) that the firm earns when
x = 0.
CHAPTER 7. OPTIMIZATION 196

Unlike the case of price-taker markets, the result above is not general. We
can easily find a case in which dp/dt > 0. Suppose that the demand curve
has constant elasticity at each price and that marginal cost is constant (its
slope is 0). Recalling that M R = p · (1 + 1/Epd), setting M R = M C implies
that p = M C/(1 + 1/Epd). Here, Epd is the price elasticity of demand.
For a price-searching firm to produce an appreciable amount of at good with
this type of demand curve, marginal revenue must be positive. That requires
that the demand curve must be elastic. That is, Eps < −1. This, in turn,
1
implies that 1 + 1/Epd < 1. Because Epd is a constant, 1+1/Epd is a constant
markup that maximizes profits. If a firm’s sale is taxed, then the tax-inclusive
1
marginal cost is another constant, c+t, and the price will rise by t· 1+1/Epd > t.
Figure 7.17 shows the impact of a $1 excise tax on the sale of products in
market with linear demand and cost curves and in a market with constant
elasticity and constant marginal cost. The details are in the workbook that
accompanies this chapter.

Figure 7.17: Tax effects

In the linear case, the price increase is clearly less than that vertical dis-
placement of the marginal cost curve (which in this illustration is $1). In
CHAPTER 7. OPTIMIZATION 197

some sense, therefore, we can say that the buyers and the single seller share
the cost of the tax. In the second example, however, the price rises by more
than the tax (twice as much in this case because we use Epd = −2), so the
question of “sharing” cannot be stated in simple terms.
One might be tempted to think that the monopolist gains from having a
tax imposed, because the price rises by more than the tax. This inference
is incorrect, because it ignores two effects of the tax. First, the monopolist
must pay the tax and, second, the monopolist’s output decreases. In this
illustrative example the monpolist’s profit (less fixed cost, which we ignore)
falls from $7.5 to $6.0.21
This illustrative case that results in dp/dt > 1 is not the only case in which
this result can occur. Bishop [2] provides the general conditions that lead to
this result. Also Bishop addresses the impact of ad valorem taxes.

7.5.6 Inventories and Reordering


Most business firms live in a world where their production is not perfectly
synchronized with their sales. Any particular firm therefore typically main-
tains some sort of inventory of unsold units of its output. There is a direct
relationship between the number of units of inventory and the cost of keeping
that inventory. Hence the firm wants to maintain as small an inventory as
possible and still be able to meet anticipated customer orders. At the same
time, however, there are costs associated with starting up production and
reordering when the firm’s inventory is depleted.
As a result, the firm must balance these two types of costs when it decides
about how large an inventory to keep and how often to reorder. A large
inventory increases inventory storage costs but reduces reorder costs. A small
inventory decreases inventory storage costs but increases reorder costs. The
optimal inventory (that inventory that minimizes the sum of storage and
reorder costs) must take both types of costs into account.
Let Q be the expected sales of the firm in units in a particular time period,
which we will designate a year. Suppose that Q = 60, 000; this implies that
21
Initially the firm sells 1.875 units for $8 per unit and incurs a variable cost (AV C =
M C) of $4. Its profit is 1.875 · 8 − 4 · 1.875 = 7.5. With a $1 tax, the profit falls to
1.2 · 10 − 1.2 · 5 = 6.
CHAPTER 7. OPTIMIZATION 198

the firm expects to sell 60,000 units over the space of the next year. Let us
further assume that these sales will be spaced evenly throughout the year,
so that 60,000/ 12 = 5000 units will be sold each month.
Storage Cost. Let U represent the number of units that the firm receives
when it reorders. This means that the average number of units the firm
has in its inventory (assuming that the sales of the units are spaced evenly
throughout time) is U/2. There are costs associated with maintaining a unit
of inventory in terms of protection, storage, and so forth. Let c represent the
cost of maintaining a unit of inventory for one year. Hence ·U/2 is the total
cost of maintaining an average inventory of U/2 units.
Reordering cost. Assume that two separate types of costs are associated
with reordering to replenish the inventory. The first type of cost is fixed in
nature and does not vary with the size of the order. The cost of recording
an order (which presumably does not depend on the size of the order) is an
example of this type of cost. We represent this fixed cost by the letter f . The
second type of cost varies directly and proportionately with the size of the
order and covers the incremental cost of shipping and packaging each unit
in the order.22 Let b refer to the incremental cost associated with reordering
each of U units.
The total cost of reordering in a specific instance is equal to the sum f +b·U .
Since a total of Q units is eventually needed for sale, and U units are reordered
each time, a total of Q/U reorders are made during the year. This means that
the total cost of reordering during the entire year is given by (f +b·U )·(Q/U ).
Total cost (storage and reordering). The total inventory cost (TIC)
associated with storing and reordering is T IC = c·U/2+U +(f +b·U )·Q/U ,
which can be rewritten as T IC = c · U/2 + f · Q/U + b · Q. Note that b is
not a coefficient of U .
To determine the cost-minimizing number of orders requires that we find the
optimal value of U (the order size), because the number of orders placed
is Q/U . Thus, we require the solution to dT IC/dU , which is dT IC/dq =
c/2 − 2 · f · Q/U 2 . Setting this expression equal
p to zero and solving yields
the expression U 2 = 2 · f · Q/c, so that U = (2 · f · Q/c). This result is
commonly called the square root law of inventory management. Because b is
22
Proportionality is not required. If the relationship between order size and order cost
were more complicated, the analysis would proceed in the same way. The exact nature of
the solution would, of course, differ somewhat.
CHAPTER 7. OPTIMIZATION 199

not a coefficient of Q, its value has no effect on the optimal inventory level
(or, equivalently, the optimal number of orders per year).
To confirm that we have found a cost-minimizing order size, evaluate the sec-
ond derivative, which is d2 T IC/dU 2 = 2QfU3
. All terms in this expression are
positive, so the value of U that satisfies the first-order condition corresponds
to a minimum value of T IC.
We determine the optimal value of U , given the following: f = 500, Q =
60000. That value is U ≈ 1732.05, which implies that the optimal number
of orders, N , is 34.64. Of course, the actual number must be an integer, and
might be constrained by provider-imposed restrictions. To see what other
values of N imply, substitute N = Q/U into T IC, so that

2N 2 f + Qc + 2N Qb
T IC = .
2N
Using the coefficient values above and b = 1 generates the relationship that
Figure 7.18 depicts.

Figure 7.18: TIC vs. N

For this particular set of parameter values, T IC is not very sensitive to N


CHAPTER 7. OPTIMIZATION 200

for N > 25 or so. Of course, this result is due to the parameters this example
and is not general.

7.5.7 The Optimizing Firm


The applications section of Chapter 6 developed an example of the output-
maximizing, cost-minimizing firm in detail. This section demonstrates how
we can use the more sophisticated tools of this chapter to extend the prob-
lems we introduced earlier. We revisit the firm that produces an output by
combining two inputs, capital (K) and labor (L). The per-unit cost of the
two inputs are r and w respectively.23
We look at the firm’s behavior from two equivalent and complementary view-
points. First, we suppose that the firm has a fixed budget (cost that it can
incur) and determine the conditions that must pertain if that budget is to
be used to produce that maximum output. Then we suppose that the firm’s
output level is pre-determined and we determine the conditions that must
pertain if that output level is to produced at the lowest possible cost. It will
come as no surprise that the conditions in the two cases are the same. This
result reflects the duality of these two approaches.
Maximizing Output Subject to a Cost Constraint
The production function for a representative firm that takes the form Q =
f (L, K), where Q =output, L = labor, and K = capital. The representative
firm has C0 dollars to spend on inputs and faces a cost constraint given by
C0 = w · L + r · K. The task of the firm is to maximize f (K, L) subject to
the constraint. We construct a Lagrangian function that reflects these facts:
W (L, K, λ) = f (L, K) − λ(C0 − w · L − r · K)
Setting the first partial derivatives of W with respect to L, K, and λ equal
to zero, we obtain three first-order conditions:

• W L = fL − λ · w = 0

• W K = fK − λ · r = 0

• Wλ = C0 − w · L − r · K = 0
23
The two-input limitation is for demonstration only. The methods easily extend to
any number of inputs.
CHAPTER 7. OPTIMIZATION 201

The first two conditions above imply that fL .w = λ and fK /w = λ, so


fL /w = fK /r. This condition implies that at the marginal the return per
$ must be the same for all inputs. Alternatively, we can state the results
as implying that fL /fK = M P L/M P K = w/r, that the marginal rate of
technical substitution between the two inputs must equal the ratio of the
marginal costs of employing the inputs.
Finally, the condition λ = fL /w = fK /r provides an interpretation of the
Lagrangian multiplier in this setting. Here, λ is the marginal product per
dollar spent. That is, it is the change in output per (small) change in cost.
Cost Minimization Given a Level of Output
Assume that the representative firm wishes to minimize the cost of producing
a certain level of output Q0. That is, the firm wishes to minimize its cost
function C = w · L + r · K subject to the constraint that it must produce
Q0 units of output, where Q = f (L, K). The Lagrangian function to be
minimized is V (L, K, µ) = w · L + r · K + µ · (Q0 − f (L, K)). Setting the
first partial derivatives of V with respect to L, K, and µ equal to zero, we
obtain this set of first-order conditions.

• VL = w − µ · fL = 0

• VK = r − µ · fK = 0

• Vµ = Q0 − w · L − r · K = 0

The first two first-order conditions yield implications that are identical to
the implications of the first-order conditions that we derived above. Also, we
can determine that µ = 1/λ = w/M P L = r/M P K. For this firm w is the
marginal input cost of labor, the cost of acquiring one more unit of labor,
and M P L, as always, the the change in output from a small change in L.
Therefore, the ratio of the two is the ratio of the cost change to the output
change, which is the marginal cost. Therefore, µ = M C.
Profit Maximization
Now, we can reexamine the implications of profit maximization. We have
established that profit maximization requires producing a quantity such that
M R = M C. We now know that M C = w/M P L = r/M P K if the firm’s
output is to be produced in a least-cost fashion. These conditions combine to
imply that the firm must employ labor and capital at rates such that M R =
CHAPTER 7. OPTIMIZATION 202

w/M P L = r/M P K. This implies that the profit-maximizing firm employs


the two inputs at rates such that w = M R · M P L and r = M R · M P K.
These are the marginal revenue products of the two inputs, M RP L and
M RP K. If the firms are price-takers in their output markets p = M R and
the employment levels are such that w = p · M P L and r = p · M P K. These
are the values of marginal products of the two inputs, V M P L and V M P K.24

7.6 Questions and Problems


1. Find the extrema, if such exist, and determine whether each extremum
is a maximum or a minimum.

a. z = f (x, y) = 2 · x3 + y 2
b. z = f (x, y) = x2 + x2 · y + y 2
c. z = f (x, y) = x3 − x · y 2
d. z = f (r, s) = r + 2 · r2 + s − s3
e. z = f (P, Y ) = 12 · P · Y − P · Y 2
f. z = f (L, K) = 10 · L0.75 · K 0.25

2. Heinz Westphal, Vintner, imports Rhein and Mosel wines. The value of
the wine V √
increases as time passes according to the following formula:
V = 6 · 2.5 t , where t = ageing time in years. The present value of
the wine P V , given a discount rate r and continuous appreciation, is
P V = V · e−r·t .

(a) How long should Westphal hold the wine before selling it in order to
maximize the present value of the wine? That is, what t maximizes
P V ? State your solution as a general expression in terms of r.
(b) If r = 0.08, what is the corresponding t that maximizes P V ?

3. The Des Moines Packing Company has a total cost (T C) function of


the form T C = f (M, L), where M = unbutchered meat in pounds and
24
If the firm is a price-searching in one or more input markets, then the conditions for
profit maximization change slightly.
CHAPTER 7. OPTIMIZATION 203

L = hours of labor. Specifically, T C = 3 · M + 7 · L. The production


function for finished, butchered meat (Q) in pounds is Q = 2M 0.5 · L0.5 .
The Des Moines Packing Company wishes to produce 10,000 pounds
of finished, butchered meat in this time period.

(a) Find the quantities of M and L that minimize the cost of doing
so.
(b) With respect to the production function, do diminishing marginal
returns exist with respect to M , L, or both?
(c) Does the production function exhibit increasing, decreasing, or
constant returns to scale? (d) Does Euler’s theorem apply?25

4. Determine (if possible) whether the following functions are monotoni-


cally increasing or decreasing. If neither, determine whether points of
inflection exist and, if so, where; or whether extreme points exist and,
if so, where; and whether the extreme points, if any, are maxima or
minima.

a. y = f (x) = 2 · x2
b. y = 6 + 0.15 · x
c. y = 6 · x2 + 2 · x + 1
d. y = a + b · x + c · x2 + d · x3
e. y = 10 + 5 · x + 2 · x2 − x3
f. y = sin(x)
g. y = sin(x2 )
h. y = 15 − x + 2 · x2 + x3
i. y = 100/x2
25
At first glance, this question should bother you. Isn’t the ratio of meat to finished
meat close to 1? Econometricians estimating production functions like this one run into
a “dominant variable” problem. One of the variables, like unbutchered meat here, varies
so closely with output that the impacts of the other variables cannot be detected very
accurately. This problem is one of detection and estimation, however, and not one of
existence. Just as butchered meat cannot be produced without unbutchered meat, it
cannot be produced without other inputs like labor and capital. It is possible that the
elasticity of subsitution is much lower than 1, which is the elasticity of substitution in this
example. Replacing this function with, say, a constant-elasticity-of-substitution function
would not change the nature of the analysis, just its difficulty.
CHAPTER 7. OPTIMIZATION 204

5. An electric power company has two generator plants, which we label


A and B. The total cost functions in each plant are given by T CA =
8 + 5QA − QA2 + O.5 · QA3 , and T CB = 2 + 2 · QB + QB 2 , so T C =
T CA + T CB, where T C = total cost. Q = QA + QB is kilowatt hours
generated (in thousands). The company wishes to minimize the cost
of generating any given amount of electricity. How should it allocate
production between the two generating plants if it must produce 10,000
kilowatt hours (Q = 10)?

6. Harold Hedonist has a utility function of the form U = QA · QB, where


U = utility, and QA and QB are quantities of two different goods or
services. Mr. Hedonist, who has $100 to spend in this time period, faces
parametric prices such that P A = $1 and P B = $2. Use the Lagrange
multiplier technique to determine the utility-maximizing quantities of
QA and QB.

7. Assume that f (x) is a monotonic transformation of x given that f (x1 ) >


f (x0 ) whenever x1 > x0 was previously greater than x0 . Monotonic
transformations are order-preserving. With respect to ordinal util-
ity maximization, the maximization of a monotonic transformation of
a utility function yields exactly the same results as the maximiza-
tion of the original utility function. Demonstrate that maximizing
U = QA2 · QB 2 subject to 100 = QA + 2 · QB yields the same re-
sults as those found in the preceding problem.

8. The West Mifflin Ford dealership expects to sell 1000 new Mustangs
during the next year. These sales will be evenly spaced throughout the
year. The cost of storing an unsold Mustang for 1 year is $1500. The
cost of placing a new order for new Mustangs is $700 plus $250 per new
automobile ordered.

(a) (a) What is the optimal size of order that West Mifflin should place
when it orders new Mustangs?
(b) How many such orders should West Mifflin place during this year?
(c) Determine total inventory cost for the two integer values nearest
to the computed value.
(d) How much would total inventory cost change if Ford allows no
more than 24 orders per year?
CHAPTER 7. OPTIMIZATION 205

9. The state of Taxonia wishes to maximize the total tax revenue T that it
receives from a per-unit tax of amount t per unit that it is going to place
on the output of Monopoly, Inc., to which Taxonia grants monopoly
status. The total revenue T R function of Monopoly, Inc., is given by
T R = 6 · Q − Q2 , while its total cost (T C) function in the absence of
a sales tax is given by T C = 2 · Q, where Q is units of output.

(a) What tax per unit will maximize total tax receipts for the state of
Taxonia?
(b) How much tax revenue (T) will this tax raise?
(c) What are Monopoly Inc.’s profit-maximizing price and output in
this situation?
Chapter 8

Integral Calculus

The previous four chapters addressed differential calculus. For a function


such as y = f (x), we learned how to find a new function, dy/dx, which we
termed the derivative of the function y with respect to the variablex We
found that the first derivative of a function is the slope of the function at a
particular point.
This chapter introduces the second main branch of the calculus: the integral
calculus. The integral calculus is distinctive in two specific ways. First, it
enables us to define and measure the concept of area; for example, the area
under a curve. As with the differential calculus, we apply the limit notion
when we work with an integral. A few of the very many applications of the
integral calculus to business and economics include: measuring consumer
surplus, determining the total amount of depreciation that a firm will realize
in a specific time span, measuring the deadweight loss due to monopoly, and
finding total product when one knows only marginal product or total variable
cost when one knows marginal cost.
The second way in which the integral calculus is notable is that the technique
of integration is operationally the inverse of differentiation. Whereas in the
differential calculus a function is given and one must find the corresponding
derivative, in the integral calculus the derivative of the function is given
and we must work backward to find the original function. This relationship
is useful, for example, when we have some knowledge of the expression for
marginal sales revenue, but seek an expression for total sales revenue. The
backward direction of this process of integration is the reason that an integral

206
CHAPTER 8. INTEGRAL CALCULUS 207

of a function is often called the antiderivative of that function.

8.1 The Definite Integral


We begin our analysis by considering the definite integral, which is defined
as follows: If F (x) is a function such that dF (x)/dx = F 0 (x) = f (x) is its
derivative for a given interval on the x axis, then F (x) can be defined to be
the antiderivative or integral of f (x).
R
The process of integration is symbolized as follows: f (x)dx) = F (x) + C.
The left-hand side of this
R equation is read, “the integral f of x with respect
to x.” The elongated symbol is an integral sign (which, as we shall soon
see, implies the summation of continuous values). We call f (x) the integrand.
That is, f (x) is the function that is being integrated. The symbol dx indicates
that we are integrating with respect to a variable x. On the right-hand side
of this equation, F (x) + C is the indefinite integral, while C itself is any
arbitrary constant of integration.
The inclusion of the dx symbol on the left-hand side of this equation may
seem superfluous. However, we cannot omit it, since it indicates the variable
with respect to which we are integrating. If dx were absent, then the equation
would be just as incomplete as a derivative that was written df (x)/d. One
might presume that the differentiation is taking place with respect to x.
However, that should not be assumed and may not be true.
In the process of finding the antiderivative or integral of a given function f (x),
we produce another function of x, namely F (x) + C. This, too, is analogous
to the process of differentiation, where we differentiated a function of x and
thereby generated a second function that was, in general, also a function of
x.
When we find the antiderivative or integral of a function, it will in general
not be unique. That is, many alternative functions could have the same
derivative. For example, when y = F (x) = 2 · x, dy/dx = 2; however, when
y = F (x) = 2 · x + 1000, dy/dx = 2 also. Thus, 2 · x is an antiderivative of
2, and 2 · x + 1000 is also an antiderivative of 2. More generally, so also is
2 · x + C where C is any constant, whether or not C’s value is specified.
Hence, an infinite number of antiderivatives (integrals) can be associated
with a particular derivative. By way of contrast, in the past four chapters,
CHAPTER 8. INTEGRAL CALCULUS 208

when we found the derivative of a function, that derivative was unique. For
example, if y = F (x) = 2 · x2 , then dy/dx = 4 · x, which is unique. There is
no other value or function that is the first derivative of the function y = 2·x2 .
We stress the non-uniqueness of an antiderivative (integral), because it is
important to the understanding of the process of integration. In general, the
indefinite integral of f (x) = 2 · x would be

dF (x) d(x2 + C)
F (x) = x2 + C, for = = 2 · x = f (x).
dx dx
Geometrically, y = x2 + C represents a family of curves that are parallel to
one another, but have a vertical displacement from one another.
Figure 8.1 illustrates such a family of curves for the function F (x) = x2 + C.
Unless we know the value of the arbitrary constant C, we cannot determine
the unique antiderivative of a given function. When additional information
is supplied concerning the value of the constant C, we state that the initial
conditions or boundary conditions have been specified. In the example de-
picted in Figure 8.1, if we are given the initial condition that x = 0, and
the value F (x) = F (0) = 3, then the value of the constant C is determined.
F (0) now equals F (x) + C = F (0) + C = 3, so 02 + C = 3, and C = 3. Thus
F (x) = x2 + C becomes x2 + 3. Figure 8.1 shows F (x) + C for three values
of C: 0, 20, and -20.

Figure 8.1: Graphical representation of the constant of integration


CHAPTER 8. INTEGRAL CALCULUS 209

The F (x) term of the indefinite integral F (x) + C is entirely a function of x


and has no definite numerical value. That is why F (x) + C is referred to as
the indefinite integral. In the absence of additional information, the value of
F (x) + C is unknown and indefinite.
You must pay careful attention to the notation you use when you are finding
an antiderivative or integral. Whereas f 0 (x) denoted a derivative of a function
in the differential calculus, F (x) is the antiderivative or original function in
the integral calculus, and f (x) now refers to the derivative.
R
In describing the process of integration in the equation f (x)dx) = F (x)+C,
we used a function of the form y = f (x). Functions involving the letter x
are customarily used to illustrate the process of integration, just as we used
functions involving the letter x to illustrate the process of differentiation.
There is nothing magical about the symbol x. We could use any other letter,
such as, s, t, u, v, or w, to illustrate integration with equal validity.

8.2 Rules and Properties Relating to the In-


tegral
Chapter 5 presented a series of rules that greatly simplified the task of finding
the derivative of a function. This section states a series of rules that will help
you integrate a wide variety of functions.
Differentiation and integration have an inverse relationship to each other, so
many of the rules relating to integration are closely related to the rules for
differentiation. Specifically, it is often the case that one need only reverse a
specific rule for differentiation to get the needed rule for integration. Never-
theless, this is not always the case. Some integrals are not easy to evaluate.
It is customary to use published tables to assist one in evaluating integrals
because of the multitude of mathematical forms that are involved.1 In the
next few pages, therefore, we consider only a few of the rules that you can
use to find integrals.
1
Enter “integral table” into your browser’s search engine to get a list of tables. Com-
puter algebra systems like Maxima contain extensive tables.
CHAPTER 8. INTEGRAL CALCULUS 210

8.2.1 The Power Rule


1
R R
Suppose that f (x) = xn . Then f (x)dx = xn dx = n+1 · xn+1 + C = F (x).
We prove this directly by taking the derivative of F (x). This proof simply
applies the definition of the integral as the antiderivative. Remembering that
the differential of y = f (x) is dy = fx dx, the differential of the right-hand
side of the power-rule equation above is

1

d · xn+1 + C
n+1 n+1 n
= · x = xn , so fx dx = xn dx.
dx n+1
The proof above correctly suggests that the derivative of an integral must
always be equal to the integrand. That is, if the correct integration has been
performed, d(F (x) + C)/dx must be equal to f (x).
Examples
R 1 d(x6 /6+C)
1. x5 dx = 6
· x6 + C Check: dx
= x5
2 /2+C)
xdx = 12 · x2 + C Check: d(x
R
2. dx
=x
R R R d(x1 /1+C)
3. dx = 1dx = x0 dx = x + C Check: dx
= x0 /1 = 1
R√ d( 2 ·x3/2 +C) √
xdx = x1/2 dx = 23 · x3/2
R
4. Check: 3 dx = x1/2 = x
R 1 R −2
5. x 2 · dx = x dx = − x1 + C
−1
Check: d(−(1/x)+C)
dx
= d(−xdx +C) = x−2 = 1/x2

The power rule of integration explicitly requires that n 6= −1. The following
example demonstrates why Rthis restriction is necessary. Let us try to find
R −1
the integral of f (x) = l/x. 1/x)dx = x dx = (1/0) · x0 . Hence, when
n = −1, the power rule no longer applies because the integral is undefined
due to division by 0. The following rule deals with this type of situation.

8.2.2 The General Logarithmic Rule


The general logarithmic rule states that x1 dx = x−1 dx = log(|x|) +
R R

C, for x 6= 0. Here, log refers to the natural logarithm.


CHAPTER 8. INTEGRAL CALCULUS 211

To prove this result, note that the differential of the right-hand side of the
equation describing the general logarithmic rule is d(log(|x|) + C), which is
equal to (l/x)dx. We started our integration with (l/x)dx on the left-hand
side, so the general logarithmic rule is proved.
The antiderivative in the general logarithmic rule constains an absolute-value
sign. This is used because logarithms do not exist for negative values of any
variable x. When working with a problem in which we are certain that
the domain of a variable consists only of positive values, we may omit the
absolute-value sign.

8.2.3 The General Exponential Rule


Suppose that
R a baseaa is raised to the x power. That is y = f (x) = ax . In
x
this case ydx = log(a) + C, where log denotes the natural logarithm. To
prove this assertion, recall that d(ax )/dx = ax · log(a).
An important special case of the general exponential rule is Rthe rule for which
a = e the base of natural logarithms. Because log(e) = 1, ex dx = ex + C.
Before we develop additional rules that facilitate dealing with exponential
functions that are not of the precise form y = ax or y = ex , we confront a
problem that often confuses people who are learning about integration. We
have already mentioned that in order to get rid of the arbitrary constant of
integration, initial conditions (boundary conditions) must be specified. In
many cases theR initial condition is the value of the arbitrary constant itself.
For example, 2·xdx = x2 +C. If the initial condition is x = 0 and F (0) = 3,
then the constant of integration would also be 3: F (0) = 02 +C = 3, so C = 3.

It is possible, however, for the initial condition to have a value other than
that of the constant of integration. Exponential functions
R x sometimes furnish
examples of this phenomenon. Consider
R the integral e dx with the initial
condition F (0) = 3. Thus F (x) = ex dx = e0 +C. Hence F (0) = e0 +C = 3,
or 1+C = 3, and so C = 2. That is, F (0) = 3 6= C = 2. Therefore we should
not always assume that the constant of integration and the initial condition
of the function are identical.
CHAPTER 8. INTEGRAL CALCULUS 212

8.2.4 Important Properties of Integration


In order to progress in our evaluation of integrals, we state some important
theorems that enable us to develop further useful techniques and rules for
integration.
The additive property is this: If both f (x) and g(x) are integrable, then
the integral of u(x) = f (x) + g(x) is the sum of the integrals of f (x) and
g(x). More generally, the integral of a sum of a finite number of functions is
equal to the sum of the integrals.
Z
(f (x) + g(x))dx = F (x) + C1 + G(x) + C2 = F (x) + G(x) + C,

where C = C1 + C2, and more generally,


n
Z X n
X
fi (x)dx = Fi (x) + C,
i=1 i=1

where
Pn Fi (x) is the antiderivative for fi (x) for all i = 1, 2, . . . , n and C =
i=1 Ci , the sum of the arbitrary constants of integration.
R R
The multiplicative property is this: K · f (x)dx = K · f (x)dx =
K · F (x) + C, where K is any constant.
The linearity property combines the additive and multiplicative properties
to this: a linear combination of n functions has an integral that equals the
linear combinations of the integrals of the individual functions.

n
Z X n
X
Ki · fi (x)dx = Ki · Fi (x) + C,
i=1 i=1

where Fi (x) is the antiderivative for fi (x) for all i = 1, 2, . . . , n; C = ni=1 Ci ,


P
the sum of the arbitrary constants of integration; and Ki is the coefficient
for fi (x).
Examples Confirm each by taking the derivative of the integral.

x2 dx = − 34 x3 + C
R R
1. −4 · x2 dx = −4 ·
CHAPTER 8. INTEGRAL CALCULUS 213
R R R R
2. (3 · x2 − 5 · x + 1)dx = 3 · x2 dx − 5 · xdx + 1dx =
x3 + C1 − 52 · x2 + C2 + x + C3 = x3 − 25 · x2 + x + C
R R
3. (8/x)dx = 8 · (1/x)dx = 8 · log(|x|) + C

4. (2 · ex + x−2 )dx = 2 · ex dx + x−2 dx = 2 · ex − x1 + C


R R R

8.2.5 Integration by Substitution


Our rules of integration, and the theorems stated above, deal with relatively
uncomplicated integrands. As we have seen, however, more complicated
integrands do exist and cannot be handled by the rules and theorems. We
need a process by which we can transform a complicated integrand into the
simple integrands utilized in the rules and theorems. One such process is
known as the substitution method of integration, which “substitutes” a new
variable of integration for the original variable. The object of the substitution
is to transform the complicated integrand into one of the simple integrands
with which our rules and theorems can deal.
The technique of integration byR substitution is applicable whenever we can
transform the original integral f [g(x)]dx as follows:
Z Z Z Z
0 du
f (x)dx = f [g(x)] · g (x)dx = f (u) · dx = f (u)du.
dx

The substitution involved requires the replacement of g(x) by u and the


0
Rreplacement of g (x)dx by
R du. This substitution transforms the operation
dx into the operation du. Now, integrating with respect to the variable
u, yields an indefinite integral that is a function of u, such as F (u) + C.
We can then transform this indefinite integral back into a function of the
original variable, x, by making the opposite substitution, that is, replacing
u with g(x) and replacing du with g 0 (x)dx. This accomplishes the needed
integration by means of substitution.
We can now rewrite our four integration rules for the cases in which substi-
tution is carried out:

un+1
R
1. The power rule: un du = n+1
+ C, f or n 6= −1
CHAPTER 8. INTEGRAL CALCULUS 214

1
R
2. The general logarithmic rule: u
du = loge |u| + C
au
R
3. The general exponential rule: au du = loge a
+C
R
4. The exponential rule, base e: eu du = eu + C
Integration by substitution is connected to the use of the chain rule in differ-
entiation. Integration is, as we have pointed out, the reverse of differentiation.
This means that when we introduce a new function u = g(x) in the process
of integration, the usual checking process (by which we ascertain whether
our integral is correct) must utilize the chain rule of differentiation. That is,
since integration by substitution involves the introduction of a new function
u, which is a function of x, the checking process must use the function-of-a-
function rule (the chain rule) in order to return us to the original function.
Examples
R
1. Evaluate the integral 2 · (e2·x + 1)2 · e2·x dx.
Let u = e2·x + 1. Then, du/dx = 2 · eR2·x or dx = 2·e12·x du.
Our integral, stated
R in terms of u is 2 · u2 e2·x · du/(2 · e2·x ).
This simplifies to u2 du = 13 · u3 + C.
This, in terms of x is 31 · (e2·x + 1)3 + C.

