.

Notes on
Differential
E quations

Robert E. Terrell

Preface
These Notes on Differential Equations are an introduction and invitation. The focus
is on
1. important models
2. calculus (review?) in applied contexts
I may point out that the title is not Solving Differential Equations; we derive them,
discuss them, review calculus background for them, apply them, sketch and compute them, and also solve them and interpret the solutions. This breadth is new to
many students.
The notes, available for many years on my web page, have evolved from lectures I
have given while teaching the Engineering Mathematics courses at Cornell University. They could be used for an introductory unified course on ordinary and partial
differential equations. There is minimal manipulation and a lot of emphasis on the
teaching of concepts by example.
For background on calculus see
Lax, P., and Terrell, M., Calculus With Applications, Springer, 2014.
The focus on key models here was influenced by the Lax Terrell book. In a few
places we assume familiarity with the divergence theorem. For further information
see:
Churchill, Ruel V., Fourier Series and Boundary Value Problems, McGraw
Hill, 1941
Hubbard, John H., and West, Beverly H., Differential Equations, a Dynamical Systems Approach, Parts 1 and 2, Springer, 1995 and 1996.
and the software discussed in Lecture 5.
Some of the exercises have the format Whats rong with this?. These are either
questions asked by students or errors taken from test papers of students in this
class, so it could be quite beneficial to study them.
Robert E. Terrell
version 5: 2014
version 1: 1997
i

Contents
1 The Bankers Equation
1.1

Slope Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2 A Gallery of Differential Equations

3 The Transport Equation

3.1

A Conservation Law . . . . . . . . . . . . . . . . . . . . . . . .

3.2

Traveling Waves . . . . . . . . . . . . . . . . . . . . . . . . . . .

4 The Logistic Population Model

11

5 Existence and Uniqueness and Software

14

6 Newtons Law of Cooling

17

6.1

Investments . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

20

7 Exact equations for Air and Steam

21

8 Eulers Numerical Method

24

9 Spring-mass oscillations

28

9.1

Conservation laws and uniqueness . . . . . . . . . . . . . . . . .

10 Applications of Complex Numbers

31
33

10.1 Exponential and characteristic equation . . . . . . . . . . . . . .

33

10.2 The Fundamental Theorem of Algebra . . . . . . . . . . . . . . .

38

10.3 A forced oscillator . . . . . . . . . . . . . . . . . . . . . . . . .

39

11 Three masses oscillate

42

12 Boundary Value Problems

47

ii

13 The Conduction of Heat

50

13.1 Walk the line . . . . . . . . . . . . . . . . . . . . . . . . . . . .

14 Initial Boundary Value Problems for the Heat Equation

53
57

14.1 Insulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

58

14.2 Product Solutions . . . . . . . . . . . . . . . . . . . . . . . . . .

59

14.3 Superposition . . . . . . . . . . . . . . . . . . . . . . . . . . . .

61

15 The Wave Equation

65

16 Application of Power Series: a Drum model

69

16.1 A new Function for the Drum model, J0 . . . . . . . . . . . . . .

73

16.2 But what does the drum Sound like? . . . . . . . . . . . . . . . .

75

17 The Euler equation for Fluid Flow, and Acoustic Waves

17.1 Sound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18 The Laplace Equation

77
81
82

18.1 Laplace leads to Fourier . . . . . . . . . . . . . . . . . . . . . . .

85

18.2 Fouriers Dilemma . . . . . . . . . . . . . . . . . . . . . . . . .

87

18.3 Fourier answered by Orthogonality . . . . . . . . . . . . . . . . .

89

19 Application to the weather?

92

iii

The Bankers Equation

T ODAY: An example involving your bank account, and nice pictures called
slope fields (or direction fields). How to read a differential equation.

Welcome to the world of differential equations! They describe many processes in

the world around you, but of course well have to convince you of that. Today
we are going to give an example, and find out what it means to read a differential
equation.
A differential equation is an equation which contains a derivative of an unknown
function. It tells something about a rate of change, from which we hope to deduce
facts about the function. Here is a differential equation.
dy
= .01y
dt
It might represent your bank account, where the balance is y(t) at a time t years
after you open the account, and the account is earning 1% interest. Regardless of
the specific interpretation, lets see what the equation says. Since we see the term
dy/dt we can tell that y is a function of t, and that the rate of change is a multiple,
namely .01, of the value of y itself. We definitely should always write y(t) instead
of just y, and we will sometimes, but it is traditional to be sloppy.
For example, if y happens to be 2000 at a particular time t, the rate of change of y
is then .01(2000) = 20, and the units of this rate are dollars/year. From calculus
we know that y is increasing whenever y 0 is positive, thus whenever y is positive. I
hope your bank balance is positive!
P RACTICE : What do you estimate the balance will be, roughly, a year from now, if it
is 2000 and is growing at 20 dollars/year?
This is not supposed to be a hard question. By the way, when I ask a question, dont
cheat yourself by ignoring it. Think about it, and future things will be easier.

Later when y is, say, 5875.33, its rate of change will be .01(5875.33) = 58.7533
which is much faster. Well sometimes refer to y 0 = .01y as the bankers equation.
Do you begin to see how you can get useful information from a differential equation
fairly easily, by just reading it carefully? One of the most important skills to learn
about differential equations is how to read them. For example in the equation
y 0 = .01y 10
1

there is a new negative influence on the rate of change, due to the 10. This 10
could represent withdrawals from the account.
P RACTICE : What must be the units of the 10?

Whether the resulting value of y 0 is actually negative depends on the current value
of y. For example, if y = 100 then y 0 = 1 10 is negative, and y must be
decreasing. If y = 1000 then y 0 = 0, while if y = 2000 then y 0 is positive
and y is increasing. It seems that we ought to maintain some minimal balance.
That is an example of reading a differential equation. As a result of this reading
skill, you can perhaps recognize that the bankers equation is very idealized: It
does not account for deposits or changes in interest rate. It didnt account for
withdrawals until we appended the 10, but even that is an unrealistic continuous
rate of withdrawal. You can think about how to modify the equation to include
those things more realistically.

1.1

Slope Fields

It is significant that you can make graphs of the solutions sometimes. In the bank
account problem we have already noticed that the larger y is, the greater the rate
of increase. This can be displayed by sketching a slope field as in Figure 1.1.
Slope fields are done as follows. First, the general form of an ordinary differential
equation is
y 0 = f (y, t)
where y(t) is the unknown function, and f is given. To make a slope field for
this equation, choose some points (y, t) and evaluate f there. According to the
differential equation, these numbers must be equal to the derivative y 0 , which is the
slope for the graph of the solution. These resulting values of y 0 are then plotted
using small line segments to indicate the slopes. For example, at the point t =
6, y = 20, the equation y 0 = .01y says that the slope must be .01(20) = .2. So we
go to this point on the graph and place a mark having this slope. Solution curves
then must be tangent to the slope marks. This can be done by hand or computer,
without solving the differential equation.
Note that we have included the cases y = 0 and y < 0 in the slope field even
though they might not apply to your bank account.
P RACTICE : Try making a slope field for y 0 = y + t. To begin, what is the slope at
(t, y) = (3, 3) if the solution y(t) goes through that point? What are all the points
(t, y) where the slope is 0?

20

15

10

-5
0

10

20

30

40

50

60

70

80

Figure 1: The slope field for y 0 = .01y as made by an octave script as on page 18. A
solution starting from y(0) = 8 is also shown.

There is also a way to explicitly solve the bankers equation. Assume we are looking for a positive solution. Then y is not zero so it is alright to divide the equation
by y, getting
1 dy
= .01
y dt
Then integrate, using the chain rule:
ln(y) = .01t + c
where c is some constant. Then
y(t) = e.01t+c = c1 e.01t
Here we have used a property of the exponential function, that ea+b = ea eb , and
set c1 = ec . The potential answer which we have found must now be checked by
substituting it into the differential equation to make sure it really works. You ought
to do this. Now. You will notice that the constant c1 can in fact be any constant, in
spite of the fact that our derivation of it seemed to suggest that it be positive.
This is common with differential equations: It is not so important what methods
you use; what is important is that you check to see whether you are right. Even
guessing answers is a highly respected method! if you check them.
You might be skeptical about any bank account that grows exponentially. If so,
good. It is clearly impossible for anything to grow exponentially forever. Perhaps it is reasonable for a limited time. The hope of applied mathematics is that
3

our models will be idealistic enough to solve while being realistic enough to be
worthwhile.
The last point we want to make about this example concerns the constant c1 . What
is the amount of money you originally deposited, y(0)? Do you see that it is the
same as c1 ? That is because y(0) = c1 e0 and e0 = 1. If your original deposit
was 300 dollars then c1 = 300. This value y(0) is called an initial condition,
and serves to pick the solution we are interested in out from among all those which
might be drawn in Figure 1.
E XAMPLE : Be sure you can do the following kind
x0
x(0)

= 3x
= 5

Like before, we get a solution x(t) = ce3t . Then x(0) = 5 = ce0 = c so

c = 5 and the answer is
x(t) = 5e3t
Check it to be absolutely sure.
P ROBLEMS
1. Make slope fields for x0 = x, x0 = t, x0 = x, x0 = x + cos t.
2. Sketch some solution curves onto the given slope field in Figure 2.
3. What general fact do you know from calculus about the graph of a function y if y 0 > 0?
Apply this fact to any solution of y 0 = y y 3 : consider cases where the values of y
lie in each of the intervals (, 1),(1, 0), (0, 1), and (1, ). For each interval, state
whether y is increasing or decreasing.
4. (continuing 3.) If y 0 = y y 3 and y(0) = 12 , what do you think will be the lim y(t)?
t
Make a slope field if youre not sure.
5. Reconsider the bankers equation y 0 = .01y. If the interest rate is 3% at the beginning
of the year and expected to rise linearly to 4% over the next two years, what would you
replace .01 with in the equation? You are not asked to solve the equation.
6. In y 0 = .01y 10, suppose the withdrawals are changed from $10 per year continuously,
to $200 every other week. Do you think it would be alright to use a smooth function of the
form a cos bt to approximate the withdrawals? What would you take for a and b?
7. A rectangular tank measures 2 meters east-west by 3 meters north-south and contains
water of depth x(t) meters, where t is measured in seconds. One pump pours water in
at the rate of .05 [m3 /sec] and a second variable pump draws water out at the rate of
.07 + .02 cos(t) [m3 /sec]. The variable pump has period 1 hour. Set up a differential
equation for the water depth, including the correct value of .

3
2
1
0
-1
-2
-3
-2

10

Figure 2: The slope field for Problem 2.

A Gallery of Differential Equations

T ODAY: A gallery is a place to look and to get ideas, and to find out how
other people view things.

Here is a list of differential equations, as a preview of things to come. Unlike

the bankers equation y 0 = .01y, not all differential equations are about money.
However, many of them are conservation laws, which track the changes in some
quantity, like the bankers equation tracks your balance.
1. y 0 = .01(600 y) This equation is a model for the heating of a pizza in a
600 degree oven. Of course the .01 is there for comparison with the bankers
interest. The physical law involved is called Newtons Law of Cooling but
it applies to heating also. Well study it on page 17. This is a conservation
law which tracks the exchange of heat energy between the pizza and the
oven.
2. Newtons law of cooling is an ordinary differential equation, ODE. There
are also partial differential equations, PDE, which means that the unknown
functions depend on more than one variable. Then partial derivatives show
up in the equation. One example is ut = uxx where the subscripts denote
partial derivatives. This is a conservation law called the heat equation, and
u(x, t) is the temperature at position x and time t, when heat is allowed to
5

conduct only along the x axis, as through a wall or along a metal bar. It is
a more detailed study than Newtons law of cooling, and we will discuss it
more starting on page 50.
3. Newton had other laws as well, one of them being the F = ma law of
inertia. You might have seen this in a physics class, but not realised that it
is a differential equation. That is because it concerns the unknown position
of a mass, and the second derivative a, of that position. In fact, the original
differential equation, the very first one over 300 years ago, was made by
Newton for the case in which F was the gravity force between the earth
and moon, and m was the moons mass. Before that, nobody knew what a
differential equation was, and nobody knew that gravity had anything to do
with the motion of the moon. (They thought gravity was what made their
physics books heavy.)
4. Maxwells equations for electric and magnetic fields in empty space are
Et = curl B,

Bt = curl E,

div E = 0,

div B = 0.

This is a system of first order partial differential equations in the six components of E and B. We wont discuss these excapt to see how they relate to
the wave equation, see Problem 10.
5. There are many wave equations. utt = uxx looks sort of like the heat conduction equation, but is very different because of the second time derivative.
When it describes musical vibrations of a guitar string it is an instance of
Newtons law of motion. When it describes light waves it is a consequence
of Maxwells equations. The equation for vibrations of a drum head is the
two dimensional wave equation, utt = uxx + uyy that we discuss starting on
page 69.
P ROBLEMS
8. Newtons gravity law says that the force between a big mass at the origin of the x axis
and a small mass at point x(t) is proportional to x2 . How would you write the F = ma
law for that as a differential equation?
9. What functions do you know about from calculus, that are equal to their second derivative? the negative of their second derivative?
10. Suppose the electric and magnetic fields are of the form E = (0, u, 0), B = (0, 0, v).
Use Maxwells equations to show that u and v depend only on x and t and that u solves
the wave equation utt = uxx .

The Transport Equation

T ODAY: A first order partial differential equation.

Here is a partial differential equation, sometimes called a transport equation, and

sometimes called a wave equation.
w
w
+3
=0
t
x
P RACTICE : We remind you that partial derivatives are the rates of change holding all
but one variable fixed. For example

x y 2 x + 2yt = 2y,
t

x y 2 x + 2yt = 1 y 2
x

What is the y partial?

Our PDE is abbreviated

wt + 3wx = 0
You can tell by the notation that w is to be interpreted as a function of both t and
x. You cant tell what the equation is about. We will see that it can describe certain
types of waves. There are many types of waves, such as water waves, electromagnetic waves, the wavelike motion of musical instrument strings, the invisible
pressure waves of sound, the waveforms of alternating electric current, and others.
This equation is a simple model.
P RACTICE : You know from calculus that increasing functions have positive derivatives. In Figure 3 a wave shape is indicated as a function of x at one particular time t.
Focus on the steepest part of the wave. Is wx positive there, or negative? Next, look
at the transport equation. Is wt positive there, or negative? Which way will the steep
profile move next?
Remember how important it is to read a differential equation.

3.1

A Conservation Law

Well derive the equation as one model for conservation of mass. You might feel
that the derivation of the equation is harder than the solving of the equation.
7

w
a

a+h

Figure 3: The wind blows sand along the surface. Some enters the segment (a, a + h) from
the left, and some leaves at the right. The net difference causes changes in the height of
the dune there.

We imagine that w represents the height of a sand dune which moves by the wind,
along the x direction. The assumption is that the sand blows along the surface,
crossing position (x, w(x, t)) at a rate proportional to w. Thus taller areas encounter more wind-blown sand. The proportionality factor is taken to be 3, which
has dimensions of velocity, like the wind.
The law of conservation of sand says that over each segment (a, a + h) you have
d
dt

a+h

w(x, t) dx = 3w(a, t) 3w(a + h, t)

a

That is the time rate of the total sand on the left side, and the sand flux on the right
side. Divide by h and take the limit.
P RACTICE : 1. Why is there a minus sign on the right hand term?
2. What do you know from the Fundamental Theorem of Calculus about
1
h

a+h

f (x) dx ?
a

The limit we need is the case in which f is wt (x, t).

We find that wt (a, t) = 3wx (a, t). Of course a is arbitrary. That concludes the
derivation.

3.2

Traveling Waves

When you first encounter PDE, it can appear, because of having more than one
independent variable, that there is no reasonable place to start working. Do I try t
first, x, or what? In this section well just explore a little. If we try something that
doesnt help, then we try something else.

1.5
1
0.5
0
-0.5
-1
-1.5
0

10

15

20

25

30

Figure 4: Graphs of cos(x 3t) at times t = 0, .5, 1. How fast is the wave moving?
P RACTICE : Find all solutions to our transport equation of the form
w(x, t) = ax + bt
In case that is not clear, it does not mean derive ax+bt somehow. It means substitute
the hypothetical w(x, t) = ax + bt into the PDE and see whether there are any such
solutions. What is required of a and b?

Those practice solutions dont look much like waves. Lets try something more
wavey.
P RACTICE : Find all solutions to our transport equation of the form
w(x, t) = c cos(ax + bt).

So far, we have seen a lot of solutions to our transport equation (if you did the
practice problems). Here are a few of them:
w(x, t) = x 3t
w(x, t) = 2.1x + 6.3t
w(x, t) = 40 cos(5x 15t)
2
w(x, t) = cos(8x 24t)
7
For comparison, that is a lot more variety than we found for the bankers equation.
Remember that the only solutions to the ODE y 0 = .01y are constant multiples
of e.01t . Now lets go out on a limb. Since our transport equation allows straight
lines of all different slopes and cosines of all different frequencies and amplitudes,
maybe it also allows other things too.
9

Try
w(x, t) = f (x 3t)
where we wont specify the function f yet. Without specifying f any further, we
cant find the derivatives we need in any literal sense, but can apply the chain rule
anyway. The intention here is that f ought to be a function of one variable, say s,
and that the number x 3t is being inserted for that variable, s = x 3t. The
partial derivatives are computed using the chain rule, because we are composing f
with the function x 3t of two variables. The chain rule here looks like this:
wt =

w
df s
=
= 3f 0 (x 3t)
t
ds t

P RACTICE : Figure out why wx = f 0 (x 3t).

