Differential Equations Notes

Lectures on Differential Equations
Philip Korman
Department of Mathematical Sciences
University of Cincinnati
Cincinnati Ohio 45221-0025
Contents
Introduction
vii
1 First Order Equations

1.1 Integration by Guess-and-Check . . . . .
1.2 First Order Linear Equations . . . . . .
1.2.1 The Integrating Factor . . . . . .
1.3 Separable Equations . . . . . . . . . . .
1.3.1 Problems . . . . . . . . . . . . .
1.4 Some Special Equations . . . . . . . . .
1.4.1 Homogeneous Equations . . . . .
1.4.2 The Logistic Population Model .
1.4.3 Bernoullis Equation . . . . . . .
1.4.4 Riccatis Equation . . . . . . . .
1.4.5 Parametric Integration . . . . . .
1.4.6 Some Applications . . . . . . . .
1.5 Exact Equations . . . . . . . . . . . . .
1.6 Existence and Uniqueness of Solution . .
1.7 Numerical Solution by Eulers method .
1.7.1 Problems . . . . . . . . . . . . .
1.8 The Existence and Uniqueness Theorem
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
3
5
10
12
15
15
18
20
21
23
24
26
31
32
33
37
2 Second Order Equations

44
2.1 Special Second Order Equations . . . . . . . . . . . . . . . . . 44
2.1.1 y is not present in the equation . . . . . . . . . . . . . 45
2.1.2 x is not present in the equation . . . . . . . . . . . . . 46
2.1.3 The Trajectory of Pursuit . . . . . . . . . . . . . . . . 47
2.2 Linear Homogeneous Equations with Constant Coefficients . . 50
2.2.1 The Characteristic Equation Has Two Distinct Real
Roots . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
ii
CONTENTS
2.2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
2.10
2.11
2.12
2.13
2.14
2.15
The Characteristic Equation Has Only One (Repeated)

Real Root . . . . . . . . . . . . . . . . . . . . . . . . .
The Characteristic Equation Has Two Complex Conjugate
Roots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 Eulers Formula . . . . . . . . . . . . . . . . . . . . .
2.3.2 The General Solution . . . . . . . . . . . . . . . . . .
2.3.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . .
Linear Second Order Equations with Variable Coefficients . .
Some Applications of the Theory . . . . . . . . . . . . . . . .
2.5.1 The Hyperbolic Sine and Cosine Functions . . . . . .
2.5.2 Different Ways to Write the General Solution . . . . .
2.5.3 Finding the Second Solution . . . . . . . . . . . . . . .
2.5.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . .
Non-homogeneous Equations. . . . . . . . . . . . . . . . . .
Non-homogeneous Equations. More on Guessing of Y (t) . . .
The Method of Variation of Parameters . . . . . . . . . . . .
The Convolution Integral . . . . . . . . . . . . . . . . . . . .
2.9.1 Differentiation of Integrals . . . . . . . . . . . . . . . .
2.9.2 Yet Another Way to Compute a Particular Solution .
Applications of Second Order Equations . . . . . . . . . . . .
2.10.1 Vibrating Spring . . . . . . . . . . . . . . . . . . . . .
2.10.2 Problems . . . . . . . . . . . . . . . . . . . . . . . . .
2.10.3 A Meteor Approaching the Earth . . . . . . . . . . . .
2.10.4 Damped Oscillations . . . . . . . . . . . . . . . . . . .
Further Applications . . . . . . . . . . . . . . . . . . . . . . .
2.11.1 Forced and Damped Oscillations . . . . . . . . . . . .
2.11.2 An Example of a Discontinuous Forcing Term . . . . .
2.11.3 Oscillations of a Pendulum . . . . . . . . . . . . . . .
2.11.4 Sympathetic Oscillations . . . . . . . . . . . . . . . . .
Oscillations of a Spring Subject to a Periodic Force . . . . . .
2.12.1 Fourier Series . . . . . . . . . . . . . . . . . . . . . . .
2.12.2 Vibrations of a Spring Subject to a Periodic Force . .
Eulers Equation . . . . . . . . . . . . . . . . . . . . . . . . .
Linear Equations of Order Higher Than Two . . . . . . . . .
2.14.1 The Polar Form of Complex Numbers . . . . . . . . .
2.14.2 Linear Homogeneous Equations . . . . . . . . . . . . .
2.14.3 Non-Homogeneous Equations . . . . . . . . . . . . . .
2.14.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . .
Oscillation and Comparison Theorems . . . . . . . . . . . . .
iii
53
54
54
55
57
59
63
64
64
66
67
68
72
74
77
77
77
78
78
81
84
86
87
87
89
90
91
92
92
96
97
101
101
102
104
105
108
iv
CONTENTS
3 Using Infinite Series to Solve Differential Equations

3.1 Series Solution Near a Regular Point . . . . . . . . . . . . .
3.1.1 Maclauren and Taylor Series . . . . . . . . . . . . .
3.1.2 A Toy Problem . . . . . . . . . . . . . . . . . . . . .
3.1.3 Using Series When Other Methods Fail . . . . . . .
3.2 Solution Near a Mildly Singular Point . . . . . . . . . . . .
3.2.1 Derivation of J0 (x) by Differentiation of the Equation
3.3 Moderately Singular Equations . . . . . . . . . . . . . . . .
3.3.1 Problems . . . . . . . . . . . . . . . . . . . . . . . .
115
. 115
. 115
. 117
. 118
. 123
128
. 129
. 135
4 The Laplace Transform

4.1 The Laplace Transform And Its Inverse
4.1.1 Review of Improper Integrals . .
4.1.2 The Laplace Transform . . . . .
4.1.3 The Inverse Laplace Transform .
4.2 Solving The Initial Value Problems . . .
4.2.1 Step Functions . . . . . . . . . .
4.3 The Delta Function and Impulse Forces
4.4 Convolution and the Tautochrone Curve
4.5 Distributions . . . . . . . . . . . . . . .
4.5.1 Problems . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
139
139
139
140
142
143
146
148
151
156
158
5 Linear Systems of Differential Equations

164
5.1 The Case of Distinct Eigenvalues . . . . . . . . . . . . . . . . 164
5.1.1 Review of Vectors and Matrices . . . . . . . . . . . . . 164
5.1.2 Linear First Order Systems with Constant Coefficients 166
5.2 A Pair of Complex Conjugate Eigenvalues . . . . . . . . . . . 170
5.2.1 Complex Valued Functions . . . . . . . . . . . . . . . 170
5.2.2 The General Solution . . . . . . . . . . . . . . . . . . 171
5.3 The Exponential of a Matrix . . . . . . . . . . . . . . . . . . 172
5.3.1 Solutions of Planar Systems Near the Origin . . . . . 177
5.3.2 Problems . . . . . . . . . . . . . . . . . . . . . . . . . 179
5.4 Floquet Theory and Masseras Theorem . . . . . . . . . . . . 184
6 Non-Linear Systems
6.1 The Predator-Prey Interaction .
6.2 Competing Species . . . . . . . .
6.3 An Application to Epidemiology
6.4 Lyapunovs Stability . . . . . . .
6.4.1 Problems . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
190
190
194
199
201
207
CONTENTS
6.5 Periodic Population Models . . . . . . . . . . . . . . . . . . . 210

7 The Fourier Series and Boundary Value Problems
219
7.1 Fourier Series for Functions of an Arbitrary Period . . . . . . 219
7.1.1 Even and Odd Functions . . . . . . . . . . . . . . . . 221
7.1.2 Further Examples and the Convergence Theorem . . . 223
7.2 Fourier Cosine and Fourier Sine Series . . . . . . . . . . . . . 226
7.3 Two Point Boundary Value Problems . . . . . . . . . . . . . 229
7.3.1 Problems . . . . . . . . . . . . . . . . . . . . . . . . . 230
7.4 The Heat Equation and the Method of Separation of Variables 234
7.5 Laplaces Equation . . . . . . . . . . . . . . . . . . . . . . . 242
7.6 Wave Equation . . . . . . . . . . . . . . . . . . . . . . . . . 246
7.6.1 Non-Homogeneous Problems . . . . . . . . . . . . . . 250
7.6.2 Problems . . . . . . . . . . . . . . . . . . . . . . . . . 252
7.7 Calculating Earths Temperature . . . . . . . . . . . . . . . . 256
7.7.1 The Complex Form of the Fourier Series . . . . . . . . 256
7.7.2 Temperatures Inside the Earth and Wine Cellars . . . 258
7.8 Laplaces Equation on Circular Domains . . . . . . . . . . . . 260
7.9 Sturm-Liouville Problems . . . . . . . . . . . . . . . . . . . . 264
7.9.1 Fourier-Bessel Series . . . . . . . . . . . . . . . . . . . 268
7.9.2 Cooling of a Cylindrical Tank . . . . . . . . . . . . . . 269
7.10 Greens Function . . . . . . . . . . . . . . . . . . . . . . . . . 271
7.10.1 Problems . . . . . . . . . . . . . . . . . . . . . . . . . 274
7.11 The Fourier Transform . . . . . . . . . . . . . . . . . . . . . . 278
7.12 Problems on Infinite Domains . . . . . . . . . . . . . . . . . . 280
7.12.1 Evaluation of Some Integrals . . . . . . . . . . . . . . 280
7.12.2 The Heat Equation for < x < . . . . . . . . . 281
7.12.3 Steady State Temperatures for the Upper Half Plane . 282
7.12.4 Using the Laplace Transform for a Semi-Infinite String 283
7.12.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . 284
8 Elementary Theory of PDE
8.1 Wave Equation: Vibrations of an Infinite String
8.2 Semi-Infinite String: Reflection of Waves . . . .
8.3 Bounded String: Multiple Reflections . . . . . .
8.4 Neumann Boundary Conditions . . . . . . . . .
8.5 Non-Homogeneous Wave Equation . . . . . . .
8.5.1 Problems . . . . . . . . . . . . . . . . .
8.6 First Order Linear Equations . . . . . . . . . .
8.6.1 Problems . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
286
286
291
294
297
299
302
305
310
vi
CONTENTS
8.7
8.8
8.9
8.10
Laplaces Equation: Poissons Integral Formula

Some Properties of Harmonic Functions . . . .
Maximum Principle . . . . . . . . . . . . . . .
Maximum Principle for Heat Equation . . . . .
8.10.1 Uniqueness on an Infinite Interval . . .
8.10.2 Problems . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
310
312
315
317
319
322
9 Numerical Computations
324
9.1 The Capabilities of Software Systems, Like Mathematica . . 324
9.2 Solving Boundary Value Problems . . . . . . . . . . . . . . . 326
9.3 Solving Nonlinear Boundary Value Problems . . . . . . . . . 330
9.4 Direction Fields . . . . . . . . . . . . . . . . . . . . . . . . . 332
10 Appendix
338
.1
The Chain Rule and Its Descendants . . . . . . . . . . . . . . 338
.2
Partial Fractions . . . . . . . . . . . . . . . . . . . . . . . . . 339
.3
Eigenvalues and Eigenvectors . . . . . . . . . . . . . . . . . . 340
Introduction
This book is based on several courses that I taught at the University of
Cincinnati. Chapters 1-4 are based on the course Differential Equations
for sophomores in science and engineering. Chapters 7 and 8 are based on
the course Fourier Series and PDE. While some theoretical material is
either quoted, or just mentioned without proof, I have tried to show all of
the details when doing problems. I have tried to use plain language, and
not to be too wordy. I think that an extra word of explanation has often as
much potential to confuse a student, as to be helpful. I have also tried not
to overwhelm students with new information. I forgot who said it first: one
should teach the truth, nothing but the truth, but not the whole truth.
I hope that the experts will find this book useful as well. It presents
several important topics that are hard to find in the literature: Masseras
theorem, Lyapunovs inequality, Picones form of Sturms comparison theorem, sideways heat equation, periodic population models, hands on
numerical solution of nonlinear boundary value problems, etc. The book
also contains new exposition of some standard topics. We have completely
revamped the presentation of the Frobenius method for series solution of
differential equations, so that the regular singular points are now hopefully in the past. In the proof of the existence and uniqueness theorem,
we have replaced the standard Picard iterations with monotone iterations,
which should be easier for students to absorb. There are many other fresh
touches throughout the book. The book contains a number of interesting
problems, including some original ones, published by the author over the
years in the Problem Sections of SIAM Review, EJDE, and other journals.
All of the challenging problems are provided with hints, making them easy
to solve for instructors.
How important are differential equations? Here is what Isaac Newton
said: It is useful to solve differential equations. And what he knew was
just the beginning. Today differential equations are used widely in science
vii
viii
INTRODUCTION
and engineering. This book presents many applications as well. Some of

these applications are very old, like the tautochrone problem considered by
Christian Huygens in 1659. Some applications, like when a drone is targeting
a car, are modern. Differential Equations is also a beautiful subject, which
lets students see Calculus in action.
I have tried to start each topic with simple examples, to keep the presentation focused, and to show the details. I think this book is suitable for
self-study. However, the instructor can help in many ways. He will present
the subject with the enthusiasm it deserves, draw pictures, talk about the
history, and his jokes will supplement the lame ones in the book.
It is a pleasure to thank Ken Meyer and Dieter Schmidt for constant
encouragement, while I was writing this book. I also wish to thank Ken
for reading the entire book, and making a number of useful suggestions,
like doing Fourier series early, with applications to periodic vibrations and
radio tuning. I wish to thank Roger Chalkley, Tomasz Adamowicz, Dieter
Schmidt, and Ning Zhong for a number of useful comments.
Chapter 1
First Order Equations

1.1
Integration by Guess-and-Check
Many problems in differential equations end with a computation of an integral. One even uses the term integration of a differential equation instead
of solution. We need to be able to compute integrals quickly.
Recall the product rule
(f g)0 = f g 0 + f 0 g .
Example
xex dx. We need to find the function, with the derivative equal
to xex . If we try a guess: xex , then its derivative

(xex )0 = xex + ex
has an extra term ex . To remove this extra term, we subtract ex from the
initial guess, so that
Z
xex dx = xex ex + c.
By differentiation, we verify that this is correct. Of course, integration by

parts may also be used.
Z
1
Example x cos 3x dx. Starting with the initial guess x sin 3x, with the
3
derivative equal to x cos 3x + 13 sin 3x, we compute
Z
1
1
x cos 3x dx = x sin 3x + cos 3x + c.
3
9
CHAPTER 1. FIRST ORDER EQUATIONS
We see that the initial guess is the product f (x)g(x), chosen in such a
way that f (x)g 0(x) gives the integrand.
Z
1
Example xe5x dx. Starting with the initial guess xe5x , we compute
5
Z
Example
1
1
xe5x dx = xe5x e5x + c.
5
25
1
x2 sin 3x dx. The initial guess is x2 cos 3x. Its derivative
3

1
x2 cos 3x
3
0
2
= x2 sin 3x x cos 3x
3
2
has an extra term x cos 3x. To remove this term, we modify our guess:
3
2
1
x2 cos 3x + x sin 3x. Its derivative
3
9
1
2
x2 cos 3x + x sin 3x
3
9
still has an extra term

Z
0
= x2 sin 3x +
2
sin 3x
9
2
sin 3x. So we make the final adjustment
9
1
2
2
x2 sin 3x dx = x2 cos 3x + x sin 3x +
cos 3x + c .
3
9
27
This is easier than integrating by parts twice.

Example
x2
+ 4 dx. We begin by rewriting this integral as
x x2 + 4
One usually computes this integral by a substitution u = x Z+ 4, with

du = 2x dx. Forgetting a constant multiple, the integral becomes u1/2 du.
Ignoring a constant multiple again, this evaluates to u3/2 . Returning to the

original variable, we have our initial guess x2 + 4
3/2
. Differentiation
3/2

1/2
d 2
x +4
= 3x x2 + 4
dx
gives us the integrand with an extra factor of 3. To fix that, we multiply

the initial guess by 13 :
Z
x x2 + 4 dx =
3/2
1 2
x +4
+ c.
3
1/2
dx.
1.2. FIRST ORDER LINEAR EQUATIONS
1
dx. Instead of using partial fractions, let us
+ 1)(x2 + 4)
try to split the integrand as
Example
(x2
1
1
.
x2 + 1 x2 + 4
We see that this is off by a factor. The correct formula is
1
1
=
2
2
(x + 1)(x + 4)
3
Then
(x2
1
1
2
2
x +1 x +4
1
1
x
1
dx = tan1 x tan1 + c .
2
+ 1)(x + 4)
3
6
2
Sometimes one can guess the splitting twice, as in the following case.
Example
x2
x2 (1
1
dx.
x2 )
1
1
1
1
1 1 1
1 1
1
= 2+
= 2+
= 2+
+
.
2
2
(1 x )
x
1x
x
(1 x)(1 + x)
x
21x 21+x
Then (for |x| < 1)

Z
1.2
x2
1
1 1
1
dx = ln(1 x) + ln(1 + x) + c .
2
(1 x )
x 2
2
First Order Linear Equations
Background
Suppose we need to find a function y(x) so that
y 0 (x) = x.
This is a differential equation, because it involves a derivative of the unknown
function. This is a first order equation, as it only involves the first derivative.
Solution is, of course,
x2
(2.1)
y(x) =
+ c,
2
where c is an arbitrary constant. We see that differential equations have

infinitely many solutions. The formula (2.1) gives us the general solution.
Then we can select the solution that satisfies an extra initial condition. For
example, for the problem
(2.2)
y 0 (x) = x
y(0) = 5
we begin with the general solution given in formula (2.1), and then evaluate
it at x = 0
y(0) = c = 5.
So that c = 5, and solution of the problem (2.2) is
y(x) =
x2
+ 5.
2
The problem (2.2) is an example of an initial value problem. If the variable

x represents time, then the value of y(x) at the initial time is prescribed to
be 5. The initial condition may be prescribed at other values of x, as in the
following example:
y0 = y
y(1) = 2e .
Here the initial condition is prescribed at x = 1, e denotes the Euler number
e 2.718. Observe that while y and y 0 are both functions of x, we do not
spell this out. This problem is still in the Calculus realm. Indeed, we
are looking for a function y(x), with the derivative equal to y(x). This is a
property of the function ex , and its constant multiples. The general solution
is
y(x) = cex,
and the initial condition gives
y(1) = ce = 2e,
so that c = 2. The solution is then
y(x) = 2ex .
We see that the main effort is in finding the general solution. Selecting
c, to satisfy the initial condition, is usually easy.
Recall from Calculus that

d g(x)
e
= eg(x) g 0(x).
dx
In case g(x) is an integral, we have
R
d R p(x) dx
e
= p(x)e p(x) dx ,
dx
(2.3)
because the derivative of the integral
1.2.1
The Integrating Factor
p(x) dx is p(x).
Let us find the general solution of the equation

(2.4)
y 0 + p(x)y = g(x),
where p(x) and g(x) are given functions. This is a linear equation, as we
have a linear function of y, and y 0 . Because we know p(x), we can calculate
the function
R
(x) = e p(x) dx ,
and its derivative

(2.5)
0 (x) = p(x)e
p(x) dx
= p(x).
We now multiply the equation (2.4) by (x), giving

(2.6)
y 0 + yp(x) = g(x).
Let us use the product rule and the formula (2.5) to calculate the derivative
d
[y] = y 0 + y0 = y 0 + yp(x).
dx
So that we may rewrite (2.6) in the form
(2.7)
d
[y] = g(x).
dx
This allows us to compute the general solution. Indeed,

we know the function
R
on the right. By integration, we express (x)y(x) = (x)g(x) dx, and then
solve for y(x).
In practice one needs to memorize the formula for (x), and the form
(2.7). When computing (x), we shall always take the constant of integration to be zero, c = 0, because the method works for any c.
Example Solve
y 0 + 2xy = x
y(0) = 2 .
Here p(x) = 2x, and g(x) = x. Compute
R
(x) = e
The equation (2.7) takes the form
2x dx
= ex .
d h x2 i
2
e y = xex .
dx
Integrate both sides, and then perform integration by a substitution u = x2
(or use guess-and-check)
x2
e y=
xex dx =
Solving for y, gives
1 x2
e + c.
2
y(x) =
1
2
+ cex .
2
y(0) =
1
+ c = 2,
2
From the initial condition
3
1 3
2
so that c = . Answer: y(x) = + ex .
2
2 2
Example Solve
1
y 0 + y = cos 2t , y(/2) = 1 .
t
Here the independent variable is t, y = y(t), but the method is, of course,
the same. Compute
R 1
(t) = e t dt = eln t = t,
and then
d
[ty] = t cos 2t.
dt
Integrate both sides, and perform integration by parts
ty =
t cos 2t dt =
1
1
t sin 2t + cos 2t + c.
2
4

Divide by t
y(t) =
1
1 cos 2t c
sin 2t +
+ .
2
4 t
t
The initial condition gives

y(/2) =
We solve for c
1 1
c
+
= 1.
4 /2 /2
1
c = /2 + ,
4
and so the solution is

(2.8)
1
1 cos 2t /2 + 14
y(t) = sin 2t +
+
.
2
4 t
t
This function y(t) gives us a curve, called the integral curve. The initial
condition tells us that y = 1 when t = /2, so that the point (/2, 1) lies
on the integral curve. What is the maximal interval on which the solution
(2.8) is valid? I.e., starting with the initial point t = /2, how far can we
continue the solution to the left and to the right of the initial point? We see
from (2.8) that the maximal interval is (0, ). At t = 0, the solution y(t)
is undefined.
Example Solve
x
dy
+ 2y = sin x ,
dx
y() = 2 .
Here the equation is not in the form (2.4), for which the theory applies. We
divide the equation by x
dy
2
sin x
+ y=
.
dx x
x
2
sin x
Now the equation is in the right form, with p(x) = and g(x) =
. As
x
x
before, we compute, using the properties of logarithms,
R
(x) = e
And then
2
x
dx
= e2 ln x = eln x = x2 .
d h 2 i
sin x
x y = x2
= x sin x.
dx
x
Integrate both sides, and perform integration by parts

x2 y =
x sin x dx = x cos x + sin x + c,
giving us the general solution

y(x) =
cos x sin x
c
+ 2 + 2.
x
x
x
The initial condition implies

y() =
c
1
+ 2 = 2.

Solve for c:
c = 2 2 + .
cos x sin x 2 2 +
+ 2 +
. This solution is valid on the
x
x
x2
interval (, 0) (that is how far it can be continued to the left and to the
right, starting from the initial point x = ).
Answer: y(x) =
Example Solve
dy
1
=
,
dx
yx
y(1) = 0 .
We have a problem: not only this equation is not in the right form, this
1
is not a linear function of y. We need a
is a nonlinear equation, because yx
little trick. Let us pretend that dy and dx are numbers, and take reciprocals
of both sides of the equation, getting
dx
= y x,
dy
or
dx
+ x = y.
dy
Let us now think of y as independent variable, and x as a function of y,

x = x(y). Then the last equation
is linear, with p(y) = 1 and g(y) = y. We
R
1 dy
proceed as usual: (y) = e
= ey , and
d y
[e x] = yey .
dy

y
3
1.0
0.5
1.5
2.0
2.5
3.0
-1
Figure 1.1: The integral curve x = y1+2ey , with the initial point marked
Integrating
ey x =
yey dy = yey ey + c,
or
x(y) = y 1 + cey .
To find c, we need an initial condition. The original initial condition tells
us that y = 0 for x = 1. For the inverse function x(y) this translates to
x(0) = 1. So that c = 2.
Answer: x(y) = y 1 + 2ey (see the Figure 1.1).
The rigorous justification of this method is based on the formula for
the derivative of an inverse function, that we recall next. Let y = y(x) be
some function, and y0 = y(x0 ). Let x = x(y) be its inverse function. Then
x0 = x(y0 ), and we have
dx
(y0 ) =
dy
1
dy
dx (x0 )
10
1.3
Separable Equations
Background
Suppose we have a function F (y), and y in turn depends on x, y = y(x). So
that, in effect, F depends on x. To differentiate F with respect to x, we use
the chain rule from Calculus:
dy
d
F (y(x)) = F 0 (y(x)) .
dx
dx
The Method
Suppose we are given two functions F (y) and G(x), and let us use the
corresponding lower case letters to denote their derivatives, so that F 0 (y) =
f (y) and G0 (x) = g(x). Our goal is to solve the following equation (to find
the general solution)
dy
f (y)
(3.1)
= g(x).
dx
This is a nonlinear equation.
We begin by rewriting this equation, using the upper case functions
F 0 (y)
dy
= G0 (x).
dx
Using the chain rule, we rewrite this as

d
d
F (y) =
G(x).
dx
dx
If derivatives of two functions are the same, these functions differ by a
constant, so that
(3.2)
F (y) = G(x) + c.
This is the desired general solution! If one is lucky, it may be possible to
solve this relation for y as a function of x. If not, maybe one can solve for
x as a function of y. If both attempts fail, one can use an implicit plotting
routine to draw the integral curves, given by (3.2).
We now describe a simple procedure, which leads from the equation (3.1)
dy
to its solution (3.2). Let us pretend that dx
is not a notation for a derivative,
but a ratio of two numbers dy and dx. Clearing the denominator in (3.1)
f (y) dy = g(x) dx.
11
1.3. SEPARABLE EQUATIONS
We have separated the variables, everything involving y is now on the left,

while x appears only on the right. Integrate both sides:
Z
f (y) dy =
g(x) dx,
which gives us immediately the solution (3.2).

Example Solve
dy
= x(1 + y 2 ).
dx
To separate the variables, we multiply by dx, and divide by (1 + y 2 )
Z
dy
dy =
1 + y2
x dx .
So that the general solution is

1
arctan y = x2 + c,
2
which we can also put into the form
1
y = tan( x2 + c).
2
Example Solve

xy 2 + x dx + ex dy = 0 .
This is an example of a differential equation, written in differentials. (Didy

xy 2 + x
viding through by ex dx, we can put it into a familiar form
=
,
dx
ex
although there is no need to do that.)
By factoring, we are able to separate the variables:
ex dy = x(y 2 + 1) dx ;
Z
dy
=
y2 + 1
xex dx ;
tan1 y = xex + ex + c .
Answer: y(x) = tan xex + ex + c .
Recall that by the fundamental theorem of Calculus,

d
dx
f (t) dt = f (x) ,
12
for any constant a. TheR integral ax Rf (t) dt gives us an antiderivative of f (x),

so that we may write f (x) dx = ax f (t) dt + c. Here we may let c be an
arbitrary constant, and a fixed, or the other way around.
R
Example Solve
dy
2
= ex y 2 ,
dx
Separation of variables
Z
y(1) = 2 .
dy
dy =
y2
ex dx
gives on the right an integral that cannot be evaluated in elementary functions. We shall change it to a definite integral, as above. We choose a = 1,
because the initial condition was given at x = 1:
Z
dy
dy =
y2
1
=
y
et dt + c ,
1
2
et dt + c .
When x = 1, we have y = 2, which means that c = 12 . Answer: y(x) =

R x t2
1
R x t2
1 . For any x, the integral 1 e dt can be quickly computed
1 e dt + 2
by a numerical integration method, e.g., by the trapezoidal method.
1.3.1
Problems
I. Integrate by Guess-and-Check.
1.
3.
xe5x dx.
1
xe 2 x dx.
x cos 2x dx.
5.
7.
(x2
4.
2.
x sin 3x dx.
Answer. ex/2 (4 2x) + c.
1
Answer. x cos 2x +
2
1 2 1
x
sin 2x + c.
2
4

p
x
dx.
Answer. x2 + 1 + c.
x2 + 1
Z
1

x
1 2
2
6.
dx.
Answer.
ln
x
+
1
ln
x
+
2
+ c.
(x2 + 1)(x2 + 2)
2
2
1
dx.
+ 1)(x2 + 9)
Answer.
1
1
x
tan1 x
tan1 + c.
8
24
3
13
1.3. SEPARABLE EQUATIONS

8.
(ln x)5
dx.
x
9.
x2 ex dx.
10.
2x
Answer.
Answer.
sin 3x dx.
1
(ln x)6 + c.
6
1 x3
e + c.
3
2x
Answer. e
2
3
sin 3x
cos 3x + c.
13
13

Hint: Look for the antiderivative in the form Ae2x sin 3x + Be2x cos 3x, and
determine the constants A and B by differentiation.
II. Find the general solution of the linear problems.
1. y 0 +
1
y = cos x.
x
2. xy 0 + 2y = ex .
3. x4 y 0 + 3x3 y = x2 ex .
4.
dy
= 2x(x2 + y).
dx
5. xy 0 2y = xe1/x .
6. y 0 + 2y = sin 3x.
c
cos x
+ sin x +
.
x
x
c
(x + 1)ex
Answer. y = 2
.
x
x2
c
(x 1)ex
Answer. y = 3 +
.
x
x3
Answer. y =
Answer. y = cex x2 1.
Answer. y = cx2 x2 e1/x .
Answer. y = ce2x +
3
2
sin 3x
cos 3x.
13
13
7. x yy 0 1 = y 2 .
Hint: Set v = y 2 . Then v 0 = 2yy 0 , and one obtains a linear equation for
v = v(x).
Answer. y 2 = 2x + cx2 .
III. Find the solution of the initial value problem, and state the maximum
interval on which this solution is valid.
1
cos(x)+x sin(x)
1. y 0 + y = cos x, y( ) = 1.
Answer. y =
; (0, ).
x
x
2
2. xy 0 + (2 + x)y = 1, y(2) = 0.
Answer. y =
(, 0).
3. x(y 0 y) = ex , y(1) =
4. (t + 2)
1
.
e
dy
+ y = 5, y(1) = 1.
dt
1
3ex2
1
+
2;
2
x
x
x
Answer. y = ex ln |x| + ex ; (, 0).

Answer. y =
5t 2
; (2, ).
t+2
14
5. ty 0 2y = t4 cos t, y(/2) = 0.
Answer. y = t3 sin t + t2 cos t t2 ; (, ). Solution is valid for all t.

2
5et 5e2
dr
+ r = 5tet , r(2) = 0.
Answer. r =
; (1, ).
6. t ln t
dt
ln t
dy
1
7.
= 2
, y(2) = 0.
dx
y +x
dx
Hint: Obtain a linear equation for
.
Answer. x = 2+4ey 2y y 2 .
dy
8 . Find a solution (y = y(t)) of y 0 + y = sin 2t, which is a periodic function.
Hint: Look for a solution in the form y(t) = A sin 2t + B cos 2t, plug this
into the equation, and determine the constants A and B.
1
2
Answer. y = sin 2t cos 2t.
5
5
9 . Show that the equation y 0 + y = sin 2t has no other periodic solutions.

Hint: Consider the equation that the difference of any two solutions satisfies.
10 . For the equation
y 0 + a(x)y = f (x)
assume that a(x) a0 > 0, where a0 is a constant, and f (x) 0 as x .
Show that any solution tends to zero as x .
Rx
Hint: Write the integrating factor as (x) = e

as x . Then express
y=
Rx
0
a(t) dt
ea0 x , so that (x)
(t)f (t) dt + c
,
(x)
and use LHospitals rule.

IV. Solve by separating the variables.
1.
dy
2
=
.
dx
x(y 3 + 1)
Answer.
2. ex dx ydy = 0, y(0) = 1.
3. (x2 y 2 + y 2 )dx yxdy = 0.
y4
+ y 2 ln x = c.
4
Answer. y = 2ex 1.
Answer. y = e
x2
2
+ln x+c
= cxe
x2
2
15
1.4. SOME SPECIAL EQUATIONS

2
.
Answer. y =
2
2 t +13
4. y 0 (t) = ty 2 (1 + t2 )1/2 , y(0) = 2.
e ex x
Answer. y =
.
e
5. (y xy + x 1)dx + x dy = 0, y(1) = 0.
0
Rx
x2
6. y = e y, y(2) = 1.
0
Answer. y = e
7. y = xy + xy, y(0) = 2.
et dt
Answer. y =
2e
x2
2
3 2e
1.4
1.4.1
x2
2
Some Special Equations

Homogeneous Equations
y
Let f (t) be a given function. If we set here t = , we obtain a function
x
y
f ( ), which is a function of two variables x and y, but it depends on them
x
in a special way. One calls functions of the form f (y/x) to be homogeneous.
y 4x
is a homogeneous function, because we can put it into
For example,
xy
the form
y 4x
y/x 4
=
,
xy
1 y/x
so that here f (t) =
t4
1t .
Our goal is to solve the homogeneous equation

(4.1)
dy
y
= f ( ).
dx
x
y
Set v = . Since y is a function of x, the same is true of v = v(x). Solving
x
for y, y = xv, we express by the product rule
dy
dv
= v+x .
dx
dx
Switching to v in (4.1), gives
dv
= f (v).
dx
This a separable equation! Indeed, after taking v to the right, we can separate the variables
Z
Z
dv
dx
dv =
.
f (v) v
x
(4.2)
v+x
16
After solving this equation for v(x), we can express the original unknown
y = xv(x).
In practice, one should try to remember the formula (4.2).
Example Solve
dy
x2 + 3y 2
=
dx
2xy
y(1) = 2 .
To see that this equation is homogeneous, we rewrite it as

dy
1x 3y
=
+
.
dx
2y
2x
Set v =
y
x
1
, or y = xv. Observing that = , we have
x
y
v
v+x
dv
11 3
=
+ v.
dx
2v 2
Simplify:
dv
11 1
1 + v2
=
+ v=
.
dx
2v 2
2v
Separating the variables, gives
x
2v
dv =
1 + v2
dx
.
x
We now obtain the solution, by doing the following steps (observe that ln c
is another way to write an arbitrary constant)
ln (1 + v 2 ) = ln x + ln c = ln cx;
1 + v 2 = cx;
v = cx 1;
y(x) = xv = x cx 1 .
From the initial condition
y(1) = c 1 = 2.
It follows that we need to select minus, and c = 5.
Answer: y(x) = x 5x 1.
17
We mention next an alternative (equivalent) definition: a function f (x, y)

is called homogeneous if
f (tx, ty) = f (x, y),
If this holds, then setting t =
for all constants t.
1
, we see that
x
y
f (x, y) = f (tx, ty) = f (1, ),
x
y
so that f (x, y) is a function of , and the old definition applies.
x
Example Solve
dy
y
=
, with x > 0 .
dx
x + xy
It is more straightforward to use the new definition to verify that the function
y
f (x, y) = x+
xy is homogeneous. For all t > 0, we have
f (tx, ty) =
y
(ty)
p
=
= f (x, y) .
x
+
xy
(tx) + (tx)(ty)
Letting y/x = v, or y = xv, we rewrite this equation as

xv
v
.
v + xv 0 =
=
x + x xv
1+ v
We proceed to separate the variables:
x
dv
v
v 3/2
v =
;
=
dx
1+ v
1+ v
Z
Z
dx
1+ v
dv
=
;
3/2
x
v
2v 1/2 + ln v = ln x + c .
The integral on the left was evaluated by performing division, and splitting
it into two pieces. Finally, we replace v by y/x:
2
x
y
+ ln = ln x + c ;
y
x
x
+ ln y = c .
y
We have obtained an implicit representation of a family of solutions.
2
When separating the variables, we had to assume that v 6= 0 (in order

to divide by v 3/2 ). In case v = 0, we obtain another solution: y = 0.
18
1.4.2
The Logistic Population Model
Let y(t) denote the number of rabbits on a tropical island at time t. The
simplest model of population growth is
y 0 = ay
y(0) = y0 .
Here a > 0 is a given constant. This model assumes that initially the number
of rabbits was equal to some number y0 > 0, while the rate of change of
population, given by y 0 (t), is proportional to the number of rabbits. The
population of rabbits grows, which results in a faster and faster rate of
growth. One expects an explosive growth. Indeed, solving the equation, we
get
y(t) = ceat .
From the initial condition y(0) = c = y0 , which gives us y(t) = y0 eat , an
exponential growth. This is the notorious Malthusian model of population
growth. Is it realistic? Yes, sometimes, for a limited time. If the initial
number of rabbits y0 is small, then for a while their number may grow
exponentially.
A more realistic model, which may be used for a long time, is the logistic
model:
y 0 = ay by 2
y(0) = y0 .
Here a, b and y0 are given positive constants, y = y(t). If y0 is small,

then for small t, y(t) is small. Then the by 2 term is negligible, and we
have exponential growth. As y(t) increases, the by 2 term is not negligible
anymore, and we can expect the rate of growth of y(t) to get smaller and
smaller, and y(t) to tend to a finite limit. (Writing the equation as y 0 =
(a b y)y, we can regard the a b y term as the rate of growth.) In case
the initial number y0 is large (when y0 > a/b), the quadratic on the right
is negative, so that y 0 (t) < 0, and the population decreases. If y0 = a/b,
then y 0 (0) = 0, and we expect that y(t) = a/b for all t. We now solve the
problem to confirm our guesses.
This problem can be solved by separation of variables. Instead, we use
another technique. Divide both sides of the equation by y 2 :
y 2 y 0 = ay 1 b.
19
1
. By the generalized
y(t)
power rule, v 0 = y 2 y 0 , so that we can rewrite the last equation as
Introduce a new unknown function v(t) = y 1 (t) =
v 0 = av b,
or
v 0 + av = b.
This a linear equation for v(t)! To solve it, we follow the familiar steps, and
then we return to the original unknown function y(t):
R
(t) = e
a dt
= eat ;
d h at i
e v = beat;
dt
at
e v=b
eat dt =
v=
y(t) =
b at
e + c;
a
b
+ ceat ;
a
1
=
v
b
a
1
.
+ ceat
To find the constant c, we use the initial condition

y(0) =
c=
y(t) =
b
a
1
= y0 ;
+c
1
b
;
y0 a
1
b
a
1
y0
b
a
eat
The problem is solved. Observe that limt+ y(t) = a/b, no matter what
initial value y0 we take. The number a/b is called the carrying capacity. It
tells us the number of rabbits, in the long run, that our island will support.
A typical solution curve, called the logistic curve is given in Figure 1.2.
20

y
2.5
2.0
1.5
1.0
0.5
0.2
0.4
0.6
0.8
1.0
1.2
1.4
Figure 1.2: The solution of y 0 = 5y 2y 2 , y(0) = 0.2
1.4.3
Bernoullis Equation
Let us solve the equation

y 0 (t) = p(t)y(t) + g(t)y n(t).
Here p(t) and g(t) are given functions, n is a given constant. The logistic
model above is just a particular example of Bernoullis equation.
We divide this equation by y n :
y n y 0 = p(t)y 1n + g(t).
Introduce a new unknown function v(t) = y 1n (t). Compute v 0 = (1
1
n)y n y 0 , so that y n y 0 = 1n
v 0 , and rewrite the equation as
v 0 = (1 n)p(t)v + (1 n)g(t).
This is a linear equation for v(t)! After solving for v(t), we calculate y(t) =
1
v 1n (t).
Example Solve
t
y0 = y + .
y
Writing this equation in the form y 0 = y + ty 1/2 , we see that this is

Bernoullis equation, with n = 1/2, so that we need to divide through
by y 1/2 . But that is the same as multiplying through by y 1/2, which we
do, obtaining
y 1/2 y 0 = y 3/2 + t .
21
We now let v(t) = y 3/2 , v 0 (t) = 23 y 1/2 y 0 , obtaining a linear equation for v,
which we solve as usual:
2 0
v = v + t;
3
(t) = e
3
e 2 t v =
3
2
dt
= e 2 t ;
3
3
v 0 v = t;
2
2
d h 3t i 3 3t
e 2 v = te 2 ;
dt
2
3
3 3t
2 3
te 2 dt = te 2 t e 2 t + c;
2
3
v = t
3
2
+ ce 2 t .
3
3
2
Returning to the original variable y, we have the answer: y = t + ce 2 t
3
1.4.4 Riccatis Equation

Let us try to solve the equation
y 0 (t) + a(t)y(t) + b(t)y 2 (t) = c(t).
Here a(t), b(t) and c(t) are given functions. In case c(t) = 0, this is
Bernoullis equation, which we can solve. For general c(t), one needs some
luck to solve this equation. Namely, one needs to guess a particular solution
p(t). Then a substitution y(t) = p(t) + z(t) produces Bernoullis equation
for z(t)
z 0 + (a + 2bp) z + bz 2 = 0 ,
which we can solve.
There is no general way to find a particular solution, which means that
one cannot always solve Riccatis equation. Occasionally one can get lucky.
Example Solve
y 0 + y 2 = t2 2t .
We have a quadratic on the right, which suggest that we look for a particular
solution in the form y = at + b. Plugging this in, we get a quadratic on the
left too. Equating the coefficients in t2 , t and constant terms, we obtain three
equations to find a and b. In general, three equations with two unknowns will
have no solutions, but this is a lucky case, with the solution a = 1, b = 1,
2/3
22
so that p(t) = t+1 is a particular solution. Substituting y(t) = t+1+v(t)

into the equation, and simplifying, we get
v 0 + 2(1 t)v = v 2 .
This is Bernoullis equation. As before, we divide through by v 2 , and then
1
v0
set z = , z 0 = 2 , to get a linear equation:
v
v
v 2 v 0 + 2(1 t)v 1 = 1;
= e
2(1t) dt
et
= et
2 2t
2 2t
z=
d h t2 2t i
2
e
z = et 2t;
dt
z 0 2(1 t)z = 1;
et
2 2t
dt .
The last integral cannot be evaluated through elementary functions (Mathematica can evaluate it through a special function, called Erfi). So we
leave this integral unevaluated. Then we get z from the last formula, after
which we express v, and finally y. We have obtained a family of solutions:
2
et 2t
y(t) = t + 1 + R t2 2t . (The usual arbitrary constant c is now inside
e
dt
the integral.) Another solution: y = t + 1 (corresponding to v = 0).
Example Solve
6
.
t2
We look for a particular solution in the form y(t) = a/t. Plugging this in,
we determine a = 2, so that p(t) = 2/t is a particular solution (a = 3/2 is
also a possibility). The substitution y(t) = 2/t + v(t) produces Bernoullis
equation
8
v 0 + v + 2v 2 = 0 .
t
7
2
Solving it, we obtain v(t) = 8
, and v = 0. Solutions: y(t) = +
ct 2t
t
7
2
, and also y = .
ct8 2t
t
Let us outline an alternative approach to the last problem. Set y = 1/z,
where z = z(t) is a new unknown function. Plugging this into (4.3), then
clearing the denominators, we have
(4.3)
y 0 + 2y 2 =
z0
1
6
+ 2 2 = 2;
2
z
z
t
23

6z 2
.
t2
This is a homogeneous equation, which can be solved as before.
z 0 + 2 =
There are some important ideas that we learned in this subsection.