A computer algebra system cannot determine what substitution works,


but it can reduce the amount of work and, more importantly, reduce
the chance of error. In this case, the following commands are entered:
y: 2*(e^(2*x) +1)^2*e^(2*x)$ ’integrate(y,x);
changevar(%, u=(e^(2*x) +1), u, x);
Note the ’ before integrate. This tells Maxima to state the result in
a “noun” form–that is, not to evaluate it. The changevar command
contains the noun form that the ’integrate command creates, the
definition of the u substitution, and the name of the original indepen-
dent variable, x.
The result consists of two items. The first is the unevaluated integral
2
2 %e2x %e2x + 1 dx (recall that %e is Maxima’s notation for the
R

constant
R e = 2.718 . . .), and the second is the integral stated in terms
of u, u2 du . If you want Maxima to evaluate the result of the out-
put above, enter this command: integrate(u^2,u);, producing the
3
result u3 . Be aware that Maxima does not generate the constant of
integration; you must keep in mind that it exists.
CHAPTER 8. INTEGRAL CALCULUS 215
R
2. Evaluate 3 · x2 · (x3 − 4)2 dx. Let u = x3 − 4, so that du/dx) = 3 · x2 ,
or dx = du/(3 · x2 ).
With
R the pertinent substitutions,
R
3 · x2 · (x3 − 4)2 dx = 3 · x2 · u2 · (du/(3 · x2 )) = u2 du .
This final expression is evaluated as u3 /3 + C = (1/3) · (x2 − 4)2 + C.
dx
3. Evaluate x−2 . Let u = x − 2, so that du/dx = 1 or du = dx.
Now, x−2 = du
R dx R
u
= log(|u|) + C = log(|x − 2|) + C.

The three examples above indicate that an appropriate substitution consists


of two parts that are related to each other. One part of the substitution is
the derivative of the other part of that substitution. It is also possible that
integration by substitution may result in a constant multiple of f (u)du. As
the examples below demonstrate, however, this does not present a problem.
The homogeneous property allows us to factor this constant multiple and to
place it in front of the integral sign.
Examples
R
1. 24·xRdx . Let u =R4 · x so du/dx = R4 or dx = du/4.
Now 24·x dx = 2 eu du/4 = (1/2) eu du
= (1/2) · eu + C = (1/2) · e4·x + C.
R dx
2. 2·x−5 Let u = 2 · x − 5, so dx = (1/2) · du.
Now 2·x−5 = u1 · 12 · du = 12 · u1 · du =
R dx R R
1
2
· log(|u| + C = 21 · log(|2 · x − 5|) + C.
R
3. K 4·x dx, where K is a constant. Let u = 4 · x so that du/dx = 4 or
dx = 41R· du.
Thus, K 4·x dx = K u · 14 · du = 41 K u du =
R R
1 1 K 4·x
4
· K u · log(K) + C = 4·log(K) + C.

In order for the technique of integration by substitution to work, we must


always completely transform the original integrand from one that involves one
variable, say x, to a completely different function involving another variable,
say u. If substitution is impossible, or is carried out improperly so that a
function or functions of two or more variables results, then we must try a new
substitution, for there is no general way in which we can find the integral of
this new quantity.
CHAPTER 8. INTEGRAL CALCULUS 216

Example

Evaluate x2·x−3
R
2 −3·x dx. Try this: Let u = 2 · x − 3 so that du/dx = 2 or
dx = Rdu/2.
Now x2·x−3
R u
2 −3·x dx = x2 −3·x
· 21 · du, which is not integrable, since the sub-
stitution created a new function with two variables, x and u. The proper
substitution should have been u = x2 − 3 · x.
du
Now du/dx = 2 · x − 3) so dx = x−3 . Complete the steps to confirm that the
2
integral is log(|u|) + C = log(|x − 3 · x|) + C.
It is often possible to decide on the appropriate substitution by simple obser-
vation of the original integrand. That ability, however, usually means that
you have acquired the knowledge and foresight that seem to come only with
experience, some trial and error, and hard work. Integration is generally
considered to be a more difficult process to master than differentiation. The
correct way to integrate a function is not always readily apparent. Also, if the
substitution is carried out improperly so that a function or functions of two
or more variables results, then you must try a new substitution. There is no
completely general way to find the needed integral by means of substitution.
All of these difficulties are reasons why tables of integrals are so useful, and
one of the reasons that a computer algebra system can be a useful asset.2
Exercises 8.1
Integrate the following integrals. Determine how many of these integrals
Maxima can evaluate directly, without the use of the changevar command.

1. (4x3 − 3x2 + 2x − 6)dx 10. (1 − x) dx


3 5
2. ( x −3x+2
x2
)dx 11. √ x 6 dx
2 (1−x )
3. (4x√− 3) dx 2
12. x (1 − x3 ) dx
2
4. 4x 2x2 + 1 dx
x 13. e−x dx
5. x2 −6
dx e1/x
√ 14. x2
dx
6. 2x + 1 dx x2 +4
15. xe dx
7. √ x dx 2 x3
1−x2
x+1
16. x e dx
8. x2 +2x+3
dx 17. e3x
dx
e3x +3
9. (2x − 5)2 dx 18. 8x
a dx

2
Many online integral tables are available. For example: http://integral-table.com/.
CHAPTER 8. INTEGRAL CALCULUS 217

8.2.6 Integration by Parts


Another method used to transform complex, seemingly unworkable inte-
grands into more workable forms is integration by parts. Just as integration
by substitution was seen to be the inverse of the chain rule for differentia-
tion, the technique of integration by parts may be viewed as the inverse of
the product rule for differentiation.
R
Integration by parts is applicable whenever the original integral g(x)dx can
be transformed as follows: f (x) · g 0 (x)dx = f (x) · g(x) − f 0 (x) · g(x)dx + C.
R

This expression can be converted into a more abbreviated form by making


the following substitutions: u = f (x), v = g(x), du = f 0 (x) dx, and dv =
g 0 (x) dx.
The expression now becomes u dv = v · u − v du + C.
In review, the procedure that we must follow when attempting to Revaluate
an integral by parts is to transform any original integral of the form g(x)dx
into an integral in which the only term to evaluate is f 0 (x)g(x) dx. If we
have made appropriate choices when we substituted for f (x) and g(x), then
the transformed integral will be easier to evaluate than the original.
Examples
R
1. Evaluate x · ex dx. Let u = x and dv = ex dx. Then du = dx and
x
Rv = e x. R
x · e dx = x · ex − ex dx = x · ex − ex + C.
The Maxima command integrate(x*%e^x, x) generates the result
(x − 1) · ex , the same as above except for the constant of integration.
R
2. Evaluate log(x)dx. Let u = log(x) and dv = dx. Then, du = 1/x
and
R v = x. R
log(x)dx = x · log(x) − x · (1/x)dx = x · log(x) − x + C.

Unfortunately no general rule dictates the best way to transform a complex


integral into a more pliable one. However, a few hints can reduce the difficulty
of what is otherwise often a frustrating procedure.
First, when we make a substitution, we know that f (x) and g 0 (x)dx are the
two terms that make up the left-hand side of the equation. g 0 (x)dx is the
differential of g(x). This means that we should choose the differential of g(x)
CHAPTER 8. INTEGRAL CALCULUS 218

for substitution purposes. However, that differential must be integrable so


that we can find g(x), which is part of the right-hand side of the equation.
Second, the object of transforming the integral is to produce an integral
that is more amenable to ordinary rules of integration. Hence we should
choose the most complicated substitution that is possible, yet integrable, for
g 0 (x)dx. That is, a more complex substitution that is integrable is preferred
to a simple substitutionR that is also integrable. The more complete the
substitution, the easier f 0 (x) · g(x)dx, the transformed function, will be
to integrate. In everyday terms, you should accomplish as much as possible
with the substitution.
Fortunately, tables of integrals can reduce the difficulty of this process,
though practice to learn the general nature of functional forms is required.
Perhaps more fortunately, computer algebra systems routinely contain large
tables of integrals and can integrate many of the expressions that otherwise
would take much time and subject one to a significant risk of error.
Exercise 7.2
Evaluate the following integrals. Confirm that Maxima can integrate all of
these expressions.
R R
1. R x · log(x)dx 5. R x2 /e3·x dx
2. R log(2 · x)dx 6. R (x + 4) · log(x)dx
2
3. R x · e 2·x+3
dx 7. x · e−x dx
4. x · e2·x dx
R p
8. log(x)/ (x)dx

8.3 Applications of the Indefinite Integral


We now look at several examples in which a marginal function, represented
by a first derivative, is known. For example, we know marginal cost, which
is dT C/(dQ. What we do not know is the total cost function itself, which a
is T C = f (Q) In brief, we shall now examine situations in which we know
the derivative of a function, and can use that knowledge to construct the
function itself.
CHAPTER 8. INTEGRAL CALCULUS 219

8.3.1 Marginal Cost and Total Variable Cost


Marginal cost MC, the addition to total cost T C (and to total variable cost
T V C) that occurs when an incremental unit of output is produced, is given
by M C = dT C/dq = dT V C/dQ.
We can find the total cost function, where TC = C(Q),R by integrating the
marginal cost function with respect to output: T C = M CdQ = T V C(Q)+
K, where K is the constant of integration. Given the economic setting of
this example, K is the firm’s per-period fixed cost.
2
As an example, let M CR = 25−10·Q+3·Q
R and fixed cost = 100. Then total
cost can be written as M CdQ = (25 − 10 · Q + 3 · Q2 )dQ = T V C(Q) + K
which implies that T C = 25 · Q − 5 · Q2 + Q3 + 100.
We can check this result by observing that M C = dT C/dQ = 25−10Q+3·Q2 .

8.3.2 Marginal Revenue and Total Revenue


Marginal revenue M R is the change in sales revenue a firm obtains when the
quantity that is sells changes by one (small) unit of output: M R = dT R/dQ,
where T R = total sales revenue in dollars. If we know the form of the
marginal revenue function, then we can find the total revenue function by
integrating
R the marginal revenue function with respect to output: T R =
M R dQ = R(Q) + C.
The arbitrary constant of integration C has a value of 0 in a total revenue
function. This recognizes the fact that usually the firm will realize no revenue
if it does not sell any of its output. We can therefore state the total revenue
function as T R = R(Q).
As an example, let the marginal revenue function
R M R = 25 − 2 · Q. Then it
follows that total revenue is given by dT R = (25−2·Q)dQ = 25·Q−Q2 +C,
which is T R = 25 · Q − Q2 because C = 0.
We can check this result by observing that dT R/dQ = 25 − 2 · Q, which is
our original M R function. Also, note that the average revenue, AR = P =
25 − Q (the inverse demand curve), has one-half the slope of the M R curve,
a characteristic of the linear demand (and marginal revenue) curve that we
have noted before.3
3
The inverse demand curve is the same as the average revenue curve when, and only
CHAPTER 8. INTEGRAL CALCULUS 220

8.3.3 Demand Functions, Total Revenue, and Price


Elasticity of Demand
Suppose that we have a good estimate of the point price elasticity of demand
Ep and that we are confident that this elasticity will remain much the same
over some relevant range of prices. Thus our analysis can proceed as if we
know that Ep = (dP/dQ) · (P/Q) = (dP/P ) · (P/dP ) = K, where K is the
constant value of Ep.

R dQ/Q = K · dP/P . We can integrate both sides


Some rearrangement yields
so that ∈ dQ/Q = K dP/P . The general logarithmic rule implies that
log(|Q|) + C1 = K · log(|P |) + C2. Both price and quantity are positive
values, so the absolute values can be replaced with actual values. Rewrite
this expression as log(Q) = K · log(P ) + C3, where C3 = C2 − C1.
Exponentiation of both sides of this expression yields Q = eC3 · P K , a general
expression for the constant-elasticity demand curve. We can rewrite this
expression as Q = C · P K , where C = eC3 and K = Ep, the elasticity of
demand.
Suppose that K = −1.5, P = 2, and Q = 25. Then 25 = C · 21.5 Solving
25 = C · 2−1.5 for C yields P ≈ 70.7. Therefore, the demand curve is
approximately Q = 70.7 · P −1.5 over the relevant range of values.4

8.4 The Definite Integral


The integral calculus was introduced in order to measure the area under a
curve. Archimedes (287 -212 B.C.) successfully utilized the “method of ex-
haustion” in order to find the approximate area contained in a region by
placing inside that region a polygonal region that more or less approximated
the original region of interest. Successive polygonal regions were then intro-
duced, each with an additional side, in order to give a closer approximation
when, each unit of the product sells for the same price. That is, if any sort of price
discrimination is practiced, this relationship does not hold.
4
Confirm that M R > 0 over the range of values for which this expression is a reasonable
representation of the demand curve. Suppose that you are advising a seller and that the
seller tells you that Ep = −0.75, what advice would you offer. Note that Ep = −0.75
implies that M R < 0.
CHAPTER 8. INTEGRAL CALCULUS 221

Figure 8.2: The closed interval [a,b]

of the original region of interest. Eventually, if the process were carried out
long enough, the method of exhaustion would lead to a close approximation
of the area of a particular region.
We use the method of exhaustion to develop an intuitive and visual idea of
how the integral calculus is used to find the area under a curve. Instead of
using a many-sided polygon, we use a rectangle (a four-sided polygon).5
Figure 8.2 illustrates the continuous function y = f (x), where the domain
of the function is the closed interval [a, b]. The problem confronting us is to
calculate the shaded area, which is the area enclosed by the curve and the
abscissa between points a and b. We refer to this area as A.
As an illustrative approximation to the area defined above, we divide the
interval [a, b] into n subintervals (where n = 4 in our example) as shown in
Figure 8.3. Part (a) approximates the area under the curve by inscribing four
rectangles below the curve between points a and b. Part (b) approximates
the area under the curve between points a and b by inscribing four rectangles
from above the curve. The left-hand boundary of each rectangle in part (a)
has a minimum height of y = f (x), whereas the right-hand boundary of each
rectangle in part (b) has a height that represents the maximum value that
y = f (x) assumes in that subinterval.
5
Later, we look at numerical methods for estimating areas that cannot be determined
CHAPTER 8. INTEGRAL CALCULUS 222

Figure 8.3: Approximating the area using rectangles: (a) approximation from
below, and (b) approximation from above.
CHAPTER 8. INTEGRAL CALCULUS 223

The area of a rectangle is given by the product of the height and the width
of that rectangle. The first rectangle in Figure 8.3 has a height of f (x0 ) and
a width of ∆x0 = x1 − x0 . To generalize, the ith rectangle in part (a) has
a height of f (xi ) and a width of ∆xi . The area of the ith rectangle is given
by Areai = f (xi )∆x˙ i . The total area in the four rectangles between points
a and b in part (a) is given by
4
X
A−
n = f (x1 ) · ∆xi .
i=0

We can see that this is an underestimate of the total area under the curve
between points a and b.
In similar fashion, we can measure the area of each rectangle in Figure 8.3.
This measure, which yields an overestimate of the area under the curve be-
tween points a and b, is equal to
5
X
A+
n = f (x1 ) · ∆xi .
i=1

The two approximations to the area under the curve between points a and b
are labeled A− +
i (underestimate) and Ai (overestimate). It is apparent that
− +
An < A < An . The unshaded portions of the rectangles under the curve in
part (a) and above the curve in part (b) are responsible for the differences
between A− +
n , A, and An .

It is possible to achieve an even better approximation of the area under the


curve in the closed interval [a, b] by further subdividing that interval. Figure
8.4 illustrates the effects of increasing the number of subdivisions from 4 to
8. As n increases from 4 to 8, and ∆xi becomes smaller, areas A− and A+
differ less from each other, and also become closer approximations of the true
area A.
In the limit, as ∆x → 0, both areas A− and A+ approach the true area A.
That is, when the area of the inscribed rectangles in part (a) is equal to
the area in the circumscribed rectangles in part (b), we have found the area
under the curve A. This area is known as the Riemann integral or definite
integral.
analytically. Those methods typically use trapezoids.
CHAPTER 8. INTEGRAL CALCULUS 224

Figure 8.4: Approximating the area using rectangles: (a) approximation from
below, and (b) approximation from above.
CHAPTER 8. INTEGRAL CALCULUS 225


Definition: Let A+n be the upper estimate and An the lower estimate of the
area under the graph of y = f (x) when the interval [a, b] is divided into n
subintervals. If

lim A+n = lim An
n→∞ n→∞

the function f (x) is said to be Riemann integrable, and A is said to be the


Riemann or definite integral of f (x) on [a, b].
Rb
The expression a f (x) dx = A is read, “the integral of f (x) from a to b is
A.” The letters a and b signify the limits of integration. That is, the lower
limit or bound of the variable x is equal to a, and the upper limitR or bound
6
of variable x is equal to b. For example, the definite integral 2 f (x) dx
indicates that we shall integrate the function y = f (x) between the values of
2 and 6 for variable x.
Four matters relating to our definition of the definite integral warrant addi-
tional discussion.
R
• First,
P the integral sign ( ) functions in place of the summation sign
( ), which we have used so often previously in a wide range of different
contexts. The integral sign indicates that the number of terms (or
rectangles) to be summed is infinite–more P precisely that n → ∞. The
integral is, therefore, a special case of a -type summation.

• Second, the symbol representing change, ∆x, has now been replaced
by the integration notation dx and represents an infinitesimal change.

• Third, the indefinite integral with which we previously worked resulted


in a function of variable x, whereas the definite integral results in a
numeric answer that represents a specific area.
Rb
• Fourth, when we evaluate a f (x) dx, the constant of integration that
we encountered with the indefinite integral now disappears. This is
the second fundamental theorem of the calculus, which we will shortly
introduce, gives us the following result:
Z b b
f (x)dx = F (x) = F (b) + C − (F (a) + C) = F (b) − F (a).

a a

That is, when we integrate f (x) over the interval from a to b, the
constant of integration disappears.
CHAPTER 8. INTEGRAL CALCULUS 226

The first fundamental theorem of the calculus relates F (x) and f (x).
This is the theorem: Given an integrable function f (t) on a closed
Rb
interval [a, b], that is, given a f (t)dt = F (t) if a ≤ x ≤ b, then the
derivative of Ft exists at each value x and is equal to f (t). That is,
F 0 (t) = f (t).

Examples
R3 3
x2
1. 0
xdx = 2
= 9/2 − 0 = 9/2
0
R2 2
x4
2. −1
3
(x − 3 · x ) = 2
4
− x = ( 16
3
4
− 8) − ( 14 + 1) = − 21
4
−1
R9 dx
3. 3
= loge (9) − loge 3 = log3 (9/3) = loge (3)
x
The loge rather than log is placed here as a reminder. As noted earlier,
Maxima uses log to mean loge and we follow this convention.

The next two examples illustrate the fact that when we use the change- of-
variable technique in order to integrate a function, that is, when we integrate
by substitution, we must always use new limits of integration.
Examples
R 15 x
1. 3 2·x−5 .
Let u = 2 · x − 5. Then du/dx = 2, so dx = du/2. Note that when
we integrate with respect to u, the new limits of integration are: When
x = 4, u = 3 and when x = 10, u = 15. Thus
Z 15
1 du 1 15
· = · log(|u|) = (log(15) − log(3))/2 =

3 2 u 2 3

log(15/3)/2 = log(5)/2.
Alternatively, before we evaluate the integral, we can convert the an-
tiderivative back from u to x and then use the original limits of 4 and
10. That is,
Z 15
1 du 1 15 1 10
· = · log(|u|) = · log(|2 · x − 5) =

3 2 u 2 3 2 4

(log(15) − log(3))/2 = log(5)/2,


as before.
CHAPTER 8. INTEGRAL CALCULUS 227
R2
2. Evaluate 0 3 · x2 · (x3 − 1)2 dx. Let u = x3 − 1 so that du/dx = 3 · x2 or
dx = (1/3) · x2 du. When we integrate with respect to u, the limits of
integration become u = −1 when x = 0 and u = 7 when x = 2. Thus,
Z 7
2 1 3 7
u du = · u = (343 + 1)/3 = 344/3.
−1 3 −1

As an exercise, convert u back to x and then use the original limits of


0 and 2 to get the same answer.

Dealing with Negative Areas

Consider the continuous function y = f (x), which was depicted in Figure


8.2. We wish to find the area that lies under the curve, but above the x axis,
between points a and b. That area is shaded in Figure 8.2.
It is possible, however, that a function may assume both positive and negative
values. This means that the graph of such a function lies below the abscissa
for some values of x, and above the x axis for other values of x. Figure 8.5
illustrates such a possibility. The area between the curve and the x axis for
the interval [c, d], which is indicated by “−”: f (x)dx, is negative in sign. It
is negative because the heights of the rectangles that are circumscribed or
inscribed in that region are negative.
Definition. A negative area is the area measured by the definite integral
that lies below the x axis and above the curve representing the function being
integrated.
When we measure the area for the interval [a, b], that is, when we find
Rb
a
f (x) dx in Figure 8.5, the positive and negative areas counteract each
other. Specifically, the area for the interval [c, d] is subtracted from the sum
of the areas for the intervals [a, c] and [(d, b]. However, if you are interested
in the numeric or absolute value of these three areas, then you must sum
the areas of the regions above the x axis, minus any areas of regions below
the x axis (−(A) = −A). That is, the total or absolute area between the
curve illustrated in Figure 8.5 and the x axis is, in the interval [a, b], given
Rb
by a |f (x)| dx.
Rb
The absolute-value sign in a |f (x)| dx implies that the graph of |f (x)| ,
which is illustrated in Figure 8.5(b), coincides with the graph of f (x) when
CHAPTER 8. INTEGRAL CALCULUS 228

Figure 8.5: Integrating negative areas. (a) Sum of the areas’ values. (b) Sum
of the areas’ absolute values.

f (x) ≥ 0. When f (x) < 0 in some intervals, as is the case in part (b), we
can obtain f (x) by finding its mirror image with respect to the x axis. The
area between the curve and the x axis in the interval [c, d] is equivalent in
absolute size in both parts (a) and (b). The area between the curve and the
x axis in the interval [c, d] in part (b) is the mirror image of the area between
the curve and the x axis in the same interval in part (a).
As we demonstrate below, the function |f (x)| is integrable on the interval
[a, b] whenever f (x) is integrable on the same interval. That is, we can
Rb
show that a |f (x)| dx is the sum of the positive areas minus the sum of the
Rb Rc Rd Rb
negative areas. Hence a f (x) dx = a f (x) dx − c f (x) dx + d f (x) dx.
Examples

1. Find the absolute value of the area bounded by the curve y = x3 − 6 ·


x2 + 8 · x and the x axis, over the range x = 0 to x = 4. (See the shaded
areas of RFigure 8.6, left panel.)
2 R4
Area = 0 (x3 − 6 · x2 + 8 · x)dx − 2 (x3 − 6 · x2 + 8 · x)dx =
CHAPTER 8. INTEGRAL CALCULUS 229
  2  4  4
x4
− 2 · x3 + 4 · x2 − x4 − 2 · x3 + 4 · x2 = 4 − (−4) = 8

4
0 2
Be aware that the definite integral over the range 0 to 4 does not
evaluate to 8. This integral is the sum of the two areas, which equals
0.

2. Find the area bounded by the curve y = x2 − 4 · x and the x axis such
that only positive values of x are permitted. See the shaded area in the
right panel of Figure 8.6. This function has roots at x = 0 and x = 4
and is negative over
 3 the range
 4defined by the two roots. Therefore,
R4 x
Area = 0 dx = 3 − 2 · x2 = (64/3) − 32 − 0 = −32/3, so the

0
absolute value is 32/3.

Figure 8.6: Graphs for Examples 1 and 2

8.4.1 Some Additional Properties of Integrals


We can now state some additional properties of integrals that are useful in
practical situations.
CHAPTER 8. INTEGRAL CALCULUS 230

1. If we interchange the limitsRof integration, the


R b sign of the definite inte-
b
gral also changes. That is, a f (x) dx = − a f (x) dx.
R3 3
An example: 0 x2 dx = x3 /3 0 = 9 − 0 = 9 , and
R0 2 0
3
x dx = x3 /3 3 = 0 − 9 = −9
Ra
2. a f (x) dx = F (a) − F (a) = 0
Rn Rb Rc Rn
3. a f (x) dx = a f (x) dx + b f (x) dx + · · · + n−1 f (x) dx.
This equation says that we can successfully divide a definite integral
into a sum of finite subintegrals. This property warrants some discus-
sion.
First, consider a point that relates to the previous property. First, the
limits of integration usually seem to suggest that one is counting some
points or areas twice. For example, b appears first as the upper limit of
the first subintegral. Then it appears as the lower limit on the second
subintegral. (Likewise, for c, . . . , n − 1.) Is this a double counting of
point b? The answer is no, because Property 2 demonstrates that the
integral of a single point is zero. Therefore it is entirely appropriate to
use a point such as b as both an upper and a lower limit in the process
of integration.
Property 3 also enables us to find the area under the graph of a function
that is discontinuous. The definition of a definite integral explicitly
states that the function R b being integrated must be continuous over the
interval [a, b], if it is a f (x) dx that we wish to find. This would seem
to pose a problem if the function is discontinuous at one or more points
within that interval [a, b]. However, Property 3 enables us to break the
original integral into several subintegrals in order to avoid the problem
of a discontinuous function.
Consider Figure 8.7. Both panels depict functions that are discontin-
uous at point c. Hence it is impossible to integrate either of these
functions over the entire interval [a, b]. The solution is to break the
overall integral for the interval [a, b] into two subintegrals, as follows:
Rb Rc Rb
a
f (x) dx = a
f (x) dx + c
f (x) dx.6

6
This discontinuity must be of either a jump or point variety in order to apply the
method described here. It cannot be an infinite discontinuity.
CHAPTER 8. INTEGRAL CALCULUS 231

Figure 8.7: Functions wth a discontinuity


CHAPTER 8. INTEGRAL CALCULUS 232

Integration over an interval requires continuity within that integral, so we


can integrate over the interval [a, c] and then separately integrate over the
interval [c, b], and sum the results. Both functions illustrated in Figure 8.7
are continuous within these subintervals, and therefore the integration can
be carried out.
Evaluate the following integrals
R3 Re
1. 0 x2 dx 9. 1
loge (x) dx
R2
2. −1 (2 · x + 3 · x2 ) dx 10.
R9
√1 dx
4 x
R3
3. 1 (x2 − 3 · x + 8) dx R1
11. −1
x2 · (x3 + 1) dx
R2
4. 0 (8 · x3 + 6 · x2 − 2 · x + 5) dx R3
R4 12. 0
e−2·x dx
5. −2 (x − 1) · (x − 2) dx
R1
R0 13. x · ex dx
6. −1 (x + 1)2 dx 0

R1
R2
7. −2 (x − 5)4 dx 14. −1
3 · x2 · (x3 − 4)2 dx
R4 3 R 10
8. 2 x4·x
4 +1 dx 15. 4
2·x−3
x2 −3·x
dx

8.4.2 Integrating the Area between Curves


An interesting and occasionally difficult problem arises when we wish to de-
termine the area of a region that is formed and enclosed by curves. For
example, consider Figure 8.8 How do we proceed when we wish to find ei-
ther the area labeled A the area labeled B? We need only to reformulate
slightly our basic approach to finding an integral in order to accommodate
this situation.
Definition: Given: (1) two functions f (x) and g(x), both of which are
integrable on the interval [a, b]. (2) f (x) ≥ g(x) in the interval [a, b]. Then
Rb
a
(f (x) − g(x)) dx is the area between the curves of these two functions.
Rb
In Figure 8.8, then, the area between functions f 2 and g2, Area B, is a (f 1−
g1) dx.7 Suppose, however, that we wish to compute the total area that f 2
7
Likewise for f 1 and g1. This illustration is constructed so that Area A and Area B
CHAPTER 8. INTEGRAL CALCULUS 233

Figure 8.8: Areas between pairs of curves

and g2 bound between a and c. If our interest is in the absolute value of the
sum of Areas B and C. In that case, integrating (f 2 − g2) will not work. We
must integrate a function |f 2 − g2|.
Definition. Given two functions f (x) and g(x), both of which are integrable
on the interval [a, b]. The
R c total absolute area between the curves of these two
functions is given by a |f (x) − g(x)| dx.
Example. The equations for the two expressions that define Areas B and C
in Figure 8.8 are f 2 = −x2 /10 + 2 · x − 5 and g2 = x2 /10 − 2 · x + 5. For these
functions Area B = 94.281, Area B + Area C = -27.86, and absolute value of
Area B plus Area C = 216.42. All values are computed in theRaccompanying
c
workbook. Maxima’s integrate command does not evaluate a |f 2−g2| dx.
If
R cb has been determined,
Rb then the absolute
Rc area can be computed as follows:
a
|f 2 − g2| dx = a |f 2 − g2| dx − b |f 2 − g2| dx.8
Whether you seek the value of the algebraic difference between two functions
are equal.
8
Also, Maxima’s romberg command, which we consider below, can evaluate the ex-
pression in terms of absolute values.
CHAPTER 8. INTEGRAL CALCULUS 234

or the absolute value depends on the question at hand. To consider an


economic application, suppose that f 1 is a marginal value function (whatever
the units of value might be), and that g1 is a marginal opportunity cost (in
the same units of value as f 1). Suppose that we wish to determine the net
value of x. Then, we seek the algebraic difference between the two functions,
not the absolute value.
Thus, if b units could be produced, the net gain would be Area A less the
small area to its left. Inspection reveals that this combined area is positive
(in fact, its value is 80.474). Suppose, however, the x can be provided only
in a quantity x = c. In that case, the total net benefit of x is Area A less
the sum of the small area to the left of Area A and Area C, which appears
to be negative (as the accompanying workbook shows, the value is -27.86.)
Producing c units of x destroys value.
Occasions sometimes arise when it is easier to find the area between two
curves if one conceptually interchanges the roles of the x and y axes. Some-
times, doing so is necessary. Consider the two curves illustrated in Figure
8.9. We wish to find the shaded area in the interval [c, d] (i.e., two values of
y) between the two curves. We must write the relationship between variables
x and y as x = g(y) for the curve farthest to the left, because x values do
not map to unique y values. Likewise the relationship between variables for
the curve to the right is written as x = f (y) The two functions g(y) and f (y)
Rd
are integrable over the interval [c, d], as follows: c .
Often, when two curves bound an area, the choice of the variable with respect
to which we integrate can be selected entirely as a matter of convenience.
Either choice will yield the same value for the area. Two examples follow.
Example 1. Determine the size of the area bounded by the curves y 2 = 2 · x
and y = 2 · x − 2. (See Figure 8.10.) The two curves intersect at (2, 2) and
(1/2, -1), which are found by solving the two equations simultaneously.
For y 2 = 2 · x, y is not a function of x, but x can be stated as a function
of y: x = y 2 /2. Even so, we can integrate with respect to x by breaking
“Area” into
√ two parts. √ Between x = 0 and x = 1/2, the area is the difference
between 2 · x and√− 2 · x Between x = 1/2 and x = 2, the area is the
difference between 2 · x and 2 · x − 2. Look at the first panel of Figure 8.10.
Thus,
R 1/2 √ √ R2 p
Area = 0 2 · x − (− 2 · x) + 1/2 (2 · x) − (2 · x − 2) dx =
CHAPTER 8. INTEGRAL CALCULUS 235

Figure 8.9: Integrating with respect to variable y

1/2  2
2 1 27
3
· (2 · x) 3/2
+ 3
3/2
· (2 · x) 2
−x +2·x = 12
= 94 .
0 1/2

Alternatively, and much more easily, we can integrate with respect to y. The
two curves are drawn with y as the independent variable in the second panel
of Figure 8.10. The integral is this:  2
R 2  y+2 y2   2
y y3
1
Area = −1 2 − 2 dy = 2 · 2 + 2 · y − 3 = 27 12
= 94 .
−1
2
Example 2. Find the area bounded by the curves y= x and y = 2 · x. (See
Figure 8.11.) Solving the two equations simultaneously, we find that the
points of intersection are (0, 0) and (2, 4). Thus
2
R2
Area = 0 (2 · x − x2 ) dx = (x2 − x3 /3) 0 = 4/3.
Alternatively, we can integrate with respect to y, adjusting for the proper
limits, so that:
R4 √  2
 4
x − y2 dy = 23 · y 3/2 − y4 = 43 .