Setting those into the transport equation we get

wt + 3wx = 3f 0 + 3f 0 = 0
That is interesting. It means that any differentiable function f gives us a solution.
Any dune shape is allowed. You see, it doesnt matter at all what f is, as long as it
is some differentiable function.
Dont forget: differential equations are a model of the world. They are not the
world itself. Real dunes cannot have just any shape f whatsoever. They are more
specialized than our model.
P RACTICE : Check the case f (s) = 22 sin(s) 10 sin(3s). That is, verify that
w(x, t) = 22 sin(x 3t) 10 sin(3x 9t)
is a solution to our wave equation.
P ROBLEMS
11. Work all the practice items in this lecture if you have not done so yet.
12. Find a lot of solutions to the wave equation
yt 5yx = 0
and tell which direction the waves move, and how fast.
1
13. Check that w(x, t) =
is one solution to the equation wt + 3wx = 0.
1 + (x 3t)2

10

14. What does the initial value w(x, 0) look like in Problem 13, if you graph it as a function
of x?
15. Sketch the profile of the dune shapes w(x, 1) and w(x, 2) in Problem 13. What is
happening? Which way is the wind blowing? What is the velocity of the dune? Can you
tell the velocity of the wind?
16. Solve ut + ux = 0 if we also want to have the initial condition u(x, 0) =
1
7 sin(4x). Sketch the wave shape for several times.

1
5

cos(2x) +

The Logistic Population Model

T ODAY: The logistic equation is an improved model for population growth.

We have seen that the bankers equation y 0 = .01y has exponentially growing
solutions. It also has a completely different interpretation from the bank account
idea. Suppose you have a population containing about y(t) individuals. The word
about is used because if y = 32.51 then we will have to interpret how many
individuals that is. Also the units could be, say, thousands of individuals, rather
than just plain individuals. The population could be anything from people on earth,
to deer in a certain forest, to bacteria in a certain Petrie dish. We can read this
differential equation to say that the rate of change of the population is proportional
to the number present. That perhaps captures some element of truth, yet we see
right away that no population can grow exponentially forever. Sooner or later there
will be a limit imposed by space, or food, or energy, or something.
The Logistic Equation
Here is a modification to the bankers equation that overcomes the previous objection.
dy
= .01y(1 y)
dt
In order to understand why this avoids the exponential growth problem we must
read the differential equation carefully. Remember that I said this is an important
skill.
Here we go. You may rewrite the right-hand side as .01(y y 2 ). You know that
when y is small, y 2 is very small. Consequently the rate of change is still about
.01y when y is small, and you will get exponential growth, approximately. After
this goes on for a while, it is plausible that the y 2 term will become important.
In fact as y increases toward 1 (one thousand or whatever), the rate of change
approaches 0. That is intended to limit the population.
11

Figure 5: A slope field for the Logistic Equation. Note that solutions starting near 0 have
about the same shape as exponentials until they get near a.

For simplicity we now dispense with the .01, and for flexibility introduce a parameter a, and consider the logistic equation
y 0 = y(a y).
If we make a slope field for this equation we see something like Figure 5.
The solutions which begin with initial conditions between 0 and a evidently grow
toward a as a limit. This in fact can be verified by finding an explicit formula
for the solution. Proceeding much as we did for the bank account problem, first
separate the variables
1
dy = dt
ay y 2
To make this easier to integrate, well use a trick which was discovered by a student
in this class, and multiply first by y 2 /y 2 . Then integrate
Z
Z
y 2
dy = dt
ay 1 1
1
ln(ay 1 1) = t + c
a
The integral can be done without the trick, using partial fractions, but that is longer.
Now solve for y(t)
ay 1 1 = ea(t+c) = c1 eat
a
y(t) =
1 + c1 eat
These manipulations would have to be done more carefully if we had not specified
that we are interested in y values between 0 and a. For example, the ln of a negative
12

number is not defined. However, we emphasize that the main point is to check any
formulas found by such manipulations. So lets check it:
y0 =

a2 c1 eat
(1 + c1 eat )2

We must compare this expression to

ay y 2 =

at ) 1
a2
a2
2 (1 + c1 e

=
a
1 + c1 eat (1 + c1 eat )2
(1 + c1 eat )2

You can see that this matches y 0 . Note that the value of c1 is not restricted to be
positive, even though the derivation above may have required it. We have seen
this kind of thing before, so checking is very important. The only restriction here
occurs when the denominator of y is 0, which can occur if c1 is negative. If you
stare long enough at y you will see that this does not happen if the initial condition
is between 0 and a, and that it restricts the domain of definition of y if the initial
conditions are outside of this interval. All this fits very well with the slope field
above. In fact, there is only one solution to the equation which is not contained in
our formula.
P RACTICE : Can you see what it is?
P ROBLEMS
17. Suppose that we have a solution y(t) for the logistic equation y 0 = y(a y). Choose
some time delay, say 3 time units to be specific, and set z(t) = y(t 3). Is z(t) also a
solution to the logistic equation?
18. The three S-shaped solution curves in Figure 5 all appear to be exactly the same
shape. In view of Problem 17, are they?
19. Prof. Verhulst made the logistic model in the mid-1800s. The US census data from
the years 1800, 1820, and 1840, show populations of about 5.3, 9.6, and 17 million. Well
need to choose some time scale t1 in our solution y(t) = a(1 + c1 eat )1 so that t = 0
means 1800, t = t1 means 1820, and t = 2t1 means 1840. Figure out c1 , t1 , and a to
match the historical data. WARNING : The arithmetic is very long. It helps if you use the
fact that ea2t1 = (eat1 )2 . ANSWER : c1 = 36.2, t1 = .0031, and a = 197.
20. Using the result of Problem 19, what population do you predict for the year 1920? The
actual population in 1920 was 106 million. The Professor was pretty close wasnt he? He
was probably surprised to predict a whole century.

13

21. Census data for 1810, 1820, and 1830 show populations of 7.2, 9.6, and 12.8 million.
Trying those as in Problem 19, it turned out that I couldnt fit the numbers due to the
numerical coincidence that
(7.2)(12.8) = (9.6)2 .
That is why I switched to the years in Problem 3. (This indicates that the fitting of real data
to a model is nontrivial.) Show that exponential functions f (t) = cekt do have a related
property:

2
f (t b)f (t + b) = f (t) .
But exponential functions dont solve the logistic model.

Existence and Uniqueness and Software

T ODAY: We learn that some equations have unique solutions, some have too
many, and some have none. Also an introduction to some of the available
software.

If you are running an experiment you would like to think that the same results will
follow from the same initial conditions each time you repeat the experiment. This
makes us feel that our differential equations ought to have unique solutions.
On the other hand, it can happen that a differential equation has no solution at all,
or a solution which is not defined for all time.
E XAMPLE : Consider x0 = x2 . If x is never 0, multiply by x2 to get
x2 x0 = 1.

0
Use the chain rule to recognize this as x1 = 1. Integrate to get
x1 = t + c, so
1
x(t) =
.
t+c
This function is not defined when t = c.
1
defines two functions, not one,
In fact we should point out that the formula t+c
the domains being (, c) and (c, ) respectively. The reason for this distinction is that the solution of a differential equation has to be differentiable, and
therefore continuous.

We therefore are interested in the following general statement about what sorts of
equations have solutions, and when they are unique, and how long, in time, these
solutions last. It is called the Fundamental Existence Theorem.
14

We remind you that the partial derivative of a function of several variables is defined to mean that the derivative is constructed by holding all other variables conf
f
stant. For example, if f (x, t) = x2 tcos(t) then
= 2xt and
= x2 +sin(t).
x
t
F UNDAMENTAL T HEOREM Consider an initial value problem of the
form
x0 (t) = f (x, t)
x(t0 ) = x0
f
where f , t0 , and x0 are given. Suppose it is true that f and
are
x
continuous functions of t and x in at least some small region containing the initial condition (x0 , t0 ). Then the conclusion is that there is
a solution to the problem, it is defined at least for a small amount of
time both before and after t0 , and there is only one such solution.
The example above, x0 = x2 , is in the form specified: f (t, x) = x2 is continuous,
f
= 2x is also continuous. Lets take the initial value to be x(0) = 2. The
and
x
theorem applies, so we look at the conclusion: there is a solution defined for some
interval of t around the initial time. Take c = 21 in our solution formula, to get
x(t) =

1
.
t 21

That has x(0) = 2. It is defined for < t < 12 , and blows up when t = 1/2. The
theorem did not predict the blowup. If you were a scientist working on something
which might blow up, you would be glad to be able to predict when or if the
explosion will occur. But this requires a more detailed analysis in each casethere
is no general theorem about it.
Software
In spite of examples we have seen so far, it turns out that it is not possible to write
down solution formulas for most differential equations. This means that we have
to draw slope fields or go to the computer for approximate solutions. We soon will
study how approximate solutions can be computed. Meanwhile we are going to
introduce you to some of the available tools.
There are several software packages available to help your study of differential
equations.
There are java applets available at
15

x = x*(1x)
2

1.5

0.5

0.5

1
2

10

Figure 6: The slope field and several solutions for the Logistic Equation. Question: Do
these curves really run into each other? Read the Fundamental Theorem again if youre
not sure.

http://math.rice.edu/dfield/dfpp.html
for a single ordinary differential equation,
http://math.rice.edu/pplane/pplane.html
for a system of 2, and the rather primitive
http://www.math.cornell.edu/bterrell/de
for a system of 1, 2, or 3. [At the time of writing, the dfield and pplane web
page was under repair.] Probably the earliest user-friendly differential equation
software was MacMath by John Hubbard. There are also java applets on partial
differential equations. These are for the heat and wave equations in one or two
space dimensions, and for the Laplace equation in two dimensions, available from
http://www.math.cornell.edu/bterrell
The other approach is to do some programing in any of several available languages.
These include matlab, its free counterpart octave available from
http://www.octave.org,
and the freeware program xpp. The script used to make several of the slope field
16

figures in these notes is listed in Figure 7. I ran it in octave. It approximates

solutions using Eulers numerical method which is explained on page 24.
P ROBLEMS
22. Find out how to download and use some of the programs mentioned in this lecture, try
a few simple things, and read some of the online help which they contain.
23. Answer the question in the caption for Figure 6.

Newtons Law of Cooling

T ODAY: The same mathematics can describe changing temperature of an
object and balance on a loan.

The differential equation

x0 = kx
says that the rate of change of x(t) is proportional to the value of x(t). This is
reasonable in some applications, such as when k represents the rate of interest on
a savings account. The equation predicts exponential growth when k is positive, or
decay when k is negative.
Newtons law of cooling is the statement that the exponential growth applies sometimes to the temperature of an object, provided that x is taken to mean the difference in temperature between the object and its surroundings. Suppose the object
has temperature T (t) at time t. Then
x(t) = T (t) E
where E is the environment temperature.
P RACTICE : If the object is hotter than the environment, will the object cool or heat?
Is x0 positive or negative?

In view of that practice (you did think about the practice right?) we will write the
differential equation as x0 = kx where k is some positive number.
P RACTICE : Check that x0 = kx is equivalent to T 0 = k(T E).

For example suppose we have placed a 100 degree pizza in a 600 degree oven. We
let x(t) be the pizza temperature at time t, minus 600. This makes x negative,
while x0 is certainly positive because the pizza is heating up.
17

function dirf(x1,x2,t1,t2,x0)
% make a direction field for x= ef(x,t) in rectangle [t1,t2] by [x1,x2]
% and compute a solution with initial value x(t1) = x0.
tmarks = linspace(t1,t2,16);
marklg = (t2-t1)/32;
xlong = (x2-x1)/32;
xmarks = linspace(x1,x2,16);
teuler = linspace(t1,t2,100);
xeuler = zeros(1,100);
tstep = (t2-t1)/100;

%
%
%
%
%
%

16 equally spaced ts for slope marks

half as long each
in case of steep slopes
16 equally spaced xs for slope marks
100 ts for Euler method
100 xs to be calculated

% draw slope marks

for i=1:16
for j=1:16
F = ef(xmarks(j),tmarks(i));
if abs(F)<1
line([tmarks(i) tmarks(i)+marklg],...
[xmarks(j) xmarks(j)+marklg*F]);
else
line([tmarks(i) tmarks(i)+xlong/F],...
[xmarks(j) xmarks(j)+xlong]);
endif
end
end

% the slope

% so mark isnt too long

% draw Euler approximation

xeuler(1) = x0;
t = t1;
for k=1:99
xeuler(k+1) = xeuler(k)+tstep*ef(xeuler(k),t);
t = t+tstep;
end
hold on
plot(teuler,xeuler,"k")
hold off
print -deps -FHelvetica:20
function val = ef(x,t)
val = x.*x-sin(t);

"dirfig.eps"

% the right hand side in x = f(x,t)

% change as needed

Figure 7: A program to draw slope fields (direction fields) in octave.

18

p = p+340260*sign(t6)
600

500

400

300

200

100

0
0

10

12

Figure 8: A pizza at temperature p(t) heats and then cools. (k = 1 here.) To change the
environment from 600 degrees to 80 degrees at time 6 the equation p0 = (p E) was
written as
p0 = p + a b sign(t 6)
with a and b selected to achieve the 600 and 80.
P RACTICE : Check in this case (heating) too, it is correct to write T 0 = k(T E),
i.e., that both sides are positive.

Therefore Newtons law of cooling is also Newtons law of heating. The solution
x(t) = Cekt , and C = x(0) = 100 600. Equivalently the solution to
T 0 = k(T 600),

T (0) = 100

is the pizza temperature

T (t) = 600 500ekt .
We dont have any way to get k using the information given. It would suffice for
example, to be told that after the pizza has been in the oven for 15 minutes, its
temperature is 583 degrees. This says that 583 = 600 500e15k . So we can
solve for k and then answer any questions about temperature at other times.
In this model, we imagine that the environment is much larger than the object so
that E doesnt change while T does change. But the environment temperature can
change if we move the pizza from the oven to the 80 degree kitchen. A plot of
the temperature history under such conditions is in Figure 8. The temperature is
continuous when the move occurs at time t = 6 but is not differentiable then.
A first order linear equation is of the form
x0 + ax = b
where a and b might be functions of t. Newtons law of cooling T 0 = k(T E)
and the exponential growth equation x0 = kx are examples.
19

6.1

Investments

Our bank account equation y 0 = .01y can be made more realistic and interesting.
Suppose we make withdrawals at a rate of $3500 per year. This can be included in
the equation as a negative influence on the rate of change.
y 0 = .01y 3500
Again we see a first order linear equation. But the equation is good for more than
an idealised bank account. Suppose you buy a car at 1% financing, paying $3500
per year. Now loosen up your point of view and imagine what the bank sees. From
the point of view of the bank, they just invested a certain amount in you, at 1%
interest, and the balance decreases by withdrawals of about $3500 per year.
So the same equation describes two apparently different kinds of investments.
E XAMPLE : A car is bought using the loan as described above. If the loan is
to be paid off in 6 years, what price can we afford?
The price is y(0). We need
y 0 = .01y 3500 with y(6) = 0.
As for Newtons law of cooling, we can write it as y 0 = .01(y 350 000)
and expect by analogy that y(t) = 350 000 + c1 e.01t . Set t = 6 to get
0 = y(6) = 350 000 + c1 e.06 , so
c1 = 350 000 e.06 = 329 618.
.
.
This implies that the price is y(0) = 350 000 329 618 = 20 382.
P ROBLEMS
24. Describe in whole sentences what the differential equation
y 0 = ky + `
could be used for. If someone in your family is interested but hasnt taken the course, what
background would you have to explain so he or she could read and understand the things
you wrote?
25. This problem outlines a method for solving first order linear equations. Suppose we
have the idea to multiply a first order linear equation x0 + ax = b by a factor f , so that the
result of the multiplication is (f x)0 = f b, i.e., that the equation becomes recognizable as
an instance of the product rule,
f 0 x + f x0 = f b.
Show that for this plan to work, you will need to require that f 0 = af . In case a is constant,
deduce that eat will be a suitable choice for f . A function f used in this manner is called
an integrating factor.

20

x
Figure 9: Here x(t) is the length of a line of people waiting to buy tickets. Is the rate of
change proportional to the amount present? Does the ticket seller work twice as fast when
the line is twice as long?
26. (continuing 25) In case a is a function of t, verify that e
for f .

a dt

will be a suitable choice

27. The temperature of an apple pie is recorded as a function of time. It begins in the
oven at 450 degrees, and is moved to an 80 degree kitchen. Later it is moved to a 40
degree refrigerator, and finally back to the 80 degree kitchen. Make a sketch somewhat
like Figure 8, which shows qualitatively the temperature history of the pie.
28. Newtons law of cooling looks like u0 = au when the surroundings are at temperature
0. This is sometimes replaced by the Stephan-Boltzmann law u0 = bu4 , if the heat is
radiated away rather than conducted away. Suppose the constants a, b are adjusted so
that the two rates are the same at some temperature, say 10 Kelvin. Which of these laws
predicts faster cooling when u < 10? u > 10?
29. Saras employer contributes $3000 per year to a retirement fund, which earns 3% interest. Set up an initial value problem to model the balance in her fund, if it began with $0
when she was hired. Use the result of Problem 25 to solve it. How much money will she
have after 20 years?
1
converts the logistic equation y 0 = .01(yy 2 )
y
of Lecture 4 to the first order equation x0 = .01(x 1), and figure out a philosophy for
why this might hold.
30. Show that the change of variables x =

31. Answer the question in the caption of Figure 9.

Exact equations for Air and Steam

T ODAY: An historically important exact differential equation happens to
be first order linear too.