Knowledge of one particular solution may help to crack open the equation,
and get other solutions. Also, the form of this particular solution depends
on the equation.
1.4.5 Parametric Integration

Let us solve the problem (here y = y(x))
y=
1 y0 2
y(0) = 1 .
This equation is not solved for the derivative y 0 (x). Solving for y 0 (x), and
then separating the variables, one may indeed find the solution. Instead, let
us assume that
y 0 (x) = sin t,
where t is a parameter. From the equation
y=
1 sin2 t =
cos2 t = cos t,
if we assume that cos t 0. Recall the differentials: dy = y 0 (x) dx, or

dx =
so that
dy
sin t dt
=
= dt,
y 0 (x)
sin t
dx
= 1, which gives
dt
x = t + c.
We obtained a family of solutions in parametric form:

x = t + c
y = cos t .
Solving for t, t = x+c, so that y = cos(x+c). From the initial condition,

we conclude that c = 0, giving us the solution y = cos x. This solution is
valid on infinitely many disjoint intervals where cos x 0 (because we see
24
from the equation that y 0). This problem admits another solution: y = 1.
For the equation
y0 + y0 = x
we do not have an option of solving for y 0 (x). Parametric integration appears
to be the only way to solve it. We let y 0 (x) = t, so that from the equation,
x = t5 + t, and dx = (5t4 + 1) dt. Then
dy = y 0 (x) dx = t(5t4 + 1) dt ,
dy
so that
= t(5t4 + 1), which gives y = 56 t6 + 12 t2 + c. We obtained a family
dt
of solutions in parametric form:
x = t5 + t
y = 56 t6 + 12 t2 + c .
1.4.6
Some Applications
Differential equations arise naturally in geometric and physical problems.

Example Find all positive decreasing functions y = f (x), with the following property: the area of the triangle formed by the vertical line going down
from the curve, the x-axis and the tangent line to this curve is constant,
equal to a > 0.
y
6
(x0 , f (x0 ))
c
@
@
@
@
@
y = f (x)
@
x1
- x
The triangle formed by the tangent line, the line x = x0 , and the x-axis
25
Let (x0 , f (x0)) be an arbitrary point on the graph of y = f (x). Draw

the triangle in question, formed by the horizontal line x = x0 , the x-axis,
and the tangent line to this curve. The tangent line intersects the x-axis
at some point x1 , lying to the right of x0 , because f (x) is decreasing. The
slope of the tangent line is f 0 (x0 ), so that the point-slope equation of the
tangent line is
y = f (x0 ) + f 0 (x0 )(x x0 ) .
At x1 , we have y = 0, so that
0 = f (x0 ) + f 0 (x0 )(x1 x0 ) .
Solve this for x1 , x1 = x0
f (x0 )
. It follows that the horizontal side of our
f 0 (x0 )
f (x0 )
, while the vertical side is f (x0 ). The area of this right
f 0 (x0 )
triangle is then
1 f 2 (x0 )
= a.
2 f 0 (x0 )
triangle is
(Observe that f 0 (x0 ) < 0, so that the area is positive.) The point x0 was
arbitrary, so that we replace it by x, and then we replace f (x) by y, and
f 0 (x) by y 0 :
1 y2
y0
1
.
0 = a; or 2 =
2y
y
2a
We solve this differential equation by taking the antiderivatives of both sides:
1
1
=
x+c.
y
2a
Answer: y(x) =
2a
.
x + 2ac
2a
.
x
Example A tank holding 10L (liters) is originally filled with water. A saltwater mixture is pumped into the tank at a rate of 2L per minute. This
mixture contains 0.3 kg of salt per liter. The excess fluid is flowing out of
the tank at the same rate (2L per minute). How much salt does the tank
contain after 4 minutes?
This is a family of hyperbolas. One of them is y =
Let t be the time (in minutes) since the mixture started flowing, and let
y(t) denote the amount of salt in the tank at time t. The derivative, y 0 (t),
gives the rate of change of salt per minute, and it is equal to the difference
between the rate at which salt flows in, and the rate it flows out. The salt
is pumped in at a rate of 0.6 kg per minute. The density of salt at time t is
26
y(t)
y(t)
(so that each liter of the solution in the tank contains
kg of salt).
10
10
y(t)
Then, the salt flows out at the rate 2
= 0.2 y(t) kg/min. The difference
10
of these two rates is y 0 (t), so that
y 0 = 0.6 0.2y .
This is a linear differential equation. Initially, there was no salt in the tank,
so that y(0) = 0 is our initial condition. Solving this equation together with
the initial condition, we have y(t) = 3 3e0.2t. After 4 minutes, we have
y(4) = 3 3e0.8 1.65 kg of salt in the tank.
Now suppose a patient has alcohol poisoning, and doctors are pumping
in water to flush his stomach out. One can compute similarly the weight of
poison left in the stomach at time t.
1.5
Exact Equations
Let us begin by recalling partial derivatives. If a function f (x) = x2 + a

depends on a parameter a, then f 0 (x) = 2x. If g(x) = x2 + y 3 , with a
dg
parameter y, we have
= 2x. Another way to denote this derivative
dx
is gx = 2x. We may also regard g as a function of two variables, g =
g(x, y) = x2 + y 3 . Then the partial derivative with respect to x is computed
g
= 2x.
by regarding y to be a parameter, gx = 2x. Alternative notation:
x
g
Similarly, a partial derivative with respect to y is gy =
= 3y 2 . It gives
y
us the rate of change in y, when x is kept fixed.
The equation (here y = y(x))
y 2 + 2xyy 0 = 0
can be easily solved, if we rewrite it in the equivalent form
d 2
xy = 0.
dx
Then xy 2 = c, and the solution is
c
y(x) = .
x
27
1.5. EXACT EQUATIONS

We wish to play the same game for general equations of the form
(5.1)
M (x, y) + N (x, y)y 0(x) = 0.
Here the functions M (x, y) and N (x, y) are given. In the above example,
M = y 2 and N = 2xy.
Definition The equation (5.1) is called exact if there is a function (x, y),
with continuous derivatives up to second order, so that we can rewrite (5.1)
in the form
d
(5.2)
(x, y) = 0.
dx
The solution of the exact equation is (c is an arbitrary constant)
(5.3)
(x, y) = c.
There are two natural questions: when is the equation (5.1) exact, and if it
is exact, how does one find (x, y)?
Theorem 1.5.1 Assume that the functions M (x, y), N (x, y), My (x, y) and
Nx (x, y) are continuous in some disc D : (x x0 )2 + (y y0 )2 < r 2 , around
some point (x0 , y0 ). Then the equation (5.1) is exact in D, if and only if,
the following partial derivatives are equal
(5.4)
My (x, y) = Nx(x, y) ,
for all points (x, y) in D.
This theorem says two things: if the equation is exact, then the partials are
equal, and conversely, if the partials are equal, then the equation is exact.
Proof:
1. Assume that the equation (5.1) is exact, so that it can be
written in the form (5.2). Performing the differentiation in (5.2), we write
it as
x + y y 0 = 0 .
But this is the same equation as (5.1), so that
x = M
y = N .
Taking the second partials
xy = My
yx = Nx .
28
We know from Calculus that xy = yx , therefore My = Nx.

2. Assume that My = Nx. We will show that the equation (5.1) is then
exact by producing (x, y). We just saw that (x, y) must satisfy
(5.5)
x = M (x, y)
y = N (x, y) .
Take the antiderivative in x of the first equation

(5.6)
(x, y) =
M (x, y) dx + h(y),
where h(y) is an arbitrary function of y. To determine h(y), plug the last

formula into the second line of (5.5)
y (x, y) =
or
(5.7)
My (x, y) dx + h0 (y) = N (x, y),
h (y) = N (x, y)
My (x, y) dx p(x, y).
Observe that we denoted by p(x, y) the right side of the last equation. It
turns out that p(x, y) does not really depend on x! Indeed, taking the partial
derivative in x,
p(x, y) = Nx (x, y) My (x, y) = 0,

x
because it was given to us that My (x, y) = Nx(x, y). So that p(x, y) is a
function of y only, or p(y). The equation (5.7) takes the form
h0 (y) = p(y).
We determine h(y) by integration, and use it in (5.6) to get (x, y).
Recall that the equation in differentials
M (x, y) dx + N (x, y) dy = 0
is an alternative form of (5.1).
Example Solve
ex sin y + y 3 (3x ex cos y)
dy
= 0.
dx
29
1.5. EXACT EQUATIONS

Here M (x, y) = ex sin y + y 3 , N (x, y) = 3x + ex cos y. Compute
My = ex cos y + 3y 2
Nx = ex cos y 3 .
The partials are not the same, this equation is not exact, and our theory
does not apply.
Example Solve (for x > 0)

Here M (x, y) =
y
+ 6x
x
dx + (ln x 2) dy = 0 .
y
+ 6x and N (x, y) = ln x 2. Compute
x
My =
1
= Nx ,
x
and so the equation is exact. To find (x, y), we observe that the equations
(5.5) take the form
y
x = + 6x
x
y = ln x 2 .
Take the antiderivative in x of the first equation
(x, y) = y ln x + 3x2 + h(y),
where h(y) is an arbitrary function of y. Plug this (x, y) into the second
equation
y = ln x + h0 (y) = ln x 2,
which gives
h0 (y) = 2.
Integrating, h(y) = 2y, and so (x, y) = y ln x + 3x2 2y, giving us the
solution
y ln x + 3x2 2y = c .
c 3x2
. Observe that when we
ln x 2
were solving for h(y), we chose the integration constant to be zero, because
at the next step we set (x, y) equal to c, an arbitrary constant.
We can solve this relation for y, y(x) =
30
Example Find the constant b, for which the equation

ye2xy + x dx + bxe2xy dy = 0
is exact, and then solve the equation with that b.

Setting equal the partials My and Nx, we have
e2xy + 2xye2xy = be2xy + 2bxye2xy .
We see that b = 1. When b = 1, the equation becomes

ye2xy + x dx + xe2xy dy = 0,
and we know that it is exact. We look for (x, y) as before:

x = ye2xy + x
y = xe2xy .
Just for practice, let us begin this time with the second equation. Taking
the antiderivative in y, in the second equation,
(x, y) =
1 2xy
e + h(x) ,
2
where h(x) is an arbitrary function of x. Plugging (x, y) into the first

equation,
x = ye2xy + h0 (x) = ye2xy + x .
This tells us that h0 (x) = x, h(x) = 12 x2 , and then (x, y) = 12 e2xy + 21 x2 .
1
1
1
Answer: e2xy + x2 = c, or y =
ln(2c x2 ).
2
2
2x
Exact equations are connected with conservative vector fields. Recall
that a vector field F(x, y) =< M (x, y), N (x, y) > is called conservative if
there is a function (x, y), called the potential, such that F(x, y) = (x, y).
Recalling that (x, y) =< x, y >, we have x = M , and y = N , the
same relations that we had for exact equations.
31
1.6. EXISTENCE AND UNIQUENESS OF SOLUTION
1.6
Existence and Uniqueness of Solution
We consider a general initial value problem

y 0 = f (x, y)
y(x0 ) = y0 ,
with a given function f (x, y), and given numbers x0 and y0 . Let us ask two
basic questions: is there a solution of this problem, and if there is, is the
solution unique?
Theorem 1.6.1 Assume that the functions f (x, y) and fy (x, y) are continuous in some neighborhood of the initial point (x0 , y0 ). Then there exists a
solution, and there is only one solution. The solution y = y(x) is defined on
some interval (x1 , x2 ) that includes x0 .
One sees that the conditions of this theorem are not too restrictive,
so that the theorem tends to apply, providing us with the existence and
uniqueness of solution. But not always!
Example Solve
y0 =
y(0) = 0 .
The function f (x, y) = y is continuous (for y 0), but its partial derivative
1
in y, fy (x, y) = 2
y , is not even defined at the initial point (0, 0). The
2
theorem does not apply. One checks that the function y = x4 solves our
initial value problem (for x 0). But here is another solution: y = 0.
(Having two different solutions of the same initial value problem is like
having two primadonnas in the same theater.)
Observe that the theorem guarantees existence of solution only on some
interval (it is not happily ever after).
Example Solve for y = y(t)
y0 = y2
y(0) = 1 .
Here f (t, y) = y 2 , and fy (t, y) = 2y are continuous functions. The theorem
1
applies. By separation of variables, we determine the solution y(t) =
.
1t
As time t approaches 1, this solution disappears, by going to infinity. This
phenomenon is sometimes called blow up in finite time.
32
1.7
Numerical Solution by Eulers method
We have learned a number of techniques for solving differential equations,

however the sad truth is that most equations cannot be solved (by a formula).
Even a simple looking equation like
y0 = x + y3
is totally out of reach. Fortunately, if you need a specific solution, say the
one satisfying the initial condition y(0) = 1, it can be easily approximated
(by the Theorem 1.6.1, such solution exists, and is unique).
In general, we shall deal with the problem
y 0 = f (x, y)
y(x0 ) = y0 .
Here the function f (x, y) is given (in the example above we had f (x, y) =
x + y 3 ), and the initial condition prescribes that the solution is equal to a
given number y0 , at a given point x0 . Fix a step size h, and let x1 = x0 + h,
x2 = x0 + 2h, . . . , xn = x0 + nh. We will approximate y(xn ), the value of
the solution at xn . We call this approximation yn . To go from the point
(xn , yn ) to the point (xn+1 , yn+1 ) on the graph of solution y(x), we use the
tangent line approximation:
yn+1 yn + y 0 (xn )(xn+1 xn ) = yn + y 0 (xn )h = yn + f (xn , yn )h .
(We have expressed y 0 (xn ) = f (xn , yn ) from the differential equation.) The
resulting formula is easy to implement, it is just one loop, starting with the
initial point (x0 , y0 ).
One continues the computations until the points xn go as far as needed.
Decreasing the step size h, will improve the accuracy. Smaller hs will require
more steps, but with the power of modern computers, that is not a problem,
particularly for simple examples, like the one above. In that example x0 = 0,
y0 = 1. If we choose h = 0.05, then x1 = 0.05, and
y1 = y0 + f (x0 , y0 )h = 1 + (0 + 13 ) 0.05 = 1.05.
Continuing, we have x2 = 0.1, and
y2 = y1 + f (x1 , y1 )h = 1.05 + (0.05 + 1.053 ) 0.05 1.11 .
33
1.7. NUMERICAL SOLUTION BY EULERS METHOD

Next, x3 = 0.15, and
y3 = y2 + f (x2 , y2 )h = 1.11 + (0.1 + 1.113 ) 0.05 1.18 .
These computations imply that y(0.05) 1.05, y(0.1) 1.11, and y(1.15)
1.18. If you need to approximate the solution on the interval (0, 0.4), you
have to make five more steps. Of course, it is better to program a computer.
Eulers method is using the tangent line approximation, or the first two
terms of the Taylor series approximation. One can use more terms of the
Taylor series, and develop more sophisticated methods (which is done in
books on numerical methods). But here is a question: if it is so easy to compute numerical approximation of solutions, why bother learning analytical
solutions? The reason is that we seek not just to solve a differential equation, but to understand it. What happens if the initial condition changes?
The equation may include some parameters, what happens if they change?
What happens to solutions in the long term?
1.7.1
Problems
I. Determine if the equation is homogeneous, and if it is, solve it.

y + 2x
dy
=
.
dx
x
dy
x2 xy + y 2
2.
=
.
dx
x2
1.
3.
Answer. y = cx + 2x ln x.
Answer. y =
x + cx + x ln x
.
c + ln x
dy
y 2 + 2x
=
.
dx
y
dy
y 2 + 2xy
2x2
=
,
y(1)
=
2.
Answer.
y
=
.
dx
x2
3 2x
y
y
5. xy 0 y = x tan .
Answer. sin = cx.
x
x
4.
6. y 0 =
x2 + y 2
, y(1) = 2.
xy
Answer. y = x2 ln x2 + 4x2 .
y + x1/2 y 3/2
7. y =
, with x > 0, y > 0.
xy
0
II. Solve the following Bernoullis equations.
Answer. 2
y
= ln x + c.
x
34
1. y 0
2.
1
y = y 2 , y(2) = 2.
x
dy
y2 + 1
=
.
dx
2y
Answer. y =
Answer. y = cex 1.
3. y + x 3 y = 3y.
2x
.
2
x2
3
2
x 1
Answer. y =
+ + ce2x , and y = 0.
3 6
Hint: When dividing the equation by 3 y, one needs to check if y = 0 is a

solution, which indeed it is.
4. y 0 + xy = y 3 .
5. The equation
dy
y 2 + 2x
=
dx
y
could not be solved in the last problem set, because

it is not homogeneous.
p
2x
Can you solve it now?
Answer. y = ce 2x 1.
III. Determine if the equation is exact, and if it is, solve it.

1. (2x + 3x2 y) dx + (x3 3y 2 ) dy = 0.
2. (x + sin y) dx + (x cos y 2y) dy = 0.
Answer. x2 + x3 y y 3 = c.
1
Answer. x2 + x sin y y 2 = c.
2
2y 3
x
dx
+
dy = 0.
Answer. x2 + y 4 = c.
x2 + y 4
x2 + y 4
4. Find a simpler solution for the preceding problem.
3.
5. (6xycos y) dx+(3x2+x sin y+1) dy = 0.

6. (2xy) dx+(2y x) dy = 0, y(1) = 2.
Answer. 3x2 yx cos y+y = c.

Answer. x2 +y 2 xy = 3.
7. Find the value of b for which the following equation is exact, and then
solve the equation, using that value of b
(yexy + 2x) dx + bxexy dy = 0 .
Answer. b = 1, y =
1
ln(c x2 ).
x
IV 1. Use parametric integration to solve

3
y0 + y0 = x .
1.7. NUMERICAL SOLUTION BY EULERS METHOD
35
3 4 1 2
t + t + c.
4
2
2. Use parametric integration to solve
Answer. x = t3 + t, y =
y = ln(1 + y 0 ) .
Answer. x = 2 tan1 t + c, y = ln(1 + t2 ). Another solution: y = 0.
3. Use parametric integration to solve
y 0 + sin(y 0 ) = x .
1
Answer. x = t + sin t, y = t2 + t sin t + cos t + c.
2
4. A tank has 100L of water-salt mixture, which initially contains 10 kg of
salt. Water is flowing in at a rate of 5L per minute. The new mixture flows
out at the same rate. How much salt remains in the tank after an hour?
Answer. 0.5 kg.
5. A tank has 100L of water-salt mixture, which initially contains 10 kg of
salt. A water-salt mixture is flowing in at a rate of 3L per minute, and each
liter of it contains 0.1 kg of salt. The new mixture flows out at the same
rate. How much salt remains in the tank after t minutes?
Answer. 10 kg.
6. Water is being pumped into patients stomach at a rate of 0.5L per minute
to flush out 300 grams of alcohol poisoning. The excess fluid is flowing out at
the same rate. The stomach holds 3L. The patient can be discharged when
the amount of poison drops to 50 grams. How long should this procedure
last?
7. Find all curves y = f (x) with the following property: if you draw a
tangent line at any point (x, f (x)) on this curve, and continue the tangent
line until it intersects the x -axis, then the point of intersection is x2 .
Answer. y = cx2 .
8. Find all positive decreasing functions y = f (x), with the following property: in the triangle formed by the vertical line going down from the curve,
the x-axis and the tangent line to this curve, the sum of two sides adjacent
to the right angle is constant, equal to a > 0.
Answer. y a ln y = x + c.
36
9. (i) Apply Eulers method to

y 0 = x(1 + y), y(0) = 1 .
Take h = 0.25, and do four steps, obtaining an approximation for y(1).
(ii) Take h = 0.2, and do five steps of Eulers method, obtaining another
approximation for y(1).
(iii) Solve the above problem analytically, and determine which one of the
two approximations is better.
10 . (From the Putnam competition, 2009) Show that any solution of
y0 =
x2 y 2
+ 1)
x2 (y 2
satisfies lim y(x) = .

x
Hint: Using partial fractions, rewrite this equation as

y0 =
1 + 1/x2
1
2.
2
y +1
x
Assume, on the contrary, that y(x) is bounded when x is large. Then y 0 (x)
exceeds a positive constant for all large x, and therefore y(x) tends to infinity,
a contradiction (observe that 1/x2 becomes negligible for large x).
11. Solve
x(y 0 ey ) + 2 = 0 .
Hint: Divide the equation by ey , then set v = ey , obtaining a linear equation for v = v(x).
Answer. y = ln(x + cx2 ).
12. Solve the integral equation
y(x) =
y(t) dt + x + 1 .
Hint: Differentiate the equation, and also evaluate y(1).

Answer. y = 3ex1 1.
V 1. Find two solutions of the initial value problem
y 0 = (y 1)1/3, y(1) = 1 .
1.8. THE EXISTENCE AND UNIQUENESS THEOREM
37
Is it good to have two solutions of the same initial value problem? What
went wrong? (Why the existence and uniqueness theorem does not apply
here?)
2. Find all y0 , for which the following problem has a unique solution
x
y0 = 2
, y(2) = y0 .
y 2x
Hint: Apply the existence and uniqueness theorem.

Answer. All y0 except 2.
3. Show that the function
x|x|
solves the problem
4
p
y 0 = |y|
y(0) = 0 ,
for all x. Can you find another solution?
Hint: Consider separately the cases when x > 0, x < 0, and x = 0.
4. Show that the problem (here y = y(t))
y 0 = y 2/3
y(0) = 0
has infinitely many solutions.
Hint: Consider y(t) that is equal to zero for t < a, and to
where a > 0 is any constant.
(ta)3
27
for t a,
1.8 The Existence and Uniqueness Theorem

For the initial value problem
(8.1)
y 0 = f (x, y)
y(x0 ) = y0 ,
we shall prove a more general existence and uniqueness theorem than the
one stated before. We define a rectangular box B around the initial point
(x0 , y0 ) to be the set of points (x, y), satisfying x0 a x x0 + a and
y0 b x y0 + b, for some positive a and b. It is known from Calculus
that in case f (x, y) is continuous on B, it is bounded on B, so that for some
constant M > 0
(8.2)
|f (x, y)| M, for all points (x, y) in B .
38
Theorem 1.8.1 Assume that f (x, y) is continuous on B, and for some

constant L > 0, we have
(8.3)
|f (x, y1) f (x, y2 )| L|y1 y2 |,
for any two points (x, y1 ) and (x, y2 ) in B. Then the initial value problem
(8.1) has a unique solution, which is defined for x on the interval (x0
b
b
b
b
M , x0 + M ), in case M < a, and on the interval (x0 a, x0 + a) if M a.
b
Proof: Assume, for definiteness, that M
< a. We shall prove the existence
of solutions first, and we shall assume that x > x0 (the case when x < x0
is similar). Integrating the equation in (8.1) over the interval (x0 , x), we
convert the initial value problem (8.1) into an equivalent integral equation
(8.4)
y(x) = y0 +
x0
f (t, y(t)) dt .
(If y(x) solves (8.4), then y(x0 ) = y0 , and by differentiation y 0 = f (x, y).)
By (8.2), M f (t, y(t)) M , and then any solution of (8.4) lies between
two straight lines
y0 M (x x0 ) y(x) y0 + M (x x0 ) .
b
For x0 x x0 + M
these lines stay in the box B. (In the other case,
b
when M a, these lines stay in B for x0 x x0 + a.) We denote
(x) = y0 + M (x x0 ), and call this function a supersolution, while (x) =
y0 M (x x0 ) is called a subsolution.
1. A special case. Let us make an additional assumption that f (x, y)

is increasing in y, so that if y2 > y1 , then f (x, y2 ) > f (x, y1 ), for any two
points (x, y1 ) and (x, y2) in B. We shall construct a solution of (8.1) as a
limit of the sequence of iterates (x), y1 (x), y2 (x), . . ., yn (x), . . ., defined as
follows
Z x
y1 (x) = y0 +
f (t, (t)) dt ,
x0
y2 (x) = y0 +
x0
f (t, y1 (t)) dt , . . . , yn (x) = y0 +
x0
f (t, yn1 (t)) dt .
We claim that for all x on the interval x0 < x x0 +

inequalities hold
(8.5)
b
M,
(x) y1 (x) y2 (x) yn (x) .
the following
39
Indeed, we have f (t, (t)) M , and then

y1 (x) = y0 +
x0
f (t, (t)) dt y0 M (x x0 ) = (x) ,
giving us the first of these inequalities. We have

y2 (x) y1 (x) =
x0
[f (t, y1 (t)) f (t, (t))] dt 0 ,
using that (x) y1 (x), and the monotonicity of f (x, y). So that y1 (x)
y2 (x), and the other inequalities in (8.5) are established similarly. Next, we
b
claim that for all x in the interval x0 < x x0 + M
, all of these iterates lie
below the supersolution (x), so that we have
(8.6)
(x) y1 (x) y2 (x) yn (x) (x) .
Indeed, we have f (t, (t)) M , so that

y1 (x) = y0 +
x0
f (t, (t)) dt y0 + M (x x0 ) = (x) .
It follows that y1 (x) stays in the box B, for x0 < x x0 +

y2 (x) = y0 +
b
M,
and then
x0
f (t, y1 (t)) dt y0 + M (x x0 ) = (x) ,
and so on for all yn (x).

b
At each x in (x0 , x0 + M
), the sequence {yn (x)} is non-decreasing,
bounded above by the number (x). Hence, this sequence has a limit which
we denote by y(x). By the monotone convergence theorem, we may pass to
the limit in the recurrence relation
yn (x) = y0 +
x0
f (t, yn1 (t)) dt ,
concluding that y(x) gives the desired solution of the integral equation (8.4).
2. The general case. Define g(x, y) = f (x, y) + Ay. If we choose the
constant A large enough, then the new function g(x, y) will be increasing in
y. Indeed, using our condition (8.3),
g(x, y2) g(x, y1) = f (x, y2 ) f (x, y1 ) + A (y2 y1 )
40

L (y2 y1 ) + A (y2 y1 ) = (A L) (y2 y1 ) > 0 ,
for any two points (x, y1 ) and (x, y2) in B, provided that A > L, and y2 > y1 .
We now consider an equivalent equation
y 0 + Ay = f (x, y) + Ay = g(x, y) .
Multiplying both sides by the integrating factor eAx , we put this equation
into the form
d h Ax i
e y = eAx g(x, y) .
dx
Set z(x) = eAx y(x), then y(x) = eAx z(x), and the new unknown function
z(x) satisfies

z 0 = eAx g x, eAx z
(8.7)
z(x0 ) = eAx0 y0 .

The function eAx g x, eAx z is increasing in z. The special case applies,

so that the solution z(x) of (8.7) exists. Then y(x) = eAx z(x) gives the
desired solution of (8.1).
Finally, we prove the uniqueness of solution. Let z(x) be another solution
b
of (8.4) on the interval (x0 , x0 + M
), so that
z(x) = y0 +
f (t, z(t)) dt .
x0
Subtracting this from (8.4), we have

y(x) z(x) =
x0
[f (t, y(t)) f (t, z(t))] dt .
Assume first that x is in [x0 , x0 +

estimate
L(x x0 )
Then using the condition (8.3), we
|y(x) z(x)|
1
2L ].
x0
|f (t, y(t)) f (t, z(t))| dt L
max
1
[x0,x0 + 2L
]
|y(x) z(x)|
x0
|y(t)) z(t)| dt
1
max |y(x) z(x)| .
1
2 [x0 ,x0 + 2L
]
It follows that
max
1
[x0 ,x0 + 2L
]
|y(x) z(x)|
1
max |y(x) z(x)| .
1
2 [x0 ,x0 + 2L
]

But then max[x0 ,x0 +
1
]
2L
41
1
].
|y(x)z(x)| = 0, so that y(x) = z(x) on [x0 , x0 + 2L
1
1
Let x1 = x0 + 2L
. We have just proved that y(x) = z(x) on [x0 , x0 + 2L
], and
in particular y(x1 ) = z(x1 ). Repeating (if necessary) the same argument on
1
[x1 , x1 + 2L
], and so on, we will eventually conclude that y(x) = z(x) on
b
(x0 , x0 + M ).
We shall need the following important tool.

Lemma 1.8.1 (The Bellman-Gronwall Lemma) Assume that for x x0 ,
the functions u(x) and a(x) are continuous, and satisfy u(x) 0, a(x) 0.
Assume that for some number K > 0 we have
(8.8)
Then
(8.9)
Proof:
u(x) K +
x0
a(t)u(t) dt, for x x0 (K > 0) .

Rx
u(x) Ke
x0
a(t) dt
, for x x0 .
Divide the inequality (8.8) by its right hand side (which is positive)
a(x)u(x)
R
a(x) .
K + xx0 a(t)u(t) dt
Integrating both sides over (x0 , x) (the numerator of the fraction on the left
is equal to the derivative of its denominator), gives

ln K +
x0
a(t)u(t) dt ln K
a(t) dt ,
x0
which implies that

K+
x0
Rx
a(t)u(t) dt Ke
x0
a(t) dt
Using the inequality (8.8) once more, we get (8.9).
In addition to the initial value problem (8.1), consider

(8.10)
z 0 = f (x, z)
z(x0 ) = z0 .
If z0 = y0 , then z(x) = y(x) for all x. Now suppose that z0 6= y0 , but

|z0 y0 | is small. If the condition (8.3) holds, then we claim that z(x) and
42
y(x) will remain close over any bounded interval (x0 , x0 + p). This is known
as continuous dependence of solutions, with respect to the initial condition.
We begin the proof of the claim by observing that z(x) satisfies
z(x) = z0 +
f (t, z(t)) dt .
x0
From this formula we subtract (8.4), and then estimate

x
z(x) y(x) = z0 y0 +
|z(x) y(x)| |z0 y0 | +

|z0 y0 | +
[f (t, z(t)) f (t, y(t))] dt ;
x0
x0
x0
|f (t, z(t)) f (t, y(t))| dt
L |z(t) y(t)| dt .
(We used the triangle inequality

for numbers:
|a + b| |a| + |b|, the triangle
R
R
inequality for integrals: | xx0 g(t) dt| xx0 |g(t)| dt, and the condition (8.3).)
By the Bellman-Gronwall Lemma
|z(x) y(x)| |z0 y0 | eL(xx0 ) |z0 y0 | eLp, for x (x0 , x0 + p) ,
so that z(x) and y(x) remain close over the interval (x0 , x0 + p), provided
that |z0 y0 | is small.
Problems
1. Assume that the function u(x) 0 is continuous for x 1, and for some
number K > 0, we have
xu(x) K +
x
1
u(t) dt, for x 1 .
Show that u(x) K, for x 1.

2. Assume that the functions a(x) 0, and u(x) 0 are continuous for
x x0 , and we have
u(x)
x0
a(t)u(t) dt, for x x0 .
Show that u(x) = 0, for x x0 . Then give an alternative proof of the

uniqueness part of the Theorem 1.8.1.
43
Hint: Let K 0 in the Bellman-Gronwall Lemma.

