Introduction to

Applied Mathematics
Fourth Edition

Alan Parks
Lawrence University
Appleton, Wisconsin
 ♠ This book is the printed fourth edition of a text for the two-term
Applied Calculus course at Lawrence University. The original project
that grew into this text was begun during the summer of 2011.

Alan
c Parks. αλασ Publishing, Appleton, Wisconsin, 2018. The principal
copy of this edition was printed on December 27, 2018.
 Contents

Introduction v
Chapter 1. The Exponential Function 1
1. Exponentials and Logarithms 1
2. Exponential Models 7
Chapter 2. Recursion 15
1. Recursive Models 17
2. Investigating Recursive Models 22
Chapter 3. The Derivative 27
1. Discovering the Derivative 27
2. The Derivative at a Point 34
3. The Derivative of a Function 37
Chapter 4. Computing the Derivative 43
1. The Power Rule, Linearity 44
2. Products and Quotients 50
3. Exponentials and Logarithms 54
4. The Chain Rule 56
Chapter 5. Interpreting and Using the Derivative 63
1. Curve Sketching 65
2. Newton’s Method 72
3. The Chain Rule Revisited 75
4. Marginals 80
Chapter 6. Linear Optimization 85
1. Simple Examples 85
2. More Complicated Examples 88
3. Shadow Prices – Lagrange Multipliers 94
iii
iv CONTENTS

Chapter 7. The Integral 99

1. Antiderivatives 99
2. The Definite Integral 107
3. Riemann Sums 111
Chapter 8. Interpreting and Using the Integral 117
1. Anti-Rates 117
2. Area 120
3. Probability 128
4. Quantities in Economics 138
Chapter 9. Matrix Algebra 145
1. Matrix Arithmetic 145
2. Applications of Matrix Algebra 152
Chapter 10. Linear Equations 163
1. Equations and Solutions 163
2. Matrix Inverse 172
3. Applications of Linear Equations 176
Chapter 11. Partial Derivatives 189
1. Partial derivatives 189
2. Higher Order Derivatives 194
3. The Chain Rule 197
Chapter 12. Non-Linear Optimization 205
1. The First Derivative Test 205
2. Lagrange Multipliers 209
3. Fitting a Model to Data 218
Chapter 13. Spreadsheet Formulas 227
1. Function Values 227
2. Recursion Calculations 231
3. Matrix Calculations 233
4. The Solver 235
Bibliography 237
Index 239
 Introduction

Mathematics has suffered from its own success. Mathematical techniques

have yielded insights into virtually every area of quantitative work, and those
techniques have grown so far and so fast and for so long that it is hard to keep
track of them, let alone understand them comprehensively. Furthermore, the
theoretical underpinnings of the subject have attracted focussed, inventive,
and sustained thought over the past 2300 years, at least. Mathematics is
very broad and very deep. How do you introduce its connections to other
disciplines? We feel rather strongly that the answer involves introducing the
reasoning that produces those connections, and then by teaching the analytical
and numerical work that proffers insights into other disciplines.
Analytical work is that done by exact techniques in the calculus and linear
algebra. The calculus was developed between about 1660 and the late 1800’s;
it centers on the idea of the limit – the computation of quantities by successive
approximations. Linear algebra uses matrices to represent sequences and to
solve linear equations. Numerical calculations are those done on a computer or
calculator, where the numbers considered are, in general, approximate. These
approaches will be intertwined. Our overall goals are (1) to provide you with
analytical tools used both to express the main ideas in quantitative disciplines
and to make deductions from that theory; (2) to provide you with numerical
tools that will enable you to investigate applied problems directly and in detail.
We also have important secondary goals: to increase your facility with the
most common and important mathematical functions; to increase your ability
to read and understand analytic descriptions of variables and their relations;
to increase your confidence in computation.

v
vi INTRODUCTION

To be more specific about subject matter: during the first term (Applied
Calculus I), we will study recursive sequences, the derivative, and linear opti-
mization problems. The second term (Applied Calculus II) will involve inte-
gration, linear algebra, multivariate derivatives, and non-linear optimization
problems.
This text will be coordinated with class work and homework; above all,
this book is meant to be read carefully. Many students reading mathematics
seriously are surprised at how long it takes to wade through a short section
of text. Perhaps that’s because lines of reasoning that seemed clear when you
scanned them quickly dissolve when you try to reproduce or apply them. And
the ability to reproduce and apply what you have read is the principal test of
mathematical understanding.
There are many problems presented during the exposition: labeled with
the letter T followed by the chapter number, then a period, and then the
problem number. There is a separate problems document with a great number
of exercises from which class examples and homework will be chosen.
Even though this book is intended to be self-contained, you are welcome
to consult other books if you find them helpful. Our bibliography (on p.237)
was taken, for the most part, from Lawrence University’s Seeley G. Mudd
Library. The books [1], [2], and [3], for example, are widely used textbooks.
Indeed, there are an enormous number of texts for courses such as this one,
and the student interested in additional problems or alternative approaches
will have no trouble finding help from libraries and online. You might be
helped occaisionally by a textbook for the general calculus, such as [13].
We hope you will be drawn into applied mathematics in a way that tran-
scends the memorization of algorithmic calculation – we hope you will see
interesting corridors to other disciplines, and, more than anything else, that
you will gain confidence in recognizing applied problems and in addressing
them effectively. We have intended for the level of this material to increase
somewhat as the course moves along. We have observed that the reader’s
facility in reading mathematics increases correspondingly.
 CHAPTER 1

The Exponential Function

1. Exponentials and Logarithms

An exponential function 1 is a function such as 2x or 3x or (1/2)x , where the

variable x is in the exponent. It turns out that we can deal with all exponential
functions by understanding a particular exponential function, where the base
is the special number e. The number e is hard to define; here is a very famous
formula for it.
1 1 1 1 1
e=2+ + + + + + ···
2 3·2 4·3·2 5·4·3·2 6·5·4·3·2
The formula goes on forever, and you might recognize the denominators as
factorials. Anyway, we can use this formula to get a good approximation to
the number e.
e ≈ 2.718281828459 . . .
There is no obvious pattern in the decimals, and, it turns out, e is an irrational
number – that is to say it can’t be written as a ratio of integers. We almost
always just use the letter e to stand for the number and leave it like that –
the way we use π for the circle number. Calculators use an approximation like
the one we just gave.
The function ex is called the exponential function; it is the most important
function in both applied and theoretical mathematics. The definite article the
in the name the exponential function identifies the use of the base e. When
1Most italicized terms will be found in the Index at the end of the book. We have
indexed both terms and symbolic expressions.
1
2 1. THE EXPONENTIAL FUNCTION

another base is used, as in 2x , the function is an exponential function. We will

see that we can use the exponential function ex to understand all the other
exponential functions. There are very good reasons for using the base e, as
we will see later in the course. For now, we want to review properties of the
exponential function and its associated logarithm. Their graphs are given on
p.13.

Properties of ex
(1) ea is a positive real number for every real number a
(2) e0 = 1
(3) ea+b = ea · eb for all real numbers a, b
(4) (ea )b = ea·b for all real numbers a, b

The listed properties are a matter of algebra when a and b are rational
numbers – when a, b can be written as ratios of integers. We will use these
properties no matter what a, b are. Here is a consequence; make sure you see
which properties (1)-(4) are being used in each step.
1
ea · e−a = ea−a = e0 = 1 so that e−a =
ea
This shows that we can switch an exponential from the numerator to the
denominator by negating the exponent. This works for moving from the de-
nominator to the numerator as well:
1 1
−3
= e3 and = e−4
e e4
This identity also shows why ea can never be 0, for 0 has no reciprocal.
There is an alternate notation used for the exponential function; we write
x
e = exp(x). This notation allows us to keep the exponent on the same level
as the function; that can make some expressions more readable. For instance,
compare
2 3
e(x −2·x )/2 = exp (x2 − 2 · x3 )/2


 1. EXPONENTIALS AND LOGARITHMS 3

The exponential function is increasing: if a < b, then ea < eb . For instance,

if x > 0, then ex > e0 = 1. Also, if x < 0, then ex < e0 = 1.
Each exponential function has an associated logarithm in the same base.
The logarithm in base e is called the natural logarithm and denoted ln(x).
Thus,

(1.1) ea = b if and only if a = ln(b)

There are two other ways to say the same thing.

(1.2) exp(ln(b)) = b for all b > 0 and ln(exp(a)) = a for all a

We see that the natural logarithm and exponential function cancel each other
when they are in sequence. This means that they are function inverses of each
other. Function inverses are not algebraic inverses: ln 6= 1/ exp; rather, the
equations (1.1) define the meaning of the inverse.
The logarithm has properties analogous to the properties of the exponential
function. We might have time to discuss one or two of these in class.

Properties of ln(x)
(1) ln(b) is defined for every positive number b
(2) ln(1) = 0
(3) ln(b · c) = ln(b) + ln(c) for all positive numbers b, c
(4) ln(ba ) = a · ln(b) for every real number a and every positive number b
These properties are not as important as the properties of the exponential
function, but properties of ln(x) will come up in various problems during the
term.
Here is our first text problem. As we mentioned in the Introduction, they
are indicated by the capital T.

Problem T1.1. Show that the function ln(x) is increasing.

4 1. THE EXPONENTIAL FUNCTION

Solution. Suppose that a < b, and we need to show that ln(a) < ln(b). (Both
a, b are positive.) If not, then either ln(a) = ln(b) or ln(a) > ln(b).
If ln(a) = ln(b), then (1.2) shows that

a = exp(ln(a)) = exp(ln(b)) = b

contrary to the assumption that a < b. Similarly, if ln(b) < ln(a), then
remembering that the exponential function is increasing, we have

b = exp(ln(b)) < exp(ln(a)) = a

again, contrary to a < b. We conclude that ln(a) < ln(b), as claimed.

Here is a typical problem that involves the inverse relationship between the
exponential function and its logarithm.

Problem T1.2. Solve for x in the equation 3 · exp(2 · x − 5) = 9.

Solution. Since the x we want is in the exponent, we will use the logarithm
as in (1.1). In order to use those equations, we need to get the exponential
function all by itself.

3 · exp(2 · x − 5) = 9
exp(2 · x − 5) = 3 now use (1.1)
2 · x − 5 = ln(3)

Now it is a short step to x:

1  
2 · x − 5 = ln(3) so that 2 · x = ln(3) + 5 which is x = · ln(3) + 5
2

The answer we obtained is exact and, therefore, perfectly acceptable math-

ematically. However, we might prefer an answer that is numerical – an ap-
proximation obtained on a calculator or computer. A calculator gives

ln(3) ≈ 1.098612289
 1. EXPONENTIALS AND LOGARITHMS 5

and so
1  
x≈ · (1.098612289) + 5 = 3.049306144
2
Mathematicians are very careful about the difference between exact answers
and numerical answers. We insist on using the approximation sign x ≈ 3.049
whenever a numerical answer occurs. There is no general rule about the num-
ber of decimal places used for numerical answers; we will be fairly loose about
that.
Exact answers are almost always preferable to numerical answers. Some-
times we need to compare numbers to see which is larger, and for that purpose
numerical answers can be better, provided they are accurate enough. We will
do a fair amount of numerical work in the course, and so we will certainly get
our fill of approximations.
We mention an algebraic detail in the previous problem. The expression
2 · x − 5 contains both a product and a difference. By the precedence rules of
algebra, the product is done first; in other words
h i h i
2 · x − 5 = 2 · x − 5 NOT 2 · x − 5

The expression on the far right would be 2 · x − 10. This applies to addition
and division, as well; for instance,
h i h i 1
3 · x2 + 4 · x + 1/x = 3 · x2 + 4 · x +
x
If you are likely to make a mistake in this type of expression, use parenthesis [or
brackets] to clarify. If you are unsure of the meaning of the string of operations
on any text problem or homework problem, please ask!

Problem T1.3. For which x is e−x < 0.01?

Solution. A previous problem showed that the logarithm is increasing. Thus,

(1.2) shows that

−x = ln(exp(−x)) < ln(0.01) so that − x < ln(0.01)

6 1. THE EXPONENTIAL FUNCTION

Multiplying both sides of the last inequality by −1, and remembering to switch
the inequality, we get
x > − ln(0.01) ≈ 4.6052

Here are two more abstract problems. First, we claimed that all exponen-
tial functions can be written using the special exponential function ex . Proving
the following identity furnishes a nice exercise in using the properties of ex .

Problem T1.4. Let a > 0, and let b be an arbitrary number. Show that

(1.3) ab = exp(b · ln(a))

Solution. We use property (4) of the logarithm, followed by (1.2).

exp(b · ln(a)) = exp(ln(ab )) = ab

For instance, 2x = exp(x · ln(2)) and 10x = exp(x · ln(10)).

We can also relate an arbitrary logarithm to the natural logarithm ln(x).

Problem T1.5. Let a > 0 and let b be an arbitrary number. Show that
ln(b)
loga (b) =
ln(a)

Solution. The definition of the logarithm: if ac = b, then c = loga (b). So, we

prove the identity by showing that

aln(b)/ ln(a) = b

To do this, we use the previous problem:

 
ln(b)/ ln(a) ln(b)
a = exp · ln(a) = exp(ln(b)) = b
ln(a)
 2. EXPONENTIAL MODELS 7

2. Exponential Models

In applications of mathematics, the word model is used very generally

for any mathematical object that represents something else. In this section
we introduce some models that use the exponential function. These models
are fairly simple; they introduce some aspects of models in general and they
provide the opportunity to begin working with the exponential function.

Exponential Population Growth A population P of bacteria at time t

grows exponentially if P = P0 · exp(k · t) for some positive constant k. The
number P0 is the initial number of bacteria – the number of bacteria when
t = 0.

Problem T1.6. If we start with 100 bacteria, growing exponentially, and we

have 5000 in an hour, how long does it take for the population to double?
Show that the doubling time does not depend on P0 but only on k.

Solution. We will use the model to get equations to solve. We will omit some
algebraic steps that use the properties (1)-(4) given above for the exponential
function; you should see if you can do them yourself, and we will discuss the
algebra in class.
If P is the population of bacteria after t hours, then we have

P = P0 · exp(k · t)

for some constant k. The phrase “If we start with 100” tells us that P0 = 100.
When t = 1 hour we have

5000 = 100 · exp(k · 1) so that k ≈ 3.912

To calculate doubling time and to show that it depends only on k, we think

about the equation
2 · P0 = P0 · exp(k · t)
8 1. THE EXPONENTIAL FUNCTION

We have a specific value of k, and we can solve for t. First, divide by P0 .

(What if P0 = 0?) We get 2 = exp(k · t). This leads to
ln(2)
t= ≈ 0.177 hours
k
Notice that this value of t does not depend on P0 but only on k.
The exact value of k in the preceding problem is k = ln(50). Remember
to use the approximation sign to distinguish numerical answers! What is the
exact doubling time?
Another issue may have occurred to some readers. Typical values of
P0 · exp(kt) in the previous model are fractional; you might ask, “Don’t pop-
ulation values have to be whole numbers?” Good question. Remember that
a model represents something applied; the model is not the population! In
some contexts, fractional numbers may furnish decent approximations to whole
numbers; in some contexts fractional numbers are not allowed at all. We are
focusing more on the arithmetic of the exponential function at this point.
Radioactive Decay Radioactive substances lose mass over time. If Y mea-
sures the mass at time t, then we can model Y = Y0 · exp(−k · t) where k is a
positive constant.
Think about the following problem; we will discuss it in class.
Problem T1.7. Radium-226 is naturally radioactive. The constant k in the
model is approximately 4.3 · 10−3 , and t is time in years. How long does it take
for a mass of radium-226 to decay and lose 90% of its original mass?
Newton’s Law of Heating and Cooling We imagine placing an object
in an environment of constant temperature. We assume that the object does
not have any internal source of heat or cooling but is gradually warmed up or
cooled off by the environment. If F is the difference between the temperature
of the object and the temperature of the environment,2 then Newton’s Law of
2The temperature of the environment is also called the ambient temperature.
 2. EXPONENTIAL MODELS 9

Heating and Cooling says that F = F0 · exp(−k · t), where t is time and k is a
positive constant, and F0 is the initial difference in temperature.

Problem T1.8. Suppose that F = 30 at some value of t and F = 20 after 40

minutes have passed. When will F = 0.01?

Solution. We use 30 as our initial value of F , so that F0 = 30 and then

F = 30 · exp(−k · t). We have F = 20 when t = 40 minutes, and so

20 = 30 · exp(−40 · k)

This equation can be solved for k; we get k ≈ 0.01. We want to find t so that
F = 0.01. We’ll see that t ≈ 801 minutes.
In the problem we just did, the fact that we start with F = 30, a positive
number, tells us that the object is warmer than the environment, and so it
is cooling off. An object colder than the environment would have a negative
value of F .

Turning Off a Circuit We imagine a very simple electrical circuit consisting

of a resistor and capacitor in series.3 A circuit has a current I associated
with it; the current is the number of electrons passing through the circuit per
unit time, measured in amperes, usually abbreviated amps. In general, I is a
function of time t. When our simple circuit is turned off, we have4
 
t
I = I0 · exp −
R·C
where R, C are positive constants associated with the resistor and capacitor,
respectively.

3It is not necessary for you to understand anything technical about electrical circuits
here.
4It may surprise you that the current doesn’t just jump to 0. According to this model,
the electrons gradually slow down.
10 1. THE EXPONENTIAL FUNCTION

Problem T1.9. Suppose we know that R = 106 ohms, that I0 = 1 amp, and
that I = 0.01 amps when t = 0.05 seconds. What is C? (The units of C are
Farads.)

Solution. As before, we substitute what we know into the equation we are

given.
 
0.05 0.05
0.01 = 1 · exp − 6 so then ln(0.01) = − 6
10 · C 10 · C
We then get
0.05
C=− ≈ 1.086 · 10−8 Farads
106 · ln(0.01)

A vocabulary term: constants involved in a model are called parameters;

the k in exp(k · t), the initial values P0 , Y0 , etc., the R and C in the circuit
problem, all these are parameters. Much experimental science has as its object
to measure parameters.

The Cobb-Douglas Model

In the 1920’s Cobb and Douglas5 developed an equation meant to explain
the production level Y of a single good as a function of capital K (monetary
value) and labor L (in person-hours):

Y = a · K b · Lc

where a, b, c are positive constants.6 One of the main ways to deal with this
equation is to take the logarithm of both sides, using the identities on p.3.

ln(Y ) = ln(a) + b · ln(K) + c · ln(L)

5Cobb, C. W.; Douglas, P. H. (1928). ”A Theory of Production”. American Economic

Review 18 (Supplement): 139–165. See also [10].
6Technically: a is the total factor productivity, and b, c are output elasticities for capital
and labor, respectively.
 2. EXPONENTIAL MODELS 11

Problem T1.10. We have the following values for a Cobb-Douglas model,

where a = 1. Find b, c.
K L Y
100 25 57
80 30 54
Solution. We use the logarithmic equation. Since a = 1, its logarithm is 0.

ln(Y ) = b · ln(K) + c · ln(L)

ln(57) = b · ln(100) + c · ln(25)
ln(54) = b · ln(80) + c · ln(30)

Working with numerical values, we get

4.04 ≈ 4.61 · b + 3.22 · c

3.99 ≈ 4.38 · b + 3.40 · c

Solving for b, c, we obtain b ≈ 0.566 and c ≈ 0.445.

Problem T1.11. In a Cobb-Douglas model, when the capital doubles, the

production increases by 23%. Find the parameter b.

Solution. We can use the original equation Y = a · K b · Lc . For capital to

double, we replace K by 2 · K. Production increases by 23%; that means that
0.23 · Y is added to Y , resulting in

Y + 0.23 · Y = 1.23 · Y

Thus, we have

1.23 · Y = a · (2 · K)b · Lc = 2b · a · K b · Lc = 2b · Y

We see that 1.23 = 2b . Taking the logarithm of both sides, ln(1.23) = b · ln(2),
and
ln(1.23)
b= ≈ 0.299
ln(2)
12 1. THE EXPONENTIAL FUNCTION

A Limited Population
In Chapter 2 we will meet with the logistic equation that describes a pop-
ulation growing by reproduction, but limited by an external factor. Here is a
generic formula for such a population P where t is time and E, c are positive
constants.
E
(1.4) P =
1 − c · exp(−E · t)
Problem T1.12. Show that P gets close to E as t gets larger and larger.7
Solution. We know that E > 0, and so E · t > 0 when t > 0. As t gets
larger, the quantity −E · t is large and negative. The graph of the exponential
function shows that exp(−E · t) gets smaller and smaller as t gets larger. As
the exponential term gradually vanishes, the fromula (1.4) for P goes to E.

7Thisfact shows the role of the constant E in this model. It is an equilibrium – studied
in the next chapter.
 2. EXPONENTIAL MODELS 13

The Exponential Function.

...
....
....
.
.....
..
....
....
...
......
.
.....
.....
......
y .
.......
.
..
.
......
......
.
.......
.......
........
............
.............
.
...
..
..
..
..........
y = ex
(0, 1) r
.
..
..
..
..
..
..
..
..
..
..
..
..
.....................
...............
................
..
..
..
..
..
.
.................................................
..............................................................................
x

The Natural Logarithm.

......
......
......
......
y .
.....
.
..
......
....
....
....
..
....
.
....
....
....
..
.....
...
....
....
...
y = ln(x)
...
...
...
...
...
...
...
...
..
.
...
...
r
(1, 0) ..
..
... x
.
..
..
..
..
...
.
..
..
..
..
...
.
..
..
..
..
...
.
..
..
..
..
...
.
..
..
..
..
...
..
..
...
...
.....
..
...
...
...
.....
.
 CHAPTER 2

Recursion

Suppose we monitor the population of a bacteria culture each hour, starting

at 8:00 a.m. on a certain day. Let P0 be the population at 8:00 a.m., let P1
be the population at 9:00 a.m., let P2 be the population at 10:00 a.m., and so
on. The sequence that results

P0 , P1 , P2 , P3 , · · ·

is a model of the bacteria population. Many applications involve sequences

of numbers in which a given term in the sequence determines the next term.
In our bacteria population, it makes sense that the population at one time
determines the population at the next time. Abstractly: Pn determines Pn+1 ,
for each n. To be specific, let’s suppose that the population increases by a
factor of r each hour.1 Thus,

Pn+1 = Pn + r · Pn

Adding r · Pn accomplishes the increase. We can write this equation like this.

(2.1) Pn+1 = (1 + r) · Pn for n ≥ 0

We have much to say about equation (2.1); we begin by calling it a recursive

equation. It is a recursive model of the population. The constant r is a
parameter.
1This simple model is called the Malthusian model, and it is closely related to the
exponential model introduced in the previous chapter. See [4, p.2]. The entire first chapter
of [6] is devoted to models of population growth, beginning with the Malthusian model.
15
16 2. RECURSION

In a recursive model, there is a sequence of numbers, each one determined

by the ones before it. If we are given a starting number, then the rest of the
sequence unfolds from there. It is important to see that equation (2.1) is really
infinitely many equations in one. Indeed, the expression “for n ≥ 0” tells us
that the equation holds for n = 0, 1, 2, 3, . . .. The number r stays the same
hour to hour. Thus, (2.1) says that

P1 = (1 + r) · P0 (2.1) with n = 0
P2 = (1 + r) · P1 (2.1) with n = 1
P3 = (1 + r) · P2
P4 = (1 + r) · P3
..
.

These equations don’t tell us directly how to get the individual values, but
they tell us how to get from one value to the next. That’s recursion.
Next we observe that if we have a specific value of r and if we know the
initial population P0 , then (2.1) gives us the rest of the sequence of populations.
For example, let r = 1, so that (2.1) tells us that Pn+1 = 2 · Pn . Suppose that
P0 = 3, and just use the recursive equation over and over, as we did above:

P1 = 2 · P0 = 2 · 3 = 6
P2 = 2 · P1 = 2 · 6 = 12
P3 = 2 · P2 = 2 · 12 = 24
P4 = 2 · P3 = 2 · 24 = 48
..
.

We can take this as far as we wish.

Given that we’re doubling each time, you might know or have seen a for-
mula for Pn in general; the point we are making is that we can produce the
 1. RECURSIVE MODELS 17

sequence from (2.1) without a formula. There are many recursive sequences for
which no direct formula is available. And, as we mentioned above, recursion
is the natural way many sequences are given in the first place. So, we want to
be familiar and comfortable with recursion!
In the next section we give practical examples.

1. Recursive Models

We begin with an example that looks mathematically the same as Malthu-

sian growth.
Simple Interest We have an amount of money M invested, so that every
so often, interest is added in – we say compounded .2 The interest is some
fraction r of the money we have at that point in time. Suppose, for example,
that interest is added each month. Let Mn be the amount of money we have
after n months. Then to get Mn+1 , we add r · Mn .
(2.2) Mn+1 = Mn + r · Mn = (1 + r) · Mn
This parameter r is the interest rate, and it is usually expressed as a percent-
age.3 If we are given a specific interest rate and an initial amount of money
M0 , then we get the sequence M1 , M2 , . . . exactly as we got the population
sequence in the Malthusian model.
Repayment of a Loan We borrow an amount of money M0 and pay it
back gradually, incurring a finance charge as well. The finance charge is a
fraction of what is owed that is added to the debt. In almost all loans, the
money is paid back monthly, and so the natural time step is one month. Let
2See [10, Section 10.1] for an approach to interest that emphasizes a formula for Mn
rather than the recursive equation.
3No mystery here. Percent means divided by 100, so 7% = 0.07. Interest rates are
usually expressed as annual rates, and they need to be adjusted for the actual time step.
If the annual rate is 6% and the time step is one month, then r = 0.06/12 = 0.005, since
there are 12 months in a year.
18 2. RECURSION

r stand for the finance charge, and let q be the monthly payment, so that r, q
are parameters. Then
(2.3) Mn+1 = (1 + r) · Mn − q
The finance charge is added in, and then the monthly payment is applied to
decrease the amount owed. If we know the finance charge r, the monthly
payment q, and the initial amount M0 , we should be able to construct the
sequence M1 , M2 , . . .. We are probably interested in getting M to be 0, so
that the loan is payed off.
The Logistic Equation On p.12 we introduced the logistic equation 4 using
the exponential function. That model imagines time moving along continu-
ously – any real number is a valid time in the model. Now we will introduce
a discrete version of the model – a version in which time moves along step by
step. For instance, we might think of time in days and make one measurement
each day.
In 1838, Verhulst suggested the logistic equation for a population P that
changes over time due to two phenomena: reproductive growth at a rate k
and an external (environmental) limitation E. The numbers k, E are positive
constants – they are parameters. Here is the discrete version of the logistic
equation.
 
Pn
(2.4) Pn+1 = Pn + k · Pn · 1 −
E
The units of E are those of population. You can see that this recursive equation
is more complicated. Let’s notice a couple of features.
An equilibrium is a sequence that repeats the same value over and over.
We would have an equilibrium in the logistic equation if the sequence
P0 , P1 , P2 , P3 , . . . looks like P, P, P, P, . . .
4Sometimes the logistic equation is called Verhulst’s equation; it goes by other names
as well, having been rediscovered several times over the years. See [4, p.12]
 1. RECURSIVE MODELS 19

Problem T2.1. Find the equilibria for sequences produced by the logistic
equation.

Solution. We replace Pn and Pn+1 in the logistic equation by P – that’s saying

that all the Pn are the same number. When we start this, we don’t know, in
advance, that there are any equilibria, but if there are, the equation with the
P will find them. Here is that equation, and some algebra:
 
P
P =P +k·P · 1−
E
 
P
0=k·P · 1−
E
Since k > 0, we see that
P
P = 0 or 1 − =0
E
The second equation says that P = E.
Remember that the equation P = 0 indicates the sequence 0, 0, 0, . . .; it is
the sequence that is the equilibrium. We found two equilibria for the logistic
equation: the sequence 0, 0, 0, . . . and the sequence E, E, E, . . .. We can also
write Pn = 0 for the first sequence, remembering that n = 0, 1, 2, . . ., and we
write Pn = E for the second sequence. The first equilibrium makes sense: if
you start with no organisms, you certainly can’t get any by reproduction! We
will discuss the second equilibrium in class.
The general behavior of (2.4) is complicated and hard to deduce directly
from the recursion. We will construct specific sequences for this equation to
get a sense of the possibilities. The equation on p.12 hints that the exponential
function is relevant. That’s true but we won’t be in a position to see that until
much later in the course. For now, work with the discrete equation exclusively.
Speaking of equilibria, does the loan repayment model have any equilibria?
To find out, write M for Mn and Mn+1 in (2.3), and solve for M . Why is the
answer to that question interesting?
20 2. RECURSION

Having introduced equilibria, you might ask whether there was an equilib-
rium in the simple interest model. It is an important question regarding just
about any recursive model.

A Predator/prey model This time we have two population sequences: a

population Pn of prey and a population Qn of predators.5 We want to keep
track of both populations and their interaction, while keeping things relatively
simple. The sequences Pn and Qn will be parallel; we’ll be thinking of them
like this:
P0 P1 P2 P3 . . .
Q0 Q1 Q2 Q3 . . .
We assume there are positive constants (parameters) a, b, c, d such that

Pn+1 = (1 + a) · Pn − b · Pn · Qn
Qn+1 = (1 − c) · Qn + d · Pn · Qn

Let’s give a brief explanation of these equations; we will say more about them
in class. Look at the equation for Pn+1 . The (1+a)·Pn term looks like Malthu-
sian growth – reproductive growth. If Qn = 0 (if there are no predators), the
prey will increase in numbers exponentially. Apparently they have an abun-
dant food supply. The −b · Pn · Qn term in Pn+1 says that the prey will tend
to decrease the more predators there are. The product term Pn · Qn imagines
random meetings between prey and predator; the greater are the numbers Pn
and Qn , the more likely these species are to encounter each other. Bad for
prey.
Now look at the equation for Qn+1 ; the (1 − c) · Qn term says that the
population of predators will decrease if left to itself (if Pn = 0, so that there
are no prey). The term d · Pn · Qn represents a tendency for the predator
population to increase in the presence of their food supply.

5See [4, p.86ff].

 1. RECURSIVE MODELS 21

To get this model started, we need the initial number P0 of prey and the
initial number Q0 of predators. Then the equations take over, producing P1
and Q1 ; then producing P2 and Q2 , and so on.
This model has two equilibria; one of them has Pn = 0 = Qn , for all n.
That makes perfect sense; if you have no prey and no predators, then you will
never get any! You should find the other equilibrium. Can you explain that
equilibrium in ecological terms?
Let’s show that the exponential model and the recursive Malthusian model
are related.

Problem T2.2. In the exponential population model introduced on p.7:

P = P0 · exp(k · t)

let t measure years. Suppose that P0 6= 0. Consider the sequence Pn =

P0 · exp(k · n); that’s the population after n years. Show that Pn follows the
Malthusian model.

Solution. We need to show that the recursive equation of the Malthusian

model holds: that Pn+1 = (1 + r) · Pn , for some number r. In other words,
the ratio Pn+1 /Pn should be a constant (the constant 1 + r). Let’s look at the
ratio:
Pn+1 P0 · exp(k · (n + 1)) exp(k · (n + 1))
= =
Pn P0 · exp(k · n) exp(kn)
exp(kn) · exp(k)
= = exp(k)
exp(kn)
We see that r = exp(k) − 1, and the sequence satisfies the Malthusian model.

When k is small, exp(k) − 1 ≈ k, and so k and r are sometimes used

interchangeably. We emphasize that this only makes sense when k (or r) is
small. You might experiment on a calculator to see this.
22 2. RECURSION

2. Investigating Recursive Models

What might we want to know about a recursive sequence? We will use

both exact formulas and numerical investigation. Regarding the latter, you
will be asked to use Excel6 on a regular basis to compute various quantities.
The Spreadsheet Appendix contains lab exercises meant to introduce you to
Excel; we will discuss this in class.
We begin with a very simple question about the loan repayment model.

Problem T2.3. In the loan repayment model, suppose that $10000 is bor-
rowed, that the (annual) interest rate is 5%, and that the monthly payment is
$500. How long does it take to pay off the loan?

Solution. As above, let Mn be the amount of money left to pay off after n
months. Then M0 = 10000. We have the recursion Mn+1 = (1 + r) · Mn − q.
Since 5% is the annual interest rate, the monthly interest rate is r = 0.05/12.
The payment q = 500.
We computed Mn for n = 1, 2, . . ., looking for the place where Mn becomes
negative. Here are the first few terms; the decimal parts are cents – they have
been rounded.
n: 0 1 2 3 4 ...
Mn : 10000 9541.67 9081.42 8619.26 8155.18 . . .
As we computed the Mn , we encountered these entries.
n: 19 20 21
Mn : 957.26 461.25 −36.83
After 20 months, there is less owed than the monthly payment, even after the
finance charge is applied. So, in the 21st month, we pay off the loan.
Here is a new model; it measures the speed of a falling object.

6Throughout this book, Excel stands for Microsoft Excel.

 2. INVESTIGATING RECURSIVE MODELS 23

Problem T2.4. We record the speed of a falling stone at each second:

V0 , V1 , V2 , ...

and we suppose that Vn+1 = (1−k)·Vn +g, where k, g are positive parameters.7
The units of speed here are feet per second. Find an equilibrium. Now let
k = 0.1 and g = 20; starting with V0 = 0, does the sequence approach the
equilibrium?

Solution. To find an equilibrium, we write V for Vn and Vn+1 and obtain the
equation

(2.5) V = (1 − k) · V + g

We solve V = g/k feet per second. The sequence that repeats g/k over and
over is an equilibrium. (The ratio g/k is called the terminal speed of the falling
object.)
Now let k = 0.1 and g = 20, so that the equilibrium is 20/0.1 = 200 feet
per second. Letting V0 = 0, we used Excel to calculate the sequence. Here are
some values we obtained.
n : 0 1 2 ··· 70 71 72
Vn : 0 20 38 · · · 199.87 199.89 199.90
The sequence increases slowly; V100 ≈ 199.99. It looks like the sequence is
approaching equilibrium.

The Babylonian Sequence Let c > 0, and define a recursion

 
1 c
an+1 = · an + for n = 0, 1, 2, . . .
2 an
Heron of Alexandra8 attributed this formula to ancient Babylonian mathe-
maticians. To find out what it is for, we experiment, starting with a0 = c.
7The number k measures the air resistance on the stone; the number g measures the
pull of gravity.
8See [14, p.423].
24 2. RECURSION

When c = 1, compute that the sequence is in equilibrium 1, 1, 1, . . .. When

c = 4, here are numerical values of the first few terms.

4, , 2.5, 2.05, 2.00061, 2.00000009

It looks like the sequence is approaching the number 2. Is that an equilibrium

when c = 4? We take c = 4 and a0 = 2, then we get 2, 2, 2, . . ..
√
Problem T2.5. Show that c is an equlibrium for the Babylonian Sequence.
√ √
Solution. We let an = c, and show that an+1 = c:
√ 2!
√ √
 
1 c 1 c
an+1 = · c+ √ = · c+ √
2 c 2 c
1 √ √
1 √ √
= · c+ c = ·2· c= c
2 2

The Babylonian Sequence is used to approximate square roots, and it gets

√
a good deal of accuracy very quickly. You should try approximating 2 for
yourself: use c = 2 = a0 .

A Markov Process We begin with a specific, fanciful example. Suppose

that we have exactly two types of weather: sunny and rainy. If today is sunny,
it is 70% likely to be sunny tomorrow and 30% likely to be rainy. If today is
rainy, it is 55% likely to be sunny tomorrow and 45% likely to be rainy.
Suppose that on day n, the probability of sunny is Sn , and the probability
of rainy is Rn . We want to calculate the probability of sun and rain for the
next day. The probabilities will be denoted Sn+1 and Rn+1 . To calculate Sn+1 ,
we consider the possible weather the day before (day n). First, it could have
been sunny and then it went from sunny to sunny. The probability of sun on
day n was Sn and the probability of sun on day n and sun on the next day
is 70% (from the previous paragraph). The elementary theory of probability
tells us that the probability of the chain of events is a product, so that the
 2. INVESTIGATING RECURSIVE MODELS 25

probability of being sunny on day n and then going to sunny on day n + 1 is

Sn · 0.7.
The other way we could be sunny on day n + 1 is to be rainy on day n and
then go from rain to sun: Rn · 0.55. Since these are the two ways we get sunny
on day n + 1, we have

Sn+1 = 0.7 · Sn + 0.55 · Rn

We add the probabilities here, since the two events (sunny day n, rainy day
n) do not overlap.
We calculate Rn+1 in a similar way. See if you can do it, and we’ll discuss
it in class.
Rn+1 = 0.3 · Sn + 0.45 · Rn
As in the predator/prey model, we have two recursive sequences that are re-
lated to each other. You will experiment with the behavior of these sequences.
A Markov process involves a sequence of time steps; at each step there are
a finite number of possibilities, exactly one of which must happen at that step.
The possibilities are usually called states. The states in the weather example
are sunny and rainy. For each states S, T , there is a definite probability asso-
ciated with having state S at time step t and state T at time step t + 1. We
will see several examples of Markov processes.

The Fibonacci Numbers In 1202, Fibonacci9 became interested in an ide-

alized problem involving populations of rabbits. Here is what he assumed.
(1) we always have many available males to mate
(2) immature female rabbits take a month to become mature enough to
reproduce

9This problem is a typical counting problem that involves recursion. Books on com-
binatorics, such as [16] contain many such problems. See [15, p.214ff] for information on
Fibonnaci.
26 2. RECURSION

(3) mature female rabbits gestate each month, become pregnant, and
produce one male and one female at the end of the month
Fibonacci was able to find a recursion for the number Fn of female rabbits at
the end of month n:
(2.6) Fn+2 = Fn+1 + Fn for n = 0, 1, 2, . . .
This is a different type of recursion than what we have considered so far. In the
previous examples, you computed a sequence term from the term just previous
to it. In (2.6) you need two previous terms.
Can you show that (2.6) is correct? You might want to consider the number
Mn of mature females at the end of month n, and the number In is immature
females. Then Fn = Mn + In . Fibonacci’s assumptions show that
Mn+1 = Mn + In
In+1 = Mn
We will complete the details in class.
If you assume you have one immature female to start with, so that M0 = 0
and I0 = 1, then you get F0 = 1 and F1 = 1, and then the recursion (2.6) takes
over. This specific example gives what are called the Fibonacci numbers; they
have a remarkable number of interesting properties and they arise in a host of
counting problems.
We have tried to keep to fairly simple models in this chapter, while pushing
in the direction of the complexity found in more realistic applications. The
reader might glance at Chapter 6 of [6], regarding gas exchange in the human
lung, to see a typical, much more complicated, example.
 CHAPTER 3

The Derivative

This chapter begins our analytic work.

