Introduction To Differential Equations: Jeffrey R. Chasnov

Introduction to Differential
Equations
Lecture notes for MATH 2351/2352
Jeffrey R. Chasnov
The Hong Kong University of

Science and Technology
The Hong Kong University of Science and Technology
Department of Mathematics
Clear Water Bay, Kowloon
Hong Kong
Copyright ○
c 2009–2014 by Jeffrey Robert Chasnov
This work is licensed under the Creative Commons Attribution 3.0 Hong Kong License.
To view a copy of this license, visit http://creativecommons.org/licenses/by/3.0/hk/
or send a letter to Creative Commons, 171 Second Street, Suite 300, San Francisco,
California, 94105, USA.
Preface
What follows are my lecture notes for a first course in differential equations,
taught at the Hong Kong University of Science and Technology. Included in
these notes are links to short tutorial videos posted on YouTube.
Much of the material of Chapters 2-6 and 8 has been adapted from the
widely used textbook “Elementary differential equations and boundary value
problems” by Boyce & DiPrima (John Wiley & Sons, Inc., Seventh Edition,
○2001).
c Many of the examples presented in these notes may be found in this
book. The material of Chapter 7 is adapted from the textbook “Nonlinear
dynamics and chaos” by Steven H. Strogatz (Perseus Publishing, ○1994).
c
All web surfers are welcome to download these notes, watch the YouTube
videos, and to use the notes and videos freely for teaching and learning. An
associated free review book with links to YouTube videos is also available from
the ebook publisher bookboon.com. I welcome any comments, suggestions or
corrections sent by email to [email protected]. Links to my website, these
lecture notes, my YouTube page, and the free ebook from bookboon.com are
given below.
Homepage:
http://www.math.ust.hk/~machas
YouTube:
https://www.youtube.com/user/jchasnov
Lecture notes:
http://www.math.ust.hk/~machas/differential-equations.pdf
Bookboon:
http://bookboon.com/en/differential-equations-with-youtube-examples-ebook
iii
Contents
0 A short mathematical review 1

0.1 The trigonometric functions . . . . . . . . . . . . . . . . . . . . . 1
0.2 The exponential function and the natural logarithm . . . . . . . 1
0.3 Definition of the derivative . . . . . . . . . . . . . . . . . . . . . 2
0.4 Differentiating a combination of functions . . . . . . . . . . . . . 2
0.4.1 The sum or difference rule . . . . . . . . . . . . . . . . . . 2
0.4.2 The product rule . . . . . . . . . . . . . . . . . . . . . . . 2
0.4.3 The quotient rule . . . . . . . . . . . . . . . . . . . . . . . 2
0.4.4 The chain rule . . . . . . . . . . . . . . . . . . . . . . . . 3
0.5 Differentiating elementary functions . . . . . . . . . . . . . . . . 3
0.5.1 The power rule . . . . . . . . . . . . . . . . . . . . . . . . 3
0.5.2 Trigonometric functions . . . . . . . . . . . . . . . . . . . 3
0.5.3 Exponential and natural logarithm functions . . . . . . . 3
0.6 Definition of the integral . . . . . . . . . . . . . . . . . . . . . . . 3
0.7 The fundamental theorem of calculus . . . . . . . . . . . . . . . . 4
0.8 Definite and indefinite integrals . . . . . . . . . . . . . . . . . . . 5
0.9 Indefinite integrals of elementary functions . . . . . . . . . . . . . 5
0.10 Substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
0.11 Integration by parts . . . . . . . . . . . . . . . . . . . . . . . . . 6
0.12 Taylor series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
0.13 Complex numbers . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1 Introduction to odes 11
1.1 The simplest type of differential equation . . . . . . . . . . . . . 11
2 First-order odes 13
2.1 The Euler method . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Separable equations . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Linear equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4.1 Compound interest . . . . . . . . . . . . . . . . . . . . . . 20
2.4.2 Chemical reactions . . . . . . . . . . . . . . . . . . . . . . 21
2.4.3 Terminal velocity . . . . . . . . . . . . . . . . . . . . . . . 23
2.4.4 Escape velocity . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4.5 RC circuit . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.6 The logistic equation . . . . . . . . . . . . . . . . . . . . . 27
v
vi CONTENTS
3 Second-order odes, constant coefficients 29

3.1 The Euler method . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 The principle of superposition . . . . . . . . . . . . . . . . . . . . 30
3.3 The Wronskian . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.4 Homogeneous odes . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.4.1 Real, distinct roots . . . . . . . . . . . . . . . . . . . . . . 32
3.4.2 Complex conjugate, distinct roots . . . . . . . . . . . . . 34
3.4.3 Repeated roots . . . . . . . . . . . . . . . . . . . . . . . . 36
3.5 Inhomogeneous odes . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.6 First-order linear inhomogeneous odes revisited . . . . . . . . . . 41
3.7 Resonance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.8 Damped resonance . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4 The Laplace transform 47

4.1 Definition and properties . . . . . . . . . . . . . . . . . . . . . . . 47
4.2 Solution of initial value problems . . . . . . . . . . . . . . . . . . 51
4.3 Heaviside and Dirac delta functions . . . . . . . . . . . . . . . . . 54
4.3.1 Heaviside function . . . . . . . . . . . . . . . . . . . . . . 54
4.3.2 Dirac delta function . . . . . . . . . . . . . . . . . . . . . 56
4.4 Discontinuous or impulsive terms . . . . . . . . . . . . . . . . . . 57
5 Series solutions 61
5.1 Ordinary points . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.2 Regular singular points: Cauchy-Euler equations . . . . . . . . . 65
5.2.1 Real, distinct roots . . . . . . . . . . . . . . . . . . . . . . 67
5.2.2 Complex conjugate roots . . . . . . . . . . . . . . . . . . 67
5.2.3 Repeated roots . . . . . . . . . . . . . . . . . . . . . . . . 67
6 Systems of equations 69
6.1 Determinants and the eigenvalue problem . . . . . . . . . . . . . 69
6.2 Coupled first-order equations . . . . . . . . . . . . . . . . . . . . 71
6.2.1 Two distinct real eigenvalues . . . . . . . . . . . . . . . . 71
6.2.2 Complex conjugate eigenvalues . . . . . . . . . . . . . . . 75
6.2.3 Repeated eigenvalues with one eigenvector . . . . . . . . . 77
6.3 Normal modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
7 Nonlinear differential equations 83

7.1 Fixed points and stability . . . . . . . . . . . . . . . . . . . . . . 83
7.1.1 One dimension . . . . . . . . . . . . . . . . . . . . . . . . 83
7.1.2 Two dimensions . . . . . . . . . . . . . . . . . . . . . . . 84
7.2 One-dimensional bifurcations . . . . . . . . . . . . . . . . . . . . 87
7.2.1 Saddle-node bifurcation . . . . . . . . . . . . . . . . . . . 87
7.2.2 Transcritical bifurcation . . . . . . . . . . . . . . . . . . . 88
7.2.3 Supercritical pitchfork bifurcation . . . . . . . . . . . . . 88
7.2.4 Subcritical pitchfork bifurcation . . . . . . . . . . . . . . 89
7.2.5 Application: a mathematical model of a fishery . . . . . . 92
7.3 Two-dimensional bifurcations . . . . . . . . . . . . . . . . . . . . 94
7.3.1 Supercritical Hopf bifurcation . . . . . . . . . . . . . . . . 94
7.3.2 Subcritical Hopf bifurcation . . . . . . . . . . . . . . . . . 95
CONTENTS vii
8 Partial differential equations 97

8.1 Derivation of the diffusion equation . . . . . . . . . . . . . . . . . 97
8.2 Derivation of the wave equation . . . . . . . . . . . . . . . . . . . 98
8.3 Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
8.4 Fourier cosine and sine series . . . . . . . . . . . . . . . . . . . . 102
8.5 Solution of the diffusion equation . . . . . . . . . . . . . . . . . . 104
8.5.1 Homogeneous boundary conditions . . . . . . . . . . . . . 104
8.5.2 Inhomogeneous boundary conditions . . . . . . . . . . . . 108
8.5.3 Pipe with closed ends . . . . . . . . . . . . . . . . . . . . 109
8.6 Solution of the wave equation . . . . . . . . . . . . . . . . . . . . 112
8.6.1 Plucked string . . . . . . . . . . . . . . . . . . . . . . . . 112
8.6.2 Hammered string . . . . . . . . . . . . . . . . . . . . . . . 114
8.6.3 General initial conditions . . . . . . . . . . . . . . . . . . 114
8.7 The Laplace equation . . . . . . . . . . . . . . . . . . . . . . . . 115
8.7.1 Dirichlet problem for a rectangle . . . . . . . . . . . . . . 115
8.7.2 Dirichlet problem for a circle . . . . . . . . . . . . . . . . 116
viii CONTENTS
Chapter 0
A short mathematical
review
A basic understanding of calculus is required to undertake a study of differential

equations. This zero chapter presents a short review.
0.1 The trigonometric functions

The Pythagorean trigonometric identity is
sin2 𝑥 + cos2 𝑥 = 1,
and the addition theorems are
sin(𝑥 + 𝑦) = sin(𝑥) cos(𝑦) + cos(𝑥) sin(𝑦),
cos(𝑥 + 𝑦) = cos(𝑥) cos(𝑦) − sin(𝑥) sin(𝑦).
Also, the values of sin 𝑥 in the first quadrant can be remembered by the rule of
quarters, with 0∘ = 0, 30∘ = 𝜋/6, 45∘ = 𝜋/4, 60∘ = 𝜋/3, 90∘ = 𝜋/2:
√︂ √︂ √︂
∘ 0 ∘ 1 ∘ 2
sin 0 = , sin 30 = , sin 45 = ,
4 4 4
√︂ √︂
3 4
sin 60∘ = , sin 90∘ = .
4 4
The following symmetry properties are also useful:
sin(𝜋/2 − 𝑥) = cos 𝑥, cos(𝜋/2 − 𝑥) = sin 𝑥;
and
sin(−𝑥) = − sin(𝑥), cos(−𝑥) = cos(𝑥).
0.2 The exponential function and the natural

logarithm
The transcendental number 𝑒, approximately 2.71828, is defined as
(︂ )︂𝑛
1
𝑒 = lim 1 + .
𝑛→∞ 𝑛
1
2 CHAPTER 0. A SHORT MATHEMATICAL REVIEW
The exponential function exp (𝑥) = 𝑒𝑥 and natural logarithm ln 𝑥 are inverse
functions satisfying
𝑒ln 𝑥 = 𝑥, ln 𝑒𝑥 = 𝑥.
The usual rules of exponents apply:
𝑒𝑥 𝑒𝑦 = 𝑒𝑥+𝑦 , 𝑒𝑥 /𝑒𝑦 = 𝑒𝑥−𝑦 , (𝑒𝑥 )𝑝 = 𝑒𝑝𝑥 .
The corresponding rules for the logarithmic function are
ln (𝑥𝑦) = ln 𝑥 + ln 𝑦, ln (𝑥/𝑦) = ln 𝑥 − ln 𝑦, ln 𝑥𝑝 = 𝑝 ln 𝑥.
0.3 Definition of the derivative

The derivative of the function 𝑦 = 𝑓 (𝑥), denoted as 𝑓 ′ (𝑥) or 𝑑𝑦/𝑑𝑥, is defined
as the slope of the tangent line to the curve 𝑦 = 𝑓 (𝑥) at the point (𝑥, 𝑦). This
slope is obtained by a limit, and is defined as
𝑓 (𝑥 + ℎ) − 𝑓 (𝑥)
𝑓 ′ (𝑥) = lim . (1)
ℎ→0 ℎ
0.4 Differentiating a combination of functions

0.4.1 The sum or difference rule
The derivative of the sum of 𝑓 (𝑥) and 𝑔(𝑥) is
(𝑓 + 𝑔)′ = 𝑓 ′ + 𝑔 ′ .
Similarly, the derivative of the difference is
(𝑓 − 𝑔)′ = 𝑓 ′ − 𝑔 ′ .
0.4.2 The product rule

The derivative of the product of 𝑓 (𝑥) and 𝑔(𝑥) is
(𝑓 𝑔)′ = 𝑓 ′ 𝑔 + 𝑓 𝑔 ′ ,
and should be memorized as “the derivative of the first times the second plus
the first times the derivative of the second.”
0.4.3 The quotient rule

The derivative of the quotient of 𝑓 (𝑥) and 𝑔(𝑥) is
(︂ )︂′
𝑓 𝑓 ′ 𝑔 − 𝑓 𝑔′
= ,
𝑔 𝑔2
and should be memorized as “the derivative of the top times the bottom minus
the top times the derivative of the bottom over the bottom squared.”
0.5. DIFFERENTIATING ELEMENTARY FUNCTIONS 3
0.4.4 The chain rule

The derivative of the composition of 𝑓 (𝑥) and 𝑔(𝑥) is
(︁ (︀ )︀)︁′
= 𝑓 ′ 𝑔(𝑥) · 𝑔 ′ (𝑥),
(︀ )︀
𝑓 𝑔(𝑥)
and should be memorized as “the derivative of the outside times the derivative
of the inside.”
0.5 Differentiating elementary functions

0.5.1 The power rule
The derivative of a power of 𝑥 is given by
𝑑 𝑝
𝑥 = 𝑝𝑥𝑝−1 .
𝑑𝑥
0.5.2 Trigonometric functions

The derivatives of sin 𝑥 and cos 𝑥 are
(sin 𝑥)′ = cos 𝑥, (cos 𝑥)′ = − sin 𝑥.
We thus say that “the derivative of sine is cosine,” and “the derivative of cosine
is minus sine.” Notice that the second derivatives satisfy
(sin 𝑥)′′ = − sin 𝑥, (cos 𝑥)′′ = − cos 𝑥.
0.5.3 Exponential and natural logarithm functions

The derivative of 𝑒𝑥 and ln 𝑥 are
1
(𝑒𝑥 )′ = 𝑒𝑥 , (ln 𝑥)′ = .
𝑥
0.6 Definition of the integral

The definite integral of a function 𝑓 (𝑥) > 0 from 𝑥 = 𝑎 to 𝑏 (𝑏 > 𝑎) is defined
as the area bounded by the vertical lines 𝑥 = 𝑎, 𝑥 = 𝑏, the x-axis and the curve
𝑦 = 𝑓 (𝑥). This “area under the curve” is obtained by a limit. First, the area is
approximated by a sum of rectangle areas. Second, the integral is defined to be
the limit of the rectangle areas as the width of each individual rectangle goes to
zero and the number of rectangles goes to infinity. This resulting infinite sum
is called a Riemann Sum, and we define
∫︁ 𝑏 𝑁
∑︁ (︀ )︀
𝑓 (𝑥)𝑑𝑥 = lim 𝑓 𝑎 + (𝑛 − 1)ℎ · ℎ, (2)
𝑎 ℎ→0
𝑛=1
where 𝑁 = (𝑏 − 𝑎)/ℎ is the number of terms in the sum. The symbols on the
left-hand-side of (2) are read as “the integral from 𝑎 to 𝑏 of f of x dee x.” The
Riemann Sum definition is extended to all values of 𝑎 and 𝑏 and for all values
of 𝑓 (𝑥) (positive and negative). Accordingly,
∫︁ 𝑎 ∫︁ 𝑏 ∫︁ 𝑏 ∫︁ 𝑏
𝑓 (𝑥)𝑑𝑥 = − 𝑓 (𝑥)𝑑𝑥 and (−𝑓 (𝑥))𝑑𝑥 = − 𝑓 (𝑥)𝑑𝑥.
𝑏 𝑎 𝑎 𝑎
Also, if 𝑎 < 𝑏 < 𝑐, then

∫︁ 𝑐 ∫︁ 𝑏 ∫︁ 𝑐
𝑓 (𝑥)𝑑𝑥 = 𝑓 (𝑥)𝑑𝑥 + 𝑓 (𝑥)𝑑𝑥,
𝑎 𝑎 𝑏
which states (when 𝑓 (𝑥) > 0) that the total area equals the sum of its parts.
0.7 The fundamental theorem of calculus

view tutorial
Using the definition of the derivative, we differentiate the following integral:
𝑥
∫︀ 𝑥+ℎ ∫︀ 𝑥
𝑓 (𝑠)𝑑𝑠 − 𝑓 (𝑠)𝑑𝑠
∫︁
𝑑 𝑎 𝑎
𝑓 (𝑠)𝑑𝑠 = lim
𝑑𝑥 𝑎 ℎ→0 ℎ
∫︀ 𝑥+ℎ
𝑓 (𝑠)𝑑𝑠𝑥
= lim
ℎ→0 ℎ
ℎ𝑓 (𝑥)
= lim
ℎ→0 ℎ
= 𝑓 (𝑥).
This result is called the fundamental theorem of calculus, and provides a con-
nection between differentiation and integration.
The fundamental theorem teaches us how to integrate functions. Let 𝐹 (𝑥)
be a function such that 𝐹 ′ (𝑥) = 𝑓 (𝑥). We say that 𝐹 (𝑥) is an antiderivative of
𝑓 (𝑥). Then from the fundamental theorem and the fact that the derivative of a
constant equals zero,
∫︁ 𝑥
𝐹 (𝑥) = 𝑓 (𝑠)𝑑𝑠 + 𝑐.
𝑎
∫︀ 𝑏
Now, 𝐹 (𝑎) = 𝑐 and 𝐹 (𝑏) = 𝑎 𝑓 (𝑠)𝑑𝑠 + 𝐹 (𝑎). Therefore, the fundamental
theorem shows us how to integrate a function 𝑓 (𝑥) provided we can find its
antiderivative:
∫︁ 𝑏
𝑓 (𝑠)𝑑𝑠 = 𝐹 (𝑏) − 𝐹 (𝑎). (3)
𝑎
Unfortunately, finding antiderivatives is much harder than finding derivatives,

and indeed, most complicated functions cannot be integrated analytically.
We can also derive the very important result (3) directly from the definition
of the derivative (1) and the definite integral (2). We will see it is convenient
0.8. DEFINITE AND INDEFINITE INTEGRALS 5
to choose the same ℎ in both limits. With 𝐹 ′ (𝑥) = 𝑓 (𝑥), we have

∫︁ 𝑏 ∫︁ 𝑏
𝑓 (𝑠)𝑑𝑠 = 𝐹 ′ (𝑠)𝑑𝑠
𝑎 𝑎
𝑁
∑︁
𝐹 ′ 𝑎 + (𝑛 − 1)ℎ · ℎ
(︀ )︀
= lim
ℎ→0
𝑛=1
𝑁 (︀ )︀
∑︁ 𝐹 (𝑎 + 𝑛ℎ) − 𝐹 𝑎 + (𝑛 − 1)ℎ
= lim ·ℎ
ℎ→0
𝑛=1
ℎ
𝑁
∑︁ (︀ )︀
= lim 𝐹 (𝑎 + 𝑛ℎ) − 𝐹 𝑎 + (𝑛 − 1)ℎ .
ℎ→0
𝑛=1
The last expression has an interesting structure. All the values of 𝐹 (𝑥) eval-
uated at the points lying between the endpoints 𝑎 and 𝑏 cancel each other in
consecutive terms. Only the value −𝐹 (𝑎) survives when 𝑛 = 1, and the value
+𝐹 (𝑏) when 𝑛 = 𝑁 , yielding again (3).
0.8 Definite and indefinite integrals

The Riemann sum definition of an integral is called a definite integral. It is
convenient to also define an indefinite integral by
∫︁
𝑓 (𝑥)𝑑𝑥 = 𝐹 (𝑥),
where F(x) is the antiderivative of 𝑓 (𝑥).
0.9 Indefinite integrals of elementary functions