3. Assume that the functions a(x) 0, and u(x) 0 are continuous for
x x0 , and we have
u(x)
x0
a(t)u2 (t) dt, for x x0 .
Show that u(x) = 0, for x x0 .
Hint: Observe that u(x0 ) = 0. When t is close to x0 , u(t) is small. But then
u2 (t) < u(t).
4. Show that if a function x(t) satisfies
0
dx
x2 for all t, and x(0) = 0,
dt
then x(t) = 0 for all t (, ).
Hint: Show that x(t) = 0 for t > 0. In case t < 0, introduce new variables
y and s, by setting x = y and t = s, so that s > 0.
Chapter 2
Second Order Equations

2.1
Special Second Order Equations
Probably the simplest second order equation is

y 00 (x) = 0.
Taking the antiderivative
y 0 (x) = c1 .
We have denoted an arbitrary constant by c1 , because we expect another arbitrary constant to make an appearance. Indeed, taking another antiderivative, we get the general solution
y(x) = c1 x + c2 .
This example suggests that general solutions for second order equations
depend on two arbitrary constants.
General second order equations for the unknown function y = y(x) can
often be written as
y 00 = f (x, y, y 0),
where f is a given function of its three variables. We cannot expect all such
equations to be solvable, as we could not even solve all first order equations.
In this section we study special second order equations, which are reducible
to first order equations, greatly increasing their chances to be solved.
44
45
2.1. SPECIAL SECOND ORDER EQUATIONS
2.1.1
y is not present in the equation
Let us solve for y(t) the equation

ty 00 y 0 = t2 .
The derivatives of y are present in this equation, but not the function y itself.
We denote y 0 (t) = v(t), and v(t) is our new unknown function. Clearly,
y 00 (t) = v 0 (t), and the equation becomes
tv 0 v = t2 .
This is a first order equation for v(t)! This equation is linear, so that we
solve it as before. Once we obtain v(t), we determine the solution y(t) by
integration. Details:
1
v0 v = t ;
t
(t) = e
1
t
dt
= e ln t = eln t =
1
;
t
d 1
v = 1;
dt t

1
v = t + c1 ;
t
y 0 = v = t2 + c1 t ;
t2
t3
+ c1 + c2 .
3
2
Here c1 and c2 are arbitrary constants.
y(t) =
Let us solve the following equation for y(x):

2
y 00 + 2xy 0 = 0 .
Again, y is missing in the equation. By setting y 0 = v, with y 00 = v 0 , we
obtain a first order equation. This time the equation for v(x) is not linear,
but we can separate the variables. We have:
v 0 + 2xv 2 = 0 ;
dv
= 2xv 2 .
dx
46
CHAPTER 2. SECOND ORDER EQUATIONS
The last equation has a solution v = 0, or y 0 = 0, giving y = c. Assuming

that v 6= 0, we continue
dv
dv =
v2
2x dx ;
1
= x2 c1 ;
v
y0 = v =
x2
1
.
+ c1
Let us now assume that c1 > 0. Then

Z
y(x) =
x2
1
1
x
dx = arctan + c2 .
+ c1
c1
c1
If c1 = 0 or c1 < 0, we get two different formulas for solutions! Indeed, in

1
1
case c1 = 0, we have y 0 = 2 , and by integration we obtain y = + c3 ,
x
x
the second family of solutions. In case c1 < 0, we can write (replacing c1 by
c21 , with a new c1 )
y0 =
1
1
1
1
=
.
2
2
x c1
2c1 x c1 x + c1

Integrating, we get the third family of solutions

y=
1
1
ln |x c1 |
ln |x + c1 | + c4 .
2c1
2c1
The fourth family of solutions is y = c.
2.1.2
x is not present in the equation
Let us solve for y(x)

3
y 00 + yy 0 = 0 .
All three functions appearing in the equation are functions of x, but x itself
is not present in the equation. On the curve y = y(x), the slope y 0 is a
function of x, but it is also a function of y. We set y 0 = v(y), and v(y) will
be our new unknown function. By the chain rule
y 00 (x) =
d
dy
v(y) = v 0 (y)
= v 0v ,
dx
dx
47
and our equation takes the form

v 0 v + yv 3 = 0 .
This a first order equation! To solve it, we begin by factoring

v v 0 + yv 2 = 0.
If the first factor is zero, y 0 = v = 0, we obtain a family of solutions y = c.
Setting the second factor to zero
dv
+ yv 2 = 0 ,
dy
we have a separable equation. We solve it by separating the variables
dv
=
v2
y dy ;
1
y2
y2
=
+ c1 =
+ c1 ;
v
2
2
dy
2
=v= 2
.
dx
y + 2c1
To find y(x) we need to solve another first order equation. Again, we separate the variables
Z
Z

y 2 + 2c1 dy = 2 dx ;
y 3 /3 + 2c1 y = 2x + c2 .
This gives us a second family of solutions.
2.1.3 The Trajectory of Pursuit

A car is moving along a highway (the x axis) with constant speed a. A
drone is flying in the skies (at a point (x, y), which depends on time t), with
constant speed v. Find the trajectory for the drone, so that the tangent
line always passes through the car. Assume that the drone starts at a point
(x0 , y0 ), and the car starts at x0 .
If X gives the position of the car, we can express the slope of the tangent
line as follows (see the figure)
(1.1)
dy
y
=
.
dx
Xx
48
We have X = x0 + at. Then we rewrite (1.1) as

x0 + at x = y
dx
.
dy
Differentiate this formula with respect to y, and simplify

dt
1 d2 x
= y 2.
dy
a dy
(1.2)
q
1
ds
, and ds = dx2 + dy 2 , so that dt = ds =
On the other hand, v =
dt
v
1q 2
2
dx + dy , and then
v
dt
1
=
dy
v
(1.3)
dx
dy
2
+ 1.
dt
< 0, so that minus is needed in front of the square root.)
dy
dx 00
d2 x
Comparing (1.2) with (1.3), and writing x0 (y) =
, x (y) = 2 , we arrive
dy
dy
at the equation of the motion for the drone
(Observe that
y x00 (y) =
a
v
x0 2 (y) + 1 .
6 c(x0 , y0 )
x0
@
@c(x, y)
@
@
R
@
@
@
@c
y = f (x)
- x
49
The unknown function x(y) is not present in this equation. We therefore

set x0 (y) = p(y), with x00 (y) = p0 (y), obtaining a first order equation for p(y),
which we solve by separating the variables
dp
aq 2
=
p + 1,
dy
v
y
Z
dp
a
=
2
v
p +1
dy
.
y
The integral on the left is computed by a substitution p = tan , giving

ln p +
p+
We write this as
p2
+1 =
a
(ln y + ln c) ,
v
p2 + 1 = cy v
(with a new c) .
p2 + 1 = cy v p ,
(1.4)
and square both sides, getting

1 = c2 y
2a
v
2cy v p .
We solve this for p = x0 (y):

a
1
1 a
x0 (y) = cy v y v .
2
2c
The constant c we determine from (1.4). At t = 0, p = x0 (y0 ) = 0, and so

a
c = y0 v . (At t = 0, the drone is pointed vertically down.) Then

x0 (y) =
1
2
y
y0
a
v
1
2
y
y0
a
v
Integrating, and using that x(y0 ) = x0 , we finally obtain (assuming v 6= a)

y0
x(y) =
2(1 + a/v)
"
y
y0
1+ a
v
y0
1
2(1 a/v)
"
y
y0
1 a
v
1 + x0 .
50
2.2
Linear Homogeneous Equations with Constant

Coefficients
We wish to find solution y = y(t) of the equation

(2.1)
ay 00 + by 0 + cy = 0 ,
where a, b and c are given numbers. This is arguably the most important
class of differential equations, because it arises when applying Newtons
second law of motion (or when modeling electrical oscillations). If y(t)
denotes displacement of an object at time t, then this equation relates the
displacement with velocity y 0 (t) and acceleration y 00 (t).
Observe that if y(t) is a solution, then so is 2y(t). Indeed, plug 2y(t)
into the equation:
a(2y)00 + b(2y)0 + c(2y) = 2 ay 00 + by 0 + cy = 0 .
The same argument shows that c1 y(t) is a solution for any constant c1 . If
y1 (t) and y2 (t) are two solutions, similar argument will show that y1 (t)+y2 (t)
and y1 (t) y2 (t) are also solutions. More generally, a linear combination of
two solutions, c1 y1 (t) + c2 y2 (t), is also a solution, for any constants c1 and
c2 . Indeed,
a (c1 y1 (t) + c2 y2 (t))00 + b (c1 y1 (t) + c2 y2 (t))0 + c (c1 y1 (t) + c2 y2 (t))
= c1 (ay100 + by10 + cy1 ) + c2 (ay200 + by20 + cy2 ) = 0 .
This is called linear superposition property of solutions. The term homogeneous refers to the right hand side of this equation being zero. Homogeneous
equations possessing linear superposition property are called linear.
We now try to find a solution of the equation (2.1) in the form y = ert ,
where r is a constant to be determined. We have y 0 = rert and y 00 = r 2 ert ,
so that plugging into the equation (2.1) gives

a(r 2 ert ) + b(rert) + cert = ert ar 2 + br + c = 0 .

Dividing by a positive quantity ert , we get
ar 2 + br + c = 0 .
This is a quadratic equation for r, called the characteristic equation. If r is
a root (solution) of this equation, then ert solves our differential equation
(2.1). When solving a quadratic equation, it is possible to encounter two
real roots, one (repeated) real root, or two complex conjugate roots. We
shall look at these cases in turn.
2.2. LINEAR HOMOGENEOUS EQUATIONS WITH CONSTANT COEFFICIENTS51
2.2.1
The Characteristic Equation Has Two Distinct Real

Roots
The roots r1 and r2 are real, and r2 6= r1 . Then er1 t and er2 t are two
solutions, and their linear combination gives us the general solution
y(t) = c1 er1 t + c2 er2 t .
As we have two constants to play with, one can prescribe two additional
conditions for the solution to satisfy.
Example Solve
y 00 + 4y 0 + 3y = 0
y(0) = 2
0
y (0) = 1 .
We prescribe that at time zero the displacement is 2, and the velocity is 1.
These two conditions are usually referred to as the initial conditions, and
together with the differential equation, they form an initial value problem.
The characteristic equation is
r 2 + 4r + 3 = 0 .
Solving it (say by factoring as (r + 1)(r + 3) = 0), we get its roots r1 = 1,
and r2 = 3. The general solution is then
y(t) = c1 et + c2 e3t .
We have y(0) = c1 + c2 . Compute y 0 (t) = c1 et 3c2 e3t , and therefore
y 0 (0) = c1 3c2 . The initial conditions tell us that
c1 + c2 = 2
c1 3c2 = 1 .
We have two equations to find two unknowns c1 and c2 . We find that
c1 = 5/2, and c2 = 1/2 (say by adding the equations).
5
1
Answer: y(t) = et e3t .
2
2
Example Solve
y 00 4y = 0 .
52
The characteristic equation is

r2 4 = 0 .
Its roots are r1 = 2, and r2 = 2. The general solution is then
y(t) = c1 e2t + c2 e2t .
More generally, for the equation
y 00 a2 y = 0
(a is a given constant),
the general solution is

y(t) = c1 eat + c2 eat .
This should become automatic, because such equations appear often.
Example Find the constant a, so that the solution of the initial value
problem
9y 00 y = 0, y(0) = 2, y 0 (0) = a
is bounded as t , and find that solution.
We begin by writing down (automatically!) the general solution
1
y(t) = c1 e 3 t + c2 e 3 t .
1
1
1
1
Compute y 0 (t) = c1 e 3 t + c2 e 3 t , and then the initial conditions give
3
3
y(0) = c1 + c2 = 2
y 0 (0) = 13 c1 + 31 c2 = a .
Solving this system of two equation for c1 and c2 (by multiplying the second
equation through by 3, and adding the result to the first equation), we get
c2 = 1 + 32 a, and c1 = 1 32 a. The solution is
1
1
3
3
y(t) = 1 a e 3 t + 1 + a e 3 t .
2
2
In order for this solution to stay bounded as t , the coefficient in front

1
of e 3 t must be zero. So that 1 + 32 a = 0, and a = 23 . The solution then
1
becomes y(t) = 2e 3 t .
Finally, recall that if r1 and r2 are roots of the characteristic equation,
then we can factor the characteristic polynomial as
(2.2)
ar 2 + br + c = a(r r1 )(r r2 ) .
2.2. LINEAR HOMOGENEOUS EQUATIONS WITH CONSTANT COEFFICIENTS53
2.2.2
The Characteristic Equation Has Only One (Repeated)

Real Root
This is the case when r2 = r1 . We still have a solution y1 (t) = er1 t . Of

course, any constant multiple of this function is also a solution, but to form
the general solution we need another truly different solution, as we saw in
the preceding case. It turns out that y2 (t) = ter1 t is that second solution,
and the general solution is then
y(t) = c1 er1 t + c2 ter1 t .
To justify that y2 (t) = ter1 t is a solution, we begin by observing that in
this case the formula (2.2) becomes
ar 2 + br + c = a(r r1 )2 .
Square out the quadratic on the right as ar 2 2ar1 r + ar12 . Because it is
equal to the quadratic on the left, the coefficients of both polynomials in r 2 ,
r, and the constant terms are the same. We equate the coefficients in r:
(2.3)
b = 2ar1 .
0
r1 t
To plug y2 (t) into the equation, we compute itsderivatives
y2 (t) = e +
r1 ter1 t = er1 t (1 + r1 t), and similarly y200 (t) = er1 t 2r1 + r12 t . Then
ay200 + by20 + cy2 = aer1 t 2r1 + r12 t + ber1 t (1 + r1 t) + cter1 t

= er1 t (2ar1 + b) + ter1 t ar12 + br1 + c = 0 .

In the last line, the first bracket is zero because of (2.3), and the second
bracket is zero because r1 solves the characteristic equation.
Example 9y 00 + 6y 0 + y = 0.
The characteristic equation
9r 2 + 6r + 1 = 0
has a double root r = 13 . The general solution is then
1
y(t) = c1 e 3 t + c2 te 3 t .
Example Solve
y 00 4y 0 + 4y = 0
y(0) = 1, y 0 (0) = 2 .
54

r 2 4r + 4 = 0
has a double root r = 2. The general solution is then
y(t) = c1 e2t + c2 te2t .
Here y 0 (t) = 2c1 e2t + c2 e2t + 2c2 te2t , and from the initial conditions
y(0) = c1 = 1
y 0 (0) = 2c1 + c2 = 2 .
From the first equation c1 = 1, and then c2 = 4. Answer: y(t) = e2t 4te2t .
2.3
2.3.1
The Characteristic Equation Has Two Complex Conjugate Roots

Eulers Formula
Recall Maclaurens formula

1 2
1
1
1
z + z3 + z4 + z5 + .
2!
3!
4!
5!
Plug in z = i , where i = 1 is the imaginary unit, and is a real

number. Calculating the powers, and separating the real and imaginary
parts, we have
ez = 1 + z +
1
1
1
1
2
3
4
5
2! (i ) + 3! (i ) + 4! (i ) + 5! (i ) +
1 2
1
1 4
1
= 1 + i 2!
3!
i 3 + 4!
+ 5!
i 5 +

1 4
1 3
1 5
1 2
1 2!
+ 4!
+ + i 3!
+ 5!
+
ei = 1 + i +
=
= cos + i sin .
We derived Eulers formula:

(3.1)
ei = cos + i sin .
Replacing by , we have
(3.2)
ei = cos() + i sin() = cos i sin .
2.3. THE CHARACTERISTIC EQUATION HAS TWO COMPLEX CONJUGATE ROOTS55

Adding the last two formulas, we express
(3.3)
cos =
ei + ei
.
2
Subtracting from (3.1) the formula (3.2), and dividing by 2i

(3.4)
2.3.2
sin =
ei ei
.
2i
The General Solution
Recall that to solve the equation

ay 00 + by 0 + cy = 0 ,
we begin with solving the characteristic equation
ar 2 + br + c = 0 .
Complex roots come in conjugate pairs: if p + iq is one root, then p iq
is the other, and we may assume that q > 0. These roots are, of course,
different, so that we have two solutions z1 = e(p+iq)t , and z2 = e(piq)t . The
problem with these solutions is that they are complex-valued. If we add
z1 + z2 , we get another solution. If we divide this new solution by 2, we get
z1 + z2
yet another solution. So that the function y1 (t) =
is a solution of
2
z1 z2
our equation, and similarly the function y2 (t) =
is another solution.
2i
Using the formula (3.3), compute
y1 (t) =
e(p+iq)t + e(piq)t
eiqt + eiqt
= ept
= ept cos qt .
2
2
This is a real valued solution of our equation! Similarly,

y2 (t) =
e(p+iq)t e(piq)t
eiqt eiqt
= ept
= ept sin qt
2i
2i
is our second solution. The general solution is then

y(t) = c1 ept cos qt + c2 ept sin qt .
Example Solve y 00 + 4y 0 + 5y = 0.
56

r 2 + 4r + 5 = 0
can be solved quickly by completing the square:
(r + 2)2 + 1 = 0; (r + 2)2 = 1;
r + 2 = i; r = 2 i .
Here p = 2, q = 1, and the general solution is
y(t) = c1 e2t cos t + c2 e2t sin t .
Example Solve y 00 + y = 0.
r2 + 1 = 0
has roots i. Here p = 0 and q = 1, and the general solution is y(t) =
c1 cos t + c2 sin t. More generally, for the equation
y 00 + a2 y = 0
(a is a given constant),
the general solution is

y(t) = c1 cos at + c2 sin at .
This should become automatic, because such equations appear often.
Example Solve
y 00 + 4y = 0,
y(/3) = 2, y 0 (/3) = 4 .
The general solution is

y(t) = c1 cos 2t + c2 sin 2t .
Compute y 0 (t) = 2c1 sin 2t + 2c2 cos 2t. From the initial conditions
2
2
1
3
y(/3) = c1 cos
+ c2 sin
= c1 +
c2 = 2
3
3
2
2
2
2
+ 2c2 cos
= 3c1 c2 = 4 .
y 0 (/3) = 2c1 sin
3
3
This gives c1 = 3 1, c2 = 3 + 1. Answer:
y(t) = ( 3 1) cos 2t + ( 3 + 1) sin 2t .
2.3. THE CHARACTERISTIC EQUATION HAS TWO COMPLEX CONJUGATE ROOTS57
2.3.3
Problems
I. Solve the second order equations, with y missing.

2
1. 2y 0 y 00 = 1.
Answer. y = (x + c1 )3/2 + c2 .
3
x2
+ c1 ln x + c2 .
2. xy 00 + y 0 = x.
Answer. y =
4
x3
3. y 00 + y 0 = x2 .
Answer. y =
x2 + 2x + c1 ex + c2 .
3
4. xy 00 + 2y 0 = (y 0 )2 , y(1) = 0, y 0 (1) = 1.
Answer. y = 2 tan1 x
.
2
II. Solve the second order equations, with x missing.

1. yy 00 3 (y 0)3 = 0.
00
0 2
2. yy + (y ) = 0.
Answer. 3 (y ln y y) + c1 y = x + c2 , and y = c.
Answer. y 2 = c1 + c2 x (this includes the y = c family).
3. y 00 = 2yy 0 , y(0) = 0, y 0 (0) = 1.
Answer. y = tan x.
4 . y 00 y = 2xy 0 , y(0) = 1, y 0 (0) = 4.

Hint: Write:
00
y yy
02
02
= (2x 1)y ;
y 00 y y 0 2
= 2x 1;
y02
y
0
y

0
= 2x 1 .
Integrating, and using the initial conditions
y
1
(2x 1)2
2
=
x
x
+
=
.
y0
4
4
4x
Answer. y = e 2x1 .
III. Solve the linear second order equations, with constant coefficients.
1. y 00 + 4y 0 + 3y = 0.
2. y 00 3y 0 = 0.
3. 2y 00 + y 0 y = 0.
4. y 00 3y = 0.
5. 3y 00 5y 0 2y = 0.
Answer. y = c1 et + c2 e3t .
Answer. y = c1 + c2 e3t.
1
Answer. y = c1 et + c2 e 2 t .
Answer. y = c1 e
3t
3t
+ c2 e
58
6. y 00 9y = 0, y(0) = 3, y 0 (0) = 3.
Answer. y = e3t + 2e3t.
7. y 00 + 5y 0 = 0, y(0) = 1, y 0 (0) = 10.
8. y 00 + y 0 6y = 0, y(0) = 2, y 0 (0) = 3.
Answer. y = 3 + 2e5t .
7
3e2t
Answer. y = e3t
.
5
5
9. 4y 00 y = 0.
Answer. y = 3et/3 2et .
10. 3y 00 2y 0 y = 0, y(0) = 1, y 0 (0) = 3.
11. 3y 00 2y 0 y = 0, y(0) = 1, y 0 (0) = a. Then find the value of a for

1
which the solution is bounded, as t .
Answer. a = .
3
IV. Solve the linear second order equations, with constant coefficients.
1. y 00 + 6y 0 + 9y = 0.
Answer. y = c1 e3t + c2 te3t .
2. 4y 00 4y 0 + y = 0.
Answer. y = c1 e 2 t + c2 te 2 t.
3. y 00 2y 0 + y = 0, y(0) = 0, y 0 (0) = 2.
Answer. y = 2tet .
V. Using Eulers
formula, compute: 1. ei
9
2i
4. e
5. 2ei 4
2. ei/2
4. 9y 00 6y 0 + y = 0, y(0) = 1, y 0 (0) = 2.
3. ei
3
4
6. Show that sin 2 = 2 sin cos , and cos2 = cos2 sin2 .
Hint: Begin with ei2 = (cos + i sin )2 . Apply Eulers formula on the left,
and square out on the right. Then equate the real and imaginary parts.
7 . Show that
sin 3 = 3 cos2 sin sin3 , and cos 3 = 3 sin2 cos + cos3 .
Hint: Begin with ei3 = (cos + i sin )3 . Apply Eulers formula on the left,
and cube out on the right. Then equate the real and imaginary parts.
VI. Solve the linear second order equations, with constant coefficients.
1. y 00 + 4y 0 + 8y = 0.
2. y 00 + 16y = 0.
Answer. y = c1 e2t cos 2t + c2 e2t sin 2t.

Answer. y = c1 cos 4t + c2 sin 4t.
3. y 00 4y 0 +5y = 0, y(0) = 1, y 0 (0) = 2.
Answer. y = e2t cos t4e2t sin t.
2.4. LINEAR SECOND ORDER EQUATIONS WITH VARIABLE COEFFICIENTS59

4. y 00 + 4y = 0, y(0) = 2, y 0 (0) = 0.
Answer. y = 2 cos 2t.

1
5. 9y 00 + y = 0, y(0) = 0, y 0 (0) = 5.
Answer. y = 15 sin t.
3
!
t
3
3
00
0
6. y y + y = 0.
Answer. y = e 2 c1 cos
t + c2 sin
t .
2
2
1
Answer. y = 8et cos t.
2
Answer. y = sin(t /4).
7. 4y 00 + 8y 0 + 5y = 0, y() = 0, y 0 () = 4.
8. y 00 + y = 0, y(/4) = 0, y 0 (/4) = 1.
VII.
1. Consider the equation (y = y(t))
y 00 + by 0 + cy = 0 ,
with positive constants b and c. Show that all of its solutions tend to zero,
as t .
2. Consider the equation
y 00 + by 0 cy = 0 ,
with positive constants b and c. Assume that some solution is bounded, as
t . Show that this solution tends to zero, as t .
2.4
Linear Second Order Equations with Variable

Coefficients
Linear Systems
Recall that a system of two equations (here the numbers a, b, c, d, g and h
are given, while x and y are unknowns)
ax+ by = g
cx+dy = h
has
a unique
solution if and only if the determinant of the system is non-zero,

a b

= ad bc 6= 0. This is justified by explicitly solving the system:
c d
x=
dg bh
,
ad bc
y=
ah cg
.
ad bc
60

a b

It is also easy to justify that a determinant is zero,
c d

= 0, if and only

if its columns are proportional, so that a = b and c = d, for some constant

.
General Theory
We consider an initial value problem for linear second order equations
(4.1)
y 00 + p(t)y 0 + g(t)y = f (t)

y(t0 ) =
y 0 (t0 ) = .
The coefficient functions p(t) and g(t), and the function f (t) are assumed to
be given. The constants t0 , and are also given, so that at some initial
time t = t0 , the values of the solution and its derivative are prescribed. It
is natural to ask the following questions. Is there a solution to this problem?
If so, is the solution unique, and how far can it be continued?
Theorem 2.4.1 Assume that the functions p(t), g(t) and f (t) are continuous on some interval (a, b) that includes t0 . Then the problem (4.1) has a
solution, and only one solution. This solution can be continued to the left
and to the right of the initial point t0 , so long as t remains in (a, b).
If the functions p(t), g(t) and f (t) are continuous for all t, then the
solution can be continued for all t, < t < . This is better than what
we had for first order equations (where blow up in finite time was possible).
Why? Because the equation here is linear. Linearity pays!
Corollary 2.4.1 Let z(t) be a solution of (4.1) with the same initial data
as y(t): z(t0 ) = and z 0 (t0 ) = . Then z(t) = y(t) for all t.
Let us now study the homogeneous equation (for y = y(t))
(4.2)
y 00 + p(t)y 0 + g(t)y = 0 ,
with given coefficient functions p(t) and g(t). Although this equation looks
relatively simple, its analytical solution is totally out of reach, in general.
(One has to either solve it numerically, or use infinite series.) In this section
we shall study some theoretical aspects. In particular, we shall prove that
a linear combination of two solutions, which are not constant multiples of
2.4. LINEAR SECOND ORDER EQUATIONS WITH VARIABLE COEFFICIENTS61

one another, gives the general solution (a fact that we had intuitively used
for equations with constant coefficients).
We shall need a concept of the Wronskian determinant of two functions
y1 (t) and y2 (t), or the Wronskian, for short:

y (t) y (t)

2
W (t) = 10
y1 (t) y20 (t)

= y1 (t)y20 (t) y10 (t)y2 (t) .

(Named in honor of Polish mathematician J.M. Wronski, 1776-1853.) Sometimes the Wronskian is written as W (y1 , y2 )(t) to stress its dependence on
y1 (t) and y2 (t). For example,

sin 2t
cos 2t
W (cos 2t, sin 2t)(t) =
2 sin 2t 2 cos 2t

= 2 cos2 2t + 2 sin2 2t = 2 .

Given the Wronskian and one of the functions, one can determine the
other one.
Example If f (t) = t, and W (f, g)(t) = t2 et , find g(t).
Solution: Here f 0 (t) = 1, and so

t g(t)
W (f, g)(t) =
1 g 0 (t)

= tg 0 (t) g(t) = t2 et .

This is a linear first order equation for g(t). We solve it as usual, obtaining
g(t) = tet + ct .
If g(t) = cf (t), with some constant c, we compute that W (f, g)(t) = 0,
for all t. The converse statement is not true. For example, the functions
f (t) = t2 and
(
t2
if t 0
g(t) =
t2 if t < 0
are not constant multiples of one another, but W (f, g)(t) = 0. This is
seen by computing the Wronskian separately in case t 0, and for t < 0.
(Observe that g(t) is a differentiable function, with g 0 (0) = 0.)
Theorem 2.4.2 Let y1 (t) and y2 (t) be two solutions of (4.2), and W (t) is
their Wronskian. Then
R
(4.3)
W (t) = ce p(t) dt .
where c is some constant.
62
This is a remarkable fact. Even though we do not know y1 (t) and y2 (t),
we can compute their Wronskian.
Proof: We differentiate the Wronskian W (t) = y1 (t)y20 (t) y10 (t)y2 (t):
W 0 = y1 y200 + y10 y20 y10 y20 y100 y2 = y1 y200 y100 y2 .
Because y1 is a solution of (4.2), we have y100 + p(t)y10 + g(t)y1 = 0, or
y100 = p(t)y10 g(t)y1 , and similarly y200 = p(t)y20 g(t)y2 . With these
formulas, we continue
W 0 = y1 (p(t)y20 g(t)y2 ) (p(t)y10 g(t)y1 ) y2
= p(t) (y1 y20 y10 y2 ) = p(t)W .
We obtained a linear first order equation for W (t). Solving it, we conclude
(4.3).
Corollary 2.4.2 We see from (4.3) that either W (t) = 0 for all t, when
c = 0, or else W (t) is never zero, in case c 6= 0.
Theorem 2.4.3 Let y1 (t) and y2 (t) be two non-trivial solutions of (4.2),
and W (t) is their Wronskian. Then W (t) = 0, if and only if y1 (t) and y2 (t)
are constant multiples of each other.
We just saw that if two functions are constant multiples of each other,
then their Wronskian is zero, while the converse statement is not true, in
general. But if these functions happen to be solutions of (4.2), then the
converse statement is true.
Proof: Assume that the Wronskian of two solutions y1 (t) and y2 (t) is zero.
In particular it is zero at any point t0 , so that

y1 (t0 ) y2 (t0 )
0
y1 (t0 ) y20 (t0 )

= 0.

When a 2 2 determinant is zero, its columns are proportional. Let us

assume the second column is equal to times the first one, where is some
number, so that y2 (t0 ) = y1 (t0 ) and y20 (t0 ) = y10 (t0 ). We may assume
that 6= 0, because otherwise we would have y2 (t0 ) = y20 (t0 ) = 0, and then
y2 (t) would be a trivial solution, contrary to our assumptions. Consider
the function z(t) = y2 (t)/. This function is a solution of the homogeneous
equation (4.2), and it has initial values z(t0 ) = y1 (t0 ) and z 0 (t0 ) = y10 (t0 ),
2.5. SOME APPLICATIONS OF THE THEORY
63
the same as y1 (t). By Corollary 2.4.1, it follows that z(t) = y1 (t), so that
y2 (t) = y1 (t), for all t.
Definition We say that two solutions y1 (t) and y2 (t) of (4.2) form a fundamental set, if for any other solution z(t), we can find two constants c01 and
c02 , so that z(t) = c01 y1 (t) + c02 y2 (t). In other words, the linear combination
c1 y1 (t) + c2 y2 (t) gives us all solutions of (4.2).
Theorem 2.4.4 Let y1 (t) and y2 (t) be two solutions of (4.2), that are not
constant multiples of one another. Then they form a fundamental set.
Proof: Let y(t) be a solution of the equation (4.2). Let us try to find the
constants c1 and c2 , so that z(t) = c1 y1 (t) + c2 y2 (t) satisfies the same initial
conditions as y(t), so that
(4.4)
z(t0 ) = c1 y1 (t0 ) + c2 y2 (t0 ) = y(t0 )

z 0 (t0 ) = c1 y10 (t0 ) + c2 y20 (t0 ) = y 0 (t0 ) .
This is a system of two linear equations to find c1 and c2 . The determinant

of this system is just the Wronskian of y1 (t) and y2 (t), evaluated at t0 . This
determinant is not zero, because y1 (t) and y2 (t) are not constant multiples of
one another. (This determinant is W (t0 ). If W (t0 ) = 0, then W (t) = 0 for
all t, by the Corollary 2.4.2, and then by the Theorem 2.4.3, y1 (t) and y2 (t)
would have to be constant multiples of one another.) It follows that the 22
system (4.4) has a unique solution c1 = c01 , c2 = c02 . The function z(t) =
c01 y1 (t) + c02 y2 (t) is then a solution of the same equation (4.2), satisfying the
same initial conditions, as does y(t). By the Corollary 2.4.1, y(t) = z(t) for
all t. So that any solution y(t) is a particular case of the general solution
c1 y1 (t) + c2 y2 (t).
Finally, we mention that two functions are called linearly independent, if

they are not constant multiples of one another. So that two solutions y1 (t)
and y2 (t) form a fundamental set, if and only if they are linearly independent.
2.5
Some Applications of the Theory
We shall give some practical applications of the theory from the last section.
But first, we recall the functions sinh t and cosh t.
64

y
15
10
y = cosh t
5
-3
-2
-1
Figure 2.1: The cosine hyperbolic function
2.5.1
The Hyperbolic Sine and Cosine Functions
One defines
cosh t =
et + et
,
2
and
sinh t =
et et
.
2
In particular, cosh 0 = 1, sinh 0 = 0. Observe that cosh t is an even function,

while sinh t is odd. Compute:
d
cosh t = sinh t
dt
and
d
sinh t = cosh t .
dt
These formulas are similar to those for cosine and sine. By squaring out,
one sees that
cosh2 t sinh2 t = 1, for all t.
(There are other similar formulas.) We see that the derivatives, and the
algebraic properties of the new functions are similar to those for cosine and
sine. However, the graphs of sinh t and cosh t look totally different: they are
not periodic, and they are unbounded, see Figures 2.1 and 2.2.
2.5.2
Different Ways to Write the General Solution
For the equation

(5.1)
y 00 a2 y = 0
you remember that the functions eat and eat form a fundamental set, and
y(t) = c1 eat + c2 eat is the general solution. But y = sinh at is also a
solution, because y 0 = a cosh at and y 00 = a2 sinh at = a2 y. Similarly, cosh at
65

y
15
10
5
-3
-2
y = sinh t
1
-1
-5
-10
-15
Figure 2.2: The sine hyperbolic function
is a solution. It is not a constant multiple of sinh at, so that together they

form another fundamental set, and we have another form of the general
solution of (5.1)
y = c1 cosh at + c2 sinh at .
This is not a new general solution, as it can be reduced to the old one, by
expressing cosh at and sinh at through the exponentials. However, the new
form is useful.
Example Solve: y 00 4y = 0, y(0) = 0, y 0 (0) = 5.
We write the general solution as y(t) = c1 cosh 2t + c2 sinh 2t. Using

that cosh 0 = 1 and sinh 0 = 0, we have y(0) = c1 = 0. With c1 = 0, we
update the solution: y(t) = c2 sinh 2t. Compute y 0 (t) = 2c2 cosh 2t, and
y 0 (0) = 2c2 = 5, giving c2 = 52 . Answer: y(t) = 52 sinh 2t.
Yet another form of the general solution of (5.1) is
y(t) = c1 ea(tt0 ) + c2 ea(tt0 ) ,
where t0 is any number.
Example Solve: y 00 9y = 0, y(2) = 1, y 0 (2) = 9.
We select t0 = 2, writing the general solution as
y(t) = c1 e3(t2) + c2 e3(t2) .
Compute y 0 (t) = 3c1 e3(t2) + 3c2 e3(t2), and use the initial conditions
c1 + c2 = 1
3c1 + 3c2 = 9 .
66
We see that c1 = 2, and c2 = 1. Answer: y(t) = 2e3(t2) + e3(t2).

We can write general solutions, centered at t0 , for other simple equations
as well. For the equation
y 00 + a2 y = 0
the functions cos a(tt0 ) and sin a(tt0 ) are both solutions, for any value of
t0 , and they form a fundamental set (because they are not constant multiples
of one another). We can then write the general solution as
y = c1 cos a(t t0 ) + c2 sin a(t t0 ) .
Example Solve: y 00 + 4y = 0, y( 5 ) = 2, y 0 ( 5 ) = 6.
We write the general solution as
y = c1 cos 2(t
) + c2 sin 2(t ) .
5
5
Using the initial conditions, we quickly compute c1 = 2, and c2 = 3.
Answer: y = 2 cos 2(t ) 3 sin 2(t ).

5
5
2.5.3
Finding the Second Solution
Next we solve Legendres equation (for 1 < t < 1)

(1 t2 )y 00 2ty 0 + 2y = 0 .
It has a solution y = t. (Lucky!) We need to find the other solution in the
fundamental set. We divide the equation by 1 t2 , to put it into the form
(4.2) from the preceding section:
2t
2
y0 +
y = 0.
2
1t
1 t2
y 00
Call the other solution y(t). By the Theorem 2.4.2, we can calculate the
Wronskian of two solutions:

t y(t)

W (t, y) =
1 y 0 (t)

R

= ce

2t
1t2
dt
We shall set here c = 1, because we need just one solution, which is not a
constant multiple of t. Then
0
ty y = e
2t
1t2
dt
= e ln(1t ) =
1
.
1 t2
67

This is a linear equation, which we solve as usual:
R
1
1
1
1/t dt
;
(t)
=
e
= e ln t = ;
y0 y =
2
t
t(1 t )
t
d 1
1
1
y = 2
; y=
2
dt t
t (1 t ) t

1
t2 (1 t2 )
dt .
We computed the last integral before, by the guess-and-check method, so

that
Z
1
1
1
y=t
dt = 1 t ln(1 t) + t ln(1 + t) .
2
2
t (1 t )
2
2
Again, we took the constant of integration c = 0, because we need just

one solution, which is not a constant multiple of t. Answer: y(t) = c1 t +
1+t
1
.
c2 1 + t ln
2
1t
2.5.4
Problems
I. 1. Find the Wronskians of the following functions.

1
7 5
(i) f (t) = e3t , g(t) = e 2 t .
Answer. e 2 t .
2
2t
2t
4t
(ii) f (t) = e , g(t) = te .
Answer. e .
(iii) f (t) = et cos 3t, g(t) = et sin 3t.
Answer. 3e2t.
2. If f (t) = t2 , and the Wronskian W (f, g)(t) = t5 et , find g(t).

Answer. g(t) = t3 et t2 et + ct2 .
3. Let y1 (t) and y2 (t) be any two solutions of
y 00 t2 y = 0 .
Show that W (y1 (t), y2 (t))(t) = constant.
II. For the following equations one solution is given. Using Wronskians, find
the second solution, and the general solution.
y1 (t) = et .
1. y 00 2y 0 + y = 0,
2. (t 2)y 00 ty 0 + 2y = 0,
y1 (t) = et .
Answer. y = c1 et + c2 t2 + 2t 2 .
3. ty 00 + 2y 0 + ty = 0,
y1 (t) =
sin t
t .
Answer. y = c1
sin t
cos t
+ c2
.
t
t
68
III. Express the solution, by using the hyperbolic sine and cosine functions.
1
Answer. y = sinh 2t.
6
1. y 00 4y = 0, y(0) = 0, y 0 (0) = 13 .
2. y 00 9y = 0, y(0) = 2, y 0 (0) = 0.
Answer. y = 2 cosh 3t.
3. y 00 y = 0, y(0) = 3, y 0 (0) = 5.
Answer. y = 3 cosh t + 5 sinh t.
IV. Solve the problem, by using the general solution centered at the initial
point.
1. y 00 + y = 0, y(/8) = 0, y 0 (/8) = 3.
Answer. y = 3 sin(t /8).
2. y 00 + 4y = 0, y(/4) = 0, y 0 (/4) = 4.
Answer. y = 2 sin 2(t /4) = 2 sin(2t /2) = 2 cos 2t.
3. y 00 2y 0 3y = 0, y(1) = 1, y 0 (1) = 7.
Answer. y = 2e3(t1) e(t1) .
V. 1. Show that the functions y1 = t, and y2 = sin t, cannot be both

solutions of
y 00 + p(t)y 0 + g(t)y = 0 ,
no matter what p(t) and g(t) are.
Hint: Consider W (y1 , y2 )(0), then the Corollary 2.4.2, and the Theorem
2.4.3.
2. Assume that the functions y1 = 1, and y2 = cos t, are both solutions of
some equation
y 00 + p(t)y 0 + g(t)y = 0 .
What is the equation?
Hint: Use the Theorem 2.4.2 to determine p(t), then argue that g(t) must
be zero. Answer. The equation is
y 00 cot ty 0 = 0 .
2.6
Non-homogeneous Equations.
We wish to solve the non-homogeneous equation

(6.1)
y 00 + p(t)y 0 + g(t)y = f (t) .
2.6. NON-HOMOGENEOUS EQUATIONS.
69
Here the coefficient functions p(t) and g(t), and the function f (t) are given.
The corresponding homogeneous equation is
(6.2)
y 00 + p(t)y 0 + g(t)y = 0 .
Assume that we know the general solution of the homogeneous equation:

y(t) = c1 y1 (t) + c2 y2 (t). Our goal is to find the general solution of the
non-homogeneous equation. Suppose that we can find somehow a particular
solution Y (t) of the non-homogeneous equation, so that
(6.3)
Y 00 + p(t)Y 0 + g(t)Y = f (t) .
From the equation (6.1) we subtract the equation (6.3):

(y Y )00 + p(t)(y Y )0 + g(t)(y Y ) = 0 .
We see that the function v = y Y is a solution the homogeneous equation
(6.2), and then v(t) = c1 y1 (t) + c2 y2 (t), for some constants c1 and c2 . We
express y = Y + v, giving
y(t) = Y (t) + c1 y1 (t) + c2 y2 (t) .
In words: the general solution of the non-homogeneous equation is equal to
the sum of any particular solution of the non-homogeneous equation, and
the general solution of the corresponding homogeneous equation.
Example Solve y 00 + 9y = 4 cos 2t.
The general solution of the corresponding homogeneous equation
y 00 + 9y = 0
is y(t) = c1 cos 3t + c2 sin 3t. We look for a particular solution in the form
Y (t) = A cos 2t. Plug this in, then simplify:
4A cos 2t + 9A cos 2t = 4 cos 2t;
5A cos 2t = 4 cos 2t ,
4
giving A = 54 , and Y (t) = 54 cos 2t. Answer: y(t) = cos 2t+c1 cos 3t+
5
c2 sin 3t.
This was an easy example, because the y 0 term was missing. If y 0 term
is present, we need to look for Y (t) in the form Y (t) = A cos 2t + B sin 2t.
70
Prescription 1 If the right side of the equation (6.1) has the form a cos t+
b sin t, with constants a, b and , then look for a particular solution in the
form Y (t) = A cos t + B sin t. More generally, if the right side of the
equation has the form (at2 + bt + c) cos t + (et2 + gt + h) sin t, then look
for a particular solution in the form Y (t) = (At2 + Bt + C) cos t + (Et2 +
Gt + H) sin t. Even more generally, if the polynomials are of higher power,
we make the corresponding adjustments.
Example Solve y 00 + 3y 0 + 9y = 4 cos 2t + sin 2t.
We look for a particular solution in the form Y (t) = A cos 2t + B sin 2t.
Plug this in, then combine the like terms
4A cos 2t 4B sin 2t + 3 (2A sin 2t + 2B cos 2t) + 9 (A cos 2t + B sin 2t)
= 4 cos 2t + sin 2t;
(5A + 6B) cos 2t + (6A + 5B) sin 2t = 4 cos 2t + sin 2t .