1. Discovering the Derivative

Calculus is about variation; it attempts to relate the change in one variable

to the change in others. Here is a standard notation for change: if the variable
x starts at the value 3 and ends up at the value 5, then we write ∆x = 5−3 = 2
to indicate the change in x. The capital delta ∆ stands for change. In general,
change is computed as

∆ = (final value) − (starting value)

The expression ∆x is not a product; it is one symbol (a word spelled with two
letters). The value of ∆x depends on two values of x; one value designated as
starting and other as final.
To work another example, suppose the variable z starts at 7 and ends up
at 3. Then
∆z = 3 − 7 = −4

Note that the change is final minus start regardless of the size of either number.
In this case, ∆z was negative – that means that the variable z decreased. In
the first example ∆x = 2 was positive, indicating that x increased.
Notice that if x is the starting value and the change is ∆x, then the ending
value is x + ∆x.
27
28 3. THE DERIVATIVE

Later we will consider the situation of an object moving along a line – say
that the line is the x-axis. The value of x at some point in time is the object’s
position. The quantity ∆x, the change in position, is called the displacement.
If we undergo a displacement of −2, we mean that we move 2 units in the
negative direction on the x-axis; that would be to the left.
The delta notation will arise frequently with two variables related by a
function. Say y = f (x). Starting with a value of x, we have the corresponding
value of y = f (x). Now say we undergo a change ∆x in x, the ending value of
x is x + ∆x, and the y-value that goes with this is f (x + ∆x). So, the starting
value of y is f (x) and the ending value of y is f (x + ∆x), and
(3.1) ∆y = f (x + ∆x) − f (x)
This equation is important; it relates the change in a variable y to values of a
function f (x) that computes that variable y.
Now we move toward this chapter’s main topic. Let’s consider two ap-
plication problems: finding the velocity of a moving object and finding the
slope of a graph. We will see that these problems are, in some sense, the same
problem; when we abstract out what the two problems have in common, we
will discover the derivative. This is very typical mathematics – a general idea
is discovered as a common feature of several concrete problems. An additional
application will be discussed in class.
Problem: Velocity.
We are interested in the discovery of a mathematical idea; it is interesting
that the approach of a particular physics text is identical. See [9, p.23ff].
We begin with some definitions that may be familiar to you – we want to
make sure everyone is on the same page. We imagine traveling along a straight
line – let’s say we are traveling along the x-axis. The speed of a moving body
measures the ratio of distance traveled to the time elapsed. A constant speed
of 30 miles per hour (MPH) predicts that we travel 30 miles each hour elapsed.
 1. DISCOVERING THE DERIVATIVE 29

The velocity of a moving body is the speed along with a plus or minus sign
that indicates the direction of the motion: positive velocity means that we are
traveling in the positive direction on the x-axis (to the right); negative velocity
means that we are traveling to the left. So, if the velocity is -10 MPH, then
the speed is 10 MPH and we are moving to the left. Notice that speed and
velocity have the same units.
If the speed is constant, then we can calculate it by dividing distance
traveled by time elapsed. Say we start at x = 3 at time t = 5, and we end
up at x = −10 when t = 7. Then we have traveled a distance of 3 + 10 = 13
miles in 2 hours, and so our speed was 13/2 = 6.5 MPH. The velocity, on
the other hand, would be the displacement divided by the time elapsed, for
displacement takes direction into account. The displacement in this example
is
∆x = (ending x) − (starting x) = (−10) − (3) = −13
and so the velocity is −13/2 = −6.5 MPH. Notice that the denominator 2 in
the case of both speed and velocity is the change in time ∆t. In particular,
notice that the ratio x/t is not relevant to either quantity: for instance, we
had x = 3 when t = 7; the ratio 3/7 doesn’t tell us about speed.
Constant velocity is very unusual. For instance, motion often starts from
a standstill, speeding up and slowing down along the way. It is the business
of calculus to explain how to calculate velocity when it is not constant. Let’s
think about that.
Suppose that the x axis measures miles, and time t is in hours. Suppose
we want to know the velocity when t = 3. Let’s assume we are at x = 10
when t = 3, and we end up at x = −80 at t = 5. The displacement is
∆x = −80 − 10 = −90 miles, and the elapsed time is ∆t = 5 − 3 = 2 hours.
Their ratio is
∆x −90
= = −45 miles per hour
∆t 2
30 3. THE DERIVATIVE

What does the number −45 tell us? It would be too much to expect that it
tells us the velocity at t = 3, for the velocity may have varied quite a bit during
the two hours elapsed. The ratio gives what is called an average velocity. An
average is meant to give one number that smooths out variation along with
way, but it doesn’t tell us much about what was happening exactly at t = 3.
Instead of measuring over an elapsed time of 2 hours, suppose we use
∆t = 0.25 hours; say we start at t = 3 and end at t = 3.25, and suppose
that the displacement is ∆x = −12 miles over this change in time. Then the
average velocity is
∆x −12 miles
= = −48 miles per hour
∆t 0.25 hours
The time interval 0.25 hours is much smaller than 2 hours, and so we expect
that −48 miles per hour is closer to the actual velocity at t = 3. Nonetheless,
there could be some variation in 0.25 hours, and so -48 miles per hour is still
an average velocity, and we still don’t know our exact velocity.
Now suppose we have ∆t = 0.01 hours, and the displacement is ∆x =
−0.47 miles. This time, the average velocity is
∆x −0.47 miles
= = −47 miles per hour
∆t 0.01 hours
Because we are measuring over one hundredth of an hour we expect even less
variation in velocity than with the previous 0.25 hours, so, this average velocity
is probably closer to the actual velocity at t = 3.
It looks like this: to get the average velocity to be close to the actual
velocity, use a very small elapsed time, so that the velocity can’t vary too
much. As the time interval gets smaller and smaller, we expect the average
velocity to be close to the actual velocity. The average velocity is
∆x
∆t
 1. DISCOVERING THE DERIVATIVE 31

and here is the technical notation for letting the time interval ∆t get very
small.
∆x
(3.2) velocity = lim
∆t→0 ∆t

The notation ∆t → 0 is meant to convey that ∆t (the elapsed time) gets

smaller and smaller; we consider the average velocity ratio for each possible
value of ∆t, and see what happens. The notation lim stands for limit, and we
read lim∆t→0 as “the limit as ∆t goes to 0.” Limits can be studied for their
own sake; in this course they will occur in a fairly restricted set of instances.
Here is a very specific example. Suppose that x = t2 , and say we want the
velocity when t = 3. The position at t = 3 is x = 32 = 9. The equation (3.2)
asks for ∆x and ∆t. If t starts at 3 and undergoes a change ∆t, then it ends
up at 3 + ∆t, and x then ends up at (3 + ∆t)2 . Thus,
∆x = ending − starting = (3 + ∆t)2 − 9
We indulge in some algebra on the average velocity. At the start, we use the
standard formula
(3.3) (a + b)2 = a2 + 2 · a · b + b2
Here goes.
∆x (3 + ∆t)2 − 9 9 + 6 · ∆t + (∆t)2 − 9
= =
∆t ∆t ∆t
6 · ∆t + (∆t)2 ∆t · (6 + ∆t)
= = cancel ∆t
∆t ∆t
= 6 + ∆t
We see that the average velocity is actually quite simple in this case. When
∆t is very close to 0, the expression 6 + ∆t is close to 6. This shows that the
exact velocity at t = 3 is 6.
32 3. THE DERIVATIVE

Problem: Slope
We are given a curve y = f (x) in the plane, and we consider the problem
of finding the slope at some point. The slope is, by definition, the slope of a
tangent line to the curve. If we are given a sketch of the curve, we can probably
determine a tangent line by eye; what we want is an exact, analytic way of
finding the slope of that line. Idea: we can find the slope of a line through two
points on the curve – such a line is a secant line. If the two points are very
close together, their secant line is approximately a tangent line.
Suppose we are interested in the slope at some point (x0 , y0 ) on the curve
y = f (x). We will imagine a secant line through (x0 , y0 ) and some other point
on the curve. We can imagine this other point as resulting from a change of
∆x in x, and then the ratio ∆y/∆x gives the slope of the secant line.
...
......
y = f (x)
......
......
r .
......
.
.......
.
 ......
 ......
secant  
.
.......
.
..
..
......
......
.......
......
∆y

.......
.......
 .
..
 .
...
.
........
 .......
........
.........
......
......  .
............
.
..
..
..
...........

r 
.......
....... ...............
.....................................

(x0 , y0 ) ∆x

If the two points at the ends of the secant are very close to each other – if
∆x is small – it makes sense that the secant line would be close to the tangent.
In other words,
∆y
tangent slope = lim
∆x→0 ∆x
Let’s work a specific example. Let’s find the slope to y = 1 − x2 at x = 2.
The point on the curve where x = 2 is (2, −3). That’s our starting point
for secant lines. Stepping ∆x, we have x-coordinate 2 + ∆x, and then the
y-coordinate is 1 − (2 + ∆x)2 . And so, again using (3.3), we compute
 1. DISCOVERING THE DERIVATIVE 33

∆y = 1 − (2 + ∆x)2 − (−3) = 1 − 4 + 4 · ∆x + (∆x)2 + 3

 

= 1 − 4 − 4 · ∆x − (∆x)2 + 3 = −4 · ∆x − (∆x)2

Then the slope of the secant line is

∆y −4 · ∆x − (∆x)2 ∆x · (−4 − ∆x)
secant slope = = = = −4 − ∆x
∆x ∆x ∆x
When ∆x is small, this expression is close to −4. The slope of the tangent to
y = 1 − x2 at the point (2, −3) is −4.
There are many other examples that can be given; each one constitutes an
application of the derivative. Here is one that we will discuss in class.

Problem: Density
Assume we have a block of wood lying along the interval 0 ≤ x ≤ 2 on
the x-axis, where the x measures yards. (So the wood is 2 yards long.) We
want to think about the weight of the wood. Notice that it doesn’t make
sense to talk about the weight at a point along the block, since a point is not
three-dimensional, and so the weight at a point is 0 (ideally). We measure
how weight accumulates as we move from left to right over the log. For each
x with 0 ≤ x ≤ 2, let W (x) be the weight in pounds of the wood from the
left-hand end up to the point x. In other words, W (x) is the weight of x yards
of wood – the left-hand x yards. We will see that this function W (x) can be
used to calculate the density of the wood – the number of pounds per yard we
have at any given point.1

1The density can vary over the block, either because the thickness of the block varies, or
because the characteristics of the wood vary. This approach to density – using the function
W that measures the weight at various points is consistent with the ideas of Isaac Newton
that led to the discovery of the general calculus. See [15]. Also, we mention that the word
density means different things in different contexts; we will point this out as we encounter
examples.
34 3. THE DERIVATIVE

At a point x = 1 on the block, we contemplate a change ∆x. We think of

∆x as giving us a small part of the block. Here is a picture, imagining that
∆x > 0.

0 1 1 + ∆x

By definition W (1) is the weight of the block up to the point x = 1, and

W (1 + ∆x) is the weight up to x = 1 + ∆x. Thus,

∆W = W (1 + ∆x) − W (1)

is the weight in pounds of the block between x = 1 and x = 1 + ∆x. The ratio
∆W
∆x
has units pounds per yard, an approximation of the density at x. Why approx-
imation? Because the thickness or shape could vary between 1 and 1 + ∆x. In
fact, the ratio ∆W/∆x is the average density on that piece of wood. Letting
∆x get smaller and smaller, we expect to get closer to the actual density at x.
You might try the following problem.

Problem T3.1. Let W (x) = 2 + x3 , and calculate the density at x = 1.

2. The Derivative at a Point

In our calculation of velocity the independent variable was time t, whereas

in the calculation of slope the independent variable was x. However, the
calculations were similar in that they involved a ratio of deltas. In both cases,
we computed something like this:
∆y
(3.4) lim
∆x→0 ∆x
 2. THE DERIVATIVE AT A POINT 35

where the steps ∆x were taken from a specific point (x0 , y0 ) on the curve
y = f (x), so that y0 = f (x0 ). The limit (3.4) is called the derivative of f (x)
at x = x0 .
Problem T3.2. Compute the derivative of 1/x at x = 4.
Solution. At the point x = 4 on the curve y = 1/x, we have y = 1/4. Thus,
x starts at 4 and y starts at 1/4. The definition (3.4) of the derivative wants
∆y:
1 1 4 − (4 + ∆x) −∆x
∆y = − = =
4 + ∆x 4 (4 + ∆x) · 4 4 · (4 + ∆x)
And so
∆y −∆x 1 −1
= · =
∆x 4 · (4 + ∆x) ∆x 4 · (4 + ∆x)
We see that when ∆x is small, the ratio is close to −1/(4 · 4). Thus, the slope
of y = 1/x at x = 4, is −1/16. Or, if y = 1/x describes an object traveling on
the y-axis, where x is time, then we just computed the velocity of the object
at time x = 4.
Problem T3.3. Compute the derivative of 3 · x − 7 at x = −2.
Solution. On y = 3 · x − 7 and x = −2, we have y = −13. Then check that
∆y = 3 · (−2 + ∆x) − 7 − (−13) = 3 · ∆x
and so
∆y 3 · ∆x
= =3
∆x ∆x
The resulting expression 3 stays at that value no matter what ∆x does. Thus,
the derivative is 3.
Let’s interpret the derivative we just calculated. That derivative is the
slope of y = 3 · x − 7; that curve is a straight line of slope 3, so it’s no wonder
that the derivative is 3.
There are many other examples. We may have time to mention some of
them in class.
36 3. THE DERIVATIVE

(1) Let v(t) be the velocity at time t of an object moving along an axis.
Then the derivative of v at t = t0 is the acceleration a(t) of the object
at time t = t0 . The units of acceleration are distance/(time)2 .
(2) If W (x) is the work done in moving an object along the x-axis to the
point x, then the derivative of W at x = x0 is the force applied at the
point x = x0 to produce that motion.
(3) If C(x) measures the cost of quantity x of some good, then the deriv-
ative of C at x = x0 is the cost per unit of x, at x = x0 . (The cost
per unit is the price; price can vary, depending on how much we buy.)
You might wonder about calculating the derivative numerically.
Problem T3.4. Approximate the derivative of ex at x = 2.
Solution. Let y = ex . To compute ∆y, we use 2 as our starting value and
then 2 + ∆x will be our ending value. Then
∆y exp(2 + ∆x) − exp(2)
∆y = exp(2 + ∆x) − exp(2) and =
∆x ∆x
We used Excel to compute the ratio for various small values of ∆x:
x: 0.1 0.05 0.01 0.005 0.001 0.0001 0.00001
∆y/∆x: 7.7711 7.5769 7.4261 7.4076 7.3928 7.3894 7.3891
It looks as if the derivative is 7.389, or so.
In the next chapter we will be able to find the exact value of the deriva-
tive of ex ; numerical calculation raises the question why we shouldn’t prefer
that method to the algebraic method we used to this point. There are two
important reasons to prefer the algebraic method: (1) in applications of the
derivative we will want to understand the behavior of the derivative over a
large set of points, all at once. Algebraic formulas disclose that behavior much
more readily than tables of values. (2) Many of the (practical!) uses of the de-
rivative involve purely symbolic quantities – there will be no specific numerical
values to compute. Having said that, we do admit that numerical calculations
are often useful, and they are necessary in situations where we do not have
 3. THE DERIVATIVE OF A FUNCTION 37

a formula for the function involved. We close by saying that the numerical
calculation of derivatives, in general, is a hard problem with many nuances.
We won’t take the time to introduce that subject in any formal way. Still,
there are a couple of problems at the end of this chapter that invite numerical
calculation.

3. The Derivative of a Function

We have been calculating the derivative at a specific point. Given a func-

tion such as x2 , it turns out to be efficient to calculate the derivative at all
points simultaneously. To do that, we imagine the curve y = x2 at a generic
point (x, x2 ). We start at x and move to x + ∆x. We can compute ∆y:

∆y = (x + ∆x)2 − x2

Expanding out:

∆y = x2 + 2 · x · ∆x + (∆x)2 − x2
= 2 · x · ∆x + (∆x)2

And so
∆y 2 · x · ∆x + (∆x)2
= = 2 · x + ∆x
∆x ∆x
In the expression 2 · x + ∆x, the x indicates s a specific point on the curve – we
just don’t want to make which point. And the ∆x is supposed to get small.
When that happens, the expression 2 · x + ∆x gets close to 2 · x – that’s the
value of the derivative of x2 , in general.

Problem T3.5. Find the slope of y = x2 at x = 5.

Solution. The slope of the tangent line is the derivative of x2 at x = 5. We

calculated the derivative 2 · x for all x, and so the derivative at x = 5 is simply
2 · 5 = 10. Thus, 10 is the slope of the tangent line at x = 5.
38 3. THE DERIVATIVE

Here is one of the notations we will use to designate the general derivative:

(x2 )0 = 2 · x

In this notation, the prime mark indicates the derivative.

Let’s see what the derivative looks like in general. Say we have a function
f (x). If y = f (x), and if we step from x to x + ∆x, then (3.1) (back on p.28)
said that
∆y = f (x + ∆x) − f (x)
And so,
∆y f (x + ∆x) − f (x)
=
∆x ∆x
To get the derivative, we need to let ∆x → 0, as in (3.4). We can write the
result in two ways:
f (x + ∆x) − f (x)
(3.5) f 0 (x) = lim
∆x→0 ∆x
or
∆y
(3.6) f 0 (x) = lim
∆x→0 ∆x

Problem T3.6. Find (1/x)0 .

Solution. With f (x) = 1/x, we imitate (3.5). To do this, it might help to

write f (x) with a placeholder instead of the x.
1 1
f () = so that f (x + ∆x) =
 x + ∆x
And (3.5) looks like this.
f (x + ∆x) − f (x) 1/(x + ∆x) − 1/x
=
∆x ∆x
To work with this, we first write the denominator as its own fraction:
 
1/(x + ∆x) − 1/x 1 1 1
= − ·
∆x x + ∆x x ∆x
 3. THE DERIVATIVE OF A FUNCTION 39

Now we continue.
 
f (x + ∆x) − f (x) 1 1 1
= − · get a common denominator
∆x x + ∆x x ∆x
x − (x + ∆x) 1
= ·
(x + ∆x) · x ∆x
−∆x 1
= · cancel ∆x
(x + ∆x) · x ∆x
−1
=
(x + ∆x) · x
Let’s not lose track of this expression; it’s nothing more nor less than ∆y/∆x.
As ∆x gets close to 0, the expression looks like −1/x2 .
 0
1 1
=− 2
x x

Here is a case where we can predict the answer before we compute it.
Problem T3.7. Compute the derivative of m · x + b, where m, b are constants.
Solution. This time we will use (3.6). Let y = m · x + b is a line of slope m.
Since the derivative computes slope, we expect that
(m · x + b)0 = m
Let’s check. We have
   
∆y = m · (x + ∆x) + b − m · x + b
= m · x + m · ∆x + b − m · x − b
= m · ∆x
And then
∆y m · ∆x
= =m
∆x ∆x
This secant slope is constant. (That’s because a secant line to a line is just
the line itself!) And so the derivative is m.
40 3. THE DERIVATIVE

We have seen that similar patterns can occur in variables with different
names. If we have f (B), then f 0 (B) asks for the derivative of f , just as f 0 (x)
does. Thus, (B 2 )0 = 2 · B, just as we already computed (x2 )0 = 2x. Once we
compute the derivative of a function abstractly, we have the derivative in the
same function no matter what variable is used. That efficiency is the point of
having general formulas.
Suppose we have y = f (x) and we know that f 0 (3) > 0. We claim that y
is increasing when x increases from x = 3. This is because ∆y/∆x → f 0 (3)
as ∆x → 0, with x = 3. Since f 0 (3) is positive, the ratios ∆y/∆x must be
positive when ∆x is small. If x increases from 3 a very small bit, then ∆x > 0,
and so for ∆y/∆x to be positive, we must have ∆y positive, as well. That ∆y
is positive is that y is increasing.
A similar argument can be made when f 0 (x) < 0. In that case, we say that
y is decreasing, which means that if x increases, then y decreases. This gives
very practical information that we will use often.
Virtually every function for which there is a formula has a derivative that
we can calculate, as we will see. But there are some (unusual) functions for
which the ratio in (3.5) does not do anything sensible as ∆x goes to 0. The
functions f (x) for which f 0 (x) exists are said to be differentiable. The most
common way for us to know that a function is differentiable is to have a formula
for its derivative; in the next few sections of this text we will develop a set of
rules for taking the derivative – rules which will apply to most every function
formula.
When we have y = f (x), the equations (3.5) and (3.6) write the derivative
two ways:
f (x + ∆x) − f (x) ∆y
(3.7) f 0 (x) = lim = lim
∆x→0 ∆x ∆x→0 ∆x
 3. THE DERIVATIVE OF A FUNCTION 41

These formulations are really the same, since

∆y = f (x + ∆x) − f (x)

The formulations represent different points of view. The ratio that uses f (x)
is thinking about the derivative as applying to a function; the ratio with ∆y
is thinking about the relationship between the variables x, y.
These two points of view are united in one more notation. It is customary
to write
dy ∆y
(3.8) = lim
dx ∆x→0 ∆x
Equation (3.6) shows that dy/dx is just another notation for the derivative:
dy
(3.9) = f 0 (x)
dx
Why have two notations for the same thing? As we mentioned, the two no-
tations represent two different points of view. It turns out that the dy/dx
notation will be useful in its own right. The dy and dx are differentials.2
Similar to the ∆x notation, dx is one symbol and not a product. It is not
trivial to explain exactly what differentials are; for our course, we will see that
they furnish a useful notation, especially when units are involved. Indeed,
suppose that x measures weight in pounds and y measures frogs, and suppose
there is a function y = f (x) that tells us how many frogs we get for so-many
pounds. (We can suppose that different frogs have different weights, so f (x)
may be quite complicated.) The units of the ratio ∆y/∆x would be frogs per
pound, that ratio coming directly from the units of y and the units of x. Since
∆y/∆x → dy/dx, the derivative dy/dx has the same units: frogs per pound.
This simple reasoning works in general. If we have an object moving along
the x-axis, where that axis measures feet, and if t is time in seconds, then the
2The differential notation dy/dx comes from Leibnitz, one of the discoverers of the
calculus. The book [15] has more information.
42 3. THE DERIVATIVE

velocity v(t) has units feet per second. The derivative formula
dx
v(t) =
dt
shows the same ratio of units: feet per second. Taking this one step further,
the derivative of v(t) could be written
dv feet per second
v 0 (t) = in units
dt second
and a little fraction algebra:
feet per second feet 1 feet
= · =
second second second second2
The units feet per second-squared are the units of acceleration, and in fact the
derivative of the velocity is acceleration, as we have mentioned.
If W = f (x) is the weight in pounds of the left-hand x feet of a steel bar,
then
dW
f 0 (x) =
pounds/foot
dx
As we know from our previous work, dW/dx is the density.
 CHAPTER 4

Computing the Derivative

Let’s summarize the notation for the derivative. Given y = f (x), we have
two ways to write the derivative:

dy
(4.1) f 0 (x) =
dx
The definition of the derivative asks us to think about a ratio; here are the
two formulations of that ratio.
∆y f (x + ∆x) − f (x)
(4.2) =
∆x ∆x
The ratio of deltas is the average change in y per change in x. When we let
∆x → 0, we get the derivative, the exact change in y per change in x:

f (x + ∆x) − f (x)
(4.3) f 0 (x) = lim
∆x→0 ∆x
A function that has a derivative is a differentiable function. In terms of dif-
ferentials:
dy ∆y
(4.4) = lim
dx ∆x→0 ∆x

The units of f 0 (x) are the units of y divided by the units of x.

We are about to learn rules for calculating the derivative of a function from
the form of the function, without using the limits (4.3) and (4.4) directly. We
have already learned one such rule: the derivative of m · x + b is the slope m.
43
44 4. COMPUTING THE DERIVATIVE

We remind you of a few other special cases that occur often.

c0 = 0 for all constants c

(x2 )0 = 2x
 0
1 1
=− 2
x x
In this chapter we will lay out the standard derivative formulas learned
by everyone taking calculus at every level. It doesn’t take a lot of verbiage
to write out all the formulas; to learn them you it will help to work as many
problems as possible for each formula. The problems document has exercises
of all types!

1. The Power Rule, Linearity

Power Rule Let n be a number. Then the derivative of xn is n · xn−1 . In

other words
dy
If y = xn then = (xn )0 = n · xn−1
dx
For instance

(x2 )0 = 2 · x and (x3 )0 = 3 · x2 and (x4 )0 = 4 · x3 and so on

Another example: since 1/x = x−1 , the Power Rule applies here, as well, for
that rule would compute
 −1 0 1
x = (−1) · x−2 = − 2
x
and this was what we got by the limit calculation of the derivative. We will
prefer writing the derivative of 1/x as −1/x2 .
We will discuss a general argument for the Power Rule later. Here are some
special cases, using the limit (3.5). We will skirt some of the details, in case
you want to try them yourself!

Problem T4.1. Show that (x3 )0 = 3 · x2 .

 1. THE POWER RULE, LINEARITY 45

Solution. If y = x3 , then1

∆y = (x + ∆x)3 − x3 = 3 · x2 · ∆x + 3 · x · (∆x)2 + (∆x)3

and so
∆y
= 3 · x2 + 3 · x · ∆x + (∆x)2
∆x
As ∆x → 0, we get (x3 )0 = 3 · x2 .

Problem T4.2. Show that (x−2 )0 = −2 · x−3 .

Solution. Let y = 1/x2 , and then

1 1
∆y = 2
− 2
(x + ∆x) x
x − (x + ∆x)2
2
=
(x + ∆x)2 · x2
−2 · x · ∆x − (∆x)2
=
(x + ∆x)2 · x2
and so
∆y −2 · x − ∆x
=
∆x (x + ∆x)2 · x2
As ∆x → 0, we get
dy −2 · x −2
= = = −2 · x−3
dx x4 x3

Here are two more rules will allow us to calculate the derivative of a wide
variety of functions.

Constant Multiple Rule If k is a constant, and f (x) is differentiable, then

0
k · f (x) = k · f 0 (x)

Problem T4.3. Compute (4 · x2 )0 .

1We are using an expansion of (x + ∆x)3 that is analogous to the expansion of (a + b)2
in (3.3). These are special cases of what is called the Binomial Theorem.
46 4. COMPUTING THE DERIVATIVE

0
0
Solution. Using the Constant Multiple Rule, we have 4 · x2 = 4 · x2 . We
know that (x2 )0 = 2 · x, and so

0
4 · x2 = 4 · (x2 )0 = 4 · 2 · x = 8 · x

Another case that comes up often: a minus sign is multiplication by −1,

so that
0
[−x3 ]0 = (−1) · x3 = (−1) · [x3 ]0 = (−1) · 3 · x2 = −3 · x2


When we take the derivative of a negative, we don’t have to convert the minus
sign to a (−1). We just pull out the minus sign; the point is that we are pulling
0
out a constant according to the Constant Multiple Rule: − f (x) = −f 0 (x).


Addition Rule If f (x) and g(x) are differentiable, then

0
f (x) + g(x) = f 0 (x) + g 0 (x)

Example. We have  0  0
1 5 1
0
+x = + x5
x x
We already know the derivative of 1/x and of x5 , and so
 0  0
1 5 1
0 1
+x = + x5 = − 2 + 5 · x4
x x x

The Constant Multiple Rule and the Addition Rule together are sometimes
called the linearity of the derivative. Here is an abstract version that combines
them: let f (x) and g(x) be differentiable functions, and let a, b be constants.
Then
 0  0  0
a · f (x) + b · g(x) = a · f (x) + b · g(x) Addition
= a · f 0 (x) + b · g 0 (x) Constant Multiple
 1. THE POWER RULE, LINEARITY 47

Example. Here we combine the power rule and linearity; underneath each
line we give the reason for that particular step.

0
4 · x5 − 3 · x3 + 7 · x + 1

0
0
0
0
= 4 · x5 + − 3 · x3 + 7 · x + 1
addition rule

0
0
0
= 4 · x5 − 3 · x3 + 7 · x + 0
constant multiple, derivative of a constant
= 4 · 5 · x4 − 3 · 3 · x2 + 7 · 1
power rule, including x0 = 1
= 20 · x4 − 9 · x3 + 7

Example. Slightly uglier functions.

9
0
0 1
0
0
3 · x2 + − x−4 = 3 · x2 + 9 · + − x−4
x x
addition rule

0 1
0
0
= 3 · x2 + 9 · + (−1) · x−4
x
constant multiple
1
= 3 · 2 · x + 9 · (−1) · 2 + (−1) · (−4) · x−5
x
power rule
9
= 6 · x − 2 + 4 · x−5
x
algebra

In the last two examples we belabored the steps to make sure you could see
exactly what we were doing. As we do more and more derivative calculations,
we will work at doing them more quickly, using the rules without thinking
48 4. COMPUTING THE DERIVATIVE

about them too much. In getting started, it will help to be self-conscious of

the rules we are using.

Problem T4.4. We are moving along the y-axis; position on that axis is
measured in feet. At time t seconds we are at y = 3t − 12t3 . What is our
velocity when t = −2?

Solution. Velocity is dy/dt feet per second. The derivative:

dy
0
= 3t − 12t3 = 3 · 1 − 12 · 3 · t2 = 3 − 36 · t2
dt
That’s the velocity at an arbitrary time t. When t = −2, we get velocity

3 − 36 · (−2)2 = 3 − 36 · 4 = −141 feet per second

Problem T4.5. An object travels along the y-axis (measuring meters in its
usual vertical orientation) such that y = 3 · t2 − 12 · t + 14 at time t seconds.
When is the object moving down?

Solution. “Moving down” means that the velocity dy/dt is negative. We

compute
dy
0
= 3 · t2 − 12 · t + 14 = 3 · 2 · t − 12 = 6t − 12 meters per second
dt
We want the velocity to be negative:

6t − 12 < 0 which is 6t < 12 that is t < 2

Before t = 2, the object is moving down.

If Q is a quantity that depends on time, then Q0 = dQ/dt measures the
rate of change of Q. Say, for example, that Q measures kg, and t is time in
seconds. Then Q0 has units kg/sec. Suppose that Q = 10 kg and Q0 = 2
kg/sec. It is common to express the rate of change 2 as a percentage of the
 1. THE POWER RULE, LINEARITY 49

value of Q. The number 2 is 2/10=20% of the number 10. The ratio 20% is
the relative rate of Q. Thus, the relative rate of Q is
Q0 1 dQ
= ·
Q Q dt
Of course, we must be assuming that Q 6= 0. Notice the units of relative rate
here: percent per second.
Many uses of the derivative involve abstract quantities for which we do not
have equations. Here is an example.

Problem T4.6. Helium gas in a closed container obeys the Ideal Gas Law :2
V · P = k · T , where P is the pressure, V the volume, T the absolute tem-
perature, and k a positive constant. In a closed container, we would have V
constant, whereas P, T could change in time. Show that P, T have the same
relative rates.

Solution. We take the derivative of both sides of the Gas Law equation, using
time t as variable. Since V, k are constants, the Constant Multiple Rule applies
to them:
dP dT
V · =k·
dt dt
To get at relative rates, we divide each side of this equation by the correspond-
ing side of the original equation:
1 dP 1 dT 1 dP 1 dT
·V · = ·k· so that · = ·
VP dt kT dt P dt T dt
This says that P, T have the same relative rate.
Since P, T have the same relative rates, if P is changing by, say, 23% per
second, then T is changing at 23% per second, as well.

2See [9, p.543].

50 4. COMPUTING THE DERIVATIVE

2. Products and Quotients

We introduce two more fundamental rules for calculating derivatives. We

will focus on understanding and using these rules. In class we will consider
the question of their validity.

Product Rule Given differentiable functions f (x) and g(x), we have

0
f (x) · g(x) = f 0 (x) · g(x) + f (x) · g 0 (x)

This rule is sometimes described as taking turns; we take the derivative of

the first function, leaving the second alone, and then we leave the first function
alone and take the derivative of the second function.
Example. We imitate the Product Rule carefully, and then apply the other
rules previously discussed. Make sure you can account for this calculation!
h i0
2
(x − 4 · x + 3) · (5 − 1/x)
= (x2 − 4 · x + 3)0 · (5 − 1/x) + (x2 − 4 · x + 3) · (5 − 1/x)0
= (2 · x − 4) · (5 − 1/x) + (x2 − 4 · x + 3) · (0 − (−1)/x2 )
= (2 · x − 4) · (5 − 1/x) + (x2 − 4 · x + 3) · (1/x2 )

The expression that results is somewhat complicated; nonetheless, we do not

try to simplify it, since we are just interested in applying the rules.
The rules of differentiation work together. Here is a calculation using the
Product Rule – but it could be done another way.

(x3 )0 = (x · x2 )0 = (x)0 · x2 + x · (x2 )0

= 1 · x2 + x · 2 · x = x2 + 2 · x2 = 3 · x 2

The answer agrees with the Power Rule; in fact, the Product Rule implies the
Power Rule in the case that the exponent is a positive integer.
Here is a practical, abstract use of the Product Rule.
 2. PRODUCTS AND QUOTIENTS 51

Problem T4.7. Suppose we sell tables, and the weekly demand D for tables
(how many we can sell) depends on the unit price p. The notation D(p)
expresses that demand is a function of price. Suppose that we don’t know D(p)
exactly, but we know that it is positive and we know that D0 (p) < 0. Under
what circumstances should we raise the price in order to increase revenue?
Solution. That D0 (p) < 0 is that D(p) decreases when p increases.3
The revenue R is the product of demand and price: R = D(p) · p; this is
the product of the number of tables and the price of each table. To have an
increase in price result in an increase in revenue, we want dR/dp > 0.
The Product Rule computes
dR
= D0 (p) · p + D(p) · 1
dp
In the inequality
D0 (p) · p + D(p) > 0
we want to divide by D0 (p). Since that quantity is negative, the division will
reverse the inequality:
D(p) D(p)
p+ 0
< 0 which is p < − 0
D (p) D (p)
Again we remind you that D0 (p) is negative, and so the fraction −D(p)/D0 (p)
is positive. The last inequality says that if the price is below the ratio −D/D0 ,
then we should raise the price to increase R (even though we will decrease
demand).
Problem T4.8. Use the Product Rule to show that (f (x)2 )0 = 2 · f (x) · f 0 (x).
Solution. Write f (x)2 = f (x) · f (x), and use the Product Rule:
h i0
f (x) · f (x) = f 0 (x) · f (x) + f (x) · f 0 (x) = 2 · f (x) · f 0 (x)

3This fact about the sign of the derivative was mentioned on p.40.
52 4. COMPUTING THE DERIVATIVE

The derivative rule for ratios of functions is somewhat similar to the Prod-
uct Rule.

Quotient Rule Given differentiable functions f (x) and g(x), we have

0
f 0 (x) · g(x) − f (x) · g 0 (x)

f (x)
=
g(x) g(x)2

The numerator of this formula should remind you of the Product Rule: we
take turns with each function. But there is a minus sign in between rather
than a plus sign. We need to take the derivative of the numerator f (x) first.

Example. The Quotient Rule used carefully.

" #0
3·x−4
x5 + 7 · x + 2
(3 · x − 4)0 · (x5 + 7 · x + 2) − (3 · x − 4) · (x5 + 7 · x + 2)0
=
(x5 + 7 · x + 2)2
3 · (x5 + 7 · x + 2) − (3 · x − 4) · (5 · x4 + 7)
=
(x5 + 7 · x + 2)2

As in our first Product Rule example, there is no need to work further with
this expression if we are just trying to apply the rule.
As with the Product Rule, the Quotient Rule is consistent with the other
rules.

Example. We calculate the derivative of x−5 using the Quotient Rule, apply-
ing the Power Rule to x5 .
 0
−5 0 1
(x ) = algebra
x5
10 · x5 − 1 · (x5 )0
= Quotient Rule
(x5 )2
 2. PRODUCTS AND QUOTIENTS 53

0 · x5 − 5 · x4
= Power Rule x5
x10
−5 · x4
= algebra
x10
= −5 · x4−10 = −5 · x−6

The final answer agrees with the Power Rule applied to x−5 .
We mentioned that the Product Rule can be used to establish the Power
Rule for xn when n is a positive integer. The previous example can be gen-
eralized to establish the Power Rule when n is a negative integer, using the
Quotient Rule. Thus, the Product Rule and Quotient Rule give the Power
Rule when the exponent is an integer.

Problem T4.9. The mass M kg of a cluster of cells satisfies M = t/(t2 + 9),

where t is time in hours. For t ≥ 0, when is the mass increasing?

Solution. Recall that M is increasing when dM/dt > 0. (By the way, the
derivative has units kg per hour.) We compute dM/dt using the Quotient
Rule.
dM 1 · (t2 + 9) − t · 2 · t
=
dt (t2 + 9)2
We need to work with this derivative to find out where it is positive. So, we
simplify the numerator:
1 · (t2 + 9) − t · 2 · t t2 + 9 − 2 · t2 −t2 + 9
= =
(t2 + 9)2 (t2 + 9)2 (t2 + 9)2
We want to know where this is positive. The denominator (t2 + 9)2 is positive,
and so the fraction is positive when the numerator is positive. That’s

−t2 + 9 > 0 which is 9 > t2

Since t ≥ 0, we see that 9 > t2 means that 3 > t. Thus, M is increasing when
t < 3.
54 4. COMPUTING THE DERIVATIVE

3. Exponentials and Logarithms

In Chapter 1 we introduced exponential functions Ax and the function

called the exponential function ex . Also, we have been enjoying the alternative
notation ex = exp(x). Be sure you remember the graph of y = ex and of the
logarithm y = ln(x); those graphs are on p.13. The prominence of ex comes
from the simplicity of its derivative.

The Exponential Function We have

0
ex = ex and exp0 (x) = exp(x)

This is a remarkable formula – it takes some experience to appreciate it.

Let’s focus on using it, at first.

Problem T4.10. For which x does y = exp(x)/(x + 3) have positive slope?

Solution. We are asking when dy/dx > 0. Use the Quotient Rule:
 x 0
e (ex )0 · (x + 3) − ex · (x + 3)0 ex (x + 3) − ex ex (x + 2)
= = =
x+3 (x + 3)2 (x + 3)2 (x + 3)2
(Notice, in the second equation, when we took the derivative of the ex , it
doesn’t look as if anything happened! That’s the simplicity of the derivative
rule for the exponential function.) We want our slope to be positive. The
exponential function is positive, and the (x+3)2 in the denominator is positive,
and so the sign of the slope is determined by x + 2. We see that the slope is
positive when x > −2.
The graph of the exponential function shows that it is always increasing. In
other words, its slope is always positive – but that slope is just the exponential
function, which is positive, so that makes sense.
On p.36, we estimated the derivative of ex at x = 2 numerically. Now
we see that the exact value of that derivative is e2 . The formula (ex )0 = ex
depends on the limit considered in the following.
 3. EXPONENTIALS AND LOGARITHMS 55

Problem T4.11. Show that (exp(∆x) − 1)/∆x ≈ 1 when ∆x is small. Then

assume that (exp(∆x) − 1)/∆x → 1 as ∆x → 0, and show that the derivative
formula for exp(x) follows.

Solution. Excel gives the following approximations:

∆x : 0.1 0.01 0.001 0.0001 0.00001
(e∆x − 1)/∆x : 1.051709 1.00502 1.000500 1.000050 1.000005
Not definitive but seemingly consistent with the claimed value 1.
Now assume the limit, and we use the properties of the exponential function
to manipulate the delta-ratio.
exp(x + ∆x) − exp(x) exp(x) · exp(∆x) − exp(x)
=
∆x ∆x
exp(x) · [exp(∆x) − 1]
=
∆x
exp(∆x) − 1
= exp(x) ·
∆x
The ratio to the right goes to 1, by assumption, and so the original fraction
goes to exp(x), and that’s the derivative of exp(x).
The limit we assumed in the previous problem requires more theoretical
knowledge of the exponential function than we are willing to pursue, so we
will be content to assume that the derivative formula for ex is correct. Later
in the text, we will see that this formula leads to the following.

The Natural Logarithm We have

1
ln0 (x) =
x

Problem T4.12. Where does x · ln(x) have derivative equal to 0?

Solution. Product rule:

h i0 1
x · ln(x) = 1 · ln(x) + x · = ln(x) + 1
x
56 4. COMPUTING THE DERIVATIVE

This is zero when ln(x) + 1 = 0, so that ln(x) = −1, and that’s x = e−1 = 1/e.

Problem T4.13. Explain why ln(4x) and ln(x) have the same derivative.

Solution. We need to remember an identity from Chapter 1:

ln(4x) = ln(4) + ln(x)

The number ln(4) is a constant, and its derivative is 0. And so,

 0  0 1
ln(4x) = ln(4) + ln(x) = 0 +
x
That’s why ln(4x) and ln(x) have the same derivative.

4. The Chain Rule

A composite function is formed by plugging one function into another. For

instance, suppose that f (y) = y 2 + 5 · y and g(x) = 3x − 2, then

f (g(x)) = g(x)2 + 5 · g(x) = (3x − 2)2 + 5 · (3x − 2)

The Chain Rule explains how to take the derivative of a composite function.

Chain Rule We have

h i0
f (g(x)) = f 0 (g(x)) · g 0 (x)

Let’s go right to an example: consider

 4 0
(x + 3 · x2 )5

The pattern f (g(x)) sees an outside function f and an inside function g. The
outside function in our example is the 5-th power function. The inside function
is x4 +3·x2 . The Chain Rule says to take the derivative of the outside function,
leaving the inside function alone. That’s the expression f 0 (g(x)) in the Chain
 4. THE CHAIN RULE 57

Rule formula. The derivative of the 5-th power function comes from the Power
Rule; here is the formula, using  as the variable.
 5 0
 = 5 · 4

In our example,  = x4 + 3 · x2 , so we have

4
f 0 (g(x)) = 5 · x4 + 3 · x2

The other factor in the Chain Rule formula is g 0 (x); that’s the derivative of
the inside function:
 4 0
x + 3 · x2 = 4 · x3 + 3 · 2 · x = 4 · x3 + 6 · x

Putting these together:

0
f (g(x)) = f 0 (g(x)) · g 0 (x)

 4 0
4
(x + 3 · x2 )5 = 5 · x4 + 3 · x2 · (4 · x3 + 6 · x)

Let’s do another one; this time we will try to focus on the outside-then-
inside pattern and not clutter things with function notation.
 0 h i0
1 4 −7
= (5 − 7 · x )
(5 − 7 · x4 )7
= −7 · (5 − 7 · x4 )−8 · (5 − 7 · x4 )0 ***
= −7 · (5 − 7 · x4 )−8 · (0 − 7 · 4 · x3 )

At the equation marked (***) we took the derivative of the (−7)th power (the
outside function), leaving the inside function 5 − 7 · x4 alone. Then we took
the derivative of the inside.

Problem T4.14. Find where is y decreasing, if

3
y=
(x2 + 5)2
58 4. COMPUTING THE DERIVATIVE

Solution. We want to find where dy/dx < 0. The expression for y is a

quotient, and we could use the Quotient Rule, but let’s notice that
3
= 3 · (x2 + 5)−2
(x2 + 5)2
We read (x2 + 5)−2 as an outside function (the (−2) power) and an inside
function x2 + 5. Here is the derivative:
0 0
3 · (x2 + 5)−2 = 3 · (x2 + 5)−2
 
constant multiple
= 3 · (−2) · (x2 + 5)−3 · (x2 + 5)0 Chain Rule
= −6 · (x2 + 5)−3 · 2 · x
−12 · x
= 2 algebra
(x + 5)3
Where is this negative? The denominator is always positive, so we would need
the numerator to be negative. We have −12 · x < 0 when x > 0.

Problem T4.15. Find the derivative of e3t .