Area = 0

0
CHAPTER 8. INTEGRAL CALCULUS 236

Figure 8.10: Area bounded by a pair of functions, two views

Exercise 8.4
Draw a sketch (either by hand or using Maxima) bounded by the follow-
ing expressions. Compute the values by hand and check your solution with
Maxima.

1. y = x3 , y = 0, x = 0, x = 2 6. y = 6 · x − x2 , y = x2 − 2 · x
2. y = 9 − x2 , y = x + 3
7. y 2 = x, y = x/2 − 3/2
2
3. y = 3 − x , y = −2 · x
8. y = y = x3 , y = 2 · x + 4, x = 0
4. y = 6 − x, y = x + 2, y = 0
5. y = 6 − x, y = x + 2, y = 8 9. y = x, y = 10−4·x, y = 0, x = 0
CHAPTER 8. INTEGRAL CALCULUS 237

Figure 8.11: Graphs of y = x2 and y = 2 · x

8.4.3 Improper lntegrals


Our examination of integration has thus far assumed a continuous function
of the form y = f (x), which is defined for the closed finite interval [a, b].
When an integral satisfies these restrictions, it is said to be a proper integral.
This section discusses integrals that are said to be improper. Specifically, we
examine situations in which the restrictions for a definite integral are relaxed.

We consider two general types of improper integrals. The first occurs when
there are infinite limits of integration. The second occurs when there is an
infinite integrand.
Case 1: Improper integral due to an infinite limit of integration
When the limits of integration areRno longer finite, for example, when we
b
wish to study the definite integral a f (x) dx as a → ∞ and/or as → −∞
we have an improper integral. In such a case, it is not possible to find a finite
value for the integral. This is because F (∞) − F (0) is meaningless, as are
F (b) − F (−∞) and F (∞) − F (−∞).
Definition. An improper integral with an infinite limit of integration is
formally symbolized by
Z ∞ Z b
b
f (x) dx = lim f (x) dx or lim F (x) a .
a b→∞ a b→∞
CHAPTER 8. INTEGRAL CALCULUS 238

Such an integral is said to be convergent when the limit exists and is finite,
whereas it is said to be divergent when the limit does not exist.
Rb
We can use the definition in any particular case by initially finding a , that
is, by finding the indefinite integral F (x). Second, we evaluate F (x) for a
and b, then find the limit as b → ∞. If the limit is finite, then the integral
exists and is convergent. If the limit is infinite, then the integral is diverging
and has no finite value.
It is not uncommon to see an improper integral written without the limit
notation in front of the integral. That is, instead of lim f (x) dx we often see
R∞ ∞ b→∞
the shorthand expression a f (x) dx = F (x) a . This shorthand notation
nevertheless must be evaluated with the limit concept held firmly in mind.
This implicit step must be carried out, since the limit may be divergent, and
if it is, the integral has no finite value.
The existence of an improper integral with an infinite limit for its upper
bound does not change the fact that we are measuring the area under a
curve. Figure 8.12 illustrates the graph of a function y = f (x) where the
Rb R∞
upper limit of integration b is infinite. That is, a f (x) dx = a f (x) dx.
If the improper integral is convergent, that is, if the limit exists, then the
shaded region under the curve is considered to be a finite area. However, if
the improper integral is divergent, then a limit does not exist and the shaded
area under the curve is infinite in size.
It is possible, of course, for the lower bound of integration to be infinite
as well. In this case,
R b the lower bound aR tends to −∞. We can define the
b
improper integral −∞ f (x) dx as lim a f (x) dx. We then use the usual
a→−∞
procedure to determine whether the improper integral is convergent or di-
vergent.
A more complicated case is the situation
R ∞ in which both limits of integration
are, infinite; that is, we wish to find −∞ f (x) dx.
Definition. An improper integral with both limits of integration infinite exists
when, for any real number C,
Z ∞ Z C Z ∞ Z b
f (x) dx = f (x) dx + f (x) dx = lim f (x) dx.
−∞ −∞ C a→−∞,b→∞ a
RC R∞
Both integrals, −∞ f (x) dx and C f (x) dx, must be convergent in order
R∞
for the improper integral −∞ f (x)dx to be convergent. If either of the two
CHAPTER 8. INTEGRAL CALCULUS 239

Figure 8.12: Improper integral: the case of an infinite limit


R∞
integrals is divergent, then the improper integral −∞
f (x)dx is divergent.

Examples

R0 R0 0
e3·x
1. −∞
e3·x dx = lim e3·x dx = lim = 1/3 − 0 = 1/3
a→−∞ a a→−∞ 3 a

R∞ Rb √ b √
2. √1 dx = lim dx/x = lim 2 ·
x = lim (2 · b − 2)
1 x b→∞ 1 →∞ b→∞
1
This integral is divergent; its limit does not exist.
R∞ Rb x b
3. −∞ ex dx = lim a
e dx = lim e x a = lim (eb − ea)
a→−∞,b→∞ a→−∞,b→∞ a→−∞,b→∞
The term eb grows without bound as b increases, so the last term does
not have a limit. Therefore, the integral is divergent.

Instructing Maxima to attempt to evaluate the three intergrals yields these


results. The command integrate(%e^(3*x),x,minf,0); yields the result
1
3
. Both the commands integrate(1/x,x,1,inf); and
integrate(%e^x,x,minf,inf); yield the same warning, “defint: integral is
CHAPTER 8. INTEGRAL CALCULUS 240

Figure 8.13: Three examples of improper integrals

divergent.”

Case 2: Improper integral due to an infinite integrand


The second type of improper integral occurs whenever the integrand becomes
infinite (due to an infinite discontinuity). It does not matter here that the
limits of integration are finite. We need to examine three instances involving
improper integrands.
The first instance involves an infinite
Rb integrand at lower limit.
Definition. Given: an integral x f (t) dt that exists for the interval a < t ≤
Rb
b. Define a new function F (x) such that: F (x) = x f (t) dt for a < t ≤ b
The function F (x) is said to be an improper integral at the point x = a and
Rb
is denoted by the symbol lim x f (t) dt.
x→a
The second instance to consider involves an infinite integrand at the upper
limit. The definition above considers the circumstance in which the integrand
becomes infinite at the lower limit of integration. It is also possible for the
integrand
Rb to become infinite at the upper limit of integration:
lim x f (t) dt for a ≤ t < b.
x→b
CHAPTER 8. INTEGRAL CALCULUS 241

Either of the improper integrals noted in the preceding paragraphs can be


evaluated in the limit to determine whether the improper integral is conver-
gent or divergent. This analysis duplicates that in which we were concerned
with infinite limits to integration, but the integrand was finite.

Examples

R3 Rb b
dx dx
1. = lim = loge (|x − 3|) = lim loge (|b − 3|) − loge (| − 3|)

0 x−3 b→3 0 x−3 b→3
0
The limit does not exist, and the integral is divergent.
R1 R 1 dx 1
2. 0 dx = lim 0 x = lim loge (|x|) = lim (loge (|1|) − loge (|a|)

a→0 a→0 a x→a
The limit does not exist, and the integral is divergent.
R1 R1 √ 1 √
3. 0 √dx = lim a √dx = lim 2 · x = lim (2 − 2 · a) = 2
(x) a→0 (x) a→0 a a→0

Case 3: Improper integral due to an infinite discontinuity


Finally, the third instance involves an infinite integral due to an infinite
discontinuity. The case sometimes arises in which a function y = f (x) is
Rb
discontinuous at some point c. Then the integral a f (x) dx is defined for
the interval [a, b], except at point c, when a < c < b. The additivity theorem
Rb Rc
relating to integrals tells us that we can write a f (x) dx = a f (x) dx +
Rb
c
f (x) dx.
Only when both of the two integrals on the right-hand Rside of this equation
b
converge can we be certain that the improper integral a f (x) dx also con-
verges. It is not sufficient for only one of the integrals on the right-hand side
of the equation to be convergent.
Rb Rc Rb
The analysis in the equation a f (x) dx = a f (x) dx + c f (x) dx relates
to a situation in which one finite discontinuity exists. We can extend this
equation to deal with the situation in which any finite number of such in-
finite discontinuities exist. For example, assume that y = f (x) is infinitely
Rdiscontinuous
b Rat
c
points c Rand d, whereR a < c and d < b. Then we have
d b
a
f (x) dx = a
f (x) dx + c f (x) dx + d f (x) dx. All three of the integrals
on the right-hand side must be convergent in order for us to assert that the
improper integral is convergent.
CHAPTER 8. INTEGRAL CALCULUS 242

Examples
R5
1. Evaluate 1 dx/(x − 2)2 . The integrand is discontinuous at x = 2 so
R5 R5 R5
we restate it as 1 dx/(x − 2)2 = 1 dx/(x − 2)2 + 1 dx/(x − 2)2 .
Evaluating these integrals reveals,
R 5 however, that both are divergent. If
either is divergent, then so is 1 dx/(x − 2)2 .
R5 √
2. Evaluate 1 dx/ x2 − 9. This integral poses two sources of difficulty.
First, for x < 3 it’s value is is a complex number. Second, it is discon-
tinuous at x = 3. Evaluating  √ the
 entire integral using M axima yields
5
this result: log (3) − i atan 2 , which is a complex value. Integrating
from 3 to 5 yields log(3). Thus, the integral converges. Evaluating
this interval by hand involves some trigonometric substitutions. See
Mitchell [?].

Exercise 8.5. Evaluate the following integrals.


R∞ 2 R∞ 2 R8
1. 0 √xx3 +1 dx 5. −∞ x · e−x dx 9. dx
−1 1/x3
R∞
6. 0 e−x dx
R3 R0 x
2. 3 dx/(3 − x) 10. −1 x2 −1
dx
R∞ R1 R∞
3. 1 dx/x2 7. 0 loge (x) dx 11. 1
x · loge (x) dx
R0 R1 R2
4. −∞ (x · ex ) dx 8. −1 dx/x4 12. √ x dx
0 4−x2

8.5 Economic Applications


The concept of the definite integral pertains to numerous economics applica-
tions. The following examples are illustrative.

8.5.1 Consumer Surplus and Producer Surplus


The concept of consumer surplus is often used in applied welfare economics
to evaluate the desirability of particular policy (or regime) options. For
example, one can use the idea of consumer surplus to measure the loss that
consumers realize as a result of the exercise of business and labor monopoly
CHAPTER 8. INTEGRAL CALCULUS 243

Figure 8.14: Consumer Surplus

power. One can also use consumer surplus to help make a decision about the
desirability of building a new highway, a new lock and dam, or of expanding
a wilderness area.9 Figure 8.14 provides a framework for examining the basic
aspects of consumer surplus.
Start with area A in Figure 8.14. The area of A (which extends off the
graph) is the numerical value of consumer surplus when consumers can buy
theR quantity that they select at a price of $5 per unit. Mechanically, this area
x
is 0 1 (p − 5) dx, where x1 is the quantity demanded when p = 5 (x1 u 35.9).
The value of p at each quantity is the price that a consumer is willing and
able to pay for that unit. Hence, the demand curve is also a willingness to
pay curve.
For this demand curve, which is x = p−3/4 , area A is approximately $17,218.
We find this area by integrating the inverse demand curve, p = (x/120)−4/3 .
The integral for the inverse of this inelastic demand curve is divergent, so we
9
The concept of consumer surplus is more complex than the present illustration indi-
cates. For details and references, see Hammock and Mixon [7]. Also, while the concept
is fairly precise, applications in evaluating projects like those cited above must involve a
significant degree of imprecision.
CHAPTER 8. INTEGRAL CALCULUS 244

integrate from x = 0.001 to x = x1 ). The exact value of area A is typically


of little interest and often cannot be interpret in a meaningful way. Suppose
that some small amount of x is a requisite for life. Then placing a value on
the first few units per time period makes no economic sense. In any event,
the question of the total value of consumer surplus is not of interest.
More interesting are areas B, C, and D. The area B + C can be determined
in either of two ways. First, we could determine the value of x when p = 3
(the lower price here) and integrate the inverse demand curve from x =
0.001 to that value of x. This would yield area A + B + C. Subtraction
provides Rour solution. Alternatively,we could integrate the demand curve as
5
follows: 3 x dp, where x is the a function of p. For this example, the area is
approximately $86.05.
Consider two reasons that the price might be $5 rather than $3, or vice versa.
First, a $2 per unit tax might be imposed on a good for which the market
price as $3. Then area B is the tax revenue and area C is the deadweight
loss (a loss to consumers but that does not accrue to anyone). Alternatively,
the higher price might reflect monopoly privilege, in which case the seller
gains area B, which the consumers lose area B + C. Again, area C depicts a
deadweight loss.
Finally, look at area D. This area that would be of interest if a park were
to be expanded, or some other service were to be provided without requiring
a payment by users (added highway lanes, for example). Integrating the
inverse demand curve over the range x = 60 to x = 80 would provide a
measure of the value to this expansion to users and provides guidance as to
whether the expansion is warranted. For this illustration, the value of area
D is approximately $41.47.
The concept of producer surplus is analogous to that of consumer surplus. For
price-taking firms, we can define producer surplus as the difference between
the price that producers receive for their product and the price that one or
more of the firms must be paid to produce the marginal unit. In fact, this
amount is the marginal cost of producing the additional output, either by
expanded the output from incumbent firms or by the entry of more firms, or
both.
In a market with price searchers, the producer surplus is the total revenue
less the integral of the marginal cost curve. We limit out attention to the
case of price takers. Before offering illustrative examples, we consider a
CHAPTER 8. INTEGRAL CALCULUS 245

central feature of markets that consist of price searchers: the quantity that
maximizes the sum of consumer and producer surplus is the equilibrium
quantity.
Let f (x) and g(x) be the inverse
R x demand and supply curves. The area be-
tween these is Surpluses = 0 (f (x) − g(x) dx = F (x) − G(x) + C, where
C is the unknown constant of integration. It is apparent upon inspection
that the first-order condition for maximizing the “Surpluses” function is that
f (x) − g(x) = 0. The remarkable conclusion is, therefore, that the equilib-
rium quantity generates maximum combined surpluses.10

We now consider two examples in which consumer surplus and producer


surplus are computed. Figure 8.15 illustrates these two examples. Also, we
consider an example in which areas under demand curves can be evaluated
and interpreted but cannot be cleanly divided into consumer surplus and
producer surplus.
Example 1. Find the consumer’s surplus given that the demand and supply
functions in a price-takers’ market, for a particular commodity are these:
Demand: p = 30 − 2 · x2 , Supply: p = 3 + x2 .
The positive solution to this set of equations, x = 3 is the equilibrium quan-
tity. Substituting that value into either the demand or supply curve yields
p = 12 as theR equilibrium price. The integrals thatR 3 we seek to evaluate are
3 2 2
these: CS = 0 (30 − 2 · x − 12) dx = 36 and PS = 0 (12 − (3 + Q ) dx) = 18.
The meaning of producer surplus requires some elaboration. Which produc-
ers receive it, and how? Consider first a constant cost industry (horizontal
long-run supply). In this case, the surplus does not exist. Next, suppose
that an industry consists for a large number of identical firms and that as
the number of firms increase, the prices of some inputs increase. Then the
producer surplus accrues to the owners of those resources, with the firms
earning zero economic profit. Finally, consider an industry of firms with dif-
ferent cost curves. Then the firms with costs lower than those of the marginal
firms can earn profits. Also, owners of resources employed by the firms in
the industry might receive part of the surplus. See [7], Chapter 9.
10
This does not imply, however, that the quantity is optimal. The inverse demand and
supply curves do not take into account external effects of producing or consuming the
product. Also, willingness to pay is one criterion in evaluating the marginal value of a
unit. Other criteria, such as paternalism, could enter one’s analysis.
CHAPTER 8. INTEGRAL CALCULUS 246

Figure 8.15: Consumer and producer surplus illustrations

Example 2. The inverse demand function for a particular commodity is P =


28 − 5 · x, and the marginal cost of producing and marketing the commodity
is M C = 2 · x + 4. We determine the consumer surplus and the producer
surplus.
We find maximum profit by setting marginal revenue equal to marginal cost
and finding the profit-maximizing output, x = 2. Inserting thisRvalue into the
3
demand curve yields p = 18. The relevant integrals are CS = 0 (28 − 5 · x −
R3
18) dx = 10 (the gray triangle in Figure 8.15) and PS = 0 (18−(2·x+4)) dx =
24 (the orange trapezoid in Figure 8.15).
It is easy to confirm that the orange area is the same size as the triangle
defined by the y axis, the marginal revenue curve, and the marginal cost
curve. This common area is the amount that producing 2 units adds to the
firm’s profit. That profit is, therefore, PS - fixed cost.
Example 3. This is a stylized response to a rhetorical question that is often
posed: What does it say about a society’s values that elite athletes earn a
multiple of the amount the school teachers earn? We will see that the correct
(if not convincing) answer is, “Nothing.” The wages paid to individuals
CHAPTER 8. INTEGRAL CALCULUS 247

Figure 8.16: Illustrative marginal value functions

roughly (often very roughly) approximate the marginal value of their services.
To determine the total value of these service (however that value is defined)
requires integrating the marginal value function. An important proposition
of elementary calculus is that one cannot determine total value by looking
at the margin (derivative).
To illustrate this point, suppose
√ that the marginal value
√ function for educa-
tors is M V edu = 50000/ edu and M V ath = 500/ ath, where edu is the
number of educators and ath is the number of elite athletes.11
For the values used here, the resulting marginal value of educators is approx-
imately $63,640 per year and that of elite athletes is approximately $273,860
per year. The total areas under the marginal value curves are approximately
$254,560,000 and $1,643,200, respectively. Thus the total value attributed to
11
This example is purely illustrative but not purely fanciful. In a recent year, the
number of K-12 teachers was about 2.4 million and the number of professional athletes
was about 12,000. The median salary of teachers was around $55,000, while that of
profession athletes was around $36,000. Accordingly, we presume that the very higly paid,
elite athletes number well below 12,000. For our purposes, we set the number at 3,000
and the number of educators at 2,000,000.
CHAPTER 8. INTEGRAL CALCULUS 248

the services of educators is about 154.92 times that attributed to the athletes
(to emphasize: this is a multiple of 154.92, not 154.92 percent). All values
are computed in Maxima. The accompanying workbook shows the details.
This example, unlike the previous two, does not offer a breakdown into con-
sumer and producer surplus. This fact reflects the nature of the two labor
markets in which these activities occur. Educators are employed by an amal-
gam of government and private agencies (mostly government). These agen-
cies’ employment and wage decisions are likely politically motivated. rather
than aimed at maximizing a relatively simple objective function.
The market for athletes is likely even more complex. The value provided by
athletes will be divided among the athletes, the owners of franchises, and
spectators. Neither the competitive model of price takers nor the monopolis-
tic model of price searchers applies well. Superstar models and tournament
models both predict very high earnings for a few participants.12

8.5.2 The Normal Distributions


One of the cornerstones of modern statistics is the normal distribution. An
astonishingly broad range of physical and human phenomena can be usefully
represented by a normal distribution. A random variable x is said to have a
normal distribution if its density function is given by the equation
!
1 −(x − µ)2
N (x) = √ exp ,
2πσ 2σ 2

where exp(. . . ) is the same as e(...) . Here µ is the mean of this population’s
values and σ is its standard deviation.
12
The model as stated implicitly assumes that all of these athletes are identical. Both
the superstar model and the tournament model indicate that, even with homogeneity, large
differences will accrue. If the athletes are not quite identical, both models predict great
rewards for relatively small differences in ability. Cyrenne [5] summarizes these models
and applies them to salaries of ice hockey players. The model in the current example
also ignores earnings differences due different individual bargaining abilities or due to
endorsement earnings.
Garicano and Rossi-Hansberg [6] is a much more ambitious application and extension
of the superstar model.
CHAPTER 8. INTEGRAL CALCULUS 249

Figure 8.17: Standard normal distribution

The characteristic bell-shaped curve that represents the normal distribution


is illustrated in Figure
R ∞ 8.17. The area under the normal curve is given by the
improper integral −∞ N (x) dx and is equal to 1.
Figure 8.17 shows a special case of the normal distribution, one in which
µ = 0 and σ = 1. The area to the left of x = −1.96 is the fraction of the
population with values less than this value; that area is approximately 0.025.
The area above x = 1.96 is the same. Therefore, about 95 percent of the
population has values between -1.96 and 1.96. Likewise, about 90 percent
of the population has values between -1.645 and 1.645. These values can be
obtained from a table of value.
Once we have entered the formula into Maxima, however, we can confirm
that these values represent integrals. Consider the following commands and
the resulting output:
float(integrate(N(x,0,1), x,-1.645,1.645 )) yields 0.90003, and
float(integrate(N(x,0,1), x,-1.96,1.96 )) yields 0.95.
We can confirm that the area under the curve equals 1. We do so with two
commands, one evaluating the area to the left of 0 and the other evaluating
CHAPTER 8. INTEGRAL CALCULUS 250

the area to the right: float(integrate(N(x,0,1), x,minf,0 )) yields 0.5,


as does float(integrate(N(x,0,1), x,0,inf )). We could have directly
entered −∞ (minf) and ∞ (inf) into the command, which would result in
1.0.
The commands above relate to the standard normal distribution. Once we
have generated our function in Maxima, we can analyze a normal distribution
without standardizing the units. Suppose, for example, that µ = 5 and σ = 0,
and that we wish to know the fraction of the population that has values less
than 1. The command float(integrate(N(x,5,3), x,minf,1 )) provides
the result, 0.091211, or 9.1211 percent.13

8.5.3 Capital Accumulation


Capital accumulation is the process of adding to a given stock of capital by
the process known as investment. The capital stock in time t is designated
by K(t). The rate at which the stock of capital is being depleted or increased
over time is given by the derivative, dK/dt. This variable is net investment,
which is the amount of capital added to the stock Rless depreciation. Thus,
dK/dt = I(t) − D(t), so that the capital stock is (I(t) − D(t)) dt, where
conceptually, integration occurs from the beginning of time.
Over a finite period, t = a to t = b the relevant expression is
Z b
b
(I(t) − D(t)) dt = K(t) a = K(a) − K(b).
a

As an example, suppose that K(0) = 100, and I(t) − D(t) = a · eg·t , where
t = 0 is the initial period. The first panel shows the investment level for
each time period, and the second panel shows the capital stock for each time
period. Let t1 = 20, where t1 replaces b in the general expression for the
capital stock.
The equation for the capital stock can be written as the command K(K0,
a,g,t1):=’’(K0 + integrate(NetInv(t,a,g),t,0,t1)), which yields this
13
Creating a function as we have done is not necessary. Maxima’s distrib module
provides these values and more. Furthermore, it does so for 25 continuous and discrete
distributions, not just normal distributions.
CHAPTER 8. INTEGRAL CALCULUS 251

Figure 8.18: Net investment and capital stock growth

output: !
%eg t1 1
K (K0 , a, g, t1 ) := a − + K0 .
g g
Recall that %e is Maxima’s notation for the constant e(= 2.718 . . .). We use
K0 = 100, a = 1, and g = 0.03. For t1 = 20, the value of the capital stock
is approximately 127.404, so the stock has grown by 27.404 units during this
twenty year period.
Observe how the value of ∆K appears in the two panels. In the flow panel
on the left, it is an area: changes per year summed over the 20 years. In the
stock panel on the right, it appears as a vertical distance.

8.5.4 The Solow Growth Model14


The preceding example shows that investment adds to a capital stock. This
relationship holds for any firm (or even household), and it holds for the
14
This material relates to the preceding application more than to the content of this
chapter. It may be omitted without loss of continuity.
CHAPTER 8. INTEGRAL CALCULUS 252

aggregate economy. This section sketches the Solow growth model, which
illustrates how production, consumption, saving, and investment interact to
determine an economy’s capital stock and per-capita income. The develop-
ment used here follows Mankiw [11].
The model begins with production. The model assumes that production is
a function of two inputs, capital (K) and labor (L). Also, the production
functions exhibits constant returns to scale, or in terms that Chapter 6 de-
velops, it is homogeneous of degree 1. Formally, Y = f (K, L) where Y is
total output. Because f (K, L) is homogeneous of degree 1, multiplying all
inputs by the same value multiplies output by the same value. We multiply,
both K and L by 1/L, which causes Y to be multiplied by 1/L. Hence,
Y /L = f (K/L, L/L) or y = f (k, 1) where lower-case letters denote per-
labor-unit values. (From now on, we refer to these as “per capita,” which
is exactly appropriate if L is proportional to population) The constant in
f (. . .) is of no consequence, so we rewrite the per-capita production function
as y = f (k).
For √
illustration, we use the simple Cobb-Douglass
√ production function Y =
A · K · L, which converts to y = A · k, where A is a technology index.
Solow [18] does not use a specific functional specification.
The production function generates a marginal product of√capital function:
mpk = dy/dk. In the illustrative example mpk = A/(2 · k). We use this
function below to determine income shares.
We limit our attention to a relatively simple model. This is a model of a
closed economy without government. Thus total output is either consumed
or saved: Y = C + S.15 We restate these values in per-unit-of-labor terms,
y = c + s, in order to relate them directly to the production relationships
stated above.
Another assumption is that not only is the economy closed in terms of trade
(no net exports), but it is also closed in terms of capital flows. This as-
sumption implies that, in equilibrium, y − c = s = i where y is per-capita
output and c is per-capita consumption, so s = y − c is per-capita saving.
Saving is the only source of funds for investment, so per-capita saving equals
per-capita investment: s = i.
15
This is not as severe an assumption as it might appear. If part of the output is diverted
to government, then government must spend on either consumption or investment.
CHAPTER 8. INTEGRAL CALCULUS 253

Figure 8.19: Solow growth model equilibrium

We can now define an investment function, which we name chgk : chg k =


s · f (k, A) − d · k, where d is the depreciation rate. (Investment is the change
in the capital stock, k.) For the system to have attained equilibrium, the
per-capita stock of capital must not be changing. This occurs only when
chg k = 0 Figure 8.19 shows the nature of equilibrium for these parameter
values: A = 3, s = 0.15, and d = 0.1. This economy saves 30 percent of
total output, and therefore invests this amount. Each year ten percent of
the capital stock must be replaced. The result is a capital stock of k = 20.25
For larger values, depreciation exceeds investment, and for smaller values
investment exceeds depreciation. Per-capita output is 13.75 units per year,
and per-capita consumption is 11.475 units per year.
We can determine that one-half of output accrues to labor and one-half to
capital, due to the nature of the Cobb-Douglas function. Thus capitalists
receive 13.5/2 = 6.75 units of output per unit of labor. Of this, they must
maintain the capital stock, so net-of-depreciation income is 4.725 times the
number of units of labor. Each unit of labor receives 6.75 units of output per
CHAPTER 8. INTEGRAL CALCULUS 254

year.16

8.5.5 Present Value


The present value (often called the discounted value) of a single payment
At to be received t periods in the function is At /(1 + i/n)n·t , where i is the
discount rate (often an interest rate at which the entity that is to received
the payment can borrow or lend), and n is the number of times that the
discounting is compounded each period. In the limit as n → ∞, this value,
V , approaches At · e−i·t .
This expression holds for any value of A at any number of periods t. There-
fore, the present value of a discrete number of such would be the sum of
the present values of the individual values. If, for example, A1 , A2 , and A3
are to be received at the end one year, two years, and three years respec-
tively, and compounding is instantaneous,Pthen the present value of this flow
is V = A1 · e−i + A1 · e−2·i + A1 · e−3·i = 3t=1 At · e−i·t .
Change the nature of the flow by allowing the flow to be continuous at a
rate At per year (or whatever period is specified). That is payment begins
immediately and is continuously made throughout the year. The summa-
tion expression
R3 in the preceding paragraph now gives way to this integral:
V = 0 At · e−i·t dt. Keep in mind that each individual payment occurs in
infinitesimally short period, with the rate being such that the sum of the
payments in year t is At .
The example below compares two scenarios. In the first, lump sum of 1000,
3000, and 2500 are received at the ends of the first, second, and third years.
In the second, the annual amounts are the same, but they are spread contin-
uously throughout the year. The resulting present values are higher in the
second case, because the payments are received earlier.
These Maxima commands apply in the first case. [A1,A2,A3] :[1000,
3000, 2500], [V1, V2, V3 ]: [A1*exp(-0.1*1), A2* exp(-0.1*2),
A3* exp(-0.1*3)], and V: V1+V2+V3;. The resulting value is 5213.1, ap-
proximately.
16
Alternatively, we can determine that the marginal product of capital, 1/3, is the
payment to each unit of capital and that (1/3) · 20.25 = 6.75 per unit of labor.
CHAPTER 8. INTEGRAL CALCULUS 255

Now, spread the payments evenly over each of the three years. The resulting
present value calculation is executed with these commands (VC for “value,
continuous”): [VC1: float( integrate( A1*exp(-0.1*t),t,0,1)),
VC2: float( integrate( A2*exp(-0.1*t),t,1,2)),
VC3: float( integrate( A3*exp(-0.1*t),t,1,2))] and
VC: VC1 + VC2 + VC3. The resulting value is 5687.5, approximately. As
predicted, the fact that the payments are made throughout each year rather
than at the ends increases their present value somewhat.
We can generalize
Rτ this expression to allow for τ (the Greek letter tau) periods.
Now V = 0 R(t) · e−d·t . Here R(t) is a function of t, so the integral of this
expression will depend on R(t)’s form. The discount rate is d. Consider a
simple case in which R is the same each period. Now
Z τ
R
V = D · e−d·t = · (1 − e−d·τ ).
0 r
Suppose that τ = 2, R = 3000, and d = 0.06. Then the present value of this
stream is (3000/0.06) · (1 − e0.12 ) ≈ 50000 · (1 − 0.8869) ≈ 5565 dollars.
Figure 8.20 generalizes this calculation by letting τ range from 0 to 50 periods.
Also, it shows the effect of increasing the discount rate from 0.06 to 0.08.
The flattening of the two present value functions as the payment period is
lengthened reflects the effect of compounding: the discounting process has
an increasingly large impact as the successive payments move farther into
the future. Likewise, the size of the discount factor becomes increasingly
important as the length of the payment period increases.
Consider an application, due to Chiang [4]. A wine dealer holds a quantity
of wine. This wine can be sold immediately for $A, but holding it and selling
it later will result in a higher price. Suppose
√ that the wine’s value increasing
(t) 17
according to this function: P = A · e .
−d·t
√ present value√ of the wine, sold in period t is V (t) = P · e
The = A·
(t) −d·t t−d·t
e ·e = A·e . For now, we assume that storage cost is zero, so the
profit-maximizing dealer must simply choose the value of t that yields maxi-
mum present value. We can convert the expression to a linear-in-logarithms
expression, log(V (t) = log(A) + t1/2 − d · t.
17
The precise functional relationship between price and age is not important. This one
is used for convenience. The reasoning in this example extends to examples like fisheries
and forests, in which the growth is a physical growth function rather that a price function.
See McAfee [10], Chapter 4.
CHAPTER 8. INTEGRAL CALCULUS 256

Figure 8.20: Present value, time, and the discount rate

The logarithmic rule for differentiation implies that


1 dV 1
· = · t−1/2 − r.
V dt 2
√ 1
Because A is finite, this condition implies that t = 2·r . Squaring both sides
of this equation shows that the optimal length of storage is 4·r1 2 . A higher
interest rate implies a shorter optimal storage period. This makes sense, for
the interest rate is either what the dealer must pay to underwrite holding
this wine or the rate that the dealer must forgo on alternative uses of funds,
while the growing price is the return to this particular investment.
Add to this model the fact that storage costs are not zero. Suppose that
each case of wine incurs a storage cost Rat a rate of s dollars per year. The
t
present value of that stream of cost is 0 e−d·t dt = (s/r) · (1 − e−d·t ). The
present value of that stream of costs must be subtracted from the present
value of the sale in time t (a single value, not a stream: a case can be sold
only once).
The present value of the sale at time t, N (t) is, therefore, N (t) = A(t)·e−d·t −
(c/r) · (1 − e−d·t ) = (A(t) + (s/r) · e−d·t − s/r. We apply the product rule to
CHAPTER 8. INTEGRAL CALCULUS 257

determine that
dN (t) dA(t) −d·t  s  −d·t
= ·e − d · A(t) + ·e .
dt dt d

This expression equals zero only if d(A(t)


dt
= d · A(t) + s. The term on the
left-hand side is the gain in value from holding the asset one more period,
i. e., the marginal benefit from postponing sale. The term on the right-hand
side is the marginal cost of postponing sale. This cost consists of increased
discount due to the postponement plus the per-period storage cost. Thus,
the familiar marginal cost=marginal benefit criterion for optimization must
be satisfied.

