" D O Y O U K N O W , " T H E D E V I L C O N F I D E D , " N O T E V E N T H E B E S T M AT H E M AT I -

C I A N S O N O T H E R P L A N E T S - A L L F A R A H E A D O F Y O U R S - H A V E S O LV E D I T ?

W H Y, T H E R E I S A C H A P O N S AT U R N - H E L O O K S S O M E T H I N G L I K E A M U S H -

R O O M O N S T I LT S - W H O S O LV E S PA R T I A L D I F F E R E N T I A L E Q U AT I O N S M E N -

TA L LY ; A N D E V E N H E ’ S G I V E N U P. "

ARTHUR PORGES, “THE DEVIL AND SIMON FLAGG”

I N O R D E R T O S O LV E T H I S D I F F E R E N T I A L E Q U AT I O N Y O U L O O K AT I T T I L L A

S O L U T I O N O C C U R S T O YO U .

G E O R G E P Ó LYA
JARED C. BRONSKI ALDO J. MANFROI

DIFFERENTIAL
E Q U AT I O N S

U N I V E R S I T Y O F I L L I N O I S M AT H E M AT I C S
Copyright © 2023 Jared C. Bronski Aldo J. Manfroi

published by university of illinois mathematics

tufte-latex.googlecode.com

Licensed under the Apache License, Version 2.0 (the “License”); you may not use this file except in com-
pliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/
LICENSE-2.0. Unless required by applicable law or agreed to in writing, software distributed under the
License is distributed on an “as is” basis, without warranties or conditions of any kind, either
express or implied. See the License for the specific language governing permissions and limitations under
the License.

First printing, November 2023

 Contents

I Ordinary Differential Equations 17

1 First Order Differential Equations 19

1.1 What is a Differential Equation? 19
1.1.1 Basic terminology 19
1.1.2 Examples of differential equations 21
1.2 Basic solution techniques. 24
1.3 Slope fields for first order equations 29
1.4 Existence-Uniqueness theorem 31
1.5 First Order Linear Inhomogeneous Equations: 35
1.6 Exact and other special first order equations 40
1.7 Autonomous equations, equilibria, and the phase line. 48
1.8 Numerical Methods 53
1.8.1 First order (Euler) method. 53

2 Higher Order Linear Equations 57

2.0.1 Existence and uniqueness theorem. 57
2.1 Linear Homogeneous Equations 58
2.1.1 Linear Independence and the Wronskian 58
2.2 Linear constant coefficient equations: the characteristic polynomial 64
2.2.1 Real distinct roots 66
2.2.2 Complex distinct roots 67
2.2.3 Multiple roots 68
 6

2.3 Non-homogeneous linear equations: Operator notation and the structure of solutions 71
2.3.1 Operator notation 72
2.4 The structure of solutions to a non-homogeneous linear differential equation 73
2.5 The method of undetermined coefficients 74
2.6 The Annihilator Method 78
2.6.1 Problems: Undetermined Coefficients 81
2.7 Variation of Parameters 82
2.8 The Laplace Transform 85
2.8.1 Laplace Transform Problems 89

3 Mechanical and electrical oscillations 91

3.1 Mechanical oscillations 91
3.1.1 Undamped mass-spring systems 91
3.1.2 Mass-spring systems with damping 93
3.1.3 Problems: Mechanical Oscillations 97
3.2 RC and RLC circuits 102
3.2.1 RC Circuits 102
3.2.2 RLC Circuits, Complex numbers and Phasors 103

4 Systems of Ordinary Differential Equations 109

4.1 Introduction 109
4.2 Existence and Uniqueness 111
4.2.1 Homogeneous linear first order systems. 112
4.3 Variation of Parameters 114
4.4 Constant coefficient linear systems. 116
4.5 Problems on Linear Algebra and Systems 123

II Boundary Value Problems, Fourier Series, and the Solution of Partial Dif-
ferential Equations. 125

5 Boundary Value Problems 127

 7

5.1 Examples of boundary value problems 127

5.2 Existence and Uniqueness 129
5.3 Eigenvalue Problems 132

6 Fourier Series 137

6.1 Background 137
6.2 The Classical Fourier Series: Orthogonality 138
6.3 Periodic, even and odd extensions 142
6.3.1 The periodic extension 142
6.3.2 The even extension 143
6.3.3 The odd reflection 143
6.4 The Fourier cosine and sine series. 144
6.5 Fourier series practice problems 147

7 Partial Differential Equations and Separation of Variables 149

7.1 The Heat Equation 149
7.1.1 The heat equation with Dirichlet boundary conditions. 150
7.1.2 Separation of Variables 151
7.1.3 The heat equation with Neumann boundary conditions. 154
7.2 The One-dimensional Wave Equation 158
7.2.1 Background 158
7.2.2 Interpretation 160
7.2.3 Solving for the coefficients 160

III Background Material and Solutions 163

8 Background Material 165

8.1 A Review of complex numbers and the Euler formula 165
8.2 A Review of linear algebra 169
8.2.1 Matrices 169
8.2.2 Matrix multiplication 170
 8

8.2.3 Determinants 171

8.2.4 Systems of Linear Equations 174
8.2.5 Eigenvalues and eigenvectors 176

9 Problem Solutions 181

List of Figures

1.1 An initial investment of $100 at 2% (r = .02) interest compounded

yearly. 22
1.2 An initial investment of $100 at 2% (r = .02) interest compounded
quarterly. 22
1.3 An initial investment of $100 at 2% (r = .02) interest compounded
continuously. 22
dy
1.4 A slope field for dt = − yt (blue) together with a solution curve (red). 29
dy t
1.5 A slope field for dt = y (blue) together with a solution curve (red). 29
dy
1.6 The slope field for the logistic equation = 2(1 − y ) ydt 30
1.7 A graph of US Oil production as a function of time from 1900 to 2020.
The red curve depicts a fit to a logistic curve. By Plazak - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=42670844
31
1.8 An illustration of the contraction mapping principle. 34
dy
1.9 The slope field and some solution curves for the logistic equation dt =
(1 − y ) y 43
dy
1.10 A plot of the slope field for dt = y2 − 1. 49
1.11 A plot of f (y) = y2 − 1 vs. y. The arrows on the y axis indicate the
dy
sign of dt = f (y). 49
dy
1.12 A plot of the phase line for dt = y2 − 1. The two equilibria are y =
−1 and y = 1. 49
dy
1.13 The phase-line for a bistable system (flip-flop) dt = y(1 − 2y)(y −
1). The equilibria y = 0 and y = 1 are stable, the equilibrium y =
1
2 unstable. 51
1.14 The bifurcation diagram for the logistic model with constant harvest-
kP2
ing, dP
dt = kP ( P0 − P ) − h. For low fishing rates, h < 4 there are
0

two equilibria, one stable and one unstable. for higher fishing rates
kP02
h> 4 there are no equilibria. 52
10

1.15 The Euler method got some positive press in the movie “Hidden Fig-
ures”, when Katherine Goble Johnson used it to calculate the orbits
of the Mercury astronauts. The trajectories of the first two manned
Mercury missions (Freedom 7, piloted by Alan Shepard and Liberty
Bell 7, piloted by Gus Grissom) were calculated entirely by hand by
Johnson and other computers. Glenn’s flight (Friendship 7) was the
first to have the orbit calculated by an electronic computer. Glenn
refused to fly the mission unless Johnson checked the results of the
electronic computation personally. NASA; restored by Adam Cuerden, Public domain, via Wiki-
media Commons 53
dy
1.16 A graph of the exact solution to dt = y2 + t2 with y(0) = 0 for
t ∈ (0, 1) together with the Euler and improved Euler approxima-
tions to the solution with N = 6 subdivisions ((∆t = 16 ). The step
size has been deliberately chosen to be large to exaggerate the dif-
ference. It is apparent that the improved Euler method does a bet-
ter job of approximating the solution than the standard Euler method,
and that the fourth order Runge-Kutta can’t be distinguished from
the exact solution at this scale. 54

2.1 Robert Hooke (1635–1703) was a natural philosopher and polymath.

Originally employed at Oxford as an organist he became assistant
to Robert Boyle. He first hypothesized the inverse square law of grav-
itation (later developed by Newton) and made contributions to me-
chanics, the theory of light, atronomy, geology, clockmaking and many
other fields. His writing could be somewhat obfuscated: he announced
what we know call Hooke’s law as the anagram “ceiiinosssttuv”. This
can be rearranged to “Ut tensio, sic vis”, Latin for “As the extension,
so the force.” It also anagrams to “Incisive Stouts”, so it is likely that
Hooke was an early beer critic as well. 68
2
3.1 The solutions to the equation ddt2x + γ dx ′
dx + 4x = 0 with x (0) = 1; x (0) =
0 for γ = 1, 4, 5 (shown in red, blue, and black respectively) rep-
resenting the underdamped, critically damped, and overdamped cases.
Note that the solution decays fastest in the critically damped case.
94
3.2 A plot of the magnitude of the complex amplitude A(ω ) as a func-
tion of ω for k = 1, m = 1 and different values of γ between γ =
0 and γ = 2 96
3.3 A plot of ϕ = arctan( k−γω
mω 2
) as a function of ω for k = 1, m = 1
and different values of γ between γ = 0 and γ = 2 96
3.4 The unit of capacitance, the farad, is named after Michael Faraday.
Despite growing up poor and leaving school at an early age to ap-
prentice to a bookbinder Faraday became one of the preeminent sci-
entists of his day. 102
 11

3.5 A simple circuit consisting of a battery, switch, resistor and capac-

itor. 102
3.6 A simple circuit consisting of a resistor, a capacitor, and a sinusoidally
varying voltage. 103
3.7 The particular solution for an RC-circuit with a sinusoidal voltage.
The current Ĩ (t) will lead the voltage by an angle between 0 and π2 .
As time increases the picture rotates counterclockwise but the an-
gle between the voltage and the current does not change. 106
3.8 The particular solution for an RL-circuit with a sinusoidal voltage.
The current Ĩ (t) will lag behind the voltage by an angle between 0
and π2 . As time increases the picture rotates counterclockwise but
the angle between the voltage and the current does not change. 106

6.1 The square wave function. 141

6.2 The square wave function together with the first twenty-six terms
of the Fourier series. 141
6.3 The function f ( x ) = x (1 − x ) for x ∈ (0, 1) and extended periodi-
cally. 141
6.4 The function from Example 6.2.2 (Black) together with first 5 terms
(Red) and first twenty five terms (Blue) of the Fourier series. 142
6.5 The function defined x ∈ (0, 1) (top) and the same function extended
periodically to the whole line (bottom). 142
6.6 The function defined x ∈ (0, 1) (top) and the even extension to the
whole line (bottom). 143
6.7 The function defined x ∈ (0, 1) (top) and the odd extension to the
whole line (bottom). 143
6.8 The function defined x ∈ (0, L) (top) and Fourier cosine series (mid-
dle) and Fourier sine series (bottom). 144
6.9 The function f ( x ) = x defined for x ∈ (0, 2) (top) together with
the even (middle) and odd (bottom) extensions. 145
6.10 The first fifty terms of the Fourier cosine (top) and sine (bottom) se-
ries for function f ( x ) = x defined for x ∈ (0, 2). 146

7.1 The temperature (in ◦ C) at the center of the rod as a function of time.
154
13
Introduction

This is a collection of lecture notes for an introductory undergraduate

differential equations class based on a course taught by the author at
the University of Illinois. The focus is primarily on linear equations
with some consideration of more qualitative results for nonlinear
equations. The text is divided into three parts. The first part covers
scalar equations, focusing on initial value problems but with some
consideration of boundary value problems near the end. The sec-
ond part considers systems of linear differential equations, mainly
constant coefficients. The third part considers eigenvalue problems,
partial differential equations and separation of variables.
 Part I

Ordinary Differential
Equations
1
First Order Differential Equations

1.1 What is a Differential Equation?

1.1.1 Basic terminology

A differential equation is any relationship between a function

(usually denoted y(t)) and its derivatives up to some order. An or-
dinary differential equation involves only ordinary derivatives – in
other words y is a function of a single variable t.

dy d2 y dk y
F (t, y, , 2 ,..., k ) = 0
dt dt dt

A partial differential equation is a relationship between the partial

derivatives of a function of two or more variables. We will mostly
deal with ordinary differential equations in the course, although we
will touch on some basic partial differential equations near the end of
the semester.

Example 1.1.1. The following are all differential equations. The first five
are ordinary differential equations, the remaining two are partial differential
20 differential equations

equations.

dy
= −ky (1.1)
dt
d2 y
= −y (1.2)
dt2
d5 y
= − y3 (1.3)
dt5
8
d y dy
+y
=y (1.4)
dt8 dt
 2 11 2
dy d y
y2 + = (1.5)
dt dt11
∂u 2
∂ u
= 2 (1.6)
∂t ∂x
∂u ∂u ∂3 u
+u + 3 = 0 (1.7)
∂t ∂x ∂x

Some basic terminology:

The order of a differential equation is the order of the high- Note that the term order is used in
est derivative which appears in the equation. The order of the equa- many different ways in mathematics.
In the context of differential equations
tions above are 1, 2, 5, 8, 11, 2 and 3 respectively. It is important to order refers exclusively to the order of
be able to determine the order of a differential equation since this the derivative. Don’t confuse the order
with the highest power that appears, or
determines the way in which the differential equation is posed. In any other usage of order.
particular it determines how many pieces of initial data we must
provide in order to have a unique solution.
In addition to the differential equation itself we typically need
to specify a certain number of initial conditions in order to find a
unique solution. The number of initial conditions that we need to
specify is generally equal to the order of the differential equation. For
instance for a second order differential equation we would typically
specify two pieces of initial data. Usually this would be the value of In many applications we specify the
the function and the value of the first derivative at the initial point t0 . values of the function and the first
k − 1 derivatives at the point t0 , so this
For instance an example might be will be assumed throughout the first
part of the text. In a few important
d2 y dy applications, however, it is necessary
= y3 − y y (0) = 1 (0) = −1. to specify function values at more
dt2 dt than one point. For instance one might
specify the function value(s) at t = 0 as
A differential equation where all of the data is specified at a single well as t = 1. This is called a boundary
point is called an initial value problem. A differential equation where value problem and will be discussed in
a subsequent section of the notes.
the data is specified at two or more points is called a boundary value
problem. We will mostly discuss initial value problems in this course,
although we will consider boundary value problems later in the
semester.
 first order differential equations 21

Exercise 1.1.1
For the initial value problem

d2 y dy
= y3 − y y (0) = 1 (0) = −1.
dt2 dt
d2 y d3 y
compute the value of dt2 (0) as well as dt3 (0).
Notice that this can be continued indefinitely: given the value of
dy
y(0) and dt (0) we can compute the value of any other derivative at
t = 0. This explains why, for a second order equation, two pieces
of initial data are the right amount of data – once we have specified
the first two derivatives we can in principle calculate all of the higher
order derivatives.

An equation is linear if the dependent variable and all its deriva-

tives enter into the equation linearly. The dependence on the inde-
pendent variable can be arbitrary. Of the seven equations in example
(1.1.1) equations (1.1), (1.2) and (1.6) are linear since the dependent
variable and its derivatives enter linearly. The third equation is non-
linear because of the y3 term, the fourth equation is nonlinear be-
dy
cause of the y dt term, and the seventh because of the u ∂u ∂x term.
A linear equation can be written in the form

dn y d n −1 y
an (t) n
+ a n −1 ( t ) n −1 + . . . a 0 ( t ) y = f ( t ).
dt dt
In the special case where the right hand side f (t) is zero

dn y d n −1 y
an (t) + a n −1 ( t ) + . . . a0 (t)y = 0.
dtn dtn−1
the equation is said to be homogeneous. If the right hand side is non-
zero the equation is said to be non-homogeneous. We will see in later
chapters that, if we are able to solve the homogeneous equation, then
we will be able to solve the non-homogeneous equation.
A great deal of time in this class will be spent learning to solve
linear differential equations. Solving nonlinear differential equations
is in general quite difficult, although certain special kinds can be
solved exactly.

1.1.2 Examples of differential equations

Differential equations arise in many physical contexts when

the rate of change of some quantity can be expressed in terms of the
quantity itself. This usually takes the form of some physical law. In
the first example we consider compound interest.
22 differential equations

Example 1.1.2 (Compound Interest). If a savings account accrues in-

terest that is compounded yearly at a rate r, and Pi denotes the amount of
Capital (Dollars)

160

money in the account after i years then (assuming that no other money is 150

added to or removed from the account in this period) Pi satisfies the equation 140

130

Pi+1 = (1 + r ) Pi . 120

110

This equation has the solution

100
Years
0 5 10 15 20 25

Pi = (1 + r )i P0 . Figure 1.1: An initial invest-

ment of $100 at 2% (r = .02)
If, instead of compounding yearly the interest is compounded twice a year interest compounded yearly.
then the amount of money Pi satisfies Capital (Dollars)

160

r
Pi+1/2 = (1 + ) Pi 150

2 140

r r
Pi+1 = (1 + ) Pi+1/2 = (1 + )2 Pi . 130

2 2 120

110

Here i is still measured in years, and r is still the yearly rate. The above 100
Years
equation has the solution 0 5 10 15 20

Figure 1.2: An initial invest-

Pi = (1 + r/2)2i P0 .
ment of $100 at 2% (r = .02)
Notice that compounding twice annually gives you a bit more in interest interest compounded quarterly.
than compounding once a year, since (1 + r/2)2 = 1 + r + r2 /4 > (1 + r ). Capital (Dollars)

2
The extra term r4 represents the fact that you are getting interest on the
160

150

interest. If the interest is compounded n times per year we find that 140

130
r ni
Pi = (1 + ) P0 120

n 110

Years
The continuum limit consists of letting n tend to infinity, and assuming 5 10 15 20

that Pi goes over to a continuous function P(t). Using the finite difference Figure 1.3: An initial invest-
P −P
approximation to the derivative, i+∆∆ i ≈ dP dt we find the differential ment of $100 at 2% (r = .02)
equation interest compounded continu-
dP
= rP. ously.
dt
It is not hard to see that this differential equation has the solution

P(t) = P0 ert .

The graphs in the side margin on the previous depict the growth of an
initial investment of $100 earning 2% interest per year over a twenty-four
year period when in the cases where the interest is compounded yearly,
quarterly (4× per year), and continuously. You can see that the graphs look
very similar.
A similar model governs radioactive decay except that in the case of
radioactive decay the constant r is negative, since one loses a fixed fraction
of the population at each step.
 first order differential equations 23

Example 1.1.3 (Newton’s Law of Cooling). Newton’s law of cooling

states that the rate of change of the temperature of a body is proportional
to the difference in temperature between the body and the surrounding
medium. What is the differential equation which governs the temperature of
the body? Newton’s law of cooling is an ap-
Let T denote the temperature of the body. The rate of change of the tem- proximation where we neglect the
distribution of temperature inside the
perature is obviously dTdt . If the temperature of the surrounding medium is body, and assume that it is all at the
denoted by T0 then we have the equation same temperature. If we are interested
in the temperature at different points
inside the body we would have to solve
dT
= −k( T − T0 ) the heat equation, a partial differential
dt equation. We will do this in a later
where −k is the constant of proportionality. We put the − sign in there chapter.
because it is clear that the temperature of the body should converge to that
of the medium: If T > T0 then the body should be getting cooler (the rate of
change should be negative

Another example comes from classical physics:

Example 1.1.4 (Newton’s Law of Motion). Let x (t) denote the position
2
of a particle. Newton’s second law of motion states that f = ma = m ddt2x .
If we assume that the force is conservative, meaning that the work done on
a particle is independent of the path, then it follows that force is given by
minus the derivative of the potential energy V (x(t)).
In this case we have
d2 x
m = F = −∇x V (x(t)).
dt2
This is, of course, a differential equation for the position x(t). It is second
order and is generally nonlinear.

Example 1.1.5 (Growth of a Raindrop). A simple model for the growth

of raindrops is this: the rate of change of the volume of the raindrop is It would be impossible to discuss
differential equations without
proportional to the surface area of the drop. The basic intuition is this: the discussing Isaac Newton, probably
drop changes its volume by (1) evaporation and (2) by merging with other the greatest mathematician/physicist
in history. One of Newton’s first
droplets. The rate at which both of these processes happen should be propor- applications of calculus was to
tional to the surface area. the problem of planetary motion,
The relevant equations are where he showed that Kepler’s
empirical laws of planetary motion
4π followed from the assumption that the
V= r ( t )3 planets move under the influence of
3 gravity. Illustration by After Godfrey
dV dr Kneller -https://exhibitions.lib.cam.
= 4πr2 =k 4πr 2
dt | {z dt}
| {z } ac.uk/linesofthought/artifacts/
Surface Area newton-by-kneller/
Rate of change of volume
dr
=k
dt
so the rate of change of the radius of the droplet is constant, so the radius
is expected to grow linearly in time.
24 differential equations

On the importance of verification: While it can be very dif-

ficult to solve a general ordinary differential equation, it is usually
pretty easy to check whether or not a given function y = f (t) solves
the equation, since one can simply compute all derivatives and check
whether or not the equation is satisfied.
Since this is a second order equation
Example 1.1.6. Verify that the function we should be able to specify two initial
conditions, the value of the function
and the value of the derivative. Thus it
y(t) = A cos(t) + B sin(t) should be no surprise that this solution
involves two arbitrary constants, A and
satisfies the equation B.
d2 y
= −y
dt2
Example 1.1.7. Check that the function This function only involves one un-
determined constant, t0 . Since this
1 equation is second order we expect that
y(t) = the general solution should involve two
cos(t − t0 )
constants. The most general solution
to this equation is something called
satisfies the equation an elliptic function. These are special
d2 y functions originally studied by Gauss,
= 2y3 − y
dt2 and are not generally expressible in
terms of the elementary transcendental
Practice Exercises: functions such as sin t, cos t, et , although
for certain special values they reduce to
elementary functions.
Exercise 1.1.2
Write down a differential equation which is different from any of the
examples from this chapter. Give the order of the equation, and state
whether it is linear or nonlinear.

Exercise 1.1.3
Check that the function y(t) = 1 − t2 satisfies the differential equation
ty′ − y = −(1 + t2 )

Exercise 1.1.4
Suppose that y(t) satisfies y′′ + y = 0 together with the initial condi-
dk y
tions y(0) = 1, y′ (0) = 0. What is dtk
(0) as a function of k?

1.2 Basic solution techniques.

There are essentially only two methods for solving ordinary

differential equations (ODEs). The first is to recognize the equation
as representing the exact derivative of something, and integrate up
 first order differential equations 25

using the fundamental theorem of calculus. The second is to guess.

This text will cover a number of techniques for solving ordinary
differential equations, but they all reduce to one of these two basic
methods.
This section will mainly focus on the first technique: recognizing
an equation as representing an exact derivative. There are a number
of types of equation that often arise in practice where this is possible.
The simplest example is when one has some derivative of y equal
to a function of t. This occasionally arises in applications. For such
equations we can simply integrate up a number of times.

Example 1.2.1. Suppose that the height y(t) of a falling body evolves
according to We see again a lesson from the previous
d2 y section: we have a second order differ-
= −g ential equation the solution of which
dt2 involves two arbitrary constants of in-
Find the height as a function of time. tegration, v and h. The general solution
of an equation of nth order will typically
Integrating up once we find that
involve n arbitrary constants. Here the
constants enter linearly, since the equa-
dy
= − gt + v tion is linear. For a nonlinear equation
dt the dependence on the constants could
be much more complicated.
where v is a constant of integration (representing the velocity of the body at
time t = 0: v = y′ (0)). Integrating up a second time gives the equation

gt2
y=− + vt + h
2
where h = y(0) is a second constant of integration representing the initial
height of the body.

In the previous example the equation was given in the form of an

exact derivative. While this is not always the case one can frequently
manipulate the equation so that it takes the form of an exact deriva-
tive.

Example 1.2.2. Solve the equation

dT
= −k( T − T0 )
dt
arising in Newton’s law of cooling.
If we divide the equation through by T − T0 we find
dT
dt
= −k
T − T0

The left-hand side is the derivative of ln | T − T0 |. Integrating up gives

ln | T − T0 | = −kt + A

where A is a constant of integration. Exponentiating give The form of this solution should not be
surprising. As t → ∞ the exponential
term decays away and we have T (t) →
T0 . This implies that, at long times, the
temperature of the body tends to the
equilibrium temperature.
26 differential equations

| T − T0 | = e−kt+ A = e A e−kt
T − T0 = ±e A e−kt
T = T0 + Ce−kt .

In the last step we have a prefactor ±e A , where the ± comes from elim-
inating the absolute value. This is, in the end, just a constant so we call
C = ±e A .

Here is a similar example that involves both the independent

variable t and the dependent variable y:

Example 1.2.3. Solve the initial value problem

dy
= ty y (0) = 1
dt
We can rewrite this as
dy
= tdt.
y
Once we have done this we are in a situation where we have only y on the
left-hand side of the equation and only t on the right. Since the equation has
been “separated” in this way we can integrate it. Integrating gives

dy
Z Z
= tdt (1.8)
y
t2
ln(y(t)) = +C (1.9)
2
Imposing the initial condition y(0) = 1 gives C + 0 = ln 1 = 0, therefore

t2
ln |y(t)| =
2
t2
y(t) = e 2 .

The most general first order equation which can be solved this way
is called a separable equation, which we define as follows:

Definition 1.2.1. A first order equation is said to be separable if it can be

written in the form
dy
= f (y) g(t)
dt
If f (y) ̸= 0 then this equation can be integrated in the following manner.

dy
Z Z
= g(t)dt + C
f (y)

This procedure gives y implicitly as a function of t via the implicit

function theorem. It may or may not be possible to actually solve for
y explicitly as a function of t, as we were able to do in the preceeding
two examples.
 first order differential equations 27

Example 1.2.4. Solve the equation

dy
= ( y2 + y ) t y (2) = 1
dt
We can solve this by writing it in the form
dy
= tdt
y2 +y
and integrating up to get
Z y(t) Z t
dy
= tdt (1.10)
y (2) y2 + y 2
Z y(t) Z t
1 1
− dy = tdt (1.11)
y (2) y y+1 2
t2 t
ln(y) − ln(y + 1) − ln(1) + ln(2) = | (1.12)
2 2
Exponentiating the above gives
y t2 1 t2
= e 2 −2−ln(2) = e 2 −2
y+1 2
This can be solved for y to give
t2
e 2 −2
y= t2
2 − e 2 −2
Another example where it is not really possible to solve for y
explicitly as a function of t is given by the following:

Example 1.2.5. Solve the equation

dy t2 + 5
= 3 y (0) = 1
dt y + 2y
One should be careful in applying this idea. In particular one must
be careful to not introduce extraneous roots, or to remove relevant
roots, by assuming that certain quantities are non-zero. Here is an
example

Example 1.2.6. Solve the differential equation

dy 1
= y3 y (0) = 0
dt
Since this is a separable equation we can write it as
dy
1
= dt
y3
Which integrates up to
3 2
y3 = t + c
2
28 differential equations

when t = 0 we have y = 0, giving c = 0. Thus we have

2t 3
y=( )2
3
There is only one problem: This is NOT the only solution - there is a second
1
solution y(t) = 0. Dividing through by y 3 implicity assumes that this
quantity is not zero, and so this procedure misses the solution y(t) = 0.
This should be a little bit unsettling. If a differential equation is modeling
a physical system, such as a trajectory, there should be a unique solution:
physically there is only one trajectory. Here we have a differential equation
with more than one solution. So what happened?
1
Well, when we divide through by y 3 we are making the implicit assump-
tion that this quantity is not identically zero. So we need to go back and
check that y = 0 is not a solution to the equation. In this case it is a solution
to the equation.
Later we will state a fundamental existence and uniqueness theorem. This
will show that for “nice” differential equations this simply does not happen:
given an initial condition there is one and only one solution.
Practice Exercises:

Exercise 1.2.1
Solve the differential equation y′ = ty ln |y|

Exercise 1.2.2
Suppose that a projectile is fired upwards at 100 meters per second,
and that the acceleration due to gravity is 10m s−2 . At what time
does the projectile hit the ground?

Exercise 1.2.3
If the projectile has a horizontal velocity of 20 meters per second how
far has it traveled horizontally when it strikes the ground?

Exercise 1.2.4
Solve the following initial value problems.
a) y′ = cos(t) + 1, y (0) = 2 b) t2 y′ = 1, y (1) = 0
1 ′
c) ty = et , y (0) = 0 d) ty′ = 2, y (1) = 3
e) tyy′ = 1, y (1) = 2 f) y′ = y cos(t) + y, y (0) = 2
y2
g) y′ = 2ty + y − 2t − 1, y(0) = h) y′ = t +1 , y (0) = 2
0
 first order differential equations 29

1.3 Slope fields for first order equations

Slope fields give a geometric interpretation for differential equa-

tions that is analogous to the familiar interpretation from calculus This tangent line interpretation will also
of the derivative as the slope of the tangent line. If one considers a be useful for deriving numerical meth-
dy ods for solving differential equations
first order differential equation dt = f (y, t) the left-hand side gives such as the Euler approximation.
the derivative of the solution, which we interpret as the slope of the
tangent line, as a function of y and t.

dy
= f (y, t)
dt | {z } 1.0
function of (t, y)
|{z}
slope

In other words the differential equation gives the slope of the

0.5

curve as a function of (t, y). This suggests a graphical construction

0.0
for solution curves. At each point (t, y) in the plane one can draw a

y
small line segment of slope f (y, t). This is known as a vector field or
-0.5

slope field.
By “following” the slope lines we can generate a solution curve to -1.0

the differential equation. This method is excellent for giving a quali- -1.0 -0.5 0.0 0.5 1.0

tative idea as to the behavior of solutions to differential equations. x

Example 1.3.1. Consider the differential equation Figure 1.4: A slope field for
dy t
dt = − y (blue) together with a
dy t solution curve (red).
=−
dt y

The slope field associated to this equation, along with one solution curve,
are shown in the margin. One can see that the curve is tangent to the line
segments. Since this equation is separable it can be solved explicitly

dy t
=−
dt y
ydy = −tdt
y2 t2
= c−
2 2
y2 + t2 = 2c

from which it follows that the solution curves are circles.

A similar looking example is provided by 1.0

dy t 0.5
=
dt y
0.0
while this differential equation looks quite similar to the previous one the
y

slope field as well as the solutions appear quite different. The solution curve
-0.5

looks like a hyperbola, and this can be verified by integrating the equation.
-1.0

-1.0 -0.5 0.0 0.5 1.0

x

Figure 1.5: A slope field for

dy
dt = yt (blue) together with a
solution curve (red).
30 differential equations

Often one would like a qualitative understanding of the behavior

of a differential equation, without necessarily needing to know the
solution formula. There are a number of equations that are amenable
to this type of qualitative analysis. One famous one is the logistic
model
Example 1.3.2. Logistic Model A common model for the growth of popu-
lations is called the “logistic model”.
The logistic model posits that the growth of a population is according to 1.0

the equation
dP P
= k(1 − ) P.
dt P0

y
0.5

Here k and P0 are positive constants. The quantity k is a growth rate and
quantity P0 is known as the carrying capacity. Note that for small
populations, P < P0 the growth rate is positive, but for populations above 0.0

the maximum sustainable one (P > P0 ) the growth rate becomes negative
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

and the population decreases. x

The slope field is shown in Figure (1.9) for parameter values k = 2 and
Figure 1.6: The slope field
P0 = 1, along with a typical solution curve (red). The solution grows in an
for the logistic equation
exponential fashion for a while but the population saturates at the carrying dy
dt = 2(1 − y ) y
capacity P0 = 1.
This model has also been applied to study the question of “peak oil.” In
this context the quantity P(t) represents the total amount of oil pumped
from the ground from the beginning of the oil industry to time t. This is as-
sumed to follow a logistic curve, with the carrying capacity Po representing
the total amount of accessible oil.
dP
= k0 ( P0 − P(t)) P(t)
dt
The logistic nature of the curve is meant to reflect the fact that as more oil
is pumped the remainder becomes harder to recover. This makes a certain
amount of sense - there is only a finite amount of oil on Earth, so it makes
sense that P(t) should asymptote to a constant value. There is considerable
debate as to whether this hypothesis is correct, and if so how to estimate the
parameters k, P0 .
One way to estimate the constant P0 , which represents the total amount
of oil, is to look at P′ (t), the rate at which oil is being pumped from the
ground, as this is something for which we have data. It is easy to see from
the graph of P(t) that P′ (t) has a single maximum (hence the phrase peak
oil) and decays away from there (in the graph above the maximum of P′ (t)
occurs at t = 0. It is not hard to calculate that P′ (t), the rate of oil produc-
tion, has a maximum when P = P0 /2, which is to say half of all the oil is
gone. The easiest way to see this is to differentiate the differential equation

P′ (t) = kP( P0 − P)
′′
P (t) = k( P0 − 2P) P′ = k2 P( P0 − P)( P0 − 2P)
 first order differential equations 31

The pumping rate P′ (t) should have a maximum when P′′ (t) = 0 which
occurs when P = 0, P = P20 or P = P0 . From the equation we have that
P′ (t) = 0 when P = 0 or when P = P0 , so the maximum of the pumping
rate occurs when half of the oil has been pumped from the ground. Thus
one way to try to estimate the total amount of oil is to look at the pumping
records, try to determine when the peak production occured (or will occur)
and conjecture that, at that point, half of all the oil has been pumped.
While the jury is still out on this (for more details see David Goodstein’s
book “Out of Gas: The End of the Age Of Oil ”) it seems that the peak oil Figure 1.7: A graph of US Oil
theory has done a very good job of predicting US oil production but not such production as a function of
a good job predicting natural gas production. time from 1900 to 2020. The red
curve depicts a fit to a logis-
The graph in the margin shows the crude oil production in the
tic curve. By Plazak - Own work, CC BY-SA 4.0,
US and a fit of the curve P′ (t) to it. It is clear that the peak is around
https://commons.wikimedia.org/w/index.php?curid=42670844
1970. A crude way to estimate the total area under the curve is to
approximate it as a triangle of base ≈ 70yrs and a height of about
3.5Gb/year giving an area of about 125Gb (= 125 Billion barrels) of
total production (By eyeball this looks like an overestimate). This
suggests that the total amount of crude produced in the US over all
time will asymptote to something like 250Gb. The US currently con-
sumes 15 Million barrels of oil a day, about 23 of which are imported.
All data and the US crude production graph from Wikipedia

1.4 Existence-Uniqueness theorem

In the previous section we saw that the differential equation

1
y′ = y 3 y (0) = 0

does not have a unique solution. There are at least two solutions sat-
isfying the same differential equation and the same initial condition:

2 3
y1 ( t ) = ( t ) 2
3
y2 (t) = 0.

This is not good for a physical problem. If one is calculating some

physical quantity - the trajectory of a rocket, the temperature of a
reactor, etc. - it is important to know that

• The problem has a solution. (Existence)

• There is only one solution. (Uniqueness)

This is the problem of existence and uniqueness of solutions. The

example above shows us that this is something that cannot just be
32 differential equations

assumed. In this section we give a theorem that guarantees the exis-

tence and uniqueness of solutions to certain differential equations.
Let’s begin by stating the following theorem, which is the funda-
mental existence result for ordinary differential equations

Theorem 1.4.1 (Existence and Uniqueness). Consider the first order

initial value problem

dy
= f (y, t) y ( t0 ) = y0
dt
• If f (y, t) is continuous in a neighborhood of the point (y0 , t0 ) then a
solution exists in some rectangle |t − t0 | < δ, |y − y0 | < ϵ.
∂f
• If, in addition, ∂y (y, t) is continuous in a neighborhood of the point
(y0 , t0 ) then the solution is unique.

The above conditions are sufficient, but not necessary. It may be

that one or the other of these conditions fails and the solution still
exists/is unique.

Let’s consider some examples:

Example 1.4.1. What can you say about the equation

dy
= (y − 5) ln(|y − 5|) + t y (0) = 5
dt
the function (y − 5) ln(|y − 5|) + t is continuous in a neighborhood of
y = 5, t = 0. However computing the first partial with respect to y gives

∂f
(y, t) = 1 + ln |y − 5|
∂y

which is NOT continuous in a neighborhood of the point (t = 0, y =

5). The above theorem guarantees that there is a solution, but does not
guarantee that it is unique. Again the above theorem gives a sufficient
∂f
condition: The fact that ∂y is not continuous does not necessarily imply
that there is more than one solution, only that there may be more than one
solution.
Note that if we had the initial condition y(0) = 1 (or anything other than
y(0) = 5 the theorem would apply, and we would have a unique solution, at
least in a neighborhood of the initial point.

The method of proof is interesting. While we won’t actually prove

it we will mention how it is proved. The method is called “Picard
iteration”. Iterative methods are important in many aspects of nu-
merical analysis and engineering as well as in mathematics, so it
is worthwhile to give some sense of how this particular iteration
 first order differential equations 33

scheme works. In this case Picard iteration is a useful theoretical tool

but is not a very good way to solve equations numerically in practice.
Numerical methods will be covered later in the notes.
The idea behind Picard iteration is as follows: suppose that one
want to solve
y′ = f (y, t) y ( t0 ) = y0 .

The idea is to construct a sequence of functions yn (t) that get closer

and closer to solving the differential equation. The initial guess, y0 (t),
will be a function satisfying the right boundary condition. The easiest
thing to take is a constant y(0) (t) = y0 . The next guess y(1) (t) is
defined by
dy(1)
( t ) = f ( y (0) ( t ) , t ) y (1) ( t 0 ) = y 0
dt
One continues in this way: the nth iterate yn (t) satisfies

dy(n)
( t ) = f ( y ( n −1) ( t ), t ) y ( n ) ( t0 ) = y0 .
dt
which can be integrated up to give
Z t
y ( n ) ( t ) = y0 + f (y(n−1) (t), t)dt
t0

Picard’s existence and uniqueness theorem guarantees that this pro-

cedure converges to a solution of the original equation.

Theorem 1.4.2. Consider the initial value problem

dy
= f (y, t) y ( t0 ) = y0 .
dt
∂f
Suppose that f (y, t) and ∂y are continuous in a neighborhood of the point
(y0 , t0 ). Then the initial value problem has a unique solution in some inter-
val [t0 − δ, t0 + δ] for δ > 0, and the Picard iterates yn (t) converge to this
solution as n → ∞.

As an example: Lets try an easy one by hand (NOTE: In practice

one would never try to do this by hand - it is primarily a theoretical
tool - but this example works out nicely.)

Example 1.4.2. Find the Picard iterates for

y′ = y y (0) = 1

The first Picard iterate is y(0) (t) = 1. The second is the solution to

dy(1)
= y (0) ( t ) = 1 y (1) ( 0 ) = 1
dt
34 differential equations

The solution to this is y(1) (t) = 1 + t. Continuing in this fashion y(2) (t)
solves
dy(2)
= y (1) ( t ) = 1 + t y (2) ( 0 ) = 1
dt
t2
so y(2) (t) = 1 + t + 2. Continuing in this fashion we find that

t2 t3
y (3) ( t ) = 1 + t + +
2 6
t2 t3 t4
y (4) ( t ) = 1 + t + + +
2 6 24
t2 t3 t4 t5
y (5) ( t ) = 1 + t + + + +
2 6 24 120

It is not difficult to see by induction that the nth Picard iterate is just the
nth Taylor polynomial for the function y = et , which is the solution to the
original differential equation.

A second example is as follows:

Example 1.4.3. Find the first few Picard iterates of the equation

y ′ = t2 + y2 y (0) = 1

They are

y (0) ( t ) = 1
y(1) (t) = 1 + t + t3 /3
first order differential equations 33

2t3 t4 2t5 t7
y (2) ( t ) = 1 + t + t 2 +
They are

+ + + y (0) ( t ) = 1

3 6 15 63
y(1) (t) = 1 + t + t3 /3
first order differential equations 33

2t3 t4 2t5 t7
y (2) ( t ) = 1 + t + t 2 +
They are

+ + + y (0) ( t ) = 1

3 6 15 63
y(1) (t) = 1 + t + t3 /3
first order differential equations 33

2t3 t4 2t5 t7
y (2) ( t ) = 1 + t + t 2 +
They are

+ + + y (0) ( t ) = 1

3 6 15 63
y(1) (t) = 1 + t + t3 /3
first order differential equations 33

2t3 t4 2t5 t7
y (2) ( t ) = 1 + t + t 2 +
They are

+ + + y (0) ( t ) = 1

3 6 15 63
y(1) (t) = 1 + t + t3 /3
first order differential equations 33

2t3 t4 2t5 t7

47t7 41t8
y (2) ( t ) = 1 + t + t 2 +
They are

+ + +
3 6 15 63
y (0) ( t ) = 1
y(1) (t) = 1 + t + t3 /3

y (2) ( t ) = 1 + t + t 2 +
2t3
+ +
t4 2t5 t7 The solution to this is y(1) (t) = 1 + t. Continuing in this fashion y(2) (t)
solves
first order differential equations 35

47t7 41t8
4 5 8 29 47t7 41t8
y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +
4
3
5
6 15
+
8
63
29 47t7 41t8
so y(2) (t) = 1 + t +
dy(2)

y (3) ( t ) = 1 + t +
dt
= y (1) ( t ) = 1 + t

t2
y (2) ( 0 ) = 1

It is not difficult to see by induction that the nth Picard iterate is just the
y (4) ( t ) = 1 + t +
t2
2. Continuing in this fashion we find that

y (5) ( t ) = 1 + t +
t2

t2
2

2
+

+ +
t3

t3
6

6 24
t4

4 5 8 29
y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 + +
+
299t9 4t10
3
184t11
6 15
t12
90 315
+
630 original differential equation.

11340
+
525
+
51975
+
4t13 t15
2268 12285 59535
+ +
Example 1.4.3. Find the first few Picard iterates of the equation
nth Taylor polynomial for the function y = et , which is the solution to the
2
+ +
t3
6 24 120
t4
+
t5

A second example is as follows:

47t7 41t8
3 6 15 90 315 630
The pattern for the coefficients in this example is not so clear.
They are

y (0) ( t ) = 1
y 0 = t2 + y2 y (0) = 1

y(1) (t) = 1 + t + t3 /3

y (2) ( t ) = 1 + t + t 2 +

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +
2t3

+
The solution to this is y(1) (t) = 1 + t. Continuing in this fashion y(2) (t)

11340
299t9
+
4t10
525
4
3
3

+
+ +

184t11
5
6
t4

51975
6
2t5
15

15
+
8
+

t12
63
t7

29
90
+
4t13
47t7
315
+
+
t15
41t8
630
solves

so y(2) (t) = 1 + t +

It is not difficult to see by induction that the nth Picard iterate is just the

original differential equation.

Example 1.4.3. Find the first few Picard iterates of the equation
dy(2)

They are

behind the proof is something called the contraction mapping prin-

We will not present the full proof here, but the basic principle The pattern for the coefficients in this example is not so clear.
nth Taylor polynomial for the function y = et , which is the solution to the

behind the proof is something called the contraction mapping prin-

A second example is as follows:

ciple. The contraction mapping principle is the simplest example of

what mathematicians call a fixed point theorem. Leaving aside tech-
nical details the contraction mapping principle says the following.
Suppose we have a set W, and for each element f 2 W we have a no-
tion of the size of each element k f k. This is what is called a normed
space. Now suppose that we have a map T from W to itself such that
for any two elements f , g 2 W we have that k T ( f ) T ( g)k  rk f gk
y (0) ( t ) = 1

for r < 1. In other words shrinks the distances between points in the
y(1) (t) = 1 + t + t3 /3
y (3) ( t ) = 1 + t +

set by a fixed contraction factor r that is strictly less than one. Then
there is a unique element h that is fixed by the map: T (h) = h, and
We will not present the full proof here, but the basic principle
y (2) ( t ) = 1 + t + t 2 +
y (4) ( t ) = 1 + t +

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +
y (5) ( t ) = 1 + t +
dt
t2
2.
= y (1) ( t ) = 1 + t

Continuing in this fashion we find that

t2

t2

t2
2

2
+

+ +

+ +
t3

t3

t3
6

6
y (2) ( 0 ) = 1

24
t4

t4
+
t5
first order differential equations 35

The pattern for the coefficients in this example is not so clear.

behind the proof is something called the contraction mapping prin-

2268 12285 59535

ciple. The contraction mapping principle is the simplest example of

what mathematicians call a fixed point theorem. Leaving aside tech-
+
11340
299t9
+
y 0 = t2 + y2

4t10
525
2t3

4
3
3

+
+ +

184t11
5
6
t4

51975
6
2

2t5
15
y (0) = 1

15
6

+
8
+

2268 12285 59535

t12
63
24 120

t7

29
90
+
4t13
47t7
315
+
+
t15
41t8
630
The solution to this is y(1) (t) = 1 + t. Continuing in this fashion y(2) (t)
solves

so y(2) (t) = 1 + t +

It is not difficult to see by induction that the nth Picard iterate is just the

original differential equation.

Example 1.4.3. Find the first few Picard iterates of the equation

nical details the contraction mapping principle says the following.

We will not present the full proof here, but the basic principle
They are

The pattern for the coefficients in this example is not so clear.

behind the proof is something called the contraction mapping prin-

ciple. The contraction mapping principle is the simplest example of
what mathematicians call a fixed point theorem. Leaving aside tech-
nical details the contraction mapping principle says the following.
Suppose we have a set W, and for each element f 2 W we have a no-
tion of the size of each element k f k. This is what is called a normed
space. Now suppose that we have a map T from W to itself such that
for any two elements f , g 2 W we have that k T ( f ) T ( g)k  rk f gk
for r < 1. In other words shrinks the distances between points in the
set by a fixed contraction factor r that is strictly less than one. Then
there is a unique element h that is fixed by the map: T (h) = h, and
nth Taylor polynomial for the function y = et , which is the solution to the

A second example is as follows:

We will not present the full proof here, but the basic principle
y (0) ( t ) = 1
y(1) (t) = 1 + t + t3 /3

y (2) ( t ) = 1 + t + t 2 +

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +

+
dy(2)

11340
y (3) ( t ) = 1 + t +

y (4) ( t ) = 1 + t +

y (5) ( t ) = 1 + t +

299t9
dt
t2
2.
= y (1) ( t ) = 1 + t

Continuing in this fashion we find that

+
y 0 = t2 + y2

4t10
525
2t3

4
3
3

+
+ +

184t11
5
6
t4

51975
6
t2

t2

t2
2

2
+

+ +

+ +

2t5
15
y (0) = 1

15
t3

t3

t3
6

+
8
+

2268 12285 59535

t12
y (2) ( 0 ) = 1

63
24

24 120

t7
t4

t4

29
90
+

+
t5

4t13
47t7
315
+
+
t15
41t8
630
first order differential equations

The solution to this is y(1) (t) = 1 + t. Continuing in this fashion y(2) (t)
solves

so y(2) (t) = 1 + t +

It is not difficult to see by induction that the nth Picard iterate is just the

original differential equation.