From our known derivatives of elementary functions, we can determine some
simple indefinite integrals. The power rule gives us
𝑥𝑛+1
∫︁
𝑥𝑛 𝑑𝑥 = + 𝑐, 𝑛 ̸= −1.
𝑛+1
When 𝑛 = −1, and 𝑥 is positive, we have
∫︁
1
𝑑𝑥 = ln 𝑥 + 𝑐.
𝑥
If 𝑥 is negative, using the chain rule we have
𝑑 1
ln (−𝑥) = .
𝑑𝑥 𝑥
Therefore, since {︂
−𝑥 if 𝑥 < 0;
|𝑥| =
𝑥 if 𝑥 > 0,
we can generalize our indefinite integral to strictly positive or strictly negative
𝑥: ∫︁
1
𝑑𝑥 = ln |𝑥| + 𝑐.
𝑥
Trigonometric functions can also be integrated:

∫︁ ∫︁
cos 𝑥𝑑𝑥 = sin 𝑥 + 𝑐, sin 𝑥𝑑𝑥 = − cos 𝑥 + 𝑐.
Easily proved identities are an addition rule:

∫︁ ∫︁ ∫︁
(︀ )︀
𝑓 (𝑥) + 𝑔(𝑥) 𝑑𝑥 = 𝑓 (𝑥)𝑑𝑥 + 𝑔(𝑥)𝑑𝑥;
and multiplication by a constant:

∫︁ ∫︁
𝐴𝑓 (𝑥)𝑑𝑥 = 𝐴 𝑓 (𝑥)𝑑𝑥.
This permits integration of functions such as
𝑥3 7𝑥2
∫︁
(𝑥2 + 7𝑥 + 2)𝑑𝑥 = + + 2𝑥 + 𝑐,
3 2
and ∫︁
(5 cos 𝑥 + sin 𝑥)𝑑𝑥 = 5 sin 𝑥 − cos 𝑥 + 𝑐.
0.10 Substitution
More complicated functions can be integrated using the chain rule. Since
𝑑 (︀
𝑓 𝑔(𝑥) = 𝑓 ′ 𝑔(𝑥) · 𝑔 ′ (𝑥),
)︀ (︀ )︀
𝑑𝑥
we have ∫︁
𝑓 ′ 𝑔(𝑥) · 𝑔 ′ (𝑥)𝑑𝑥 = 𝑓 𝑔(𝑥) + 𝑐.
(︀ )︀ (︀ )︀
This integration formula is usually implemented by letting 𝑦 = 𝑔(𝑥). Then one

writes 𝑑𝑦 = 𝑔 ′ (𝑥)𝑑𝑥 to obtain
∫︁ ∫︁
𝑓 ′ 𝑔(𝑥) 𝑔 ′ (𝑥)𝑑𝑥 = 𝑓 ′ (𝑦)𝑑𝑦
(︀ )︀
= 𝑓 (𝑦) + 𝑐
(︀ )︀
= 𝑓 𝑔(𝑥) + 𝑐.
0.11 Integration by parts

Another integration technique makes use of the product rule for differentiation.
Since
(𝑓 𝑔)′ = 𝑓 ′ 𝑔 + 𝑓 𝑔 ′ ,
we have
𝑓 ′ 𝑔 = (𝑓 𝑔)′ − 𝑓 𝑔 ′ .
Therefore, ∫︁ ∫︁
′
𝑓 (𝑥)𝑔(𝑥)𝑑𝑥 = 𝑓 (𝑥)𝑔(𝑥) − 𝑓 (𝑥)𝑔 ′ (𝑥)𝑑𝑥.
0.12. TAYLOR SERIES 7
Commonly, the above integral is done by writing

𝑢 = 𝑔(𝑥) 𝑑𝑣 = 𝑓 ′ (𝑥)𝑑𝑥
𝑑𝑢 = 𝑔 ′ (𝑥)𝑑𝑥 𝑣 = 𝑓 (𝑥).
Then, the formula to be memorized is
∫︁ ∫︁
𝑢𝑑𝑣 = 𝑢𝑣 − 𝑣𝑑𝑢.
0.12 Taylor series

A Taylor series of a function 𝑓 (𝑥) about a point 𝑥 = 𝑎 is a power series rep-
resentation of 𝑓 (𝑥) developed so that all the derivatives of 𝑓 (𝑥) at 𝑎 match all
the derivatives of the power series. Without worrying about convergence here,
we have
𝑓 ′′ (𝑎) 𝑓 ′′′ (𝑎)
𝑓 (𝑥) = 𝑓 (𝑎) + 𝑓 ′ (𝑎)(𝑥 − 𝑎) + (𝑥 − 𝑎)2 + (𝑥 − 𝑎)3 + . . . .
2! 3!
Notice that the first term in the power series matches 𝑓 (𝑎), all other terms
vanishing, the second term matches 𝑓 ′ (𝑎), all other terms vanishing, etc. Com-
monly, the Taylor series is developed with 𝑎 = 0. We will also make use of the
Taylor series in a slightly different form, with 𝑥 = 𝑥* + 𝜖 and 𝑎 = 𝑥* :
𝑓 ′′ (𝑥* ) 2 𝑓 ′′′ (𝑥* ) 3
𝑓 (𝑥* + 𝜖) = 𝑓 (𝑥* ) + 𝑓 ′ (𝑥* )𝜖 + 𝜖 + 𝜖 + ....
2! 3!
Another way to view this series is that of 𝑔(𝜖) = 𝑓 (𝑥* + 𝜖), expanded about
𝜖 = 0.
Taylor series that are commonly used include
𝑥2 𝑥3
𝑒𝑥 = 1 + 𝑥 + + + ...,
2! 3!
𝑥3 𝑥 5
sin 𝑥 = 𝑥 − + − ...,
3! 5!
𝑥2 𝑥4
cos 𝑥 = 1 − + − ...,
2! 4!
1
= 1 − 𝑥 + 𝑥2 − . . . , for |𝑥| < 1,
1+𝑥
𝑥2 𝑥3
ln (1 + 𝑥) = 𝑥 − + − . . . , for |𝑥| < 1.
2 3
A Taylor series of a function of several variables can also be developed. Here,
all partial derivatives of 𝑓 (𝑥, 𝑦) at (𝑎, 𝑏) match all the partial derivatives of the
power series. With the notation
𝜕𝑓 𝜕𝑓 𝜕2𝑓 𝜕2𝑓 𝜕2𝑓
𝑓𝑥 = , 𝑓𝑦 = , 𝑓𝑥𝑥 = , 𝑓𝑥𝑦 = , 𝑓𝑦𝑦 = , etc.,
𝜕𝑥 𝜕𝑦 𝜕𝑥2 𝜕𝑥𝜕𝑦 𝜕𝑦 2
we have
𝑓 (𝑥, 𝑦) = 𝑓 (𝑎, 𝑏) + 𝑓𝑥 (𝑎, 𝑏)(𝑥 − 𝑎) + 𝑓𝑦 (𝑎, 𝑏)(𝑦 − 𝑏)

1 (︀
𝑓𝑥𝑥 (𝑎, 𝑏)(𝑥 − 𝑎)2 + 2𝑓𝑥𝑦 (𝑎, 𝑏)(𝑥 − 𝑎)(𝑦 − 𝑏) + 𝑓𝑦𝑦 (𝑎, 𝑏)(𝑦 − 𝑏)2 + . . .
)︀
+
2!
0.13 Complex numbers

view tutorial: Complex Numbers
view tutorial: Complex Exponential Function
We define the imaginary number 𝑖 to be one of the two numbers that √ satisfies
the rule (𝑖)2 = −1, the other number being −𝑖. Formally, we write 𝑖 = −1. A
complex number 𝑧 is written as
𝑧 = 𝑥 + 𝑖𝑦,
where 𝑥 and 𝑦 are real numbers. We call 𝑥 the real part of 𝑧 and 𝑦 the imaginary
part and write
𝑥 = Re 𝑧, 𝑦 = Im 𝑧.
Two complex numbers are equal if and only if their real and imaginary parts
are equal.
The complex conjugate of 𝑧 = 𝑥 + 𝑖𝑦, denoted as 𝑧¯, is defined as
𝑧¯ = 𝑥 − 𝑖𝑦.
Using 𝑧 and 𝑧¯, we have

1 1
Re 𝑧 = (𝑧 + 𝑧¯) , Im 𝑧 = (𝑧 − 𝑧¯) .
2 2𝑖
Furthermore,
𝑧 𝑧¯ = (𝑥 + 𝑖𝑦)(𝑥 − 𝑖𝑦)
= 𝑥2 − 𝑖2 𝑦 2
= 𝑥2 + 𝑦 2 ;
and we define the absolute value of 𝑧, also called the modulus of 𝑧, by
|𝑧| = (𝑧 𝑧¯)1/2
√︀
= 𝑥2 + 𝑦 2 .
We can add, subtract, multiply and divide complex numbers to get new
complex numbers. With 𝑧 = 𝑥 + 𝑖𝑦 and 𝑤 = 𝑠 + 𝑖𝑡, and 𝑥, 𝑦, 𝑠, 𝑡 real numbers,
we have
𝑧 + 𝑤 = (𝑥 + 𝑠) + 𝑖(𝑦 + 𝑡); 𝑧 − 𝑤 = (𝑥 − 𝑠) + 𝑖(𝑦 − 𝑡);
𝑧𝑤 = (𝑥 + 𝑖𝑦)(𝑠 + 𝑖𝑡)
= (𝑥𝑠 − 𝑦𝑡) + 𝑖(𝑥𝑡 + 𝑦𝑠);
𝑧 𝑧𝑤¯
=
𝑤 𝑤𝑤¯
(𝑥 + 𝑖𝑦)(𝑠 − 𝑖𝑡)
=
𝑠2 + 𝑡2
(𝑥𝑠 + 𝑦𝑡) (𝑦𝑠 − 𝑥𝑡)
= 2 +𝑖 2 .
𝑠 + 𝑡2 𝑠 + 𝑡2
0.13. COMPLEX NUMBERS 9
Furthermore,
√︀
|𝑧𝑤| = (𝑥𝑠 − 𝑦𝑡)2 + (𝑥𝑡 + 𝑦𝑠)2
√︀
= (𝑥2 + 𝑦 2 )(𝑠2 + 𝑡2 )
= |𝑧||𝑤|;
and
𝑧𝑤 = (𝑥𝑠 − 𝑦𝑡) − 𝑖(𝑥𝑡 + 𝑦𝑠)

= (𝑥 − 𝑖𝑦)(𝑠 − 𝑖𝑡)
= 𝑧¯𝑤.
¯
Similarly
⃒𝑧⃒ |𝑧| 𝑧 𝑧¯
⃒ ⃒= , ( )= .
⃒ ⃒
𝑤 |𝑤| 𝑤 𝑤¯
Also, 𝑧 + 𝑤 = 𝑧 + 𝑤. However, |𝑧 + 𝑤| ≤ |𝑧| + |𝑤|, a theorem known as the
triangle inequality.
It is especially interesting and useful to consider the exponential function of
an imaginary argument. Using the Taylor series expansion of an exponential
function, we have
(𝑖𝜃)2 (𝑖𝜃)3 (𝑖𝜃)4 (𝑖𝜃)5
𝑒𝑖𝜃 = 1 + (𝑖𝜃) + + + + ...
2! )︂3! (︂ 4! 3 5!5
2 4
(︂ )︂
𝜃 𝜃 𝜃 𝜃
= 1− + − ... + 𝑖 𝜃 − + + ...
2! 4! 3! 5!
= cos 𝜃 + 𝑖 sin 𝜃.
Therefore, we have
cos 𝜃 = Re 𝑒𝑖𝜃 , sin 𝜃 = Im 𝑒𝑖𝜃 .
Since cos 𝜋 = −1 and sin 𝜋 = 0, we derive the celebrated Euler’s identity
𝑒𝑖𝜋 + 1 = 0,
that links five fundamental numbers, 0, 1, 𝑖, 𝑒 and 𝜋, using three basic mathe-
matical operations, addition, multiplication and exponentiation, only once.
Using the even property cos (−𝜃) = cos 𝜃 and the odd property sin (−𝜃) =
− sin 𝜃, we also have
𝑒−𝑖𝜃 = cos 𝜃 − 𝑖 sin 𝜃;
and the identities for 𝑒𝑖𝜃 and 𝑒−𝑖𝜃 results in the frequently used expressions,
𝑒𝑖𝜃 + 𝑒−𝑖𝜃 𝑒𝑖𝜃 − 𝑒−𝑖𝜃
cos 𝜃 = , sin 𝜃 = .
2 2𝑖
The complex number 𝑧 can be represented in the complex plane with Re 𝑧
as the 𝑥-axis and Im 𝑧 as the 𝑦-axis. This leads to the polar representation of
𝑧 = 𝑥 + 𝑖𝑦:
𝑧 = 𝑟𝑒𝑖𝜃 ,
where 𝑟 = |𝑧| and tan 𝜃 = 𝑦/𝑥. We define arg 𝑧 = 𝜃. Note that 𝜃 is not unique,
though it is conventional to choose the value such that −𝜋 < 𝜃 ≤ 𝜋, and 𝜃 = 0
when 𝑟 = 0.
Useful trigonometric relations can be derived using 𝑒𝑖𝜃 and properties of the
exponential function. The addition law can be derived from
𝑒𝑖(𝑥+𝑦) = 𝑒𝑖𝑥 𝑒𝑖𝑦 .
We have
cos(𝑥 + 𝑦) + 𝑖 sin(𝑥 + 𝑦) = (cos 𝑥 + 𝑖 sin 𝑥)(cos 𝑦 + 𝑖 sin 𝑦)

= (cos 𝑥 cos 𝑦 − sin 𝑥 sin 𝑦) + 𝑖(sin 𝑥 cos 𝑦 + cos 𝑥 sin 𝑦);
yielding
cos(𝑥 + 𝑦) = cos 𝑥 cos 𝑦 − sin 𝑥 sin 𝑦, sin(𝑥 + 𝑦) = sin 𝑥 cos 𝑦 + cos 𝑥 sin 𝑦.
De Moivre’s Theorem derives from 𝑒𝑖𝑛𝜃 = (𝑒𝑖𝜃 )𝑛 , yielding the identity
cos(𝑛𝜃) + 𝑖 sin(𝑛𝜃) = (cos 𝜃 + 𝑖 sin 𝜃)𝑛 .
For example, if 𝑛 = 2, we derive
cos 2𝜃 + 𝑖 sin 2𝜃 = (cos 𝜃 + 𝑖 sin 𝜃)2

= (cos2 𝜃 − sin2 𝜃) + 2𝑖 cos 𝜃 sin 𝜃.
Therefore,
cos 2𝜃 = cos2 𝜃 − sin2 𝜃, sin 2𝜃 = 2 cos 𝜃 sin 𝜃.
With a little more manipulation using cos 𝜃 + sin2 𝜃 = 1, we can derive
2
1 + cos 2𝜃 1 − cos 2𝜃
cos2 𝜃 = , sin2 𝜃 = ,
2 2
which are useful formulas for determining
∫︁ ∫︁
1 1
cos2 𝜃 𝑑𝜃 = (2𝜃 + sin 2𝜃) + 𝑐, sin2 𝜃 𝑑𝜃 = (2𝜃 − sin 2𝜃) + 𝑐,
4 4
from which follows
∫︁ 2𝜋 ∫︁ 2𝜋
2
sin 𝜃 𝑑𝜃 = cos2 𝜃 𝑑𝜃 = 𝜋.
0 0
Chapter 1
Introduction to odes
A differential equation is an equation for a function that relates the values of

the function to the values of its derivatives. An ordinary differential equation
(ode) is a differential equation for a function of a single variable, e.g., 𝑥(𝑡), while
a partial differential equation (pde) is a differential equation for a function of
several variables, e.g., 𝑣(𝑥, 𝑦, 𝑧, 𝑡). An ode contains ordinary derivatives and a
pde contains partial derivatives. Typically, pde’s are much harder to solve than
ode’s.
1.1 The simplest type of differential equation

view tutorial
The simplest ordinary differential equations can be integrated directly by finding
antiderivatives. These simplest odes have the form
𝑑𝑛 𝑥
= 𝐺(𝑡),
𝑑𝑡𝑛
where the derivative of 𝑥 = 𝑥(𝑡) can be of any order, and the right-hand-side
may depend only on the independent variable 𝑡. As an example, consider a mass
falling under the influence of constant gravity, such as approximately found on
the Earth’s surface. Newton’s law, 𝐹 = 𝑚𝑎, results in the equation
𝑑2 𝑥
𝑚 = −𝑚𝑔,
𝑑𝑡2
where 𝑥 is the height of the object above the ground, 𝑚 is the mass of the
object, and 𝑔 = 9.8 meter/sec2 is the constant gravitational acceleration. As
Galileo suggested, the mass cancels from the equation, and
𝑑2 𝑥
= −𝑔.
𝑑𝑡2
Here, the right-hand-side of the ode is a constant. The first integration, obtained
by antidifferentiation, yields
𝑑𝑥
= 𝐴 − 𝑔𝑡,
𝑑𝑡
11
12 CHAPTER 1. INTRODUCTION TO ODES
with 𝐴 the first constant of integration; and the second integration yields
1
𝑥 = 𝐵 + 𝐴𝑡 − 𝑔𝑡2 ,
2
with 𝐵 the second constant of integration. The two constants of integration 𝐴
and 𝐵 can then be determined from the initial conditions. If we know that the
initial height of the mass is 𝑥0 , and the initial velocity is 𝑣0 , then the initial
conditions are
𝑑𝑥
𝑥(0) = 𝑥0 , (0) = 𝑣0 .
𝑑𝑡
Substitution of these initial conditions into the equations for 𝑑𝑥/𝑑𝑡 and 𝑥 allows
us to solve for 𝐴 and 𝐵. The unique solution that satisfies both the ode and
the initial conditions is given by
1
𝑥(𝑡) = 𝑥0 + 𝑣0 𝑡 − 𝑔𝑡2 . (1.1)
2
For example, suppose we drop a ball off the top of a 50 meter building. How
long will it take the ball to hit the ground? This question requires solution of
(1.1) for the time 𝑇 it takes for 𝑥(𝑇 ) = 0, given 𝑥0 = 50 meter and 𝑣0 = 0.
Solving for 𝑇 ,
√︂
2𝑥0
𝑇 =
𝑔
√︂
2 · 50
= sec
9.8
≈ 3.2sec.
Chapter 2
First-order differential
equations
Reference: Boyce and DiPrima, Chapter 2
The general first-order differential equation for the function 𝑦 = 𝑦(𝑥) is written
as
𝑑𝑦
= 𝑓 (𝑥, 𝑦), (2.1)
𝑑𝑥
where 𝑓 (𝑥, 𝑦) can be any function of the independent variable 𝑥 and the depen-
dent variable 𝑦. We first show how to determine a numerical solution of this
equation, and then learn techniques for solving analytically some special forms
of (2.1), namely, separable and linear first-order equations.
2.1 The Euler method

view tutorial
Although it is not always possible to find an analytical solution of (2.1) for
𝑦 = 𝑦(𝑥), it is always possible to determine a unique numerical solution given
an initial value 𝑦(𝑥0 ) = 𝑦0 , and provided 𝑓 (𝑥, 𝑦) is a well-behaved function.
The differential equation (2.1) gives us the slope 𝑓 (𝑥0 , 𝑦0 ) of the tangent line
to the solution curve 𝑦 = 𝑦(𝑥) at the point (𝑥0 , 𝑦0 ). With a small step size
Δ𝑥, the initial condition (𝑥0 , 𝑦0 ) can be marched forward in the x-coordinate
to 𝑥 = 𝑥0 + Δ𝑥, and along the tangent line using Euler’s method to obtain the
y-coordinate
𝑦(𝑥0 + Δ𝑥) = 𝑦(𝑥0 ) + Δ𝑥𝑓 (𝑥0 , 𝑦0 ).
This solution (𝑥0 + Δ𝑥, 𝑦0 + Δ𝑦) then becomes the new initial condition and is
marched forward in the x-coordinate another Δ𝑥, and along the newly deter-
mined tangent line. For small enough Δ𝑥, the numerical solution converges to
the exact solution.
13
14 CHAPTER 2. FIRST-ORDER ODES
2.2 Separable equations

view tutorial
A first-order ode is separable if it can be written in the form
𝑑𝑦
𝑔(𝑦) = 𝑓 (𝑥), 𝑦(𝑥0 ) = 𝑦0 , (2.2)
𝑑𝑥
where the function 𝑔(𝑦) is independent of 𝑥 and 𝑓 (𝑥) is independent of 𝑦. Inte-
gration from 𝑥0 to 𝑥 results in
∫︁ 𝑥 ∫︁ 𝑥
′
𝑔(𝑦(𝑥))𝑦 (𝑥)𝑑𝑥 = 𝑓 (𝑥)𝑑𝑥.
𝑥0 𝑥0
The integral on the left can be transformed by substituting 𝑢 = 𝑦(𝑥), 𝑑𝑢 =