8.6 Questions and Problems


1. An economy currently uses 500 million barrels of petroleum per year.
With current technology, the use of petroleum is expected to grow at a
rate of 8 percent per year, so that the growth pattern can be phrased as
P (t) = 500·e0.08·t . A promising computerized injection system promises
to reduce the annual growth rate to 6 percent per year.
a. State the new growth path mathematically.
b. Plot the two growth paths using Maxima.
c. Determine the amount of petroleum that the new technology promises
to save over the next ten years.
2. The nonlinear
√ inverse demand√ and supply curves for a product are
pd = 225 − 5 · x and ps = 36 + 1.8 · x. Use Maxima’s find root
command to confirm that the equilibrium quantity is approximately
27.794 units per time period and that the equilibrium price is approx-
imately $9.2752 per unit. Then determine the amount of surplus that
accrues to consumers and to producers.
3. Megacorp is the only seller for a product, the inverse demand for which
is pd = 144 − x2 . Megacorp’s marginal cost is mc = 48 + x2 /2.
a. Confirm that Megacorp will maximize its profits if it sells approxi-
mately 5.2372 units per time period for a price of approximately
$116.57.
CHAPTER 8. INTEGRAL CALCULUS 258

b. Assuming that Megacorps does maximize its profits, compute the


values for Consumer Surplus and for Producer Surplus
c. Determine that the efficient quantity, at which p = mc, is 8 units
Also, integrate pd − mc over the range 52372 to 8 to determine the
deadweight loss due to Megacorp’s producing its chosen quantity
rather than x = 8.

4. The maker of a generic household appliance has this marginal cost


function for the appliance: mc = 0.00003 · x2 − 0.03 · x + 20, where x
is the number of units produced each period. Furthermore, it incurs a
fixed cost of $15,000 per production period. It can sell each unit for
$30 (it cannot affect this price by changing x).

a. Determine the total cost function per period.


b. Determine the quantity that maximizes this firm’s per-period profit.
c. Given the answer to (b), find the area between the price and the
marginal cost function. Use the available information to calculate
the firm’s per-period profit.
d. Calculate total revenue. Use the total cost function from (a) and
total revenue to confirm the solution in (c).

5. Currently 100,000 cars per hour use a stretch of highway at rush hours.
Over the next few years, this value will grow, following this growth
function: g(t) = 10000

0.4t
. To what value will the number have grown in 3
years?

6. Consider this marginal revenue functions that apply over the relevant
ranges for product y: MRy = 10/(1 + y)2 .

a. Integrate MRy to determine the total revenue function. Use eco-


nomic analysis to calculate the constant of integration.
b. Determine the average revenue (inverse demand) function. Graph
demand and marginal revenue over the range y = 2 to y = 10.
Chapter 9

Matrix Algebra

Chapter 1 used Stigler’s diet problem to demonstrate the power of mathe-


matics in formulating and solving an important problem. The crux of the
diet problem is to find the least expensive combination of 80 foods available
to a consumer that will satisfy nine recommended daily dietary allowances,
as established by the Food and Nutrition Board of the National Academy of
Sciences.
Formally, the problem is expressed as follows: Minimize C = P1 · X1 + C2 ·
X2 +... · · ·+P80 ·X80 , where C is the cost of a diet that consists of 80 possible
food items. This minimization problem is subject to a set of nine constraints,
each of which corresponds to a minimum recommended dietary requirement
i. The P ’s are prices of the goods, and the X’s are the quantities.
The dietary restrictions are stated as a set of linear equations:

a11 · X1 + a12 · X2 + · · · a1,80 · X80 = R1


a21 · X1 + a22 · X2 + · · · a2,80 · X80 = R2
.. .. .. ..
. . . .
a91 · X1 + a92 · X2 + · · · a9,80 · X80 = R9

Each term aij relates the number of units of requirement i that are provided
by one unit of food j.
We have a system that consists of of 10 equations, the objective function that
we are trying to minimize plus the nine constraint equations. The system

259
CHAPTER 9. MATRIX ALGEBRA 260

contains 80 unknown values, the quantities of the 80 different foods that can
be consumed. This systems is an example of applied linear algebra.
Definition: Linear algebra is the study of systems of linear equations and the
attempt to find a simultaneous solution for the unknowns of those equations,
if such a solution exists.
It is important to note that linear algebra deals with linear equations. Lin-
ear equations are generally easier to deal with than are nonlinear equations.
Nonlinear equations and nonlinear models often cannot be solved without
the help of a computer.1 It is also true, however, that we can usefully ap-
proximate many business and economics relationships with linear functional
forms. Hence we are not severely disadvantaged by the fact that matrix alge-
bra is restricted to the study, manipulation, and solution of linear equations.

9.1 Matrices and Vectors: Definitions


Begin with a very general case involving a system of m linear equations in n
variables:

a11 · x1 + a12 · x2 + ··· a1n · xn = c1


a21 · x1 + a22 · x2 + ··· a2n · xn = c2
.. .. .. ..
. . . .
am1 · x1 + am2 · x2 + · · · amn · xn = cm

The n variables, x1 x2 , . . . , xn above are specifically aligned in a particular


fashion. The variable denoted x1 is the first variable and must appear in the
first column. The variable x2 is the second variable and must appear in the
second column. Any variable xj must appear in the j th column.
Similarly, the subscript aij is definitive with respect to location in the system
of equations. For example, a24 represents the coefficient of the variable that
appears in the second row and the fourth column of the system of equations.
In general, aij refers to the coefficient of the variable located in the ith row
and the j th column.
1
Computer algebra systems like Maxima provide tools for solving systems of nonlinear
equations. Also, they provide tools for evaluating the behavior via simulations of systems
that cannot be solved analytically.
CHAPTER 9. MATRIX ALGEBRA 261

Finally, the parameters c1 , c2 , . . . , cm are m in number and are unattached


to any of the variables xj . The constant c3 belongs in the third row such that
n
X
a3j · xj = c3 .
j=1

More generally, for any particular row, i,


n
X
aij · xj = ci .
j=1

Definition. A matrix is a rectangular, ordered array of elements or entries


consisting of numbers, parameters, or variables.
The existence of a matrix is usually signaled by the use of brackets [ ] or
parentheses ( ). We use the bracket notation [ ] in this text.
Even though it is written in standard notation, the system of equations as
expressed above can become cumbersome and unwieldy. Fortunately, this
system of equations can be simply rewritten using a shorthand notation. Let
A · X = C represent the system of equations that is displayed above, with

     
a11 a12 ··· a1n x1 c1
 a21 a22 ··· a2n   x2   c2 
A =  .. X =  ..  and C =  .. 
     
.. .. 
 . . .  . .
am1 am2 · · · amn xn cn

A, X, and C are matrices, and AX = C.2 The elements in a particular


matrix are not separated by commas, but by blank spaces. It is customary
to symbolize a matrix by an upper-case (capital) letter, such as A, X, or
C, whereas the elements in a particular matrix are customarily denoted by
lower-case (small) letters, such as a, aij , and b. It is possible for the elements
of the matrix to be numeric values, such as 5, 7, or 11.24. In this case, the
numeric values are used in preference to lower-case letters.
2
The same relationship could be written as X · A = C. Doing so, however, would
require rephrasing X and C as follows: X = [x1 , x2 , · · · , xn ] and C = [c1 , c2 , · · · , cn ] .
CHAPTER 9. MATRIX ALGEBRA 262

The matrix labeled A above represents the coefficients of the variables in the
system of equations. The A matrix has m rows and n columns. This can be
contrasted with the variable matrix, labeled X, which consists of n rows and
only one column. In general, there is no relationship between the number
of rows and the number of columns in a matrix. The number of rows is not
necessarily related to the number of columns, and vice versa. What is the
case, however, is that the number of rows and the number of columns define
the dimension (or order) of a matrix. For example, matrix A has m rows and
n columns and is therefore said to be an m × n matrix (which is read, “m by
n matrix ”). The dimension of a matrix is always read rows first, columns
second. A 5 × 7 matrix has five rows and seven columns, not vice versa. In
an important special case in which m = n, for example, a 5 X 5 matrix, one
is dealing with a square matrix.
We occasionally encounter the notation A = [aij ], which represents a matrix
composed of the elements that take the form aij . The number of rows and
columns is unspecified. Note well that [aij ], which represents a matrix, is not
equivalent to aij , which represents a specific element in a matrix. That is,
[aij ] 6= aij unless the dimensions of the matrix are 1 × 1.
A small comment on notation is in order. Either aij or ai,j may be used to
indicate the value in the matrix element in the row i and column j. If we
instruct Maxima to create a matrix of a’s, using the command genmatrix(a,
3, 3), the resulting matrix is
 
a1,1 a1,2 a1,3
a2,1 a2,2 a2,3  .
a3,1 a3,2 a3,3

This differs from  


a11 a12 a13
a21 a22 a23 
a31 a32 a33
only in terms of notation. The two are equivalent ways of saying the same
thing. We use the latter, more compact, notation when doing so does not
introduce a possibility of ambiguity.
It is instructive to rewrite Stigler’s diet problem in matrix notation. We can
denote the objective function that we seek to minimize by P · X = C, and
we can represent the constraint equations by A · X = R, where
CHAPTER 9. MATRIX ALGEBRA 263
 
P = P1 P2 · · · P80 is a 1 × 80 matrix,


X1
 X2 
X =  ..  is an 80 x 1 matrix,
 
 . 
X80

C = [C] is a 1 × 1 matrix,
 
a1,1 a1,2 · · · a1,80
a2,1 a2,2 · · · a2,80 
A =  ..
 
.. .. .. 
 . . . . 
a9,1 a9,2 · · · a9,80

is a 9 × 80 matrix, and

 
R1
R2 
R =  .. 
 
 . 
R9

is a 9 × 1 matrix.
Matrices X and C have the dimensions 80 × 1 and 9 × 1 respectively. Both
matrices have only one column and are referred to as column vectors. Matrix
P has the dimensions 1 × 80 and is referred to as a row vector.
We can use the concept of a vector to view a matrix as a series of related
row and/or column vectors. Consider the matrix
 
a11 a12 ··· a1n
 a21 a22 ··· a2n 
A =  ..
 
.. .. 
 . . . 
am1 am2 · · · amn
CHAPTER 9. MATRIX ALGEBRA 264

We can consider this matrix to be an ordered set of m row vectors,3 namely,

   
A1 a1,1 a1,2 ··· a1,n
 A2   a2,1 a2,2 ··· a2,n 
A =  ..  =  ..
   
.. .. 
 .   . . . 
Am am,1 am,2 · · · am,n

It is absolutely essential that an m × n matrix be read as m rows by n


columns. An m × n matrix is not equivalent to an n × m matrix except in
the special circumstances in which m = n and we have a square matrix. For
example, we define matrices J and K as follows:
 
  1 4
1 2 3
J= and K = 2 5.
4 5 6
3 6

Matrices J and K do not have identical dimensions. J is a 2 × 3 matrix,


whereas K is a 3 × 2 matrix.
A matrix is an ordered array of elements, according to our definition. Each
element of the matrix has an assigned location in the matrix. Any alteration
of that assigned location will, in general, alter the matrix and the system of
equations it represents. Consider the following system of equations.

8 · x1 + 10 · x2 + 12 · x3 = 1
3 · x1 + + 2 · x3 = 0
x1 − 2 · x2 − 5 · x3 = −5

The coefficient matrix of this system of equations is given by this matrix.


 
8 10 12
A = 3 0 2
1 −2 −5

The element in the second row and second column (a2 2) of the coefficient
matrix is 0 and must be included. Further, should we interchange the el-
ements in the first and second columns of the first row, that is, should we
3
We could also view the matrix as an ordered set of n column vectors.
CHAPTER 9. MATRIX ALGEBRA 265

interchange a11 and a12 , then the matrix would become the one below, A∗.
 
10 8 12
A∗ =  3 0 2
1 −2 −5

Matrix A∗ now represents the coefficients of this system of equations.

10 · x1 + 8 · x2 + 12 · x3 = 1
3 · x1 + + 2 · x3 = 0
x1 − 2 · x2 − 5 · x3 = −5

Matrices A and A∗ are not the same; they represent different sets of coeffi-
cients.
We must finally observe that a matrix has no numeric value per se. One
cannot state that a matrix has a value of 5, 7, 14, or any other number. A
matrix is simply a shorthand, efficient method of writing an array of elements.

9.2 Matrix Operations


We have already seen that a matrix is a compact and logical way to write
an array of elements. We represented a system of m linear equations in n
variables as A · X = C.
We have yet to indicate how one matrix is related to another. We have not
touched on such questions as: Why did we choose to write X and C as column
vectors rather than as row vectors? How do we multiply matrices? When are
two matrices equal? Do the laws governing the addition and subtraction of
real numbers also hold for matrices? This section addresses these and other
matters.
More specifically, this section considers the following matrix algebra con-
cepts: (1) equality, (2) addition, (3) subtraction, (4) the commutative and
associative laws of addition and subtraction, (5) scalar multiplication, (6)
matrix multiplication, and (7) the commutative and associative laws of mul-
tiplication.
CHAPTER 9. MATRIX ALGEBRA 266

9.2.1 Matrix Equality


Two matrices A = [aij ] and B = [bij ] are said to be equal, such that A = B,
if and only if A and B have the same dimensions and all corresponding
elements in their arrays are identical. That is, A = B if and only if aij = bij
for all i and j. The four examples that follow illustrate the nature of matrix
equality.
   
1 0 1 0
1. A = and B = A 6= B because a21 6= b21
0 1 1 1
   
1 2 3 1 2 3
2. A = and B = A=B
4 5 6 4 5 6
   
1 2 1 2
3. A = 3
 4 and B = 3 4
5 6 5 6
A 6= B because A and B have different dimensions
   
x1 1
4. A = x2
  and B = 0

x3 8
A = B if and only if x1 = 1, x2 = 0, and x3 = 8

9.2.2 Addition and Subtraction of Matrices


We can add two matrices, A = [aij ] and B = [bij ], if and only if A and B
have the same dimensions. A + B = C such that [aij ] + [bij ] = [cij ], where
cij = aij + bij for all i and j. Matrix C has the same dimensions as A and B.
We can add two matrices of the same dimension, but we cannot add two
matrices of different dimensions. For example, we can add a 2 × 3 matrix to
another 2 × 3 matrix. We cannot, however, add a 2 × 3 matrix to a 3 × 2
matrix. The definition also tells us that addition involves adding correspond-
ing elements in each matrix. That is, given that the two matrices are of the
same dimension, we may add them by summing the corresponding elements
of each matrix. This means that we add the element that is in the first row,
first column of matrix A to the element that is in the first row, first column
CHAPTER 9. MATRIX ALGEBRA 267

of matrix B. The result is the element that appears in the first row, first
column of the summed matrix C.
Likewise, we pair and add the elements in the first row, second column of
each matrix, and so forth. Formally, we can write this process as follows:
   
a1,1 a1,2 · · · a1,n b1,1 b1,2 · · · b1,n
 a2,1 a2,2 · · · a2,n   b2,1 b2,2 · · · b2,n 
..  +  .. ..  =
   
 .. .. ..
 . . .   . . . 
am,1 am,2 · · · am,n bm,1 bm,2 bm,3 bm,n
 
b1,1 + a1,1 b1,2 + a1,2 · · · b1,n + a1,n
 b2,1 + a2,1 b2,2 + a2,2 · · · b2,n + a2,n 
 
 .. .. .. 
 . . . 
bm,1 + am,1 bm,2 + am,2 · · · bm,n + am,n

To subtract matrix B from matrix A, replace the + with − in each of the


cells.
The following six examples illustrate the processes of addition and subtrac-
tion of matrices.
     
1 2 5 6 6 8
1. A = , and B = , so A + B =
3 4 7 8 10 12
     
1 2 3 4 5 6 5 7 9
2. A = , and B = , so A + B =
4 5 6 1 2 3 5 7 9
 
  1 4
1 2 3
3. A = , and B = 6 2 5 , so A + B does not exist
4 5 6
3 6
     
4 8 1 2 3 6
4. A = , and B = , so A − B =
10 12 3 4 7 8
     
1 0 1 1 2 3 0 −2 −2
5. A = , and B = , so A − B =
0 1 0 4 5 6 −4 −4 −6
 
  2 1
2 5
6. A = , and B = 4 6 , so A − B does not exist
4 6
3 5
CHAPTER 9. MATRIX ALGEBRA 268

Adding and subtracting matrices is governed by the Commutative and As-


sociative Laws for Matrix Addition and Subtraction.
The Commutative Law. The order in which we add or subtract matrices
is irrelevant. The commutative law demonstrates this fact. Observe that
we treat subtraction as the addition of a negative number, so the proof for
addition applies directly to subtraction.
Proof : A + B = [aij + bij ] = [aij + bij ] = [bij + aij ] = [bij ] + [aij ] = B + A
The Associative Law. The associative law deals with situations in which
three or more matrices are being added. We can apply the associative law
to subtraction by considering subtraction to be the addition of a negative
number. The law can be stated in terms of three matrices:
Given matrices A, B, and C, A + (B + C) = (A + B) + C = A + B + C.
Once the resulting matrix is defined, this process can be extended to more
matrices, so the law applies to the addition or subtraction of any finite num-
ber of matrices.
Proof : A + (B + C) = [aij ] + [bij + cij ] = [aij + bij ] + [cij ] = (A + B) + C =
[aij + bij + cij ] = A + B + C
An example:
    
1 0 1 0 0 1
A= B= C=
0 1  1 0   0 1 
1 0 1 1 2 1
A + (B + C) = + = =A+B+C
0 1 1 1 1 2
2 1
Confirm that (A + B) + C = .
1 2
Scalar Multiplication. The process of multiplying a matrix by a number
(called a scalar in matrix because it changes the scale of the matrix by
changing the size of all elements proportionately) is referred to as scalar
multiplication.
A formal definition of scalar multiplication is this: Given a matrix A = [aij ]
and a scalar k, the scalar multiplication of k and A, written k · A, is defined
to be  
k · a11 k · a12 · · · k · a1n
 k · a21 k · a22 · · · k · a2n 
k · a = [k · aij ] =  .. ..  .
 
..
 . . . 
k · am1 k · am1 · · · k · amn
CHAPTER 9. MATRIX ALGEBRA 269

Consider two examples:


   
1 2 5 10
1. Let k = 5 and A = . Then k · A = 5 · A = .
3 4 15 20
   
1 −2 3 −3 6 −9
2. Let k = −3 and B = . Then k · B = .
4 5 −6 −12 −15 18

Matrix Multiplication Like matrix addition and subtraction, matrix mul-


tiplication also has attendant dimensional requirements. However, these re-
quirements differ from those imposed in matrix addition and subtraction.
Two matrices A and B can be multiplied together to form the product A · B
if and only if the column dimension of A is equal to the row dimension of
B.4
Thus, we may multiply an m × n matrix by an n × p matrix. An m × n
matrix A is said to be postmultiplied by an n × p matrix B in order to form
a new n × p matrix C. Or we could equivalently state that n × p matrix B is
premultiplied by m × n matrix A, once again yielding an m × p matrix C.
We can formally define the process of matrix multiplication as follows. Given:
A = [aik ], an m × n matrix, and B = [bkj ], an n × p matrix, where aik , is
any element of A and bkj is any element of B. Then: A · B = C, an m × p
matrix whose elements are
Xn
cij = aik · bkj
k=1

for all i = 1, 2 . . . , m and j = 1, 2 . . . , p.


Consider the matrices A and B, which the Maxima commands A: genmatrix(a,
2, 2); B: genmatrix(b,2,1); created:
   
a1,1 a1,2 b1,1
.
a2,1 a2,2 b2,1
The product A · B, generated by the command A.B;, is matrix C, a 2 × 1
matrix that consists of two sums:
 
a1,2 b2,1 + a1,1 b1,1
.
b2,1 a2,2 + b1,1 a2,1
4
Note the use of the dot. Matrix multiplication of the sort discussed here is often
called “dot multiplication” and the result is the “dot product.”
CHAPTER 9. MATRIX ALGEBRA 270

The command B.A; instructs Maxima to premultiply a 2 × 2 matrix by a


2 × 1 matrix, which it cannot do. Accordingly, Maxima’s response to this
command is
MULTIPLY MATRICES: attempt to multiply nonconformable matrices.
– an error. To debug this try: debugmode(true);.

You must clearly understand the following:

• The product of matrices A and B, A · B is read, “B is premultiplied


by A” or “A is postmultiplied by B.”
• In order to form the matrix product A · B, the column dimension of A
must be equal to the row dimension of B.
• If A · B is defined, then the result is a new matrix C that exhibits the
row dimension of A and column dimension of B.
• That the product A · B is defined does not imply that the product B · A
must also be defined.

Consider three examples.

1. If A is a 2 × 3 matrix while B is a 3 × 3 matrix, the A · B is a 2 × 3


matrix and B · A is not defined.
2. If A is a 1 × 3 matrix while B is a 3 × 1 matrix, the A · B is a 1 × 1
matrix and B · A is a 3 × 3 matrix. Even though both A · B and B · A
are defined, they do not have the same dimensions.
3. Matrices A, B, and A · B are generated by these Maxima commands:
[A: matrix( [2,5], [7,1], [8,3] ),
B:matrix([4,6,7], [9,10,11]), A.B]. The matrices are these:
   
2 5   53 62 69
7 1 · 4 6 7 = 37 52 60 .
9 10 11
8 3 59 78 89

Name the third, matrix C. Then we can note that c32 = 2k=1 a3k ·bk2 =
P
8 · 4 + 3 · 9 = 59. Choose 2 or 3 other values in C and confirm that they
are generated in like fashion.
CHAPTER 9. MATRIX ALGEBRA 271

The preceding examples indicate the dimensions of the matrix that results
from matrix multiplication. Our approach to matrix multiplication has so far
been mechanical. We shall now develop an intuitive understanding of matrix
multiplication as well. In particular, we return to linear equations systems,
with which this chapter began. We do so in order to address the logic of
matrix algebra, or linear algebra.
Recall the general system
a11 · x1 + a12 · x2 + ··· a1n · xn = c1
a21 · x1 + a22 · x2 + ··· a2n · xn = c2
.. .. .. ..
. . . .
am1 · x1 + am2 · x2 + · · · amn · xn = cm
We previously learned how to abbreviate this system of linear equations
A·X = C. Our definition of matrix equality enables us to state that
Pnmatrix C
equals the product A·X if and only if element ci is given by ci = k=1 aik ·xk
for i = 1, 2, . . . , m.
The subscripts of the terms in this system lead intuitively to the definition
of matrix multiplication. Specifically, we observe that the subscript k is used
in both the aik and the xk terms. This ensures that the number of columns
in matrix A is the same as the number of rows in matrix X. In more detailed
form, the matrix multiplication A · X = C involves the following:
 Pn 
a11 · x1 + a12 · x2 + . . . a1n · xn Pnk=1 a 1k · x k c1
a21 · x1 + a22 · x2 + . . . a2n · xn =  a
 k=1 2k k  · x  = c2
.. .. ..  ..  ..
. . . 
Pn .
 .
am1 · x1 + am2 · x2 + . . . amn · xn k=1 amk · xk cm

We now need to go from the somewhat familiar case above to the general
case in which A is once again an m × n matrix and X P is an n × p matrix.
Any element cij of the new matrix A · X = C is given by nk=1 aik · xjk , where
i = 1, 2, . . . , m and j = 1, 2, . . . , p.
In detail, the product A · X = C is given by the equalities below:
 Pn Pn 
k=1 a 1k · x k1 · · · k=1 a 1k · x kp
 Pn a2k · xk1 · · · Pn a2k · xkp 
 k=1 k=1
[aik ] · [xkj ] =   = [cij ],

.. ..

Pn . Pn . 
k=1 amk · xk1 · · · k=1 amk · xkp
CHAPTER 9. MATRIX ALGEBRA 272

as i = 1, 2, · · · , m and j = 1, 2, . . . p.
In this case, the matrix of coefficients is applied to a matrix of veriables.
Each column of the latter is the same length, but not (necessarily) the same
variables. Suppose, for example, the coefficients relate to demographic data
like age, years of schooling, and income. The age and location variables would
likely be well-defined, but the analyst might have little reason to choose
between two competing measures of income. The first column in the X
matrix might have pre-tax income, and the second column might have post-
tax income. Multiplying the (fixed-value) coefficient matrix A to the two-
column X matrix would result in a two-column C matrix. The values in
these columns would differ due to the different income measures.
Before addressing the nature of this issue more formally, we consider a sim-
ple hypothetical example. Consider three individuals who are the same age,
have the same level of schooling, and have the same pre-tax and post-tax in-
come, but behave somewhat differently. Specifically, suppose that their level
of consumption of some product is defined by the following three equations:
z1: 10 + 5*age - 2*years + 3*income,
z2: 12 + 4*age - 1.5*years+2.5*income, and
z3: 11 + 4.5*age - 1.8*years + 2.75*income. Each of these is 30 years
of age and has 12 years of schooling. Each also has a post-tax income of 35
and a pre-tax income of 40 (presumably in $1000’s per year). The three ma-
data and the resulting values of z for these people.5
trices below contain this 
  1 1  
10 5 −2 3 30 30 241 256
12 4 0.5 2.5  ·  12 12 = 225.5 238.0 .
  
1 4.5 −1.8 2.75 210.65 224.4
35 40
Note the two 1’s in the X matrix. This is the value by which the constant
term is multiplied. The A matrix is 3 × 4, and the X matrix is 4 × 2, so the
C matrix is 3 × 2. We have two predicted levels of z for each person. The
first column uses post-tax income to predict z,a nd the second column uses
pre-tax income.
We now develop a more general treatment, one that provides a relatively
simple way to remember what is supposed to be multiplied and what is
5
The “people” are likely to representatives of some groups, perhaps identified by region
or ethnicity. Coefficients would likely be estimates that have come from econometric
studies.
CHAPTER 9. MATRIX ALGEBRA 273

Figure 9.1: Matrix Multiplication

supposed to be added. Note that the number of columns in general matrix


A = [aij ] is n, while the number of rows in matrix X = [xkj ] is also n. Hence
the multiplication is defined. The result of the multiplication C = [cij ] is a
matrix of m rows and p columns. Matrix A must have i rows; matrix X must
have jPcolumns. The individual elements in matrix C are given by
cij = ni=1 aik · xkj , with i = 1, 2, . . . m and j = 1, 2, . . . p.
Thus, each element cij depends on the elements in row i of matrix A and the
elements in column j of matrix X. Figure 9.1 demonstrates this relationship
visually. The ith row of matrix A times the j th column of matrix X results
in element cij (ith row, j th column) in matrix C. The ith row in matrix A,
the j th column in matrix X, and element cij in matrix C are all enclosed in
shaded boxes in Figure 9.1.
Consider the following three examples.
   
1 2 1 0 2
1. A = B=
3 4 5 6 3   
1·1+2·5 1·0+2·6 1·2+2·3 11 12 8
A·B = =
3·1+4·5 3·0+4·6 3·2+4·3 23 24 18
B.A is not defined.
 
  3 1  
2. A = 1 2 3 B = 2 0 A · B = 10 7
1 2
B · A is not defined.
CHAPTER 9. MATRIX ALGEBRA 274
 
  3  
3. A = 1 2 3 B = 2
 A · B = 10 , a scalar
  1
3 6 9
B · A = 2 4 6

1 2 3

9.2.3 Commutative, Associative, and Distributive Laws


of Matrix Multiplication
Commutative Law. Matrix multiplication requires that the column dimen-
sion of the premultiplier be equal to the row dimension of the postmultiplier
in order for multiplication to be defined. This requirement virtually elimi-
nates the property of commutability. In general, even if the products A · B
and B · A are both defined, AB 6= BA. Hence they do not “commute,”
and the commutative law does not hold.6 Example 3 above illustrates this
circumstance.
Associative Law. The associative law of matrix multiplication tells us that
A · (B · C) = (A · B) · C = A · B · C, provided that the dimensional re-
quirements for multiplication are satisfied. Figure 9.2 shows the dimensional
requirements that must be satisfied.

Figure 9.2: Conditions for associattive property to hold

Exercise 9.1.

6
You might have noted an exception to this rule. When scalar k and matrix A are
multiplied, then k · A = A · k. The case of an identity matrix (to be defined shortly) is
another exception. Actually, k = k · 1, and 1 is a degenerate identity matrix, so this is
actually just one exception. More later.
CHAPTER 9. MATRIX ALGEBRA 275

1. Find the coefficient matrix for each of the follow systems of linear equa-
tions.

a. 3 · x1 + 2 · x2 + 4 · x3 = 17, x1 + 2 · x2 + x3 = 4, and
5 · x1 + x2 + 3 · x3 = −2
b. x1 + x2 = 4 and 3 · x1 + 2 · x2 = 0
c. x + 2 · y + 4 · z − 2 = −6, −4 · x + 2 · w = 7,
3 · y + z − 4 · w = 0, and −x − y + z = 6
d. x + y − z = 10, −5 · y + 3 · z = 4, and −3 · x + 2 · y = −3

2. State thedimensions
 of the following matrices.
1 2 1    
1 2 3 1
(a) A = 3 4 0
  (b)B = (c) C =
4 5 6 2
0 0 2
 
0 19 9    
0 26 12 9 12 15 2 3
(d) D =   (e) E = (f) F =
0 33 15 −9 5 1 4 5
2 7 4
 
1 2 −1 2 1
  3 0 −6 0 1
(g) G = −9 5 1 (h) H = 0 0

8 0 1
2 −11 10 15 5
For the next three sets of matrices, perform the indicated matrix opera-
tions whenever the matrices meet the required dimensional constraints.
 