Solution. The function e3t is a composite; this is easier to see if we use the
notation exp(3t); the outside function is the exponential function and the
inside function is 3t. Thus, we use the Chain Rule:
0
exp(3t) = exp0 (3t) · (3t)0 = exp(3t) · 3 = 3e3t


The derivative formula exp0 = exp was used, leaving the inside 3t alone.
The previous problem can be done another way; write e3t = (et )3 , and then
h i0
(et )3 = 3 · (et )2 · (et )0 = 3e2t · et = 3e3t

Same answer!
We take the previous problem a little further. Many of the exponential
models in Chapter 1 were of the form y = y0 · exp(k · t), where y0 , k are
constant. We can explain these models as showing that y changes at a rate
 4. THE CHAIN RULE 59

proportional to y. Indeed, the rate of change of y is

dy h i0
= y0 · exp(k · t) = y0 · exp(k · t) · (k · t)0 = y0 · exp(k · t) · k = k · y
dt
The expression k · y is proportional to y, with k being the constant of pro-
portionality. Thus, y changes at a rate proportional to y. Furthermore, the
relative rate of y is constant:
1 dy 1
· = ·k·y =k
y dt y
For example, suppose we have a population that grows at a relative rate of 10%
per year. Then k = 10% = 0.1, and the population P satisfies the equation
P = P0 · exp(0.1 · t), where t is time in years.
Problem T4.16. Where is the slope to y = exp(−x2 ) positive, and where is
it negative?
Solution. The slope is the derivative, and the derivative uses the Chain Rule,
as in the last problem.
y 0 = exp0 (−x2 ) · (−x2 )0 = exp(−x2 ) · (−2 · x) = −2 · x · exp(−x2 )
The exponential function is always positive, and so it doesn’t affect the sign
of y 0 . We see that y 0 > 0 when −2x > 0, and this happens when x < 0. Thus,
the curve y = exp(−x2 ) is increasing when x < 0. We have y 0 < 0 when x > 0,
and this is where y = exp(−x2 ) is decreasing.
Problem T4.17. An object moves along the y-axis (in meters) such that
y = ln(t2 − t + 1), where t is time in seconds. When is the object moving up?
Solution. (Note: the quantity t2 −t+1 is always positive, and so the logarithm
is defined.) We are moving up when the velocity dy/dt is positive.
dy  0
= ln(t2 − t + 1) = ln0 (t2 − t + 1) · (t2 − t + 1)0
dt
1 2t − 1
= 2 · (2t − 1) = 2
t −t+1 t −t+1
60 4. COMPUTING THE DERIVATIVE

As we mentioned, the quantity t2 −t+1 is positive, and so the sign is determined

by 2t − 1. We have 2t − 1 > 0 when t > 1/2. After 1/2 second we are moving
up.

Problem T4.18. Show that the relative rate of f (t) is the derivative of
ln(f (t)), assuming that f (t) is positive.

Solution. Chain Rule again!

0 1
ln(f (t)) = ln0 (f (t)) · f 0 (t) = · f 0 (t)

f (t)
as needed.
We have made the point previously that the differentiation rules are related
to each other.

Problem T4.19. Show that the Chain Rule and the Product Rule imply the
Quotient Rule.

Solution. Let f (x), g(x) be differentiable and write

f (x)
= f (x) · g(x)−1
g(x)
To take the derivative, we apply the Product Rule and then the Chain Rule.
 0 h i0
f (x) −1
= f (x) · g(x) product
g(x)
0
= f 0 (x) · g(x)−1 + f (x) · g(x)−1 chain on g −1


= f 0 (x) · g(x)−1 + f (x) · (−1) · g(x)−2 · g 0 (x) algebra

0 0 0
f (x) f (x) f (x)g(x) − f (x)g (x)
= − =
g(x) g(x)2 g(x)2
Why is the Chain Rule true? We will not give a complete proof, but we
can present the main idea of the proof. In order to compute [f (g(t))]0 , we need
to consider the ratio
f (g(t + ∆t)) − f (g(t))
∆t
 4. THE CHAIN RULE 61

We will rewrite the numerator, letting x = g(t), so that ∆x = g(t + ∆t) − g(t),
and then
f (g(t + ∆t)) − f (g(t)) f (x + ∆x) − f (x)
=
∆t ∆t
Supposing that ∆x 6= 0, we have
f (x + ∆x) − f (x) f (x + ∆x) − f (x) ∆x
= ·
∆t ∆x ∆t
We get the derivative of f (g(t)) if we let ∆t → 0.
Letting ∆t → 0, the ratio ∆x/∆t goes to dx/dt = g 0 (t). Also, it makes
sense that if ∆t → 0, then ∆x → 0, and so
f (x + ∆x) − f (x)
→ f 0 (x) = f 0 (g(t))
∆x
Putting the two factors back together, we get
0 f (x + ∆x) − f (x) ∆x
= f 0 (g(t)) · g 0 (t)

f (g(t)) = lim ·
∆t→0 ∆x ∆t
and that’s the Chain Rule.
The foregoing raises two technical issues; although we will not resolve either
of them, honesty compels us to point them out. First, we have to consider
what happens when ∆x = 0. Second, we need the fact that ∆x → 0 when
∆t → 0. That second fact is the continuity of x = g(t); you can look up
continuity in Section 10-3 of [1], if you are curious.
The Chain Rule can be used to explain the Power Rule. Recall the formula
(1.3) on p.6; it says that
xn = exp(n · ln(x))
(We need x > 0 so that ln(x) is defined.) Taking the derivative with respect
to x:
n n
(xn )0 = exp0 (n · ln(x)) · (n · ln(x))0 = exp(n · ln(x)) ·
= xn · = n · xn−1
x x
and that’s the Power Rule, at least when x > 0. We’ll be content with that
case.
 CHAPTER 5

Interpreting and Using the Derivative

We begin with an algebraic preliminary: the problem of determining where

a function is positive, where it is negative, and where it is zero. For a rational
function (a polynomial or a ratio of polynomials), we can answer the question
by factoring.

Problem T5.1. Let f (x) = x4 −x3 −6·x2 . Find out for where f (x) is positive
and where it is negative.

Solution. The roots of f (x) are the x values where f (x) is 0. It is a theorem
of elementary algebra that x = a is a root of f (x) if and only if x − a is a factor
of f (x). Thus, to find where f (x) is 0, we factor, starting with noticing the x2
on each term. For the other factors, there are several methods you may have
learned; we will be happy to help you individually if you need some review.

x4 − x3 − 6 · x2 = x2 · (x2 − x − 6) = x2 · (x + 2) · (x − 3)

The factors show that the roots of f (x) are x = 0, −2, 3. We graph these on
a number line, and this chops the real numbers into four intervals:

x < −2, −2 < x < 0, 0 < x < 3, 3<x

On each of these intervals f (x) is not 0, and so on each interval its sign cannot
change.1 There is a fairly simple way to figure out these signs, working from
1The assertion that f cannot change sign between roots follows from the fact already
enunciated that roots come from factors. There is also a calculus theorem called the Inter-
mediate Value Theorem that is relevant to this situation. You are encouraged to look up
that result online or in a calculus text.
63
64 5. INTERPRETING AND USING THE DERIVATIVE

right to left along the intervals. For the rightmost interval 3 < x, the sign is
determined by the leading coefficient of the polynomial – the coefficient on the
highest power of x. That highest power is x4 and the coefficient is 1; thus, the
sign on 3 < x is positive. If you wish, choose various values of x greater than
3; plug each of them into the polynomial to observe that you get a positive
value each time.
As we move to the left, across each root, we look at the exponent of that
root as a factor of f (x). When we cross x = 3, going from 3 < x to 0 < x < 3,
the factor is x − 3 to the first power. The exponent is 1 is an odd number, and
so the sign changes as we cross x = 3. Thus, the sign of 0 < x < 3 is negative.
Moving left across x = 0, the factor is x2 and the exponent 2 is even. When
we cross a root with an even exponent, the sign stays the same. The sign of
0 < x < 3 was negative, and so the sign of −2 < x < 0 is negative, too.
When we cross x = −2, we encounter the factor x+2 which has exponent 1,
an odd exponent, so that the sign changes. Here is the resulting sign diagram.
+ − − +
−2 0 3
For any number x, the diagram tells us quickly whether f (x) is positive, neg-
ative, or zero. For instance f (−1) < 0 and f (17) > 0 and f (1) < 0.
Problem T5.2. Find the signs of (x2 − 4 · x − 5)/(x − 7).
Solution. Factor!
x2 − 4 · x − 5 (x + 1) · (x − 5)
=
x−7 x−7
The roots are x = −1 and x = 5. The denominator root x = 7 is a singularity
– a place where the function is not defined. We include singularities in the
determination of sign, in the same way as roots. The leading coefficient on the
function is 1, and each exponent is 1. Here is the sign diagram.
− + − +
−1 5 7
 1. CURVE SKETCHING 65

Problem T5.3. Find the signs of 10 · x4 − 5 · x3 − 5 · x5 .

Solution. We rearrange terms to go from the highest power of x to the lowest;

then we see we can pull out −5 · x3 .

10 · x4 − 5 · x3 − 5 · x5 = −5 · x5 + 10 · x4 − 5 · x3
= −5 · x3 (x2 − 2 · x + 1)
= −5 · x3 · (x − 1)2

Here is the sign diagram; make sure you can explain each of the signs!
+ − −
0 1

We will work other examples in class.

1. Curve Sketching

We show how to make a quick sketch of a curve y = f (x). We will see that
a great deal of information can be observed from such a graph.
We know that dy/dx = f 0 (x) is the slope of the curve y = f (x). Also,
y = f (x) is increasing when f 0 (x) > 0, and it is decreasing when f 0 (x) < 0. If
f 0 (x) = 0 when x = a, then x = a is called a critical point. At each critical
point, the tangent has slope 0; the tangent is horizontal.
By a quick sketch we mean a sketch that plots critical points and that
indicates where a curve is increasing and where is it decreasing. The in-
creasing/decreasing feature of a curve is its oscillation. These features are
determined by the signs of the derivative.

Problem T5.4. Draw a quick sketch of y = x2 − 6 · x + 8.

Solution. We find the sign diagram for the derivative, as in the last section.
dy
= 2 · x − 6 = 2 · (x − 3)
dx
66 5. INTERPRETING AND USING THE DERIVATIVE

This expression is 0 when x = 3, it is positive when x > 3, and it is negative

when x < 3. These are signs of the derivative. To relate them back to the
curve: when x > 3 we have y 0 > 0, so that y is increasing. When x < 3, we
have y 0 < 0, so that y is decreasing.
The point x = 3 where y 0 = 0 is a critical point. When x = 3 we have
y = −1, so the critical point (3, −1) is on the graph. Here is a sketch of the
graph, using the information just obtained.

...... .....
...... .....
...... .....
...... ......
......
....... .
..
.......
.
.
....... .......
........ .......
......... ........
.......... .........
........... ..........
r
............... ...
..
..
..
..
............................................................

(3, −1)

Problem T5.5. Make a quick sketch of y = (x2 + x + 4)/(x + 1).

Solution. The derivative uses the Quotient Rule; the first equation below
applies that rule carefully. We skip lightly over the rest of the algebra:
(2 · x + 1) · (x + 1) − (x2 + x + 4) · 1
y0 =
(x + 1)2
x2 + 2 · x − 3 (x + 3) · (x − 1)
= 2
=
(x + 1) (x + 1)2
The critical points: x = −3 and x = 1. Here is the sign diagram, including the
resulting information about the graph. Note carefully the use of y 0 as opposed
to y. We use the abbreviations inc and dec for increasing and decreasing,
respectively.
y0 : + − − +
−3 −1 1
y: inc dec dec inc
 1. CURVE SKETCHING 67

As we said, x = −1 is a singularity – the function is undefined there. Thus, the

graph looks like this. (The vertical line at x = −1 is not part of the graph.)

... .
... ...
... ...
... ...
...
... ...
... ...
..
...
... ...
.
... ..
.... ...
....
r
....
........ .............
.........

(−3, −5) (1, 3)

r
...........
....... ............
y = (x2 + x + 4)/(x + 1)
.... ....
.... ...
.... ...
...
... ...
... ...
... ...
... ...
... ...
.... ...
...
... .

x = −1

Problem T5.6. Draw a quick sketch of y = x4 − 8 · x3 + 18 · x2 − 20.

Solution. The derivative has x as a factor:
y 0 = 4 · x3 − 24 · x2 + 36 · x
= 4 · x · (x2 − 6 · x + 9)
= 4 · x · (x − 3)2
The critical points x = 0 and x = 3. When x = 3, y = 7; when x = 0,
y = −20. Here are the signs.
y0 : − + +
0 3
y: dec inc inc
Can you draw the shape? There are horizontal tangents at (0, −20) and (3, 7).
68 5. INTERPRETING AND USING THE DERIVATIVE

Using a Quick Sketch

We mentioned that there are several features of a graph that are apparent
from its oscillations. One such feature is the number of roots that the function
has – the number of times the graph of y = f (x) crosses the x-axis.

Problem T5.7. How many real roots does x4 − 8 · x + 3 have?

Solution. Let y = x4 − 8 · x + 3, so that the polynomial has a root when the

curve crosses the x-axis. The graph should tell! The derivative and its factors:

y 0 = 4 · x3 − 8 = 4 · (x3 − 2)
√
This has only one critical point: x3 −2 = 0 gives x = 3 2. You may or may not
remember how to factor x3 − 2; let’s show that we don’t need to factor it. We
√
can determine the signs of y 0 by choosing points on either side of 3 2 ≈ 1.26
√
and observe the sign of y 0 . Taking x = 1 to the left of 3 2, we see that y 0 = −4;
√
thus, y 0 < 0 to the left of the cube root. Taking x = 2 to the right of 3 2, we
compute y 0 = 24, so that y 0 > 0 to the right of the critical point. This shows
√
us that y = x4 − 8 · x + 3 has a minimum as x = 3 2. Plugging in, we get
√
y ≈ −4.56. Since the point ( 3 2, −4.56) is below the x-axis, and since y rises
√
to the left and to the right2 of x = 3 2, we see that x4 − 8 · x + 3 has two real
roots.
Here is another kind of problem for which a quick sketch is relevant.

Problem T5.8. Find the maximum and minimum of y = −2x3 +3x2 +12x+10
for 0 ≤ x ≤ 3.

Solution. Factor y 0 :

−6x2 + 6x + 12 = −6(x2 − x − 2) = −6(x − 2)(x + 1)

2We assume you know that the values of a polynomial get large without bound as the
x-values get large and positive or large and negative. For a polynomial of degree 4, y will
get large and positive as x gets large in both directions.
 1. CURVE SKETCHING 69

and the sign diagram.

y0 : − + −
−1 2
Now we consider the values of x between 0 and 3.
y0 : + −
0 2 3
The graph shows that y has a maximum at x = 2, where y = 30. The minimum
of y on this interval must occur at one of the two endpoints. When x = 0,
y = 10 and when x = 3, y = 19. The minimum occurs at x = 0.
Problem T5.9. Using the same function as the previous problem, find the
maximum and minimum for −2 < x ≤ 0.
Solution. We use the same sign diagram for the critical points, superimposing
the endpoints −2, 0.
y0 : − +
−2 −1 0
We see that the minimum occurs at x = −1, y = 3. The maximum would have
to be an endpoint. When x = −2, y = 14, and when x = 0, y = 10. Thus,
x = −2 is the winner, but notice that x = −2 is not allowed in the specified
interval −2 < x ≤ 0. It follows that there is no maximum.
Later in the course we will take an extended look at applied problems
involving maximums and minimums.

The Second Derivative. The derivative measures whether a function is

increasing or decreasing. The derivative of the derivative, called the second
derivative, measures whether the derivative is increasing or decreasing. The
second derivative of f (x) is
 0 0
f (x) denoted f 00 (x)
In terms of the graph, f 00 (x) tells us the concavity of the curve. When f 00 (x) >
0, we say that y = f (x) is concave up; when f 00 (x) < 0, we say that y = f (x) is
70 5. INTERPRETING AND USING THE DERIVATIVE

concave down. Between the two possibilities for the sign of f 0 (x) and the two
possibilities for the sign of f 00 (x), there are four possibilities. The following
circle picture gives the shape of the curve in each case.3 We will discuss this
in class.

...........................
r
............................................................
...............
............... ...........
........... ..........
.
............... ........
...... .......
.... .......
....... ....
..
... ....
... ....
.. ....

..
.
.
.
..
...
y0 > 0 y0 < 0 ...
...
...
...
..
. ...
.... ...

...
...
...
y 00 < 0 y 00 < 0 ...
..
..
..
... ..
..
... ..
... ..
...
... ...
r...
...
..
..
r
..
...
..
.. ....
.. .
.
.. ..
.. ..
.. ..
..
..
..
..
...
y0 < 0 y0 > 0 ...
.
.
..
..

... ...
... ..
..
y 00 > 0 y 00 > 0
... .
.
...
... ...
... ...
.... ...
.... ......
.... ...
.... ....
....... ......
....... .......
........
.......... ..
. ..
..........
............ ......
............
.................
r
.............................................................................................

Problem T5.10. Graph y = x · exp(−x) plotting critical points and showing

oscillation and concavity.

Solution. We use the Product Rule and the Chain Rule to compute

y 0 = −(x − 1) · exp(−x)
y 00 = (x − 2) · exp(−x)

The exponential function is always positive, no matter that its argument is

−x, and so it doesn’t contribute any signs. Here is the sign diagram, including

3We are using the arcs on the circle to indicate the shape. Of course, we are not saying
that all curves are circular.
 1. CURVE SKETCHING 71

both derivatives. We abbreviate concave down to dn and concave up to up.

y0 : + − −
y 00 : − − +
1 2
y: inc dec dec
dn dn up

And here is the graph; notice the change in shape at x = 2. (In drawing the
picture, we have exaggerated that change somewhat to make it prominent.)

(1,.r1/e)
.....
.......... ..............
..... ....
....

.
.
..
....
...
....
...
...
...
y = x · exp(−x)
... ...
...
... ...
... ...
... ...
.
..
.. ...
..... r (2, 2/e2 )
......
......
.......
.......
........
........
.........
...........
...............
..................

Many functions that occur in economics have a positive derivative and a

negative second derivative. Here is a made-up example: suppose your expected
grade G(h) on an exam depends on the number of hours h you study, according
to the following table. (Grades are given on the 4-point scale, so 1=D, 2=C,
etc.)
h: 0 1 2 3
G(h) : 1 2.75 3.5 3.75
Your grade increases when h increases; that’s hardly surprising: it says that
the derivative G0 (h) is positive. But notice that when h changes from 0 to 1,
G(h) increases by 1.75, whereas when h changes from 1 to 2, G(h) increases
by only 0.75. The increase from the first additional hour of study is greater
than the increase from the second hour. And for the third hour, the increase
is only 0.25. It looks as if the derivative G0 (h) might be decreasing: G00 (h) is
negative.
72 5. INTERPRETING AND USING THE DERIVATIVE

A function with a positive derivative and a negative second derivative ex-

hibits diminishing returns; the return on investing an additional hour of study,
in our example, is greater for the first hour than for the second hour, and
greater for the second hour than for the third hour.
Problem T5.11. Let A, b be positive constants with b < 1. Show that the
function y = A · xb exhibits diminishing returns for x > 0.
Solution. We use the Power Rule on xb .
y 0 = A · b · xb−1 and y 00 = A · b · (b − 1) · xb−2
Since x > 0, the terms xb−1 and xb−2 are positive. Thus, y 0 > 0. Since b < 1,
we have b − 1 < 0, and so y 00 < 0.

2. Newton’s Method

There are many ways to find numerical solutions to equations of the form
f (x) = 0. We will discuss a recursive method called Newton’s Method or the
Newton-Raphson Method .4
We want to approximate a solution to the equation f (x) = 0. We start
with a number R0 , and compute
f (Rn )
Rn+1 = Rn −
f 0 (Rn )
Of course, we need f 0 (Rn ) 6= 0. This method is not infallible, but, remarkably
often, if the initial R0 is at all close to a solution, the sequence gives very
accurate approximations to the solution.
√
Problem T5.12. Use Newton’s Method to approximate 2.
√
Solution. We need 2 to be a solution to an equation of the form f (x) = 0;
let’s use x2 − 2 = 0, so that f (x) = x2 − 2 in the notation of the method. Then
4Asthe names imply, the method goes back to Newton and/or Raphson, although the
connection with calculus was, apparently, not elucidated at first.
 2. NEWTON’S METHOD 73

f 0 (x) = 2x. Thus, the recursion is this:

Rn2 − 2
(5.1) Rn+1 = Rn −
2Rn
There is no single way to come up with the initial approximation R0 . Since
√
2 is between 1 and 2, we let R0 = 1. Here is the sequence up to R5 .
n: 0 1 2 3 4 5
Rn : 1 1.5 1.4167 1.4142157 1.4142136 1.4142136

The Rn look constant from R5 onward. Comparing to Excel’s sqrt(2) (an-

other approximation!), we see that we have apparent accuracy.
The recursion expression just used can be simplified significantly:
Rn2 − 2 Rn 1 Rn 1
Rn+1 = Rn − = Rn − + = +
2Rn 2 Rn 2 Rn
You might recognize the resulting recursion as being an example of the Baby-
lonian sequence! The Babylonians did not have Newton’s Method in general,
but their calculation of square roots was a special case of Newton’s Method.
The expression we just obtained for the square root approximation is sim-
pler than the recursion (5.1), and so it might be preferable. Such algebra is
sometimes useful for employing Newton’s Method, but we will not insist on it.
When you use the method, feel free to use the expression Rn − f (Rn )/f 0 (Rn )
without simplifying.

Problem T5.13. The equation ex = c has solution x = ln(c). Use Newton’s

Method to approximate ln(c), where c = 2, 3, 4, 5.

Solution. Newton’s Method needs an equation of the form f (x) = 0, and so

we write the equation ex = c as ex − c = 0. Thus, f (x) = ex − c. Here is the
recursion:
exp(Rn ) − c
Rn+1 = Rn −
exp(Rn )
74 5. INTERPRETING AND USING THE DERIVATIVE

Taking R0 = 1 in each case, here are some of the approximations.

n: 1 2 3 4 5
ln(2) ≈ 0.735759 0.694042 0.693148 0.693147 0.693147
ln(3) ≈ 1.103638 1.098625 1.098612 1.098612 1.098612
ln(4) ≈ 1.471518 1.389825 1.386301 1.386294 1.386294
ln(5) ≈ 1.839397 1.633963 1.609736 1.609438 1.609438
In each case, R5 looks like a good approximation to the logarithm, as can be
seen by checking the values in Excel.
We have introduced the recursion for Newton’s Method; let’s explain how
it comes about. Imagine a graph y = f (x) at a point x = Rn near a root
f (c) = 0. On the curve, this point is (Rn , f (Rn )). To find the root x = c, we
would ideally like to follow the graph y = f (x) to the point (c, 0) on the x-
axis. The idea behind Newton’s Method is to use the tangent line to y = f (x)
at x = Rn as a substitute for the curve. In other words, we find the point
(Rn+1 , 0) where the tangent line meets the x-axis. (See the picture below.)
We would want the two points (Rn , f (Rn )) and (Rn+1 , 0) both to be on the
tangent line – that means that the slope of the tangent line can be found from
these points:
f (Rn ) − 0
slope =
Rn − Rn+1
0
On the other hand, the slope is f (Rn ), and so we have
f (Rn )
f 0 (Rn ) =
Rn − Rn+1
When we solve that equation for Rn+1 we get the recursive equation that
defines Newton’s Method.
...
r (Rn , f (Rn ))
....
....
....
..
....
....

.....
.....
..
..
.......
..
......
......
.......
(c, 0) ..
..
..
.
.......
.......
.

........
.........

.
..
..
.
r .........
...........
...........
r
 x
.
......
..
.
.............
..............
...............

(Rn+1 , 0)
y = f (x)
 3. THE CHAIN RULE REVISITED 75

Here is another example.

Problem T5.14. Find where the tangent line to y = x4 + x2 + x + 3 is
horizontal.
Solution. For the tangent line to be horizontal, its slope needs to be 0. Thus,
we want to solve dy/dx = 0:
4 · x3 + 2 · x + 1 = 0
We don’t know how to factor this cubic; graphing it shows that it has one
root; we use Newton’s Method:
4 · Rn3 + 2 · Rn + 1
Rn+1 = Rn −
12 · Rn2 + 2
Starting with R0 = 0, we compute R6 ≈ R7 ≈ −0.3855. This gives us the
x-value, and the y-value ≈ 2.7852.

3. The Chain Rule Revisited

3.1. Differentials. It is interesting to write the Chain Rule in terms of

differentials. Let y = f (x) and x = g(t), so that y = f (g(t)). Here are the
various derivatives:
dy dx dy  0
= f 0 (x) = g 0 (t) = f (g(t))
dx dt dt
Here is the Chain Rule, with x = g(t) substituted in.
0
f (g(t)) = f 0 (g(t)) · g 0 (t) = f 0 (x) · g 0 (t)


We replace the derivatives on the far left and far right by differential expres-
sions, and we get the differential version of the Chain Rule.
Chain Rule – Differential Version Given x = g(t) and y = f (x), where
g(t) and f (x) are differentiable, we have
dy dy dx
(5.2) = ·
dt dx dt
76 5. INTERPRETING AND USING THE DERIVATIVE

Equation (5.2) looks like the dx factors were canceled. When we introduced
differentials, we said we would not give an explicit definition – in any case
they are not numbers, and so the dx factors don’t cancel literally. But the
Chain Rule does say that the equation (5.2) is true, so the point is that we can
apparently manipulate differentials as if they were numbers. We claimed before
that differentials would prove useful; the Chain Rule justifies that promise.

3.2. Inverse Functions. We will consider a very simple example: y = x5 .

The differential notation dy/dx asks for the derivative of y as a function of x.
Thus,
dy
= 5 · x4
dx
We can rewrite the equation y = x5 by taking the fifth root: y 1/5 = x. In
this equation x is a function of y, and the derivative dx/dy would ask for the
derivative of that function.
dx 1
= · y −4/5
dy 5
This derivative is defined as long as y 6= 0, since the negative exponent indi-
cates a term in the denominator.
The Chain Rule showed that differentials seem to behave algebraically.
Let’s show that they behave that way in this context. Compute
dy dx 1
· = 5 · x4 · · y −4/5
dx dy 5
4 −4/5
=x ·y canceling 5’s
= x4 · (x5 )−4/5 using that y = x5
= x4 · x−4 multiplying consecutive exponents
= x4−4 = x0 = 1 adding exponents

Here is the upshot:

dy dx
(5.3) · =1
dx dy
 3. THE CHAIN RULE REVISITED 77

As with the equation (5.2), it looks as if the differentials are playing algebra
– we keep saying that is why they are useful.
The function x5 and y −1/5 are called inverse functions; they undo each
other:
(x5 )1/5 = x5/5 = x1 = x and (y 1/5 )5 = y 5/5 = y 1 = y
You have seen this pattern before: y = exp(x) if and only if ln(y) = x. And
it occurs in other situations. Let’s show that (5.3) holds whenever we have
differentiable inverse functions. If y = f (x) and x = g(y) are inverses, then

x = g(y) = g(f (x))

Using the ordinary Chain Rule on both sides:

1 = g 0 (f (x)) · f 0 (x)

Now differentials: g 0 (f (x)) = g 0 (y) = dx/dy, and f 0 (x) = dy/dx. We see that
(5.3) holds.
But there is an additional fact. It turns out that equation (5.3) holds even
if we do not have an explicit formula for the inverse function, as long as the
derivatives involved in (5.3) are not 0. This fact is called the Inverse Function
Theorem.5 Here is an applied example.

Problem T5.15. We observe that demand x for pianos depends on price p

(thousands of dollars). Suppose that
44 · 103 30
x= + +3
p3 p
An economist might like to think of price as a function of demand. What
is dp/dx when p = 10? (We will see later that this derivative is called the
marginal price.)
5Technically, if y = f (x) has a non-zero derivative on an open interval, then the y values
will traverse an open interval, as well, and there will be an inverse function x = g(y) for y
in its open interval.
78 5. INTERPRETING AND USING THE DERIVATIVE

Solution. The definition of the expression dp/dx requires us to find p = f (x),

and then dp/dx = f 0 (x). We have an equation x as a function of p, but it
is not clear that we can solve for p in this equation. The Inverse Function
Theorem tells us that we don’t have to! Using the equation analogous to (5.3)
in the present context, we see that
dp dx
· =1
dx dp
and this equation can be written
dp 1
(5.4) =
dx dx/dp
as long as dx/dp 6= 0. We use the equation for x as a function of p to compute
dx/dp, remembering that 1/p3 = p−3 .
0
44 · 103 30 3 · 44 · 103 30

dx
= + + 3 = − − 2
dp p3 p p4 p
When p = 10, we see that
dx 3 · 44 · 103 30
=− − = −13.5 pianos per thousand dollars
dp 104 102
Equation (5.4) shows that
dp 1 1
= =− thousand dollars per piano
dx dx/dp 13.5

Notice that both derivatives were negative – increasing price decreases

demand, and so increasing demand decreases price.
We can use this reasoning to derive the formula for (ln(x))0 from the formula
for (ex )0 . Let y = ln(x), so that x = ey . Then
 0 dy 1 1 1 1
ln(x) = = = y 0 = y =
dx dx/dy (e ) e x
 3. THE CHAIN RULE REVISITED 79

Previously, we stated that ln0 (x) = 1/x, but we didn’t explain where that
formula came from. Now we see that it comes from the formula exp0 (x) =
exp(x) and the Inverse Function Theorem.
Here is a problem that combines (5.2) and (5.3).
Problem T5.16. A particle traces a curve in the xy-plane. At time t, we
have x = (t2 − 1)/(t2 + 1) and y = 2t/(t2 + 1). What is the slope of the path
of the particle when t = 2? (Slope is dy/dx, as always!)
Solution. We see that we can calculate dx/dt and dy/dt. Treating differentials
algebraically, and not worrrying about reasoning for the moment, we might
expect that
dy dy/dt
=
dx dx/dt
provided that dx/dt 6= 0. Compute dy/dt using the Quotient Rule.
dy 2(t2 + 1) − 2t(2t) 2 − 2t2
= =
dt (t2 + 1)2 (t2 + 1)2
Similarly,
dx 2t(t2 + 1) − (t2 − 1)2t 4t
= 2 2
= 2
dt (t + 1) (t + 1)2
Then
dy dy/dt (2 − 2t2 )/(t2 + 1)2 2 − 2t2
= = =
dx dx/dt 4t/(t2 + 1)2 4t
When t = 2, we would have dy/dx = −6/8 = −3/4. That would be the slope
of the path when t = 2.
We worked the previous problem in a typical way – people don’t usually
worry too much about justifying differential calculations. For completeness,
let’s show that what we did is valid. We had dx/dt = 4t/(t2 + 1)2 , and that’s
not zero when t = 2. The Inverse function Theorem then says that t is a
function of x near that point, and (5.3) kicks in to show that
dt 1 (t2 + 1)2
= =
dx dx/dt 4·t
80 5. INTERPRETING AND USING THE DERIVATIVE

Theoretically, t is a function of x, and the equation y = 2t/(t3 + 1) shows that

y is a function of t. Then y is a function of x and the Chain Rule shows that
dy dy dt
= ·
dx dt dx
This shows that our calculation of dy/dx was justfiied.
As we move forward, we won’t stop to anaylze the reasoning behind our use
of differentials. As long as we don’t divide by zero derivatives, the algebraic-
like equations (5.2) and (5.3) will win through.

4. Marginals

Suppose that the cost C of making something is a function of the number

x made. Not unreasonable. If C, x have specific values, the marginal cost is
the change in cost when x increases by 1. In other words, marginal cost is ∆C
when ∆x = 1. Since
∆C dC
≈
∆x dx
we have ∆C ≈ dC/dx when ∆x = 1. Because the derivative dC/dx can
often be computed, that derivative is usually taken for the marginal cost.6 We
might only be interested in the sign of dC/dx, since that tells us whether C
will increase or decrease if x increases.
If M is a quantity (particularly, an economic quantity) that depends on
the level x of production of something, then marginal M is the change in M
when the one extra unit of production – generally taken to be dM/dx. There
is sometimes a little insignificant confusion about the units of marginal M .
For instance, if M is dollars and x is pianos, then dM/dx is dollars per piano
– a unit cost. However, we are taking dM/dx as an approximation for ∆M ,
and ∆M would be just dollars. This distinction is usually immaterial; we will
use the units of the derivative to be consistent with our previous terminology.
6Alternatively,some sources define marginal cost to be dC/dx, but in any case the idea
is to use the derivative to estimate ∆C.
 4. MARGINALS 81

If we remember that ∆M was actually ∆M/∆x, with ∆x = 1, then we arrive

at the derivative units: dollars per piano.

Problem T5.17. The cost of producing x units of a good is the sum of three
quantities: a constant overhead of $6000 for machinery, a production cost of
$50 per unit, and a quantity discount $1000 times the square root of x. If we
currently produce 120 units, what is the marginal cost?

Solution. The description of the cost shows that

√
C = 6000 + 50 · x − 1000 · x

The marginal cost is none other than dC/dx:

dC 1 500
= 50 − 1000 · · x−1/2 = 50 − √
dx 2 x
When x = 120, we have C 0 ≈ 4.36. We expect the cost to increase by roughly
$4.36 if one extra unit is produced.

Problem T5.18. In the previous problem, the total value is x · C. What is

the marginal total value when 100 units are produced?

Solution. The total value T = x · C where C is cost. By the Product Rule,

dT
= C + x · C0
dx
0
We have the formula for C in the previous problem. Plugging in x = 100, we
compute that T 0 ≈ 1000 (dollars per unit).
The word marginal can indicate various other sorts of derivatives. Suppose,
for example, that the production P is determined by the amount of labor L
used; let’s write P = f (L). The derivative
dP
= f 0 (L)
dL
is called the marginal product of labor. The abstract phrase marginal Y of X
is simply dY /dX.
82 5. INTERPRETING AND USING THE DERIVATIVE

Here is a typically convoluted problem. Algebra and calculus sort every-

thing out.
Problem T5.19. It takes 3 hours of labor to produce one table. The produc-
tion cost of x tables is 100x − x2 /10 dollars. What is the marginal production
cost of hours of labor when we make 10 tables?
Solution. Let L be the number of hours of labor, and let V be the production
cost. The marginal production cost of labor is simply dV /dL, and that’s what
we want.
The problem statement tells us that V = 100x − x2 /10, and since it takes
3 hours to make one table we have L = 3x. Thus, we can easily get these
derivatives:
dV 2 dL
= 100 − x = 100 − x/5 and =3
dx 10 dx
The derivative we want can be found using the differentials:
dV dV /dx 100 − x/5
= =
dL dL/dx 3
Now we are ready for specific numbers. With x = 10, the marginal value
of labor is
100 − 10/5 98
= dollars per hour
3 3

Let’s interpret the answer, 98/3 dollars per hour, of the previous problem.
Given that we now make 10 tables (using 30 hours of labor), if could increase
labor by one hour, our production costs would go up by roughly 98/3 dollars.
Problem T5.20. Suppose that the supply S of pencils (sold in boxes of 1000)
satisfies S = 2 + 1.3 · p, where p is the price of a box in dollars. Suppose that
the demand D for pencils satisfies D = 35 − m · p, where m is a constant.
Describe m as a marginal. If m increases, what happens to the price where
supply and demand are equal? What happens to revenue when m = 2 and
increases?
 4. MARGINALS 83

Solution. We see that dD/dp = −m, so that −m is the marginal demand of

price.
We are interested in the price where S = D; let’s call that price q. (We
use q rather than p, since q is a specific value of the variable p.) Using the
expressions for S and D, we have

2 + 1.3 · q = 35 − m · q

Moving the q-terms together, this is

1.3 · q + m · q = 35 − 2 which is (1.3 + m) · q = 33

And we have
33
q=
1.3 + m
To measure the change in q as m increases, we take the derivative
dq
0 33
= 33 · (1.3 + m)−1 = −33 · (1.3 + m)−2 = −
dm (1.3 + m)2
Since the derivative is negative, we see that q decreases as m increases.
As for revenue, that is R = D · p. When p = q, we get this formula for R:

R = D · q = (35 − m · q) · q = 35 · q − m · q 2

We can take the derivative with respect to m, remembering that we already

have a formula for dq/dm.
dR dq dq
= 35 · − 1 · q 2 − m · 2q ·
dm dm dm
When m = 2, we see that q = 33/3.3 = 10 and dq/dm = −33/3.32 ≈ −3.03,
so that
dR
(5.5) ≈ 35 · (−3.03) − 100 − 40 · (−3.03) = −84.85
dm
If m increases, R will decrease dramatically.
In the previous problem, suppose it would cost us $7 to decrease m by 0.1.
(Perhaps, advertising would make demand less sensitive to price.) Should we
84 5. INTERPRETING AND USING THE DERIVATIVE

spend the $7? The formula (5.5) estimates

∆R ≈ −84.85 · ∆m = −84.85 · (−0, 1) = 8.485
We can increase revenue by $8.49 by spending $7; seems like a good idea.
There are many terms involving marginals. It is not the purpose of this
course to learn the terminology, but here is a typical problem.
Problem T5.21. Suppose that p = 1000−x1/3 , with price p and demand x for
a certain good. The price elasticity of demand is the ratio of relative change
of demand to relative change of price. Find that elasticity when x = 100.
Solution. The set-up has x as independent variable, and so the relative change
in p is
1 dp 1 1 1
· = · (−1) · · x−2/3 = −
p dx p 3 3 · p · x2/3
The relative change in x is
1 dx 1
· =
x dx x
The elasticity is the ratio of the relative changes (prepare for some algebra):
1/x 3·p
2/3
= − 1/3
−1/(3 · p · x ) x
When x = 100, we get p ≈ 995.36, and the elasticity is roughly −625.
 CHAPTER 6

Linear Optimization

An optimization problem seeks the maximum or minimum value of a func-

tion. For a simple function of one variable, this information can be obtained
from the graph – from the signs of the derivative. More practical optimization
problems usually involve many variables, along with complicated conditions
on those variables that make the one-variable graphing approach impractical.
Among the more complicated problems there is a fundamental dichotomy be-
tween linear optimization problems, which are introduced in this chapter, and
the non-linear optimization problems introduced in Chapter 12 (in the second
term of this course). We will use simple examples to introduce linear prob-
lems; once we understand the basic form and terminology, we will consider
more complicated (interesting!) problems. We will also study decision-making
based on the properties of solutions.
We will be able to use graphing to handle simple examples; later we will
be using Excel to obtain numerical solutions.

1. Simple Examples

For our first example we will focus strictly on the mathematical details of
a typical, small problem.

Problem T6.1. Find the maximum of Z = 6 · x + 7 · y, where 3 · x + 5 · y ≤ 15

and 2 · x + y ≤ 4 and x, y ≥ 0.
85
86 6. LINEAR OPTIMIZATION

Solution. The quantity Z being optimized (in this case, maximized) is called
the objective.1 The objective in this problem is a function of the variables x, y;
those are the problem variables. The other conditions on the variables are the
constraints.
The constraints may look complicated, but they are easy to graph. To
graph the inequality 3 · x + 5 · y ≤ 15, we solve for y:
1 3
3 · x + 5 · y ≤ 15 yields y ≤ · (15 − 3 · x) = 3 − · x
5 5
The inequality y ≤ 3 − 35 · x refers to points on or below the line y = 3 − 35 · x.
Similarly, the inequality 2 · x + y ≤ 4 comes to y ≤ 4 − 2 · x, and that’s the
points on or below the line y = 4 − 2 · x. The other two constraints x, y ≥ 0 say
we are in the first quadrant. Here, to the left, is a picture of the constraints;
we included the x-intercepts of the two lines, as well as the y-intercepts.

(0, 4)
A
(0, 3) HAA (0, 3) HHr(5/7, 18/7)
H
AHH A
A H A
A HH A
(2, 0) (5, 0) (2, 0)
The problem requires all the constraints to hold. That means that the
constraints define the set of points on and inside the small quadrilateral defined
by the picture; that quadrilateral is pictured to the right. We have caculated
the intersection point between the two lines. We claim that the maximum of
Z cannot occur interior to the quadrilaterial. This is easy to see: at an interior
point, we can increase x or y, or both, and Z = 6 · x + 7 · y will increase.

1Noticethat this use of the word objective is not the usual one. Usually, an objective is a
goal. The word objective in an optimization problem refers to the quantity being optimized,
as opposed to the desire to optimize that quantity. This possibly confusing usage has become
standard.
 1. SIMPLE EXAMPLES 87

We also claim that the maximum of Z occurs at one of the corner points of
the quadrilateral. This will best be shown in class: briefly, along each boundary
line we can write Z as a linear function of one variable. A linear function has
a maximum at one of its endpoints – the endpoints of the boundary lines are
the corners in the picture. We will give more specific details in class.
To find the maximum of Z, we have only to calculate its values at each
corner:
point: (0, 0) (2, 0) (0, 3) (5/7, 18/7)
Z 0 12 21 156/7
Because 156/7 ≈ 22.3, we see that the maximum of Z is 156/7, occuring at
(5/7, 18/7).
In the problem just worked, the point (5/7, 18/7) gives values of the prob-
lem variables: x = 5/7 and y = 18/7. We call this a solution to the problem.
To repeat: the solution gives the values of the problem variables, not the value
of the objective. The value of the objective Z = 156/7 at the solution is the
maximum of the objective.
Here is a similar problem that we will work in class.