𝑦 ′ (𝑥)𝑑𝑥, and changing the lower and upper limits of integration to 𝑦(𝑥0 ) = 𝑦0
and 𝑦(𝑥) = 𝑦. Therefore,
∫︁ 𝑦 ∫︁ 𝑥
𝑔(𝑢)𝑑𝑢 = 𝑓 (𝑥)𝑑𝑥,
𝑦0 𝑥0
and since 𝑢 is a dummy variable of integration, we can write this in the equivalent
form ∫︁ 𝑦 ∫︁ 𝑥
𝑔(𝑦)𝑑𝑦 = 𝑓 (𝑥)𝑑𝑥. (2.3)
𝑦0 𝑥0
A simpler procedure that also yields (2.3) is to treat 𝑑𝑦/𝑑𝑥 in (2.2) like a fraction.
Multiplying (2.2) by 𝑑𝑥 results in
𝑔(𝑦)𝑑𝑦 = 𝑓 (𝑥)𝑑𝑥,
which is a separated equation with all the dependent variables on the left-side,
and all the independent variables on the right-side. Equation (2.3) then results
directly upon integration.
𝑑𝑦
Example: Solve 𝑑𝑥 + 21 𝑦 = 32 , with 𝑦(0) = 2.
We first manipulate the differential equation to the form
𝑑𝑦 1
= (3 − 𝑦), (2.4)
𝑑𝑥 2
and then treat 𝑑𝑦/𝑑𝑥 as if it was a fraction to separate variables:
𝑑𝑦 1
= 𝑑𝑥.
3−𝑦 2
We integrate the right-side from the initial condition 𝑥 = 0 to 𝑥 and the left-side
from the initial condition 𝑦(0) = 2 to 𝑦. Accordingly,
∫︁ 𝑦 ∫︁ 𝑥
𝑑𝑦 1
= 𝑑𝑥. (2.5)
2 3−𝑦 2 0
2.2. SEPARABLE EQUATIONS 15
dy/dx + y/2 = 3/2

6
3
y
0
0 1 2 3 4 5 6 7
x
𝑑𝑦
Figure 2.1: Solution of the following ode: 𝑑𝑥 + 21 𝑦 = 32 .
The integrals in (2.5) need to be done. Note that 𝑦(𝑥) < 3 for finite 𝑥 or the
integral on the left-side diverges. Therefore, 3 − 𝑦 > 0 and integration yields
]︀𝑦 1 ]︀𝑥
− ln (3 − 𝑦) 2 = 𝑥 0 ,
2
1
ln (3 − 𝑦) = − 𝑥,
2
− 21 𝑥
3−𝑦 =𝑒 ,
1
𝑦 = 3 − 𝑒− 2 𝑥 .
Since this is our first nontrivial analytical solution, it is prudent to check our
result. We do this by differentiating our solution:
𝑑𝑦 1 1
= 𝑒− 2 𝑥
𝑑𝑥 2
1
= (3 − 𝑦);
2
and checking the initial conditions, 𝑦(0) = 3 − 𝑒0 = 2. Therefore, our solution

satisfies both the original ode and the initial condition.
𝑑𝑦
Example: Solve 𝑑𝑥 + 12 𝑦 = 32 , with 𝑦(0) = 4.
This is the identical differential equation as before, but with different initial
conditions. We will jump directly to the integration step:
∫︁ 𝑦
1 𝑥
∫︁
𝑑𝑦
= 𝑑𝑥.
4 3−𝑦 2 0
Now 𝑦(𝑥) > 3, so that 𝑦 − 3 > 0 and integration yields

]︀𝑦1 ]︀𝑥
− ln (𝑦 − 3) 4
=
𝑥 ,
2 0
1
ln (𝑦 − 3) = − 𝑥,
2
1
𝑦 − 3 = 𝑒− 2 𝑥 ,
1
𝑦 = 3 + 𝑒− 2 𝑥 .
The solution curves for a range of initial conditions is presented in Fig. 2.1.
All solutions have a horizontal asymptote at 𝑦 = 3 at which 𝑑𝑦/𝑑𝑥 = 0. For
𝑦(0) = 𝑦0 , the general solution can be shown to be 𝑦(𝑥) = 3+(𝑦0 −3) exp(−𝑥/2).
𝑑𝑦
Example: Solve 𝑑𝑥 = 23+2𝑦
cos 2𝑥
, with 𝑦(0) = −1. (i) For what values of
𝑥 > 0 does the solution exist? (ii) For what value of 𝑥 > 0 is 𝑦(𝑥)
maximum?
Notice that the solution of the ode may not exist when 𝑦 = −3/2, since 𝑑𝑦/𝑑𝑥 →
∞. We separate variables and integrate from initial conditions:
(3 + 2𝑦)𝑑𝑦 = 2 cos 2𝑥 𝑑𝑥
∫︁ 𝑦 ∫︁ 𝑥
(3 + 2𝑦)𝑑𝑦 = 2 cos 2𝑥 𝑑𝑥
−1 0
]︀𝑦 ]︀𝑥
3𝑦 + 𝑦 2 −1 = sin 2𝑥 0
𝑦 2 + 3𝑦 + 2 − sin 2𝑥 = 0
1 √
𝑦± = [−3 ± 1 + 4 sin 2𝑥].
2
Solving the quadratic equation for 𝑦 has introduced a spurious solution that
does not satisfy the initial conditions. We test:
{︂
1 -1;
𝑦± (0) = [−3 ± 1] =
2 -2.
Only the + root satisfies the initial condition, so that the unique solution to the
ode and initial condition is
1 √
𝑦 = [−3 + 1 + 4 sin 2𝑥]. (2.6)
2
To determine (i) the values of 𝑥 > 0 for which the solution exists, we require
1 + 4 sin 2𝑥 ≥ 0,
or
1
sin 2𝑥 ≥ − . (2.7)
4
Notice that at 𝑥 = 0, we have sin 2𝑥 = 0; at 𝑥 = 𝜋/4, we have sin 2𝑥 = 1;
at 𝑥 = 𝜋/2, we have sin 2𝑥 = 0; and at 𝑥 = 3𝜋/4, we have sin 2𝑥 = −1 We
therefore need to determine the value of 𝑥 such that sin 2𝑥 = −1/4, with 𝑥 in
the range 𝜋/2 < 𝑥 < 3𝜋/4. The solution to the ode will then exist for all 𝑥
between zero and this value.
2.3. LINEAR EQUATIONS 17
(3+2y) dy/dx = 2 cos 2x, y(0) = −1

0
−0.2
−0.4
−0.6
−0.8
y
−1
−1.2
−1.4
−1.6
0 0.5 1 1.5
x
Figure 2.2: Solution of the following ode: (3 + 2𝑦)𝑦 ′ = 2 cos 2𝑥, 𝑦(0) = −1.
To solve sin 2𝑥 = −1/4 for 𝑥 in the interval 𝜋/2 < 𝑥 < 3𝜋/4, one needs to
recall the definition of arcsin, or sin−1 , as found on a typical scientific calculator.
The inverse of the function
𝑓 (𝑥) = sin 𝑥, −𝜋/2 ≤ 𝑥 ≤ 𝜋/2
is denoted by arcsin. The first solution with 𝑥 > 0 of the equation sin 2𝑥 = −1/4
places 2𝑥 in the interval (𝜋, 3𝜋/2), so to invert this equation using the arcsine
we need to apply the identity sin (𝜋 − 𝑥) = sin 𝑥, and rewrite sin 2𝑥 = −1/4 as
sin (𝜋 − 2𝑥) = −1/4. The solution of this equation may then be found by taking
the arcsine, and is
𝜋 − 2𝑥 = arcsin (−1/4),
or (︂ )︂
1 1
𝑥= 𝜋 + arcsin .
2 4
Therefore the solution exists for 0 ≤ 𝑥 ≤ (𝜋 + arcsin (1/4)) /2 = 1.6971 . . . ,
where we have used a calculator value (computing in radians) to find arcsin(0.25) =
0.2527 . . . . At the value (𝑥, 𝑦) = (1.6971 . . . , −3/2), the solution curve ends and
𝑑𝑦/𝑑𝑥 becomes infinite.
To determine (ii) the value of 𝑥 at which 𝑦 = 𝑦(𝑥) is maximum, we examine
(2.6) directly. The value of 𝑦 will be maximum when sin 2𝑥 takes its maximum
value over the interval where the solution exists. This will be when 2𝑥 = 𝜋/2,
or 𝑥 = 𝜋/4 = 0.7854 . . . .
The graph of 𝑦 = 𝑦(𝑥) is shown in Fig. 2.2.
2.3 Linear equations

view tutorial
The first-order linear differential equation (linear in 𝑦 and its derivative) can be
written in the form
𝑑𝑦
+ 𝑝(𝑥)𝑦 = 𝑔(𝑥), (2.8)
𝑑𝑥
with the initial condition 𝑦(𝑥0 ) = 𝑦0 . Linear first-order equations can be inte-
grated using an integrating factor 𝜇(𝑥). We multiply (2.8) by 𝜇(𝑥),
[︂ ]︂
𝑑𝑦
𝜇(𝑥) + 𝑝(𝑥)𝑦 = 𝜇(𝑥)𝑔(𝑥), (2.9)
𝑑𝑥
and try to determine 𝜇(𝑥) so that
[︂ ]︂
𝑑𝑦 𝑑
𝜇(𝑥) + 𝑝(𝑥)𝑦 = [𝜇(𝑥)𝑦]. (2.10)
𝑑𝑥 𝑑𝑥
Equation (2.9) then becomes
𝑑
[𝜇(𝑥)𝑦] = 𝜇(𝑥)𝑔(𝑥). (2.11)
𝑑𝑥
Equation (2.11) is easily integrated using 𝜇(𝑥0 ) = 𝜇0 and 𝑦(𝑥0 ) = 𝑦0 :
∫︁ 𝑥
𝜇(𝑥)𝑦 − 𝜇0 𝑦0 = 𝜇(𝑥)𝑔(𝑥)𝑑𝑥,
𝑥0
or (︂ ∫︁ 𝑥 )︂
1
𝑦= 𝜇0 𝑦0 + 𝜇(𝑥)𝑔(𝑥)𝑑𝑥 . (2.12)
𝜇(𝑥) 𝑥0
It remains to determine 𝜇(𝑥) from (2.10). Differentiating and expanding (2.10)
yields
𝑑𝑦 𝑑𝜇 𝑑𝑦
𝜇 + 𝑝𝜇𝑦 = 𝑦+𝜇 ;
𝑑𝑥 𝑑𝑥 𝑑𝑥
and upon simplifying,
𝑑𝜇
= 𝑝𝜇. (2.13)
𝑑𝑥
Equation (2.13) is separable and can be integrated:
∫︁ 𝜇 ∫︁ 𝑥
𝑑𝜇
= 𝑝(𝑥)𝑑𝑥,
𝜇0 𝜇 𝑥0
∫︁ 𝑥
𝜇
ln = 𝑝(𝑥)𝑑𝑥,
𝜇0 𝑥0
(︂∫︁ 𝑥 )︂
𝜇(𝑥) = 𝜇0 exp 𝑝(𝑥)𝑑𝑥 .
𝑥0
Notice that since 𝜇0 cancels out of (2.12), it is customary to assign 𝜇0 = 1. The

solution to (2.8) satisfying the initial condition 𝑦(𝑥0 ) = 𝑦0 is then commonly
written as (︂ ∫︁ 𝑥 )︂
1
𝑦= 𝑦0 + 𝜇(𝑥)𝑔(𝑥)𝑑𝑥 ,
𝜇(𝑥) 𝑥0
with (︂∫︁ 𝑥 )︂
𝜇(𝑥) = exp 𝑝(𝑥)𝑑𝑥
𝑥0
the integrating factor. This important result finds frequent use in applied math-
ematics.
2.3. LINEAR EQUATIONS 19
𝑑𝑦
Example: Solve 𝑑𝑥 + 2𝑦 = 𝑒−𝑥 , with 𝑦(0) = 3/4.
Note that this equation is not separable. With 𝑝(𝑥) = 2 and 𝑔(𝑥) = 𝑒−𝑥 , we
have
(︂∫︁ 𝑥 )︂
𝜇(𝑥) = exp 2𝑑𝑥
0
= 𝑒2𝑥 ,
and
(︂ ∫︁ 𝑥 )︂
3
𝑦 = 𝑒−2𝑥 + 𝑒2𝑥 𝑒−𝑥 𝑑𝑥
4 0
(︂ ∫︁ 𝑥 )︂
3
= 𝑒−2𝑥 + 𝑒𝑥 𝑑𝑥
4 0
(︂ )︂
3
= 𝑒−2𝑥 + (𝑒𝑥 − 1)
4
(︂ )︂
1
= 𝑒−2𝑥 𝑒𝑥 −
4
(︂ )︂
1
= 𝑒−𝑥 1 − 𝑒−𝑥 .
4
𝑑𝑦
Example: Solve 𝑑𝑥 − 2𝑥𝑦 = 𝑥, with 𝑦(0) = 0.
This equation is separable, and we solve it in two ways. First, using an inte-
grating factor with 𝑝(𝑥) = −2𝑥 and 𝑔(𝑥) = 𝑥:
(︂ ∫︁ 𝑥 )︂
𝜇(𝑥) = exp −2 𝑥𝑑𝑥
0
−𝑥2
=𝑒 ,
and ∫︁ 𝑥
2 2
𝑦 = 𝑒𝑥 𝑥𝑒−𝑥 𝑑𝑥.
0
The integral can be done by substitution with 𝑢 = 𝑥2 , 𝑑𝑢 = 2𝑥𝑑𝑥:

∫︁ 𝑥 ∫︁ 𝑥2
2 1
𝑥𝑒−𝑥 𝑑𝑥 = 𝑒−𝑢 𝑑𝑢
0 2 0
1 ]︀𝑥2
= − 𝑒−𝑢 0
2
1 (︁ 2
)︁
= 1 − 𝑒−𝑥 .
2
Therefore,
1 𝑥2 (︁ 2
)︁
𝑦= 𝑒 1 − 𝑒−𝑥
2
1 (︁ 𝑥2 )︁
= 𝑒 −1 .
2
Second, we integrate by separating variables:
𝑑𝑦
− 2𝑥𝑦 = 𝑥,
𝑑𝑥
𝑑𝑦
= 𝑥(1 + 2𝑦),
∫︁ 𝑦𝑑𝑥 ∫︁ 𝑥
𝑑𝑦
= 𝑥𝑑𝑥,
0 1 + 2𝑦 0
1 1
ln (1 + 2𝑦) = 𝑥2 ,
2 2
2
1 + 2𝑦 = 𝑒𝑥 ,
1 (︁ 𝑥2 )︁
𝑦= 𝑒 −1 .
2
The results from the two different solution methods are the same, and the choice
of method is a personal preference.
2.4 Applications
2.4.1 Compound interest
view tutorial
The equation for the growth of an investment with continuous compounding
of interest is a first-order differential equation. Let 𝑆(𝑡) be the value of the
investment at time 𝑡, and let 𝑟 be the annual interest rate compounded after
every time interval Δ𝑡. We can also include deposits (or withdrawals). Let 𝑘 be
the annual deposit amount, and suppose that an installment is deposited after
every time interval Δ𝑡. The value of the investment at the time 𝑡 + Δ𝑡 is then
given by
𝑆(𝑡 + Δ𝑡) = 𝑆(𝑡) + (𝑟Δ𝑡)𝑆(𝑡) + 𝑘Δ𝑡, (2.14)
where at the end of the time interval Δ𝑡, 𝑟Δ𝑡𝑆(𝑡) is the amount of interest
credited and 𝑘Δ𝑡 is the amount of money deposited (𝑘 > 0) or withdrawn
(𝑘 < 0). As a numerical example, if the account held $10,000 at time 𝑡, and
𝑟 = 6% per year and 𝑘 = $12,000 per year, say, and the compounding and
deposit period is Δ𝑡 = 1 month = 1/12 year, then the interest awarded after
one month is 𝑟Δ𝑡𝑆 = (0.06/12) × $10,000 = $50, and the amount deposited is
𝑘Δ𝑡 = $1000.
Rearranging the terms of (2.14) to exhibit what will soon become a deriva-
tive, we have
𝑆(𝑡 + Δ𝑡) − 𝑆(𝑡)
= 𝑟𝑆(𝑡) + 𝑘.
Δ𝑡
The equation for continuous compounding of interest and continuous deposits
is obtained by taking the limit Δ𝑡 → 0. The resulting differential equation is
𝑑𝑆
= 𝑟𝑆 + 𝑘, (2.15)
𝑑𝑡
which can solved with the initial condition 𝑆(0) = 𝑆0 , where 𝑆0 is the initial
capital. We can solve either by separating variables or by using an integrating
2.4. APPLICATIONS 21
factor; I solve here by separating variables. Integrating from 𝑡 = 0 to a final

time 𝑡,
∫︁ 𝑆 ∫︁ 𝑡
𝑑𝑆
= 𝑑𝑡,
𝑆0 𝑟𝑆 + 𝑘 0
(︂ )︂
1 𝑟𝑆 + 𝑘
ln = 𝑡,
𝑟 𝑟𝑆0 + 𝑘
𝑟𝑆 + 𝑘 = (𝑟𝑆0 + 𝑘)𝑒𝑟𝑡 ,
𝑟𝑆0 𝑒𝑟𝑡 + 𝑘𝑒𝑟𝑡 − 𝑘
𝑆= ,
𝑟
𝑘
𝑆 = 𝑆0 𝑒𝑟𝑡 + 𝑒𝑟𝑡 1 − 𝑒−𝑟𝑡 ,
(︀ )︀
(2.16)
𝑟
where the first term on the right-hand side of (2.16) comes from the initial
invested capital, and the second term comes from the deposits (or withdrawals).
Evidently, compounding results in the exponential growth of an investment.
As a practical example, we can analyze a simple retirement plan. It is
easiest to assume that all amounts and returns are in real dollars (adjusted for
inflation). Suppose a 25 year-old plans to set aside a fixed amount every year of
his/her working life, invests at a real return of 6%, and retires at age 65. How
much must he/she invest each year to have $8,000,000 at retirement? We need
to solve (2.16) for 𝑘 using 𝑡 = 40 years, 𝑆(𝑡) = $8,000,000, 𝑆0 = 0, and 𝑟 = 0.06
per year. We have
𝑟𝑆(𝑡)
𝑘= ,
𝑒𝑟𝑡 − 1
0.06 × 8,000,000
𝑘= ,
𝑒0.06×40 − 1
= $47,889 year−1 .
To have saved approximately one million US$ at retirement, the worker would
need to save about HK$50,000 per year over his/her working life. Note that the
amount saved over the worker’s life is approximately 40 × $50,000 = $2,000,000,
while the amount earned on the investment (at the assumed 6% real return) is
approximately $8,000,000 − $2,000,000 = $6,000,000. The amount earned from
the investment is about 3× the amount saved, even with the modest real return
of 6%. Sound investment planning is well worth the effort.
2.4.2 Chemical reactions