−2    
3. Given that A = 0  B = 1 1 2
 C= 2 6 0 ,
4
find (a), A · B (b) B · A, and (c) A · (B + C) = A · B + A · C.
     
2 3 3 5 5 4
4. Given that A = B= C=
4 5 2 0 0 −4
find (a) A + B , (b) A · (B + C), (c) A · B · C, and (d) C · B · A.
 
  2 1 4
5 6 1 3 1 3 1 
5. Given that A = and B =  1 0 −1 find A · B.

1 2 0 −1
0 2 −1
CHAPTER 9. MATRIX ALGEBRA 276

9.3 Special Types of Matrices


Certain types of matrices occur with such frequency that they require ad-
ditional study. Other matrices have elements or dimensions that sometimes
present problems. This section considers the following specific situations:
the identity matrix, the diagonal matrix, the scalar matrix, the null matrix,
and the transpose of a matrix.

9.3.1 Identity Matrix


In the real number system, the number one (1) has the unique property that
for any number a, we have 1 · a = a · 1 = a, and in particular 1 · 1 = 1. In
matrix algebra, the identity or unit matrix plays the same role as does the
number 1 in ordinary algebra.
The formal defintion of the identiy matrix is this. The identity matrix of
order n, denoted by the symbol I or In is a square matrix whose main, or
principal, diagonal contains no other elements except the number 1. All other
elements in the identity matrix are zeros. Thus, In is an identity matrix if
 
1 0 0 ··· 0
0 1 0 · · · 0
 
0 0 1 · · · 0
 .
 .. .. .. 
 . . .
0 0 0 ··· 1

The subscript n on the identity matrix In , indicates the dimension of the


matrix. We can see that the main diagonal of the identity matrix consists of
the elements beginning with the element in the upper left-hand corner of the
matrix and proceeding diagonally to the lower right-hand corner of the ma-
trix. The main diagonal therefore contains the elements a11 , a22 , a33 , . . . , ann .
In the case at hand, every element of the main diagonal is the number 1.
The identity matrix is sometimes written in shorthand as
(
1, i = j
I = [δij ], where δij =
0, i 6= j
CHAPTER 9. MATRIX ALGEBRA 277

and δij is known as Kronecker’s delta.


 
  1 0 0
1 0
Two examples: I1 = and I3 = 0 1 0.
0 1
0 0 1
If A is a square matrix of order n, and In is the identity matrix, then I · A =
A · 1 = A. The example below uses a 2 × 2 matrix A and a 2 × 2 identity
matrix. The first line of command creates the two matrices. The second line
contains a command to create a matrix that consists of four matrices and
some text material. The resulting matrix is really a table and should not be
used for analytical purposes.
I: matrix([1,0],[0,1])$ A:matrix([1,2],[3,4])$
matrix(["I","A","A.I","I.A"],[I,A,A.I,I.A]);
 
 I   A   A.I   I.A 
 1 0 1 2 1 2 1 2 
0 1 3 4 3 4 3 4

When A is not a square matrix, say m × n, I · A = A · I = A still holds, but


in a particular and limited way. The identity matrix in the term I · A is not
the same identity matrix in the term A · I. The dimensional requirements
for an identity matrix differ, depending on whether we premultiply (I · A) or
postmultiply (A · I). In the example below, A is a 3 × 2 matrix. It can be
postmultiplied by a 2 × 2 identity matrix or premultiplied by a 3 × 3 matrix.
In each case the result is the original matrix.
 
 A  I2  I3  A.I2 I3.A
 1 2   1 0 0 1 2 1 2 
3 4 1 0 0
 
1 0 3 4  3 4
0 1
4 6 0 0 1 4 6 4 6

Another circumstance of interest occurs when A = I. Then I · I = I, which


by iteration implies that I k = I for k = 1, 2, 3, . . . , n. Thus an identity
matrix raised to any power is in fact equal to itself. Such a matrix is an
idempotent matrix. By definition, an idempotent matrix is one that, when
raised to any power, does not change in value. A is an idempotent matrix if
A · A = A. Idempotent matrices are central to the development of regression
analysis.
CHAPTER 9. MATRIX ALGEBRA 278

9.3.2 Diagonal Matrices


The concept of a diagonal matrix is directly related to the idea of the prin-
cipal diagonal of a matrix. By definition, a diagonal matrix is (usually) a
square matrix in which all elements both above and below the main diagonal
are zero.7 Matrix D is a general expression for a diagonal matrix:

···
 
a11 0 0
 0 a22 ··· 0 
 .. 
D= 0 . ··· 0 .
 
 . .. .. 
 .. . ··· . 
0 0 ··· ann

Consider four examples.


 
1 0
1. is a diagonal matrix.
0 2
 
1 0 0
2. 0 0 0 is a diagonal matrix.
0 0 2
Only one of nonzero value is required on the main diagonal.
 
1 0 0
3. 0 1 0 is not a diagonal matrix.
−1 −1 0
Confirm that this is an idempotent matrix.
 
1 0
4. is a diagonal matrix.
0 1
The identity matrix is an important special case of a diagonal matrix.
7
The main diagonal or principal diagonal of a matrix consists of the elements that lie
on the diagonal that runs from top left to bottom right. We limit our attention to square
matrices.
CHAPTER 9. MATRIX ALGEBRA 279

9.3.3 Scalar Matrix


A scalar matrix is any square matrix S such that S = λ · I = λ[·Sij ], where
λ is any scalar. For example, the following matrix is a scalar matrix
   
3 0 0 1 0 0
S = 0 3 0 = 3 · 0 1 0 = 3 · I
0 0 3 0 0 1

9.3.4 Null Matrix


In the real number system, zero has the unique property that for any number
a, we have 0 · a = a · 0 = 0. Also, a + 0 = 0 + a = a. In matrix algebra,
the null or zero matrix plays a role similar to that of zero in the real number
system. The definition of a null matrix is this: A null or zero matrix consists
of elements that are all equal to zero. For example, all of the following three
matrices are null.

 
  0
0 0 0  
O= O=  O= 0 0 0 0 0
0 0 0
0

As the second and third examples above indicate, the null matrix is not
restricted to being a square matrix, as are the identity, diagonal, and scalar
matrices. A square null matrix is idempotent.
The null matrix, like the number 0 in the real number system, has several
unique qualities. For instance, the commutative law for addition of matrices
holds when the null matrix and another matrix A are added if both the null
matrix and the A matrix satisfy the usual dimensional requirements. That
is,
CHAPTER 9. MATRIX ALGEBRA 280

The commutative law with respect to the multiplication of matrices presents


a more difficult problem when a null matrix is involved. The two products
A · O = O and O · A = O both result in a null matrix. However, if matrix A
is not square, then the products (A · O and O · A) will not commute. That
is,

On the other hand, if both matrix A and the null matrix are square, then
the two products will commute. Thus,

9.3.5 A Digression on Matrix Algebra


In three instances the intuition we have developed with reference to the
algebra of numbers interferes with the proper application of matrix algebra.
CHAPTER 9. MATRIX ALGEBRA 281

Case 1. We have already discussed the situation in which the product A · B


does commute in the algebra of numbers, but does not commute in matrix
algebra. For example, 5 · 6 = 6 · 5 = 30, while
           
1 0 2 3 2 3 2 3 1 0 8 3
· = 6= · = .
2 1 2 4 6 10 2 4 2 1 10 4

Case 2. Given two real numbers a and b, we know from number algebra
that if a · b = 0, then a = 0 and/or b = 0. However, in matrix algebra, the
product A · B = O does not imply that A = O and/or that B = O. The
following two examples illustrate this point.
     
0 0 0 3 0 0
A= B= A·B = =O
−2 3 0 2 0 0

     
1 3 −3 6 0 0
A= B= A·B = =O
2 6 1 −2 0 0

Case 3. Given three real numbers a, b, and c, we know from number algebra
that when a · b = a · c (with a 6= 0), then b = c. Once again, however, this
relationship does not hold in matrix algebra. For example, given matrices
A, B, and C such that A · B = A · C, it does not follow that B and C are
identical matrices such that B = C, as this example shows.
       
1 3 2 2 −4 2 8 14
A= B= C= A·C = = A · B,
2 6 2 4 4 4 16 28

but B 6= C.

9.3.6 The Transpose of a Matrix


We are now familiar with the meaning of the dimension of a matrix. The oc-
casion sometimes arises, especially where matrix multiplication is concemed,
when we need to form a new matrix whose rows and columns are interchanged
in such a way that they reverse the dimension of the original matrix. Such a
new matrix is referred to as the transpose of the original matrix.
CHAPTER 9. MATRIX ALGEBRA 282

More formally, given an m × n matrix labeled A, the transpose of A, denoted


by A0 or AT , is a new matrix whose rows are the columns of A and whose
columns are the rows of A. Thus, the new matrix has the dimension n × m.
Thatis, if 
a11 a12 · · · a1n
 a21 a22 · · · a2n 
A =  .. ..  = [aij ] then
 
..
 . . . 
a a · · · amn
 m1 m2 
a11 a21 · · · am1
 a12 a22 · · · am2 
A0 =  .. ..  = [aji ].
 
..
 . . . 
a1n a2n · · · amn
To repeat, the original matrix A has the dimension m×n, while the transpose
of A, A0 or AT , has the dimension n×m, as the next four examples illustrate.
 
1 2  
0 1 3 5
1. A = 3 4 , so A =
  .
2 4 6
5 6
A is a 3 × 2 matrix; A’ is a 2 × 3 matrix.
 
1
B = 1 2 3 so B 0 = 2.
 
2.
3
B is a 1 × 3 matrix, while B’ is a 3 × 1 matrix. The transpose of an
n-dimensional row vector is an n-dimensional column vector and vice
versa.
   
1 0 0 1 0 0
3. I = 0 1 0 so I 0 = 0 1 0. That is, I = I 0 .
0 0 1 0 0 1
 
a11 0 0
4. Likewise, for any square diagonal matrix: If A =  0 a22 , then
  0 0 a33
a11 0 0
A0 =  0 a22 , or A = A0 .
0 0 a33
CHAPTER 9. MATRIX ALGEBRA 283

Transposed matrices exhibit four properties that are of immediate interest.


Property 1: (A0 )0 = A. That is, the transposed matrix of an already trans-
posed matrix is the original matrix. This proposition is easily demonstrated:
A0 = [aij ]0 = [aji ] and (A0 )0 = [aji ]0 = aij .
Example:
     
1 5 0 1 −3 0 0 1 5
A= implies that A = and that (A ) = =A.
−3 7 5 7 −3 7
Property 2: A ± B)0 = A0 ± B 0 . More specifically, If A is an m × n matrix,
and B is also an m × n matrix, then both A ± B)0 and A0 ± B 0 are also n × m
matrices that have the same elements. In words, this property asserts that
the transpose of a sum of two matrices is the sum of the transposes of those
two matrices.
To prove this assertion, let A = [aij ] and B = [bij ], so that C = A ± B = [cij ].
Then [cij ]0 = [cji ] = [aji ] ± [bji ] = [aji ]0 ± [bji ]0 .
We can extend this result to the addition or subtraction of any finite number
of matrices such that (A1 ± A2 ± · · · ± An )0 = A01 ± A02 ± · · · ± A0n
The next
 example
 illustrates
  this property.
 
1 5 1 0 2 5
A = 0 8 B =  4 3 A + B =  3 10
2 1 −4 2 −2 2
and    
0 2 3 −2 0 1 0 2
(A + B) = A =
 5 10 2  5 8 1
1 3 −4 2 3 −2
B0 = A0 + B 0 =
0 2 1 5 10 2
Property 3: (A · B)0 = B 0 · A0 . Matrix A has the dimensions m × n, and
matrix B has the dimensions n × p. Then (A · B)0 = B 0 · A0 . That is, the
transpose of the product of these two matrices is the product of the individual
transposes of these matrices in reverse order.
Proof: Let A = [aik ], B = [bjk ], and C = [cij ] = A · B. Then
X X X X
cij = aik · bkj and c0ij = cji = ajk · bki = a0kj · b0ik = b0ik · a0kj .
k k k k

We can extend result to include the product of any finite number of matrices:
A1 · A2 · · · · · An )0 = A0n · · · · · A2 · A1 .
CHAPTER 9. MATRIX ALGEBRA 284

The next three examples illustrate this property. The first example was
created as the text was being printed. That is, the computations were done
by hand. In the second and third, we used Maxima to create the matrices and
to carry out the computations. In each of these two examples the Maxima
output consists of three lists. The first list is the original matrices (A and B
or A, B, and C). The second list contains the products of the original lists,
the transpose of that product, and the transposes of the original matrices.
The third list contains a single item, B 0 · A0 or C · B 0 · A0 . The accompanying
workbook shows the commands.
   
3 2 2 0
1. A = B=
0 5 1 3
Then        
8 6 0 8 5 0 3 0 0 2 1
A·B = (A · B) = A = B =
5 15   6 15  2 5 0 3
8 5 8 5
and (B · A)0 = . Thus (A · B)0 = = B 0 · A0
6 15 6 15
2. A and B    
1 3 −1 1 0
[2 0 0  , −1 2]
0 −1 6 1 3
A · B, (A · B)0 , A0 , and B 0
   
−3 3   1 2 0  
−3 2 7  1 −1 1
[ 2 0 ,
  , 3 0 −1 ,
 ]
3 0 16 0 2 3
7 16 −1 0 6
A · B)0  
−3 2 7
3 0 16
3. A, B, and C
 
    3 2 1
2 1 0 3 −1 
[ , , 2 −1 0 ]
3 4 1 0 2
1 0 −1
A · B · C, (A · B · C)0 , (A · B · C)0 , A0 , B 0 , and C0
     
  15 35   0 1 3 2 1
15 −4 1 2 3 
[ , −4 −1 ,
  , 3 0 , 2 −1 0 ]
35 −1 −1 1 4
1 −1 −1 2 1 0 −1
CHAPTER 9. MATRIX ALGEBRA 285

C 0 · B 0 · A0  
15 35
−4 −1
1 −1

Property 4: When the transpose of a square matrix results in the original


matrix, the original matrix is said to be symmetric about its main diagonal.
That is, if A = A0 , then we have a symmetric matrix. The elements in matrix
A that are above the main diagonal are a mirror image of the elements of
matrix A that are below the main diagonal.
   
1 2 4 1 0 0
For example A = 2 3 5 as is I = 0 1 0. All identity matrices are
4 5 6 0 0 1
symmetrical.

Exercise 9-1
Find the transpose of each of the  following seven matrices.
 
  −12 5   4
1 3 4 4 −2
1. 2.  0 8  3. 4. 0 

5 −1 −1 −2 1
−5 4   −3
  1 2 3
  1 −1 2
5. 1 2 5 6. 7. 2 3 4
0 3 4
   4 4 4  
2 4 1 3 1 0
8. Given that A = , B= , and C = show that
1 2 0 5 −1 3
(a) (A + B)’ = A’ + B’, (b) (A · B)0 = B 0 · A0 ,
(c) (A · B · C)0 = C 0 · B 0 · A0 and (d) (A0 )0 = A.

9.4 Determinants
Previous sections have demonstrated how it to write a linear equation system
in shorthand by means of matrix algebra. For example, we developed the
shorthand A · X = C to represent a typical system of linear equations. It’s
nice to be able to write a large system of equations in a concise, shorthand
notation. However, the premium is on being able to solve that system of
CHAPTER 9. MATRIX ALGEBRA 286

equations for the values of the unknown variables represented by the vector
X.
We can find solutions in a large number of situations. For example, when
we have two linear equations in two unknowns, we can find the solution val-
ues of the unknown variables by setting one unknown variable equal to the
other, substituting, and solving. It is apparent, nonetheless, that the process
of substitution becomes exceedingly complex when many equations and un-
knowns are involved. Therefore we must further develop our matrix-algebra
tools so that we can find the solution values for a large set of simultaneous
linear equations.
The first step we must take is to master the concept of the determinant of a
matrix. Once we have found the value of the determinant, we ordinarily know
whether or not we can solve the system of equations in question, and we often
can find the precise solution values. The determinant of a square matrix is
a uniquely defined scalar (number) that is characteristic of that particular
matrix.
Determinants are denoted by vertical straight lines. Thus,
a11 a12 · · · a1n

a21 a22 · · · a2n
|A| = ..

.. ..
. . .

an1 an2 · · · ann
is a scalar (number). This scalar is said to the the determinant of of the nth
order.
The determinant is calculated by summing products of the matrix’s terms in
a specific fashion. Consider a 2 × 2 matrix
 
a1,1 a1,2
.
a2,1 a2,2

Its determinant is a1,1 a2,2 − a1,2 a2,1 , the product of the terms in the diagonal
less the product of the two off-diagonal terms.
This cross-multiplication process can be extended to a 3 × 3 matrix as the
third example below illustrates. The third-order determinant consists of six
terms, three of which are added and three of which are subtracted in the
process of cross-multiplication. Using the same notation as in the 2 × 2
matrix, the determinant of a 3 × 3 determinant is
a11 ·a22 ·a33 +a12 ·a23 ·a31 +a13 ·a21 ·a32 −(a31 ·a22 ·a13 +a32 ·a23 ·a11 +a33 ·a21 ·a12 ).
CHAPTER 9. MATRIX ALGEBRA 287

Figure 9.3: Calculating the value of a third-order determinant

Figure 9.3 illustrates how we can find the various products when a third-
order determinant is involved. The solid lines in Figure 9.3 form a cross
product of three elements, beginning in each case with an element in the
top row and including two other elements that are each from a different row
and column. The dashed lines also form a cross product of three elements,
beginning in each case with an element from the bottom row and including
two other elements, each of which is from a different row and column. The six
products together determine the value of the determinant, with the solid-line
products to be added and the dashed-line products to be subtracted.
Examples
 
4 2
1. A = , so |A| = 4 · 5 − 1 · 2 = 20 − 2 = 18.
1 5
 
−3 −4
2. A = , so |A| = −3 · 5 − 1 · (−4) = −15 + 4 = −11.
1 5
CHAPTER 9. MATRIX ALGEBRA 288
 
1 0 0
3. A = 3 2 4, so |A| = (1 · 2 · 3 + 0 · 4 · 4 + 0 · 3 · 1) −
4 1 3
(4 · 2 · 0 + 1 · 4 · 1 + 3 · 3 · 0) = 2.

Evaluating Determinants of Orders Higher than Three. The cross-


multiplication methods of evaluating determinants of orders two and three
cannot be directly applied to determinants of orders higher than three. We
can use another procedure to evaluate determinants of the fourth (and higher)
orders. This new procedure is known as expansion by cofactors, and it
operates on the principle of reducing a higher-order determinant to a se-
ries of second- or third-order determinants that we can evaluate by cross-
multiplication.
We illustrate the general nature of this approach. We do not develop it fully,
however, because computer algebra systems offer a much quicker and less
error-prone approach to determining the value of the determinant of a large
matrix. We illustrate this alternative immediately following our sketch of
expansion by cofactors.
Consider once more a 3 × 3 matrix. We show that the determinant can be
recast as a set of three smaller (2 × 2 matrices, each multiplied by one term
in a given row or column. In our expansion, we use the terms of row 1. Any
row or column could be used in this same way.
 
a1,1 a1,2 a1,3
|A| = a2,1 a2,2 a2,3  ,
a3,1 a3,2 a3,3
which we have determined to be
a1,1 a2,2 a3,3 −a1,2 a2,1 a3,3 −a1,1 a2,3 a3,2 +a1,3 a2,1 a3,2 +a1,2 a2,3 a3,1 −a1,3 a2,2 a3,1 .
We now extract the terms that we will use in our process (a1,1 , a1,2 , and a1,3 ),
yielding
a1,1 (a2,2 a3,3 − a2,3 a3,2 ) − a1,2 (a2,1 a3,3 − a2,3 a3,1 ) + a1,3 (a2,1 a3,2 − a2,2 a3,1 ) .
From the definition of the determinant, we can recast this expression in terms
of three 2 × 2 matrices, each multiplied by the relevant cofactor:
     
a2,2 a2,3 a2,1 a2,3 a2,1 a2,2
a1,1 · − a1,2 · + a1,3 · .
a3,2 a3,3 a3,1 a3,3 a3,1 a3,2
CHAPTER 9. MATRIX ALGEBRA 289

Label these smaller matrices |A1,1 |, |A1,2 |, and |A1,3 |. We can now restate |A|
as follows:
X 3
|A| = (−1)1+j · a1,j · |A1,j |.
j=1

The term (−1)i+j ·|Ai,j | is the cofactor of ai,j . When i+j is an even number, a
value is added to the sum; when i + j is an odd number, a value is subtracted
from the sum. As noted above, any row value (i, not just i = 1) can be used.
Also, the summation could be over the rows in a specified column.
The terms |Ai,j | are called minors. The terms (−1)( i + j) + |Ai,j | are called
signed minors or cofactors. We can define each cofactor as |Ci,j | = (−1)i+j ·
|Ai, j|. Then a general expression for the solution of a determinant by the
method of cofactors can be written as
n
X
|A| = ai,j · |Ci,j |
j=1

for any row i or


n
X
|A| = ai,j · |Ci,j |
i=1

for any column j.


The use of cofactors can greatly simplify the computation of some determi-
nants. Look at Example 3 above. All of the terms in row 1 except the first
equal. Therefore, we can see, almost by inspection, that the determinant
is 1 · (2 · 3 − 1 · 4) = 2 Judicious selection of row or column can make the
computation much easier than the preceding discussion might suggest that it
is. Even so, for n > 3, the use of a computer algebra system is recommended.
Consider this 5 × 5 matrix
1 1 1 1 1
2 3 4 5 6
2 1 2 1 2
 33 2
3
5
1
3
3
7
3

 44 5 2 7 8
.
 2 4 1 4
5 3 7 2 9
5 5 5 5 1
6 7 8 9 2

No row or column contains any zeros, and the computation, even with the
application of the method cofactors will be arduous and subject to mistakes.
CHAPTER 9. MATRIX ALGEBRA 290

1
The solution, given almost instantaneously by Maxima is 560105280000
.8
The site http://www.purplemath.com/modules/minors.htm offers a slightly
more expansive discussion of this topic. This site shows how to compute the
value of a 4 × 4 determinant. It also offers some tips on how to manipulate
a matrix so as to generate cells that have 0 as a value. These tips involve
using the properties of determinants that we state below.
A digression on the differences between matrices and determinants.
Matrices and determinants are not the same thing. A matrix, denoted by
brackets or parentheses, has no numeric value. A matrix is a rectangular
array of numbers, variables, and parameters. A determinant, on the other
hand, does have a numeric value. A determinant is defined to be a scalar
(number).
Matrices can be of any dimension and need not be square. Determinants
must be square. A 2 × 3 determinant does not exist.
Properties of determinants. We can usefully apply the following prop-
erties when we work with determinants. These properties apply to determi-
nants of any dimension.
Property 1. The determinant of a matrix A has the same value as the deter-
minant of its transpose A0 . Let
   
a b 0 a c
A= . Then A = .
c d b d
For both of these matrices, the determinant is the same, a · d − b · c.
Property 2. Interchanging any two rows (or any two columns) of a determi-
nant does not alter the absolute value of that determinant. It does, however,
change the sign of the determinant.

a b b a
For example = a · d − b · c but
= b · c − a · d. Confirm that
c d d c
interchanging rows 1 and 2 of the initial matrix has the same effect as inter-
changing the columns.
Property 3. A determinant in which any two rows (or any two columns) are
identical, or a determinant in which any two rows (or any two columns) are
multiples of each other, has a value of zero.
8
Three commands result in the creation of this matrix and the compuation of its
determinant: h[i,j]:= i/(i+j), genmatrix(h,5,5), and determinant(%).
CHAPTER 9. MATRIX ALGEBRA 291

a b
For example = a · b · k − k · a · b = 0.
k · a k · b
Property 4. A determinant in which any row or any column has all zero
elements has a value of zero.

a b
For example = 0 · a − 0 · b = 0.
0 0
Property 5. Adding (or subtracting) a multiple of one row of a determinant
to (from) another row of that determinant, or adding (or subtracting) a
multiple of one column of a determinant to (from) another column of that
determinant does not change the value of the determinant.

a b
For example = a · d + a · k · b − (c · b + k · a · b = a · d − c · b).
c + k · a d + k · b
Property 6. If every element in one row (or one column) is multiplied by a
constant k, then the value of the determinant is also multiplied by k. For
example, 
a b a b
For A = , |A| = a·d−c·b = and = k ·a·d−k ·c·b = k ·|A|.
c d k · c k · d
By iteration, if we multiply the elements of two columns of a matrix M by k,
then the determinant of the new matrix is k 2 ·|M |. Further, if we multiply all
elements of an n × n matrix M by k the determinant of the resulting matrix
is k n · |M |.
Exercise 9.3. Evaluate each of the following matrices. Confirm your solu-
tions with M axima.
     
1 2 3 1 3 4 4 1 6
1. 2 3 4 2. 2 0 7 3. 7 2 9
1 5 7 5 6 9 3 0 8
 
  2 1 −3 4  
2 1 5 −4 7 −2  1 1
4. 5.   6.
3 4 4 0 6 −3 −3 −3
3 −2 5 2
 
    1 2 −2 3
1 1 1 2 1 1 3 −1 5 0
7. 0 1 1 8. 0 5 −2 9.  
1 7 2 −3
0 0 1 1 −3 4
4 0 2 1
CHAPTER 9. MATRIX ALGEBRA 292

9.5 The Inverse of a Matrix


Much of our work in this chapter has dealt with a system of n linear equations
in n unknowns, such as
 
a11 · x1 + a12 · x2 + · · · +a1n · xn = c1
 a21 · x1 + a22 · x2 + · · · +a2n · xn = c2 
 
 .. .. .. .. 
 . . . .
an1 · x1 + an2 · x2 + · · · +ann · xn = cn

We have seen that we can write this system of equations as A · X = C. We


shall soon discuss the problem of solving such an equation system.
Solving a linear equation system of the form A · X = C seems easy at first
glance. We are tempted to divide both sides of the equation system by A,
giving us X = C/A. Unfortunately, we cannot do this. In matrix algebra,
the division operator is not defined. That is, we cannot divide a matrix C by
another matrix A. Instead, we must use a technique that involves finding the
inverse of a matrix. Symbolically, in ordinary algebra, the follow equalities
hold: C/A = C · A−1 = A−1 · C. That is, division can be restated as a
multiplication process, where 1/C is the inverse of C. The process in matrix
algebra is analagous to that in ordinary algebra, but it must be developed
carefully.
To repeat, in matrix algebra, we must use the inverse of a matrix instead
of dividing one matrix by another. Moreover, we must be careful in doing
so, for (as we have already seen) matrix multiplication does not commute.
Given
A · X = C
n×n n×1 n×1
if an inverse does exist for A, then the solution for the X matrix is

X = A−1 · C
.
n×n n×n n×1

This is in general not equivalent to X = C · A−1 .


We proceed by defining the inverse matrix and then considering how to go
about determining its contents. Definition: If it exists, the inverse of a square
CHAPTER 9. MATRIX ALGEBRA 293

matrix A is another square matrix, denoted A−1 , that satisfies the relation
A−1 · A = A · A−1 = I .
This definition of the inverse matrix is consistent with ordinary algebraic
rules. For example, in ordinary algebra, a · a−1 = a−1 · a = 1. In the
case of matrix algebra, it makes no difference whether A is premultiplied
or postmultiplied by A−1 . The product that results is always the identity
matrix I.
Our definition of the inverse of a matrix has two noteworthy implications.
We state these without proof. See Perlis [17].

1. If an inverse of a matrix does exist, then this inverse is unique. That


is, A−1 is the only matrix that, when multiplied by A, results in the
identity matrix I.

2. An inverse matrix is that the commutative law of multiplication does


hold when we multiply a square matrix by its inverse. That is, A·A−1 =
A−1 · A = I.
This is an important exception to the rule that, in general, matrix
multiplication is not commutative.

9.5.1 Finding the Inverse of a Matrix if It Exists


The following theorem yields both the necessary and sufficient conditions for
determining whether or not an inverse matrix exists, and how one can find
it. Theorem: A square matrix A has an inverse matrix A−1 if and only if the
determinant |A| =
6 0, in which case A is said to be nonsingular. The inverse
−1
is given by A = (1/|A| · adjA, where “adjA” refers to the adjoint of matrix
A.
We emphasize three points concerning this theorem. First, a square matrix
is a necessary, but not a sufficient condition, for an inverse matrix to exist.
Second, if matrix A does have an inverse, A is said to be nonsingular. If
matrix A does not have an inverse, then A is said to be singular. (We shall
return to the concepts of singularity and nonsingularity when we solve sys-
tems of simultaneous equations.) Third, the condition |A| = 6 0 is a sufficient
condition for an inverse to exist.
CHAPTER 9. MATRIX ALGEBRA 294

The theorem above uses the term adjoint of matrix A. We now define this
new concept. Definition: The adjoint of matrix A, denoted by adjA, is the
transpose of the cofactor matrix of A, which we encountered when computing
determinants. More formally, let the cofactor matrix of matrix A be given
by C = [ |Aij | ]. Then the adjoint of A is given as
 0  
|C11 | |C12 | · · · |C1n | |C11 | |C21 | · · · |Cn1 |
 |C21 | |C22 | · · · |C2n |   |C12 | |C22 | · · · |Cn2 | 
..  = adjA =  .. ..  .
   
 .. .. ..
 . . .   . . . 
|Cn1 | |Cn2 | · · · |Cnn | |C1n | |C2n | · · · |Cnn |

Remember that a cofactor of an element is a signed minor given by |Cij | =


(−1)i+j · |Aij |. The cofactor is not the element of a particular row and
column multiplied by the signed minor. That is, |Cij | =6= aij · (−1)i+j · |Aij |.
Only when we expand by a particular row or column in order to find the
determinant do we need to multiply the element and cofactor together.
   
2 1 4 −3
Consider a 2 × 2 matrix, A = . Then |A| = 5, [ |Cij | ] = ,
  3 4   −1 2
4 −1 4 −1
and adjA = . Thus A−1 = 51 · . To confirm that we have
−3 2  −3 2     
4/5 −1/5 2 1 1 0
found the inverse, compute A−1 · A = · =· .
−3/5 2/5 3 4 0 1
Now we consider a 3 × 3 matrix
 
1 3 4
M = 2 0 7 .
5 6 9

To analyze this matrix, we use Maxima, which computes the determinant of


matrix M . The value is 57. We use the command adjM: adjoint(M) to
define the adjoint matrix and assign it a name.9 The result is
 
−42 −3 21
adjM =  17 −11 1  .
12 9 −6
9
Maxima does not have a command to createthe matrix of cofactors, but this is just
the transpose of the adjoint.
CHAPTER 9. MATRIX ALGEBRA 295

We can either divide the adjoint matrix by the determinant or use the com-
mand invert(M) to determine the inverse matrix, which is
 14 1 7

− 19 − 19 19
M −1 =  57
17
− 11
57
1 
57
.
4 3 2
19 19
− 19
Finally, we can compute either M −1 · M or M · M −1 to confirm that the
result is a 3 × 3 identity matrix.
Suppose that our matrix is
 
1 0 4
2 −3 1 .
6 −9 3
For this matrix (for which the values in the third row are 3 times their
counterparts in the second row), the determinant is 0. Thus we cannot divide
the elements of the adjoint matrix by the determinant in order to compute
the elements of the inverse matrix. This matrix is singular and does not have
an inverse matrix.
The three examples above illustrate two important points. First, the defi-
nition of an inverse matrix requires that |A| =6 0. Not only does this mean
that matrix A is nonsingular, but also it recognizes that division by zero is
undefined. Therefore |A| 6= 0 is a sufficient condition for an inverse matrix
to exist. Second, it is always possible to check whether the theorem con-
cerning inverse matrices has been applied correctly. One need only multiply
the alleged inverse and the original matrix. If the theorem has been applied
correctly, the result must be the identity matrix.
Properties of Inverse Matices. Three properties of inverse matrices war-
rant mention.