Example 1.4.3. Find the first few Picard iterates of the equation

They are

The pattern for the coefficients in this example is not so clear.

behind the proof is something called the contraction mapping prin-

ciple. The contraction mapping principle is the simplest example of
what mathematicians call a fixed point theorem. Leaving aside tech-
nical details the contraction mapping principle says the following.
Suppose we have a set W, and for each element f 2 W we have a no-
tion of the size of each element k f k. This is what is called a normed
space. Now suppose that we have a map T from W to itself such that
for any two elements f , g 2 W we have that k T ( f ) T ( g)k  rk f gk
for r < 1. In other words shrinks the distances between points in the
set by a fixed contraction factor r that is strictly less than one. Then
there is a unique element h that is fixed by the map: T (h) = h, and
nth Taylor polynomial for the function y = et , which is the solution to the

A second example is as follows:

We will not present the full proof here, but the basic principle
y (0) ( t ) = 1
y(1) (t) = 1 + t + t3 /3

y (2) ( t ) = 1 + t + t 2 +

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +

+
dy(2)

11340
y (3) ( t ) = 1 + t +

y (4) ( t ) = 1 + t +

y (5) ( t ) = 1 + t +

299t9
dt
t2
2.
= y (1) ( t ) = 1 + t

Continuing in this fashion we find that

+
y 0 = t2 + y2

4t10
525
2t3

4
3
3

+
+ +

184t11
5
6
t4

51975
6
t2

t2

t2
2

2
+

+ +

+ +

2t5
15
y (0) = 1

15
t3

t3

t3
6

+
8
+

2268 12285 59535

t12
y (2) ( 0 ) = 1

63
24

24 120

t7
t4

t4

29
90
+

+
t5

4t13
47t7
315
+
+
t15
41t8
630
first order differential equations

Figure 1.8: An illustration of the

contraction mapping principle.
The solution to this is y(1) (t) = 1 + t. Continuing in this fashion y(2) (t)
solves

so y(2) (t) = 1 + t +

It is not difficult to see by induction that the nth Picard iterate is just the

original differential equation.

Example 1.4.3. Find the first few Picard iterates of the equation

They are

The pattern for the coefficients in this example is not so clear.

behind the proof is something called the contraction mapping prin-

ciple. The contraction mapping principle is the simplest example of
what mathematicians call a fixed point theorem. Leaving aside tech-
nical details the contraction mapping principle says the following.
Suppose we have a set W, and for each element f 2 W we have a no-
tion of the size of each element k f k. This is what is called a normed
space. Now suppose that we have a map T from W to itself such that
for any two elements f , g 2 W we have that k T ( f ) T ( g)k  rk f gk
for r < 1. In other words shrinks the distances between points in the
set by a fixed contraction factor r that is strictly less than one. Then
there is a unique element h that is fixed by the map: T (h) = h, and
nth Taylor polynomial for the function y = et , which is the solution to the

A second example is as follows:

We will not present the full proof here, but the basic principle
y (0) ( t ) = 1
y(1) (t) = 1 + t + t3 /3

y (2) ( t ) = 1 + t + t 2 +

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +

+
dy(2)

11340
y (3) ( t ) = 1 + t +

y (4) ( t ) = 1 + t +

y (5) ( t ) = 1 + t +

299t9
dt
t2
2.
= y (1) ( t ) = 1 + t

Continuing in this fashion we find that

+
y 0 = t2 + y2

4t10
525
2t3

4
3
3

+
+ +

184t11
5
6
t4

51975
6
t2

t2

t2
2

2
+

+ +

+ +

2t5
15
y (0) = 1

15
t3

t3

t3
6

+
8
+

2268 12285 59535

t12
y (2) ( 0 ) = 1

63
24

24 120

t7
t4

t4

29
90
+

+
t5

4t13
47t7
315
+
+
t15
41t8
630
first order differential equations

Figure 1.8: An illustration of the

contraction mapping principle.
They are

The pattern for the coefficients in this example is not so clear.

behind the proof is something called the contraction mapping prin-

ciple. The contraction mapping principle is the simplest example of
what mathematicians call a fixed point theorem. Leaving aside tech-
nical details the contraction mapping principle says the following.
Suppose we have a set W, and for each element f 2 W we have a no-
tion of the size of each element k f k. This is what is called a normed
space. Now suppose that we have a map T from W to itself such that
for any two elements f , g 2 W we have that k T ( f ) T ( g)k  rk f gk
for r < 1. In other words shrinks the distances between points in the
set by a fixed contraction factor r that is strictly less than one. Then
there is a unique element h that is fixed by the map: T (h) = h, and
this fixed point can be found by taking any starting point and iterat-
ing the map: lim N !• T N ( f ) = h. This theorem is illustrated by the
margin figure. The proof of the Picard existence theorem amounts
to showing that the Picard iterates form (for small enough neigh-
borhood of (t0 , y0 ) a contraction mapping, and thus converge to a
unique fixed point.

proven we will now apply it to some examples.

Example 1.4.4. Consider the equation

The function f (t, y) = t2 +y2 is NOT continuous in a neighborhood of the

origin (Why?), so the theorem does not apply. On the other hand if we had
the initial condition

then the function f (t, y) = t2 +y2 is continuous in a neighborhood of the

point y = 0, t = 1, so the equation would have a unique solution.
We will not present the full proof here, but the basic principle

Having some idea of how the existence and uniqueness theorem is

y (0) ( t ) = 1
y(1) (t) = 1 + t + t3 /3

y (2) ( t ) = 1 + t + t 2 +

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +

+
11340
299t9

dy

dy
dt

dt
+

ty
4t10

= 2

= 2
525
2t3

4
3
3

ty
t + y2

t + y2
+
+ +

ty

ty
184t11
5
6
t4

51975
6
2t5
15

15

y (0) = 0

y (1) = 0
+
8
+

2268 12285 59535

t12
63
t7

29
90
+
4t13
47t7
315
+
+
t15
41t8
630
first order differential equations

Figure 1.8: An illustration of the

contraction mapping principle.
They are

The pattern for the coefficients in this example is not so clear.

behind the proof is something called the contraction mapping prin-

ciple. The contraction mapping principle is the simplest example of
what mathematicians call a fixed point theorem. Leaving aside tech-
nical details the contraction mapping principle says the following.
Suppose we have a set W, and for each element f 2 W we have a no-
tion of the size of each element k f k. This is what is called a normed
space. Now suppose that we have a map T from W to itself such that
for any two elements f , g 2 W we have that k T ( f ) T ( g)k  rk f gk
for r < 1. In other words shrinks the distances between points in the
set by a fixed contraction factor r that is strictly less than one. Then
there is a unique element h that is fixed by the map: T (h) = h, and
this fixed point can be found by taking any starting point and iterat-
ing the map: lim N !• T N ( f ) = h. This theorem is illustrated by the
margin figure. The proof of the Picard existence theorem amounts
to showing that the Picard iterates form (for small enough neigh-
borhood of (t0 , y0 ) a contraction mapping, and thus converge to a
unique fixed point.

proven we will now apply it to some examples.

Example 1.4.4. Consider the equation

The function f (t, y) = t2 +y2 is NOT continuous in a neighborhood of the

origin (Why?), so the theorem does not apply. On the other hand if we had
the initial condition

then the function f (t, y) = t2 +y2 is continuous in a neighborhood of the

point y = 0, t = 1, so the equation would have a unique solution.
We will not present the full proof here, but the basic principle

Having some idea of how the existence and uniqueness theorem is

y (0) ( t ) = 1
y(1) (t) = 1 + t + t3 /3

y (2) ( t ) = 1 + t + t 2 +

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +

+
11340
299t9

dy

dy
dt

dt
+

ty
4t10

= 2

= 2
525
2t3

4
3
3

ty
t + y2

t + y2
+
+ +

ty

ty
184t11
5
6
t4

51975
6
2t5
15

15

y (0) = 0

y (1) = 0
+
8
+

2268 12285 59535

t12
63
t7

29
90
+
4t13
47t7
315
+
+
t15
41t8
630
first order differential equations

Figure 1.8: An illustration of the

contraction mapping principle.
They are

The pattern for the coefficients in this example is not so clear.

behind the proof is something called the contraction mapping prin-

ciple. The contraction mapping principle is the simplest example of
what mathematicians call a fixed point theorem. Leaving aside tech-
nical details the contraction mapping principle says the following.
Suppose we have a set W, and for each element f 2 W we have a no-
tion of the size of each element k f k. This is what is called a normed
space. Now suppose that we have a map T from W to itself such that
for any two elements f , g 2 W we have that k T ( f ) T ( g)k  rk f gk
for r < 1. In other words shrinks the distances between points in the
set by a fixed contraction factor r that is strictly less than one. Then
there is a unique element h that is fixed by the map: T (h) = h, and
this fixed point can be found by taking any starting point and iterat-
ing the map: lim N !• T N ( f ) = h. This theorem is illustrated by the
margin figure. The proof of the Picard existence theorem amounts
to showing that the Picard iterates form (for small enough neigh-
borhood of (t0 , y0 ) a contraction mapping, and thus converge to a
unique fixed point.

proven we will now apply it to some examples.

Example 1.4.4. Consider the equation

The function f (t, y) = t2 +y2 is NOT continuous in a neighborhood of the

origin (Why?), so the theorem does not apply. On the other hand if we had
the initial condition

then the function f (t, y) = t2 +y2 is continuous in a neighborhood of the

point y = 0, t = 1, so the equation would have a unique solution.
We will not present the full proof here, but the basic principle

Having some idea of how the existence and uniqueness theorem is

y (0) ( t ) = 1
y(1) (t) = 1 + t + t3 /3

y (2) ( t ) = 1 + t + t 2 +

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +

+
11340
299t9

dy

dy
dt

dt
+

ty
4t10

= 2

= 2
525
2t3

4
3
3

ty
t + y2

t + y2
+
+ +

ty

ty
184t11
5
6
t4

51975
6
2t5
15

15

y (0) = 0

y (1) = 0
+
8
+

2268 12285 59535

t12
63
t7

29
90
+
4t13
47t7
315
+
+
t15
41t8
630
first order differential equations

Figure 1.8: An illustration of the

contraction mapping principle.
They are

The pattern for the coefficients in this example is not so clear.

behind the proof is something called the contraction mapping prin-

ciple. The contraction mapping principle is the simplest example of
what mathematicians call a fixed point theorem. Leaving aside tech-
nical details the contraction mapping principle says the following.
Suppose we have a set W, and for each element f 2 W we have a no-
tion of the size of each element k f k. This is what is called a normed
space. Now suppose that we have a map T from W to itself such that
for any two elements f , g 2 W we have that k T ( f ) T ( g)k  rk f gk
for r < 1. In other words shrinks the distances between points in the
set by a fixed contraction factor r that is strictly less than one. Then
there is a unique element h that is fixed by the map: T (h) = h, and
this fixed point can be found by taking any starting point and iterat-
ing the map: lim N !• T N ( f ) = h. This theorem is illustrated by the
margin figure. The proof of the Picard existence theorem amounts
to showing that the Picard iterates form (for small enough neigh-
borhood of (t0 , y0 ) a contraction mapping, and thus converge to a
unique fixed point.

proven we will now apply it to some examples.

Example 1.4.4. Consider the equation

The function f (t, y) = t2 +y2 is NOT continuous in a neighborhood of the

origin (Why?), so the theorem does not apply. On the other hand if we had
the initial condition

then the function f (t, y) = t2 +y2 is continuous in a neighborhood of the

point y = 0, t = 1, so the equation would have a unique solution.
We will not present the full proof here, but the basic principle

Having some idea of how the existence and uniqueness theorem is

y (0) ( t ) = 1
y(1) (t) = 1 + t + t3 /3

y (2) ( t ) = 1 + t + t 2 +

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +

+
11340
299t9

dy

dy
dt

dt
+

ty
4t10

= 2

= 2
525
2t3

4
3
3

ty
t + y2

t + y2
+
+ +

ty

ty
184t11
5
6
t4

51975
6
2t5
15

15

y (0) = 0

y (1) = 0
+
8
+

2268 12285 59535

t12
63
t7

29
90
+
4t13
47t7
315
+
+
t15
41t8
630
first order differential equations

Figure 1.8: An illustration of the

contraction mapping principle.
They are

The pattern for the coefficients in this example is not so clear.

behind the proof is something called the contraction mapping prin-

ciple. The contraction mapping principle is the simplest example of
what mathematicians call a fixed point theorem. Leaving aside tech-
nical details the contraction mapping principle says the following.
Suppose we have a set W, and for each element f 2 W we have a no-
tion of the size of each element k f k. This is what is called a normed
space. Now suppose that we have a map T from W to itself such that
for any two elements f , g 2 W we have that k T ( f ) T ( g)k  rk f gk
for r < 1. In other words shrinks the distances between points in the
set by a fixed contraction factor r that is strictly less than one. Then
there is a unique element h that is fixed by the map: T (h) = h, and
this fixed point can be found by taking any starting point and iterat-
ing the map: lim N !• T N ( f ) = h. This theorem is illustrated by the
margin figure. The proof of the Picard existence theorem amounts
to showing that the Picard iterates form (for small enough neigh-
borhood of (t0 , y0 ) a contraction mapping, and thus converge to a
unique fixed point.

proven we will now apply it to some examples.

Example 1.4.4. Consider the equation

then the function f (t, y) = t2 +y2 is continuous in a neighborhood of the

point y = 0, t = 1, so the equation would have a unique solution.
The function f (t, y) = t2 +y2 is NOT continuous in a neighborhood of the
origin (Why?), so the theorem does not apply. On the other hand if we had
the initial condition
We will not present the full proof here, but the basic principle

Having some idea of how the existence and uniqueness theorem is

y (0) ( t ) = 1
y(1) (t) = 1 + t + t3 /3

y (2) ( t ) = 1 + t + t 2 +

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +

+
11340
299t9

dy

dy
dt

dt
+

ty
4t10

= 2

= 2
525
2t3

4
3
3

ty
t + y2

t + y2
+
+ +

ty

ty
184t11
5
6
t4

51975
6
2t5
15

15

y (0) = 0

y (1) = 0
+
8
+

2268 12285 59535

t12
63
t7

29
90
+
4t13
47t7
315
+
+
t15
41t8
630
first order differential equations

Figure 1.8: An illustration of the

contraction mapping principle.
firstorderdifferentialequations35Theyarey(0)(t)=1y(1)(t)=1+t+t3/3y(2)(t)=1+t+t2+2t33+t46+2t515+t763y(3)(t)=1+t+t2+43t3+56t6+815t5+2990t6+47t7315+41t8630+299t911340+4t10525+184t1151975+t122268+4t1312285+t1559535Thepatternforthecoefficientsinthisexampleisnotsoclear.Wewillnotpresentthefullproofhere,butthebasicprinciplebehindtheproofissomethingcalledthecontractionmappingprin-ciple.Thecontractionmappingprincipleisthesimplestexampleofwhatmathematicianscallafixedpointtheorem.Leavingasidetech-nicaldetailsthecontractionmappingprinciplesaysthefollowing.SupposewehaveasetW,andforeachelementf2Wwehaveano-tionofthesizeofeachelementkfk.Thisiswhatiscalledanormedspace.NowsupposethatwehaveamapTfromWtoitselfsuchthatforanytwoelementsf,g2WwehavethatkT(f)T(g)krkfgkforr<1.Inotherwordsshrinksthedistancesbetweenpointsinthesetbyafixedcontractionfactorrthatisstrictlylessthanone.Thenthereisauniqueelemenththatisfixedbythemap:T(h)=h,andthisfixedpointcanbefoundbytakinganystartingpointanditerat-ingthemap:limN!•TN(f)=h.Thistheoremisillustratedbythemarginfigure.TheproofofthePicardexistencetheoremamountstoshowingthatthePicarditeratesform(forsmallenoughneigh-borhoodof(t0,y0)acontractionmapping,andthusconvergetoauniquefixedpoint.
Figure1.8:Anillustrationofhecontractionmappingprinciple.Havingsomeideaofhowtheexistenceanduniquenesstheoremisprovenwewillnowapplyittosomeexamples.Example1.4.4.Considertheequationdydt=tyt2+y2y(0)=0Thefunctionf(t,y)=tyt2+y2isNOTcontinuousinaneighborhoodoftheorigin(Why?),sothetheoremdoesnotapply.Ontheotherhandifwehadtheinitialconditiondydt=tyt2+y2y(1)=0thenthefunctionf(t,y)=tyt2+y2iscontinuousinaneighborhoodofthepointy=0,t=1,sotheequationwouldhaveauniquesolution.
firstorderdifferentialequations35Theyarey(0)(t)=1y(1)(t)=1+t+t3/3y(2)(t)=1+t+t2+2t33+t46+2t515+t763y(3)(t)=1+t+t2+43t3+56t6+815t5+2990t6+47t7315+41t8630+299t911340+4t10525+184t1151975+t122268+4t1312285+t1559535Thepatternforthecoefficientsinthisexampleisnotsoclear.Wewillnotpresentthefullproofhere,butthebasicprinciplebehindtheproofissomethingcalledthecontractionmappingprin-ciple.Thecontractionmappingprincipleisthesimplestexampleofwhatmathematicianscallafixedpointtheorem.Leavingasidetech-nicaldetailsthecontractionmappingprinciplesaysthefollowing.SupposewehaveasetW,andforeachelementf2Wwehaveano-tionofthesizeofeachelementkfk.Thisiswhatiscalledanormedspace.NowsupposethatwehaveamapTfromWtoitselfsuchthatforanytwoelementsf,g2WwehavethatkT(f)T(g)krkfgkforr<1.Inotherwordsshrinksthedistancesbetweenpointsinthesetbyafixedcontractionfactorrthatisstrictlylessthanone.Thenthereisauniqueelemenththatisfixedbythemap:T(h)=h,andthisfixedpointcanbefoundbytakinganystartingpointanditerat-ingthemap:limN!•TN(f)=h.Thistheoremisillustratedbythemarginfigure.TheproofofthePicardexistencetheoremamountstoshowingthatthePicarditeratesform(forsmallenoughneigh-borhoodof(t0,y0)acontractionmapping,andthusconvergetoauniquefixedpoint.firstorderdifferentialequations31Theyarey(0)(x)=1y(1)(x)=1+x+x3/3y(2)(x)=1+x+x2+2x33+x46+2x515+x763y(3)(x)=1+x+x2+43x3+56x6+815x5+2990x6+47x7315+41x8630+299x911340+4x10525+184x1151975+x122268+4x1312285+x1559535Thepatternforthecoefficientsinthisexampleisnotsoclear.Wewillnotpresentthefullproofhere,butthebasicprinciplebehindtheproofissomethingcalledthecontractionmappingprin-ciple.Thecontractionmappingprincipleisthesimplestexampleofwhatmathematicianscallafixedpointtheorem.Leavingasidetech-nicaldetailsthecontractionmappingprinciplesaysthefollowing.SupposewehaveasetW,andforeachelementf2Wwehaveano-tionofthesizeofeachelementkfk.Thisiswhatiscalledanormedspace.NowsupposethatwehaveamapTfromWtoitselfsuchthatforanytwoelementsf,g2WwehavethatkT(f)T(g)krkfgkforr<1.Inotherwordsshrinksthedistancesbetweenpointsinthesetbyafixedcontractionfactorrthatisstrictlylessthanone.Thenthereisauniqueelemenththatisfixedbythemap:T(h)=h,andthisfixedpointcanbefoundbytakinganystartingpointanditerat-ingthemap:limN!•TN(f)=h.Thistheoremisillustratedbythemarginfigure.TheproofofthePicardexistencetheoremamountstoshowingthatthePicarditeratesform(forsmallenoughenigh-borhoodof(x0,y0)acontractionmapping,andthusconvergetoauniquefixedpoint.firstorderdifferentialequations31Theyarey(0)(x)=1y(1)(x)=1+x+x3/3y(2)(x)=1+x+x2+2x33+x46+2x515+x763y(3)(x)=1+x+x2+43x3+56x6+815x5+2990x6+47x7315+41x8630+299x911340+4x10525+184x1151975+x122268+4x1312285+x1559535Thepatternforthecoefficientsinthisexampleisnotsoclear.Wewillnotpresentthefullproofhere,butthebasicprinciplebehindtheproofissomethingcalledthecontractionmappingprin-ciple.Thecontractionmappingprincipleisthesimplestexampleofwhatmathematicianscallafixedpointtheorem.Leavingasidetech-nicaldetailsthecontractionmappingprinciplesaysthefollowing.SupposewehaveasetW,andforeachelementf2Wwehaveano-tionofthesizeofeachelementkfk.Thisiswhatiscalledanormedspace.NowsupposethatwehaveamapTfromWtoitselfsuchthatforanytwoelementsf,g2WwehavethatkT(f)T(g)krkfgkforr<1.Inotherwordsshrinksthedistancesbetweenpointsinthesetbyafixedcontractionfactorrthatisstrictlylessthanone.Thenthereisauniqueelemenththatisfixedbythemap:T(h)=h,andthisfixedpointcanbefoundbytakinganystartingpointanditerat-ingthemap:limN!•TN(f)=h.Thistheoremisillustratedbythemarginfigure.TheproofofthePicardexistencetheoremamountstoshowingthatthePicarditeratesform(forsmallenoughenigh-borhoodof(x0,y0)acontractionmapping,andthusconvergetoauniquefixedpoint.firstorderdifferentialequations31Theyarey(0)(x)=1y(1)(x)=1+x+x3/3y(2)(x)=1+x+x2+2x33+x46+2x515+x763y(3)(x)=1+x+x2+43x3+56x6+815x5+2990x6+47x7315+41x8630+299x911340+4x10525+184x1151975+x122268+4x1312285+x1559535Thepatternforthecoefficientsinthisexampleisnotsoclear.Wewillnotpresentthefullproofhere,butthebasicprinciplebehindtheproofissomethingcalledthecontractionmappingprin-ciple.Thecontractionmappingprincipleisthesimplestexampleofwhatmathematicianscallafixedpointtheorem.Leavingasidetech-nicaldetailsthecontractionmappingprinciplesaysthefollowing.SupposewehaveasetW,andforeachelementf2Wwehaveano-tionofthesizeofeachelementkfk.Thisiswhatiscalledanormedspace.NowsupposethatwehaveamapTfromWtoitselfsuchthatforanytwoelementsf,g2WwehavethatkT(f)T(g)krkfgkforr<1.Inotherwordsshrinksthedistancesbetweenpointsinthesetbyafixedcontractionfactorrthatisstrictlylessthanone.Thenthereisauniqueelemenththatisfixedbythemap:T(h)=h,andthisfixedpointcanbefoundbytakinganystartingpointanditerat-ingthemap:limN!•TN(f)=h.Thistheoremisillustratedbythemarginfigure.TheproofofthePicardexistencetheoremamountstoshowingthatthePicarditeratesform(forsmallenoughenigh-borhoodof(x0,y0)acontractionmapping,andthusconvergetoauniquefixedpoint.firstorderdifferentialequations31Theyarey(0)(x)=1y(1)(x)=1+x+x3/3y(2)(x)=1+x+x2+2x33+x46+2x515+x763y(3)(x)=1+x+x2+43x3+56x6+815x5+2990x6+47x7315+41x8630+299x911340+4x10525+184x1151975+x122268+4x1312285+x1559535Thepatternforthecoefficientsinthisexampleisnotsoclear.Wewillnotpresentthefullproofhere,butthebasicprinciplebehindtheproofissomethingcalledthecontractionmappingprin-ciple.Thecontractionmappingprincipleisthesimplestexampleofwhatmathematicianscallafixedpointtheorem.Leavingasidetech-nicaldetailsthecontractionmappingprinciplesaysthefollowing.SupposewehaveasetW,andforeachelementf2Wwehaveano-tionofthesizeofeachelementkfk.Thisiswhatiscalledanormedspace.NowsupposethatwehaveamapTfromWtoitselfsuchthatforanytwoelementsf,g2WwehavethatkT(f)T(g)krkfgkforr<1.Inotherwordsshrinksthedistancesbetweenpointsinthesetbyafixedcontractionfactorrthatisstrictlylessthanone.Thenthereisauniqueelemenththatisfixedbythemap:T(h)=h,andthisfixedpointcanbefoundbytakinganystartingpointanditerat-ingthemap:limN!•TN(f)=h.Thistheoremisillustratedbythemarginfigure.TheproofofthePicardexistencetheoremamountstoshowingthatthePicarditeratesform(forsmallenoughenigh-borhoodof(x0,y0)acontractionmapping,andthusconvergetoauniquefixedpoint.firstorderdifferentialequations31Theyarey(0)(x)=1y(1)(x)=1+x+x3/3y(2)(x)=1+x+x2+2x33+x46+2x515+x763y(3)(x)=1+x+x2+43x3+56x6+815x5+2990x6+47x7315+41x8630+299x911340+4x10525+184x1151975+x122268+4x1312285+x1559535Thepatternforthecoefficientsinthisexampleisnotsoclear.Wewillnotpresentthefullproofhere,butthebasicprinciplebehindtheproofissomethingcalledthecontractionmappingprin-ciple.Thecontractionmappingprincipleisthesimplestexampleofwhatmathematicianscallafixedpointtheorem.Leavingasidetech-nicaldetailsthecontractionmappingprinciplesaysthefollowing.SupposewehaveasetW,andforeachelementf2Wwehaveano-tionofthesizeofeachelementkfk.Thisiswhatiscalledanormedspace.NowsupposethatwehaveamapTfromWtoitselfsuchthatforanytwoelementsf,g2WwehavethatkT(f)T(g)krkfgkforr<1.Inotherwordsshrinksthedistancesbetweenpointsinthesetbyafixedcontractionfactorrthatisstrictlylessthanone.Thenthereisauniqueelemenththatisfixedbythemap:T(h)=h,andthisfixedpointcanbefoundbytakinganystartingpointanditerat-ingthemap:limN!•TN(f)=h.Thistheoremisillustratedbythemarginfigure.TheproofofthePicardexistencetheoremamountstoshowingthatthePicarditeratesform(forsmallenoughenigh-borhoodof(x0,y0)acontractionmapping,andthusconvergetoauniquefixedpoint.firstorderdifferentialequations31space.NowsupposethatwehaveamapTfromWtoitselfsuchthatforanytwoelementsf,g2WwehavethatkT(f)T(g)krkfgkforr<1.Inotherwordsshrinksthedistancesbetweenpointsinthesetbyafixedcontractionfactorrthatisstrictlylessthanone.Thenthereisauniqueelemenththatisfixedbythemap:T(h)=h,andthisfixedpointcanbefoundbytakinganystartingpointanditerat-ingthemap:limN!]inftyTN(f)=h.Thistheoremisillustratedbythemarginfigure.TheproofofthePicardexistencetheoremamountstoshowingthatthePicarditeratesform(forsmallenoughenigh-borhoodof(x0,y0)acontractionmapping,andthusconvergetoauniquefixedpoint.firstorderdifferentialequations31space.NowsupposethatwehaveamapTfromWtoitselfsuchthatforanytwoelementsf,g2WwehavethatkT(f)T(g)krkfgkforr<1.Inotherwordsshrinksthedistancesbetweenpointsinthesetbyafixedcontractionfactorrthatisstrictlylessthanone.Thenthereisauniqueelemenththatisfixedbythemap:T(h)=h,andthisfixedpointcanbefoundbytakinganystartingpointanditerat-ingthemap:limN!]inftyTN(f)=h.Thistheoremisillustratedbythemarginfigure.TheproofofthePicardexistencetheoremamountstoshowingthatthePicarditeratesform(forsmallenoughenigh-borhoodof(x0,y0)acontractionmapping,andthusconvergetoauniquefixedpoint.firstorderdifferentialequations31space.NowsupposethatwehaveamapTfromWtoitselfsuchthatforanytwoelementsf,g2WwehavethatkT(f)T(g)krkfgkforr<1.Inotherwordsshrinksthedistancesbetweenpointsinthesetbyafixedcontractionfactorrthatisstrictlylessthanone.Thenthereisauniqueelemenththatisfixedbythemap:T(h)=h,andthisfixedpointcanbefoundbytakinganystartingpointanditerat-ingthemap:limN!]inftyTN(f)=h.Thistheoremisillustratedbythemarginfigure.TheproofofthePicardexistencetheoremamountstoshowingthatthePicarditeratesform(forsmallenoughenigh-borhoodof(x0,y0)acontractionmapping,andthusconvergetoauniquefixedpoint.firstorderdifferentialequations31space.NowsupposethatwehaveamapTfromWtoitselfsuchthatforanytwoelementsf,g2WwehavethatkT(f)T(g)krkfgkforr<1.Inotherwordsshrinksthedistancesbetweenpointsinthesetbyafixedcontractionfactorrthatisstrictlylessthanone.Thenthereisauniqueelemenththatisfixedbythemap:T(h)=h,andthisfixedpointcanbefoundbytakinganystartingpointanditerat-ingthemap:limN!]inftyTN(f)=h.Thistheoremisillustratedbythemarginfigure.TheproofofthePicardexistencetheoremamountstoshowingthatthePicarditeratesform(forsmallenoughenigh-borhoodof(x0,y0)acontractionmapping,andthusconvergetoauniquefixedpoint.firstorderdifferentialequations31space.NowsupposethatwehaveamapTfromWtoitselfsuchthatforanytwoelementsf,g2WwehavethatkT(f)T(g)krkfgkforr<1.Inotherwordsshrinksthedistancesbetweenpointsinthesetbyafixedcontractionfactorrthatisstrictlylessthanone.Thenthereisauniqueelemenththatisfixedbythemap:T(h)=h,andthisfixedpointcanbefoundbytakinganystartingpointanditerat-ingthemap:limN!]inftyTN(f)=h.Thistheoremisillustratedbythemarginfigure.TheproofofthePicardexistencetheoremamountstoshowingthatthePicarditeratesform(forsmallenoughenigh-borhoodof(x0,y0)acontractionmapping,andthusconvergetoauniquefixedpoint.Figure1.8:AnIllustrationofthecon-tractionmappingprinciple.Havingsomeideaofhowtheexistenceanduniquenesstheoremisprovenwewillnowapplyittosomeexamples.Example1.4.4.Considertheequationdydx=xyx2+y2y(0)=0Thefunctionf(x,y)=xyx2+y2isNOTcontinuousinaneighborhoodoftheorigin(Why?),sothetheoremdoesnotapply.Ontheotherhandifwehadtheinitialconditiondydx=xyx2+y2y(1)=0thenthefunctionf(x,y)=xyx2+y2iscontinuousinaneighborhoodofthepointy=0,x=5,sotheequationwouldhaveauniquesolution.FirstOrderLinearInhomogeneousEquations:NextIwanttoconsiderequationsofthefollowingform:dydx+P(x)y=Q(x)whichis,ifyou’llrecall,afirstorderlinearinhomogeneousequa-tion.AslongasP(x)andQ(x)arecontinuousfunctionsthenP(x)y+Q(x)isacontinuousfunction,andhencetheaboveequationhasauniquesolution.Itturnsoutthattheaboveequationcanalwaysbesolvedbywhatiscalledthe“integratingFactor”method,whichiskindofaspecialcaseofwhatisknownas“variationofparameters”.Tobeginwithwe’llconsiderthespecialcaseQ(x)=0(thisiscalledthe“linear”or“linearhomogeneous”problem.Inthiscasetheequationisseparabledydx+P(x)y=0withsolutiony(x)=y0eRP(x)dxFigure1.8:AnIllustrationofthecon-tractionmappingprinciple.Havingsomeideaofhowtheexistenceanduniquenesstheoremisprovenwewillnowapplyittosomeexamples.Example1.4.4.Considertheequationdydx=xyx2+y2y(0)=0Thefunctionf(x,y)=xyx2+y2isNOTcontinuousinaneighborhoodoftheorigin(Why?),sothetheoremdoesnotapply.Ontheotherhandifwehadtheinitialconditiondydx=xyx2+y2y(1)=0thenthefunctionf(x,y)=xyx2+y2iscontinuousinaneighborhoodofthepointy=0,x=5,sotheequationwouldhaveauniquesolution.FirstOrderLinearInhomogeneousEquations:NextIwanttoconsiderequationsofthefollowingform:dydx+P(x)y=Q(x)whichis,ifyou’llrecall,afirstorderlinearinhomogeneousequa-tion.AslongasP(x)andQ(x)arecontinuousfunctionsthenP(x)y+Q(x)isacontinuousfunction,andhencetheaboveequationhasauniquesolution.Itturnsoutthattheaboveequationcanalwaysbesolvedbywhatiscalledthe“integratingFactor”method,whichiskindofaspecialcaseofwhatisknownas“variationofparameters”.Tobeginwithwe’llconsiderthespecialcaseQ(x)=0(thisiscalledthe“linear”or“linearhomogeneous”problem.Inthiscasetheequationisseparabledydx+P(x)y=0withsolutiony(x)=y0eRP(x)dxFigure1.8:AnIllustrationofthecon-tractionmappingprinciple.Havingsomeideaofhowtheexistenceanduniquenesstheoremisprovenwewillnowapplyittosomeexamples.Example1.4.4.Considertheequationdydx=xyx2+y2y(0)=0Thefunctionf(x,y)=xyx2+y2isNOTcontinuousinaneighborhoodoftheorigin(Why?),sothetheoremdoesnotapply.Ontheotherhandifwehadtheinitialconditiondydx=xyx2+y2y(1)=0thenthefunctionf(x,y)=xyx2+y2iscontinuousinaneighborhoodofthepointy=0,x=5,sotheequationwouldhaveauniquesolution.FirstOrderLinearInhomogeneousEquations:NextIwanttoconsiderequationsofthefollowingform:dydx+P(x)y=Q(x)whichis,ifyou’llrecall,afirstorderlinearinhomogeneousequa-tion.AslongasP(x)andQ(x)arecontinuousfunctionsthenP(x)y+Q(x)isacontinuousfunction,andhencetheaboveequationhasauniquesolution.Itturnsoutthattheaboveequationcanalwaysbesolvedbywhatiscalledthe“integratingFactor”method,whichiskindofaspecialcaseofwhatisknownas“variationofparameters”.Tobeginwithwe’llconsiderthespecialcaseQ(x)=0(thisiscalledthe“linear”or“linearhomogeneous”problem.Inthiscasetheequationisseparabledydx+P(x)y=0withsolutiony(x)=y0eRP(x)dxFigure1.8:AnIllustrationofthecon-tractionmappingprinciple.Havingsomeideaofhowtheexistenceanduniquenesstheoremisprovenwewillnowapplyittosomeexamples.Example1.4.4.Considertheequationdydx=xyx2+y2y(0)=0Thefunctionf(x,y)=xyx2+y2isNOTcontinuousinaneighborhoodoftheorigin(Why?),sothetheoremdoesnotapply.Ontheotherhandifwehadtheinitialconditiondydx=xyx2+y2y(1)=0thenthefunctionf(x,y)=xyx2+y2iscontinuousinaneighborhoodofthepointy=0,x=5,sotheequationwouldhaveauniquesolution.FirstOrderLinearInhomogeneousEquations:NextIwanttoconsiderequationsofthefollowingform:dydx+P(x)y=Q(x)whichis,ifyou’llrecall,afirstorderlinearinhomogeneousequa-tion.AslongasP(x)andQ(x)arecontinuousfunctionsthenP(x)y+Q(x)isacontinuousfunction,andhencetheaboveequationhasauniquesolution.Itturnsoutthattheaboveequationcanalwaysbesolvedbywhatiscalledthe“integratingFactor”method,whichiskindofaspecialcaseofwhatisknownas“variationofparameters”.Tobeginwithwe’llconsiderthespecialcaseQ(x)=0(thisiscalledthe“linear”or“linearhomogeneous”problem.Inthiscasetheequationisseparabledydx+P(x)y=0withsolutiony(x)=y0eRP(x)dxFigure1.8:AnIllustrationofthecon-tractionmappingprinciple.Havingsomeideaofhowtheexistenceanduniquenesstheoremisprovenwewillnowapplyittosomeexamples.Example1.4.4.Considertheequationdydx=xyx2+y2y(0)=0Thefunctionf(x,y)=xyx2+y2isNOTcontinuousinaneighborhoodoftheorigin(Why?),sothetheoremdoesnotapply.Ontheotherhandifwehadtheinitialconditiondydx=xyx2+y2y(1)=0thenthefunctionf(x,y)=xyx2+y2iscontinuousinaneighborhoodofthepointy=0,x=5,sotheequationwouldhaveauniquesolution.FirstOrderLinearInhomogeneousEquations:NextIwanttoconsiderequationsofthefollowingform:dydx+P(x)y=Q(x)whichis,ifyou’llrecall,afirstorderlinearinhomogeneousequa-tion.AslongasP(x)andQ(x)arecontinuousfunctionsthenP(x)y+Q(x)isacontinuousfunction,andhencetheaboveequationhasauniquesolution.Itturnsoutthattheaboveequationcanalwaysbesolvedbywhatiscalledthe“integratingFactor”method,whichiskindofaspecialcaseofwhatisknownas“variationofparameters”.Tobeginwithwe’llconsiderthespecialcaseQ(x)=0(thisiscalledthe“linear”or“linearhomogeneous”problem.Inthiscasetheequationisseparabledydx+P(x)y=0withsolutiony(x)=y0eRP(x)dxFigure1.8:AnIllustrationofthecon-tractionmappingprinciple.Havingsomeideaofhowtheexistenceanduniquenesstheoremisprovenwewillnowapplyittosomeexamples.Example1.4.4.Considertheequationdydx=xyx2+y2y(0)=0Thefunctionf(x,y)=xyx2+y2isNOTcontinuousinaneighborhoodoftheorigin(Why?),sothetheoremdoesnotapply.Ontheotherhandifwehadtheinitialconditiondydx=xyx2+y2y(1)=0thenthefunctionf(x,y)=xyx2+y2iscontinuousinaneighborhoodofthepointy=0,x=1,sotheequationwouldhaveauniquesolution.Figure1.8:AnIllustrationofthecon-tractionmappingprinciple.Havingsomeideaofhowtheexistenceanduniquenesstheoremisprovenwewillnowapplyittosomeexamples.Example1.4.4.Considertheequationdydx=xyx2+y2y(0)=0Thefunctionf(x,y)=xyx2+y2isNOTcontinuousinaneighborhoodoftheorigin(Why?),sothetheoremdoesnotapply.Ontheotherhandifwehadtheinitialconditiondydx=xyx2+y2y(1)=0thenthefunctionf(x,y)=xyx2+y2iscontinuousinaneighborhoodofthepointy=0,x=1,sotheequationwouldhaveauniquesolution.Figure1.8:AnIllustrationofthecon-tractionmappingprinciple.Havingsomeideaofhowtheexistenceanduniquenesstheoremisprovenwewillnowapplyittosomeexamples.Example1.4.4.Considertheequationdydx=xyx2+y2y(0)=0Thefunctionf(x,y)=xyx2+y2isNOTcontinuousinaneighborhoodoftheorigin(Why?),sothetheoremdoesnotapply.Ontheotherhandifwehadtheinitialconditiondydx=xyx2+y2y(1)=0thenthefunctionf(x,y)=xyx2+y2iscontinuousinaneighborhoodofthepointy=0,x=1,sotheequationwouldhaveauniquesolution.Figure1.8:AnIllustrationofthecon-tractionmappingprinciple.Havingsomeideaofhowtheexistenceanduniquenesstheoremisprovenwewillnowapplyittosomeexamples.Example1.4.4.Considertheequationdydx=xyx2+y2y(0)=0Thefunctionf(x,y)=xyx2+y2isNOTcontinuousinaneighborhoodoftheorigin(Why?),sothetheoremdoesnotapply.Ontheotherhandifwehadtheinitialconditiondydx=xyx2+y2y(1)=0thenthefunctionf(x,y)=xyx2+y2iscontinuousinaneighborhoodofthepointy=0,x=1,sotheequationwouldhaveauniquesolution.Figure1.8:AnIllustrationofthecon-tractionmappingprinciple.Havingsomeideaofhowtheexistenceanduniquenesstheoremisprovenwewillnowapplyittosomeexamples.Example1.4.4.Considertheequationdydx=xyx2+y2y(0)=0Thefunctionf(x,y)=xyx2+y2isNOTcontinuousinaneighborhoodoftheorigin(Why?),sothetheoremdoesnotapply.Ontheotherhandifwehadtheinitialconditiondydx=xyx2+y2y(1)=0thenthefunctionf(x,y)=xyx2+y2iscontinuousinaneighborhoodofthepointy=0,x=1,sotheequationwouldhaveauniquesolution.Figure1.8:Anillustrationofhecontractionmappingprinciple.Havingsomeideaofhowtheexistenceanduniquenesstheoremisprovenwewillnowapplyittosomeexamples.Example1.4.4.Considertheequationdydt=tyt2+y2y(0)=0Thefunctionf(t,y)=tyt2+y2isNOTcontinuousinaneighborhoodoftheorigin(Why?),sothetheoremdoesnotapply.Ontheotherhandifwehadtheinitialconditiondydt=tyt2+y2y(1)=0thenthefunctionf(t,y)=tyt2+y2iscontinuousinaneighborhoodofthepointy=0,t=1,sotheequationwouldhaveauniquesolution.
35
35
35
35
35
35
35
35

Suppose we have a set W, and for each element f 2 W we have a no-

tion of the size of each element k f k. This is what is called a normed
space. Now suppose that we have a map T from W to itself such that
for any two elements f , g 2 W we have that k T ( f ) T ( g)k  rk f gk
Figure 1.8: An illustration of the
Figure 1.8: An illustration of the
contraction mapping principle.
Figure 1.8: An illustration of the
contraction mapping principle.

contraction mapping principle.

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +
299t9 4t10 184t11 t12 4t13 t15
ciple. The contraction mapping principle is the simplest example of
what mathematicians call a fixed point theorem. Leaving aside tech-
for r < 1. In other words shrinks the distances between points in the
set by a fixed contraction factor r that is strictly less than one. Then
there is a unique element h that is fixed by the map: T (h) = h, and

nical details the contraction mapping principle says the following.

4 5 8 29
+ + + + + +
Suppose we have a set Ω, and for each element f ∈ Ω we have a no-
tion of the size of each element ∥ f ∥. This is what is called a normed
Figure 1.8: An illustration of the
contraction mapping principle.

+
11340 525 51975 2268 12285 59535
space. Now suppose that we have a map T from Ω to itself such that
for any two elements f , g ∈ Ω we have that ∥ T ( f ) − T ( g)∥ ≤ ρ∥ f − g∥
for ρ < 1. In other words T shrinks the distances between points in
the set by a fixed contraction factor ρ that is strictly less than one.
Then there is a unique element h that is fixed by the map: T (h) = h,

The pattern for the coefficients in this example is not so clear.

and this fixed point can be found by taking any starting point and
iterating the map: lim N →∞ T N ( f ) = h. This theorem is illustrated by

47t7 41t8
3 6 15 90 315 630
the margin figure. The proof of the Picard existence theorem amounts
to showing that the Picard iterates form (for small enough neigh-
borhood of (t0 , y0 ) a contraction mapping, and thus converge to a
unique fixed point.

We will not present the full proof here, but the basic principle
proven we will now apply it to some examples.
Having some idea of how the existence and uniqueness theorem is

Example 1.4.4. Consider the equation

dy

behind the proof is something called the contraction mapping prin-

dt
= 2
t + y2
ty
y (0) = 0

The function f (t, y) = t2 +y2 is NOT continuous in a neighborhood of the

ty

origin (Why?), so the theorem does not apply. On the other hand if we had

ciple. The contraction mapping principle is the simplest example of

the initial condition
dy
dt
= 2
t + y2
ty
y (1) = 0

299t9 4t10 184t11 t12 4t13 t15

then the function f (t, y) = t2 +y2 is continuous in a neighborhood of the
ty

what mathematicians call a fixed point theorem. Leaving aside tech-

point y = 0, t = 1, so the equation would have a unique solution.

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +
nical details the contraction mapping principle says the following.
Suppose we have a set Ω, and for each element f ∈ Ω we have a no-

+ + + + + +
Figure 1.8: An illustration of the
tion of the size of each element ∥ f ∥. This is what is called a normed

4 5 8 29
contraction mapping principle.

11340 525 51975 2268 12285 59535

space. Now suppose that we have a map T from Ω to itself such that

+
for any two elements f , g ∈ Ω we have that ∥ T ( f ) − T ( g)∥ ≤ ρ∥ f − g∥
for ρ < 1. In other words T shrinks the distances between points in
the set by a fixed contraction factor ρ that is strictly less than one.
Then there is a unique element h that is fixed by the map: T (h) = h,
and this fixed point can be found by taking any starting point and

The pattern for the coefficients in this example is not so clear.

iterating the map: lim N →∞ T N ( f ) = h. This theorem is illustrated by
the margin figure. The proof of the Picard existence theorem amounts

3 6 15 90 315 630
to showing that the Picard iterates form (for small enough neigh-
borhood of (t0 , y0 ) a contraction mapping, and thus converge to a
unique fixed point.
Having some idea of how the existence and uniqueness theorem is

We will not present the full proof here, but the basic principle
proven we will now apply it to some examples.

Example 1.4.4. Consider the equation

dy ty
= 2 y (0) = 0
dt t + y2

behind the proof is something called the contraction mapping prin- ty

The function f (t, y) = t2 +y2 is NOT continuous in a neighborhood of the

4 5 8 29
origin (Why?), so the theorem does not apply. On the other hand if we had
the initial condition

ciple. The contraction mapping principle is the simplest example of

dy ty
= 2 y (1) = 0
dt t + y2

299t9 4t10 184t11 t12 4t13 t15

then the function f (t, y) = t2 +y2 is continuous in a neighborhood of the
ty

point y = 0, t = 1, so the equation would have a unique solution.

what mathematicians call a fixed point theorem. Leaving aside tech-

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +
nical details the contraction mapping principle says the following.
Suppose we have a set Ω, and for each element f ∈ Ω we have a no-

+ + + + + +
Figure 1.8: An illustration of the
tion of the size of each element ∥ f ∥. This is what is called a normed
contraction mapping principle.
space. Now suppose that we have a map T from Ω to itself such that

11340 525 51975 2268 12285 59535

+
for any two elements f , g ∈ Ω we have that ∥ T ( f ) − T ( g)∥ ≤ ρ∥ f − g∥
for ρ < 1. In other words T shrinks the distances between points in
the set by a fixed contraction factor ρ that is strictly less than one.
Then there is a unique element h that is fixed by the map: T (h) = h,
and this fixed point can be found by taking any starting point and

The pattern for the coefficients in this example is not so clear.

iterating the map: lim N →∞ T N ( f ) = h. This theorem is illustrated by
the margin figure. The proof of the Picard existence theorem amounts

3 6 15 90 315 630
to showing that the Picard iterates form (for small enough neigh-
borhood of (t0 , y0 ) a contraction mapping, and thus converge to a
unique fixed point.
Having some idea of how the existence and uniqueness theorem is

We will not present the full proof here, but the basic principle
proven we will now apply it to some examples.

Example 1.4.4. Consider the equation

dy ty
= 2 y (0) = 0
dt t + y2

behind the proof is something called the contraction mapping prin- ty

The function f (t, y) = t2 +y2 is NOT continuous in a neighborhood of the
origin (Why?), so the theorem does not apply. On the other hand if we had
the initial condition

ciple. The contraction mapping principle is the simplest example of

dy ty
= 2 y (1) = 0

y (3) ( t ) = 1 + t + t 2 + t 3 + t 6 + t 5 + t 6 +
dt t + y2

299t9 4t10 184t11 t12 4t13 t15

ty
then the function f (t, y) = t2 +y2 is continuous in a neighborhood of the
point y = 0, t = 1, so the equation would have a unique solution.

what mathematicians call a fixed point theorem. Leaving aside tech-

nical details the contraction mapping principle says the following.
Suppose we have a set Ω, and for each element f ∈ Ω we have a no-

+ + + + + +
Figure 1.8: An illustration of the
tion of the size of each element ∥ f ∥. This is what is called a normed
contraction mapping principle.