Suppose that two chemicals 𝐴 and 𝐵 react to form a product 𝐶, which we write
as
𝑘
𝐴 + 𝐵 → 𝐶,
where 𝑘 is called the rate constant of the reaction. For simplicity, we will use
the same symbol 𝐶, say, to refer to both the chemical 𝐶 and its concentration.
The law of mass action says that 𝑑𝐶/𝑑𝑡 is proportional to the product of the
concentrations 𝐴 and 𝐵, with proportionality constant 𝑘; that is,
𝑑𝐶
= 𝑘𝐴𝐵. (2.17)
𝑑𝑡
Similarly, the law of mass action enables us to write equations for the time-
derivatives of the reactant concentrations 𝐴 and 𝐵:
𝑑𝐴 𝑑𝐵
= −𝑘𝐴𝐵, = −𝑘𝐴𝐵. (2.18)
𝑑𝑡 𝑑𝑡
The ode given by (2.17) can be solved analytically using conservation laws. We
assume that 𝐴0 and 𝐵0 are the initial concentrations of the reactants, and that
no product is initially present. From (2.17) and (2.18),
𝑑
(𝐴 + 𝐶) = 0 =⇒ 𝐴 + 𝐶 = 𝐴0 ,
𝑑𝑡
𝑑
(𝐵 + 𝐶) = 0 =⇒ 𝐵 + 𝐶 = 𝐵0 .
𝑑𝑡
Using these conservation laws, (2.17) becomes
𝑑𝐶
= 𝑘(𝐴0 − 𝐶)(𝐵0 − 𝐶), 𝐶(0) = 0,
𝑑𝑡
which is a nonlinear equation that may be integrated by separating variables.
Separating and integrating, we obtain
∫︁ 𝐶 ∫︁ 𝑡
𝑑𝐶
=𝑘 𝑑𝑡
0 (𝐴0 − 𝐶)(𝐵0 − 𝐶) 0
= 𝑘𝑡. (2.19)
The remaining integral can be done using the method of partial fractions. We
write
1 𝑎 𝑏
= + . (2.20)
(𝐴0 − 𝐶)(𝐵0 − 𝐶) 𝐴0 − 𝐶 𝐵0 − 𝐶
The cover-up method is the simplest method to determine the unknown coeffi-
cients 𝑎 and 𝑏. To determine 𝑎, we multiply both sides of (2.20) by 𝐴0 − 𝐶 and
set 𝐶 = 𝐴0 to find
1
𝑎= .
𝐵 0 − 𝐴0
Similarly, to determine 𝑏, we multiply both sides of (2.20) by 𝐵0 − 𝐶 and set
𝐶 = 𝐵0 to find
1
𝑏= .
𝐴0 − 𝐵 0
Therefore,
(︂ )︂
1 1 1 1
= − ,
(𝐴0 − 𝐶)(𝐵0 − 𝐶) 𝐵 0 − 𝐴0 𝐴0 − 𝐶 𝐵0 − 𝐶
and the remaining integral of (2.19) becomes (using 𝐶 < 𝐴0 , 𝐵0 )
∫︁ 𝐶 (︃∫︁ )︃
𝐶 ∫︁ 𝐶
𝑑𝐶 1 𝑑𝐶 𝑑𝐶
= −
0 (𝐴0 − 𝐶)(𝐵0 − 𝐶) 𝐵0 − 𝐴0 0 𝐴0 − 𝐶 0 𝐵0 − 𝐶
(︂ (︂ )︂ (︂ )︂)︂
1 𝐴0 − 𝐶 𝐵0 − 𝐶
= − ln + ln
𝐵0 − 𝐴0 𝐴0 𝐵0
(︂ )︂
1 𝐴0 (𝐵0 − 𝐶)
= ln .
𝐵0 − 𝐴0 𝐵0 (𝐴0 − 𝐶)
Using this integral in (2.19), multiplying by (𝐵0 − 𝐴0 ) and exponentiating, we

obtain
𝐴0 (𝐵0 − 𝐶)
= 𝑒(𝐵0 −𝐴0 )𝑘𝑡 .
𝐵0 (𝐴0 − 𝐶)
Solving for 𝐶, we finally obtain
𝑒(𝐵0 −𝐴0 )𝑘𝑡 − 1

𝐶(𝑡) = 𝐴0 𝐵0 ,
𝐵0 𝑒(𝐵0 −𝐴0 )𝑘𝑡 − 𝐴0
which appears to be a complicated expression, but has the simple limits

{︃
𝐴0 , if 𝐴0 < 𝐵0 ,
lim 𝐶(𝑡) =
𝑡→∞ 𝐵0 , if 𝐵0 < 𝐴0
= min(𝐴0 , 𝐵0 ).
As one would expect, the reaction stops after one of the reactants is depleted;
and the final concentration of product is equal to the initial concentration of
the depleted reactant.
2.4.3 Terminal velocity

view tutorial
Using Newton’s law, we model a mass 𝑚 free falling under gravity but with
air resistance. We assume that the force of air resistance is proportional to the
speed of the mass and opposes the direction of motion. We define the 𝑥-axis to
point in the upward direction, opposite the force of gravity. Near the surface
of the Earth, the force of gravity is approximately constant and is given by
−𝑚𝑔, with 𝑔 = 9.8 m/s2 the usual gravitational acceleration. The force of air
resistance is modeled by −𝑘𝑣, where 𝑣 is the vertical velocity of the mass and
𝑘 is a positive constant. When the mass is falling, 𝑣 < 0 and the force of air
resistance is positive, pointing upward and opposing the motion. The total force
on the mass is therefore given by 𝐹 = −𝑚𝑔 − 𝑘𝑣. With 𝐹 = 𝑚𝑎 and 𝑎 = 𝑑𝑣/𝑑𝑡,
we obtain the differential equation
𝑑𝑣
𝑚 = −𝑚𝑔 − 𝑘𝑣. (2.21)
𝑑𝑡
The terminal velocity 𝑣∞ of the mass is defined as the asymptotic velocity after
air resistance balances the gravitational force. When the mass is at terminal
velocity, 𝑑𝑣/𝑑𝑡 = 0 so that
𝑚𝑔
𝑣∞ = − . (2.22)
𝑘
The approach to the terminal velocity of a mass initially at rest is obtained by

solving (2.21) with initial condition 𝑣(0) = 0. The equation is both linear and
separable, and I solve by separating variables:

∫︁ 𝑣 ∫︁ 𝑡
𝑑𝑣
𝑚 =− 𝑑𝑡,
0 𝑚𝑔 + 𝑘𝑣 0
(︂ )︂
𝑚 𝑚𝑔 + 𝑘𝑣
ln = −𝑡,
𝑘 𝑚𝑔
𝑘𝑣
1+ = 𝑒−𝑘𝑡/𝑚 ,
𝑚𝑔
𝑚𝑔 (︁ )︁
𝑣=− 1 − 𝑒−𝑘𝑡/𝑚 .
𝑘
Therefore, 𝑣 = 𝑣∞ 1 − 𝑒−𝑘𝑡/𝑚 , and 𝑣 approaches 𝑣∞ as the exponential term
(︀ )︀
decays to zero.
As an example, a skydiver of mass 𝑚 = 100 kg with his parachute closed
may have a terminal velocity of 200 km/hr. With
2 2
𝑔 = (9.8 m/s )(10−3 km/m)(60 s/min)2 (60 min/hr)2 = 127, 008 km/hr ,
one obtains from (2.22), 𝑘 = 63, 504 kg/hr. One-half of the terminal velocity
for free-fall (100 km/hr) is therefore attained when (1 − 𝑒−𝑘𝑡/𝑚 ) = 1/2, or 𝑡 =
𝑚 ln 2/𝑘 ≈ 4 sec. Approximately 95% of the terminal velocity (190 km/hr ) is
attained after 17 sec.
2.4.4 Escape velocity

view tutorial
An interesting physical problem is to find the smallest initial velocity for a
mass on the Earth’s surface to escape from the Earth’s gravitational field, the
so-called escape velocity. Newton’s law of universal gravitation asserts that the
gravitational force between two massive bodies is proportional to the product of
the two masses and inversely proportional to the square of the distance between
them. For a mass 𝑚 a position 𝑥 above the surface of the Earth, the force on
the mass is given by
𝑀𝑚
𝐹 = −𝐺 ,
(𝑅 + 𝑥)2
where 𝑀 and 𝑅 are the mass and radius of the Earth and 𝐺 is the gravitational
constant. The minus sign means the force on the mass 𝑚 points in the direction
of decreasing 𝑥. The approximately constant acceleration 𝑔 on the Earth’s
surface corresponds to the absolute value of 𝐹/𝑚 when 𝑥 = 0:
𝐺𝑀
𝑔= ,
𝑅2
and 𝑔 ≈ 9.8 m/s2 . Newton’s law 𝐹 = 𝑚𝑎 for the mass 𝑚 is thus given by
𝑑2 𝑥 𝐺𝑀
2
=−
𝑑𝑡 (𝑅 + 𝑥)2
𝑔
=− , (2.23)
(1 + 𝑥/𝑅)2
where the radius of the Earth is known to be 𝑅 ≈ 6350 km.

A useful trick allows us to solve this second-order differential equation as

a first-order equation. First, note that 𝑑2 𝑥/𝑑𝑡2 = 𝑑𝑣/𝑑𝑡. If we write 𝑣(𝑡) =
𝑣(𝑥(𝑡))—considering the velocity of the mass 𝑚 to be a function of its distance
above the Earth—we have using the chain rule
𝑑𝑣 𝑑𝑣 𝑑𝑥
=
𝑑𝑡 𝑑𝑥 𝑑𝑡
𝑑𝑣
=𝑣 ,
𝑑𝑥
where we have used 𝑣 = 𝑑𝑥/𝑑𝑡. Therefore, (2.23) becomes the first-order ode
𝑑𝑣 𝑔
𝑣 =− ,
𝑑𝑥 (1 + 𝑥/𝑅)2
which may be solved assuming an initial velocity 𝑣(𝑥 = 0) = 𝑣0 when the mass
is shot vertically from the Earth’s surface. Separating variables and integrating,
we obtain ∫︁ 𝑣 ∫︁ 𝑥
𝑑𝑥
𝑣𝑑𝑣 = −𝑔 .
𝑣0 0 (1 + 𝑥/𝑅)2
The left integral is 21 (𝑣 2 − 𝑣02 ), and the right integral can be performed using
the substitution 𝑢 = 1 + 𝑥/𝑅, 𝑑𝑢 = 𝑑𝑥/𝑅:
∫︁ 𝑥 ∫︁ 1+𝑥/𝑅
𝑑𝑥 𝑑𝑢
2
= 𝑅
0 (1 + 𝑥/𝑅) 1 𝑢2
]︂1+𝑥/𝑅
𝑅
=−
𝑢 1
𝑅2
=𝑅−
𝑥+𝑅
𝑅𝑥
= .
𝑥+𝑅
Therefore,
1 2 𝑔𝑅𝑥
(𝑣 − 𝑣02 ) = − ,
2 𝑥+𝑅
which when multiplied by 𝑚 is an expression of the conservation of energy (the
change of the kinetic energy of the mass is equal to the change in the potential
energy). Solving for 𝑣 2 ,
2𝑔𝑅𝑥
𝑣 2 = 𝑣02 − .
𝑥+𝑅
The escape velocity is defined as the minimum initial velocity 𝑣0 such that
the mass can escape to infinity. Therefore, 𝑣0 = 𝑣escape when 𝑣 → 0 as 𝑥 → ∞.
Taking this limit, we have
2 2𝑔𝑅𝑥
𝑣escape = lim
𝑥→∞ 𝑥+𝑅
= 2𝑔𝑅.
2 √
With 𝑅 ≈ 6350 km and 𝑔 = 127 008 km/hr , we determine 𝑣escape = 2𝑔𝑅 ≈
40 000 km/hr. In comparison, the muzzle velocity of a modern high-performance
rifle is 4300 km/hr, almost an order of magnitude too slow for a bullet, shot
into the sky, to escape the Earth’s gravity.

i
(a)
(b)
+
C
ߝ
_
Figure 2.3: RC circuit diagram.
2.4.5 RC circuit
view tutorial
Consider a resister 𝑅 and a capacitor 𝐶 connected in series as shown in Fig. 2.3.
A battery providing an electromotive force, or emf ℰ, connects to this circuit
by a switch. Initially, there is no charge on the capacitor. When the switch
is thrown to a, the battery connects and the capacitor charges. When the
switch is thrown to b, the battery disconnects and the capacitor discharges,
with energy dissipated in the resister. Here, we determine the voltage drop
across the capacitor during charging and discharging.
The equations for the voltage drops across a capacitor and a resister are
given by
𝑉𝐶 = 𝑞/𝐶, 𝑉𝑅 = 𝑖𝑅, (2.24)
where 𝐶 is the capacitance and 𝑅 is the resistance. The charge 𝑞 and the current
𝑖 are related by
𝑑𝑞
𝑖= . (2.25)
𝑑𝑡
Kirchhoff’s voltage law states that the emf ℰ in any closed loop is equal to
the sum of the voltage drops in that loop. Applying Kirchhoff’s voltage law
when the switch is thrown to a results in
𝑉𝑅 + 𝑉𝐶 = ℰ. (2.26)
Using (2.24) and (2.25), the voltage drop across the resister can be written in
terms of the voltage drop across the capacitor as
𝑑𝑉𝐶
𝑉𝑅 = 𝑅𝐶 ,
𝑑𝑡
and (2.26) can be rewritten to yield the first-order linear differential equation
for 𝑉𝑐 given by
𝑑𝑉𝐶
+ 𝑉𝐶 /𝑅𝐶 = ℰ/𝑅𝐶, (2.27)
𝑑𝑡
with initial condition 𝑉𝐶 (0) = 0.
The integrating factor for this equation is
𝜇(𝑡) = 𝑒𝑡/𝑅𝐶 ,
and (2.27) integrates to

∫︁ 𝑡
𝑉𝐶 (𝑡) = 𝑒−𝑡/𝑅𝐶 (ℰ/𝑅𝐶)𝑒𝑡/𝑅𝐶 𝑑𝑡,
0
with solution (︁ )︁
𝑉𝐶 (𝑡) = ℰ 1 − 𝑒−𝑡/𝑅𝐶 .
The voltage starts at zero and rises exponentially to ℰ, with characteristic time
scale given by 𝑅𝐶.
When the switch is thrown to b, application of Kirchhoff’s voltage law results
in
𝑉𝑅 + 𝑉𝐶 = 0,
with corresponding differential equation
𝑑𝑉𝐶
+ 𝑉𝐶 /𝑅𝐶 = 0.
𝑑𝑡
Here, we assume that the capacitance is initially fully charged so that 𝑉𝐶 (0) = ℰ.
The solution, then, during the discharge phase is given by
𝑉𝐶 (𝑡) = ℰ𝑒−𝑡/𝑅𝐶 .
The voltage starts at ℰ and decays exponentially to zero, again with character-
istic time scale given by 𝑅𝐶.
2.4.6 The logistic equation

view tutorial
Let 𝑁 = 𝑁 (𝑡) be the size of a population at time 𝑡 and let 𝑟 be the growth
rate. The Malthusian growth model (Thomas Malthus, 1766-1834), similar to
a compound interest model, is given by
𝑑𝑁
= 𝑟𝑁.
𝑑𝑡
Under a Malthusian growth model, a population grows exponentially like
𝑁 (𝑡) = 𝑁0 𝑒𝑟𝑡 ,
where 𝑁0 is the initial population size. However, when the population growth
is constrained by limited resources, a heuristic modification to the Malthusian
growth model results in the Verhulst equation,
(︂ )︂
𝑑𝑁 𝑁
= 𝑟𝑁 1 − , (2.28)
𝑑𝑡 𝐾
where 𝐾 is called the carrying capacity of the environment. Making (2.28)
dimensionless using 𝜏 = 𝑟𝑡 and 𝑥 = 𝑁/𝐾 leads to the logistic equation,
𝑑𝑥
= 𝑥(1 − 𝑥),
𝑑𝜏
where we may assume the initial condition 𝑥(0) = 𝑥0 > 0. Separating variables
and integrating ∫︁ 𝑥 ∫︁ 𝜏
𝑑𝑥
= 𝑑𝜏.
𝑥0 𝑥(1 − 𝑥) 0
The integral on the left-hand-side can be done using the method of partial
fractions:
1 𝑎 𝑏
= +
𝑥(1 − 𝑥) 𝑥 1−𝑥
𝑎 + (𝑏 − 𝑎)𝑥
= ;
𝑥(1 − 𝑥)
and equating the coefficients of the numerators proportional to 𝑥0 and 𝑥1 , we

have 𝑎 = 𝑏 = 1. Therefore,
∫︁ 𝑥 ∫︁ 𝑥 ∫︁ 𝑥
𝑑𝑥 𝑑𝑥 𝑑𝑥
= +
𝑥0 𝑥(1 − 𝑥) 𝑥0 𝑥 𝑥0 (1 − 𝑥)
𝑥 1−𝑥
= ln − ln
𝑥0 1 − 𝑥0
𝑥(1 − 𝑥0 )
= ln
𝑥0 (1 − 𝑥)
= 𝜏.
Solving for 𝑥, we first exponentiate both sides and then isolate 𝑥:
𝑥(1 − 𝑥0 )
= 𝑒𝜏 ,
𝑥0 (1 − 𝑥)
𝑥(1 − 𝑥0 ) = 𝑥0 𝑒𝜏 − 𝑥𝑥0 𝑒𝜏 ,
𝑥(1 − 𝑥0 + 𝑥0 𝑒𝜏 ) = 𝑥0 𝑒𝜏 ,
𝑥0
𝑥= . (2.29)
𝑥0 + (1 − 𝑥0 )𝑒−𝜏
We observe that for 𝑥0 > 0, we have lim𝜏 →∞ 𝑥(𝜏 ) = 1, corresponding to
lim 𝑁 (𝑡) = 𝐾.
𝑡→∞
The population, therefore, grows in size until it reaches the carrying capacity of
its environment.
Chapter 3
Second-order linear
differential equations with
constant coefficients
The general second-order linear differential equation with independent variable