1. If A−1 exists, the (A−1 )−1 = A. That is, the inverse of an inverse
matrix, if it exists, is the original matrix. For an example, consider
M −1 in Example 2 above. Its inverse is
 
1 3 4
2 0 7 ,
5 6 9
which is matrix M .
CHAPTER 9. MATRIX ALGEBRA 296

2. If A−1 and B −1 both exist, then (A · B)−1 = B −1 · A−1 . That is, the
inverse of the product of two matrices is equal to the product of their
inverses in reverse order. This property generalizes to any number of
matrices, so that (A · B · · · Z)−1 = Z −1 · · · B −1 · · · A−1 .

3. If A−1 exists, then (A0 )−1 = (A−1 )0 . That is, the inverse of the transpose
is the transpose of the inverse.

Summary and Review: Finding the Inverse of a Matrix


By hand:

1. Determine whether or not the inverse matrix exists. That is, find |A|
If |A| = 0, then there is no inverse matrix.

2. Find the cofactor matrix. That is, find C = [|Cij |].

3. Find the adjoint of matrix A. That is, take the transpose of the cofactor
matrix such that C 0 = adjA.

4. Divide adjA by the determinant of A. That is,


1
A−1 = · adjA.
|A|

With Maxima: Use the command invert(M), where M is the name that you
have assigned to the matrix.
Exercise 9.4. For each of the following matrices, find the inverse if it exists.
     
  1 1 1 2 1 1 2 1 3
2 1
1. 2. 0 1 1 3. 0 5 −2 4.  3 0 1
0 5
0 0 1 1 −3 4 −1 1 4
     
−1 0 2 1 0 −2 7 6 5
5.  3 1 −6 6. −3 −1 6  7. 1 2 1
−2 −1 5 2 1 −5 3 −2 1
     
3 −5 0 5 2 4
8. 9. 10.
−1 2 6 4 −3 −6
CHAPTER 9. MATRIX ALGEBRA 297

9.6 Solving Simultaneous Linear Equations


By this juncture, we have become quite familiar with a simultaneous linear
equation system of the form

A · X = C
.
n×n n×n n×1

If the inverse matrix A−1 does exist, then premultiplying both sides of A·X =
C by A−1 yields A−1 · X = A−1 · C, or

X = A−1 · C = D
.
n×1 n×n n×1 n×1

Our definition of matrix equality tells us that the left side n × 1 column
vector of unknown variables represented by X must be equal to the right side
n × 1 column vector of solution values represented by D if the two sides are
indeed equal.
We have also found that we can find an inverse matrix (such as A−1 ), only
if the matrix A is square. We stated this requirement by asserting that
A−1 · A = A · A−1 = I, the identity matrix. This means that the number of
equations is equal to the number of unknown variables.
Recall that when the value of the determinant of a matrix is zero, then you
cannot find an inverse for that matrix. That is, A−1 = |A|
1
· adjA and A|| =
6 0.
Thus, when |A| =6 0, there is a unique solution for a linear equation system.
Nonsingularity implies that an inverse can be found. When an inverse ma-
trix can be found, then there is a unique solution. We may summarize the
relationship between nonsingularity and the existence of a unique solution
as follows: Nonsingularity implies the existence of an inverse and a unique
solution and vice versa.
Consider two examples.

1. Given: y−4·x = 12 and y+3·x = 5. The commands [A:matrix([1,-4


], [1, 3 ]), C:transpose(matrix([12,5])), invA : invert(A),
CHAPTER 9. MATRIX ALGEBRA 298

D: invA.C]; create the commands that generate the following: A, the


vector C, A−1 , and D. The resulting output follows:
     3 4  
1 −4 12 8
, , 7 1 71 ,
1 3 5 −7 7 −1

So, this system’s solution is y = 8 and x = −1.

2. Given: 2 · x1 − x2 − x3 = 0, −x1 + 4 · x2 − x3 = 0, x1 + x2 = 8. Then


the X matrix, the C vector, X −1 and D are as below (created with the
same set of commands as above).
     1 1 5
  
2 −1 −1 0 8
− 8 8
5
−1 4 −1 , 0 , − 1 1 3  , 3
8 8 8
1 1 0 8 − 58 − 38 78 7

The solution to this system is x1 = 5, x2 = 3, and x3 = 7

Application: Ordinary Least Squares Estimation


In your statistics course you encountered the “normal equations.” These
equations solved for values of the parameters of a linear equation and were
phrased in terms of the observed values of the independent variable(s) and
the dependent variable, where the relationship is linear. That is, yt = b0 +
b1 · x1,t + b2 · x2,t + . . . + bk · xk,t + et . A set of n values of the dependent
and dependent variables are collected (n > k + 1). We can place these data
into two matrices. The et are error terms that result from a random process;
they are the “noise” while the first part of the equation is the “signal.”
The first matrix, X consists of n rows and n+1 columns (the +1 is to accom-
modate the constant term, as our illustrative example below demonstrates).
The second matrix, C consists of a single column of n values. In matrix no-
tations, the relationship between X and Y is as follows: Y = B ·X +E where
E is a vector of n error terms. Statistics textbooks show that the ordinary
least squares estimators for the values of the b terms (the coefficients) are
derived as follows: Bols = (X 0 · X)−1 · X 0 · Y . That is, the transpose of X,
which is k × n) is muliplied by X, which is n × k, yielding a square k × k
matrix, for which the inverse is computed. This inverse is multiplied by the
product of the X 0 · X and Y matrices, which is a k × 1 vector. The product
of the inverse, a k × k matrix and this vector is another k × 1 vector, which
is the OLS estimators.
CHAPTER 9. MATRIX ALGEBRA 299

To derive a specific set of estimates (numerical values) from this set of esti-
mators (rules), we require a data set. To illustrate the process, we look at a
hypothetical example in which yt = b0 + b1 · xt + b2 · x2,t + et . The et terms
are not observable. They can be estimated after the parameters have been
estimated. The data are as follows:
   
1 2 5 1
1 5 8 2.1
   
1 6 9 ,  3  .
   
1 7 7 4.5
1 6.5 8 7

The first matrix contains the x values. This matrix includes a column of
1’s which attach to the constant term, the estimated value of b0 . We are
estimating three parameters, b0 , b1 , and b2 . We have five data points. Me-
chanically, this is enough to provide estimates. It is not nearly enough to
provide reliable estimates, but it does illustrate the process.
Now we transpose X, and then mulitply this transpose by X. The result
in the 3 × 3 matrix in the middle. Then we determine the inverse of that
matrix.
     
1 1 1 1 1 5 26.5 37 6.63 0.259 −1.05
2 5 6 7 6.5 , 26.5 1.56102 2.05102  ,  0.259 0.139 −0.135
5 8 9 7 8 37 2.05102 283 −1.05 −0.135 0.239

The first of the following pair of of matrices shows the result of multiplying
X 0 by Y . Multiplying (X 0 · X)−1 by this matrix creates our set of OLS
estimates for the coefficients: (X 0 · X)−1 · (X 0 · Y ).
   
17.6 0.806
1.07102  ,  1.16  .
1.36102 −0.466

Thus our estimated relationship is yt = 0.806 + 1.16 · x1 − 0.466 · x2 .10


Once we have computed the estimates for the coefficients, we can apply them
to the data in order to see what values the models estimate. The final set of
10
Be aware of the word “estimated.” We do not know the true relationship (if any)
between x and y and yielded these observed values.
CHAPTER 9. MATRIX ALGEBRA 300

output shows the estimates. It also repeats the Y vector for comparison, and
it shows the size of the residuals–the differences between the OLS estimates
and the actual values in this sample.
 
OLS  estimates
 Actual
 values
 Residuals
 

 0.802 1 0.198  
  2.89  2.1 −0.792
      
  3.59  3 −0.588
      
  5.68  4.5  −1.18 
4.64 7 2.36

Exercise 9.5. Write each of the following as matrices A, X and C, such


that A · X = C. For each, create the X −1 matrix and solve the system of
equations. Confirm your solutions with Maxima.

1. x1 + 3 · x2 = 15 6. x1 + 2 · x2 − 3 · x3 = −1
x1 − 2 · x2 = −3 3 · x1 − x2 + 2 · x3 = 7
5 · x1 + 3 · x2 − 4 · x3 = 2
2. 2 · x1 + 3 · x2 = 10
− 4 · x1 + x2 = −6 7. 2 · x1 + x2 − 2 · x3 = 10
3. 2 · x1 − 3 · x2 = 7 3 · x1 + 2 · x2 − 2 · x3 = 1
3 · x 1 + 5 · x2 = 1 5 · x1 + 4 · x2 + 3 · x3 = 4

4. 10 · x1 − x2 − x3 = 0 8. x1 + 2 · x2 − 3 · x3 = 6
− x1 + 12 · x2 − 2 · x3 = 0 2 · x1 − x2 + 4 · x3 = 2
x1 + 2 · x2 = 24 4 · x1 + 3 · x2 − 2 · x3 = 14

5. 12 · x1 − 2 · x2 − x3 = 0 9. x1 + 3 · x2 − 2 · x3 = 0
12 · x1 − 6 · x2 − x5 = 0 2 · x1 − 3 · x2 + x3 = 0
x1 + x2 = 16 3 · x1 − 2 · x2 + 2 · x3 = 0

9.7 Maxima and Minima: Functions of n In-


dependent Variables
Chapter 7 dealt initially with the means of finding extreme points for func-
tions of one independent variable, and subsequently with the same consid-
eration for functions of two independent variables. The case(s) of three or
CHAPTER 9. MATRIX ALGEBRA 301

more independent variables were said to flow directly from the one- and two-
independent variable cases. This assertion was not formally demonstrated.
With the help of matrix algebra, however, we can see how to identify extreme
points when we deal with functions that have n independent variables.

9.7.1 First-Order (Necessary) Conditions


Consider a function of the form z = f (x1 , x2 , . . . , xn ), with the first partial
derivatives of the function given by f1 , f2 , . . . , fn . For z to have extreme
points, whether maxima or minima, it must be at a “stationary” position.
That is, it is necessary that f1 = f2 = . . . = fn = 0. This is the first-order
condition for finding extreme points.
Chapter 7 demonstrated the first-order condition for extreme point(s) graph-
ically for a function of one independent variable. This graphical represen-
tation had intuitive attractiveness, for it involved drawing a tangent to the
curve of the function at all points at which the slope of that graph was equal
to 0.
The analogous geometry for the case of two independent variables involves
a three-dimensional diagram. A six-dimensional diagram would be needed
to illustrate extreme points when five independent variables are involved. In
general, we need n + 1 dimensions to illustrate the case which involves n
independent variables. It is difficult to draw intelligible three-dimensional
diagrams; four or more dimensions strain both graphical talents and under-
standing. For that reason, we cannot illustrate the geometry of maxima and
minima where n dimensions are involved.

9.7.2 Second-Order (Sufficient) Condition


We state without proof the second-order condition for an extremum of a
function. Given that the first partial derivatives of z = f (x1 , x2 , . . . , xn )
exist and have been set equal to 0 for solution purposes, we must find the
Hessian determinant (or simply Hessian) relating to the function. This Hes-
sian determinant of a function z = f (x1 , x2 , . . . , xn ) is denoted by |H|, and is
composed of elements that are second-order partial derivatives of the function
CHAPTER 9. MATRIX ALGEBRA 302

such that  
f11 f12 · · · f1n
 f21 f22 · · · f2n 
|H| =  .. ..  .
 
..
 . . . 
fn1 fn2 · · · fn

Once we have found the Hessian determinant, one of the following conditions
must hold:

a. When |H1 |, |H2 |, . . . , |Hn | > 0, we have a minimum at the critical point.

b. When |H1 < 0, |H2 | > 0, |H3 | < 0,. . . , (i. e., alternating signs) we have
a maximum at the critical point.

c. When neither (a) nor (b) is true, the test fails, and we must examine the
function in the neighborhood of the critical point in order to determine
whether an extreme point exists.

The terms |H1 |, |H2 |, . . . , |Hn | are principal minors of a Hessian determinant.
The Hessian determinant that is used in the second-order condition is a
symmetric determinant. That is, the main diagonal of the Hessian consists
of all second-order partial derivatives of the function with respect to the
variables of the function; for example, f11 , f22 , . . . , fnn . The off-diagonal
elements in the Hessian are composed of all mixed or cross-partial derivatives
of the function, for example, f12 or f36 , where, according to Young’s theorem,
fij = fji .
The process for defining the Hessian minors is illustrated here for |H1 |, |H2 |,
and |H3 |. The process continues up to and including Hn . (The | | indicates
determinants, not absolute values.)

f11 f12 f11 f21
|H1 | = |f11 | = f11 |H2 | = =
f21 f22 f12 f22

f11 f12 f13 f11 f21 f31

|H3 | = f21 f22 f23 = f12 f22 f32 . . .
f31 f32 f33 f13 f23 f33
and so forth through |Hn |.
CHAPTER 9. MATRIX ALGEBRA 303

We can see that these results are consistent with the conditions that Chapter
7 developed for a single variable. In that case the Hessian contains a single
element, f11 . When |H1 | > 0, we have found a local minimum, and when
|H1 | < 0, we have found a local maximum. This is the same condition that
Chapter 7 develped, that f 00 (x) > 0 indicates a minimum and f 00 (x) < 0
indicates a maximum.
Analogously, a function of two independent variables satisfied the second-
order condition for a minimum in Chapter 7. If fxx · fyy − (f xy)2 > 0 and
fxx , fyy > 0 at the critical point. A maximum existed if fxx · fyy − (fxy )2 > 0
and fxx , fyy < 0 at the critical point. This is precisely what the second-order
condition for the Hessian determinants requires:

fxx fxy
|H1 | = |fxx | = fxx > 0 and |H2 | = = fxx · fyy − (fxy )2 > 0
fyx fyy
for a minimum. Similarly,

fxx fxy
|H1 | = |fxx | = fxx < 0 and |H2 | = = fxx · fyy − (fxy )2 > 0
fyx fyy
are required to establish a maximum.
A function with only one independent variable has only one principal minor.
A function of two independent variables has only two principal minors. A
function with n independent variables has n principal minors. We must
examine each of those principal minors when we seek to determine whether
an extreme point exists. If any one of those principal minors is found to have
an incorrect sign, then we need go no further with the evaluation process.
It is wise to begin with H1 , proceed to H2 , and so forth. If, for example, we
are testing for the existence of a maximum at a critical point, then the signs
of the principal minors will alternate, beginning with a negative. If H1 < 0
and H2 > 0, but H3 > 0, then a maximum point may not exist. In this case,
the test fails, and we must examine the function in the neighborhood of the
critical point in order to determine whether a maximum exists (a process in
which a computer algebra system is quite useful). In any case, we need not
go beyond |H3 | to determine that the test has failed. Finally, if there are
n independent variables, and n is an even number, then the sign of the nth
principal minor must be positive if a maximum exists. If n is odd, then the
sign of the nth principal minor must be negative for a maximum to exist.
Consider two examples.
CHAPTER 9. MATRIX ALGEBRA 304

1. Find the extreme point(s) for this function: z = x2 + x · y + y 2 − 3 · x +2.


1st order conditions: zx = 2 · x + y − 3 = 0 and zy = x + 2 · y.
z zxy
2nd order condition: |H1 | = zxx = 2 > 0 and |H2 | = xx =
z yx z + yy
2 1
1 2 = 3 > 0, so z reaches a minimum point at (2,-1).

2. Find the extreme value(s) for z = x31 + 3 · x1 · x3 + 2 · x2 − x22 − 3 · x23 .


z1 = −3 · x21 + 3 · x3 = 0, z2 = 2 − 2 · x2 = 0, and 3 · x1 − 6 · x3 = 0.
1st order conditions: z1 = −3 · x21 + 3 · x3 = 0, z2 = 2 − 2 · x2 = 0,
and 3 · x1 − 3 · x3 = 0.
The critical points are (0,1,0) and (1/2,1,1/4).

2nd order conditions for (1/2,1,1/4):


|H1 | = |z1 | = | − 6 · x1 | = −6 · x1 = −3 < 0
z z −6 · x1 0
|H2 | = 11 12| = = 12 · x1 = 6 > 0
z21 z 22 0 −2
−3 0 3

|H3 | = 0 −2 0 = −18. The critical value is a maximum.
3 0 −6
2nd order conditions for (0,1,0): |H1 | = | − 6 · x2 | = 0.
The test fails for this critical point. We must, therefore, go back to the
function and examine around the neighborhood of the critical point.
The critical point is neither a maximum nor a minimum.

Exercise 9.6. Determine the critical points, if any, that correspond to local
maxima or minima for the following functions.

1. z = 2 · x2 + y 2 − 2 · x · y + 5 · x − 3 · y + 1

2. z = 2 · x1 + x1 · x2 + 4 · x2 + x1 · x3 + x23 + 8

3. z = 4 · x1 · x2 + 3 · x3 · x21 + x2 · x3

4. z = x21 + x22 + 8 · x23 − x1 · x2 + 10


CHAPTER 9. MATRIX ALGEBRA 305

5. z = x21 + x22 + x23 + x1 · x2 + x2 · x3 − 3 · x1 − 8

6. z = x3 + y 3 + z 3 − 3 · x · y · z

7. z = 5 · x3 − 2 · x · y + 3 · y 2

9.8 Optimization: Maxima and Minima Sub-


ject to Constraints
We now extend the results of the previous section to deal with the situation in
which we wish to identify extreme points in functions that have n independent
variables, and in which the functions are subject to one or more constraints.
The existence of one or more constraints implies two things: the relevant
extreme value cannot be reached, and trade-offs exist. The latter is especially
pertinent to economics, for it implies that an objective function z cannot
typically be maximized by choosing any xi level such that ∂z/∂xi = 0.11

9.8.1 The Case of One Constraint


Consider a function of n variables of the form z = f (x1 , x2 , . . . , xn ) subject
to a constraint given by g(x1 , x2 , . . . , xn ) = 0. We now form a new objective
function of the form

L = L(x1 , x2 , ·, xn , λ) = f (x1 , x2 , . . . , xn ) − λ · g(x1 , x2 , . . . , xn ),

where λ is a Lagrangian multiplier whose value is to be determined by the


maximization (minimization) process.
First-order (necessary) condition
We differentiate the new objective function given in the expression above
with respect to the n + 1 variables x1 , x2 , . . . , xn , and λ, set these partial
11
Dwight Lee, in personal conversation, characterizes this result as follows: Anything
worth doing is worth doing half-assed.
CHAPTER 9. MATRIX ALGEBRA 306

derivatives equal to 0, and solve for their critical roots. Only critical-root
values can be extreme points. However, a critical-root value is not always an
extreme point.
We need a second-order (sufficient) condition in order to make a firm judg-
ment.
Second-order (sufficient) condition
Given that the first partial derivatives of L exist and have been set equal to
0 for solution purposes, we must find the bordered Hessian determinant
relating to the function and its constraints. The bordered Hessian determi-
nant of a function z = f (x1 , x2 , . . . , xn ), subject to g(x1 , x2 , . . . , xn ) = 0, is
denoted by |H B | and is composed of all second-order partial derivatives of
the constraint
such that
0 g1
g2 · · · gn
g1 L11 L12 · · · L1n

|H B | = g2 L21 L22 · · · L2n .

.. .. .. ..
. . . .

gn Ln1 Ln2 · · · Lnn

If all the first-order partial derivatives of the constraint and all the second-
order partial derivatives of the function L exist at the critical point(s), then
one of the following conditions must hold.
(a) When |H2B |, |H3B |, . . . , |HnB | < 0, we have a minimum at the critical point.
(b) When [H2B > 0|, |H3B l < 0, |F4B | > 0, . . ., we have a maximum at the
critical point.
(c) When neither (a) nor (b) is met, the test fails, and we must examine the
function in the neighborhood around the critical point in order to determine
whether a constrained extreme value exists.
A bordered Hessian determinant is a symmetric determinant. It is simply
a Hessian determinant that is bordered by the first partial derivatives of
the constraint, and 0. The symmetry follows from the fact that a Hessian
determinant, which is the major part of a bordered Hessian determinant, is
also symmetric.
It is customary to denote the ith bordered principal minor of a bordered
Hessian determinant by the symbol |HiB |. Among the bordered principal
CHAPTER 9. MATRIX ALGEBRA 307

minors of the bordered Hessian are the following:



0 g1 g2 g3
0 g1 g2
B
B
g1 L11 L12 L13
|H2 | = g1 L11 L12 |H3 | = .
g2 L21 L22 L23

g2 L21 L22
g3 L31 L32 L33

The notation |H2B | means that we must take the Hessian determinant |H2 |
and place around it the appropriate border.
We should note carefully that the process of evaluating bordered Hessian
determinants begins with |H2B |, not with |H1B | We do not evaluate |H1B |
when we maximize or minimize subject to a single constraint. We shall
shortly state a general rule that deals with this situation.
Consider two examples.

1. Find the maximum or minimum value(s) for the function z = x21 −10·x22 ,
subject to this constraint: x1 − x2 = 18.
Our Lagrangian expression is L = z = x21 − 10 · x22 + λ · (x1 − x2 − 18).
The first-order conditions are 2 · x1 + λ = 0, −20 · x2 − λ = 0, and x1 −
x2 − 18 = 0, so the critical values are x1 = 20, x2 = 2, and λ = −40.
Our second-order condition is

0 1 −1

|H2B | = 1 2 0 = 18 > 0,
−1 0 −20

so the constrained maximum value of z occurs when x1 = 20 and


x2 = 2.
2. Find the values of x, y and z that yield a maximum or or minimum
value of w = 5 · x2 + 10 · y 2 + z 2 − 4 · x · y − 2 · x · z − 36 · y.
The first order conditions are
−2 · z − 4 · y + 10 · x + µ = 0
20 · y − 4 · x + 2µ − 36 = 0
2 · z − 2 · x + 4µ = 0
4 · z + 2 · y + x − 12.
The solution to the first order condition is
58 101 82 12
x = ,y = ,z = ,µ = − .
49 49 49 49
CHAPTER 9. MATRIX ALGEBRA 308

The second order conditions are



0 1 2

|H2B | = 1 10 −4 = −76
2 −4 20

and
0 1 2 4

B
1 10 −4 −2
|H3 | =

= −3528.
2 −4 20 0
4 −2 0 2
Thus, we determine that a constrained minimum point on w occurs at
the critical values of x, y, and z.

9.8.2 The Case of Two or More Constraints


The problem of maximizing or minimizing a function of n independent vari-
ables subject to two or more constraints is analogous to the situation in the
preceding section, in which there was only one constraint. In terms of a bor-
dered Hessian determinant, we simply add another border for each effective
constraint.
Consider a function of the form z = f (x1 , x2 , . . . , xn ), which is subject to m
constraints (m < n) given by g(x1 , x2 , . . . , xn ) = 0, h(x1 , x2 , . . . , xn ) = 0, . . . ,
k(x1 , x2 , . . . , xn ) = 0. In order to find any extreme points that might exist, we
form a new objective function of the form L = L(x1 , x2 , . . . , xn , λ1 , λ2 , . . . , λn )
= f (x1 , x2 , . . . , xn ) + λ1 · g(x1 , x2 , . . . , xn ) + λ2 · h(x1 , x2 , . . . , xn ) + . . . , λn ·
k(x1 , x2 , . . . , xn ), where the λ’s are Lagrange multipliers.
First-Order (Necessary) Condition
This new objective function, L, is differentiated with respect to each of the
n + m variables (x1 , x2 , . . . , xn , λ1 , λ2 , . . . , λm ). The resulting partial deriva-
tives are set equal to 0 and solved for critical points. These critical points
may or may not identify actual extreme points. We must apply a second-
order (sufficient) condition to identify those critical points that are, indeed,
extreme points.
Second-Order (Sufficient) Condition
Given that the first partial derivatives of L exist and have been set equal to
CHAPTER 9. MATRIX ALGEBRA 309

zero for solution purposes, we must find the appropriate bordered Hessian
determinant. In this case, the bordered Hessian is given by

0 · · · · · · 0 k1 k2 · · · kn

. .. .. .. ..
.
. . . . .
· · · · 0 0 h1 h2 · · · hn

0 · · · 0 0 g1 g2 · · · gn

|H B | = .
k1 · · · h1 g1 L11 L12 · · · L1n
k2 · · · h2 g2 L21 L22 · · · L2n

. .. .. .. .. ..
.. . . . . .

k · · · h g L L ··· L
n n n n1 n2 nn

If all the first-order partial derivatives of the constraints and all the second-
order partial derivatives of the function L exist at the critical point(s), then
one of the following conditions must hold:

B B B
(a) When |Hm+1 |, |Hm+2 |, . . . , |Hm+n | all have the same sign, namely (−1)n ,
we have a constrained minimum at the critical points.
B B B B
(b) When |Hm+1 |, |Hm+2 |, . . . , |Hm+n | alternate in sign, where |Hm+1 | has
m+1
the sign (−1) , we have a constrained maximum at the critical point.

(c) When the requirements of neither (a) nor (b) aremet, the test fails and
we must examine the function in the neighborhood around the critical
point in order to determine whether a constrained extremum exists.

The practice of beginning the evaluation of the bordered Hessians with some-
thing other than |H1 | continues. The rule that guides this behavior requires
that we begin with a bordered Hessian whose size is one bigger than the num-
ber of constraints. Hence, when m constraints exist, we begin our analysis of
B
the bordered Hessians with |Hm+1 |. We can now look back to our previous
work and explain our previous choices in this regard. When m = 0 and no
constraint exists, we begin with |H1 | When m = 1, we begin with |H2B | and
so on.
The signs that are required for the successive bordered Hessian determinants
follow a definite order. When we evaluate critical point(s) with respect to
a maximum, the bordered Hessians must alternate in sign. In the case in
CHAPTER 9. MATRIX ALGEBRA 310

B
which two constraints exist, the sign of |Hm+1 | = |H3B | must be negative, the
sign of |H4B | must be positive, and so forth. The rule is that the sign is given
by (−1)m+1 . Hence, if m = 2, the sign of |H3B | is (−1)3 = −1, and |H3B | is
negative.
The sign determination for the case of a constrained minimum differs from
that of a constrained maximum. When m = 0, all bordered Hessians must
be positive. When m = 1, all bordered Hessians must be negative. When
m = 2, all bordered Hessians must once again be positive. In general, the
sign of all bordered Hessians must be the same as the sign of (−1)m if a
minimum exists.
As an example, suppose that five constraints apply. Then all bordered Hes-
sian determinants must have a negative sign, for (−1)5 = −1, which is nega-
tive. We can see that when the number of constraints is odd, the signs must
all be negative, whereas when the number of constraints is even, the signs
must all be positive.

Exercise 8.7. Use Lagrangian multipliers to maximize or minimize the


following functions subject to the indicated constraints.

1. z = 2 · x1 + x2 + 2 · x1 · x2 , subject to the constraint 2 · x1 + x2 = 100.


2. z = 25 · x1 · x2 · x3 , subject to the constraint x1 + 2 · x2 + 4 · x3 = 180.
3. w = loge (x)+loge (y)+loge (z), subject to the constraint 5·x+2·y +z =
120.
4. z = x21 + 4 · x22 + x23 − 4 · x1 · x2 − 6 · x3 , subject to the constraint
x1 + x2 + x3 = 21

9.9 Questions and Problems


1. Write the following system of equations in matrix notation.
a11 · x1 + a12 · x2 + a13 · x3 + · · · +a1n · xn = K1
a21 · x1 + a22 · x2 + a23 · x3 + · · · +a2n · xn = K2
.. .. .. .. ..
. . . . .
an1 · x1 + an2 · x2 an3 · x3 + · · · +ann · xn = Kn
CHAPTER 9. MATRIX ALGEBRA 311

2. Given the following national income equations:

1) (1) Y = C + I + G,
(2) C=200 + 0.7 · Y ,
(3) I = 75 + 0.1 · Y ,
(4) G = 100

(a) Find the equilibrium values of C, I, G, and Y .


(b) What is the numeric value of the multiplier that we would use to
determine the effects of autonomous changes in I on Y ?
(c) Suppose that autonomous investment increases from 75 to 100.
What increase in Y results? (Show this result algebraically and
also by means of matrix algebra.)
(d) Demonstrate that your numeric solutions to parts (a) and (c) are
general equilibrium solutions in the sense that they are internally
consistent with each other in magnitude.

3. Find the equilibrium prices (Pi ) and quantities (Qi )for goods A, B,
and C using inverse matrix algebra. Then use Maxima to check your
answer.
QD,A = 8 − 2 · PA + 3 · PB − PC QD,B = 4 − 4 · PB + PA + 3 · PC
QD,C = 6 − PC + 3 · PA + 3 · PB
QS.A = 10 QS,B = 2 · PA + 2 QS,C = 8 + PC

4. Find the equilibrium position of a consumer, given the following utility


function and budget information:
U = log(x) + log(y) + log(z)
M = $120 Px = $4 Py = $2 Pz =4
M is money income per period, (x, y, z) is the per-period consumption
levels of the three goods, and U is the utility level. The prices are
Px , Py , and Pz .
Chapter 10

Linear Programming and


Input-Output Analysis

Matrix algebra that has several important uses beyond those demonstrated in
Chapter 9. In particular, matrix algebra underlies two powerful quantitative
techniques, linear programming and input-output analysis. Indeed, these are
just extension of matrix algebra, but we devote a separate chapter to them
only as a means of focusing attention on their importance to economists.
We begin by considering linear programming in a matrix algebra framework,
and subsequently examine input-output analysis in the same fashion.

10.1 Linear Programming


Linear programming is a mathematical technique used to derive economically
efficient solutions to problems that arise in a wide range of situations. Lin-
ear programming always involves the maximization or minimization of some
function, subject to various constraints. For example, a school district may
wish to minimize the cost of busing students from one location to another,
with the students’ initial locations considered as given. A firm might wish to
maximize its output given that it faces certain input prices and has a finite
limit on how much money it can spend on those inputs. Or a government
agency may seek to minimize the cost of feeding, clothing, and equipping

312
CHAPTER 10. LINEAR PROGRAMMING 313

an infantry division, subject to constraints concerning the location of the


resources used and how much of each resource is required.
Linear programming is a mathematical technique whereby one maximizes or
minimizes a linear function subject to various constraints, or side conditions,
which are stated in the form of linear inequalities.