+
space. Now suppose that we have a map T from Ω to itself such that

11340 525 51975 2268 12285 59535 for any two elements f , g ∈ Ω we have that ∥ T ( f ) − T ( g)∥ ≤ ρ∥ f − g∥
for ρ < 1. In other words T shrinks the distances between points in
the set by a fixed contraction factor ρ that is strictly less than one.
Then there is a unique element h that is fixed by the map: T (h) = h,
and this fixed point can be found by taking any starting point and

The pattern for the coefficients in this example is not so clear.

iterating the map: lim N →∞ T N ( f ) = h. This theorem is illustrated by

3 6 15 90 315 630
the margin figure. The proof of the Picard existence theorem amounts
to showing that the Picard iterates form, for small enough neigh-
borhood of (t0 , y0 ) a contraction mapping, and thus converge to a
unique fixed point.
Having some idea of how the existence and uniqueness theorem is
proven we will now apply it to some examples.

We will not present the full proof here, but the basic principle Example 1.4.4. Consider the equation

dy
= 2
ty
y (0) = 0
dt t + y2

behind the proof is something called the contraction mapping prin- ty

The function f (t, y) = t2 +y2 is NOT continuous in a neighborhood of the
origin (Why?), so the theorem does not apply. On the other hand if we had
the initial condition

ciple. The contraction mapping principle is the simplest example of

dy ty
= 2 y (1) = 0
t + y2

299t9 4t10 184t11 t12 4t13 t15

dt
ty
then the function f (t, y) = t2 +y2 is continuous in a neighborhood of the
point y = 0, t = 1, so the equation would have a unique solution.

what mathematicians call a fixed point theorem. Leaving aside tech-

nical details the contraction mapping principle says the following.
Suppose we have a set Ω, and for each element f ∈ Ω we have a no-

+ + + + + + tion of the size of each element ∥ f ∥. This is what is called a normed

Figure 1.8: An illustration of the
contraction mapping principle.

11340 525 51975 2268 12285 59535

space. Now suppose that we have a map T from Ω to itself such that
for any two elements f , g ∈ Ω we have that ∥ T ( f ) − T ( g)∥ ≤ ρ∥ f − g∥
for ρ < 1. In other words T shrinks the distances between points in
the set by a fixed contraction factor ρ that is strictly less than one.
Then there is a unique element h that is fixed by the map: T (h) = h,
and this fixed point can be found by taking any starting point and

The pattern for the coefficients in this example is not so clear.

iterating the map: lim N →∞ T N ( f ) = h. This theorem is illustrated by
the margin figure. The proof of the Picard existence theorem amounts
to showing that the Picard iterates form, for small enough neigh-
borhood of (t0 , y0 ) a contraction mapping, and thus converge to a
unique fixed point.
Having some idea of how the existence and uniqueness theorem is
proven we will now apply it to some examples.

We will not present the full proof here, but the basic principle Example 1.4.4. Consider the equation

dy
= 2
ty
y (0) = 0
dt t + y2

behind the proof is something called the contraction mapping prin- ty

The function f (t, y) = t2 +y2 is NOT continuous in a neighborhood of the
origin (Why?), so the theorem does not apply. On the other hand if we had
the initial condition

ciple. The contraction mapping principle is the simplest example of

dy ty
= 2 y (1) = 0
dt t + y2
ty
then the function f (t, y) = t2 +y2 is continuous in a neighborhood of the
point y = 0, t = 1, so the equation would have a unique solution.

what mathematicians call a fixed point theorem. Leaving aside tech-

nical details the contraction mapping principle says the following.
Suppose we have a set Ω, and for each element f ∈ Ω we have a no-
Figure 1.8: An illustration of the
tion of the size of each element ∥ f ∥. This is what is called a normed
contraction mapping principle.
space. Now suppose that we have a map T from Ω to itself such that
for any two elements f , g ∈ Ω we have that ∥ T ( f ) − T ( g)∥ ≤ ρ∥ f − g∥
for ρ < 1. In other words T shrinks the distances between points in
the set by a fixed contraction factor ρ that is strictly less than one.
Then there is a unique element h that is fixed by the map: T (h) = h,
 first order differential equations 35

and this fixed point can be found by taking any starting point and
iterating the map: lim N →∞ T N ( f ) = h. This theorem is illustrated by
the margin figure. The proof of the Picard existence theorem amounts
to showing that the Picard iterates form, for small enough neigh-
borhood of (t0 , y0 ), a contraction mapping, and thus converge to a
unique fixed point.
Having some idea of how the existence and uniqueness theorem is
proven we will now apply it to some examples.

Example 1.4.4. Consider the equation

dy ty
= 2 y (0) = 0
dt t + y2
ty
The function f (t, y) = t2 +y2 is NOT continuous in a neighborhood of the
origin (Why?), so the theorem does not apply. On the other hand if we had
the initial condition
dy ty
= 2 y (1) = 0
dt t + y2
ty ∂f
then the function f (t, y) = t2 +y2 and the partial derivative ∂y are continu-
ous in a neighborhood of the point y = 0, t = 1, so the equation would have
a unique solution.

Exercise 1.4.1
Determine if the Theorem of Existence and Uniqueness guarantees
a solution and a unique solution to the following initial value prob-
lems.
y2 y2
a) y′ + t2
= ty, y (0) = 1 b) y′ + t2
= ty, y (3) = 1
y+t y+t
c) y′ = y−t , y (1) = 2 d) y′ = y−t , y (2) = 2
y1/3 ty
e) y′ = t2
, y (1) = 0 f) y′ + cos(y)
= 0, y (0) = 0

1.5 First Order Linear Inhomogeneous Equations:

Most of the equations considered in these notes will be linear equa-

tions, since linear equations often arise in practice, and are typically
much easier to solve than nonlinear equations. A first order linear
inhomogeneous initial value problem takes the form

dy
+ P(t)y = Q(t) y ( t0 ) = y0 .
dt
36 differential equations

If P(t) and Q(t) are continuous functions of t then f (y, t) = − P(t)y +

∂f
Q(t) and ∂y = − P(t) are continuous functions and hence Picard
existence and uniqueness theorem guarantees a unique solution.
For linear equations, due to the simple dependence on y , one can
do a little better than the Picard theorem. In fact one can show the
following.

Theorem 1.5.1. Consider the first order linear inhomogeneous initial value
problem
dy
+ P(t)y = Q(t) y ( t0 ) = y0 .
dt
If P(t) and Q(t) are continuous on a interval ( a, b) containing t0 then the
initial value problem has a unique solution on ( a, b).

This is stronger than the Picard theorem because it guarantees that

the solution exists and is unique in the entire interval over which
P(t), Q(t) are continuous. The Picard theorem generally only guaran-
tees existence and uniqueness in some small interval [t0 − δ, t0 + δ].
It turns out that the above equation can always be solved by what
is called the integrating factor method, which is a special case of
what is known as variation of parameters.
To begin with we’ll consider the special case Q(t) = 0 (this is
called the “linear homogeneous” problem.)

dy
+ P(t)y = 0
dt
In this case the equation is separable with solution
P(t)dt
R
y ( t ) = y0 e −

P(t)dt .
R
We’re going to introduce a new variable w(t) = y(t)e Note
that
dw dy
= ( + P(t)y)e P(t)dt
R

dt dt
Taking the equation

dy
+ P(t)y = Q(t)
dt
P(t)dt
R
and multiplying through by e gives

dy R
P(t)dt
R
P(t)dt
( + P(t)y)e = Q(t)e
| dt {z }
dw
dt

dw R
P(t)dt
= Q(t)e
dt
The latter can be integrated up, since it is an exact derivative. This
Remark 1.5.1. Here the mathematics
terminology differs somewhat from the
engineering terminology. In engineering
it is customary to split the solution into
two parts, the zero input and zero state
functions. The zero state solution is a par-
ticular solution that satisfies zero boundary
conditions at the initial point t0 . This can
be achieved by taking t0 as the lower limit
in the integral. The zero state solution is
 first order differential equations 37

gives Z
P(t)dt
R
w= Q(t)e +c
P(t)dt
R
since w(t) = y(t)e we can solve for y and get
Z
P(t)dt P(t)dt
+ |ce− {zP(t)dt}
R R R
y = e− Q(t)e (1.13)
| {z } homogeneous
particular

This is called the “integrating factor” method. The basic algorithm is

this: given the equation

dy
+ P(t)y = Q(t)
dt
we should
P(t)dt .
R
• Calculate the integrating factor µ(t) = e

• Multiply through by µ(t), recognize the lefthand side as an exact

derivative

• Integrate up.

There are a couple of things to note about this solution. These

are not so important now, but they are going to be important later
when we talk about the general solution to a linear inhomogeneous
equation.

• The solution splits into a particular solution plus a solution to the

homogeneous problem.

• The particular solution can be realized as an integral of the right-

hand side times the homogeneous solution. This is a special case
of what we will later see as the variation of parameters formula.

Another way to say the first point is as follows: the difference

between any two solutions to the linear inhomogeneous problem is
a solution to the homogeneous problem. To see this note that if y1
solves
dy1
+ P ( t ) y1 = Q ( t )
dt
and y2 solves
dy2
+ P ( t ) y2 = Q ( t )
dt
then y1 − y2 solves

d ( y1 − y2 )
+ P(t)(y1 − y2 ) = 0
dt
so the difference between any two solutions of the inhomogeneous
problem solves the homogeneous problem.
38 differential equations

Example 1.5.1. Find the general solution to the equation

y′ + ty = t3

The integrating factor in this case is

tdt t2
R
µ(t) = e =e2

and we can finish the problem using Equation (1.13)

We can do this another way, by guessing a particular solution. Notice
that if y is a polynomial then y′ and ty are both polynomials. Assuming that
y = at2 + bt + c then we have

y′ + ty = at3 + bt2 + (c + 2a)t + b

for this to equal t3 one should pick a = 1, b = 0, c = −2. This gives a

particular solution y = t2 − 2. The above calculation implies that ANY
solution is given by the sum of a particular solution plus the solution to the
homogeneous problem. The solution to the homogeneous problem is
t2
y = y0 e − 2

so the general solution is

t2
y = y0 e − 2 + t2 − 2

Example 1.5.2. Find the general solution to

dy
+ tan(t)y = cos(t)
dt
The integrating factor is, in this case
sin(t)
dt 1
R
µ(t) = e cos(t) = e− ln(cos(t)) =
cos(t)

Multiplying through gives

dy
sec(t) + sec(t) tan(t)y = 1
dt
dy d
recognizing the righthand side as sec(t) dt + sec(t) tan(t)y = dt (sec( t ) y )
we have
d
(sec ty) = 1
dt
sec ty = t + C
y = t cos t + C cos(t)

Example 1.5.3 (A Mixture Problem). A tank contains 10 kg of salt

dissolved in 100 liters of water. Water containing 0.001 kg of salt per
 first order differential equations 39

liter flows into the tank at 10 L min−1 and well mixed water flows out at
10 L min−1 . What is the differential equation governing the amount of salt
in the tank at time t? Solve it.
Let Y (t) represent the total amount of salt (in kilograms) in the wa-
ter. The total amount of water in the tank is a constant 100 liters. The
amount of salt flowing into the tank is .01 kg/min. The amount flow-
ing out depends on the concentration of the salt. If the amount of salt is
Y (t) then the concentration is Y (t)/100, and the amount flowing out is
10Y (t)/100 = Y (t)/10. So the differential equation is
dY Y
= .01 −
dt 10
1 t
10 dt
R
The integrating factor is e = e 10 giving
d t t
(Ye 10 ) = .01e 10 (1.14)
dt
t t
(Ye 10 ) = .1e 10 + c (1.15)
t
− 10
Y = .1 + ce (1.16)
(1.17)

Initially we have Y (0) = 10 (initially there are 10 kg of salt) thus

t
Y = .1 + 9.9e− 10

and the total amount of salt is asymptotic to 0.1 kg. Can you see why this
must be so? Can you show it without explicitly solving the equation?
A more interesting problem occurs when the water flows into the
tank at a different rate than it flows out. For instance:

Exercise 1.5.1
Suppose now that the water flows out of the tank at 5 L min−1 . What
is the equation governing the amount of salt in the tank? Solve it! At
what time t is the salt content in the tank at a minimum?

Exercise 1.5.2
Use the integrating factor method to solve the following first order
linear equations. If an initial condition is given then solve for the
constant, otherwise find the general solution.
dy
a) dt + 2y = 1 + t y (0) = 0
dy 2 sin(t)
b) dt + t y = t2
dy
c) t dt + y = sin(t) y(π ) = 0
dy 2t cos(t)
d) dt + t2 +1 y = t2 +1 y (0) = 1
dy sin(t) sin(t)
e) dt + cos(t) y = cos3 (t) y (0) = 0
40 differential equations

1.6 Exact and other special first order equations

There are a few special types of equations that often come up in

applications. One of these, which frequently arises in applications, is
the exact equation:

dy
N ( x, y) + M( x, y) = 0 (1.18)
dx
where M, N are related in a particular way. In order to be exact there
must be some function ψ( x, y) (called the potential) such that

∂ψ
= N ( x, y) (1.19)
∂y
∂ψ
= M( x, y). (1.20)
∂x
Exact equations are connected with vector calculus and the prob-
lem of determining if a vector field is a gradient, so you may have
previously encountered this idea in vector calculus or physics. To be slightly more careful the equality
An obvious first question is how to recognize an exact equation. of mixed partials is guaranteed by
Clairaut’s theorem.
For general functions N ( x, y) and M ( x, y) there is not always such
a potential function ψ( x, y). A fact to recall from vector calculus is Theorem 1.6.1 (Clairaut). Suppose that
ϕ( x, y) is defined on D, an open subset of
∂2 ψ ∂2 ψ
the equality of mixed partials: ∂x∂y = ∂y∂x . Differentiating the above R2 . Further suppose that the second order
equations with respect to x and y respectively gives mixed partials ϕxy and ϕyx exist and are
continuous on R2 . Then the mixed partials
ϕxy and ϕyx must be equal on D.
∂2 ψ ∂N
= (1.21)
∂x∂y ∂x
∂2 ψ ∂M
= (1.22)
∂y∂x ∂y

So a necessary condition is that Nx = My . It turns out that this is also

a sufficient condition.1 If we can find such a function ψ then by the 1
On simply connected domains.
chain rule the above equation is equivalent to From the point of view of vector calcu-
lus we can define the potential function
by the line integral
dy Z
ψy ( x, y) + ψx ( x, y) = 0 (1.23) ψ( x, y) = N ( x, y)dy + M ( x, y)dx
dx γ

d where γ is any curve beginning at an

(ψ( x, y)) = 0 (1.24)
dx arbitrary point (say (0, 0)) and ending
at ( x, y). Because the vector field is
ψ( x, y) = c (1.25) conservative the line integral does not
depend on the path. In the example we
This makes geometric sense – Equation (1.18) is equivalent to are taking a piecewise path consisting
of a line segment from (0, 0) to (0, y)
followed by a line segment from (0, y)
∇ψ · (dx, dy) = 0
to ( x, y).
 first order differential equations 41

so the directional derivative of ψ along the curve is zero. This is

equivalent to saying that ψ is constant along the curve ( x, y( x )), or
that the curves ( x, y( x )) are level sets of ψ.
Given that we can recognize an exact equation how do we find
ψ? The inverse to partial differentiation is partial integration. This is
easiest to illustrate with an example

Example 1.6.1. Solve the equation

(y3 + 3x2 y)dy + (3y2 x )dx = 0

We’d like to find a function ψ( x, y) such that

∂ψ
= y3 + 3x2 y (1.26)
∂y
∂ψ
= 3y2 x (1.27)
∂x
Well, first we have to check that this is possible. The necessary condition
is that the mixed partials ought to be equal. Differentiating gives

∂N ∂  3 
= y + 3x2 y = 6xy
∂x ∂x
∂M ∂  
= 3xy2 = 6xy
∂y ∂y

Thus we know that there exists such a ψ. Now we have to find it. We can
start by integrating up

y4 3x2 y2
Z
ψ= y3 + 3x2 ydy = + + f ( x ).
4 2
Here we are integrating with respect to y, so the undetermined constant is
a function of x. This makes sense: if we take the partial derivative of f ( x )
with respect to y we will get zero. We still don’t know f ( x ), but we can use
the other equation
ψx = 3xy2 + f ′ ( x )
comparing this to the above gives f ′ ( x ) = 0 so f ( x ) = c. Thus the solution
is given by
y4 3x2 y2
ψ( x, y) = + =c
4 2
Note that this solution is IMPLICIT. You can actually solve this for y as
a function of x (The above is quadratic in y) but you can also leave it in this
form.

Another example is the following:

Example 1.6.2.
dy −2xy
= 2
dx x + y2
42 differential equations

This is similar to an example we looked at previously, as an illustration of

the existence/uniqueness theorem. We can rewrite this as

( x2 + y2 )dy + (2xy)dx = 0

This equation is exact (check this!) so we can find a solution ψ( x, y) = c.

We compute ψ( x, y) as outlined above

ψx = 2xy
ψ = x2 y + g(y)

To compute the undetermined function g(y) we use the other equation

ψ = x2 y + g(y)
ψy = x2 + g′ (y) = x2 + y.2
y3
From this we can conclude that g′ (y) = y2 or g(y) = 3 . Thus the solution
is
y3
+ x2 y = c
3

Exercise 1.6.1
Classify each of the following equations as exact or not exact. If they
are exact find the solution in the form ψ( x, y) = c. You are given an
initial condition so you should be able to determine c.
dy
a) (5x + 2y + 2) dx + (2x + 5y + 1) = 0; y (1) = 3
dy
b) (3x + 7y + 2) dx + (2x + 5y + 1) = 0; y (2) = 8
dy
c) (−3x2 + 10xy − 3y2 − e x−y ) dx + (3x2 − 6xy + 5y2 + e x−y ) =
0; y (0) = 1
dy
d) (3xy − 3y2 ) dx + (3x2 + 3y2 ) = 0; y (2) = −1
dy
e) (6xy − 3y2 ) dx + (3x2 + 3y2 ) = 0; y(−1) = −2
dy
f) (2x + y) dx + (2y − x ) = 0; y(−1) = 0
dy
g) (3y2 + xe xy ) dx + (2x + ye xy ) = 0; y (0) = 1

A second important class of exactly solvable first order equa-

tions are the separable equations. These are equations that can be
written in the form
dy
= f (t) g(y)
dt
Such equations can be solved by separating the dependent and
independent variables. If we manipulate the equation to get the inde-
pendent variable on one side and the dependent variable on the other
we find that
dy
= f (t)dt.
g(y)
 first order differential equations 43

This can be integrated to give

dy
Z Z
= f (t)dt.
g(y)

Example 1.6.3 (Logistic Equation). The logistic equation

dy
= y (1 − y )
dt

is a common model for growth of a population y(t) in a situation where

there is a maximum sustainable population, in this case chosen to be 1 .
This equation might model a situation in which there is a fixed amount of
resource necessary to sustain the population. In the case where the initial
population y is small the equation behaves like

dy
= y − y2 ≈ y.
dt

This is the standard exponential growth model. As y approaches 1, the max-

imum sustainable population, the growth rate goes to zero. If the population
is above 1 then the population will actually decrease towards the maximum
sustainable population.
We can solve this explicitly by using the fact that the equation is separa-
ble. With a little algebra we find that

dy
= y (1 − y )
dt
dy
= dt
y (1 − y )
1 1
 
+ dy = dt
y 1−y

where the last step follows from the method of partial fractions, from calcu-
lus. Assuming that the population y is between 0 and 1 we can integrate
this up to find

ln y − ln(1 − y) = t + c
y
= et+c
1−y
y = et+c − et+c y
et+c
y=
1 + et+c

The side margin shows the slope field and some solutions to the logistic
equation for different values of c but they all show the same characteristic
shape: populations with positive initial conditions all tend to the asymptote
y = 1, the maximum sustainable population.
Figure 1.9: The slope field and
some solution curves for the
dy
logistic equation dt = (1 − y)y
44 differential equations

Exercise 1.6.2
Find a general solution for each equation.
cos(t)
a) y′ = y2
b) yy′ = t + ty2
1
c) t2 y′ = cos(y)
d) t(y2 + 1)y′ = y

There are a number of other types of equations that can be

integrated by various tricks. One class of equations is called scale
invariant and is of the form,

dy y
= F( )
dt t
y
These can be solved by the substitution v = t or y = tv. This leads to

dy dv
= t + v = F (v)
dt dt
which is a separable equation.
An example of this type is

dy ty
= 2
dt t + y2

Making the substitution y = tv gives

dv t2 v v
t +v = 2 2 2
=
dt t +t v 1 + v2
Which is separable.

Exercise 1.6.3
Find a general solution for each equation.
y2 t y
a) yy′ = 2t + t b) y′ = 2( y + t )
+ t
t 2t−y
c) ty′ = y + y d) y′ = t+y
sin( t )

Sometimes the appearance of a particular linear pattern

in the equation may suggest different a substitution.

Example 1.6.4. Consider the equation

dy t2 + 2ty + y2
=−
dt 1 + ( t + y )2

Well, it is clear that the righthand side is only a function of the sum t + y.
This suggests making the substitution v = y + t. The equation above
 first order differential equations 45

becomes
dv v2
−1 = − (1.28)
dt 1 + v2
dv 1
= (1.29)
dt 1 + v2
2
(1 + v )dv = dt (1.30)
Integrating this up gives the equation
v3
v+ = t+c (1.31)
3
1
y + t + ( y + t )3 = t + c (1.32)
3

Exercise 1.6.4
Find a general solution for each equation.
1
a) y′ = ey+2t − 2 b) y′ = (3y − t)2 + 3

Another example occurs for equations of second order

when either the dependent variable y or the independent variable t is
missing. These equations are reducible.
Example 1.6.5. Consider the problem
y′′ + y′ /t = t5
This is a second order equation, but there is no y term. Thus we can make
the substitution v = y′ . This gives
v′ + v/t = t5
This is a first order linear (inhomogeneous) equation, and thus can be solved
by multiplying through by an appropriate integrating factor. In this case the
dt/t = eln(t) = t Thus
R
integrating factor is e
d
(tv) = t6
dt
Another similar example occurs when the independent variable t
is missing.
F (y, y′ , y′′ ) = 0
This can be simplified by the substitution
y′ = p(y)
(in other words, instead of thinking of t as the independent variable
we think of y as the independent variable. Differentiating the above
gives
dp dp
y′′ = y′ =p
dy dy
for example
46 differential equations

Example 1.6.6. Solve the equation

y′′ y2 = y′

Making the change of variables above gives

dp
y2 p =p
dy
either p = 0 or dividing through by p gives
dp 1
= 2
dy y
or p(y) = − 1y + c but this is the same as

dy 1 cy − 1
= c− = (1.33)
dt y y
y
dy = dt (1.34)
cy − 1
which can be integrated up to get
y log(1 − cy)
+ = t − t0 .
c c2
Here we have denoted the second constant of integration by t0 . Note that
as this is a second order equation we have two arbitrary constants of inte-
gration (c and t0 ) but they do not enter linearly. Also note that we have an
implicit representation for y as a function of t but we cannot solve explicitly
for y as a function of t.

Exercise 1.6.5
Find a general solution for each equation.
a) y′′ + 3y′ = te−3t b) y′′ + 2ty′ = 0
c) ty′′ = y′ d) y′′ = 2yy′
e) ey y′′ = y′ f) yy′′ = (y′ )2

One more case: A Bernoulli equation is a first order equation of

the form
dy
+ P(t)y = Q(t)y a
dt
for some real number a ̸= 0, 1 (If a = 0 or a = 1 the equation is first
order linear.) We can divide the equation through by y a to get
dy
y− a + P ( t ) y 1− a = Q ( t ).
dt
dy
If we define w = y1− a ; dw
dt = (1 − a ) y
−a
dt we see that the new
dependent variable w(t) satisfies a first order linear equation
1 dw
+ P ( t ) w = Q ( t ).
1 − a dt
 first order differential equations 47

Example 1.6.7. Find the general solution to the nonlinear first order equa-
tion
dy
+ y = cos(t)y5 .
dt
dy
Making the change of variables w(t) = y(t)−4 ; dw
dt = −4y
−5
dt gives the
following differential equation for w(t):

dw
−1/4 + w = cos(t).
dt
This is first order linear and can be solved: the general solution is

16 4
w(t) = Ae−4t + cos(t) − sin(t) = y(t)−4
17 17
which gives
− 1
16 4

4
y(t) = Ae−4t + cos(t) − sin(t)
17 17

Exercise 1.6.6
Find a general solution for each equation.
t
a) y′ + y = y b) ty′ = 3y + ty2
√ 2y
c) ty′ + 2y = t3 y d) y′ = ty3 − t

Exercise 1.6.7
Solve the following initial value problems.

a) ty′ + 2y = t3 y2 , y (1) = 1
y2 sin( x )−2xy−ey
b) y′ = x2 + xey +2y cos( x )
, y (0) = 1
13 ′ 5
c) ty′′ + 2y′ = t, y (1) = 12 , y (1) = 4

d) (t + 1)y′ = y2 , y (0) = 1
e) y′′ = y(y′ )2 , y(0) = 0, y′ (0) = 1
2t−y
f) y′ = y+t , y (0) = −1
√ 6y
g) y′ = 2t y − t , y (1) = 4
h) t2 y′ = y + 3
2x −3x2 y2 −y3
i) y′ = 2x3 y+3xy2
, y (1) = 1
48 differential equations

1.7 Autonomous equations, equilibria, and the phase line.

In this section we will consider first order autonomous

equations, where the equation does not explicitly involve the in-
dependent variable. While such equations are always separable, and
thus solvable, the explicit solutions may be complicated and not par-
ticularly useful. It is always possible, on the other hand, to get very
concrete and useful qualitative information on the solution.
The general first order autonomous equation is given by

dy
= f ( y ). (1.35)
dt
Again these equations are always solvable in principle, as they are
separable and can be reduced to integration, but such a representa-
tion may not be very useful. Even assuming that the integration can
be done explicitly the method naturally gives t as a function of y.
It may not be possible to explicitly invert the relation to find y as a
function of t.
The ideas in this section amount to an elaboration on the idea of
a slope field. Equation (1.35) relates the slope of the curve y(t) to the
value of f (y). A special role is played by the points where the slope
is zero. These are called equilibria.
dy
Definition 1.7.1. An equilibrium is any constant solution to dt= f ( y ).
dy dy0
If y(t) = y0 is a solution to Equation (1.35) then dt = dt = 0 = f ( y0 ),
hence an equilibrium is a solution to f (y0 ) = 0.

One of the main ideas of this section is this: if we consider draw-

ing the slope field for an autonomous equation it is clear that the
slope does not depend on the independent variable t. For this rea-
son we do not really need to draw the whole plane, as any vertical
segment is repeated over and over. We can collapse all of the infor-
mation into a graphical device known as the phase line as follows.
We draw a vertical line representing the y-axis. Along this axis we
put an arrow pointing upward where f (y) is positive and an arrow
pointing downwards where f (y) is negative. Points where f (y) = 0
are, of course, equilibria.
We refer to this kind of diagram as the phase line. Fig. (1.12) is an
example of a phase line but oriented horizontally. See the following
example.

Example 1.7.1. Consider the autonomous equation

dy
= y2 − 1
dt
 first order differential equations 49

The equilibria are values of y for which y2 − 1 = 0, or y = ±1. In

drawing the phase line we see that y2 − 1 < 0 if y ∈ (−1, 1), so this
1.5

interval would be populated with downward arrows. If, on the other hand,
y > 1 or y < −1 then y2 − 1 > 0, and these two regions would be filled 1.0

with upwards arrows. This is illustrated in the three margin figures. Figure 0.5

1.10 depicts the slope field. It is clear that the slopes do not depend on t, only 0.0

one y, so we don’t really need a two dimensional figure – all we really need
-0.5
to know is how the slopes depend on y.
dy
In the next figure, (Fig. (1.11) we have plotted the slope dt = y2 − 1 as a -1.0

function of y. Note that this plot is turned 90 degrees from the slope field: in -1.5

the slope field t is taken to be the independent variable, here y is taken to be -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

the independent variable.

dy Figure 1.10: A plot of the slope
On the y axis we have plotted arrows representing the sign of dt : arrows
dy
pointing to the left indicate
dy
< 0, or y decreasing, while arrows pointing field for dt = y2 − 1.
dt
dy
to the right indicate dt > 0, or y increasing. We can clearly see the equilib-
ria at y = ±1. We can condense this plot to just a one-dimensional field of
-1. 1.
arrows, as we have done in Fig. (1.12).
This figure tells us a lot about the qualitative behavior of y(t): we can see
that y is decreasing in the interval y ∈ (−1, 1) and increasing outside that
interval. Further we can see that there is a real difference in the behavior of
the solution near the equilibria y = ±1. The equilibrium at y = −1 has the
property that y values that start near it move towards it. The equilibrium Figure 1.11: A plot of
at y = 1 has the property that y values that start near it move away. This f (y) = y2 − 1 vs. y. The ar-
motivates the next definition. rows on the y axis indicate the
dy
Definition 1.7.2. An equilibrium is said to be stable if nearby initial sign of dt = f (y).
points converge to the equilibrium. An equilibrium is unstable if nearby -1. 1.

initial points move away from the equilibrium.

Stability is a very important concept from the point of view of Figure 1.12: A plot of the phase
dy
applications, since it tells us something about the robustness of the line for dt = y2 − 1. The two
solution. Stable equilibria are in some sense “self-correcting”. If we equilibria are y = −1 and y = 1.
start with an initial condition that is close to the equlibrium then the
dynamics will naturally act in such a way as to move the solution
closer to the equilibrium. If the equilibrium is unstable, on the other
hand, the dynamics tends to move the solution away from the equi-
librium. This means that unstable equilibria are less robust – we have
to get things just right in order to observe them.
It would be nice to have a method to decide whether a given equi-
librium is stable or unstable. This is the content of the next theorem.

Theorem 1.7.1 (Hartman-Grobman). Suppose that we have the equation

dy
= f (y)
dt
and y∗ is an equilibrium ( f (y∗ ) = 0). Then
50 differential equations

• If f ′ (y∗ ) > 0 then the equilibrium is unstable - solutions initially near

y∗ move away from y∗ exponentially.

• If f ′ (y∗ ) < 0 then the equilibrium is stable - solutions initially near y∗

move towards y∗ exponentially.

• If f ′ (y∗ ) = 0 further analysis is required.

We won’t give a proof of this theorem, which is extremely impor-

tant in the study of dynamical systems, but we will try to give the
idea behind it. If y is near the equilibrium y∗ then we can try to find
a solution y(t) = y∗ + ϵv(t), where ϵ is small. If we substitute this
into the equation we find that

dv
ϵ = f (y∗ + ϵv).
dt
So far this is exact – all that we have done is to rewrite the equation.
At this point, however, we make an approximation. Specifically we
will Taylor expand the function f to first order in ϵ and use the fact
that f (y∗ ) = 0:

dv
ϵ = f (y∗ + ϵv) ≈ f (y∗ ) + ϵ f ′ (y∗ )v + O(ϵ2 ) = ϵ f ′ (y∗ )v.
dt
y−y∗
So the quantity v = ϵ satisfies the approximate equation

dv
= f ′ (y∗ )v.
dt
We know how to solve this equation: v exhibits exponential growth if
f ′ (y∗ ) > 0 and exponential decay if f ′ (y∗ ) < 0.

Example 1.7.2 (Logistic Growth Revisited). Previously we considered

dP
= P( P0 − P)
dt
The equilibria are P = 0 and P = P0 . In this case f ′ ( P) = P0 − 2P, and
we have f ′ (0) = P0 > 0 and f ′ ( P0 ) = − P0 < 0, so the equilibrium
with zero population is unstable and the equilibrium with P = P0 is stable.
The zero population is unstable, so a small population will tend to grow
exponentially. As the population approaches P0 the growth rate slows and
the population approaches an asymptote. If the population begins above P0
then the population will decrease towards the equilibrium population P0 .
//
The first flip-flop was built in 1918
Example 1.7.3 (Bistability). The differential equation
by W. Eccles and F.W. Jordan using a
pair of vacuum tubes. Vacuum tube
dy flip-flops were used in the Colossus
= y(1 − 2y)(y − 1)
dy code-breaking computer in Bletchly
Park during World War II.
Photo of William Eccles. Uncredited from a photograph that

appeared on p. 228 of the June 1925 issue of Radio Broadcast

magazine. , Public domain, via Wikimedia Commons

 first order differential equations 51

has a property called “bistability” . In electrical engineering circuits with

bistable behavior are commonly called “flip-flops” and have been an impor-
tant design element since the earliest days of electrical engineering. In this
example there are three equilibria, y = 0, y = 12 , y = 1. The middle equilib- 0 0.5 1.

rium y = 12 is unstable, while y = 0 and y = 1 are stable. The phase line for
this equation is depicted in the margin.
Flip-flops are ubiquitous in consumer electronics, as they form the basis
of logic gates and RAM memory, and one can buy any number of integrated Figure 1.13: The phase-line for
circuits with various numbers and types of flips-flops on a single chip. a bistable system (flip-flop)
dy
Bistability means that they will remain in one of the two stable states dt = y(1 − 2y)(y − 1). The
(either 0 or 1) until some external force comes along to change the state. equilibria y = 0 and y = 1 are
stable, the equilibrium y = 21
Exercise 1.7.1 unstable.

For each of the following autonomous equations sketch a slope field,

draw a phase line, and determine the equilibria and identify their
stability.
dy
a) dt = 6y − 2y2
dy
b) dt = y2 − 4y
dy
c) dt = y2 − y3
In the study of differential equations and their applications one is
often led to consider how the qualitative aspects of an autonomous
differential equation – the number and stability of the equilibria – are
influenced by various physical parameters in the problem. It turns
out that while generally small changes in the physical parameters
typically do not change the qualitative properties of the equation
there are certain values of the parameters where small changes in the
parameters can produce dramatic changes in the qualitative prop-
erties of the solutions. This is a circle of ideas known as bifurcation
theory. We will explore a relatively simple case of this in the next
example.

Example 1.7.4. Fish

In population biology people often try to understand how various forces
can effect populations. For instance, if one is trying to manage a resource
such as a fishing ground one would like to weigh the economic benefits of
increased fishing against the long-term sustainability of the resource. In the
absence of fishing the simplest model for the growth of a population of fish
would be a logistic type model
dP
= kP( P0 − P).
dt
As we have observed before this equation has a stable equilibrium popu-
lation of P = P0 – any initially positive population will converge exponen-
52 differential equations

Population P
tially to the stable equilibrium P0 – and an unstable equilibrium population P0

P = 0.
One might try to model the effect of fishing on a fish population by a
model of the following form 0
k P20
Fishing Rate h
4
0

dP
= kP( P0 − P) − h. (1.36) Figure 1.14: The bifurcation di-
dt
Here the first term is the standard logistic term and the second term is a agram for the logistic model
constant, which is meant to represent fish being caught at a constant rate. with constant harvesting,
dP
Let’s assume that h, the rate at which fish are caught, is something that we dt = kP( P0 − P) − h. For
kP2
can set by a policy decision. We’d like to understand how the fish population low fishing rates, h < 40 there
is influenced by h, and how large a fishing rate we can allow before it ceases are two equilibria, one stable
to be sustainable. and one unstable. for higher
The righthand side is a quadratic, so there are two roots and thus two kP02
fishing rates h > 4 there are
equilibria. The quadratic formula gives the equilibria as no equilibria.
q
P0 ± P02 − 4h/k
P=
2
√
so if h/k is less than P0 /2 there are two positive real roots. The lower
one has positive derivative and the upper one has negative derivative so the
√
upper one is stable and the lower one is unstable. So if h/k < P20 there is
always a stable fixed point with a population greater than P20 .
√ kP2
Now consider what happens if h/k is greater than P0 /2, or h > 40 .
√
P ± P02 −4h/k
In this case the roots P = 0 2 are complex, and so there are
no equilibria. In this case the righthand side of Equation (1.36) is always
negative. This implies that the population rapidly decays to zero and the
population crashes.
Consider this from a policy point of view. Imagine that, over time, the
kP2
fishing rate is slowly increased until h is just slightly less than 40 . As the
fishing rate is increased the stable equilibrium population decreases, but it is
always larger than P20 , and there is a second, unstable equilibrium just below
P0
2 . There is no obvious indication that the population is threatened, but
if h is further increased then the two equilibria vanish and the population
crashes.
This suddent disappearance of a stable and an unstable equilibrium is
known as the “saddle-node bifurcation”’. It is illustrated in the margin
Figure (1.14), which plots the locations of the two equilibria (the stable and
the unstable) as a function of the fishing rate h. The stable branch is depicted
by a solid line, the unstable branch by a dashed line. For small values of h
kP02
there is one stable equilibrium and one unstable one, but at h = 4 the two
kP02
equilibria collide and vanish, and for h > 4 there are no equilibria.

We will discuss equilibria and qualitative methods further in a

later section of the notes. The Hartman-Grobman theorem extends
 first order differential equations 53

to higher dimensional systems, although the classification of the

equilibria is somewhat more involved, and is no longer simply a
question of stable or unstable.

1.8 Numerical Methods

We close the chapter on first order equations with an introduction

to numerical methods. Just as in calculus, where not every integral
can be expressed in terms of elementary functions, we cannot expect
to be able to analytically solve every differential equation. Even in
the case where the equation can be solved explicitly it may happen
that the solution is defined implicitly or in some form that may not
be convenient to use. For this reason it is desirable to have numerical
methods. In the case of integration you probably learned about the
Riemann sum, the trapezoidal rule, Simpson’s rule, etc. There are
natural analogs of these methods for solving differential equations. Figure 1.15: The Euler method
got some positive press in the
movie “Hidden Figures”, when
1.8.1 First order (Euler) method.
Katherine Goble Johnson used
it to calculate the orbits of the
The simplest numerical method is what is known as a first Mercury astronauts. The trajec-
order or Euler method. Suppose that one would like to solve the first tories of the first two manned
order ordinary differential equation Mercury missions (Freedom 7,
piloted by Alan Shepard and
y′ = f (y, t) y (0) = y0 .
Liberty Bell 7, piloted by Gus
One can formally solve this equation by integrating and applying the Grissom) were calculated en-
fundamental theorem of calculus. If we integrate this equation from tirely by hand by Johnson and
t = 0 to t = ∆t we find that other computers. Glenn’s flight
Z ∆t (Friendship 7) was the first to
y(∆t) − y(0) = f (y(t), t)dt. have the orbit calculated by an
0 electronic computer. Glenn re-
This formula is correct, but it is not terribly useful – it expresses fused to fly the mission unless
y(∆t) in terms of y(t) for t ∈ (0, ∆t) in the integral. So we would Johnson checked the results
have to know y(t) in order to find y(t). of the electronic computation
However, if ∆t is small then we expect that y(t) will not change personally. NASA; restored by Adam Cuerden,
very much in the interval (0, ∆t). In this case we can just approxi- Public domain, via Wikimedia Commons

mate y(t) ≈ y(0) inside of f . This gives

Z ∆t
y(∆t) − y(0) = f (y(t), t)dt
0
Z ∆t
≈ f (y(0), 0)dt
0
y(∆t) ≈y(0) + f (y(0), 0)∆t
54 differential equations

We can, of course, continue this process: we can use our approxima-

tion of y(∆t) to find an approximation for y(2∆t), and so on. This
method is known as the first order explicit scheme, or (explicit) Euler
scheme.

Method 1.8.1. The first order Euler scheme for the differential equation
dy
= f (y, t) y ( a ) = y0 (1.37)
dt
on the interval ( a, b) is the following iterative procedure. For some choice
of N we divide the interval ( a, b) into N subintervals with ∆t = b− a
N and
ti = t0 + i∆t. We then define yi by the following iteration

yi+1 = yi + f (yi , ti )∆t.

Then the solution to Equation (1.37) at position ti is approximately given by

yi :
y(ti ) ≈ yi + O(∆t).
If you think about this from the point
Of course we get a better approximation if we take smaller steps, of the slope field what we are really
doing is replacing the smooth curve
∆t, but it comes at a cost since we have to do more steps to compute with a polygonal path. We pick a
the function. The Euler scheme is first order because the error scales point (t0 , y0 ), follow the tangent line
like the first power of the step size, (∆t)1 . It is often more desirable to for a distance ∆t to find a new point
(t0 + ∆t, y0 + f (y0 , t0 )∆t) and continue
use what is known as a higher order method, where the error scales this process to get a polygonal curve
like (∆t)k , with k > 1, especially if one requires more accuracy. To made up of a bunch of straight line
segments.
see this imagine that three different methods have errors ∆t, ∆t2 and
∆t4 , and one would like the error to be 10−6 . The first method would
require N = 1000000 steps, the second would require N = 1000 steps,
3
and the third would require N = 10 2 ≈ 32 steps. This is a bit naive
– the higher order methdos generally require more work, and there
are generally different constants multiplying the error – but it is often
more efficient to use a higher order method.
We briefly mention a couple of popular higher order methods. If 0.4
. Exact

the Euler method is the equivalent of the left-hand rule for Riemann
■ Euler
☺.
...
..
0.3 ... ▼ Heun
.
....

sums then the equivalent of the midpoint rule is the following rule:
.
.
.... ■
..
.... ☺ RK4
...
0.2 .▼...
.☺
....
....
..
....... ■
.
......
yi+1 = yi + f (yi + f (yi , ti )∆t/2, ti + ∆t/2)∆t. 0.1
....... ■
.........
.........
☺......
.▼
.......

...........▼
☺
..
...
...
...
...
. ■
☺▼ ▼................................☺
.■
■.................................☺ ......
.▼
■
0.2 0.4 0.6 0.8 1.0

This rule is also called the midpoint rule, and you can see why. In-
stead of evaluating f at (ti , yi ) it is evaluated at the midpoint between Figure 1.16: A graph of the ex-
dy
this and the next point. This is second order accurate: we have that act solution to dt = y2 + t2 with
yi − y(ti ) = O(∆t2 ). You can also see that this rule requires a bit more y(0) = 0 for t ∈ (0, 1) together
computational work: we have to evaluate the function f (y, t) twice with the Euler and improved
for each step that we take. A different second order method is called Euler approximations to the
Heun’s method or the improved Euler method. It is defined by solution with N = 6 subdivi-
sions ((∆t = 61 ). The step size
has been deliberately chosen to
yi+1 = yi + f (yi , ti )∆t/2 + f (yi + f (yi , ti )∆t, ti + ∆t)∆t/2
be large to exaggerate the dif-
ference. It is apparent that the
improved Euler method does
a better job of approximating
the solution than the standard
Euler method, and that the
fourth order Runge-Kutta can’t
be distinguished from the exact
 first order differential equations 55

and is the analog of the trapezoidal rule for numerical integration.

A more complicated rule is called the fourth order Runge-Kutta
method, or RK4. This is defined by

k1 = f (yi , ti )∆t
k2 = f (yi + k1 /2, ti + ∆t/2)∆t
k3 = f (yi + k2 /2, ti + ∆t/2)∆t
k4 = f (yi + k3 , ti + ∆t)∆t
1
y i +1 = y i + (k + 2k2 + 2k3 + k4 )
6 1
and is the natural analog to Simpson’s rule for numerical integration.
RK4 is a fourth order method, yi − y(ti ) = O(∆t4 ). There are many
other methods that one can define – each has its own advantages and
disadvantages.
2
Higher Order Linear Equations

2.0.1 Existence and uniqueness theorem.

One important class of differential equations is that of linear differen-
tial equations. Recall that a linear differential equation of order n is
one that can be written in the form

dn y d n −1 y d n −2 y
n
+ a n −1 ( t ) n −1 + a n −2 ( t ) n −2 + . . . a 0 ( t ) y = f ( t )
dt dt dt
In other words a linear differential equation is a linear relationship
between y and its first n derivatives. It can depend on the indepen-
dent variable t in an arbitrary way. To begin with we first mention
the fundamental existence and uniqueness theorem. Since the differ-
ential equation depends linearly on y and its derivatives the existence
theorem becomes simpler, and can be stated solely in terms of the
coefficients ak (t) and f (t).

Theorem 2.0.1 (Existence and Uniqueness). Consider the general linear

initial value problem

d n y n −1 dk y ( n −1)
y(t0 ) = y0 , y′ (t0 ) = y0′ , . . . y(n−1) (t0 ) = y0
dtn k∑
+ a k ( t ) = f (t)
=0 dtk

−1
If the functions { ak (t)}nk= 0 , f ( t ) are all continuous in an open set t ∈ ( a, b )
containing t0 then the given initial value problem has a unique solution
defined for all t ∈ ( a, b) .

Note that this result is much stronger than the general (nonlinear)
existence/uniqueness result. The nonlinear result only guarantees the
existence of a solution in some small neighborhood about the initial
condition. We don’t really know a priori how big that interval might
be. For linear equations we know that as long as the coefficients are
well-behaved (continuous) then a unique solution exists.
It is worth defining a couple of pieces of terminology that will be
important.
58 differential equations

Definition 2.0.1. A linear differential equation in which the forcing term

f (t) is zero is called “homogeneous”:

d n y n −1 dk y
n
+ ∑ ak (t) k = 0
dt k =0 dt
A linear differential equation in which the forcing term f (t) is non-zero is
called “non-homogeneous” or ”inhomogeneous”.
Inhomogeneous Homogeneous

dy dy
− 5y = t − 3y = 0
Example 2.0.1. dt dt
d2 y dy d2 y dy
+ 2 + y = cos t + et − + 5y = 0
dt2 dt dt2 dt
4 d3 y
d y 1 d4 y d3 y
+ cos t 3 + y = + sin t 3 + et y = 0
dt 4 dt 1 + t3 dt 4 dt

Exercise 2.0.1
Without actually solving the problems, determine in which interval I
we are guaranteed a unique solution for each initial value problem.
t2 ′ cos (t) sin (t)
a) y′′′ − t −2 y + t +3 y = t2
, y(−1) = 2, y′ (−1) = 3, y′′ (−1) =
4
t2 ′ cos (t) sin (t)
b) y′′′ − t −2 y + t +3 y = t2 , y(1) = 2, y′ (1) = 3, y′′ (1) = 4
t2 ′ cos (t) sin (t)
c) y′′′ − t− 2 y + t +3 y = t2 , y(3) = 2, y′ (3) = 3, y′′ (3) = 4
t +1
d) (t − 2)y′′ + sin (t)
y = e t , y (1) = 0, y′ (1) = 1

2.1 Linear Homogeneous Equations

2.1.1 Linear Independence and the Wronskian

Linear equations are important for a couple of reasons. One is that
many of the equations that arise in science and engineering are lin-
ear. A second reason is that solutions of linear equations have a lot of
structure – much more so than solutions of some arbitrary differen-
tial equation. In particular they form a linear vector space. That is the
basic idea that we wish to develop in this section. The first important
aspect of this structure that we consider is the idea of superposition.
Theorem 2.1.1 (Superposition Principle). Consider the linear homoge-
neous ( f (t) = 0) differential equation

d n y n −1 dk y
n
+ ∑ ak (t) k = 0
dt k =0 dt
 higher order linear equations 59

If y1 (t), y2 (t), . . . ym (t) are all solutions to the above equation then any
arbitrary linear combination, y(t) = c1 y1 (t) + c2 y2 (t) + . . . cm ym (t) is also
a solution.

The proof is simple – it just amounts to noticing that the derivative

d
is a linear operator: dt (c1 y1 (t) + c2 y2 (t)) = c1 dy dy2
dt + c2 dt , together
1

with the fact that the derivatives of y enter into the differential equa-
tion linearly.