𝑡 and dependent variable 𝑥 = 𝑥(𝑡) is given by
𝑥
¨ + 𝑝(𝑡)𝑥˙ + 𝑞(𝑡)𝑥 = 𝑔(𝑡), (3.1)
where we have used the standard physics notation 𝑥˙ = 𝑑𝑥/𝑑𝑡 and 𝑥 ¨ = 𝑑2 𝑥/𝑑𝑡2 .
A unique solution of (3.1) requires initial values 𝑥(𝑡0 ) = 𝑥0 and 𝑥(𝑡
˙ 0 ) = 𝑢0 .
The equation with constant coefficients—on which we will devote considerable
effort—assumes that 𝑝(𝑡) and 𝑞(𝑡) are constants, independent of time. The
second-order linear ode is said to be homogeneous if 𝑔(𝑡) = 0.
3.1 The Euler method

view tutorial
In general, (3.1) cannot be solved analytically, and we begin by deriving an
algorithm for numerical solution. Consider the general second-order ode given
by
𝑥
¨ = 𝑓 (𝑡, 𝑥, 𝑥).
˙
We can write this second-order ode as a pair of first-order odes by defining
𝑢 = 𝑥,
˙ and writing the first-order system as
𝑥˙ = 𝑢, (3.2)
𝑢˙ = 𝑓 (𝑡, 𝑥, 𝑢). (3.3)
The first ode, (3.2), gives the slope of the tangent line to the curve 𝑥 = 𝑥(𝑡);
the second ode, (3.3), gives the slope of the tangent line to the curve 𝑢 = 𝑢(𝑡).
Beginning at the initial values (𝑥, 𝑢) = (𝑥0 , 𝑢0 ) at the time 𝑡 = 𝑡0 , we move
along the tangent lines to determine 𝑥1 = 𝑥(𝑡0 + Δ𝑡) and 𝑢1 = 𝑢(𝑡0 + Δ𝑡):
𝑥1 = 𝑥0 + Δ𝑡𝑢0 ,
𝑢1 = 𝑢0 + Δ𝑡𝑓 (𝑡0 , 𝑥0 , 𝑢0 ).
29
30 CHAPTER 3. SECOND-ORDER ODES, CONSTANT COEFFICIENTS
The values 𝑥1 and 𝑢1 at the time 𝑡1 = 𝑡0 + Δ𝑡 are then used as new initial values
to march the solution forward to time 𝑡2 = 𝑡1 + Δ𝑡. As long as 𝑓 (𝑡, 𝑥, 𝑢) is a
well-behaved function, the numerical solution converges to the unique solution
of the ode as Δ𝑡 → 0.
3.2 The principle of superposition

view tutorial
Consider the second-order linear homogeneous ode:
𝑥
¨ + 𝑝(𝑡)𝑥˙ + 𝑞(𝑡)𝑥 = 0; (3.4)
and suppose that 𝑥 = 𝑋1 (𝑡) and 𝑥 = 𝑋2 (𝑡) are solutions to (3.4). We consider
a linear combination of 𝑋1 and 𝑋2 by letting
𝑋(𝑡) = 𝑐1 𝑋1 (𝑡) + 𝑐2 𝑋2 (𝑡), (3.5)
with 𝑐1 and 𝑐2 constants. The principle of superposition states that 𝑥 = 𝑋(𝑡)

is also a solution of (3.4). To prove this, we compute
(︁ )︁
𝑋¨ + 𝑝𝑋˙ + 𝑞𝑋 = 𝑐1 𝑋 ¨ 1 + 𝑐2 𝑋¨ 2 + 𝑝 𝑐1 𝑋˙ 1 + 𝑐2 𝑋˙ 2 + 𝑞 (𝑐1 𝑋1 + 𝑐2 𝑋2 )
(︁ )︁ (︁ )︁
= 𝑐1 𝑋 ¨ 1 + 𝑝𝑋˙ 1 + 𝑞𝑋1 + 𝑐2 𝑋 ¨ 2 + 𝑝𝑋˙ 2 + 𝑞𝑋2
= 𝑐1 × 0 + 𝑐2 × 0
= 0,
since 𝑋1 and 𝑋2 were assumed to be solutions of (3.4). We have therefore shown

that any linear combination of solutions to the second-order linear homogeneous
ode is also a solution.
3.3 The Wronskian

view tutorial
Suppose that having determined that two solutions of (3.4) are 𝑥 = 𝑋1 (𝑡) and
𝑥 = 𝑋2 (𝑡), we attempt to write the general solution to (3.4) as (3.5). We must
then ask whether this general solution will be able to satisfy the two initial
conditions given by
𝑥(𝑡0 ) = 𝑥0 , 𝑥(𝑡
˙ 0 ) = 𝑢0 . (3.6)
Applying these initial conditions to (3.5), we obtain
𝑐1 𝑋1 (𝑡0 ) + 𝑐2 𝑋2 (𝑡0 ) = 𝑥0 ,
𝑐1 𝑋˙ 1 (𝑡0 ) + 𝑐2 𝑋˙ 2 (𝑡0 ) = 𝑢0 , (3.7)
which is observed to be a system of two linear equations for the two unknowns
𝑐1 and 𝑐2 . Solution of (3.7) by standard methods results in
𝑥0 𝑋˙ 2 (𝑡0 ) − 𝑢0 𝑋2 (𝑡0 ) 𝑢0 𝑋1 (𝑡0 ) − 𝑥0 𝑋˙ 1 (𝑡0 )

𝑐1 = , 𝑐2 = ,
𝑊 𝑊
3.4. HOMOGENEOUS ODES 31
where 𝑊 is called the Wronskian and is given by

𝑊 = 𝑋1 (𝑡0 )𝑋˙ 2 (𝑡0 ) − 𝑋˙ 1 (𝑡0 )𝑋2 (𝑡0 ). (3.8)
Evidently, the Wronskian must not be equal to zero (𝑊 ̸= 0) for a solution to
exist.
For examples, the two solutions
𝑋1 (𝑡) = 𝐴 sin 𝜔𝑡, 𝑋2 (𝑡) = 𝐵 sin 𝜔𝑡,
have a zero Wronskian at 𝑡 = 𝑡0 , as can be shown by computing
𝑊 = (𝐴 sin 𝜔𝑡0 ) (𝐵𝜔 cos 𝜔𝑡0 ) − (𝐴𝜔 cos 𝜔𝑡0 ) (𝐵 sin 𝜔𝑡0 )
= 0;
while the two solutions
𝑋1 (𝑡) = sin 𝜔𝑡, 𝑋2 (𝑡) = cos 𝜔𝑡,
with 𝜔 ̸= 0, have a nonzero Wronskian at 𝑡 = 𝑡0 ,
𝑊 = (sin 𝜔𝑡0 ) (−𝜔 sin 𝜔𝑡0 ) − (𝜔 cos 𝜔𝑡0 ) (cos 𝜔𝑡0 )
= −𝜔.
When the Wronskian is not equal to zero, we say that the two solutions
𝑋1 (𝑡) and 𝑋2 (𝑡) are linearly independent. The concept of linear independence
is borrowed from linear algebra, and indeed, the set of all functions that satisfy
(3.4) can be shown to form a two-dimensional vector space.
3.4 Second-order linear homogeneous ode with

constant coefficients
view tutorial
We now study solutions of the homogeneous, constant coefficient ode, written
as
𝑎¨
𝑥 + 𝑏𝑥˙ + 𝑐𝑥 = 0, (3.9)
with 𝑎, 𝑏, and 𝑐 constants. Such an equation arises for the charge on a capacitor
in an unpowered RLC electrical circuit, or for the position of a freely-oscillating
frictional mass on a spring, or for a damped pendulum. Our solution method
finds two linearly independent solutions to (3.9), multiplies each of these solu-
tions by a constant, and adds them. The two free constants can then be used
to satisfy two given initial conditions.
Because of the differential properties of the exponential function, a natural
ansatz, or educated guess, for the form of the solution to (3.9) is 𝑥 = 𝑒𝑟𝑡 , where
𝑟 is a constant to be determined. Successive differentiation results in 𝑥˙ = 𝑟𝑒𝑟𝑡
and 𝑥 ¨ = 𝑟2 𝑒𝑟𝑡 , and substitution into (3.9) yields
𝑎𝑟2 𝑒𝑟𝑡 + 𝑏𝑟𝑒𝑟𝑡 + 𝑐𝑒𝑟𝑡 = 0. (3.10)
Our choice of exponential function is now rewarded by the explicit cancelation
in (3.10) of 𝑒𝑟𝑡 . The result is a quadratic equation for the unknown constant 𝑟:
𝑎𝑟2 + 𝑏𝑟 + 𝑐 = 0. (3.11)
Our ansatz has thus converted a differential equation into an algebraic equation.
Equation (3.11) is called the characteristic equation of (3.9). Using the quadratic
formula, the two solutions of the characteristic equation (3.11) are given by
1 (︁ √︀ )︁
𝑟± = −𝑏 ± 𝑏2 − 4𝑎𝑐 .
2𝑎
There are three cases to consider: (1) if 𝑏2 − 4𝑎𝑐 > 0, then the two roots are
distinct and real; (2) if 𝑏2 −4𝑎𝑐 < 0, then the two roots are distinct and complex
conjugates of each other; (3) if 𝑏2 − 4𝑎𝑐 = 0, then the two roots are degenerate
and there is only one real root. We will consider these three cases in turn.
3.4.1 Real, distinct roots

When 𝑟+ ̸= 𝑟− are real roots, then the general solution to (3.9) can be written
as a linear superposition of the two solutions 𝑒𝑟+ 𝑡 and 𝑒𝑟− 𝑡 ; that is,
𝑥(𝑡) = 𝑐1 𝑒𝑟+ 𝑡 + 𝑐2 𝑒𝑟− 𝑡 .
The unknown constants 𝑐1 and 𝑐2 can then be determined by the given initial
conditions 𝑥(𝑡0 ) = 𝑥0 and 𝑥(𝑡
˙ 0 ) = 𝑢0 . We now present two examples.
Example 1: Solve 𝑥
¨ + 5𝑥˙ + 6𝑥 = 0 with 𝑥(0) = 2, 𝑥(0)
˙ = 3, and find the
maximum value attained by 𝑥.
view tutorial
We take as our ansatz 𝑥 = 𝑒𝑟𝑡 and obtain the characteristic equation
𝑟2 + 5𝑟 + 6 = 0,
which factors to
(𝑟 + 3)(𝑟 + 2) = 0.
The general solution to the ode is thus
𝑥(𝑡) = 𝑐1 𝑒−2𝑡 + 𝑐2 𝑒−3𝑡 .
The solution for 𝑥˙ obtained by differentiation is
𝑥(𝑡)
˙ = −2𝑐1 𝑒−2𝑡 − 3𝑐2 𝑒−3𝑡 .
Use of the initial conditions then results in two equations for the two unknown
constant 𝑐1 and 𝑐2 :
𝑐1 + 𝑐2 = 2,
−2𝑐1 − 3𝑐2 = 3.
Adding three times the first equation to the second equation yields 𝑐1 = 9; and
the first equation then yields 𝑐2 = 2 − 𝑐1 = −7. Therefore, the unique solution
that satisfies both the ode and the initial conditions is
𝑥(𝑡) = 9𝑒−2𝑡 − 7𝑒−3𝑡

(︂ )︂
7
= 9𝑒−2𝑡 1 − 𝑒−𝑡 .
9
Note that although both exponential terms decay in time, their sum increases
initially since 𝑥(0)
˙ > 0. The maximum value of 𝑥 occurs at the time 𝑡𝑚 when
𝑥˙ = 0, or
𝑡𝑚 = ln (7/6) .
The maximum 𝑥𝑚 = 𝑥(𝑡𝑚 ) is then determined to be
𝑥𝑚 = 108/49.
¨ − 𝑥 = 0 with 𝑥(0) = 𝑥0 , 𝑥(0)

Example 2: Solve 𝑥 ˙ = 𝑢0 .
Again our ansatz is 𝑥 = 𝑒𝑟𝑡 , and we obtain the characteristic equation
𝑟2 − 1 = 0,
with solution 𝑟± = ±1. Therefore, the general solution for 𝑥 is
𝑥(𝑡) = 𝑐1 𝑒𝑡 + 𝑐2 𝑒−𝑡 ,
and the derivative satisfies
𝑥(𝑡)
˙ = 𝑐1 𝑒𝑡 − 𝑐2 𝑒−𝑡 .
Initial conditions are satisfied when
𝑐1 + 𝑐2 = 𝑥0 ,
𝑐1 − 𝑐2 = 𝑢0 .
Adding and subtracting these equations, we determine
1 1
𝑐1 = (𝑥0 + 𝑢0 ) , 𝑐2 = (𝑥0 − 𝑢0 ) ,
2 2
so that after rearranging terms
𝑒 + 𝑒−𝑡 𝑒 − 𝑒−𝑡
(︂ 𝑡 )︂ (︂ 𝑡 )︂
𝑥(𝑡) = 𝑥0 + 𝑢0 .
2 2
The terms in parentheses are the usual definitions of the hyperbolic cosine and
sine functions; that is,
𝑒𝑡 + 𝑒−𝑡 𝑒𝑡 − 𝑒−𝑡
cosh 𝑡 = , sinh 𝑡 = .
2 2
Our solution can therefore be rewritten as
𝑥(𝑡) = 𝑥0 cosh 𝑡 + 𝑢0 sinh 𝑡.
Note that the relationships between the trigonometric functions and the complex
exponentials were given by
𝑒𝑖𝑡 + 𝑒−𝑖𝑡 𝑒𝑖𝑡 − 𝑒−𝑖𝑡
cos 𝑡 = , sin 𝑡 = ,
2 2𝑖
so that
cosh 𝑡 = cos 𝑖𝑡, sinh 𝑡 = −𝑖 sin 𝑖𝑡.
Also note that the hyperbolic trigonometric functions satisfy the differential
equations
𝑑 𝑑
sinh 𝑡 = cosh 𝑡, cosh 𝑡 = sinh 𝑡,
𝑑𝑡 𝑑𝑡
which though similar to the differential equations satisfied by the more com-
monly used trigonometric functions, is absent a minus sign.
3.4.2 Complex conjugate, distinct roots

view tutorial
We now consider a characteristic equation (3.11) with 𝑏2 − 4𝑎𝑐 < 0, so the roots
occur as complex conjugate pairs. With
𝑏 1 √︀
𝜆=− , 𝜇= 4𝑎𝑐 − 𝑏2 ,
2𝑎 2𝑎
the two roots of the characteristic equation are 𝜆 + 𝑖𝜇 and 𝜆 − 𝑖𝜇. We have
thus found the following two complex exponential solutions to the differential
equation:
𝑍1 (𝑡) = 𝑒𝜆𝑡 𝑒𝑖𝜇𝑡 , 𝑍2 (𝑡) = 𝑒𝜆𝑡 𝑒−𝑖𝜇𝑡 .
Applying the principle of superposition, any linear combination of 𝑍1 and 𝑍2 is
also a solution to the second-order ode. We can then form two different linear
combinations that are real, given by
𝑍1 + 𝑍2
𝑋1 (𝑡) =
2
𝑒𝑖𝜇𝑡 + 𝑒−𝑖𝜇𝑡
(︂ )︂
𝜆𝑡
=𝑒
2
= 𝑒𝜆𝑡 cos 𝜇𝑡,
and
𝑍1 − 𝑍2
𝑋2 (𝑡) =
2𝑖
𝑒 − 𝑒−𝑖𝜇𝑡
(︂ 𝑖𝜇𝑡 )︂
𝜆𝑡
=𝑒
2𝑖
= 𝑒𝜆𝑡 sin 𝜇𝑡.
Having found the two real solutions 𝑋1 (𝑡) and 𝑋2 (𝑡), we can then apply the
principle of superposition a second time to determine the general solution 𝑥(𝑡):
𝑥(𝑡) = 𝑒𝜆𝑡 (𝐴 cos 𝜇𝑡 + 𝐵 sin 𝜇𝑡) . (3.12)
It is best to memorize this result. The real part of the roots of the characteristic
equation goes into the exponential function; the imaginary part goes into the
argument of cosine and sine.
¨ + 𝑥 = 0 with 𝑥(0) = 𝑥0 and 𝑥(0)
˙ = 𝑢0 .
view tutorial
The characteristic equation is
𝑟2 + 1 = 0,
with roots
𝑟± = ±𝑖.
The general solution of the ode is therefore
𝑥(𝑡) = 𝐴 cos 𝑡 + 𝐵 sin 𝑡.

The derivative is
𝑥(𝑡)
˙ = −𝐴 sin 𝑡 + 𝐵 cos 𝑡.
Applying the initial conditions:
𝑥(0) = 𝐴 = 𝑥0 , 𝑥(0)
˙ = 𝐵 = 𝑢0 ;
so that the final solution is
𝑥(𝑡) = 𝑥0 cos 𝑡 + 𝑢0 sin 𝑡.
¨ − 𝑥 = 0 as 𝑥(𝑡) =
Recall that we wrote the analogous solution to the ode 𝑥
𝑥0 cosh 𝑡 + 𝑢0 sinh 𝑡.
¨ + 𝑥˙ + 𝑥 = 0 with 𝑥(0) = 1 and 𝑥(0)
˙ = 0.
𝑟2 + 𝑟 + 1 = 0,
with roots √
1 3
𝑟± = − ± 𝑖 .
2 2
(︃ √ √ )︃
− 12 𝑡 3 3
𝑥(𝑡) = 𝑒 𝐴 cos 𝑡 + 𝐵 sin 𝑡 .
2 2
The derivative is
(︃ √ √ )︃
1 1 3 3
𝑥(𝑡)
˙ = − 𝑒− 2 𝑡 𝐴 cos 𝑡 + 𝐵 sin 𝑡
2 2 2
√ (︃ √ √ )︃
3 −1𝑡 3 3
+ 𝑒 2 −𝐴 sin 𝑡 + 𝐵 cos 𝑡 .
2 2 2
Applying the initial conditions 𝑥(0) = 1 and 𝑥(0)

˙ = 0:
𝐴 = 1,
√
−1 3
𝐴+ 𝐵 = 0;
2 2
or √
3
𝐴 = 1, 𝐵= .
3
Therefore,
(︃ √ √ √ )︃
− 21 𝑡 3 3 3
𝑥(𝑡) = 𝑒 cos 𝑡+ sin 𝑡 .
2 3 2
3.4.3 Repeated roots

view tutorial
Finally, we consider the characteristic equation,
𝑎𝑟2 + 𝑏𝑟 + 𝑐 = 0,
with 𝑏2 − 4𝑎𝑐 = 0. The degenerate root is then given by
𝑏
𝑟=− ,
2𝑎
yielding only a single solution to the ode:
(︂ )︂
𝑏𝑡
𝑥1 (𝑡) = exp − . (3.13)
2𝑎
To satisfy two initial conditions, a second independent solution must be found
with nonzero Wronskian, and apparently this second solution is not of the form
of our ansatz 𝑥 = exp (𝑟𝑡).
One method to determine this missing second solution is to try the ansatz
𝑥(𝑡) = 𝑦(𝑡)𝑥1 (𝑡), (3.14)
where 𝑦(𝑡) is an unknown function that satisfies the differential equation ob-
tained by substituting (3.14) into (3.9). This standard technique is called the
reduction of order method and enables one to find a second solution of a ho-
mogeneous linear differential equation if one solution is known. If the original
differential equation is of order 𝑛, the differential equation for 𝑦 = 𝑦(𝑡) reduces
to an order one lower, that is, 𝑛 − 1.
Here, however, we will determine this missing second solution through a
limiting process. We start with the solution obtained for complex roots of the
characteristic equation, and then arrive at the solution obtained for degenerate
roots by taking the limit 𝜇 → 0.
Now, the general solution for complex roots was given by (3.12), and to
properly limit this solution as 𝜇 → 0 requires first satisfying the specific initial
conditions 𝑥(0) = 𝑥0 and 𝑥(0)˙ = 𝑢0 . Solving for 𝐴 and 𝐵, the general solution
given by (3.12) becomes the specific solution
(︂ )︂
𝑢0 − 𝜆𝑥0
𝑥(𝑡; 𝜇) = 𝑒𝜆𝑡 𝑥0 cos 𝜇𝑡 + sin 𝜇𝑡 .
𝜇
Here, we have written 𝑥 = 𝑥(𝑡; 𝜇) to show explicitly that 𝑥 depends on 𝜇.
Taking the limit as 𝜇 → 0, and using lim𝜇→0 𝜇−1 sin 𝜇𝑡 = 𝑡, we have
lim 𝑥(𝑡; 𝜇) = 𝑒𝜆𝑡 𝑥0 + (𝑢0 − 𝜆𝑥0 )𝑡 .
(︀ )︀
𝜇→0
The second solution is observed to be a constant, 𝑢0 − 𝜆𝑥0 , times 𝑡 times the

first solution, 𝑒𝜆𝑡 . Our general solution to the ode (3.9) when 𝑏2 − 4𝑎𝑐 = 0 can
therefore be written in the form
𝑥(𝑡) = (𝑐1 + 𝑐2 𝑡)𝑒𝑟𝑡 ,
where 𝑟 is the repeated root of the characteristic equation. The main result to
be remembered is that for the case of repeated roots, the second solution is 𝑡
times the first solution.
3.5. INHOMOGENEOUS ODES 37
Example: Solve 𝑥
¨ + 2𝑥˙ + 𝑥 = 0 with 𝑥(0) = 1 and 𝑥(0)
˙ = 0.
𝑟2 + 2𝑟 + 1 = (𝑟 + 1)2
= 0,
which has a repeated root given by 𝑟 = −1. Therefore, the general solution to
the ode is
𝑥(𝑡) = 𝑐1 𝑒−𝑡 + 𝑐2 𝑡𝑒−𝑡 ,
with derivative
𝑥(𝑡)
˙ = −𝑐1 𝑒−𝑡 + 𝑐2 𝑒−𝑡 − 𝑐2 𝑡𝑒−𝑡 .
Applying the initial conditions, we have
𝑐1 = 1,
−𝑐1 + 𝑐2 = 0,
so that 𝑐1 = 𝑐2 = 1. Therefore, the solution is
𝑥(𝑡) = (1 + 𝑡)𝑒−𝑡 .
3.5 Second-order linear inhomogeneous ode