Problem T6.2. Find the maximum of Z = 3 · a + 2 · b where 2 · a + 7 · b ≤ 14

and 7 · a + 5 · b ≤ 35 and a, b ≥ 0.
In each of the two problems we considered, the objective (quantity) is a
linear function of the problem variables. Also, the constraints involve linear
expressions in the problem variables. In the examples we used, the constraints
were non-strict inequalities. They can also be equations, as we will see in more
complicated examples. The presence of only linear expressions in the objective
and constraints is what makes an optimization problem linear.
It is possible for the constraints of a linear optimization problem to con-
tradict each other. In that case, we say that the problem is infeasible. It
may seem silly to consider constraints that don’t make sense, but that can
88 6. LINEAR OPTIMIZATION

easily happen if we try to impose a complicated set of conditions on an ap-

plied problem. If there are values of the problem variables that satisfy the
constraints, then the problem is feasible. A feasible problem may or may not
have a solution.

Problem T6.3. Can we find the maximum of Z = x where x + y = 7 and

x ≥ 0?

Solution. The problem is feasible, since x = 0, y = 7 satisfies the constraints.

But the problem has no solution, since the objective Z has no maximum. We
can make Z = x as large as we wish and have y = 7 − x to keep the constraints
happy. (Note that we are allowed to let y be negative.)
A feasible problem with a solution may have more than one solution.

Problem T6.4. Minimize Z = −3 · x − 4 · y where 6 · x + 8 · y ≤ 48 and

x, y ≥ 0.

Solution. As before, we graph the constraints. Remembering that x, y ≥ 0,

the line 6 · x + 8 · y = 9 runs between (0, 9/8) and (3/2, 0). The value of Z is
−9/2 at both the corner points, and it is 0 at (0, 0). Thus, the minimum of Z
is −9/2, occurring at (0, 9/8) and (3/2, 0). Actually, Z = −9/2 at all points
on the line segment joining (0, 9/8) and (3/2, 0).

2. More Complicated Examples

Typical optimization problems involve many variables and constraints, and

so we cannot hope to solve general problems by graphing. We hope that graph-
ing simple problems gives you some idea what to expect in general. Moving
to more realistic problems, we will emphasize setting up problems, and then
we will use Excel to find numerical answers.
We recommend the book [12], because it is an excellent source for a wide
array of application problems and because it teaches an overall approach to
 2. MORE COMPLICATED EXAMPLES 89

linear problems. The author, George Dantzig, discovered a fundamental nu-

merical algorithm for solving linear optimization problems, and his algorithm
is the engine of many software solvers – for instance, the Excel Solver. We
will use the Solver to obtain numerical answers; in class we will demonstrate
this.

2.1. Mixing. There are a host of linear optimization problems that in-
volve mixing ingredients together to make something: choosing foods to con-
struct a diet, choosing metals to manufacture an alloy, etc. In the problems
we consider there are many ways to mix the ingredients – that’s probably a
surprising feature, since you are used to situations where there is one recipe
for the desired mix. (e.g. There is only one way to make a water molecule out
of hydrogen and oxygen.)
Problem T6.5. We manufacture four models of desks.2 Each desk is con-
structed in the carpentry shop and then sent to the finishing shop. The num-
ber of hours of labor required for each model at each stage is given in the table
below. We also list the per desk profit realized from the sale of each model.
How many desks of each type should we make to maximize profit, if we have
up to 6000 hours available for carpentry and 4000 for finishing? (Note: the
mix here is the mix of numbers of models scheduled to be constructed.)
model 1 model 2 model 3 model 4
carpentry 4 9 7 10
finishing 1.5 1.5 3 40
unit profit 12 20 18 40
Solution. The objective is profit Q, and it is a function of the number of each
desk model we make. Thus, we have four variables: M1 , M2 , M3 , M4 giving us
the number of each model made. Then profit is easily computed:
Q = 12 · M1 + 20 · M2 + 18 · M3 + 40 · M4
2Modified from an example in [12, p.50]
90 6. LINEAR OPTIMIZATION

Notice that we know the objective profit Q as a function – we don’t have

specific numerical values of Q.
Constraints? The variables are non-negative. We have up to 6000 hours of
carpentry labor; reading the table we see the number of hours for each model:
4 · M1 + 9 · M2 + 7 · M3 + 10 · M4 ≤ 6000
We pause to mention that inequality constraints are very common in optimiza-
tion problems. We might assume that if we want to maximize profit, we will
need to use all 6000 hours here, and so we would actually have an equation
rather than an inequality. That may be the case; the solution will disclose the
situation. When a problem has many constraints, they may not all be able to
produce equality in every inequality.
The finishing hours constraint is similar to the constraint for carpentry.
1.5 · M1 + 1.5 · M2 + 3 · M3 + 40 · M4 ≤ 4000
The Excel Solver gives this solution to maximize Q:
M1 ≈ 1379, M2 ≈ 0, M3 ≈ 0, M4 ≈ 42.28, Q ≈ 18482.76

We might wonder about the decimal answer for M4 in the last problem.
Can we make fractional numbers of desks? In general, it is a significant compli-
cation to require integer values in an optimization problem. The Solver allows
us to constrain some or all of the variables to be integers, and, if the problem
is not too large, then the Solver will find the constrained solution. However,
for larger problems the Solver can work for a long time without getting a
solution, and there’s no way to tell in advance how long it will take. When
we constrain the desk problem variables to be integers, we get M1 = 1380 and
M4 = 48 and M2 = M3 = 0, and the maximum of Q is 18480. Note that
this solution is not obtained merely by rounding the decimal values in the first
solution – that just does not work in general. In this text, we do not want to
 2. MORE COMPLICATED EXAMPLES 91

introduce the many nuances possible when variables have to be integers, and
so we will allow decimal answers without worrying. But you should be aware
that there is a major issue here. We also remind you that we are dealing with
models, and the answer, decimal or integer, might be an approximation in any
case.

2.2. Transportation. We imagine a single good – say pianos. We have

a bunch of warehouses, with varying numbers of pianos in each. And we have
store locations from which various numbers of pianos will be sold. We want
to ship pianos from the warehouses to the stores at minimal cost.
Let’s suppose we have 3 warehouses and 4 stores, with unit shipping costs
from each warehouse to each store indicated in the following table.3 The last
column of the table shows how many pianos must be shipped to each store.
The bottom row of the table gives the supply of pianos at each warehouse.
warehouse: A B C demand
store 1: 10 50 20 60
store 2: 5 10 30 70
store 3: 20 10 6 120
store 4: 30 30 30 100
supply: 100 200 80 –
The key to this problem is to realize that we need 12 variables: one variable
for each warehouse, store pair; the value of the variable will give the number
of pianos shipped from the particular warehouse to the particular store. Let
X[A1] be the number shipped from warehouse A to store 1; let X[A2] be from
warehouse A to store 2, and so on. The objective is the cost Z:

Z = 10 · X[A1] + 5 · X[A2] + · · · + 6 · X[C3] + 30 · X[C4]

The coefficients come from the cost table. We have omitted some of the terms;
there are 12 of them altogether!
3Shipping costs can vary due to many factors such as distance.
92 6. LINEAR OPTIMIZATION

As for constraints, each store needs to receive the correct number of pianos.
Store 1 needs 60 pianos, and so

X[A1] + X[B1] + X[C1] = 60

There is a similar constraint for the other three stores. Also, each warehouse
can send no more pianos than it has in stock. Warehouse A has 100 pianos,
and so

X[A1] + X[A2] + X[A3] + X[A4] ≤ 100

Notice the inquality here. There are 380 pianos, total, at the warehouses, and
the stores want a total of 350, so not all pianos will be shipped. We get three
warehouse constraints. Finally, all the variables are non-negative, since we
cannot ship negative numbers of pianos!
In the Solver, the variables will be laid out in a 4 × 3 table, corresponding
to the shipping cost table, as we will see in class. The minimum cost is
Z = 4980.

2.3. Time Sequence. We consider a problem in which there is a sequence

of variable values in time. For instance, each week we produce a certain amount
of a quantity. The problem we consider is a step toward more realistic (larger!)
problems.

Problem T6.6. During a given week, over a four week period, we have workers
and trainees. It takes one week to turn a trainee into a worker. Workers are
divided into producers (making a product) and those who are idle. Here is a
table of revenue produced for each producer, the cost per idle worker, and the
cost per trainee. There is a maximum number of producers for each week in
the table, as well. Assuming we begin with 10 workers, figure out how many
producers, idle, and trainees we need each week to maximize total revenue.
 2. MORE COMPLICATED EXAMPLES 93

week: 1 2 3 4
producer (unit revenue) 12 10 10 12
idle (unit cost) 5 4 2 4
trainee (unit cost) 6 7 4 -
maximum producers 8 30 25 40
Solution. As we have been advocating, we turn everything into a variable!
We’ll have workers Wn during weeks 1, 2, 3, 4, and trainees Tn during weeks
1, 2, 3. (Because there are only four weeks, it doesn’t make sense to pay for
trainees during week 4.) We’ll have producers Pn for weeks 1, 2, 3, 4 and idle
In for weeks 1, 2, 3, 4. The objective R is revenue: production minus costs.
R = 12 · P1 + 10 · P2 + 10 · P3 + 12 · P4
− 5 · I1 − 4 · I2 − 2 · I3 − 4 · I4
− 6 · T1 − 7 · T2 − 4 · T3
Even though this is complicated, it is just a sumproduct expression in Excel.
We want to maximize R as a function of all these variables.
For constraints, we just read through the problem carefully. Trainees turn
into workers the next week. Thus,
W2 = W1 + T1 W3 = W2 + T2 W4 = W3 + T3
These equations are called time sequence equations, since they show how the
variables step through the time intervals. Next, workers are divided into pro-
ducers and idle.
W1 = P1 + I1 W2 = P2 + I2
W3 = P3 + I3 W4 = P4 + I4
These equations are sometimes called material balance equations, since they
show how various groups are divided up – they provide a kind of accounting
of the workers. It is not crucial to use the terms times sequence equations
and material balance equations, but we think they help in setting up the more
94 6. LINEAR OPTIMIZATION

complicated problems. Time sequence equations show how to get from one
time step to the next one; material balance equations show how items are
categorized at a single time step.
There is a maximum number of producers each week:
P1 ≤ 8 P2 ≤ 30 P3 ≤ 25 P4 ≤ 40
And we begin with W1 = 10. As usual, all the variables are non-negative, and
we have our problem set-up!
In class, we will describe using Excel to solve this problem. Here is a table
showing numbers of workers each week.
week: 1 2 3 4
workers: 10 30 30 40
Our solution shows that we will be training workers during weeks 1 and 3, and
we will have idle workers during those weeks, as well.

3. Shadow Prices – Lagrange Multipliers

Linear optimization problems contain a multitude of coefficients; some of

these values may be negotiable. Let’s go back to the first problem worked in
this chapter and express one of the coefficients as a parameter that can be
changed.
Problem T6.1 – revisited. Find the maximum of Z = 6 · x + 7 · y, where
3 · x + 5 · y ≤ A and 2 · x + y ≤ 4 and x, y ≥ 0. The parameter A starts at 15,
but it might change; what happens to the maximum value of Z in that case?

Solution. We have replaced the right side 15 in the first constraint by a

variable right side A. Currently A = 15, but we want to know what will
happen if A changes. Notice that the objective Z is a function of the problem
variables x, y; it doesn’t mention A at all. The change in A doesn’t affect the
objective Z, it affects the maximum of Z. Let’s see this.
 3. SHADOW PRICES – LAGRANGE MULTIPLIERS 95

The solution to this problem occurs where the two lines meet; we solve
those equations using our parameter A.
20 − A 2 · A − 12
3 · x + 5 · y = A and 2 · x + y = 4 gives x = ,y=
7 7
This solution gives maximum objective
8 · A + 36
Z̄ =
7
We have used the notation Z̄ to distinguish the maximum value of the objective
from Z as a function of x, y. Notice that Z̄ is a function of the parameter A.
When A = 15, we get the solution originally obtained: x = 5/7 and y = 18/7
and Z̄ = 156/7.
It is easy to use our formula for Z̄ to get information about ∆Z̄. Since Z̄
is a linear function of the parameter A, we see that
∆Z̄ dZ̄ 8
= =
∆A dA 7
If A increases from 15 to 16, so that ∆A = 1, then
8 dZ̄ ∆Z̄
= = = ∆Z̄
7 dA ∆A
We expect the maximum objective to increase by 8/7.
Let’s call attention to the shift in point of view here. While we are working
the optimization problem the objective Z is a function of the problem variables
x, y, and A is the constant 15. In other words, A is fixed while we do the
optimization problem. Once we do the problem, the maximum objective Z̄ is
a function of the parameter A, without the problem variables.
The derivative dZ̄/dA = 8/7 is a special case of the derivative of the
optimal objective with respect to a parameter. This derivative is identified
by several technical terms, for it arises in several contexts. Mathematicians
and practitioners of some other disciplines call it a Lagrange multiplier . In
economics problems, it may be called a shadow price – we will explain that
term momentarily. People also say that the derivative measures the sensitivity
96 6. LINEAR OPTIMIZATION

of the maximum with respect to the parameter A. We will use the term
Lagrange multiplier, while nodding to other terms as they make sense.
Here is a typical applied problem.
Problem T6.7. We make (cheap) chairs and tables. It costs $10 to make
each chair, and $50 to make each table, and we can spend up to $500 to make
them. Each chair takes 3 hours of labor and each table 2 hours; we have 60
hours of labor available. We get $22 profit from each chair and $35 from each
table. How many chairs and tables should we make to maximize profit? What
would it be worth to us to increase the available hours of labor by 1?
Solution. Let x be the number of chairs we make, and let y be the number
of tables. Cost: 10 · x + 50 · y ≤ 500. Labor: 3 · x + 2 · y ≤ 60. Also, x, y ≥ 0.
The objective is profit P = 22 · x + 35 · y.
Let’s think about the second question: the question about what happens if
the amount of labor changes. The amount of labor L is a parameter, currently
60. We see that the objective profit P is a function of the variables x, y;
the profit does not mention the parameter L. The maximum value P̄ of P ,
however, will be a function of L. The Lagrange multiplier would be dP̄ /dL.
The Solver finds the solution
x = 15.4, y = 6.9, so that P̄ = 580.8
When the Solver reports that it has found a solution, the right side of the
dialogue box shows three options: Answer and Sensitivity and Limits. If
you select Sensitivity, then a new worksheet will be produced that shows
the Lagrange Multipliers (or it may call them shadow prices) – one multiplier
for each constraint. Write L for the amount of labor – currently 60 hours. The
Lagrange Multiplier for L goes with the constraint 3 · x + 2 · y ≤ L. Excel
calculated the Lagrange multiplier for the labor constraint to be 5.8. Thus,
dP̄
= 5.8
dL
 3. SHADOW PRICES – LAGRANGE MULTIPLIERS 97

An additional hour of labor is ∆L = 1. Then ∆P̄ ≈ 5.8; an additional hour

of labor brings in an additional profit around $5.80. That’s how much the
additional hour of labor is worth to us.
Here is a decision-making problem arising from what we just did: suppose
we could increase the available labor hours by one hour for only $5.00. The
shadow price (Lagrange Multiplier) estimates an increase of $5.80 in profit, so
we come out ahead by 80/c. Looks like a good deal.
Notice the units of the Lagrange Multiplier in the previous problem. Maxi-
mum profit P̄ is in dollars, and L is in hours (of labor). Thus, dP̄ /dL has units
dollars per hour. Those units are a price – the price of an additional hour of
labor in this problem. Because the units of the Lagrange Multiplier are price,
the multiplier is called a shadow price. Some people use the term shadow price
for every Lagrange Multiplier because of the prevalence of problems in which
we are thinking about the cost of changing some parameter. Notice that in the
language of marginals, the shadow price in the present problem is the marginal
maximum profit of labor hours.
You will do several problems that use Lagrange Multipliers to evaluate the
cost of resources to make decisions concerning them.
 CHAPTER 7

The Integral

1. Antiderivatives

Recall that a function that has a derivative is called differentiable. All

functions that have algebraic formulas, or that involve the exponential function
or the natural logarithm, are differentiable on every open interval where they
are defined. Not all functions are differentiable – this is a difficult and technical
question that we will not address in general.
We have worked extensively with the problem of finding the derivative
0
f (x) for given functions f (x). Now we want to reverse that process, for there
are many situations in which the derivative is given, and we want to find the
function. For instance, we might know the velocity v(t) of an object moving
on the x-axis and be seeking the position x(t). The notation g(x) = f 0 (x) says
that g(x) is the derivative of f (x); to reverse the description we say that f (x)
is an antiderivative of g(x). For instance, since
 2 0 1
3x − ln(x) = 6x −
x
2
we have that 3x − ln(x) is an antiderivative of 6x − 1/x.
The derivative of a constant is 0. So, if we add a constant k to the anti-
derivative:
3x2 − ln(x) + k
then we get the same derivative
 2 0 1
3x − ln(x) + k = 6x − + 0
x
99
100 7. THE INTEGRAL

Taking k = 6 as a specific example, both 3x2 − ln(x) and 3x2 − ln(x) + 6 are
antiderivatives of 6x − 1/x.
It turns out that all the antiderivatives of a given function can be described
in this way: find a particular antiderivative, and the rest of them are obtained
by adding constants.1 Thus, since (x3 )0 = 3x2 , we have
If f 0 (x) = 3x2 then f (x) = x3 + k for some constant k
We introduce the most common notation for the antiderivative. We will
concentrate on using this notation; it will be explained more fully in the next
section. We write Z
f (x) · dx
R
for an antiderivative of f (x), also called the indefinite integral of f (x). The
is the integral sign; it originated as a German S, standing for summe, recog-
nizable as the the German word for sum. The function f (x) is the integrand ,
and the differential dx indicates the variable. Since (x2 )0 = 2 · x, we write
Z
2
(7.1) x = 2 · x · dx

We have (x2 + 3)0 = 2 · x, and so

Z
2
(7.2) x +3= 2 · x · dx

As we saw before, different functions can have the same derivative; that’s why
the indefinite integral is indefinite! We will see that ambiguity is an advantage
in this setting. Some texts write
Z
2 · x · dx = x2 + k where k is a constant

This point of view regards the indefinite integral as standing for all antideriva-
tives of 2 · x simultaneously. Our preference is to use the indefinite integral
1Thisfact is a consequence of the Mean Value Theorem. You might look up a statement
of that result in a standard calculus text. When the Mean Value Theorem is expressed in
terms of the secant line on a graph, it seems obvious to most people.
 1. ANTIDERIVATIVES 101

to stand for some antiderivative rather than all of them. The thing to remem-
ber is that if you have one antiderivative of a function, you can get the rest of
them by adding constants.
Each rule for calculating derivatives can be turned into a rule for calculating
antiderivatives. We will begin with the power rule.

Inverse Power Rule Let n be a number with n 6= −1. Then

xn+1
Z
xn · dx =
n+1

Explanation. We use the Power Rule for derivatives to compute

 n+1 0
x xn
= (n + 1) · = xn
n+1 n+1
and this proves that our Inverse Power Rule is correct.
We disallow n = −1 in the Inverse Power Rule for the simple reason that
when n = −1, the n + 1 in the denominator would be 0. The case n = −1 has
its own formula. Remember that x−1 = 1/x.

Logarithm Rule We have

Z
dx
= ln |x| when x 6= 0
x

Explanation. Remember that ln(a) is defined only when a > 0. When x > 0,
we know that
Z
0 1 dx
ln (x) = and so ln(x) = when x > 0
x x
We will need to have an antiderivative for 1/x in the case that x < 0, as well.
In that case, notice that −x > 0, and so ln(−x) is defined. Use the Chain
Rule to compute:
0 1 1
ln(−x) = ln0 (−x) · (−x)0 =

· (−1) =
−x x
102 7. THE INTEGRAL

Thus,
Z
dx
ln(−x) = when x < 0
x
When −x < 0, we can write −x = |x|; when x > 0, we have x = |x|, so we
have the Logarithm Rule in both cases.
Don’t forget the absolute value in ln |x|. We will need to allow x to be
negative, and the logarithm of a negative is undefined, but |x| is always positive
if x 6= 0, and so ln |x| is defined.
The exponential function is its own derivative, and so it is its own anti-
derivative.

Exponential Rule We have

Z
ex · dx = ex

Recall the Addition Rule and Constant Multiple Rule for derivatives:
0 0
f (x) + g(x) = f 0 (x) + g 0 (x) and c · f (x) = c · f 0 (x)
 

where f (x) and g(x) are differentiable, and c is a constant. It follows that
we have an Addition Rule and a Constant Multiple Rule for the indefinite
integral.

Addition Rule, Constant Multiple Rule If f (x) and g(x) are differen-
tiable, and if c is a constant, then
Z Z Z
 
(7.3) f (x) + g(x) · dx = f (x) · dx + g(x) · dx

and
Z Z
(7.4) c · f (x) · dx = c · f (x) · dx

Here is a sample use.

 1. ANTIDERIVATIVES 103

Example. We compute
Z   Z Z
x 5 x 5
2 · e + 3 · dx = 2 · e · dx + · dx addition rule
x x3
Z Z
= 2 · e · dx + 5 · x−3 · dx
x
constant multiple

x−2
= 2 · ·ex + 5 · exp, inverse power
−2
5
= 2 · ex −
2x2

In the following problem, pay attention to the use of the unknown constant
K.

Problem T7.1. A steel bar 2 feet long has density 10x2 − x3 + 2ex pounds
per foot at the point x feet from one end of the bar. What is the weight of
the bar?

Solution. Measure density and weight W (x) from the left hand end of bar.
Then W 0 (x) is the density, so that W 0 (x) = 10x2 − x3 + 2ex , and so
Z
10x2 − x3 + 2ex · dx
 
W (x) =

Let’s use the various rules on the right side, and then we’ll bring W (x) back
into the picture. Make sure you understand which rules are being used in each
line of the following.
Z
10x2 −x3 + 2ex · dx
 

Z Z Z
= 10x · dx − x · dx + 2ex · dx
2 3

Z Z Z
= 10 · x · dx − x · dx + 2 · ex · dx
2 3

x3 x4 10 1
= 10 · − + 2 · ex = x3 − x4 + 2ex
3 4 3 4
104 7. THE INTEGRAL

The function we ended up with is an antiderivative of the density, and W (x)

is an antiderivative of the density. It follows that
10 1
(7.5) W (x) = x3 − x4 + 2ex + K
3 4
for some constant K. We can figure out what K is by plugging in any particular
value of x. Go back to the definition of W (x) to see that W (0) = 0. Let x = 0
in (7.5) and we have

0=2+K so that −2=K

This shows that

10 3 1 4
W (x) = x − x + 2ex − 2
3 4
Finally, the weight of the bar is W (2):
10 1
W (2) = · 8 − · 16 + 2e2 − 2 ≈ 35.44 pounds
3 4

In the previous problem, we used the antiderivative rules to find the indef-
inite integral of the density. The weight function is a definite antiderivative.
The constant K synchronizes the indefinite antiderivative to the definite one
of the problem.
Sometimes we are asked to verify an antiderivative formula: that’s just
asking us to take the derivative.

Problem T7.2. Show that

Z
x
x + 7 · ln |x − 7| = · dx for x 6= 7
x−7

Solution. As we said, we are just being asked to take the derivative; we need
to remember that (ln |x|)0 = 1/x no matter whether x > 0 and x < 0.
Compute
 0 7
x + 7 · ln |x − 7| = 1 +
x−7
 1. ANTIDERIVATIVES 105

We get the integrand with a little algebra.

7 x−7 7 x
1+ = + =
x−7 x−7 x−7 x−7
as needed.
There is an antiderivative rule that corresponds to the Chain Rule. While
we will not pursue that correspondence in general, we will use the following
special case.
Linear Inside Rule Suppose that a, b are constants with a 6= 0, and suppose
that f 0 (x) is differentiable. Then
Z
1
f 0 (ax + b) · dx = · f (ax + b)
a
Explanation Take the derivative of the right side, using the Chain Rule.
 0
1 1 1
· f (ax + b) = · f 0 (ax + b) · (ax + b)0 = · f 0 (ax + b) · a = f 0 (ax + b)
a a a

The Linear Inside Rule is somewhat abstract. If we ignore the linear term
ax + b, we see the pattern f 0 (x) · dx = f (x). The Rule says to imitate this
R

pattern, carrying the ax + b along the way we do in the Chain Rule. But we
modify the antiderivative with the fraction 1/a, dividing by the slope of the
linear term ax + b. Let’s see some uses of this rule.
Problem T7.3. Compute each of the following.
Z Z Z p
4
(a) (3x − 2) · dx (b) exp(−2t) · dt (c) 3y + 1 · dy

Solution. For (a), we use that

x5
Z
x4 · dx =
5
In the notation of the Linear Inside Rule, f 0 (x) = x4 . Notice that the integral
in (a) involves the fourth power function with 3x − 2 inside – that’s the linear
106 7. THE INTEGRAL

inside, the ax + b expression. We compute the antiderivative of (3x − 2)4 by

using the antiderivative of the fourth power, dividing by the slope a = 3.
1 (3x − 2)5
Z
(3x − 2)4 · dx = ·
3 5
For (b), the linear inside on exp(−2t) is −2t, which has slope a = −2.
Thus,
Z
1
exp(−2t) · dt = · exp(−2t)
−2
For (c), we treat the square root the way we did the fourth power in (a):
1 (3y + 1)3/2
Z p Z
2
3y + 1 · dy = (3y + 1)1/2 · dy = · = · (3y + 1)3/2
3 3/2 9

For many integrals, there is more than one way to get the answer. For
instance, consider
Z
(5 − 2 · t) · dt

The most direct approach is to use the rules for polynomials:

t2
Z
(5 − 3 · t) · dt = 5 · t − 2 · = 5 · t − t2
2
There is another way! We see that (5−2·t)1 is linear inside the first power,
and so, remembering to divide by the slope
1 (5 − 2 · t)2
Z
1
(5 − 3 · t) · dt = = − · (5 − 2 · t)2
−2 2 4
It is instructive to compare the two answers. Multiplying out the second one:
1 1 25
− · (5 − 2 · t)2 = − · (25 − 20 · t + 4 · t2 ) = − + 5 · t − t2
4 4 4
This is the first answer, with a constant added. We can always add a constant
to an antiderivative to get another one.
 2. THE DEFINITE INTEGRAL 107

2. The Definite Integral

We introduce the definite integral , one of the most important and com-
monly occurring mathematical objects in all of science. We will discuss the
definition and computation of the integral, and we will give the two main ways
it occurs in applications.
There are many approaches to this material. We want to start with the
computation of the definite integral. That approach can be presented by build-
ing on the indefinite integral studied in the previous section.
First, we remind you about the closed interval notation. For instance,
[1, 4] stands for the set of all numbers x with 1 ≤ x ≤ 4. In general, if a, b are
numbers with a ≤ b, the closed interval [a, b] is the set of all numbers x with
a ≤ x ≤ b.
Second, we introduce notation that will help us evaluate functions at var-
ious values. Given F (x) and numbers a, b, we define
b
F (x) = F (b) − F (a)

a

Plug in the top number first, then subtract what you get when you plug in
the bottom number. Thus,
3 2
x  = 32 − 22 = 5 and
2
exp(x/2) = exp(1) − exp(0) = e − 1

2 0
b
You might recognize that F (x) is ∆F . Hold that thought.

a
You have seen the indefinite integral
Z
f (x) · dx

as a notation for an antiderivative of f (x). If f (x) is defined for every value of

x on a closed interval [a, b], the definite integral of f (x) over [a, b] is written
Z b
f (x) · dx
a
108 7. THE INTEGRAL

As in the indefinite integral, the function f (x) is called the integrand and
the differential dx indicates the variable x for which f (x) is a function. The
number a is the lower limit of integration, and the number b is the upper limit
of integration.
We have many things to say about the definite integral; we begin with how
it is computed in most cases. First, find an antiderivative2 F (x) for f (x) on
[a, b]. Then,
Z b b
f (x) · dx = F (x) = F (b) − F (a)

a a

For instance, the Inverse Power Rule says that

Z
1
(7.6) x2 · dx = · x3
3
And we compute the definite integral:
Z 4
1 4 1 1 56
x2 · dx = · x3  = · 43 − · 23 =

2 3 2 3 3 3
We know that x2 has other antiderivatives. For instance,
Z
1 3
· x + 5 = x2 · dx
3
What if we had used this antiderivative rather than just x3 /3?
Z 4  
2 1 3 4
x · dx = ·x +5 
2 3 2
   
1 3 1 3
= ·4 +5 − ·2 +5
3 3
64 8 56
= +5− −5=
3 3 3
2To get very technical, the mere existence of an antiderivative for f (x) is not enough
to ensure that the definite integral exists. If the function f (x) is differentiable, however,
then the definite integral will exist, as will an antiderivative for f (x), and our linking of the
antiderivative and the definite integral will be correct. All of the functions we will consider
are differentiable.
 2. THE DEFINITE INTEGRAL 109

We see that the constant 5 cancels! This is what happens in general, for
suppose we have an antiderivative F (x) for f (x), so that F 0 (x) = f (x). In
Section 1 we said that any other antiderivative for f (x) is of the form F (x)+K
where K is a constant. Using F (x) + K in place of F (x) in the definition of
the definite integral we get
Z b b
f (x) · dx = (F (x) + K)

a a

= (F (b) + K) − (F (a) + K)
= F (b) + K − F (a) − K = F (b) − F (a)

Thus, we compute the same value for the definite integral no matter which
antiderivative is used.
We see that the value of the definite integral is ∆F , where F 0 (x) = f (x)
is the integrand. That leads to many applications. Here is an example.

Problem T7.4. An object is moving along the x-axis with velocity v(t) =
5 − 2 · t2 , for t ≥ 0. What is the displacement as t goes from 1 to 3?

Solution. Recall that displacement is ∆x where x is the position. Write x(t)

to remind us that x is a function of t. Since t goes from 1 to 3, we are asking
for
∆x = x(3) − x(1)

We know that velocity v(t) = x0 (t), and so x(t) is an antiderivative of v(t).

Thus,
3 Z 3 Z 3
x(3) − x(1) = x(t) = v(t) · dt = (5 − 2 · t2 ) · dt

1 1 1
t3 3  2 2 22
= 5 · t − 2 ·  = 5 · 3 − · 33 − 5 · 1 −
 
=−
3 1 3 3 3
110 7. THE INTEGRAL

There is a subtlety in the solution to the previous problem. Our calculation

of the definite integral used the antiderivative
t3
Z
(7.7) 5 · t − 2 · = (5 − 2 · t2 ) · dt
3
The integrand is velocity, and x(t) is an antiderivative of velocity. Does that
mean that x is the left side of (7.7)? Not necessarily! It would be the case
that
t3
x(t) = 5 · t − 2 ·
+K
3
for some constant K, but the constant does not affect the calculation of the
definite integral. The antiderivative we used and the antiderivative x(t) give
the same answer.
Here are some basic properties of the definite integral. Throughout, each
integrand is assumed to be defined and differentiable between the limits of
integration. The first two properties say that the integral is linear, that inte-
gration distributes through sums and across constants. These identities follow
directly from the corresponding identities (7.3) and (7.4) for indefinite inte-
grals, as we will point out in class. The third property shows how to chain
integrals together. These properties are fairly natural, and we will usually use
them without calling attention.
Z b Z b Z b
(7.8) (f (x) + g(x)) · dx = f (x) · dx + g(x) · dx
a a a
Z b Z b
(7.9) C · f (x) · dx = C · f (x) · dx
a a
where C is constant
Z b Z c Z c
(7.10) f (x) · dx + f (x) · dx = f (x) · dx
a b a
 3. RIEMANN SUMS 111

We will verify (7.10); let F 0 (x) = f (x).

Z b Z c b c
f (x) · dx + f (x) · dx = F (x) + F (x)
 
a b a b

= F (b) − F (a) + F (c) − F (b)

Z c
= F (c) − F (a) = f (x) · dx
a

and (7.10) holds.

3. Riemann Sums

We have computed the definite integral from an antiderivative for the inte-
grand. There are many functions for which there is no obvious formula for the
antiderivative – for instance, integrals involving the function exp(−x2 ) come
up often in statistics. How are we to compute the definite integral when we
can’t find an antiderivative? If you accept the statement we made to the effect
that the integral occurs often, you might find this question compelling.
There is another problem with our approach to the integral. Many appli-
cations of the integral do not involve an antiderivative in any obvious way.3
The link from those applications to the integral involves looking at the inte-
gral differently. It turns out that this alternative point of view answers the
question posed in the last paragraph about the existence of antiderivatives.
Let’s introduce this new idea via a very specific example. We need a
function defined on a closed interval; let’s use x2 − 3 on [1, 3]. Now we divide
the interval [1, 3] into some number of equal subintervals. Let’s divide it into
5 subintervals; the length of [1, 3] is 3 − 1 = 2, and so each of the 5 pieces will
have length 2/5 = 0.4. Here are the subintervals.

1 1.4 1.8 2.2 2.6 3

3For instance, see the application example on p.139 of Chapter 8.

112 7. THE INTEGRAL

Next we add up the function values at the right hand endpoint of each
subinterval. Our function was x2 − 3, and here are the values:
x: 1.4 1.8 2.2 2.6 3
x2 − 3 : −1.04 0.24 1.84 3.76 6
And their sum:
−1.04 + 0.24 + 1.84 + 3.76 + 6 = 10.8

Finally, we multiply by the width of the subintervals; in this case, the width
is 0.4. The product of the sum and width is called a Riemann sum.

10.8 · 0.4 = 4.32

A single Riemann sum is not impressive, and the calculation of specific ex-
amples dissolves into arithmetic. What is interesting is what happens as the
number of intervals used gets larger and larger. Here are the Riemann sums
for x2 − 3 on [1, 3] when we divide up the interval into n pieces for various n.
(In class we will explain how the sums were computed; the answers have been
rounded.)
n: 5 10 100 1000 106
Riemann sum : 4.32 3.48 2.75 2.675 2.667
Our table shows only a few sums, but as the number n of subintervals gets
larger and larger, the Riemann sums get closer and closer to (about) 2.667.
That number turns out to be the value of the definite integral using the same
function x2 − 3 and the same interval [1, 3]:
Z 3
x3
 
2
3 8 8
(x − 3) · dx = − 3 · x = 0 − − = ≈ 2.67

1 3 1 3 3
It turns out that Riemann sums can be made close to the definite integral by
using more and more subintervals.
Let’s describe what we just did more generally. We are given a function
f (x) defined on a closed interval [a, b]. Given a positive integer n, there is a
 3. RIEMANN SUMS 113

Riemann sum which we will denote4 R(f (x), a, b, n). The sum is formed by
dividing the interval [a, b] into n equal subintervals, each of width (b − a)/n.
Denote the endpoints of the subintervals like this.
(7.11) a = x0 < x1 < x2 < · · · < xn−1 < xn = b
We write a = x0 and b = xn to make the notation for the subintervals more
uniform.5 Here are the subintervals:
[x0 , x1 ], [x1 , x2 ], [x2 , x3 ], · · · [xn−1 , xn ]
Here is the Riemann sum:
b−a h i
(7.12) R(f (x), a, b, n) = · f (x1 ) + f (x2 ) + f (x3 ) + · · · + f (xn )
n
If f (x) is differentiable, then it turns out that as n gets larger and larger,
the sums approach a specific number; this number is the definite integral. In
notation
Z b
(7.13) lim R(f (x), a, b, n) = f (x) · dx
n→∞ a
The notation n → ∞ simply means that n gets larger and larger, without
bound. We are saying that the larger n gets, the closer the Riemann sums get
to the definite integral.
In our Riemann sums, the terms f (xj ) evaluate the function f (x) at the
right endpoint of each of the subintervals into which the main interval is par-
titioned. It is just as acceptable to use the left endpoints, or, in fact, any
point in each of the subintervals. The points can be chosen systematically or
randomly, as long as there is a term f (cj ) for some point cj in the j-th interval.
Computer programs that calculate Riemann sums numerically use a variety of
schemes. And those schemes are important, since one use of (7.13) is to give
4There is no universally used notation for Riemann sums. The notation we introduce
will be used only in this course.
5We won’t need a precise formula for the x , although, in case you are interested,
j
xj = a + j · (b − a)/n.
114 7. THE INTEGRAL

numerical approximations to definite integrals. We will provide a web function

for that purpose.

Problem T7.5. Demonstrate (7.13) when f (x) = 2 · x + 5 on [0, 6].

Solution. We need to pick various n and form the Riemann sum (7.12). The
problem doesn’t tell us how to do this. We’ll start with n = 8 and compute
the sum by hand. When the interval [0, 6] is divided into 8 subintervals, each
has width 6/8 = 0.75. Here are the endpoints, as in (7.11).

0 < 0.75 < 1.5 < 2.25 < 3 < 3.75 < 4.5 < 5.25 < 6

The Riemann sum (7.12) involves these function values:

x : 0.75 1.5 2.25 3 3.75 4.5 5.25 6
2 · x + 5 : 6.5 8 9.5 11 12.5 14 15.5 17
We need to add up the function values and multiply by the width 0.75.
h i
R(2x + 5, 0, 6, 8) = 0.75 · 6.5 + 8 + 9.5 + 11 + 12.5 + 14 + 15.5 + 17
= 70.5

We computed the sums for other values of n numerically.

n: 100 1000 5000 10000 50000
Riemann sum : 66.36 66.036 66.0072 66.0036 66.00072
Numerical evidence will never be absolutely convincing, but it looks as if the
Riemann sums are getting close to 66.
We know how to calculate the definite integral:
Z 6 6
(2 · x + 5) · dx = x2 + 5 · x = 66 − 0 = 66

0 0

Here is an integral for which the approximation is essential, since we don’t

know how to find an antiderivative for the integrand.
 3. RIEMANN SUMS 115

Problem T7.6. Approximate the following integral.

Z 2
x2 · exp(−x2 ) · dx
−2

Solution. We computed Riemann sums for the following values of n.

n: 100 1000 5000
Riemann Sum : 0.84539 0.84545 0.84545
We are pretty confident that we have 5 decimal places for the integral.
There is an important theoretical point here. Riemann sums can be used
to approximate the integral from the integrand – you don’t need to know
an antiderivative. It turns out that Riemann sums can be used to define an
antiderivative for a given differentiable function f (x). Indeed, we fix a left
endpoint a, and then, for each t ≥ a, we define
Z t
F (t) = f (x) · dx
a
where the integral is defined by Riemann sums, using (7.13). It can be shown
that F 0 (x) = f (x); this is that F (x) is an antiderivative for f (x). The point is
that a function that is differentiable on an interval always has an antiderivative
there, whether we can find a nice formula for that antiderivative, or not.
The link between antiderivatives and Riemann sums is called the Fun-
damental Theorem of Calculus. This theorem summarizes our work in this
chapter.

Fundamental Theorem of Calculus. Let f (x) be a differentiable

function on [a, b]. Then the equation (7.13) holds. Furthermore, f (x) has an
antiderivative F (x), defined for all x on the interval [a, b], and
Z b
f (x) · dx = F (b) − F (a)
a
 CHAPTER 8

Interpreting and Using the Integral

1. Anti-Rates

If y is a function of time t, then the rate of change of y is the derivative

dy/dt. If we are given the rate of change g(t) of y, then
Z
y = g(t) · dt

We lightheartedly refer to y as the anti-rate of g(t). Also remember that ∆y

is a definite integral. Both these ideas nurture applications.

Problem T8.1. A population grows at rate 10 · (10 − t) individuals per year,

where t is time in years. Assuming that the population starts at 0, what is
the population after 15 years? When does the population reach 0?

Solution. For the first question, we notice that if the population is P (t), then
the rate of growth is P 0 (t). Thus,

P 0 (t) = 10 · (10 − t)

The population after 15 years is P (15). Since P (0) = 0 (population starts at

0), we see that
Z 15
P (15) = P (15) − P (0) = P 0 (t) · dt
0
And we compute
Z 15 15
t2 15
Z  
0
P (t) · dt = 10 · (10 − t) · dt = 10 · 10 · t −  = 375
0 0 2 0
Thus, P (15) = 375.
117
118 8. INTERPRETING AND USING THE INTEGRAL

To see when the population is 0, we solve for the time t = c.