Example 2.1.1. Consider the equation

d2 y
+y = 0
dt2
It is not hard to guess that the functions y1 (t) = cos(t), y2 (t) = sin(t)
both satisfy the equation. By the superposition theorem y = c1 cos(t) +
c2 sin(t) is also a solution. If one is given the equation with boundary condi-
tions
y′′ + y = 0 y(0) = y0 y′ (0) = y0′
we can try the solution y = c1 cos(t) + c2 sin(t) and see if we can solve for
c1 , c2 . Substituting t = 0 gives

y(0) = c1 cos(0) + c2 sin(0) = c1 = y0

Similarly differentiating y = c1 cos(t) + c2 sin(t) and substituting t = 0

gives
dy
(0) = −c1 sin(0) + c2 cos(0) = c2 = y0′
dt
This gives a solution to the problem, y = y0 cos(t) + y0′ sin(t), which, by
the existence and uniqueness theorem, must be the ONLY solution, since
solutions are unique.

The same thing holds in general. We’ll state this as a “guiding

principle”.
Guiding Principle: If we can find n different solutions to a nth The important thing here is “different.”
order linear homogeneous differential equation, then the general In the previous example we could have
taken a linear combination of sin t
solution will be given by a linear combination of those solutions. and 2 sin t. This would have solved
This may remind you of the following fact from linear algebra: if the differential equation. However
this would not have been the most general
we have n linearly independent vectors in an n-dimensional vector solution. For instance the solution to
space then they form a basis, and any vector in the space can be d2 y
+y = 0 y (0) = 1 y ′ (0) = 0
expressed as a linear combination of these vectors. In fact it is easy to dt2
see that the solutions to a linear homogeneous differential equation is y = cos(t), which is not a linear
satisfy all of the axioms of an abstract vector space. So when we combination of sin(t) and 2 sin(t).

are solving linear homogeneous differential equations we are really

doing linear algebra.
The next important idea is to distinguish when a collection of so-
lutions are really “different”. This is not always so clear, especially in
60 differential equations

the case of higher order equations. We’ll need some facts from linear
algebra regarding the solvability of a linear system of equations.
The first is the idea of linear independence. A set of vectors
⃗v1 , ⃗v2 . . . , ⃗vk is linearly indpendent if ∑ ci⃗vi = 0 implies that ci = 0
for all i. In other words a set of vectors is linearly independent if the
only linear combination of the vectors that adds up to the zero vector
is when the coefficient of each vector individually is zero. We extend
this same definition to functions

Definition 2.1.1. A set of functions y1 (t), y2 (t), . . . , yn (t) is linearly

independent if ∑ ci yi (t) = 0 implies that ci = 0 for all i. A set of functions
y1 (t), y2 (t), . . . , yn (t) is linearly dependent if there is a linear combination
∑ ci yi (t) = 0 where the coefficients ci are not all zero.
Referring back to the earlier example it is easy to see that the func-
tions y1 (t) = cos t and y2 (t) = sin t are linearly independent. One
way to see this is to use the trigonometric identity
c2
q
c1 cos t + c2 sin t = c21 + c22 cos(t − ϕ) ϕ = arctan( ).
c1
A second, and probably easier, way is to take the equation c1 cos t +
c2 sin t = 0 and substitute in t = 0 and t = π2 to find

c1 = 0
c2 = 0.

Similarly the functions y1 (t) = sin t and y2 (t) = 2 sin t are linearly
dependent, since 2y1 − y2 = 0.
The second fact we need from lineal algebra is the following one
about the solutions of a set of n linear equations in n unknowns,
which is a standard fact from any first course in linear algebra.

Theorem 2.1.2. Given a set of n linear equations in n unknowns ai

M11 M12 M13 . . . M1n c1 y0

    
  y′
M M M . . . M c
  
 21 22 23 2n   2   0 
 .
 . .. .. ..    ..  = 
   .. 
 . . . .  .   .


( n −1)
Mn1 M2n M3n . . . Mnn cn y0

These equations have a unique solution if and only if the determinant is

nonzero:
M11 M12 M13 . . . M1n
M21 M22 M23 . . . M2n
.. .. .. .. ̸= 0.
. . . .
Mn1 M2n M3n . . . Mnn
If the determinant is zero then the equations either have no solution or have
infinitely many solutions.
 higher order linear equations 61

As a practical matter if we have n solutions y1 (t), y2 (t), . . . yn (t) to

an nth order linear differential equation

d n y n −1 dk y
dtn
+ ∑ a k ( t )
dtk
=0
k =0

then what we would really like is to be able to express the solution to

the equation with arbitrary initial conditions

d n y n −1 dk y ( n −1)
y(t0 ) = y0 ; y′ (t0 ) = y0′ ; . . . y(n−1) (t0 ) = y0
dtn k∑
+ a k ( t ) =0
=0 dtk

as a linear combination of our basis, y(t) = c1 y1 (t) + c2 y2 (t) +

. . . cn yn (t). If one tries to solve for the coefficients c1 , c2 . . . , cn then we
find a system of linear equations

y1 ( t0 ) y2 ( t0 ) y3 ( t0 ) . . . y n ( t0 ) c1 y0
    

 y1′ (t0 ) y2′ (t0 ) y3′ (t0 ) . . . y′n (t0 ) 
 c2
 
  y0′ 

 .. .. .. ..  .. = .. 

 . . . .

 .
 
  .


( n −1) ( n −1) ( n −1) ( n −1) ( n −1)
y1 ( t0 ) y2 ( t0 ) y3 ( t0 ) . . . y n ( t0 ) cn y0

We know that this system has a unique solution if the determinant

of this matrix is non-zero. This motivates the following definition,
which formalizes the idea that y1 (t), y2 (t), . . . yn (t) should all be
“different”.

Definition 2.1.2. The Wronskian determinant (or Wronskian) of a set of n

functions y1 (t), y2 (t), . . . yn (t) is defined to be the determinant

y1 ( t ) y2 ( t ) y3 ( t ) . . . yn (t)
y1′ (t) y2′ (t) y3′ (t) . . . y′n (t)
W (y1 , y2 , . . . , yn )(t) = .. .. .. ..
. . . .
( n −1) ( n −1) ( n −1) ( n −1)
y1 ( t ) y2 ( t ) y3 (t) . . . yn (t)

The following observation shows that the Wronskian is a way to

detect the dependence or independence of a set of functions.

Observation 2.1.1. Suppose that a set of functions y1 (t), y2 (t), . . . , yn (t)

are linearly dependent. Then the Wronskian is identically zero,

Proof. If y1 (t), y2 (t), . . . , yn (t) are linearly dependent then there ex-
ists constants c1 , c2 , . . . cn not all zero such that c1 y1 (t) + c2 y2 (t) +
. . . cn yn (t) = 0. Differentiating with respect to t shows that c1 y1′ (t) +
c2 y2′ (t) + . . . cn y′n (t) = 0, and similarly all higher derivatives. Thus we
have a linear combination of the columns of the matrix which sums
to zero, and thus the determinant is zero.
62 differential equations

It is important that, given a linear differential equation and an

arbitrary set of initial data, we be able to find a solution satisfying
that initial data. The following theorem shows that the Wronskian is
also the correct tool for determining when this is possible. The first
theorem states that if the Wronskian is non-zero at a point then we
can find a linear combination satisfying any given initial condition.

Theorem 2.1.3. If y1 (t), y2 (t), . . . , yn (t) are n solutions to an nth order

linear homogeneous differential equation

d n y n −1 di y
n
+ ∑ ai ( t ) i = 0
dt i =0 dt

where the coefficients ai (t) are all continuous in a neighborhood of t0 and

and the Wronskian W (y1 , y2 , . . . , yn )(t0 ) ̸= 0 then any solution to the
initial value problem

d n y n −1 di y
dtn
+ ∑ a i ( t )
dti
=0
i =0
y ( t0 ) = y0
y′ (t0 ) = y0′
..
.
( n −1)
y ( n −1) ( t 0 ) = y 0

can be written as a linear combination of y1 (t), y2 (t), . . . , yn (t).

To see this we simply apply the existence and uniqueness theo-

rem. We know that solutions to differential equations are, under the
hypothesis above, unique. Since we can always find a solution in the
form of a linear combination of y1 (t), y2 (t) . . . yn (t) then all solutions
are of this form. There is a bit of subtlety here. Note that the Wron-
skian is a function of the independent variable, in this case t. If the
functions {yi (t)}in=1 are linearly dependent then the Wronskian is
identically zero – in other words zero for all t. The first theorem says
that we can find a linear combination that satisfies any initial con-
dition at t0 provided that the Wronskian is not zero at t0 . The next
thing that we will show is that if the functions yi (t) solve a linear dif-
ferential equation with continuous coefficients then the Wronskian is
either NEVER zero or it is ALWAYS zero. So a basis for solutions at
one point will also be a basis for solutions at a different point.

Theorem 2.1.4 (Abel). Suppose that y1 (t), y2 (t)...yn (t) are solutions to
the linear homogeneous linear differential equation

dn y d n −1 y d n −2 y
n
+ a n −1 ( t ) n −1 + a n −2 ( t ) n −2 + . . . a 0 ( t ) y = 0
dt dt dt
 higher order linear equations 63

Then the Wronskian W (t) solves the FIRST ORDER homogeneous linear
differential equation
W ′ + a n − 1 ( t )W = 0

Proof. We will prove this only for n = 2, which follows from a

straightforward calculation. we have

W (t) = y1 y2′ − y1′ y2

W ′ (t) = y1′ y2′ + y1 y2′′ − y1′ y2′ − y1′′ y2
= y1 y2′′ − y1′′ y2

and so

W ′ + a1 (t)W = y1 (y2′′ + a1 (t)y2′ ) − y2 (y1′′ + a1 (t)y1′ )

= y1 (− a2 (t)y2 ) − y2 (− a2 (t)y1 ) = 0

Corollary 2.1.1. Suppose that y1 (t), y2 (t)...yn (t) are solutions to the linear
homogeneous linear differential equation

dn y d n −1 y d n −2 y
+ a n − 1 ( t ) + a n − 2 ( t ) + . . . a0 ( t ) y = 0
dtn dtn−1 dtn−2
with ai (t) continuous on the whole real line. Then the Wronskian is either
identically zero or it is never zero,

Proof. We know that W ′ + an−1 (t)W = 0, so we have that W (t) =

Rt
a (s)ds Rt
W ( t 0 ) e t0 n −1 . Since an−1 (t) is continuous we have that t an−1 (s)ds
Rt 0
a (s)ds
is finite. The exponential function is never zero, so e t0 n−1 is
never zero. Thus the only way that W (t) can be zero is if the constant
W (t0 ) = 0 in which case the Wronskian is identically zero.

Example 2.1.2. The differential equation

y′′ + y = 0

has solutions y1 (t) = cos(t) and y2 (t) = sin(t). It follows from Abel’s
theorem that the Wronskian solves

W′ = 0

in other words the Wronskian is a constant independent of t. We already

know this to be the case – it is easy to compute that W (t) = cos2 (t) +
sin2 (t) = 1.
Thus we know that y1 (t) = cos(t) and y2 (t) = sin(t) can satisfy an
arbitrary set of initial conditions at any point t.
64 differential equations

Example 2.1.3. It is easy to verify that two solutions to

4 6
y′′ − y′ + 2 y = 0
t t
are y1 = t2 and y2 = t3 . Notice that the coefficients are continuous
everywhere except t = 0. It follows that the Wronskian solves

4
W′ − W = 0
t
which has the solution W = ct4 . Computing the Wronskian of y1 and y2
gives
W = y1 y2′ − y1′ y2 = t2 (3t2 ) − (2t)(t3 ) = t4
The Wronskian is zero at t = 0 and non-zero everywhere else. Thus we can
satisfy an arbitrary initial condition everywhere EXCEPT at t = 0

Exercise 2.1.1
Determine if the following set of solutions are linearly independent
or dependent by calculating the Wronskian.

a) e3t , te3t b) e3t , 2e3t , te3t

c) e2t , e3t d) t + 1, t2 , 3t − t2
e) t − t2 , 4t − 1, 4t2 − 1

2.2 Linear constant coefficient equations: the characteristic poly-

nomial

We saw in the previous section that if one can construct n linearly

independent solutions y1 (t), y2 (t) . . . yn (t) to an nth order linear equa-
tion then the general solution is simply a linear superposition of
these solutions y(t) = ∑i ci yi (t). This brings us to the question of
how we find such solutions. Unfortunately we will generally not be
able to solve most differential equations in terms of the familiar func-
tions of calculus. However there is one important situation in which
we CAN always solve differential equations in terms of functions like
trigonometric functions, exponentials, and polynomials. This is the
case of "constant coefficient" linear differential equations, in which
the coefficients do not depend on the independent variable but are
instead constant. The most general constant coefficient linear nth
order equation is given by

dn y d n −1 y dy
n
+ p n −1 n −1 + . . . p 1 + p 0 y = 0
dt dt dt
 higher order linear equations 65

with p0 , p1 , . . . pn−1 constants.

To begin with we can look for a solution of a particular form, that
of an exponential function
y = ert .
If we substitute this into the above differential equation the condition
that y = ert be a solution is that

( r n + p n −1 r n −1 + p n −2 r n −2 + . . . p 1 r + p 0 ) = 0

The above is called the characteristic equation (sometimes the charac-

teristic polynomial).
Definition 2.2.1. Given the nth order constant coefficient linear differential
equation
dn y d n −1 y dy
n
+ p n − 1 n 1
+ . . . p1 + p0 y = 0
dt dt − dt
the characteristic equation is defined to be

( r n + p n −1 r n −1 + p n −2 r n −2 + . . . p 1 r + p 0 ) = 0

The utility of the characteristic equation is the for constant coefficient

equations the solution can be reduced to the problem of finding the roots of a
polynomial. To illustrate this we first begin with a simple example.
Example 2.2.1. Find the solution to the initial value problem

y′′ − 3y′ + 2y = 0
y (0) = 1
y ′ (0) = 4

Looking for a solution to y′′ − 3y′ + 2y = 0 in the form y = ert gives a

characteristic equation r2 − 3r + 2 = (r − 1)(r − 2) = 0. This gives two
roots, r = 2 and r = 1, and thus two solutions y1 = et and y2 = e2t .
The general solution is given by y = Aet + Be2t and y′ = Aet + 2Be2t .
Imposing the conditions y(0) = 1, y′ (0) = 4 gives two equations

y (0) = A + B = 1
y′ (0) = A + 2B = 4

The simultaneous solution to these two equations is A = −2, B = 3. The

solution to the initial value problem is then y(t) = −2et + 3e2t
We know from the fundamental theorem of algebra that an nth
degree polynomial always has exactly n roots, counted according
to multiplicity. These roots will give us the n linearly independent
solutions to the differential equation. However it is a bit tricky to
deal with the cases of complex roots and roots of higher multiplicity,
so we begin with the simplest possible case, where the polynomial
has n real distinct roots.
66 differential equations

2.2.1 Real distinct roots

If the characteristic polynomial has n distinct real roots then one
obviously has n solutions y1 = er1 t , y2 = er2 t , . . .. This motivates the
following:

Theorem 2.2.1. Suppose that the characteristic equation

( r n + p n −1 r n −1 + p n −2 r n −2 + . . . p 1 r + p 0 ) = 0

has n real distinct roots r1 , r2 , r3 , . . . , rn . Then the constant coefficient linear

differential equation

dn y d n −1 y dy
n
+ p n −1 n −1 + . . . p 1 + p 0 y = 0
dt dt dt
has solutions y1 = er1 t , y2 = er2 t , . . ..

It may not be completely obvious, but these solutions are linearly

independent, as is shown in the next theorem:

Theorem 2.2.2. If y1 (t) = er1 t , y2 (t) = er2 t , . . . , yn (t) = ern t with

r1 , r2 . . . rn distinct, then y1 (t), y2 (t) . . . , yn (t) are linearly independent. In
particular the Wronskian is given by
n −1 n
W (y1 (t), y2 (t), . . . , yn (t)) = e(r1 +r2 +...rn )t ∏ ∏ (r j − ri )
i =1 j = i +1

We won’t prove this, but it follows from an identity for the deter-
minant of what is known as a Vandermonde matrix. The Wronskian
is given as the product of an exponential (which never vanishes)
times a product over the differences of roots. If all of the roots are
distinct the difference between any two roots is non-zero, and thus
the product is non-zero, and so the Wronskian is never zero.

Example 2.2.2. Find three linearly independent solutions to

y′′′ + 6y′′ + 3y′ − 10y = 0

The characteristic equation is given by

r3 + 6r2 + 3r − 10 = 0

It is easy to see that the roots are r = 11 ,r2 = −2 and r3 = −5. This gives

y1 ( t ) = e t y2 (t) = e−2t y3 (t) = e−5t .

Since the roots are distinct the solutions are guaranteed to be linearly inde-
pendent. In fact we know that the Wronskian satisfies

W (y1 , y2 , y3 ) = e(1−2−5)t (−2 − 1)(−5 − 1)(−5 − 2)

= (−3)(−6)(−3)e−6t = −54e−6t
 higher order linear equations 67

Since the characteristic equation for an nth order linear constant

coefficient equation is a nth degree polynomial, it will have n roots
counted according to multiplicity. However the roots may be com-
plex, and there may be roots of higher multiplicity – we may have to
count certain roots more than once. Recall the following definition of
multiplicity

Definition 2.2.2. A number r0 is a root of the polynomial P(r ) of mul-

tiplicity k if (r − r0 )k divides P(r ) and (r − r0 )k+1 does not divide P(r )
.

Exercise 2.2.1
Below are a list of polynomials together with a root of that polyno-
mial. Find the multiplicity of the root.

a) P(r ) = (1 − r2 ) for r = 1
b) P(r ) = (1 − r )2 for r = 1
c) P(r ) = r3 − 3r2 + 4 for r = 2
d) P(r ) = r3 − 3r2 + 4 for r = 1
e) P(r ) = r5 + r3 for r = 0

There are some subtleties connected with complex roots and multi-
ple roots. The next two subsections illustrate this.

2.2.2 Complex distinct roots

The case of complex roots is not much more difficult than that of real
roots, if we remember the Euler formula.

Theorem 2.2.3 (Euler). The complex exponential eiσt can be expressed in

terms of trigonometric functions as follows:

eiσt = cos σt + i sin σt.

More generally we have that

e(µ+iσ)t = eµt (cos σt + i sin σt) = eµt cos σt + i eµt sin σt.

If r is a complex root of the characteristic equation then one can

either work directly with the complex exponential ert , or one can
work with the real and imaginary parts.

Theorem 2.2.4. Suppose that the characteristic equation

( r n + p n −1 r n −1 + p n −2 r n −2 + . . . p 1 r + p 0 ) = 0
68 differential equations

has a pair of complex conjugate roots r = µ + iω and r̄ = µ − iω. Then the

constant coefficient linear differential equation

dn y d n −1 y dy
n
+ p n −1 n 1
+ . . . p1 + p0 y = 0
dt dt − dt

has solutions y1 (t) = e(µ+iω )t , y2 (t) = e(µ−iω )t . Alternatively one can

work with the real and imaginary parts

y1 (t) = eµt cos(ωt) y2 (t) = eµt sin(ωt)

There are a couple of idealizations
being made here. First we are neglect-
Example 2.2.3 (Harmonic Oscillator). The differential equation ing damping and loss terms. We are
also assuming a perfectly Hooke’s law
spring – one in which the restoring
my′′ + ky = 0, force is proportional to the displace-
ment. Real springs have mechanical loss
is an idealized equation for a mass-spring system called the simple harmonic due to friction, air resistance, etc, and
are not perfectly linear. Nevertheless it
oscillator. Here m represents the mass and k the linear spring constant. The is a very common and useful model.
characteristic equation is

mr2 + k = 0,

with roots given by

r r
k k
r= − = ±i .
m m
q
k
This gives two solutions: y1 = eiωt , y2 = e−iωt , with ω = m These
are perfectly acceptable complex solutions to the differential equation. In
engineering, particularly in electrical engineering, it is standard to work
with complex valued solutions. If one would prefer to have real-valued
solutions then one can use the fact that any linear combination of solutions
is also a solution. Taking the special linear combinations

Figure 2.1: Robert Hooke (1635–

1 iωt
(e + e−iωt ) = cos ωt 1703) was a natural philosopher
2
and polymath. Originally em-
1 iωt
(e − e−iωt ) = sin ωt ployed at Oxford as an organist
2i
he became assistant to Robert
shows that y1 (t) = cos ωt and y2 (t) = sin ωt are (linearly independent) Boyle. He first hypothesized
real solutions to the differential equation my′′ + ky = 0. the inverse square law of grav-
itation (later developed by
The other complication that can arise is when the characteristic
Newton) and made contribu-
polynomial has roots of higher multiplicity.
tions to mechanics, the theory
of light, atronomy, geology,
2.2.3 Multiple roots clockmaking and many other
The other exceptional case occurs when the characteristic equation fields. His writing could be
has multiple roots. As always we begin with an example. somewhat obfuscated: he an-
nounced what we know call
Hooke’s law as the anagram
“ceiiinosssttuv”. This can be re-
arranged to “Ut tensio, sic vis”,
Latin for “As the extension, so
the force.” It also anagrams to
“Incisive Stouts”, so it is likely
 higher order linear equations 69

Example 2.2.4. Consider the differential equation

y′′ − 2y′ + y = 0.

looking for a solution in the form y = ert gives the characteristic equation

r2 − 2r + 1 = (r − 1)2 = 0

so r = 1 is a root of multiplicity two. This gives one solution as y1 = et .

We need to find a second linearly independent solution. This can be done by
the following trick: we can factor the differential operator as follows: if we
define w = y′ − y then y′′ − 2y′ + y = 0 is equivalent to w′ − w = 0. The
solution to w′ − w = 0 is given by w = Aet . The equation y′ − y = w is
then equivalent to y′ − y = Aet . This can be solved by the integrating factor
method detailed in the first chapter to give y = Atet + Bet .

Theorem 2.2.5. Constant coefficient equations Suppose that

dn y d n −1 y dy
n
+ p n −1 n −1 + . . . p 1 + p 0 y = 0
dt dt dt
is a constant coefficient nth order differential equation, and that

( r n + p n −1 r n −1 + p n −2 r n −2 + . . . p 1 r + p 0 ) = 0

is the characteristic polynomial. Recall that the total number of roots of the
polynomial counted according to mutliplicity is n. For each simple root
(multiplicity one) ri we get a solution

yi ( t ) = eri t

If ri is a repeated root of the characteristic equation of multiplicity k then the

differential equation has k linearly independent solutions given by

y1 ( t ) = eri t
y2 (t) = teri t
y3 ( t ) = t2 eri t
..
.
y k ( t ) = t k −1 e r i t

If r = µ + iσ, r ∗ = µ − iσ are a complex conjugate pair of roots one can

either use the complex exponential solutions

y1 (t) = e(µ+iσ)t
y2 (t) = e(µ−iσ)t

or the real form of the solutions

y1 (t) = eµt cos(σt)

y2 (t) = eµt sin(σt)
70 differential equations

Example 2.2.5. Solve the initial value problem

y′′′ − 3y′′ + 3y′ − y = 0 y (0) = 0 y ′ (0) = 0 y′′ (0) = 1

The characteristic equation is r3 − 3r2 + 3r − 1 = (r − 1)3 = 0. There

is a root r = 1 of multiplicity k = 3. This gives three linearly independent
solutions y1 (t) = et , y2 (t) = tet , y3 (t) = t2 et . Taking the general solution
as y(t) = Aet + Btet + ct2 et we get

y(t) = Aet + Btet + Ct2 et

y′ (t) = Aet + Btet + Bet + Ct2 et + 2Ctet
y′′ (t) = Aet + Btet + 2Bet + Ct2 et + 4Ctet + 2Cet

This gives three equations in three unknowns

y (0) = A = 0
y ′ (0) = A + B = 0
y′′ (0) = A + 2B + 2C = 1

1
which can be solved to find A = 0, B = 0, C = 2 and the solution to the
initial value problem y(t) = 21 t2 et

Example 2.2.6. Solve the initial value problem

y′′′′ + 2y′′′ − 2y′ − y = 0 y (0) = 0 y ′ (0) = 0 y′′ (0) = 0 y′′′ (0) = 1

The characteristic equation is

r4 + 2r3 − 2r − 1 = 0

The roots are r = −1 with multiplicity k = 3 and r = 1 with multiplicity

k = 1. Therefore a linearly independent set of solutions is given by

y1 ( t ) = e − t
y2 (t) = te−t
y3 ( t ) = t2 e − t
y4 ( t ) = e t

Writing the general solution as y(t) = Ae−t + Bte−t + Ct2 e−t + Det we
 higher order linear equations 71

can solve for the coefficients as follows:

y(t) = Ae−t + Bte−t + Ct2 e−t + Det

y (0) = A + D = 0
y′ (t) = − Ae−t − Be−t t + Be−t − Ce−t t2 + 2Ce−t t + Det
y ′ (0) = − A + B + D = 0
y′′ (t) = Ae−t + Be−t t − 2Be−t + Ce−t t2 − 4Ce−t t + 2Ce−t + Det
y′′ (0) = A − 2B + 2C + D = 0
y′′′ (t) = − Ae−t − Be−t t + 3Be−t − Ce−t t2 + 6Ce−t t − 6Ce−t + Det
y′′′ (0) = − A + 3B − 6C + D = 1

Solving this system of four equations for the unknowns A, B, C, D gives

A = − 18 , B = − 14 , C = − 14 , D = 18 .

Exercise 2.2.2
Find general solutions to the following differential equations

a) y′′ + y′ − 6y = 0
b) y′′′ − 6y′′ + 9y′ = 0
c) y′′′ + 4y′ = 0
d) y′′′ + 3y′′ + 3y′ + y = 0
e) y′′′′ + 2y′′ + y = 0
f) y′′′ − 3y′′ + 4y = 0
g) y′′′ − 5y′′ + y′ − 5y = 0
h) y′′′′ − 8y′′′ + 16y′′ = 0
i) y′′′ + 4y′′ + 6y′ = 0

2.3 Non-homogeneous linear equations: Operator notation and

the structure of solutions

In the previous section we found a method for solving a general

constant coefficient linear homogeneous equation by finding the
roots of the characteristic polynomial. We next consider the non-
homogeneous case, where the forcing term f (t) is non-zero. We will
begin with the method of undetermined coefficients, and the closely
related Annihilator Method. These methods only work for certain
forcing terms f (t), but when they do work they are typically much
easier than method of variation of parameters, which we will learn
later.
72 differential equations

2.3.1 Operator notation

We are going to use the following notation. L will be a linear differ-
ential operator: something that acts on functions. For instance if we
define the linear operator L to be

d2 d
L= 2
+5 +4
dt dt
then the operator L is something that acts on functions and returns
another function. For instance L acting on a function y gives

Ly = y′′ + 5y′ + 4y.

So

Let = et + 5et + 4et = 10et ,

Le−2t = 4e−2t − 10e−2t + 4e−2t = −2et ,
L sin t = − sin t + 5 cos t + 4 sin t = 5 cos t + 3 sin t

Similarly if we define the linear operator L to be

d2 d
L = (1 + t2 ) 2
+ 9t − et
dt dt
then
Ly = (1 + t2 )y′′ + 9ty′ − et y

Note that in each of these cases the operator L has the property
that L( ay1 ( a) + by2 (t)) = aLy1 (t) + bLy2 (t). This is called a linear
operator. This follows from the fact that the derivative (and hence
the nth derivative) is a linear operator, and so a linear combination
of these is also a linear operator. Notice that a linear homogeneous
differential equation can always be written in the form

Ly = 0

for some choice of linear operator L, while a linear non-homogeneous

equation can be written in the form

Ly = f (t)

It will often be useful to use the operator notation, particularly

when we are discussing linear differential equations in the abstract
and don’t necessarily want to discuss a particular equation.
 higher order linear equations 73

2.4 The structure of solutions to a non-homogeneous linear dif-

ferential equation

This section is short, but contains an important observation about the

structure of the solution to a non-homogeneous linear equation.

Theorem 2.4.1. Suppose that the non-homogeneous differential equation

Ly = f (t)

has a solution y part (t) (called the particular solution). This solution need
not involve any arbitrary constant. Then the general solution to

Ly = f (t)

is given by It is common in engineering texts,

y(t) = y part (t) + yhomog (t) though not in mathematics texts, to
further separate the solution into
where yhomog (t) is the general solution to "zero input response" and "zero state
response" functions. The zero input
response function solves the homoge-
Ly = 0. neous problem Ly = 0 together with
the appropriate initial conditions. The
Proof. The proof is easy, given the operator notation. Suppose that zero state response function solves the
y(t) is the general solution to non-homogeneous problem Ly = f
together with zero initial conditions. In
Ly = f (t) other words the zero input solution is
the homogeneous solution satisfying
the correct initial conditions with zero
while y part (t) is a particular solution, Ly part (t) = f (t). Let’s look at forcing term and the zero state response
y(t) − y part (t). We have that solution is a particular solution satis-
fying zero initial conditions and the
L(y(t) − y part (t)) = Ly(t) − Ly part (t) = f (t) − f (t) = 0 correct forcing term.

so y(t) − y part (t) = yhomog (t) is a solution to the homogeneous

problem.

In summary, if we can solve the homogeneous problem and find

one particular solution then we know the general solution.

Example 2.4.1. Find the general solution to

y′′ + y = t

It isn’t hard to see that if y(t) = t; y′ (t) = 1; y′′ (t) = 0 and so y′′ + y =
0 + t = t. So the general solution is the particular solution y p (t) = t plus a
solution to the homogeneous problem. The homogeneous problem is

y′′ + y = 0

which has the solution yhomog (t) = A sin t + B cos t. Therefore the general
solution to
y′′ + y = t
is given by y = t + A sin t + B cos t.
74 differential equations

Example 2.4.2. Find the zero input response and zero state response to

y′ + y = 1 y (0) = 2

The zero input response satisfies the differential equation

y′zi + yzi = 0 yzi (0) = 2

which has the solution

yzi (t) = 2e−t .

The zero state response satisfies the differential equation

y′zs + yzs = 1 yzs (0) = 0

It is not hard to guess that a particular solution is y part (t) = 1 which gives
the zero state response as
yzs = 1 − e−t .

The complete solution is the sum

y(t) = (1 − e−t ) + 2e−t = 1 + e−t

The question now is how to find particular solutions. The first

method that we will present is the the method of undetermined co-
efficients, which works for constant coefficient differential equations
with a righthand side taking a particular form.

2.5 The method of undetermined coefficients

The method of undetermined coefficients is a way to solve certain

non-homogeneous linear differential equations of the form

Ly = f (t)

where L is a constant coefficient differential operator - in other words

dn d n −1
L= + p 1 + . . . pn
dtn dtn−1
with p1 , p2 , . . . pn constant, and f (t) is a forcing term that can be
written as an

• Polynomial in t

• sin(ωt) or cos(ωt)

• Exponential e at ,
 higher order linear equations 75

or sums and products of terms of this form (but not compositions or

ratios!). For instance it would work for f (t) of any of the following
forms:

f (t) = cos(t)
f ( t ) = t5
f (t) = e−6t
f (t) = et cos(t) + t11 sin(t) − 256 sin3 (3t)
f (t) = t3 e−t cos(t) + t2 sin(5t) − 11
f (t) = et − 22t cos(5t)

But it would not work for f (t) of the following forms

2
f (t) = et
f (t) = tan(t)
cos(t)
f (t) =
t

The basic idea can be summarized as follows:

It takes a bit of experience to be able
Method 2.5.1 (Undetermined Coefficients (Naive)). To solve a constant to see what is the right form. We will
shortly talk about the annihilator
coefficient linear differential equation of the form method, which will always give you
the correct form to guess, but is a little
n −1
dn di bit more involved to apply in practice.
Ly = (
dtn
+ ∑ p i
dti
)y = f (t) Note that it is not a problem to include
i =0 extra terms, as you will find in the
course of applying the method that the
where f (t) is given by sums and products of coefficients of the unnecessary terms
are zero.
• Exponentials e at

• Sine or Cosine sin ωt or cos ωt

• Polynomial Pn (t) = ∑ ai ti .

You should

• Make a guess (ansatz) for y(t) in the same form as f (t) , with unde-
termined coefficients A1 , A2 , . . . An .

• Substitute your guess into the differential equation.

• Solve for the coefficients Ai . These should be constants.

If you include a sin t term you should
As an example consider the differential equation always include the cos t term, and
vice-versa. Basically this is because
the form is closed under derivatives –
y′′ + 3y′ − 4y = sin t.
the derivative of a linear combination
of sin t and cos t is again a linear
combination of sin t and cos t.
76 differential equations

The forcing term f (t) = sin t, so we guess something in the same

form, a trigonometric function y(t) = A1 sin t + A2 cos t. Substituting
this guess into the original equation gives

y = A1 sin t + A2 cos t
y′ = A1 cos t − A2 sin t
y′′ = − A1 sin t − A2 cos t
y′′ + 3y′ − 4y = −(5A1 + 3A2 ) sin t + (−5A2 + 3A1 ) cos t.

We want to have y′′ + 3y′ − 4y = sin t, so we need

− (5A1 + 3A2 ) = 1
(−5A2 + 3A1 ) = 0

since that will give us the result that we want. We can solve for A1
5 3
and A2 to find A1 = − 34 ; A2 = − 34 .

Exercise 2.5.1
Suppose that in the example above one tried a solution of the form
y = A1 sin t + A2 cos t + A3 e2t . Show that you must have A3 = 0.

It can sometimes be a little confusing as to the proper form of the

guess, so here is a table that you might find helpful. Note that we
will see in a second that these forms occasionally need to be modified
depending on the solutions to the homogeneous equation.

(Usually!) The form of a solution for the method of undetermined

coefficients.
f (t) y(t)
f (t) = tk y(t) = A0 + A1 t + A2 t2 . . . Ak tk = Pk (t)
f (t) = eσt y(t) = Aeσt
f (t) = sin ωt or f (t) = cos ωt y(t) = A sin ωt + B cos ωt
f (t) = tk sin ωt or f (t) = tk cos ωt y = Pk (t) sin ωt + Qk (t) cos ωt
f (t) = eσt sin ωt or f (t) = eσt cos ωt y(t) = A eσt sin ωt + B eσt cos ωt
f (t) = tk eσt y = Pk (t) eσt
f (t) = t e sin ωt or f (t) = tk eσt sin ωt
k σt y = Pk (t) e sin ωt + Qk (t) eσt cos ωt
σt

Exercise 2.5.2
(1) State the form for the solution by undetermined coefficients and
(2) find the solution. Does the “naive” version work for the last prob-
lem?

a) y′′′ + 5y′ − 6y = e5t cos(4t)

b) y′′ + 2y′ + y = cos(t) + t2 et
 higher order linear equations 77

c) y′′′ + 5y′ − 6y = e2t

d) y′′ + 3y′ + 2y = e−t
Unfortunately the “naive” version of undetermined coefficients
does not always work. The last example in the previous exercise
shows that sometimes the naive method as described here fails: the
solution is not quite of the same form as the righthand side. Rather
than being an exponential it is t times an exponential, te−t . This
suggests the following rule of thumb.
Method 2.5.2 (Better Undetermined Coefficients Method). To solve a
constant coefficient differential equation

L(y) = f (t)
guess y(t) of the same form as f (t) unless one or more of the terms of your
guess is itself a solution to the homogeneous equation

Ly = 0
in this case multiply these terms by the smallest power of t such that none of
the terms in your guess satisfy the homogeneous equation.
Example 2.5.1. Consider the differential equation
d2 y
− y = t2 et + sin t
dt2
Normally when we see t2 et we would guess y = At2 et + Btet + Cet +
D sin t + E cos t. In this case the solutions to the homogeneous problem
d2 y
dt2
− y = 0 has two linearly indpendent solutions y1 = et and y2 = e−t .
That means that we should try a solution of the form y = At3 et + Bt2 et +
Ctet + D sin t + E cos t. Note that we don’t multiply the sin t or cos t terms
by t, as they are not solutions to the homogeneous equation.
Example 2.5.2. Consider the differential equation
d4 y d3 y d2 y dy
− 2 + 2 − 2 + y = t + 5 sin t + 3et + 2tet
dt4 dt3 dt2 dt
The characteristic equation for the homogeneous problem is

r4 − 2r3 + 2r2 − 2r + 1 = 0

This has four roots. r = ±i are simple roots and r = 1 is a double root. This
gives four solutions to the homogeneous problem

y1 (t) = sin t
y2 (t) = cos t
y3 ( t ) = e t
y4 (t) = tet
78 differential equations

Normally for the right-hand side f (t) = t + 5 sin t + 3et + 2tet we would
guess something of the following form

y(t) = A + Bt + C sin t + D cos t + Eet + Ftet .

| {z } | {z } | {z }
t 5 sin t 3et +2tet

Here the underbraces show the term(s) in f (t) that are responsible for the
terms in y(t). However we have some exceptions to make here: some of the
terms in our guess are solutions to the homogeneous problem: The func-
tions sin t, cos t, et and tet all solve the homogeneous equation. We should
mulitply the solutions by the smallest power of t so that no terms in the
guess solve the homogeneous equation. The function A + Bt doesn’t
solve the homogeneous problem, so we don’t need to change these terms. The
functions sin t, cos t do, but t sin t and t cos t do not, so we multiply these
terms by t. The functions et and tet both solve the homogeneous equation.
If we multiply by t we get Etet + Ft2 et . One of these terms still solves the
homogeneous problem. If we multiply by t2 we get Et2 et + Ft3 et , none of
which solves the homogeneous problem. Thus we should guess

y(t) = A + Bt + Ct sin t + Dt cos t + Et2 et + Ft3 et .

| {z } | {z } | {z }
t 5 sin t 3et +2tet

The particular solution works out to be

5 1 1
y(t) = 2 + t + t sin t + t2 et + t3 et .
4 4 6
Finding the correct form of the solution in the method of undeter-
mined coefficients becomes a bit cumbersome when the characteristic
equation has roots of high multiplicity. There is a variation of this
method, usually called the annihilator method. This method is a little
more work but it is always clear what the correct form of the solution
should be. This is the subject of the next section.

2.6 The Annihilator Method

Now to present this in a slightly different perspective: the right-

hand sides that are allowed are those for which the derivatives are
in the same form: the derivative of a polynomial is a polynomial,
the derivative of an exponential is an exponential, the derivative of
A1 cos(t) + A2 sin(t) is a linear combination of cos(t) and sin(t),
and the same for their products. Since all of these functions have the
property that the derivative is of the same form it follows that a con-
stant coefficient linear combination of them is also of the same form.
So it makes sense to look for a solution that is also in the same form.
 higher order linear equations 79

Another way to think of these functions is as the functions f (t)

that are allowed are those which can arise from solving a constant co-
efficient homogeneous differential equation. This is the basic idea of
the method of annihilators: one finds an “annihilator”, a differential
operator that annihilates the righthand side.
The basic algorithm is as follows:

Method 2.6.1 (Annihilator Method). To solve the linear constant coeffi-

cient inhomogeneous differential equation

Ly = f (t)

where f (t) is given by sums and products of polynomials, exponentials,

sines and cosines.

1. Find a set of linearly independent solutions to the homogeneous problem

Ly = 0.

2. Find a linear constant coefficient operator L̃ such that L̃ f = 0

3. Act on both sides of the equation with L̃ to get

L̃Ly = L̃ f = 0

4. Find a set of linearly independent solutions to the homogeneous problem

L̃Ly = 0.

5. The solution will be in the form of a linear combination of all functions

in (4) that DO NOT APPEAR IN (1)

Example 2.6.1. Solve the equation

d2 y
+ y = et
dt2
using annihilators.
To do this we first find the solution to the homogeneous problem. The
d2 y
characteristic equation for dt2 + y = 0 is r2 + 1 = 0. Solving for r gives
r = ±i or y1 (t) = sin t, y2 (t) = cos t
We next need to find an annihilator, something that “kills” the righthand
d
side. The operator dt − 1 does the trick: if we act on the function et with this
d
operator we get zero. Acting on the above equation with dt − 1 gives

y′′′ − y′′ + y′ − y = 0

The characteristic equation is r3 − r2 + r − 1 = 0. This has three roots

r = ±i, r = 1. Three linearly independent solutions are y1 = sin t; y2 =
cos t; y3 = et . The first two are solutions to the original equation, but the
80 differential equations

third is not. Thus we should try a particular solution of the form y = Aet .
Substituting into the original equation gives
y′′ + y = 2Aet = et
Thus 2A = 1 and A = 21 . Note that if we mistakenly included the other
terms B sin t + C cos t we would find that B and C are arbitrary.
Example 2.6.2. Find the correct form of the solution for the method of
undetermined coefficients for the equation
d6 y d4 y d2 y
+ 3 + 3 + y = 2 sin t + 5 cos t
dt6 dt4 dt2
The homogeneous equation is
d6 y d4 y d2 y
+ 3 + 3 +y = 0
dt6 dt4 dt2
so that the characteristic equation is
r6 + 3r4 + 3r2 + 1 = 0
which can be written as (r2 + 1)3 = 0, so r = ±i are roots of multiplicity 3.
This means that the six linearly independent solutions are
y1 = cos t
y2 = t cos t
y3 = t2 cos t
y4 = sin t
y5 = t sin t
y6 = t2 sin t
d2
The annihilator for A sin t + B cos t is L̃ = dt2
+ 1. Acting on both sides
of the equation with the annihilator gives
d2 d6 y d4 y d2 y
( + 1 )( + 3 + 3 + y) = 0.
dt2 dt6 dt4 dt2
The characteristic equation is
(r 2 + 1)4 = 0
This has eight linearly independent solutions:
y1 = cos t
y2 = t cos t
y3 = t2 cos t
y4 = t3 cos t
y5 = sin t
y6 = t sin t
y7 = t2 sin t
y8 = t3 sin t
 higher order linear equations 81

d 2 d6 y d4 y d2 y
There are two solutions of ( dt 2 + 1)( dt6 + 3 dt4 + 3 dt2 + y ) = 0 that do not
d6 y d4 y d2 y
solve dt6 + 3 dt4 + 3 dt2 + y = 0. These are t3 sin t and t3 cos t. Therefore
the particular solution can be assumed to take the form At3 sin t + Bt3 cos t.

For most cases this method is overkill – it is typically easier to use

the naive method of undetermined coefficients – but it can sometimes
be tricky to figure out what the correct guess should be. Here is a list
of various forcing terms f (t) and their annihilators.

f (t) Annihilator
d
1 dt
d k +1
Pk (t) dtk+1
d
e at dt − a
d2
A sin ωt + B cos ωt dt2
+ ω2
Ae at sin ωt + Be at cos ωt ( dtd − a)2 + ω 2
2
Pk (t) sin ωt + Qk (t) cos ωt ( dtd 2 + ω 2 )k+1
Pk (t)e at sin ωt + Qk (t)e at cos ωt (( dtd − a)2 + ω 2 )k+1

Note that Pk (t) and Qk (t) denote polynomials of degree k. Note

that the coefficients of the polynomials P, Q could be different.

2.6.1 Problems: Undetermined Coefficients

Exercise 2.6.1
Find particular solutions to the following differential equations
dy dy
a) dt + y = et b) dt + 3y = sin(t) + e2t
dy d2 y
c) dt + y = t sin(t) + e−t d) dt2
+ y = te−t
d2 y
e) dt2
+ 2 dy
dt + 5y = e
−t + cos( t )

Exercise 2.6.2
Find annihilators for the following forcing functions

a) f (t) = t2 + 5t
b) f (t) = sin(3t)
c) f (t) = 2 cos(2t) + 3t sin(2t)
d) f (t) = t2 et + sin(t)

Exercise 2.6.3
Find the correct form of the guess for the method of undetermined
coefficients. You do not need to solve for the coefficients (although
you may– answers will be given in the answer section.)
82 differential equations

a) y′′ + y = et + 1 b) y′ + y = sin(t)
d7 y 5
c) y′′ + y = sin(t) d) dt7
+ 32 ddt5y − 18 dy
dt + 11y =
3 2
11t + t + 7t + 2
d7 y 5 2 d5 y 4 3 d2 y
e) dt7
+ 32 ddt5y − 18 ddt2y = 180t3 f) dt5
+ 3 ddt4y + 3 ddt3y + dt2
=
12t2 + 6t + 8 − 3e−t

Exercise 2.6.4
Find a particular solution to the following equations using the
method of undetermined coefficients.

a) y′′ − y′ − 3y = 3t2 + 8t − 6 b) y′′ + 4y′ + 3y = −3t2 + t + 16

c) y′′ − 2y = −6t2 + 4t + 6 d) y′′ + 4y′ + 2y = −29 cos(3t) +
33 sin(3t)
e) y′′ − 3y = −4 sin(t) f) y′′ + 5y′ − 2y = 8 cos(t) +
2 sin(t)
g) y′′ + y = 4 cos(t) − 6 sin(t) h) y′′ + 4y = 12 cos(2t) −
4 sin(2t)
i) y′′′ − 3y′′ + 4y = 9e−t + 6e2t j) y′′′ = 3y′ + 2y = 27e−2t + 6et

2.7 Variation of Parameters

The method of undetermined coefficients is usually the easiest way to

solve a non-homogeneous linear differential equation when it is ap-
plicable, but there are many equations to which it does not apply. A
more general method is variation of parameters. The method of vari-
ation of parameters will solve any non-homogeneous linear equation
provided we can solve the homogeneous equation.
For second order equations the basic formula is given by the fol-
lowing theorem:

Theorem 2.7.1. Consider the differential equation

d2 y dy
+ a1 ( t ) + a0 ( t ) y = f ( t )
dt2 dt
Suppose that y1 (t) and y2 (t) are two linearly independent solutions to the
homogeneous problem. Then a particular solution to the non-homogeneous
problem is given by
Z t Z t
y1 ( s ) f ( s ) y2 ( s ) f ( s )
y p ( t ) = y2 ( t ) ds − y1 (t) ds
a W (s) a W (s)

where W (s) is the Wronskian of y1 , y2 . There is some freedom in choosing

the lower limit of integration a as to make the integrals finite.
 higher order linear equations 83

Proof. The basic idea is to look for a solution in the form

y = A ( t ) y1 ( t ) + B ( t ) y2 ( t )

Since we now have two unknowns, A(t) and B(t), but only one equa-
tion, we will need a second equation in order to have a unique solu-
tion. Differentiating gives

y′ = A′ (t)y1 (t) + A(t)y1′ (t) + B′ (t)y2 (t) + B(t)y2′ (t).

We impose the condition A′ (t)y1 (t) + B′ (t)y2 (t) = 0. This leaves us

with

y′ = A(t)y1′ (t) + B(t)y2′ (t)

y′′ = A′ (t)y1′ (t) + A(t)y1′′ + B′ (t)y2′ (t) + B(t)y2′′ (t)

Substituting into the original equation gives

A′ (t)y1′ (t) + B′ (t)y2′ (t) = f (t).

Combining this with the earlier condition we get a system of two

equations for the two unknowns A′ , B′

A ′ ( t ) y1 ( t ) + B ′ ( t ) y2 ( t ) = 0
A′ (t)y1′ (t) + B′ (t)y2′ (t) = f (t)

Solving this for A′ (t), B′ (t) gives

f ( t ) y2 ( t )
A′ (t) = − (2.1)
W ( y1 , y2 )
f ( t ) y1 ( t )
B′ (t) = . (2.2)
W ( y1 , y2 )

Equations (2.1) and (2.2) can be integrated up to find A(t) and

B(t). Note that the constants of integration give the homogeneous
solution.

This method actually extends to linear equations of any order,

provided that we are able to solve the corresponding homogeneous
problem.