Some vector fields are gradients, some are not, and some differential equations are
21

Figure 10: How fast can it go?

said to be exact. An exact differential mdx + ndy is one that can be written
as df = fx dx + fy dy for some function f (x, y). That is, it is associated with a
gradient vector field f = fx~ + fy~. [Associated in this sense: start at (x, y)
where f (x, y) is the value. Take a step dx East, and a step dy North, then f
changes approximately df . Instead of choosing dx and dy independently take your
step dx~ + dx~ in the direction of f to get the largest possible increase in f . In
fact df is the dot product of f with the step.] The differentials are often used
in thermodynamics, while the gradient fields are used in many subjects, such as
gravity. In this lecture we are just going to do one example.
It was in the early days of steam engines, when people first found out that there
was a new invention on which they could travel at 25 miles per hour. No human
had ever gone nearly that fast except on a horse, or on ice skates. Can you imagine
the thrill?
It was an outgrowth of the coal mining industry. Coal is fuel, but unfortunately
for the miners, the mines tended to fill with water. A pump was made to fix this
problem and it was driven by an engine which ran on, well, it ran on coal! But
people being as they are, it wasnt long before somebody attached wheels to the
engine and they started competing to see who could go fastest.
At about this time people noticed that every new train went faster than the last one.
The natural question was whether there was any limit to the speed. So M. Carnot
studied this and found that he could keep track of the temperature and pressure of
the steam, but that neither of those was equal to the energy of the moving train.
Eventually it was worked out that the heat energy added to the steam by the fuel
was indeed related to the temperature and pressure. They called the new rule the
first law of thermodynamics. It looked something like this, although the numbers
Im using here are for air, not steam:
heat added = 717 dT + 287 T

dV
V

[Joule/kg]

is supposed to hold whenever a process occurs that makes a small change in the
temperature T [Kelvin] and the specific volume V [m3 /kg] of the gas. Here, the
pressure comes in, again for air, through the ideal gas law P = 287T , V = 1/.
22

A main point discovered: that expression for heat added is not a differential of any
function of T and V . This was so important that they even made a special symbol
for the heat added: d/Q which survives to this day in some books.
P RACTICE : Using simpler numbers and variables, check that
7 dx + 2

x
dy 6= df
y

for any function f (x, y). If it were df = fx dx + fy dy, then you would have
7 = fx

and 2

x
= fy
y

See why that cant be true? What do you know about mixed partial derivatives, fxy
and fyx ?

This relates to the gradient because an alternate way to express that is: For every
function f
x
7~ + 2 ~ 6= f (x, y)
y
Anyway, the big disappointment to the steam engine builders was that the energy
added in the process was not a differential, which meant that you could not make
a table of values for how fast you are going to go, based only on temperature and
pressure.
But, the big discovery was that if you divide the heat added by T , you can make a
table of that.
P RACTICE :
7

dx
dy
+2
= d(something)
x
y

and so for air you also have

717

dT
dV
heat added
+ 287
=
= d(something)
T
V
T

dV
The something is called entropy, denoted S, and for air, 717 dT
T + 287 V = dS.
In Problem 32 you can figure out S from that. There are tables of the entropy of
steam in the back of your thermodynamics book.

So what did Carnot come up with? Is there a limit to how fast the train can go?
Well, he thought of an idealized engine cycle where for part of the time you have
23

dT = 0 and for the other part, dS = 0. Since it is possible to relate the V changes
to the mechanical work of the engine, that allows a computation to proceed. He
worked out how fast that ideal train can go. Youll have to read about it in your
thermo book.
P ROBLEMS
32. Do the practice items if you havent yet. Work out the entropy of air as a function of T
and V .
33. Using the ideal gas law for air, P = 287 T , work out the entropy of air as a function
of T and P .
34. In the isentropic case, meaning entropy doesnt change, we can think of V as a
function of T and write the equation
717

dT
dV
+ 287
=0
T
V

as a first order linear equation

dV
717 1
+
V = 0.
dT
287 T
Solve that. To simplify the numbers, 717/287 = 2.5.
35. Redo Problem 34 thinking of T as a function of V .
36. Use the ideal gas law in the result of Problem 34 to show that P = k1.4 for some
constant k. Well use this in Lecture 17.

Eulers Numerical Method

T ODAY: A numerical method for solving differential equations either by
hand or on the computer, several ways to run it, and how your calculator
works.

Today we return to one of the first questions we asked. If your bank balance
dy
y(t) is $2000 now, and
= .01y so that its rate of change is $20 per year now,
dt
about how much will you have in one year? Hopefully you guess that $2020 is a
reasonable first approximation, and then realize that as soon as the balance grows
even a little, the rate of change goes up too. The answer is therefore somewhat
more than $2020.
The reasoning which lead you to $2020 can be formalised as follows. We consider
x0 = f (x, t)
24

x(t0 ) = x0
Choose a stepsize h and look at the points t1 = t0 + h, t2 = t0 + 2h, etc. We
plan to calculate values xn which are intended to approximate the true values of
the solution x(tn ) at those times. The method relies on knowing the definition of
the derivative
x(t + h) x(t)
x0 (t) = lim
.
h0
h
We make the approximation
. xn+1 xn
x0 (tn ) =
h
Then the differential equation is approximated by the difference equation
xn+1 xn
= f (xn , tn )
h

E XAMPLE : Suppose the bank gives 2.8% interest. With h = 1 it takes only
one step to cover the first year. The bank account equation becomes
y 0 = .028y,
approximated by

yn+1 yn
= .028yn
h

or
yn+1 = yn + .028hyn
This leads to y1 = y0 +.028hy0 = 1.028y0 = 2056. For a better approximation we may take h = .2, but then 5 steps are required to reach the one-year
mark. We calculate successively
y1 = y0 + .028hy0 = 1.0056y0 = 2011.200000
y2 = y1 + .028hy1 = 1.0056y1 = 2022.462720
y3 = y2 + .028hy2 = 1.0056y2 = 2033.788511
y4 = y3 + .028hy3 = 1.0056y3 = 2045.177726
y5 = y4 + .028hy4 = 1.0056y4 = 2056.630721
Look, you get more money if you calculate more accurately!

25

Here the bank has calculated interest 5 times during the year. Continuously Compounded interest means taking h close to 0, so that you are in the limiting situation
of calculus.
We know the answer to this problem. It is y(1) = 2000e.028(1) = 2056.791369....
Continuous compounding gets you the most money. Usually we do not have such
formulas for solutions, and then we have to use this method or some other numerical method.
This method is called Eulers method, in honour of Leonard Euler, a Swiss mathematician of the 18th century. He worked out many things, and in later life he was
blind. Maybe you know that the e in e = 2.718 . . . does not stand for exponential. He also invented some things which go by other peoples names. So show
some respect, and pronounce his name correctly, oiler.
Now well do one for which the answer is not as easily known ahead of time.
Assume that p(t) is the proportion of a population which carries but is not affected
by a certain disease virus, initially 8%. The rate of change is influenced by two
factors. First, each year about 5% of the carriers get sick, so are no longer counted
in p. Second, the number of new carriers each year is about .02 of the population
but varies a lot seasonally. The differential equation is
p0 = .05p + .02(1 + sin(2t))
p(0) = .08
The solution in Figure 11 was computed using Eulers method.
There are more sophisticated methods than Eulers. One of them is built into
octave under the name lsode. You can type help lsode in octave to
get information on it, or try the example:
function xdot = ef(x,t)
xdot = -.05*x+.02*(1+sin(2*pi*t));
end;
t = linspace(0,10,200);
x = lsode("ef",.08,t);
plot(t,x)

We will show one more example to convince you that these computations come
close to things you already know. Look again at the simple equation x0 = x, with
x(0) = 1. You know the solution to this by now, right? Eulers method with step h
gives
xn+1 = xn + hxn
26

0.4

0.35

0.3

0.25

0.2

0.15

0.1

0.05
0

10

20

30

40

50

60

Figure 11: You can see the seasonal variation plainly, and there appears to be a trend to
level off. This is a dangerous disease, apparently.

This implies that

x1 = (1 + h)x0 = 1 + h
x2 = (1 + h)x1 = (1 + h)2

xn = (1 + h)n
Thus to get an approximation for x(1) = e in n steps, we put h = 1/n and receive

1 n
.
e= 1+
n
Lets see if this looks right. With n = 2 we get (3/2)2 = 9/4 = 2.25. With
.
n = 6 and some arithmetic we get (7/6)6 = 2.521626, and so forth. The point
is that these calculations can be done without a scientific calculator. You can even
use a grocery store calculator that only does +*/, and use it to compute important
things.
Did you ever wonder how your scientific calculator works? Sometimes people
think all the answers are stored in there somewhere. But really it uses ideas and
methods like the ones here to calculate many things based only on +*/. Isnt that
nice?

27

E XAMPLE : Well estimate some cube roots by starting with a differential

equation for x(t) = t1/3 . Then x(1) = 1 and x0 (t) = 13 t2/3 . These give
the differential equation
1
x0 = 2
3x
Then Eulers method says xn+1 = xn + 3xhn 2 , and we will use x0 = 1,
h = .1:
.1
= 1.033333...
x1 = 1 +
3
.
Therefore (1.1)1/3 = 1.0333.
x2 = x1 +

h .
= 1.0645
3x1 2

.
Therefore (1.2)1/3 = 1.0645 etc. In fact, (1.0645)3 = 1.206 . . .. For better
accuracy, h can be decreased.
P ROBLEMS

37. What does Eulers method give for 2, if you approximate it by setting x(t) = t and
solving
1
x0 =
with x(1) = 1
2x
Use 1, 2, and 4 steps, i.e., h = 1, .5, .25 respectively.
38. Solve

y 0 = (cos y)2

with y(0) = 0

for 0 t 3 numerically.
39. Solve the differential equation in Problem 38 by separation of variables.
40. Compare your answers to Problems 38 and 39. Is it true that you just computed
tan 1 (3) using only +*/ and cosine? Figure out a way to compute tan 1 (3) using
only +*/.
41. Solve the carrier equation p0 = .05p + .02(1 + sin(2t)) using the integrating factor
method. [The integral is pretty hard, but you can do it.] Predict from your solution, the
proportion of the population at which the number of carriers levels off after a long time,
remembering from Figure 11 that there will continue to be fluctuations about this value.
Does your number seem to agree with the picture?

Spring-mass oscillations
T ODAY: Forced and unforced frictionless oscillations. Natural frequency.

28

The prototype for todays subject is x00 = x. You know the solutions to this already, though you may not realize it. Think about the functions and derivatives you
know from calculus. In fact, here is a good method for any differential equation,
not just this one. Make a list of the functions you know, starting with the very
simplest. Your list might be
0
1
c
t
tn
et
cos(t)
...
Now run down the list trying things in the differential equation. In x00 = x try 0.
Well! what do you know? It works. The next few dont work. Then cos(t) works.
Also sin(t) works. Frequently, as here, you dont need to use a very long list before
finding something. As it happens, cos(t) and sin(t) are not the only solutions to
x00 = x. You wouldnt think of it right away, but 2 cos(t) 5 sin(t) also works,
and in fact any linear combination c1 cos(t) + c2 sin(t) is a solution.
P RACTICE : Find similar solutions to x00 = 9x.

The equations x00 = cx occur frequently enough that you should know all their
solutions.
P RACTICE : Find all solutions to x00 = 0. This is the case c = 0 of x00 = cx.

Consider the equations

x00 + 3x = 0
y 00 + 3y = sin(2t)
The first one is called the homogeneous form of the second one, or the second is
called a forced form of the first. Mechanically what they mean is as follows. Since
we know the solutions to the first one (dont we?) are

x(t) = c1 cos 3t + c2 sin 3t

this first equation is about something vibrating or oscillating. It can be interpreted
as a case of Newtons F = ma law, if you write it as 3x = 1x00 . Here x is
the position of a unit mass, x00 is its acceleration, and there is a force 3x which
29

Figure 12: Unforced and forced springmass systems.

opposes the displacement x. We call this a springmass system. It can be drawn

as in Figure 12, where x is measured up.
The 3x is interpreted as a spring force because it is in the direction opposite x:
if you pull the spring 1.5 units up, then x = 1.5 and the force is 4.5, or 4.5 units
downward. This is also a system without friction, and without gravity, as we see
from the fact that there are no other forces except for the spring force, and that the
oscillation continues
undiminished forever. Note that the natural frequency of

2
:
this system is 23 cycles/second, in the sense that the period of x is
3

2
2
2
+ c2 sin
x t + = c1 cos
3 t+
3 t+
3
3
3

= c1 cos
3t + 2 + c2 sin
3t + 2 = x(t).
The forced equation involves an additional force, as you can see if you write it as
3y + sin(2t) = 1y 00 . The picture in this case is like the right side of Figure 12.
Now we turn to solution methods for the forced equation. We are guided by the
physics.
What will happen with a system which wants to vibrate at a frequency

3
2
of
, and somebody reaches in and shakes it at a frequency of
? Part of the
2
2
motion could be at each of these frequencies. Lets try that. Assume
y(t) = x(t) + A sin(2t)
where x is the solution given above for the unforced equation. Then
y 00 + 3y = x00 + 3x 22 A sin(2t) + 3A sin(2t) = A sin(2t)
We want this to equal sin(2t), so A = 1. Notice how the terms involving x
dropped out. Our solution becomes

y(t) = c1 cos( 3t) + c2 sin( 3t) sin(2t)

30

Figure 13: The top function is the sum of the other two!

One such solution, sin( 3t) sin(2t), is graphed in Figure 13, together with
the individual terms sin( 3t) and sin(2t). In Problem 44 you can explore the
patterns there.

9.1

Conservation laws and uniqueness

Sometimes a second order equation can be integrated once to yield a first order
equation.
For example, lets pretend that we dont know how to solve the equation x00 = x.
You can try to integrate this equation with respect to t. Look what happens:
Z
Z
00
x dt = x dt

P RACTICE : You can do the left side, getting x0 , but what happens on the right?

Multiply the equation x00 = x by x0 . You get

x0 x00 = xx0
P RACTICE : See whether you can integrate it now.

Now xx0 is the derivative of 12 (x)2 , and x0 x00 is the derivative of 21 (x0 )2 . So integrating, you get
1 0 2
1
(x ) = c (x)2
2
2
There is a physical interpretation for this first order equation, which is conservation
of energy. Conservation of energy means the following: x is the position and x0
31

the velocity of an oscillating particle. The energy is the sum of kinetic energy
and potential energy. The kinetic energy 12 mv 2 is 12 (x0 )2 , and the potential energy
1
1 2
2
2 kx is 2 x since here k and m are 1. So what is c? It is the total energy of the
oscillator. The energy is periodically transferred from to potential to kinetic and
back.
Here is an example of the power of the conservation law.
U NIQUENESS T HEOREM There is no other (real-valued) solution to
x00 = x than the ones you already know about.
You probably wondered whether anything besides the sine and cosine had that
property. Of course there are the linear combinations of those. But maybe we just
arent smart enough to figure out others. The Theorem says no: thats all there are.
P ROOF : Suppose the initial values are x(0) = a and x0 (0) = b, and we write
down the answer x(t) that we know how to do. Then suppose your friend claims
there is a second answer to the problem, called y(t). Set u(t) = x(t) y(t) for
the difference which we hope to prove is 0. Then u00 = u. We know from the
conservation law idea that then
1
1 0 2
(u ) = c (u)2
2
2
What is c? The initial values of u are 0, so c = 0. That makes u identically 0. QED
P RACTICE : Do you see why c being zero makes u identically 0?
P ROBLEMS
42. The solutions c1 cos t + c2 sin t of x00 = x really are sinusoids: they can be written
in the form
c sin(t + d).
Use one of the addition formulas
sin(a + b) = sin(a) cos(b) + cos(a) sin(b)
cos(a + b) = cos(a) cos(b) sin(a) sin(b)

to find equations connecting the unknown c and d with the known c1 and c2 .
Note that here we were combining sinusoids having the same frequency.
43. Use the addition formulas for the sine and cosine to combine sinusoids of different
frequencies: show that
c sin(a + b) + d sin(a b) = (c + d) sin(a) cos(b) + (c d) cos(a) sin(b).

32

44. Use the result of Problem 43 to find the period of the slow repetition (beats) in
Figure 13.
45. Find a solution of x00 +4x = sin 3t of the form A sin(3t), and discuss what goes wrong
when you try the same method on x00 + 4x = sin 2t.
46. Find a conservation law for the equation x00 + x3 = 0.
47. Do you think there are any conservation laws for x00 + x0 + x = 0?
48. Whats rong with this? x00 + 4x = 0, x = cos(2t) + sin(2t) + C

10

Applications of Complex Numbers

T ODAY: A method for solving homogeneous linear equations introduces the
exponential of complex numbers. We also use that to solve some forced
equations.

10.1

Exponential and characteristic equation

The motivation for this method is that exponential functions have appeared several
times in the equations which we have been able to solve. Trying x = ert in
ax00 + bx0 + cx = 0
we find ar2 ert + brert + cert = (ar2 + br + c)ert This will be zero only if
ar2 + br + c = 0
since the exponential is never 0. This is called the characteristic equation. For
example, the characteristic equation of x00 + 4x0 3x = 0 is r2 + 4r 3 = 0. See
the similarity? We converted a differential equation to an algebraic equation that
looks abstractly similar.
E XAMPLE : x00 + 4x0 3x = 0 has characteristic equation
r2 + 4r 3 = 0.

Using the quadratic formula,

the roots are r = 2
7. So we have found

two solutions x = e(2 7)t and x = e(2+ 7)t . Check that any linear
combination

x(t) = c1 e(2 7)t + c2 e(2+ 7)t

is a solution too.

33

product

sum
1+2i

1+2i
3+i

3+i
4

Figure 14: Addition of complex numbers is the same as for vectors. Multiplication adds
the angles while multiplying the lengths. The left picture illustrates the sum of 3 + i and
1 + 2i, while the right illustrates the product.
E XAMPLE : x0 + 3x = 0 has characteristic equation r + 3 = 0, so a solution
is e3t . Check that any multiple c1 e3t is too.
P RACTICE : x0000 = 16x gives r4 = 16. One root is r = 2, so one solution is e2t . Are
there others?

E XAMPLE : You already know x00 = x very well, right? But our method
gives the characteristic equation r2 +1 = 0. This does not have real solutions,
so well have to work more to understand this one.

In view of that Example, lets talk about complex numbers.