Equating the corresponding coefficients
5A + 6B = 4
6A + 5B = 1 .
19
26
Solving this system, A = 26
61 and B = 61 , so that Y (t) = 61 cos 2t
19
61 sin 2t. The general solution of the corresponding homogeneous equation
y 00 + 3y 0 + 9y = 0
3
3
26
is y = c1 e 2 t cos 3 2 3 t + c2 e 2 t sin 3 2 3 t. Answer: y(t) = cos 2t
61
3 3
3 3
19
32 t
32 t
cos
sin
sin 2t + c1 e
t + c2 e
t.
61
2
2
Example Solve y 00 + 2y 0 + y = t 1.
On the right we see a linear polynomial. We look for particular solution
in the form Y (t) = At + B. Plugging this in, gives
2A + At + B = t 1 .
Equating the corresponding coefficients, we have A = 1, and 2A+B = 1, so
that B = 3. Then, Y (t) = t 3. The general solution of the corresponding
homogeneous equation
y 00 + 2y 0 + y = 0
2.6. NON-HOMOGENEOUS EQUATIONS.
71
is y = c1 et + c2 tet . Answer: y(t) = t 3 + c1 et + c2 tet .

Example Solve y 00 + y 0 2y = t2 .
On the right we have a quadratic polynomial (two of its coefficients
happened to be zero). We look for a particular solution as a quadratic
Y (t) = At2 + Bt + C. Plugging this in, gives
2A + 2At + B 2(At2 + Bt + C) = t2 .
Equating the coefficients in t2 , t, and the constant terms, we have
2A = 1
2A 2B = 0
2A + B 2C = 0 .
From the first equation, A = 12 , from the second one, B = 21 , and from
the third, C = A + 12 B = 34 . So that Y (t) = 12 t2 12 t 43 . The general
solution of the corresponding homogeneous equation is y(t) = c1 e2t + c2 et .
1
1
3
Answer: y(t) = t2 t + c1 e2t + c2 et .
2
2
4
The last two examples lead to the following prescription.
Prescription 2 If the right side of the equation (6.1) is a polynomial of
degree n: a0 tn + a1 tn1 + + an1 t + an , look for a particular solution
as a polynomial of degree n: A0 tn + A1 tn1 + + An1 t + An , with the
coefficients to be determined.
And on to the final possibility.
Prescription 3 If the right side of the equation (6.1) is a polynomial
of de
gree n, times an exponential : a0 tn + a1 tn1 + + an1 t + an et, look
for a particular solution as a polynomial of degree n times the same exponential: A0 tn + A1 tn1 + + An1 t + An et , with the coefficients to be
determined.
Example Solve y 00 + y = te2t .
We look for a particular solution in the form Y (t) = (At + B)e2t . Compute
Y 0 (t) = Ae2t 2(At + B)e2t = (2At + A 2B)e2t , Y 00 (t) = 2Ae2t
2(2At + A 2B)e2t = (4At 4A + 4B)e2t . Plug Y (t) in:
4Ate2t 4Ae2t + 4Be2t + Ate2t + Be2t = te2t .
72
Equate the coefficients in te2t , and in e2t

5A = 1
4A + 5B = 0 ,
which gives A = 1/5 and B = 4/25, so that Y (t) = ( 51 t +
1
4
y(t) = ( t + )e2t + c1 cos t + c2 sin t.
5
25
4
2t
.
25 )e
Answer:
Example Solve y 00 4y = t2 + 3et , y(0) = 0, y 0 (0) = 2.

Reasoning as before, we can find Y (t) as a sum of two pieces, Y (t) = Y1 (t) +
Y2 (t), where Y1 (t) is any particular solution of
y 00 4y = t2 ,
and Y2 (t) is any particular solution of
y 00 4y = 3et .
Using our prescriptions, Y1 (t) = 1/4t2 1/8, and Y2 (t) = et . The general
solution is y(t) = 1/4t2 1/8 et + c1 cosh 2t + c2 sinh 2t. We calculate:
c1 = 9/8, and c2 = 3/2.
Answer: y(t) = 1/4t2 1/8 et + 9/8 cosh 2t + 3/2 sinh 2t.
2.7
Non-homogeneous Equations. More on Guessing of Y (t)
The prescriptions from the previous section do not always work. In this
section we sketch a fix. More details can be found in the book of W.E.
Boyce and R.C. DiPrima. A more general approach for finding Y (t) will be
developed in the next section.
Example Solve y 00 + y = sin t.
We try Y (t) = A sin t+B cos t, according to the Prescription 1. Plugging
Y (t) in, we get
0 = sin t ,
which is impossible. Why did we strike out? Because A sin t and B cos t are
solutions of the corresponding homogeneous equation. Let us multiply the
initial guess by t, and try Y = At sin t + Bt cos t. Calculate Y 0 = A sin t +
2.7. NON-HOMOGENEOUS EQUATIONS. MORE ON GUESSING OF Y (T )73

At cos t + B cos t Bt sin t, and Y 00 = 2A cos t At sin t 2B sin t Bt cos t.
Plug Y into our equation, and simplify:
2A cos t At sin t 2B sin t Bt cos t + At sin t + Bt cos t = sin t ;
2A cos t 2B sin t = sin t .
We conclude that A = 0, and B = 1/2, so that Y = 21 t cos t.

Answer: y = 12 t cos t + c1 sin t + c2 cos t.
This example prompts us to change the strategy. We now begin by
solving the corresponding homogeneous equation. The Prescriptions from
the previous section are now the Initial Guesses for the particular solution.
The whole story If any of the functions, appearing in the Initial Guess,
is a solution of the corresponding homogeneous equation, multiply the entire
Initial Guess by t, and look at the new functions. If some of them are still
solutions of the corresponding homogeneous equation, multiply the entire
Initial Guess by t2 . This is guaranteed to work. (Of course, if none of the
functions appearing in the Initial Guess is a solution of the corresponding
homogeneous equation, then the Initial Guess works.)
In the preceding example, the Initial Guess involved the functions sin t
and cos t, both solutions of the corresponding homogeneous equation. After
we multiplied the Initial Guess by t, the new functions t sin t and t cos t are
not solutions of the corresponding homogeneous equation, and the new guess
worked.
Example Solve y 00 + 4y 0 = 2t 5.
The fundamental set of the corresponding homogeneous equation
y 00 + 4y 0 = 0
consists of the functions y1 (t) = 1, and y2 (t) = e4t . The Initial Guess,
according to the Prescription 2, Y (t) = At + B, is a linear combination of
the functions t and 1, and the second of these functions is a solution of the
corresponding homogeneous equation. We multiply the Initial Guess by t,
obtaining Y (t) = t(At + B) = At2 + Bt. This is a linear combination of t2
and t, both of which are not solutions of the corresponding homogeneous
equation. Plugging in Y (t), gives
2A + 4(2At + B) = 2t 5 .
74
Equating the coefficients in t, and the constant terms, we have

8A = 2
2A + 4B = 5 ,
giving A = 1/4, and B = 11/8. The particular solution is Y (t) =
t2
11
Answer: y(t) =
t + c1 + c2 e4t .
4
8
t2
4
11
8 t.
Example Solve y 00 + 2y 0 + y = tet .

The fundamental set of the corresponding homogeneous equation consists of
the functions y1 (t) = et , and y2 (t) = tet . The Initial Guess, according to
the Prescription 3, (At + B)et = Atet + Bet , is a linear combination of
the same two functions. We multiply the Initial Guess by t: t(At + B)et =
At2 et +Btet . The new guess is a linear combination of the functions t2 et
and tet . The first of these functions is not a solution of the corresponding
homogeneous equation, but the second one is. Therefore, we multiply the
Initial Guess by t2 : Y = t2 (At + B)et = At3 et + Bt2 et . It is convenient
to write Y = (At3 + Bt2 )et , and we plug this in:
(6At + 2B)et 2(3At2 + 2Bt)et + (At3 + Bt2 )et

+2 (3At2 + 2Bt)et (At3 + Bt2 )et + (At3 + Bt2 )et = tet .
We divide both sides by et , and simplify:
6At + 2B = t .
It follows that
6A = 1
2B = 0 ,
giving A = 1/6, and B = 0. Answer: y(t) =
2.8
1 3 t
t e + c1 et + c2 tet .
6
The Method of Variation of Parameters
We now present a more general way to find a particular solution of the

non-homogeneous equation
(8.1)
y 00 + p(t)y 0 + g(t)y = f (t) .
2.8. THE METHOD OF VARIATION OF PARAMETERS
75
We assume that y1 (t) and y2 (t) form a fundamental solution set for the
corresponding homogeneous equation
(8.2)
y 00 + p(t)y 0 + g(t)y = 0 .
(So that c1 y1 (t) + c2 y2 (t) is the general solution of (8.2).) We look for a
particular solution of (8.1) in the form
(8.3)
Y (t) = u1 (t)y1 (t) + u2 (t)y2 (t) ,
with some functions u1 (t) and u2 (t), that we shall choose to satisfy the
following two equations
(8.4)
u01 (t)y1 (t) + u02 (t)y2 (t) = 0

u01 (t)y10 (t) + u02 (t)y20 (t) = f (t) .
We have a system of two linear equations to find u01 (t) and u02 (t). Its determinant

y (t) y (t)
1

2
W (t) = 0

y1 (t) y20 (t)
is the Wronskian of y1 (t) and y2 (t). We know that W (t) 6= 0 for all t,
because y1 (t) and y2 (t) form a fundamental solution set. By Cramers rule
(or by elimination), we solve
(8.5)
u01 (t) =
u02 (t) =
f (t)y2 (t)
,
W (t)
f (t)y1 (t)
.
W (t)
We compute u1 (t) and u2 (t) by integration. We shall show that Y (t) =

u1 (t)y1 (t) + u2 (t)y2 (t) is a particular solution of the non-homogeneous equation (8.1). We compute the derivatives of Y (t), in order to plug Y (t) into
our equation (8.1). We have
Y 0 (t) = u01 (t)y1 (t)+u02 (t)y2 (t)+u1 (t)y10 (t)+u2 (t)y20 (t) = u1 (t)y10 (t)+u2 (t)y20 (t).
Here the first two terms disappeared, thanks to the first formula in (8.4).
Next:
Y 00 (t) = u01 (t)y10 (t) + u02 (t)y20 (t) + u1 (t)y100 (t) + u2 (t)y200 (t)
= f (t) + u1 (t)y100 (t) + u2 (t)y200 (t) ,
76
by using the second formula in (8.4). Then

Y 00 + pY 0 + gY = f (t) + u1 y100 + u2 y200 + p(u1 y10 + u2 y20 ) + g(u1 y1 + u2 y2 )
= f (t) + u1 (y100 + py10 + gy1 ) + u2 (y200 + py20 + gy2 )
= f (t) ,
which means that Y (t) is a particular solution of (8.1). (Both brackets are
zero, because y1 (t) and y2 (t) are solutions of the homogeneous equation
(8.2).)
In practice, one begins by writing down the formulas (8.5).
Example y 00 + y = tan t.
The fundamental set of the corresponding homogeneous equation is y1 =
sin t and y2 = cos t. Their Wronskian

sin t
W (t) =
cos t
cos t
sin t
and the formulas (8.5) give

= sin2 t cos2 t = 1 ,

u01 (t) = tan t cos t = sin t

u02 (t) = tan t sin t =
sin2 t
.
cos t
Integrating, u1 (t) = cos t. We set the constant of integration to zero, because we only need one particular solution. Integrating the second formula,
u2 (t) =
sin2 t
dt =
cos t
1 cos2 t
dt =
cos t
( sec t + cos t) dt
= ln | sec t + tan t| + sin t .

We have a particular solution (Y = u1 y1 + u2 y2 )
Y (t) = cos t sin t+( ln | sec t+tan t|+sin t) cos t = cos t ln | sec t+tan t| .
Answer: y(t) = cos t ln | sec t + tan t| + c1 sin t + c2 cos t.
77
2.9. THE CONVOLUTION INTEGRAL
2.9
2.9.1
The Convolution Integral

Differentiation of Integrals
If g(t, s) is a function of two variables, then the integral ab g(t, s) ds depends

on a parameter t (s is a dummy variable). We differentiate this integral as
follows
Z b
Z b
d
g(t, s) ds =
gt (t, s) ds ,
dt a
a
R
where
gt(t, s) denotes the partial derivative in t. To differentiate the integral
Rt
g(s)
ds, we use the fundamental theorem of Calculus:
a
d
dt
g(s) ds = g(t) .
The integral at g(t, s) ds depends on a parameter t, and it has t as its upper

limit. One can show that
R
d
dt
g(t, s) ds =
gt (t, s) ds + g(t, t) ,
so that, in effect, we combine the previous two formulas. Let now z(t) and
f (t) be some functions, then the last formula gives
(9.1)
2.9.2
d
dt
z(t s)f (s) ds =
z 0 (t s)f (s) ds + z(0)f (t) .
Yet Another Way to Compute a Particular Solution
We consider the non-homogeneous equation

(9.2)
y 00 + py 0 + gy = f (t) ,
where p and g are given numbers, and f (t) is a given function. Let z(t)
denote the solution of the corresponding homogeneous equation, satisfying
(9.3)
z 00 + pz 0 + gz = 0, z(0) = 0, z 0 (0) = 1 .
Then we can write a particular solution of (9.2) as a convolution integral

(9.4)
Y (t) =
z(t s)f (s) ds .
78
To justify this formula, we compute the derivatives of Y (t), by using the

formula (9.1), and the initial conditions z(0) = 0 and z 0 (0) = 1:
Y 0 (t) =
Y 00 (t) =
z 0 (t s)f (s) ds + z(0)f (t) =
z 00 (t s)f (s) ds + z 0 (0)f (t) =
z 0 (t s)f (s) ds ;
z 00 (t s)f (s) ds + f (t) .
Then
Y 00 (t) + pY 0 (t) + gY (t)
=
Rt
0
[z 00 (t s) + pz 0 (t s) + gz(t s)] f (s) ds + f (t) = f (t) .
Here the integral is zero, because z(t) satisfies the homogeneous equation
(with constant coefficients p and g) at all values of its argument t, including
t s.
Example Let us now revisit the equation
y 00 + y = tan t .
Solving
z 00 + z = 0, z(0) = 0, z 0 (0) = 1 ,
we get z(t) = sin t. Then
Y (t) =
sin(t s) tan s ds .
Writing sin(ts) = sin t cos scos t sin s, and integrating, it is easy to obtain
the solution we had before.
2.10
Applications of Second Order Equations
One of the main applications of differential equations is to model mechanical

and electrical oscillations.
2.10.1
Vibrating Spring
Let y = y(t) denote the displacement of a spring from its natural position.
Its motion is governed by Newtons second law
ma = f .
79
2.10. APPLICATIONS OF SECOND ORDER EQUATIONS
The acceleration a = y 00 (t). We assume that the only force f , acting on the
spring, is its own restoring force, which by Hookes law is f = ky, for small
displacements. Here the physical constant k > 0 describes the stiffness (or
the hardness) of the spring. We have
my 00 = ky .
2
Divide
p both sides by the mass m of the spring, and denote k/m = (or
= k/m), obtaining
y 00 + 2 y = 0 .
The general solution, y(t) = c1 cos t + c2 sin t, gives us the harmonic

motion of the spring. Let us write the solution y(t) as
y(t) =
c21 + c22 q
c1
c21
+ c22
c2
cos t + q
sin t
2
2
c1 + c2
c1
c2
=A
cos t + sin t
A
A

where we have denoted A = c21 + c22 . Observe that ( cA1 )2 +( cA2 )2 = 1, which
means that we can find an angle , such that cos = cA1 , and sin = cA2 .
Then our solution takes the form
y(t) = A (cos t cos + sin t sin ) = A cos(t ) .
So that any harmonic motion is just a shifted cosine curve of amplitude
q
2
A = c21 + c22 , and of period
. We see that the larger is , the smaller
is the period, and the oscillations are more frequent. So that gives us the
frequency of oscillations. The constants c1 and c2 can be computed, once
the initial displacement y(0), and the initial velocity y 0 (0) are given.
Example Solving the initial value problem
y 00 + 4y = 0, y(0) = 3, y 0 (0) = 8
one gets y(t) = 3 cos 2t 4 sin 2t. This solution is a periodic function, with
the amplitude 5, the frequency 2, and the period .
The equation
(10.1)
y 00 + 2 y = f (t)
80
models the case when an external force, with acceleration equal to f (t), is
applied to the spring. Indeed, the equation of motion is now
my 00 = ky + mf (t) ,
from which we get (10.1), dividing by m.
Let us consider the case of a periodic forcing term
(10.2)
y 00 + 2 y = a sin t ,
where a > 0 is the amplitude of the external force, and is the forcing frequency. If 6= , we look for a particular solution of this non-homogeneous
a
equation in the form Y (t) = A sin t. Plugging this in, gives A = 2
.
2
a
The general solution of (10.2), which is y(t) = 2
sin t + c1 cos t +
2
c2 sin t, is a superposition (sum) of the harmonic motion, and the response
a
term ( 2
sin t) to the external force. We see that the solution is still
2
bounded, although not periodic anymore, in general (such functions are
called quasiperiodic).
A very important case is when = , so that the forcing frequency is
the same as the natural frequency. Then a particular solution has the form
Y (t) = At sin t + Bt cos t, so that solutions become unbounded, as time
t increases. This is the case of resonance, when a bounded external force
produces unbounded response. Large displacements will break the spring.
Resonance is a serious engineering concern.
Example y 00 + 4y = sin 2t, y(0) = 0, y 0 (0) = 1.
Both the natural and forcing frequencies are equal to 2. The fundamental
set of the corresponding homogeneous equation consists of sin 2t and cos 2t.
We search for a particular solution in the form Y (t) = At sin 2t + Bt cos 2t.
As before, we compute Y (t) = 41 t cos 2t. Then the general solution is
y(t) = 14 t cos 2t + c1 cos 2t + c2 sin 2t. Using the initial conditions, we
calculate c1 = 0 and c2 = 58 , so that y(t) = 14 t cos 2t + 58 sin 2t. The term
14 t cos 2t introduces oscillations, with the amplitude 14 t increasing without
bound, as time t . (It is customary to call this unbounded term a
secular term, which seems to imply that the harmonic terms are divine.)
81
y
6
y = - t cos 2t
4
4
2
10
15
20
25
30
-2
-4
-6
Figure 2.3: The graph of the secular term y = 41 t cos 2t, oscillating between
the lines y = 14 t, and y = 14 t
2.10.2
Problems
I. Solve the non-homogeneous equations.

1. 2y 00 3y 0 +y = 2 sin t.
Answer. y = c1 et/2 +c2 et +
3 cos t 1
sin t.
5
5
2. y 00 + 4y 0 + 5y = sin 2t + 3 cos 2t.

23
11
Answer. y =
sin 2t +
cos 2t + c1 e2t cos t + c2 e2t sin t.
65
65
3. y 00 2y 0 + y = t + 2.
4. y 00 +4y = t2 3t+1.
Answer. y = t + 4 + c1 et + c2 et t.

1 2
Answer. y =
2t 6t + 1 +c1 cos 2t+c2 sin 2t.
8
5. y 00 4y = te3t , y(0) = 0, y 0 (0) = 1.
1
13e2t e2t 6e3t
Answer. y = e3t t
+
.
5
50
2
25
6. 4y 00 + 8y 0 + 5y = 2t sin 3t.
2
16
31
24
t
t
Answer. y = t
+
sin 3t +
cos 3t + c1 et cos + c2 et sin .
5
25 1537
1537
2
2
7. y 00 + y = 2e4t + t2 .
2
Answer. y = e4t + t2 2 + c1 cos t + c2 sin t.
17
00
0
8. y y = 2 sin t cos 2t.
1
Answer. y = cos t sin t + cos 2t + sin 2t + c1 + c2 et .
2
82
II. Solve the non-homogeneous equations (using the modified prescriptions).

1. y 00 + y = 2 cos t.
2. y 00 + y 0 6y = e2t .
3. y 00 + 2y 0 + y = 2e3t.
4. y 00 2y 0 + y = tet .
5. y 00 4y 0 = 2 cos t.
6. 2y 00 y 0 y = t + 1.
Answer. y = t sin t + c1 cos t + c2 sin t.

1
Answer. y = e2t t + c1 e3t + c2 e2t .
5
3t
e
Answer. y =
+ c1 et + c2 tet .
8
t3 et
Answer. y =
+ c1 et + c2 tet .
6
t
cos t 4 sin t
+
+ c1 e4t + c2 .
Answer. y = +
2
17
17
1
Answer. y = t + c1 e 2 t + c2 et .
III. Write down the form in which one should look for a particular solution,
but DO NOT compute the coefficients.
1. y 00 4y 0 + 4y = 3te2t + sin 4t t2 .
2. y 00 9y 0 + 9y = e3t + te4t + sin t 4 cos t.

IV. Find a particular solution, by using the method of variation of parameters, and then write down the general solution.
1. y 00 2y 0 + y =
et
.
1 + t2
1
Answer. y = c1 et + c2 tet et ln(1 + t2 ) + tet tan1 t.
2
00
0
t
2. y 2y + y = 2e .
Answer. y = t2 et + c1 et + c2 tet .
et
.
Answer. y = tet + tet ln t + c1 et + c2 tet .
t
4. y 00 + y = sec t.
Answer. y = c1 cos t + c2 sin t + ln(cos t) cos t + t sin t.
1
5. y 00 + 2y 0 + y = 2e3t.
Answer. y = e3t + c1 et + c2 tet .
8
00
0
6. y + 4y = sin 2t, y(0) = 0, y (0) = 1.
1
5
Answer. y = t cos 2t + sin 2t.
4
8
3. y 00 + 2y 0 + y =
V. Verify that the functions y1 (t) and y2 (t) form a fundamental solution
set for the corresponding homogeneous equation, and then use variation of
parameters, to find the general solution.

1. t2 y 00 2y = t3 1.
83
y1 (t) = t2 , y2 (t) = t1 .
Hint: begin by putting this equation into the right form.

Answer. y =
1 1 3
1
+ t + c1 t2 + c2 .
2 4
t
2. ty 00 (1 + t)y 0 + y = t2 e3t.
Answer. y =
y1 (t) = t + 1, y2 (t) = et .
1 3t
e (2t 1) + c1 (t + 1) + c2 et .
12
3 . (3t3 + t)y 00 + 2y 0 6ty = 4 12t2 .
y1 (t) =
1
, y2 (t) = t2 + 1.
t
Hint: Use Mathematica to compute the integrals.

1
Answer. y = 2t + c1 + c2 (t2 + 1).
t
VI.
1. Use the convolution integral, to solve
y 00 + y = t2 , y(0) = 0, y 0 (0) = 1 .
Answer. y = t2 2 + 2 cos t + sin t.
2. Show that y(t) =
equation
(t s)n1
f (s) ds gives a solution of the n-th order
(n 1)!
y (n) = f (t) .
(This formula lets you compute n consecutive antiderivatives at once.)

Hint: Use the formula (9.1).
VII.
1. A spring has natural frequency = 2. Its initial displacement is 1, and
the initial velocity is 2. Find its displacement y(t) at any time t. What is
the amplitude of the oscillations?
Answer. y(t) = sin 2t cos 2t, A = 2.

2. A spring of mass 2 lb is hanging down from the ceiling, and its stiffness
constant is k = 18. Initially, the spring is pushed up 3 inches, and is given
velocity of 2 inch/sec, directed downward. Find the displacement of the
spring y(t) at any time t, and the amplitude of oscillations. (Assume that
the y axis is directed down from the equilibrium position.)
Answer. y(t) =
2
3
sin 3t 3 cos 3t, A =
85
3 .
84
3. A spring has natural frequency = 3. An outside force, with acceleration

f (t) = 2 cos t, is applied to the spring. Here is a constant, 6= 3. Find
the displacement of the spring y(t) at any time t. What happens to the
amplitude of oscillations in case is close to 3?
2
Answer. y(t) =
cos t + c1 cos 3t + c2 sin 3t.
9 2
4. Assume that = 3 in the preceding problem. Find the displacement of

the spring y(t) at any time t. What happens to the spring in the long run?
1
Answer. y(t) = t sin 3t + c1 cos 3t + c2 sin 3t.
3
5. Consider dissipative (or damped) motion of a spring
y 00 + y 0 + 9y = 0 .
Write down the solution, assuming that < 6. What is the smallest value
of the dissipation constant , which will prevent the spring from oscillating?
Answer. No oscillations for 6.
6. Consider forced vibrations of a dissipative spring
y 00 + y 0 + 9y = sin 3t .
Write down the general solution for
(i) = 0
(ii) 6= 0.
What does friction do to the resonance?
2.10.3
A Meteor Approaching the Earth
Let r = r(t) denote the distance of some meteor from the center of the Earth.
The motion of the meteor is governed by Newtons law of gravitation
(10.3)
mr 00 =
mM G
.
r2
Here m is the mass of the meteor, M denotes the mass of the Earth, and
G is the universal gravitational constant. Let a be the radius of the Earth.
If an object is sitting on Earths surface, then r = a, and r 00 = g, so that
from (10.3)
MG
g= 2 .
a
85
Then M G = ga2 , and we can rewrite (10.3) as

r 00 = g
(10.4)
a2
.
r2
We could solve this equation by letting r 0 = v(r), because the independent

variable t is missing. Instead, let us multiply both sides of the equation by
r 0 , and write it as
a2
r 0 r 00 + g 2 r 0 = 0 ;
r
d
dt
(10.5)
1 02
a2
r g
2
r
= 0;
1 02
a2
r (t) g
= c.
2
r(t)
1 2
a2
, remains constant at
So that the energy of the meteor, E(t) = r 0 (t)g
2
r(t)
all time. (That is why the gravitational
force field is called conservative.) We
q
2a2
0
can now express r (t) = 2c + g r(t) , and calculate the motion of meteor
r(t) by separation of variables. However, as we are not riding on the meteor,
this seems to be not worth the effort. What really concerns us is the velocity
of impact, as the meteor hits the Earth.
Let us assume that the meteor begins its journey with zero velocity
r 0 (0) = 0, and at a distance so large that we may assume r(0) = . Then
the energy of the meteor at time t = 0 is zero, E(0) = 0. As the energy
remains constant at all time, the energy at the time of impact is also zero.
At the time of impact, we have r = a, and the velocity of impact we denote
by v (r 0 = v). Then from (10.5)
1 2
a2
v (t) g
= 0,
2
a
and the velocity of impact is
v=
2ga .
Food for thought: the velocity of impact is the same, as it would have been
achieved by free fall from height a.
Let us now revisit the harmonic oscillations of a spring:
y 00 + 2 y = 0 .
86
Similarly to the meteor case, multiply this equation by y 0 :

y 0 y 00 + 2 yy 0 = 0;
d
dt
E(t) =
1 02 1 2 2
y + y = 0;
2
2

1 02 1 2 2
y + y = constant .
2
2
With the energy E(t) conserved, no wonder the motion of the spring was
periodic.
2.10.4
Damped Oscillations
We add an extra term to our model of spring motion:

my 00 = ky k1 y 0 ,
where k1 is another positive constant. We have an additional force, which
is directed in the opposite direction, and is proportional to the velocity of
motion y 0 . This can be either air resistance or friction. Denoting k1 /m =
> 0, and k/m = 2 , we have
y 00 + y 0 + 2 y = 0 .
(10.6)
Let us see what effect the extra term y 0 has on the energy of the spring,
E(t) = 12 y 0 2 + 21 2 y 2 . We differentiate the energy, and express from the
equation (10.6), y 00 = y 0 2 y, obtaining
2
E 0 (t) = y 0 y 00 + 2 yy 0 = y 0 (y 0 2 y) + 2 yy 0 = y 0 .
We see that E 0 (t) 0, and the energy decreases. This is an example of a
dissipative motion. We expect the amplitude of oscillations to decrease with
time. We call the dissipation (or damping) coefficient.
To solve the equation (10.6), we write down its characteristic equation
r 2 + r + 2 y = 0 .
2 4 2
.
2
There are three cases to consider.
The roots are r =
87
2.11. FURTHER APPLICATIONS
(i) 2 4 2 < 0. (The dissipation coefficient is small.) The roots are
q
complex. If we denote 2 4 2 = q 2 , the roots are i . The general
2
2
solution
q
q
y(t) = c1 e 2 t sin t + c2 e 2 t cos t
2
2
exhibits damped oscillations (the amplitude of oscillations tends to zero, as
t ).
(ii) 2 4 2 = 0. There is a double real root . The general solution

2
y(t) = c1 e 2 t + c2 te 2 t
tends to zero as t , without oscillating.
(iii) 2 4 2 > 0. The roots are real and distinct. If we denote q =
q
+ q
2 4 2 , then the roots are r1 =
, and r2 =
. Both roots
2
2
are negative, because q < . The general solution
y(t) = c1 er1 t + c2 er2 t
tends to zero as t , without oscillating. We see that large enough
dissipation kills the oscillations.
2.11
Further Applications
2.11.1
Forced and Damped Oscillations
It turns out that even a little damping is enough to avoid resonance. We

consider the model
(11.1)
y 00 + y 0 + 2 y = sin t .
Our theory tells us to look for a particular solution in the form Y (t) =
A1 cos t + A2 sin t. Once the constants A1 and A2 are determined, we can
use trigonometric identities to put this solution into the form
(11.2)
Y (t) = A sin(t ) ,
with the constants A and depending on A1 and A2 . So, let us look for
a particular solution directly in the form (11.2). We transform the forcing
term as a linear combination of sin(t ) and cos(t ):
sin t = sin ((t ) + ) = sin(t ) cos + cos(t ) sin .
88
We now plug Y (t) into the equation (11.1):

A 2 sin(t ) + A cos(t ) + A 2 sin(t ) =
sin(t ) cos + cos(t ) sin .
Equating the coefficients in sin(t ) and cos(t ), we have
(11.3)
A( 2 2 ) = cos
A = sin .
We now square both of these equations, and add the results

A2 ( 2 2 )2 + A2 2 2 = 1 ,
which allows us to calculate A:
A= p
( 2
1
.
2 )2 + 2 2
To calculate , we divide the second equation in (11.3) by the first
, or = tan1 2
.
tan = 2
2

2
We computed a particular solution
1
Y (t) = p 2
sin(t ), where = tan1 2
.
2
2
2
2
2
( ) +
We now make a physically reasonable assumption that the damping coefficient is small, so that 2 4 2 < 0. Then the characteristic equation for
the homogeneous equation corresponding to (11.1),
r 2 + r + 2 = 0 ,
has a pair of complex roots 2 i, where = 42 . The general

solution of (11.1) is then
1
y(t) = c1 e 2 t cos t + c2 e 2 t sin t + p 2
sin(t ) .
( 2 )2 + 2 2
The first two terms of this solution are called the transient oscillations,
because they quickly tend to zero, as the time t goes on (sic transit gloria
mundi). So that the second term, Y (t), describes the oscillations in the
long run. We see that oscillations of Y (t) are bounded, no matter what is
the frequency of the forcing term. The resonance is gone! Moreover, the
largest amplitude of Y (t) occurs not at = , but at a slightly smaller
value of . Indeed, the maximal amplitude happens when the quantity in
the denominator, ( 2 2 )2 + 2 2 , is the smallest. This quantity
is a
q
2
quadratic in 2 . Its minimum occurs when 2 = 2 2 , or =
2
2 .
89
2.11.2
An Example of a Discontinuous Forcing Term
We now consider equations with a jumping force function. A simple function

with a jump at some number c, is the Heaviside step function
uc (t) =
0 if t c
.
1 if t > c
Example For t > 0, solve the problem

y 00 + 4y = f (t)
y(0) = 0, y 0 (0) = 3 ,
where
f (t) =
0
if t /4
.
t + 1 if t > /4
(No external force is applied to the spring before the time t = /4, and the
force is equal to t + 1 afterwards. The forcing function can be written as
f (t) = u/4 (t)(t + 1).)
The problem naturally breaks down into two parts. When 0 < t /4,
we are solving
y 00 + 4y = 0 .
Its general solution is y(t) = c1 cos 2t + c2 sin 2t. From the initial conditions,
we find that c1 = 0, c2 = 32 , so that
3
sin 2t, for t /4 .
2
At later times, when t > /4, our equation is
(11.4)
y(t) =
y 00 + 4y = t + 1 .
(11.5)
But what are the new initial conditions at the time t = /4? Clearly, we
can get them from (11.4):
(11.6)
y(/4) =
3
, y 0 (/4) = 0 .
2
The general solution of (11.5) is y(t) = 14 t + 41 + c1 cos 2t + c2 sin 2t. Calculating c1 and c2 from the initial conditions in (11.6), we find that y(t) =
1
1
1
5
4 t + 4 + 8 cos 2t + ( 4 16 ) sin 2t. Answer:

y(t) =
3
2
sin 2t,
1
4t
1
4
+ 18 cos 2t + ( 45
if t /4
16 ) sin 2t,
if t > /4
90
2.11.3
Oscillations of a Pendulum
Assume that a small ball of mass m is attached to one end of a rigid rod of
length l, while the other end of the rod is attached to the ceiling. Assume
also that the mass of the rod itself is so small, that it can be neglected.
Clearly, the ball will move on an arch of a circle of radius l. Let = (t)
be the angle that the pendulum makes with the vertical line, at the time
t. We assume that > 0 if the pendulum is to the left of the vertical line,
and < 0 to the right of the vertical. If the pendulum moves by an angle
radians, it covers the distance l = l(t). It follows that l0 (t) gives the
velocity, and l00 (t) the acceleration. We assume that the only force acting
on the mass is the force of gravity. Only the projection of this force on
the tangent line to the circle is active, which is mg cos( 2 ) = mg sin .
Newtons second law of motion gives
ml00 (t) = mg sin .
(Minus, because the force works to decrease the angle , when > 0.)
Denoting g/l = 2 , we have the pendulum equation
00 (t) + 2 sin (t) = 0 .
a
@
@
@
l
@
@
@
@dm
Gravity acting on a pendulum, =
If the oscillation angle (t) is small, then sin (t) (t), giving us again
a harmonic oscillator
00 (t) + 2 (t) = 0 ,
this time as a model of small oscillations of a pendulum.
91
2.11.4
Sympathetic Oscillations
Suppose that we have two pendulums hanging from the ceiling, and they are
coupled (connected) through a weightless spring. Let x1 denote the angle
the left pendulum makes with the vertical line. We consider this angle to
be positive if the pendulum is to the left of the vertical, and x1 < 0 if
the pendulum is to the right of the vertical. Let x2 be the angle the right
pendulum makes with the vertical, with the same assumptions on its sign.
We assume that x1 and x2 are small in absolute value, which means that
each pendulum separately can be modeled by a harmonic oscillator. For
coupled pendulums the model is
x001 + 2 x1 = k(x1 x2 )
(11.7)
x002 + 2 x2 = k(x1 x2 ) ,
where k > 0 is a physical constant, the stiffness of the coupling spring. If

x1 > x2 > 0, then the coupling spring is extended, so that the spring tries
to contract, and in doing so it pulls back the left pendulum, while pulling
forward (accelerating) the right pendulum. (Correspondingly, the forcing
term is negative in the first equation, and positive in the second one.) In
case 0 < x1 < x2 , the spring is compressed, and as it tries to expand,
it accelerates the first (left) pendulum, and slows down the second (right)
pendulum. We shall solve the system (11.7), together with the simple initial
conditions
(11.8)
x1 (0) = a, x01 (0) = 0, x2 (0) = 0, x02 (0) = 0 ,
which correspond to the first pendulum beginning with a small displacement
angle a, and zero initial velocity, while the second pendulum is at rest.
x1

d
x2
Two pendulums connected by a weightless spring

We add the equations in (11.7), and call z1 = x1 + x2 . Obtain
z100 + 2 z1 = 0, z1 (0) = a, z10 (0) = 0 .
92
The solution of this initial value problem is z1 (t) = a cos t. Subtracting

the second equation in (11.7) from the first one, and calling z2 = x1 x2 ,
we have z200 + 2 z2 = 2kz2 , or
z200 + ( 2 + 2k)z2 = 0, z2 (0) = a, z20 (0) = 0 .
Denoting 2 + 2k = 12 , or 1 = 2 + 2k, we have z2 (t) = a cos 1 t.

Clearly, z1 + z2 = 2x1 . Then
z1 + z2
a cos t + a cos 1 t
1
1 +
(11.9) x1 =
=
= a cos
t cos
t,
2
2
2
2
using a trig identity for the last step. Similarly,
z1 z2
a cos t a cos 1 t
1
1 +
(11.10) x2 =
=
= a sin
t sin
t.
2
2
2
2
We now analyze the solution, given by the formulas (11.9) and (11.10).
If k is small (the coupling is weak), then 1 is close to , and so their
1
1
difference 1 is small. It follows that both cos
t and sin
t
2
2
change very slowly with time t. We now rewrite the solution as
1 +
1 +
x1 = A cos
t, and x2 = B sin
t,
2
2
1
1
t, and B = a sin
t, as slowly varywhere we regard A = a cos
2
2
ing amplitudes. We interpret this by saying that the pendulums oscillate
1 +
with the frequency
, and with slowly varying amplitudes A and B.
2
1
At the times t, when cos
t is zero, and the first pendulum is at rest
2
1
(x1 = 0), the amplitude of the second pendulum satisfies | sin
t| = 1,
2
obtaining its largest possible absolute value. We have a complete exchange
of energy: when one of the pendulums is doing the maximal work, the other
one is resting. We have sympathy between the pendulums. Observe also
that A2 + B 2 = a2 . This means that at all times t, the point (x1 (t), x2 (t))
lies on a circle of radius a in the (x1 , x2 ) plane.
2.12
Oscillations of a Spring Subject to a Periodic

Force
2.12.1
Fourier Series
For vectors in three dimensions, one of the central notions is that of the
scalar product (also known as the inner product or the dot product).
2.12. OSCILLATIONS OF A SPRING SUBJECT TO A PERIODIC FORCE93
x1
y1
Namely, if x = x2 and y = y2 , then their scalar product is

x3
y3
(x, y) = x1 y1 + x2 y2 + x3 y3 .
Scalar product can be used to compute the length of a vector, ||x|| =

and the angle between the vectors x and y
cos =
(x, x),
(x, y)
.
||x|| ||y||
In particular, the vectors x and y are orthogonal (perpendicular) if (x, y) =

0. If i, j and k are the unit coordinate vectors, then (x, i) = x1 , (x, j) = x2 ,
and (x, k) = x3 . Writing x = x1 i + x2 j + x3 k, we express
(12.1)
x = (x, i)i + (x, j)j + (x, k)k .
This formula gives probably the simplest example of Fourier Series.