Theorem 2.7.2. Consider the nth order linear differential equation

dn y d n −1 y
n
+ a n −1 ( t ) n −1 + . . . + a 0 ( t ) y = f ( t )
dt dt
Suppose that y1 (t), y2 (t), . . . yn (t) are n linearly independent solutions to
the homogeneous problem. Then a particular solution to the non-homogeneous
problem is given by
y p ( t ) = ∑ Ai ( t ) yi ( t )
84 differential equations

where the functions Ai (t) solve the following system of equations

y1 y2 ... yn A1′ (t) 0

    
 ′   . 
y1′ y2′ ... y′n   A2 (t)   .. 

.

.. .. .. ..  .  = 
 .  

. . . .  .   0 
 

y1n−1 y2n−1 ... ynn−1 A′n (t) f (t)

Example 2.7.1. Consider the differential equation

d2 y 1 dy 4
+ − 2y = t
dt2 t dt t
The homogeneous equation is equidimensional and can be solved by looking
for a solution of the form y(t) = tα . Substituting in the homogeneous
equation we get

α(α − 1)tα−2 + αtα−2 − 4tα−2 = 0

which has solutions α = 2 and α = −2. Therefore two linearly independent

solutions of the homogeneous equation are y1 (t) = t2 and y2 (t) = t−2
Using variation of parameters we now look for a particular solution of the
form
y = A ( t ) y1 ( t ) + B ( t ) y2 ( t )

The Wronskian of y1 and y2 is given by W = −4t−1 and using (2.1) and

(2.2) we have
1 t4
A′ (t) = , B′ (t) = −
4 4
Integrating and substituting back into y p we get

t3
yp =
5

Exercise 2.7.1
Use variation of parameters to find a particular solution for each
equation. If the particular solution can also be found using undeter-
mined coefficients, use both methods to check your answer.

a) y′′ + y′ − 6y = et
b) y′′ + 1t y′ − 9
t2
y = 7t2 (See example above)

c) y′′ + y′ − 2y = 10 cos (t)

1
d) y′′ + y = cos (t)

e) y′′ − 6y′ + 9y = 2e3t

f) y′′′ − y′′ − 2y′ = 4t (Note that this is third order)
 higher order linear equations 85

2.8 The Laplace Transform

There is another method available for solving constant coeffi-

cient differential equations, called the Laplace transform. The Laplace
transform L is defined as follows

Definition 2.8.1. If f (t) is a function defined on [0, ∞) that grows at most

exponentially fast then the Laplace transform of f is defined to be
Z ∞
F (s) = L( f ) = e−st f (t)dt.
0

The somewhat surprising fact, which we will not prove here, is

that the Laplace transform is invertible. The inversion formula in-
volves a certain integral in the complex plane, called the Mellin inte-
gral, which we will not discuss here. Most important for us is the fact
that the transformation is invertible.
Let us begin with a couple of examples.

Example 2.8.1. We have the following Laplace transform pairs

f (t) = e at F (s) = s−1 a
f (t) = cos(bt) F (s) = s2 +s b2
b
f (t) = sin(bt) F (s) = s2 + b2
To see the first note that we have

e−(s− a)t ∞
Z ∞ Z ∞
1
F (s) = L(e at ) = e−st e at dt = e−(s− a)t dt = |0 =
0 0 −(s − a) s−a

assuming that s > a in order that the integral is convergent.

The second two formulae follow from the first together with the Euler
formula. Taking a = ib in the first identity we have that
Z ∞
1 s + ib
e−st eibt dt = = 2
0 s − ib s + b2

Taking the real part of this equation gives

Z ∞
s
e−st cos(bt)dt =
0 s2 + b2
while taking the imaginary part gives
Z ∞
b
e−st sin(bt)dt =
0 s2 + b2
assuming in each case that s > 0.

We have one more important observation, namely that derivatives

transform nicely under the Laplace transform.
86 differential equations

Lemma 2.8.1. The Laplace transform of f ′ (t) satsifies

L( f ′ ) = sL( f ) − f (0)

This is easily seen by integration by parts. We have

Z ∞
′
L( f ) = e−st f ′ (t)dt Q = e−st ; dP = f ′ (t)dt; dQ = −se−st dt; P = f (t)
0
Z
= f (t)e−st |0∞ − (−s) f (t)e−st dt = sF (s) − f (0)

Note that this identity can be iterated to give higher derivatives:

L( f ′′ ) = sL( f ′ ) − f ′ (0) = s(sL( f ) − f (0)) − f ′ (0) = s2 L( f ) − s f (0) − f ′ (0)

If we have the constant coefficient linear differential equation

My = f (t)

where M is a constant coefficient linear differential operator then

taking the Laplace transform will lead to an equation of the form

P ( s )Y ( s ) + Q ( s ) = F ( s )

where

• P(s) is the characteristic polynomial, of degree n

• Q(s) is a polynomial of degree at most n that arises from the initial

conditions after applying identities like L( f ′ ) = sL − f (0).

• F (s) is the Laplace transform of f (t).

We can solve this for Y (s) to find that

F (s) Q(s)
Y (s) = −
P(s) P(s)
Transforming back gives

F (s) Q(s)
y ( t ) = L −1 ( ) − L −1 ( )
P(s) P(s)
These two terms are precisely the “zero input” and “zero state” solu-
tions. The quantity
Q(s)
yh (t) = −L−1 ( )
P(s)
satsifies the homogeneous equation My = 0 along with the given
initial conditions. The term
F (s)
y p ( t ) = L −1 ( )
P(s)
solves My = f along with zero initial conditions. As a first example
we will solve a constant coefficient linear homogeneous equation.
 higher order linear equations 87

Example 2.8.2. Solve

y′′ (t) + y(t) = 0 y(0) = 1; y′ (0) = 0

using the Laplace transform. Taking the Laplace transform gives

L(y′′ ) + L(y) = 0
s2 L(y) − sy(0) − y′ (0) + L(y) = 0
s2 L(y) − s + L(y) = 0
s
Y (s) = L(y) = 2
s +1

From our very small table of Laplace transforms we know that L(cos(bt) =
s
s2 + b2
. Taking b = 1 we can conclude that

L(cos(t)) = L(y) 7→ y(t) = cos(t)

More commonly the Laplace transform is used to solve inhomoge-

neous problems. Here is an example. It will be helpful to know the
identity
k!
L(tk e at ) =
( s − a ) k +1
Example 2.8.3. Solve

y′′ + 2y′ + y = te−t y(0) = 0; y ′ (0) = 0

using the Laplace transform. Taking the Laplace transform of both sides
gives

1
(s2 + 2s + 1)Y (s) = L(te−t ) =
( s + 1)2
1
Y (s) =
( s + 1)4

From the identity above we have that

3! 6
L(t3 e−t ) = =
( s + 1)4 ( s + 1)4

and so we can conclude that

1
L(y) = L(t3 e−t )
6

and thus that y(t) = 16 t3 e−t .

Below is a list of Laplace transform identities that might prove

useful. The list includes both abstract properties as well as the
Laplace transform of particular functions.
88 differential equations

Function Laplace Transform

f (t) F (s)
1
1 s
k!
tk s k +1
k!
tk e− at ( s + a ) k +1
b
sin(bt) b2 + s2
s
cos(bt) s2 + b2
b
e− at sin(bt) b2 +(s+ a)2
s+ a
e− at cos(bt) b2 +(s+ a)2
f (t) F (s)
f (t) + g(t) F (s) + G (s)
f ′ (t) sF (s) − f (0)
f ′′ (t) s2 F ( s ) − s f (0) − f ′ (0)
dk f
dtk
s k F ( s ) − s k −1 f (0 ) − s k −2 f ′ (0 ) − . . . f ( k −1) (0 )
t f (t) − F ′ (s)
tk f (t) (−1)k F (k) (s)
at
e f (t) F (s − a)
1
R∞
t f (t) s F ( σ ) dσ
This table gives enough information to find the inverse transform
of any rational function. The basic observation is the same one that
allows one to compute the integral of any rational function: namely
that any rational function can be written as a linear combination
of functions of the form (s−1a)k , ((s− as)−2 a+b2 )k and ((s− a)12 +b2 )k . If the
roots are simple (multiplicity one) then k may always be taken to be
zero although if there are roots of higher multiplicity one will need
expressions of this form for k > 1. We next give an example where
we must use partial fractions to find the solution

Example 2.8.4. Solve the initial value problem

y′′ + 2y′ + 2y = et y(0) = 0; y ′ (0) = 1

Taking the Laplace transform gives

  1
s2 Y (s) − sy(0) − y′ (0) + 2 (sY (s) − y(0)) + 2Y (s) =
s−1
1
(s2 + 2s + 2)Y (s) − 1 =
s−1
1 1
Y (s) = +
s2 + 2s + 1 (s − 1)(s2 + 2s + 2)

1
At this point we must re-express (s−1)(s2 +2s+2)
using partial fractions.
 higher order linear equations 89

We have the partial fractions expansion

1 A Bs C
= + 2 + 2
(s − 1)(s2 + 2s + 2) s − 1 s + 2s + 2 s + 2s + 2
1 = A(s2 + 2s + 2) + B(s)(s − 1) + C (s − 1)

One can write down equations corresponding to the s2 , s and 1 terms but
it is easier to plug in a couple of s values. If substitute s = 1 in the above we
find that
1
5A = 1 A= .
5
Similarly if we substitute s = 0 in we find that
2 3
1 = 2A − C C = 2A − 1 = −1 = −
5 5
Finally if we take the s2 terms we find that
1
0 = A+B B=−
5
This gives
1 1 1 s 3 1 1 1 1 1 s 2 1
Y (s) = L(y) = − − + = − +
5 s − 1 5 s2 + 2s + 2 5 s2 + 2s + 2 s2 + 2s + 2 5 s − 1 5 s2 + 2s + 2 5 s2 + 2s + 2
Note that s2 + 2s + 2 = (s + 1)2 + 1 by completing the square. From the
table we can see that
s+a b
L(e at cos(bt)) = L(e at sin(bt)) =
( s + a )2 + b2 ( s + a )2 + b2
We will need to slightly rewrite one of the terms
1 s 1 s+1−1 1 s+1 1 1
− =− 2 =− 2 +
5 s2 + 2s + 2 5 s + 2s + 2 5 s + 2s + 2 5 s2 + 2s + 2
This gives
1 1 1 s+1 3 1 1 1 3
Y (s) = L(y) = − + = L(et ) − L(e−t cos(t)) + L(e−t sin(t))
5 s − 1 5 ( s + 1)2 + 1 5 ( s + 1)2 + 1 5 5 5
From this we can see that
1 1 3
y(t) = et − e−t cos(t) + e−t sin(t)
5 5 5

2.8.1 Laplace Transform Problems

Exercise 2.8.1
Given y(t) solving the specified differential equation find the Laplace
transform Y (s). You do not need to compute the inverse transform.
a) y′ + y = sin(t) + 6 y (0) = 2
b) y′′ + y = e2t + t y(0) = 3; y ′ (0) = 2
c) y′′′′ + y = te3t y(0) = 0; y′ (0) = 0; y′′ (0) = 0; y′′′ (0) = 0
90 differential equations

Exercise 2.8.2
Find the inverse Laplace transform of the following functions:
1 s
a) Y (s) = s −1 b) Y (s) = s2 +1
1 5s+7
c) Y (s) = ( s +3)5
d) Y (s) = s2 +4
4s+3 2s+5
e) Y (s) = ( s −2)2 +1
f) Y (s) = s2 −1
4s2 −9s−4
g) Y (s) = s(s−1)(s+2)

Exercise 2.8.3
Solve each initial value problem using the Laplace Transform

a) y′′ − 4y′ + 3y = e2t , y(0) = 1, y ′ (0) = 2

b) y′′ + 4y′ + 3y = 9t, y(0) = 0, y ′ (0) = −5
3
Mechanical and electrical oscillations

3.1 Mechanical oscillations

3.1.1 Undamped mass-spring systems

Mechanical systems are ultimately governed by Newton’s

law, F = ma. For simple (one degree of freedom) systems this means
that the governing equations are second order, since the acceleration
is the second derivative of the position. The force in an undamped
mass-spring system is provided by Hooke’s law, which says that the
force is proportional to the displacement. This gives the equation
for a mass m and spring with spring constant k with no damping or
external forcing as
mx ′′ = −kx.
The general solution takes the form
r ! r !
k k
x (t) = A cos t + B sin t
m m

This is what we expect: in the absence of damping the

qsolutions
k
oscillate periodically, with a natural frequency ω0 = m , a quantity
that will be important in our later analysis. A couple of things to note about this
In the case where there is an external forcing f (t)the equation formula. First a bit of units analysis.
The spring constant k has units kg/s2 =
becomes N/m while the mass m has units kg, so
mx ′′ = −kx + f (t). that ω0 has units s−1 . This is what we
expect: the quantity ω0 t should be in
Often one can assume that the forcing is some periodic function. For radians, which are unitless.
It is also good to think
q about the
instance for an automobile suspension one source of forcing is the
k
scaling: since ω0 = m this implies
rotation of the wheels, which is a periodic phenomenon. Similarly if
that large masses oscillate more slowly
one drives over a grooved road at a constant speed this generates a (lower frequency) as do smaller spring
periodic forcing. It is simplest to model this by a simple sin(ωt) or constants. This jibes with our intuition
about how an oscillating mass should
cos(ωt) term, so we get a model like behave.

mx ′′ = −kx + sin(ωt)
92 differential equations

Let’s try to solve this. We know from the method of undetermined

coefficients that the correct guess should (usually!) be to try to find a
particular solution of the form

xpart (t) = A1 cos(ωt) + A2 sin(ωt)

Substituting this into the equation give

′′
mxpart + kxpart = (k − mω 2 )( A1 cos(ωt) + A2 sin(ωt))
= sin(ωt)

1
Thus we need to choose A1 = 0 and A2 = k−mω 2 . The general
solution is the sum of the particular solution and the solution to the
homogeneous problem, so
r ! r !
k k sin(ωt)
x (t) = A cos t + B sin t +
m m (k − mω 2 )

You should notice a few things about this solution.

• The solution has positive amplitude for ω 2 small - it is in phase

with the driver.

• The solution has negative amplitude for ω 2 large - it is π out of

phase with the driver.
q
k
• When the driving frequency ω is close to m the amplitude of the
response becomes large.
q
k
• The amplitude is undefined at ω = ± m = ± ω0 .
For some dramatic incidents of un-
This last case is called “resonance”. In engineering and the sci- planned resonance see the footage of
ences resonance is an extremely important phenomenon. In some the collapse of the Tacoma Narrows
Bridge in Washington (AKA gallop-
situations resonant is a good thing; if one is designing a concert hall ing Gerty) or the oscillations of the
or a musical instrument or a speaker cabinet having the correct res- Millenium Bridge in London.
onant frequencies can greatly improve the sound. When designing
structures such as buildings that are subject to external forces reso-
nance is usually something to be avoided. Either way it is important
to understand the phenomenon and how to cause it (or prevent it).
While it may look like it from the formula it is not actually true that
the amplitude is infinite when ω 2 = mk . Rather, since the forcing term
is itself a solution to the homogenous equation we need to look for
a solution of a different form. We know what to guess in this case: a
linear combination of t sin(t) and t cos(t). It is easy to check that this
works.
 mechanical and electrical oscillations 93

Exercise 3.1.1
Verify that a particular solution to
r !
k
mx ′′ = −kx + F0 sin t
m

is given by
q 
k
t cos mt
xpart (t) = − F0 √
2 km

Exercise 3.1.2
Suppose that the suspension of a car with bad shock absorbers can
be modelled as a mass-spring system with no damping. Assume
that putting a 100kg mass into the trunk of the car causes the car to
sink by 2cm, and that the total mass of the car is 1000kg. Find the
resonant frequency of the suspension, in s−1 .
You will first need to find the spring constant of the suspension.
For simplicity take the gravitational acceleration to be g = 10ms−2 .

3.1.2 Mass-spring systems with damping

In real physical systems one usually has some form of damping

as well. In the context of a mass-spring system damping is usually
modelled by a term proportional to the velocity of the system, dx
dt ,
but having the opposite sign. This acts to slow the movement of the
mass and decrease oscillations. In mechanical systems a dashpot is
a component that provides a damping force, usually assumed to be
linear in velocity and of opposite sign. Screen doors on homes, for
example, frequently have a dashpot to keep the door from slamming
shut. As a second example the suspension of a car has two main
components, springs and shocks. The springs are basically Hooke’s
law springs, and shocks are dashpots that provide a damping force.
In the case where we consider a mass-spring system with linear
damping Newton’s law becomes

d2 x dx
m = −kx − γ + f (t)
dt2 dt

where, as always, f (t) represents some external forcing term. Again it is worthwhile doing some
First we look at the homogeneous equation. dimensional analysis. The coefficient γ,
when multiplied by a velocity, should
give a force, so γ should have units
d2 x dx kg s−1 . Notice that from the three
m + γ + kx = 0.
dt2 dt quantities m, γ, k we can form the
quantity √γ , which is dimensionless.
mk
This is a measure of how important
damping is in the system. If √γ is
mk
small then damping is less important
compared with inertial (mass-spring)
effects. If √γ is large it means that
mk
inertial effects are small compared with
damping. This will become important
in our later analysis.
94 differential equations

The characteristic equation is

mr2 + γr + k = 0

which has roots

γ2 − 4km
p
−γ ±
r= .
2m

There are two cases here, with very different qualitative behaviors.
Figure 3.1: The solutions to the
in the case γ2 < 4km, or equivalently √γ < 2, the characteristic 2
km equation ddt2x + γ dx
dx + 4x = 0
polynomial has a complex conjugate pair of roots. In this situation
with x (0) = 1; x ′ (0) = 0 for
the two linearly independent solutions are given by
γ = 1, 4, 5 (shown in red, blue,
r ! and black respectively) rep-
γ
− 2m t k γ2 resenting the underdamped,
x1 ( t ) = e cos − t (3.1)
m 4m2 critically damped, and over-
damped cases. Note that the
r !
γ
− 2m t k γ2
x2 ( t ) = e sin − t (3.2) solution decays fastest in the
m 4m2
critically damped case.

In this case the solutions to the homogeneous problem consist of ex-

ponentially decaying oscillations. This is called the “underdamped”
case due to the presence of these sustained oscillations – in mechani-
cal systems such sustained oscillations are usually unwanted.
In the case γ2 ≥ 4km, or √γ ≥ 2, the roots of the characteristic
km
polynomial are real and negative, and the solutions to the homoge-
neous problem consist of decaying exponentials, with no sustained
oscillations. When γ2 > 4km the roots are distinct, and this is known
as the “overdamped” case. When γ2 = 4km the roots are equal, and
this is known as the “critically damped” case (See Exercise 3.1.4 for
the exact form of the solutions.) In most mechanical situations one
would like to be in the critically damped or slightly overdamped
case. In engineering
√
the reciprocal quantity
Q = γkm is sometimes denoted as
the quality factor or Q-factor of the
system, and measures the quality of the
resonator, with a high Q representing a
Exercise 3.1.3 system with strong resonance. Exam-
ples of mechanical systems where one
Let us revisit an exercise from the previous subsection. Suppose that would like to have a high Q include
musical instruments, tuning forks, etc.
the suspension of a car can be modeled as a mass-spring system The Q for a tuning fork is typically of
with damping. Assume that putting a 100kg mass into the trunk of the order of 1000, so it will vibrate for a
the car causes the car to sink by 2cm, and that the total mass of the long time before the sound dies out.

car is 1000kg. Assume that for best performance one would like the
suspension of the car to be critically damped. How should the value
of the damping coefficient γ, in kg s−1 , be?
 mechanical and electrical oscillations 95

Exercise 3.1.4
Verify that, in the critically damped case, the two solutions to the
homogeneous problem are given by
γ
x1 (t) = e− 2m t
γ
x2 (t) = te− 2m t

Now let’s consider the inhomogeneous damped harmonic oscilla-

tor with a sinusoidal external forcing term

d2 x dx
m 2
+ γ + kx = F0 sin(ωt).
dt dx
If we look for a solution of the form

x = A1 cos(ωt) + A2 sin(ωt)

and substitute this in to the equation we get the following set of

equations for A1 and A2 :

(k − mω 2 ) A1 + γωA2 = 0 (3.3)
2
(k − mω ) A2 − γωA1 = F0 . (3.4)

The first equation comes from requiring that the coefficients of the
cos(ωt) terms sum to zero, the second from demanding that the
coefficients of the sin(ωt) terms sum to F0 . The solution to these two
linear equations is given by
 
(k − mω 2 )2 + γ2 ω 2 A1 = F0 γω (3.5)
 
(k − mω 2 )2 + γ2 ω 2 A2 = F0 (k − mω 2 ) (3.6)

This solution is often expressed in a somewhat different way. One

can use the angle addition formulas

sin(ωt + ϕ) = sin(ωt) cos(ϕ) + cos(ωt) sin(ϕ)

cos(ωt + ϕ) = cos(ωt) cos(ϕ) − sin(ωt) sin(ϕ)

to express a linear combination of sin(ωt) and cos(ωt) in the form

A1 cos(ωt) + A2 sin(ωt) = A sin(ωt + ϕ).

The amplitude A and phase ϕ are given by

q
A = A21 + A22
ϕ = arctan( A1 /A2 ).
96 differential equations

One can think of the mechanical system as transforming the forc-

ing term by changing the amplitude and phase of the sinusoid. This
point of view is particularly prevalent when talking about RLC cir-
cuits, which we will cover next. In our case using the expression for
A1 , A2 derived above we find that
Figure 3.2: A plot of the magni-
F0 tude of the complex amplitude
A= p
(k − mω 2 )2 + γ2 ω 2 A(ω ) as a function of ω for
k = 1, m = 1 and different
 
γω
ϕ = arctan .
k − mω 2 values of γ between γ = 0 and
γ=2
Figure 3.2 shows the plot of A as a function of ω for k = 1, m = 1
and various values of γ between γ = 0 and γ = 2. Notice that there
are two different behaviors depending on the damping coefficient γ.
√
When γ < 2km the curveqhas a local minimum at ω = 0 and a
− γ2
√
global maximum at ω = ± 2km 2m2
. When γ > 2km the curve has
a global maximum at ω = 0 and it decreases as |ω | increases. The
√
magenta curve represents the value γ = 2km.
Figure 3.3 shows a graph of the phase shift ϕ = arctan( k−γω mω 2
) as
a function of ω for k = 1, m = 1 and various values of the damping
γ. The phase shift varies from 0 to π. It is always ϕ = π2 at frequency
q
k
ω = m , and for small damping (high Q) the phase looks approxi-
q
mately like a step function – it is close to zero for ω < mk and close
q
to π for ω > mk .
In the case of weak damping ( √γ ≪ 1) this means that if the
km Figure 3.3: A plot of ϕ =
system is driven at frequencies below the "natural" frequency ω =
q
k
arctan( k−γω
mω 2
) as a function
then the response will be in phase with the forcing term, while if
m of ω for k = 1, m = 1 and differ-
the system is driven at a frequency above the natural frequency then
ent values of γ between γ = 0
the response will be 180◦ out of phase with the forcing.
and γ = 2

Exercise 3.1.5
Suppose that a car with bad shocks can be modeled as a mass-spring
system with a mass of m = 750kg, a spring constant of k = 3.75 × 105
N/m, and a damping coefficient of γ = 2 × 104 kg s−1 , and that the
car is subject to a periodic forcing of f (t) = 1000N sin(150t) due to an
unbalanced tire. Find the particular solution. What is the amplitude
of the resulting oscillations?
 mechanical and electrical oscillations 97

3.1.3 Problems: Mechanical Oscillations

Exercise 3.1.6
Find the homogeneous solutions

a) y′′ + 4y = 0
b) y′′ + y = 0
c) y′′ + 6y′ + 10y = 0
d) y′′ + 5y′ + 6y = 0

Exercise 3.1.7
Solve the following initial value problems

a) y′′ + 2y′ + y = 0 y (0) = 0 y ′ (0) = 1

b) y′′ + 3y′ + 2y = 0 y (0) = 1 y ′ (0) = 0
c) y′′ + 6y′ + 10y = 0 y (0) = 0 y ′ (0) = 1
d) y′′ + y′ + 2y = 0 y (0) = 1 y ′ (0) = 1

Exercise 3.1.8
Solve the initial value problems

a) y′′ + 4y = cos(t) y (0) = 0 y ′ (0) = 0

b) y′′ + y = cos(2t) y (0) = 1 y ′ (0) = 0
c) y′′ + 5y = cos(t) y (0) = 0 y ′ (0) = 1
d) y′′ + 6y = cos(3t) y (0) = 0 y ′ (0) = 1

Exercise 3.1.9
Solve the following initial value problems

a) y′′ + 2y′ + y = cos(t) y (0) = 0 y ′ (0) = 0

b) y′′ + 3y′ + 2y = sin(t) y (0) = 1 y ′ (0) = 0
c) y′′ + 6y′ + 10y = sin(t) + cos(t) y (0) = 0 y ′ (0) = 0
d) y′′ + y′ + 2y = 4 sin(t) y (0) = 1 y ′ (0) = 1

Exercise 3.1.10
Find the general solution to the homogeneous damped harmonic
oscillator equation
d2 y dy
m 2 +γ + ky = 0
dt dt
for the following parameter values. In each case classify the equation
as overdamped, underdamped or critically damped.
98 differential equations

a) m = 20.0 kg, γ = 40.0 Ns/m, k = 100.0 N/m

b) m = 25.0 kg, γ = 50.0 Ns/m, k = 25.0 N/m
c) m = 10.0 kg, γ = 50.0 Ns/m, k = 60.0 N/m
d) m = 10.0 kg, γ = 10.0 Ns/m, k = 30.0 N/m

Exercise 3.1.11
Solve the following initial value problem

d2 y dy
m + γ + ky = 0
dt2 dt
for the following sets of initial values and parameters.

a) m = 20.0 kg, γ = 40.0 Ns/m, k = 100.0 N/m y (0) =

0 m, y′ (0) = 5 m/s
b) m = 25.0 kg, γ = 50.0 Ns/m, k = 25.0 N/m y (0) =
2 m, y′ (0) = 0 m/s
c) m = 10.0 kg, γ = 50.0 Ns/m, k = 60.0 N/m y (0) =
1 m, y′ (0) = 1 m/s
d) m = 10.0 kg, γ = 10.0 Ns/m, k = 30.0 N/m y (0) =
2 m, y′ (0) = −1 m/s

Exercise 3.1.12
Suppose that a car suspension can be modeled as damped mass-
spring system with m = 1500 kg and spring constant k = 40 000 Ns/m.

a) What are the units of γ?

b) How large should the damping coefficient γ be chosen so that
the system is critically damped?

Exercise 3.1.13
Suppose that a car suspension can be modeled as damped mass-
spring system with m = 2000 kg. Also suppose that if you load 600 kg
in the car the height of the suspension sinks by 1 cm. How large
should the damping coefficient be so that the suspension is critically
damped for the unloaded car? Recall Hooke’s law, F = −k∆x, and be
careful to keep consistent units.

Exercise 3.1.14
Consider the damped, driven harmonic oscillator

d2 y dy
4 + + 4y = cos(ωt)
dt2 dt
 mechanical and electrical oscillations 99

what frequency ω produces the largest amplitude response (largest

amplitude of the particular solution).

Exercise 3.1.15

Consider the differential equation

d2 y
+y = 0 y(0) = cos(s) y′ (0) = − sin(s)
dt2

a) Show that y(t) = cos(t + s) is a solution by checking that it

satisfies both the differential equation and the initial conditions.

b) Show that y(t) = cos(s) cos(t) − sin(s) sin(t) is a solution to the

same equation by taking the general solution A cos(t) + B sin(t)
and solving for the coefficients A, B.

c) Since solutions are unique conclude the trigonometric addition

formula cos(t + s) = cos(s) cos(t) − sin(s) sin(t)

d) Use the same style of argument to prove that sin(t + s) =

cos(t) sin(s) + sin(t) cos(s)

Exercise 3.1.16

Consider the damped driven harmonic oscillator

my′′ + γy′ + ky = cos(ωt)

a) Show that the particular solution is given by y(t) = ℜ( Aeiωt ),

where A = iγω +(k1−mω2 ) . Here ℜ denotes the real part.

b) The amplitude A is largest in magnitude where the the denom-

inator is smallest in magnitude. Show that the squared magni-
tude of the denominator is given by f (ω ) = γ2 ω 2 + (k − mω 2 )2

c) Show that f (ω ) always has a critical point at ω = 0, and has

two more if γ2 < 2km
100 differential equations

Exercise 3.1.17

[A]

[B]
 mechanical and electrical oscillations 101

[C]
The figures above depict the amplitude, as a function of frequency ω,
for the particular solution to

my′′ + γy′ + ky = cos(ωt)

for three different sets of parameter values:

a) m = 1, γ = 1, k = 4
b) m = 1, γ = 2, k = 4
c) m = 1, γ = 4, k = 4

identify which graph corresponds to which list of parameters,

102 differential equations

3.2 RC and RLC circuits

3.2.1 RC Circuits

Another important application for second order constant

coefficient differential equations is to RLC circuits, circuits containing
resistance (R), inductance (L) and capacitance (C). As we will see the
role that these quantities play in circuits is analogous to the role that
damping coefficient γ, the mass m and the spring constant k play in
mass-spring systems.
We begin with a very simple circuit consisting of a resistor and a
capacitor in series with a battery and a switch, as in Figure 3.5. Sup-
pose that the battery supplies a voltage V0 and that at time t = 0
the switch is closed, completing the circuit. We would like to derive
the differential equation governing the system, and solve it. To begin
with let VR denote the voltage drop across the resistor and VC de-
note the voltage drop across the capacitor. These should add up (by Figure 3.4: The unit of capaci-
Kirchhoff’s law) to the voltage supplied by the battery: tance, the farad, is named after
Michael Faraday. Despite grow-
VR + VC = V0 . ing up poor and leaving school
at an early age to apprentice to
Now both VR and VC can be expressed in terms of the current. If a bookbinder Faraday became
I (t) denotes the current at time t then we have that one of the preeminent scientists
VR = IR of his day.
Z t
1
VC = I (s)ds
C 0

In order to get a differential equation we must differentiate VR +

VC = V0 and apply the fundamental theorem of calculus to find that

dVR dV dI 1 dV
+ C = R + I = 0 = 0,
dt dt dt C dt
since the battery supplies a constant voltage. Thus the current satis-
1
fies R dI
dt + C I = 0. To find the initial condition we note that at time
1 0
t = 0 the voltage drop across the capacitor is VC = C 0 I (t)dt = 0,
R
Figure 3.5: A simple circuit
and thus VR (0) = RI (0) = V0 , and the differential equation becomes consisting of a battery, switch,
dI 1 V0 resistor and capacitor.
R + I=0 I (0) = .
dt C R
This can be solved to give

V0 − t
I (t) = e RC ,
R
from which the other quantities can be derived. For instance VR =
t
IR = V0 e− RC . The voltage drop across the capacitor, VC can be found
 mechanical and electrical oscillations 103

1
Rt
either by using VC + VR = V0 or by taking VC = C 0 I ( s ) ds and
t
− RC
doing the integral. Both methods give VC = V0 (1 − e ). Note that
a resistance R ohms times a capacitance of C Farads gives time in
seconds. This is usually called the RC time constant – one common
application of RC circuits is to timing.

Exercise 3.2.1
A circuit consists of a 5V battery, a 5 kΩ resistor, a 1000µF capacitor
and a switch in series. At time t = 0 the switch is closed. At what
time does VC , the voltage drop across the capacitor, equal 4V?
A somewhat more interesting problem is when the voltage is not
constant (DC) but instead varies in time, for instance, the voltage
varies sinusoidally as V (t) = V0 sin(ωt), as in Figure 3.6. Essentially
the same derivation above gives the equation Figure 3.6: A simple circuit
consisting of a resistor, a capac-
dI 1 dV (t) itor, and a sinusoidally varying
R + I= = V0 ω cos(ωt).
dt C dt voltage.
This can be solved by the method of undetermined coefficients, look-
ing for a particular solution of the form A1 cos(ωt) + A2 sin(ωt).
Alternatively one can use the formula derived in the section on mass
spring systems with the replacements m → 0, γ → R, k → C1 and
F0 → V0 ω. Either way we find the solution

V0 RC2 ω 2 V0 ωC
Ipart (t) = 2
sin(ωt) + cos(ωt).
(ωRC ) + 1 (ωRC )2 + 1
t
The homogeneous solution Ihomog (t) = Ae− RC is exponentially In many problems with damping such
as this the homogeneous solution is
decaying, so we will assume that enough time has passed that this
exponentially decaying. This is often
term is negligibly small. called the "transient response" and can,
From here any other quantity of interest can be found. For in- in many situations, be neglected.

stance VC , the voltage drop across the capacitor, is given by Again note that ω has units s−1 and
RC has units s so ωRC is dimension-
less. Dimensionless quantities are an
VC = VR − V (t) = RI (t) − V0 sin(ωt)
important way to think about a system,
V0 sin(ωt) V RωC cos(ωt) since they do not depend on the sys-
= − 0 tem of units used. If ωRC is small it
1 + (ωRC )2 1 + (ωRC )2
means that the period of the sinusoid
V0 is much longer than the time-constant
= p sin(ωt + ϕ)
1 + (ωRC )2 of the RC circuit, and it should be-
have like a constant voltage. In this
case, where ωRC ≈ 0 it is not hard to
where ϕ = arctan(−ωRC ).
see that VC ≈ V0 sin(ωt). If ωRC is
large, on the other hand, the voltage
drop across the capacitor will be small,
3.2.2 RLC Circuits, Complex numbers and Phasors V cos(ωt)
VC ≈ − 0 ωRC . This is the simplest
example of a low-pass filter. Low fre-
quencies are basically unchanged, while
There is a third type of passive component that is interesting high frequencies are damped out.
from a mathematical point of view, the inductor. An inductor is a
104 differential equations

coil of conducting wire: the voltage drop across an inductor is pro-

portional to the rate of change of the current through the inductor.
The proportionality constant, denoted by L, is the inductance and
is measured in Henrys. The derivation of the response of an RLC
circuit to an external voltage is similar to that given for the RC cir-
cuit. Kirchhoff’s las implies that the total voltage drop across all of
the components is equal to the imposed voltage. Denoting the volt-
age drops across the inductor, resistor and capacitor by VI , VR , VC
respectively we have that

dI (t) 1
Z
VL + VR + VC = L + RI (t) + I = V ( t ).
dt C
Taking the derivative gives a second order differential equation for
the current i (t) It is again worth thinking a bit about
d2 I ( t ) dI 1 dV units. The quantity RC has units of
L 2
+R + I = . (3.7) time – seconds if R is measured in
dt dt C dt Ohms and C in Farads. The quantity
L
Note that all of the results of the section on mechanical oscillations R is also a unit of time, seconds if L is
measured in Henrys and R in Ohms.
translates directly to this equation, with inductance being analo- 2
The quantity RLC is dimensionless. The
gous to mass, resistance to damping coefficient, the reciprocal of equation is overdamped, with exponen-
capacitance to the spring constant, current to displacement and the 2
tially decaying solutions, if RLC > 4 and
derivative of voltage to force. Therefore we will not repeat those cal- is underdamped, with solutions in the
form of an exponentially damped sine
culations here. Rather we will take the opportunity to introduce a or cosine, if R2 C
< 4.
L
new, and much easier way to understand these equations through
the use of complex numbers and what are called "phasors". While
it requires a little bit more sophistication the complex point of view
makes the tedious algebraic calculations that we did in the previous
section unnecessary: all that one has to be able to do is elementary
operations on complex numbers.
In many situations one is interested in the response of a circuit to
a sinusoidal voltage. The voltage from a wall outlet, for instance, is
sinusoidal. If the voltage is sinusoidal, V (t) = V0 cos(ωt) then from
(3.7) the basic differential equation governing the current I (t) is

d2 I ( t ) dI 1 dV
L 2
+R + I = = −ωV0 sin(ωt).
dt dt C dt
We could use the method of undetermined coefficients here, as we
did in the previous section, and look for a particular solution in the
form I (t) = A cos(ωt) + B sin(ωt) but it is a lot easier to use the
Euler formula and complex arithmetic. Recall the Euler formula says

eiωt = cos(ωt) + i sin(ωt).

We will begin by replacing the real voltage V (t) = V0 cos(ωt) with

a complex voltage Ṽ (t) = V0 eiωt , so that Re(V0 eiωt ) = V0 cos(ωt).
Here Re denotes the real part. This seems strange but it is just a
 mechanical and electrical oscillations 105

mathematical trick: since the equations are linear one can solve for
a complex current, Ĩ, and then take the real part, and that gives the
particular solution. In other words the equation

d2 I ( t ) dI 1 dV 
iωt

L + R + I = = V0 Re iωe
dt2 dt C dt
is the same as
 2
d Ĩ (t) d Ĩ (t) 1 dV
  
Re L 2
+ R + Ĩ ( t ) = = Re iωV0 eiωt
dt dt C dt
and because it is linear we can instead solve
d2 Ĩ (t) d Ĩ (t) 1 dṼ
L +R + Ĩ (t) = = iV0 ωeiωt
dt2 dt C dt
and then take the real part of the complex current Ĩ (t). This will be a
lot simpler for the following reason: instead of looking for a solution
as a linear combination of cos(ωt) and sin(ωt) we can just look for a
solution in the form Ĩ (t) = Aeiωt . The constant A will be complex but
we won’t have to solve any systems of equations, etc – we will just
have to do complex arithmetic. When we are done we just have to
take the real part of the complex solution Ĩ (t) and we get the solution
to the original problem.
To begin we consider the case of an RC circuit with imposed volt-
age V (t) = V0 cos(ωt) :
dI 1 dV
R + I= = −ωV0 cos(ωt).
dt C dt

Complexifying the voltage Ṽ (t) = V0 eiωt we get ddtṼ = iωV0 eiωt and
thus
d Ĩ 1 dṼ
R + Ĩ = = iωV0 eiωt .
dt C dt
Using undetermined coefficients we look for a solution in the form
Ĩ (t) = I0 eiωt . Since this guess already contains both the sin(ωt) and
cos(ωt) terms we do not need any additional terms. This is what
makes the method easier. Substituting this into the equation gives It is worth noting here that RCω is
1
dimensionless, so ωC has units of
1 Ohms.
(iωR + ) I0 eiωt = iωV0 eiωt
C
which is the same as
iωV0
I0 =
iωR + C1
V0
= i
R− ωC
i
R+ ωC
= V0 1
R2 + ω 2 C2
106 differential equations

So the complex current is just the complex voltage multiplied by the

i
R+ ωC
complex number 1 . This should remind you of Ohm’s law:
R2 +
ω 2 C2
V = IR or I = V i
R . In fact we have V0 = ( R − ωC ) I0 so it looks
like Ohm’s law with a complex resistance, called impedance. If the
1
capacitance is really big, so ωC is really small then this is exactly the
usual Ohm’s law.
Recall that complex numbers can be identified with points in the
plane. The function V0 eiωt represents a point that rotates counter-
clockwise around a circle of radius V0 at a uniform angular fre-
quency. Also recall that when we multiply complex numbers the
magnitudes multiply and the arguments add. The complex num- Figure 3.7: The particular so-
i
ber (impedance) ( R − ωC ) lies in the fourth quadrant. Since V0 = lution for an RC-circuit with a
i
( R − ωC ) I0 this means that the voltage always lags behind the cur- sinusoidal voltage. The current
rent, or equivalently the current leads the voltage. Remember that we Ĩ (t) will lead the voltage by an
set things up so that the voltage and current rotate counterclockwise. angle between 0 and π2 . As time
i
If R = 0 then V0 = (− ωC ) I0 and the voltage is exactly π2 behind the increases the picture rotates
current. If C is large then V0 ≈ I0 R and the voltage and the current counterclockwise but the angle
are in phase. between the voltage and the
Next let’s consider an LR circuit, which is dominated by resistance current does not change.
and inductance, again with a voltage Ṽ (t) = V0 cos(ωt). This time we
are solving
d2 Ĩ d Ĩ
L 2 +R = iV0 ωeiωt .
dt dt
Again looking for a solution of the form Ĩ (t) = I0 eiωt we find

(−ω 2 L + iωR) I0 eiωt = iωV0 eiωt

or
V0 R − iωL
V0 = ( R + iωL) I0 or I0 = = 2 V0
( R + iωL) R + ω 2 L2
Since the complex number (impedance) ( R − iωL) lies in the fourth
quadrant this implies that the voltage leads the current, or equiva-
lently the current lags the voltage. This is illustrated in the marginal
figure.
Of course any real circuit will have inductance, resistance and Figure 3.8: The particular so-
capacitance. In this case we have lution for an RL-circuit with a
sinusoidal voltage. The current
d2 Ĩ d Ĩ 1
L 2
+ R + Ĩ = iV0 ωeiωt . Ĩ (t) will lag behind the voltage
dt dt C
by an angle between 0 and π2 .
Looking for a solution of the form Ĩ (t) = I0 eiωt we find that As time increases the picture
1 1 rotates counterclockwise but
(−ω 2 L + iωR + ) I0 = iωV0 or V0 = I0 ( R + i (ωL − )).
C ωC the angle between the volt-
The impedance ( R + i (ωL − ωC1
)) can either be in the fourth or the age and the current does not
first quadrant, depending on the frequency ω and the sizes of the change.
 mechanical and electrical oscillations 107

1
inductance L and the capacitance C. If ωL > ωC or ω 2 LC > 1 then
1
( R + i (ωL − ωC )) lies in the first quadrant and the voltage leads the
current. If ω 2 LC < 1 then the voltage lags the current. If ω 2 LC = 1
then the impedance is real and the voltage and the current are in
phase. We have encountered this condition before – this is (one)
definition of effective resonance.
The relative phase between the voltage and the current in an AC
system is important, and is related to the "power factor". For reasons
of efficiency it is undesirable to have too large of an angle between
the voltage and the current in a system: the amount of power that
can be delivered to a device is Vrms Irms cos(θ ), where θ is the angle
between the current and the voltage. If this angle is close to π2 then
very little power can be delivered to the device since cos(θ ) is small.
Many industrial loads, such as electric motors, have a very high There is a mnemonic "ELI the ICE
inductance due to the windings. This high inductance will usually be man", to remember the effects of induc-
tance and capacitance. In an inductive
offset by a bank of capacitors to keep the angle between the voltage circuit (L is inductance) the voltage
and the current small. E leads the current I. In a capacitive
circuit (C is capacitance) the current I
Example 3.2.1. Solve the linear constant coefficient differential equation leads the voltage E.

d3 y dy
+ 4 + 5y = cos(2t)
dt3 dt
using complex exponentials.
We can replace this with the complex differential equation

d3 z dz
+ 4 + 5z = e2it
dt3 dt

and then take the real part. Looking for a solution in the form z(t) = Ae2it
we find that

A((2i )3 + 4(2i ) + 5)e2it = e2it

5Ae2it = e2it
1
A= .
5
Notice that in this case the constant A worked out to be purely real, al-
though in general we expect it to be complex. This gives a particular solu-
tion z = 15 e2it . Taking the real part gives y(t) = Re(z(t)) = Re(e2it /5) =
1
5 cos(2t ).
The homogeneous solution is given by the solution to

d3 z dz
+ 4 + 5z = 0.
dt3 dt
The characteristic polynomial is

r3 + 4r + 5 = 0
108 differential equations

√
1± 19 i
which has the three roots r = −1, r = 2 . Thus the general solution to

d3 y dy
3
+ 4 + 5y = cos(2t)
dt dt
is
√ √
1 t 19t t 19t
y = cos(2t) + A1 e−t + A2 e 2 cos( ) + A3 e 2 sin( )
5 2 2
or alternatively in terms of complex exponentials as
√ √
1 1+ i 19 1−i 19
y= cos(2t) + A1 e−t + A2 e 2 t
+ A3 e 2 t
5
Example 3.2.2. Solve the differential equation

d3 y d2 y dy
− 3 +2 = sin(t)
dt3 dt2 dt
As before we can solve this by solving the complex equation

d3 z d2 z dz
3
− 3 2 + 2 = eit
dt dt dt
and taking the imaginary part. Looking for a solution in the form z(t) =
Aeit we find that

A((i )3 − 3(i )2 + 2(i ))eit = eit (3.8)

A (3 + i ) = 1 (3.9)
1 3− i
A= 3+ i = 10 . (3.10)

The particular solution is thus

3 − i it
y(t) = Im( e )
10
3−i
= Im( (cos(t) + i sin(t))
10
3 1
= sin(t) − cos(t)
10 10

The characteristics polynomial for the homogeneous equation is given by

r3 − 3r2 + 2r = 0

which has roots r = 0, r = 1, r = 2. This gives the general solution as

3 1
y(t) = sin(t) − cos(t) + A1 + A2 et + A3 e2t
10 10
4
Systems of Ordinary Differential Equations

4.1 Introduction

Systems of differential equations arise in many ways. In some

cases the differential equations of physics and engineering are most
naturally posed as systems of equations. For instance in the absence
of charges and currents Maxwell’s equations for the electric field
E( x, t) and the magnetic field B( x, t) are given by Throughout this section quantities in
lower-case bold will generally represent
∂B vector quantities, while those in upper
= −∇ × E case bold represent matrices. In this
∂t case, however, we keep with tradition
∂E and represent the electric and mag-
µ 0 ϵ0 = ∇ × B.
∂t netic fields with upper case E and B
respectively.

This is a system of six partial differential equations relating the first

partial derivatives of the 3-vectors E( x, t)and B( x, t).
Another example from physics concerns the vibrations of a molecule,
for instance carbon dioxide, which is an example of a linear tri-
atomic molecule. Treating the molecule classically and the bonds
as linear springs leads to the following system of three second order
differential equations for the positions of the carbon atom (x2 ) and
the two oxygen atoms (x1,3 ):

d2 x1
mO = − k ( x1 − x2 )
dt2
d2 x2
m C 2 = − k ( x2 − x1 ) − k ( x2 − x3 )
dt
d2 x3
mO 2 = − k ( x 3 − x 2 ) .
dt
In epidemiology a basic set of models for the spread of a conta-
gious disease are the SIR models. There are many different variations
of the SIR models, depending on various modeling assumptions, but
110 differential equations

a one common set of models is the following:

dS
= − βIS
dt
dI
= βIS − kI
dt
dR
= kI
dt

In this model the quantities S(t), I (t), R(t) represent the population
of susceptible, infected and recovered people in the population. The
constants β and k are constants representing the transmission rate
and recovery rate of for the disease in question, and together the
solution to this system of equations provides a model for under-
standing how a disease propagates through a population.
In other situations it may be desirable to write a higher order
equation as a system of first order equations. Take, for instance, the
second order equation

d2 y dy
2
+ p ( t ) + q ( t ) y = f ( t ).
dt dt

dy
If we introduce the new variable z(t) = dt and using the fact that
dz d2 y
dt = then the second order equation above can be written as the
dt2
system of equations

dy
=z
dt
dz
+ p ( t ) z + q ( t ) y = f ( t ).
dt

This system of equations can be written as

dv
= A ( t ) v + g ( t ).
dt

Here the vector quantities v(t), g (t) and the matrix A(t) are given by
" # " # " #
y(t) 0 0 1
v(t) = g (t) = A(t) = .
z(t) f (t) −q(t) − p(t)

More generally if we have an nth order equation

dn y d n −1 y dy
+ p n − 1 ( t ) + . . . + p1 ( t ) + p0 ( t ) y = f ( t )
dtn dtn−1 dt

we can do the change of variables to rewrite this as n first order

 systems of ordinary differential equations 111

equations
dy
z1 ( t ) =
dt
dz1 d2 y
z2 ( t ) = (= 2 )
dt dt
..
.
dzn−2 d n −1 y
z n −1 ( t ) = (= n−1 )
dt dt
dzn−1
= − p n −1 ( t ) z n −1 − p n −2 z n −2 − . . . − p 1 z 1 − p 0 y + f ( t ).
dt
This is a set of n first order equations in y, z1 , z2 , . . . zn−1 . There are,
of course, many ways to rewrite an nth order equation as n first order
equations but this is in some sense the most standard. In many situ-
ations, particularly if one must solve the equations numerically, it is
preferable to write a single higher order equation as a system of first
order equations.