Z c
t2 c c2
   
0 = 10 · (10 − t) · dt = 10 · 10 · t −  = 10 · 10 · c −
0 2 0 2
Dividing by 10 and rearranging, we have 10c = c2 /2. We see that c = 0 is one
solution; not surprising, since we start with population 0. If c 6= 0, we can
divide by c, and get
c
10 = so that c = 20
2
After 20 years, the population has died off.
To compute P (15) in the previous problem, we used a definite integral –
that worked because we could write
P (15) = P (15) − P (0) = ∆P
Let’s sketch an alternative way to do the same problem: we have P 0 (t) =
10 · (10 − t), and so we can write P (t) as an indefinite integral:
t2
Z  
P (t) = 10 · (10 − t) · dt = 10 · 10 · t − + K = 100 · t − 50 · t2 + K
2
We include the constant K to synchronize the antiderivative we found from the
power rule with P – a specific anti-derivative. Since we know that P (0) = 0,
we can compute K by plugging in t = 0. We get 0 = P (0) = K. Now we have
the more specific formula P (t) = 100 · t + 50 · t. The problem asks for P (15),
and we get that by plugging in t = 15.
Problem T8.2. An income stream is defined by a rate at which money comes
in. Say we receive 1000 exp(−0.03t) $/year for 7 years.1 What is the total
amount of money received?
Solution. If A(t) is the amount of money received up to time t, then A0 (t) is
the rate at which money is received. In other words, A0 (t) = 1000 exp(−0.03t).
1Notice
that the rate is not constant during a year. Thus, this model does not imagine
that we are paid each year, but continually as time goes along. That’s why the units $ per
year indicate a derivative.
 1. ANTI-RATES 119

Also, A(0) = 0, and the total amount received is A(7). Thus,

Z 7
A(7) = A(7) − A(0) = A0 (t) · dt
0
Z 7
1000 7
= 1000 · exp(−0.03t) · dt = · exp(−0.03t)

0 −0.03 0

1000 · exp(−0.03 · 7) 1000 · exp(0)

=− + ≈ $6313.86
0.03 0.03

As with the population problem, we can do the income stream problem

with an indefinite integral instead of a definite one. We may discuss this in
class.
√
Problem T8.3. An object moving along the y-axis has acceleration 3 t. The
initial velocity of the object is -2. What is its displacement over 1 ≤ t ≤ 4?
Solution. If v is the velocity, then dv/dt is the acceleration. Thus
Z √
t3/2
v = 3 t · dt = 3 · + C = 2 · t3/2 + C
3/2
where C is a constant. When t = 0, we have v = −2 (the initial velocity), and
so C = −2.
Writing y(t) to indicate that y is a function of t, we have y 0 (t) = v. The
problem asks for the displacement of y when 1 ≤ t ≤ 4. That displacement is
Z 4
∆y = y(4) − y(1) = y 0 (t) · dt
1
Z 4 Z 4
 3/2 
= v · dt = 2t − 2 · dt
1 1
t5/2 4 4 4
=2· − 2t = · t5/2 − 2t
 
5/2 1 5 1
   
4 5/2 4 94
= ·4 −8 − −2 =
5 5 5
120 8. INTERPRETING AND USING THE INTEGRAL

Problem T8.4. We have an empty cylindrical container of radius 2 inches.

Water is poured in at a rate of 10 − t cubic inches per second, for 0 ≤ t ≤ 10.
What is the height of the water at the end?
Solution. The rate in cubic inches per second is the derivative dV /dt where
V is the volume. Thus, V is an antiderivative of 10 − t. Also, V (0) = 0, since
the container starts empty. At time t = 10, we have
Z 10 Z 10
0
V (10) = V (10) − V (0) = V (t) · dt = (10 − t) · dt
0 0
t2 10 100
= 10t −  = 100 − = 50 cubic inches
2 0 2
The volume of a cylinder satisfies V = π · r2 · h, where r is the radius and h is
the height. We have r = 2 and V = 50, and we can solve for h:
V 50
h= 2
= ≈ 3.98 inches
π·r π·4

There is another way to do the previous problem; we can focus on height

h from the get-go. We have V = π · r2 · h; since r = 2, this is V = 4 · π · h.
Taking the derivative with respect to time t:
dV dh
=4·π·
dt dt
The problem tells us dV /dt:
dh 10 − t dh
10 − t = 4 · π · so that =
dt 4·π dt
We can find h with an antiderivative, just as we found V .

2. Area

Let f (x) be a differentiable and non-negative function on the interval [a, b].
Let R be the plane figure2 bounded by the x-axis, the curve y = f (x), the line
x = a, and the line x = b. Here is a generic picture of R.
2A plane figure is nothing more than a set of points in the plane.
 2. AREA 121

............................................
.............................. ................
............... ............
............ ...........
..
..
..
............
. ..........
..
..
..
......
. .........
........
..
..
.......
. ........
.......
..
..
......
. .......
.....
.
......
.
..
.
R ........
..........
..........
...........
y = f (x)
........... ..
............. ............
.............. ............
................. ..............
......................... .................
.....................................................................................
r r
a b

It turns out that

Z b
(8.1) area of R = f (x) · dx
a
We want to sketch an explanation for this very important formula.
During the early 1600’s Fermat and others were able to calculate the ar-
eas of various plane figures, particularly when the boundary of the figure is
a polynomial function. It was noticed that the formulas obtained seemed to
involve antiderivatives of the boundary functions. See [11, Chapter 4]. New-
ton’s teacher Barrow drew up lectures on this subject – lectures that he turned
over to Newton to complete. Newton developed a much more comprehensive
and succinct theory, and he was able to sketch out a convincing argument
for (8.1). Newton’s idea was a key factor in the development of the general
calculus. Leibnitz developed similar ideas at about the same time as Newton,
and the integral notation we use goes back to Leibnitz. See [11, Chapter 5].
We choose a number z between a and b, and let A(z) be the area between
y = f (x) and the x-axis and between x = a and x = z. Here is a generic
picture.
.......................................................
........................ ...............
.............. ............
............ ..........
..
..
..
...........
. ..........
..
......
. .........
..
.. ........
..
..
..
...... ........
.
....... .......
..
. .......
.....
.
.....
.
..
. ........
..........
.......... y = f (x)
A(z) ...........
...........
.............
..............
..........
............
..............
.................
.......................... ..
..
..
..
..
..
..
..
..
................
.
......................................................................
r r r
a z b
122 8. INTERPRETING AND USING THE INTEGRAL

What Newton realized is that ∆A is itself an area. Here is a picture in the

case that ∆z > 0.

..................................................................................
................
............. ............ (z, f (z))
.
...
.........
.
..
..
..
............
...........
r
..........
.........
........
..
..
.........
. ........
.......
..
..
......
. .......
..... .......
.
.......
.
..
. .........
..........
.......... y = f (x)
A(z) ∆A ...........
............
.............
................
........
...........
..............
.................... ...............
............................................................................................................

r r r r
a z z + ∆z b

Imagining a very thin such area, ∆A is almost a rectangle with base ∆z

and height f (z), for f (z) is the height of the left-hand vertical side of ∆A. In
other words,
∆A
≈ f (z) ∆A ≈ f (z) · ∆z which is
∆z
Newton assumed that if we let ∆z → 0 the approximation becomes an equa-
tion. More modern mathematics can be used to show that this is, in fact, the
case. Thus, we get
(8.2) A0 (z) = f (z)
The derivative of the area function is the boundary function. Notice how the
variable x in f (x) changes to a z when we take the derivative of the area
function. The variable z is a specific value of the variable x.
Let us see how the foregoing leads to (8.1) in a special case.
Problem T8.5. Find the area bounded by the x-axis, the curve y = 1/x, for
1 ≤ x ≤ 3.
Solution. Observe that 1/x is positive for x between 1 and 3. Let A(z) be
the area under y = 1/x, and 1 ≤ x ≤ z. Notice that A(1) = 0, since that’s
the area of a line segment. Also notice that A(3) is the area we want. Thus,
Z 3
(8.3) area = A(3) = A(3) − A(1) = A0 (x) · dx
1
 2. AREA 123

The reason for involving the derivative of A is that (8.2) shows that
1 1
A0 (z) = and so A0 (x) =
z x
changing variable from z to x. Thus, (8.3) shows that
Z 3
1
area = · dx
1 x

and that is the formula (8.1).

So much for theory; how about calculating the area! Now that we have
our integral formula, we have only to compute the integral:
Z 3
1 3
area = · dx = ln |x| = ln(3) − ln(1) = ln(3)

1 x 1

In the solution just done, we emphasized the use of Newton’s function A(z).
The point of that function was to obtain the equation (8.1) and calculate the
area by integration. Going forward, we will use (8.1) without appealing to the
apparatus of Newton’s area function.
Because ln |x| is an antiderivative of 1/x, we see that areas under y = 1/x
are logarithms. Perhaps that is an unexpected interpretation of the logarithm.

Problem T8.6. Write the area of a circle of radius r as an integral, and

compute the integral numerically when r = 1.

Solution. The equation of a circle of radius r with center at (0, 0) is x2 + y 2 =

√
r2 . The top half of the circle is defined by solving for y ≥ 0 to get y = r2 − x2 ,
for −r ≤ x ≤ r. The area between the top half of the circle and the x-axis
(diameter) is
Z r √
r2 − x2 · dx
−r
This is the area of a semi-circle of radius r, and so the area of the circle is
twice this integral.
124 8. INTERPRETING AND USING THE INTEGRAL

Our Riemann sum calculator gives the area when r = 1 as approximately

3.1416. (That’s supposed to be π.)
√
We do not know how to find an antiderivative for r2 − x2 . You have been
told that the area of a circle of radius r is π · r2 . If we define π by the area of
a circle with r = 1, then, as in the previous problem,
Z 1√
π =2· 1 − x2 · dx
−1

Let’s show how this integral leads to the π · r2 formula. Let

Z √
R(x) = 2 · 1 − x2 · dx

Even though we don’t have a formula for R(x), we will use it to good effect.
We have
1
(8.4) π = R(x) = R(1) − R(−1)

−1
For r > 0, use the Chain Rule to compute
d x x 1 r x2 1
R = R0 · = 1− 2 ·
dx r r r r r
r √
r 2 − x2 1 r 2 − x2 1 2 √ 2
=2· · = 2 · · = · r − x2
r2 r r r r2
This shows us that
x Z √
2
r ·R =2· r2 − x2 · dx
r
and so the area of a circle of radius r is this:
Z r√  x  r
2· r2 − x2 · dx = r2 · R

r

−r −r

= r2 · (R(1) − R(−1))
Looking at (8.4), we see that the area of the circle of radius r is π · r2 .
The previous argument involves the abstract existence of an antiderivative
– a consequence of the Fundamental Theorem of Calculus. We will not pursue
this sort of argument very often – there will be one more use in this chapter
 2. AREA 125

– but we need to say that this kind of calculation has some important uses in
applied mathematics.
We computed the area between the graph of a function and the x-axis,
assuming that the function was above the x-axis. More generally, we can
compute the area between y = f (x) and y = g(x) for a ≤ x ≤ b, assuming
that f (x) ≥ g(x).

Area between cuves If f (x) ≥ g(x) for a ≤ x ≤ b, then the area between
y = f (x) and y = g(x) on this interval is
Z b
(f (x) − g(x)) · dx
a

Explanation We can add a constant c to g(x) so that g(x) + c > 0; that puts
y = g(x) + c above the x-axis. Since f (x) ≥ g(x), it follows that y = f (x) + c
is also above the x-axis. The area between y = f (x) and y = g(x) is the same
as the area between y = f (x) + c and y = g(x) + c, since the latter area is just
the former area moved up in the plane.
If you draw a generic picture, you will see that the area between y = f (x)+c
and y = g(x) + c is the difference between the area under y = f (x) + c and the
x-axis over the interval [a, b] and the area under y = g(x) + c and the x-axis
over [a, b]. Thus, the area we are trying to compute is this:
Z b Z b
(f (x) + c) · dx − (g(c) + c) · dx
a a
Z b Z b Z b Z b
= f (x) · dx + c · dx − g(x) · dx − c · dx
a a a a
Z b Z b
= f (x) · dx − g(x) · dx
a a
Z b
= (f (x) − g(x)) · dx
a

We see that the constant c disappears.

126 8. INTERPRETING AND USING THE INTEGRAL

The area formula shows that we integrate the upper curve minus the lower
curve. We usually determine which is which from the graphs. In the next few
problems we deliberately do not draw the graphs in the text; we want to do
that interactively in class, or have you do it for yourself.
Problem T8.7. Find the area bounded by y = x2 and y = 2 · x + 8.
Solution. It is not hard to see how a parabola and line intersect. The area
bounded by these curves must lie between them, and because the parabola
is concave up, we expect the line to be the upper curve. Let’s see: to find
the boundaries of the area, we need to figure out where they intersect. The
equation x2 = 2 · x + 8 leads to x2 − 2 · x + 8 = 0, which is (x + 2) · (x − 4) = 0.
We see that the figure is bounded by the interval [−2, 4].
It is not hard to figure out which is the upper curve from the sign diagram
for the difference. Or we can plug in a number in the interval [−2, 4] (say
x = 0). One way or another we see that y = 2 · x + 8 is above y = x2 . Our
area is
Z 4 h i Z 4 h i
2 2
(2 · x + 8) − x · dx = 2 · x + 8 − x · dx
−2 −2

2 1 3 4
=x +8·x− ·x 
3 −2
h 64 i h 8i
= 16 + 32 − − 4 − 16 +
3 3
= 36

Problem T8.8. Find the area bounded by y = 2 · x2 + 2 and y = x2 + 3 · x

and x ≤ 3.
Solution. This one is harder to see, because both curves are concave up
parabolas. First, let’s find the intersection between the curves:
2 · x2 + 2 = x2 + 3 · x leads to (x − 1) · (x − 2) = 0
 2. AREA 127

The curves intersect at x = 1 and x = 2, and so they trap area between them
there. Over that interval, plugging in x = 1.5 (or looking at the sign pattern),
we see that y = x2 + 3 · x is the upper curve, and y = 2 · x2 + 2 is the lower
curve. The area of the bubble between the curves is
Z 2 Z 2
 2 2 1
3 · x − x2 − 2 · dx =
  
(x + 3 · x) − (2 · x + 2) · dx =
1 1 6
But there’s more! The boundary x = 3 must also be taken into account.
Between x = 2 and x = 3, the curves trap another bubble of area. In that
range, the curve y = 2 · x2 + 2 is the upper curve. (How did we know that?)
The additional area is this:
Z 3 Z 3
2 2 5
x2 + 2 − 3 · x · dx =
   
(2 · x + 2) − (x + 3 · x) · dx =
2 2 6
The total area is (1/6) + (5/6) = 1.

Problem T8.9. Find the area bounded by y = ex and y = 2 − x and x ≥ 0.

Solution. From the graph, we see that y = 2 − x is the upper curve and
y = ex is the lower curve, until they meet somewhere to the right of the y-axis.
To find the intersection point, we need to solve ex = 2 − x. This equation
does not have a nice, algebraic solution, and so we use Newton’s Method to
approximate the solution.3 The equation is ex + x − 2 = 0. Starting with
x = 0, we obtain the approximate solution α = 0.44285.
The boundary x ≥ 0 gives the left-hand limit of integration. The area is
then
α
x2 α2
Z α
x x
(2 − x − e ) · dx = 2 · x − −e  =2·α− − eα + 1 ≈ 0.2305
0 2 0 2

3You might wish to review Newton’s Method, since it is going to be used several times
in this chapter. See p.72.
128 8. INTERPRETING AND USING THE INTEGRAL

3. Probability

A random variable is a variable that can take on a range of values and

such that various sets of values have a probability assigned to them.4 Random
variables that have finitely many possible values are called discrete;5 random
variables taking values from an interval are called continuous. In statistics,
continuous random variables are used all the time; even when discrete variables
occur they are often approximated by continuous variables. Calculations on
continuous random variables involve the calculus; we want to give typical
examples. Our emphasis here is on the calculation of probability and related
quantities via the integral; in a statistics course, the emphasis would be on the
interpretation of those probabilities and quantities.
Probabilities for continuous random variables are computed using what is
called a probability density function. The units of such a function are chance
per unit of measurement.6 For example, suppose that x is a continuous random
variable that represents the thickness in inches of a piece of steel produced by
a somewhat random process. Suppose that 9 ≤ x ≤ 11 and f (x) = 3 − (x/4)
is the density function in chance per inch. (Inch, since that’s x’s units.) As we
pointed out, this means that f (x) is the derivative of chance with respect to x
(inches), and so we compute chance (probability) by integration. For instance,
the chance that 9 ≤ x ≤ 10 is the integral
Z 10 
x x2 10 5
3− · dx = 3 · x −  =
9 4 8 9 8
The answer 5/8 is a probability: there is a 5/8 chance that one of the pieces
of steel has thickness between 9 and 11 inches. Notice that the limits of
integration are the boundaries of the values of x.
4The
meaning of probability is an interesting subject. We will take a computational
approach to show how the calculus is involved.
5Example: a coin flip can come out in two possible ways: heads and tails.
6Units in ratio: a probability density function is a derivative.
 3. PROBABILITY 129

If we integrate over the entire range of x, we expect to get 1 = 100%, since

all values of x are between 9 and 11:
Z 11 
x x2 11
3− · dx = 3 · x −  = 1
9 4 8 9
Students are often surprised by the notion of a probability density function;
they arise from histograms, which represent probabilities as areas – we might
have time to say something about that in class. Our immediate purpose is
to introduce a small number of famous types of continuous random variables.
You are invited to search online to get an introduction to their uses, and if you
take a course in statistics or research methods you will study them in detail.
Our concern here is to use the integral to calculate probabilities.
Constant distribution About as simple as you can get: the probability
density function is constant on the interval [a, b]. What this means is that the
random variable x is equally likely to any particular value on that interval. If
the constant density is c, then we would need to have
Z b b
1= c · dx = c · x = c · (b − a)

a a

and this shows that c = 1/(b − a). Thus, for the constant distribution, we can
compute the density.
Problem T8.10. For a constant distribution on [−3, 5], what is the probabil-
ity that x > 0?
Solution. The constant is 1/(5 + 3) = 1/8. The probability is then
Z 5
1 x 5 5
· dx =  =
0 8 8 0 8
This is easy to interpret: the statement that x > 0 is the statement that
0 < x ≤ 5. Since that interval has length 5, and the entire interval [−3, 5] has
length 8, the probability that x > 0 is 5/8. Sadly, when the density is not
constant, we cannot use such straightforward reasoning.
130 8. INTERPRETING AND USING THE INTEGRAL

The closed interval [0, 5] computes the probability that 0 ≤ x ≤ 5. We

wanted x > 0, so we want to integrate over 0 < x ≤ 5. The chance that x = 0,
R0
however is 0 (dx/8) = 0, so the integral over 0 < x ≤ 5 is the same as that
over 0 ≤ x ≤ 5. There is a mild paradox here: in a continuous probability
distribution, the probability that the random variables takes any particular
value is 0. The probability becomes positive only over intervals.

Exponential distribution We imagine standing by the side of a street and

watching the cars go by. When a car passes, we start a stop watch, and stop it
when the next car passes, recording the elapsed time. We regard the elapsed
time as a continuous random variable x. The variable x is non-negative, and
so its interval of values is [0, ∞). The ∞ right endpoint just means that there
is no right endpoint! The variable x can be arbitrarily large – theoretically.7
Such a continuous random variable is often thought to follow the exponential
distribution, whose probability density function looks like this:

(8.5) λ · exp(−λ · x)

where λ is a positive parameter – we wil see its significance momentarily.

Since x ranges over the non-negative reals, we should have
Z ∞
(8.6) 1= λ · exp(−λ · x) · dx
0
We can actually compute this integral without too much trouble. The up-
per limit ∞ simply means that we let the upper limit get larger and larger.
Technically,
Z ∞ Z b
(8.7) λ · exp(−λ · x) · dx = lim λ · exp(−λ · x) · dx
0 b→∞ 0
An integral with an infinite upper limit is called an improper integral ; they
can be studied for their own sake, and there are several nuances that can be
7Remember that a mathematical model represents aspects of something else. Having
infinity as an endpoint means that we don’t want to worry about how large x can be.
 3. PROBABILITY 131

observed. Our use of improper integrals will be confined to some probability

calculations, and so we will take an ad-hoc approach, mostly in the interest of
time. The integral in (8.7) with b as an endpoint can be calculated:
Z b
1 b
λ · exp(−λ · x) · dx = · λ · exp(−λ · x)

0 −λ 0
b
= − exp(−λ · x) = 1 − exp(−λ · b)

0

Remembering that λ is positive, if we let b get larger and larger, the graph of
the exponential function shows that exp(−λ · b) → 0. Our improper integral
becomes 1, and so, as expected, the formula (8.6) holds.

Problem T8.11. In the car watching experiment, suppose that time x is in

seconds and λ = 1/3. What is the probability that two cars will pass less than
4 seconds apart?

Solution. We are asking for the probability that 0 ≤ x < 4. As in a a previous

problem, letting x = 4 does not change the answer, and so we integrate over
the closed interval [0, 4]. We can do the integral by hand.
Z 4
1  x 1 1  x  4
· exp − · dx = · · exp −

0 3 3 −1/3 3 3 0

 x  4
= − exp −

3 0

= − exp(−4/3) + exp(0)
≈ 0.736

Our answer says that about 73.6% of the time, there will be less than 4 seconds
between cars.

Problem T8.12. In an exponential distribution with λ = 1/5, what is the

probability that x > 6?
132 8. INTERPRETING AND USING THE INTEGRAL

Solution. Let P be the probability we want. Remembering that x comes from

[0, ∞), the values x > 6 are in the interval (6, ∞). It makes sense to compute
Z ∞
1  x
P = · exp − · dx
6 5 5
We treat the infinite upper limit as we did in (8.7). The antiderivative can be
done as before:
1 1  x  ∞  x  ∞
P = · · exp − = − exp −
 
−1/5 5 5 6 5 6
 

As we said, the infinite upper limit means that we think about what hap-
pens when x gets larger and larger (we can write x → ∞). We see that
exp(−x/5) → 0 as x → ∞, and so

P = 0 + exp(−6/5) ≈ 0.3012

About 30.1% of the values of x are greater than 6.

There is an interesting alternative way to do the previous problem. The
values of x can be divided into two sets with no overlap: x in [0, 6] and x
in (6, ∞). Since x has to come from one of these two intervals, if Q is the
probability that x is in [0, 6], and P the probability that x is in (6, ∞), then
Q + P = 1. In other words,
Z 6
1  x
P =1−Q=1− · exp − · dx
0 5 5
You should compute that this gives the same answer for P as before. Notice
that this alternative method avoids an infinite endpoint.

Standard normal distribution The standard normal distribution occurs

in a host of statistical calculations, all involving the so-called bell-shaped curves.
The continuous random variable x ranges over the entire set of real numbers:
(−∞, ∞). Here is its probability density function:
 2
1 x
√ · exp −
2π 2
 3. PROBABILITY 133
√
Obviously, the 2π is mysterious, not to say surprising. We would have
Z ∞  2
1 x
1= √ · exp − · dx
−∞ 2π 2
This formula is quite difficult to prove theoretically. We will be content to
work numerically and observe approximations, as is done in practice.

Problem T8.13. Suppose that the variable x satisfies the standard normal
distribution. Find the probability that −1 ≤ x ≤ 1. (This is a famous
probability: the chance of landing within one standard deviation of the mean.)

Solution. We don’t know an antiderivative for exp(−x2 /2), and so we use a

Riemann sum approximation to integrate the density function. We computed
the sum using 5000 sub-intervals in [−1, 1].
Z 1  2
1 x
√ · exp − · dx ≈ 0.6826
2π −1 2
The reader can do an online search for a standard normal table to verify that
this number is correct to the decimals shown.

The Mean Many of the typical statistical quantities associated with prob-
abilities are computed via integration. We don’t want to get carried away
into statistics proper, but we mention two important examples, both of which
measure a kind of average for a continuous random variable.
Let x be a random variable in the interval [a, b] with probability density
function f (x). (Remember that it doesn’t really matter whether the endpoints
a, b are included, or not. For instance, we can have b = ∞.) The mean of x is
meant to be the best approximation to x by a constant. When we discuss least
squares at the end of the course, we can go into this in more detail, but the
idea is that the constant approximation is an average.8 Here is the formula for

8The word average is informal.

134 8. INTERPRETING AND USING THE INTEGRAL

the mean
Z b
(8.8) mean of x = x · f (x) · dx
a

This formula is an example of the occurrence of an integral where there is no

obvious anti-derivative. Riemann sums are involved here, starting with the
simple formula for the mean of a finite list of numbers. We may have time
to say something about this in class, but we will focus on using the integral
formula.
The mean of the constant distribution is the average of the endpoints.

Problem T8.14. Show that (a + b)/2 is the mean of the constant distribution
on [a, b].

Solution. The probability density function is the constant 1/(b − a). The
mean is then computed using (8.8).
Z b
1 x2 1 b 1
x· · dx = ·  = · (b2 − a2 )
a b−a 2 b − a a 2(b − a)
1 b+a
= · (b − a)(b + a) =
2(b − a) 2

In the exponential distribution with density function as in (8.5), it turns

out that the mean is 1/λ. (And that’s the significance of the parameter λ!)

Problem T8.15. Let λ = 1/5 in the exponential distribution. Verify that the
mean is 5.

Solution. Let µ be the mean. If x follows the exponential distribution with

λ = 1/5, then x ≥ 0. Thus, there is no upper limit to the values of x, and
(8.8) looks like this:
Z ∞
1  x
µ= x· · exp − · dx
0 5 5
 3. PROBABILITY 135

We don’t know how to find an antiderivative here. We experiment with Rie-

mann sums, using higher and higher upper limit of integration to simulate
going to infinity. We see that once x ≥ 300 (or so), the integral does not
change. Thus,
Z 300
1  x
µ≈ · exp − x·· dx
0 5 5
With 5000 sub-intervals, the Riemann sum is roughly 4.9994; with 10000 sub-
intervals, the sum rounds to 4.99985. Close to 5, as expected.
Let’s use a little trickery to demonstrate an important fact:

Problem T8.16. The mean of the standard normal distribution is 0.

Solution. We defined the probability density function for this distribution. If

µ is the mean, then
Z ∞  2
1 x
µ= x · √ · exp − · dx
−∞ 2π 2
We divide this up into two pieces: (−∞, 0] and [0, ∞).
Z 0  2 Z ∞  2
1 x 1 x
(8.9) µ = x · √ · exp − · dx + x · √ · exp − · dx
−∞ 2π 2 0 2π 2
We will show that the first integral is the opposite of the second! To do
this, let Z  2
1 x
F (x) = x · √ · exp − · dx
2π 2
It will be convenient to have F (0) = 0, as well. If we are given a random
antiderivative F (x), we can add a constant to it to get this property. Now
compute
Z b  2
1 x
x · √ · exp − · dx = F (b) − F (0) = F (b)
0 2π 2
and we get
Z ∞  2
1 x
(8.10) x · √ · exp − · dx = lim F (b)
0 2π 2 b→∞
136 8. INTERPRETING AND USING THE INTEGRAL

Use the Chain Rule to compute

d
F (−x) = F 0 (−x) · (−1)
dx
(−x)2
 
1
= (−1) · (−x) · √ · exp −
2π 2
 2
1 x
= x · √ · exp −
2π 2
0
= F (x)

The functions F (−x) and F (x) have the same derivative, and so they differ
by a constant: F (−x) = F (x) + C for some constant C. Let x = 0, and we
see that C = 0. Thus, F (−x) = F (x). Here is what follows:
Z 0  2
1 x
x · √ · exp − · dx = F (0) − F (−b) = −F (−b) = −F (b)
−b 2π 2
Thus,
Z 0  2
1 x
(8.11) x · √ · exp − · dx = lim −F (b)
−∞ 2π 2 b→∞

Taking (8.10) and (8.11) to (8.9), we see that

µ = lim F (b) + lim −F (b) = 0

b→∞ b→∞

Here is a more straightforward problem.

Problem T8.17. We have a variable x with 0 ≤ x ≤ 1 and the probability

density is 3 · x2 . Find the mean.

Solution. From (8.8):

Z 1 Z 1
2 3
x · 3 · x · dx = 3 · x3 · dx =
0 0 4
 3. PROBABILITY 137

The Median For some random variables, the mean doesn’t make sense as
an average. An example: in the distribution of income, the mean is typically
too large due to the influence of very high incomes that aren’t balanced by
very low incomes. It makes more sense to compute the value that divides the
random variables by probability. The median m of a random variable x has
the property that the probability of x < m is 50%, and the probability that
x > m is 50%. The mean and median are not always the same; each gives its
own idea of an average.
Problem T8.18. Find the median m for the random variable x with 0 ≤ x ≤ 1
and probability density function 3 · x2 .
Solution. The probability that 0 ≤ x ≤ m is 50%:
Z m
1
= 3 · x2 · dx = m3
2 0
√
3
and so m = 1/ 2 ≈ 0.7937.
We computed the mean µ = 3/4 for this distribution; the median was not
the same. As we said, each of them can be used as an average depending on
the context.
Here is a typical problem: the median is the solution to an algebraic equa-
tion.
Problem T8.19. Find the median of the random variable x on [1, 3] with
probability density (3 · x2 + 2)/30.
Solution. If m is the median, then
Z m
1 3 · x2 + 2 x3 + 2 · x m
= · dx =
2 30 30

1 1
3
m +2·m 3
= −
30 30
Multiplying through by 30, we get
15 = m3 + 2 · m − 3 which is m3 + 2 · m − 18 = 0
138 8. INTERPRETING AND USING THE INTEGRAL

We use Newton’s Method, starting at 2, since that’s in the range of the variable,
and we get m ≈ 2.367.

4. Quantities in Economics

There are many quantitites that occur in economics that are computed as
integrals. We survey a few of them.
Suppose we have an annual interest rate r on savings, constant for the
foreseeable future.9 Continuous compounding of savings is an exponential
model, where an initial amount P dollars saved for t years results in P ·exp(r·t)
dollars at the end of the t years. Write Q for the amount at the end of t years,
so that
Q = P · exp(r · t)
Turning time around, we regard Q as the given and solve for P .
Q · exp(−r · t) = P · exp(r · t) · exp(−r · t) = P · exp(0) = P
This value of P is called the present value of Q dollars t years from now.
(8.12) present value of Q = Q · exp(−r · t)

Problem T8.20. What is the present value of $1 million 50 years from now.
Assume an interest rate of 5%.
Solution. We have r = 0.05 (per year, as usual), and t = 50 years. Recall
that $1 million is 106 .The present value is given in (8.12)
$106 · exp(−0.05 · 50) ≈ $82, 084

It is straightforward to interpret present value. If we have $82,084 now,

and if we invest it at 5% compounded continuously, then we have $1 million
at the end of 50 years.
9Such an interest rate is often called the discount rate.
 4. QUANTITIES IN ECONOMICS 139

Present Value of an Income Stream. We introduced the idea of an

income stream on p.118: a rate at which income is accumulated. Let f (t)
dollars per year be an income stream, where t is time in years with 0 ≤ t ≤ T .
Suppose that an annual interest rate r on savings is available over that same
interval. The present value of the income stream is the total present value of
all the money from the stream. We have
Z T
Present value of income stream = f (t) · exp(−r · t) · dt
0

Explanation. This example was mentioned earlier in the book; the integral
comes about from Riemann sums rather than by an anti-derivative. We divide
the time interval [0, T ] into n subintervals, each of width T /n (years elapsed).
Let [p, q] be one of these subintervals. Over this time, the stream produces
roughly f (q) · T /n dollars. That’s since f (t) measures dollars per year and
T /n measures the years in [p, q]. The number f (q) · T /n only approximates
the amount of money, since f (t) may vary over the subinterval. The amount
f (q) · T /n is earned about q years from now, and so its present value is given
by (8.12)
T
f (q) ·
· exp(−r · q)
n
This term has the form f (q)·exp(−r·q) times the width T /n of the subinterval.
When these terms are added up, they form a Riemann sum for the function
f (t) · exp(−r · t) on the interval [0, T ]. As n gets larger and larger, (7.13) shows
that the Riemann sum approaches the definite integral in the equation for the
present value of the stream.

Problem T8.21. Find the present value of the income stream 10 · t · (20 − t)
dollars per year over 20 years, if the annual interest rate is 5%.

Solution. If you graph y = 10 · t · (20 − t) you will see that the rate of income
increases over the first 10 years, and then it decreases over the final 10 years.
140 8. INTERPRETING AND USING THE INTEGRAL

Notice that 5% = 0.05. Our formula above shows that

Z 20
Present value = 10 · t · (20 − t) · exp(−0.05 · t) · dt
0

We used a Riemann sum to approximate this integral: ≈ 8291.07.

Here is the meaning of the answer: if we have $8291 right now, we can gen-
erate the same income that the stream generates over the next 10 years. This
assumes continuous compounding of interest and continuous income from the
stream. Actual compounding and income would only approximate continuity.
Present value is often used to compare different ways to make the same
money.

Problem T8.22. Show that each of these income streams produces a total
income of $10000 over 10 years: 1000 per year, and 200 · t per year. Given
an interest rate of 3%, which one is preferable in terms of having the smaller
present value?

Solution. The statements about income are easy to verify:

Z 10 Z 10
1000 · dt = 10000 and 200 · t · dt = 10000
0 0

Present value of the first stream:

Z 10
1000
1000 · exp(−0.03 · t) · dt = · (1 − e−0.3 ) ≈ $8639.39
0 0.03
Second stream:
Z 10
200 · t · exp(−0.03 · t) · dt ≈ $8208.07
0

(Obtained by the Riemann sum calculator.) The second stream has the smaller
present value.
Is it clear why we prefer the smaller present value? Because it is the
cheaper way to produce the income stream.
 4. QUANTITIES IN ECONOMICS 141

Consumer Surplus. (This is a standard topic in many texts; see [10], for
example.10) A demand curve relates the unit price of some good to the demand
– how many will be sold. It is natural to think of the demand x as a function
of the unit price p, but our calculation will be simpler if we think of p as a
function of x. So, we imagine p = f (x). We usually suppose that f (x) is
decreasing and concave up,11 so that f 0 (x) < 0 and f 00 (x) > 0.
Suppose that the current unit price of apples12 is p0 , and demand is x0 , so
that p0 = f (x0 ). Some consumers are willing to pay more than p0 for apples;
the consumer surplus measures the total amount of money saved because some
consumers are paying less than they are willing to pay. On the graph p = f (x),
these consumers are to the left of x0 , where f (x) is higher than p0 .
To estimate consumer surplus, let n be a positive integer and divide the
interval [0, x0 ] into n subintervals of width ∆x = x0 /n. Let [x1 , x2 ] be one of
these intervals, so that x2 − x1 = ∆x. We think of this interval as standing
for the ∆x consumers who were willing to pay f (x2 ) for apples. Each of these
consumers paid p0 for the apples; each saved f (x2 ) − p0 . Their surplus is thus
[f (x2 ) − p0 ] · ∆x (the amount each saved times the number of consumers).
The total surplus for all consumers is the sum of all these terms, and that’s a
Riemann sum for the function f (x) − p0 over the interval [0, x0 ].
Letting the number n get larger and larger, we arrive at the following.
Z x0
 
consumer surplus = f (x) − p0 · dx
0

10The concept of consumer surplus is a kind of lightning rod, as well. See the article by
Currie, Murphy, and Schmitz, The Concept of Economic Surplus and Its Use in Economic
Analysis, The Economic Journal, 1971, pp.741-799. Available on the internet.
11This is a form of diminishing returns, for the rate of decrease of p is decreasing (in
absolute value).
12For a unit of apples, imagine a large crate.
142 8. INTERPRETING AND USING THE INTEGRAL

We can separate out the p0 term.

Z x0 Z x0 Z x0
 
f (x) − p0 · dx = f (x) · dx − p0 · dx
0 0 0
Since p0 is constant, the right most integral evaluates to p0 · x0 . We end up
with this formula.
Z x0
(8.13) consumer surplus = f (x) · dx − p0 · x0
0
where p = f (x) is the demand curve.
Problem T8.23. Suppose that a demand curve is p = x/10−10·ln(x+1)+50,
in dollars, for 0 ≤ x ≤ 100. At price $15, find the consumer surplus.
Solution. We have p0 = 15, and we need to find x0 , where
p0 = x0 /10 − 10 · ln(x0 + 1) + 50
We need to solve the equation
x0 x0
15 = − 10 · ln(x0 + 1) + 50 which is − 10 · ln(x0 + 1) + 35 = 0
10 10
Starting at x0 ≈ 1, Newton’s Method gives x0 ≈ 58.357.
Now we use (8.13) to get the surplus C.
Z 58.357 h
x i
C≈ − 10 · ln(x + 1) + 50 · dx − 15 · 58.357 ≈ 326.28
0 10

Problem T8.24. The per unit tax on an item is supposed to generate revenue
√
equal to 10% of the consumer surplus. If the demand curve is p = 100 − x,
in dollars, for 0 ≤ x ≤ 1000, and the current price is $75, what should the per
item tax be?
Solution. When p0 = 75, we solve for x0 = 252 = 625. The consumer surplus
from (8.13):
Z 625
√
(100 − x) · dx − 75 · 625 ≈ $5208.33
0
 4. QUANTITIES IN ECONOMICS 143

The tax revenue is 10% of this surplus: $520.83. The demand is 625, and so
the per item tax is
$520.83
≈ 83/c
625

Relating taxes to surplus is intended to limit how much taxes cut into
spending. Good luck.
 CHAPTER 9

Matrix Algebra

1. Matrix Arithmetic

This section introduces matrices and algebraic operations on them; perhaps

it goes without saying that you will need to learn these operations.
A matrix is a table of numbers. Such a table organizes its entries into hori-
zontal rows and vertical columns. The entries are often enclosed in parentheses
or brackets    
1 2 3 1 2 3
=
4 5 6 4 5 6
This matrix has 2 rows and 3 columns, and so its size is 2 × 3. The rows
and columns are numbered in an obvious way: rows from top to bottom
and columns from left to right. Thus,  the second row of the given matrix

1
is 4 5 6 and the first column is . The entry in row i and column j is
4
called the i, j-entry. We often used square brackets to indicate entries: A[1, 2]
stands for the entry in row 1 and column 2 of matrix A. It is also common
to use subscripts Ai,j , but we prefer the bracket notation, since it is more
readable.
Two matrices are the same if they have the same size and the same cor-
responding entries. Thus, A = B means that both A, B are m × n, for some
m, n, and also A[i, j] = B[i, j] for all relevant i, j.
Matrices have an arithmetic that is somewhat similar to number arithmetic,
although there are important differences that we will note.

145
146 9. MATRIX ALGEBRA

1.1. Matrix Addition. Matrix addition is simply entry by entry. The

two matrices added have to have the same size, so the entries of one will
correspond to the entries of the other. For instance,
     
1 2 3 −2 3 7 −1 5 10
+ =
4 5 6 5 −6 −6 9 −1 0
It is very easy to see that addition of matrices has properties similar to
those of number addition. Let A, B, C be matrices all of the same size.
(1) Commutative Law A + B = B + A
(2) Associative Law (A + B) + C = A + (B + C)

Problem T9.1. Show that the Commutative Law is true.

Solution. The definition of A + B and B + A shows that they are m × n,

since A, B are both m × n. Thus, A + B and B + A have the same size. To
compare [i, j] entries, we use the definition of matrix addition to write

(A + B)[i, j] = A[i, j] + B[i, j]

Because real addition is commutative, we have A[i, j]+B[i, j] = B[i, j]+A[i, j].
And the definition of matrix addition shows that

B[i, j] + A[i, j] = (B + A)[i, j]

We conclude that (A + B)[i, j] = (B + A)[i, j] for all i, j. The entries of A + B

and B + A are same, and we see that they are the same matrix.
In class, we will discuss the significantly different meaning of the plus sign
in (A + B)[i, j] and in A[i, j] + B[i, j].
It is easy to take such things as the commutative and associative properties
for granted; that is precisely the point of having them. These identities allow
us to perform various matrix additions using familiar properties of numbers.
For each size m × n that a matrix can have, we define the matrix Om×n
of that size and having all its entries 0. This is called the m × n zero matrix.
 1. MATRIX ARITHMETIC 147

Because matrix addition is done entry by entry, it is obvious that adding the
zero matrix to a matrix A does not change A.
Zero Matrix If A is an m × n matrix, then
A + Om×n = A = Om×n + A

1.2. Scalar Multiplication. The next thing we want to do is to multiply

a matrix by a number. When numbers are used this way, they are often called
scalars. To multiply a matrix by a scalar, we simply multiply each entry in
the matrix by the scalar. For example,
   
1 2 −2 −4
−2 · −1 −2 =  2 4 
0 5 0 −10
As with matrix addition, the properties of this operation are fairly natural,
and they are not hard to verify. Let A, B be matrices of the same size, and let
α, β be scalars.
(1) 0 · A = O
(2) β · (A + B) = (β · A) + (β · B)
(3) (α + β) · A = (α · A) + (β · A)
(4) (α · β) · A = α · (β · A)
We will demonstrate at least one of these in class. The argument is similar
to the one we gave for matrix addition being commutative, in that we use the
definition of the matrix operations and familiar properties of the real numbers.
It is customary to write (−1) · A as −A. Observe that each entry of −A is
the negative of the corresponding entry of A. Then A + (−A) = O.
Matrix addition and scalar multiplication allow us to solve simple matrix
equations in exactly the same way as we solve simple number equations. For
example, suppose that X is an unknown matrix and we have
   
1 2 7 9
3·X + =
3 4 −1 6
148 9. MATRIX ALGEBRA

We can add the negative of a matrix to both sides to isolate 3 · X.

       
1 2 1 2 7 9 1 2
3·X + − = −
3 4 3 4 −1 6 3 4
   
0 0 6 7
3·X + =
0 0 −4 2
 
6 7
3·X =
−4 2
In doing the previous calculation, we used the associative property of matrix
addition and the defining property of the zero matrix. Now we scalar multiply
both sides by 1/3.
 
1 1 6 7
·3·X = ·
3 3 −4 2
 
2 7/3
X=
−4/3 2/3
The point of this example is that there is nothing profound involved. We solved
for the matrix X as we would have solved a simple equation for a number X.
We tend to use matrix arithmetic without being very self-conscious.