We now consider the general second-order linear inhomogeneous ode (3.1):
𝑥
¨ + 𝑝(𝑡)𝑥˙ + 𝑞(𝑡)𝑥 = 𝑔(𝑡), (3.15)
with initial conditions 𝑥(𝑡0 ) = 𝑥0 and 𝑥(𝑡
˙ 0 ) = 𝑢0 . There is a three-step solution
method when the inhomogeneous term 𝑔(𝑡) ̸= 0. (i) Find the general solution
of the homogeneous equation
𝑥
¨ + 𝑝(𝑡)𝑥˙ + 𝑞(𝑡)𝑥 = 0. (3.16)
Let us denote the homogeneous solution by
𝑥ℎ (𝑡) = 𝑐1 𝑋1 (𝑡) + 𝑐2 𝑋2 (𝑡),
where 𝑋1 and 𝑋2 are linearly independent solutions of (3.16), and 𝑐1 and 𝑐2
are as yet undetermined constants. (ii) Find any particular solution 𝑥𝑝 of the
inhomogeneous equation (3.15). A particular solution is readily found when
𝑝(𝑡) and 𝑞(𝑡) are constants, and when 𝑔(𝑡) is a combination of polynomials,
exponentials, sines and cosines. (iii) Write the general solution of (3.15) as the
sum of the homogeneous and particular solutions,
𝑥(𝑡) = 𝑥ℎ (𝑡) + 𝑥𝑝 (𝑡), (3.17)
and apply the initial conditions to determine the constants 𝑐1 and 𝑐2 . Note that
because of the linearity of (3.15),
𝑑2 𝑑
𝑥
¨ + 𝑝𝑥˙ + 𝑞𝑥 = (𝑥ℎ + 𝑥𝑝 ) + 𝑝 (𝑥ℎ + 𝑥𝑝 ) + 𝑞(𝑥ℎ + 𝑥𝑝 )
𝑑𝑡2 𝑑𝑡
= (¨
𝑥ℎ + 𝑝𝑥˙ ℎ + 𝑞𝑥ℎ ) + (¨
𝑥𝑝 + 𝑝𝑥˙ 𝑝 + 𝑞𝑥𝑝 )
=0+𝑔
= 𝑔,
so that (3.17) solves (3.15), and the two free constants in 𝑥ℎ can be used to
satisfy the initial conditions.
We will consider here only the constant coefficient case. We now illustrate
the solution method by an example.
¨ − 3𝑥˙ − 4𝑥 = 3𝑒2𝑡 with 𝑥(0) = 1 and 𝑥(0)

Example: Solve 𝑥 ˙ = 0.
view tutorial
First, we solve the homogeneous equation. The characteristic equation is
𝑟2 − 3𝑟 − 4 = (𝑟 − 4)(𝑟 + 1)
= 0,
so that
𝑥ℎ (𝑡) = 𝑐1 𝑒4𝑡 + 𝑐2 𝑒−𝑡 .
Second, we find a particular solution of the inhomogeneous equation. The form
of the particular solution is chosen such that the exponential will cancel out of
both sides of the ode. The ansatz we choose is
𝑥(𝑡) = 𝐴𝑒2𝑡 , (3.18)
where 𝐴 is a yet undetermined coefficient. Upon substituting 𝑥 into the ode,

differentiating using the chain rule, and canceling the exponential, we obtain
4𝐴 − 6𝐴 − 4𝐴 = 3,
from which we determine 𝐴 = −1/2. Obtaining a solution for 𝐴 independent of

𝑡 justifies the ansatz (3.18). Third, we write the general solution to the ode as
the sum of the homogeneous and particular solutions, and determine 𝑐1 and 𝑐2
that satisfy the initial conditions. We have
1
𝑥(𝑡) = 𝑐1 𝑒4𝑡 + 𝑐2 𝑒−𝑡 − 𝑒2𝑡 ;
2
and taking the derivative,
𝑥(𝑡)
˙ = 4𝑐1 𝑒4𝑡 − 𝑐2 𝑒−𝑡 − 𝑒2𝑡 .
Applying the initial conditions,

1
𝑐1 + 𝑐2 − = 1,
2
4𝑐1 − 𝑐2 − 1 = 0;
or
3
𝑐1 + 𝑐2 = ,
2
4𝑐1 − 𝑐2 = 1.
This system of linear equations can be solved for 𝑐1 by adding the equations
to obtain 𝑐1 = 1/2, after which 𝑐2 = 1 can be determined from the first equa-
tion. Therefore, the solution for 𝑥(𝑡) that satisfies both the ode and the initial
3.5. INHOMOGENEOUS ODES 39
conditions is given by
1 4𝑡 1 2𝑡
𝑥(𝑡) = 𝑒 − 𝑒 + 𝑒−𝑡
2 2
1 4𝑡 (︀
= 𝑒 1 − 𝑒−2𝑡 + 2𝑒−5𝑡 ,
)︀
2
where we have grouped the terms in the solution to better display the asymptotic
behavior for large 𝑡.
We now find particular solutions for some relatively simple inhomogeneous
terms using this method of undetermined coefficients.
¨ − 3𝑥˙ − 4𝑥 = 2 sin 𝑡.
Example: Find a particular solution of 𝑥
view tutorial
We show two methods for finding a particular solution. The first more direct
method tries the ansatz
𝑥(𝑡) = 𝐴 cos 𝑡 + 𝐵 sin 𝑡,
where the argument of cosine and sine must agree with the argument of sine in
the inhomogeneous term. The cosine term is required because the derivative of
sine is cosine. Upon substitution into the differential equation, we obtain
(−𝐴 cos 𝑡 − 𝐵 sin 𝑡) − 3 (−𝐴 sin 𝑡 + 𝐵 cos 𝑡) − 4 (𝐴 cos 𝑡 + 𝐵 sin 𝑡) = 2 sin 𝑡,
or regrouping terms,
− (5𝐴 + 3𝐵) cos 𝑡 + (3𝐴 − 5𝐵) sin 𝑡 = 2 sin 𝑡.
This equation is valid for all 𝑡, and in particular for 𝑡 = 0 and 𝜋/2, for which
the sine and cosine functions vanish. For these two values of 𝑡, we find
5𝐴 + 3𝐵 = 0, 3𝐴 − 5𝐵 = 2;
and solving for 𝐴 and 𝐵, we obtain

3 5
𝐴= , 𝐵=− .
17 17
The particular solution is therefore given by
1
𝑥𝑝 = (3 cos 𝑡 − 5 sin 𝑡) .
17
The second solution method makes use of the relation 𝑒𝑖𝑡 = cos 𝑡 + 𝑖 sin 𝑡 to
convert the sine inhomogeneous term to an exponential function. We introduce
the complex function 𝑧(𝑡) by letting
𝑧(𝑡) = 𝑥(𝑡) + 𝑖𝑦(𝑡),
and rewrite the differential equation in complex form. We can rewrite the equa-
tion in one of two ways. On the one hand, if we use sin 𝑡 = Re{−𝑖𝑒𝑖𝑡 }, then the
differential equation is written as
𝑧¨ − 3𝑧˙ − 4𝑧 = −2𝑖𝑒𝑖𝑡 ; (3.19)

and by equating the real and imaginary parts, this equation becomes the two
real differential equations
¨ − 3𝑥˙ − 4𝑥 = 2 sin 𝑡,
𝑥 𝑦¨ − 3𝑦˙ − 4𝑦 = −2 cos 𝑡.
The solution we are looking for, then, is 𝑥𝑝 (𝑡) = Re{𝑧𝑝 (𝑡)}.

On the other hand, if we write sin 𝑡 = Im{𝑒𝑖𝑡 }, then the complex differential
equation becomes
𝑧¨ − 3𝑧˙ − 4𝑧 = 2𝑒𝑖𝑡 , (3.20)
which becomes the two real differential equations
¨ − 3𝑥˙ − 4𝑥 = 2 cos 𝑡,
𝑥 𝑦¨ − 3𝑦˙ − 4𝑦 = 2 sin 𝑡.
Here, the solution we are looking for is 𝑥𝑝 (𝑡) = Im{𝑧𝑝 (𝑡)}.

We will proceed here by solving (3.20). As we now have an exponential
function as the inhomogeneous term, we can make the ansatz
𝑧(𝑡) = 𝐶𝑒𝑖𝑡 ,
where we now expect 𝐶 to be a complex constant. Upon substitution into the

ode (3.20) and using 𝑖2 = −1:
−𝐶 − 3𝑖𝐶 − 4𝐶 = 2;
or solving for 𝐶:
−2
𝐶=
5 + 3𝑖
−2(5 − 3𝑖)
=
(5 + 3𝑖)(5 − 3𝑖)
−10 + 6𝑖
=
34
−5 + 3𝑖
= .
17
Therefore,
𝑥𝑝 = Im{𝑧𝑝 }
{︂ }︂
1
= Im (−5 + 3𝑖)(cos 𝑡 + 𝑖 sin 𝑡)
17
1
= (3 cos 𝑡 − 5 sin 𝑡).
17
¨ + 𝑥˙ − 2𝑥 = 𝑡2 .
view tutorial
The correct ansatz here is the polynomial
𝑥(𝑡) = 𝐴𝑡2 + 𝐵𝑡 + 𝐶.
Upon substitution into the ode, we have
2𝐴 + 2𝐴𝑡 + 𝐵 − 2𝐴𝑡2 − 2𝐵𝑡 − 2𝐶 = 𝑡2 ,

3.6. FIRST-ORDER LINEAR INHOMOGENEOUS ODES REVISITED 41
or
−2𝐴𝑡2 + 2(𝐴 − 𝐵)𝑡 + (2𝐴 + 𝐵 − 2𝐶)𝑡0 = 𝑡2 .
Equating powers of 𝑡,
−2𝐴 = 1, 2(𝐴 − 𝐵) = 0, 2𝐴 + 𝐵 − 2𝐶 = 0;
and solving,
1 1 3
𝐴=− , 𝐵=− , 𝐶=− .
2 2 4
The particular solution is therefore
1 1 3
𝑥𝑝 (𝑡) = − 𝑡2 − 𝑡 − .
2 2 4
3.6 First-order linear inhomogeneous odes re-

visited
The first-order linear ode can be solved by use of an integrating factor. Here I
show that odes having constant coefficients can be solved by our newly learned
solution method.
Example: Solve 𝑥˙ + 2𝑥 = 𝑒−𝑡 with 𝑥(0) = 3/4.

Rather than using an integrating factor, we follow the three-step approach: (i)
find the general homogeneous solution; (ii) find a particular solution; (iii) add
them and satisfy initial conditions. Accordingly, we try the ansatz 𝑥ℎ (𝑡) = 𝑒𝑟𝑡
for the homogeneous ode 𝑥˙ + 2𝑥 = 0 and find
𝑟 + 2 = 0, or 𝑟 = −2.
To find a particular solution, we try the ansatz 𝑥𝑝 (𝑡) = 𝐴𝑒−𝑡 , and upon substi-
tution
−𝐴 + 2𝐴 = 1, or 𝐴 = 1.
Therefore, the general solution to the ode is
𝑥(𝑡) = 𝑐𝑒−2𝑡 + 𝑒−𝑡 .
The single initial condition determines the unknown constant 𝑐:
3
𝑥(0) = = 𝑐 + 1,
4
so that 𝑐 = −1/4. Hence,
1
𝑥(𝑡) = 𝑒−𝑡 − 𝑒−2𝑡
(︂ 4 )︂
−𝑡 1
=𝑒 1 − 𝑒−𝑡 .
4
3.7 Resonance
view tutorial
Resonance occurs when the frequency of the inhomogeneous term matches the
frequency of the homogeneous solution. To illustrate resonance in its simplest
embodiment, we consider the second-order linear inhomogeneous ode
¨ + 𝜔02 𝑥 = 𝑓 cos 𝜔𝑡,

𝑥 𝑥(0) = 𝑥0 , 𝑥(0)
˙ = 𝑢0 . (3.21)
Our main goal is to determine what happens to the solution in the limit 𝜔 → 𝜔0 .
The homogeneous equation has characteristic equation
𝑟2 + 𝜔02 = 0,
so that 𝑟± = ±𝑖𝜔0 . Therefore,
𝑥ℎ (𝑡) = 𝑐1 cos 𝜔0 𝑡 + 𝑐2 sin 𝜔0 𝑡. (3.22)
To find a particular solution, we note the absence of a first-derivative term,

and simply try
𝑥(𝑡) = 𝐴 cos 𝜔𝑡.
Upon substitution into the ode, we obtain
−𝜔 2 𝐴 + 𝜔02 𝐴 = 𝑓,
or
𝑓
𝐴= .
𝜔02 − 𝜔 2
Therefore,
𝑓
𝑥𝑝 (𝑡) = cos 𝜔𝑡.
𝜔02 − 𝜔2
Our general solution is thus
𝑓
𝑥(𝑡) = 𝑐1 cos 𝜔0 𝑡 + 𝑐2 sin 𝜔0 𝑡 + cos 𝜔𝑡,
𝜔02 − 𝜔 2
with derivative
𝑓𝜔
𝑥(𝑡)
˙ = 𝜔0 (𝑐2 cos 𝜔0 𝑡 − 𝑐1 sin 𝜔0 𝑡) − sin 𝜔𝑡.
𝜔02 − 𝜔 2
Initial conditions are satisfied when

𝑓
𝑥0 = 𝑐1 + ,
𝜔02 − 𝜔 2
𝑢0 = 𝑐2 𝜔0 ,
so that
𝑓 𝑢0
𝑐1 = 𝑥0 − , 𝑐2 = .
𝜔02 − 𝜔 2 𝜔0
3.7. RESONANCE 43
Therefore, the solution to the ode that satisfies the initial conditions is
(︂ )︂
𝑓 𝑢0 𝑓
𝑥(𝑡) = 𝑥0 − 2 2
cos 𝜔0 𝑡 + sin 𝜔0 𝑡 + 2 cos 𝜔𝑡
𝜔0 − 𝜔 𝜔0 𝜔0 − 𝜔 2
𝑢0 𝑓 (cos 𝜔𝑡 − cos 𝜔0 𝑡)
= 𝑥0 cos 𝜔0 𝑡 + sin 𝜔0 𝑡 + ,
𝜔0 𝜔02 − 𝜔 2
where we have grouped together terms proportional to the forcing amplitude 𝑓 .

Resonance occurs in the limit 𝜔 → 𝜔0 ; that is, the frequency of the inhomo-
geneous term (the external force) matches the frequency of the homogeneous
solution (the free oscillation). By L’Hospital’s rule, the limit of the term pro-
portional to 𝑓 is found by differentiating with respect to 𝜔:
𝑓 (cos 𝜔𝑡 − cos 𝜔0 𝑡) −𝑓 𝑡 sin 𝜔𝑡

lim 2 2
= lim
𝜔→𝜔0 𝜔0 − 𝜔 𝜔→𝜔0 −2𝜔
(3.23)
𝑓 𝑡 sin 𝜔0 𝑡
= .
2𝜔0
At resonance, the term proportional to the amplitude 𝑓 of the inhomogeneous

term increases linearly with 𝑡, resulting in larger-and-larger amplitudes of oscil-
lation for 𝑥(𝑡). In general, if the inhomogeneous term in the differential equation
is a solution of the corresponding homogeneous differential equation, then the
correct ansatz for the particular solution is a constant times the inhomogeneous
term times 𝑡.
To illustrate this same example further, we return to the original ode, now
assumed to be exactly at resonance,
¨ + 𝜔02 𝑥 = 𝑓 cos 𝜔0 𝑡,
𝑥
and find a particular solution directly. The particular solution is the real part
of the particular solution of
𝑧¨ + 𝜔02 𝑧 = 𝑓 𝑒𝑖𝜔0 𝑡 ,
and because the inhomogeneous term is a solution of the corresponding homo-

geneous equation, we take as our ansatz
𝑧𝑝 = 𝐴𝑡𝑒𝑖𝜔0 𝑡 .
We have
𝑧˙𝑝 = 𝐴𝑒𝑖𝜔0 𝑡 (1 + 𝑖𝜔0 𝑡) , 𝑧¨𝑝 = 𝐴𝑒𝑖𝜔0 𝑡 2𝑖𝜔0 − 𝜔02 𝑡 ;
(︀ )︀
and upon substitution into the ode
𝑧¨𝑝 + 𝜔02 𝑧𝑝 = 𝐴𝑒𝑖𝜔0 𝑡 2𝑖𝜔0 − 𝜔02 𝑡 + 𝜔02 𝐴𝑡𝑒𝑖𝜔0 𝑡

(︀ )︀
= 2𝑖𝜔0 𝐴𝑒𝑖𝜔0 𝑡
= 𝑓 𝑒𝑖𝜔0 𝑡 .
Therefore,
𝑓
𝐴= ,
2𝑖𝜔0
and
𝑓 𝑡 𝑖𝜔0 𝑡
𝑥𝑝 = Re{ 𝑒 }
2𝑖𝜔0
𝑓 𝑡 sin 𝜔0 𝑡
= ,
2𝜔0
the same result as (3.23).
¨ − 3𝑥˙ − 4𝑥 = 5𝑒−𝑡 .
view tutorial
If we naively try the ansatz
𝑥 = 𝐴𝑒−𝑡 ,
and substitute this into the inhomogeneous differential equation, we obtain
𝐴 + 3𝐴 − 4𝐴 = 5,
or 0 = 5, which is clearly nonsense. Our ansatz therefore fails to find a solution.

The cause of this failure is that the corresponding homogeneous equation has
solution
𝑥ℎ = 𝑐1 𝑒4𝑡 + 𝑐2 𝑒−𝑡 ,
so that the inhomogeneous term 5𝑒−𝑡 is one of the solutions of the homogeneous
equation. To find a particular solution, we should therefore take as our ansatz
𝑥 = 𝐴𝑡𝑒−𝑡 ,
with first- and second-derivatives given by
𝑥˙ = 𝐴𝑒−𝑡 (1 − 𝑡), ¨ = 𝐴𝑒−𝑡 (−2 + 𝑡).