4.2 Existence and Uniqueness

We will briefly describe some existence and uniqueness results for

systems of differential equations. These should look very familiar, as
they are basically vectorized versions of the results that we presented
earlier for scalar differential equations. We first present the nonlinear
result.

Theorem 4.2.1. Consider the nonlinear system

dy
= f (y, t) y(t0 ) = y0
dt
There exists a solution in some interval (t0 − ϵ, t0 + ϵ) if the (vector-valued)
function f (y, t) is continuous in a neighborhood of the point (y0 , t0 ). The
∂f
solution is unique if the partial derivatives ∂yi are continuous in a neigh-
j
borhood of the point (y0 , t0 ). The notation here is that f is a vector with
components f 1 , f 2 , . . . and that y is a vector with components yi . Thus in
order to prove existence and uniqueness we must check the continuity of the
∂f
n components of f and the n2 partial derivatives ∂yi .
j

As in the scalar case we will not spend much time on nonlinear

equations of order greater than first. From this point on we will
mainly focus on linear equations. The most general linear equation
can be written
dv
= A ( t ) v + g ( t ). (4.1)
dt
All of the terminology and results from the section on higher order
linear equations carries over naturally to systems of linear first order
112 differential equations

equations: in fact in many cases the definitions and formulae look

more natural in the setting of a first order system. As is the case for
scalar equations the existence and uniqueness theorem for linear
equations is somewhat simpler than that for nonlinear equations.

Theorem 4.2.2. Consider the initial value problem

dv
= A(t)v + f (t) v(t0 ) = v0 . (4.2)
dt
Suppose that A(t), f (t) depends continuously on t in some interval I con-
taining t0 . Then (4.2) has a unique solution in the interval I.

4.2.1 Homogeneous linear first order systems.

As in the case of scalar equations a homogeneous equation is one for
which there is no forcing term – no term that is not proportional to y.
If the case of a first-order system a homogeneous equation is one of
the form
dv
= A(t)v. (4.3)
dt
From the previous existence and uniqueness theorem we have a
solution as long as the matrix-valued function A(t) is continuous.
Next we define two important notions, the idea of a Wronskian and
linear independence and the idea of a fundamental solution matrix.

Definition 4.2.1. Suppose that v1 (t), v2 (t), v3 (t) . . . vn (t) are n solutions
to equation 4.3. The Wronskian W (t) of these solutions is defined to be the
determinant of the matrix with columns v1 (t), v2 (t), v3 (t) . . . vn (t). We say
that the vectors are linearly independent if W (t) ̸= 0.

At this point it is worth doing an example to verify that this def-

inition of the Wronskian agrees with our previous definition of the
Wronskian.

Example 4.2.1. Suppose that we have a second order equation

d2 y dy
+ p(t) + q(t)y = 0
dt2 dt
with y1 (t) and y2 (t) two linearly independent solutions. Previously we
defined the Wronskian to be W (t) = y1 (t)y2′ (t) − y1′ (t)y2 (t). We can write
this second order equation as a pair of first order equations by introducing a
dy
new variable z(t) = dt . As we showed earlier y(t), z(t) satisfy

dy
=z
dt
dz
= − p(t)z − q(t)y.
dt
 systems of ordinary differential equations 113

Since y1 (t), y2 (t) are two solutions to the original second order equation we
have that two solutions to the system of equations are given by
! ! ! !
y1 ( t ) y1 ( t ) y2 ( t ) y2 ( t )
v1 (t) = = v2 (t) = = .
z1 ( t ) y1′ (t) z2 ( t ) y2′ (t)

The Wronskian of these two solutions is given by

!
y1 ( t ) y2 ( t )
W (t) = det = y1 (t)y2′ (t) − y1′ (t)y2 (t)
y1′ (t) y2′ (t)

in agreement with the previous definition. More generally if we take an nth

order linear equation and rewrite it as a system of n first order equations
in the usual way then the Wronskian of a set of solutions to the nth order
equation will agree with the Wronskian of the corresponding solutions to the
system.
Definition 4.2.2. Suppose that v1 (t), v2 (t), v3 (t) . . . vn (t) are n solu-
tions to equation 4.3 whose Wronskian is non-zero. We define a funda-
mental solution matrix M (t) to be a matrix whose columns are given by
v1 (t), v2 (t), v3 (t) . . . vn (t).
We note a few facts about M (t).
• The determinant of a fundamental solution matrix M (t) is the
Wronskian W (t). A fundamental solution matrix is invertible since
the Wronskian is non-zero.

• There are many choices of a fundamental solution matrix, depend-

ing on how one chooses the set of solutions. Perhaps the most
convenient choice is the one where M (t0 ) = I, the identity ma-
trix. Given any fundamental solution matrix the one which is the
identity at t = t0 is given by M (t) M −1 (t0 ).

• The solution to the homogeneous system v′ = A(t)v; v(t0 ) = v0

is given by v(t) = M (t) M −1 (t0 )v0 .
The next theorem that we need is Abel’s theorem. Recall that if A
is a matrix then the trace of A, denoted Tr( A), is equal to the sum of
the diagonal elements of A.
Theorem 4.2.3 (Abel’s Theorem). Suppose that v1 (t), v2 (t), . . . vn (t) are
solutions to the system
dv
= A(t)v.
dt
Let W (t) be the Wronskian of those solutions. Then W (t) satisfies the first
order equation
dW
= Tr( A(t))W.
dt
Here Tr( A(t) is the trace of the matrix A == the sum of the diagonal ele-
ments.
114 differential equations

In the case where the system comes from a higher order linear
equation of the form

dn y d n −1 y d n −2 y dy
n
+ p n − 1 ( t ) n 1
+ p n − 2 ( t ) n 2
+ · · · + p1 ( t ) + p0 ( t ) y = 0
dt dt − dt − dt
the matrix A(t) takes the form

0 1 0 0... 0
 
 0 0 1 0... 0 
 
 . .. .. ..
 ..
A(t) = 

. . .


 0 0 ... 0 1
 

− p0 ( t ) − p1 ( t ) − p 2 ( t ) . . . − p n −2 ( t ) − p n −1 ( t )

so that Tr ( A(t)) = − pn−1 (t) and Abel’s theorem implies that dWdt =
− pn−1 (t)W.
As in the case of Abel’s theorem for a nth order equation the main
utility of Abel’s theorem is the following corollary, which tells us
that (assuming that the coefficients are continuous) the Wronskian is
either never zero or it is always zero.

Corollary 4.2.1. Suppose that v1 (t), v2 (t), . . . vn (t) are solutions to the
system
dv
= A(t)v.
dt
in some interval I in which A(t) is continuous. Then the Wronskian W (t)
is either never zero in the interval I or it is always zero in the interval I.

Proof. The proof follows simply from Abel’s theorem. Suppose that
there exists a point t0 ∈ I such that W (t0 ) = 0. From Abel’s theorem
the Wronskian satisfies the differential equation

dW
= Tr( A(t))W W ( t0 ) = 0
dt
One solution is obviously W (t) = 0 but A(t) is continuous, and
therefore by our existence and uniqueness theorem this is the only
solution.

It is also worth noting once again that the Wronskian is non-zero

at a point t0 if and only if any vector x can be respresented as a linear
combination of v1 (t0 ), v2 (t0 ), . . . vn (t0 ).

4.3 Variation of Parameters

Generally if one is able to solve a homogeneous linear differential

equation then one can solve the corresponding inhomogeneous equa-
tion. This fact is known as the "variation of parameters" formula.
 systems of ordinary differential equations 115

Variation of parameters is one instance where the results for systems

of first order equations look much simpler than the corresponding
results for higher order equations. Here we can see one of the benefits of
solving a first order system of equa-
tions, rather than a single nth order
Theorem 4.3.1 (Variation of Parameters). Consider the non-homogeneous
equation. Both the solution formula
linear equation itself as well as the proof greatly resem-
ble the integrating factor method for a
dv first order linear inhomogeneous equa-
= A(t)v + g (t) v(t0 ) = v0 . (4.4) tion. The derivation of the variation of
dt
parameters formula for a higher-order
scalar equation, on the other hand, is
Suppose that one can find n linearly independent solutions v1 (t), v2 (t) . . . vn (t) much less transparent.
to the corresponding homogeneous equation

dv
= A(t)v.
dt

Define the matrix M (t) to be the matrix with column {vi (t)}in=1 . Then the
general solution to (4.4) is given by
Z t
v(t) = M (t) M −1 (s) g (s)ds + M (t) M (t0 )−1 v0
t0

Proof. The proof is quite similar to the integrating factor method for
a first order scalar equation. We have the equation

dv
= A(t)v + g (t) v(t0 ) = v0
dt

and the fundamental solution matrix M (t), satisfying

dM
= A(t) M.
dt

We define a new variable w(t) via v(t) = M (t)w(t). From this

definition it follows that

dv dM dw
= w(t) + M (t)
dt dt dt
dw
= A(t) Mw(t) + M (t)
dt
dw
= A(t)v + M (t)
dt

Substituting this into the differential equation for v(t) gives

dw
A(t)v + M (t) = A(t)v + g (t)
dt
dw
M (t) = g (t)
dt
dw
= M −1 ( t ) g ( t )
dt
116 differential equations

Integrating this equation from t0 to t and applying the fundamental

theorem gives
Z t
w ( t ) − w ( t0 ) = M −1 (s) g (s)ds
t0
Z t
w(t) = M −1 (s) g (s)ds + w(t0 )
t0

Using the fact that v = Mw or w = M −1 v gives

Z t
v(t) = M (t) M −1 (s) g (s)ds + M (t) M −1 (t0 )v0
t0

4.4 Constant coefficient linear systems.

There are, unfortunately, very few systems of ordinary differential

equations for which one can write down the complete analytical solu-
tion. One important instance when we can write down the complete
solution is the case of constant coefficients.

Definition 4.4.1. A constant coefficient homogeneous linear system of

differential equations is one that can be written in the form

dv
= Av. (4.5)
dt
where the matrix A does not depend on t.
We can find a number of solutions to a constant coefficient equa-
tion, perhaps even all of them, using basic linear algebra.

Lemma 4.4.1. Suppose that the matrix A has a set of eigenvectors {uk }
with corresponding eigenvalues λk :

Auk = λk uk

Then the functions

vk (t) = uk eλk t

are solutions to (4.5). If the eigenvalues are distinct then the solutions vk (t)
are linearly independent.

Example 4.4.1. Consider the constant coefficient linear homogeneous

equation
!
dv 1 2
= v.
dt 2 1
 systems of ordinary differential equations 117

The matrix A has eigenvalues of λ1 = −1, with eigenvector (1, −1) T and
λ2 = 3, with eigenvector (1, 1) T . This gives
!
1
v1 (t) = e−t
−1
!
1 3t
v2 (t) = e
1

Since the eigenvalues are distinct these solutions are linearly independent.

Unfortunately matrices do not in general always have a complete

set of eigenvectors. One example of this is the matrix
!
1 1
A= .
0 1

The characteristic equation is (λ − 1)2 = 0 so λ = 1 is an eigenvalue

of (algebraic) multiplicity 2, but there is only one linearly indepen-
dent vector such that Au = u, namely u = [0, 1] T . This should not be
too surprising – we have already seen that for higher order equations
if the characteristic polynomial has a root of multiplicity higher than
one then we may have to include terms like teλt , t2 eλt , etc. For sys-
tems the annihilator method does not really apply, so instead we give
an algorithm for directly computing the fundamental solution matrix.
First a definition

Definition 4.4.2. For a constant coefficient homogeneous linear differential

equation
dv
= Av
dt
we define the matrix exponential M (t) = e At to be the fundamental solution
matrix satisfying the initial condition M(0) = I, where I is the n × n
identity matrix. Alternatively the matrix exponential can be defined by the
power series
∞ k k
A2 t2 A3 t3 t A
e At = I + At + + +... = I + ∑
2! 3! k =1
k!

where A2 = AA, A3 = AAA, and so on.

It is usually not practical to directly compute the matrix expo-
nential from the power series, but for simple examples it may be
possible.

Example 4.4.2. Find the matrix exponential for the matrix

!
0 −1
A=
1 0
118 differential equations

using the power series definition. We begin by computing the first few
powers.
!
−1 0
A2 = = −I
0 −1
!
0 1
A3 = = −A
−1 0
!
1 0
A4 = =I
0 1
!
0 −1
A5 = =A
1 0
...

To summarize even powers of A are ± I and odd powers of A are ± A. This

suggests separating the power series into even and odd powers. This gives

t2 2 t3 3 t4 4
etA = I + tA + A + A + A +...
2 6 24
t2 t4 t3 t5
= (1 − + + . . .) I + (t − + + . . .) A
2 24 6 120
= cos(t) I + sin(t) A
!
cos(t) − sin(t)
= .
sin(t) cos(t)

This example notwithstanding direct evaluation of the power se-

ries is usually not a practical method for finding the matrix exponen-
tial. We give two algorithms for computing the matrix exponential:
diagonalization and Putzer’s method. The diagonalization method
works only for the case when the matrix A is diagonalizable – when
it has a complete set of eigenvectors. This is not always true but it
does hold for some important classes of matrices. If the eigenvalues
of A are all distinct, for instance, or if A is a symmetric matrix, i.e.
A T = A, then we know that A is diagonalizable.

Method 4.4.1 (Diagonalization). Suppose that the matrix A has n lin-

early independent eigenvectors u1 , u2 , . . . un with corresponding eigenvalues
λ1 , λ2 , . . . λn . The eigenvalues may have higher multiplicity as long as the
number of linearly independent eigenvectors is the same as the multiplic-
ity of the eigenvalue. Let U denote the matrix with columns given by the
eigenvectors, u1 , u2 , . . . un , and U −1 denote the inverse. Note that since the
eigenvectors are assumed to be linearly independent we know that U must
be invertible. Then the matrix exponential is given by The eigenvalues can be in any order
but the same order should be used
throughout. Each eigenvalue should
be listed according to its multiplicity –
in other words if 2 is an eigenvalue of
multiplicity five then it should appear
five times in the list.
 systems of ordinary differential equations 119

e λ1 t
 
0 0 0 ... 0
 0 e λ2 t 0 0 ... 0
 

 0 0 e λ3 t 0 ... 0
 

 −1
etA =U . .. .. .. U

 .. . . . 
 
 0 0 ... 0 e λ n −1 t 0
 

0 0 ... 0 0 eλn t

Notice that in this case, since there are "enough" eigenvectors

we are basically just finding n linearly independent solutions and
writing down a particular fundamental matrix, the one that satisfies
the initial condition M (0) = I. This method covers a lot of cases
that arise in practice, but not all of them. For instance if the system is
derived from a higher order scalar equation that has multiple roots
then A will never have a full set of eigenvectors. The second method Again the order of the eigenvalues is
is a little more involved to describe but it will always work and does not important but the order here should
be the same as when you calculated the
not require one to calculate the eigenvectors, only the eigenvalues. matrices Bk . Note the pattern that holds
for every function after k = 1 : each
Method 4.4.2 (Putzer’s Method). Given an n × n matrix A with eigen- function satisfies a first order linear
−1 equation with the previous function as
values λ1 , λ2 , λ3 , . . . , λn define the matrices {Bk }nk= 0 as follows: the forcing term.

B0 = I
B1 = A − λ1 I
B2 = ( A − λ2 I ) ( A − λ1 I ) = ( A − λ2 I ) B1
B3 = ( A − λ3 I ) ( A − λ2 I ) ( A − λ1 I ) = ( A − λ3 I ) B2
..
.
Bn−1 = ( A − λn−1 I ) ( A − λn−2 I ) . . . ( A − λ1 I ) = ( A − λn−1 I ) Bn−2 .

As a check if you compute Bn = ( A − λn I ) Bn−1 it should be zero due to a

result in linear algebra known as the Cayley-Hamilton theorem.
Define a sequence of functions {rk (t)}nk=1

dr1
= λ1 r1 r1 (0) = 1
dt
dr2
= λ2 r2 + r1 ( t ) r2 (0) = 0
dt
dr3
= λ3 r3 + r2 ( t ) r3 (0) = 0
dt
..
.
drn
= λ n r n + r n −1 ( t ) r n (0) = 0
dt
Then the matrix exponential is given by

n −1
etA = r1 (t) B0 + r2 (t) B1 + . . . rn (t) Bn−1 = ∑ rk+1 (t) Bk .
k =0
120 differential equations

The Putzer method tends to require more computations, but the

computations required are easy – matrix multiplication and solution
of a first order linear differential equation. It also works for all matri-
ces. The diagonalization method tends to require less computations,
but the computations that are required use somewhat more linear
algebra: calculation of eigenvectors and inversion of a matrix. It may
also not work if A is non-symmetric with eigenvalues of higher mul-
tiplicity.

Example 4.4.3. Compute the matrix exponential etA where A is the matrix
 
1 1 0
A = 1 1 0
 
0 0 2

using the diagonalization method and Putzer’s method.

The eigenvalues of A are λ1 = 0, λ2 = 2, λ3 = 2. There is a repeated
eigenvalue – λ = 2 has multiplicity two – but since A is symmetric we are
guaranteed that there are two linearly independent eigenvectors. The three
eigenvectors are
     
1 1 0
u1 =  −1 u2 = 1 u3 = 0 .
     
0 0 1

and so the matrix U is given by

 
1 1 0
U =  −1 1 0
 
0 0 1

and the inverse by

1
− 21
 
0
 21
U −1 = 1
0

2 2
0 0 1

we have that the matrix exponential is given by

1
− 12
   
1 1 0 1 0 0 2 0
etA =  −1 1 0 0 e2t 0   21 1
0
   
2
0 0 1 0 0 e2t 0 0 1
 2t
e +1 e2t −1

0
 2t2 2
e2t +1
=  e 2−1 0

2
0 0 e2t
 systems of ordinary differential equations 121

Using Putzer’s method we have that the matrices B0,1,2 are given by
 
1 0 0
B0 = 0 1 0
 
0 0 1
 
1 1 0
B1 = A − 0I = A = 1 1 0
 
0 0 2
  
−1 1 0 1 1 0
B2 = ( A − 2I ) B1 =  1 −1 0 1 1 0
  
0 0 0 0 0 2
 
0 0 0
= 0 0 0
 
0 0 0

In this case B2 happens to be zero. Computing the functions ri (t) we have

dr1
= 0r1 r1 (0) = 1
dt
r1 (t) = e0t = 1
dr2
= 2r2 (t) + 1 r2 (0) = 0
dt
e2t 1
r2 ( t ) = −
2 2
Ordinarily we would have to compute the solution to

dr3 e2t 1
= 2r3 + r2 (t) = 2r3 + − r3 (0) = 0
dt 2 2

(you can check that the solution is r3 (t) = 2t e2t + 14 e2t − 41 ) but in this case
since it is being multiplied by B2 which is the zero matrix we don’t really
need to calculate this. Thus we have
 2t
e +1 e2t −1

2t 2 2 0
e −1
etA = 1 · B0 +
 2t 2t
B1 =  e 2−1 e 2+1 0 

2
0 0 e2t

Example 4.4.4. Find the matrix exponential etA for

 
3 2 1
A =  −1 0 0
 
−1 −1 0

The characteristic polynomial for this matrix is (λ − 1)3 = 0 so λ = 1 is

an eigenvalue of multiplicity three. There is only one linearly independent
eigenvector, u = [−1, 1, 0] T , so this matrix is not diagonalizable, so Putzer’s
122 differential equations

method is the best option. Computing the B matrices.

 
1 0 0
B0 = I = 0 1 0
 
0 0 1
 
2 2 1
B1 = A − I = −1 −1 0 
 
−1 −1 −1
 
1 1 1
B2 = ( A − I ) B1 = −1 −1 −1
 
0 0 0

In this case all of the B matrices are non-zero, but if we multiply B2 by

A − I you will see that we get zero, as implied by the Cayley-Hamilton
theorem. Next we need to compute the functions r (t).

dr1
= r1 r1 (0) = 1
dt
r1 ( t ) = e t
dr2
= r2 ( t ) + e t r2 (0) = 0
dt
r2 (t) = tet
dr3
= r3 (t) + tet r3 (0) = 0
dt
t2
r2 ( t ) = e t
2
The matrix exponential is given by

t2
etA = et B0 + tet B1 + et B2
2
2 2 2
(1 + 2t + t2 )et (2t + t2 )et (t + t2 )et
 
2 2 2
=  −(t + t2 )et (1 − t − t2 )et − t2 et 
 

−tet −tet (1 − t ) e t
 systems of ordinary differential equations 123

4.5 Problems on Linear Algebra and Systems

Exercise 4.5.1
Find the characteristic polynomial and the eigenvalues for the follow-
ing matrices
! ! !
3 2 5 4 0 1
a) b) c)
2 3 0 2 −1 0
   
! 3 2 1 −2 1 1
2 4
d) e) 0 2 3 f)  0 −1 1 
   
−9 2
0 3 2 1 1 −2
 
3 1 2
g) 0 4 1 
 
0 0 −2

Exercise 4.5.2
Find the eigenvectors for the matrices in problem 4.5.1

Exercise 4.5.3
Solve the following initial value problems
! !
dv 4 2 3
a) dt = v; v (0) =
2 4 5
! !
1 2 2
b) dv
dt = 0 2 v; v (0) =
2
! !
dv 1 2 2
c) dt = v; v (0) =
0 1 3

Exercise 4.5.4
Find the matrix exponential etA for the matrices given in problem
4.5.1

Exercise 4.5.5
Find the eigenvalues and eigenvectors for the two matrices
   
2 1 1 2 1 1
0 2 4 0 2 0
   
0 0 2 0 0 2

How many linearly independent eigenvectors did you find in each

case?
124 differential equations

Exercise 4.5.6
Find the matrix exponentials for the two matrices given in problem
4.5.5
 Part II

Boundary Value Problems,

Fourier Series, and the
Solution of Partial
Differential Equations.
5
Boundary Value Problems

5.1 Examples of boundary value problems

Prior to this point we have been exclusively focused on initial value

problems. An initial value problem is one where all of the informa-
tion is prescribed at some initial time, say t = 0. An example of an
initial value problem is

d3 y dy dy d2 y
+ y2 + cos(y) = 5 y(0) = 1; (0) = 2; (0) = −2.
dt3 dt dt dt2
For initial value problems we have an existence and uniqueness the-
orem that (under some mild technical conditions) guarantees that
there exists a unique solution in some small neighborhood of the
initial data.
A differential equation problem where values are specified at more
than one point is called a boundary value problem. For instance the
problem
π
y′′ + y = 0 y(0) = 1 y( ) = 0
2
is a boundary value problem. In some cases, such as the above, there
is a unique solution (in the above case the solution is y = cos(t)).
In other cases there may be no solution, or the solution may not be
unique. While many applications of differential equations are posed
as initial value problems there are also applications which give rise to
two point boundary value problems.
One example of a boundary value problem arises in computing the
deflection of a beam:

Example 5.1.1. The equation for the deflection y( x ) of a beam with con-
stant cross-sectional area and elastic modulus is given by

d4 y
EI = g( x )
dx4
where E is a constant called the “elastic modulus”, and I is the second mo-
ment of the cross-sectional area of the beam. A larger value of E represents a
128 differential equations

stiffer beam that is more resistant to bending. The function g( x ) represents

load on the beam: the force per unit of length.
Since this is a fourth order equation it requires four boundary conditions.
If the ends of the beam are located at x1 and x2 then one typically requires
two boundary conditions at x1 and two at x2 . The nature of the boundary
conditions depends on the way in which the end of the beam is supported:
Simply supported/pinned y( xi ) = 0; y′′ ( xi ) = 0
Fixed/clamped y( xi ) = 0; y′ ( xi ) = 0
Free y′′ ( xi ) = 0; y′′′ ( xi ) = 0
One can “mix-and-match” these boundary conditions. For instance the
equation for a beam which is fixed at x = 0 and simply supported at x = L
and is subjected to a constant load w is
d4 y
EI = w y(0) = 0; y′ (0) = 0; y( L) = 0; y′′ ( L) = 0
dx4
Integrating this equation up four times gives
wx4 Ax3 Bx2
y( x ) = + + + Cx + D
24 6 2
imposing the four boundary conditions y(0) = 0; y′ (0) = 0; y( L) =
0; y′′ ( L) = 0 gives four equations for the four unknowns A, B, C, D.
D=0 y (0) = 0
C=0 y ′ (0) = 0
BL2 AL3 wL4
D + CL + + + =0 y( L) = 0
2 6 24
wL2
B + AL + =0 y′′ ( L) = 0
2
The solution of these four linear equations is given by
5Lw L2 w
A=− B= C=0 D=0
8EI 8EI
So the deflection of the beam is given by
wL4 x 4  x 3  x 2 
  
y( x ) = 2 −5 +3 .
48EI L L L
Since the beam is supported at both x = 0 and x = L one expects that the
maximum deflection should occur somewhere between those points. To find
the point of maximum deflection we can take the derivative and set it equal
to zero:
wL4
 3
x x2 x wL3  x  x 2 x
    
′
y (x) = 8 4 − 15 3 + 6 2 = 8 − 15 + 6 = 0
48EI L L L 48EI L L L
√ √
The roots of this cubic are Lx = 0, Lx = 15−16 33 ≈ 0.58 and Lx = 15+16 33 ≈
1.30. The beam has length L so the root Lx = 1.30 is outside of the domain.
The root Lx = 0 is one of the support points: we know that the deflection
there is zero, so the maximum deflection must occur at x ≈ 0.58L.
 boundary value problems 129

5.2 Existence and Uniqueness

Unlike initial value problems which, under some mild assumptions,

always have unique solutions a boundary value problem may have
a unique solution, no solution, or an infinite number of solutions.
Let’s consider two examples, which appear quite similar. The two
examples are

y′′ + π 2 y = 0 y (0) = 0 y (1) = 0 (5.1)

′′ 2
y + π y = sin(πx ) y (0) = 0 y (1) = 0 (5.2)

In the first example we can find that the characteristic equation is

given by
r2 + π 2 = 0
so r = ±iπ and the general solution to be y( x ) = A cos(πx ) +
B sin(πx ) If we impose the conditions y(0) = 0 and y(1) = 0 this
gives two equations

A cos(0) + B sin(0) = A = 0
A cos(π ) + B sin(π ) = − A = 0 .

In this case we get two equations that are not linearly independent,
but they are consistent. We know from linear algebra that there are
an infinite number of solutions – A = 0 and B is undetermined.
In the second case we can solve the inhomogeneous equation
using the method of undetermined coefficients or of variation of
parameters. We find that the general solution is given by

x cos(πx )
y( x ) = A cos(πx ) + B sin(πx ) − .
2π
In this case when we try to solve the boundary value problem we
find
0 cos(0)
y(0) = A cos(0) + B sin(0) − =A=0
2π
π cos(π ) 1
y(1) = A cos(π ) + B sin(π ) − = − A + = 0.
2π 2
In this case the boundary conditions lead to an inconsistent set of
linear equations: A = 0 and A = 12 , for which there can be no
solution.
The following theorem tells us that these are the only possibilities:
a two point boundary value problem can have zero, one, or infinitely
many solutions. For simplicity we will state and prove the theorem
for second order two point boundary value problems; the proof for
higher order boundary value problems is basically the same, with
minor notational changes.
130 differential equations

Theorem 5.2.1. Consider the second order linear two point boundary value
problem

d2 y dy
+ p( x ) + q( x )y = Ly = f ( x ) (5.3)
dx2 dx
α1 y ( a ) + α2 y ′ ( a ) = A
β 1 y(b) + β 2 y′ (b) = B,

where p( x ), q( x ), f ( x ) are continuous on [ a, b]. The corresponding homoge-

neous problem is given by

d2 y dy
+ p( x ) + q( x )y = Ly = 0 (5.4)
dx2 dx
α1 y ( a ) + α2 y ′ ( a ) = 0
β 1 y(b) + β 2 y′ (b) = 0,

Then the following holds:

1. Problem (5.4) has either a unique solution u( x ) = 0 or an infinite

number of solutions.

2. If Problem (5.4) has a unique solution then Problem (5.3) has a unique
solution.

3. If Problem (5.4) does not have a unique solution then Problem (5.3)
either has no solutions or an infinite number of solutions.

Proof. This theorem may remind you of the following fact from linear
algebra: a system of linear equations

Mx = b

either has no solutions, a unique solution, or infinitely many solu-

tions. It has a unique solution if the only solution to the homoge-
neous equation Mx = 0 is x = 0. If Mx = 0 has at least one non-zero
solution then Mx = b either has no solutions or infinitely many
solutions. We will use this fact of linear algebra to prove this result.
We can assume without loss of generality that A = 0 = B. If
this is not the case one can make the change of variables y( x ) =
w( x ) + u( x ), where w( x ) is any infinitely differentiable function that
satisfies the boundary conditions

α1 w ( a ) + α2 w ′ ( a ) = A β 1 w( a) + β 2 w′ ( a) = B,

This will lead to a boundary value problem for u( x ) with homoge-

neous boundary conditions.
We know that the solution to equation (5.4) can be written as a
linear combination of two independent solutions

y( x ) = Cy1 ( x ) + Dy2 ( x ).
 boundary value problems 131

Imposing the boundary conditions gives us the following system of

equations for C, D

(α1 y1 ( a) + α2 y1′ ( a))C + (α1 y2 ( a) + α2 y2′ ( a)) D = 0

( β 1 y1 (b) + β 2 y1′ (b))C + ( β 1 y2 (b) + β 2 y2′ (b)) D = 0.

We know from linear algebra that typically the only solution to this
set of homogeous linear equations is C = 0, D = 0. There is a non-
zero solution if and only if the matrix
!
α1 y1 ( a) + α2 y1′ ( a) (α1 y2 ( a) + α2 y2′ ( a))
M= (5.5)
( β 1 y1 (b) + β 2 y1′ (b)) ( β 1 y2 (b) + β 2 y2′ (b))

is singular, in which case there are an infinite number of solutions.

The inhomogeneous case is similar. There the general solution is

y( x ) = Cy1 ( x ) + Dy2 ( x ) + y part ( x ),

where y part ( x ) is the particular solution. Applying the boundary

conditions leads to the equations for C, D

(α1 y1 ( a) + α2 y1′ ( a))C + (α1 y2 ( a) + α2 y2′ ( a)) D = −α1 y part ( a) + α2 y′part ( a)

( β 1 y1 (b) + β 2 y1′ (b))C + ( β 1 y2 (b) + β 2 y2′ (b)) D = −α1 y part (b) + α2 y′part (b)

or equivalently
! !
C −(α1 y part ( a) + α2 y′part ( a))
M = (5.6)
D −(α1 y part (b) + α2 y′part (b))

Note that the M arising here is the same as the M arising in the solu-
tion to the homogeneous problem. Therefore the system of equations
(5.6), and hence the boundary value problem, has a unique solution
if and only if the matrix M is non-singular, which is true if and only
if the homogeneous boundary value problem (5.4) has a unique solu-
tion y( x ) = 0.

Exercise 5.2.1
Determine if the following boundary value problems have no solu-
tions, one solution, or infinite solutions. If there are solutions, find
them.

a) y′′ + y = 0 y (0) = 1 y(π ) = 0

b) y′′ + y = 0 y (0) = 1 y( π2 ) = 0
c) y′′ + y = 0 y (0) = 1 y′ (π ) = 1
d) y′′ + y = 2 y (0) = 0 y(π ) = 4
e) y′′ + y = 2 y (0) = 1 y(π ) = 0
f) y′′ + 4y = 0 y (0) = 3 y(π ) = 3
132 differential equations

Exercise 5.2.2
Determine the values of A for which the following boundary value
problems have solutions.

a) y′′ + y = x y (0) = 2 y(π ) = A

b) y′′ + 4y = 0 y ′ (0) = A y′ (π ) = 1
c) y′′ + 9y = 9x y (0) = 0 y(π ) = A

5.3 Eigenvalue Problems

A special kind of boundary value problem is something called an

eigenvalue problem. An eigenvalue problem is a boundary value
problem involving an auxiliary parameter. The goal is to determine
the values of the parameter for which the problem has a non-trivial
solution. These special values of the parameter are called “eigenval-
ues” and the corresponding non-trivial solutions are called “eigen-
functions” or “eigensolutions”.

Definition 5.3.1. A second order two point eigenvalue problem consists of

a second order linear differential equation involving an unknown parameter
λ, called the eigenvalue parameter

L(λ)y = 0,

where L(λ) is a second order linear differential operator, together with

homogeneous two point boundary conditions:

α1 y ( a ) + α2 y ′ ( a ) = 0 β 1 y(b) + β 2 y′ (b) = 0.

This equation always has the solution y( x ) = 0, which we call the trivial
solution. We know from Theorem 5.2.1 that such an equation will either
have a unique solution or an infinite number of solutions. The eigenvalues
are the values of λ for which the boundary value problem has an infinite
number of solutions. The non-trivial solutions are called eigenfunctions.
More generally an nth order eigenvalue problem would consist of an nth
order linear differential equation involving an unknown parameter, together
with n homogeneous boundary conditions. Most commonly these would
be specified at two points, although one could specify boundary conditions
at more than two points. In the theory of elastic beams, for instance, one
often has to solve a fourth order eigenvalue problem, with two boundary
conditions specified at each end of the beam.

We now present several example of eigenvalue problems, along

with their solution.
 boundary value problems 133

Example 5.3.1. For what values of λ does the boundary value problem

y′′ + λy = 0 y (0) = 0 y ′ (1) = 0

have a non-trivial solution (a solution other than y( x ) = 0)?

We will consider three cases:
Case 1: λ > 0
Let’s write λ = k2 where k is a non-zero real number. Then the equation
becomes y′′ + k2 y = 0 which has solutions

y = A cos(kx ) + B sin(kx )

Using the boundary conditions we have that y(0) = A = 0, and y′ (1) =

Bk cos(k) = 0. Since k ̸= 0 the only solutions are B = 0 or cos(k) = 0. We
are only interested in non-zero solutions so we choose cos(k) = 0 which is
true for k = π2 , 3π 5π 1 2 1 2 2
2 , 2 , . . . = ( n + 2 ) π. So for λ = k = ( n + 2 ) π we
have non-trivial solutions.
Case 2: λ = 0
The equation is now y′′ = 0 which has solutions

y = A + Bx

Imposing y(0) = 0 tells us that A = 0. Imposing y′ (1) = 0 tells us that

B = 0. So λ = 0 is not an eigenvalue because we only have the trivial
solution.
Case 3: λ < 0
Let’s write λ = −k2 where k is a non-zero real number. Then the equa-
tion becomes y′′ − k2 y = 0 which has solutions

y = Aekx + Be−kx

Using the boundary conditions we have that y(0) = A + B = 0, and

y′ (1) = Akek − Bke−k = 0. Because k ̸= 0 the only solution in this case
is A = B = 0 which gives the trivial solution. So there are no negative
eigenvalues.
In summary the eigenvalues for this problem, the values of λ for which
we can have non-trivial solutions, are given by λ = (n + 12 )2 π 2 where n is
an integer.
Example 5.3.2. Find all of the eigenvalues and the corresponding eigen-
functions for the boundary value problem

y′′ + λy = 0 y (0) = 0 y( L) = 0

We will consider three cases:

Case 1: λ > 0
We rewrite λ = k2 where k is a non-zero real number. The equation is
now y′′ + k2 y = 0 which has solutions

y = A cos(kx ) + B sin(kx )
134 differential equations

Using the first boundary condition we have that y(0) = A = 0. Using this
result in the second boundary condition we have y( L) = B sin(kL) = 0.
Therefore either B = 0 or sin(kL) = 0. We are only interested in non-zero
solutions so we choose sin(kL) = 0 which is true for kL = π, 2π, 3π, . . . =
2 2
nπ. So the eigenvalues are λn = k2 = n Lπ2 and the corresponding eigen-
functions are yn ( x ) = sin( nπx
L ). We use the subscript in λn and yn ( x ) to
indicate that eigenfunctions correspond to specific eigenvalues.
Case 2: λ = 0
The equation is now y′′ = 0 which has solutions

y = A + Bx

Imposing y(0) = 0 tells us that A = 0. Imposing y( L) = 0 tells us that

B = 0. So λ = 0 is not an eigenvalue because we only have the trivial
solution.
Case 3: λ < 0
We rewrite λ = −k2 where k is a non-zero real number. The equation is
now y′′ − k2 y = 0 which has solutions

y = Aekx + Be−kx

Using the boundary conditions we have that y(0) = A + B = 0, and

y( L) = AekL + Be−kL = 0. Because kL ̸= 0 the only solution in this case
is A = B = 0 which gives the trivial solution. So there are no negative
eigenvalues.
2 2
In summary the eigenvalues for this problem are given by λn = n Lπ2
and the corresponding eigenfunctions are yn ( x ) = sin( nπx
L ) where n is an
integer.

Example 5.3.3. Find all of the eigenvalues and the corresponding eigen-
functions for the boundary value problem

y′′ + y′ + λy = 0 y (0) = 0 y (1) = 0

The characteristics equation is given by

r2 + r + λ = 0
q
which has roots r = − 12 ± 1
4 − λ. It is convenient to separate this into
three cases:

• λ > 14 , when the characteristic equation has complex roots.

• λ = 14 , when the characteristic equation has a double root.

• λ < 14 , when the characteristic equation has real and distinct roots.

Case 1: λ > 14
Let’s write λ = 14 + k2 where k is a non-zero real number. The general
solution is given by y( x ) = Ae− x/2 cos(kx ) + Be− x/2 sin(kx ). The
 boundary value problems 135

boundary condition y(0) = 0 implies that A = 0. The boundary condition

that y(1) = 0 implies that B sin(k ) = 0. Choosing B = 0 would mean
we have only the zero solution, so instead we need sin(k) = 0. This implies
that k = nπ and λn = 41 + n2 π 2 with corresponding eigenfunctions
yn ( x ) = e− x/2 sin(nπx )
Case 2: λ = 14
In this case r = − 21 is a double root of the characteristic equation, so the
general solution is y = Ae− x/2 + Bxe− x/2 . The condition y(0) = 0 implies
that A = 0, and the condition y(1) = Be−1/2 = 0 implies that B = 0. So
λ = 1/4 is not an eigenvalue.
Case 3: λ < 41
Let’s write λ = 41 − k2 where k is a non-zero real number. In this case the
general solution is y = Aer1 x + Ber2 x , with r1 = − 12 + k and r2 = − 12 − k.
Imposing the boundary conditions gives two equations

y (0) = A + B = 0
y(1) = Aer1 + Ber2 = 0

This can be written as

! ! !
1 1 A 0
=
er1 er2 B 0

The only solution is A = 0, B = 0 unless the determinant of the matrix is

zero. The determinant is given by er2 − er1 which is not zero since r1 ̸= r2 .
In summary the eigenvalues of this problem are given by λn = 14 + n2 π 2
with corresponding eigenfunctions yn ( x ) = e− x/2 sin(nπx ).

Example 5.3.4 (Guitar String). The equation for a one dimensional vibrat-
ing string, such as a guitar string, is

∂2 y 2
2∂ y
= c y(0, t) = 0 y( L, t) = 0.
∂t2 ∂x2
Here c is a constant representing the wave speed of the string. If we look for
a “pure tone” or harmonic we look for a solution of the form

y = f ( x ) cos(ωt)

Substituting this into the equation gives the following equation for f ( x )

c2 f xx + ω 2 f = 0 f (0) = f ( L) = 0.

This is an eigenvalue problem for ω, the frequency. The fact that this equa-
tion has non-zero solutions only for specific values of ω tells us that the
string will only vibrate at certain specific frequencies, the eigenvalues. The
general solution to the above is given by
ω ω
f ( x ) = A cos( x ) + B sin( x )
c c
136 differential equations

as long as ω ̸= 0. Imposing the boundary conditions we find that

ω ω
f (0) = A = 0 f ( L) = A cos( L) + B sin( L) = 0.
c c
Substituting the first equality into the second tells us that B sin( ωc L) = 0,
so either B = 0 or sin ωc L = 0. The first gives us the zero solution. The
second case happens when ωc L = π, 2π, 3π, . . . These are the frequencies at
which the string can vibrate. The lowest frequency ω = πc L is often called
the fundamental while the higher frequencies are known as harmonics or
overtones.

Exercise 5.3.1
Calculate all of the eigenvalues and the corresponding eigenfunctions
for the following boundary value problems.

a) y′′ + λy = 0 y ′ (0) = 0 y ′ (5) = 0

b) y′′ + λy = 0 y ′ (0) = 0 y (3) = 0
c) y′′ + λy = 0 y(−π ) = 0 y(π ) = 0
d) y′′ + 2y′ + λy = 0 y (0) = 0 y(π ) = 0
e) y′′ − 2y′ + λy = 0 y ′ (0) = 0 y′ ( L) = 0
6
Fourier Series

6.1 Background

In science and engineering it is often desirable to have an efficient

way to represent some function of interest. In linear algebra one often
works with eigenvectors. The eigenvectors of a matrix A are the non-
zero vectors vi for which
Avi = λi vi .

This should be very reminiscent of section 5.2, on boundary value

problems. In each case we have an equation depending on a param-
eter λ which has a non-zero solution only for certain values of that
parameter. In the case of the eigenvalues of the matrix these are the
roots of the characteristic polynomial, while in the case of a differ-
ential boundary value problem these are typically the roots of some
more complicated function, but the idea is similar. It is worthwhile to
recall a couple of results from linear algebra.

Theorem 6.1.1. Suppose that the matrix A is a real n × n symmetric

matrix. The eigenvalue problem

Av = λv

has n linearly independent eigenvectors {vi }in=1 . These vectors are linearly
independent, form a basis for Rn , and can be chosen to be orthonormal:
(
0 i ̸= j
vi · vj =
1 i=j

One of the advantages of an orthonormal basis is that it makes

many of the operations of linear algebra much simpler. The matrix
A itself looks much simpler in the eigenbasis. Recall the following
additional results from linear algebra:

Theorem 6.1.2. Suppose that {vi }in=1 is an orthonormal basis for Rn . For
138 differential equations

any w ∈ Rn we have that

n
w= ∑ αi vi αi = vi · w
i =1

This theorem shows that it is easy to express a given vector in

terms of an orthonormal basis: the coefficients are simply the dot
product of the vector with the basis vectors. We will give a proof
here, since it is simple and will be important for what is to follow

Proof. We are given that {vi }in=1 is an orthonormal basis for Rn , so

for any w there must exists coefficient αi such that
n
w= ∑ αi vi = α1 v1 + α2 v2 + . . . αn vn (6.1)
i =1

To find (for instance) α1 we can take the dot product with v1 to get
n
v1 · w = ∑ αi vi · v1 = α1 v1 · v1 + α2 v1 · v2 + . . . αn v1 · vn .
i =1

However the basis is orthonormal, so v1 · v1 = 1 and vj · v1 = 0 for

j ̸= 1, so this reduces to
w · v1 = α1 .
In the same spirit taking the dot product of Equation (6.1) with vj for
j ≥ 2 gives α j = vj · w

Theorem 6.1.3. Suppose that the matrix A is a real n × n symmetric

matrix. Assume that the n linearly independent eigenvectors {vi }in=1 of A
are chosen to be orthonormal. Let U be the matrix whose ith column is vi .
Then the matrix Λ defined by

Λ = U T AU

is diagonal, with the eigenvalues λ1 , λ2 , . . . , λn along the diagonal.

The topic of most of the rest of the course will be Fourier series.
The classical Fourier series, along with the Fourier sine and cosine
series, are the analog of the eigenvector expansion for certain sec-
ond order boundary value problems. The functions {sin( nπx ∞
L )}n=1 or
nπx ∞
{cos( L )}n=0 are analogous to the eigenvectors, and most "reason-
able" functions can be decomposed in terms of these basis functions.

6.2 The Classical Fourier Series: Orthogonality

Suppose that we have a function f ( x ) that is periodic with period 2L:

f ( x + 2L) = f ( x ). It is often convenient to be able to express f ( x ) in
terms of known functions, such as sines and cosines. This is the idea
 fourier series 139

behind the classical Fourier series. To begin with we state a result

about definite integrals that is related to the the orthogonality that we
discussed in the previous section.

Lemma 6.2.1. The functions 1, {sin( nπx ∞ nπx ∞

L )}n=1 , {cos( L )}n=1 are orthog-
onal on the interval (0, L). Letting δn,m denote the Kronecker delta symbol
(
0 n ̸= m
δn,m =
1 n=m

then we have the identities (assuming n, m are integers ≥ 1)

Z 2L  nπx   mπx 
cos cos dx = L δn,m
0 L L
Z 2L  nπx  mπx 
sin cos dx = 0
0 L L
Z 2L  nπx   mπx 
sin sin dx = L δn,m
0 L L
Z 2L  nπx 
cos dx = 0
0 L
Z 2L  nπx 
sin dx = 0
0 L
Z 2L
1dx = 2L.
0

Note that the functions are orthogonal, not orthonormal! Now if

one knew in advance that a given function was expressible as a linear
combination of the functions 1, {sin( nπx ∞ nπx ∞
L )}n=1 , {cos( L )}n=1 then
it is relatively simple to find what the coefficients must be. Suppose
that one knows that
∞  nπx  ∞  nπx 
f ( x ) = A0 + ∑ An cos
L
+ ∑ Bn sin L
.
n =1 n =1

Then, to find (for instance) A5 we would multiply the above by

cos( 5πx
L ) and integrate over the interval (0, 2L ). This would give

Z 2L
5πx
 
f ( x ) cos dx
0 L
Z 2L ∞ Z 2L
5πx 5πx  nπx 
   
= A0 cos dx + ∑ An cos cos dx
0 L n =1 0 L L
∞ Z 2L
5πx  nπx 
 
+ ∑ Bn cos sin dx. (6.2)
n =1 0 L L

Next we should notice that all of the terms on the righthand side of the
140 differential equations

equation are zero except for one! We have that

Z 2L
5πx
 
cos dx = 0
0 L
Z 2L
5πx  nπx 
 
cos sin dx = 0 for all n
0 L L
Z 2L
5πx  nπx 
 
cos cos dx = 0 for all n ̸= 5
0 L L
Z 2L
5πx 5πx
   
cos cos dx = L.
0 L L

Using these to simplify the righthand side of equation (6.2) we find

Z 2L
5πx
 
f ( x ) cos dx = A5 L,
0 L
R 2L
or A5 = L1 0 f ( x ) cos 5πx dx. There is, of course, nothing special


L
about n = 5: one can do the same calculation for any value of n. This
leads to the following result:

Theorem 6.2.1. Suppose that f ( x ) defined for x ∈ (0, 2L) can be repre-
sented as a series of the form
∞  nπx  ∞  nπx 
f ( x ) = A0 + ∑ An cos
L
+ ∑ Bn sin
L
. (6.3)
n =1 n =1

Then the coefficients are given by

1 2L
Z
A0 = f ( x )dx (6.4)
2L 0
Z 2L
1  nπx 
An = f ( x ) cos dx n≥1 (6.5)
L 0 L
Z 2L
1  nπx 
Bn = f ( x ) sin dx (6.6)
L 0 L
The series in Equation (6.3) of Theorem 6.2.1 is called a Fourier
series. The somewhat surprising fact is that under some very mild
assumptions all periodic functions can be represented as a Fourier
series.