10.1.1 Stigler’s Diet Problem Once Again


Chapter 1 considered a classic economic allocation exercise, Stigler’s diet
problem. We now recast the diet problem as a linear-programming problem
that uses matrix algebra. The crux of the diet problem is to find the least
expensive combination of foods available to consumers that will allow these
same consumers to satisfy recommended daily dietary allowances established
by the Food and Nutrition Board of the National Academy of Sciences. That
is, Stigler’s objective was to find the least expensive menu that would give
the consumer required minimum levels of calories, vitamins, and so forth.1
Stigler allowed the consumer to choose a menu from among 80 possible foods.
Let Xj refer to the quantity of food j that an individual consumes per time
period. We can then represent the quantities of the 80 different foods con-
sumed by the individual as X1 , X2 , . . . , X30 . The Xj , with j = 1, 2, . . . , 80,
are decision variables, or unknowns, that are to be determined by solving the
linear- programming problem.
Any attempt to determine the least expensive way to accomplish a given end
must necessarily consider the prices of various alternatives. Let Pj represent
the dollar price of food type j. Thus the prices of the 80 different foods that
can be selected are represented by P1 , P2 , · · · , P80 . For example, P1 might
be $0.02 and represent the price of peanut butter per ounce. It follows that
P1 · X1 , the price of peanut butter per ounce times the number of ounces
purchased, is the consumer’s total expenditure on peanut butter.
The equation below defines the total expenditure C that the consumer makes
on the 80 different foods. This equation is referred to as the objective func-
tion; it is this function that we seek to minimize as a part of Stigler’s diet
1
Recall that, as a matter of historical accuracy, the formal technique of linear program
was invented a few years after Stigler’s research. Stigler used a series of approximations
to derive a result that direct application of linear program would have yielded more easily.
CHAPTER 10. LINEAR PROGRAMMING 314

problem. This objective function is linear in the XjP ’s, which we call the
decision variables.Thus, our objective function is C = 801 Pj · X j .
2

An obvious way to minimize C is to not spend any money on food, but this
is not permissible, for such a menu plan would not satisfy the recommended
daily dietary allowances, the constraints. More bluntly, it would kill the
consumer. Let R symbolize a dietary requirement. Thus R1 might represent
the recommended intake of calories per individual per day. In Stigler’s 1945
version of the diet problem, the recommended intake of calories per individual
per day was 3000. Fewer calories would presumably be detrimental to the
consumer’s health. We represent the nine different dietary requirements of
Stigler’s problem by the variable names R1 , R2 , . . . , R9 .
The diet problem is inherently challenging because two different foods seldom
yield the same amount of nutrient per ounce of food. For example, in Stigler’s
1945 exercise, 1 ounce of uncooked bacon yielded 186 calories, whereas 1
ounce of uncooked sirloin steak yielded only 88 calories. Let the symbol aij
represent the number of units of nutrient i that are provided by 1 ounce of
food j. Hence the aij for uncooked bacon (in terms of calories) is 186, while
the analogous aij for uncooked sirloin steak is 88.
A consumer can satisfy a nutrient requirement by eating many different foods.
The term aij · Xj represents the total number of units of nutrient i that
are obtained when one consumes a given number of ounces of food j. For
example, if one consumes 6 ounces of uncooked bacon, then Xj = 6, and
since aij = 186, aij · Xj = 186 · 6) = 1116 calories.
We have previously noted that Stigler assumed that the individual must
obtain at least 3000 calories per day from the foods consumed. The consumer
can choose among the 80 foods in order to satisfy this requirement. We can
write this constraint as follows, stating that the sum of all the calories the
consumer derives from consuming various foods must be 3000 or greater:
a1,1 · X1 + a1,2 · X2 + · · · + a1,80 · X80 ≥ 3000.
2
Nonlinear objective functions are not permissible in a linear programming problem.
If the researcher attempts to represent the underlying phenomenon by a linear equation
when it is actually nonlinear, then the results obtained will be inaccurate and unreli-
able. Fortunately, nonlinear programming techniques are now accessible, given advances
in mathematics and in computing power. Even spreadsheet programs like Excel typically
offer ways to address nonlinear systems. We illustrate nonlinear programming at the end
of this section.
CHAPTER 10. LINEAR PROGRAMMING 315

The nine daily dietary requirements that Stigler imposed on the consumer in
his diet problem were reported in Chapter 1. Each of these nine requirements
constitutes a constraint on the consumer’s activities that takes the form of a
linear inequality similar to the expression above that addresses caloric intake.
The nine constraints are
a1,1 · X1 + a1,2 · X2 + · · · +a1,80 · X80 ≥ R1
a2,1 · X2 + a2,2 · X2 + · · · +a2,80 · X80 ≥ R2
.. .. .. ..
. . . .
a9,1 · X9 + a9,1 · X2 + · · · +a9,80 · X80 ≥ R9

The constraints or requirements n this set of equations are, like the objective
function, linear in the decision variables Xj .
We must include one additional, seemingly obvious set of constraints, that
Xj ≥ 0 for all j = 1, 2, . . . , 80. These restrictions are known as nonnega-
tivity constraints. They explicitly restrict the solution values of the decision
variables to be either zero or positive, thus eliminating the possibility of a
nonsensical solution that might (for example) allow the individual to consume
-5 ounces of bacon.
A formal expression of the problem at hand is that we must minimize our
objective function C subject to the nine inequalities and the eighty nonneg-
ativity conditions being satisfied.
We now show a variant of the Stigler solution, which also appears in Chapter
1. For purposes of exposition, Stigler focused on a subset of five foods and
eight constraints. The solution to this subset yielded results that were close
to that of the larger model. The problem consists of the following nine
equations, with the first on being the objective function.

x5 + x4 + x3 + x2 + x1 (z)

26.9x5 + 1.1x4 + 2.6x3 + 8.4x2 + 44.7x1 >=3 (c1)


1691x5 + 106x4 + 125x3 + 422x2 + 1411x1 >=70 (c2)
11.4x5 + 4x3 + 15.1x2 + 2x1 >=0.8 (c3)
792x5 + 138x4 + 36x3 + 9x2 + 365x1 >=12 (c4)
918.4x4 + 7.2x3 + 26x2 >=5 (c4)
CHAPTER 10. LINEAR PROGRAMMING 316

38.4x5 + 5.7x4 + 9x3 + 3x2 + 55.4x1 >=1.8 (c5)


24.6x5 + 13.8x4 + 4.5x3 + 23.5x2 + 33.3x1 >=2.7 (c6)
217x5 + 33x4 + 26x3 + 11x2 + 441x1 >=18 (c7)
2755x4 + 5369x3 + 60x2 >=75. (c8)
The names of the expressions are in at the right. The expression z is the ob-
jective (cost) function, for which we seek a minimum value. The expressions
c1 · · · c8 are the constraints. The next set of commands loads the simplex
program (more on this below) and enters the relevant command, which iden-
tifies the objective function, the eight constraints, and the nonnegativity
conditions: load(simplex)$ minimize lp(z, [c1, c2, c3, c4, c5,
c6, c7, c8, x1>=0, x2>=0, x3>=0, x4>=0, x5>=0]).
The model yields the output in a list. The first item in the list is the value
of z. The second item in the list is an embedded list of five x values:
[0.10904, [x5 = 0.048628, x4 = 0.0051128, x3 = 0.01125,
x2 = 0.0085915, x1 = 0.035456]].
The results show that the cost is $0.10904 (which would correspond to about
$0.80 in 2016 dollars ).

10.1.2 A Formal Definition of a Linear-Programming


Model
The previous example, which stated Stigler’s diet problem as a linear-programming
problem, exhibits the three fundamental properties of any linear-programming
problem:

1. The object of the problem is to find optimal values for the decision
variables or unknowns in the problem.
2. The optimal values of the decision variables are such that they either
minimize or maximize an explicit linear objective function.
3. The minimization or maximization solution of the objective function
must be feasible. That is, the values of the decision variables in the
optimal solution must satisfy both the linear inequality constraints and
the nonnegativity constraints.
CHAPTER 10. LINEAR PROGRAMMING 317

Our general linear-programming model has the following structure:]


Maximize or Minimize:
Z=b1 · X1 + b2 · X2 + . . . + bn · Xn
Subject to:
a11 · X1 + a12 · X2 + · · · +a1n · Xn (≤, =≥) c1
a21 · X2 + a22 · X2 + · · · +a2n · Xn (≤, =≥) c2
.. ..
. .
am1 · Xn + am2 · X2 + · · · +amn · Xn (≤, =≥) cm
and
Xj ≥ 0 for all j = 1, 2, . . . , n,
where n = the number of decision variables, and m = the number of con-
straints or side conditions.
Each of the decision variables (Xj ’s) must appear in the objective function
of a linear-programming problem, Z. Only one sign (≤, =, or ≥) can hold
in any particular constraint. The bj , aij , and ci are the parameters of the
model. The ci ’s are the amounts of scarce resource i available for allocation
(i = 1, 2, . . . , m). The parameter aij represents the amount of resource i
that is consumed by, or allocated to, each unit of decision variable j (for
i = 1, 2, . . . , m and j = 1, 2, . . . , n). The change in Z that results from a unit
increase in Xj is represented by bj . For example, if bj = −5, then a one-unit
increase in Xj decreases Z by five units.
We can write the general linear-programming model more compactly using
matrix notation as follows:
Maximize or Minimize: Z = B · X
Subject to:
A · X (≤, =≥), C and X ≥ 0, where,
     
X1 c1 a11 a12 · · · a1n
 X2     c2   a21 a22 · · · a2n 
X =  ..  , B = b1 b2 · · · bn , C =  ..  , A =  ..
     
.. .. 
 .   .   . . . 
Xn cm am1 am2 · · · amn

10.1.3 A Graphical Illustration


Although a graphical approach to solving linear programming problems is
seldom very efficient in terms of either time or effort, the graphical approach
CHAPTER 10. LINEAR PROGRAMMING 318

does offer a clear understanding of what linear programming actually in-


volves. We use the graphical approach in this section to solve a relatively
simple linear-programming problem. This should enable you to visualize
what is happening when you encounter a more complex linear-programming
solution.
We demonstrate the graphical approach to linear programming by a problem
concerning a firm, Acme Manufacturing Company, that produces two prod-
ucts, X1 and X2 . Acme is a price-taker in the market for both goods. The
prices are $20 per unit for X1 and $15 per unit for X2 .
Acme’s production is limited by three resource constraints. Producing both
X1 and X2 requires three inputs, which we label a, b, and c. Thus, Acme
has two production functions: X1 = X1 (a, b, c) and X2 = X2 (a, b, c). For
the relevant production period, Acme has only 60 units of input a, 24 units
of input b, and 84 units of input c. Acme cannot augment the quantities of
these inputs during the current production period.
For each unit of good X1 that it produces, Acme uses 5 units of input a, 3
units of input b, and 12 units of input c. The production of a single unit of
good X2 requires 15 units of input a, 4 units of input b, and 7 units of input
c. Acme cannot alter these input-output relationships during the current
production period. It need not use all of any one of the inputs.
Acme’s objective is to maximize the revenue R that it receives from the
sale of goods X1 and X2 , while at the same time not violating any resource
constraint that it faces. We can state this problem in linear-programming
terms as follows:
Maximize: R = P1 · X1 + P2 · X2 = 20 · X1 1 + 15 · X2 ,
subject to:
5X1 + 15X2 ≤ 60 (input constraint a),
3 · X1 + 4 · X2 ≤ 24 (input constraint b),
12 · X1 + 7 · X2 ≤ 84 (input constraint c), and
X1 , X2 ≥ 0 (nonnegativity constraints).
Once we have formulated the linear programming problem, the next step
in the graphical approach is to delineate the solution space. We do this by
graphing the constraints (as shown in Figure 10.1). The constraints, when
graphed, enclose a set of feasible (possible) solutions that constitute the solu-
tion space. Linear programming selects the point within that solution space,
CHAPTER 10. LINEAR PROGRAMMING 319

Figure 10.1: Graphic representation of the Acme production problem

or feasible region, that maximizes (minimizes) the value of the objective func-
tion.
If we ignore, for the time being, the less-than sign in the linear constraints of
Acme’s linear programming problem, and therefore treat these equations as
if they had equals signs then we can begin to provide a visual representation
of the linear-programming problem. From Chapter 2, we know that we can
graph a straight line without great difficulty if we are given two points on that
line. For example, if we know the horizontal intercept (abscissa) and vertical
intercept (ordinate) of a line, then we can connect these two intercepts with
a straight line and obtain the needed graph. We use this technique to define
the solution space for Acme. Figure 10.1 illustrates this procedure.
Figure 10.1 also shows an iso-revenue line (for R = $200–any value will do).
This line lies outside the feasible region, so this level of revenue cannot be
attained.
The shaded area in Figure 10.1 has four corners (not counting the origin,
where R = 0). Two of the points are interior and involve the production of
positive amounts of both products. Two points are on the axes and involve
the production of just one of the two goods. Figure 10.2 replicates Figure
10.1 and focuses on the feasible region.
CHAPTER 10. LINEAR PROGRAMMING 320

Figure 10.2: The feasible region of the Acme problem

The accompanying workbook shows that the interior solutions occur at (24/5,
12/5) and (56/9,4/3), or approximately (4.8, 2.4) and (6.22, 1.33).
An infinite number of points lie either within or on the boundary of feasible
region. This means that an infinite number of possibilities confront us when
we attempt to identify the solution that is optimal. To illustrate the search
technique that linear programming performs, we consider a few specific points
within the solution space. At the origin, as we have seen, R = 0. Acme can,
however, produce positive quantities of goods X1 and X2 that generate some
sales revenue, so the solution at the origin is not optimal. We can do better.
We assert that any movement from the origin that involves production of
more of one good and no less of the other adds to revenue. Such moves
are possible whenever the (X1 , X2 ) combination is inside the bondary of the
feasible region. Therefore, any point inside the solution space represents less
production (and therefore less total sales revenue) than at least one point on
the boundary of the solution space. Thus, the optimal solution lies on the
boundary of the solution space, not inside it.
When we are considering points as candidates for the optimal solution, we
can ignore any point inside the solution space. Knowledge of this fact sub-
stantially reduces the number of possible solutions with which we must con-
CHAPTER 10. LINEAR PROGRAMMING 321

tend. We proceed therefore to consider only those solution points that are
on the boundary. We begin with the values, X1 = 7, X2 = 0 which yields
R = $140. From this point, we move counter-clockwise to (6.22,1.33), at
which R = $144.35. The next move, to (4.8, 2.4) generates revenue of
$132.00. Finally, when no X1 is produced and X2 = 4, the revenue level
falls to $60. Therefore, among the corners, the largest revenue occurs where
X1 = 6.22 and X2 = 1.33.
One of the corner points of the solution space is always an optimal solution to
a linear-programming problem, so we are justified in ignoring the linear seg-
ments between these corners. This assertion may not be intuitively obvious,
however, and requires more explanation.
Consider Figure 10.3, which graphs the constraints of the Acme linear pro-
gramming problem. The solution space (feasible region) is indicated by the
black border. Graphs that correspond to five values of the objective function,
R = $20X1 + $15X2 also appear. Four of the five revenue levels are feasible;
they are the values calculated above. For the highest of these five revenue
values, the iso-revenue line is tangent to the feasible region boundary. Thus,
a single value on that line is feasible. For R > $144.44, no point is in the
feasible region. Thus, the iso-revenue line for R = $160, shown in yellow,
cannot be attained.
In general, a corner of the solution space will always be an optimal solution to
a linear programming problem. Only when the slope of the objective function
is the same as the slope of a binding constraint will there be more than one
optimal solution. In the unlikely event that the slope of the objective function
in Figure 10.3 were the same as the slope of any one of the three line segments,
then all the points on that line segment (including the corner points) would
be optimal. Thus, the value of the objective function at the corner point is
at least as high (or low, if the program involves minimization) as the points
on the line segment, so only corner points need be considered.
An efficient linear-programming solution technique is ordinarily used to iden-
tify the corner points in a problem, evaluate them, and select the optimal
solution from among those possibilities. The simplex algorithm, one such
technique that is frequently used, is the one that Maxima used above to solve
the subset of the Stigler diet problem. This algorithm is a search technique
that identifies and evaluates the corner points in a problem. It repeatedly
strives to find a better solution than the one at hand. When it reaches an
CHAPTER 10. LINEAR PROGRAMMING 322

Figure 10.3: The feasible region and revenue levels for Acme

optimal point, such as the one that we identified in the Acme problem, it
stops. A movement to any other corner would result in a worse solution.
The table below shows the objective function R and the three constraints.
 
R 15X2 + 20X1
C1 15X2 + 5X1 ≤ 60
 
C2 4X2 + 3X1 ≤ 24 
C3 7X2 + 12X1 ≤ 84

The following commands load the simplex module and execute the command
to execute the linear programming tool: load(simplex)$
float(maximize lp( R, [C1,C2,C3,X1>=0,X2>=0]));. The maximize lp
command is embedded in a float command to generate easily-interpreted
values rather than exact values. It is optional. The results are consistent
with our computations above: [144.44,[X2=1.3333,X1=6.2222]]. The
first item that the command reports is the value of the objective function.
The second item is a list of the two production levels.
CHAPTER 10. LINEAR PROGRAMMING 323

10.1.4 The Dual Problem


We can actually view every linear-programming problem as consisting of
two separate problems. The original linear-programming problem is called
the primal problem, while a second formulation of this original problem is
known as the dual problem. The dual problem in linear programming often
yields results that are quite useful to analyst.
Also, the dual problem is sometimes easier to solve than the original (primal)
problem. Therefore, when the primal problem is intractable or just difficult
to solve, we can sometimes solve the dual problem, then use the information
from that dual to solve the primal problem.
Suppose that the original (primal) linear-programming problem is this:
Maximize: Z = b1 · X1 + b2 · X2 + · · · + bn · Xn
Subject to m constraints:
a11 · X1 + a12 · X2 + · · · + a1n · Xn ≤ C1
a11 · X1 + a12 · X2 + · · · + a1n · Xn ≤ C1
.. .. .. ..
. . . .
a11 · X1 + a12 · X2 + · · · + a1n · Xn ≤ C1
and
Xj ≥ 0 for all j = 1, 2, . . . , n.
The dual problem associated with this primal problem requires minimization.
The problem is this:
Minimize: W = d1 · Y1 + d2 · Y2 + · · · + dm · Ym ,
subject to
a11 · Y1 + a21 · Y2 + · · · +am1 · Ym ≥ b1
a22 · Y2 + a22 · Y2 + · · · +am2 · Ym ≥ b2
.. .. .. ..
. . . .
a1n · Y1 + a2n · Y2 + · · · +amn · Ym ≥ bn
where Y1 , Y2 , . . . , Ym are the dual variables.
A definite symmetry exists between a primal problem and its dual. When
the primal problem involves the maximization of a function, then the dual
problem involves the minimization of some function. When the constraints on
the primal objective function require the firm to arrange its activities so that
its input use and expenditures are less than or equal to certain constants,
the constraints on the dual objective function require that certain of the
CHAPTER 10. LINEAR PROGRAMMING 324

firm’s activities be equal to or exceed certain constant levels. For example,


if the primal problem is to maximize output given a cost constraint of $200,
then the dual problem is to minimize the cost of producing a certain level
of output, perhaps 10 units. We shall be more specific about the symmetry
between the primal and dual linear-programming problems in a moment.
One of the most interesting results of many dual problems in linear program-
ming is the idea of a shadow price. Decision-makers often wish to estimate
the value of contributions that various inputs make to the optimal solution.
On occasion, not all the available units of a particular input are used in the
optimal solution. In such a circumstance, the decision-maker has little need
for additional units of such inputs. Therefore the shadow price of such inputs
is 0, because an additional unit of such an input would not alter the optimal
solution. For example, the value of an extra seat in a classroom that already
has many empty seats is 0. Similarly, the value of a second textbook to a
student who already has one is 0 unless the student loses the first textbook.
Thus, a shadow price indicates the value at the margin that an input has for
the objective function’s activities.
We can, in many situations, interpret the shadow price of an input as the
value of the marginal product of that input. This is an interesting and quite
useful result, particularly in decision situations in which the input in question
is not purchased in the market, or in which the connection between that input
and the ultimate output seems quite distant at best. The shadow price in
this case indicates the price that the decision-maker would be willing to pay
for additional units of this input.
Economic planners in nonmarket economies such as the former Soviet Union
have made considerable use of the shadow prices of linear programming.
Since market prices often did not exist in the Soviet Union, it was difficult for
planners to cost, price, and value things efficiently. Shadow prices provided
some guidance in the absence of market signals.
Similarly, large organizations (the military services are a prime example) that
do not vend their wares in a conventional fashion can use shadow-pricing
techniques to increase the efficiency of many activities, such as purchasing,
routing, and intra-branch transfers.
CHAPTER 10. LINEAR PROGRAMMING 325

The Symmetry between the Primal and the Dual Problems

1. If the primal problem involves maximization, then the dual problem


involves minimization, and vice versa.

2. If the primal problem involves ≤ constraints, then the dual problem


involves ≥ constraints.

3. The coefficients of the variables in the primal objective function are the
right-hand constants of the constraint equations in the dual problem.

4. The coefficients of the variables in the dual objective function are the
right-hand constants of the constraint equations in the primal problem.

5. A new set of variables Yi appears in the dual objective function and


constraint equations. These Yi values are the shadow prices of the
inputs.

6. If the primal problem consists of n decision variables and m constraint


equations, then the dual problem consists of m variables and n con-
straints.

7. The coefficients of the constraint equations in the primal problem are


the same as the coefficients of the constraint equations in the dual
problem except that the rows and columns are interchanged. That is,
each aij now becomes afi. In matrix notation,
A = [aij ], for i = 1, 2, . . . , m and j = 1, 2, . . . , n
in the primal problem, then A0 = [aji ] is associated with the dual
problem.

8. The nonnegativity constraints apply to all variables in both the primal


and the dual problems.

9. The optimal solution is identical for both the primal and the dual
problems.

Examples

1. Given that the primal problem is the Acme Manufacturing situation is


maximize: R = 20 · X1 + 15 · X2
CHAPTER 10. LINEAR PROGRAMMING 326

subject to:
5 · X1 + 15 · X2 ≤ 60
3 · X1 + 4 · X2 ≤ 24
12 · X1 + 7 · X2 ≤ 84 and
X1.X2 ≥ 0,
then the dual problem is
minimize: C = 60va + 24 · vb + 84 · vc
subject to:
5 · va + 3 · vb + 12 · vc ≥ 20
(The “value” to the firm of selling a unit of X1 is attributed to the
inputs, which could be used in producing X2 . Likewise for X2 below.)
15 · va + 4 · vb + 7 · vc ≥ 20
va , vb ≥ 0
The optimal solution is: X1 = 6.22, X2 = 1.33 R = 144.35 va = 0,
vb = 1.47, and vc = 1.30
The v’s are shadow prices. The objective is to minimize cost, C, based
on these shadow prices. The value of va is 0, because units of input a
remain unused after the optimal solution has been implemented.
Here is the relevant information in the form of Maxima commands:
Cost: 60*va + 24*vb + 84*vc; Ca: 5*va + 3*vb +12*vc >= 20;
Cb: 15*va + 4*vb + 7*vc >= 15;. The command minimize lp(
Cost, [Ca,Cb,va>=0, vb>=0, vc>=0]); implements the simplex method
and yields the following results (a float command was invoked also):
[144.44, [vc = 1.2963, vb = 1.4815, va = 0.0]].
The values differ slightly from those above, because Maxima’s output
involves less rounding error.
A digression: We can return to the primal problem look at the shadow
price in a slightly different way, one that relates to the Lagrangian
multipliers that we encountered earlier. The commands below, sequen-
tially add 1 unit of inputs a, b, and c, holding the other two at the
initial levels: C1alt: 15*X2+5*X1<=61$ C2alt:4*X2 + 3*X1 <=
25$ C3alt:7*X2 + 12*X1<=85$. The results, after invoking a float
command are these:
 
[144.44, [X2 = 1.3333, X1 = 6.2222]]
 [145.93, [X2 = 1.7778, X1 = 5.963]] 
[145.74, [X2 = 1.2222, X1 = 6.3704]]
CHAPTER 10. LINEAR PROGRAMMING 327

Adding a unit of a changes nothing. Adding either b or c changes


everything: X1 , X2 , and R all increase. Subtracting the initial R,
144.44, from each of the two higher values, yields the shadow prices of
b and c.
2. Given that the primal problem is
minimize: W = 2 · X1 + 5 · X2
subject to:
5 · X1 + 6 · X2 ≥ 12
−3 · X1 + 4 · X2 ≥ 10
X1 + 5 · X2 ≥ 8
2 · X1 + X2 ≥ 3 and
X1 , X2 ≥ 0
then the dual problem is
maximize:
U = 12 · Y1 + 10 · Y2 + 8 · Y3 + 3 · Y4
subject to:
5 · Y1 − 3 · Y2 + Y3 + 2 · Y4 ≤ 2
6 · Y1 + 4 · Y2 + 5 · Y3 + Y4 ≤ 5
Y1 , Y2 , Y3 , Y4 ≥ 0

Exercise: Draw the four constraints in the primal problem for Example 2
and shade the feasible region. Use the graph to estimate the values of X1
and X2 . Which constraints are binding? See the accompanying workbook to
determine the exact values of X1 and X2 and to confirm that two of the Y
values are zero.

10.1.5 Nonlinear Programming


We noted earlier that when one or more constraints, or the objective function,
is not linear, then a linear programming solution may either be impossible or
misleading. Fortunately, modern computers can examine nonlinear models,
often quite rapidly. This is so, despite the fact that nonlinear programming
typically involves many more computations. Maxima offers the nonlinear
programming option cobyla (constrained optimization by linear algebra).
Bradley, et al. [3] provides a detailed overview of the nature of nonlinear
programming and of the types of approaches to implementing such program-
ming.
CHAPTER 10. LINEAR PROGRAMMING 328

We provide a brief overview in two steps. First, we revisit the diet problem,
comparing the simplex solution to that provided by cobyla. Then we set up
and solve a simple nonlinear problem that appears in Bradley et al.
A nonlinear problem cannot be addressed with linear programming, but a
linear problem can be addressed with nonlinear programming. We saw earlier
that the simplex method could be implemented with the following commnds:
load(simplex)$
minimize lp(z, [c1,c2,c3,c4,c5,c6,c7,c8,
x1>=0,x2>=0,x3>=0,x4>=0,x5>=0]);
where z is the objective function, and the brackets contain a list of con-
straints. The resulting output is [0.10904, [x5 = 0.048628, x4 = 0.0051128,
x3 = 0.01125, x2 = 0.0085915, x1 = 0.035456].
The interpretation appears early in this section
The cobyla counterparts are the following input: load(fmin cobyla);
fmin cobyla(z, [x1,x2,x3,x4,x5], [0.031,0.01,0.01,0.005,0.005],
constraints = [c1,c2,c3,c4,c5,c6,c7,c8,
x1>=0,x2>=0,x3>=0,x4>=0,x5>=0],iprint=1);.
This input differs from the simplex input only in minor details, with one
exception. The minor details are that the constraints are explicitly identified
as such and that a code iprint=1 is added in order to control the amount of
detail that is reported. The more substantive difference is that this command
requires a list of initial guesses of the values of x1 , x2 , . . .. In general, the
better guesses one can provide the better the method will work. In particular,
close guesses reduce the number of iterations required to obtain the desired
results.
As with the input, the output differs in detail from the simplex output:
[[x1 = 0.035456, x2 = 0.0085915, x3 = 0.01125,
x4 = 0.0051128, x5 = 0.048628], 0.10904, 42, 0].
The output appears with thex values first, followed the the value of z. The
number 42 is the number iterations that were required. The 0 is a code that
indicates that the process was completed without error.
We now consider a simple example that is not amenable to linear program-
ming. We wish to maximize the function z = 2 ∗ x1 − x12 + x2. Un-
fortunately cobyla is limited to minimization, but we can minimize −z =
−(2 ∗ x1 − x12 + x2). The constraints are these: x2 ≤ 1.8, x21 + x22 ≤ 4,
x1 ≥ 0, and x2 ≥ 0. The command that we enter is load(fmin cobyla)$
CHAPTER 10. LINEAR PROGRAMMING 329

Figure 10.4: Feasible region and objective function, nonlinear

fmin cobyla(-(2*x1 - x1^|2 + x2), [x1,x2], [1,1.8],


constraints = [x2<=1.8,x1^2 + x2^2<=4,x1>=0,x2>=0], iprint=1);. We
used Figure 10.4 to guess that x1 would be near to 1 and that x2 would equal
1.8.3
The resulting output,

[[x1 = 0.87178, x2 = 1.8], −2.7836, 20, 0]

confirms our expections. The value of −z is -2.7836, so z = 2.7836 is the


highest attainable value of our objective function.