1.3. Matrix Multiplication. We need one more operation on matrices,

a kind of multiplication. Its formula is not easy to motivate. For our immediate
purposes, it will be better just to introduce multiplication and to get used to
computing it.
To give the definition we start with the dot product of a row and a column.
We need to have a row and a column with the same number of entries; we
multiply corresponding entries and add up the products:
 
3

4
1 0 −2 3 ·   1  = (1 · 3) + (0 · 4) + (−2 · 1) + (3 · (−5))

−5
= 3 + 0 − 2 − 15 = −14
 1. MATRIX ARITHMETIC 149

Here is another example.


  −1
4 3 −2 ·  2  = 4 · (−1) + 3 · 2 + (−2) · 1
1
= −4 + 6 − 2 = 0

Now we are ready to define matrix multiplication. There are several nu-
ances in this definition – make sure you understand the details. First, the
matrices A and B can be multiplied to form the product A · B if A has the
same number of columns as B has rows. The best way to check this is to put
the sizes of A and B side by side (keep A’s size on the left). Suppose that A
is 4 × 3 and B is 3 × 5:
4×3 3×5
We need the inside numbers to match: A has 3 columns and B has 3 rows.
This tells us that we can form the product A · B. The outside numbers give
us the size of the result; the matrix A · B will be 4 × 5.
If A is 9 × 9 and B is 7 × 9, then, putting the sizes next to each other:

9×9 7×9

we see that A · B is not defined. Notice, however, that if we put B first we get

7×9 9×9

we see that B · A is defined ! And B · A will be 7 × 9. In matrix multiplication

the order of the matrices is crucial.
Let’s introduce the calculation of A · B via an example. Suppose
 
1 2  
−2 7
A = 3 4  , B =
1 −3
5 6
Looking at the sizes:
A B
3×2 2×2
150 9. MATRIX ALGEBRA

we see that A · B is 3 × 2. Here is a template for a 3 × 2 matrix.

1,1 1,2
2,1 2,2
3,1 3,2

To get each entry in the product, we compute the dot product of a row of
A and a column of B, the row and column indicated by the entry. Thus, to
get the 1, 2 entry of A · B, we want the dot product of row 1 of A and column
2 of B:

(A · B)[1, 2] = (row 1 of A) · (col 2 of B)

 
  7
= 1 2 ·
−3
= 1 · 7 + 2 · (−3) = 1

Similarly, the 3, 1 entry of A · B is this.

(A · B)[3, 1] = (row 3 of A) · (col 1 of B)

 
  −2
= 5 6 ·
1
= 5 · (−2) + 6 · 1 = −4

Here is the entire product.

 
0 1
A · B = −2 9 
−4 17

Problem T9.2. Let

 
1 2 3 0  
1 −1 2
A = 4 5 6 0 and B =
3 0 4
7 8 9 0
See if you can compute A · B and B · A.
 1. MATRIX ARITHMETIC 151

Solution. The matrix A has 4 columns, and B has 3 rows, so A · B is not

defined. As for B · A:
B A
2×3 3×4
Thus, B · A makes sense; it is 2 × 4.
We set up a template for the answer.
 
  1 2 3 0  
1 −1 2  1,1 1,2 1,3 1,4
· 4 5 6 0 =
3 0 4 2,1 2,2 2,3 2,4
7 8 9 0
Here is the answer.
 
 1 2 3 0  
1 −1 2  11 13 15 0
· 4 5 6 0 =
3 0 4 31 38 45 0
7 8 9 0

Multiplication is not commutative in general. As we have seen, we might

have A · B defined but B · A not defined; then A · B 6= B · A. If A is 2 × 3
and B is 3 × 2, then A · B is 2 × 2, and B · A is 3 × 3, so that both A · B and
B · A are defined, but they are not the same size, and so they cannot be equal.
Even if A · B and B · A are both defined and of the same size, they are often
not equal. As we mentioned before, in matrix multiplication the order of the
matrices must be observed carefully.
The identity matrices act like the number 1 in multiplication. For each
positive integer n, there is an n × n matrix In defined like this.
 
  1 0 0 0
  1 0 0
1 0  0 1 0 0
I2 = , I3 = 0 1 0 , I4 =  , ...
0 1 0 0 1 0
0 0 1
0 0 0 1

The matrix In has In [i, i] = 1 for each i; we say that the 1’s are on the diagonal .
The other entries of In are 0.
152 9. MATRIX ALGEBRA

Identity Matrices If A is m × n, then

Im · A = A = A · In

You should check that the size of Im · A is m × n and that the size of A · In
is m × n. In class, we will discuss these identities; they follow directly from
the formula for matrix multiplication.
We have emphasized that matrix multiplication is not commutative. How-
ever, it does have many properties similar to those of numbers. We list these
properties to be clear and complete in our introduction to matrix algebra. As
we mentioned previously, we will use these properties without calling attention
to them. Throughout the list of rules, A, B, C are matrices.
(1) Associative Law If A · B and B · C are defined, then

(A · B) · C = A · (B · C)

(2) Distributive Law If A · B is defined, and if B, C have the same size,

then
A · (B + C) = (A · B) + (A · C)
(3) Distributive Law If A · C is defined, and if A, B have the same size,
then
(A + B) · C = (A · C) + (B · C)
(4) Scalars If A · B is defined, and if k is a constant, then

k · (A · B) = (k · A) · B = A · (k · B)

2. Applications of Matrix Algebra

In Section 3 of the Spreadsheet Appendix, we discuss the Excel formulas

for matrix arithmetic. We will make use of those formulas as needed in the
following.
 2. APPLICATIONS OF MATRIX ALGEBRA 153

2.1. Recursive Sequences. We consider a pair of recursive sequences,

as in Chapter 2. Suppose we have An and Bn , for n = 0, 1, 2, . . ., with these
equations.
An+1 = An − 2 · Bn A0 = 1
Bn+1 = −An + 3 · Bn B0 = 2
We know that the recursive equations give us further values of both sequences.
What we want to see is that the recursive equations are matrix equations:
     
An+1 1 −2 A
(9.1) = · n
Bn+1 −1 3 Bn
Compare this equation carefully with the recursion.
Equation (9.1) suggests using matrix multiplication to compute the se-
quences An , Bn . Here is a possible spreadsheet layout.
A B C D E
1 1 -2
2 -1 3
3 n: 0 1 2
4 A[n] 1 Λ1 Γ1 · · ·
5 B[n] 2 Λ2 Γ2 · · ·
The matrix in (9.1) appears in A1:B2 of the spreadsheet. The initial values of
the sequences appear in B4:B5. The entries marked with Λ1 , Λ2 are selected
when the following matrix multiplication formula is entered in C4:
mmult($A$1:$B$2,B4:B5)
As a result, A1 , B1 appear in C4:C5. If we fill-right from C4:C5 to the
right, we will get subsequent terms of the sequences. For instance, the entries
marked Γ1 , Γ2 will have the formula:
mmult($A$1:$B$2,C4:C5)
The dollar signs around the reference to the 2×2 matrix in A1:B2 will maintain
that reference as the sequence terms are computed.
154 9. MATRIX ALGEBRA

Problem T9.3. Use spreadsheet matrix multiplication to compute the first

10 terms of the sequences An , Bn , where

An+1 = An + Bn A0 = 4
Bn+1 = 3 · An − Bn B0 = 1

Solution. The recursion is a matrix equation, like (9.1).

     
An+1 1 1 A
= · n
Bn+1 3 −1 Bn
Here is a possible spreadsheet layout.
A B C D E
1 matrix 1 1
2 3 -1
3 n: 0 1 2
4 A[n] 4 Λ1
5 B[n] 1 Λ1
where the cells marked Λ1 were selected when the the formula

=mmult($B$1:$C$2,B4:B5)

was placed in cell C4. We can fill-right to produce the remaining 9 terms
of the sequences. We get A9 = 1280 and B9 = 2816.
Here is another type of recursion that can be written as matrix multipli-
cation.

Problem T9.4. Write a matrix equation for this recursion:

Cn+2 = −6 · Cn+1 − 13 · Cn for n = 0, 1, 2, . . .

and C0 = 1 and C1 = 0.

Solution. The key idea is to pair Cn and Cn+1 . Let Bn = Cn+1 for n ≥ 0,
and then B0 = C1 = 0. The recursive equation is now

Bn+1 = −6 · Bn − 13 · Cn
 2. APPLICATIONS OF MATRIX ALGEBRA 155

Here are the two recursive equations for Bn and Cn :

Cn+1 = Bn = 0 · Cn + 1 · Bn
Bn+1 = −13 · Cn − 6 · Bn
Here is the corresponding matrix equation:
     
Cn+1 0 1 C
= · n
Bn+1 −13 −6 Bn
We can compute this sequence the way we computed the previous two.

2.2. Leslie’s Growth Model. Patrick Leslie used matrices to model

changes in animal populations over time.1 We present a simplified example to
illustrate the main idea.
We divide a female population into age categories, each category spanning
a specific number of years. Each of the categories has two numbers associated
with it. The birth rate is the average number of female offspring that an animal
in that category will bear during the time for that category. The survival rate
is the likelihood that an animal in that category will survive into the next
category.
Here is a table for female New Zealand sheep.2 Each time category is three
years long.
age birth rate survival rate
0-3 0.436 0.795
4-6 1.502 0.787
7-9 1.513 0.462
10-12 1.313 0
Let An , Bn , Cn , Dn give the population in each age category after 3 · n years.

1One of Leslie’s papers on the subject is The use of matrices in certain population
mathematics appeared in 1945 in the journal Biometrika, Vol. 33(3), pp183-212.
2This table is a simplified version of the more realistic (and larger) table in the paper
Parameters for Seasonally Breeding Populations, by G. Caughley from the journal Ecology,
vol. 48, 1967, pp. 834-839
156 9. MATRIX ALGEBRA

Thus, A2 is the number of female sheep aged 0-3 after 6 years; B4 is the
number of female sheep aged 4-6 after 12 years, and so on. The birth rate
puts new animals into the An sequence. The survival rate puts animals in the
next category. In light of this, here are recursive equations for the populations.

An+1 = 0.436 · An + 1.502 · Bn + 1.513 · Cn + 1.313 · Dn

Bn+1 = 0.795 · An
Cn+1 = 0.787 · Bn
Dn+1 = 0.462 · Cn

In matrices:
     
An+1 0.436 1.502 1.513 1.313 An
 Bn+1  0.795 0 0 0   Bn 
(9.2) · 
 Cn+1  =  0
  
0.787 0 0   Cn 
Dn+1 0 0 0.462 0 Dn
Here is an example problem.

Problem T9.5. In the sheep model just considered, suppose we start with 10
sheep in category 1 and none in the other categories. How long until the total
population reaches 100 sheep?

Solution. Equation (9.2) can be used in a spreadsheet to produce the se-

quence, starting with    
A0 10
 B0   0 
 = 
 C0   0 
D0 0
The total population is the sum of the populations for each category. Here
are the total populations for certain values of n. (As usual, the model is
approximate – this one predicts fractional numbers.)
n: 3 4 5 6
population: 37.3 59.9 97.7 158.3
 2. APPLICATIONS OF MATRIX ALGEBRA 157

The population hits 100 somewhere between n = 5 and n = 6; that’s between

15 and 18 years from the start.

2.3. Markov Processes. These processes were introduced in Chapter 2.

We have a finite number of states; at each time step we have to be in exactly
one of the states. We are interested in probability vectors – a list showing the
probability that we are in each of the possible states. For example, suppose
there are three states: 1,2,3. A probability vector Q might look like this:
 
0.2
Q = 0.5
0.3

Since Q[1] = 0.2, the probability is 20% that we are in state 1. State 2:
50% and state 3: 30%. If we are in state 2 for certain, then Q[2] = 1 and
Q[1] = Q[3] = 0. A probability vector has non-negative entries, one for each
state, that add up to 1=100%, since we have to be in exactly one of the states.
The key feature of a Markov process is its transition matrix ; that matrix
shows the probability of going from one state, at a particular time, to the
various states at the next time. Suppose, for example, we have three states
1,2,3, with the following transition matrix.
 
0.1 0.4 0.2
(9.3) A = 0.6 0.1 0.7
0.3 0.5 0.1

Each column of A imagines we are in the state of that column and tells us the
chance that we move to each of the three states at the next time step. For
instance, A[3, 2] = 0.5. If we are in state 2, then the chance that we will be in
state 3 at the next time is 50%.
If Qn holds the probability of being in each state at time n, then

(9.4) Qn+1 = A · Qn
158 9. MATRIX ALGEBRA

gives the probabilities at time n + 1. We will discuss this equation in class;

obviously it gives us yet another example of recursion.
It is a theorem that if the entries in A are all positive, then no matter what
Q0 is, the recursive sequence (9.4) converges to a probability distribution Q̄
such that
(9.5) Q̄ = A · Q̄
The probability vector Q̄ is an equilibrium for the recursion (9.4).
Problem T9.6. For the transition matrix A in (9.3) above, find the equilib-
rium probability vector.
Solution. The theorem to which we referred says that we can get the equi-
librium probability vector Q̄ by starting with an arbitrary probability vector
Q0 and running the recursion out. We started with
   
1 0.258
Q0 = 0 and obtained Q̄ ≈ Q13 ≈ 0.421
0 0.320

It is instructive to start with other choices for Q0 ; we should obtain the

same Q̄.
The equilibrium equation (9.5) can be written in a way that is discussed
in the next chapter. Suppose that we have n states so that A is n × n. Then
Q̄ = A · Q̄ is In · Q̄ = A · Q̄
and we can move Q̄ to the left.
(9.6) (In − A) · Q̄ = On×1
Because Q̄ is a probability vector, there are two more conditions: first, the
entries in Q̄ are non-negative; second, the entries in Q̄ add up to 1. This is
an example of a system of linear equations; those are discussed in the next
chapter.
 2. APPLICATIONS OF MATRIX ALGEBRA 159

2.4. Leontief ’s Linear Economy. This macro-economic model is dis-

cussed in [5] and in [10]. In class, we will work with a specific example. Here,
we describe the model abstractly.
We imagine an economy with n goods, and we want to keep track of the
production and consumption of those goods. We have an n × n table A where
A[i, j] gives the number of units of good i needed to make one unit of good
j. The matrix A is called the technology matrix . Let X[j] hold the number
of units of good j that is produced; we call X the production schedule. Then
A[i, j] · X[j] is the number of units of good i needed to produce the X[j]
units of good j. Let C[i] be the number of units of good i that is consumed
after production; the matrix C is a consumption vector . Thus, the X[i] units
of good i are used in two ways: for production of various goods and to be
consumed.
The production X has to cover both uses. In class, we will see that this
leads to this inequality:

(9.7) A·X +C ≤X

The meaning of inequality here is that it applies to each entry on both sides.
For instance, the third entry on the left is less than or equal to the third entry
on the right. And so for all entries. If (9.7) holds, we say that X is a feasible
production schedule.
Later in the course we will see how to find a feasible production schedule,
if such a schedule exists. At this point, we introduce a related concept. If we
are given a feasible production schedule X0 , then we can form the recursive
sequence

(9.8) Xn+1 = A · Xn + C for n = 0, 1, 2, . . .

and it turns out that each Xn is a feasible production schedule. Furthermore,

this sequence moves steadily toward a schedule Y that satisfies (9.7) as an
160 9. MATRIX ALGEBRA

equation.
(9.9) A·Y +C =Y
The equation shows that production Y exactly covers usage in production and
consumption. In other words, Y is a maximally efficient schedule.
Let’s interpret equation (9.8). The right side A · Xn + C meaures the goods
used in the economy: used for production and used for consumption. Setting
Xn+1 to this means that each producer produces next year exactly what was
used this year. In many situations, when each person adjusts his own activity
to make it more efficient, the efficiency of the group is unpredictable. For
instance, if everyone comes early to a concert to get a good seat, the task of
finding a good seat will be no easier than it was when everyone came on time.
But in Leontief’s model, individual adjustments lead to macro-efficiency. As
we converge to (9.9), the economy becomes maximally efficient: everything
that is produced is used.
Problem T9.7. We have a technology matrix A, consumption vector C, and
feasible production schedule X0 , as follows
     
0 0.15 0.225 0.075 30 150
 1.5 0.075 0.0075 0.15   10 
   300
A=  0 and C = and X0 =  
0.3 0 0.075 100 250
0.75 0.225 0.075 0.15 50 300
Find the sequence Xn described above and estimate the schedule Y such that
(9.9) holds.
Solution. We have only to use the recursion (9.8). Excel shows that X11 ≈
X12 , and  
140
286
X12 =  
207
277
This should be a good estimate of Y .
 2. APPLICATIONS OF MATRIX ALGEBRA 161

We will discuss the spreadsheet layout of the previous problem in class.

For now, we indicate what the recursion (9.8) looks like. Say, to be specific
that the 4 × 4 matrix A appears in the spreadsheet as A1:D4, and suppose
that C is F1:F4. Now suppose that X0 occurs at C6:C9 and that we want X1
to appear at D6:D9. Then, selecting D6:D9, we enter this formula in D6:
mmult($A$1:$D$4,C6:C9)+$F$1:$F$4
Using fill-right we can produce as many of the Xn as we wish.
 CHAPTER 10

Linear Equations

1. Equations and Solutions

1.1. Definitions. In a linear equation a sum of multiples of the variables

is set equal to a number. Such as,

3 · x1 − 2 · x2 + 5 · x3 + x 4 = 1

A solution to this equation is a set of values of the variables. For instance, let

x1 = 1, x2 = 4, x3 = 0, x4 = 6

and observe that these values make the linear equation true:1

3·1−2·4+5·0+6=1

A system of linear equations is a finite list of linear equations, all in the

same variables. Example: we have variables a, b, c and equations
a − 2b − c = −2
(10.1) 2a + 3b + 3c = 15
a+b+c=6
Such a system can be written as a matrix product:2
     
1 −2 −1 a −2
(10.2) 2 3 3  ·  b  =  15 
1 1 1 c 6

1Don’t worry about how this solution was found; we’ll discuss that later.
2This form provides a partial explanation for the unexpected definition of matrix
multiplication.
163
164 10. LINEAR EQUATIONS

It is crucial that you see how the matrix product (10.2) corresponds directly
to the equations in (10.1). There is one row of the matrix on the left for each
equation. There is one column of the matrix on the left for each variable.
A solution to a system of equations is a solution to each of the equations
simultaneously. For the system (10.1) we can let
   
a 3
 b  = 2
c 1
and observe that (10.2) now gives us a true matrix equation:
     
1 −2 −1 3 −2
2 3 3 · 2 = 15 
   
1 1 1 1 6
as you should verify! The single matrix equation is really three equations: one
for each entry – one for each of the three equations that the system requires
to hold simultaneously.
Abstractly, a system of linear equations looks like this:
(10.3) A·X =B
where A is a matrix of numbers – it is called the coefficient matrix . The
symbol X stands for a column of unknowns (variables), as many variables as
A has columns. The matrix B gives the right sides of each equation, and so B
is called the right side matrix . A solution to this matrix equation is a specific
matrix C, corresponding to X, such that
A·C =B
The j-th entry of C tells the value of the j-th variable in X.
It is not hard to solve a system of linear equations. The main idea is
to use one of the equations to solve for one of the variables; then you can
substitute for that variable in the remaining equations, reducing the problem
by one equation and one variable, and then continuing in the same way. We
 1. EQUATIONS AND SOLUTIONS 165

could take a rather ad-hoc approach to the solution of linear equations, but
we choose to be systematic for at least two reasons. First, there is information
hidden in the equation AX = B that is relevant to general applications of
matrices; a careful solution technique will disclose this information. Second, a
systematic approach to solutions will allow us to reach a solution in an efficient
way that avoids dead ends.

1.2. Elimination. As we mentioned above, we want to give an official

algorithm for solving systems of linear equations, so that we can check our
work, and so that we will be sure to get an answer. We will use the name
Elimination for the algorithm. To explain it, we consider an example.
     
−1 0 1 a 1
(10.4)  2 0 2 · b = 2
   
1 3 3 c 3

Notice that each column of the coefficient matrix A corresponds to an un-

known. Each row of the coefficient matrix corresponds to an equation. We
put the coefficient matrix A with the right side matrix B to form a single
matrix [A|B]. (The vertical line is just meant to separate the A from the B
so that we do not confuse this juxtaposition with the matrix product AB.)

 
a b c =
−1 0 1 1
[A|B] = 
2

0 2 2
1 3 3 3

The matrix [A|B] is the augmented matrix of the system. The augmented
matrix is just a table that contains all the numbers we are working with. We
have labeled the columns of A with the unknowns, so we don’t forget them!
And we labeled the last column with an equals sign – it is the right side B.
166 10. LINEAR EQUATIONS

Elimination involves three types of operations performed on the augmented

matrix. These operations are the elementary operations.3
(1) Interchange two of the rows.
(2) Multiply a row by a non-zero number.
(3) Add a multiple of one row to another (leaving the first row the same).
Here is the key property of elementary operations: if we have matrices
A, B, C and A · C = B, and if we perform an elementary operation on the
augmented matrix [A|B] to obtain the matrix [A0 |B 0 ], then A0 · C = B 0 . In
other words, if C is a solution to the equation A · X = B, then C is a solution
to the equation A0 · X = B 0 .
Furthermore, for each elementary operation there is an elementary opera-
tion that undoes it. If we interchange rows 2 and 5, say, then interchanging
rows 2 and 5 again brings us back to where we started. If we multiply row 7
by 9, then multiplying row 7 by 1/9 brings us back where we started. And if
we add -2 times row 3 to row 1, then adding +2 times row 3 to row 1 gets us
back to where we started.4 So, if we have the augmented matrix [A0 |B 0 ], as in
the previous paragraph, then we can perform an elementary operation to get
the matrix [A|B] back again. By what we said in the previous paragraph: if
C is a solution to A0 · X = B 0 , then C is a solution to A · X = B.
The last two paragraphs prove this: if we are given the equation A·X = B,
and if we apply the same elementary operation to A and B to get the equation
A0 · X = B 0 , then the second equation has exactly the same solutions as the
first equation.
Elimination will apply elementary operations to transform a given system
into a form in which the solutions (or lack of solutions) will be obvious. Since

3These
operations are also called row operations and elementary row operations.
4The
claim about this last elementary operation is a little harder to see; look at a couple
of examples.
 1. EQUATIONS AND SOLUTIONS 167

the operations do not change the set of solutions, the solutions of the given
system will be obvious from the transformed system.
Let’s go back to the example linear equation to demonstrate the Elimi-
nation algorithm. Then we will describe Elimination in general. In trying
the algorithm on our example, observe the notation we use for the elementary
operations. Rows are referred to by roman numerals. We use −(1/3) · II for
multiplying row 2 by -1/3, and −4 · I + II for adding -4 times row 1 to row 2.
Here is the augmented matrix we had:
 
a b c =
−1 0 1 1
[A|B] = 
2

0 2 2
1 3 3 3

We begin with the −1 at position 1, 1. This entry is called a pivot. It will

be convenient for our pivots to be 1’s, and so we multiply row one by −1 to
change the pivot into a 1.
 
a b c =  
−1 1 0 −1 −1
0 1 1

2
 −1·I 2 0 2 2
0 2 2
1 3 3 3
1 3 3 3

(We drop the top row to avoid writing it over and over.) Now we use the pivot
1 in the [1, 1] position to eliminate the occurrence of the variable a in all the
other equations. To do this, we add multiples of row one to the other rows,
clearing the entries below a. Watch, and notice that we do not change row
one.
   
1 0 −1 −1 1 0 −1 −1
2 0 2 −2 · I + II
2 0 0 4 4
−1 · I + III
1 3 3 3 0 3 4 4
Do you see why we added −2 times row I to row II? We wanted to clear the
2 at [2, 1].
168 10. LINEAR EQUATIONS

Next we move to column 2 (the column of the variable b). We already

have a pivot in row one, and so we look below that row. We will use the 3 in
position [3, 2], since it’s non-zero. We bring that entry to row two by switching
rows two and three.
   
1 0 −1 −1 1 0 −1 −1
0 0 4 4 II ↔ III 0 3 4 4
0 3 4 4 0 0 4 4
Now we divide row two by the pivot 3 to turn it into a 1, as we did before.
   
1 0 −1 −1 1 0 −1 −1
0 3 4 4  1/3 · II 0 1 4/3 4/3
0 0 4 4 0 0 4 4
Next, we would eliminate variable b from the other equations, but it’s already
gone! We move to column 3 (the variable c), and row three. Our pivot is the
4 in [3, 3]. Divide row 3 by 4, and then clear above. Here goes.
   
1 0 −1 −1 1 0 −1 −1
0 1 4/3 4/3 1/4 · III 0 1 4/3 4/3
0 0 4 4 0 0 1 1
 
1 0 0 0
−4/3 · III + II 
0 1 0 0
1 · III + I
0 0 1 1
We have found pivots in all three rows, and so there are no more pivots to
find. The resulting form is called row-echelon form. The row-echelon form
still gives equations in a, b, c:

a=0 b=0 c=1

Because these equations were obtained from the original equations (10.4) by
elementary operations, the resulting equations have the same solutions as the
original. In other words, a = b = 0 and c = 1 is the unique solution to (10.4)!
Now we describe Elimination in general, given the system AX = B.
 1. EQUATIONS AND SOLUTIONS 169

Elimination
Apply Steps 1-4, with row 1, 2, . . ., in turn, as the current row, until the bottom
row is reached or Step 1 fails.
Step 1. Find the leftmost column in the coefficient part of the augmented
matrix having a non-zero entry at the current row or below. If there is no such
entry, Step 1 fails. Otherwise, choose one such entry (this entry is a pivot).
Step 2. Switch rows, if necessary, to bring the pivot to the current row.
Step 3. Multiply the current row by the inverse of the pivot (so that the
pivot now has value 1).
Step 4 Add multiples of the current row to rows above and below it so that
the pivot is the only non-zero entry in its column.
Example Elimination Let’s solve the system of equations given by the
following augmented matrix, where the unknowns are listed along the top.
 
x1 x2 x3 x4 x 5
 2 -3 1 4 1 17 
 
 -4 6 -1 -6 0 -27
 
 1 1 2 7 -1 20 
-4 1 -4 -16 3 -50
Step 1 might choose the 2 at the 1,1-entry as pivot. Step 2 is not needed,
and Step 3 and Step 4 compute
   
2 -3 1 4 1 17 (1/2) · I 1 -3/2 1/2 2 1/2 17/2
-4 6
 -1 -6 0 -27 4 · I + II
 0 0
 1 2 2 7 
1 1 2 7 -1 20  -1 · I + III 0 5/2 3/2 5 -3/2 23/2
-4 1 -4 -16 3 -50 4 · I + IV 0 -5 -2 -8 5 -16
With row 2 as current row, the leftmost column at row 2 or below is column
2. We use 5/2 as pivot; Step 2 interchanges rows 2 and 3:
170 10. LINEAR EQUATIONS
   
1 -3/2 1/2 2 1/2 17/2 1 -3/2 1/2 2 1/2 17/2
0 0 1 2 2 7  0 5/2 3/2 5 -3/2 23/2
0 5/2 3/2 5 -3/2 23/2 II ↔ III
   
0 0 1 2 2 7 
0 -5 -2 -8 5 -16 0 -5 -2 -8 5 -16
Then Step 3 and Step 4 come along.
   
1 -3/2 1/2 2 1/2 17/2 1 0 7/5 5 -2/5 77/5
0 5/2 (2/5) · II
 3/2 5 -3/2 23/2  (3/2) · II + I
0
 1 3/5 2 -3/5 23/5

0 0 1 2 2 7  0 0 1 2 2 7 
5 · II + IV
0 -5 -2 -8 5 -16 0 0 1 2 2 7
Now row 3 is the current row. We use the 1 at entry 3,3 as pivot. Step 2 and
Step 3 are skipped. Step 4:
   
1 0 7/5 5 -2/5 77/5 1 0 0 11/5 -16/5 28/5
0 1 3/5 2 -3/5 23/5 -(3/5) · III + II 0 1 0 4/5 -9/5 2/5 
  -(7/5) · III + I  
0 0 1 2 2 7  0 0 1 2 2 7 
-1 · III + IV
0 0 1 2 2 7 0 0 0 0 0 0
Continue to row 4 and Step 1 fails. We are done.
Recall that the coefficient matrix you get at the end of Elimination is in
row-echelon form; the echelons referred to are the columns with their pivots
organized left to right and top to bottom. The number of pivots (pivoted
variables) in Elimination is called the rank of the coefficient matrix.
The last equation in this system simply says 0 = 0, which is always true
and, therefore, has no effect on solutions. Let’s write out the three equations
that matter.

x1 + (11/5)x4 − (16/5)x5 = 28/5

x2 + (4/5)x4 − (9/5)x5 = 2/5
x3 + 2x4 + 2x5 = 7

If x4 and x5 are chosen arbitrarily, then x1 and x2 and x3 are uniquely deter-
mined in a solution to the system. Thus, there are infinitely many solutions,
 1. EQUATIONS AND SOLUTIONS 171

and, in choosing some particular solution, x4 and x5 are arbitrary. These ar-
bitrary variables did not get pivots in Elimination, whereas the determined
variables did get pivots. When a system is consistent (has solutions), the piv-
oted variables are determined by the non-pivoted variables. The non-pivoted
variables are said to be free since their values are arbitrary. The pivoted vari-
ables are called basic.. The number of basic variables is the number of pivots
is the rank of the coefficient matrix.

Problem T10.1. Solve the system of equations given by the following aug-
mented matrix, in which the unknowns are listed along the top.
 
x1 x2 x3 =
1 2 3 1
 
4 5 6 0
7 8 9 0

Solution. We put the relevant step from the Elimination algorithm in a square
at the top of each cluster of calculations. As we will see, we find pivots at
entries [1, 1] and [2, 2].
   
1 2 3 1 4 1 2 3 1
4 5 6 0 −4 · I + II 0 −3 −6 −4
7 8 9 0 −7 · I + III 0 −6 −12 −7
   
1 2 3 1 4 1 0 −1 −5/3
3 0 1 2 4/3 −2 · II + I 0 1 2 4/3 
−1/3 · II
0 −6 −12 −7 6 · II + III 0 0 0 1
We are done, since there are no further pivots in the coefficient matrix. As
before, the solutions to the final equation are the same as the original. Row 3
says that
0x1 + 0x2 + 0x3 = 1
and this is impossible. Thus, the equation in this problem has no solutions at
all. We say the system is inconsistent. This example shows the general form
172 10. LINEAR EQUATIONS

of an inconsistent system at the end of Elimination. There will be a row of

zeros in the coefficient matrix and a non-zero right side.
In the previous example, even though we did not find any solutions, we
still get the rank of the coefficient matrix. That rank is 2. Why?
Summary Given a system of linear equations A · X = B, we form the
 
augmented matrix A|B , where A is the coefficient part and B is the right
side. We perform Elimination, noting which variable get pivots, and ending
up with row-echelon form. If the row-echelon form has a row with 0’s in the
coefficient part and a non-zero right side, then the equations are inconsistent –
they have no solutions. Otherwise, the equations are consistent, and the row-
echelon form shows how to write the basic variables (that get pivots) in terms
of the free variables (that did not get pivots). If the equations are consistent
and there are no free variables, then there is a unique solution.
Here is one for you to do, perhaps in class.
Problem T10.2. Solve this system of equations.
− a + 2 · c − d + e = −8
4 · a + 2 · b − 10 · c − 10 · e = 10
2 · b − 2 · c − 4 · d − 6 · e = −22
a+b−3·c−2·e=3

1.3. Numerical Elimination. In class we will discuss the numerical so-

lution of equations. This is, in general, a tricky business, but we will show
how to get good information in a wide variety of problems, using the material
from the next section.

2. Matrix Inverse

The numerical equation 3 · x = 5 has a simple solution x = 5/3. This

equation comes about from multiplying both sides of 3 · x = 5 by the inverse
 2. MATRIX INVERSE 173

of 3. If, in the linear equation A · X = E, the matrix A has an inverse A−1 ,

then we might use that matrix in the same way:

A−1 · A · X = A−1 · E so that X = A−1 · E

It looks as if this solves the linear equation! The trouble is that not all matrices
have an inverse, as we will see.
Because matrix multiplication is, in general, not commutative, we need to
be pickier in the argument we just gave. Consider the equation A · X = E,
and suppose there is a matrix B such that B · A is an identity matrix. Let’s
not worry about sizes at this point, and just write B · A = I. If X = D is a
solution to A · X = E, so that A · D = E, then we can multiply on the left by
B, and get this:

D = I · D = (B · A) · D = B · (A · D) = B · E

This shows: if D is a solution to A · X = E, then D = B · E. In other words,

the equation A · X = E can have at most one solution.
Observe that
 
  1 2  
3 −2 0  1 0
(10.5) · 1 3 =
1 1 0 0 1
−4 2
Write the two matrix on the left side P, Q, so that P · Q = I2 . We can apply
the reasoning of the previous paragraph to Q: the equation Q · X = E has at
most one solution. If you do Elimination on Q, you will see that it has rank
2, so that each unknown in Q · X = E gets a pivot and a solution would have
to be unique. Sadly, the equation Q · X = E can be inconsistent; here is an
example for you to check.
   
1 2 1
 1 3 · X = 0
−4 2 0
174 10. LINEAR EQUATIONS

We are saying that equation (10.5) does not find a solution to the present
equation, since the present equation doesn’t have any solution at all.
If B · A is an identity matrix, then A · X = E has at most one solution for
each E, but sometimes there may be no solution at all.. Let’s see how to get
the existence of a solution: if there is a matrix C such that A · C is an identity
matrix, then we claim that C · E is a solution to A · X = E. Indeed,
A · (C · E) = (A · C) · E = I · E = E
Going back to (10.5), we had the equation P · Q = I2 , and so P · X = E
has a solution for all E. If you do Elimination on P , you will see that it has
rank 2, so that every row gets a pivot and there cannot be an inconsistency.
However, because there is a free variable (in the third column), the solution
will never be unique.
We can put the two ideas together: if there are matrices B, C such that
B · A and A · C are identity matrices, then every equation A · X = E will have
a unique solution. In this case, we say that A is invertible. The situation with
B, C is a lot simpler than it looks, as we will see by embarking on an argument
to establish a circle of facts that will tell us which matrices are invertible and
how to find the B, C.
Suppose that A is invertible, and get matrices B, C so that B · A and A · C
are identity matrices. Let A be m × n. (Soon we will see that m = n, but, for
now, we work in general.) The matrix B · A has n columns, and so In has to
be the identity matrix equal to that product: B · A = In . This shows that B
is n × m. Similarly, A · C has m rows, and so A · C = Im , and so C is n × m.
Next we show that B = C! Indeed, using the associative law, we compute
B = B · Im = B · (A · C) = (B · A) · C = In · C = C
 2. MATRIX INVERSE 175

This is important: if A is invertible, then the same, unique matrix multiplies

to give an identity matrix on each side of A. We write A−1 for B = C; that
matrix is the inverse of A. And it is clear that A is the inverse of A−1 .
Here is our circle of facts.

(a) Suppose that A is m × n and has an inverse. Then m = n and A has rank
n.
(b) Suppose that A is an n × n matrix of rank n. Then the row-echelon form
of A is In
(c) Suppose that the matrix A has row-echelon form equal to In . Then A has
an inverse.

Proof of (a): Consider the equation A · X = Om×1 . As we showed above,

the presence of an inverse shows that the equation A · X = O has a unique
solution. It follows that there can be no free variables in that equation, and
so A has rank n. Since the rank is at most the number of rows of A, we see
that n ≤ m.
Now apply the same argument to A−1 , which is n × m and has inverse A.
Then A−1 has rank m. It follows that m ≤ n, and we have m = n.

Proof of (b): The first pivot goes in row 1, the second in row 2, and so on.
Since the rank is n, each row gets a pivot. There are n columns, as well, and so
each column gets a pivot. The pivots, which are 1’s in row-echelon form, move
left to right as we go down the rows, and it is easy to see that the pivots have
to be at the [i, i] entries for i = 1, 2, . . . , n. This shows that the row-echelon
form of A is In .

Proof of (c): Starting with the equation A · X = In , we apply Elimination

to [A|In ]. Row-echelon form of A is In , and so Elimination gives this: [In |C].
The equation In ·X = C has the same solutions as A·X = In . We have X = C
is a solution to In · X = C, and so we have A · C = In .
176 10. LINEAR EQUATIONS

Now we think about the matrix C. Since A · C = In , we showed above

that the equation C · X = On×1 has exactly one solution X = On×1 . Thus, in
Elimination, the matrix C has rank n. By statement (b), its row-echelon form
is In . By the first part of the argument for (c) (replacing A by C), there is a
matrix D such that C · D = In . This shows that C is invertible, with A, D as
the matrices that give identity matrices. By what we showed above, it follows
that A = D, and so A · C = In and C · A = In . The matrix A is invertible!
Here is a convenient summary.
How to find an inverse Given the m × n matrix A:
(1) If m 6= n, then A does not have an inverse.
(2) If m = n, do Elimination on [A|In ] with A as coefficient matrix. If A
does not have rank n, then it does not have an inverse.
(3) If A has rank n, then the Elimination in (2) results in [In |A−1 ].
As mentioned in the Spreadsheet Appendix, if we are given an n×n matrix
A, Excel can tell us (approximately) whether A has an inverse, and it can find
the inverse, using its functions mdeterm and minverse. We keep reminding you
that all numerical work is approximate; Excel can be wrong about whether a
matrix has an inverse. Usually, if there is such an error, it is the error of finding
an inverse for a matrix that doesn’t actually have one. If the determinant
mdeterm is very small, we should be careful.

3. Applications of Linear Equations

3.1. Leontief ’s model revisited. Recall Leontief’s model economy, in-

troduced on p.159.
 3. APPLICATIONS OF LINEAR EQUATIONS 177

Problem T10.3. In a Leontief model, let the technology matrix be A and

consumption matrix C, where
   
0.1 0.5 0.1 15
A = 0.3 0 0.8  and C = 20
0.2 0.3 0.01 30
Is there a production schedule X such that A · X + C = X?

Solution. Recall that the equation for X says that it is efficient in the sense
that all that is produced is used. We can place the X as an unknown in a
linear equation:

A · X + C = I3 · X so that C = (I3 − A) · X

Excel told us the the matrix I3 − A is invertible, and it computed

X = (I3 − A)−1 · C:
 
88
X ≈ 112
82
Since the entries of X are positive, these are realistic production numbers, and
we have our solution.

3.2. Curve Fitting. We consider a general geometric problem: is there

a curve of a certain sort passing through a given set of points?

Problem T10.4. Find coefficients a, b, c, so that (1, 4), (2, 4), and (3, 10) all
lie on the curve y = a · x2 + b · x + c.

Solution. Plug the points into the equation to see what equations we need:

(1, 4) : 4 = a · 12 + b · 1 + c
(2, 5) : 5 = a · 22 + b · 2 + c
(3, 10) : 10 = a · 32 + b · 3 + c
178 10. LINEAR EQUATIONS

We get the following system of linear equations.

 
a b c
1 1 1 4
(10.6) 
4 2

1 5
9 3 1 10
Elimination (by hand) gives this row-echelon form.
 
1 0 0 2
(10.7) 0 1 0 −5
0 0 1 7
and our curve is y = 2 · x2 − 5 · x + 7.
The row-echelon form (10.7) of the coefficient matrix in the last problem
has no rows of 0’s, and so the equation (10.6) is consistent no matter what the
right side is – no matter what are the y-values of the points considered. In
other words, the existence of a solution seems to depend on the x-values.
The previous problem is a special case of a general fact about polynomials:
if we have n + 1 points (xi , yi ) with distinct x-coordinates, then there is a
unique (polynomial) curve of the form

y = a0 + a1 · x + · · · + an · xn

passing through the points. In the curve, the coefficients aj are the unknowns.
In the case n = 1, we are saying that there is a unique curve y = a0 + a1 · x
passing through two points with distinct x-coordinates – that’s a unique line
through two points. We prove the general fact in two worked-out problems.

Problem T10.5. Let x1 , x2 , . . . , xn+1 be distinct real numbers. Define the

(n + 1) × (n + 1) matrix
1 x1 x21 ··· xn1
 
1 x2 x22 ··· xn2 
V =
 ... .. .. .. .. 
. . . . 
1 xn+1 x2n+1 · · · xnn+1
 3. APPLICATIONS OF LINEAR EQUATIONS 179

Show that V has an inverse.5

Solution. The circle of facts on p.175 shows that V will have an inverse if its
rank is n + 1. That would mean that the equation V · X = O would have a
unique solution. The matrix X is (n + 1) × 1, write it

a0
 
 a1 
X=
 ... 


an

and we see that the equation V · X = O amounts to this:

a0 + a1 x1 + a2 x21 + · · · + an xn1 = 0
a0 + a1 x2 + a2 x22 + · · · + an xn2 = 0
...

a0 + a1 xn+1 + a2 x2n+1 + · · · + an xnn+1 = 0

This tells us that the polynomial a0 + a1 x + · · · + an xn has n + 1 roots: the

xj . A polynomial of degree at most n cannot have more than n roots, unless
the polynomial is the zero polynomial. In other words, all the aj have to be
0. Thus, V · X = O has only one solution: X = O, and so V has rank n + 1
and is invertible.

Problem T10.6. Let (x1 , y1 ), . . . , (xn+1 , yn+1 ), where the xj are distinct.
Then there is a unique curve y = a0 + a1 x + · · · + an xn passing through
these points.