Complex numbers Complex numbers are expressions of the form a + bi where
a and b are real numbers. You add, subtract, and multiply them just the way you
think you do, except that i2 = 1. So for example,
(2.5 + 3i)2 = 6.25 + 2(2.5)(3i) + 9i2 = 3.25 + 15i
If you plot these points on a plane, plotting the point (x, y) for the complex number
x + yi, you will see that the angle from the positive x axis to 2.5 + 3i gets doubled
when you square, and the length gets squared. Addition and multiplication are in
fact both very geometric, as you can see from Figure 14.
P RACTICE : Use the geometric interpretation of multiplication in Figure 14 to find a
square root of i.

34

Division of complex numbers is best accomplished by using this formula for reciprocals:
1
a bi
= 2
a + bi
a + b2
P RACTICE : Verify that this reciprocal formula is correct, i.e., that if a and b are not
both 0 then
a bi
(a + bi) = 1.
a2 + b2

E XAMPLE : Solve for a:

3
(2 + i)a = 7.
i
We find
a=
= (7 + 3i)

7 3i
7 + 3i
=
(2 + i)
2+i

2i
14 6i + 7i + 3i2
17 + i
=
=
.
2
+1
5
5

22

If z = a + bi and a and b are real, then the real and imaginary parts are re(z) = a
and im(z) = b. (not bi.) Two complex numbers are equal by definition when the
real and imaginary parts are equal.
P RACTICE : Check that 1 + 0i = 0 + (i)i. Why does this not contradict that last
sentence?

In the example x00 = x we found that we needed to understand

ecomplex .
We define e(s+ti) = es eti by analogy with known properties of real exponentials,
but this still requires a definition of eti . We claim that the only reasonable choice
is cos(t) + i sin(t). The reason is as follows. The whole process of solving second
order equations by the characteristic equation method depends on the formula
d(ert )
= rert
dt
Lets require that this hold also when r = i. Writing eit = f (t)+ig(t) this requires
f 0 + ig 0 = i(f + ig) = g + if
35

so f 0 = g and g 0 = f . These should be solved using the initial conditions

e0 = 1 = f (0) + ig(0). These give f (0) = 1, f 0 (0) = 0 and f 00 = f . The only
solution is f (t) = cos(t) and g(t) = sin(t). Therefore our definition becomes
es+ti = es (cos t + i sin t).
Now it turns out that we actually get
lem 53.)

d(ert )
= rert for all complex r. (See Probdt

Now we can do more examples.

E XAMPLE : x0000 = 16x again. We find solutions ert if
r4 = 16,

r2 = 4, 4,

so
r = 2, 2, 2i, 2i.
These give real valued solutions x = e2t , x = e2t , and complex valued
solutions x = e2it , x = e2it . The most general solution is
x(t) = c1 e2t + c2 e2t + c3 e2it + c4 e2it .

The last time you went into the lab, all the instruments were probably showing
real numbers, werent they? So it is common to rewrite our solutions in a real
form. The following new idea is required: If x(t) is a complex solution to a linear
differential equation with real coefficients then the real and imaginary parts of x
are also solutions!
For example, e3it is a solution to x00 = 9x. The real and imaginary parts are
respectively cos(3t) and sin(3t), and these are certainly solutions also.
To see why this works in general, suppose that x = u+iv solves ax00 +bx0 +cx = 0.
This says that
a(u00 + iv 00 ) + b(u0 + iv 0 ) + c(u + iv) = (au00 + bu0 + cu) + i(av 00 + bv 0 + cv) = 0.
Assuming that a, b, c, u, and v are all real, you can conclude that au00 + bu0 + cu
and av 00 + bv 0 + cv are also 0.
P RACTICE : The same idea works for equations of any order. Try a first order one.

P RACTICE : The result does not work for the equation x00 + x2 = 0. Why?

36

So lets rework the previous example. One solution to x0000 = 16x is e2it . The real
and imaginary parts are cos(2t) and sin(2t). In fact we can write the previously
given solution
x(t) = c1 e2t + c2 e2t + c3 e2it + c4 e2it
as

= c1 e2t + c2 e2t + c3 cos(2t) + i sin(2t) + c4 cos(2t) i sin(2t)

= c1 e2t + c2 e2t + (c3 + c4 ) cos(2t) + (c3 c4 )i sin(2t)
= c1 e2t + c2 e2t + c5 cos(2t) + c6 sin(2t).
E XAMPLE : Solve

y 00 + 2y 0 + 2y = 0

The characteristic
equation is r2 + 2r + 2 = 0, so by the quadratic formula

2 48
r=
= 1 i. One solution is e(1+i)t = et (cos(t) + i sin(t)).
2
Taking the real and imaginary parts, the solution is
y(t) = c1 et cos(t) + c2 et sin(t)
Dont forget to check it.
P ROBLEMS
49. Solve r2 6r + 10 = 0, r3 6r2 + 10r = 0, and r2 6r 10 = 0.
50. Solve y 00 + 3y 0 + 4y = 0.
51. Find a, if a +

1
a

= 0; if a +

1
a

= i.

52. Sketch the graphs of the functions et cos(t), et cos(3t), and e4t sin(3t). Do these
fit your idea of an oscillation with friction?
53. Use the definition

e(a+bi)t = eat (cos(bt) + i sin(bt))

to verify that

d(ert )
= rert
dt
when r = a + bi is any complex number.
54. For each t, e(1+i)t is a complex number, which is a point in the plane. So as t varies, a
curve is traced in the plane. Sketch it.
55. A polynomial r2 + br + c = 0 has roots r = 2 i. Find b and c.
56. A characteristic equation r2 + br + c = 0 has roots r = 2 i. What was the
differential equation?
57. Whats rong with this? y 00 + y 2 = 0, r2 + 1 = 0, r = 1, y = et + c. There are at
least 3 errors.

37

10.2 The Fundamental Theorem of Algebra

Think about this: A new number written i was invented to solve the equation x2 +
1 = 0. Of course you have noticed that the people who invented it were not very
happy about it: contrast complex and imaginary with real and rational.
But soon the following theorem was discovered.
By introducing the new number i, not only can you solve x2 + 1 = 0, but you can
also solve at least in principle all these:
x2 + 2 = 0,

x3 i = 0,

and even
x6 (3 2i)x4 x3 + ix 39.1778i + 43.2 = 0.
F UNDAMENTAL T HEOREM OF A LGEBRA Let
p(x) = a0 + a1 x + a2 x2 + + an1 xn1 + xn
be a polynomial of degree n > 0 with any complex coefficients ak .
Then there are complex numbers r1 , r2 , . . . rn which are roots of p
and p factors as
p(x) = (x r1 )(x r2 ) (x rn )
As a practical matter, we can handle the cases a0 + x and a0 + a1 x + x2 easily. We
all know the quadratic formula. There is a cubic formula that most people dont
know for solving the n = 3 case, and a quartic formula that hardly anybody knows
for the n = 4 case, which takes about a page to write. Then something interesting
happens.
A BEL S T HEOREM Let n 5. Suppose you figure out every formula
that could ever be written in terms of the coefficients ak and the operations of addition, subtraction, multiplication, division, and extraction
of roots. Then none of those formulas give you the roots of the polynomial.
So: the roots exist, but if you need them then you will usually have to approximate
them numerically.
P ROBLEMS

38

58. In the Fundamental Theorem we took the coefficient of xn to be 1 just for convenience.
But if you dont do that the factorization must be written differently. For example,
6 + x + x2 = (x 2)(x + 3)
is correct. Figure out the number c in the case
30 + 5x + 5x2 = c(x 2)(x + 3)
59. How must the factorization be written in general if you dont assume the coefficient of
xn is 1?
60. You know that in the factorization 6 + x + x2 = (x 2)(x + 3) you have 6 =
(2)(3). In general what is the product of the roots r1 r2 rn in terms of parameters
appearing in the Fundamental Theorem?
61. You know that in the factorization 6 + x + x2 = (x 2)(x + 3), the x term comes
from x = 2x + 3x when you multiply the right hand side. In general what is the sum of
the roots r1 + r2 + rn in terms of parameters appearing in the Fundamental Theorem?

10.3 A forced oscillator

Lets look at a spring and mass with friction and sinusoidal forcing.
Take F = ma in the form
kx cx0 + cos(t) = mx00
for example
x00 + x0 + x = cos t.
We point out the following.
1. x is the real part of the solution to
z 00 + z 0 + z = eit
2. It is easy to find a solution z(t).
3. Therefore a solution x can be found in two steps.
Lets see why these are true. 1. To say that x is the real part of z means that
z(t) = x(t) + iy(t)
for some real-valued y. Use the definition eit = cos t + i sin t and differentiate z
to get
z 00 + z 0 + z = x00 + iy 00 + x0 + iy 0 + x + iy = cos t + i sin t.
39

The real parts must match:

x00 + x0 + x = cos t
(and the imaginary parts too: y 00 + y 0 + y = sin t). 2. To find z we guess that the
response will be proportional to the forcing,
z(t) = Aeit .
Substituting,
z 00 + z 0 + z = (i2 A + iA + A)eit = iAeit .
We need this to be eit , so A = i1 = i. We have found
z(t) = ieit = i(cos t + i sin t) = sin t i cos t.
Since x is the real part, we found a solution
x(t) = sin t.

P RACTICE : Check it.

Note: We made a guess and found a solution, but did not find all of them. We
found, in a sense, the most important one. The reason is that all other solutions
approach this one as time goes by. (See Problem 63.)
The force cos t shakes the oscillator back and forth once every 2 seconds. Now
lets see what happens if we shake the oscillator at various frequencies. Remember
that our definition
ei = cos + i sin
means: the point at angle on the unit circle of complex numbers.
P RACTICE : Where on the unit circle will I find the number ei ?

Please do not read any further until you understand that Practice problem, because
I want you to understand what comes after it. Now eit = cos(t) + i sin(t) runs
around the unit circle at a constant speed which depends on . So the real part
oscillates between 1, right?
Next, consider the function Aeit where A is complex. Multiplication by A adds
the angle of A, and rescales the size by |A|. So Aeit runs around the circle of
radius |A|.
40

P RACTICE : So the real part of Aeit oscillates between what two real numbers?

Consider the equation

x00 + x0 + x = cos(t)
which is like before except for the frequency parameter . Do our two step method:
x = re z
where z 00 + z 0 + z = eit . Try
z = Aeit .
Then

z 00 + z 0 + z = (i)2 + i + 1 Aeit =? eit .

We need ( 2 + i + 1)A = 1 or
A=

1
.
1 2 + i

If you did the Practice problem (eh?) then you know that x(t) oscillates between
|A| and |A|. This depends on , so lets see how so:
|A| =

1
1
=p
.
|1 2 + i|
(1 2 )2 + 2

P RACTICE : A quick check on magnitudes of complex numbers: Is |3 + 4i| equal

to 32 + 42 or to 32 42 ? According to Pythagoras, which one gives the correct

distance to the origin?

Figure 15 shows the graph of the amplitude of x(t) as a function of . This is

called a response curve. It shows that the largest oscillations occur if the applied
force has frequency about 1. There is very little response if you shake it fast.
P RACTICE : Try an experiment of that type, say with your keychain hanging from a
rubber band. Jiggle the upper end of the rubber band fast and see whether the keys
move very far.
P ROBLEMS

62. Find a solution to y 00 + y 0 + 3y = sin(t) in the form y(t) = im Aeit .

41

1.2

0.8

0.6

0.4

0.2

0
1

Figure 15: Graph of the amplitude (maximum over t) of x(t) as a function of the forcing
frequency .
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
u

Figure 16: Three equal masses m, equal spring constants k. Their positions u, v, and w
are measured from the three equilibrium positions.
63. Check that if x1 and x2 are any two solutions of x00 + x0 + x = cos t then the difference
x = x1 x2 satisfies x00 + x0 + x = 0. Then use the characteristic equation to see that
x(t) tends to 0 as t increases.
64. Plot the response curves for x00 + bx0 + x = cos(t) for small and large values of b.
Do they have the same shape?

11

Three masses oscillate

T ODAY: Three masses, three frequencies, three modes of oscillation.

Newtons ma = F for the left mass in Figure 16 is

mu00 = ku k(u v).
You wont have any trouble understanding the term ku which is the force due to
the left spring, but the k(u v) for the middle spring needs some thought.
P RACTICE : If u and v move the same distance to the right, what is u v and what is
the force on the left mass due to the middle spring?

42

P RACTICE : If v > u at some time, is the force to the right or left? Does this match
the expression k(u v)?

After thinking carefully about that, you will be able to write the Newtons laws for
the middle and right masses:
mv 00 = k(v u) k(v w),
mw00 = k(w v).

E XAMPLE : Suppose the masses are located at u = 1, v = 2, w = 4.

What are the forces? The left spring is stretched by 1 unit, the middle one by
2 1 = 1, the right by 2. So there is no net force on the left mass, k to the
right on the middle mass, and 2k, that is 2k to the left on the right mass.

To simplify the equations,

p take m = k = 1. Or, if you prefer, the time t can be
rescaled by a factor of m/k. The system becomes
u00 = 2u + v
v 00 = 2v + u + w
w00 = w + v.
Lets combine the 3 equations. From the u00 equation we get v in terms of u:
v = u00 + 2u.
From the v 00 equation we get w = v 00 + 2v u. Substitute to get w in terms of u:

00

w = u00 + 2u + 2 u00 + 2u u = u0000 + 4u00 + 3u.

Substitute these expressions into the w00 equation:
0000
00

u + 4u00 + 3u = u0000 + 4u00 + 3u + (u00 + 2u)

Combine terms to get a single equation for u:
u000000 + 5u0000 + 6u00 + u = 0.
Figure 17 shows how complicated a solution u to this equation can be. Our job is
to make sense of that!
43

3
2
1
0
-1
-2
-3
0

20

40

60

80

100

Figure 17: An example graph of u. Too complicated to understand (but less complicated
than, say, a seismograph). It suggests that the smallest disturbances are equally spaced in
time. That is a clue about the highest of the 3 frequencies.

Lets find simple solutions u. Try u(t) = ert as usual. We need

r6 + 5r4 + 6r2 + 1 = 0.
According to the fundamental theorem of algebra there will be 6 roots r. To locate
them, notice there are no odd degree terms, so set r2 = a. We need
a3 + 5a2 + 6a + 1 = 0.
This polynomial in the variable a is alternately 1, 1, 1 when a = 2, 1, and 0,
as we find by experimenting. Since the leading term is a3 , the values are negative
for a large and negative and positive for a positive. According to the intermediate
value theorem then there are roots between 0 and 1, between 1 and 2, and
less than 2. These are illustrated in Figure 18.
Since the three roots a = r2 are negative, the six roots r are pure imaginary, and
they are nearly
r = 1.802i, 1.247i, .444i.
Therefore the solution u is a linear combination
u(t) = c1 cos(1.802t) + c2 sin(1.802t)
+c3 cos(1.247) + c4 sin(1.247t) + c5 cos(.444t) + c6 sin(.444t)

44

2.5
2
1.5
1
0.5
0
-0.5
-1
-1.5
-2
-3.5

-3

-2.5

-2

-1.5

-1

-0.5

Figure 18: Graph of the cubic polynomial a3 + 5a2 + 6a + 1. The roots are approximately
3.2, 1.6, .3.

and this corresponds to the complexity of Figure 17. What we want instead is to
understand the individual modes. Take
u(t) = cos(t)
where is any one of the three angular frequencies 1.802, 1.247, or .444. Use our
previous substitutions to find
v(t) = u00 + 2u = ( 2 + 2)u
and
w(t) = u0000 + 4u00 + 3u = ( 4 4 2 + 3)u
What does that mean? w and v have different amplitude, and sometimes opposite
directions from u, but the same frequency. See Figure 19.

2 + 2
4 4 2 + 3
u
v
w

1.802
1.247
.555
cos(1.802t)
1.247u
.555u

1.247
.444
.802
cos(1.247t)
.444u
.802u

.444
1.802
2.250
cos(.444t)
1.802u
2.250u

The general solution consists of a combination of these three modes, which are
hidden in Figure 17.
These complicated motions contain simpler motions at just 3 frequencies. This
example relates to our later study of Fourier series: Fouriers project was to find all
the frequencies hidden within functions.
45

11111111111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111111111
u

11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111

11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111

Figure 19: The three modes illustrated. At each time when u(t) = cos(t) is maximum,
v and w are as shown. = .444, 1.247, 1.802 from top to bottom. Vertical lines mark the
equilibrium positions.
P ROBLEMS
65. If y1 is a solution to y 000 + 3y 0 + 2y = 0, define y2 = y10 . Differentiate the differential
equation to explain why y2 and y1 + 8y2 are also solutions. Based on this idea, can you
explain why u, v, and w all satisfy the same 6-th order equation?
66. The oscillation graphed in Figure 20 is a sum cos(at) + cos(bt) + .2 cos(ct). Estimate
the angular frequencies a, b, and c.

46

2.5
2
1.5
1
0.5
0
-0.5
-1
-1.5
0

10

12

Figure 20: The graph in Problem 66. Remember 6 is approximately 2.

12

Boundary Value Problems

T ODAY: A change from initial conditions. Boundary values. We find that we
dont know everything about y 00 = y after all.

First you have to know what a boundary is. It is nearly the same thing in mathematics as on a map: the boundary of a cube consists of its six faces, the boundary
of Puerto Rico is its shore line, the boundary of the interval [a, b] consists of the
two points a and b. The concept of a boundary value problem is to require that
some conditions hold at the boundary while a differential equation holds inside the
set. Here is an example.
E XAMPLE :
y 00 = y
y(0) = 0
y(2) = 0
We are asked to solve a very familiar differential equation, but under very
unfamiliar conditions. The function y is supposed to be 0 at 0 and . The
differential equation here has solutions y = A cos(t) + B sin(t). We apply
the first boundary condition, giving y(0) = 0 = A. So we must take A = 0.
That was easy! Now apply the second boundary condition, giving y(2) =
0 = B sin(2). Well, it just so happens that the sin(2) is zero. So the
second boundary condition is fulfilled no matter what B is. Answer: y(t) =
B sin(t), B arbitrary.