We shall now consider functions f (t) that are periodic, with period 2.
Such functions show us everything they got on any interval of length 2. So
let us consider them on the interval (, ). Given two functions f (t) and
g(t), we define their scalar product as
(f, g) =
f (t) g(t) dt .
We
call the functions orthogonal if (f, g) = 0. For example, (sin t, cos t) =
R
sin t cos t dt = 0, so that sin t and cos t are orthogonal. (Observe that the
orthogonality of these functions has nothing to do with the angle at which
their graphs intersect.) The notion of scalar product allows us to define the
norm of a function
||f || =
(f, f ) =
sZ
f 2 (t) dt .
For example,
sZ
sZ
1 1
|| sin t|| =
sin t dt =
cos 2t dt = .
2
Similarly,
for any positive integer n, || sin nt|| = , || cos nt|| = , and
||1|| = 2.
94

We now consider an infinite set of functions
1, cos t, cos 2t, . . . , cos nt, . . . , sin t, sin 2t, . . . , sin nt, . . . .
They are all mutually orthogonal. This is because

(1, cos nt) =
cos nt dt = 0;
(1, sin nt) =
sin nt dt = 0;
(cos nt, cos mt) =
(sin nt, sin mt) =
(sin nt, cos mt) =
cos nt cos mt dt = 0, for all n 6= m;

sin nt sin mt dt = 0, for all n 6= m;
sin nt cos mt dt = 0, for any n and m .
The last three integrals are computed by using trig identities. If we divide these functions by their norms, we shall obtain an orthonormal set of
functions
1
cos t cos 2t
cos nt
sin t sin 2t
sin nt
, , , ... , , ... , , , ... , , ... ,
2
which is similar to the vectors i, j and k. It is known that these functions form a complete set, so that any function f (t) can be represented
as their linear combination. Similarly to the formula for vectors (12.1), we
decompose an arbitrary function f (t) as
X
1
cos nt
sin nt
f (t) = 0 +
n + n
2 n=1
where
1
1
0 = (f (t), ) =
2
2
cos nt
1
n = (f (t), ) =
sin nt
1
n = (f (t), ) =
f (t) dt ,
f (t) cos nt dt ,
f (t) sin nt dt .
2.12. OSCILLATIONS OF A SPRING SUBJECT TO A PERIODIC FORCE95
It is customary to denote a0 = 0 / 2, an = n / , and bn = n / , so

that the Fourier Series takes the final form
f (t) = a0 +
(an cos nt + bn sin nt) ,
n=1
with
a0 =
1
2
f (t) dt ,
1
an =
f (t) cos nt dt ,

Z
1
f (t) sin nt dt .
bn =

Z
Example Let f (t) be a function of period 2, which is equal to t + on

the interval (, ). This is the saw-tooth function. It is not defined at the
points n and n, with n odd, but this does not affect the integrals that
we need to compute. Compute
1
a0 =
2
(t + ) dt =
Using guess-and-check
1
an =

1

(t + )2 = .

1
sin nt cos nt
(t + ) cos nt dt =
(t + )
+ 2
= 0,

n
n
because sin n = 0, and cosine is an even function. Similarly

1
bn =

1
cos nt
sin nt
(t + ) sin nt dt =
(t + )(
)+ 2

n
n
2
2
2
= cos n = (1)n = (1)n+1 .
n
n
n
(Observe that cos n is equal to 1 for even n, and to 1 for odd n, which
may be combined as cos n = (1)n.) We obtained a Fourier series for the
function f (t):
X
2
(1)n+1 sin nt ,
f (t) = +
n
n=1
which is valid on (, ) (with the exception of points n and n, with

n odd). Restricting to the interval (, ), we have
t+ = +
X
2
n
n=1
(1)n+1 sin nt, for < t < .
96
It might look that we did not accomplish much by expressing a simple

function t + through an infinite series. However, we can now express
solutions of differential equations through Fourier series.
y
6
c
- t
The saw-tooth function

2.12.2
Vibrations of a Spring Subject to a Periodic Force
As before, the model is

y 00 + 2 y = f (t) ,
where y = y(t) is the displacement, > 0 is a constant (the natural
frequency), and f (t) is a given function of period 2, the external force.
This equation also models oscillations in electrical circuits. Expressing f (t)
by its Fourier series, we have
y 00 + 2 y = a0 +
(an cos nt + bn sin nt) .
n=1
Let us assume that 6= n, for any integer n (to avoid resonance). According
to our theory, we look for a particular solution in the form Y (t) = A0 +
P
n=1 (An cos nt + Bn sin nt). Plugging this in, we find
Y (t) =
X
an
bn
a0
+
cos nt + 2
sin nt
2
2
2
n
n2
n=1
The general solution is then

y(t) =
X
a0
an
bn
+
cos nt + 2
sin nt + c1 cos t + c2 sin t .
2
2 n2
n2
n=1

We see that the coefficients in the m-th harmonics (in cos mt and sin mt)
are large, provided that the natural frequency is selected to be close to m.
97
2.13. EULERS EQUATION
That is basically what happens, when one is turning the tuning knob of a
radio set. (The knob controls , while your favorite station broadcasts at a
frequency m, so that its signal has the form f (t) = am cos mt + bm sin mt.)
2.13
Eulers Equation
Preliminaries
What is the meaning of 3 2 ? Or, more generally, what is the definition of

r
the function tr , where r is any real number? Here it is: tr = eln t = er ln t .
We see that the function tr is defined only for t > 0. The function |t|r is
defined for all t 6= 0, but what is the derivative of this function?
More generally, if f (t) is a differentiable function, the function f (|t|) is
differentiable for all t 6= 0 (|t| is not differentiable at t = 0). Let us define
the following step function
sign(t) =
1
if t > 0
.
1 if t < 0
Observe that
d
|t| = sign(t), for all t 6= 0 ,
dt
as follows by considering separately the cases t > 0, and t < 0.
By the chain rule, we have
(13.1)
In particular,
d
f (|t|) = f 0 (|t|)sign(t), for all t 6= 0 .
dt
d
r
dt |t|
= r|t|r1 sign(t).
Observe also that t sign(t) = |t|, and (sign(t))2 = 1, for all t 6= 0.

The Important Class of Equations
Eulers equation has the form (here y = y(t))
(13.2)
t2 y 00 + aty 0 + by = 0 ,
where a and b are given numbers. Assume first that t > 0. We look for
solution in the form y = tr , with the constant r to be determined. Plugging
this in, gives
t2 r(r 1)tr2 + atrtr1 + btr = 0 .
98
Dividing by a positive quantity tr

(13.3)
r(r 1) + ar + b = 0
gives us a quadratic equation, called the characteristic equation. There are

three possibilities with respect to its roots, which we consider next.
Case 1 There are two real and distinct roots r1 6= r2 . Then tr1 and tr2
are two solutions, which are not constant multiples of each other, and the
general solution (valid for t > 0) is
y(t) = c1 tr1 + c2 tr2 .
If r1 is either an integer, or a fraction with an odd denominator, then tr1
is also defined for t < 0. If the same is true for r2 , then the above general
solution is valid for all t 6= 0. For other r1 and r2 , this solution is not even
defined for t < 0. In such a case, by using the differentiation formula (13.1),
we see that y(t) = |t|r1 gives a solution of Eulers equation, which is valid
for all t 6= 0. Just plug it in, and observe that sign(t)sign(t) = 1, and
tsign(t) = |t|. So that
y(t) = c1 |t|r1 + c2 |t|r2
gives a general solution valid for all t 6= 0.
Example t2 y 00 + 2ty 0 2y = 0.
r(r 1) + 2r 2 = 0
has roots r1 = 2, and r2 = 1. Solution: y(t) = c1 t2 + c2 t, valid for all
t 6= 0. Another general solution valid for all t 6= 0 is y(t) = c1 |t|2 + c2 |t| =
c1 t2 + c2 |t|. This is a truly different solution! Why such an unexpected
complexity? If one divides this equation by t2 , then the functions p(t) = 2/t
and g(t) = 2/t2 from our general theory, are both discontinuous at t = 0.
We have a singularity at t = 0, and, in general, solution y(t) is not even
defined at t = 0 (as we see in this example), and that is the reason for the
complexity. However, when solving initial value problems, it does not matter
which form of the general solution one uses. For example, if we prescribe
some initial conditions at t = 1, then both forms of the general solution
are valid only on the interval (, 0), and on that interval both forms are
equivalent.
We now turn to the cases of equal roots, and of complex roots, for the
characteristic equation. We could proceed similarly to the linear equations
2.13. EULERS EQUATION
99
with constant coefficients. Instead, we make a change of independent variables from t to a new variable s, by letting t = es , or s = ln t. By the chain
rule
dy
dy ds
dy 1
=
=
.
dt
ds dt
ds t
Using the product rule, and then the chain rule,
d2 y
d dy 1 dy 1
d2 y ds 1 dy 1
d2 y 1
dy 1
=
(
)
.
2
2
2
2
2
2
dt
dt ds t
ds t
ds dt t
ds t
ds t
ds t2
Then Eulers equation (13.2) becomes
dy
d2 y dy
+ a + by = 0 .
2
ds
ds
ds
This is a linear equations with constant coefficients! We can solve it for
any a and b. Let us use primes again to denote the derivatives in s in this
equation. Then we have
(13.4)
y 00 + (a 1)y 0 + by = 0 .
Its characteristic equation

(13.5)
r 2 + (a 1)r + b = 0
is exactly the same as (13.3).

We now return to Eulers equation, and its characteristic equation (13.3).
Case 2 r1 is a double root of the characteristic equation (13.3), i.e., r1 is a
double root of (13.5). Then y = c1 er1 s + c2 ser1 s is the general solution of
(13.4). Returning to the original variable t, by substituting s = ln t:
y(t) = c1 tr1 + c2 tr1 ln t .
This general solution of Eulers equation is valid for t > 0. More generally,
it is straightforward to verify that
y(t) = c1 |t|r1 + c2 |t|r1 ln |t|
gives us the general solution of Eulers equation, valid for all t 6= 0.
Case 3 p iq are complex roots of the characteristic equation (13.3). Then
y = c1 eps cos qs + c2 seps sin qs is the general solution of (13.4). Returning to
100
the original variable t, by substituting s = ln t, we get the general solution

of Eulers equation
y(t) = c1 tp cos(q ln t) + c2 tp sin(q ln t) ,
valid for t > 0. Replacing t by |t|, will give us the general solution of Eulers
equation, valid for all t 6= 0.
Example t2 y 00 3ty 0 + 4y = 0, t > 0.
r(r 1) 3r + 4 = 0
has a double root r = 2. The general solution: y = c1 t2 + c2 t2 ln t.
Example t2 y 00 3ty 0 + 4y = 0, y(1) = 4, y 0 (1) = 7.
Using the general solution from the preceding problem, we calculate c1 = 4
and c2 = 1. Answer: y = 4t2 t2 ln t.
Example t2 y 00 + ty 0 + 4y = 0, y(1) = 0, y 0 (1) = 3.

r(r 1) + r + 4 = 0
has a pair of complex roots 2i. Here p = 0, q = 2, and the general solution,
valid for both positive and negative t, is
y(t) = c1 cos(2 ln |t|) + c2 sin(2 ln |t|) .
From the first initial condition, y(1) = c1 = 0, so that y(t) = c2 sin(2 ln |t|).
To use the second initial condition, we need to differentiate y(t) at t =
1. Observe that for negative t, |t| = t, and so y(t) = c2 sin(2 ln(t)).
Then y 0 (t) = c2 cos(2 ln(t)) 2t , and y 0 (1) = 2c2 = 3, giving c2 = 3/2.
3
Answer: y(t) = sin(2 ln |t|).
2
Example t2 y 00 3ty 0 + 4y = t 2, t > 0.
This is a non-homogeneous equation. We look for a particular solution in the
1
form Y = At+B. Plugging this in, we compute Y = t . The fundamental
2
solution set of the corresponding homogeneous equation is given by t2 and
1
t2 ln t, as we saw above. The general solution is y = t + c1 t2 + c2 t2 ln t.
2
2.14. LINEAR EQUATIONS OF ORDER HIGHER THAN TWO
101
2.14
Linear Equations of Order Higher Than Two
2.14.1
The Polar Form of Complex Numbers
For a complex number x + iy, one can use the point (x, y) to represent it.
This turns the usual plane into the complex plane. The point (x, y) can also
be identified by its polar coordinates (r, ). We shall always take r > 0.
Then
z = x + iy = r cos + ir sin = r(cos + i sin )
gives us polar form of a complex number z. Using Eulers formula, we can
3
also write z = rei . For example, 2i = 2ei 2 , because the point (0, 2)
has polar coordinates (2, 3

2ei 4 , and 1 = ei (real
2 ). Similarly, 1 + i =
numbers are just particular cases of complex ones).
There are infinitely many ways to represent a complex number using polar coordinates z = rei(+2m) , where m is any integer (positive or negative).
We now compute the n-th root(s) of z:
(14.1)
z 1/n = r 1/n ei( n +
2m
)
n
, m = 0, 1, . . . , n 1 .
Here r 1/n is the positive n-th root of the positive number r. (The high
school n-th root.) When m varies from 0 to n 1, we get different values,
and then the roots repeat themselves. There are n complex n-th roots of
any complex number (and in particular, of any real number). All of these
roots lie on a circle of radius r 1/n around the origin, and the difference in
polar angles between any two neighbors is 2/n.
Example Solve the equation: z 4 + 16 = 0.
We need the four complex roots of 16 = 16ei(+2m) . The formula (14.1)
gives
m
(16)(1/4) = 2ei( 4 + 2 ), m = 0, 1, 2, and 3 .
When m = 0, we have ei 4 = 2(cos 4 + i sin 4 ) = 2 + i 2. We compute the
other roots
similarly.
They
come
in
two
complex
conjugate
pairs:
2
i
2
and 2 i 2. In the complex plane, they all lie on circle of radius 2, and
the difference in polar angles between any two neighbors is /2.
Example Solve the equation: z 3 + 8 = 0.
We need the three complex roots of 8. One of them is z1 = 2 = 2ei ,
and the other two lie on
an angle 2/3 away, so
the circle of radius 2, at
that z2 = 2ei/3 = 1 + 3i, and z3 = 2ei/3 = 1 3i. (Alternatively, the
102
root z = 2 is easy to guess. Then factor z 3 + 8 = (z + 2)(z 2 2z + 4), and

set the second factor to zero, to find the other two roots.)
6
'$
b
b
2
b
b
&%
The four complex fourth roots of 16, on the circle of radius 2
2.14.2
Linear Homogeneous Equations
Let us consider the fourth order equations

(14.2)
a0 y 0000 + a1 y 000 + a2 y 00 + a3 y 0 + a4 y = 0 ,
with given numbers a0 , a1 , a2 , a3 , and a4 . Again, we search for solution in

the form y(t) = ert , with a constant r to be determined. Plugging this
in, and dividing by the positive exponent ert, we obtain the characteristic
equation
(14.3)
a0 r 4 + a1 r 3 + a2 r 2 + a3 r + a4 = 0 .
If we have an equation of order n
(14.4)
a0 y (n) + a1 y (n1) + + an1 y 0 + an y = 0 ,
then the corresponding characteristic equation is

a0 r n + a1 r n1 + + an1 r + an = 0 .
The fundamental theorem of Algebra says that any polynomial of degree
n has n roots in the complex plane, counted according to their multiplicity
(double root is counted as two roots, and so on). The characteristic equation
(14.3) has four complex roots.
The theory is similar to the second order case. We need four different
solutions of (14.2), so that every solution is not a linear combination of the
other three (for the equation (14.4) we need n different solutions). Every root
of the characteristic equation must pull its weight. If the root is simple,
it brings in one solution, if it is repeated twice, then two solutions. (Three
103
solutions, if the root is repeated three times, and so on.) The following cases
may occur for the n-th order equation (14.4).
Case 1 r1 is a simple real root. Then it brings er1 t into the fundamental
set.
Case 2 r1 is a real root repeated s times. Then it brings the following s
solutions into the fundamental set: er1 t , ter1 t , . . ., ts1 er1 t .
Case 3 p+iq and piq are simple complex roots. They contribute: ept cos qt
and ept sin qt into the fundamental set.
Case 4 p + iq and p iq are repeated s times each. They bring the following
2s solutions into the fundamental set: ept cos qt and ept sin qt, tept cos qt and
tept sin qt, . . ., ts1 ept cos qt and ts1 ept sin qt.
Example y 0000 y = 0. The characteristic equation is
r4 1 = 0 .
We solve it by factoring
(r 1)(r + 1)(r 2 + 1) = 0 .
The roots are 1, 1, i, i. The general solution: y(t) = c1 et + c2 et +
c3 cos t + c4 sin t.
Example y 000 3y 00 + 3y 0 y = 0. The characteristic equation is
r 3 3r 2 + 3r 1 = 0 .
This is a cubic equation. You probably did not study how to solve it by
Cardanos formula. Fortunately, you must remember that the quantity on
the left is an exact cube:
(r 1)3 = 0 .
The root r = 1 is repeated 3 times. The general solution: y(t) = c1 et +

c2 tet + c3 t2 et .
Let us suppose you did not know the formula for cube of a difference.
Then one can guess that r = 1 is a root. This means that the cubic polynomial can be factored, with one factor being r 1. The other factor is
then found by the long division. The other factor is a quadratic polynomial,
whose roots are easy to find.
Example y 000 y 00 + 3y 0 + 5y = 0. The characteristic equation is
r 3 r 2 + 3r + 5 = 0 .
104
We need to guess a root. The procedure for guessing a root (for textbook
examples) is a simple one: try r = 0, r = 1, r = 2, and then give up.
We see that r = 1 is a root. It follows that the first factor is r + 1, and
the second one is found by the long division:
(r + 1)(r 2 2r + 5) = 0 .
The roots of the quadratic are r2 = 1 2i, and r3 = 1 + 2i. The general
solution: y(t) = c1 et + c2 et cos 2t + c3 et sin 2t.
Example y 0000 + 2y 00 + y = 0. The characteristic equation is
r 4 + 2r 2 + 1 = 0 .
It can be solved by factoring
(r 2 + 1)2 = 0 .
(Or one could set z = r 2 , and obtain a quadratic equation for z.) The
roots are i, i, each repeated twice. The general solution: y(t) = c1 cos t +
c2 sin t + c3 t cos t + c4 t sin t.
Example y 0000 + 16y = 0. The characteristic equation is
r 4 + 16 = 0 .
Its solutions
are
the four complex roots of 16, computed earlier:

and 2 i 2. The general solution:
2t
y(t) = c1 e
2 i 2,
cos( 2 t)+c2 e 2 t sin( 2 t)+c3 e 2 t cos( 2 t)+c4 e 2 t sin( 2 t) .
Example y (5) + 9y 000 = 0. The characteristic equation is

r 5 + 9r 3 = 0 .
Factoring r 3 (r 2 + 9) = 0, we see that the roots are: 0, 0, 0, 3i, 3i. The
general solution: y(t) = c1 + c2 t + c3 t2 + c4 cos 3t + c5 sin 3t.
2.14.3
Non-Homogeneous Equations
The theory is parallel to the second order case. Again, we need a particular
solution.
Example y (5) + 9y 000 = 3t sin 2t.
105
We just found the general solution of the corresponding homogeneous equation. We produce a particular solution in the form Y (t) = Y1 (t) + Y2 (t),
where Y1 (t) is a particular solution of y (5) + 9y 000 = 3t, and Y2 (t) is a
particular solution of y (5) + 9y 000 = sin 2t. We guess that Y1 (t) = At4 ,
1
1
and compute A = 72
, and that Y2 (t) = B cos 2t, which gives B = 40
. We
1 4
1
have Y (t) = 72 t 40 cos 2t.
Answer: y(t) =
2.14.4
1 4
72 t
1
40
cos 2t + c1 + c2 t + c3 t2 + c4 cos 3t + c5 sin 3t.
Problems
I. Solve the non-homogeneous equations with discontinuous forcing function.

1. y 00 + 9y = f (t), where f (t) = 0 for 0 < t < , and f (t) = t for t > ,
y(0) = 0, y 0 (0) = 2.
Answer:
y(t) =
3 sin 3t,
9t + 9
cos 3t
if t
17
27
sin 3t, if t >
2. y 00 + y = f (t), where f (t) = 0 for 0 < t < , and f (t) = t for t > ,
y(0) = 2, y 0 (0) = 0.
II. Find the general solution, valid for t > 0.
1. t2 y 00 2ty 0 + 2y = 0.
2. t2 y 00 + ty 0 + 4y = 0.
Answer. y = c1 t + c2 t2 .
Answer. y = c1 cos(2 ln t) + c2 sin(2 ln t).
3. t2 y 00 + 5ty 0 + 4y = 0.
Answer. y = c1 t2 + c2 t2 ln t.
4. t2 y 00 + 5ty 0 + 5y = 0.
Answer. y = c1 t2 cos(ln t) + c2 t2 sin(ln t).
5. t2 y 00 3ty 0 = 0.
1
6. y 00 + t2 y = 0.
4
Answer. y = c1 + c2 t4 .
Answer. y = c1 t + c2 t ln t.
III. Find the general solution, valid for all t 6= 0.
1. t2 y 00 + ty 0 + 4y = 0.
Answer. y = c1 cos(2 ln |t|) + c2 sin(2 ln |t|).
2. 2t2 y 00 ty 0 + y = 0.
Answer. y = c1 |t| + c2 t.
3. 2t2 y 00 ty 0 + y = t2 3.
106
Hint: look for a particular solution as Y = At2 + Bt + C.

q
1
Answer. y = t2 3 + c1 |t| + c2 t.
3
2 00
0
4. 2t y ty + y = t3 . Hint: look for a particular solution as Y = At3 .
q
1 3
Answer. y =
t + c1 |t| + c2 t.
10
IV. Solve the following initial value problems.
1. t2 y 00 2ty 0 + 2y = 0, y(1) = 2, y 0 (1) = 5.
Answer. y = t + 3t2 .
2. t2 y 00 3ty 0 +4y = 0, y(1) = 1, y 0 (1) = 2.
Answer. y = t2 4t2 ln |t|.
3. t2 y 00 +3ty 0 3y = 0, y(1) = 1, y 0 (1) = 2.
3
1
Answer. y = t3 t.
4
4
4. t2 y 00 ty 0 + 5y = 0, y(1) = 0, y 0 (1) = 2.
Answer. y = t sin(2 ln t).
V. Solve the polynomial equations.

1. r 3 1 = 0.
Answer. Roots are 1, ei
2. r 3 + 27 = 0.
4
3. r 16 = 0.
3
3
2
Answer. 3,
2
3
3 3
3
2 i, 2
Answer. 2 and 2i.
4. r 3r + r + 1 = 0.
Answer. 1, 1
i 4
i 3
4
i 5
4
, ei
+
4
3
3 3
2 i.
2, 1 +
i 7
4
2.
5. r 4 + 1 = 0.
Answer. e , e
6. r 4 + 4 = 0.
Answer. 1 + i, 1 i, 1 + i, 1 i.
7. r 4 + 8r 2 + 16 = 0.
,e
,e
Answer. 2i and 2i are both double roots.
8. r 4 + 5r 2 + 4 = 0.
Answer. i and 2i.
VI. Find the general solution.

000
1. y y = 0.
t/2
Answer. y = c1 e +c2 e
2. y 000 3y 00 + y 0 + y = 0.
3. y (4) 8y 00 +16y = 0.
3
3
t/2
cos
t+c3 e
sin
t.
2
2
2)t
Answer. y = c1 et + c2 e(1
+ c3 e(1+
2)t
Answer. y = c1 e2t +c2 e2t +c3 te2t +c4 te2t .
107
4. y (4) + 8y 00 + 16y = 0.
Answer. y = c1 cos 2t + c2 sin 2t + c3 t cos 2t + c4 t sin 2t.
5. y (4) + y = 0.
Answer. y = c1 e
t
2
6. y 000 y = t2 .
t
t
t
t
t
t
t
cos + c2 e 2 sin + c3 e 2 cos + c4 e 2 sin .
2
2
2
2
3
3
2
t
t/2
t/2
Answer. y = t +c1 e +c2 e
cos
t+c3 e
sin
t.
2
2
7. y (6) y 00 = 0.
Answer. y = c1 +c2 t+c3 et +c4 et +c5 cos t+c6 sin t.
8. y (8) y (6) = sin t.

Answer. y =
1
sin t + c1 + c2 t + c3 t2 + c4 t3 + c5 t4 + c6 t5 + c7 et + c8 et .
2
9. y 0000 + 4y = 4t2 1.
1
Answer. y = t2 + c1 et cos t + c2 et sin t + c3 et cos t + c4 et sin t.
4
VII. Solve the following initial value problems.
1. y 000 + 4y 0 = 0, y(0) = 1, y 0 (0) = 1, y 00 (0) = 2.
Answer. y =
3
2
1
2
cos 2t 12 sin 2t.
2. y 0000 + 4y = 0, y(0) = 1, y 0 (0) = 1, y 00 (0) = 2, y 000(0) = 3.
Answer. y = 18 et (cos t 5 sin t) + 38 et (3 cos t sin t).

3. y 000 + 8y = 0, y(0) = 0, y 0 (0) = 1, y 00 (0) = 2.
Answer. y = 13 e2t + 13 et cos 3 t.
VIII.
1. Find the linear homogeneous differential equation of the lowest possible
order, which has the following functions as its solutions: 1, e2t , sin t.
Answer. y 0000 + 2y 000 + y 00 + 2y 0 = 0.
2. Find the general solution of
(t + 1)2 y 00 4(t + 1)y 0 + 6y = 0 .
Hint: Look for the solution in the form y = (t + 1)r .
108
Answer. y = c1 (t + 1)2 + c2 (t + 1)3 .

3. Solve the following nonlinear equation
2
2y 0 y 000 3y 00 = 0 .
(This equation is connected to the Schwarzian derivative, defined as Sf =
y000
y0
3
2

y00 2
.)
y0
Hint: Write this equation as

3 y 00
y 000
=
,
y 00
2 y0
then integrate, to get
y 00 = c1 y 0
(3/2)
Let y 0 = v, and obtain a first order equation.

1
Answer. y =
+ c3 , and also y = c4 t + c5 . (Equivalently, the answer
c1 t + c2
c1 t + c2
can be written as y =
.)
c3 t + c4
4. Consider the nonlinear equation
y 00 (t) + a(t)y 3 (t) = 0, 0 t < ,
where a(t) is a continuously differentiable function, and a(t) a0 > 0 for
all t 0, a0 is a constant. Assume that either a0 (t) > 0, or a0 (t) < 0 for all
t [0, ). Show that any solution remains bounded on [0, ).
Hint: Consider the energy function E(t) = 12 y 0 2 (t) + 14 a(t)y 4 (t). Using
the equation, E 0 (t) = 14 a0 (t)y 4 (t). In case a0 (t) < 0 for all t, the energy
is decreasing, and so y(t) is bounded. In case a0 (t) > 0 for all t, we have
E 0(t)
a0(t)
E 0 (t) = 14 a0 (t)y 4 (t) a0 (t) E(t)
a(t) , or E(t) a(t) . Integrating this over (0, t),
1 4
we get E(t) a(t) E(0)
a(0) , which implies that 4 y (t)
2.15
E(0)
a(0)
= constant.
Oscillation and Comparison Theorems
The equation
y 00 + n2 y = 0
has a solution y(t) = sin nt. The larger is n, the more roots this solution has,
and so it oscillates faster. In 1836, J.C.F. Sturm discovered the following
theorem.
2.15. OSCILLATION AND COMPARISON THEOREMS
109
Theorem 2.15.1 (The Sturm Comparison Theorem.) Let y(t) and v(t) be
respectively non-trivial solutions of the following equations
(15.1)
y 00 + b(t)y = 0 ,
(15.2)
v 00 + b1 (t)v = 0 .
Assume that the given continuous functions b(t), and b1 (t) satisfy
(15.3)
b1 (t) b(t) for all t.
In case b1 (t) = b(t) on some interval, assume additionally that y(t) and v(t)
are not constant multiples of one another. Then v(t) has a root between any
two consecutive roots of y(t).
Proof:
Let t1 < t2 be two consecutive roots of y(t), y(t1 ) = y(t2 ) = 0.
We may assume that y(t) > 0 on (t1 , t2 ) (in case y(t) < 0 on (t1 , t2 ), we
may consider y(t), which is also a solution of (15.1)). Assume, contrary to
what we want to prove, that v(t) has no roots on (t1 , t2 ). We may assume
that v(t) > 0 on (t1 , t2 ).
Multiply the equation (15.2) by y(t), and subtract from that the equation
(15.1), multiplied by v(t). The result may be written as
0
v 0 y vy 0 + (b1 b)yv = 0 .
Integrating this over (t1 , t2 ), gives
0
(15.4) v(t2 )y (t2 ) + v(t1 )y (t1 ) +
t2
t1
[b1 (t) b(t)] y(t)v(t) dt = 0 .
All three terms on the left are non-negative. If b1 (t) > b(t) on some subinterval of (t1 , t2 ), then the third term is strictly positive, and we have
a contradiction. We now consider the case when b1 (t) = b(t) for all t
(t1 , t2 ). We claim that v(t) cannot vanish at either t1 or t2 , so that v(t1 ) >
0, and v(t2 ) > 0. Indeed, in case v(t1 ) = 0, we consider the function
y 0 (t1 )
z(t) = 0
v(t). This function is a solution of (15.1), and z(t1 ) = y(t1 ) =
v (t1 )
0, z 0 (t1 ) = y 0 (t1 ), so that by the uniqueness of solutions for initial value
problems, z(t) = y(t) for all t, and then y(t) and v(t) are constant multiples
of one another, which is not allowed. It follows that v(t1 ) > 0, and similarly
we prove that v(t2 ) > 0. The uniqueness theorem for initial value problems
also implies that y 0 (t1 ) > 0, and y 0 (t2 ) < 0 (otherwise, y(t) = 0). Then the
110
first two terms in (15.4) are strictly positive, and we have a contradiction in
(15.4).
v(t)
y(t)
t1
t2
- t
The functions y(t) and v(t)

In case y(t) and v(t) are two solutions of the same equation (15.1),
which are not constant multiples of one another, the theorem implies that
their roots interlace, which means that between any two roots of one of the
solutions there is a root of the other one.
By a similar argument, one proves the following version of the Sturm
Comparison Theorem, involving a differential inequality.
Lemma 2.15.1 Assume that the functions u(t) and v(t) are twice continuously differentiable, and they satisfy
v 00 + q(t)v = 0, v(a) = 0 ,
(15.5)
u00 + q(t)u 0, u(a) = 0, u0 (a) > 0 ,
on some interval (a, b) (with a given continuous function q(t)). Then v(t)
oscillates faster than u(t), provided that both functions are positive. Namely,
if v(t) > 0 on (a, b), then u(t) > 0 on (a, b).
If, on the other hand, u(t) > 0 on (a, b) and u(b) = 0, then v(t) must
vanish on (a, b].
Proof:
(15.6)
As before, we have
0
v 0 u vu0 0,
for x (a, b) .
Assume, contrary to what we want to prove, that u() = 0 at some (a, b).
Case 1. The inequality in (15.5) is strict on some sub-interval of (a, ). The
the same is then true for the inequality (15.6). Integrating (15.6) over (a, ),
we have
v()u0 () < 0 ,
111
which is a contradiction (because v() > 0, u0 () 0).

Case 2. We have u00 + q(t)u = 0 on (a, ). Then u(t) and v(t) are solutions
of the same equation on (a, ), and u(a) = v(a) = 0. It follows that u(t) and
v(t) are constant multiples of one another, but u() = 0, while v() > 0, a
contradiction, proving the first part of the lemma.
The second statement is proved similarly.
We shall need the following Calculus formula, discovered by Mauro Picone in 1909.
Lemma 2.15.2 (Picones Identity) Assume that the functions a(t) and a1 (t)
are differentiable, the functions u(t) and v(t) are twice differentiable, and
v(t) > 0 for all t. Then

0

u
vau0 ua1 v 0
v
Proof:
2
u2
u
= u(au ) (a1 v 0 )0 + (a a1 )u0 + a1 u0 v 0
v
v
0 0
2
The derivative on the left is equal to
u

u0 v uv 0
u
vau0 ua1 v 0 + v(au0 )0 u(a1 v 0 )0 +(aa1 ) u0 v 0 A+B+C .
2
v
v
v
The middle term B is the same as the first two terms on the right in Picones
identity. It remains to prove that A + C is equal to the sum of the last two
terms on the right in Picones identity. We expand A and C, and after a
cancellation, we get
0
A+C = (aa1 )u +a1
"
2 #
u
2 u
u 2 u0 v 0 + v 0
v
v
02
completing the proof.
u 0
v
v
We now turn to the general second order equations

(15.7)
= (aa1 )u0 +a1 u0
p(t)u00 + q(t)u0 + r(t)u = 0 .
The given functions p(t), q(t) and r(t) are assumed to be differentiable, and
p(t) > 0 for all t. We divide this equation by p(t)
u00 +
q(t) 0 r(t)
u +
u = 0,
p(t)
p(t)
R
and then multiply by the integrating factor a(t) = e

r(t)
a(t)
, we arrive at
p(t)
0
(15.8)
a(t)u0 + b(t)u = 0 ,
q(t)
p(t)
dt
. Denoting b(t) =
2
112
which is known as the self-adjoint form of (15.7).