Theorem 6.2.2. Suppose that f ( x ) is piecewise C2 (twice differentiable)

and 2L-periodic: f ( x + 2L) = f ( x ), with jump discontinuities at the points
of discontinuity. Then the Fourier series
∞  nπx  ∞  nπx 
A0 + ∑ An cos
L
+ ∑ Bn sin L
n =1 n =1

1
converges to f ( x ) at points of continuity of f ( x ), and to 2 ( f ( x − ) + f ( x + ))
at the jump discontinuities.
 fourier series 141

Example 6.2.1 (Square Wave). Let f ( x ) be defined to be

x ∈ (0, 12 )
(
1
f (x) =
0 x ∈ ( 21 , 1)

and repeated periodically (see the margin figure). Find the Fourier series for
f ( x ).
The period in this case is 2L = 1. The Fourier coefficients are easy to
compute in this case, and are given as follows:
Z 1 1 Figure 6.1: The square wave
1
Z
2
A0 = f ( x )dx = dx = function.
0 0 2
1
sin(2πnx ) 12
Z
2
An = 2 cos(2nπx )dx = |0 = 0
0 πn
1
cos(2πnx ) 21 1 − cos(πn)
Z
2
Bn = 2 sin(2nπx )dx = − |0 =
0 πn πn
Given the fact that cos(πn) = 1 if n is even and cos(πn) = −1 if n is odd
Bn can also be written as
2
(
n odd
Bn = πn
0 n even

The margin figure depicts the square wave function f ( x ) together with
the first twenty-six terms of the Fourier series
Figure 6.2: The square wave
Example 6.2.2. Find the Fourier series for the function f ( x ) defined as function together with the first
follows twenty-six terms of the Fourier
series.
f ( x ) = x (1 − x ) x ∈ (0, 1)
f ( x + 1) = f ( x )

The graph of this function on the interval (−1, 2) is presented in the side
margin: the function continues periodically to the rest of the real line.
This is a Fourier series with 2L = 1 The Fourier coefficients can be
calculated by integration by parts as

1
Z2L 1 Z
1
A0 = f ( x )dx = x (1 − x )dx = Figure 6.3: The function f ( x ) =
2L 0 0 6
1 2L
Z  nπx  Z 1 x (1 − x ) for x ∈ (0, 1) and ex-
An = f ( x ) cos dx = 2 x (1 − x ) cos (2nπx ) dx =
L 0 L 0 tended periodically.
1 x (1 − x ) (1 − 2x ) cos(2nπx ) 1
 
=2 − sin (2nπx ) |10 + |0
4n3 π 3 2nπ 4n2 π 2
1
=− 2 2
n π
Z 1
Bn = 2 x (1 − x ) sin(2nπx )dx = 0.
0
142 differential equations

The easiest way to see that all coefficients Bn must be zero is to note that
sin(2nπx ) is an odd function and the function f ( x ) is an even function, so
they must be orthogonal. This gives the series
∞
1 cos(2πnx )
f (x) = −∑
6 n =1 n2 π 2

The margin figure shows a comparison between f ( x ), (in black) the first five
terms of the Fourier series (in red) , and the first twenty-five terms of the
Fourier series (in blue). It is clear that even five terms of the Fourier series
gives a good approximation to the original function (though with visible
error), and that it is difficult to distinguish between the original function
and the first twenty-five terms of the Fourier series– one can see a very small
deviation between the two near x = 0 and x = 1, the endpoints. Note
that since f ( x ) is a piecewise C2 continuous function the Fourier series
converges to the function at all points.

Figure 6.4: The function from

6.3 Periodic, even and odd extensions Example 6.2.2 (Black) together
with first 5 terms (Red) and
It is common to have a function that is defined on a finite domain
first twenty five terms (Blue) of
[0, L] that we wish to extend to the whole line. There are a number
the Fourier series.
of different ways to do this, and these are connected with different
variations of the Fourier series and to different boundary conditions
for partial differential equations. In this section we discuss three of
the most common ways to extend a function.

6.3.1 The periodic extension

If we are given a function f ( x ) defined on the domain (0, L) the
simplest way to extend it is periodically: we define f ( x ) on the whole
line by f ( x + L) = f ( x ). For instance if we want to define f ( x ) in
the interval ( L, 2L) we can use the equation f ( x + L) = f ( x ): if
x ∈ (0, L) then x + L ∈ [ L, 2L). Similarly we can use the equation
f ( x + L) = f ( x ) again to relate values of the function in [2L, 3L) to
values of the function in ( L, 2L). Graphically the periodic extension
consists of taking the graph of the function f ( x ) in the interval [0, L)
and "repeating" it in the intervals (− L, 0), ( L, 2L), (2L, 3L), . . .. Notice
that this procedure may introduce discontinuities in the graph if
limx→0+ f ( x ) ̸= limx→ L− ( x ).
This procedure is illustrated in the marginal figure. Depicted is
a function defined on (0, L), together with the function extended
periodically to the whole line. Note that the resulting function is, by Figure 6.5: The function defined
construction periodic with period L and even if the function f ( x ) is x ∈ (0, 1) (top) and the same
continuous on [0, L] the resulting periodic extension will typically function extended periodically
have jump discontinuities at x = nL. to the whole line (bottom).
 fourier series 143

6.3.2 The even extension

Recall that a function is even if f (− x ) = f ( x ). An even function is
invariant under reflection across the y-axis. More generally we say
that f ( x ) is "even about the point x0 " if f (2x0 − x ) = f ( x ), with the
case x0 = 0 being the usual definition of even. Such a function is
invariant under reflection across the line x = x0 . A second method of
extending a function f ( x ) defined for x ∈ (0, L) to the whole line is
to extend it so it is even across both boundaries, x = 0 and x = L. In
other words given ( x ) defined for x ∈ (0, L) we would define f ( x ) for
x ∈ (− L, 0) by f (− x ) = f ( x ), we would define it for x ∈ ( L, 2L) by
f (2L − x ) = f ( x ), etc.
Graphically this amounts to taking the graph of f ( x ) for x ∈ (0, L)
and repeatedly reflecting it across the boundary points x = nL. This
is illustrated in the marginal figure for a typical function f ( x ). Figure 6.6: The function defined
There are a couple of things to notice here. First note that if the x ∈ (0, 1) (top) and the even
original function f ( x ) is continuous on the closed interval [0, L] then extension to the whole line
the even extension is also a continuous function. This differs from (bottom).
the periodic extension, since the periodic extension of a continuous
function on [0, L] won’t be continuous unless f (0) = f ( L). There is,
however, typically a jump in the derivative of the function across the
boundary. Second notice that the resulting function is periodic, but
the period is 2L, not L. This is because the function is reflected across
the boundaries. It requires two reflections to get back the original
function.

6.3.3 The odd reflection

The last case that we will discuss is the case of odd reflections. Recall
that a function is odd if f (− x ) = − f ( x ), and that an odd function
is invariant under reflection across the y-axis followed by reflection
across the x-axis. More generally a function is "odd about the point
x0 " if f (2x0 − x ) = − f ( x ). A function is odd if it is invariant under
reflection across the line x = x0 followed by reflection across the
x-axis. A second method of extending a function f ( x ) defined for
x ∈ (0, L) to the whole line is to extend it so it is odd across both
boundaries, x = 0 and x = L. In other words given f ( x ) defined for
x ∈ (0, L) we would define f ( x ) for x ∈ (− L, 0) by f (− x ) = − f ( x ),
we would define it for x ∈ ( L, 2L) by f (2L − x ) = − f ( x ), etc.
Graphically this amounts to taking the graph of f ( x ) for x ∈ (0, L)
and repeatedly doing odd reflections across the boundary points
x = nL. This is illustrated in the marginal figure for a typical function
f ( x ).
Note that, as in the case of even reflections the resulting function
is periodic with a period of 2L, twice the width of the original do-

Figure 6.7: The function de-

fined x ∈ (0, 1) (top) and the
odd extension to the whole line
(bottom).
144 differential equations

main. Again this is because it requires two reflections to get back the
original function. Also note that the resulting function will typically
have jump discontinuities unless the original function tends to zero at
x = 0 and x = L.

6.4 The Fourier cosine and sine series.

The Fourier cosine and Fourier sine series are connected with the
even and the odd extensions of a function defined on (0, L) to the
whole line. We’ll begin with the even extension. We saw in the pre-
vious section that the even extension of a function defined on (0, L)
results in a function with a period of 2L. Thus we can expand this
function in a Fourier series of period 2L. Because the extended func-
tion is even we expect that this series will only involve the cosine
terms, since cosine is an even function. The formulas for the Fourier
coefficients become

1
Z 2L
A0 = f ( x )dx
2L 0
1 2L  nπx 
Z
An = f ( x ) cos dx
L 0 L
Z 2L
1  nπx 
Bn = f ( x ) sin dx Figure 6.8: The function defined
L 0 L
∞  nπx   nπx  x ∈ (0, L) (top) and Fourier co-
f ( x ) = A0 + ∑ An cos + Bn sin
L L sine series (middle) and Fourier
n =1
sine series (bottom).
Now the original function is defined only on (0, L) so it would be
preferable to express everything in terms of the function values in
the interval (0, L). For the An terms since f ( x ) and cos( nπx
L ) are both
even functions the integrals over (0, L) and over ( L, 2L) are equal. For
the Bn terms, on the other hand, f ( x ) is even and sin(( nπx
L ) is odd, so
the integrals over (0, L) and over ( L, 2L) are equal in magnitude and
opposite in sign and cancel. This gives

1 L
Z
A0 = f ( x )dx (6.7)
L 0
2 L  nπx 
Z
An = f ( x ) cos dx (6.8)
L 0 L
∞  nπx 
f ( x ) = A0 + ∑ An cos (6.9)
n =1
L

This is the Fourier cosine series for a function f ( x ) defined on (0, L).
The Fourier sine series is similar, but we make the odd extension
of f ( x ) across all boundaries. This again results in a function with
 fourier series 145

period 2L, which can be expanded in a Fourier series

1
Z 2L
A0 = f ( x )dx
2L 0
Z 2L
1  nπx 
An = f ( x ) cos dx
L 0 L
1 2L  nπx 
Z
Bn = f ( x ) sin dx
L 0 L
∞  nπx   nπx 
f ( x ) = A0 + ∑ An cos + Bn sin .
n =1
L L

This time when we reduce the integrals to integrals over the interval
(0, L) we find that when computing An the contributions from the
integration over (0, L) and the integration over ( L, 2L) are equal
in magnitude but opposite in sign, since f ( x ) is odd and cos even.
This means that all of the An terms are zero. For the Bn terms the
contributions from the integration over (0, L) and the integration over
( L, 2L) are the same, so the integral is just twice the integral over
(0, L). This results in
Z L
2  nπx 
Bn = f ( x ) sin dx (6.10)
L 0 L
∞  nπx 
f ( x ) = ∑ Bn sin . (6.11)
n =1
L

This is the Fourier sine series.

The difference between the Fourier cosine and sine series is il-
lustrated in the margin figure. The top graph depicts the original
function, which is defined only on (0, L). The middle figure depicts
the Fourier cosine series, and the bottom the Fourier sine series. Note
that all three functions agree in the original domain (0, L) but they
differ outside of this region.

Theorem 6.4.1 (Fourier Sine and Cosine Series). Given a function f ( x )

which is C2 on the interval [0, L]. The Fourier sine and cosine series are
defined as follows:
Sine Series

∞  nπx  Figure 6.9: The function

f (x) = ∑ Bn sin L
f ( x ) = x defined for x ∈ (0, 2)
n =1 (top) together with the even
2 L  nπx 
(middle) and odd (bottom)
Z
Bn = f ( x ) sin dx
L 0 L extensions.
The series converges to f ( x ) for x ∈ (0, L) and to 0 for x = 0, x = L.

Cosine Series
146 differential equations

∞  nπx 
f ( x ) = A0 + ∑ An cos
L
n =1
1 L
Z
A0 = f ( x )dx
L 0
2 L  nπx 
Z
An = f ( x ) cos dx
L 0 L

The series converges to f ( x ) for x ∈ [0, L].

Example 6.4.1. We consider the Fourier sine and cosine series for the
function f ( x ) = x on the interval x ∈ (0, 2). The Fourier cosine series is
given by

1 2
Z
A0 = xdx = 1
2 0
Z 2
2  nπx  4
An = x cos dx = 2 2 (cos(nπ ) − 1)
2 0 2 n π
∞  nπx 
f ( x ) = A0 + ∑ An cos
n =1
2
∞
4  nπx 
= 1 + ∑ 2 2 (cos(nπ ) − 1) cos
n =1 n π
2

Note that the integral for An is easily done by parts. Similarly the Fourier
sine series is given by

Z 2
2  nπx 
Bn = x sin dx
2 0 L
4
=− cos(nπ )
nπ
∞
(−1)n+1  nπx 
f (x) = ∑ sin .
n =1
nπ L

The Fourier cosine and Fourier sine series for f ( x ) = x defined for
x ∈ (0, 2) is depicted in margin figure 6.10. The graphs depict the first
fifty terms of the cosine series (top) and sine series (bottom). The fifty term
cosine series is essentially indistinguishable from the even extension of the Figure 6.10: The first fifty terms
function. The fifty term sine series is a good approximation but one can of the Fourier cosine (top) and
see a noticeable oscillation. This is the typical artifact associated with a sine (bottom) series for function
jump discontinuity, and is known as Gibbs phenomenon. This same effect f ( x ) = x defined for x ∈ (0, 2).
can sometimes be seen in image compression – the jpg algorithm uses a
discrete cosine series to do compression. At high compression rates one can
sometimes observe this ringing phenomenon in regions where the image has
a sharp transition from light to dark.
 fourier series 147

6.5 Fourier series practice problems

Exercise 6.5.1
Calculate the Fourier series expansions for the given functions de-
fined over one period. Pay attention to the interval in which they are
defined and adjust the limits of integration accordingly.

a) f ( x ) = 42 for x ∈ (0, π )
(
x for x ∈ (0, 2)
b) f ( x ) =
4−x for x ∈ (2, 4)
c) f ( x ) = 5 − 2x for x ∈ (−3, 3)
d) f ( x ) = 10 cos(2x ) for x ∈ (−π, π )
(
x for x ∈ (0, 1)
e) f ( x ) =
1 for x ∈ (1, 2)
f) f ( x ) = x2 for x ∈ (−5, 5)

Exercise 6.5.2
Calculate the Fourier sine series and the Fourier cosine series expan-
sions for the given functions.

a) f ( x ) = 42 for x ∈ (0, π )
(
x for x ∈ (0, 2)
b) f ( x ) =
4−x for x ∈ (2, 4)
c) f ( x ) = 5 − 2x for x ∈ (0, 3)
d) f ( x ) = 10 cos(2x ) for x ∈ (0, π )
(
x for x ∈ (0, 1)
e) f ( x ) =
1 for x ∈ (1, 2)
f) f ( x ) = x2 for x ∈ (0, 5)
7
Partial Differential Equations and Separation of Vari-
ables

7.1 The Heat Equation

The heat equation or diffusion equation governs phenomenon such

as the propagation of heat in a solid body or diffusion of a passive
tracer (such as a dye) in a liquid. In three spatial dimensions the heat
equation is the following partial differential equation.
 2
∂ u ∂2 u ∂2 u

∂u
=σ + 2+ 2 .
∂t ∂x2 ∂y ∂z

Here σ is a constant known as the (thermal) diffusivity, and u( x, y, z, t)

represents the termperature at location ( x, y, z) at time t or, in the
case of diffusion, the concentration of the tracer at location ( x, y, z) at
time t. In one spatial dimension the heat equation would be

∂u ∂2 u
= σ 2.
∂t ∂x
The one dimensional heat equation governs propagation of heat in a
one-dimensional medium such as a thin rod.
Let’s think a little about what this equation means. Imagine that
at some time t the temperature profile is given by u( x, t). If u( x, t)
has a local minimum (a cool spot) then u xx > 0 and hence ut > 0.
Conversely if u( x, t) has a local maximum, a hot spot, then u xx < 0
and hence ut < 0. So at any given time the hot spots will be getting
cooler and the cold spots will be getting warmer. The heat equation
is the mathematical expression of the idea that a body will tend to
move towards thermal equilibrium.
In solving ordinary differential equations we need to specify ini-
tial conditions. For the heat equation we need to specify an initial
temperature distribution, u( x, 0) = u0 ( x ), to specify what the initial
distribution of temperature in the rod looks like. Finally when we
150 differential equations

are solving the heat equation we usually are doing so on some finite
domain. We generally need to say something about what happens at
the boundary of the domain. There are different types of boundary
conditions. The two most important types are called Dirichlet and
Neumann conditions. A Dirichlet condition consists of specifying
u( x, t) at the boundary. This would correspond to specifying the tem-
perature at the boundary. We could imagine, for instance, putting
one end of the rod in a furnace at a fixed temperature and studying
how the rest of the rod heats up. The end of the rod in the furnace
would satisfy a Dirichlet condition.
A Neumann condition means that u x is specified at the boundary.
Physically specifying u x amounts to specifying the rate at which heat
is entering or leaving the body through the boundary. A homoge-
neous Neumann boundary condition, u x = 0, means that no heat
is entering or leaving the body through this boundary (the end is
insulated). For instance the heat equation

∂u ∂2 u
=σ 2
∂t ∂x
u( x, 0) = 0
u( L, t) = T0
u x (0, t) = 0

describes a rod of length L. Initially the rod is at zero temperature

(u( x, 0) = 0). One end of the rod is maintained at a constant temper-
ature T0 (u( L, t) = T0 ) while the other end of the rod is insulated so
that no heat is gained or lost through the end (u x (0, t) = 0).

7.1.1 The heat equation with Dirichlet boundary conditions.

In this section we consider the solution to the heat equation with
Dirichlet boundary conditions

∂u ∂2 u
=σ 2 (7.1)
∂t ∂x
u( x, 0) = u0 ( x ) (7.2)
u(0, t) = TLeft (7.3)
u( L, t) = Tright . (7.4)

Again this equation describes the evolution of the temperature as a

function of x and t, u( x, t), given that the initial temperature distri-
bution in the rod is given by u0 ( x ) and the ends are maintained at
temperatures TLeft and TRight respectively. The first step in solving
this equation is to do a change of variables to make the boundary
conditions homogeneous. The best way to think about this is as the
 partial differential equations and separation of variables 151

"equilibrium solution". An equilibrium solution is, by definition, one

that is not changing in time. Thus it should satisfy the following two
point boundary value problem:

∂2 u
=0 (7.5)
∂x2
u(0, t) = TLeft (7.6)
u( L, t) = Tright . . (7.7)

Note that this is just Equations (7.1)–(7.4) with ut = 0 (we are looking
for an equilibrium, so it should not change in time) and the initial
condition removed. This is easy to solve, as it is basically just an
ODE. The solution is
TRight − TLeft
uequilibrium ( x ) = TLeft + x.
L
If we now define the new function
TRight − TLeft
 
v( x, t) = u( x, t) − uequilibrium ( x ) = u( x, t) − TLeft + x
L
then it is straightforward to see that the function v( x, t) satisfies

∂v ∂2 v
=σ 2 (7.8)
∂t ∂x
TRight − tLeft
 
v( x, 0) = u0 ( x ) − TLeft + x (7.9)
L
v(0, t) = 0 (7.10)
v( L, t) = 0. (7.11)

So it is enough to be able to solve the heat equation with homoge-

neous Dirichlet boundary conditions. We will do this by the method
of separation of variables.

7.1.2 Separation of Variables

We begin with the heat equation with homogeneous Dirichlet bound-
ary conditions

∂v ∂2 v
=σ 2 (7.12)
∂t ∂x
v(0, t) = 0 (7.13)
v( L, t) = 0. (7.14)

We will incorporate the initial condition later. We are going to look

for a solution of a special form. We look for a solution that can be
written as a product of a function of x and a function of t:

v( x, t) = T (t) X ( x ).
152 differential equations

Substituting this into equation 7.12 gives the equation

dT d2 X
(t) X ( x ) = σT (t) 2
dt dx
dT d2 X
dt ( t ) dx2
(x)
=σ ,
T (t) X (x)

where the second equation follows from dividing through by X ( x ) T (t).

Notice that at this point we have the variables separated: all of the t
dependence is on one side of the equation and all of the x depen-
dence is on the other side. At this point we make an important ob-
servation: if we have a function of t equal to a function of x then both
must be constant. Considering the righthand side we set

d2 X
dx2
(x)
= −λ,
X (x)
d2 X
( x ) = −λX ( x )
dx2
Here the minus sign is not necessary, we have included it simply
for convenience. Now we would like to incorporate the boundary
condition that v = T (t) X ( x ) vanish at x = 0 and x = L. This implies
that X ( x ) must vanish at x = 0 and x = L. This gives us the two
point boundary value problem

d2 X
dx2
(x)
= −λ,
X (x)
d2 X
( x ) = −λX ( x )
dx2
X (0) = 0
X ( L) = 0.

This is an eigenvalue problem, one that we have already solved, in

fact. The solution is
nπx
X ( x ) = B sin( )
L
n2 π 2
λ= .
L2
This means that T (t) solves

dT n2 π 2
= −σλT = −σ 2
dt L
2 π2
−σ n t
which has the solution T (t) = e L2 . Thus the separated solution is
2 π2
−σ n t nπx
v( x, t) = Be L2 sin( ).
L
 partial differential equations and separation of variables 153

Note that this is a solution for every integer n. We can use super-
position to get a more general solution: since the sum of solutions is
a solution we have that
∞ 2 π2
−σ n t nπx
v( x, t) = ∑ Bn e L2 sin(
L
) (7.15)
n =1

is a solution.
2 π2
−σ n t
To summarize the function v( x, t) = ∑∞ n=1 Bn e
L2 sin( nπx
L )
satisfies the heat equation with homogeneous boundary conditions

vt = σv xx
v(0, t) = 0
v( L, t) = 0.

The only remaining condition is the initial condition v( x, 0) = v0 ( x ).

Substituting t = 0 into equation (7.2.2) gives the condition
∞
nπx
v( x, 0) = ∑ Bn sin( L
) = v0 ( x ). (7.16)
n =1

Note that we know how to solve this problem, and to compute the
coefficients Bn so the equation (7.16) holds. This is a Fourier sine
series, and thus we know that
Z L
2 nπx
Bn = v0 ( x ) sin( )dx.
L 0 L
We will summarize these results in the form of a theorem.

Theorem 7.1.1. The solution to the heat equation with homogeneous

Dirichlet boundary conditions

vt = σv xx
v( x, 0) = v0 ( x )
v(0, t) = 0
v( L, t) = 0

is given by the Fourier sine series

∞
nπx −σ n2 π22 t
v( x, t) = ∑ Bn sin( L
)e L
n =1
Z L
2 nπx
Bn = v0 ( x ) sin( )dx
L 0 L
Example 7.1.1. A steel rod is 2 m long. The thermal diffusivity of steel is
about σ = 20mm2 s−1 . The rod is initially at temperature v( x, 0) = 300
◦ C and the ends are cooled to 0◦ C. Find the temperature profile in the rod as
154 differential equations

a function of time. Assuming that heat is lost only through the ends of the
rods how long until the maximum temperature in the rod is 100◦ C?
The heat equation is

vt = σv xx
v( x, 0) = 300◦ C x ∈ (0, 2)
◦
v(0, t) = 0 C
v (2 , t ) = 0◦ C

It is easiest to work consistently in meters: a thermal diffusivity of σ =

20mm2 s−1 is equal to 2 × 10−5 m2 s−1 . The Fourier coefficients are given
by

2
Z2m nπx
Bn = 300◦ C sin( )dx
2m 0 2m
1 2m nπx 2 m
=− cos( )|
1 m nπ 2m 0
600◦ C
= (1 − cos(nπ ))
nπ
and so the solution is given by
∞
600◦ C 2 2
−2×10−5 m2 s−1 n π2 t nπx
v( x, t) = ∑ nπ
(1 − cos(nπ )) e 4 m sin(
2m
).
n =1

This can be simplified a bit if we notice that 1 − cos(πn) is equal to 0 for n

even and 2 for n odd, so the above expression is the same as
∞
1200◦ C −2×10−5 m2 s−1 (2n−1)22 π2 t (2n − 1)πx
v( x, t) = ∑ (2n − 1)π
e 4 m sin(
2m
).
n =1

It is clear that the maximum temperature in the rod will be at the center.
Plotting the function
Figure 7.1: The temperature (in
∞
1200◦ C −2×10−5 m2 s−1 (2n−1)22 π2 t 1 ◦ C) at the center of the rod as a
v(1 m, t) = ∑ e 4 m sin((n − )π )
( 2n − 1 ) π 2
n =1 function of time.
(Figure 7.1) we find the depicted graph of the temperature at the center of
the rod. One can see that the temperature at the center of the rod first dips
below 100◦ C at time t ≈ 20, 000s.

7.1.3 The heat equation with Neumann boundary conditions.

In the previous section we saw that the heat equation with

Dirichlet (zero temperature) boundary conditions could be solved
by the method of separation of variables. The same basic technique
will apply to many different kinds of boundary conditions. Perhaps
 partial differential equations and separation of variables 155

the next most important type of boundary condition is Neumann,

or zero flux boundary conditions. In the heat equation context a ho-
mogeneous (zero) Neumann boundary condition means that there is
zero flux of heat through the boundary. In other words the boundary
is insulated.
A typical heat problem with Neumann boundary conditions might
be something like the following: It is worth thinking about how the so-
lution should behave. It the boundaries
are insulated then after a long time
vt = σv xx
the temperature should approach the
v( x, 0) = v0 ( x ) average of the initial temperature. We
will see later that this is true.
v x (0, t) = 0
v x ( L, t) = 0

We can solve this by the same method, separation of variables, that

we used before. Again we begin by looking for a solution in the form
of a product v( x, t) = X ( x ) T (t). Substituting this in to the equation
vt = v xx gives

T ′ (t) X ( x ) = σT (t) X ′′ ( x ) (7.17)

T ′ (t) X ′′ ( x )
=σ (7.18)
T (t) X (x)

As in the previous section we notice that if we have a function of

only t equal to a function of only x then they must both be constant.
This tells us that X ′′ ( x )/X ( x ) = −λ. At this point we impose the We have absorbed the factor of σ into
boundary conditions, which gives us an eigenvalue problem for our definition of the constant λ, as
well as included the minus sign, for
X ( x ): convenience.

X ′′ ( x ) = −λX ( x )
X ′ (0) = 0
X ′ ( L) = 0

As before it is easiest to separate this into three cases: λ > 0, λ = 0

and λ < 0 depending on the nature of the roots of the characteristic
equation. In the case λ > 0 the general solution is
√ √
X ( x ) = A cos( λx ) + B sin( λx ).

Taking the derivative and substituting x = 0 gives

√
X ′ (0) = B λ = 0,

and since λ > 0 we must have that B = 0. Taking the derivative and
substituting x = L gives
√ √ √ √ √ √
X ′ ( L) = − A λ sin( λL) + B λ cos( λL) = − A λ sin( λL) = 0,
156 differential equations

Here we have used the fact that B = 0. Once again since λ > 0 we
√
must have that either A = 0 or sin( λL) = 0. We are interested in
finding a non-zero solution: if A = 0 then our solution is identically
√ √
zero. So we require that sin( λL) = 0. This implies that λL = nπ
2 2
or λ = n Lπ2 . This gives X ( x ) = A cos( nπx
L ). Substituting this back
into Equation (7.18) we find that
T ′ (t) X ′′ ( x ) n2 π 2
=σ = −σ 2
T (t) X (x) L
n2 π 2
T ′ (t) = −σ T (t)
L2
2 π2 t
− σn
So the separated solutions for λ > 0 are T (t) X ( x ) = An e L2 cos( nπx
L ).
For λ = 0 we have X ′′ ( x ) = 0 and so X ( x ) = A + Bx. Imposing the
boundary conditions gives
X ′ (0) = B = 0
X ′ ( L) = B = 0.
So there are no conditions on A. Thus X ( x ) = A is a solution and the
corresponding T (t) solves
T ′ (t) X ′′ ( x )
=σ =0
T (t) X (x)
and so T (t) is also constant. Putting them together we have T (t) X ( x ) =
A0 .
Finally for λ < 0 the general solution is
√ √
−λx λx
X ( x ) = Ae + Be .
It is easy to check that there are no eigenvalues in this case: the only
solution satisfying X ′ (0) = 0 and X ′ ( L) = 0 is X ( x ) = 0.
Now that we have found all of the separated solutions we can
combine them: the equation is linear, so any linear combination of
solutions is a solution. This gives us a more general solution
nπx −σ n2 π22 t
v( x, t) = A0 + ∑ An cos( )e L .
L
This function satisfies the partial differential equation vt = σv xx
along with the boundary conditions v x (0, t) = 0 and v x ( L, t) = 0. The
only thing remaining is the initial condition:
nπx
v( x, 0) = A0 + ∑ An cos( ) = v0 ( x ).
L
This is a Fourier cosine series problem! We already know that
1 L
Z
A0 = v0 ( x )dx
L 0
Z L
2 nπx
An = v0 ( x ) cos( )dx.
L 0 L
We state this as a theorem
 partial differential equations and separation of variables 157

Theorem 7.1.2. The solution to the heat equation with Neumann boundary
conditions

vt = σv xx
v( x, 0) = v0 ( x )
v x (0, t) = 0
v x ( L, t) = 0

is given by Note the behavior when the time t

gets large. The exponential terms
decay to zero, and all that remains is
nπx −σ n2 π22 t the constant term. So for long times
v( x, t) = A0 + ∑ An cos( )e L t ≫ 1 we have that v( x, t) ≈ A0 =
L 1 L
L 0 v0 ( x ) dx – the temperature profile
R
1 L
Z
A0 = v0 ( x )dx tends to the average of the initial
L 0 temperature profile. This is exactly
2 L
Z
nπx what we would expect.
An = v0 ( x ) cos( )dx.
L 0 L
158 differential equations

Exercise 7.1.1

Determine the solution to the following diffusion equations with the

given initial conditions and boundary conditions using separation of
variables.

a) b)

vt = 5 v xx vt = 2 v xx
v( x, 0) = x v( x, 0) = 5 − x
v(0, t) = 0 v x (0, t) = 0
v(3, t) = 0 v x (5, t) = 0

c)

vt = 3 v xx
v( x, 0) = 10
v x (0, t) = 0
v(π, t) = 0

7.2 The One-dimensional Wave Equation

7.2.1 Background

The wave equation is a partial differential equation that governs

many types of wave propagation. In one spatial dimension the wave
equation is given by
∂2 v 2
2∂ v
= c .
∂t2 ∂x2
This is the equation that governs the vibrations of a string, such as
a guitar or piano string. Here v( x, t) represents the displacement
of the string at location x and time t. As is the case with the heat
equation this needs to be supplemented with some initial conditions
and some boundary conditions. Since the equation is second order
in time we expect to have to impose two initial conditions. Generally
one specifies v( x, 0) and vt ( x, 0), representing the displacement and
velocity of the string at time t = 0. Since the equation is second order
in space we also expect to have two boundary conditions. For the
case of a vibrating string the most natural boundary conditions are
Dirichlet, implying that the string is fixed at the ends. This leads to
 partial differential equations and separation of variables 159

the following partial differential equation

vtt = c2 v xx (7.19)
v(0, t) = 0 (7.20)
v( L, t) = 0 (7.21)
v( x, 0) = f ( x ) (7.22)
vt ( x, 0) = g( x ). (7.23)

We could go through the separation of variables argument that we

used for the heat equation again, but it is perhaps nicer to use the
fact that we already know the eigenfunctions of the Dirichlet eigen-
value problem. We know that the eigenvalue problem

u xx + λu = 0
u(0, t) = 0
u( L, t) = 0

nπ 2
has the solutions un ( x ) = sin( nπx
L ) and λn = . This suggests


L
that we expand the solution in terms of the eigenfunctions. Looking
for a solution in the form
∞
nπx
v( x, t) = ∑ βn (t) sin( L
)
n =1

and substituting it into (7.19) we find that

∞ ∞
d2 β n nπx n2 π 2 c2 nπx
∑ 2
sin (
L
) = ∑ − 2
β n sin(
L
).
n=1 dt n =1 L

The functions {sin( nπx ∞

L )}n=1 are orthogonal, which implies that they
are linearly independent. From this it follows that the coefficients
satisfy the ordinary differential equation

d2 β n n2 π 2 c2
2
=− βn .
dt L2
This is a constant coefficient linear differential equation, the solution
is given by
nπct nπct
β n (t) = Bn cos( ) + Cn sin( )
L L
This gives
∞
nπx nπct nπx nπct
v( x, t) = ∑ Bn sin( L
) cos(
L
) + Cn sin(
L
) sin(
L
) (7.24)
n =1

The constants Bn , Cn are determined by the initial data, but before

we discuss this we first comment a bit on the physical interpretation.
160 differential equations

7.2.2 Interpretation
One thing that we notice about the solution to the wave equation
given in equation (7.24) is that the time dependence is sinusoidal.
This is in contrast to the solution to the heat equation , where the
time dependence is decaying exponential. This reflects the different
physics behind the heat and wave equations: heat problems tend
to decay to some kind of equilibrium. This decay is reflected in the
(decaying) exponential dependence on time. The wave equation, on
the other hand, supports sustained oscillations. This is reflected in
the sinusoidal dependence on time.
In general a solution to the wave equation is given by a superpo-
sition of terms of the form sin( nπx nπct
L ) cos( L ). These building blocks
n
are called standing waves. The wavenumber of these waves is k = 2L
nc
and the angular frequency is ω = 2L . The tone produced by a vi-
brating string will contain of a number of different frequencies. The
lowest frequency is n = 1: this is called the fundamental. The next
is n = 2: this has a frequency twice the fundamental frequency. This
is exactly one octave above the fundamental. The next mode, n = 3,
produces has a frequency three times the fundamental. In music a
pair of tones in a perfect fifth have frequencies in the ratio 3:2, so
n = 3 and n = 2 form a perfect fifth. Similarly n = 4 and n = 3
form a fourth, etc. Much of our music theory developed the way it
did because the first instruments were stringed instruments.

7.2.3 Solving for the coefficients

The solution given in (7.24) satisfies the wave equation with homoge-
neous Dirichlet boundary conditions.
∂2 v ∂2 v
2
= c2 2
∂t ∂x
v(0, t) = 0
v( L, t) = 0.

The only thing that remains is to satisfy the initial conditions v( x, 0) =

f ( x ) and vt ( x, 0) = g( x ). This amounts to a Fourier sine series prob-
lem. Substituting t = 0 into (7.24) we find that
∞
nπx nπct nπx nπct
v( x, t) = ∑ Bn sin( L
) cos(
L
) + Cn sin(
L
) sin(
L
)
n =1
∞
nπx
v( x, 0) = ∑ Bn sin( L
) = f (x)
n =1

We know from orthogonality that the solution is given by

Z L
2 nπx
Bn = f ( x ) sin( )dx.
L 0 L
 partial differential equations and separation of variables 161

Similarly if we take the partial derivative with respect to t and evalu-

ate at t = 0 we find that

∞
nπx nπct nπx nπct
v( x, t) = ∑ Bn sin( L
) cos(
L
) + Cn sin(
L
) sin(
L
)
n =1
∞
∂v nπc nπx nπct nπc nπx nπct
( x, t) = ∑ − Bn sin( ) sin( )+ Cn sin( ) cos( )
∂t n =1
L L L L L L
∞
∂v nπc nπx
( x, 0) = ∑ Cn cos( ) = g( x )
∂t n =1
L L

This is again a Fourier sine series problem, but this time the coeffi-
cient is nπc
L Cn . Thus we have

nπc 2 L nπx
Z
Cn = g( x ) sin( )dx
L L 0 L
Z L
2 nπx
Cn = g( x ) sin( )dx
nπc 0 L

Example 7.2.1. Solve the wave equation

vtt = v xx
v(0, t) = 0
v(1, t) = 0
v( x, 0) = 0
vt ( x, 0) = 1

corresponding to a string with no initial displacement and an initial velocity

of 1. The Fourier coefficients are given by

Bn = 0
Z L
2 nπx 2L
Cn = sin( )dx = 2 2 (1 − cos(nπ ))
nπc 0 L n π c

given the solution

∞
2L nπx nπt
v( x, t) = ∑ n 2 π2
(1 − cos(nπ )) sin(
L
) sin(
L
)
n =1

Exercise 7.2.1

Determine the solution to the following wave equations with the

given initial conditions and boundary conditions using separation of
variables.
162 differential equations

a) b)

vtt = 9 v xx vtt = 16 v xx
v(0, t) = 0 v(0, t) = 0
v(2, t) = 0 v(3, t) = 0
v( x, 0) = x v( x, 0) = 6
vt ( x, 0) = 0 vt ( x, 0) = 5 sin(πx )
 Part III

Background Material and

Solutions
8
Background Material

8.1 A Review of complex numbers and the Euler formula

In this we review some background material on complex numbers

and the Euler formula. Recall that the imaginary number i is defined
√
as i = −1 or i2 = −1. A complex number is defined to be a number
of the form a + bi where a and b are real numbers. Complex numbers
are important for a number of reasons. The first is the fundamental
theorem of algebra: The concept of multiplicity is im-
portant. The multiplicity of a root xi
Theorem 8.1.1. Let Pn ( x ) be a polynomial of degree n. Then Pn ( x ) of a polynomial P( x ) is the largest
power k such that ( x − xi )k divides
has exactly n roots counted according to multiplicity: there are n com- P( x ). For instance the polynomial
plex numbers x1 , x2 , . . . xn (not necessarily distinct) such that Pn ( x ) = P( x ) = x3 − 3x + 2 has three roots,
A( x − x1 )( x − x2 ) . . . ( x − xn ). x = −2 has multiplicity one and
x = 1 has multiplicity two, since
x3 − 3x + 2 = ( x − 1)2 ( x + 2). We
The fundamental theorem of algebra requires that we work over
sometimes call roots of multiplicity one
the complex numbers: there are polynomials with no real roots. "simple".

Example 8.1.1. Find the roots of P( x ) = x4 − 1.

Using the identity ( a2 − b2 ) = ( a − b)( a + b) we have that p( x ) =
( x − 1) = ( x2 − 1)( x2 + 1) = ( x − 1)( x + 1)( x2 + 1). The product is zero
4

if and only if one of the factors is zero. The first factor is zero if x − 1 = 0
or x = 1. The second is zero if ( x + 1) = 0 or x = −1. The third factor
√
is zero if x2 + 1 = 0 or x = ± −1 = ±i. Thus the four roots are
x = 1, x = −1, x = i, x = −i.

Of course we need to be able to do algebra with complex numbers.

In particular we need to be able to add, subtract, multiply and divide
them. Addition and subtraction are defined component-wise:

( a + ib) + (c + id) = ( a + c) + i (b + d) ( a + ib) − (c + id) = ( a − c) + i (b − d)

Multiplication is not defined pointwise, but rather it is defined to

be consistent with the distributive property: we FOIL it out

( a + ib)(c + id) = ac + iad + ibc + (ib)(id) = ac + iad + ibc − bd = ( ac − bd) + i ( ad + bc).

166 differential equations

Division is a bit harder to define. We first define the complex conju-

gate. If z = a + ib is a complex number then z̄ (sometimes denoted
z∗ is defined to be z̄ = a − ib. Note that zz̄ = ( a + ib)( a − ib) =
a2 − (ib)2 = a2 + b2 ≥ 0. Now we can define division by

a + ib a + ib c − id ( a + ib)(c − id)
= = .
c + id c + id c − id c2 + d2

For instance to compute 52+ +3i = 2+3i 5−7i = 31+i = 31 + 1 i. The

7i 5+7i 5−7i 25+49 74 74
modulus of a complex number, denoted |z| = | a + ib| is defined by
√ p
|z| = | a + ib| = zz̄ = a2 + b2 .

There are a couple of important geometric interpretations of complex

numbers. First is the complex plane

Definition 8.1.1 (Complex Plane). Given a complex number z = a + ib

we can identify this complex number with the point ( a, b) in the x − y
plane.

Under this definition complex numbers add in the way that vec-
tors normally do: they add component-wise. We will give an inter-
pretation to multiplication shortly, but first we introduce the polar
representation. First we need to recall the Euler formula, which tells
us how to exponentiate complex numbers.

Theorem 8.1.2 (Euler). The complex exponential is defined by

eiθ = cos(θ ) + i sin(θ )

More generally if z = a + ib then we have

ez = e a+ib = e a (cos(b) + i sin(b)).

The Euler formula links the transcendental functions e x , cos x and

sin x through the imaginary number i. A second way to represent
complex numbers is through the polar representation. This is essen-
tially polar coordinates on the x − y plane.

Definition 8.1.2. Given a complex number z = a + ib we can represent it

in polar form as
z = reiθ

The quantities r and θ, referred to as the modulus and the argument, are
defined by
p b
r= a2 + b2 = | z | tan(θ ) =
a
a = r cos θ b = r sin θ
 background material 167

Going from (r, θ ) to ( a, b) is straightforward. Going from ( a, b) to (r, θ )

is straightforward EXCEPT that one must be careful to choose the correct
branch of the arctangent. There will be two solutions that differ by π, and
you must be careful to choose the correct one. Note that θ is only defined up
to multiples of 2π.

Example 8.1.2. Express the complex number −5 − 5i in polar form.

√
The modulus is given by r = (−5)2 + (−5)2 = 5 2. The argument θ
p

is defined by
−5
tan θ = =1
−5
Here we have to be careful: there are two angles θ such that tan(θ ) = 1,
θ = π4 , in the first quadrant and θ = 5π
4 , in the third quadrant. The original
point is −5 − 5i = (−5, −5) in the third quadrant, so
√ 5π
(−5 − 5i ) = 5 2e 2 i

The argument arg(z) = θ and the modulus |z| have some nice
properties under multiplication:

Lemma 8.1.1. If z1 and z2 are complex numbers and arg(zi ) = θi is the

argument or angle in the polar representation then

|z1 z2 | = |z1 ||z2 | and arg(z1 z2 ) = arg(z1 ) + arg(z2 )

in the second equation we have the understanding that the argument is only
defined up to multiples of 2π. In other words when we multiply complex
numbers the absolute values multiply and the angles add.

Lemma 8.1.2. From Euler’s formula

eiθ = cos(θ ) + i sin(θ )

we also have
e−iθ = cos(θ ) − i sin(θ )
and

1  iθ 
cos(θ ) = e + e−iθ (8.1)
2
1  iθ 
sin(θ ) = e − e−iθ (8.2)
2i
We can actually prove these two formulas using differential equa-
tions. The left side of (8.1), cos(θ ), solves the initial value problem

y′′ + y = 0, y(0) = 1, y ′ (0) = 0

The right side of (8.1), 12 eiθ + e−iθ , also solves this initial value

problem. But the Theorem of Existence and Uniqueness guarantees

168 differential equations

the uniqueness of the solution, therefore these two solutions must be

the same, which proves (8.1). Formula (8.2) can be similarly proven
using the initial value problem

y′′ + y = 0, y(0) = 0, y ′ (0) = 1

 background material 169

8.2 A Review of linear algebra

This is a short review of material from linear algebra that we will

need in the course.

8.2.1 Matrices
A matrix M is just a rectangular array of numbers, M jk . Some exam-
ples include
 
1
M1 =  2 
 
−7
 
3 2
M2 =  1 8 
 
0 −5
 
2 1 3 −2
M3 =  − 4 6 0 8 
 
1 0 0 −5
There are a number of algebraic operations that we can perform
on matrices. Firstly we can multiply a matrix by a constant (scalar) a.
This simply amounts to multiplying each entry of the matrix by the
same scalar. For example

 
3 2
M2 =  1 8 
 
0 −5
 
9 6
3 · M2 = 3 24 
 
0 −15
 
2 1 3 −2
M3 =  − 4 6 0 8 
 
1 0 0 −5
 
−8 −4 −12 8
− 4 · M3 =  16 −24 0 −32
 
−4 0 0 20
Next we can add two matrices provided they have the same dimen-
sions! Addition is done termwise – we add the corresponding matrix
entries. For example
     
3 2 1 7 4 9
1 8  + 0 −1 = 1 7 
     
0 −5 2 −3 2 −8
170 differential equations

Obviously multiplication by a scalar and matrix addition respect

the usual rules of algebra, for instance M + 3M = 4M,

8.2.2 Matrix multiplication

Matrices of compatible shapes can be multiplied. Compatible shapes
means that if we want to take the product MN then the number of
columns of M (or number of elements in a row of M) must be the
same as the number of rows of N (or number of elements in a col-
umn of N), and the resulting matrix will have the same number of
rows as M and the same number of columns as N. Note that this
means the order in which we multiply the matrices is important. For
instance if M is a 3 × 7 matrix and N is a 7 × 4 matrix then MN is
well-defined and is a 3 × 4 matrix, but the product NM is NOT DE-
FINED, since N is 7 × 4 and M is 3 × 7. This is the first instance of
something that will be important going forward. Unlike multiplica-
tion of real numbers and complex numbers, multiplication of matrices
is not commutative – the order matters.

Definition 8.2.1. If M is a k × n matrix and N is an n × m matrix then

the product MN is a k × m matrix whose entries are defined as follows
n
(MN) jl = ∑ Mjk Nkl
k =1

If you have been exposed to the "dot product" or "scalar product" in one of
your other classes the ( j, l ) entry of MN is given by the dot product of the
jth row of M with the l th column of N.

Again we want to emphasize the fact that matrix multiplication

is generally not commutative. Even if MN and NM are both defined
and have the same dimensions they will generally not be the same.

Example 8.2.1. Consider the matrices

 
1 −1 !
3 1 2
M = 2 4  N=
 
−1 5 4
7 2

The product MN is a 3 × 3 matrix given by

 
1 · 3 + (−1) · (−1) 1 · 1 + (−1) · 5 1 · 2 + (−1) · 4
MN =  2 · 3 + 4 · (−1) 2·1+4·5 2·2+4·4 
 
7 · 3 + 2 · (−1) 7·1+2·5 7·2+2·4
 
4 −4 −2
MN =  2 22 20 
 
19 17 22
 background material 171

In this case the product NM is also well defined. Check that it is a 2 × 2

matrix given by !
19 5
NM =
37 29

Exercise 8.2.1
Find pairs of matrices M and N such that the following hold
a) M + N is well-defined but MN is not.
b) MN is well-defined but M + N is not.
c) M + N , MN and NM are all well-defined but MN ̸= NM.
One very important special matrix is the identity matrix, usually
denoted I or In×n . The n × n matrix I has entries of 1 down the main
diagonal and 0 everywhere else.
Example 8.2.2.
 
! 1 0 0
1 0
I2×2 = I3 × 3 = 0 1 0
 
0 1
0 0 1
The identity matrix has the property that if I is the n × n identity
matrix and N is any n × m matrix then

IN = N

Similarly if M is any k × n matrix then

MI = M

Matrix multiplication includes the special case of vectors – a col-

umn vector is an n × 1 matrix, while a row vector is a 1 × n matrix. In
this text we will mostly deal with vectors and square (n × n) matrices.
Note that two square matrices M and N can always be multiplied
in either order, although the results will generally differ. For square
(n × n) matrices we have an important quantity associated to the
matrix called the determinant.