𝑥
Substitution into the differential equation yields
𝐴𝑒−𝑡 (−2 + 𝑡) − 3𝐴𝑒−𝑡 (1 − 𝑡) − 4𝐴𝑡𝑒−𝑡 = 5𝑒−𝑡 .
The terms containing 𝑡 cancel out of this equation, resulting in −5𝐴 = 5, or

𝐴 = −1. Therefore, the particular solution is
𝑥𝑝 = −𝑡𝑒−𝑡 .
3.8 Damped resonance

view tutorial
A more realistic study of resonance assumes an additional damping term. The
forced, damped harmonic oscillator equation may be written as
𝑚¨
𝑥 + 𝛾 𝑥˙ + 𝑘𝑥 = 𝐹 cos 𝜔𝑡, (3.24)
where 𝑚 > 0 is the oscillator’s mass, 𝛾 > 0 is the damping coefficient, 𝑘 >
0 is the spring constant, and 𝐹 is the amplitude of the external force. The
homogeneous equation has characteristic equation
𝑚𝑟2 + 𝛾𝑟 + 𝑘 = 0,
3.8. DAMPED RESONANCE 45
so that
𝛾 1 √︀ 2
𝑟± = − ± 𝛾 − 4𝑚𝑘.
2𝑚 2𝑚
When 𝛾 2 − 4𝑚𝑘 < 0, the motion of the unforced oscillator is said to be under-
damped; when 𝛾 2 − 4𝑚𝑘 > 0, overdamped; and when 𝛾 2 − 4𝑚𝑘 = 0, critically
damped. For all three types of damping, the roots of the characteristic equa-
tion satisfy Re(𝑟± ) < 0. Therefore, both linearly independent homogeneous
solutions decay exponentially to zero, and the long-time asymptotic solution
of (3.24) reduces to the (non-decaying) particular solution. Since the initial
conditions are satisfied by the free constants multiplying the (decaying) homo-
geneous solutions, the long-time asymptotic solution is independent of the initial
conditions.
If we are only interested in the long-time asymptotic solution of (3.24), we
can proceed directly to the determination of a particular solution. As before,
we consider the complex ode
𝑧 + 𝛾 𝑧˙ + 𝑘𝑧 = 𝐹 𝑒𝑖𝜔𝑡 ,
𝑚¨
with 𝑥𝑝 = Re(𝑧𝑝 ). With the ansatz 𝑧𝑝 = 𝐴𝑒𝑖𝜔𝑡 , we have
−𝑚𝜔 2 𝐴 + 𝑖𝛾𝜔𝐴 + 𝑘𝐴 = 𝐹,
or
𝐹
𝐴= .
(𝑘 − 𝑚𝜔 2 ) + 𝑖𝛾𝜔
√︀
To simplify, we define 𝜔0 = 𝑘/𝑚, which corresponds to the natural frequency
of the undamped oscillator, and define Γ = 𝛾𝜔/𝑚 and 𝑓 = 𝐹/𝑚. Therefore,
𝑓
𝐴=
(𝜔02 − 𝜔 2 ) + 𝑖Γ
(︂ )︂ (3.25)
𝑓 (︀ 2
(𝜔0 − 𝜔 2 ) − 𝑖Γ .
)︀
= 2 2 2 2
(𝜔0 − 𝜔 ) + Γ
To determine 𝑥𝑝 , we utilize the polar form of a complex number. The complex
𝑖𝜑
number 𝑧 = 𝑥 + 𝑖𝑦 can
√︀ be written in polar form as 𝑧 = 𝑟𝑒 , where 𝑥 = 𝑟 cos 𝜑,
𝑦 = 𝑟 sin 𝜑, and 𝑟 = 𝑥2 + 𝑦 2 , tan 𝜑 = 𝑦/𝑥. We therefore write
(𝜔02 − 𝜔 2 ) − 𝑖Γ = 𝑟𝑒𝑖𝜑 ,
with
Γ
√︁
𝑟= (𝜔02 − 𝜔 2 )2 + Γ2 , tan 𝜑 = .
(𝜔 2 − 𝜔02 )
Using the polar form, 𝐴 in (3.25) becomes
(︃ )︃
𝑓
𝐴 = √︀ 2 𝑒𝑖𝜑 ,
(𝜔0 − 𝜔 2 )2 + Γ2
and 𝑥𝑝 = Re(𝐴𝑒𝑖𝜔𝑡 ) becomes
(︃ )︃
𝑓 (︁ )︁
𝑥𝑝 = √︀ Re 𝑒𝑖(𝜔𝑡+𝜑)
(𝜔02 − 𝜔 2 )2 + Γ2
(︃ )︃ (3.26)
𝑓
= √︀ cos (𝜔𝑡 + 𝜑).
(𝜔02 − 𝜔 2 )2 + Γ2
The particular solution given by (3.26), with 𝜔02 = 𝑘/𝑚, Γ = 𝛾𝜔/𝑚, 𝑓 = 𝐹/𝑚,
and tan 𝜑 = 𝛾𝜔/𝑚(𝜔 2 − 𝜔02 ), is the long-time asymptotic solution of the forced,
damped, harmonic oscillator equation (3.24).
We conclude with a couple of observations about (3.26). First, if the forcing
frequency 𝜔 is equal to the natural frequency 𝜔0 of the undamped oscillator, then
𝐴 = −𝑖𝐹/𝛾𝜔0 , and 𝑥𝑝 = (𝐹/𝛾𝜔0 ) sin 𝜔0 𝑡. The oscillator position is observed
to be 𝜋/2 out of phase with the external force, or in other words, the velocity
of the oscillator, not the position, is in phase with the force. Second, the value
of the forcing frequency 𝜔𝑚 that maximizes the amplitude of oscillation is the
value of 𝜔 that minimizes the denominator of (3.26). To determine 𝜔𝑚 we thus
minimize the function 𝑔(𝜔 2 ) with respect to 𝜔 2 , where
𝛾 2 𝜔2
𝑔(𝜔 2 ) = (𝜔02 − 𝜔 2 )2 + .
𝑚2
Taking the derivative of 𝑔 with respect to 𝜔 2 and setting this to zero to determine
𝜔𝑚 yields
2 𝛾2
2(𝜔𝑚 − 𝜔02 ) + 2 = 0,
𝑚
or
2 𝛾2
𝜔𝑚 = 𝜔02 − .
2𝑚2
We can interpret this result by saying that damping lowers the “resonance”
frequency of the undamped oscillator.
Chapter 4
The Laplace transform

The Laplace transform is most useful for solving linear, constant-coefficient

ode’s when the inhomogeneous term or its derivative is discontinuous. Although
ode’s with discontinuous inhomogeneous terms can also be solved by adopting
already learned methods, we will see that the Laplace transform technique pro-
vides a simpler, more elegant solution.
4.1 Definitions and properties of the forward

and inverse Laplace transforms
view tutorial
The main idea is to Laplace transform the constant-coefficient differential equa-
tion for 𝑥(𝑡) into a simpler algebraic equation for the Laplace-transformed func-
tion 𝑋(𝑠), solve this algebraic equation, and then transform 𝑋(𝑠) back into
𝑥(𝑡). The correct definition of the Laplace transform and the properties that
this transform satisfies makes this solution method possible.
An exponential ansatz is used in solving homogeneous constant-coefficient
odes, and the exponential function correspondingly plays a key role in defining
the Laplace transform. The Laplace transform of 𝑓 (𝑡), denoted by 𝐹 (𝑠) =
ℒ{𝑓 (𝑡)}, is defined by the integral transform
∫︁ ∞
𝐹 (𝑠) = 𝑒−𝑠𝑡 𝑓 (𝑡)𝑑𝑡. (4.1)
0
The improper integral given by (4.1) diverges if 𝑓 (𝑡) grows faster than 𝑒𝑠𝑡 for
large 𝑡. Accordingly, some restriction on the range of 𝑠 may be required to
guarantee convergence of (4.1), and we will assume without further elaboration
that these restrictions are always satisfied.
The Laplace transform is a linear transformation. We have
∫︁ ∞
𝑒−𝑠𝑡 𝑐1 𝑓1 (𝑡) + 𝑐2 𝑓2 (𝑡) 𝑑𝑡
(︀ )︀
ℒ{𝑐1 𝑓1 (𝑡) + 𝑐2 𝑓2 (𝑡)} =
0
∫︁ ∞ ∫︁ ∞
= 𝑐1 𝑒−𝑠𝑡 𝑓1 (𝑡)𝑑𝑡 + 𝑐2 𝑒−𝑠𝑡 𝑓2 (𝑡)𝑑𝑡
0 0
= 𝑐1 ℒ{𝑓1 (𝑡)} + 𝑐2 ℒ{𝑓2 (𝑡)}.
47
48 CHAPTER 4. THE LAPLACE TRANSFORM
𝑓 (𝑡) = ℒ−1 {𝐹 (𝑠)} 𝐹 (𝑠) = ℒ{𝑓 (𝑡)}
1. 𝑒𝑎𝑡 𝑓 (𝑡) 𝐹 (𝑠 − 𝑎)
1
2. 1
𝑠
1
3. 𝑒𝑎𝑡
𝑠−𝑎
𝑛!
4. 𝑡𝑛
𝑠𝑛+1
𝑛!
5. 𝑡𝑛 𝑒𝑎𝑡
(𝑠 − 𝑎)𝑛+1
𝑏
6. sin 𝑏𝑡
𝑠2 + 𝑏2
𝑠
7. cos 𝑏𝑡
𝑠2 + 𝑏2
𝑏
8. 𝑒𝑎𝑡 sin 𝑏𝑡
(𝑠 − 𝑎)2 + 𝑏2
𝑠−𝑎
9. 𝑒𝑎𝑡 cos 𝑏𝑡
(𝑠 − 𝑎)2 + 𝑏2
2𝑏𝑠
10. 𝑡 sin 𝑏𝑡
(𝑠2 + 𝑏2 )2
𝑠2 − 𝑏2
11. 𝑡 cos 𝑏𝑡
(𝑠2 + 𝑏2 )2
𝑒−𝑐𝑠
12. 𝑢𝑐 (𝑡)
𝑠
13. 𝑢𝑐 (𝑡)𝑓 (𝑡 − 𝑐) 𝑒−𝑐𝑠 𝐹 (𝑠)
14. 𝛿(𝑡 − 𝑐) 𝑒−𝑐𝑠
15. 𝑥(𝑡)
˙ 𝑠𝑋(𝑠) − 𝑥(0)
16. 𝑥
¨(𝑡) 𝑠2 𝑋(𝑠) − 𝑠𝑥(0) − 𝑥(0)
˙
Table 4.1: Table of Laplace Transforms

4.1. DEFINITION AND PROPERTIES 49
There is also a one-to-one correspondence between functions and their Laplace

transforms. A table of Laplace transforms can therefore be constructed and used
to find both Laplace and inverse Laplace transforms of commonly occurring
functions. Such a table is shown in Table 4.1 (and this table will be distributed
with the exams). In Table 4.1, 𝑛 is a positive integer. Also, the cryptic entries
for 𝑢𝑐 (𝑡) and 𝛿(𝑡 − 𝑐) will be explained later in S4.3.
The rows of Table 4.1 can be determined by a combination of direct inte-
gration and some tricks. We first compute directly the Laplace transform of
𝑒𝑎𝑡 𝑓 (𝑡) (line 1):
∫︁ ∞
ℒ{𝑒𝑎𝑡 𝑓 (𝑡)} = 𝑒−𝑠𝑡 𝑒𝑎𝑡 𝑓 (𝑡)𝑑𝑡
0
∫︁ ∞
= 𝑒−(𝑠−𝑎)𝑡 𝑓 (𝑡)𝑑𝑡
0
= 𝐹 (𝑠 − 𝑎).
We also compute directly the Laplace transform of 1 (line 2):

∫︁ ∞
ℒ{1} = 𝑒−𝑠𝑡 𝑑𝑡
0
]︂∞
1 −𝑠𝑡
=− 𝑒
𝑠 0
1
= .
𝑠
Now, the Laplace transform of 𝑒𝑎𝑡 (line 3) may be found using these two results:
ℒ{𝑒𝑎𝑡 } = ℒ{𝑒𝑎𝑡 · 1}
1 (4.2)
= .
𝑠−𝑎
The transform of 𝑡𝑛 (line 4) can be found by successive integration by parts. A
more interesting method uses Taylor series expansions for 𝑒𝑎𝑡 and 1/(𝑠 − 𝑎). We
have
{︃ ∞ }︃
∑︁ (𝑎𝑡)𝑛
𝑎𝑡
ℒ{𝑒 } = ℒ
𝑛=0
𝑛!
∞
(4.3)
∑︁ 𝑎𝑛
= ℒ{𝑡𝑛 }.
𝑛=0
𝑛!
Also, with 𝑠 > 𝑎,

1 1
=
𝑠−𝑎 𝑠(1 − 𝑎𝑠 )
∞
1 ∑︁ (︁ 𝑎 )︁𝑛
= (4.4)
𝑠 𝑛=0 𝑠
∞
∑︁ 𝑎𝑛
= .
𝑛=0
𝑠𝑛+1
Using (4.2), and equating the coefficients of powers of 𝑎 in (4.3) and (4.4), results
in line 4:
𝑛!
ℒ{𝑡𝑛 } = 𝑛+1 .
𝑠
The Laplace transform of 𝑡𝑛 𝑒𝑎𝑡 (line 5) can be found from line 1 and line 4:
𝑛!
ℒ{𝑡𝑛 𝑒𝑎𝑡 } = .
(𝑠 − 𝑎)𝑛+1
The Laplace transform of sin 𝑏𝑡 (line 6) may be found from the Laplace transform
of 𝑒𝑎𝑡 (line 3) using 𝑎 = 𝑖𝑏:
ℒ{sin 𝑏𝑡} = Im ℒ{𝑒𝑖𝑏𝑡 }
{︀ }︀
{︂ }︂
1
= Im
𝑠 − 𝑖𝑏
{︂ }︂
𝑠 + 𝑖𝑏
= Im
𝑠2 + 𝑏2
𝑏
= 2 .
𝑠 + 𝑏2
Similarly, the Laplace transform of cos 𝑏𝑡 (line 7) is
ℒ{cos 𝑏𝑡} = Re ℒ{𝑒𝑖𝑏𝑡 }
{︀ }︀
𝑠
= 2 .
𝑠 + 𝑏2
The transform of 𝑒𝑎𝑡 sin 𝑏𝑡 (line 8) can be found from the transform of sin 𝑏𝑡
(line 6) and line 1:
𝑏
ℒ{𝑒𝑎𝑡 sin 𝑏𝑡} = ;
(𝑠 − 𝑎)2 + 𝑏2
and similarly for the transform of 𝑒𝑎𝑡 cos 𝑏𝑡:
𝑠−𝑎
ℒ{𝑒𝑎𝑡 cos 𝑏𝑡} = .
(𝑠 − 𝑎)2 + 𝑏2
The Laplace transform of 𝑡 sin 𝑏𝑡 (line 10) can be found from the Laplace trans-
form of 𝑡𝑒𝑎𝑡 (line 5 with 𝑛 = 1) using 𝑎 = 𝑖𝑏:
ℒ{𝑡 sin 𝑏𝑡} = Im ℒ{𝑡𝑒𝑖𝑏𝑡 }
{︀ }︀
{︂ }︂
1
= Im
(𝑠 − 𝑖𝑏)2
(𝑠 + 𝑖𝑏)2
{︂ }︂
= Im
(𝑠2 + 𝑏2 )2
2𝑏𝑠
= 2 .
(𝑠 + 𝑏2 )2
Similarly, the Laplace transform of 𝑡 cos 𝑏𝑡 (line 11) is
ℒ{𝑡 cos 𝑏𝑡} = Re ℒ{𝑡𝑒𝑖𝑏𝑡 }
{︀ }︀
(𝑠 + 𝑖𝑏)2
{︂ }︂
= Re
(𝑠2 + 𝑏2 )2
𝑠2 − 𝑏2
= 2 .
(𝑠 + 𝑏2 )2
4.2. SOLUTION OF INITIAL VALUE PROBLEMS 51
We now transform the inhomogeneous constant-coefficient, second-order, lin-

ear inhomogeneous ode for 𝑥 = 𝑥(𝑡),
𝑎¨
𝑥 + 𝑏𝑥˙ + 𝑐𝑥 = 𝑔(𝑡),
making use of the linearity of the Laplace transform:
𝑥} + 𝑏ℒ{𝑥}
𝑎ℒ{¨ ˙ + 𝑐ℒ{𝑥} = ℒ{𝑔}.
To determine the Laplace transform of 𝑥(𝑡)
˙ (line 15) in terms of the Laplace
transform of 𝑥(𝑡) and the initial conditions 𝑥(0) and 𝑥(0),
˙ we define 𝑋(𝑠) =
ℒ{𝑥(𝑡)}, and integrate ∫︁ ∞
𝑒−𝑠𝑡 𝑥𝑑𝑡
˙
0
by parts. We let
𝑢 = 𝑒−𝑠𝑡 𝑑𝑣 = 𝑥𝑑𝑡
˙
−𝑠𝑡
𝑑𝑢 = −𝑠𝑒 𝑑𝑡 𝑣 = 𝑥.
Therefore,
∫︁ ∞ ]︀∞
∫︁ ∞
𝑒−𝑠𝑡 𝑥𝑑𝑡
˙ = 𝑥𝑒−𝑠𝑡 0
+𝑠 𝑒−𝑠𝑡 𝑥𝑑𝑡
0 0
= 𝑠𝑋(𝑠) − 𝑥(0),
where assumed convergence of the Laplace transform requires
lim 𝑥(𝑡)𝑒−𝑠𝑡 = 0.
𝑡→∞
Similarly, the Laplace transform of 𝑥

¨(𝑡) (line 16) is determined by integrating
∫︁ ∞
𝑒−𝑠𝑡 𝑥¨𝑑𝑡
0
by parts and using the just derived result for the first derivative. We let
𝑢 = 𝑒−𝑠𝑡 𝑑𝑣 = 𝑥
¨𝑑𝑡
−𝑠𝑡
𝑑𝑢 = −𝑠𝑒 𝑑𝑡 𝑣 = 𝑥,
˙
so that
∫︁ ∞ ]︀∞
∫︁ ∞
−𝑠𝑡
𝑒 𝑥
¨𝑑𝑡 = ˙ −𝑠𝑡 0
𝑥𝑒 +𝑠 𝑒−𝑠𝑡 𝑥𝑑𝑡
˙
0 0
(︀ )︀
= −𝑥(0)
˙ + 𝑠 𝑠𝑋(𝑠) − 𝑥(0)
= 𝑠2 𝑋(𝑠) − 𝑠𝑥(0) − 𝑥(0),
˙
−𝑠𝑡
where similarly we assume lim𝑡→∞ 𝑥(𝑡)𝑒
˙ = 0.
4.2 Solution of initial value problems

We begin with a simple homogeneous ode and show that the Laplace transform
method yields an identical result to our previously learned method. We then
apply the Laplace transform method to solve an inhomogeneous equation.
¨ − 𝑥−2𝑥
Example: Solve 𝑥 ˙ = 0 with 𝑥(0) = 1 and 𝑥(0)
˙ = 0 by two different
methods.
view tutorial
The characteristic equation of the ode is determined from the ansatz 𝑥 = 𝑒𝑟𝑡
and is
𝑟2 − 𝑟 − 2 = (𝑟 − 2)(𝑟 + 1) = 0.
𝑥(𝑡) = 𝑐1 𝑒2𝑡 + 𝑐2 𝑒−𝑡 .
To satisfy the initial conditions, we must have 1 = 𝑐1 + 𝑐2 and 0 = 2𝑐1 − 𝑐2 ,

requiring 𝑐1 = 31 and 𝑐2 = 32 . Therefore, the solution to the ode that satisfies
the initial conditions is given by
1 2𝑡 2 −𝑡
𝑥(𝑡) = 𝑒 + 𝑒 . (4.5)
3 3
We now solve this example using the Laplace transform. Taking the Laplace
transform of both sides of the ode, using the linearity of the transform, and
applying our result for the transform of the first and second derivatives, we find
[𝑠2 𝑋(𝑠) − 𝑠𝑥(0) − 𝑥(0)]

˙ − [𝑠𝑋(𝑠) − 𝑥(0)] − [2𝑋(𝑠)] = 0,
or
(𝑠 − 1)𝑥(0) + 𝑥(0)
˙
𝑋(𝑠) = .
𝑠2 − 𝑠 − 2
Note that the denominator of the right-hand-side is just the quadratic from the
characteristic equation of the homogeneous ode, and that this factor arises from
the derivatives of the exponential term in the Laplace transform integral.
Applying the initial conditions, we find
𝑠−1
𝑋(𝑠) = . (4.6)
(𝑠 − 2)(𝑠 + 1)
We have thus determined the Laplace transformed solution 𝑋(𝑠) = ℒ{𝑥(𝑡)}.