With the help of his clever identity, M. Picone was able to give the
following generalization of the Sturm Comparison Theorem.
Theorem 2.15.2 Let u(t) and v(t) be respectively solutions of (15.8), and
0
a1 (t)v 0 + b1 (t)v = 0 .
Assume that the given differentiable functions a(t), a1 (t), and continuous
functions b(t), b1 (t) satisfy
b1 (t) b(t), and 0 < a1 (t) a(t)
for all t.
In case a1 (t) = a(t) and b1 (t) = b(t) for all t, assume additionally that u(t)
and v(t) are not constant multiples of one another. Then v(t) has a root
between any two consecutive roots of u(t).
Proof:
The proof is similar to that of the Sturm Comparison Theorem.
Let t1 < t2 be two consecutive roots of u(t), u(t1 ) = u(t2 ) = 0. As before,
we may assume that u(t) > 0 on (t1 , t2 ). Assume, contrary to what we
want to prove, that v(t) has no roots on (t1 , t2 ). We may assume, as before,
that v(t) > 0 on (t1 , t2 ). We apply Picones identity. Expressing from the
corresponding equations, (a(t)u0 )0 = b(t)u and (a1 (t)v 0 )0 = b1 (t)v, we
rewrite Picones identity as

0

u
vau0 ua1 v 0
v
= (b1 b)u2 + (a a1 )u0 + a1 u0
u 0
v
v
2
Integrating this over (t1 , t2 ), gives

0=
t2
t1
(b1 b)u dt +
t2
t1
(a a1 )u dt +
t2
t1
a1
u
u v0
v
0
2
dt .
(In case v(t1 ) = 0, we have v 0 (t1 ) > 0 by the uniqueness of solutions for
u2
2uu0
initial value problems. Then lim
= lim
= 0. Similarly, in case
tt1 v
tt1 v 0
v(t2 ) = 0, the upper limit vanishes for the integral on the left.) The integrals
on the right are non-negative. We obtain an immediate contradiction, unless
a1 (t) = a(t) and b1 (t) = b(t) for all t (t1 , t2 ). In such a case, we must also
u
have u0 v 0 = 0 on (t1 , t2 ) (so that all three integrals vanish). But then
v
u0
v0
= , and integrating we see that u(t) and v(t) are constant multiples of
u
v
one another, contradicting our assumption.
113
We define q + (t) = max(q(t), 0), the positive part of the function q(t).
Similarly, one defines q (t) = min(q(t), 0), the negative part of the function
q(t). Clearly, q(t) = q + (t) + q (t). Our next goal is the famous Lyapunov
inequality. It will follow from the following lemma.
Lemma 2.15.3 Assume that u(t) is twice continuously differentiable, and
it satisfies the following conditions on some interval (0, b) (here q(t) is a
given continuous function)
u00 + q(t)u = 0, u(0) = u(b) = 0, u(t) > 0 on (0, b) .
(15.9)
Then
Proof:
q + (t) dt >
4
.
b
From (15.9) it follows that

u00 + q + (t)u = q (t)u 0 .
Let v(t) be the solution of (c is any positive number)

v 00 + q + (t)v = 0, v(0) = 0, v 0 (0) = c > 0 .
(15.10)
By Lemma 2.15.1, v(t) must vanish on (0, b]. Let t2 (0, b] be the first root
of v(t), so that v(t) > 0 on (0, t2 ). (In case q (t) 0, we have t2 = b,
because v(t) is a constant multiple of u(t).) Integrating (15.10) (treating
q + (t)v as a known quantity),
(15.11)
v(t) = ct
(t s)q + (s)v(s) ds,
1
From v(t2 ) = 0, it follows that c =
t2
back into (15.11), we express
v(t) =
t
t2
t2
t2 v(t) =
t2
(t2 s)q + (s)v(s) ds. Plugging this
(t2 s)q + (s)v(s) ds
Breaking the first integral,

Z
R t2
0
Rt
0
+
+
R t2
t
for t [0, t2 ] .
(t s)q + (s)v(s) ds .
, we continue
[t(t2 s) t2 (t s)] q (s)v(s) ds + t
= (t2 t)
sq + (s)v(s) ds + t
t2
t2
(t2 s)q + (s)v(s) ds
(t2 s)q + (s)v(s) ds .
114
Let t0 be the point of maximum of v(t) on (0, t2 ), v(t0 ) > 0. Estimate

v(s) < v(t0 ) in both integrals on the right, then evaluate the last formula at
t0 , and cancel v(t0 ). Obtain
t2 < (t2 t0 )
<
t0
t0
sq + (s) ds + t0
(t2 s)sq (s) ds +
t2
t0
t2
t0
(t2 s)q + (s) ds
s(t2 s)q (s) ds =
t2
(t2 s)sq + (s) ds .
Dividing by t2 , gives
1<
t2
s
(1 )sq + (s) ds
t2
s
(1 )sq + (s) ds <
b
b +
q (s) ds ,
4
which implies our inequality. (On the last step we estimated the function
s
b
(1 )s by its maximum value of .)
b
4
Theorem 2.15.3 (Lyapunovs inequality) If a non-trivial solution of the
equation
u00 + q(t)u = 0
has two roots on an interval (a, b), then
Z
q + (t) dt >
4
.
ba
Proof:
Let a t1 < t2 b be two consecutive roots of u(t). We may
assume that u(t) > 0 on (t1 , t2 ), and use the above lemma
Z
q + (t) dt
t2
t1
q + (t) dt >
4
4
.
t2 t1
ba
(We may declare the point t1 to be the origin, to use the above lemma.)
Chapter 3
Using Infinite Series to Solve

Differential Equations
3.1
3.1.1
Series Solution Near a Regular Point

Maclauren and Taylor Series
Infinitely differentiable functions can often be represented by a series

(1.1)
f (x) = a0 + a1 x + a2 x2 + + an xn + .
Letting x = 0, we see that a0 = f (0). If we differentiate (1.1), and then

let x = 0, we have a1 = f 0 (0). Differentiating (1.1) twice, and then letting
f 00 (0)
f (n) (0)
x = 0, gives a2 =
. Continuing this way, we see that an =
,
2
n!
giving us the Maclauren series
f (x) = f (0) + f 0 (0)x +
X
f 00 (0) 2
f (n) (0) n
f (n) (0) n
x ++
x +=
x .
2!
n!
n!
n=0
It is known that for each f (x) there is a number R, so that the Maclauren
series converges inside the interval (R, R), and diverges when |x| > R. R
is called the radius of convergence. For some f (x), we have R = (for
example, for sin x, cos x, ex ), while for some series we have R = 0, and in
general 0 R .
Computing the Maclauren series for some specific functions, gives:
sin x = x
X
x3 x5 x7
x2n+1
+
+ =
(1)n
,
3!
5!
7!
(2n + 1)!
n=0
115
116CHAPTER 3. USING INFINITE SERIES TO SOLVE DIFFERENTIAL EQUATIONS

X
x2
x4 x6
x2n
cos x = 1
+
+ =
(1)n
,
2!
4!
6!
(2n)!
n=0
ex = 1 + x +
X
xn
x2 x3 x4
+
+
+ =
,
2!
3!
4!
n!
n=0
1
= 1 + x + x2 + + xn + .
1x
The last series, called the geometric series, converges on the interval (1, 1),
so that R = 1.
Maclaurens series gives an approximation of f (x), for x close to zero.
For example, sin x x gives a reasonably good approximation for |x| small.
x3
If we add one more term of the Maclauren series: sin x x , then, say
6
on the interval (0.3, 0.3), we get an excellent approximation.
If one needs the Maclauren series for sin x2 , one begins with a series for
sin x, and then replaces each x by x2 , obtaining
sin x2 = x2
X
x6 x10
x4n+2
+
+ =
(1)n
.
3!
5!
(2n + 1)!
n=0
One can split Maclaurens series into a sum of series with either even or
odd powers:
X
f (n) (0)
n=0
n!
X
f (2n)(0)
xn =
n=0
(2n)!
x2n +
X
f (2n+1) (0)
n=0
(2n + 1)!
x2n+1 .
In the following series only the odd powers have non-zero coefficients
X
1 (1)n
n=1
xn =
2
2
2
x + x3 + x5 +
1
3
5
X
2
1
x2n+1 = 2
x2n+1 .
2n
+
1
2n
+
1
n=0
n=0
All of the series above were centered at 0. One can replace zero by any
number a, obtaining Taylors series
f (x) = f (a) + f 0 (a)(x a) +
=
f 00 (a)
f (n) (a)
(x a)2 + +
(x a)n +
2!
n!
X
f (n) (a)
n=0
n!
(x a)n .
3.1. SERIES SOLUTION NEAR A REGULAR POINT
117
It converges on an interval (a R, a + R), centered at a. The radius of

convergence satisfies 0 R , as before. Taylors series allows us to
approximate f (x), for x close to a. For example, one can usually expect
00
(but this is not always true) that f (x) f (2) + f 0 (2)(x 2) + f 2!(2) (x 2)2,
for x close to 2, say for 1.8 < x < 2.2.
3.1.2
A Toy Problem
Let us begin with the equation (here y = y(x))

y 00 + y = 0 ,
for which we know the general solution. Let us denote by y1 (x) the solution
of the initial value problem
y 00 + y = 0, y(0) = 1, y 0 (0) = 0 .
(1.2)
By y2 (x) we denote the solution of the same equation, together with the
initial conditions y(0) = 0, y 0 (0) = 1. Clearly, y1 (x) and y2 (x) are not
constant multiples of each other. Therefore, they form a fundamental set,
giving us the general solution y(x) = c1 y1 (x) + c2 y2 (x).
Let us now compute y1 (x), the solution of (1.2). From the initial conditions, we already know the first two terms of its Maclauren series
y(x) = y(0) + y 0 (0)x +
y 00 (0) 2
y (n) (0) n
x ++
x + .
2
n!
To get more terms, we need to compute the derivatives of y(x) at zero. From
the equation (1.2), y 00 (0) = y(0) = 1. We now differentiate the equation
(1.2): y 000 + y 0 = 0, and then set x = 0:
y 000(0) = y 0 (0) = 0 .
Differentiating again, we have y 0000 + y 00 = 0, and setting x = 0,
y 0000(0) = y 00 (0) = 1 .
On the next step:
y (5)(0) = y 000 (0) = 0 .
We see that all derivatives of odd order vanish at x = 0, while the derivatives
of even order alternate between 1 and 1. We write down the Maclauren
series:
x2
x4
y1 (x) = 1
+
= cos x .
2!
4!

Similarly, we compute the series representation for y2 (x):
y2 (x) = x
x3 x5
+
= sin x .
3!
5!
We shall solve the equations with variable coefficients

P (x)y 00 + Q(x)y 0 + R(x)y = 0 ,
(1.3)
where continuous functions P (x), Q(x) and R(x) are given. We shall always
denote by y1 (x) the solution of (1.3) satisfying the initial conditions y(0) = 1,
y 0 (0) = 0, and by y2 (x) the solution of (1.3) satisfying the initial conditions
y(0) = 0, y 0 (0) = 1. If one needs to solve (1.3), together with the given
initial conditions
y(0) = , y 0 (0) = ,
then the solution is
y(x) = y1 (x) + y2 (x) .
3.1.3
Using Series When Other Methods Fail
Let us try to find the general solution of the equation

y 00 + xy 0 + 2y = 0 .
This equation has variable coefficients, and none of the previously considered
methods will apply here.
We shall derive a formula for y (n) (0), and use it to calculate y1 (x) and
y2 (x). From the equation we express y 00 (0) = 2y(0). We now differentiate
the equation
y 000 + xy 00 + 3y 0 = 0 ,
which gives y 000(0) = 3y 0 (0). Differentiate the last equation again
y 0000 + xy 000 + 4y 00 = 0 ,
and get: y 0000 (0) = 4y 00 (0). It is clear that, in general, we get a recurrence
relation
y (n) (0) = ny (n2) (0), n = 3, 4, . . . .
Let us begin with the computation of y2 (x), for which we use the initial
conditions y(0) = 0, y 0 (0) = 1. Then using, the recurrence relation,
y 00 (0) = 2y(0) = 0 ,
119

y 000 (0) = 3y 0 (0) = 3 1 ,
y 0000(0) = 4y 00 (0) = 0 .
It is clear that all derivatives of even order are zero at x = 0. Let us continue
with the derivatives of odd order:
y (5)(0) = 5y 000 (0) = (1)2 5 3 1 ,
y (7)(0) = 7y 000(0) = (1)3 7 5 3 1 .
And in general,
y (2n+1) (0) = (1)n (2n + 1) (2n 1) 3 1 .
Then the Maclauren series for y2 (x) is
y2 (x) =
X
y (n) (0)
n!
n=0
=x+
(1)n
n=1
=x+
X
y (2n+1) (0)
n=1
(2n + 1)!
x2n+1
(2n + 1) (2n 1) 3 1 2n+1

x
(2n + 1)!
(1)n
n=1
xn = x +
1
x2n+1 .
2n (2n 2) 4 2
One can also write this solution as y2 (x) =
(1)n
n=0
1
2n n!
x2n+1 .
To compute y1 (x), we use the initial conditions y(0) = 1, y 0 (0) = 0. Similarly to the above, we see from the recurrence relation that all derivatives
of odd order vanish at x = 0, while the even ones satisfy
y (2n)(0) = (1)n 2n (2n 2) 4 2 , for n = 1, 2, 3, . . . .
This leads to
y1 (x) = 1 +
n=1
(1)n
1
x2n .
(2n 1) (2n 3) 3 1
The general solution:

y(x) = c1 1 +
n=1
X
(1)n
1
x2n + c2
(1)n n x2n+1 .
(2n 1)(2n 3) 3 1
2
n!
n=0

We shall need a formula for repeated differentiation of a product of two
functions. Starting with the product rule (f g)0 = f 0 g + f g 0 , we have
(f g)00 = f 00 g + 2f 0 g 0 + f g 00 ,
(f g)000 = f 000 g + 3f 00 g 0 + 3f 0 g 00 + f g 000 ,
and in general, for the n-th derivative:
(f g)(n) = f (n) g + nf (n1) g 0 +
(1.4)
n(n 1) (n2) 00
f
g +
2
n(n 1) 00 (n2)
f g
+ nf 0 g (n1) + f g (n) .
2
(Convention: f (0) = f .) This formula is similar to the binomial formula for

the expansion of (x + y)n . Using the summation notation, we can write it
as
!
n
X
n (nk) (k)
(n)
(f g) =
f
g ,
k
k=0
where
n
k
n!
are the binomial coefficients.
k!(n k)!
The formula (1.4) simplifies considerably in case f (x) = x, or if f (x) = x2 :

(xg)(n) = ng (n1) + xg (n) ;
(x2 g)(n) = n(n 1)g (n2) + 2nxg (n1) + x2 g (n) .
We shall use infinite series to solve linear second order equations with
variable coefficients
P (x)y 00 + Q(x)y 0 + R(x)y = 0 ,
where the continuous functions P (x), Q(x) and R(x) are given. A number
a is called a regular point if P (a) 6= 0. If P (a) = 0, then x = a is called a
singular point.
Example (2 + x2 )y 00 xy 0 + 4y = 0.
For this equation, any a is a regular point. Let us find the general solution
as an infinite series, centered at a = 0, the Maclauren series for y(x). We
differentiate both sides of this equation n times. When we use the formula
(1.4) to differentiate the first term, only three terms are non-zero, because
121
the derivatives of 2 + x2 , of order three and higher, are zero. When we

differentiate xy 0 , only two terms survive. We have
n(n 1) (n)
2y + n(2x)y (n+1) + (2 + x2 )y (n+2) ny (n) xy (n+1) + 4y (n) = 0 .
2
Set here x = 0. Several terms vanish. Combining the like terms, we get

2y (n+2) (0) + n2 2n + 4 y (n)(0) = 0 ,

which gives us the recurrence relation
y
(n+2)
n2 2n + 4 (n)
(0) =
y (0) .
2
This relation is too involved to get a neat formula for y (n) (0) as a function
of n. However, it can be used to crank out the derivatives at zero, as many
as you wish. To compute y1 (x), we use the initial conditions y(0) = 1 and
y 0 (0) = 0. We see from the recurrence relation that all of the derivatives of
odd order are zero at x = 0. Setting n = 0 in the recurrence relation, we
have
y 00 (0) = 2y(0) = 2 .
When n = 2,
y 0000(0) = 2y 00 (0) = 4 .
Using these derivatives in the Maclauren series, we have
1
y1 (x) = 1 x2 + x4 + .
6
To compute y2 (x), we use the initial conditions y(0) = 0 and y 0 (0) = 1.
We see from the recurrence relation that all of the derivatives of even order
are zero. When n = 1, we get
3
3
y 000(0) = y 0 (0) = .
2
2
Setting n = 3, we have
7
21
y (5)(0) = y 000 (0) =
.
2
4
Using these derivatives in the Maclauren series, we conclude
1
7 5
y2 (x) = x x3 +
x + .
4
160

The general solution:
1
y(x) = c1 y1 (x) + c2 y2 (x) = c1 1 x2 + x4 +
6

+c2
1
7 5
x + .
x x3 +
4
160

Suppose that we wish to solve the above equation, together with the
initial conditions: y(0) = 2, y 0 (0) = 3. Then y(0) = c1 y1 (0) + c2 y2 (0) =
c1 = 2, and y 0 (0) = c1 y10 (0) + c2 y20 (0) = c2 = 3. It follows that y(x) =
2y1 (x) + 3y2 (x). If we need to approximate y(x) near x = 0, say on the
interval (0.3, 0.3), then
1
1
7 5
y(x) 2 1 x + x4 + 3 x x3 +
x
6
4
160

will provide an excellent approximation.

Example Approximate the general solution of Airys equation
y 00 x y = 0 ,
near x = 1. This equation was encountered in 1838 by G.B. Airy, in his
study of optics.
X
y (n)(1)
(x1)n .
We need to compute Taylors series about a = 1: y(x) =
n!
n=0
From the equation,
y 00 (1) = y(1) .
To get higher derivatives, we differentiate our equation n times, and then
set x = 1
y (n+2) ny (n1) xy (n) = 0;
y (n+2) (1) = ny (n1) (1) + y (n)(1), n = 1, 2, . . . .
To compute y1 (x), we use the initial conditions y(1) = 1, y 0 (1) = 0.

Then
y 00 (1) = y(1) = 1 .
Setting n = 1 in the recurrence relation:
y (3)(1) = y(1) + y 0 (1) = 1 .
When n = 2,
y (4)(1) = 2y 0 (1) + y 00 (1) = 1 .
3.2. SOLUTION NEAR A MILDLY SINGULAR POINT
123
Then, for n = 3,
y (5)(1) = 3y 00(1) + y 000 (1) = 4 .
We have
y1 (x) = 1 +
(x 1)2 (x 1)3 (x 1)4 (x 1)5

+
+
+
+ .
2
6
24
30
Again, it does not seem possible to get a simple formula for the coefficients.
To compute y2 (x), we use the initial conditions y(1) = 0, y 0 (1) = 1.
Then y 00 (1) = y(1) = 0. Setting n = 1 in the recurrence relation: y (3)(1) =
y(1) + y 0 (1) = 1. When n = 2, y (4)(1) = 2y 0 (1) + y 00 (1) = 2. Then, for n = 3,
y (5)(1) = 3y 00 (1) + y 000(1) = 1. We have
y2 (x) = x 1 +
(x 1)3 (x 1)4 (x 1)5

+
+
+ .
6
12
120
The general solution is, of course, y(x) = c1 y1 (x) + c2 y2 (x).
3.2
Solution Near a Mildly Singular Point
We consider again the equation

P (x)y 00 + Q(x)y 0 + R(x)y = 0 ,
with given functions P (x), Q(x) and R(x) that are continuous near a point a,
X
y (n) (a)
at which we want to compute solution as a series y(x) =
(x a)n.
n!
n=0
If P (a) = 0, we have a problem: one cannot compute y 00 (a) from the equation
(and the same problem for higher derivatives). However, if a is a simple
root of P (x), it turns out that one can still use series to produce a solution.
Namely, we assume that P (x) = (x a)P1 (x), with P1 (a) 6= 0. We call
x = a a mildly singular point. Dividing the equation by P1 (x), and calling
q(x) = PQ(x)
, r(x) = PR(x)
, we put it into the form
1 (x)
1 (x)
(x a)y 00 + q(x)y 0 + r(x)y = 0 .
The functions q(x) and r(x) are continuous near a. In case a = 0, we have
(2.1)
xy 00 + q(x)y 0 + r(x)y = 0 .
It turns out that for this equation we cannot expect to obtain two linearly
independent solutions, by prescribing y1 (0) = 1, y10 (0) = 0, and y2 (0) = 0,

y20 (0) = 1, the way we did before. This equation is singular at x = 0 (the
r(x)
functions q(x)
x and x are discontinuous at x = 0, and so the existence and
uniqueness theorem does not apply).
Example Let us try to solve:
xy 00 y 0 = 0, y(0) = 0, y 0 (0) = 1 .
Multiplying through by x, we obtain Eulers equation, which we solve, to
get the general solution: y(x) = c1 x2 +c2 . Then y 0 (x) = 2c1 x, and y 0 (0) = 0.
This initial value problem has no solution. However, if we change the initial
conditions, and consider the problem
xy 00 y 0 = 0, y(0) = 1, y 0 (0) = 0 ,
then we have infinitely many solutions y = 1 + c1 x2 .
We therefore lower our expectations, and we shall be satisfied to compute
just one series solution of (2.1). It turns out that in most cases it is possible
P
n
to calculate a series solution of the form y(x) =
n=0 an x , starting with
1
a0 = 1. (If a0 6= 1, then a0 y(x) is another solution of (2.1), which begins
with 1. So that we shall assume that a0 = 1. The possibility of a0 = 0 is
considered later.)
Example Find a series solution, centered at a = 0, of
xy 00 + 3y 0 2y = 0 .
It is convenient to multiply this equation by x:
x2 y 00 + 3xy 0 2xy = 0 .
We look for solution in the form
y=
an xn .
n=0
Then
0
y =
an nxn1 ,
n=1
y 00 =
n=2
an n(n 1)xn2 .
125

Observe that each differentiation kills a term. Plugging y =
into the last equation, gives
(2.2)
n=2
an n(n 1)xn +
n=1
3an nxn
n=0
an xn
2an xn+1 = 0 .
n=0
The third series is not lined up with the other two. We therefore replace
n by n 1 in that series, obtaining
2an xn+1 =
n=0
2an1 xn .
n=1
We rewrite (2.2):
(2.3)
n=2
an n(n 1)xn +
n=1
3an nxn
2an1 xn = 0 .
n=1
We shall use the following fact: if n=1 bn xn = 0 for all x, then bn = 0 for
all n = 1, 2, . . .. Our goal is combine the three series in (2.3) into a single
one, so that we can set all of the resulting coefficients to zero. The x term is
present in the second and the third series, but not in the first. However, we
can start the first series at n = 1, because at n = 1 the coefficient is zero.
So that we have
P
n=1
an n(n 1)xn +
n=1
3an nxn
2an1 xn = 0 .
n=1
Now for all n 1, the xn term is present in all three series, so that we can
combine these series into one series. We therefore just lift the coefficients:
an n(n 1) + 3an n 2an1 = 0 .
We solve for an :
an =
2
an1 , n 1 .
n(n + 2)
Starting with a0 = 1, compute a1 =

a2 =
2
,
13
2
22
23
a1 =
=
,
24
(1 2)(3 4)
2! 4!
a3 =
2
24
a2 =
,
35
3! 5!

2n+1
.
n! (n + 2)!
X
X
2n+1
2n+1
n
x =
xn .
Answer: y(x) = 1 +
n!
(n
+
2)!
n!
(n
+
2)!
n=0
n=1
and, in general, an =
Example Find a series solution, centered at a = 0, of

xy 00 3y 0 2y = 0 .
This is a small modification of the preceding problem, so that we can quickly

derive the recurrence relation:
(2.4)
an =
2
an1 , n 1 .
n(n 4)
If we start with a0 = 1, and proceed as before, then at n = 4 the denominator

is zero, and the computation stops! To avoid the trouble at n = 4, we look
for the solution in the form y =
an xn . Plugging this series into the
n=4
equation
x2 y 00 3xy 0 2xy = 0
(which is the original equation, multiplied by x), gives
(2.5)
n=4
an n(n 1)xn
n=4
3an nxn
2an xn+1 = 0 .
n=4
The coefficient in x4 , which is a4 (4 3 3 4), is zero for any choice of a4 .

Shifting n n 1 in the last series in (2.5), we see that the recurrence
relation (2.4) holds for n 5. We choose a4 = 1, and use the recurrence
relation (2.4) to calculate a5 , a6 , etc.
2
2
2
22
22
Compute: a5 =
a4 =
, a6 =
a5 =
= 24
, and in
51
51
62
6521
6!2!
n4
2
general, an = 24
.
n!(n 4)!
Answer: y(x) = x4 + 24
n=5
X
2n4
2n4
xn = 24
xn .
n!(n 4)!
n!(n
4)!
n=4
Our experience with the previous two problems can be summarized as

follows.
Theorem 3.2.1 Consider the equation (2.1). Assume that the functions
p(x) and q(x) have convergent Maclauren series expansions on some interval (, ). If q(0) is not a non-positive integer (q(0) is not equal to
127

0, 1, 2, . . .), one can find a series solution of the form y(x) =
an xn ,
n=0
starting with a0 = 1. In case q(0) = k, where k is a non-negative integer,

one can find a series solution of the form y(x) =
an xn , starting with
n=k+1
ak+1 = 1. In both cases, the series for y(x) is convergent on (, ).

Example x2 y 00 + xy 0 + (x2 2 )y = 0.
This is the Bessel equation, of great importance in Mathematical Physics!
It depends on a real parameter . It is also called Bessels equation of order
. Its solutions are called Bessels functions of order . We see that a = 0
is not a mildly singular point, for 6= 0. (Zero is a double root of x2 .) But
in case = 0, we can cancel x, putting Bessels equation of order zero into
the form:
(2.6)
xy 00 + y 0 + xy = 0 ,
so that a = 0 is a mildly singular point. We shall find the solution of this
equation, as a series centered at a = 0, the Maclauren series for y(x).
We put the equation (2.6) back into the form
x2 y 00 + xy 0 + x2 y = 0 ,
and look for the solution in the form y =
in:
n=2
an n(n 1)xn +
n=0
an nxn +
n=1
an xn , with a0 = 1. Plug this
an xn+2 = 0 .
n=0
In the last series we replace n by n 2:
n=2
an n(n 1)xn +
an nxn +
n=1
an2 xn = 0 .
n=2
None of the series has a constant term. The x term is present only in the
second series. Its coefficient is a1 , and so
a1 = 0 .
The terms involving xn , starting with n = 2, are present in all series, so that
(after combining the series)
an n(n 1) + an n + an2 = 0 ,

1.0
J0 HxL
0.8
0.6
0.4
0.2
10
15
-0.2
-0.4
Figure 3.1: The graph of Bessels function J0 (x)
giving
1
an2 .
n2
This recurrence relation tells us that all odd coefficients are zero, a2n+1 =
1
1
1
0. Starting with a0 = 1, compute a2 = 2 a0 = 2 , a4 = 2 a2 =
2
2
4
1
1
1
(1)2 2 2 , a6 = 2 a4 = (1)3 2 2 2 , and in general,
2 4
6
2 4 6
an =
1
1
1
n
= (1)n
2 = (1) 22n(n!)2 .
22 42 62 (2n)2
(2 4 6 2n)
a2n = (1)n
We then have
y =1+
n=1
a2n x2n = 1 +
n=1
(1)n
X
1
1
2n
x
=
(1)n 2n
x2n .
2
22n (n!)2
2
(n!)
n=0
We have obtained Bessels function of order zero of the first kind, with the
X
1
customary notation: J0 (x) =
(1)n 2n
x2n .
2
2
(n!)
n=0
3.2.1 Derivation of J0 (x) by Differentiation of the Equation

We differentiate our equation (2.6) n times, and then set x = 0:
ny (n+1) + xy (n+2) + y (n+1) + ny (n1) + xy (n) = 0 ;
3.3. MODERATELY SINGULAR EQUATIONS
129
ny (n+1) (0) + y (n+1) (0) + ny (n1) (0) = 0 .

(It is not always true that xy (n+2) 0 as x 0. However, in case of the
initial conditions y(0) = 1, y 0 (0) = 0 that is true, as was justified in authors
paper [13].) We get the recurrence relation
y (n+1) (0) =
n (n1)
y
(0) .
n+1
We use the initial conditions y(0) = 1, y 0 (0) = 0. Then all derivatives of

odd order vanish, while
y (2n)(0) =
= (1)n
Then
2n 1 (2n2)
2n 1 2n 3 (2n4)
y
(0) = (1)2
y
(0) = . . .
2n
2n 2n 4
(2n 1)(2n 3) 3 1
(2n 1)(2n 3) 3 1
y(0) = (1)n
.
2n(2n 2) 2
2n n!
y(x) =
X
y (2n) (0)
n=0
(2n)!
2n
n=0
(1)n
1
x2n .
22n (n!)2
We obtained again Bessels function of order zero of the first kind, J0 (x) =
X
1
(1)n 2n
x2n .
2
2
(n!)
n=0
In case of the initial conditions y(0) = 0 and y 0 (0) = 1, there is no
solution (the recurrence relation above is not valid, because the relation
xy (n+2) 0 as x 0 is not true in this case). In fact, the second solution
of Bessels equation cannot be continuously differentiable
at x = 0. Indeed,
R
c
x1 dx
the Wronskian of any two solutions is equal to ce
= . We have
x
W (y1 , y2 ) = y1 y20 y10 y2 = xc . The solution J0 (x) satisfies J0 (0) = 1, J00 (0) =
0. Therefore, the other solution, or its derivative, must be discontinuous
at x = 0. It turns out that the other solution, called Bessels function of
the second type, and denoted Y0 (x), has a term involving ln x in its series
representation, see e.g., [3].
3.3
Moderately Singular Equations
We shall deal only with the Maclauren series, so that a = 0, and the general
case is similar. We consider the equation
(3.1)
x2 y 00 + xp(x)y 0 + q(x)y = 0 ,

where p(x) and q(x) are given infinitely differentiable functions, that can be
represented by the Maclauren series
1
p(x) = p(0) + p0 (0)x + p00 (0)x2 + ,
2
1
q(x) = q(0) + q 0 (0)x + q 00 (0)x2 + .
2
(Observe the special form of the coefficient function in front of y 0 .) If it so
happens that q(0) = 0, then q(x) has a factor of x, and we can divide
the equation (3.1) by x, to obtain a mildly singular equation. So that
the difference with the preceding section is that we now allow the case of
q(0) 6= 0. Observe also that in case p(x) and q(x) are constants, the equation
(3.1) is Eulers equation, that we have studied before. This connection with
Eulers equation is the guiding light of the theory that follows.
We change to a new unknown function v(x), by letting y(x) = xr v(x),
with a constant r to be specified. With y 0 = rxr1 v + xr v 0 , and
y 00 = r(r 1)xr2 v + 2rxr1 v 0 + xr v 00 , we plug y(x) into (3.1), obtaining
(3.2)
xr+2 v 00 + xr+1 v 0 (2r + p(x)) + xr v [r(r 1) + rp(x) + q(x)] = 0 .
We now choose r, to satisfy the following characteristic equation

(3.3)
r(r 1) + rp(0) + q(0) = 0 .
The quantity in the square bracket in (3.2) then becomes

1
1
rp0 (0)x + r p00 (0)x2 + + q 0 (0)x + q 00 (0)x2 + ,
2
2
so that it has a factor of x. We take this factor out, and divide the equation
(3.2) by xr+1 , obtaining a mildly singular equation, that we analyzed in the
preceding section.
Example
2x2 y 00 xy 0 + (1 + x)y = 0.
To put this equation into the right form, we divide it by 2:

(3.4)
1
1 1
x2 y 00 xy 0 + ( + x)y = 0 .
2
2 2
Here p(x) = 12 and q(x) =
1
2
+ 21 x. The characteristic equation is then
1
1
r(r 1) r + = 0 .
2
2
131

Its roots are r = 21 , and r = 1.
1
The case r = 12 . We know that the substitution y = x 2 v will produce a

mildly singular equation for v(x). Plugging this into our equation (3.4) (or
using (3.2)), we get:
1
1
x5/2 v 00 + x3/2 v 0 + x3/2 v = 0 .
2
2
Dividing by x3/2 , we have a mildly singular equation
1
1
xv 00 + v 0 + v = 0 .
2
2
We multiply this equation by 2x, for convenience,
2x2 v 00 + xv 0 + xv = 0 ,
(3.5)
and look for solution in the form v(x) =

(3.5), we get as before
n=2
2an n(n 1)x +
n=0
an nx +
n=1
an xn . Plugging v(x) into
an xn+1 = 0 .
n=0
To line up the powers, we shift n n 1 in the last series. The first series
we may begin at n = 1, instead of n = 2, because its coefficient at n = 1 is
zero. We have
n=1
2an n(n 1)xn +
an nxn +
n=1
an1 xn = 0 .
n=1
We can now combine these series into a single series. Setting its coefficients
to zero, we have
2an n(n 1) + an n + an1 = 0 ,
which gives us the recurrence relation
an =
1
an1 .
n(2n 1)

a3 =
1
1
1
, a2 =
a1 = (1)2
,
11
23
(1 2) (1 3)
1
1
a2 = (1)3
, and in general
35
(1 2 3) (1 3 5)
an = (1)n
1
.
n! 1 3 5 (2n 1)

We have the first solution:
y1 (x) = x
1/2
v(x) = x
1/2
"
1
xn .
1+
(1)
n!
(2n
1)
n=1
n
The case r = 1. We set y = xv. Plugging this into (3.4), and simplifying
3
1
x3 v 00 + x2 v 0 + x2 v = 0 .
2
2
Dividing by x2 , we have a mildly singular equation
3
1
xv 00 + v 0 + v = 0 .
2
2
We multiply this equation by 2x, for convenience,
2x2 v 00 + 3xv 0 + xv = 0 ,
(3.6)
and look for solution in the form v =

we get:
2an n(n 1)xn +
2an n(n 1)xn +
n=2
n=0
an xn . Plugging v(x) into (3.6),
3an nxn +
n=1
an xn+1 = 0 .
n=0
As before, we start the first series at n = 1, and make a shift n n 1 in

the last series:
n=1
3an nxn +
n=1
an1 xn = 0 .
n=1
Setting the coefficient of xn to zero,

2an n(n 1) + 3an n + an1 = 0 ,
gives us the recurrence relation
an =
1
an1 .
n(2n + 1)

a3 =
1
1
1
, a2 =
a1 = (1)2
,
13
25
(1 2) (1 3 5)
1
1
a2 = (1)3
, and in general
37
(1 2 3) (1 3 5 7)
an = (1)n
1
.
n! 1 3 5 (2n + 1)
133

We have the second solution:
"
1
y2 (x) = x 1 +
(1)n
xn .
n!
(2n
+
1)
n=1
The general solution is, of course, y(x) = c1 y1 + c2 y2 .
1
x2 y 00 + xy 0 + (x2 )y = 0.
9
This is Bessels equation of order 13 . Here p(x) = 1, and q(x) = x2 19 . The
characteristic equation
1
r(r 1) + r = 0
9
Example
has roots r = 13 , and r = 31 .

1
The case r = 13 . We set y = x 3 v. Compute y 0 = 13 x 3 v + x 3 v 0 ,

1
7
4
y 00 = 49 x 3 v 32 x 3 v 0 + x 3 v 00 . Plugging this in and simplifying, we get a
mildly singular equation
1
xv 00 + v 0 + xv = 0 .
3
We multiply by 3x
(3.7)
3x2 v 00 + xv 0 + 3x2 v = 0 ,

(3.7), gives
n=2
3an n(n 1)xn +
n=0
an xn . Plugging this series into
an nxn +
n=1
3an xn+2 = 0 .
n=0
We shift n n 2 in the last series:
n=2
3an n(n 1)xn +
n=1
an nxn +
3an2 xn = 0 .
n=2
The x term is present only in the second series. Its coefficient must be zero,
so that
(3.8)
a1 = 0 .
The term xn , with n 2, is present in all three series. Setting its coefficient
to zero
3an n(n 1) + an n + 3an2 = 0 ,

gives us the recurrence relation
an =
3
an2 .
n(3n 2)
All odd coefficients are zero (because of (3.8)), while for the even ones, our
recurrence relation gives
3
a2n2 .
a2n =
2n(6n 2)
Starting with a0 = 1, we compute a2 =
a6 = (1)3
3
32
, a4 = (1)2
,
24
(2 4) (4 10)
32
, and in general,
(2 4 6) (4 10 16)
a2n = (1)n
We have the first solution:

y1 (x) = x
1/3
"
3n
.
(2 4 2n) (4 10 (6n 2))
#
3n
1+
(1)
x2n .
(2
2n)
(4
10
(6n
2))
n=1
The case r = 31 . We set y = x 3 v. Compute y 0 = 31 x 3 v + x 3 v 0 ,

1
5
2
y 00 = 29 x 3 v + 32 x 3 v 0 + x 3 v 00 . Plugging this in and simplifying, we get a
mildly singular equation
5
xv 00 + v 0 + xv = 0 .
3
We multiply by 3x
3x2 v 00 + 5xv 0 + 3x2 v = 0 ,
n
n=0 an x . Plugging this in, we
conclude as before that a1 = 0, and the following recurrence relation holds
3
an =
an2 .
n(3n + 2)
It follows that all odd coefficients are zero, while the even ones satisfy
3
1
a2n =
a2n2 =
a2n2 .
2n(6n + 2)
2n(2n + 2/3)
We then derive the second solution
y2 (x) = x
1/3
"
1
1+
(1)
x2n .
(2
2n)
((2
+
2/3)
(4
+
2/3)
(2n
+
2/3))
n=1
n
135
3.3.1
Problems
I. Find the Maclauren series of the following functions, and state their radius
of convergence
1
3
1. sin x2 .
2.
.
3. ex .
1 + x2
II. 1. Find the Taylor series of f (x) centered at a.
(ii) f (x) = ex , a = 1.
(i) f (x) = sin x, a = 2 .

2. Show that
X
1 + (1)n
n2
n=1
3. Show that
xn =
(iii) f (x) =
1
, a = 1.
x
1X
1 2n
x .
2 n=1 n2
X
n+3
n+2
xn+1 =
xn .
n!(n
+
1)
(n
1)!n
n=0
n=1
4. Expand the n-th derivative: (x2 + x)g(x)

h
5. Find the n-th derivative: (x2 + x)e2x
i(n)
i(n)
III. Find the general solution, using power series centered at a (find the
recurrence relation, and two linearly independent solutions).
1. y 00 xy 0 y = 0, a = 0.
Answer. y1 (x) =
X
x2n
n=0
2n n!
, y2 (x) =
2n n!
x2n+1 .
(2n
+
1)!
n=0
2. y 00 xy 0 + 2y = 0, a = 0.
1
1 5
Answer. y1 (x) = 1 x2 , y2 (x) = x x3
x .
6
120
3. y 00 xy 0 y = 0, a = 1.
1
1
1
Answer. y1 (x) = 1 + (x 1)2 + (x 1)3 + (x 1)4 + ,
2
6
6
1
1
1
2
3
y2 (x) = (x 1) + (x 1) + (x 1) + (x 1)4 + .
2
2
4
4. (x2 + 1)y 00 + xy 0 + y = 0, a = 0.
Answer. The recurrence relation: y (n+2) (0) = (n2 + 1)y (n)(0).
1
1
5 4
y1 (x) = 1 12 x2 + 24
x , y2 (x) = x x3 + x5 .
3
6

IV. 1. Find the solution of the initial value problem, using power series
centered at 2
y 00 2xy = 0, y(2) = 1, y 0 (2) = 0 .
1
2
Answer. y = 1 + 2(x 2)2 + (x 2)3 + (x 2)4 + .
3
3
2. Find the solution of the initial value problem, using power series centered
at 1
y 00 + xy = 0, y(1) = 2, y 0 (1) = 3 .
V. Find one series solution of the following mildly singular equations (here
a = 0)
1. 2xy 00 + y 0 + xy = 0.
Answer. y = 1
=1+
x4
x6
x2
+
+
2 3 2 4 3 7 2 4 6 3 7 11
(1)n x2n
.
2n n! 3 7 11 (4n 1)
n=1
2. xy 00 + y 0 y = 0.
Answer. y =
X
xn
n=0
(n!)2
3. xy 00 + 2y 0 + y = 0.
Answer. y =
(1)n
xn .
n!(n
+
1)!
n=0
4. xy 00 + y 0 2xy = 0.
X
1
1
2n
Answer. y = 1 +
x =
x2n .
n (n!)2
n (n!)2
2
2
n=1
n=0
VI. 1. Find one series solution in the form
n=5
singular equation (here a = 0)
xy 00 4y 0 + y = 0 .
Answer. y = x5 + 120
X
(1)n5
n!(n 5)!
n=6
xn .
an xn of the following mildly
137
2. Find one series solution of the following mildly singular equation (here
a = 0)
xy 00 2y 0 2y = 0 .
3. Find one series solution of the following mildly singular equation (here
a = 0)
xy 00 + y = 0 .
Hint: Look for solution in the form
an xn , starting with a1 = 1.
n=1
Answer. y =
(1)n1
n!(n 1)!
n=1
xn .
4. Recall that J0 (x) is a solution of

xy 00 + y 0 + xy = 0 .
2
Show that the energy E(x) = y 0 (x) + y 2 (x) is a decreasing function.

Conclude that each maximum value of J0 (x) is greater than the absolute
value of the minimum value that follows it, which in turn is larger than the
next maximum value, and so on (see the graph of J0 (x)).
5. Show that the absolute value of the slope of the tangent line decreases at
each consecutive root of J0 (x).
Hint: Use the energy function E(x) from the preceding problem.
VII. 1. Verify that the Bessel equation of order 1/2
x2 y 00 + xy 0 + (x2 1/4)y = 0
has a moderate singularity at zero. Write down the characteristic equation,
and find its roots.
(i) Corresponding to the root r = 1/2, perform a change of variables y =
x1/2 v, and obtain a mildly singular equation for v(x). Solve that equation,
to obtain one of the solutions of the Bessel equation.
Answer. y = x
x1/2 sin x.
1/2
"
1+
X
(1)n x2n
n=1
(2n + 1)!
= x
1/2
"
x+
X
(1)n x2n+1
n=1
(2n + 1)!
(ii) Corresponding to the root r = 1/2, perform a change of variables

y = x1/2 v, and obtain a mildly singular equation for v(x). Solve that
equation, to obtain the second solution of the Bessel equation.