47

Notice how different this example was from our experience with initial conditions,
that there were infinitely many solutions. Just the opposite thing can happen too:
E XAMPLE :
y 00
y(0)
y(6)

= y
= 0
= 0

This seems close to the previous example since only 2 has been changed to
6. This begins as before with A = 0. Then it just so happens that the sin(6)
is not zero. So unlike the previous case, we have to set B = 0. Answer:
y(t) = 0. There is no other possibility.

What is a physical interpretation of these problems? We know that the equation

y 00 = y describes an oscillator, a rock hanging from a Slinky, vibrating up and
1
down at frequency 2
. The condition y(0) = 0 says you want the rock at the
origin at time 0. That just means you start the clock when it is there. Since the
1
, the rock will be back at the origin after 2 seconds.
frequency of oscillation is 2
The condition y(2) = 0 happens automatically. But if you require instead that
y(6) = 0, then you are asking for the impossible. The rock takes 2 = 6.283 . . .
seconds to get back, period. The only exception is the special function 0, signifying
that the rock never moved at all. In that case it will certainly be back in 6 seconds,
since it is there already. That is why there is no solution in the second case except
for the 0 solution.
There is a tendency to use x as the independent variable when one is discussing
boundary value problems rather than t because usually in practice it is position
rather than time, in which one is interested. Well do that in the next example.
There is also the concept of eigenvalue in connection with boundary value problems.
E XAMPLE :
y 00 = cy
y(0) = 0
y() = 0

Finding c is part of the problem too.

The eigenvalue here is c. We will assume we are looking for positive values
of c. Problem 67 will analyze the cases in which c might be 0 or negative.
To solve this equation y 00 = cy
we recognize that
we are again looking at
sines and cosines: y(x) = A cos( cx) + B sin( cx), right? The boundary

48

condition y(0) = 0 = A cos(0) = A determines A = 0 as before. Then

we have to deal with the other boundary condition y() = 0 = B sin( c).
Certainly if B = 0 this is satisfied, and we have found a solution y(x) = 0.
But we have another parameter c to play with, so lets not give up too easily.
Question: Is it possible that

sin( c) = 0
for some values of c? If so, B can be anything and such c will be our
eigenvalues. Well, what do you know about where sine is zero? Certainly
sin(0) = 0 but if c = 0 then we are back at y(x) = 0 which is uninteresting
by now. Where
else is the sine zero? The sine is zero at , for example.
That gives c = , or c = 1. There is a good solution to our problem!
y(x) = B sin(x), B arbitrary, c = 1. But there are other
places where the

sine is zero. For example sin(2) = 0. That gives c = 2, or c = 4.

So there is another solution y(x) = B sin(2x), B arbitrary, c = 4. Also
sin(3) = 0, giving a solution y(x) = B sin(3x), B arbitrary, c = 9. In
general we have solutions
y(x) =
c =

B sin(nx) B arbitrary
n2 n = 1, 2, 3, . . .

So the eigenvalues for this problem are 1, 4, 9, 16, 25, . . .

P RACTICE : For which spring mass systems
my 00 = ky
with y(0) = 0 is the mass located again at 0 after seconds?

Eigenvalues are all around you. Well mention them again in connection with string
vibrations on page 68.
P ROBLEMS
67. Find all the solutions to the c 0 cases of
y 00 = cy
y(0) = 0
y() = 0
68. Find all the positive eigenvalues for the problem
y 00 = y
y(0) =0
y(6) =0

49

13

The Conduction of Heat

T ODAY: A derivation of the heat equation from physical principles. Two, in
fact.

The heat equation looks like

ut = auxx .
It is an abbreviation for

u
2u
(x, t) = a 2 (x, t)
t
x
It describes the temperature of an object in which heat energy is allowed to flow
by conduction in the x direction. Heat can also move with a flowing fluid but that,
known as convection, is not what this is about. Heat can also move as radiation,
like the warming you feel near a fire even though the air itself may be cold. It is
not about that either. It can describe the changing temperature along a metal bar or
through a thick wall.
The equation concerns the temperature u(x, t) of a metal bar along which heat is
conducted. You can imagine one end of the bar in the fire and the other in the
blacksmiths hand, to emphasize that the temperature may change with time and
with position along the bar. The number a is a physical constant which we will
explain while we derive the heat equation.
Now lets read the heat equation carefully to see whether it makes any sense or not.
Suppose you graph the temperature as a function of x, at a particular time. Maybe
the graph is concave up as in Figure 21 on the left. What do you know from calculus
about functions which are concave up? Isnt the second derivative positive? The
heat equation contains the second derivative, and says that the more positive it is,
the bigger the time derivative will be. What do you know from calculus about the
first derivative being positive? Doesnt it mean that the function is increasing? So
we have to imagine what that means, and conclude that the temperature must be
rising at any point of the bar where the temperature graph is concave up. Similarly
for the right side of Figure 21. If the temperature graph is concave down, the
temperature must be decreasing with time. Does that seem correct? There is no
hope of understanding partial differential equations unless you think about things
such as that.
As you will see from our discussion, it is not necessary to think of a metal bar. It is
also possible to think of the conduction of heat in the x direction, where this axis
passes through a door or wall, or any similar situation in which heat energy flows

50

0
1
0
1
0
1

Figure 21: Temperature versus position along a metal bar. The arrows show the time
dependence, according to the heat equation. On the left the temperature graph is concave
up and the bar is warming, and on the right it is cooling. The shaded slab is used in the
derivation.

in one dimension only. If you want to talk about heat flow along the surface of a
sheet of metal you will have to use the two-dimensional heat equation
ut = a(uxx + uyy )
If you want to talk about heat flow throughout a solid block of metal or though still
air, you use the three-dimensional heat equation
ut = a(uxx + uyy + uzz )
Our plan is to derive the one-dimensional heat equation by showing what physical
principles are behind it, and by using what you might call elementary mathematics,
i.e. calculus. Later we will show a different derivation for the three-dimensional
heat equation using the exact same physical principles, but more sophisticated
mathematics, the Divergence Theorem.
Consider a little slab cut from the bar between coordinates x and x + x, and
suppose the density of the bar is mass per length. We are going to account for the
heat energy contained within this little slab. We have
x = length of the slab
x = mass of the slab
physical principle # 1: There is something called specific heat, c, which reflects
the experimental fact that a pound of wood and a pound of steel at the same temperature do not contain the same amount of heat energy. According to this principle
we have
cux = the heat energy content of the slab
physical principle # 2: Heat flows from hot to cold, and more specifically there is
something called conductivity, k, which reflects the fact that copper conducts heat
better than feathers do. According to this principle we have
kux (x, t) = the rate heat enters the left side of the slab
kux (x + x, t) = the rate heat enters the right side
51

Figure 22: A region R inside the main block of material, with some heat energy flowing
between R and the rest of the block. Energy conducts into R through its boundary surface
S. There is no heat source or sink, chemical reaction etc., inside R. You can place R
anywhere within the material.

These two principles give us two different expressions for the rate at which heat
energy enters the slab. This is wonderful! Whenever you have two different expressions for the same thing, you are about to discover something important. Equate
them:
(cux)t = kux (x + x, t) kux (x, t).
Then divide by x to get
cut =

kux (x + x, t) kux (x, t)

.
x

Finally take the limit as x 0 to get

cut = kuxx .
This gives the heat equation, and shows us that
a=

k
.
c

A second derivation may be accomplished in three dimensions by using the divergence theorem. We imagine heat conduction in a solid block of some material, as
in Figure 22. Fix any region R inside the solid and account for heat flow into R.
The heat content is
ZZZ
cu dV
R

52

and its rate of change is

ZZZ
R

cut dV.

The rate of heat flow through the boundary surface S is

ZZ
~ ~n dA.
k u
S

By the divergence theorem this last integral is equal to

ZZZ
~ (k u)
~ dV.

Thus we have

ZZZ

ZZZ
R

~ (k u)
~ dV.

cut dV =

Subtract to get
ZZZ
R

(cut k(uxx + uyy + uzz )) dV = 0.

for all R. Now if a function happens to integrate to zero it means nothing, except
that here it is true for all R. This turns out to be enough (See Problem 72) to insure
that the integrand is zero, and we get
cut = k(uxx + uyy + uzz )

13.1 Walk the line

Lets see what happens when we allow heat to conduct along the whole number
line. Remember a derivative is a limit of difference quotients, like
f 0 (a) = lim

h0

f (a + h) f (a)
.
h

Well use that for time derivative:

ut (x, t)

u(x, t + t) u(x, t)
.
t

To simplify that, consider the arrangement in Figure 23. In that notation the time
derivative approximation becomes
ut |point 2

unew u2
.
t

53

new

t
u

Figure 23: A stencil.

For the second space derivative uxx , remember it is the x derivative of ux . So we

can do this:
u3 u2
u1
u2x
uxx (x, t) x
x
u3 2u2 + u1
=
.
(x)2
The heat equation becomes the statement
unew = u2 +

t
(u3 2u2 + u1 )
(x)2

at each point within the (x, t) plane where x is in the material and t 0. In fact
t
1
the simplest case is when
= . Then the u2 s cancel and the whole thing
2
(x)
2
simplifies to
1
unew = (u3 + u1 )
2
We start with temperature u = 0 everywhere along the line except u = 1 at the
origin.
P RACTICE : Take x = 1, t = 12 , u1 = u(1, 0) = 0, u2 = u(0, 0) = 1,
u3 = u(1, 0) = 0. Figure out unew which approximates u(0, .5).

1
Then we use unew = (u3 + u1 ) to work forward in time (upward in Figures 23
2
and 24), like in the practice problem, and find the values shown in the table. In
Problem 73 you fill values marked with asterisks (*) at time 5t.
In Figure 24 the first and last lines of the table have been plotted with dots connected to show the calculated temperatures at times 0 and 6t.
54

6t
5t
4t
3t
2t
t
0

1
64

6
15
0
0
0
64
64
0
*
0
*
0
*
1
4
0
0
0
0
16
16
1
1
0
0
0
0
8
8
1
0
0
0
0
0
4
1
0
0
0
0
0
2
0
0
0
0
0
0
6x 5x 4x 3x 2x x

20
64

6
16

0
2
4

0
1
0

15
64

1
0
64
*
0
4
1
0
0
16
16
1
1
0
0
0
0
8
8
1
0
0
0
0
0
4
1
0
0
0
0
0
2
0
0
0
0
0
0
x 2x 3x 4x 5x 6x

0
*
0

2 x

0
*
0

6
64

4 x

Figure 24: Top: A table of temperatures calculated using the stencil in Figure 23 with
t = 12 (x)2 . Center: A sketch of the calculated temperatures u as function of position,
at times 0 and 6t. using values given in the first and last lines of the table. Bottom: The
2
2
graphs of y = ex /4 and y = 21 ex /16 .

Now: the numerical method leads, with a bit of wishful thinking, to the functions
graphed at the bottom of the figure. The comparison suggests temperature functions of the form
2
u(x, t) = aebx
where a and b could be functions of t. Lets try that! We find derivatives
2

ux = 2xabebx ,

uxx = 2abebx + 4x2 ab2 ebx

and

ut = a0 ebx ab0 x2 eb(t)x .

In order to have ut = uxx , we need
a0 = 2ab,

ab0 = 4ab2 .

Since those are not familiar, lets see what part we can recognize. In the first
equation we could solve for a(t) if b were constant or if we knew b. In the second
equation we can cancel a, giving
b0 = 4b2 .
55

We can solve that one because it is separable. If b 6= 0 divide to get

b2 b0 = 4,
and integrate to get
b1 = 4t + c.
Then b(t) =

1
. Take the simplest case, where c = 0. Then
4t + c
b(t) =

1
,
4t

a0 =

1
a.
2t

Now the a equation is separable too.

1
P RACTICE : Check that a(t) = is a solution to that last equation.
t

We have found

1
2
u(x, t) = ex /4t .
t
Of course it is only defined for t > 0 because of the square root.
2
1
P RACTICE : Check derivatives again to verify that ex /4t is a solution to ut =
t
uxx .

P ROBLEMS
69. Suppose the same temperature distribution is present in two metal bars at a particular
time, where the first bar has ut = 3uxx and the second has ut = 4uxx , due to different
material. Which bar cools faster?
70. A new kind of physics is discovered on Mars. Their heat equation says ut = auxx u.
Do Mars bars cool faster or slower than Earth bars?
71. Amazingly enough, we have discovered that the heat equation used by the Klingons is
different from either Earth or Mars. They have ut = auxx + u, ut = a(uxx + uyy ) + u
etc. Sometimes their pizza doesnt cool off at all. Why?
72. There is a way of thinking about continuous functions which is a little beyond the level
of these notes, but you might be interested to hear it: Suppose continuous function f is
not zero at some particular point a, f (a) 6= 0. Maybe it is 0.266 there. Then you can
take some other points x sufficiently close to a that f (x) is near 0.266, say f (x) > 0.200.
This will hold for any point x in at least some small neighborhood of a. Knowing this,
we take
Z ZRZ in Figure 22 to be located right in that neighborhood. Then what can you say
about

f dV ? This shows that if these integrals are zero for all R, then f must be
R

indentically 0.

56

73. Use the numerical method unew = 12 (u3 + u1 ) to fill in the missing (*) values in the
table in Figure 24.

14

Initial Boundary Value Problems for the Heat Equation

T ODAY: The meaning of boundary values. Steady state solutions. Product
solutions.

There are many, many solutions to the heat equation, by comparison with any
ordinary differential equation. For example we found a solution
1
2
ex /4t .
t
More simply, all constants are solutions. A metal bar can certainly have a constant temperature. Another special class of solutions which are easy to find are the
steady-state solutions, i.e. the ones which do not depend on the time. Everyday
experience suggests that there are such cases; consider heat conduction through
the wall of a refrigerator. The kitchen is at a constant temperature, the inside of the
refrigerator is at a constant temperature, and there is a continual flow of heat into
the refrigerator through the walls. To find such solutions analytically we assume
that u is a function of x only. What happens to the heat equation then? We have
ut = 0 and ux = u0 (x), so the heat equation becomes u00 = 0. Integrating once we
get u0 = a. Integrating again gives u = ax + b. All straight lines are the graphs
of steady-state temperature distributions. These straight-line solutions include the
constant solutions.
So far we have not said much about boundary conditions or initial conditions. It
is typical of partial differential equations that there are many possible choices of
conditions, and it takes a lot of work to decide what the reasonable ones are. Lets
say for now that we are interested in heat conduction along a bar of length l which
is located in 0 x l. Then one set of conditions which specifies a solvable
problem is as follows. You may specify the temperature at the ends of the bar, and
the initial temperature along the bar. For example here is one such problem.
ut = uxx

the heat equation

u(0, t) = 40

a boundary condition at the left end

u(5, t) = 60

boundary condition at right end; the length is 5

u(x, 0) = 40 + 4x

the initial temperature

57

You may consider whether these conditions seem physically reasonable or not.
The bar is going to have one end kept at 40 degrees, the other at 60 degrees, and
the initial temperature is known. Should this kind of information be adequate to
determine the future temperatures in the bar? This is a hard type of question in
general, but in this case the answer turns out to be yes. In fact we can solve this
particular problem with one of our steady-state solutions.
P RACTICE : Find the steady-state solution to that problem.

Next consider the problem

ut
u(0, t)
u(5, t)
u(x, 0)

=
=
=
=

uxx
0
0
40

Here we have a bar of length 5 which begins uniformly at 40 degrees, and at time
zero somebody presses ice cubes against both ends of the bar and holds them there
forever. What should happen? In this case the initial condition does not match the
boundary conditions, so the solution, if it exists, cannot be a steady-state solution.
In this case, it is not possible to just quickly reason out whether the problem has
any solution, or whether it might have more than one.

14.1

Insulation

There is another popular boundary condition known as insulation. What does insulation do? If you run out of fuel in the winter, your house cools, because energy
conducts outward to the environment. It takes longer to cool, the better the insulation is. So we model insulation by the assumption that no energy passes through it.
From our derivation of the heat equation, we know that the rate of energy conduction is a multiple of ux . You might want to reread the derivation on this particular
point. So: The boundary condition at an insulated end is ux = 0.
Another way to think about insulation is like this: heat flows from hot to cold.
So for heat to flow, there must be a temperature difference. On the other hand if
ux = 0 somewhere, then the graph of temperature has slope 0, and microscopically
there is no temperature difference for small position changes there.
Now what is the following problem about?
ut = uxx
58

ux (0, t) = 0
u(l, t) = 60
u(x, 0) = 40
This is a bar of length l, insulated at the left end and held at 60 degrees at the right
end. The initial temperature is 40 degrees all along its length.

14.2 Product Solutions

There is another class of solutions to the heat equation which can be found by a
method we have used many times before. We try a solution of the form u = cert .
This idea worked on ordinary differential equations. Now we need to allow c to be a
function of x. Substituting into the heat equation ut = auxx gives rcert = ac00 ert .
These will agree provided that ac00 = rc. We know how to solve ac00 = rc. It is
a second order linear ordinary differential equation for c, and it is one of the easy
kind with no c0 term. To simplify it a little lets write r = aw2 , then what we
need is c00 = w2 c. The solutions are c = A sin(wx) + B cos(wx). Thus we have
found many solutions to the heat equation of the form
2

u = (A sin(wx) + B cos(wx))eaw t .

E XAMPLE : Determine the values of w such that there is a product solution

to
ut = auxx
having frozen boundary conditions
u(0, t) = u(, t) = 0
2

We compute u(0, t) = Beaw t in the product solution above. This will be

2
zero for all t only if B = 0. So far we have u = A sin(wx)eaw t . Then we
2
compute u(, t) = A sin(w)eaw t . This will be zero for all t if we choose
w to be any of the zeros of the sine function. These make w =1, 2, 3,....We
have found a list of solutions:
2

un (x, t) = An sin(nx)ean t , n = 1, 2, 3, ....