10.2 Input-Output Analysis


The system of markets in the United States contains millions of separate
and independent economic decision units–households, business firms, not-
3
As an exercise, repeat this with guess of 10 and 10 or some other value. Confirm that
the number of iterations increases.
CHAPTER 10. LINEAR PROGRAMMING 330

for-profit organizations, and government agencies. Each decision unit is in-


terested primarily achieving its own set of purposes, and seemingly pays
little heed to the survival and behavior of most of the other decision units.
Nonetheless, as Adam Smith persuasively demonstrated more than two cen-
turies ago, the self-serving efforts of millions of independent economic decision-
makers are somehow harnessed and drawn together by the functioning of an
economic system.
Day in, day out, this economic system provides approximately the correct
quantities of food, clothing, shelter, and other goods that consumers wish to
purchase. (Correct, in the sense that persistent shortages and surpluses are
seldom observed.) This is Smith’s “invisible hand ” at work, for no central
planning agency wills this to take place. The individual decision-makers,
perhaps intending only personal good, unwittingly (or otherwise) does public
good as well.
How and why does the economic system hang together? What are the inter-
relationships between inputs and outputs that affect our everyday lives? The
answers to questions like these are provided by general equilibrium analysis.
Adam Smith was an early practitioner of general equilibrium analysis, which
explicitly includes and analyzes reactions and feedback effects among large
numbers of variables. This analysis is based on the assumption that all
markets and all decision-makers are affected by one another’s actions to some
degree. In general equilibrium analysis, the price of oil affects the price of
gasoline, but it also affects the price of automobiles, plastic drinking cups,
and the temperature at which you choose to heat your home.
This section introduces input-output analysis, a general equilibrium approach
to production. Input-output analysis is entirely empirical in nature. It de-
scribes the relationship among various inputs and various outputs. Demand
conditions play no part in input-output models, because input-output anal-
ysis examines only the relationship between inputs to the productive process
and the outputs that result. The prices and quantities at which we might
produce and sell such inputs and outputs and such factors as price elasticity
of demand are not considered.4
4
Attempts to generalize this analysis by allowing for demand considerations and for
some elasticity of substitution in production constitute a set of techniques given the name
“computable general equilibrium” models. See Mitra-Kahn [15] for an overview and history
of these models.
CHAPTER 10. LINEAR PROGRAMMING 331

The father of input-output analysis, Wassily Leontief, Leontief used input-


output analysis to show how the production of one sector of the economy de-
pends, to some degree, on the production of all other sectors of the economy.
The input-output“tables” that Leontief produced described the numeric re-
lationships between inputs used and the outputs produced in the American
economy. This set of relationships was stated in physical terms (for example,
tons of steel or gallons of gasoline) rather than monetary terms. The price
of a ton of steel was not considered.
Most input-output models rest on three assumptions. The first is that no two
commodities are produced jointly. Each firm or market is assumed to produce
only one homogeneous product. Second, all inputs are employed in rigidly
fixed proportions in production. The law of diminishing returns does not
apply because input proportions never vary. Also, this assumption implies
constant returns to scale in all production. Third, no external economies
or diseconomies exist for any firm or market. Thus, the production by one
firm cannot affect the technology that governs the production of any other
product in either a positive of negative fashion.
Let the total production of any single industry during a particular period be
represented by Xi . Industry i’s production can be used as inputs in other
productive processes, or it can be used to satisfy final consumption demand.
If there are n different industries, then potentially n different industries can
use output Xi ,as an input. Therefore, we can write Xi = Xi1 + Xi2 + . . . +
Xin + di , where Xi = output of industry i, Xij j = output of industry i used
as an input in industry j, and di = final demand for the finished goods and
services of industry i.
We pointed out above that input-output analysis assumes that all production
takes place under conditions of rigidly fixed proportions. Thus the amount
of steel required to produce one car does not change, regardless of the num-
ber of cars produced. Let aij represent a technical coefficient of production.
Specifically, aij is the (constant) number of units of input j that are required
to produce one unit of output i. We can therefore express the production of
industry i as follows:
n
X
Xi = aij · Xj + di i = 1, 2, . . . , n,
j=1

where Xi = output of industry l, (1,7 = number of units of input j needed


CHAPTER 10. LINEAR PROGRAMMING 332

to produce one unit of output i, Xj = output of industry j, and di = final


demand for the finished goods and services of industry i.
In an economy that has n industries, there are n × n technical coefficients
of production to consider, since each industry can potentially provide inputs
to every other industry (including itself). Therefore, we can write a full
system of linear equations that describes the input-output relationships for
an economy composed of n industries:
X1 = +a11 · X1 + +a12 · X2 + . . . +a1n · Xn + d1
X2 = +a21 · X1 + +a22 · X2 + · · · +a2n · Xn + d2
.. .. .. .. ..
. . . . .
Xn = +an1 · X1 + +an2 · X2 + · · · +ann · Xn + dn

These equations could equivalently be written in terms of the final demands


for goods and services, the dis. This version of the input-output model
indicates where the final goods and services in the economy are produced:
d1 = (1 − a11 ·)X1 − a12 · X2 − ··· −a1n · Xn
d2 = −a21 · X1 + (1 − a22 ) · X2 − · · · −a2n · Xn
.. .. .. ..
. . . .
di = −ai1 · X1 − ai2 · X2 − ··· +(1 − aii ) · Xi + · · · − ain · Xn
.. .. .. ..
. . . .
dn = −an1 · X1 − an2 · X2 − ··· +(1 − ann ) · Xn

In matrix notation, this system of linear equations is expressed as


     
(1 − a11 −a12 ··· −ain X1 d1
 −a21 (1 − a22 · · · −a2n   X2   d2 
   
 ·  ..  =  ..  ,
 
 .. .. ..
 . . .   .   .
−an1 −an2 · · · (1 − ann ) Xn dn

or (I − A) · X = d
In this equation I is an n × n identity matrix, A is the technical coefficient
matrix, X is the n-industry variable matrix, and d is the final demand matrix.
We frequently refer to the matrix (I − A) as a Leontief matrix. Using matrix
inversion, if I − A is nonsingular, then we can find (I − A)−1 . This means
that the unique solution for the X matrix is X = (I − A)−1 · d.
CHAPTER 10. LINEAR PROGRAMMING 333

An illustrative example. The table on the next page lists the sources
of inputs, and the destinations of outputs, in a hypothetical economy. We
could represent th this table’s contents as a system of 11 simultaneous linear
equations in 11 unknown values. Movements along any row show the output
of an industry and where that output goes. For example, Row 6 addresses
industry F. Two units its output go to industry A, six units to industry
B, and so forth. Column 7 reveals that two units of industry D’s output
constitute the accumulation of inventories in industry F itself. Column 12
shows that he total production of industry F is 46 units.
A movement down any column in this table lists the inputs that each industry
or sector receives from other industries or sectors. For example, column
5 indicates the inputs that industry E receives from other industries and
sectors. Thus industry E uses five units of industry A’s output, three units
of industry B’s output, five units of industry C’s output, and so forth.
The “processing sector” of an input-output table (rows 1 through 6 and
columns 1 through 6 contains all those industries that produce salable goods
and services, such as cars, furniture, and toothpaste. The processing sector
of most input-output tables is highly developed and may contain as many as
500 industries.
Columns 7 through 11 contain the “final demand” sector. For example,
household purchases of goods and services, in column 11, total 14 units from
industry A, 17 units from industry B, and so forth. Rows 7 through 11 contain
the “payments sector” of the table. This sector shows the contribution of
various owners of factor inputs (for example, households) to the production
of each output. For example, households provide 19 units of their inputs,
predominantly labor, to industry A, as recorded in row 11, column 1.
Compare this sector to the Keynesian aggregate expenditures equation Y =
C+I+G+X−M , where Y is total spending, C is private-sector consumption,
I is private-sector investment, G is government spending, X is exports, and
M is imports.
The table also records total gross outlays (in row 12) and Total Gross Output
(in column 12). The total gross outlay of inputs and the total gross output
of goods and services are not equivalent to gross Domestic Product, which
deliberately excludes intermediate outputs and inputs and concentrates only
on the value of final goods and services. In constrast, total gross outlay and
total gross input involve repeated double counting. This is not bad, however,
CHAPTER 10. LINEAR PROGRAMMING 334

because the purpose of input-output analysis is to illustrate the connections


of the economy, not to provide a measure of the value of total inputs used or
outputs produced.
CHAPTER 10. LINEAR PROGRAMMING 335

We can use this table to see how one computes the technical coefficients of
production that we discussed earlier. Each technical coefficient of production
should show the number of units of input j required to produce one unit of
output i. Thke next table consists of technical coefficients of production de-
rived from the input-output matrix in the preceding table. Consider industry
C: It receives a total of 40 units of inputs, one of which comes from the de-
pletion of its own inventories. Seven of these 40 units come from industry F.
Therefore the technical coefficient of production is 7/39 ≈ 0.18. This tells us
that every unit of output produced by industry C requires 0.18 unit of the
output of Industry F. (Note that we subtract inventory depletion from total
gross outlay before computing the technical coefficient of production.)

Inputs purchased from industries:


A B C D E F
Outputs A 0.16 0.26 0.03 0.05 0.13 0.13
pro- B 0.08 0.07 0.18 0.03 0.08 0.18
duced C 0.11 0.04 0.21 0.03 0.13 0.07
by D 0.17 0.02 0.05 0.21 0.15 0.09
indus- E 0.06 0.00 0.03 0.36 0.08 0.04
tries F 0.03 0.11 0.18 0.15 0.05 0.13

Technical coefficients of production can be useful to a researcher or forecaster


if they are based on up-to-date data that accurately portray the actual pro-
ductive processes being surveyed. For example, one can use technical coef-
ficients of production to determine the probable effects of a decrease in the
output of steel on the output of cars, on apartment construction, and even
on Christmas toys.
We can also trace the effects of public policies such as road building, in-
creased defense expenditures, changes in international trade policy and the
like by means of technical coefficients of production. The U. S. Department
of Commerce has long maintained a sophisticated input-output model to as-
sist it in predicting the consequences of a wide range of public and private
actions. Regional development authorities and larger corporations have also
made extensive use of input-output models.5
5
See Miller and Blair [14] for a full development of the technique and for an overview
of its applications. Also, a Web search of “input-output” analysis will yield a large number
of useful sites, including many studies by the International Monetary Fund.
CHAPTER 10. LINEAR PROGRAMMING 336

It is possible to derive input-output multipliers that show the total change in


output that will occur as a result of a change in the output of one industry.
According to the table above, industry D receives 0.36 units of output from
industry E whenever industry D expands its output by one unit. Suppose
that industry E’s output increases initially by one unit; this initially causes
industry D’s output to rise by 0.36 units.
Observe in turn that when industry D’s output rises by one unit, industry E’s
output rises by 0.15 unit. Thus a 0.36-unit increase in industry D’s output
has a feedback effect that increases industry E’s output by 0.36×(0.15 = 0.05
units. But this increase in industry E’s output once again requires additional
inputs from industry D, and so forth. The original expansion in industry
E’s output has set off a chain reaction of secondary increases in output.
This multiplier process, which is like the “national income multiplier” of
Keynesian models, can be quantified, so that we can isolate and analyze the
final, terminal effects of a given action.

10.3 Questions and Problems


1. Given the following linear-programming problem.
Maximize: 3 · X1 + 4 · X2
Subject to: 2 · X1 + X2 ≤ 12
3 · X1 + 2 · X2 ≤ 20
X1 , X2 ≥ 0

(a) What are the decision variables in the problem?


(b) Graph the problem, both by hand and using Maxima.
(c) What quantities of X1 and X2 maximize the objective function
subject to the two constraints? What is the value of the objective
function given these values?
(d) What if the nonnegativity constraint were removed? What dif-
ference would this make in the solution? Would such a solution
be sensible? What happens when you remove these constraints in
Maxima. How do interpret the error message?
(e) Define and solve the dual problem.
CHAPTER 10. LINEAR PROGRAMMING 337

2. Hashimi, Richmond, and Blaylock (HR&B) operate an accounting firm


that performs tax audits and also completes tax returns. The trio has
a good enough reputation that they can conduct as many audits and
returns as they choose. In a typical week, HR& B devote 115 hours of
time to performing audits and doing tax returns.
They have decided that, of these 115 hours, 75 are production hours
and 40 are review hours. Each time an audit is performed, 10 hours
of production time and 4 hours of review time are used. Each time
a tax return is completed, 3 hours of production time and 2 hours of
review time are used. An audit is priced at $1000, while a tax return
is priced at $400. HR&B wishes to maximize the revenues that they
receive from performing audits and completing tax returns.
(a) On average, how many audits and how many tax returns should
HR&B conduct? How many staff hours are devoted to audits, and
how many to tax returns? What is the average weekly income?
(b) Select any other feasible solution, and demonstrate that it gener-
ates less total revenue than the solution identified in (a).
(c) Define the dual to this maximization program and determine the
shadow prices of production and review.
(d) Based on (c) suppose that a retired accountant offers 10 hours of
service and that this accountant has the same skills as the HR&B
partners. If they hire the accountant, to which activity should
they assign her? (The partners will still provide production and
review hours as if the part-time accountant were not employed.)
(e) If they hire the accountant and use her as indicated in (d), by how
much does their weekly revenue change?
3. Tony and Jim open a lemonade stand on their front sidewalk. They
can make ordinary lemonade, or they can make a Lemon Fizz for their
customers. The ingredients per liter for each of these drinks are as
follows:

Lemonade Lemon Fizz

0.25 liters of sugar 0.75 liters of sugar


2 lemons 3 lemons
1 liter of water 1 liter of ginger ale
CHAPTER 10. LINEAR PROGRAMMING 338

Water is free and available in any quantity required. A total of 4 liters


of sugar, 25 lemons, and 5 liters of ginger ale are available. Tony
and Jim believe that they can sell each 0.5-liter glass of lemonade for
$0.75, while each 0.5-liter Lemon Fizz will sell for $1.25. They wish to
maximize the revenue that they realize from the sale of lemonade and
fizzes. How much of each drink should Tony and Jim sell? How much
revenue will they earn?
Appendix A

Additional Review Questions

These review questions are selected from a final example in Professor Os-
trosky’s course at Illinois State University, “Introduction to Mathematical
Economics,” which closely parallels the material presented in this book.

1. Assume that consumption is a linear function of income. The marginal


propensity to consume is 0.90. If income were zero, this function would
imply dissaving of $180. What is the explicit consumption function?
What is the multiplier for this problem if consumption is the only
component of aggregate expenditures (C + I + G + X - M) that is
affected by income?
2. Given the CES production function
Q = γ · [δ · K −ρ + (1 − δ) · L−ρ ]−ν/ρ
where Q = output, L = labor, K = capital, and γ, δ, ρ, andνare param-
eters. Determine (a) dQ and (b) the degree of homogeneity.
3. Given the demand curve q = a − b · P , where a, b > 0 are parameters,
what is the price elasticity of demand at each intercept? Prove your
answers using the formula for elasticity.
4. The production function for a firm’s product is Q = 12 · L + 20 · K −
L2 − 2 · K 2 . The per-unit cost to the firm of L and K is $4 and $8 per
unit, respectively.

a. Derive the equations for the marginal products of L and K, M P L


and M P K.

339
APPENDIX A. ADDITIONAL REVIEW QUESTIONS 340

b. Use the condition M P L/M P K = 20/40 to determine the optimal


relatiionship between the employment of these two resources.
c. Suppose that the firm wants the total cost of inputs to be $88.
Find the greatest output possible subject to this cost constraint,
using the Lagrangian multiplier approach. Confirm that the result
agrees with your solution in (b).

5. Assume that the demand per week for the NoFuzz Cable is 10,000
subscribers when the price is $60 per unit, and 20,000 subscribers when
the price is $40.

a. Determine the demand equation, assuming that it is linear.


b. What is the elasticity at a price of $60?
c. What advice would you give to the Cable Company based on the
information contained in this demand curve?

6. Blinko Company is the sole producer of artificial lightning bugs. Man-


agement has determined that the company’s total revenue function is
T R = 100 · Q − Q3 . Determine the point elasticity of demand for
artificial lightning bugs when Q = 5.

7. Assume the following national-income model:


Y =C +I +G
C = a0 + a1 · (Y − T ) a0 > 0 and 0 < a1 < 1
I = b0 + b1 · Y + b2 · i b0 > 0, b1 > 0, b2 > 0
G = G0 a constant
i = i0 the interest rate, a constant
T = t0 + t1 · Y t0 > 0, 0 < t1 < 1
a. List all parameters, endogenous variables, and exogenous variables.
b. Solve for the equilibrium income level using matrix algebra.

8. A firm’s total cost function is T C = x2 /4 + 3 · x + 400, where x is the


number of units produced. At what level of output will average cost
per unit be a minimum?

9. Find the extreme value(s) of W = −x3 + 3 · x · z + 2 · y − y 2 − 3 · z 2 .


APPENDIX A. ADDITIONAL REVIEW QUESTIONS 341

10. How much of goods x, y, and z should a person consume so as to


maximize utility, where the utility function is given by U = 10 · x · y · z,
and where the price of x, is $1, the price of y is $2, the price of z is $4,
and the available budget is $120?

11. Solve the following set of equations using inverse matrix algebra:
x1 + 2 · x3 + x4 = 4
x1 − x2 + 2 · x4 = 12
2 · x1 + x2 + x4 = 12
x1 + 2 · x2 + x3 + x4 = 12

12. Suppose that A and B are the only two firms in the market selling the
same product (we say that they are duopolists). The industry’s inverse
demand function for the product is P = 92 − qA − qB , where qA and qB
denote the output produced and sold by A and B, respectively. For A
the cost function is CA = 10 · qA and for B it is CB = qB2 /4. Suppose
that the firms enter into an agreement on output and price control by
jointly acting as a monopoly.

a. Express profit as a function of qA and qB , and determine how


they should allocate output so as to maximize the profit of the
monopoly.
b. Determine the price that they should charge and their joint profit
level.
b. Determine the price that they should charge and their joint profit
level.
c Confirm that, at the values of Q, qB , and qB that you have deter-
mined, the two firms’ marginal cost levels are the same.

13. Given the Cobb-Douglas production function Q = A · K α Lβ , where


A, α, β are parameters, show that the expansion path of a firm is equal
to a ray (K = c · L, where c is a constant), implying that the optimal
input ratio should be the same at all output levels. Remember that the
expansion path of a firm describes the least-cost combinations required
to produce varying levels of Q. Assume that the cost of production
is given by C = PK · K + PL · L, where PK and PL are constants (i.
e.,they are not affected by the firm’s employment levels of K and L.
APPENDIX A. ADDITIONAL REVIEW QUESTIONS 342

14. The Sweet-Tooth Candy Company produces two delectable varieties of


candy, A and B, for which the average costs of production are constant
at $7.00 and $8.00 per kilogram, respectively. The quantities qA and
qB (in kilograms) of A and B that can be sold each week are given by
the joint-demand functions

qA = 24 · (PB − PA ) and qB = 20 · (30 + PA − 2 · PB ),

where PA and PB are the selling prices (in dollars per kilogram) of A
and B, respectively. Determine the selling prices that will maximize
Sweet Tooth’s profit.

15. Suppose that a price-searching firm is practicing price discrimination


by selling the same product in two separate markets at different prices.
Let qA be the number of units sold in market A, where the demand
function is PA = f (qA ), and let qB be the number of units sold in
market B, where the demand function is PB = g(qB ). Assuming that all
units are produced at one plant and that transportation and marketing
costs are the same in both markets, let the cost function for producing
q = qA + qB units be C = C(q). [Hint: Keep in mind that total
revenue from market A is solely a function of qA , and total revenue
from market B is solely a function of qB .] Set up and determine only
the first-order conditions for the monopolist to maximize profits with
respect to outputs qA and qB . Interpret your results.

16. Given that    


1 −1 2 44 0 −3
A= ,B =
0 3 4 −1 −2 3
and    
2 −3 0 1 2
C = 5 −1 −4 2 , D = −1
  
−1 0 0 3 3
find A + B, 3 · A − 4 · B, A · D, D0 · D, D · D0 .
Show the following:
(A+B)’=A’+B’ and (A + B) · D = A · D + B · D.
APPENDIX A. ADDITIONAL REVIEW QUESTIONS 343

2
3 −1 2 1
0
1 −1 1 2
17. Evaluate the determinant 0 0 −1 2 3 .
0 0 0 1 4

0 0 0 2 5
 
1 1 1
18. Find the inverse for the matrix 0 1 1.
0 0 1
19. Solve the following equations using matrix algebra.
x1 + x2 + x3 = 0 , x1 + 3 · x3 = 1 and 2 · x2 + 2 · x1 = 0

20. Find the extreme value(s) of Z = 2 · x21 + x1 · x2 + 4 · x22 + x1 · x3 + x23 + 2.

21. Determine the values of x1 , x2 , and x3 that maximize or minimize the


function Z = x21 + x22 + 7 · x23 − 2 · x1 · x3 + 10 subject to the constraint
x1 + 2 · x2 + 3 · x3 = 0. Evaluate Z at this set of xi values.
Bibliography

[1] Bassi LJ (1976) The Diet Problem Revisited. American Economist, 20:
35–39.

[2] Bishop RL (1968) The Effects of Specific and Ad Valorem Taxes, Quar-
terly Journal of Economics, 82:198–218.

[3] Bradley SP, Hax AC and Magnanti TL (1977). Applied Math-


ematical Programming. Addison-Wesley, New York. Available at:
http://web.mit.edu/15.053/www/

[4] Chiang AC (2016). Fundamental Methods of Mathematical Economics,


3rd ed. McGraw-Hill, New York.

[5] Cyrenne P (2014). Salary Inequality, Team Success and the Super-
star Effect. Available at: ftp:// ftp.repec.org/opt/ReDIF/RePEc/win/
winwop/2014-02.pdf.

[6] Garicano L and Rossi-Hansberg L (2015). Knowledge-Based Hierar-


chies: Using Organizations to Understand the Economy. Available at:
https://www.princeton.edu/ erossi/KBH.pdf.

[7] Hammock MR and Mixon JW (2013). Microeconomic Theory and Com-


putation. Springer, New York.

[8] Jerison D. (2006)Single Variable Calculus. Massachusetts Institute


of Technology: MIT OpenCourseWare, https://ocw.mit.edu. License:
Creative Commons BY-NC-SA.

[9] Leydold J and M Petry (2011) Introduction to Maxima for Economics.


Available at http://statmath.wu.ac.at/ leydold/maxima/.

344
BIBLIOGRAPHY 345

[10] McAfee RP (2007). Introduction to Economic Analysis. Available at


http://www.mcafee.cc/Introecon/IEA2007.pdf.

[11] Mankiw GN (2016). Macroeconomics, 9th ed. Macmilan, New York.

[12] Meglicki Z (2001). Maxima, Maple, and Mathematica: the Summary.


http://beige.ucs.indiana.edu/ P573/node35.html.

[13] Miller DE (2012) Using wxMaxima for Basic Set Operations. Available
at http:// andrejv.github.io/wxmaxima/help.html.

[14] Miller RE, Blair PD (2009). Input-Output Analysis: Foundations and


Extensions. Cambridge, Cambridge University Press.

[15] Mitra-Kahn BH (2008) Debunking the Myths of Computable Gen-


eral Equilibrium Models. Available at http://econpapers.repec.org/
paper/epacepawp/2008-1.htm

[16] Ostrosky AL, Koch JV (1979). Introduction to Mathematical Eco-


nomics. Waveland Press Inc, Prospect Heights IL.

[17] Perlis S (2012). Theory of Matrices. Literary Licensing, Whitefish MT.

[18] Solow, RM (1956). A Contribution to the Theory of Economic Growth.


Quarterly Journal of Economics, 70: 65–94. doi:10.2307/1884513. JS-
TOR 1884513. Pdf.

[19] Stigler GJ (1945) The Cost of Subsistence, Journal of Farm Economics,


27:303–314.
Index

[optimization], 158 map (a function), 30


matrix
abscissa, 36 as a list of lists, 32
antiderivative, 207 max, 31
see integral, 208 min, 31
arguments in Maxima function, 37 smax, 31
average product function, 110 smin, 31
average revenue solve, 32
and the demand curve, 219 subst, 30
boundary conditions, 208, 211 wxdraw2d, 35
xaxis and yaxis, 35
capital accumulation xlabel and yabel, 35
as an integral, 250 xtics, 35
cardinal variables, 12 comparative statics, 194
Cartesian coordinates, 35 comparative statics analysis, 194
CAS (computer algebra system), 5 composite functions, 42
CES function, 80 computable general equilibrium
chain rule for differentiation, 103 models
Cobb-Douglas function, 80 and input-output analysis, 330
cofactor, 294 computer algebra systems, 5
command concavity, 162
descriptive, 31 constant, 14
elementp, 20 π = 3.14159 . . ., 14
commands as a coefficient, 14
apply, 31 e = 2.71828. . . , 14
descriptive (module), 31 parametric, 14
label, 35 constant of integration, 207
length, 31 consumer surplus
limit, 72 as an integral, 243
map, 29 continuity, 81, 82

346
INDEX 347

and periodic compounding, 84 quotient rule, 99


coordinate system rules, 96
quadrants, 36 discontinuity
units, 37 “ugly”, 85
coordinates infinite, 84
rectangular coordinates, 35 jump, 84
cost and production, 112 removable, 82
cost curves, 51 types, 82
critical root, 14 domain, 38
cross price-elasticity of demand, 135 of inverse functions, 62
cubic cost function, 113 domain and range, 39
cubic function dot multiplication, 269
derivatives of, 107 dot product, 269
dual, 144, 200
definite integral, 207, 223 duopoly, 341
demand, revenue, & elasticity, 116
dependent variables, 38 elasticity
derivative, 87, 96 and partial derivative, 135
and difference quotient, 88 elasticity of substitution, 150
chain rule, 103 equation, 14
continuity requirement, 93 critical root(s), 14
diff command, 90 equi-marginal condition, 201
geometric interpretation, 91 equilibrium, 15
inverse function, 105 Euler’s Theorem, 148
linear function, 88 exponential function
quadratic, 88 derivative, 99
trigonometric functions, 103 exponential functions, 53
determinants of matrices, 286 extreme value, 159
evaluation, 288 multiple extreme values, 161
diagonal matrix, 278
differential, 124 finding extreme values, 165
as a linear approximation, 124 two independent variables, 170
composite function, 125 first derivative test, 166
differentiation first fundamental theorem of calculus, 226
chain rule, 103 first-order conditions
higher-order derivatives, 106 multivariate, 301
multiple variables, 121 fraction
product rule, 99 as a rational number, 18
INDEX 348

function matrix
and relation, 38 use to build tables, 32
definition, 37 inferior goods, 135
dependent & indepdendent variables, 38 functions, 62
inferse
domain, 38 inflection bputs and concavity, 161
explicit, 38 inflection point
form, 43 and derivatives, 164
polynomial, 44 necessary and sufficient conditions, 164
functions inframarginal firm, 150
composite, 42 inital conditions, 208
inverse, 61, 62 initial conditions, 211
monotonic, 61 initital conditions
types, 34 see boundary conditions, 208
input-output analysis, 330
Hessian determinants, 302 applications, 335
Hessians multipliers, 336
bordered, 309 technical coefficients, 335
homethetic function, 146 integral
homogeneous function, 141 additive property, 212
idempotent matrix, 277 definite, 221, 223
identity matrix, 276 graphical representation, 208
Kronecker’s delta, 277 linearity property, 212
imperfect competition, 187 multiplicative property, 212
implicit function, 129 Riemann, 223
partial derivative, 131 integral calculus, 206
Implicit Function Theorem, 130 integrand, 207
implicit function theorem, 195 integration, 206
implicit functions, 39 and summation, 225
improper integral, 237 area between two curves, 232, 234
infinite discontinuity, 241 boundary conditions, 208
infinite integrand, 240 by parts, 217
infinite limit(s) of integration, 237 by substitution, 213, 226
income determination model, 16 constant of integration, 207
income elasticity of demand, 135 discontinuous function, 230
indefinite integral, 207, 209 general exponential rule, 211
independent variables, 38 general logarithmic rule, 210
index important properties, 213
indefinite integral, 209
INDEX 349

interchanging limits of integration, linear


230 programming, 4
inverse of differentiation, 207 corner points, 321
negative areas, 227 dual, 323
power rule, 210 dual and primal symmetry, 325
proper & improper integrals, 237 feasible region, 319
rules, 209 graphical representation, 318
sum of finite subintegrals, 230 nonnegativity constraints, 315
with Maxima, 214 objective function, 316
intergration primal, 323
definite integral, 207 shadow price, 324
inventory solution space, 319
square root rule, 198 structure, 317
inventory model, 197 list
inverse function describing, 31
derivative, 105 lists, 28
inverse functions, 61 logarithmic function
inverse of a matrix, 292 derivative, 96
computation of, 293 logarithms, 57
properties, 295 common, 58
iso-revenue line, 319 natural, 58
isocost line, 144 logistic functions, 55
isoquant, 144
mapping, 62
L’Hopital’s Rule, 154 marginal analysis
Lagrange multiplier, 178 and differentiation, 95
Lagrangian multiplier marginal product, 69
as marginal cost, 201 marginal product function, 110
Lagrangian multipliers, 307 marginal rate of technical substitution, 146
Leontief, Wassily, 331 matrices
limit, 71 and extreme values, 301
and rational expressions, 79 and OLS analysis, 298
direction of approach, 72 and vectors, 263
expression as a limit, 73 cofactors, 288
infinite limits, 77 determinants, 286
limits Hessians and extreme values, 302
and monotonic functions, 79 inverse, 292
linear algebra, 260 notation, 261
Linear programming, 313 solving systems of linear equations, 297
INDEX 350

systems of equations, 260 necessary conditions


transposed matrices, 281 multivariate, 301
matrix nonlinear programming, 327
addition & subtraction, 266 cobyla, 329
and vectors, 264 normal distribution
associative law, 268 integration, 248
cofactor, 294 normal goods, 135
commutative law, 268 normative analysis, 141, 158
compact notation, 261 null matrix, 279
definition, 261
diagonal, 278 OLS (ordinary least squares), 298
equality, 266 optimization, 144
idempotent, 277 1st order condition, 306
multiplication, 269 2nd order condition, 306
null, 279 first-order conditions, 178
postmultiplication, 269 function of a single variable, 159
premultiplication, 269 Lagrange multiplier, 178
scalar, 279 maximization subject to constraint(s), 176
scalar multiplication, 268 minimization subject to contraint(s), 176
square, 262 necessary condition, 166
matrix multiplication positive vs. normative view, 158
dimensional requirements, 271 price searching firms, 187
laws of, 274 production and cost, 200
Maxima CAS, 5 second order condition, 181
assignment of names, 16 sufficient condition, 168, 172
expressions, 16 two independent variables, 170
method of exhaustion ordered pairs, 27
and definite integral, 221 vs. unordered pairs, 27
model ordinal variables, 12
and variables, 13 ordinate, 36
income determination, 14 output elasticity, 154
models parametric function, 192
and theory, 11 parametric relationships, 39
monopsony, 149 partial derivative, 121
monotonic functions, 61 cross, 133
multiplication graphical representation, 123
homogeneity property, 65 higher order, 132
INDEX 351

implicit function, 131 range


Young’s theorem, 133 of inverse functions, 62
polynomial range and domain, 39
derivative, 97 real number line, 16, 17
polynomials, 44 irrational numbers, 18
degree, 44 rational numbers, 18
examples, 45 whole numbers, 18
fitting to points, 46 rectangular coordinates, 35
positive analysis, 158 Riemann integral, 223
present value root, 14
and integration, 254
price and total value, 247 saddle point, 171
price elasticity of demand scalar matrix, 279
and total revenue, 220 second derivate test, 168
producer surplus, 244 second order condition, 172
as an integral, 243 second-order conditions
product multivariate, 302
derivative of, 99 set algebra, 20, 26
production and cost, 112, 190 difference, 22
production function, 69, 144 intersection, 22
average product, 189 union, 21
Cobb Douglas, 110 sets
elasticity of substitution, 150 concept, 18
Euler’s Theorem, 148 elements, 20
homogeneous, 146 membership, 20
homothetic, 146 null set, 20
isoquants, 146 Maxima notation, 22
marginal product, 189 roster method, 19
output elasticity, 154 set-builder method, 19
returns to scale, 146 subsets, 21
profit maximization, 118 universal, 23
price-searching firms, 187 Venn diagram, 23
price-taking firms, 184 shadow price, 324
and Lagrangian multipler, 327
quadrants of graph, 36 Smith, Adam, 330
quadratic functions, 50 Solow growth model, 252
quotient solution value, 14
derivative, 99 square matrices, 277
INDEX 352

Stigler, diet problem, 2, 260, 313 representation, 12


updated, 5 vector
Stigler, George, 2 column, 263
sufficient conditions row , 263
multivariate, 302 Venn diagram, 23
summation
homogeneity property, 65 wxMaxima, 6
telescoping property, 65 Young’s Theorem, 133
tax incidence
competitive markets, 142
price searching market, 194
taxation
price searchers, 194
theory
and models, 11
and probabilities, 11
purpose, 11
total cost
as an integral, 219
total revenue
as an integral, 219
transpose of a matrix, 281
trapezoids
and numerical integration, 221

value marginal product, 69


value of marginal product, 202
variables
and models, 13
continuous vs. discrete, 12
definition, 12
dependent, 38
endogeneous vs. exogenous, 16
endogenous vs. exogenous, 13
independent, 38
ordinal vs. cardinal, 12
parameters, 16

View publication stats

You might also like