Answer. y = x1/2 cos x.
(iii) Find the general solution.
Answer. y = c1 x1/2 sin x + c2 x1/2 cos x.
2. Find the general solution of the Bessel equation of order 3/2
x2 y 00 + xy 0 + (x2 9/4)y = 0 .
Chapter 4
The Laplace Transform

4.1
4.1.1
The Laplace Transform And Its Inverse

Review of Improper Integrals
The mechanics of computing the integrals, involving infinite limits, is similar

to that for integrals with finite end-points. For example,
Z
1
1
e2t dt = e2t |0 = .
2
2
Here we did not plug in the upper limit t = , but rather computed the
limit as t (the limit is zero). This is an example of a convergent
integral. On the other hand, the integral
Z
dt = ln t |1
t
is divergent, because ln t has an infinite limit as t . When computing

improper integrals, we use the same techniques of integration, in essentially
the same way. For example
Z
1
1
te2t dt = te2t e2t |0 = .
2
4
4

Here the antiderivative is computed by the guess-and-check method (or by

integration by parts). The limit at infinity is computed by LHospitals rule
to be zero.
139
140
CHAPTER 4. THE LAPLACE TRANSFORM
4.1.2
The Laplace Transform
Let the function f (t) be defined on the interval [0, ). Let s > 0 be a
positive parameter. We define the Laplace transform of f (t) as
Z
F (s) =
est f (t) dt = L(f (t)),
provided that this integral converges. It is customary to use the corresponding capital letters to denote the Laplace transform (so that the Laplace
transform of g(t) is denoted by G(s), of h(t) by H(s), etc.). We also use the
operator notation for the Laplace transform: L(f (t)).
We now build up a collection of Laplace transforms.
L(1) =
L(t) =
st
est dt =
est 1
| = ;
s 0
s
"
est t est 1
t dt =
2 |0 = 2 .
s
s
s
Using integration by parts (n is a positive integer)

n
L(t ) =
est tn n
t dt =
|0 +
s
s
st n
(limt est tn = limt
est tn1 dt =
n
L(tn1 ) .
s
tn
est
= 0, after n applications of LHospitals rule.)

2
2 1
2
With this recurrence relation, we now compute L(t2 ) = L(t) = 2 = 3 ,
s
s s
s
3
3!
3
2
L(t ) = L(t ) = 4 , and in general
s
s
L(tn ) =
n!
sn+1
The next class of functions are the exponentials eat , where a is some number:
at
L(e ) =
est eat dt =
1 (sa)t
1
e
|0 =
, provided that s > a.
sa
sa
Here we had to assume that s > a, to obtain a convergent integral.

Next we observe that for any constants c1 and c2
(1.1)
L(c1 f (t) + c2 g(t)) = c1 F (s) + c2 G(s),
4.1. THE LAPLACE TRANSFORM AND ITS INVERSE
141
because a similar property holds for integrals (and the Laplace transform is
an integral). This expands considerably the set of functions for which we
can write down the Laplace transform. For example, with a > 0,
1
1
1 1
1 1
s
L(cosh at) = L( eat + eat ) =
+
= 2
, for s > a .
2
2
2sa 2s+a
s a2
Similarly,
a
, for s > a .
s2 a2
We can compute the Laplace transform of any polynomial. For example,
L(sinh at) =
L 2t5 3t2 + 5 = 2L(t5 ) 3L(t2 ) + 5L(1) =

Compute
L(cos at) =
est cos at dt =
240
6
5
3+ .
s6
s
s
est (a sin at s cos at)

s
|0 = 2
.
2
2
a +s
s + a2
(One guesses that the antiderivative of est cos at is of the form Aest cos at+
Best sin at, and then evaluates the constants A and B by differentiation.)
Similarly,
L(sin at) =
est sin at dt =
a
est (s sin at + a cos at)
|0 = 2
.
2
2
a +s
s + a2
If c is some number, then

L(ect f (t)) =
est ect f (t) dt =
e(sc)t f (t) dt = F (s c) .
This is called the shift formula:

(1.2)
L(ect f (t)) = F (s c) .
For example,
L(e5t sin 3t) =
Another example:
3
.
(s 5)2 + 9
L(e2t cosh 3t) =
s+2
.
(s + 2)2 9
In the last example c = 2, so that s c = s + 2. Similarly,

L(et t5 ) =
5!
.
(s 1)6
142
4.1.3
The Inverse Laplace Transform
This is just going from F (s) back to f (t). We denote it as f (t) = L1 (F (s)).
We have
L1 (c1 F (s) + c2 G(s)) = c1 f (t) + c2 g(t) .
This is just the formula (1.1), read backward. Each of the formulas for the
Laplace Transform leads to the corresponding formula for its inverse:
tn
)
=
,
sn+1
n!
s
L1 ( 2
) = cos at ,
s + a2
1
1
L1 ( 2
) = sin at ,
2
s +a
a
1
L1 (
) = eat ,
sa
L1 (
and so on. To compute L1 , one often uses partial fractions, as well as the
inverse of the shift formula (1.2)
L1 (F (s c)) = ect f (t) ,
(1.3)
which we shall also call the shift formula.

3s 5
Example Find L1 ( 2
).
s +4
Breaking this fraction into a sum of two fractions, we have
L1 (
s
1
5
3s 5
) = 3L1 ( 2
) 5L1 ( 2
) = 3 cos 2t sin 2t .
s2 + 4
s +4
s +4
2
Example Find L1 (
2
).
(s 5)4
We recognize that a shift by 5 is performed in the function
2
. We begin
s4
2
t3
) = , and then account for the shift,
4
s
3
according to the shift formula (1.3):
by inverting this function, L1 (
L1 (
3
2
5t t
)
=
e
.
(s 5)4
3
143
4.2. SOLVING THE INITIAL VALUE PROBLEMS

s+7
).
s6
We factor the denominator, and use partial fractions
Example Find L1 (
s2
s2
s+7
s+7
2
1
=
=
,
s6
(s 3)(s + 2)
s3 s+2
which gives
L1 (
s+7
) = 2e3t e2t .
s2 s 6
2s 1
).
+ 2s + 5
The denominator cannot be factored, so we complete the square
Example Find L1 (
s2
s2
2s 1
2(s + 1) 3
2s 1
=
=
,
2
+ 2s + 5
(s + 1) + 4
(s + 1)2 + 4
and then adjust the numerator, so that it involves the same shift (as in the
2s 3
denominator). Without the shift, we have the function 2
, with the
s +4
inverse Laplace transform equal to 2 cos 2t 32 sin 2t. By the shift formula
L1 (
4.2
s2
2s 1
3
) = 2et cos 2t et sin 2t .
+ 2s + 5
2
Solving The Initial Value Problems
Integrating by parts,
0
L(f (t)) =
st 0
st
f (t) dt = e
f (t) |0 +s
est f (t) dt .
Let us assume that f (t) does not grow too fast, as t , so that |f (t)|
beat , for some positive constants a and b. If we now choose s > a, then the
limit as t is zero, while the lower limit gives f (0). We conclude
L(f 0 (t)) = f (0) + sF (s) .
(2.1)
To compute the Laplace transform of f 00 (t), we use the formula (2.1) twice
(2.2) L(f 00 (t)) = L((f 0 (t))0 ) = f 0 (0)+sL(f 0 (t)) = f 0 (0)sf (0)+s2 F (s) .
In general,
(2.3)
L(f (n) (t)) = f (n1) (0) sf (n2) (0) sn1 f (0) + sn F (s) .
144
Example Solve y 00 + 3y 0 + 2y = 0, y(0) = 1, y 0 (0) = 4.

Apply the Laplace transform to both sides of the equation. Using the linearity of the Laplace transform, and that L(0) = 0, we have
L(y 00 ) + 3L(y 0 ) + 2L(y) = 0 .
By the formulas (2.1), (2.2), and our initial conditions (L (y(t)) = Y (s)):
y 0 (0) sy(0) + s2 Y (s) + 3 (y(0) + sY (s)) + 2Y (s) = 0 ;
4 + s + s2 Y (s) + 3 (1 + sY (s)) + 2Y (s) = 0 ;
Solve for Y (s):
s2 + 3s + 2 Y (s) + s 1 = 0 .
1s
.
+ 3s + 2
To get the solution, it remains to find the inverse Laplace transform y(t) =
L1 (Y (s)). We factor the denominator, and use partial fractions
Y (s) =
s2
s2
1s
1s
2
3
=
=
.
+ 3s + 2
(s + 1)(s + 2)
s+1 s+2
Answer: y(t) = 2et 3e2t .

Of course, this problem could also be solved without using the Laplace
transform. The same is true for all other problems that we shall consider.
The Laplace transform gives an alternative solution method, and in addition,
it provides a tool that can be used in more involved situations, for example,
to solve partial differential equations.
Example Solve y 00 4y 0 + 5y = 0, y(0) = 1, y 0 (0) = 2.
Apply the Laplace transform to both sides of the equation. Using our initial
conditions, we get
2 s + s2 Y (s) 4 (1 + sY (s)) + 5Y (s) = 0 ;
s6
.
4s + 5
To invert the Laplace transform, we complete the square in the denominator,
and then produce the same shift in the numerator
Y (s) =
Y (s) =
s2
s2
s6
(s 2) 4
=
.
4s + 5
(s 2)2 + 1
145

Using the shift formula, we get the answer: y(t) = e2t cos t 4e2t sin t.
Example Solve
y 00 + 2 y = 5 cos 2t, 6= 2,
y(0) = 1, y 0 (0) = 0 .
This problem models a spring, with the natural frequency , subjected to

an external force of frequency 2. Applying the Laplace transform to both
sides of the equation, and using the initial conditions, we get
s + s2 Y (s) + 2 Y (s) =
Y (s) =
5s
;
+4
s2
5s
s
+
.
(s2 + 4)(s2 + 2 ) s2 + 2
The second term is easy to invert. To find the inverse Laplace transform of
the first term, we use the guess-and-check method (or partial fractions)
1
s
s
s
= 2
.
(s2 + 4)(s2 + 2 )
4 s2 + 4 s2 + 2

5
(cos 2t cos t) + cos t.
4
When approaches 2, the amplitude of the oscillations becomes large.
Answer: y(t) =
To treat the case of resonance, when = 2, we need one more formula.

Differentiate in s both sides of the formula
F (s) =
to obtain
0
F (s) =
or
est f (t) dt ,
est tf (t) dt = L(tf (t)) ,
L(tf (t)) = F 0 (s) .

For example,
(2.4)
L(t sin 2t) =
2
4s
d
= 2
.
2
ds s + 4
(s + 4)2
Example Solve (a case of resonance)

y 00 + 4y = 5 cos 2t,
y(0) = 0, y 0 (0) = 0 .
146
Using the Laplace transform, we have

s2 Y (s) + 4Y (s) =
Y (s) =
5s
;
+4
s2
5s
.
(s2 + 4)2
5
Then, using (2.4), y(t) = t sin 2t. We see that the amplitude of oscillations
4
(which is equal to 54 t) tends to infinity with time.
Example Solve the initial value problem for the fourth order equation
y 0000 y = 0,
y(0) = 1, y 0 (0) = 0, y 00 (0) = 1, y 000(0) = 0 .

Applying the Laplace transform, and using our initial conditions, we have
y 000 (0) sy 00 (0) s2 y 0 (0) s3 y(0) + s4 Y (s) Y (s) = 0 ;
(s4 1)Y (s) = s3 + s ;
Y (s) =
s3 + s
s(s2 + 1)
s
=
= 2
.
4
2
2
s 1
(s 1)(s + 1)
s 1
We conclude that y(t) = cosh t.
4.2.1
Step Functions
Sometimes an external force acts only over some time interval. One uses
step functions to model such forces. The basic step function is the Heaviside
function uc (t), defined for any positive constant c by
uc (t) =
0 if t < c
1 if t c
uc (t)
b
The Heaviside step function uc (t)
-t
147
(Oliver Heaviside, 1850 - 1925, was a self-taught English electrical engineer.)

Using uc (t), we can build up other step functions. For example, the function
u2 (t) u4 (t) is equal to 1 for 2 t < 4, and is zero otherwise. The function
(u2 (t) u4 (t)) t2 models a force, which is equal to t2 for 2 t < 4, and is
zero for other t.
Compute:
L(uc (t)) =
est uc (t) dt =
L1
ecs
s
est dt =
ecs
;
s
= uc (t) .
Similarly, we shall compute the Laplace transform of the following shifted

function, which begins at t = c (it is zero for 0 < t < c):
L (uc (t)f (t c)) =
est uc (t)f (t c) dt =
est f (t c) dt .
(We define uc (t)f (t c) to be zero on (0, c), even though f (t c) is not

defined on (0, c).) In the last integral we change the variable t z, by
setting t c = z. We have dt = dz, and the integral becomes
Z
es(c+z) f (z) dz = ecs
esz f (z) dz = ecs F (s) .
We have another pair of shift formulas:

L (uc (t)f (t c)) = ecs F (s) ;
L1 ecs F (s) = uc (t)f (t c) .
(2.5)
1
For example, L
2s
1
s4
1
u2 (t)(t 2)3 .
6
Example Solve
y 00 + 9y = u2 (t) u4 (t),
y(0) = 1, y 0 (0) = 0 .
Here the forcing term is equal to 1, for 2 t < 4, and is zero for other t.
Taking the Laplace transform, then solving for Y (s), we have
s2 Y (s) s + 9Y (s) =
e2s e4s
;
s
s
148

Y (s) =
s2
s
1
1
+ e2s 2
e4s 2
.
+9
s(s + 9)
s(s + 9)
Using the guess-and-check method (or partial fractions)

1
1 1
s
=
,
s(s2 + 9)
9 s s2 + 9

and therefore
L1
1
s(s2 + 9)
1 1
cos 3t .
9 9
Using (2.5), we conclude

1 1
1 1
cos 3(t 2) u4 (t)
cos 3(t 4) .
y(t) = cos 3t + u2 (t)
9 9
9 9

Observe that the solution undergoes jumps in its behavior at t = 2, and

at t = 4, which corresponds to the force being switched on at t = 2, and
switched off at t = 4.
4.3
The Delta Function and Impulse Forces
Imagine a rod, which is so thin that we may consider it to be one dimensional,

and so long that we assume it to extend for < t < , along the t axis.
Assume that the function (t) gives the density of the rod (weight per unit
length). If we subdivide the interval (N, N ), using the points t1 , t2 , . . . , tn ,
at a distance t apart, then the weight of the piece i can be approximated
P
by (ti)t, and ni=1 (ti )t gives an approximation of the total weight.
Passing to the limit, letting n, N , we get the exact value of the weight:
m = lim
t0
n
X
i=1
(ti )t =
(t) dt .
Assume now that the rod is moved to a new position in the (t, y) plane,
with each point (t, 0) moved to a point (t, f (t)), where f (t) is a given function. What is the work needed for this move? For the piece i, the work is
approximated by f (ti )(ti)t. The total work is then
W = lim lim
N t0
n
X
i=1
f (ti )(ti)t =
(t)f (t) dt .
Assume now that the rod has unit weight, m = 1, and the entire weight
is pushed into a single point t = 0. The resulting distribution of weight is
4.3. THE DELTA FUNCTION AND IMPULSE FORCES
149
called the delta distribution or the delta function, and is denoted (t). It has
the following properties:
(i) (t) = 0, for t 6= 0,
(ii)
(t) dt = 1 (unit weight),
(iii)
(t)f (t) dt = f (0).
The last formula holds, because work is expended only to move the weight 1
at t = 0, the distance of f (0). Observe that (t) is not a usual function, like
the ones studied in Calculus. (If a usual function is equal to zero, except at
one point, its integral is zero, over any interval.) One can think of (t) as
the limit of the functions
f (t) =
as 0. (Observe that
1
2
f (t) dt
if < t <
for other t
= 1.)
For any number t0 , the function (t t0 ) gives a translation of the delta

function, with the unit weight concentrated at t = t0 . Correspondingly, its
properties are
(i) (t t0 ) = 0, for t 6= t0 ,
(ii)
(iii)
(t t0 ) dt = 1,
(t t0 )f (t) dt = f (t0 ).
Using the properties (i) and (iii), we compute the Laplace transform,
for any t0 0,
L ((t t0 )) =
(t t0 )est dt =
(t t0 )est dt = est0 .
In particular,
L ((t)) = 1 .
Correspondingly,

L1 est0 = (t t0 ), and L1 (1) = (t) .
150
y
1.5
1.0
0.5
Figure 4.1: Springs response to an impulse force
In the following example we consider forced vibrations of a spring, with

the external force concentrated at t = 2. In other words, an external impulse
force is applied at t = 2.
Example Solve the initial value problem
y 00 + 2y 0 + 5y = 6 (t 2), y(0) = 0, y 0 (0) = 0 .
Applying the Laplace transform, then solving for Y (s), we have
(s2 + 2s + 5)Y (s) = 6e2s ;
Y (s) =
s2
e2s
6
= e2s
.
+ 2s + 5
(s + 1)2 + 4
1
Recalling that by shift formula L

clude, using (2.5), that
1
(s + 1)2 + 4
1 t
e sin 2t, we con2
y(t) = 3u2 (t)e(t2) sin 2(t 2) .

Before the time t = 2, the external force is zero. Coupled with zero initial
conditions, this leaves the spring at rest for t 2. The impulse force at t = 2
sets the spring in motion, but the vibrations quickly die down, because of
the heavy damping, see the graph of y(t) in Figure 4.1.
4.4. CONVOLUTION AND THE TAUTOCHRONE CURVE
4.4
151
Convolution and the Tautochrone Curve
The problem
y 00 + y = 0,
y(0) = 0, y 0 (0) = 1
has solution y = sin t. If we now add a forcing term g(t), and consider
y 00 + y = g(t),
then the solution is
(4.1)
y(t) =
y(0) = 0, y 0 (0) = 0 ,
t
sin(t v)g(v) dv,
as we saw in the section on convolution integrals. Motivated by this formula,

we now define the concept of convolution of two functions f (t) and g(t)
f g =
f (t v)g(v) dv .
The formula (4.1) can be written as

y(t) = sin t g(t) .
Another example:
t t2 =
(t v)v 2 dv = t
t
0
v 2 dv
v 3 dv =
t4
t4
t4
=
.
3
4
12
If you compute t2 t, the answer is the same. More generally,

gf = f g.
Indeed, making a change of variables v u, by letting u = t v, we express
gf =
g(t v)f (v) dv =

=
g(u)f (t u) du
f (t u)g(u) du = f g .
It turns out that the Laplace transform of a convolution is equal to the

product of the Laplace transforms:
L (f g) = F (s)G(s) .
152
Indeed,
L (f g) =
est
f (t v)g(v) dv dt =
ZZ
est f (t v)g(v) dv dt ,
where the double integral on the right hand side is taken over the region D
of the tv-plane, which is an infinite wedge 0 < v < t in the first quadrant.
We now evaluate this double integral by using the reverse order of repeated
integrations:
(4.2)
ZZ
st
e
D
f (t v)g(v) dv dt =
g(v)
Z
st
f (t v) dt dv .
For the integral in the brackets, we make a change of variables t u, by

letting u = t v,
Z
st
f (t v) dt =
es(v+u) f (u) du = esv F (s) ,
and then the right hand side of (4.2) is equal to F (s)G(s).

v
6
@
@ @
@ @
@ @
@
D
@ @
@ @
@
@ @
@- t
@ @ @
The infinite wedge D
We conclude a useful formula

L1 (F (s)G(s)) = (f g)(t) .
For example,
1
s2
(s2 + 4)2
= cos 2t cos 2t =
cos 2(t v) cos 2v dv .
Using that
cos 2(t v) = cos 2t cos 2v + sin 2t sin 2v ,
153
we conclude
1
s2
(s2 + 4)2
= cos 2t
cos2 2v dv + sin 2t
sin 2v cos 2v dv
1
1
t cos 2t + sin 2t .
2
4
Example Consider the vibrations of a spring at resonance
=
y 00 + y = 3 cos t,
y(0) = 0, y 0 (0) = 0 .
Taking the Laplace transform, we compute

Y (s) = 3
Then
(s2
s
.
+ 1)2
3
y(t) = 3 sin t cos t = t sin t .
2
The Tautochrone curve
y
d (x, y)
(x1 , v)
-x
The Tautochrone curve
Assume that we have a curve through the origin in the first quadrant of the
xy-plane, and a particle slides down this curve, under the influence of the
force of gravity. The initial velocity at the starting point is assumed to be
154
zero. We wish to find the curve so that the time T it takes to reach the bottom at (0, 0) is the same, for any starting point (x, y). This historical curve,
called the tautochrone (which means loosely the same time in Latin), was
found by Christian Huygens in 1673. He was motivated by the construction
of a clock pendulum whose period is independent of its amplitude.
Let (x1 , v) be any intermediate position of the particle, v < y. Let
s = f (v) be the length of the curve from (0, 0) to (x1 , v). Of course, the
ds
gives the speed of the particle.
length s depends also on the time t, and
dt
The kinetic energy of the particle at (x1 , v) is due to the decrease of its
potential energy (m is the mass of the particle):
1
ds
m
2
dt

By the chain rule,
2
= mg(y v) .
ds
ds dv
dv
=
= f 0 (v) , so that
dt
dv dt
dt
1
dv
f 0 (v)
2
dt

f 0 (v)
2
= g(y v) ;
p
dv
= 2g y v .
dt
(Minus, because the function v(t) is decreasing.) We separate the variables,

and integrate
Z Tp
Z y
p
f 0 (v)
(4.3)
dv =
2g dt = 2g T .
y v
0
0
(Over the time interval (0, T ), the particle descends from v = y to v = 0.)
To find the function f 0 , we need to solve the integral equation (4.3). We
may rewrite it as
p
(4.4)
y 1/2 f 0 (y) = 2g T .

Recall that in one of our homework problems we had L t 2 =

terms of the variable y
L y 2 =
(4.5)
, or in
s
.
s
We now apply the Laplace transform to the equation (4.4), and get
r
p
1
L f 0 (y) = 2g T .
s
s
155
Solving for L (f 0 (y)), gives

L f 0 (y) =
where we denoted a =
(4.6)
Tp
2g

= a
s
,
s
T2
2g. Using (4.5) again
2
f 0 (y) = ay 1/2 .
s
ds
= 1+
We have ds = dx2 + dy 2 , and so f 0 (y) =
dy
expression in (4.6):
s
2
1
dx
1+
(4.7)
= a .
dy
y
p
dx
dy
2
. We use this
dx
, and then separate the
dy
variables. But it seems easier to use the parametric integration technique.
To do that, we solve this equation for y
This is a first order equation. We could solve it for
(4.8)
y=
1+
and set
(4.9)
dx
dy
2 ,
dx
1 + cos
=
,
dy
sin
where is a parameter. Using (4.9) in (4.8), we express

y=a
sin2
sin2
a 1 cos2
a
=
a
=
= (1 cos ) .
2
2
2
+
2
cos
2
1
+
cos
2
sin + (1 + cos )
It follows that dy =
a
sin d. Then from (4.9), we get
2
dx =
1 + cos
a
dy = (1 + cos ) d .
sin
2
We compute x by integration, obtaining

x=
a
( + sin ),
2
y=
a
(1 cos ) ,
2
which is a parametric representation of the tautochrone. The name of this

curve is cycloid, and it appears in many of other applications.
156
4.5
Distributions
A function f (t) converts a number t into a number f (t). (Sometimes we do

not plug in all real t, but restrict f (t) to its domain.) Functionals convert
functions into numbers. We shall allow only nice functions to be plugged
into functionals.
A function (t) is said to be of compact support if it is equal to zero
outside of some bounded interval (a, b). Functions that are infinitely differentiable, and of compact support, are called test functions. We shall reserve
writing (t) exclusively for test functions.
Definition Distribution is a linear functional on test functions. Notation:
(f, ). (A test function (t) goes in, the number (f, ) is the output.)
Example Let f (t) be any continuous function. Define the distribution
(5.10)
(f, ) =
f (t)(t) dt .
Convergence is not a problem here, because the integrand vanishes outside

of some bounded interval. This functional is linear, because
(f, c1 1 + c2 2 ) =
f (t) (c1 1 + c2 2 ) dt
= c1
f (t)1 dt + c2
f (t)2 dt = c1 (f, 1 ) + c2 (f, 2 ) ,
for any two constants c1 and c2 , and any two test functions 1 (t) and 2 (t).
This way usual functions can be viewed as distributions. The formula
(5.10) lets us consider f (t) in the sense of distributions.
Example The Delta distribution. Define
((t), ) = (0) .
(Compare this with (5.10), and the intuitive formula

(t)(t) dt = (0)
that we had before.) We see that in the realm of distributions the delta
function (t) sits next to usual functions, as an equal member of the club.
Assume that f (t) is a differentiable function. Viewing f 0 (t) as a distribution, we have

(f 0 , ) =
f 0 (t)(t) dt =
f (t)0 (t) dt = (f, 0 ) ,
157
4.5. DISTRIBUTIONS
using integration by parts. Motivated by this formula, we now define the

derivative of any distribution f :
(f 0 , ) = (f, 0 ) .
In particular,
( 0 , ) = (, 0 ) = 0 (0) .
(So that 0 is another distribution. We know how it acts on test functions.)
Similarly, ( 00, ) = ( 0 , 0 ) = 00 (0), and in general
( (n) , ) = (1)n (n) (0) .
We see that all distributions are infinitely differentiable! In particular, all
continuous functions are infinitely differentiable, if we view them as distributions.
Example The Heaviside function H(t) has a jump at t = 0. Clearly, H(t)
is not differentiable at t = 0. But, in the sense of distributions, we have
H 0 (t) = (t) .
Indeed,
(H 0 (t), ) = (H(t), 0) =
0 (t) dt = (0) = ((t), ) .
Example The function |t| is not differentiable at t = 0. But, in the sense

of distributions,
|t|0 = 2H(t) 1 .
Indeed,
0
(|t| , ) = (|t|, ) =
(t) (t) dt
t0 (t) dt .
Integrating by parts in both integrals, we continue

0
(|t| , ) =
(t) dt +
(t) dt = (2H(t) 1, ) .
158
4.5.1
Problems
I. Find the Laplace transform of the following functions.

1. 5 + 2t3 e4t .
5 12
1
+
.
s s4
s+4
6
6
Answer. 2
4.
s +9 s
Answer.
2. 2 sin 3t t3 .
3. cosh 2t e4t .
4. e2t cos 3t.
5.
t3 3t
.
t
6. e3t t4 .
Answer.
s2
(s 2)2 + 9
2
3
.
3
s
s
24
Answer.
(s + 3)5
Answer.
II. Find the inverse Laplace transform of the following functions.

1.
2.
3.
4.
5.
6.
1
2
1
3.
Answer. sin 2t t2 .
+4 s
2
2
s
.
s2 9 s + 3
1
.
Answer. 1 et .
2
s +s
1
.
Answer. 1 cos t.
3
s +s
1
1
1
.
Answer. e3t .
s2 3s
3
3
1
1
1
.
Answer. sin t sin 2t.
2
2
(s + 1)(s + 4)
3
6
s2
1
.
+ 2s + 10
1
8. 2
.
s +s2
7.
9.
10.
s2
s2
s
.
+s+1
s1
.
s2 s 2
1 t
e sin 3t.
3
1
1
Answer. et e2t .
3
3
"
#
3
1
3
12 t
Answer. e
cos
t sin
t .
2
2
3
Answer.
Answer.
1 2t 2 t
e + e .
3
3
159
4.5. DISTRIBUTIONS
11.
s
.
2
4s 4s + 5
Answer. e 2 t
1
1
cos t + sin t .
4
8

III. Using the Laplace transform, solve the following initial value problems.
1. y 00 + 3y 0 + 2y = 0, y(0) = 1, y 0 (0) = 2.
Answer. y = e2t .
2. y 00 + 2y 0 + 5y = 0, y(0) = 1, y 0 (0) = 2.
Answer.
1 t
2e
(2 cos 2t sin 2t).
3. y 00 +y = sin 2t, y(0) = 0, y 0 (0) = 1.
Answer. y =
5
1
sin t sin 2t.
3
3
4. y 00 + 2y 0 + 2y = et , y(0) = 0, y 0 (0) = 1.
1
1
Answer. y = et et (cos t 3 sin t).
5
5
5. y 0000 y = 0, y(0) = 0, y 0 (0) = 1, y 00 (0) = 0, y 000 (0) = 0.
Answer. y =
1
1
sin t + sinh t.
2
2
6. y 0000 16y = 0, y(0) = 0, y 0 (0) = 2, y 00 (0) = 0, y 000(0) = 8.

Answer. y = sinh 2t.
IV.
1. (a) Let s > 0. Show that
sx2
dx =
Hint: Denote I =
.
s
esx dx. Then
I =
sx2
dx
sy2
dy =
es(x
2 +y2 )
dA .
This is a double integral over the entire xy-plane. Evaluate it by using polar
coordinates, to obtain I 2 = s .

(b) Show that L t 2 =
.
s
160

12
Hint: L t
1
2
t 2 est dt. Make a change of variables t x, by

letting x = t . Then L t 2 = 2
esx dx =
esx dx =
.
s
sin x
2. Show that
dx = .
x
2
0
Z
sin tx
dx, and calculate the Laplace transform
Hint: Consider f (t) =
x
0
F (s) = 2 1s .
Z
3. Solve the following system of differential equations

dx
= 2x y, x(0) = 4
dt
dy
= x + 2y, y(0) = 2 .
dt
Answer. x(t) = et + 3e3t, y(t) = et 3e3t.
4. Solve the following non-homogeneous system of differential equations
x0 = 2x 3y + t, x(0) = 0
y 0 = 2x + y, y(0) = 1 .
Answer. x(t) = et
1
3
1
9 4t
e + (7 + 4t), y(t) = et + e4t + (3 + 4t).
16
16
8
8
V.
1. A function f (t) is equal to 1 for 1 t 5, and is equal to 0 for all
other t 0. Represent f (t) as a difference of two step functions, and find
its Laplace transform.
Answer. f (t) = u1 (t) u5 (t), F (s) =
es
s
e5s
s .
2. Find the Laplace transform of t2 2u4 (t).
Answer. F (s) =
2
s3
4s
2e s .
3. Sketch the graph of the function u2 (t) 2u3 (t), and find its Laplace
transform.
4. A function g(t) is equal to 1 for 0 t 5, and is equal to 2 for t > 5.
Represent g(t) using step functions, and find its Laplace transform.
Answer. g(t) = 1 3u5 (t), G(s) =
1
s
5s
3e s .
161
4.5. DISTRIBUTIONS
5. Find the inverse Laplace transform of
Answer. 2u1 (t)(t 1) 3u4 (t)(t 4).

1 s
4s
2e
3e
.
s2
6. Find the inverse Laplace transform of e2s

Answer. u2 (t) 3 cos 2(t 2)
1
sin 2(t 2) .
2

7. Find the inverse Laplace transform of es

Answer. u1 (t)
3s 1
.
s2 + 4
1 2t2 1 3t+3
e
e
.
5
5

s2
1
.
+s6
1
8. Find the inverse Laplace transform of e 2 s 2
, and simplify the
s + 2s + 5
answer.
1
Answer. u/2 (t)et+/2 sin 2t.
2
9. Solve
y 00 + y = 2u1 (t) u5 (t), y(0) = 0, y 0 (0) = 2.
10. Solve
y 00 + 3y 0 + 2y = u2 (t), y(0) = 0, y 0 (0) = 1.
Answer. y(t) = e2t et + u2 (t)
1
1
e(t2) + e2(t2) .
2
2

VI.

1. Show that L u0c (t) = L ((t c)).
This problem shows that u0c (t) = (t c).

2. Find the Laplace transform of (t 4) 2u4 (t).
3. Solve
y 00 + y = (t ), y(0) = 0, y 0 (0) = 2.
Answer. y(t) = 2 sin t u (t) sin t.
4. Solve
y 00 + 2y 0 + 10y = (t ), y(0) = 0, y 0 (0) = 0.
1
Answer. y(t) = u (t)et+ sin 3t.
3
5. Solve
4y 00 + y = (t), y(0) = 0, y 0 (0) = 0.
162
Answer. y(t) =
1
1
sin t.
2
2
6. Solve
4y 00 + 4y 0 + 5y = (t 2), y(0) = 0, y 0 (0) = 1.
1
1
1
Answer. y(t) = e 2 t sin t + u2 (t) e 2 (t2) sin t.
4
VII.
1. Show that sin t 1 = 1 cos t. (Observe that sin t 1 6= sin t.)
2. Show that f (t) (t) = f (t), for any f (t).
(So that the delta function plays the role of unity for convolution.)
3. Find the convolution t sin at.
Answer.
at sin at
.
a2
4. Find the convolution cos t cos t.

Hint: cos(t v) = cos t cos v + sin t sin v.
Answer.
1
1
t cos t + sin t.
2
2
5. Using convolutions, find the inverse Laplace transform of the following

functions
1
t2
t2
(a) 3 2
.
Answer.
sin t =
+ cos t 1.
s (s + 1)
2
2
s
.
(s + 1)(s2 + 9)
s
(c)
.
2
(s + 1)2
(b)
(d)
(s2
1
.
+ 9)2
Answer.
et
1
3
+
cos 3t +
sin 3t.
10
10
10
Answer.
1
t sin t.
2
Answer.
1
3
sin 3t t cos 3t.
54
54
6. Solve the following initial value problem at resonance

y 00 + 9y = cos 3t,
y(0) = 0, y 0 (0) = 0.
1
t sin 3t.
6
7. Solve the initial value problem with a given forcing term g(t)
Answer. y(t) =
y 00 + 4y = g(t),
Answer. y(t) =
1
2
y(0) = 0, y 0 (0) = 0.
sin 2(t v)g(v) dv.
163
4.5. DISTRIBUTIONS
VIII.
1. Find the second derivative of |t| in the sense of distributions.
Answer. 2(t).
2. Find f (t), such that
f (n) (t) = (t) .
Answer. f (t) =
3. Let f (t) =
0
tn1
(n1)!
if t < 0
.
if t > 0
t2
if t < 0
.
2
t + 5 if t > 0
Show that f 0 (t) = 2t + 5(t).

Hint: f (t) = t2 + 5H(t).
Chapter 5
Linear Systems of
Differential Equations
5.1
5.1.1
The Case of Distinct Eigenvalues

Review of Vectors and Matrices
Recall that given two vectors
a1
b1
C1 = a2 , C2 = b2 ,
a3
b3
a1 + b1
we can add them as C1 + C2 = a2 + b2 , or multiply by a constant x:

a3 + b3
xa1
xC1 = xa2 . More generally, we can compute the linear combination

xa3
x1 a1 + x2 b1
x1 C1 + x2 C2 = x1 a2 + x2 b2 ,
x1 a3 + x2 b3
for any two constants x1 and x2 .
We shall be dealing only with the square matrices, like the following 33
matrix
a11 a12 a13
(1.1)
A = a21 a22 a23 .
a31 a32 a33
164
165
5.1. THE CASE OF DISTINCT EIGENVALUES

We shall view A as a row of column vectors A =
C1 C2 C3 , where
a11
a12
a13
C1 = a21 , C2 = a22 , C3 = a23 .

a31
a32
a33
x1
The product of a matrix A and of a vector x = x2 is defined as the

x3
vector
Ax = C1 x1 + C2 x2 + C3 x3 .
(This definition is equivalent to the more traditional one, that you might
have seen before.) We get
a11 x1 + a12 x2 + a13 x3
Ax = a21 x1 + a22 x2 + a23 x3 .

a31 x1 + a32 x2 + a33 x3
Two vectors C1 and C2 are called linearly dependent if one of them is a

constant multiple of the other, so that C2 = aC1 , for some number a. (Zero
vector is linearly dependent with any other vector.) Linearly dependent
vectors C1 and C2 go along the same line. If the vectors C1 and C2 do not
go along the same line, they are linearly independent. Three vectors C1 , C2
and C3 are called linearly dependent if one of them is a linear combination
of the others, e.g., if C3 = aC1 + bC2 , for some numbers a, b. This means
that C3 lies in the plane determined by C1 and C2 , so that all three vectors
lie in the same plane. If C1 , C2 and C3 do not lie in the same plane, they
are linearly independent.
A system of 3 equations with 3 unknowns
(1.2)
a11 x1 + a12 x2 + a13 x3 = b1

a21 x1 + a22 x2 + a23 x3 = b2
a31 x1 + a32 x2 + a33 x3 = b3
can be written in the matrix form

Ax = b ,
b1
where A is the 3 3 matrix above, and b = b2 is the given vector of the

b3
right hand sides. Recall that the system (1.2) has a unique solution for any
vector b if and only if the columns of the matrix A are linearly independent.
(In that case, the determinant |A| =
6 0, and the inverse matrix A1 exists.)
166CHAPTER 5. LINEAR SYSTEMS OF DIFFERENTIAL EQUATIONS
5.1.2
Linear First Order Systems with Constant Coefficients
We wish to find the functions x1 (t), x2 (t) and x3 (t), which solve the following
system of equations, with given constant coefficients a11 , . . . , a33,
x01 = a11 x1 + a12 x2 + a13 x3
(1.3)
x02 = a21 x1 + a22 x2 + a23 x3

x03 = a31 x1 + a32 x2 + a33 x3 ,
subject to the given initial conditions
x1 (t0 ) = , x2 (t0 ) = , x3 (t0 ) = .
We may write this system using matrix notation:
x0 = Ax, x(t0 ) = x0 ,
(1.4)
x1 (t)
where x(t) = x2 (t) is the unknown vector function, A is the 3 3

x3 (t)
matrix (1.1) of the coefficients, and x0 = is the vector of initial
conditions.
Indeed,
on
the
left
in
(1.3)
we
have
components of the vector
0
x1 (t)
x0 (t) = x02 (t) , while on the right we see the components of the vector
x03 (t)
Ax.
Let us observe that given two vector functions y(t) and z(t), which are
solutions of the system x0 = Ax, their linear combination c1 y(t) + c2 z(t) is
also a solution of the same system, for any constants c1 and c2 . Our system
of differential equations is linear.
We now search for solution of (1.4) in the form
(1.5)
x(t) = et ,
with a constant , and a vector , with entries independent of t. Plugging

this into (1.4), we have

et = A et ,
A = .
167
So that if is an eigenvalue of A, and the corresponding eigenvector, then

(1.5) gives us a solution of the problem (1.4). Observe that the same is true
for any square n n matrix A. Let 1 , 2 and 3 be the eigenvalues of our
3 3 matrix A. There are several cases to consider.
Case 1 The eigenvalues of A are real and distinct. It is known that the
corresponding eigenvectors 1 , 2 and 3 are then linearly independent. We
know that e1 t1 , e2 t2 and e3 t 3 are solutions of our system, so that their
linear combination
x(t) = c1 e1 t 1 + c2 e2 t 2 + c3 e3 t 3
(1.6)
solves our system. We claim that (1.6) gives the general solution of our
system, meaning that we can determine c1 , c2 , and c3 to satisfy any initial
conditions:
(1.7)
x(t0 ) = c1 e1 t0 1 + c2 e2 t0 2 + c3 e3 t0 3 = x0 .
This is a system of three linear equations, with three unknowns c1 , c2 , and

c3 . The matrix of this system is non-singular, because its columns, e1 t0 1 ,
e2 t0 2 and e3 t0 3 , are linearly independent (observe that these columns
are constant multiples of the linearly independent vectors 1 , 2 and 3 ).
Therefore, we can find a unique solution triple c1 , c2 , and c3 of the system
(1.7). Then x(t) = c1 e1 t 1 + c2 e2 t 2 + c3 e3 t 3 is the desired solution of
our initial value problem (1.4).
Example Solve
0
x =
"
2 1
1 2
"
x, x(0) =
1
2
This is a 2 2 system, so that we have only two terms in (1.6). "We compute
#
1
the eigenvalue 1 = 1, and the corresponding eigenvector 1 =
, and
1
"
1
. The general solution
1
"
1
1
2 = 3, with the corresponding eigenvector 2 =

is then
t
x(t) = c1 e
"
1
1
3t
+ c2 e
or in components
x1 (t) = c1 et + c2 e3t
x2 (t) = c1 et + c2 e3t .