Compare the solutions found in the Example to the eigenvalue problem discussed
in Lecture 12.
P ROBLEMS

59

74. Describe in whole sentences what the following problem is about.

ut =
ux (0, t) =
ux (l, t) =
u(x, 0) =

uxx
0
0
400

Determine whether there is a steady-state solution to this problem.

75. Describe what this problem is about.
ut
u(0, t)
u(l, t)
u(x, 0)

=
=
=
=

uxx
0
0
400

Determine whether there is a steady-state solution to this problem.

76. Find a product solution to
ut
u(0, t)
u(, t)
u(x, 0)

=
=
=
=

5uxx
0
0
2 sin(x)

77. Determine whether there is a product solution to

ut
ux (0, t)
ux (, t)
u(x, 0)

= uxx
= 0
= 0
= 3

78. Suppose that u1 and u2 are solutions to

ut = uxx
u(0, t) = 0
u(, t) = 0
Note that no initial condition has been specified. Set s = u1 + u2 . Is it true that s is also a
solution to this system?
79. Find a solution to
ut
u(0, t)
u(, t)
u(x, 0)

=
=
=
=

uxx
0
0
3 sin(x) + 5 sin(2x)

which is a sum of two product solutions. If you need help, look ahead to the next lecture,
for which this problem is intended to be a preview.

60

Figure 25: Ice cubes at the ends of the bar. There is insulation around the bar to make sure
we have heat conduction in the x direction only.
2

80. Plot the graphs of the product functions un (x, t) = sin(nx)ean t versus x for several
values of t, for the cases n = 1 and n = 2. Describe what is going on in your graphs, in
terms of temperature in a metal bar.
81. In this problem
ut = uxx
u(0, t) = 0
u(, t) = 0
suppose u(x, t) = eat y(x), i.e. an exponential function of t times a function of x. What
boundary value problem, of the type considered in Lecture 12 must y solve?
82. Lets relate the heat equation to Newtons law of cooling. Take the product solution
u = sin(x)et
to ut = uxx , and define the average temperature of the bar to be
Z
1
T (t) =
u(x, t) dt.
0
Show that T 0 = T , a version of Newtons law if you imagine the bar isolated in a frozen
2
environment. More generally for any product solution u = sin(nx)ean t show that the
average temperature obeys T 0 = an2 T . (But not all heat solutions obey Newtons law of
cooling; see Problem 86 on page 64.)

14.3

Superposition

We are going to analyze a heat conduction problem which happens to be solvable

using some of the product solutions found previously. Our emphasis is not on formulas, but on understanding what the solutions mean, and thinking about the extent
to which this heat equation corresponds with our daily experience of temperature.
Consider the problem in which a metal bar with frozen ends has temperature initially as in Figure 26. We want to find out how this temperature changes as time
goes by. What do you think will happen? We know from daily experience that heat
61

1.2

0.8

0.6

0.8
0.4

0.6
0.2

0.4
0

0.2

0
0

0.2

0.5

1.5

2.5

0.4
0

3.5

0.5

1.5

2.5

3.5

Figure 26: Left: Our initial conditions. Note the spot at which the bar is hottest. What do
you think will happen to the hot spot as time goes by, i.e. will it cool off, will it move to
the left, to the right, etc.? Right: The initial conditions are a sum of these two terms.

flows from hot to cold, and in fact a quantitative version of that statement was a
major part of our derivation of the heat equation. So probably the hot area to the
left will supply energy to the cooler parts.
Our problem is as follows.
ut = uxx
u(0, t) = 0
u(, t) = 0
u(x, 0) = sin(x) +

1
sin(2x)
3

Before attacking the problem analytically lets just read it carefully to see what
it is about. That may be an obvious suggestion, but it is also a way to prevent
ourselves from doing something idiotic. We see from the first line that we are
solving the case of the heat equation in which the physical constant a has been set
to 1 for convenience. We cant see what the domain is until we read the boundary
conditions. The second two lines say that we are dealing with the ice cube case,
where the bar ends are at 0 and on the x axis and are kept frozen.
Finally the fourth line of the problem says that the initial temperature is the sum of
two terms. Graphed separately, they look like the left part of Figure 26.
If you did problem 7 of Lecture 13, you may suspect why we have included the
graph on the right side of Figure 26. Solutions to the heat equation may be added to
produce new solutions. If you didnt do that problem yet, go do it now. In mathematics this is called linearity, and in physics it is sometimes called superposition
to crystallize the idea that two or more things are going on at the same time and
place.
Now we are ready to solve this problem. There are product solutions to the heat
2
equation of the form u = A sin(wx)eaw t for all values of w. Perhaps the ones
62

we need are the cases of the sine function which appear in our initial conditions.
Lets try that. Our candidates are
u1 (x, t) = A1 sin(x)et
and
u2 (x, t) = A2 sin(2x)e4t
Now check to be sure that these solve the heat equation and the boundary conditions. Do that now. Does it work? Ok. Notice that it doesnt matter what the
coefficients Aj are yet, and that the sum
u(x, t) = A1 sin(x)et + A2 sin(2x)e4t
is also a solution to the heat equation with these frozen boundary conditions.
The next step is to attack the initial conditions. We need to compare
u(x, 0) = A1 sin(x)e0 + A2 sin(2x)e0
from our tentative solution, with the desired initial condition
u(x, 0) = sin(x) +

1
sin(2x)
3

Now the question is, can this be made to work? Well, yes, it is really not hard to
see at this point, since we did all the hard work already. You just take A1 = 1 and
A2 = 31 . This makes the initial conditions work, and we have checked everything
else. So we have a solution to our problem. It is
u(x, t) = sin(x)et +

1
sin(2x)e4t .
3

The next step is to understand what this solution means and how it behaves as time
goes by. The best way to do this is for you to sketch the graphs of the solution for
various times, thinking about what happens to each of the two terms in the solution.
If you do that yourself, and it is a good idea to do so, you should get something
like the left side of Figure 27 for the individual terms, with the solution on the right
side. If you study the solution formula and the pictures you will see how significant
the n2 is in the time exponential part of the product solutions.
P ROBLEMS

63

1.2

0.8

0.6

0.8
0.4

0.6
0.2

0.4
0

0.2

0.2

0.4
0

0.5

1.5

2.5

0
0

3.5

0.5

1.5

2.5

3.5

Figure 27: Left:The two terms in our solution, plotted for t = 0 and t = .3 Notice that the
term involving sin(2x) has decayed a lot more than the sin(x) term. Why is that? Right:
The solution temperature at times t = 0, .3, .6, .9, 1.2. Why does the hot spot move toward
the center?
83. Think about what should happen to a metal bar which has an initial temperature distribution consisting of alternating hot, cold, hot, etc., in a lot of narrow bands along the bar.
Think of a bar which is red-hot in six places and cooler in between. Do you think it will
take very long for the temperature to change? Solve
ut
u(0, t)
u(, t)
u(x, 0)

=
=
=
=

uxx
0
0
sin(12x)

and sketch the solution for t = 0, .3, .6.

84. Solve
ut
u(0, t)
u(, t)
u(x, 0)

= uxx
= 0
= 0
= sin(x) + .5 sin(3x) + .25 sin(5x)

Sketch the initial condition.

85. Solve
ut =
ux (0, t) =
ux (, t) =
u(x, 0) =

uxx
0
0
cos(x) + .5 cos(3x) + .25 cos(5x)

Note that the initial temperature satisfies the insulated boundary conditions.
86. Take u = sin(x)et + sin(3x)e9t and define an average temperature of the bar to be
Z
1
T (t) =
u(x, t) dx.
0
Show that Newtons law of cooling T 0 = T is not quite true.

64

20

10

-10

-20

0.5

1.5

2.5

Figure 28: The one dimensional wave equation can describe vibrations of a string, shown
here at several times.

15

The Wave Equation

T ODAY: Sums of traveling waves

We have now solved several heat conduction problems and are about to turn to another partial differential equation known as the wave equation. Notice that as we
were working on the heat equation we solved a fairly large number of problems
distinguished from one another by boundary and initial conditions, while the equation itself never changed. This is typical of partial differential equations, that there
are relatively few of them which are considered to be important, and the conditions
can be changed to permit these few important equations to apply to many different
settings. The wave equation can apply to many things also, such as the vibrations
of a guitar string, but we are mainly going to talk about waves on an infinitely long
string, the whole x axis. The reason for this choice is that guitar strings can be
solved by methods similar to those we used for the heat equation, and we prefer to
introduce new ideas here.
The wave equation is
utt = c2 uxx
utt = c2 (uxx + uyy )
utt = c2 (uxx + uyy + uzz )
in one, two, or three dimensions, respectively. Note that it is second order with
respect to the time, like Newtons law. In fact it is Newtons law in disguise, sometimes. We know that second order equations can describe oscillations, so that at
least is reassuring. In one dimension u may represent the vertical position of a
vibrating string. In two dimensions or three, the wave equation can represent vibrations of a drum head or of sound in the air, and other things like electromagnetic
waves. It is a sound and light show all by itself.
65

Figure 29: The tension in the string pulls at different angles on the two ends of any little
piece of it, depending on the curvature.

Lets read the one dimensional equation again. If u is position, what is utt ? Think
about it. Did you think about it yet? It is acceleration. Also when you draw the
graph of u as a function of x as in Figure 28 what does uxx represent? Just as
for the heat equation, it is the curvature of the graph. The wave equation says that
when the graph is concave up, the acceleration must be positive. Does that seem
to fit with reality? Think about swinging a heavy jumprope. Better still, get a rope
and try it. See whether you think the acceleration direction matches the curvature
in this way.
The wave equation for a string may be derived as follows. We apply Newtons law
to a bit of the string, accounting for vertical forces.
The tension T in the string has a vertical component. On the left it is approximately
T ux (x, t) since the slope ux is the tangent of the angle, which is near the sine of
the angle if the angle is small. On the right it is T ux (x + x, t). The difference of
these must be the mass times the acceleration, so Newtons law becomes
T ux (x + x, t) T ux (x, t) = xutt
where is the mass per length of the string. Dividing by x and taking the limit
as x approaches 0 gives
T
uxx = utt

which is the wave equation with c2 =

T
.

Now lets find some solutions to the wave equation, using our usual exponential
method. We try
u = f (x)ert
in utt = c2 uxx . It becomes
r2 f (x)ert = c2 f 00 (x)ert .
Cancelling the time exponentials and setting r = ac gives f 00 = a2 f . Again an
exponential trial gives solutions f (x) = eax and f (x) = eax . We have found
66

10
9
8
7
6
5
4
3
2
1
0
2

Figure 30: The wave u(x, t) = e(xct) , graphed for various times. Note that to keep
the expression x ct constant as t increases you have to increase x also, so this wave is
moving to the right. It could make a surfer cry.

solutions
u = eax eact
where a could be any number, even complex. For example if a = 1 we have found
the solutions u = ex ect and u = ex ect . To understand what these are, we graphed
the second one in Figure 30.
These solutions certainly suggest that there are a lot of solutions which are functions of x + ct, or equivalently of x ct. Suppose you have a function
u(x, t) = g(x ct) + h(x + ct)
not necessarily built from exponentials at all. If you have read Lecture 3 you have
seen waves traveling in one direction there. These go left or right. Taking derivatives by the chain rule we find
ut = cg 0 (x ct) + ch0 (x + ct)
utt = c2 g 00 + c2 h00
ux = g 0 (x ct) + h0 (x + ct)
uxx = g 00 + h00
Look what happened there. We got utt = c2 uxx without assuming anything about
g and h except that they be differentiable. Doesnt that mean you can add waves of
nearly any shape moving left and right? Yes it does.
P ROBLEMS
1
are solutions to utt = c2 uxx for some c. Find c.
(x + 2t)2
Graph these waves for t = 0, 1, and 2.
87. Show that cos(x 2t) and

67

88. The function sin(x + t) + sin(x t) consists of 2 traveling waves and solves
utt
u(0, t)
u(, t)
u(x, 0)
ut (x, 0)

=
=
=
=
=

uxx
0
0
2 sin(x)
0

Sketch the two traveling waves, both separately and together, and try to see why the sum
has the 0 boundary condition at x = 0, even though neither of the traveling waves does so
separately.
89. Use the addition formula for the sine to show that the function in problem 88 is equal
1
to 2 sin(x) cos(t). It therefore represents a vibrating string having frequency 2
. Then find
a number k so that 2 sin(2x) cos(kt) solves the problem:
utt
u(0, t)
u(, t)
u(x, 0)
ut (x, 0)

=
=
=
=
=

uxx
0
0
2 sin(2x)
0

Sketch the motion for this case, and convince yourself that this vibration produces a sound
one octave up from that of Problem 88, i.e. has twice the frequency.
Note that 2 sin(3x) cos(3t) sounds a musical fifth above 2 sin(2x) cos(2t).
Also: the product solutions found in Problems 88 and 89 make use of the eigenfunctions
y(x) = sin(nx) that we previously found for the problem
y 00 = y
y(0) =0
y() =0
with eigenvalues = n2 , n = 1, 2, 3, . . ..
90. This problem is about using the wave equation to model an echo. Consider a function
of the form u(x, t) = f (x + ct) + f ((x ct)) where that is the same f traveling in both
directions. Show that u satisfies the condition ux (0, t) = 0. Think of sound waves moving
in 0 x and a wall at the point x = 0. Sketch a graph of u for several times using a
function f of your choice, and try to convince yourself that what you have sketched is a
representation of waves bouncing off a wall.
91. Repeat Problem 90 for u(x, t) = f (x + ct) f ((x ct)). This time you should
get the boundary condition u(0, t) = 0, which involves a different kind of bouncing at
the wall. Try to see it in your sketch. Which of Problems 90 or 91 is more like a string
vibration, with the string tied down at the origin? Which one is more like water waves at
the edge of a swimming pool?

68

Figure 31: A wave traveling initially to the left is shown at six times, t increasing from top
to bottom in the picture. It is reflected downward and reversed left-right when it bounces
off a wall at the left side. Does this fit Problem 90 or Problem 91 better?
92. Whats rong with this? The function u(x, t) = cos(t) sin(x) couldnt be a solution to
the wave equation because it isnt a traveling wave.

16

Application of Power Series: a Drum model

T ODAY: We set up a model for the vibrations of a drum head. That requires
a power series.

Our solutions so far have either been done numerically, or consisted of formulas
involving familiar functions that can be found on any scientific calculator. For
drum vibrations we need a new function, which is known, has a name, and you can
google it, but there is no button for it on your calculator.
The Drum
Our model of a drumhead has radius 1, and it only vibrates in circular symmetry.
That means you can only hit it in the center, so to speak. Using polar coordinates,
there is dependence on r and t but not on . We write
u(r, t)
for the upward displacement of the drumhead out of its horizontal resting plane,
and assume that u is small and that
u(1, t) = 0
since the edges of the head are pulled down against the rim of the drum.
69

r+dr
r

T
r

r+dr

Figure 32: Left: A drum. u(r, t) is the vertical displacement of the drum head from
equilibrium. Center: An annulus of drumhead material viewed from above. Right: A cut
through the annular region, viewed on edge. The vertical displacement is greatly exaggerated. Were accounting for the vertical components of the tension forces.

We will assume that there is a tension T [N/m] uniformly throughout the drumhead. In Problem 93 you can convince yourself that this is reasonable.
Now write Newtons F = ma law for vertical forces on any annular region of
the drumhead. We take the annulus to have radius running from r to r + dr. In
working out the forces due to the tension T , it is best if you have already worked
on the wave equation for string vibrations. The reason is that while the forces here
are quite similar to those acting on the string, the geometry here is a little more
involved because of the curved shape of the annulus.
The length of the inner curved edge of the annulus is 2r. There are forces directed
inward all along that edge. We are interested only in the vertical resultant of those
forces. That is T (2r) times the sine of the small angle with the horizontal plane.
The slope is ur (r, t). The slope ur is nearly the same as the sine of the angle for
small displacements. See Problem 95 if you arent sure about that. Similarly for
the outer edge.
The net force vertically is then
T (2)(r + dr)ur (r + dr, t) T (2r)ur (r, t)
The acceleration of the annulus is utt . Mr. Newton says that we alsoneed its mass.
Themass of the segment is its area times its density [kg/m2 ], or (r + dr)2
r2 By Newton then
T (2)(r + dr)ur (r + dr, t) T (2r)ur (r, t) = (2r dr + (dr)2 )utt
Divide by dr and let dr tend to zero, giving

70

(T rur )
= rutt . So our equation is
r

rutt =

T
(rur )r or equivalently

T
utt =

1
urr + ur
r

We will find some solutions to this having u(1, t) = 0 at the rim of the drum.
The solutions will be vibrations of our drum model, and hopefully will resemble
vibrations of a real drum.
1
coefficient
r
of ur is not defined there. What should the slope ur (0, t) be at the center, so that the
shape of the drumhead is smooth there?
P RACTICE : Polar coordinates are awkward at the origin; for example the

Separation of Variables
Maybe our wave equation has some product solutions of the form u(r, t) = R(r)T (t).
Since this is supposed to represent music, lets try
u(r, t) = R(r) cos(t)
or sine rather than cosine. Set that into the PDE to find

1 0
T
00
2
R + R cos(t)
cos(t)R =

r
We need

1
2
R00 + R0 +
R=0
r
T
For any such function we can find, there will be drum vibrations R(r) cos(t) and
R(r) sin(t)

In the next section we plan to find a nice solution R. However, it will be easier
2
if we dont have to deal with the parameters T all the time. Let a new variable
x = ar and choose a to clean up the R equation. Write R(r) = R( xa ) = f (x). We
have R0 (r) = af 0 (x) and R00 (r) = a2 f 00 (x). The DE for R becomes
a2 f 00 +
q
Of course we set a =

T.

a 0 2
af +
f = 0.
x
T

Our new differential equation becomes

f 00 +

1 0
f +f =0
x
71

(0,b)

T
T

(0,0)

(a,0)
T

Figure 33: A triangle of the drumhead. See Problem 93.