5The matrix V is called a Vandermonde matrix .

180 10. LINEAR EQUATIONS

Solution. The fact that the curves passes through the points is that these
equations hold.

a0 + a1 x1 + · · · + an xn1 = y1
..
.
a0 + a1 xn+1 + · · · + an xnn+1 = yn

The coefficient matrix for this system is the matrix V of the previous prob-
lem. By that problem, V has an inverse, and this gives us the unique curve
coefficients ai that satisfy the equations.
Here is a problem in which there are less than n + 1 points for a curve of
degree n.

Problem T10.7. Find a curve y = a0 + a1 x + x2 x2 + x3 x3 , where a3 6= 0,

through (2.3, 6.1), (3.7, 5.3), (4.2, 3).

Solution. The equations give the following augmented matrix.

 
a0 a1 a2 a3 =
 1 2.3 5.29 12.167 6.1
 
 1 3.7 13.69 50.653 5.3
1 4.2 17.64 74.088 3
The coefficient matrix is 3 × 4, and so it cannot have an inverse. But the
determinant of the first three columns is not zero, and we can therefore set a3
to be anything we want (it is free) and solve for the other variables. Multiplying
on the left by the inverse of the 3 × 3 matrix to the left, we get the following
 
a0 a1 a2 a3 =
 1 0 0 35.742 −10.629
 
 0 1 0 −33.71 12.150 
0 0 1 10.2 −2.120
This discloses the infintely many solutions to the equations. If we pick a3 6= 0,
then that free variable determines the basic variables a0 , a1 , a2 .
 3. APPLICATIONS OF LINEAR EQUATIONS 181

When the parameters are exponents, we can use the logarithm to bring
them down as linear factors.
Problem T10.8. Fit the curve w = xa · y b · z c , where a, b, c are parameters,
to these points:
x y z w
2 3 3 4
2 5 5 4
3 4 7 6
To fit the curve, get a system of linear equations from the logarithm of the
equation.
Solution. Compute that
ln(w) = a · ln(x) + b · ln(y) + c · ln(z)
We used Excel to solve the resulting system of equations, for the coefficient
matrix has an inverse.
a ≈ 2, b ≈ 0.724541939, c ≈ −0.724541939

3.3. Current in Electrical Circuits. There are many applications of

linear equations to electronics. We introduce one application that generalizes
to transportation networks.
Mathematically, an electrical circuit is a discrete graph. This is a different
use of the word graph than in graphing a curve. In this context, a graph
consists of finite many vertices and edges. We usually think of vertices as
points; in a circuit they are junctions. An edge connects a pair of vertices. We
usually picture the edges as line segments or curve segments; in a circuit they
are the wires6 that connect junctions. Here is a pictured example; for now,
just look at the picture on the left.
6The word wire is euphemistic. A wire can stand for any electronic component, simple
or complicated, that is part of the circuit.
182 10. LINEAR EQUATIONS

e d
e d
b c
b c
a
a

Our example has three vertices (junctions) and five edges (wires). The
wires have been labeled a,b,c,d,e. Each wire has a current associated with
it – we defined this real number on p.9; its absolute value is the number of
electrons passing through that wire per unit time. To indicate the direction of
the electrons, we put an arrow on each wire. That’s the picture on the right.
Let Ia be the current in wire a. If the current Ia is going in the direction of
the arrow, then Ia > 0; if the current is going in the opposite direction, then
Ia < 0. It doesn’t matter which way our arrows go, the current values can be
adjusted appropriately.
Now we have currents Ia , Ib , and so on. As mentioned earlier in this book,
the units of current are amperes, usually abbreviated amps.
One of the basic principles of electrical circuits is called Kirchoff ’s Current
Rule.7 According to Kirchoff’s Rule, the sum of the currents coming into each
junction is equal to the sum of the currents leaving that junction.8 Here are
the equations for Kirchoff’s Rule. They are linear equations with right side 0.

top vertex Ie + Ib + Ic + Id = 0
left vertex: 0 = Ie + Ib + Ia
right vertex: Ia = Ic + Id

It turns out that we can predict which variable currents can be free in
Kirchoff’s equations and which can be basic. To do this, we choose a tree in
the graph – a set of wires that does not contain a loop – a path from one
7See [9, p.785].
8Whether a current is coming in or leaving is determined by the arrow.
 3. APPLICATIONS OF LINEAR EQUATIONS 183

junction back to that same junction. (The path does not have to follow the
arrow directions.) We give our tree as many edges as possible – it is a maximal
tree. For instance, a, c form a maximal tree, for if we try to add an additional
edge, say b, then we have a loop (b to a to c). It turns out that the variables
associated with a maximal tree can be basic in Kirchoff’s equations. We chose
a, b, and so Ia , Ic can be basic variables, and Ib , Id , Ie would be free. Here is the
augmented matrix for the equations, putting the columns for Ia and Ic first, so
they will be found basic in Elimination. The second matrix is the row-echelon
form from Elimination.
 
Ia Ic Ib Id Ie =  
0 1 0 1 0 1 0
1 1 1 1 0 0 1 1 1 1 0

−1 0 −1 0 −1 0 
 =⇒
0 0 0 0 0 0
1 −1 0 −1 0 0
Sure enough, Ia , Ic are basic. Remember that different sets of variables can be
basic. A different maximal tree would give a different set of basic variables.

3.4. Discrete Heat Distribution. We imagine a square metal plate and

suppose that the temperature is held fixed along each of the four sides –
a different temperature along each side. Thermal physics tells that that at
equilibrium the temperature at a given point on the plate is a kind of average
of the temperatures around it. We can model this using a 5 by 5 grid. We
write A, B, C, D for the constant temperatures on the sides. There are 9 points
interior to the grid, and we label those points by temperature variables. (Each
label is below and to the left of its grid point.)
184 10. LINEAR EQUATIONS

x7 x8 x9
C A
x4 x5 x6

x1 x2 x3

D
Now we interpret the condition that the temperature x1 is the average
of temperatures around it. We use the temperatures x2 , x4 , and the side
temperatures C, D:
1 h i
x1 = · x2 + x4 + C + D
4
This gives a linear equation:

4x1 − x2 − x4 = C + D

We have used C + D as the right side, since C, D are given constants. Each
of the nine grid points has such an equation, and we arrive at a system of
nine equations in nine variables. Here is the augmented matrix of this system.
Each row i comes from the average equation for grid point xi .

x1 x2 x3 x4 x5 x 6 x7 x8 x9 =
 
 4 −1 0 −1 0 0 0 0 0 C + D
−1 4 −1 0 −1 0 0 0 0 D 
 
 0 −1 4 0 0 −1 0 0 0 A + D
 
−1 0 0 4 −1 0 −1 0 0 C 
 
 0 −1 0 −1 4 −1 0 −1 0 0 
 
0
 0 −1 0 −1 4 0 0 −1 A  
0
 0 0 −1 0 0 4 −1 0 B + C 

0 0 0 0 −1 0 −1 4 −1 B 
0 0 0 0 0 −1 0 −1 4 A + B
 3. APPLICATIONS OF LINEAR EQUATIONS 185

The elimination macro can be applied to the coefficient matrix to see that
it has rank 9. In other words, our equations have a unique solution once
A, B, C, D are given.
Our model can be any sort of graph. Here is an example we like.

e f g

x7 x8 x9 d

x4 x5 x6

x1 x2 x3 c

a b

Figure 1. Heat Graph

The seven lettered vertices a, b, c, d, e, f, g are given values, and the vari-
ables xj should be determined by averaging. For instance, x1 is connected to
x4 , x5 , a, and so its equation is this:

1  
x1 = · x4 + x5 + a which is 3 · x1 − x4 − x5 = a
3
Notice that x5 is connected to six vertices, and so its value is an average of six
numbers. We will write down all the equations in class or on homework.

3.5. Transshipment. We have some warehouses holding pallets of build-

ing materials. We need to ship some of the pallets around to adjust the dis-
tribution of the pallets. Here is a discrete graph of the warehouses showing
allowable shipping routes between them.
186 10. LINEAR EQUATIONS

g
1 4
d
a e f
b
2 3
c

Here is a table giving the current supply of pallets at each vertex and the
eventual supply, accomplished by shipping.
warehouse: 1 2 3 4
current 5 25 0 50
eventual 15 5 40 20
(Notice that there are 80 pallets total in both supply lists.) We want to
determine how to ship the pallets around to get the eventual supply numbers.
We have seven routes, and we let their labels stand for the amount shipped:
a, b, . . . , g. At warehouse 1, we have 5 pallets and we need 15, so we will have
a net 10 pallets coming in. Pallets can come in along route g. Does that mean
g = 10? Maybe we ship pallets out along routes a, d. The net number of
pallets we have after shipping must be 10: −a − d + g = 10. We get similar
equations at warehouses 2,3,4. Here are all the resulting equations.
− a − d + g = 10
a − b + c = −20
b − c + d − e + f = 40
e − f − g = −30
 3. APPLICATIONS OF LINEAR EQUATIONS 187

Here is the augmented matrix of this system.

 
a b c d e f g =
−1 0 0 −1 0 0 1 10 
 
 1 −1 1 0 0 0 0 −20
 
0 1 −1 1 −1 1 0 40 
0 0 0 0 1 −1 −1 −30
Insight: we saw this type of matrix before when we discussed Kirchoff’s
equations. There, we said that we can get basic variables from a set of edges
that form a maximal tree. Thus, we know we will get a solution. But be-
cause the variables represent quantities shipped, the solution needs to be non-
negative. Here is the row-echelon form obtained by the Elimination macro.
We omitted the last row, since it was all 0.
 
a b c d e f g =
1
 0 0 1 0 0 −1 −10
0 1 −1 1 0 0 −1 10 
0 0 0 0 1 −1 −1 −30
Our free variables are c, d, f, g. We want to choose non-negative values for
them such that the basic variables a, b, e will also be non-negative. There are
several ways to do this. For instance, we have let c = d = f = 0 and g = 30.
Then a = 20 and b = 40 and e = 0.
There are many other applications of linear equations, some of which are
introduced in the problems.
 CHAPTER 11

Partial Derivatives

1. Partial derivatives

When we see an expression such as a · x2 , we are used to thinking of x as

the variable, and a as a constant. The derivative would be 2 · a · x. However,
since we are saying that a, x are both unknown, it makes sense to regard both
of them as variables. Going back to our calculation of the derivative 2 · a · x, we
were thinking of x as varying and holding a constant. We might just as well
let a be the variable and hold x constant. When we do that, the derivative
of a · x2 is x2 , since (a)0 = 1, and x2 is a constant coefficient. To distinguish
between the two possible derivatives, we call 2 · a · x the partial derivative with
respect to x, and we call x2 the partial derivative with respect to a. Here is the
notation.
∂ ∂
a · x2 = 2 · a · x and a · x2 = x2



∂x ∂a
The symbol ∂ is called a del – it is a cursive Greek d, and it is a differential
for partial derivatives.1
Problem T11.1. Compute
∂ 2 ∂ 2
y + y · z − ez y + y · z − ez



and
∂z ∂y
Solution. The partial derivative notation ∂/∂z instructs us to think of z as
the variable, and to hold the other variable y constant. To help us in looking
1We feel that the distinction between the ∂ for partial derivatives and the d for ordinary
derivatives is not as important as remembering what are the variables in a given context.
However, most people are careful about using d only when there is one variable.
189
190 11. PARTIAL DERIVATIVES

at the function this way, it might help to block out expressions in y’s.
y 2 + y · z − ez = +  · z − ez
Here is the derivative, with an explicit reminder of the basic rules for calcu-
lating it.
∂ ∂ ∂ ∂ z

+ · z − ez =




+ ·z − e sum rule
∂z ∂z ∂z ∂z
We continue:
∂
∂ z

=0+ ·z − e derivative of constant
∂z ∂z
∂
∂ z

=· z − e constant multiple rule
∂z ∂z
=  · 1 − ez derivative of z and ez
Now we remember that the  stood for y, and we have
∂ 2
y + y · z − ez = y − ez


∂z
You may or may not want to block out other variables; the point is to focus
on the variable z and ignore everything else.
To compute the other derivative, we regard y as variable and hold z con-
stant. This time, we’ll leave the constant z in the expression rather than block
it out. Notice that ez will also be constant.
∂ 2 ∂ 2
∂ ∂ z

y + y · z − ez =



y + y·z − e
∂y ∂y ∂y ∂y
∂

=2·y+z· y −0
∂y
=2·y+z
Be sure you can explain the steps here!
The calculation of partial derivatives involves exactly the same rules as
the calculation of ordinary derivatives; this section is essentially a review of
differentiation. We say that the function f of several variables is differentiable
if it has a valid partial derivative with respect to each of its variables.
 1. PARTIAL DERIVATIVES 191

Problem T11.2. Assume that x, t, y are independent variables. Compute

∂ ∂
exp(−x3 · t2 ) and exp(−x3 · t2 )
∂t ∂y
Solution. The partial derivative with respect to t holds x constant, and so
x3 is constant. We’ll use the blocking out approach;  will stand for x3 . This
derivative involves the Chain Rule:
∂
0
exp(− · t2 ) = exp0 (− · t2 ) · −  · t2
∂t
As always, the derivative of the exponential function is itself, and so we get
∂
exp(− · t2 ) = exp(− · t2 ) · −  · 2 · t


∂t
Now we replace the square with x3 to get the answer.
∂
exp(−x3 · t2 ) = exp(−x3 · t2 ) · − x3 · 2 · t


∂t
As for the derivative with respect to y, we hold x, t constant. Since there
is no y mentioned in exp(−x3 · t2 ), the entire expression is constant, and so its
derivative is 0.
Problem T11.3. At position x on the interval [0, 5], measuring cm, and at
time t seconds, suppose there is heat energy H = x · (1 − x)/(1 + t) ergs.2
What are the units of ∂H/∂x and of ∂H/∂t? An object is at x = 3 when
t = 2. In which direction should the object move to heat up? As time goes on
from t = 2, what happens to H at x = 3?
Solution. The partial derivative ∂H/∂x is an ordinary derivative, with t held
fixed. In other words, we have the definition of the derivative we had previ-
ously:
∂H ∆H
= lim with ∆t = 0
∂x ∆x→0 ∆x
The units of ∆H/∆x are ergs per cm, and so those are the units of ∂H/∂x.
Similarly, the units of ∂H/∂t are ergs per second.
2The erg is a unit of energy, equivalent to one gram·cm2 /sec2 .
192 11. PARTIAL DERIVATIVES

At x = 3 and t = 2 we contemplate moving along the axis with time

stopped. The change in H will be measured by ∂H/∂x. We write
x − x2
H=
1+t
and compute
x − x2
 
∂H ∂ 1−2·x
= =
∂x ∂x 1+t 1+t
(Notice that the denominator is constant in this calculation, so we do not use
the quotient rule. The 1 + t is a constant coefficient.) When x = 3 and t = 2
the partial derivative is −5/3 ergs per cm. Since the derivative is negative, H
is decreasing – decreasing as x increases and t stays the same. Thus, to make
H go up, we should move to the left.
Now we imagine keeping x = 3 and increasing t from 2.
∂ x − x2 x − x2
 
∂H
= =−
∂t ∂t 1 + t (1 + t)2
(For this calculation, the numerator is a constant coefficient.) When x = 3
and t = 2, the partial derivative is 6/9 ergs per second. The heat will be
increasing as time elapses from t = 2.

Problem T11.4. A plate lies over the square defined by 0 ≤ x ≤ 1 and

0 ≤ y ≤ 1 in the plane. (The variables x and y measure meters.) At point
(x, y) on the plate, the density in kg/meter2 is T = x2 − x · y − 3 · y 2 + 5. Where
is ∂T /∂x > 0 and what does that mean?

Solution. We calculate
∂T
= 2x − y kg/m3
∂x
(Make sure you understand why we get the units we do!) The inequality
∂T /∂x > 0 is seen to be 2x − y > 0, and that’s 2x > y. The inequality
describes a set of points on the square – the points below the line y = 2x.
 1. PARTIAL DERIVATIVES 193

Meaning? We are thinking about changing x but not y; we are imagining

a move horizontally. That the partial derivative is positive is that T increases
when x increases: when we move right, the density begins to increase.
Let’s review some rules of differentiation.

Problem T11.5. Find these derivatives.

∂ ln |z 2 + t| ∂ ln |z 2 + t|
   
and
∂z z·t+2 ∂t z · t + 2

Solution. We use the blocking out method on the first one, writing for t.
We need to remember that the derivative of ln |x| is 1/x.

∂ ln |z 2 + |
 

∂z z· +2
(ln |z 2 + |)0 · (z · + 2) − ln |z 2 + | · (z · + 2)0
=
(z · + 2)2
quotient rule
1  2 0 2 0

= · (ln |z + |) · (z · + 2) − ln |z + | · (z · + 2)
(z · + 2)2
rewrite
(z + )0
 2 
1 2 0
= · · (z · + 2) − ln |z + | · (z · + 2)
(z · + 2)2 z2 +
Chain, derivative of logarithm
 
1 2·z 2
= · · (z · + 2) − ln |z + | ·
(z · + 2)2 z 2 +
Power Rule, constant multiple
 
1 2·z 2
= · · (z · t + 2) − ln |z + t| · t
(z · t + 2)2 z 2 + t
unblock t
194 11. PARTIAL DERIVATIVES

We’ll do the other derivative without blocking and without giving reasons.
Make sure you understand each step, as always.
∂ ln |z 2 + t|
 

∂t z · t + 2
(ln |z 2 + t|)0 · (z · t + 2) − ln |z 2 + t| · (z · t + 2)0
=
(z · t + 2)2
 
1 1 2
= · · (z · t + 2) − ln |z + t| · z
(z · t + 2)2 z 2 + t

2. Higher Order Derivatives

Just as we sometimes need the second derivative of a function of a single

variable, so we sometimes need to take the partial derivative more than once.
This is straightforward, although the notation gets a little tricky.
For example, given f (x, y), we write
∂ 2f ∂ ∂f
2
= ·
(∂x) ∂x ∂x
to indicate taking the partial derivative with respect to x twice.
Problem T11.6. Let f (x, y) = (x + 2 · y) · exy and calculate
∂ 2f
(∂y)2
Solution. First we calculate
∂f
= 2 · exy + (x + 2 · y) · exy · x
∂y
Then we take the partial derivative with respect to y again:
∂ 2f ∂  xy xy

= 2 · e + (x + 2 · y) · e · x
(∂y)2 ∂y
= 2 · exy · x + 2 · exy · x + (x + 2 · y) · exy · x2
 2. HIGHER ORDER DERIVATIVES 195

Sometimes we want to change variables in the second derivative.

∂ 2f ∂ ∂f
= ·
∂x∂y ∂x ∂y
Notice that the denominator differentials are applied right to left.3
Problem T11.7. Let f (a, b, z) = (a + b)/(b + z 2 ). Find
∂ 2f
∂a∂z
Solution. Compute
   
∂ ∂ a+b ∂ a+b 1
2
= − 2 2
· 2z = − · 2z
∂a ∂z b + z ∂a (b + z ) (b + z 2 )2

There are times when we need more than two partial derivatives. Once
you understand the notation for the second derivative, higher derivatives are
not hard to interpret.
Problem T11.8. Let G(v, x, y) = xv 3 + yv 4 . Compute
∂ 3G
∂v∂x∂v
Solution. We see that
∂ 3G ∂ ∂ ∂G
= · ·
∂v∂x∂v ∂v ∂x ∂v
∂ ∂ 
3xv 2 + 4yv 3

= ·
∂v ∂x
∂  2
= 3v = 6v
∂v

Many of the classical mathematical models are expressed via partial differ-
ential equations, referred to as PDE’s. The unknown of a PDE is a function of
3It turns out that it usually doesn’t matter in which order this second derivative is
performed, so if you forget the ordering, you will usually be ok.
196 11. PARTIAL DERIVATIVES

several variables. The PDE gives an equation that must be satisfied by various
partial derivatives of the function. Here is an example.

Problem T11.9. Show that exp(−x · y) is a solution z to the PDE

∂z ∂z
x· −y· =0
∂x ∂y

Solution. The unknown z is a function of x, y; we know that by observing

the partial derivatives in the equation. To show that exp(−x · y) is a solution
is to show that z = exp(−x · y) satisfies the equation. We are being asked to
plug in the function! Here goes. We will be a little breezy in calculating the
partial derivatives. Make sure you can obtain the same functions.
∂z ∂
= exp(−x · y) = −y · exp(−x · y)
∂x ∂x
and
∂z ∂
= exp(−x · y) = −x · exp(−x · y)
∂y ∂y
Here is the left side of the PDE:
∂z ∂z
x· −y· = x · (−y · exp(−x · y)) − y · (−x · exp(−x · y))
∂x ∂y
= −x · y · exp(−x · y) + y · x · exp(−x · y) = 0

We get 0, as the equation says we should.

There are many famous PDE’s. We won’t try to learn a bunch of them,
but we have chosen famous examples for the problems we will do in class and
on homework.

Problem T11.10. Show that ln |x2 + y 2 | satisfies Laplace’s PDE

∂ 2z ∂ 2z
+ =0
(∂x)2 (∂y)2
 3. THE CHAIN RULE 197

Solution. Remember that to show that something is a solution of an equation,

you just plug in. We let z = ln |x2 + y 2 |; the calculation goes on for a bit.
∂2 2 2 ∂ 2·x
ln |x + y | =
(∂x)2 ∂x x2 + y 2
2 · (x2 + y 2 ) − 2 · x · (2 · x) 2 · y 2 − 2 · x2
= =
(x2 + y 2 )2 (x2 + y 2 )2
The calculation of the other second derivative is similar.
∂2 2 2 2 · x2 − 2 · y 2
ln |x + y | =
(∂y)2 (x2 + y 2 )2
The two second derivatives add to 0, as needed for Laplace’s PDE.

3. The Chain Rule

We will need a Chain Rule for functions of several variables. There is a

very general Chain Rule that actually looks very much like the one variable
Chain Rule – with matrix multiplication replacing number multiplication. For
our purposes, we will need only a special case of the general formula.
Suppose we have a function f (x, y) and that x, y are functions of the vari-
able t. Since t determines x, y, and x, y determine f , we see that f is a
(composite) function of t. Thus, it makes sense to ask for df /dt.

Chain Rule for Several Variables If f (x, y) is differentiable and if x, y

are differentiable functions of t, then
df ∂f dx ∂f dy
(11.1) = · + ·
dt ∂x dt ∂y dt

Explanation. To get the derivative on the left, we need to consider ∆f /∆t.

Given ∆t, there are changes ∆x and ∆y. In class we will show how to write
∆f as a sum of two changes: one with just x changing and not y, and the
other with y changing but not x. We will end up with a formula

∆f = ∆x f + ∆y f
198 11. PARTIAL DERIVATIVES

where the subscript indicates the variable that it changing. Then

∆f ∆x f ∆y f
= +
∆t ∆t ∆t
Inserting ∆x and ∆y and ignoring that they might be 0, we write
∆f ∆x f ∆x ∆y f ∆y
= · + ·
∆t ∆x ∆t ∆y ∆t
If we let ∆t → 0, we will see (11.1) emerge.
We will give a few more details in class, but a precise argument is beyond
us at this point. You should notice how the partial derivatives arise: we change
f by changing x first and then y.

Problem T11.11. Let f (x, y) = x2 + x · y 3 , and x = t2 + 2 and y = 1 − 3t.

Compute df /dt.

Solution. We just imitate the Chain Rule (11.1).

df ∂
d 2 ∂
d
x2 + xy 3 · x2 + xy 3 · (1 − 3t)


= t +2 +
dt ∂x dt ∂y dt
3 2
= (2x + y ) · (2t) + 3xy · (−3)

Remembering that there are formulas for x, y in terms of t, we can write the
answer as a function of t:
df
= (2(t2 + 2) + (1 − 3t)3 ) · (2t) + 3(t2 + 2)(1 − 3t)2 · (−3)
dt
As long as we remember that x, y are functions of t, either form of the answer
is fine.
We have stated the Chain Rule using two variables: x, y. Any number of
variables can be used, and it does not matter what they are called. Thus, if
we have differentiable f (x1 , . . . , xn ), and if each xi is a differentiable function
of t, then the Chain Rule asserts that
df ∂f dx1 ∂f dx2 ∂f dxn
= · + · + ··· ·
dt ∂x1 dt ∂x2 dt ∂xn dt
 3. THE CHAIN RULE 199

Problem T11.12. Given F (a, b, c), suppose that a, b, c are functions of the
variable w. Find dF/dw.

Solution. We use a, b, c instead of x1 , x2 , x3 and w in place of t. Then

dF ∂F da ∂F db ∂F dc
= · + · + ·
dw ∂a dw ∂b dw ∂c dw

Here is a rather typical manipulation.

Problem T11.13. Suppose that we have g(x, y) and we define h(x) = g(x, x).
Find h0 (x) in terms of the partial derivatives of g.

Solution. Let x = t and y = t, so that x, y are functions of t. Then

dg ∂g dx ∂g dy ∂g ∂g
= · + · = ·1+ ·1
dt ∂x dt ∂y dt ∂x ∂y
Since x = t = y, we have g(x, y) = h(x) in this case. Thus, h0 (x) = dg/dt, and
so
∂g ∂g
h0 (x) = +
∂x ∂y

You might try a specific instance of the previous problem. Let g(x, y) =
x + y 2 , so that g(x, x) = x + x2 . You can verify that
∂(x + y 2 ) ∂(x + y 2 )
(x + x2 )0 = + where y = x
∂x ∂y
Suppose we are standing at (3, 2) in the xy-plane. We contemplate moving
away from this point in some direction. One natural way to indicate a direction
is to describe the changes in x and y that would occur in one time unit. If the
changes are to be (a, b), then consider

x=3+a·t
y = 2 + b · t where t ≥ 0
200 11. PARTIAL DERIVATIVES

Notice that when t = 0 we have x = 3, y = 2, so we start at the point we

are standing on. When t = 1, we have x = 3 + a, so that ∆x = a, and we
have y = 2 + b, so that ∆y = b. Thus, a, b are the changes at time one, as we
intended. As t moves from 0 to 1, we move along the line segment from (3, 2)
to (3 + a, 2 + b).
We belabor the idea of direction a little more. Thinking of the x-axis and
y-axis as compass directions, let’s suppose we are at (5, 6) and we want to go
northeast. Here are some points that are northeast from (5, 6):

(5.5, 6.5) (6, 7) (6.23, 7.23) ···

It is easy to see how to get such points: add the same positive quantity to both
5 and 6. The general such point would be (5+c, 6+c) for some positive number
c. Notice that this idea doesn’t involve time, as did the idea of the previous
paragraph. That idea suggests moving along x = 5 + c · t and y = 6 + c · t. The
various values of c tell how far we want to go in one time unit; the larger c
is, the farther we go. What this really means is that the specific c determines
our speed.
Now suppose we have f (x, y) and we want to know how f changes as we
move northeast from (5, 6). Let x = 5 + c · t and y = 6 + c · t. Then

df ∂f dx ∂f dy
= · + ·
dt ∂x dt ∂y dt
∂f ∂f
= ·c+ ·c
∂x ∂y
 
∂f ∂f
= + ·c
∂x ∂y

We see that the answer depends on c, as is not surprising. If all we are

interested in is whether f increases or decreases, then since c > 0, we get the
same sign of df /dt no matter what c we use.
 3. THE CHAIN RULE 201

Problem T11.14. Suppose that F = x2 · y + y 3 measures the temperature in

degrees Celsius at each point (x, y) in the plane, where x, y measure meters.
Suppose we are at (1, −1). As we move in the direction (−3, 4), what happens
to the temperature?

Solution. To move from (1, −1) in direction (−3, 4), we let x = 1 − 3 · t and
y = −1 + 4 · t. The derivative dF/dt is the change in temperature with respect
to t. At t = 0, that derivative measures the change in F at the point (1, −1).
Thus, we compute
dF ∂F dx ∂F dy
= · + · = 2xy · (−3) + (x2 + 3y 2 ) · 4
dt ∂x dt ∂y dt
When t = 0 we have x = 1 and y = −1 and dF/dt = 22. We see that F will
increase.
It is instructive to rework the previous problem using a different set of
changes in the same direction as (−3, 4). For instance, we can move in the
direction (−6, 8) or in the direction (−1.5, 2). You should see that you get
different values of dF/dt but that the sign stays the same.

Problem T11.15. The x, y-plane represents a plot of land. The equation

z = 100 − x2 + 4x − y 2 describes the height in yards at the point (x, y) on the
land, where x, y measure yards, as well. Suppose we are standing at (3, 2). In
which directions (a, b) will we move down the hill?

Solution. We let

x = 3 + a · t and y = 2 + b · t for t ≥ 0

With these equations, z is a function of x, y and x, y are functions of t, and so

the Chain Rule applies:
dz
= (4 − 2x) · a + (−2y) · b
dt
202 11. PARTIAL DERIVATIVES

At the point (3, 2), we see that

dz
= −2 · a − 4 · b
dt
The problem asks us for directions that move down the hill: that means
that dz/dt < 0.
1
−2 · a − 4 · b < 0 which is − 2 · a < 4 · b and that’s
·a<b −
2
Remembering that a is change in x and b is change in y, we can graph the
inequality as direction points above the line b = −a/2.

Problem T11.16. A particle moves around the circle x2 + y 2 = 4, where x, y

are functions of time t. The direction of motion is defined to be
 
dx/dt | dy/dt

Show that the direction of motion is always tangent to the circle.

Solution. We let z = x2 + y 2 , and then

dz dx dy
= 2x · + 2y ·
dt dt dt
On the other hand, z = 4, since we stay on the circle. Thus, dz/dt = 0.
Therefore,
dx dy
2x ·
+ 2y · =0
dt dt
6 0, then we can solve for the slope of the direction:
If y =
dy/dt x
=−
dx/dt y
When (x, y) are on the circle, the slope of the radius from the center (0, 0) to
(x, y) is y/x. Then −x/y is perpendicular to the radius: tangent to the circle.4
4If
x = 0, so that y/x is not defined, then we are at one of the points (0, ±2) on the circle,
where the tangent is horizontal. In that case, −x/y = 0 and the direction is horizontal, as
well.
 3. THE CHAIN RULE 203

If y = 0, then we see that x = ±2, and the Chain Rule equation says
that 2x(dx/dt) = 0. Thus, dx/dt = 0. The tangent to the circle at (±2, 0) is
vertical.
Given a function of several variables, its derivative is the matrix row of its
partial derivatives. The derivative of the function f is often denoted Df ; we
don’t usually employ the prime notation f 0 that was used for functions of one
variable. If f (x, y, z) = x2 + y 2 + x · z, then the derivative Df is a 1 × 3 matrix:
h i 
Df = ∂f ∂f ∂f

∂x ∂y ∂z = 2x + z 2y x

There isn’t much to this beyond a convenient notation that we will use in
the next chapter.
 CHAPTER 12

Non-Linear Optimization

In Chapter 6 we considered linear optimization problems; in this chapter

we consider non-linear problems. In general, those problems are harder: there
are many more possible nuances and there is no universal algorithm for solving
them, even numerically. As before, the objective of an optimization problem
is a function of the variables of the problem. The typical problem also has a
list of constraints on the variables. A constraint is an equation or inequality
in one or more of the variables. A constraint is closed if it is an equation or
a non-strict inequality (one of the forms A = B and A ≤ B and A ≥ B). A
constraint is open if it is a strict inequality (A < B or A > B).
As in Chapter 6, there are a host of applications. For example, see [10,
Sections 8.3 and 14.2] for optimization problems in economics. In class we will
discuss using the Excel Solver to obtain approximate solutions. Its use here is
reminiscent of its use on the linear problems of the previous course, but there
are important differences.

1. The First Derivative Test

We deal first with optimization problems for which all the constraints are
open, or for which there are no constraints at all. Going back to the case of
a single variable, the minimum or maximum of a function in an open interval
occurs at a critical point – a point where the derivative is 0. Remember that
the converse is false: we can have derivative zero without having an extreme.
(e.g. y = x3 at x = 0)

205
206 12. NON-LINEAR OPTIMIZATION

Throughout this chapter we will write x for a list of variables

x = (x1 , . . . , xn )

Then f (x) = f (x1 , . . . , xn ) will be a function of those variables. We know that

Df (x) is a 1 × n matrix of partial derivatives. A critical point is a specific
point p such that Df (p) = O1×n . This equation is really n equations, one for
each of the partial derivatives:
∂f
(p) = 0 for j = 1, 2, . . . , n
∂xj

Problem T12.1. Find any critical points for this function.

f (x1 , x2 , x3 ) = x31 − x1 + 2 · x22 − 2 · x2 · x3 + 3 · x23 − 4 · x2 − 18 · x3 + 5

Solution. We need to solve the equation Df = O. Here are the three equa-
tions, one for each partial derivative.
∂f
0= = 3 · x21 − 1
∂x1
∂f
0= = 4 · x2 − 2 · x3 − 4
∂x2
∂f
0= = −2 · x2 + 6 · x3 − 18
∂x3
√
The first equation gives x1 = ±1/ 3. The second and third equations yield
x2 = 3, x3 = 4. Thus, there are two critical points:
√ √
(1/ 3, 3, 4) and (−1/ 3, 3, 4)

As with functions of one variable, so with functions of several variables: if

f (x) has an extreme (maximum or minimum) at the point x = p, in the pres-
ence of open constraints, then Df (p) = O. This is called the First Derivative
Test. As with functions of one variables, we can have Df (p) = O without
x = p being an extreme.
 1. THE FIRST DERIVATIVE TEST 207

First Derivative Test If f (x) is a differentiable function and has an ex-

treme at x = p, subject to open constraints, then Df (p) = O.
The fact that the logic goes only one way restricts the use of the Test.
Problem T12.2. Could a = 1 and b = −2 give the maximum of ab2 + 2ab?
Solution. We have no constraints, and so the constraints are open. Compute
D(ab2 + 2ab) = b2 + 2b | 2ab + 2a
 
 
At a = 1 and b = −2, we have D(ab2 + 2ab) = 0 −2 . Since this is not O,
we cannot have a maximum. (Or a minimum, either!)
The First Derivative Test suggests solving Df (x) = O; just remember
that the solutions are not necessarily meaningful. But consider the following
problem.
Problem T12.3. Suppose we know that the following problem has a solution:
Minimize
Z = 3 · x4 − 16 · x3 + 18 · x2 + 6 · y 2 + 36 · y + 2 such that x < 2
Find that solution.
Solution. The constraint is open, and so the First Derivative Test applies.
Compute
∂Z ∂Z
= 12 · x3 − 48 · x2 + 36 · x and = 12 · y + 36
∂x ∂y
To find the critical points, we set the partial derivatives equal to 0. Here is
what we get: x = 0, 1, 3 and y = −3. Remembering the constraint x < 2, we
see that x = 0, 1 and y = −3. This gives us two critical points (0, −3) and
(1, −3). We know that the mimimum of Z has to be one of these, but which
one? We have only to compare the values of Z at each point:
Z(0, −3) = −52, Z(1, −3) = −47
The winner is Z = −52 at (0, −3).
208 12. NON-LINEAR OPTIMIZATION

It was very significant that we knew there was a solution in the previous
problem. Otherwise, the values we calculated at the critical points did not
have to be extremes. For a problem with open constraints, this can be a
tricky business. For a problem all of whose constraints are closed, there is
a very important theorem that is relevant. To give that theorem, we need a
term: we say that the constraints of an optimization problem are bounded if
each variable is confined to a closed interval.
The Extreme Value Theorem If f (x) is a differentiable function sub-
ject to closed and bounded constraints, then f (x) has both a maximum and
minimum subject to the constraints.
Problem T12.4. Consider the problem: Maximize x4 ·y 5 , where x2 +2·y 2 ≤ 9.
Does the Extreme Value Theorem apply to this problem?
Solution. The single constraint is closed; it implies that x2 ≤ 9, so that
√ √
−3 ≤ x ≤ 3. Similarly, we have 2 · y 2 ≤ 9, so that −3/ 2 ≤ y ≤ 3/ 2, so
that the constraints are bounded.1 Thus, the Extreme Value Theorem applies.
The problem necessarily has a solution.
Let’s take the previous problem a little further. The constraint defines
an ellipse and its interior. First suppose that the maximum occurs inside the
ellipse – where the constraint is open: x2 +2·y 2 < 9. Then the First Derivative
Test applies and the solution occurs at a critical point, where D(x4 · y 5 ) = O.
On the other hand, the maximum might occur on the ellipse – where the
constraint is x2 + 2 · y 2 = 9. In the next section we will discuss a method for
dealing with both of these cases simultaneously.
It is instructive to use the Solver on the previous problem. When we
started with x = y = 0, the Solver found x = y = 0 as solution. That’s
obviously not correct – perhaps the Solver was confused by the fact that all
1
√Note that we don’t have to show that y can actually attain all the values between
±3/ 2; it is enough to show that y’s values have to come from that closed interval.
 2. LAGRANGE MULTIPLIERS 209

points on both the x-axis and y-axis are critical points. When we started
x = 1 = y, the Solver found x = 2, y = 1.581 and objective 158.1 as solution.
That looks more plausible. We notice that this point is on the ellipse.
We use the Exteme Value Theorem in all the problems we consider in the
rest of this chapter.

2. Lagrange Multipliers

We describe Lagrange’s equation – an equation that usually holds at the

solution of an optimization problem. This equation involves the Lagrange
multipliers that we introduced in Chapter 6. It is hard to motivate the equation
in general, and so we will focus on using it.

2.1. Equation Constraints. Let’s begin with an example.

Problem T12.5. Minimize Z = x2 + y 2 where (x, y) is on the line

y = (13 − 2 · x)/3.

Solution. Notice that the single constraint is closed, but it is not bounded.
Thus, the Extreme Value Theorem does not apply. We will discuss the exis-
tence of a solution later; for now, we will assume that the minimum exists.
Now to Lagrange’s method. We write this constraint as a function set
equal to a constant:
3 · y + 2 · x = 13
In this form, the constraint gets a Lagrange multiplier λ. (It is customary to
use the Greek letter lambda as a multiplier; L for Lagrange, we think.) The
multiplier is an additional variable in the problem. The Lagrange equation
says, at a solution to the problem, that the derivative of the objective is the
multiplier times the derivative of the function side of the constraint.

(12.1) DZ = λ · D(3 · y + 2 · x)
210 12. NON-LINEAR OPTIMIZATION

Here are those derivatives.

   
DZ = 2 · x 2 · y and D(3 · y + 2 · x) = 2 3
Matrix equation (12.1) is a scalar multiple equation. Here is what it says:
2·x=λ·2
2·y =λ·3
Along with these equations, we have the line equation 3 · y + 2 · x = 13. There
is no one way to solve these equations. One method that sometimes works in
harder problems is to solve for x, y in terms of λ. We get x = λ and y = 3·λ/2.
These expressions can be substituted into the line equation:
3
3 · · λ + 2 · λ = 13 which is λ = 2
2
Then x = λ = 2 and y = 3 · λ/2 = 3, and we have (x, y) = (2, 3), and
Z = 32 + 22 = 13. This is the solution given by Lagrange’s equation.
In the previous problem, notice that Z is the square of the distance from
(x, y) to (0, 0), and so the point of minimum Z is the point closest to (0, 0).
You may remember from geometry how to find that closest point: construct
a perpendicular from (0, 0) to the line. The line in question has slope −2/3,
and so a perpendicular has slope 3/2. If we intersect the perpendicular line
with the constraint line, we get the point (2, 3) as the closest point. Thus,
in this example we have a way to show that the Lagrange equation solution
is, in fact, the actual solution to the problem. It is usually the case that the
problem solution has to satisfy the Lagrange equation. Unfortunately, it is
quite technical to give conditions for the usually. Most applied problems are
done under the assumption that the solution satisfies the Lagrange equation,
without worrying about technicalities.
In the solution to the last problem we hinted that there is no one universal
method for solving the Lagrange equations. We won’t do many problems by
hand, in any case. Numerical solvers have their own tricks.
 2. LAGRANGE MULTIPLIERS 211

Let’s do an example with two constraints.

Problem T12.6. Minimize x2 + 2 · y 2 + z 2 where x + 3 · z = y + 6 and

2 · x + y − z = −7/2.

Solution. Write Q = x2 + 2 · y 2 + z 2 for the objective. The constraints are

closed; each needs to be written as a function set equal to a constraint. For
the first constraint we move the y to the left. Now the function side of each
constraint gets a Lagrange multiplier.

λ for x − y + 3 · z
µ for 2 · x + y − z

Here is the Lagrange equation in this case.

(12.2) DQ = λ · D(x − y + 3 · z) + µ · D(2 · x + y − z)

Taking the partial derivatives in x, y, z, in turn, to calculate the derivatives,

we get these equations.

2·x=λ·1+µ·2=λ+2·µ
4 · y = λ · (−1) + µ · 1 = µ − λ
2 · z = λ · 3 + µ · (−1) = 3 · λ − µ

As we mentioned, there is no universal method for solving the Lagrange equa-

tions. In this problem, we solve the equations for x, y, z and then substitute
into the two constraints. Here is what results.
λ+2·µ µ−λ 3·λ−µ
x= y= z=
2 4 2
λ+2·µ µ−λ 3·λ−µ
− +3· =6
2 4 2
λ+2·µ µ−λ 3·λ−µ 7
2· + − =−
2 4 2 2
212 12. NON-LINEAR OPTIMIZATION

Clearing all the fractions we get these equations.