Turning to the initial conditions,
x1 (0) = c1 + c2 = 1
x2 (0) = c1 + c2 = 2 .
We calculate c1 = 3/2, c2 = 1/2. Answer:

x1 (t) = 32 et + 12 e3t
x2 (t) = 23 et + 12 e3t .
Case 2 The eigenvalue 1 is double, 2 = 1 , 3 6= 1 , however, 1 has two

linearly independent eigenvectors 1 and 2 . The general solution is given
again by the formula (1.6), with 2 replaced by 1 . We see from (1.7) that
we get a linear system, which has a unique solution, because its matrix has
linearly independent columns. (Linearly independent eigenvectors is the key
here!)
Example Solve
2 1 1
1
x0 = 1 2 1 x , x(0) = 0 .
1 1 2
4
We calculate that this matrix has a double eigenvalue
1 = 1, which
has
1
0
two linearly independent eigenvectors 1 = 0 and 2 = 1 . The

1
1
other eigenvalue is 3 = 4, with the corresponding eigenvector 3 = 1 .

1
1
0
1
x(t) = c1 et 0 + c2 et 1 + c3 e4t 1 .
1
1
1
Or in components
x1 (t) = c1 et + c3 e4t
x2 (t) = c2 et + c3 e4t
x3 (t) = c1 et + c2 et + c3 e4t .
169
Using the initial conditions, we calculate c1 = 2, c2 = 1, and c3 = 1.

Answer:
x1 (t) = 2et e4t
x2 (t) = et e4t
x3 (t) = 3et e4t .

Recall that if an n n matrix A is symmetric (aij = aji , for all i and j),
then all of its eigenvalues are real. The eigenvalues of a symmetric matrix
may be repeated, but there is always a full set of n linearly independent
eigenvectors. So that one can solve the initial value problem (1.4) for any
system with a symmetric matrix.
Case 3 The eigenvalue 1 has multiplicity two (1 is a double root of the
characteristic equation), 3 6= 1 , but 1 has only one linearly independent
eigenvector . We have one solution e1 t . By analogy with the second order
equations, we try te1 t for the second solution. However, this vector function is a scalar multiple of the first solution, so that it is linearly dependent
with it. We modify our guess:
x(t) = te1 t + e1 t .
(1.8)
We now choose a vector , to obtain a second linearly independent solution.

Plugging this x(t) into our system (1.4), and using that A = 1 , we get
e1 t + 1 te1 t + 1 e1 t = 1 te1 t + e1 t A.
Cancelling a pair of terms, and dividing by e1 t , we simplify this to
(1.9)
(A 1 I) = .
Even though the matrix A 1 I is singular (its determinant is zero), it

can be shown that the linear system (1.9) always has a solution , called
generalized eigenvector. Using this in (1.8), gives us the second linearly
independent solution, corresponding to = 1 . (Observe that c is not a
generalized eigenvector for any constant c 6= 1.)
Example Solve
x0 =
"
1 1
1
3
x.
This matrix has a double"eigenvalue

1 = 2 = 2, and only one linearly
" inde#
#
1
1
.
pendent eigenvector =
. We have one solution: x1 (t) = e2t
1
1

"
The system (1.9) to determine the vector =
1
2
takes the form
1 2 = 1
1 + 2 = 1 .
We discard the second equation, because it is a multiple of the first. The
first equation has infinitely many solutions. But all we need is just one
solution, that is not a multiple of . So we set 2 = 0, which gives 1 = 1.
We have computed the second linearly independent solution
2t
x2 (t) = te
"
1
1
+e
2t
"
2t
"
1
0
2t
"

2t
x(t) = c1 e
5.2
5.2.1
"
1
1
+ c2 te
1
1
+e
1
0
#!
A Pair of Complex Conjugate Eigenvalues

Complex Valued Functions
Recall that one differentiates complex valued functions much the same way,
as the real ones. For example,
d it
e = ieit ,
dt
where i = 1 is treated the same way as any other constant. Any complex
valued function f (t) can be written in the form f (t) = u(t)+iv(t), where u(t)
and v(t) are real valued functions. It follows by the definition of derivative
that f 0 (t) = u0 (t) + iv 0 (t). For example, using Eulers formula,
d it
d
e = (cos t + i sin t) = sin t + i cos t = i(cos t + i sin t) = ieit .
dt
dt
Any complex valued vector function x(t) can also be written as x(t) =
u(t) + iv(t), where u(t) and v(t) are real valued vector functions. Again, we
have x0 (t) = u0 (t) + iv 0 (t). If x(t) is a solution of our system (1.4), then
u0 (t) + iv 0 (t) = A(u(t) + iv(t)) .
Separating the real and imaginary parts, we see that both u(t) and v(t) are
real valued solutions of our system.
171
5.2. A PAIR OF COMPLEX CONJUGATE EIGENVALUES
5.2.2
The General Solution
Assume that the matrix A has a pair of complex conjugate eigenvalues +i

and i. They need to contribute two linearly independent solutions. The
eigenvector corresponding to + i is complex valued, which we may write
as + i, where and are real valued vectors. Then x(t) = e(+i)t ( + i)
is a solution of our system. To get two real valued solutions, we take the
real and the imaginary parts of this solution. We have
x(t) = et (cos t + i sin t)( + i)
= et(cos t sin t ) + iet (sin t + cos t ) .
So that
u(t) = et(cos t sin t ) ,
v(t) = et (sin t + cos t )
give us two real valued solutions. In case of a 2 2 matrix (when there are
no other eigenvalues), the general solution is
(2.1)
x(t) = c1 u(t) + c2 v(t) .
(If one uses the other eigenvalue i, and the corresponding eigenvector,
the answer is the same.) We show in the exercises that one can choose the
constants c1 and c2 to satisfy any initial condition x(t0 ) = x0 .
Example Solve
x0 =
"
1 2
2
1
"
x, x(0) =
2
1
We calculate the eigenvalues

"
# 1 = 1 + 2i and 2 = 1 2i. The eigenvector
i
corresponding to 1 is
. So that we have a complex valued solution
1
e(1+2i)t
"
i
. We rewrite it as
1
et (cos 2t + i sin 2t)
"
i
1
= et
"
sin 2t
cos 2t
+ iet
"
cos 2t
sin 2t
The real and imaginary parts give us two linearly independent solutions, so
that the general solution is
t
x(t) = c1 e
"
sin 2t
cos 2t
+ c2 e
"
cos 2t
sin 2t

In components
x1 (t) = c1 et sin 2t + c2 et cos 2t
x2 (t) = c1 et cos 2t + c2 et sin 2t .
From the initial conditions

x1 (0) = c2 = 2
x2 (0) = c1 = 1 ,
so that c1 = 1, and c2 = 2. Answer:
x1 (t) = et sin 2t + 2et cos 2t
x2 (t) = et cos 2t + 2et sin 2t .
We see from the form of the solutions, that if a matrix A has all eigenvalues which are either negative or have negative real parts, then limt x(t) =
0 (all components of the vector x(t) tend to zero).
5.3
The Exponential of a Matrix
In matrix notation, a linear system, with a square matrix A,

(3.1)
x0 = Ax,
x(0) = x0
looks like a single equation. In case A and x0 are constants, the solution of
(3.1) is
(3.2)
x(t) = eAt x0 .
In order to write the solution of our system in the form (3.2), we shall define
the notion of the exponential of a matrix. First, we define powers of a matrix:
A2 = A A, A3 = A2 A, and so on. Starting with the Maclauren series
ex = 1 + x +
X
x2 x3 x4
xn
+
+
+ =
,
2!
3!
4!
n!
n=0
we define (I is the identity matrix)

eA = I + A +
X
A2 A3
A4
An
+
+
+ =
.
2!
3!
4!
n!
n=0
So that eA is the sum of infinitely many matrices, and each entry of eA is

an infinite series. It can be shown that all of these series are convergent for
173
5.3. THE EXPONENTIAL OF A MATRIX
any matrix A, so that we can compute eA for any square matrix A (we shall
prove this fact for diagonalizable matrices). If O is a square matrix with
zero entries, then eO = I.
We have
eAt = I + At +
A2 t2
A3 t3
A4 t4
+
+
+ ,
2!
3!
4!
and then
A3 t2
A4 t3
d At
e = A + A2 t +
+
+ = AeAt .
dt
2!
3!
We conclude that the formula (3.2) gives the solution of the initial-value
problem (3.1). (Observe that x(0) = eO x0 = x0 .)
Example Let A =
"
an 0
0 bn
A
e =
"
a 0
, where a and b are constants. Then An =
0 b
, and we have
"
1+a+
0
Example Let A =
At
"
a2
2!
"
a3
3!
1+b+
b2
2!
b3
3!
0
+
"
ea 0
0 eb
0 1
. Then for any constant t
1
0
1 t2 /2! + t4 /4! + t + t3 /3! t5 /5! +

t t3 /3! + t5 /5! +
1 t2 /2! + t4 /4! +
"
cos t sin t
sin t
cos t
The last example lets us express the solution of the system

x01 =
(3.3)
x2 , x1 (0) =
x02 = x1 ,
x2 (0) =
in the form
"
x1 (t)
x2 (t)
= eAt
"
"
"
cos t sin t
sin t
cos t
#
#"
which is rotation of the initial vector

by the angle t, counterclockwise.
We see that the integral curves of our system are circles in the (x1 , x2 ) plane.

This makes sense, because the velocity vector
perpendicular to the position vector
"
x1
x2
"
x01
x02
"
x2
x1
is always
Observe that the system (3.3) is equivalent to the equation

x001 + x1 = 0 ,
modeling a harmonic oscillator.
Recall that we can write a square matrix
a11 a12 a13
A = a21 a22 a23

a31 a32 a33
a11
as a row of column vectors A = [C1 C2 C3 ], where C1 = a21 , C2 =

a31
a12
a13
a22 , and C3 = a23 . Then the product of the matrix A and of a

a32
a33
x1
vector x = x2 is defined as the vector

x3
Ax = C1 x1 + C2 x2 + C3 x3 .
If B is another 3 3 matrix, whose columns are the vectors K1 ,K2 and K3 ,

we define the product of the matrices A and B as follows:
AB = A [K1 K2 K3 ] = [AK1 AK2 AK3 ] .
(This definition is equivalent to the traditional one, which is giving the ijelement of the product: (AB)ij =
3
X
k=1
aik bkj .) In general, AB 6= BA.
1 0 0
Let be a diagonal matrix = 0 2 0 . Compute the product

0 0 3
0
0
1

A = A 0 A 2 A 0 = [1 C1 2 C2 3 C3 ] .
0
0
3
175
So that multiplying a matrix A from the right by a diagonal matrix , results

in the columns of A being multiplied by the corresponding entries of .
Assume now that the matrix A has 3 linearly independent eigenvectors
E1 , E2 , and E3 , so that AE1 = 1 E1 , AE2 = 2 E2 , and AE3 = 3 E3 (the
eigenvalues 1 , 2 and 3 are not necessarily different). Form the matrix
S = [E1 E2 E3 ]. Observe that S has an inverse matrix S 1 . We have
A S = [A E1 A E2 A E3 ] = [1 E1 2 E2 3 E3 ] = S .
Multiplying both sides from the left by S 1 , we have
(3.4)
S 1 A S = .
Similarly,
(3.5)
A = S S 1 .
One refers to the formulas (3.4) and (3.5) as giving the diagonalization of the
matrix A. We see that a matrix with a complete set of 3 linearly independent
eigenvectors can be diagonalized. An n n matrix A is diagonalizable, if it
has a complete set of n linearly independent eigenvectors.
If A is diagonalizable, so that the formula (3.5) holds, then A2 = A A =
S S 1 S S 1 = S 2 S 1 , and in general An = S n S 1 . We then have
(for any real scalar t)
eAt =
X
An tn
n=0
n!
X
S n tn S 1
n=0
n!
=S
X
n tn
n=0
n!
S 1 = S et S 1 .
The following important theorem holds for any n n matrix.

Theorem 5.3.1 Assume that all eigenvalues of the matrix A are either
negative or have negative real parts. Then all solutions of the system
x0 = A x
tend to zero as t .
We can prove this theorem in the case when A is diagonalizable. Indeed,
then we have
x(t) = eAt x(0) = S et S 1 x(0) .
The matrix et is a diagonal one, and its entries, e1 t , e2 t , . . . , en t , tend
to zero as t , so that x(t) 0 as t .

In general, eA+B 6= eA eB . This is because AB 6= BA for matrices,
in general. (One way to show that ex+y = ex ey is to expand all three
exponentials in power series, and show that the series on the left is the same
as the one on the right. In the process, we use that xy = yx for numbers.)
aI A
But eaI+A
(a is any number). For example,
" = e e# , because (aI)A"= A(aI) #
2 1
0 1
if A =
, then A = 2I +
, and we have
1
2
1
0
At
2tI
"
=e
e2t
0
0 e2t
"
#"
0 t
t
0
2tI
=e
cos t sin t
sin t
cos t
"
cos t sin t
sin t
cos t
= e2t
"
cos t sin t
sin t
cos t
Non-Homogeneous Systems
We shall solve the initial value problem
x0 = Ax + f (t), x(t0 ) = x0 ,
(3.6)
where f (t) is a given vector function, and A is a constant square matrix.

The solution is
Z
x(t) = eA(tt0 ) x0 + eAt
(3.7)
eAs f (s) ds .
t0
How does one think of this formula? In case of one equation (when A is a
number) we have an easy linear equation (with = eAt ), for which (3.7)
d A(tt0 )
e
= AeA(tt0 ) to justify this formula
gives the solution. We use that dt
for matrices.
Example Solve
x0 =
"
0 1
1
0
x+
"
0
3
"
1
t
, x(0) =
"
0
3
By (3.7)
At
x(t) = e
At
+e
As
"
1
s
ds .
We have
At
"
cos t sin t
sin t
cos t
As
, e
"
cos s sin s
sin s cos s
177

As
"
"
ds =
"
1
s
cos s sin s
sin s cos s
#"
1
s
"
cos s + s sin s
sin s + s cos s
"
Then
Z
eAs
"
1
s
At
Rt
(cos s + s sin s) ds
Rt 0
0 ( sin s + s cos s) ds
As
"
1
s
"
t + 2 sin t
2 2 cos t
ds =
"
t cos t + 2 sin t
2 + 2 cos t + t sin t
t + 2 sin t
2 2 cos t
We conclude that
x(t) =
"
3 sin t
3 cos t
"
t sin t
2 + cos t
or, in components, x1 (t) = t sin t, x2 (t) = 2 + cos t.

An easier approach for this particular system is to convert it to a single
equation
x001 + x1 = t, x1 (0) = 0, x01 (0) = 2 ,
for which x1 (t) = t sin t, and then x2 = x01 + 1 = 2 + cos t.
5.3.1
Solutions of Planar Systems Near the Origin
We now describe the solution curves in the x1 x2 -plane, near the origin (0, 0),
of the system
(3.8)
x0 = Ax ,
with a constant 2 2 matrix A =
"
a11 a12
a21 a22
, and x =
"
x1
x2
If the eigenvalues of A are real and distinct, 1 6= 2 , with the corresponding eigenvectors 1 and 2 , we know that the general solution is
x(t) = c1 e1 t 1 + c2 e2 t 2 .
We distinguish the following cases.
(i) Both eigenvalues are negative, 1 < 2 < 0. Then
x(t) c2 e2 t 2 ,
for large t > 0 .
All solution curves (x1 (t), x2(t)) tend to the origin (0, 0) as t , and they
are tangent to the vector 2 . We have a stable node at the origin.

(ii) Both eigenvalues are positive, 1 > 2 > 0. Then
x(t) c2 e2 t 2 ,
as t .
All solution curves (x1 (t), x2(t)) emerge from the origin (0, 0), and they are
tangent to the vector 2 . We have an unstable node at the origin.
(iii) The eigenvalues have different sign, 1 > 0 > 2 . In case the initial
point lies along the vector 2 (so that c1 = 0), the solution curve (x1 (t), x2 (t))
tends to the origin (0, 0), as t . All other solutions (when c1 6= 0) tend
to infinity, and they are tangent to the vector
t . We have a
" 1 , as #
1 0
saddle at the origin. For example, if A =
, we have 1 = 1,
0 1
1
c1 c2
1
2 = 1, and x1 = c1 et , x2 = c2 et . Express: x2 = c2 t = c2
=
.
e
x1 /c1
x1
c
Denoting c = c1 c2 , we see that solutions are the hyperbolas x2 = , which
x1
form a saddle near the origin.
Turning "to the case
# of complex eigenvalues, we begin with a special
p q
matrix B =
, with the eigenvalues p iq, and consider the system
q p
y 0 = By .
(3.9)
Its solutions are
Bt
y(t) = e
"
c1
c2
pt
=e
"
"
cos qt sin qt
sin qt cos qt
#"
c1
c2
c1
If p = 0, then any initial vector
is rotated around the origin (infinitely
c2
many times, as t ). We say that the origin is a center. If p < 0, the
solutions spiral into the origin. We have a stable spiral at the origin. If
p > 0, we have an unstable spiral.
Assume now that A is a 2 2 matrix with complex eigenvalues p iq,
q 6= 0. Let = u + iv be an eigenvector corresponding to p + iq, where u and
v are real vectors. We have A(u+iv) = (p+iq)(u+iv) = puqv +i(pv +qu),
and separating the real and imaginary parts
(3.10)
Au = pu qv, and Av = qu + pv .
179
We form a 2 2 matrix P = [u v], which has the vectors u and v as its

columns. Then, using (3.10)
AP = [Au Av] = [pu qv qu + pv] = [u v]
"
p q
q p
= PB ,
where B is the special matrix considered above. We now make a change of

variables x = P y, transforming (3.8) to
P y 0 = AP y = P By ,
and then to the system (3.9), that we have analyzed above (the inverse of
P exists). We conclude that the origin is a center if p = 0, a stable spiral if
p < 0, and an unstable spiral if p > 0.
5.3.2
Problems
I. Solve the following systems of differential equations:

0
1. x =
"
3 1
1 3
x, x(0) =
"
1
3
. Answer:
x1 (t) = 2e2t e4t
x2 (t) = 2e2t e4t .

"
"
0 1
2
2. x =
x, x(0) =
. Check your answer by reducing this
1 0
1
system to a second order equation for x1 (t). Answer:
0
x1 (t) = 21 et + 32 et
x2 (t) = 21 et + 32 et .
2 2 3
0
3. x0 = 2 3 2 x, x(0) = 1 . Answer:
4 2 5
2
x1 (t) = 3et 2e2t + 5e3t
x2 (t) = 4e2t + 5e3t
x3 (t) = 3et + 5e3t .
4 0 1
2
4. x0 = 2 1 0 x, x(0) = 5 . Answer:
2 0 1
0
x1 (t) = 2e2t 4e3t
x2 (t) = 5et 4e2t + 4e3t

x3 (t) = 4e2t + 4e3t .
2 1 1
2
5. x0 = 1 2 1 x, x(0) = 3 . Answer:
1 1 2
2
x1 (t) = 3et + e4t
x2 (t) = 2et + e4t
x3 (t) = et + e4t .
1
4 0
2
0
6. x = 4 7 0 x, x(0) = 6 . Answer:
0
0 5
1
x1 (t) = 2e3t (1 + 8t)
x2 (t) = 2e3t (3 + 8t)

x3 (t) = e5t .
"
"
0 2
2
7. x =
x, x(0) =
. Check your answer by reducing this
2
0
1
system to a second order equation for x1 (t). Answer:
0
x1 (t) = 2 cos 2t sin 2t

x2 (t) = cos 2t 2 sin 2t .
8. x =
"
3 2
2
3
x, x(0) =
"
0
1
. Answer:
x1 (t) = e3t sin 2t

x2 (t) = e3t cos 2t .
1
2
2
1
9. x0 = 1
1
0 x, x(0) = 1 . Answer:
2
0 2 1
x1 (t) = cos t + 5 sin t
181

x2 (t) = et + 2 cos t + 3 sin t
x3 (t) = et + cos t 5 sin t .
10. Solve the non-homogeneous system

0
x =
"
1 1
1
1
x+
"
1
t
, x(0) =
Answer: x1 = 2t + 5et cos t 12 et sin t 1, x2 =

II.
0
1. Show that all solutions of the system x =

constants a and b, satisfy
"
"
1t
2
4
1
+ 12 et (cos t + 10 sin t).
a
b
b a
x, with positive
lim x1 (t) = 0, and lim x2 (t) = 0 .
2. Write the equation (here b and c are constants, y = y(t))

y 00 + by 0 + cy = 0
as a system of two first order equations, by letting x1 (t) = y(t), x2 (t) =
y 0 (t). Compute the eigenvalues for the matrix of this system. Show that
all solutions tend to zero, as t , provided that b and c are positive
constants.
3. Consider the system
x01 = ax1 + bx2
x02 = cx1 + dx2 ,
with given constants a, b, c and d. Assume that a + d < 0 and ad bc > 0.
Show that all solutions tend to zero, as t (x1 (t) 0, and x2 (t) 0,
as t ).
Hint: Show that the eigenvalues for the matrix of this system are either
negative, or have negative real parts.
4. Consider the system
dx
= ax + by
dt
dy
= mx + ny ,
dt
where a,b,m and n are real constants.
(3.11)

(i) Put this system into the form
(3.12)
Hint: Express
(mx + ny) dx (ax + by) dy = 0 .

dy
dx .
(ii) Assume that (0, 0) is a center for (3.11). Show that the equation (3.12)
is exact, and solve it.
Hint: One needs n = a (and also that b and m have the opposite signs),
in order for the matrix of (3.11) to have purely imaginary eigenvalues.
Answer. mx2 by 2 = c, a family of ellipses.
5. Let A be a real 3 3 constant matrix. Suppose that all solutions of
x0 = Ax are bounded as t +, and as t . Show that every
solution is periodic, and there is a common period for all solutions.
Hint: One of the eigenvalues of A must be zero, and the other two purely
imaginary.
6. Let x(t) and y(t) be two solutions of the system
x0 = Ax ,
with an n n matrix A. Show that 5x(t), and x(t) + y(t) are also solutions.
Show that the same is true for c1 x(t) + c2 y(t), with any numbers c1 and c2 .
Are the above conclusions true if the entries of A depend on t ?
7. Show that the series for eA converges for any diagonalizable matrix A.
Hint: If A = SS 1 , then eA = Se S 1 .
8. (i) Suppose that +i is an eigenvalue of A, and +i is a corresponding
eigenvector. Show that i is also an eigenvalue of A, and i is a
corresponding eigenvector. (A is an n n matrix with real entries.)
(ii) Show that and are linearly independent. (They are not constant
multiples of each other.)
Hint: Linear dependence of and would imply linear dependence of the

eigenvectors + i and i.
(iii) Show that the formula (2.1) gives the general solution of the system
x0 = Ax, so that we can choose c1 and c2 , with x(t0 ) = x0 , for any initial
condition.
Hint: Decompose x0 as a linear combination of and , and then find c1
and c2 .
183

9. Consider a 2 2 system
x0 = Ax .
Assume that 1 is repeated eigenvalue of A (2 = 1 ), which has two linearly

independent eigenvectors. Show that the solutions are x(t) = e1 t c, which
are straight lines through the origin in the x1 x2 -plane (c is an arbitrary
vector). (If 1 < 0, all these lines tend to the origin, as t , and we
say that the origin is a stable degenerate node. If 1 > 0, the origin is an
unstable degenerate node.)
III.
1. Let A =
"
0 1
. Show that
0 0
At
"
1 t
0 1
0 1 0
2. Let A = 0 0 1 . Show that

0 0 0
eAt
1 t 21 t2
= 0 1 t .
0 0 1
0 1 0
3. Let A = 1
0 0 . Show that
0
0 2
eAt
cos t sin t 0
= sin t
cos t 0 .
0
0 e2t
3 1 0
4. Let A = 1
3 0 . Show that
0
0 2
eAt
e3t cos t e3t sin t 0

3t
= e sin t
e3t cos t 0 .
2t
0
0 e
0 1
0
5. Let A = 0
0
0 . Show that
0
0 2
eAt
6. Let A =
"
1 t 0
= 0
1 0
.
0
0 e2t
0 1
. Show that
1 0
At
"
cosh t sinh t
sinh t cosh t
7. Show that for any square matrix A

AeAt = eAt A .
Is it true that A2 eAt = eAt A2 ?
8. Show that for any square matrix A, and any constants t and s
eAt eAs = eAs eAt = eA(t+s) .
5.4
Floquet Theory and Masseras Theorem
Logs of Negative Numbers, and Logs of Matrices

We wish to give a meaning to the natural logarithm ln(2). We regard 2
as a complex number 2 + 0 i. Writing it in the polar form 2 = 2ei =
2ei(+2n) , where n is any integer, suggests that
ln(2) = ln 2 + i( + 2n),
n = 0, 1, 2, . . . .
If z = rei 6= 0 is any complex number, we define

ln z = ln r + i( + 2n),
n = 0, 1, 2, . . . .
Observe that ln z is a multi-valued function.

Given any non-singular square matrix C (so that the determinant |C| =
det C 6= 0), it is known that one can find a square matrix B, such that
eB = C. It is natural to write: B = ln C. In case C is a diagonal matrix, just
5.4. FLOQUET THEORY AND MASSERAS THEOREM
185
take the logs of the diagonal entries, to compute ln C. If C is diagonalizable,

so that C = SS 1, and the diagonal matrix has non-zero entries, then
ln C = S ln S 1 . Observe that the entries of ln C are complex valued,
and that ln C is not unique. For a general matrix C, one needs the Jordan
normal form to compute ln C, which is outside of the scope of this book.
Linear Dependence and Independence of Vectors
Given n-dimensional vectors x1 , x2, . . . , xk , we play the following game:
choose the numbers c1 , c2, . . . , ck so that
(4.1)
c1 x1 + c2 x2 + + ck xk = 0 .
There is an easy way to win: take all ci = 0. We declare this way to be

illegal (like off-side in soccer, or false start in football). If we can choose the
constants so that at least one of them is not zero, and (4.1) holds, we call
the vectors x1 , x2 , . . . , xk to be linearly dependent. Otherwise, if the only
way to make (4.1) hold is by cheating, or by taking c1 = c2 = = ck = 0,
then the vectors are called linearly independent. Assume that the vectors
x1 , x2 , . . . , xk are linearly dependent, and say c1 6= 0. Then from (4.1)
x1 =
c2
ck
x2 xk ,
c1
c1
so that one of the vectors is a linear combination of the others.

Recall that if the columns of n n matrix A are linearly independent,
then the determinant det(A) 6= 0, and the inverse matrix A1 exists. In
such a case, the system
(4.2)
Ax = b
is solvable for any vector b (x = A1 b). In case det(A) = 0, the system (4.2)
is solvable only for lucky b, which we describe next. Consider the system
(4.3)
AT v = 0 ,
where AT is the transpose matrix. In case det(A) = 0, the system (4.2) is

solvable if and only if (b, v) = 0, where v is any solution (4.3). (Observe
that det(AT ) = det(A) = 0, so that (4.3) has non-trivial solutions.) This is
known as the Fredholm alternative. Here (b, v) = bT v is the scalar (inner)
product. (The book by G. Strang [23] has more details.)
Recall also that (Ax, y) = (x, AT y) for any square matrix A, and vectors
x and y.

The Fundamental Solution Matrix
We consider systems of the form (here x = x(t) is an n-dimensional vector)
(4.4)
x0 = A(t)x
where the nn matrix A(t) depends on t. Let the vectors x1 (t), x2(t), . . . , xn(t)
be linearly independent (at all t) solutions of this system. We use these vectors as columns of the matrix X(t) = [x1 (t) x2 (t) . . . xn (t)]. The matrix
X(t) is called a fundamental solution matrix or a fundamental matrix, for
short. If, moreover, X(0) = I (the identity matrix), we call X(t) the normalized fundamental matrix. We claim that the general solution of (4.4)
is
x(t) = c1 x1 (t) + c2 x2 (t) + + cn xn (t) = X(t)c ,
where c is the column vector c = [c1 c2 . . . cn ]T , and ci s are arbitrary constants. Indeed, if y(t) is any solution of (4.4), we choose the vector c0 , so
that X(0)c0 = y(0) (or c0 = X 1 (0)y(0)). Then the two solutions of (4.4),
X(t)c0 and y(t), have the same initial values at t = 0. By the uniqueness
theorem (which holds for systems too), y(t) = X(t)c0.
Let Y (t) be another fundamental matrix Y (t) = [y1 (t) y2 (t) . . . yn (t)].
Its first column y1 (t) is a solution of (4.4), and so y1 (t) = X(t)d1 , where
d1 is a constant n-dimensional vector. Similarly, y2 (t) = X(t)d2, and so on.
Form a matrix D = [d1 d2 . . . dn ], with constant entries. Then
(4.5)
Y (t) = X(t)D ,
by the rules of matrix multiplication. Observe that the matrix D is nonsingular (D = X 1 (t)Y (t), and D 1 = Y 1 (t)X(t)).
Any fundamental matrix X satisfies the equation (4.4), so that
X 0 = A(t)X .
Indeed, the first column on the left, which is x01 , is equal to the first column
on the right, which is Ax1 , etc. Using the last formula, it is straightforward
to justify that the general solution of the non-homogeneous system
x0 = A(t)x + f (t)
is given by
x(t) = X(t)c + X(t)
X 1 (s)f (s) ds .
If X(t) is the normalized fundamental matrix, then c = x(0), and we have

(4.6)
x(t) = X(t)x(0) + X(t)
X 1 (s)f (s) ds .
187
Periodic Systems
We now consider n n systems with periodic coefficients
(4.7)
x0 = A(t)x, with A(t + p) = A(t) .
We assume that all entries of the n n matrix A(t) are functions of the
period p. Any solution x(t) of (4.7) satisfies this system at all times t, in
particular at t + p, x0 (t + p) = A(t + p)x(t + p), which implies that
x0 (t + p) = A(t)x(t + p) ,
so that x(t + p) is also a solution of (4.7). Let X(t) be a fundamental matrix
of (4.7), then so is X(t + p), and by (4.5)
(4.8)
X(t + p) = X(t)D ,
with some non-singular n n matrix D, with constant entries. Let the

matrix B be such that eBp = D, so that B = 1p ln D. Define the matrix
P (t) = X(t)eBt. Then X(t) = P (t)eBt . We claim that P (t + p) = P (t).
Indeed, using (4.8),
P (t + p) = X(t + p)eB(t+p) = X(t)DE BpeBt = X(t)eBt = P (t) .
We have just derived the following Floquet Theorem.
Theorem 5.4.1 Any fundamental matrix of the system (4.7) is of the form
X(t) = P (t)eBt , where the matrix P (t) is p-periodic, and B is a constant
matrix.
The eigenvalues of the matrix eBp are called the Floquet multipliers.
Assume that i is a Floquet multiplier, and ci is a corresponding eigenvector
of eBp. Consider x(t) = X(t)ci = P (t)eBtci , a solution of our system (4.7).
It satisfies
x(t + p) = P (t + p)eB(t+p)ci = P (t)eBt eBpci = i P (t)eBt ci = i x(t) ,
so that
x(t + p) = i x(t) .
In particular, the system (4.7) has a periodic solution (with x(t + p) = x(t)),
exactly when one of the Floquet multipliers is equal to 1.
The general solution of the periodic system (4.7) is
x(t) = P (t)eBt c .

The matrix P (t) is periodic, and therefore it is bounded. We see that x(t)
0, as t , exactly when the eigenvalues of B are either negative or have
negative real parts. The eigenvalues i s of B are called the characteristic
multipliers. Recall that is are the eigenvalues of eBp, so that i = ei p ,
1
or i = ln i . It follows that if the (complex) modulus of all Floquet
p
multipliers is < 1, then all i are either negative or have negative real parts,
and so all solutions of the system (4.7) tend to zero, as t .
We now consider non-homogeneous periodic systems
(4.9)
x0 = A(t)x + f (t), with A(t + p) = A(t), f (t + p) = f (t) ,
where f (t) is a given p-periodic vector function. The following spectacular

theorem is due to the Uruguayan mathematician J.L. Massera (published in
1950).
Theorem 5.4.2 Assume that the system (4.9) has a bounded solution, for
t 0. Then the system (4.9) has a p-periodic solution.
Observe that this theorem provides a strong conclusion, with the minimum of assumptions.
Proof:
Let z(t) be a bounded solution, whose existence is assumed in
the statement of the theorem, and let X(t) be the normalized fundamental
matrix of x0 = A(t)x, so that by (4.6)
z(t) = X(t)z(0) + X(t)
X 1 (s)f (s) ds .
In particular,
z(p) = X(p)z(0) + b ,
where we have denoted b = X(p) 0p X 1 (s)f (s) ds. By the periodicity of
our system, z(t+p) is also a solution of (4.9), which is equal to z(p) at t = 0.
Therefore, using (4.6) again,
R
z(t + p) = X(t)z(p) + X(t)
X 1 (s)f (s) ds .
Then
z(2p) = X(p)z(p) + b = X(p) (X(p)z(0) + b) + b = X 2 (p)z(0) + X(p)b + b .
By induction, for any integer m > 0,
(4.10)
z(mp) = X m (p)z(0) +
m1
X
k=0
X k (p)b .
189
For any solution of (4.9),

x(t) = X(t)x(0) + X(t)
X 1 (s)f (s) ds .
We obtain a p-periodic solution, with x(p) = x(0), provided that the initial
vector x(0) satisfies
(4.11)
(I X(p)) x(0) = b ,
where, as before, b = X(p)
Rp
0
X 1 (s)f (s) ds.
Assume, contrary to what we want to prove, that the system (4.9) has no
p-periodic solutions. Then the system (4.11) has no solutions. This implies
that det (I X(p)) = 0, and then det (I X(p))T = det (I X(p)) = 0. It
follows that the system
(4.12)
(I X(p))T v = 0
has non-trivial solutions, and, by the Fredholm alternative, we can find a
non-trivial solution v0 of (4.12), which satisfies
(4.13)
(b, v0 ) 6= 0 .
(Otherwise, the system (4.11) would have solutions.) From (4.12), v0 =

X(p)T v0 , then X(p)T v0 = X 2 (p)T v0 , which gives v0 = X 2 (p)T v0 , and inductively we get
(4.14)
v0 = X k (p)T v0 , for all positive integers k.
We now take the scalar product of (4.10) with v0 , and use (4.14):
(z(mp), v0) = (X m(p)z(0), v0) +
m1
X
(X k (p)b, v0)
k=0
= (z(0), X m(p)T v0 ) +
m1
X
k=0
(b, X k(p)T v0 ) = (z(0), v0) + m(b, v0) ,
as m , in view of (4.13). But z(t) is bounded (and so z(mp) is bounded).

We have a contradiction.

Differential Equations Notes

Uploaded by

Copyright:

Available Formats

Differential Equations Notes

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Differential Equations Notes

Uploaded by

Copyright:

Available Formats

Lectures on Differential Equations

1 First Order Equations

1.8 The Existence and Uniqueness Theorem

2 Second Order Equations

The Characteristic Equation Has Only One (Repeated)

3 Using Infinite Series to Solve Differential Equations

4 The Laplace Transform

5 Linear Systems of Differential Equations

6.5 Periodic Population Models . . . . . . . . . . . . . . . . . . . 210

Laplaces Equation: Poissons Integral Formula

and engineering. This book presents many applications as well. Some of

First Order Equations

to xex . If we try a guess: xex , then its derivative

By differentiation, we verify that this is correct. Of course, integration by

CHAPTER 1. FIRST ORDER EQUATIONS

still has an extra term

This is easier than integrating by parts twice.

+ 4 dx. We begin by rewriting this integral as

One usually computes this integral by a substitution u = x Z+ 4, with

Ignoring a constant multiple again, this evaluates to u3/2 . Returning to the

original variable, we have our initial guess x2 + 4

gives us the integrand with an extra factor of 3. To fix that, we multiply

1.2. FIRST ORDER LINEAR EQUATIONS

Then (for |x| < 1)

First Order Linear Equations

CHAPTER 1. FIRST ORDER EQUATIONS

where c is an arbitrary constant. We see that differential equations have

The problem (2.2) is an example of an initial value problem. If the variable

1.2. FIRST ORDER LINEAR EQUATIONS

Recall from Calculus that

because the derivative of the integral

The Integrating Factor

Let us find the general solution of the equation

and its derivative

We now multiply the equation (2.4) by (x), giving

This allows us to compute the general solution. Indeed,

CHAPTER 1. FIRST ORDER EQUATIONS

Solving for y, gives

From the initial condition

1.2. FIRST ORDER LINEAR EQUATIONS

The initial condition gives

and so the solution is

CHAPTER 1. FIRST ORDER EQUATIONS

Integrate both sides, and perform integration by parts

x sin x dx = x cos x + sin x + c,

giving us the general solution

The initial condition implies

Let us now think of y as independent variable, and x as a function of y,

1.2. FIRST ORDER LINEAR EQUATIONS

CHAPTER 1. FIRST ORDER EQUATIONS

Using the chain rule, we rewrite this as

1.3. SEPARABLE EQUATIONS

We have separated the variables, everything involving y is now on the left,

which gives us immediately the solution (3.2).

So that the general solution is

This is an example of a differential equation, written in differentials. (Didy

Recall that by the fundamental theorem of Calculus,

CHAPTER 1. FIRST ORDER EQUATIONS

for any constant a. TheR integral ax Rf (t) dt gives us an antiderivative of f (x),