P ROBLEMS
93. Imagine the tension forces on a small triangle of material located anywhere in the
drumhead, as in Figure 33. Suppose a triangle with vertices (0, 0), (a, 0), and (0, b) has
a force applied uniformly along each edge, perpendicular to the edge, in the plane of the
triangle, with magnitude T times that edgelength. Show that the triangle is in equilibrium,
i.e. that the sum of forces on it is 0, assuming T is a constant.
94. Verify that none of the functions
erx ,

cos x,

sin x,

ax2 + bx + c

1 0
f + f = 0.
x
95. Suppose a very acute triangle has vertices (0, 0), (1, 0), (1, m) so that the slope is m.
solve the equation f 00 +

(1,m)

(1,0)

(0,0)

Convince yourself that m is the tangent of the acute angle exactly, and the sine of the acute
angle is nearly equal to m. You could fiddle with the geometry, or try your calculator, or
try the power series for the sine and cosine, or think of something else to try.
That is why we feel it is justified to replace the sine of the angles by the ur values in our
derivation. Of course, if it turns out that the model doesnt act like a real drum then we
have to rethink this decision.

72

96. If you are not clear about the limit we found, try this. First tell what this limit is:
lim

h0

Then how about

lim

h0

f (x + h) f (x)
h

(x + h)g(x + h) xg(x)
?
h

Finally the one we need

lim

dr0

(r + dr)ur (r + dr, t) rur (r, t)

dr

16.1 A new Function for the Drum model, J0

We learn a new function which might not be on your calculator, and use it to describe vibrations of the drum head.
So far we have derived a wave equation

T
1
utt =
urr + ur

r
for the displacement of our drumhead, and we know that we need to solve f 00 +
1 0
x f + f = 0. Having such a function f , we will get many drum vibrations
r
u(r, t) = f (

r) cos(t)
T

A new function, J0
Try a power series
f (x) = f0 + f2 x2 + f4 x4 +
If you did the practice item on page 71, you know that we dont want a term f1 x.
We are guessing that we dont need any of the odd degree terms. To keep things as
clean as possible lets also take f0 = 1. We can rescale things later as needed. We
have

1
f 00 + f 0 + f = 2f2 + 4 3f4 x2 + 6 5f6 x4 + 8 7f8 x6 +
x

+ 2f2 + 4f4 x2 + 6f6 x4 + 8f8 x6 + + 1 + f2 x2 + f4 x4 + f6 x6 +

= (4f2 + 1) + (4 4f4 + f2 )x2 + (6 6f6 + f4 )x4 + (8 8f8 + f6 )x6 +
73

0.8

0.6

0.4

0.2

-0.2

-0.4
0

10

15

20

25

Figure 34: Graph of Bessell function J0 (x) = 1

30

35

40

x2
x4
x6
+

+
22
22 42
22 42 62

For that to be identically 0, we need all the coefficients to be 0, so

f2 =

1
,
22

f4 =

f2
1
= 2 2,
2
4
2 4

f6 =

f4
1
= 2 2 2,
2
6
2 4 6

f8 =

1
22 42 62 82

etc. So

x2
x4
x6
x8
+

+

22
22 42 22 42 62 22 42 62 82
Isnt that a beautiful series?
f (x) = 1

This function is sufficiently useful that it has a name, even though you wont find it
on every calculator. It is called the Bessell function J0 . So f (x) = J0 (x) or, if we
need a multiple of that, f (x) = f0 J0 (x) for some constant f0 . We need to know
what this function is like, to see whether it gives us plausible drumhead vibrations
r

r) cos(t)
u(r, t) = f0 J0 (
T
The series tells us some of the properties that we need, and the differential equation
J000 +

1 0
J + J0 = 0
x 0

itself gives other properties.

First the series: The series is reminiscent of that for the cosine function, so maybe
J0 has many zeros, and oscillates, and is periodic. Well, 2 out of 3 isnt bad: J0
has infinitely many zeros and oscillates somewhat like the cosine function, but is
not periodic. In a homework problem you will use the series to find that J0 (4) is
negative. Since we know J0 (0) = 1, there must be a number x1 between 0 and 4
where J0 (x1 ) = 0. In fact the first root x1 is roughly 2.5 and the second root x2 is
near 5.5, x3 is about 8.5, and there are infinitely many others.
74

From the differential equation we see this: suppose you have an interval like
[x1 , x2 ] where J0 is 0 at each end, and nonzero between. At any local min or
max where J00 = 0, the DE then tells us that J000 = J0 there. The graph cant
be concave up at a local max, nor concave down at a minimum. So the graph can
only have simple humps like the cosine does, no complicated zig-zags between the
roots.
After analyzing J0 like this for some time, people eventually published tables of
its values, and many years later someone built it into matlab and octave so that
we could look at the graph in Figure 34. Isnt that nice?

16.2

But what does the drum Sound like?

We dont know yet. We dont know what sounds the drum can make.
We have two boundary conditions. The first is that ur (0, t) = 0, so that the drum
shape is not pointy in the center. This is automatic since it is a property of J0 .
P RACTICE : How do you know from the series
J0 (x) = 1

x2
+
4

that J00 (0) = 0?

The second boundary condition is that at the rim we have

r

u(1, t) = f0 J0 (
1) cos(t) = 0
T
To acheive this, we need

= x1
T

or x2 or some root of J0 . Then the lowest sound we hear will correspond to the
case
s
T
u(r, t) = f0 J0 (x1 r) cos(
x1 t)

The ordinary cosine function has period 2, so this cosine will go through one
cycle as
s
T
x1 t

75

goes from 0 to 2. One cycle in every

q2
T
x
1

q
seconds. The frequency is

T x1
2 .

For larger roots x2 etc you get shorter cycles, higher pitched sounds.
P RACTICE : When you tighten up the drum tension T , do the frequencies go up or
down? Also the drummer might change to a heavier material so that increases.
What does that do to the frequencies?

E XAMPLE : Take = T for simplicity. We have a vibration

u1 (r, t) = J0 (x1 r) cos(x1 t)
where x1 is approximately 2.5. This is a solution to the boundary value
problem
1
urr + ur
r
u(1, t) = 0
ur (0, t) = 0
utt

(0 < r < 1)

having initial position u1 (r, 0) = J0 (x1 r). Another solution is

u2 (r, t) = J0 (x2 r) sin(x2 t)
where x2 is approximately 5.5. This has a different initial shape, and a different frequency, and a different initial velocity.

P RACTICE : What is the initial velocity of u1 ?

Any linear combination of these various solutions, for different roots xn , and using
both cosines and sines, is a solution to our wave equation.
P RACTICE : Check a specific case, that
u(r, t) = 3J0 (x1 r) cos(x1 t) + 2J0 (x3 r) sin(x3 t)
is a solution to utt = urr + 1r ur with u(1, t) = 0. [It saves a lot of work if you
remember that J000 (x) + x1 J00 (x) + J0 (x) = 0.]

So our model predicts what sounds the drum will make. Good.
76

P ROBLEMS
97. In this problem you can confirm part of what is shown in the graph of J0 , that J0 (4) < 0
so there must be a root x1 somewhere in the interval (0, 4). Write the partial sum of J0 (x)
through sixth powers as s6 (x)
s6 (x) = 1

x2
x4
x6
+ 2 2 2 2 2.
2
2
2 4
2 4 6

First check that s6 (4) is negative. It saves a little work if you notice that the second and
third terms cancel.
98. In Problem 97 the tail of the series is so small that J0 (4) itself is also negative. That
means,
J0 (4) = s6 (4) + E
where E is not very big and the sum is negative. Try to estimate E by observing that the
tail is an alternating series, if you remember those from calculus.
99. Make a sketch of the graphs of J0 (x1 r) and J0 (x2 r) for 0 r 1, so that you can
see what some of the drum waves look like. You can figure these out from Figure 34 by
rescaling.
100. For a particular real drum, it might be feasible to measure if you could weigh
a sample of the drumhead material, but T could present more problems. Suppose you
compare the sound of the drum to a piano or guitar and decide that the fundamental tone is
near the note A2 = 110 [cycles/sec], the next to lowest string on a guitar. Figure out T for
this drum.
101. Remind yourself that the first three tones of a string have frequencies in the proportion
1 : 2 : 3, i.e., there are solutions cos(t) sin(x), cos(2t) sin(2x), and cos(3t) sin(3x) to the
wave equation ytt = yxx . But what about a drum? Check the graph of J0 and find out
approximately how the lowest three frequencies of a drum are related. This is why pianos
and drums dont sound the same.

17

The Euler equation for Fluid Flow, and Acoustic Waves

T ODAY: A model for compressible air flow, including sound. This example
is a little more advanced than the rest of the lectures, so there are several
practice items here to guide your reading.

The purpose of this lecture is to derive three PDEs which describe compressible
flow of air in one dimension, say in a pipe. You might imagine pumping air into a
bicycle tire through a hose, for example. The main assumptions are:
There is no conduction of heat through the air. There is no friction of the air
with the pipe, or any exchange of heat between the air and the pipe.
77

1
0
0
1
0
1
0
1
0
1
0
1
0
1
0
1

1
0
0
1
0
1
0
1
0
1
0
1
0
1
0
1

W
L

Figure 35: A portion of air moving in the pipe. The pipe has cross-sectional area A. The
ends L and R dont have to move at the same speed because this is compressible flow.

You know about Newtons law F = ma, and about the idea of conservation
of mass. And of course you know calculus.
Two of the three equations are known as the Euler equations for isentropic flow.
[You do not have to know what isentropic means, but if you have read Lecture 7
you will have touched the idea.] Then having the Euler equations, we can further
derive that certain approximate solutions are solutions to the wave equation. These
solutions are very familiarthey are sound waves. Thus we also get, almost free,
another derivation of the wave equation.
We have already derived the wave equation for vibrations of a string, and indicated
a derivation for light waves (in Problem 10). Thus you see that there are three very
different situations giving rise to the same mathematics.
The method used in this derivation is rather common in science. There are many
assumptions behind every statement. So, to understand the derivation, you have to
ask yourself at every step, Why should that be true? What is being assumed? This
careful questioning can lead to good insights.
The Euler equations
The pipe is full of air. The air velocity is u(x, t) [m/sec], and density (x, t)
[kg/m3 ]. For air, the density is typically around 1.2 near the surface of the earth,
but the velocity could vary over a large range. We can think of x and u positive
toward the right.
We focus on a portion W of the air. The letters L and R in the figure refer to the
left and right ends of the mass W . W is mathematically just a moving interval.
The left coordinate L is moving at velocity uL , and to be more precise,
L0 (t) = u(L(t), t)

and uL is an abbreviation for this

and similarly for the right end R at velocity uR .

Besides velocity u(x, t) and density (x, t), we must keep track of the air pressure
p(x, t) [N/m2 ]. It is very easy to mistake and p typographically, but this wont
78

happen to you because you are reading very slowly and thinking hard, right? The
pressure provides forces pL A and pR A on the left and right ends of W . The isentropic assumption is that the pressure is related to the density by p = a , where
= 1.4 and a is constant throughout the flow. [If you have worked Problem 36
on page 24 you will have seen this related to the no heat added assumption.]
Aside from this thermodynamic assumption, the rest of our derivation will consist
of ordinary mechanics.
Newton: The rate of change of momentum of the matter in W is equal to the sum
of applied forces.
Z
d R
uA dx = pR A + pL A
dt L
d
P RACTICE : Is this really F = ma or is it more like dt
(mv) = F ? When does the
m in Newtons law have to be included inside the derivative? Of the quantities and
R
expressions , u, A, and dx, which constitute v and which m? Why do we multiply
pressure by area? Why is there a minus sign on pR A?

In the left side of Newtons law we need to remember that

Z
Z f (t)
d f (t)
h(x, t) dx =
ht (x, t) dx + h(f (t), t)f 0 (t) h(g(t), t)g 0 (t)
dt g(t)
g(t)

P RACTICE : Derive that from the Fundamental Theorem of Calculus and the Chain
Rule.

So, if we cancel all the As, Newtons law and the isentropic assumption give
Z

(u)t dx + (u)R uR (u)L uL = pR + pL = aR + aL

Now we may be able to study a limit of this for very short intervals. If it is possible
to divide by R L and take L R, then we get
(u)t + (u2 + a )x = 0
This is a partial differential equation which of course assumes the various functions
are differentiable. [We assume this. But it is important to know that the limit has
been found not in agreement with all experiments. We wont go into shock waves
at all, but you should know that there are such possibilities.]
79

P RACTICE : Why does

fR fL
RL
tend to fx as R and L tend to x? Assume f differentiable.

We notice that we have two unknown functions, u and . Probably one equation
isnt going to be enough, so we look for another equation. Turning to conservation
Z R
of mass, we figure the mass in W to be
A dx. This doesnt change as the
L

fluid moves because our set W is moving with the fluid. We have conservation of
mass
Z
d R
A dx = 0
dt L
Doing the same transformations as we did for Newton gives
Z

t A dx + (u)R A (u)L A = 0

and in the limit

t + (u)x = 0
P RACTICE : Use t + (u)x = 0 in the Newton PDE to find
ut + uux = a2 x

These two equations are the results of our work, the one-dimensional isentropic
Euler equations,
t + (u)x =0
ut + uux = a2 x

We can identify a physical meaning for the combination of terms ut + uux which
occurs in the equation of motion. In Problem 102 you can show that it is the
acceleration of the particle which is passing through the point x at time t.
P RACTICE : Can there be any acceleration at a point where ut = 0?

80

17.1 Sound
Here we start from the Euler equations, and imagine a small disturbance superimposed over an ambient stillness.
P RACTICE : Is it true that the ambient stillness u = 0, = (constant) is a solution to
the Euler equations?

We then look for approximate solutions of the form

u(x, t) =v(x, t),
(x, t) =1 + w(x, t)

where is a small number. The constant 1 could be taken to be a typical air density
at sea level on earth.
P RACTICE : Substitute our assumed u and into the Euler equations and ignore
everything containing 2 or higher powers. Show that you find for the 1 terms,
vt = a2
wx
1
wt + 1 vx = 0

Combining the two PDEs in that practice problem, we have

wtt = 1 vxt = a11 wxx ,
or
wtt = c2 wxx
where c2 = a1
1 . So the pressure disturbance satisfies the wave equation, with
c the sound speed.
It is said that Newton himself underestimated the speed of sound because he assumed (who wouldnt?) that the temperature doesnt change when tiny sound
waves pass by. But it is the entropy which does not change.
P ROBLEMS
102. Suppose a particle of fluid has position given by a function x(t). Then we have two
expressions for the velocity of this particle, x0 (t), and u(x(t), t). These must be equal.
Differentiate to show that the composition (ut + uux )(x(t), t) is equal to the acceleration
of the particle.

81

103. Air pressure at sea level on the earth is about 105 [N/m2 ]. Using this and previous
information, estimate the speed of sound. Remember that traveling waves f (x ct) are
solutions to the wave equation. If someone claps his hands at a distance of 3 football fields
from you, how long before you hear it?

18

The Laplace Equation

T ODAY: Two roads lead here. The heat conduction and drum vibration models both morph into the Laplace equation under certain conditions. Many
roads lead away. The one we take will be the Fourier road.

The Laplace equation looks like this in two dimensions

uxx + uyy = 0
or in one u00 = 0, or in three uxx + uyy + uzz = 0. In two dimensions using polar
coordinates not at the origin it looks like
1
1
urr + ur + 2 u = 0.
r
r
Solutions are called harmonic functions. In every case it says that the divergence
of the gradient of u is 0, div(u) = 0. It is abbreviated u = 0 by mathematicians
and 2 u = 0 by engineers. But what does it mean? We approach that from two
directions, as follows.
I. The Laplace equation can apply to a steady state temperature. We know from
Lecture 13 that the temperature in a region is modeled by the heat equation ut =
u, which keeps track of the conduction of heat energy from hot places to cold
places. If the temperature is not changing with time, that gives the Laplace equation. It doesnt mean there is no energy flow. A refrigerator can maintain a constant
temperature but only by forcing out the energy that seeps in through the walls. We
will consider the polar coordinate case in a disk with the temperature u specified
on the boundary circle.
II. Our drum discussion in Lecture 16 can be modified. There we had a dynamic
situation urr + 1r ur = utt with circular symmetry. If you instead analyze a circular
membrane stretched statically over a varying rim height, you find for small angles
that the same Laplace equation urr + 1r ur + r12 u = 0 applies to the displacement
u. What is added is the dependence including nonzero boundary conditions, and
what is removed is the time dependence.
82

Figure 36: A membrane stretched, or the graph of a steady temperature.

So we are solving the boundary value problem

1
1
urr + ur + 2 u = 0
r
r

(0 < r < 1, 0 2)

with a specified value at the boundary circle

u(1, ) = f ().
P RACTICE : What can we try? We have traveling waves, product solutions, power
series. Anything else?

Lets look for product solutions u(r, ) = R(r)F (). You get
1
1
R00 F + R0 F + 2 RF 00 = 0.
r
r
We need to divide by F at least:
1
1 F 00
R00 + R0 + 2 R
= 0.
r
r
F
That is not quite separated. Multiply by

r2
R:

r2 R00 + rR0 F 00
+
= 0.
R
F
Now what do we have? The R part doesnt look at all familiar. How about the
F ? Well, there is no doubt about that. We need F 00 to be a constant multiple
of F , and we know that equation very well by now. The possibilities for F ()
are cos(a), sin(a), ea , ea , and a + b, for various numbers a, b, or certain
linear combinations of those. Since we want u continuous in if possible, it has to
connect smoothly when loops around to 2 which is the same angle as 0. So we
better not use a + b unless a = 0. A constant is ok, say b = 1 for definiteness.
83

Also any of the sin(n) or cos(n) for n 1 would be alright. So far our list of
acceptable functions F is
1, cos(), sin(), cos(2), sin(2), cos(3), sin(3),
and these will be enough for our purposes. Well abbreviate the list as
cos(n),
Then

sin(n),

(n 0).