8.2.3 Determinants
We next introduce the idea of a determinant. For any n × n matrix M
we define a scalar quantity, linear in the rows of the matrix, called the
determinant. The determinant is a measure of the independence of
the rows of the matrix. The determinant can be defined axiomatically,
by the following axioms. Suppose that r1 , r2 . . . rn are the n rows of
the matrix, and let det(r1, r2, . . . , rn ) denote the determinant. Then
the determinant is the unique function satisfying the following three
properties.
172 differential equations

1. Linearity: The determinant is linear in each row, e.g. det( ar1 +

br1′ , r2 , r3 , . . . rn ) = a · det(r1 , r2 , r3 , . . . rn ) + b · det(r1′ , r2 , r3 , . . . rn ).

2. Behavior under permutation: If we switch two rows then the sign of

the determinant switches, e.g. det(r1 , r2 , r3 , . . . rn ) = − det(r2 , r1 , r3 , . . . rn ).

3. Normalization: the determinant of the identity matrix is 1.

Note that the property that the sign switches if we swap two rows
implies that the determinant is equal to zero if two rows are equal,
since that implies that det(r1 , r1 , r3 , . . . rn ) = − det(r1 , r1 , r3 , . . . rn ) and
thus det(r1 , r1 , r3 , . . . rn ) = 0. Combining this with linearity we see
that if any row is equal to a sum of other rows then the determinant
is zero.
Here are some useful properties of the determinant:

Observation 8.2.1. The determinant has the following properties:

• The determinant is zero if and only if the rows are linearly dependent.

• The determinant is zero if and only if the columns are linearly dependent.

• The determinant is equal to the (signed) volume of the parallelepiped

spanned by r1 , r2 , . . . , rn

Henceforth we will generally denote the determinant of a matrix

by writing the matrix with vertical bars:
!
a b a b
= det .
c d c d

It is always possible to compute the determinant of a matrix by

applying elementary row operations and using the three axioms, but
this is very tedious in practice. Here are some better ways to proceed.

Observation 8.2.2. The determinants of a general 2 × 2 and 3 × 3 matrix

are given as follows
!
a b
det = a·d−b·c
c d

and
 
a b c
det  d e f = a · e ·i + b · f · g + c · d · h − g · e · c − h · f · a −i · d · b
 
g h i

Note: if you have noticed a pattern here be aware that it doesn’t general-
ize in the most obvious way to 4 × 4 matrices. The “down diagonals minus
up diagonals” formula does not work for 4 × 4 matrices.
 background material 173

For larger matrices, especially ones with a large number of zero

entries, it is easiest to use “expansion by minors”, which is a way to
reduce the determinant of an N × N matrix to a sum of N smaller
determinants.

Definition 8.2.2. Given an N × N matrix M the i, j minor of M is the

determinant of the matrix obtained by striking out the ith row and the jth
column. We will denote the i, j minor determinant by detij (M).

Example 8.2.3. If M is the matrix

 
8 1 −7
M = 3 2 4 
 
6 0 4

then the (2, 3) minor would be the determinant of the matrix obtained by
removing the 2nd row and the 3rd column.

8 1 −7
8 1
3 2 4 = = 8 · 0 − 6 · 1 = −6
6 0
6 0 4

Method 8.2.1 (Minor Expansion). The determinant has an expansion

in minors as follows. Given any row or column we sum over all the entries
in the row or column the entry Mi,j times the i, j minor determinant times
(−1)i+ j .
det(M) = ∑(−1)i+ j Mij det
ij
(M)
i

The above formula gives the expansion by column minors, exchanging the
role of i, j gives the formula for row minors.

Example 8.2.4. Compute the determinant of the matrix

 
8 1 −7
M = 3 2 4 
 
6 0 4

using expansion in the second row, (3 2 4). We get

8 1 −7 8 1 −7 8 1 −7
det(M) = (−1)3 · 3 · 3 2 4 + (−1)4 · 2 · 3 2 4 + (−1)5 · 4 · 3 2 4
6 0 4 6 0 4 6 0 4
1 −7 8 −7 8 1
= (−1) · 3 · +2· + (−1) · 4 ·
0 4 6 4 6 0
= (−3) · 4 + 2 · (8 · 4 + 6 · 7) − 4 · (−6)
= 160
174 differential equations

Here are some tips for computing determinants. Firstly note that
while the minors expansion works for any row or column it is eas-
iest to choose a row or a column that contains many zero entries,
since one does not need to compute the corresponding minor deter-
minants. If, for instance, we had done the expansion in the second
column instead of the second row we would not need to compute the
(3, 2) minor determinant, as it would be multiplied by zero. Secondly
notice that the factor (−1)i+ j alternates between +1 and −1. To get
the correct sign draw a checkerboard pattern with + in the upper
left-hand corner and alternating + and − signs. The the entries give
the sign of (−1)i+ j .  
+ − +
− + −
 
+ − +
so the signs going across the second row are (− + −). This factor
is in addition any signs on the matrix coefficients themselves, of
course.

Example 8.2.5. Applying these tricks to the previous example we can see
that it would be easier to expand in either the second column or the third
row. Expanding in the third row gives

1 −7 8 1
det(M) = +6 +4 = 6 · (4 − (−14)) + 4 · (16 − 3) = 108 + 52 = 160
2 4 3 2

in agreement with the previous calculation.

8.2.4 Systems of Linear Equations

We consider a system of N linear equations in N unknowns, which
we can write as a matrix equation.

Mx = b

where M is a square matrix. The main theorem governing such sys-

tems is the following:

Theorem 8.2.1. There is a unique solution to the linear system

Mx = b

if and only if det(M) ̸= 0. Assuming that det(M) ̸= 0 the solution to the

set of linear equations is given by Cramer’s rule:
| Mi |
xi =
|M|

where Mi represents the matrix obtained by replacing the ith column of M

with the righthand side b.
 background material 175

Example 8.2.6. Solve the linear system

8x1 + x2 − 7x3 = 11
3x1 + 2x2 + 4x3 = 0
6x1 + 4x3 = 5

This is equivalent to the matrix system

    
8 1 −7 x1 11
3 2 4   x2  =  0 
    
6 0 4 x3 5

Note that in the previous example we computed this determinant to be

|M| = 160. By Cramer’s rule we have the solution

11 1 −7
0 2 4
5 0 4 178 89
x1 = = =
8 1 −7 160 80
3 2 4
6 0 4
8 11 −7
3 0 4
6 5 4 −133
x2 = =
8 1 −7 160
3 2 4
6 0 4
8 1 11
3 2 0
6 0 5 −67
x3 = =
8 1 −7 160
3 2 4
6 0 4

Exercise 8.2.2
Check that the determinants in the numerators of the expressions for
x1,2,3 are equal to 178, −133 and −67 respectively.
A closely related result is the following:

Theorem 8.2.2. Suppose that M is a square matrix. The linear homoge-

neous equation
Mx = 0
has a non-zero solution x if and only if det(M) = 0. Note that the set of
solutions is closed under addition – if x and y are two solutions to Mx = 0
176 differential equations

then any linear combination z = αx + βy also solves Mz = 0. The set of all

solutions to Mx = 0 is a vector space called the null-space of M.

8.2.5 Eigenvalues and eigenvectors

Given a square matrix M there are special vectors associated to M,
called "eigenvectors".

Definition 8.2.3. Given a square matrix M then v ̸= 0 is an eigenvector of

M and λ is an eigenvalue of M if

Mv = λv.

Example 8.2.7. For the matrix

 
4 1 2
0 2 1
 
0 1 2

there are three eigenvalues, λ = 4, λ = 3, λ = 1. The corresponding

eigenvectors are given by
     
1 −3 −1
v1 = 0 v2 =  1  v3 = −3
     
0 1 3

To see this note that

      
4 1 2 1 4 1
0 2 1 0 = 0 = 4 0
      
0 1 2 0 0 0

and similarly for the other two vectors.

Eigenvectors are nice because nice because the action of the ma-
trix on an eigenvector is particularly simple – the vectors just get
"stretched" by a scalar factor, the eigenvalue.
Next we outline how to find the eigenvalues and eigenvectors.

Theorem 8.2.3. The characteristic polynomial of a square matrix M is

defined to be
P(λ) = det(M − λI)

where I is the identity matrix. The eigenvalues of M are the roots of the
characteristic polynomial.

Proof. Recall that the eigenvalue equation is

Mv = λv.
 background material 177

This can be rewritten as

(M − λI)v = 0.

We know from Theorem 8.2.2 that the only solution to this equation
is v = 0 unless det(M − λI) = 0. Since we require a non-zero solution
λ must be a root of the characteristic polynomial.

Example 8.2.8. Find the eigenvalues of the matrix M where

 
5 2 −1
M = 0 3 3 
 
0 3 3

Computing the characteristic polynomial we find that

5−λ 2 −1
3−λ 3
P(λ) = 0 3−λ 3 = (5 − λ ) = (5 − λ)(λ2 − 6λ)
3 3−λ
0 3 3−λ

so the eigenvalues are λ = 0, λ = 5, λ = 6.

Having found the eigenvalues we can then use this information to

find the eigenvectors. To do this we solve the linear equation

(M − λI)v = 0

This equation is consistent (v = 0 is a solution, so we are guaranteed

that a solution exists. In fact we know that there are many solutions,
since any constant multiple is a solution.

Example 8.2.9. Find the eigenvector of

 
5 2 −1
M = 0 3 3 
 
0 3 3

corresponding to the eigenvalue λ = 6,

The eigenvector equation becomes
  
5−6 2 −1 v1
Mv =  0 3−6 3   v2  = 0
  
0 3 3−6 v3
  
−1 2 −1 v1
Mv =  0 −3 3  v2  = 0
  
0 3 −3 v3
178 differential equations

This is equivalent to the three equations

− v1 + 2v2 − v3 = 0
− 3v2 + 3v3 = 0
3v2 − 3v3 = 0

Note that the second and third equations are linearly dependent so we only
need one of these. Thus we must solve

− v1 + 2v2 − v3 = 0
− 3v2 + 3v3 = 0

The second equation implies that v2 = v3 . Substituting this into the first
equation shows that −v1 + v2 = 0 or v1 = v2 . Thus all three components of
v must be equal. We could choose, for instance, v3 = 1. This would give
 
1
v = 1
 
1

Another choice might be to take v3 = √1 , since then v3 is a unit vector:

3
 
1
1  
v = √ 1
3
1

Note that the characteristic polynomial always has n roots, counted

by multiplicity, but if we have a root of multiplicity k we do not al-
ways have k linearly independent eigenvectors. The following exam-
ple illustrates this phenomenon, called deficiency.

Example 8.2.10. Find the eigenvectors of

 
1 2 0
M = 0 1 1
 
0 0 1

The characteristic polynomial is (λ − 1)3 , so λ = 1 is a root of the polyno-

mial with multiplicity 3. To find the eigenvector we must solve
  
0 2 0 v1
(M − I)v = 0 0 1  v2  = 0
  
0 0 0 v3

which is equivalent to the three equations

2v2 = 0
v3 = 0
0 = 0.
 background material 179

The third is obviously always true, so we conclude that v2 = 0 and v3 =

0. Since there are no conditions on v1 it can be chosen arbitrarily, so the
eigenvector is any multiple of  
1
0
 
0
Even though λ = 1 is a root of multiplicity 3 there is only one linearly
independent eigenvector.
9
Problem Solutions

Solutions to the Exercises

Solution 1.1.1

Given the values of the first and second derivatives at t = 0 we

can find the value of the second derivative by substituting into the
equation

d2 y dy
( 0 ) = y 3 ( 0 ) − y ( 0 ) = 13 − 1 = 0 y (0) = 1 (0) = −1.
dt2 dt

To find the value of the third derivative we first differentiate the

entire equation and then evaluate at t = 0.

d2 y
= y3 − y
dt2
d3 y dy dy
3
= 3y2 −
dt dt dt
d3 y dy dy
(0) = 3y2 (0) (0) − (0) = 3 · 12 (−1) − (−1) = −2
dt3 dt dt

Solution 1.1.3

y = 1 − t2
dy
= −2t
dt
dy
t − y = −2t2 − (1 − t2 ) = −(1 + t2 )
dt
182 differential equations

Solution 1.1.4

y (0) = 1
y ′ (0) = 0
y′′ (0) = −1
y′′′ (0) = 0
y′′′′ (0) = 1
dk y
It is easy to guess the pattern here – dtk
(0) = 0 if k is odd and
dk y k
dtk
(0) = (−1) if k is even.
2

Solution 1.2.1
Solve the differential equation y′ = ty ln |y|
This equation is separable as
dy
= tdt
y ln |y|
Integrating both sides we get
t2
ln(ln(y)) = +A
2
and solving for y:
t2
y(t) = eCe 2
where we called C = e A

Solution 1.2.2
We use the equation we found for the vertical position of a falling
body:
gt2
y=− + vt + h
2
where v is the initial velocity and h is the initial height. Setting v =
10, h = 0, and g = −10, we get

y = −5t2 + 100t

Setting y = 0 (when the projectile is at height 0) and solving for t we

get two solutions: t = 0, which is the initial time, and y = 20, which
is when the projectile comes back to the ground.

Solution 1.2.3
The horizontal velocity is given to us as v x = 20 and it is unaffected
by any force (we are neglecting drag). Therefore the distance traveled
is just 20 ∗ 20 = 400 meters.
 problem solutions 183

Solution 1.2.4

a) y′ = cos(t) + 1, y(0) = 2. y(t) = sin(t) + t + 2

1
b) t2 y′ = 1, y(1) = 0. y(t) = 1 − t
1 ′ t
c) ty = e , y(0) = 0. y(t) = tet − et + 1
d) ty′ = 2, y(1) = 3. y(t) = ln(t2 ) + 3
e) tyy′ = 1, y(1) = 2. y(t) = ln(t2 ) + 4
p

f) y′ = y cos(t) + y, y(0) = 2. y(t) = 2esin(t)+t

2 +t
g) y′ = 2ty + y − 2t − 1, y(0) = 0. y(t) = 1 − et
y2 2
h) y′ = t +1 , y(0) = 2. y(t) = 1−2 ln(t+1)

Solution 1.4.1

y2
a) y′ + t2
= ty, y(0) = 1. No guaranteed solution.
y2
b) y′ + t2
= ty, y(3) = 1. Guaranteed unique solution.
y+t
c) y′ = y−t , y(1) = 2. Guaranteed unique solution.
y+t
d) y′ = y−t , y(2) = 2. No guaranteed solution.
y1/3
e) y′ = t2
, y(1) = 0. Guaranteed to exist but not guaranteed to
be unique.
ty
f) y′ + cos(y)
= 0, y(0) = 0. Guaranteed unique solution.

Solution 1.5.1
It is strongly recommended that one keep track of units here. In any
problem of this type the basic principle is that the rate of change
of any quantity is the rate at which the quantity enters the system
minus the rate at which it leaves.
Let W (t) denote the total amount of liquid in the tank. Liquid
enters at a rate of 10 L min−1 and leaves at a rate of 5 L min−1 . There
are initially 100 L of liquid, so W (t) satisfies the equation
dW
= 5 L min−1 W (0) = 100 L.
dt
Integrating up shows that W (t) = 100 L + 5 L min−1 · t
Next we need to write down the differential equation for the
amount of salt in the tank. The liquid enters the tank at 10 L min−1 ,
and the liquid contains 0.0001 kg L−1 . Thus the rate at which salt
enters the tank is

Rate in = 10 L min−1 · 0.0001 kg L−1 = 0.001 kg min−1 .

184 differential equations

Let S(t) denote the total amount of salt in the tank. The rate at
which salt leaves is slightly trickier. Liquid leaves at 5 L min−1 . The
salt content of this liquid (in kg L−1 ) will be the total amount of salt
(in kg) divided the total amount of liquid (in L), so that rate at which
salt leaves:

5 L min−1 S(t)
Rate out = 5 L min−1 · S(t)/W (t) =
100 L + 5 L min−1 · t
S(t)
=
20 min + t
Initially the total amount of salt in the tank is 10 kg, so we have
that the differential equation for the amount of salt is

dS S(t)
= 0.001 kg min−1 − S(0) = 10 kg
dt 20 min + t
Since this equation is first order linear it can be solved with an inte-
grating factor. Writing it as

dS S(t)
+ = 0.001 kg min−1 S(0) = 10 kg
dt 20 min + t
the integrating factor is
dt
= elog(20 min+t) = 20 min + t
R
µ(t) = e 20 min+t

Multiplying through by the integrating factor the equation becomes

d
((20 min + t)S(t)) = 0.001 kg min−1 · (20 min + t)
dt
= 0.02 kg + 0.001 kg min−1 · t.

Integrating this up gives

t2
((20 min + t)S(t)) = C + 0.02 kg · t + 0.001 kg min−1 ·
2
In order to solve for the constant C we substitute t = 0 to find

20 min · S(0) = 20 min · 10 kg = C

Finally we can solve for S(t), the salt content at time t

200 kg min + 0.02 kg · t + 0.0005 kg min−1 · t2

S(t) =
20 min + t
Graphing the function or doing a little calculus shows that the salt
content reaches an absolute minimum at around t ≈ 612.14 min,
when there is about 0.632 kg of salt in the tank, and about 3160.7 L of
liquid.
 problem solutions 185

Solution 1.5.2

dy 1
a) + 2y = 1 + t y(0) = 0 y(t) = 2t − e−2t + 1


dt 4
dy sin(t) cos(t)
b) dt + 2t y = t2
y(t) = A
t2
− t2
dy 1+cos(t)
c) t dt + y = sin(t) y(π ) = 0 y(t) = − t
dy 2t cos(t) 1+sin(t)
d) dt + t2 +1
y = t2 +1
y (0) = 1 y(t) = 1+ t2
dy sin(t) sin(t) 1
e) y y (0) = 0 y(t) = sec2 (t) − cos(t)


dt + cos(t)
= cos3 (t) 3

Solution 1.6.1

a) Exact: x2 + 5xy + x + y2 + 2y = 32
b) Not exact.
c) Exact: x3 − 3x2 y + 5xy2 − y3 + e x−y = −1 + e−1
d) Not exact.
e) Exact: x3 + 3y2 x − y3 = −5
x2 y2 1
f) Exact: 2xy − 2 + 2 = 2

g) Exact: x2 + y3 + e xy = 2

Solution 1.6.2
Find a general solution for each equation.
cos(t)
a) y′ = y2
. y3 = 3 sin(t) + C
2
b) yy′ = t + ty2 . y2 = Cet − 1
1 1
c) t2 y′ = cos(y)
. sin(y) = C − t

y2
d) t(y2 + 1)y′ = y. 2 + ln( yt ) = C

Solution 1.6.3
Find a general solution for each equation.

y2
a) yy′ = 2t + t . y2 = t2 ln(t4 ) + C

b) y′ = + yt . y2 + ty = t2 (ln(t) + C )
t
2( y + t )
y

c) ty′ = y + t y . cos t = C − ln(t)
sin( ) t
2t−y √
d) y′ = t+y . y = −t ± 3t2 + C
186 differential equations

Solution 1.6.4
Find a general solution for each equation.

a) y′ = ey+2t − 2. y = − ln(C − t) − 2t
b) y′ = (3y − t)2 + 13 . y = 1
C −9t + t
3

Solution 1.6.5
Find a general solution for each equation.
 
a) y′′ + 3y′ = te−3t . y = − 16 t2 − 19 t e−3t + Ae−3t + B
Rt 2
b) y′′ + 2ty′ = 0. y = A t e−s ds + B
0

c) ty′′ = y′ . y = At2 + B
d) y′′ = 2yy′ . y = A tan( At + B)
 At+b 
e) ey y′′ = y′ . y = ln 1+eA or y = A

f) yy′′ = (y′ )2 . y = Be At

Solution 1.6.6
Find a general solution for each equation.
1
a) y′ + y = yt . y2 = t − 2 + Ce−2t
4t3
b) ty′ = 3y + ty2 . y= C − t4
√ 2
t3

C
c) ty′ + 2y = t3 y. y= 8 + t
2y 1
d) y′ = ty3 − t . y2 = t2 +Ct4

Solution 1.6.7

1
a) Bernoulli equation. y = 2t2 −t3

b) Exact equation. xey + y2 cos( x ) + x2 y = 1

t3
c) Reducible equation. y = 12 − 1t + 2
1
d) Separable equation. y = 1−ln(t+1)
Ry s2
e) Reducible equation. 0 e− 2 ds = t
√
f) Scale invariant equation. y = −t − 1 + 3t2
 2  2
1
g) Bernoulli equation. y = 25 t2 + t93 or y = 1
25 t2 − 11
t3
1
 
h) Separable equation. y = 3 e1− t − 1

i) Exact equation. x3 y2 − x2 + xy3 = 1

 problem solutions 187

Solution 1.7.1

dy
a) dt = 6y − 2y2 . The equilibria are y = 0 (unstable) and y = 3
(stable).
dy
b) dt = y2 − 4y. The equilibria are y = 0 (stable) and y = 4
(unstable).
dy
c) dt = y2 − y3 . The equilibria are y = 0 (semi-stable) and y = 1
(stable)

Solution 2.0.1
Without actually solving the problems, determine in which interval I
we are guaranteed a unique solution for each initial value problem.
t2 ′
a) y′′′ − t −2 y+ cos (t)
t +3 y =
sin (t)
t2
, y(−1) = 2, y′ (−1) = 3, y′′ (−1) =
4; I = (−3, 0)
t2 ′
b) y′′′ − t −2 y+ cos (t)
t +3 y =
sin (t)
t2
, y(1) = 2, y′ (1) = 3, y′′ (1) =
4; I = (0, 2)
t2 ′
c) y′′′ − t −2 y+ cos (t)
t +3 y =
sin (t)
t2
, y(3) = 2, y′ (3) = 3, y′′ (3) =
4; I = (2, ∞)
t +1
d) (t − 2)y′′ + sin (t)
y = et , y(1) = 0, y′ (1) = 1; I = (0, 2)

Solution 2.1.1

a) e3t , te3t ; W = e6t , independent

b) e3t , 2e3t , te3t ; W = 0, dependent
c) e2t , e3t ; W= e5t , independent
d) t + 1, t2 , 3t − t2 ; W = −6, independent
e) t − t2 , 4t − 1, 4t2 − 1; W = 0, dependent

Solution 2.2.1

a) P(r ) = (1 − r2 ) = (1 − r )(1 + r ) so r = 1 has multiplicity 1

b) P(r ) = (1 − r )2 The polynomial is divisible by (r − 1)2 therefore
r = 1 has multiplicity 2
c) P(r ) = r3 − 3r2 + 4 = (r − 2)2 (r + 1) therefore r = 2 has
multiplicity 2
d) P(r ) = r3 − 3r2 + 4 = (r − 2)2 (r + 1) therefore r = 1 has
multiplicity 1
e) P(r ) = r5 + r3 = r3 (r2 + 1) therefore r = 0 has multiplicity 3
188 differential equations

Solution 2.2.2
Find general solutions to the following differential equations

a) y′′ + y′ − 6y = 0; y = c1 e2t + c2 e−3t

b) y′′′ − 6y′′ + 9y′ = 0; y = c1 + c2 e3t + c3 te3t
c) y′′′ + 4y′ = 0; y = c1 + c2 cos (2t) + c3 sin (2t)
d) y′′′ + 3y′′ + 3y′ + y = 0; y = c1 e−t + c2 te−t + c3 t2 e−t
e) y′′′′ + 2y′′ + y = 0; y = c1 cos (t) + c2 sin (t) + c3 t cos (t) +
c4 t sin (t)
f) y′′′ − 3y′′ + 4y = 0; y = c1 e−t + c2 e2t + c3 te2t
g) y′′′ − 5y′′ + y′ − 5y = 0; y = c1 e5t + c2 cos (t) + c3 sin (t)
h) y′′′′ − 8y′′′ + 16y′′ = 0; y = c1 + c2 t + c3 e4t + c4 te4t
√ √
i) y′′′ + 4y′′ + 6y′ = 0; y = c1 + c2 e−2t cos ( 2t) + c3 e−2t sin ( 2t)

Solution 2.5.2

a) y′′′ + 5y′ − 6y = e5t cos(4t) y p = Ae5t cos(4t) + Be5t sin(4t)

b) y′′ + 2y′ + y = cos(t) + t2 et y p = A cos(t) + B sin(t) + cet +
Dtet + Et2 et
c) y′′′ + 5y′ − 6y = et y p = Ae2t
d) y′′ + 3y′ + 2y = e−t y p = Ate−t since e−t solves the
homogeneous problem.

Solution 2.6.1
Find particular solutions to the following differential equations
dy et
a) dt + y = et ; yp = 2
dy 2t e2t 3 sin (t) cos (t)
b) dt + 3y = sin( t ) + e ; yp = 5 + 10 − 10
dy −t t
c) dt + y = t sin( t ) + e ; y p = te−t + 2 (sin (t) − cos (t)) +
cos (t)
2
d2 y −t
d) dt2
+ y = te−t ; y p = ( t + 1) e 2
d2 y
e) dt2
+ 2 dy
dt + 5y = e
−t + cos( t ); yp = e−t
4 + cos (t)
5 + sin (t)
10

Solution 2.6.2

d3
a) dt3
d2
b) dt2
+9
 problem solutions 189

2
d2

c) dt2
+4
 3  2 
d d
d) dt −1 dt 2 + 1

Solution 2.6.3

a) y = Aet + B A = 21 , B = b) y = A cos(t) + B sin(t)

1 A = − 12 , B = 12

c) y = At cos(t) + Bt sin(t) d) y = At3 + Bt2 + Ct + D

A = − 21 , B = 0 A = 1, B = 5, C = 17, D = 28

e) y = At5 + Bt4 + Ct3 + Dt2 f) y = At3 e−t + Bt4 + Ct3 +

A = − 21 , B = 0, C = 0, D = Dt2 A = − 21 , B = 1, C =
− 160
3 −11, D = 67

Solution 2.6.4

a) y p (t) = −t2 − 2t + 2 b) y p (t) = −t2 + 3t + 2

c) y p (t) = 3t2 − 2t d) y p (t) = − cos(3t) − 3 sin(3t)
e) y p (t) = sin(t) f) y p (t) = − cos(t) + sin(t)
g) y p (t) = 3t cos(t) + 2t sin(t) h) y p (t) = t cos(2t) + 3t sin(2t)
i) y p (t) = t2 e2t + te−t j) y p (t) = t2 et + 3te−2t

Solution 2.7.1
Use variation of parameters to find a particular solution for each
equation. If the particular solution can also be found using undeter-
mined coefficients, use both methods to check your answer.
t
a) y′′ + y′ − 6y = et ; y p = − e4 , we can use undetermined
coeff.s

b) y′′ + 1t y′ − 9
t2
y = 7t2 ; y p = t4

c) y′′ + y′ − 2y = 10 cos (t); y p = sin (t) − 3 cos (t), we can use

undetermined coeff.s
1
d) y′′ + y = cos (t)
; y p = cos (t) ln (cos (t)) + t sin (t)

e) y′′ − 6y′ + 9y = 2e3t ; y p = t2 e3t , we can use undetermined

coeff.s

f) y′′′ − y′′ − 2y′ = 4t; y p = t − t2 , we can use undetermined

coeff.s
190 differential equations

Solution 2.8.1

2 6
a) Y (s) = s +1 + s ( s +1)
+ (s+1)(1s2 +1) . The inverse transform is
1 1 7 −t
y(t) = 2 sin( t ) − 2 cos( t ) − 2 e + 6.
3s+2
b) Y (s) = s2 +1
+ (s−2)(11+s2 ) + s2 (11+s2 ) The inverse transform is
y(t) = t + 15 e2t + 14 3
5 cos( t ) + 5 sin( t ).
1
c) Y (s) = (s4 +1)(s−3)2
. The inverse transform is a mess.

We have not chosen to simplify Y (s) in any way. Your answers may
look slightly different if you chose to put on a common denominator,
apply partial fractions, or otherwise manipulate the result.

Solution 2.8.2
Problems a) – e) can be done by looking them up in a table, possibly
after having done some elementary simplifications. Problems f) and
g) require the use of partial fractions in the given form to express
Y (s) in a form that can be looked up in the table.

a) y(t) = et
b) y(t) = cos(t)
1 4 −3t
c) y(t) = 24 t e
d) y(t) = 5 cos(2t) + 27 sin(2t)
e) y(t) = 4e2t cos(t) + 11e2t sin(t)
f) Y (s) = A B
s−1 + s+1 giving y ( t ) = 27 et − 32 e−t
A B C
g) Y (s) = s + s−1 + s+2 giving y(t) = 2 − 3et + 5e−2t

Solution 2.8.3
Solve each initial value problem using the Laplace Transform method

a) y′′ − 4y′ + 3y = e2t , y(0) = 1, y′ (0) = 2; y(t) = et + e3t − e2t

b) y′′ + 4y′ + 3y = 9t, y(0) = 0, y′ (0) = −5; y(t) = 2e−t + 2e−3t +
3t − 4

Solution 3.1.6
Find the homogeneous solutions

a) y′′ + 4y = 0 y = A cos(2t) + B sin(2t)

b) y′′ +y = 0 y = A cos(t) + B sin(t)
c) y′′ + 6y′ + 10y = 0 y = Ae−3t cos(t) + Be−3t sin(t)
d) y′′ + 5y′ + 6y = 0 y = Ae−3t + Be−2t
 problem solutions 191

Solution 3.1.7
Solve the following initial value problems

a) y′′ + 2y′ + y = 0 y (0) = 0 y ′ (0) = 1 y = te−t

b) y′′ + 3y′ + 2y = 0 y (0) = 1 y ′ (0) = 0 y = 2e−t − e−2t
c) y′′ + 6y′ + 10y = 0 y (0) = 0 y ′ (0) = 1 y = 4e−3t sin(t) +
e−3t cos(t)
t
 √ √ √ 
d) y′′ + y′ + 2y = 0 y (0) = 1 y ′ (0) = 1 y = e− 2 cos( 27t ) + 3 7
7 sin( 7t
2 )

Solution 3.1.8
Solve the initial value problems
1
a) y′′ + 4y = cos(t) y (0) = 0 y ′ (0) = 0 y= 3 (cos(t) − cos(2t))
1
b) y′′ + y = cos(2t) y (0) = 1 y ′ (0) = 0 y= (4 cos(t) − cos(2t))
3
√ √ 
c) y′′ + 5y = cos(t) y(0) = 0 y ′ (0) = 1 y = 55 sin 5t +
√ 
1 1
4 cos ( t ) − 4 cos 5t
1
√
d) y′′ + 6y = cos(3t) y(0) = 0 y ′ (0) = 1 y= 3 cos( 6t) −
1
√ √
3 cos(3t ) + 6 sin( 6t )

Solution 3.1.9
Solve the following initial value problems
1
a) y′′ + 2y′ + y = cos(t) y (0) = 0 y ′ (0) = 0 y= 2 sin(t) −
1 −t
2 te
5 −t
b) y′′ + 3y′ + 2y = sin(t) y(0) = 1 y ′ (0) = 0 y = 2e −
6 −2t 1 3
5e + 10 sin(t) − 10 cos(t)
c) y′′ + 6y′ + 10y = sin(t) + cos(t) y(0) = 0 y′ (0) = 0 y=
1 −3t cos( t ) + 8e−3t sin( t )
39 cos( t ) + 5 sin( t ) − e

d) y′′ + y′ + 2y = 4 sin(t) y(0) = 1 y′ (0) = 1 y =

  √  √  √ 
1 − t/2 7t 7t
14 35 sin ( t ) − 21 cos ( t ) + e 35 cos 2 − 7 sin 2

Solution 3.1.10
Find the general solution to the homogeneous damped harmonic
oscillator equation
d2 y dy
m 2 +γ + ky = 0
dt dt
for the following parameter values. In each case classify the equation
as overdamped, underdamped or critically damped.
192 differential equations

a) m = 20.0 kg, γ = 40.0 Ns/m, k = 100.0 N/m

γ2 − 4km = −6400 kg2 /s2 underdamped
y = A exp(−1.0 s−1 t) cos(2.0 s−1 t) + B exp(−1 s−1 t) sin(2 s−1 t)
b) m = 25.0 kg, γ = 50.0 Ns/m, k = 25.0 N/m
γ2 − 4km = 0.0 kg2 /s2 critically damped
y = A exp(−1.0 s−1 t) + Bt exp(−1.0 s−1 t)
c) m = 10.0 kg, γ = 50.0 Ns/m, k = 60.0 N/m
γ2 − 4km = 100.0 kg2 /s2 overdamped
y = A exp(−2.0 s−1 t) + B exp(−3.0 s−1 t)
d) m = 10.0 kg, γ = 10.0 Ns/m, k = 30.0 N/m
γ2 − 4km = −1100.0 kg2 /s2 underdamped
y = A exp(−0.5 s−1 t) cos(1.66 s−1 t) + B exp(−0.5 s−1 t) sin(1.66 s−1 t)

Solution 3.1.11
Solve the following initial value problem
d2 y dy
m 2
+ γ + ky = 0
dt dt
for the following sets of initial values and parameters.
a) m = 20.0 kg, γ = 40.0 Ns/m, k = 100.0 N/m y (0) =
0 m, y′ (0) = 5 m/s
y = 2.5 m exp(−1.0 s−1 t) sin(2.0 s−1 t)
b) m = 25.0 kg, γ = 50.0 Ns/m, k = 25.0 N/m y (0) =
′
2 m, y (0) = 0 m/s
y = 2.0 m exp(−1.0 s−1 t) + 2.0 m/s t exp(−1.0 s−1 t)
c) m = 10.0 kg, γ = 50.0 Ns/m, k = 60.0 N/m y (0) =
1 m, y′ (0) = 1 m/s
y = 4.0 m exp(−2.0 s−1 t) − 3.0 m exp(−3.0 s−1 t)
d) m = 10.0 kg, γ = 10.0 Ns/m, k = 30.0 N/m y (0) =
2 m, y′ (0) = −1 m/s
y = 2.0 m exp(−0.5 s−1 t) cos(1.66 s−1 t)

Solution 3.1.12
dy
The quantity γ dt should have units of force, since this represents
force
the drag force. This means γ has units of velocity , or Ns/m. This is
√
equivalent to kg/s. The critical damping satisfies γ = 4km or
γ ≈ 15 492 Ns/m.

Solution 3.1.13
First we have to compute the spring constant k of the suspension,
then use this to compute the critical damping. The extra force on the
 problem solutions 193

suspension due to the load is 600 kg × 9.8 m/s2 = 5880 N. The spring
F
constant is k = − ∆x = − 5880 N
−6 cm = 98 000 N/m. Finally since we know
that k = 98 000 N/m and m = 2000 kg we have that γ = 28 000 Ns/m

Solution 3.1.14
The easiest way to solve this is to solve the complexified equation

4y′′ + y′ + 4y = eiωt

The solution is given by

eiωt
y=
4 − 4ω 2 + iω
The amplitude is the absolute value of the complex number A =
1
4−4ω 2 +iω
so we can maximize the amplitude by minimizing the ab-
solute value of the denominator. So we want to minimize |4 − 4ω 2 +
iω |2 = ω 2 + (4 − 4ω 2 )2 = f (ω ). This is a pretty straightforward
calculus problem: we have that f ′ (√ω ) = 64ω 3 − 62ω. The three crit-
ical points are ω = 0 and ω = ± 862 ≈ 0.984. The first is a local
minimum and the other two are local maxima. At the local maxima
63
the amplitude of the particular solution is | A| = 64

Solution 3.1.15
If y(t) = cos(t + s) then we have

y(t) = cos(t + s)
dy
dt = − sin(t + s)
d2 y
dt2
= − cos(t + s)
d2 y
Adding the first and the third equation shows that dt2 + y = 0. As
far as initial conditions we have that y(0) = cos(0 + s) = cos(s) and
y′ (0) = − sin(0 + s) = − sin(s). So cos(t + s) satisfies

d2 y
+y = 0 y(0) = cos(s) y′ (0) = − sin(s)
dt2
However we can also solve this equation by taking the general so-
lution y = A cos(t) + B sin(t) and solving for A and B. This gives
y = cos(t) cos(s) − sin(t) sin(s). Since solutions are unique we must
have cos(t) cos(s) − sin(t) sin(s) = cos(t + s).

Solution 3.1.16
Firstly note that if y = Aeiωt then

my′′ + γy′ + ky = (k − mω 2 + iγω ) Aeiωt

194 differential equations

1
so that A = k −mω 2 +iγω
. The squared magnitude of the denominator
is |z|2 = a + b = (k − mω 2 )2 + γ2 ω 2 = f (ω ). Differentiating
2 2

this gives f ′ (ω ) = ω (2m2 ω 2 + γ2 − 2km). This has one real root if

γ2 − 2km > 0 and three real roots if γ2 − 2km < 0.

Solution 3.1.17
The smaller the damping coefficient γ the higher the resonance peak,
so graph [C] corresponds to the smallest damping γ = 1, [B] cor-
responds to damping k = 2 and [A] corresponds to the strongest
damping γ = 4. Note that two of these cases have γ2 < 2km so they
will have three critical points, while one has γ2 > 2km so it has one
critical point.

Solution 4.5.1

a) λ2 − 6λ + 5; eigenvalues b) λ2 − 7λ + 10; eigenvalues

λ = 1, 5 λ = 2, 5
c) λ2 + 1; eigenvalues λ = d) λ2 − 4λ + 40; eigenvalues
i, λ = −i λ = 2 + 6i, λ = 2 − 6i
e) −λ3 + 7λ2 − 7λ − 15; eigen- f) −λ3 − 5λ2 − 6λ; eigenvalues
values λ = 5, 3, −1 λ = −3, −2, 0
g) −λ3 + 5λ2 + 2λ − 24; eigen-
values λ = 4, 3, −2

Solution 4.5.2
These are one possible set of solutions: recall that any (non-zero)
multiple of an eigenvector is an eigenvector.
! !
1 1
a) v1 = ; v2 =
1 −1
! !
1 4
b) v1 = ; v2 =
0 −3
! !
1 1
c) v1 = ; v2 =
i −i
! !
2 2
d) v1 = ; v2 =
3i −3i
     
3 1 1
e) v1 = 2 ; v2 = 0; v3 = −4
     
2 0 4
 problem solutions 195

     
1 1 1
f) v1 =  1  ; v2 = −1; v3 = 1
     
−2 1 1
     
1 1 11
g) v1 = 1 ; v2 = 0; v3 =  5 
     
0 0 −30

Solution 4.5.3

Parts a) and b) can be done either by finding the eigenvalues and

eigenvectors and constructing two linearly independent solutions
or by construcing the matrix exponential. Part c) requires one to
construct the matrix exponential using Putzer’s method, since this
is the only way that we know to solve a linear system with a deficit
matrix.

4e6t − e2t
!
a) v =
4e6t + e2t

4e2t − 2et
!
b) v =
2e2t
!
6tet + 2et
c) v =
3et

Solution 4.5.4

a)
et e5t e5t et
!
2 + 2 2 − 2
e5t et et e5t
2 − 2 2 + 2

b)
4 2t
e5t e3t − 1

!
3e
0 e2t

c)
!
cos(t) sin(t)
− sin(t) cos(t)
196 differential equations

d)
2 2t
e2t cos(6t)
!
3 e sin(6t )
− 23 e2t sin(6t) e2t cos(6t)

e)
−t 5e3t 3e5t e−t 7e3t 3e5t
e3t − e8 −
 
8 + 4 8 − 8 + 4
e−t e5t e5t e−t
 0 2 + 2 2 − 2
 

e 5t e−t e − t e5t
0 2 − 2 2 + 2

f)
e−3t e−2t 1 1 e−2t 1 e−3t
 
3 + 2 + 6 2 − 2 3 − 3
e − 3t e − 2t 1 e−2t 1 1 e−3t
3 − 2 + 6 2 + 2 3 −
 
 3 
− 3t − 2t − 2t 2e−3t
− 2e3 + e 2 + 16 1
2 − 2
e
3 + 31

g)
1 −2t
e3t e3t et − 1 6e5t + 5e6t − 11


30 e


1 −2t 6t
 0 e4t 6e e −1



0 0 e−2t

Solution 4.5.5
Both matrices have the same characteristic equation, (2 − λ)3 = 0,
so λ = 2 is an eigenvalue of multiplicity three in each case. The first
matrix has only one linearly independent eigenvector,
 
1
v1 = 0
 
0

Any non-zero multiple of this vector also works, of course. The sec-
ond matrix has two linearly indendent eigenvectors,
 
1
v1 = 0
 
0

and 
0
v2 =  1 
 
−1
(once more up to a non-zero multiplicative constant).
 problem solutions 197

Solution 4.5.6
For the first matrix we find that
 
1 0 0
B0 = I = 0 1 0 r1 (t) = e2t
 
0 0 1
 
0 1 1
B1 = A − 2I = 0 0 4 r2 (t) = te2t
 
0 0 0
 
0 0 4
1
B2 = ( A − 2I )( A − 2I ) = 0 0 0 r2 (t) = t2 e2t
 
2
0 0 0
which gives the matrix exponential
e2t te2t (2t2 + t)e2t
 

 0 e2t 4te2t .
 
0 0 e 2t

The second matrix works out a bit differently. We find that

 
1 0 0
B0 = I = 0 1 0 r1 (t) = e2t
 
0 0 1
 
0 1 1
B1 = A − 2I = 0 0 0 r2 (t) = te2t
 
0 0 0
 
0 0 0
1
B2 = ( A − 2I )( A − 2I ) = 0 0 0 r3 (t) = t2 e2t .
 
2
0 0 0
Note that we did not really need to compute r3 (t), as it will be multi-
plied by B2 , which is the zero matrix. This gives the matrix exponen-
tial as
e2t te2t te2t
 

 0 e2t 0 .
 
0 0 e2t
The difference between these two examples – the fact that the second
example has B2 = 0 and subsequently has no t2 e2t terms – is con-
nected to some ideas from linear algebra, the Jordan normal form
and the minimal polynomial of a matrix. These topics are usually
covered in a second undergraduate linear algebra text.

Solution 5.2.1
Determine if the following boundary value problems have no solu-
tions, one solution, or infinite solutions. If there are solutions, find
198 differential equations

them.

a) y′′ + y = 0 y (0) = 1 y(π ) = 0; No solutions.

b) y′′ + y = 0 y (0) = 1 y( π2 ) = 0; One solution: y( x ) =
cos( x )

c) y′′ + y = 0 y(0) = 1 y′ (π ) = 1; One solution:

y( x ) = cos( x ) − sin( x )

d) y′′ + y = 2 y(0) = 0 y(π ) = 4 Infinite solutions:

y( x ) = −2 cos( x ) + B sin( x ) + 2

e) y′′ + y = 2 y (0) = 1 y(π ) = 0; No solutions.

f) y′′ + 4y = 0 y(0) = 3 y(π ) = 3 Infinite solutions:
y( x ) = 3 cos(2x ) + B sin(2x )

Solution 5.2.2

Determine the values of A for which the following boundary value

problems have solutions.

a) y′′ + y = x y (0) = 2 y(π ) = A; A = π−2

b) y′′ + 4y = 0 y ′ (0) = A y′ (π ) = 1; A=1
c) y′′ + 9y = 9x y (0) = 0 y(π ) = A; A=π

Solution 5.3.1

Calculate all of the eigenvalues and the corresponding eigenfunctions

for the following boundary value problems.

a) y′′ + λy = 0 y ′ (0) = 0 y′ (5) = 0; λ0 = 0, y0 ( x ) =

n2 π 2
1; λn = 25 ,
yn ( x ) = cos( nπx
5 ).
In this case λ = 0 is an
eigenvalue and the corresponding eigenfunction is a constant
which we pick to be 1 for simplicity.

b) y′′ + λy = 0 y′ (0) = 0 y(3) = 0; λn = (n +

π2
 
1 2
2) 9 , yn ( x ) = cos (n + 21 ) πx
3

c) y′′ + λy = 0 y(−π ) = 0 λ n = n2 , y n ( x ) =
y(π ) = 0;
 
sin(nx ); and also λm = (m + 12 )2 , ym ( x ) = cos (m + 12 ) x

d) y′′ + 2y′ + λy = 0 y(0) = 0 y(π ) = 0; λn = 1 +

n2 , yn ( x ) = e− x sin(nx )

e) y′′ − 2y′ + λy = 0 y ′ (0) = 0 y′ ( L) = 0; λn = 1 +

n2 π 2
, yn ( x ) = ex sin( nπx nπ
cos( nπx


L2 L )− L L )
 problem solutions 199

Solution 6.5.1

a) 42
b)
∞ ∞
8 1 kπx 8 1 (2k − 1)πx
   
1− 2
π ∑ k2
cos
2
= 1− 2
π ∑ (2k − 1)2 cos 2
k =1, odd k =1

c)
12 ∞ (−1)k kπx
 

π k∑
5+ sin
=1
k 3

d) 10 cos(2x )
e)
∞ ∞ ∞ ∞
3 2 1 1 1 3 2 1 1 1
−
4 π2 ∑ k2
cos(kπx ) +
π ∑ k
sin(kπx ) = − 2
4 π ∑ ( 2k − 1 ) 2
cos((2k − 1)πx ) +
π ∑ k
sin(kπx )
k =1,odd k =1 k =1 k =1

f)
25 100 ∞ (−1)k kπx
 
+ 2 ∑ cos
3 π k =1 k 2 5

Solution 6.5.2

a) Sine:
168 ∞ 1
π k=∑
sin(kx )
1,odd
k

Cosine:
42

b) Sine:
∞
16 (−1)k (2k − 1)πx
 
− 2
π ∑ (2k − 1)2 sin 4
k =1
Cosine:
∞ ∞
32 1 (4k − 2)πx 8 1 (2k − 1)πx
   
1− 2
π ∑ (4k − 2)2 cos 4
= 1− 2
π ∑ (2k − 1)2 cos 2
k =1 k =1
200 differential equations

c) Sine:
∞
2 (−1)k + 5 kπx
 

π ∑ k
sin
3
k =1
Cosine:
∞
24 1 (2k − 1)πx
 
2+ 2
π ∑ (2k − 1)2 cos 3
k =1

d) Sine:
40 ∞ 2k − 1
π k∑
sin ((2k − 1) x )
=1
( 2k − 1)2 − 4
Cosine:
10 cos(2x )

e) Sine:
∞
2(−1)k
" #
4 kπ kπx
   
∑ k π2
2
sin
2
−
kπ
sin
2
k =1

Cosine:
∞
3 4 kπ kπx
     

4 k∑
+ cos − 1 cos
=1
k2 π 2 2 2

f) Sine:

∞
50(−1)k
" #
100  kπx
  
k
∑ −
kπ
+ 3 3 (−1) − 1 sin
k π 5
k =1

Cosine:
25 100 ∞ (−1)k kπx
 
+ 2 ∑ cos
3 π k =1 k 2 5

Solution 7.1.1

a)
∞
6 (−1)n − 5n2 π2 t  nπx 
v( x, t) = −
π ∑ n e 9 sin
3
n =1
 problem solutions 201

b)
∞
5 10 1 − (−1)n − 2n2 π2 t  nπx 
v( x, t) = + 2
2 π ∑ n2
e 25 cos
5
n =1

c)
40 ∞ (−1)n − 3(2n−1)2 t (2n − 1) x
 

π n∑
v( x, t) = − e 4 cos
=1 2n − 1 2

Solution 7.2.1

a)
∞
4 (−1)n 3nπt  nπx 
 
v( x, t) = −
π ∑ n cos 2
sin
2
n =1

b)

12 ∞ 1 − (−1)n 4nπt  nπx  5

  
v( x, t) = ∑
π n =1 n
cos
3
sin
3
+
4π
sin(4πt) sin(πx )