21 · λ − 3 · µ = 24
−3 · λ + 11 · µ = −14
We get λ = 1 and µ = −1, which gives x = −1/2 and y = −1/2 and z = 2
and Q = 19/4.
When we used the Solver on this problem, we obtained the same answers,
approximately. Remember that the Lagrange multipliers are available on the
Sensitivity page.
Equations (12.1) and (12.2) are special cases of the general, abstract form
of Lagrange’s equation. Each constraint gets a Lagrange multiplier. The
derivative of the objective is the sum of the derivatives of the constraints –
each constraint derivative multiplied by its multiplier.

2.2. Inequality Constraints. Inequality constraints are very common;

let’s show how to handle them. The short version: each constraint gets a La-
grange multiplier, as before, and Lagrange’s equation holds, as before. There
is an additional condition, however. Watch for it.
Problem T12.7. Maximize x2 + y 2 such that y − x2 ≤ 10 and x + y = 3.
Solution. Graphing the constraints shows that they are closed and bounded,
and so the Extreme Value Theorem tells us that there is a maximum. We are
going to focus on the Lagrange equations, rather than on solving the problem.
We turn the inequality y − x2 ≤ 10 into an equation: y − x2 + t2 = 10.
The variable t is a new variable being introduced temporarily; it will go away
eventually. Since we have t2 ≥ 0, the equation y − x2 + t2 = 10 is equivalent
to the inequality y − x2 ≤ 10. Later, we will use the fact that if y − x2 < 10,
then t 6= 0. The variable t is called a slack variable.
Here is the modified problem: maximize x2 + y 2 such that y − x2 + t2 = 10
and x + y = 3. Each constraint gets a multiplier: λ1 for y − x2 + t2 = 10 and
 2. LAGRANGE MULTIPLIERS 213

λ2 for x + y = 3. We have three variables (x, y, t) and the Lagrange equation

is this:
D(x2 + y 2 ) = λ1 · D(y − x2 + t2 ) + λ2 · D(x + y)
We compute the derivatives
     
(12.3) 2x 2y 0 = λ1 · −2x 1 2t + λ2 · 1 1 0

Look at the third entry: 0 = λ1 · (2t) + λ2 · 0 = 2 · λ1 · t, so that

0 = 2 · λ1 · t

and that says that either λ1 = 0 or t = 0. We noted above that if y − x2 < 10,
then t 6= 0, and so λ1 = 0 in this case.
These conditions can be stated without reference to the slack variable t.
We go back to the original variables x, y, and write (12.3) in terms of those
variables alone.
     
(12.4) 2x 2y = λ1 · −2x 1 + λ2 · 1 1

And we add the condition that if y − x2 < 10, then λ1 = 0.

The problem we just did shows how to handle inequality constraints: they
get a multiplier just as equation constraints do. If the inequality is a proper
inequality, then the multiplier is 0.
The Solver found this solution to the problem:

x ≈ 1.5, y ≈ 1.5, λ1 ≈ 0, λ2 ≈ 3

The maximum objective was ≈ 4.5. At the solution, the constraint y −x2 ≤ 10
is a proper inequality, and so its multiplier is 0. Observe that the equation
(12.4) holds at the solution.
Here is a problem that involves pretty much everything we have discussed.
We remind you that we will not do many problems by hand, but it will be
helpful to see the Lagrange equation method in action.
214 12. NON-LINEAR OPTIMIZATION

Problem T12.8. Find the maximum and minimum of 2x − y, where we have

y − x2 ≥ 0 and y − x ≤ 2. How do we know there is a maximum and minimum,
by the way?

Solution. Let Q = 2x−y, so that Q is the objective. It is not hard to see that
the constraints are closed and bounded. They consist of the region between
y = x2 and the line y = 2 + x. Thus, the Extreme Value Theorem says that
the maximum and minimum exist.
Let λ1 be the multiplier for y − x2 ≥ 0 and let λ2 be the multiplier for
y − x ≤ 2. Then

D(2x − y) = λ1 · D(y − x2 ) + λ2 · D(y − x)

     
2 −1 = λ1 · −2x 1 + λ2 · −1 1

We also have these conditions: if y − x2 > 0, then λ1 = 0 and if y − x < 2,

then λ2 = 0.
The conditions suggest cases:

Case 1 Let y − x2 > 0 and y − x < 2.

We have λ1 = 0 = λ2 , and the Lagrange equation is
   
2 −1 = 0 0

That is impossible: there are no points given in this case.

Case 2 Let y − x2 = 0 and y − x < 2.

We have λ2 = 0, and the equation:
   
2 −1 = λ1 · −2x 1

The second entry says that λ = −1, and the first entry equation is 2 =
(−1) · (−2x) = 2x. Thus, x = 1, and so y − x2 = 0 says that y = 1. The
objective is Q = 1 in this case.
 2. LAGRANGE MULTIPLIERS 215

Case 3 Let y − x2 > 0 and y − x = 2.

In this case, λ1 = 0 and the equation is
   
2 −1 = λ2 · −1 1

The entries disagree on the value of λ2 . There are no points in this case.

Case 4 Let y − x2 = 0 and y − x = 2.

We substitute y = x2 into the second equation: x2 − x = 2, so that
x2 − x − 2 = 0 and that’s (x − 2)(x + 1) = 0. Then x = 2, −1. When x = 2,
y = 4, and the objective is Q = 0. When x = −1, we have y = 1 and the
objective is Q = −3.
Looking at all the cases, we see that the maximum objective is 1, at (1, 1),
and the minimum objective is −3 at (−1, 1).
Notice, in Case 4, when x = 2 and y = 4, the Lagrange equations are
satisfied, but this point gives neither a maximum nor a minimum for the
objective.
The Solver finds both the maximum and minimum, and it finds the La-
grange multipliers in each case. It is interesting to observe that Lagrange’s
equation holds in each case.

When do the Lagrange equations hold?

As we mentioned, this issue is usually ignored. And it’s somewhat compli-
cated to give a complete set of conditions under which the equations have to
hold. Here is a fairly simple criterion: assume we have k equation constraints:
g1 (x, y, z) = c1
g2 (x, y, z) = c2
..
.
gk (x, y, z) = ck
where c1 , . . . , ck are constants. We have used three variables; it does not matter
how many we have.
216 12. NON-LINEAR OPTIMIZATION

We bundle the derivative of each gj into one matrix:

Dg1
 
Dg2 
G=  ... 


Dgk
If the rank of G is k at the maximum or minimum of the objective, then the
Lagrange equations have to hold. Thus, to conduct a complete use of the
Lagrange equations, we need to check points where Dg has rank less than
k, plugging them into the objective to see if the minimum/maximum occurs.
Typically, there are no such points, but, to be complete, this case should be
considered.

2.3. Shadow Prices. Just as in the case of linear optimization problems,

the Lagrange multipliers tell what happens when a parameter is changed. Let’s
see how this works in a simple problem.
Let p be a constant initially 16, and we want to maximize the product of
two positive numbers whose sum is p.
If x, y are the two positive numbers, our objective Z = x · y. We have
x + y = p and x, y > 0. In getting the Lagrange equations, we only deal with
the closed constraint. Let λ be the multiplier, and the Lagrange equation is
this:
DZ = λ · D(x + y)
and that’s
y = λ and x = λ
The constraint is then λ + λ = p, so that
p
(12.5) λ=
2
This gives x = p/2 and y = p/2 and Z = p2 /4. Notice that x, y are positive,
so we have satisfied all the constraints.
 2. LAGRANGE MULTIPLIERS 217

We write Z̄ = p2 /4 for the maximum value of Z. The shadow price of the

constraint x + y = p is the derivative
dZ̄ p
=
dp 2
We said that p = 16 initially; that means that we expect Z̄ to increase by 8
units for each unit increase in p.
Equation (12.5) shows that the shadow price is the Lagrange multiplier. We
claim that this is what happens in general. In linear optimization problems,
the shadow price is the exact amount of increase in the optimum objective. In
non-linear problems, it is an approximation.
Summary: In an optimization problem, we imagine that the constant side
of one of the constraints is a parameter that might change. Then the Lagrange
multiplier for that constraint, at a solution to the problem, gives the derivative
of the optimal objective with respect to that parameter.

Problem T12.9. We sell two related products A and B. Here are the demand
curves for each; they show the demand as a function of the price. The subscript
1 goes with A, 2 goes with B.
p2 1
x1 = 6 · 105 · · 2 and x2 = 50 · exp(−p2 /200)
p2 + 400 p1 + 10
It takes 10 hours to make one unit of A and 40 hours for one unit of B. We
have up to 1000 hours for production of A and 500 hours for B.
(a) Find the prices and demand that maximize revenue.
(b) What would we be willing to pay for 10 more hours for producing A?

Solution. The objective is revenue R, where

R = x1 · p 1 + x2 · p 2

Constraints? The two demand functions, and the production inequalities:

10 · x1 ≤ 1000 and 40 · x2 ≤ 500

218 12. NON-LINEAR OPTIMIZATION

In the Solver we set the variables p1 , p2 to be non-negative. Our solution

was this.
p1 ≈ 49.88 x1 ≈ 100 p2 ≈ 285.33 x2 ≈ 12
with maximum revenue R̄ ≈ 8414.
Regarding as a parameter h the 1000 hours for production of A, an addi-
tional 10 hours is ∆h = 10. The Solver tells us that the Lagrange multiplier
of the constraint 10 · x1 ≤ h is 2.484. Thus
∆R̄
≈ 2.484
∆h
Since ∆h = 10, we see that ∆R̄ ≈ 24.84. That is what we would be willing to
pay for the additional hours.

3. Fitting a Model to Data

On p.17 of Chapter 2 we introduced a simple recursion to monitor the value

of an investment under a fixed interest rate. Suppose we have an investment
which is supposed to be following that model according to a daily interest
rate. We’ll assume that we don’t know the exact interest rate; its existence is
a theoretical fact. Suppose we start with $100, and let Mn be the value of the
investment after n days. (Then M0 = 100.) If the daily interest rate is r, then
we know that

M1 = (1 + r) · M0 , M2 = (1 + r) · M1 , M3 = (1 + r) · M2 , ...

Now suppose we observe the actual value of the investment each day. Here
is what we observe over the first few days.
day n : 0 1 2 3 4 5
value Mn : 100 101 100 105 110 111
We want to know the interest rate r. We are supposed to have

M1 = (1 + r) · M0 which is 101 = (1 + r) · 100

 3. FITTING A MODEL TO DATA 219

This gives r = 0.01. That was easy. But we also want to have

M2 = (1 + r) · M1 which is 100 = (1 + r) · 101

which solves as r ≈ −0.001. (Apparently, day 2 was a bad day.) Then again

M3 = (1 + r) · M2 which is 105 = (1 + r) · 100

and this is r = 0.05. There is no one value of r that works in each equation!
How are we supposed to find r, given that the observed data don’t agree on
its value?
Let’s back up and realize that we should have expected something like this
to happen. Remember that we said that the existence of r is a theoretical
fact – that probably means that the value of the investment is determined by
several things and that the interest rate is an important factor, maybe the
most important factor – but perhaps it’s not the only factor. In that case, the
interest rate wouldn’t completely determine the sequence, and so we should
not be surprised at the inconsistencies. Here is another possible explanation:
suppose that the observed values are only estimates (approximations) of the
actual values. There might be errors in those estimates that make the equa-
tions inconsistent.
The first sort of error – that theory only approximates fact – suggests that
the theoretical r is approximate; the second sort of error – that observations
can be rough – suggests that the observations might not reveal the actual
r. Either way we are stuck with trying to find a theoretical parameter from
inconsistent data.
The situation we have just described is typical of virtually all experiments
in both the natural and social sciences. Some kind of underlying theory pre-
dicts what should happen, but not exactly; our observations show us what
actually happened, but not exactly. Experience shows that even when the
theory is rock solid and the observations are as precise as we can make them,
220 12. NON-LINEAR OPTIMIZATION

some sort of discrepancy between them inevitably occurs. If the theory is

important to us, we are still interested in doing the best we can to fit the ob-
served data to the theoretical model. This means getting a kind of best guess
as to the value of various parameters. We are going to describe how to do this.
First, we need to measure a kind of distance between two sequences. Second,
we describe an optimization problem to find the best fit of a theoretical model
to observed data. This subject dovetails with statistics, in particular with
regression analysis, and many of you will see that connection in other courses.

3.1. Square Norm. We need to define a kind of distance between two

sequences of the same length. Given

P0 P1 P2 · · · Pn
and Q0 Q1 Q2 · · · Qn

we define their square norm

(P0 − Q0 )2 + (P1 − Q1 )2 + (P2 − Q2 )2 + · · · + (Pn − Qn )2

Notice that this distance is a sum of squares – that is to say it is a sum of

non-negative numbers. If a sum of non-negative numbers is 0, each of the
numbers is 0. In other words, if the square norm is 0, then each term Pj − Qj
is 0, so that Pj = Qj , for each j; the sequences are the same. If the sequences
are not the same, then some difference Pj − Qj is not 0, and the square norm
will be positive. The greater the difference between the sequences, the larger
will be the square norm.
The word norm is a general word used for measures of distance between
two objects. The formula for the square norm originates in both geometry and
statistics; we will not try to explain how. Here is an example that illustrates
the most common use of the square norm.
 3. FITTING A MODEL TO DATA 221

Example. Going back to our investment example, here is what we observed

over the first few days.
day: 0 1 2 3 4 5
value: 100 101 100 105 110 111
Suppose we hypothesize a daily interest rate of 1% and we want to measure
the discrepancy between that hypothesis and our observation. Here is the 1%
sequence; we regard this sequence as predicted or theoretical.
day: 0 1 2 3 4 5
value: 100 101 102.01 103.03 104.06 105.1
Here is the square norm between the predicted sequence and the observed
sequence.

(100 − 100)2 + (101 − 101)2 + (102.01 − 100)2

+ (103.03 − 105)2 + (104.06 − 110)2 + (105.1 − 111)2
≈ 78.01

We are not yet in a position to think about the size of the norm (78 may seem
like a large number). The point is that 78.01 measures the distance between
the predicted sequence and the observed sequence.
When, as in the previous example, the square norm measures the distance
between predicted values and observed values, the norm is often called the
squares error .

Example. We bring back the Logistic Equation from p.18 of Chapter 2. We

had a population sequence Pn governed by the recursive equation
 
Pn
Pn+1 = Pn + k · Pn · 1 −
E
where the numbers k and E are constants. Suppose we know that P0 = 10,
and the claim is made that k = 0.1 and E = 50. With those values of k, E,
222 12. NON-LINEAR OPTIMIZATION

here is the predicted sequence.

n: 0 1 2 3 4 5 6
claimed Pn : 10 10.8 11.65 12.54 13.48 14.64 15.49
Now suppose we observe the following population sequence:
n: 0 1 2 3 4 5 6
actual Pn : 10 11 12 13 13 15 16
We can compute the squares error (the square norm) between the predicted
sequence and the observed sequence to measure the plausibility of the claim
about the values of k, E. The squares error is 0.99 in this case. As in the pre-
vious example, we are not yet in a position to do anything with that number.

Example. We think that the weight w (lbs) of a seven year old boy is predicted
by his height h (inches) via the formula w = 5 · h − 190. (So, this equation
is theoretical.) As an experiment, we record the height and weight of four
seven-year-olds in this chart:
height : 47 49 45 46
observed weight: 45 50 38 45
The equation w = 5h − 190 predicts the following weights:
height : 47 49 45 46
predicted weight: 45 55 35 40
The squares error measures the distance between the observed weights and the
theoretical weights: (0)2 + (5)2 + (3)2 + (5)2 = 59.

3.2. Least Squares. We have described the situation where we have some
sort of observed sequence and a theoretical prediction that it should be ex-
plained by particular parameters. In the height/weight example, we had a
particular theoretical curve w = 5 · h − 190; let’s only assume that the curve
should be a line w = m · h + b and we’ll try to find the parameters m, b.
 3. FITTING A MODEL TO DATA 223

To do this, we imagine picking a particular (but unknown) value of m, b and

computing the predicted weights from the given heights:
height : 47 49 45 46
weight: 47m + b 49m + b 45m + b 46m + b
To test the fitness of a particular m, b, we compute the squares error with the
observed data.
(47m + b − 45)2 + (49m + b − 50)2 + (45m + b − 38)2 + (46m + b − 45)2
Notice that this is a function of the parameters m, b. Call it E(m, b). To find a
best possible m, b, we minimize E(m, b). If we are able to do this, the solution
m, b is called the least squares solution. The least squares solution finds the
theoretical model that best fits the observed data.2
Minimizing E(m, b) is a non-linear optimization problem with no con-
straints. One can prove that there is a minimum, and so the First Derivative
Test says that DE(m, b) = O at the solution. The resulting equations (two
equations, one for each partial derivative) are called the normal equations. We
usually allow a numerical solver to find the solution rather than solving the
normal equations by hand.
Problem T12.10. Find m, b that minimize the squares error in the
height/weight problem.
Solution. We develop numerical formulas to imitate what we just did. Here
is what the spreadsheet might look like, showing the row and column labels.
A B C D E F
1 m b
2 0 0
3 h: 47 49 45 46
4 w(observe): 45 50 38 45
5 w(predict): =$A$2*B3+$B$2 ? ? ?
6 delta: =B4-B5 ◦ ◦ ◦ 
2See [4, Section 8.2].
224 12. NON-LINEAR OPTIMIZATION

This is not easy to read; study it carefully! The value of m is stored in A2 and
the value of b in B2. The formula in B5 computes w = m · h + b for h = 47 in
B3. The cells marked ? were obtained by fill-right from B5. That’s why
the references to m, b have dollar signs. Cell B6 holds the difference between
the predicted w (in B5, for h = 47) and the observed w in B4. The entries in
row 6 marked with ◦ were obtained by fill-right from B6. The objective
square norm is in cell F6; the  symbol stands for this:

=sumproduct(B6:E6,B6:E6)

We asked the Solver to minimize the cell F6, by changing the variables
A2:B2 (the parameters m, b), and it came back with

m ≈ 2.685, b ≈ −81.01

The minimum squares error was 9.89. (Note: when we invoked the Solver,
we were careful to uncheck the box for non-negative variables, since m, b could
be negative.)
The line that results from the solution to the previous problem is this:

w = 2.685 · h − 81.01

This line is called the line of best fit or the least squares line or the regression
line. It would be the business of statistics to argue from the minimum squares
error back to the plausibility of the theoretical curve. In this course, we will
stick with calculating the minimums.

Problem T12.11. An investment of $100 is supposed to produce a sequence of

values Mn with recursion Mn+1 = (1 + r) · Mn where r is a constant parameter.
Given the following observed sequence values, find the least squares value of
r.
n: 0 1 2 3 4
Mn : 100 130 160 190 230
 3. FITTING A MODEL TO DATA 225

Solution. We have M0 = 100. Given the number r, we have a predicted

sequence that satisfies the recursion. We set up such a sequence, using r as
a variable, which we initialize to 0. Here is what our spreadsheet formulas
look like, showing the row and column labels. Once the formulas are entered,
the spreadsheet shows number entries, as usual. This time we lay out our
spreadsheet with the data running down a column.
B C D E F
1 r n M[n]predicted M[n]observed del-squared
2 0 0 100 100 =(D2-E2)∧2
3 1 =(1+B$2)*D2 130 =D3-E3
4 2 =(1+B$2)*D3 160 =D4-E4
5 3 =(1+B$2)*D4 190 =D5-E5
6 4 =(1+B$2)*D5 230 =D6-E6
7 ?
The entry marked ? is =sumproduct(F2:F6,F2:F6). We can use fill-down
to obtain the computed entries in columns D and F, up to row 6. We have
initialized r = 0. In the Solver we ask to minimize cell F7 (where the squares
error is), using the variable B2 (the value of r). The minimum occurs when
r ≈ 0.24.
In the previous problem, suppose that M0 is regarded as a parameter, as
well. In that case, our least squares calculation includes M0 and r as variables.
If you work the problem that way, you obtain a different solution than the
one we obtained by working only with r. So, is it correct to have M0 as a
parameter? That depends on the purpose of the model within the subject it
came from. We are sticking with the mathematics involved.
 CHAPTER 13

Spreadsheet Formulas

A numerical solution is a solution obtained on a calculator or computer.

Numerical solutions always involve approximation, since computed arithmetic
is approximate – even if it is pretty accurate. In the Introduction, we men-
tioned that this course will involve both numerical and analytic calculations.
For the numerical calculations, we will often employ Excel.1 Many campus
computers have Excel; of course, you might want your own copy. Among the
many online and print resources for learning Excel, we mention the book [7].
The sections of this chapter introduce Excel techniques for performing a
wide variety of calculations. We will assume no prior exposure, but even if
you are familiar with Excel, you should read this material carefully.
The instructions we give here will assume that Excel is open with the Home
tab selected at the very top of the screen. The page before you is called a
spreadsheet. Its columns are labeled by letters and its rows by numbers. Each
rectangle is a cell , labeled by the letter and number for its column and row.

1. Function Values

We want to compute the function y = x2 at various values of x.

1Microsoft Excel. Among the many, many software tools available for numerical work,
we have chosen Excel because of its general availability and usage. Apple iWork Numbers
and Google Docs Spreadsheet, for instance, have many of the functions of Excel, but they
lack the ability to handle the non-linear problems that arise in many applications. For
what it’s worth, as of 2015 Excel is more flexible and more widely used than its current
competitors.
227
228 13. SPREADSHEET FORMULAS

First we explain how to compute a single value of this function. Enter the
following symbols and numbers in the indicated cells. You can use the arrow
keys or return to indicate that you are finished entering something.
A B C ···
1 x y
2 4 =A2*A2
..
.
The expression =A2*A2 is a formula. The equals sign at the beginning
tells Excel to compute the value of this expression rather than to display the
formula itself. The A2 in the formula refers to the contents of that cell. Since
4 is in cell A2, cell B2 shows 16 when the formula is entered. If you select B2
and look at the formula strip just above the spreadsheet grid, you will see the
formula.
Change cell A2 to various other values, and notice that B2 shows the square
of A2 each time.
When you are typing in the formula, you can click on A2 instead of typing
it. Try it by re-entering =A2*A2.
The expression A2*A2 could also be written A2∧2 The symbol ∧ is expo-
nentiation. Try changing B2 to =A2∧2 and make sure you are still computing
the square after the formula is entered.
Let’s compute a list of function values, using a more complicated function.
√
We’ll use y = x3 · exp(− x). Change B2 to

=(A2∧3)*exp(-sqrt(A2))

Notice how x3 is denoted; exp is the exponential function and sqrt is the
square root function. For the square root, we could also have written

A2∧0.5 or A2∧(1/2)

Put the following values in cells A2,A3,A4

 1. FUNCTION VALUES 229

A B C ···
1 x y
2 4 =(A2∧3)*exp(-(sqrt(A2))
3 3
4 1
√
Cell B2 should have 8.66. . .. We want to compute x3 exp(− x) for the values
3 and 1, as well. We can do this without rewriting the formula over and over.
Select B2 and place the cursor over the small square at the lower right corner
of that cell. The cursor should turn into a solid cross. (Not a double-lined
cross and not the hand symbol.) Once you have the solid cross, click and hold
down the cursor. Drag the cursor down to the cells B3,B4 below B2 and let
go! Here is what you should see.
A B C ···
1 x y
2 4 8.661
3 3 4.776
4 1 0.367
(You may have more decimal places.) Click on cell B3, and look at the formula
strip above the spreadsheet: you will see that the formula that was in B2 has
been copied down, but the reference to A2 has been changed to A3. The
operation of pulling down with the solid cross is called a fill-down.
We will understand fill-down by using it, but here is an abstract de-
scription: when you fill-down, the cell row references (these references are
numbers) are increased by one for each row drag down the column.
Let’s get a list of function values where the function has a parameter. Let
y = ln(m · x − 2), where m is a parameter (constant). We want to compute y
when x = 1, 2, 3, 4. We’ll start with m = 10; later we will change it. Set up
the following cell contents. (Notice that we have started in cell E4; we’re just
making the point that the fun can start anywhere!)
230 13. SPREADSHEET FORMULAS

E F G H
4 m 10
5 x y
6 1 =ln(F$4*E6-2)
7 2
8 3
9 4
In the entry F$4, ignore the dollar sign temporarily; that refers to cell F4,
which holds the value of m. When you enter the formula in F6, the number
2.079 is displayed; that’s the y-value when x = 1 (and m = 10).
Use fill-down from cell F6 to F9 to get the other y-values for m = 10.
Then look at the formulas in those cells. The E-column changes, row by row
– running over the four different x-values. But notice that the reference F$4
does not change. That’s the purpose of the dollar sign: to keep row 4 constant
as the fill-down is done.
Now we want to change m from 10 to 11, and then to 12. One way to do
this is simply to change the entry in F4 to 11. Try that, and notice that all
the y-values change as well! The y-values update to the new value of m.
We want to show how to change m while keeping the previous values of m
around – say to study what happens as m changes. Change F4 back to 10 and
enter 11 in G4 and 12 in H4.
Next we need to change the formula in F6; change it to this:
=ln(F$4*$E6-2)
We will explain the dollar sign on E momentarily. Then, fill-down from F6
to F9.
Now click on F6 and shift-click on F9; the cells F6,F7,F8,F9 should all be
highlighted. Place the cursor over the small square at the lower right corner
of the box of cells, hold down, and drag to the right, so that the box extends
to the right. End up at H9. Now release the cursor, and all the cells between
 2. RECURSION CALCULATIONS 231

G6 and H9 should fill in with numbers. This is called a fill-right.

Click on G7 and look at the formula strip. You should see =ln(G$4*$E6-2).
Notice that the m-value referred to is G$4. As you fill-right, the column
references change but the row references stay the same. The F in F$4 did not
have a dollar sign, and so it changed to G in the fill right. If the F had had a
dollar sign $F$4, it would not have changed. On the other hand, the x-value
$E6 does have a dollar sign, and so that column stayed the same in the fill
right. Each column refers to the same list of x-values in column E, but they
refer to the m-value at the head of their particular column.
To summarize: we have introduced formulas to calculate function values,
fill-down and fill-right as ways of replicating formulas, and the dollar
sign to prevent a fill operation from changing a particular cell reference. We
will get lots of practice with these operations as we use Excel throughout the
course.

2. Recursion Calculations

Let’s see how to compute a recursive sequence. We’ll start with a very
simple example:

Qn+1 = 1.5 · Qn for n = 0, 1, 2, . . .

We are going to compute the values of Q in column B, starting with Q0 in B1,

and then Q1 in B2, and then Q2 in B3, and so on. Let’s have Q0 = 2; enter 2
in B1. In B2 we want Q1 , and the recursive equation is Q1 = 1.5 · Q0 . Enforce
this via the following formula in B2:

=1.5*B1

Now B2 shows 3 = 1.5 · 2, which is Q1 . For the rest of the sequence, we simply
fill-down: select B2, get the solid cross in the lower right corner, and drag
down to B10. Cell B10 should show 76.88, or so. (That’s Q9 .)
232 13. SPREADSHEET FORMULAS

Suppose we wish to add up the Qn , for n = 0, 1, . . . , 9. These values are

in cells B1 through B10; we refer to this block of cells as B1:B10. In cell B12,
write the formula =sum(B1:B10); the sum of the terms should appear: 226.66,
or so.
Let’s do the same formula a different way. Select B13, and type
=sum(
Now click on cell B1, and then shift-click on cell B10. Notice that B1:B10
appears in the formula strip. Type a right parenthesis and press return, and
you should have the same formula as in B12.
Here is a more complicated set-up that will be discussed in class. The
recursion involves parameters x, y:
Fn+2 = x · Fn+1 + y · Fn for n = 0, 1, 2, . . .
We let x = −1 and y = 2 and F0 = 3 and F1 = 7, but we want to be able to
change these values.
A B C
1 x -1
2 y 2
3 n F[n]
4 0 3
5 1 7
6 =A5+1 =B$1*B5+B$2*B4
7
Select A6 and shift click on B6, so that A6:B6 is selected. Now fill-down to
get more values of n and more terms of the sequence. The references B1 and
B2 to x, y, respectively, will not change. You should notice how the recursion
updates as you go down.
To change x, y, we can simply change them in cells B1,B2. If we want to
keep the previous values of x, y, we can put new values in column C, and use
fill right from row 4 down to update the sequence.
 3. MATRIX CALCULATIONS 233

3. Matrix Calculations

In Excel, the rectangle of cells with upper left corner at A1 and lower right
corner at C4 is referred to as A1:C4. This block of cells will hold a 4×3 matrix,
as you can see by selecting that range on a spreadsheet.
Excel matrix addition is straightforward. Suppose we want to compute the
sum of two 2 × 3 matrices, one in block A2:C3 and one in block D4:F5 We
select a block of 2 × 3 cells in which to store the answer: say B7:D8. In the
upper left cell B7 we type the formula
=A2:C3+B7:D8
What you do next depends on the type of your computer. On a PC running
Windows2, press CONTROL+SHIFT+ENTER; on a Mac3, press APPLE+RETURN. We
call this kind of return a matrix return. The matrix return will fill the block
with the product sum. Typing an ordinary RETURN will not give you the entire
matrix sum.
Scalar multiplication is similarly easy. Say we want to compute A − 3 · B
where A, B are 3 × 4. Assume that A is in cells A1:D3 and B is in F1:I3. If
we want to put the answer in B5:E7, then we select those cells, and type the
following formula in B5
=A1:D3-3*F1:I3
Do a matrix return to fill B5:E7 with the answer.
The formula mmult multiplies matrices. Here is an example: we want to
multiply the 3 × 4 matrix in A1:D3 by the 4 × 3 matrix in F1:H4.

2Microsoft Windows
3Apple Macintosh, iMac, MacAir, etc.
234 13. SPREADSHEET FORMULAS

A B C D E F G H
1 1 0 -1 2 1 3 -1
2 1 1 4 6 0 1 5
3 0 2 -3 4 -2 3 -2
4 1 0 9
Now we compute the product. It doesn’t matter where it goes, but we
need room for the product matrix. Our product is 3 × 3, so maybe we move
down a ways, starting in row 11. We select a 3 × 3 block of cells, starting with
the upper left cell in the block. Say we select A11:C13. Here is formula typed.
A B C
11 =mmult(A1:D3,F1:H4)
12
13
Don’t forget to do a matrix return after typing the mmult formula. The
matrix product should appear.
A B C
11 5 0 9
12 -1 16 50
13 10 -7 52
To compute the inverse of a matrix, you use the minverse formula. Here
is a 2 × 2 example.
A B C D E
1 2 -4 =minverse(A1:B2)
2 7 6
The minverse formula is typed while D1:E2 is selected, using a matrix return,
as you would expect. If the matrix that you try to invert does not have an
inverse, each cell in the selected area will display #NUM!, indicating an error.
The mdeterm formula determines whether a square matrix has an inverse.
(Recall that a non-square matrix cannot have an inverse.) This formula gives a
single number, the determinant of the matrix. A square matrix has an inverse
if and only if its determinant is not 0. In the previous example, we could
 4. THE SOLVER 235

compute =mdeterm(A1:B2). This formula computes a number, so it is entered

with an ordinary return. This matrix has determinant 40.
An important note: You cannot edit any entry in a matrix computed from
a matrix return. If you do that by mistake, you will get an error message; you
will have to cancel the formula by clicking the delete-x on the formula text
field.

4. The Solver

The Solver is an optional add-in in Excel. If it does not appear on the

Tools menu, it should be available to be downloaded through the Tools op-
tions. The Solver is set up to solve optimization problems, but it can be used
to solve linear equations, as well. We will demonstrate its use in class. There
are many online examples and tutorials.4

4For instance, see http://www.solver.com/excel-solver-help

 Bibliography

Textbooks for applied calculus

[1] Raymond A. Barnett, Michael R. Ziegler, Karl E. Byleen, College Mathematics for
Business, Economics, Life Sciences, and Social Sciences, 12th edition, Prentice Hall,
2011.
[2] Geoffrey Berresford, Andrew Rockett, Applied Calculus, 6th edition, Brooks/Cole,
2013.
[3] Larry J. Goldstein, David C. Lay, David I. Schneider, Nakhle Asmar, Calculus and Its
Applications, 11th ed., Prentice-Hall, 2007.

Applications of this course to various disciplines

[4] Elizabeth S. Allman and John A. Rhodes, Mathematical Models in Biology, Cambridge
University Press, 2004.
[5] Robert Dorfman, Paul A. Samuelson, and Robert M. Solow, Linear Programming and
Economic Analysis, New York: McGraw-Hill, 1958, p.283ff.
[6] Frank C. Hoppensteadt and Charles Peskin, Mathematics in Medicine and the Life
Sciences, New York: Springer-Verlag, 1992.
[7] Matthew MacDonald, Excel 2013: The Missing Manual, O’Reilly Media, 2013.
[8] J.D. Murray, Mathematical Biology, New York: Springer, 1989.
[9] Raymond A. Serway and John W. Jewett, Jr. Physics for Scientists and Engineers with
Modern Physics, 7th edition, Thomson Brooks/Cole, 2008.
[10] Knut Sydsaeter and Peter J. Hammond, Essential Mathematics for Economic Analysis,
London: Pearson Education Limited, 2008.

Technical mathematics related to applied calculus

[11] Carl B. Boyer, The History of the Calculus and its Conceptual Development, New York:
Dover Publications, 1959.
[12] George B. Dantzig, Linear Programning and Extensions, Princeton: Princeton Univer-
sity Press, 1963.
[13] C. Henry Edwards and David E. Penney, Calculus, 6th edition, Pearson, 2002.
[14] Sir Thomas L. Heath, A Manual of Greek Mathematics, Mineola, New York: Dover
Publications, 2009.
[15] D.E. Smith, History of Mathematics, Volume 1, New York: Dover Publications, 1923.
[16] Alan Tucker, Applied Combinatorics, 6th edition, John Wiley and Sons, 2012.

237
 Index

(a + b)2 formula, 31 analysis, mathematical, v

A[i, j], the ij entry of matrix A, 145 antiderivative, 99
A−1 , inverse of matrix A, 175 area, as an application the integral, 120
b
F a , difference notation, 107 augmented matrix, 165
In , the n × n identity matrix, 151 average density, 34
[a, b], closed interval, 107 average velocity, 30
∆, change, 27
≈, used for all numerical answers, 5 Babylonian method for square roots, 23
exp, the exponential function (base e), 2 basic variables, 171
∂ Binomial Theorem, 45
∂x , partial derivative with respect to x,
189 block of cells in Excel, 233
R
f (x) · dx, the indefinite integral, an bounded constraints, 208
antiderivative, 100
R calculus, v
, the integral, 100
Rb capacitor, 9
a
f (x) · dx, the definite integral, 107 cell, in an Excel spreadsheet, 228
limx→a , the limit as x goes to a, cell, in Excel, 227
considered informally, 31 Chain Rule, 56
ln(x), the natural logarithm (base e), 3 Chain Rule, differential version, 75
O, the zero matrix of various sizes, 146 Chain Rule, for functions of several
R, Riemann sum, 113 variables, 197
∂, del, 189 closed constraint, 205
e, base of the exponential function, 1 closed interval, 107
f 00 (x), the second derivative of f , 69 Cobb-Douglas model, 10
f 0 (x), the derivative of f (x), 38 coefficient matrix, 164
n → ∞, n gets large without bound, 113 composite function, 56
compounded, 17
acceleration, 36, 42 concave down, 70
Addition Rule for derivatives, 46 concave up, 69
amp, abbrev. ampere, 9 concavity, of a curve, 69
ampere, 9, 182 consistent system of equations, 171
239
240 INDEX

constant distribution, 129 elementary operations, 166

Constant Multiple Rule for derivatives, elementary row operations, 166
45 Elimination, 169
constraints, in an optimization problem, entry of a matrix, 145
86 equilibrium, for a sequence, 18
consumer surplus, as an application of erg, 191
the integral, 141 exponent, 1
consumption vector, in Leontief’s model, exponential distribution, 130
159 exponential function, 1
continuous compounding, 138 exponential function, arbitrary base, 2
continuous random variable, 128 exponential function, in general, 1
critical point, 65 exponential population growth, 7
critical point, in several variables, 206 Exponential Rule, for antiderivatives, 102
current in a circuit, 182 Extreme Value Theorem, for several
current, electrical, 9 variables, 208
curve fitting, 177
factorial, 1
dec, decreasing, 66 Farads, 10
definite integral, 107 feasible production schedule, 159
definite integral, as limit of Riemann feasible, optimization problem, 88
sums, 113 Fibonacci Numbers, 25
del, partial derivative differential, 189 fill right, spreadsheet operation, 231
demand, 51 fill-down, spreadsheet operation, 229
demand curve, 141 finance charge, 17
density, 33 First Derivative Test, for several
derivative, 28 variables, 206
derivative at a point, 35 formula, in Excel, 228
derivative, of a function of several free, variables, 171
variables, 203 Fundamental Theorem of Calculus, 115
determinant, 234
graph, discrete, 181
diagonal, of a matrix, 151
differentiable function, 40, 43, 99 histogram, 129
differentiable function of several
variables, 190 Ideal Gas Law, 49
differentials, 41 identity matrices, 151
diminishing returns, 72 improper integral, 130
discount rate, 138 inc, increasing, 66
discrete graph, 181 income stream, 118
discrete, moving in definite steps, 18 inconsistent system of equations, 171
displacement, 28, 109 increasing function, 3
dot product, 148 indefinite integral, 100
infeasible, optimization problem, 87
edges of a graph, 181 initial value: when time is 0, 7
 INDEX 241

integral sign, 100 material balance equations (in

integrand, 100, 108 optimization constraints), 94
interest rate, 17 matrix, 145
Inverse Function Theorem, in one matrix addition, 145
variable, 77 matrix multiplication, 148
inverse functions, 77 matrix multiplied by a scalar, 147
inverse of matrix: how to find, 176 matrix return, in Excel, 233
Inverse Power Rule, for antiderivatives, maximal tree, among edges of a graph,
101 183
inverse, of a matrix, 175 mdeterm, Excel determinant, 235
invertible, matrix, 174 mean, 133, 134
irrational number, 1 median, 137
minverse, Excel matrix inverse, 234
junctions, in a circuit, 181 mmult, Excel matrix multiplication, 233
Kirchoff’s Current Rule, 182 model, 7
MPH, miles per hour, 28
Lagrange multiplier, 95
Lagrange multipliers, 209 natural logarithm, 3
Lagrange’s equation, 209 Newton’s law of heating and cooling, 9
Laplace’s PDE, 196 Newton’s Method, 72
leading coefficient, 64 Newton-Raphson Method, 72
least squares, 223, 224 norm, as in square norm, 220
Leontief’s linear economy, 159 normal equations, 223
Leslie’s population growth model, 155 numerical solution, 227
limit, v, 31 numerical work, v
line of best fit, 224 numerical, a decimal calculation, 4
linear algebra, v
linear equation, 163 objective, in an optimization problem,
linear inside rule for antiderivatives, 105 86, 205
linearity of the derivative, 46 open constraint, 205
linearity, of the definite integral, 110 optimization, 85
loan repayment, 17 oscillation, 65
Logarithm Rule, for antiderivatives, 101
logistic equation, 18 parameter, in a model, 10
lower limit, of integration, 108 partial derivative, 189
PDE, partial differential equation, 195
Malthusian (exponential) population pivot, in Elimination, 167
growth, 15 plane figure, 120
marginal cost, 80 position, 28
marginal Y of X, 81 Power Rule for derivatives, 44
marginal, derivative with respect to precedence rules, 5
production, 80 predator/prey model, 20
Markov process, 24, 157 present value, 138
242 INDEX

present value, as an application of the simple interest, 17

integral, 139 singularity, 64
price, 36 size of a matrix, 145
price elasticity of demand, 84 slack variable, 212
probability density function, 128 solution, system of linear equations, 164
probability vectors, in a Markov process, solution, to a linear equation, 163
157 solution, to an optimization problem, 87
problem variables, in an optimization Solver, in Excel, 235
problem, 86 speed, 28
Product Rule for derivatives, 50 spreadsheet, in Excel, 227
production schedule, in Leontief’s square norm, 220
model, 159 squares error, 221
standard normal distribution, 132
quick sketch: term used for depicting the state, of a Markov process, 25
oscillations of a curve, 65 system of linear equations, 163
Quotient Rule for derivatives, 52
tangent line, 32
radioactive decay, 8 technology matrix, in Leontief’s model,
random variable, 128 159
rank, of a matrix, 170 terminal speed, 23
rate of change, 48 time sequence equations (in optimization
rational function, 63 constraints), 94
recursion, in matrix form, 153 transition matrix, for a Markov process,
recursive equation, 15 157
regression line, 224 transshipment, on a discrete graph, 185
relative rate, 49 tree, in a graph, 182
resistor, 9
upper limit, of integration, 108
revenue, 51
Riemann sum, 112 Vandermonde matrix, 179
right side matrix, 164 velocity, 29
roots of a polynomial, 63 Verhulst’s equation, 18
row operations, 166 vertices of a graph, 181
row-echelon form, 168

scalar multiplication, 147

scalar, a number, 147
secant line, 32
second derivative, 69
sensitivity, of a parameter in an
optimization problem, 95
shadow price, 95, 97
shadow prices, 216
sign diagram, 64