F 00
= n2 .
F

P RACTICE : 1. Convince yourself of that n2 .

2. Would u(r, ) = sin( 14 ) be continuous on the x axis if we are thinking of running
from 0 to 2?

With those choices for F , the requirement on R is that

r2 R00 + rR0 n2 R = 0

r2 R00 +rR0
R

n2 = 0, or

(n 0)

Now what? This is a second order linear differential equation, but look at the coefficients. They are not constants. So we cant use the characteristic equation method;
that is based on trying exponentials and it wont work with variable coefficients.
P RACTICE : In case that isnt clear, substitute R(r) = ekr to see why it doesnt work.

We can still try something else. We havent tried a power series for a while. As it
happens, the solution for this is the easiest power series youve ever tried. It is left
for the Problems. In the next section well continue with u.
P ROBLEMS
104. Having no other good ideas about the unfamiliar R equation r2 R00 + rR0 n2 R = 0,
we could try a power series
R(r) = a0 + a1 r + a2 r2 +
But wait. It works out so nicely I would like to save you some work. Try just
R(r) = rq
If that doesnt work, you can always try something harder. You will like this one.
105. Combine your solution for R in Problem 104 with its corresponding F () that we
found, and check it in the Laplace equation to make sure that you really do have a solution.
Try it for a couple of different values of n.

84

18.1 Laplace leads to Fourier

We are solving the boundary value problem
1
1
urr + ur + 2 u = 0
r
r

(0 < r < 1, 0 2)

with a specified value at the boundary circle

u(1, ) = f ()
and so far you have found (you did Problem 105, no?) a list of solutions to the
Laplace equation:
rn cos(n),

rn sin(n),

(n > 0)

and also 1.

Now we have to think about the boundary conditions.

P RACTICE : What does the boundary look like for the case u(r, ) = r sin()? What
is r sin() anyway?

E XAMPLE : Suppose the boundary condition is that

u(1, ) = f () = 6 cos(3)
Of our list of solutions to Laplace, the one which stands out is
r3 cos(3)
It is almost what we need. Would it be alright to multiply it by 6? (yes, see
Problem 106.) So one solution to the BVP is
u(r, ) = 6r3 cos(3)

Some people call the graph in Figure 36 a saddle, and this r3 cos(3) a monkey
saddle. Can you see why?
E XAMPLE : If the boundary condition is now
u(1, ) = f () = 6 cos(3) + 3 sin(6)
what will the solution be? We can try a linear combination
u(r, ) = c1 r3 cos(3) + c2 r6 sin(6)

85

At least that will be a solution to the Laplace equation (see Problem 108).
Then we have to choose the coefficients ck if possible to get the boundary
values. It will work to take c1 = 6 and c2 = 3. Answer:
u(r, ) = 6r3 cos(3) + 3r6 sin(6)
Look, you just throw the rn factors in. After all this buildup, does that seem
strange?

Fourier
M. Fourier observed the examples weve seen. He felt that something is missing.
Of course you probably feel too that our boundary conditions are somewhat artificial, made to fit the product solutions we found. Fourier wondered if you could
work with some such boundary conditions as
u(1, ) = f () = 2
which has nothing to do with the cosines and sines. It is an interesting question
to ask. Could it be possible that you could somehow expand f in terms of cosines
and sines? Could a function f () which is not expressed in terms of the cos(n)
and sin(n) actually have those hidden within it?
The question is more strange too, because it is nearly the opposite of what we are
used to, like
4
2
+

cos() = 1
2
24
That goes the wrong waywe wanted 2 in terms of trig functions. Eventually we
will achieve this on page 92. But if the answer to the question is yes, then you can
solve the membrane problem by just putting in the powers of r as we did above:
If f () = c0 + a1 cos() + b1 sin() + a2 cos(2) + b2 sin(2) +
then you can solve the boundary value problem as follows, assuming convergence:

u(r, ) = c0 + r a1 cos() + b1 sin() + r2 a2 cos(2) + b2 sin(2) +

Isnt that remarkable?
P ROBLEMS

86

106. We need to check carefully that when u is a solution to the Laplace equation, then
so is 6u, or cu for any constant c. How can that be checked? Start with the first order
derivative
u
r
cu
u
= c ? Then how about the second derivatives, say (cu) ? Is that the
Is it true that
r
r
same as c(u )? That is the essential idea behind the fact that
2 (cu) = c2 u
Why does that prove what we need?
107. We also need to know that a sum of two harmonic functions is harmonic. Reason that
out similarly to Problem 106.
108. We also need to understand that any linear combination of harmonic functions is
harmonic. Part of the calculation goes like this.
2 (c1 u1 + c2 u2 ) = 2 (c1 u1 ) + 2 (c2 u2 ) = c1 2 (u1 ) + c2 2 (u2 )
Which part of that is by Problem 106? 107? Why does that prove what we need?

18.2

Fouriers Dilemma

We now know that it is easy to solve the boundary value problem

urr +

1
1
ur + 2 u = 0
r
r

(0 < r < 1, 0 2)

with a specified value at the boundary circle u(1, ) = f () provided that we can
express f () in terms of a series made from our list
cos(n),

sin(n),

(n > 0)

and also 1.

Fourier wanted to know how to extract the frequencies hidden in f .

P RACTICE : Try to estimate the coefficient c visually if the top function graphed in
Figure 37 is
cos() + c cos(6).

You see from the practice problem that you can sometimes visually estimate the
hidden information if it is not too deeply hidden. Fourier wanted to dig further.
Well dig all the way to the bottom in the next section. For now we take special
identities that you know about.
87

10

15

20

10

15

20

10

15

20

10

15

20

-1

-1

-1
1

-1

Figure 37: Top: What frequencies could be hidden in this function? Second: How about
now? The function has been modified a little to make some features stand out. Compare
to the last two. Third: The graph of cos(). Last: The graph of cos(6).

88

E XAMPLE : 1. You probably remember cos2 () = 12 21 cos(2).

2. From the complex exponential identity e3i = (ei )3 we can derive that
cos3 () =

3
1
cos() + cos(3)
4
4

So there are at least two functions not originally given as a linear combination
of the cos(n) and sin(n), for which we have a Fourier series. There are
just two terms in each.

We can think of the numbers 1 and 3 as the hidden (circular) frequencies and the
coefficients 34 and 14 are hidden amplitudes in the function cos3 (). We got those
examples sort of by luck, by trig identities. In the next lecture we will find out a
systematic method for getting this information.
P ROBLEMS
109. Suggest a reason why c in the practice problem must be positive.
110. Sketch a graph of the cos3 () and sketch the two terms in its Fourier series separately
to see how they fit together. A linear combination of sinusoids is not usually a sinusoid.

18.3

Fourier answered by Orthogonality

We have seen only two cases where a function which was not originally expressed
as a linear combination of sinusoids in fact is one. One example was
cos3 () =

1
3
cos() + cos(3)
4
4

But we also know there are good reasons to find these hidden frequencies if they
exist, to solve the Laplace equation for example. If there were only two examples,
we wouldnt bother.
After Fouriers time a concept was developed to extract the hidden information
from other functions. It is called orthogonality of functions. It works like this.
People knew that
Z

cos() sin() d = 0

because an antiderivative is

1
2

sin2 ().

P RACTICE : Check that.

And there are a lot of other similar integrals which are 0.

89

E XAMPLE : You dont really need an antiderivative, because cos() sin()

is an odd function: the graph has the kind of symmetry where each positive
part to the right of 0 is balanced by a negative part to the left of 0, and vice
versa. So it must integrate to 0. By the same argument,
Z
cos(n) sin(m) d = 0

for all n and m.

Functions f and g are called orthogonal on [, ] if

Z
f ()g() d = 0.

You can compare this to the dot product of vectors:

a1~ + a2~ + a3~k and b1~ + b2~ + b3~k are orthogonal if
a1 b1 + a2 b2 + a3 b3 = 0
The integral is a sum, or at least a limit of sums, and the values of the functions
play the role of coordinates. This was a very new idea, to use geometric language
to discuss functions.
We saw above that each of the cos(n) is orthogonal to all the sin(m). It is also
true (see Problem 116) that any two of the cosines are orthogonal to each other.
Also any two of the sines are orthogonal to each other, and all those trig functions
are orthogonal to the constant function 1. So there is a long list of functions, any
two orthogonal,
1, cos(), sin(), cos(2), sin(2), cos(3), sin(3), cos(4), sin(4),
sort of like ~, ~, ~k in 3-dimensional space. We have a lot more than 3 orthogonal
functions though, since in effect we are now working in an infinite number of
dimensions. This is a new concept, part of our new language.
But what else does that get you? It gets you everything. Why: Suppose it is
possible to write a function f as
f () = c0 + a1 cos() + b1 sin() + a2 cos(2) + b2 sin(2) +
but that we do not yet know what the hidden coefficients c0 , an , bn are. Here is a
way to find them. Think about the vector case, where a vector ~v can be written
~v = a1~ + a2~ + a3~k
90

but somehow we do not know what the coefficients are yet. Dot with ~, for example.
You find the number a2 by doing a dot product:
~v ~ = a1~ ~ + a2~ ~ + a3~k ~ = 0 + a2 + 0 = a2
by the orthogonality and the fact that ~ is a unit vector. So we can try the same
thing with functions, for example
Z

f () cos(2) d =

c0 cos(2) d +

a1 cos() cos(2) d +

a2 cos(2) cos(2) d+

b1 sin() cos(2) d

b2 sin(2) cos(2) d+

= 0 + 0 + 0 + a2

a3 cos(3) cos(2) d+

cos(2) cos(2) d + 0 + 0 +

R
We still need to know that remaining integral cos(2) cos(2) d. Its value is
:
Z
Z

1
1 + cos(4) d;
cos2 (2) d =
2

the 1 integrates to and the cos(4) integrates to zero: it is orthogonal to 1.

These ideas lead to the Fourier coefficient formulas
Z
1
an =
f () cos(n) d,

Z
1
bn =
f () sin(n) d,

and

1
c0 =
2

f () d

Of course we only derived the a2 case but the ideas work for all. We have shown
that if f has a Fourier series then these must be the coefficients. It is a uniqueness
argument. It is for another course to study which functions actually have Fourier
series.
P ROBLEMS
111. Take a short Fourier series such as
f () = cos() + 0.3 sin(5).
Sketch the graph of the two terms and of f to see how they fit together and how the sum
no longer has a graph like that of a trig function.

91

112. Same as Problem 111 for

f () = cos(9) + cos(10)
This one is hard to do by hand, but is interesting because you get to see an example of
amplitude modulation, AM. Compare Figure 13 on page 31.
113. Work out the Fourier coefficient integrals for the function f () = 2 . Youll need to
integrate by parts twice. The result ought to be the series
2 =

4
4
4
4
2
4 cos() + 2 cos(2) 2 cos(3) + 2 cos(4) 2 cos(5) +
3
2
3
4
5

It is for another course to discuss the convergence of Fourier series, but this series converges to 2 for . Of course this series cannot be correct for all values of
because the series is periodic, and 2 is not. This series answers Fouriers question on
page 86.
114. Use the series in Problem 113 to solve the problem
u =0
u(1, ) =2

in the unit disc, where .

115. [Harder] Show that the drum vibrations
u1 (r, t) = J0 (x1 r) cos(x1 t),

u2 (r, t) = J0 (x2 r) cos(x2 t)

are orthogonal on the drumhead whenever x1 and x2 are two roots of J(x) = 0. That
means
ZZ
u1 (r, t)u2 (r, t) rdrd = 0.
unitdisk

116. Use the complex exponential eit = cos t + i sin t in the form cos t =
to show that
Z

1
2

it

e + eit

cos(3) cos(4) d = 0.

The point is that exponentials are easy to integrate. Argue that for the same reasons any
two of sin(n), sin(m) are orthogonal.

19

Application to the weather?

T ODAY: Nonlinear systems in space. Chaos. Why we cant predict the
weather.

92

20

15

10

-5

-10

-15

-20
0

10

20

30

40

50

Figure 38: A chaotic motion.

There are systems for which the study of linear equations does not prepare you.
Notice in Figure 38 that the function plotted there is more complicated than those
in Figures 20 (page 47) and 37 (page 88); in a sense we cant discern what all the
frequencies are. This lecture is only the briefest kind of introduction to this subject.
We begin with a linear example in which your knowledge of linear systems does
help you understand a 3D system. Consider the system
x0 = y
y0 = x
z0 = z
These equations may be read as follows. The equation for z does not involve x
or y, and vice versa. In fact the z equation is a simple equation for exponential
decay. The x and y equations are a system corresponding to our favorite second
order equation x00 = x whose solutions are sines and cosines. Consequently we
can think of this system as having a circular motion in the (x, y) plane combined
with an exponential decay in the z direction. So the solutions are curves in 3-space
moving around circular cylinders centered along the z axis, and which approach
the (x, y) plane as time goes by. Any solution already in that plane remains there,
circling.
In contrast with that system let us look at the Lorentz system
x0 = 10(y x)
y 0 = xz + 28x y
z 0 = xy 83 z
See Figure 39. This very interesting system was made by Edward Lorentz who
was a meteorologist at MIT, as a very simplified model of circulation in the atmosphere. The system is nonlinear because of the xz and xy terms, and is derived
93

45

40

35

30

25

20

15

10

5
20

15

10

10

15

20

Figure 39: The Lorentz system. A solution curve x(t), y(t), z(t) in 3 dimensions is here
projected onto a plane. How do you suppose it occured to Lorentz, to think of the effect of
a butterfly on the weather?

from much more complicated equations used in fluid mechanics. You may be interested to know that the weather predictions we all hear on the radio and TV are
to some extent derived from calculations based on observational data and the real
fluid mechanics equations.
There seem to be two critical points about which the solutions circulate. One finds
that the solution turns a few times around the left one, then moves over to the right
one for a while, then back, and so on. The only problem is that it is not easy to
predict how many times it will circle each side before moving to the other! Lorentz
came upon this fact as follows. He was approximating solutions to this system on a
computer in 1963, and noticed the pattern of randomlooking jumps from one side
to the other. These were very interesting solutions, so he reran the computation
from a point part way along, to study it in more detail. In the process, he typed
in as initial conditions the numbers which had been printed out by his program.
These numbers were slightly rounded by comparison with the precision to which
they had been computed. The result made history: the qualitative features were
still there, but the solution in detail was very different, in that after a few turns the
numbers of cycles around the left and right sides came out different, just because
of rounding off a few decimal places in the initial conditions. That means that if
you try to make a picture like the one above, it will look qualitatively the same,
but may be different in detail. It also means something about weather prediction.
Lorentz asked, Does the Flap of a Butterflys Wings in Brazil Set Off a Tornado in
Texas? This is a restatement of the effect of rounding off, i.e. slightly changing,
the initial conditions. Everybody has known for years that weather prediction is
an art, that you really cant predict the weather very well. Part of what Lorentz

94

did was to explain why you cant predict the weather: you can never know enough
about the initial conditions no matter how many weather stations you build, when
the equations of motion are so sensitive to initial conditions.
We also have a shift in the way the universe is conceived, as a result of studying
this and other chaotic systems. On the one hand there is the uniqueness theorem
which says that the future is determined uniquely by the initial conditions. This is
the Newtonian approach. On the other hand we now know about chaotic systems in
which yes, solutions might be uniquely determined by their initial conditions, but
they may be so sensitive to these conditions that as a practical matter we cannot use
the predictions very far into the future. Isnt that interesting? It is a really major
idea.
P ROBLEMS
117. Check out the java applet de which is mentioned on page 16 and run some chaotic
systems on it. Several are built in as examples.
118. Solve a 3D system in octave. Set up a function file to compute your vector field, like
% file ef.m
function wdot = ef(w,t)
wdot(1)=10*(w(2)-w(1));
wdot(2)=-w(1)*w(3)+28*w(1)-w(2);
wdot(3)=w(1)*w(2)-8*w(3)/3;
Then give commands like
octave:1> t = linspace(0,50,1000);
octave:2> w = lsode("ef",[5 7 9],t);
octave:3> plot(w(:,1),w(:,3))
This will compute the solution to the Lorentz system for 0 t 50 with initial conditions
(5, 7, 9), and make a plot. The symbol w was used here for (x, y, z), so what was plotted
was x versus z. (w(:, 1) means x.) Naturally you can play with the plot command to get
different views. Change these files to solve a system of your choice. Note that you are not
restricted to just three dimensions either.
119. Change the plot command in Problem 118 to plot x versus t. This is the method by
which Figure 38 was made.
120. Change the ef.m file of Problem 118 to solve the linear system described in the text.
You will have to play with the plot command to get a nice perspective view, for example
you can plot x versus y + 2z or something like that. Does the picture fit the description
given in the text?

95

Index
bank, 1

response curve, 41

chain rule, 10
characteristic equation, 33
complex
exponential, 36
number, 34

separate variables
Laplace equation, 83
ode, 3, 12, 14
slope field, 2
octave program, 18

differential equation, 1
partial, 5
reading, 1

wave equation, 65

eigenvalue, 48
first order linear, 19
fundamental theorem
algebra, 38
calculus, 8
differential equations, 15
gravity, 6
heat equation, 50
ideal gas, 22
initial condition, 4
integrating factor, 20
Laplace equation, 82
linear combination, 29
logistic equation, 11
Maxwell, 6
octave, 16
oscillation
forced, 39
mode, 46
pizza, 5
96