We now need to compute the inverse Laplace transform 𝑥(𝑡) = ℒ−1 {𝑋(𝑠)}.
However, direct inversion of (4.6) by searching Table 4.1 is not possible, but
a partial fraction expansion may be useful. In particular, we write
𝑠−1 𝑎 𝑏
= + . (4.7)
(𝑠 − 2)(𝑠 + 1) 𝑠−2 𝑠+1
The cover-up method can be used to solve for 𝑎 and 𝑏. We multiply both sides
of (4.7) by 𝑠 − 2 and put 𝑠 = 2 to isolate 𝑎:
]︂
𝑠−1
𝑎=
𝑠+1 𝑠=2
1
= .
3
4.2. SOLUTION OF INITIAL VALUE PROBLEMS 53
Similarly, we multiply both sides of (4.7) by 𝑠 + 1 and put 𝑠 = −1 to isolate 𝑏:

]︂
𝑠−1
𝑏=
𝑠 − 2 𝑠=−1
2
= .
3
Therefore,
1 1 2 1
𝑋(𝑠) = · + · ,
3 𝑠−2 3 𝑠+1
and line 3 of Table 4.1 gives us the inverse transforms of each term separately
to yield
1 2
𝑥(𝑡) = 𝑒2𝑡 + 𝑒−𝑡 ,
3 3
identical to (4.5).
Example: Solve 𝑥
¨ + 𝑥 = sin 2𝑡 with 𝑥(0) = 2 and 𝑥(0)
˙ = 1 by Laplace
transform methods.
Taking the Laplace transform of both sides of the ode, we find
𝑠2 𝑋(𝑠) − 𝑠𝑥(0) − 𝑥(0)

˙ + 𝑋(𝑠) = ℒ{sin 2𝑡}
2
= 2 ,
𝑠 +4
where the Laplace transform of sin 2𝑡 made use of line 6 of Table 4.1. Substi-
tuting for 𝑥(0) and 𝑥(0)
˙ and solving for 𝑋(𝑠), we obtain
2𝑠 + 1 2
𝑋(𝑠) = + 2 .
𝑠 + 1 (𝑠 + 1)(𝑠2 + 4)
2
To determine the inverse Laplace transform from Table 4.1, we perform a partial
fraction expansion of the second term:
2 𝑎𝑠 + 𝑏 𝑐𝑠 + 𝑑
= 2 + . (4.8)
(𝑠2 2
+ 1)(𝑠 + 4) 𝑠 + 1 𝑠2 + 4
By inspection, we can observe that 𝑎 = 𝑐 = 0 and that 𝑑 = −𝑏. A quick

calculation shows that 3𝑏 = 2, or 𝑏 = 2/3. Therefore,
2𝑠 + 1 2/3 2/3
𝑋(𝑠) = 2
+ 2 − 2
𝑠 + 1 𝑠 + 1 (𝑠 + 4)
2𝑠 5/3 2/3
= 2 + 2 − 2 .
𝑠 + 1 𝑠 + 1 (𝑠 + 4)
From lines 6 and 7 of Table 4.1, we obtain the solution by taking inverse Laplace
transforms of the three terms separately, where 𝑏 = 1 in the first two terms, and
𝑏 = 2 in the third term:
5 1
𝑥(𝑡) = 2 cos 𝑡 + sin 𝑡 − sin 2𝑡.
3 3
4.3 Heaviside and Dirac delta functions

The Laplace transform technique becomes truly useful when solving odes with
discontinuous or impulsive inhomogeneous terms, these terms commonly mod-
eled using Heaviside or Dirac delta functions. We will discuss these functions
in turn, as well as their Laplace transforms.
4.3.1 Heaviside function

view tutorial
The Heaviside or unit step function, denoted here by 𝑢𝑐 (𝑡), is zero for 𝑡 < 𝑐 and
is one for 𝑡 ≥ 𝑐; that is,
{︂
0, 𝑡 < 𝑐;
𝑢𝑐 (𝑡) = (4.9)
1, 𝑡 ≥ 𝑐.
The precise value of 𝑢𝑐 (𝑡) at the single point 𝑡 = 𝑐 shouldn’t matter.

The Heaviside function can be viewed as the step-up function. The step-
down function—one for 𝑡 < 𝑐 and zero for 𝑡 ≥ 𝑐—is defined as
{︂
1, 𝑡 < 𝑐;
1 − 𝑢𝑐 (𝑡) = (4.10)
0, 𝑡 ≥ 𝑐.
The step-up, step-down function—zero for 𝑡 < 𝑎, one for 𝑎 ≤ 𝑡 < 𝑏, and zero
for 𝑡 ≥ 𝑏—is defined as
⎧
⎨ 0, 𝑡 < 𝑎;
𝑢𝑎 (𝑡) − 𝑢𝑏 (𝑡) = 1, 𝑎 ≤ 𝑡 < 𝑏; (4.11)
0, 𝑡 ≥ 𝑏.
⎩
The Laplace transform of the Heaviside function is determined by integra-

tion:
∫︁ ∞
ℒ{𝑢𝑐 (𝑡)} = 𝑒−𝑠𝑡 𝑢𝑐 (𝑡)𝑑𝑡
∫︁0 ∞
= 𝑒−𝑠𝑡 𝑑𝑡
𝑐
−𝑐𝑠
𝑒
= ,
𝑠
and is given in line 12 of Table 4.1.

The Heaviside function can be used to represent a translation of a function
𝑓 (𝑡) a distance 𝑐 in the positive 𝑡 direction. We have
{︂
0, 𝑡 < 𝑐;
𝑢𝑐 (𝑡)𝑓 (𝑡 − 𝑐) =
f(t-c), 𝑡 ≥ 𝑐.
4.3. HEAVISIDE AND DIRAC DELTA FUNCTIONS 55
x x=f(t)
Figure 4.1: A linearly increasing function which turns into a sinusoidal function.
The Laplace transform is

∫︁ ∞
ℒ{𝑢𝑐 (𝑡)𝑓 (𝑡 − 𝑐)} = 𝑒−𝑠𝑡 𝑢𝑐 (𝑡)𝑓 (𝑡 − 𝑐)𝑑𝑡
0
∫︁ ∞
= 𝑒−𝑠𝑡 𝑓 (𝑡 − 𝑐)𝑑𝑡
∫︁𝑐 ∞
′
= 𝑒−𝑠(𝑡 +𝑐) 𝑓 (𝑡′ )𝑑𝑡′
0
∫︁ ∞
′
−𝑐𝑠
=𝑒 𝑒−𝑠𝑡 𝑓 (𝑡′ )𝑑𝑡′
0
= 𝑒−𝑐𝑠 𝐹 (𝑠),
where we have changed variables to 𝑡′ = 𝑡 − 𝑐. The translation of 𝑓 (𝑡) a distance

𝑐 in the positive 𝑡 direction corresponds to the multiplication of 𝐹 (𝑠) by the
exponential 𝑒−𝑐𝑠 . This result is shown in line 13 of Table 4.1.
Piecewise-defined inhomogeneous terms can be modeled using Heaviside
functions. For example, consider the general case of a piecewise function defined
on two intervals:
{︂
𝑓1 (𝑡), if 𝑡 < 𝑡* ;
𝑓 (𝑡) =
𝑓2 (𝑡), if 𝑡 ≥ 𝑡* .
Using the Heaviside function 𝑢𝑡* , the function 𝑓 (𝑡) can be written in a single
line as (︀ )︀
𝑓 (𝑡) = 𝑓1 (𝑡) + 𝑓2 (𝑡) − 𝑓1 (𝑡) 𝑢𝑡* (𝑡).
This example can be generalized to piecewise functions defined on multiple
intervals.
As a concrete example, suppose the inhomogeneous term is represented by
a linearly increasing function, which then turns into a sinusoidal function for
𝑡 > 𝑡* , as sketched in Fig. 4.1. Explicitly,

{︂
𝑎𝑡, )︀ if 𝑡 < 𝑡* ;
𝑓 (𝑡) = (︀
𝑎𝑡* + 𝑏 sin 𝜔(𝑡 − 𝑡* ) , if 𝑡 ≥ 𝑡* .
We can rewrite 𝑓 (𝑡) using the Heaviside function 𝑢𝑡* (𝑡):

(︂ )︂
(︀ )︀
𝑓 (𝑡) = 𝑎𝑡 + 𝑏 sin 𝜔(𝑡 − 𝑡* ) − 𝑎(𝑡 − 𝑡* ) · 𝑢𝑡* (𝑡).
This specific form of 𝑓 (𝑡) enables a relatively easy Laplace transform. We can
write
𝑓 (𝑡) = 𝑎𝑡 + ℎ(𝑡 − 𝑡* )𝑢𝑡* (𝑡),
where we have defined
ℎ(𝑡) = 𝑏 sin 𝜔𝑡 − 𝑎𝑡.
Using line 13, the Laplace transform of 𝑓 (𝑡) is
𝐹 (𝑠) = 𝑎ℒ{𝑡} + ℒ{ℎ(𝑡 − 𝑡* )𝑢𝑡* (𝑡)}

= 𝑎ℒ{𝑡} + 𝑒−𝑡* 𝑠 ℒ{ℎ(𝑡)},
and where using lines 4 and 6,
1 𝑏𝜔 𝑎
ℒ{𝑡} = , ℒ{ℎ(𝑡)} = − 2.
𝑠2 𝑠2 + 𝜔 2 𝑠
4.3.2 Dirac delta function

view tutorial
The Dirac delta function, denoted as 𝛿(𝑡), is defined by requiring that for any
function 𝑓 (𝑡),
∫︁ ∞
𝑓 (𝑡)𝛿(𝑡)𝑑𝑡 = 𝑓 (0).
−∞
The usual view of the shifted Dirac delta function 𝛿(𝑡 − 𝑐) is that it is zero
everywhere except at 𝑡 = 𝑐, where it is infinite, and the integral over the Dirac
delta function is one. The Dirac delta function is technically not a function,
but is what mathematicians call a distribution. Nevertheless, in most cases of
practical interest, it can be treated like a function, where physical results are
obtained following a final integration.
There are many ways to represent the Dirac delta function as a limit of a
well-defined function. For our purposes, the most useful representation makes
use of the step-up, step-down function of (4.11):
1
𝛿(𝑡 − 𝑐) = lim (𝑢𝑐−𝜖 (𝑡) − 𝑢𝑐+𝜖 (𝑡)).
𝜖→0 2𝜖
Before taking the limit, the well-defined step-up, step-down function is zero
except over a small interval of width 2𝜖 centered at 𝑡 = 𝑐, over which it takes
the large value 1/2𝜖. The integral of this function is one, independent of the
value of 𝜖.
4.4. DISCONTINUOUS OR IMPULSIVE TERMS 57
The Laplace transform of the Dirac delta function is easily found by inte-
gration using the definition of the delta function:
∫︁ ∞
ℒ{𝛿(𝑡 − 𝑐)} = 𝑒−𝑠𝑡 𝛿(𝑡 − 𝑐)𝑑𝑡
0
= 𝑒−𝑐𝑠 .
This result is shown in line 14 of Table 4.1.
4.4 Discontinuous or impulsive inhomogeneous

terms
We now solve some more challenging ode’s with discontinuous or impulsive
inhomogeneous terms.
𝑥 + 𝑥˙ + 2𝑥 = 𝑢5 (𝑡) − 𝑢20 (𝑡), with 𝑥(0) = 𝑥(0)

Example: Solve 2¨ ˙ =0
The inhomogeneous term here is a step-up, step-down function that is unity
over the interval (5, 20) and zero elsewhere. Taking the Laplace transform of
the ode using Table 4.1,
𝑒−5𝑠 𝑒−20𝑠
2 𝑠2 𝑋(𝑠) − 𝑠𝑥(0) − 𝑥(0)
(︀ )︀
˙ + 𝑠𝑋(𝑠) − 𝑥(0) + 2𝑋(𝑠) = − .
𝑠 𝑠
Using the initial values and solving for 𝑋(𝑠), we find
𝑒−5𝑠 − 𝑒−20𝑠
𝑋(𝑠) = .
𝑠(2𝑠2 + 𝑠 + 2)
To determine the solution for 𝑥(𝑡) we now need to find the inverse Laplace
transform. The exponential functions can be dealt with using line 13 of Table
4.1. We write
𝑋(𝑠) = (𝑒−5𝑠 − 𝑒−20𝑠 )𝐻(𝑠),
where
1
𝐻(𝑠) = .
𝑠(2𝑠2 + 𝑠 + 2)
Then using line 13, we have
𝑥(𝑡) = 𝑢5 (𝑡)ℎ(𝑡 − 5) − 𝑢20 (𝑡)ℎ(𝑡 − 20), (4.12)
where ℎ(𝑡) = ℒ−1 {𝐻(𝑠)}.

To determine ℎ(𝑡) we need the partial fraction expansion of 𝐻(𝑠). Since the
discriminant of 2𝑠2 + 𝑠 + 2 is negative, we have
1 𝑎 𝑏𝑠 + 𝑐
= + 2 ,
𝑠(2𝑠2 + 𝑠 + 2) 𝑠 2𝑠 + 𝑠 + 2
yielding the equation
1 = 𝑎(2𝑠2 + 𝑠 + 2) + (𝑏𝑠 + 𝑐)𝑠;

or after equating powers of 𝑠,

2𝑎 + 𝑏 = 0, 𝑎 + 𝑐 = 0, 2𝑎 = 1,
yielding 𝑎 = 21 , 𝑏 = −1, and 𝑐 = − 12 . Therefore,
1/2 𝑠 + 12
𝐻(𝑠) = − 2
𝑠 2𝑠 + 𝑠 + 2
𝑠 + 21
(︂ )︂
1 1
= − .
2 𝑠 𝑠2 + 12 𝑠 + 1
Inspecting Table 4.1, the first term can be transformed using line 2, and the
second term can be transformed using lines 8 and 9, provided we complete the
square of the denominator and then massage the numerator. That is, first we
complete the square:
(︂ )︂2
1 1 15
𝑠2 + 𝑠 + 1 = 𝑠 + + ;
2 4 16
and next we write
√︁
𝑠 + 14 + √115 15
(︀ )︀
𝑠 + 21 16
2 1 = )︀2 15 .
𝑠 + 2𝑠 + 1 1
(︀
𝑠 + 4 + 16
Therefore, the function 𝐻(𝑠) can be written as
⎛ √︁ ⎞
(︀ 1
)︀ (︂ )︂ 15
1 ⎝1 𝑠+ 4 1 16
𝐻(𝑠) = − − √ ⎠.
2 𝑠 (𝑠 + 41 )2 + 15 16 15 (𝑠 + 1 2
4 ) + 15
16
The first term is transformed using line 2, the second term using line 9 and the
third term using line 8. We finally obtain
√ √
(︂ (︂ )︂)︂
1 1
ℎ(𝑡) = 1 − 𝑒−𝑡/4 cos ( 15𝑡/4) + √ sin ( 15𝑡/4) , (4.13)
2 15
which when combined with (4.12) yields the rather complicated solution for
𝑥(𝑡).
We briefly comment that it is also possible to solve this example without
using the Laplace transform. The key idea is that both 𝑥 and 𝑥˙ are continuous
functions of 𝑡. Clearly from the form of the inhomogeneous term and the initial
conditions, 𝑥(𝑡) = 0 for 0 ≤ 𝑡 ≤ 5. We then solve the ode between 5 ≤ 𝑡 ≤ 20
with the inhomogeneous term equal to unity and initial conditions 𝑥(5) = 𝑥(5)
˙ =
0. This requires first finding the general homogeneous solution, next finding a
particular solution, and then adding the homogeneous and particular solutions
and solving for the two unknown constants. To simplify the algebra, note that
the best ansatz to use to find the homogeneous solution is 𝑥(𝑡) = 𝑒𝑟(𝑡−5) , and
not 𝑥(𝑡) = 𝑒𝑟𝑡 . Finally, we solve the homogeneous ode for 𝑡 ≥ 20 using as
boundary conditions the previously determined values 𝑥(20) and 𝑥(20),
˙ where
we have made use of the continuity of 𝑥 and 𝑥. ˙ Here, the best ansatz to use
is 𝑥(𝑡) = 𝑒𝑟(𝑡−20) . The student may benefit by trying this as an exercise and
attempting to obtain a final solution that agrees with the form given by (4.12)
and (4.13).
4.4. DISCONTINUOUS OR IMPULSIVE TERMS 59
𝑥 + 𝑥˙ + 2𝑥 = 𝛿(𝑡 − 5) with 𝑥(0) = 𝑥(0)

Example: Solve 2¨ ˙ =0
Here the inhomogeneous term is an impulse at time 𝑡 = 5. Taking the Laplace
transform of the ode using Table 4.1, and applying the initial conditions,
(2𝑠2 + 𝑠 + 2)𝑋(𝑠) = 𝑒−5𝑠 ,
so that
(︂ )︂
1 −5𝑠 1
𝑋(𝑠) = 𝑒
2 𝑠2 + 21 𝑠 + 1
(︂ )︂
1 −5𝑠 1
= 𝑒
2 (𝑠 + 41 )2 + 15
⎛ √︁16 ⎞
√︂ 15
1 16 −5𝑠 ⎝ 16
= 𝑒 ⎠.
2 15 (𝑠 + 41 )2 + 15
16
The inverse Laplace transform may now be computed using lines 8 and 13 of
Table 4.1:
2 (︀√
𝑥(𝑡) = √ 𝑢5 (𝑡)𝑒−(𝑡−5)/4 sin 15(𝑡 − 5)/4 .
)︀
(4.14)
15
It is interesting to solve this example without using a Laplace transform.
Clearly, 𝑥(𝑡) = 0 up to the time of impulse at 𝑡 = 5. Furthermore, after the
impulse the ode is homogeneous and can be solved with standard methods. The
only difficulty is determining the initial conditions of the homogeneous ode at
𝑡 = 5+ .
When the inhomogeneous term is proportional to a delta-function, the solu-
tion 𝑥(𝑡) is continuous across the delta-function singularity, but the derivative
of the solution 𝑥(𝑡)
˙ is discontinuous. If we integrate the second-order ode across
the singularity at 𝑡 = 5 and consider 𝜖 → 0, only the second derivative term of
the left-hand-side survives, and
∫︁ 5+𝜖 ∫︁ 5+𝜖
2 𝑥
¨𝑑𝑡 = 𝛿(𝑡 − 5)𝑑𝑡
5−𝜖 5−𝜖
= 1.
˙ − ) = 1/2. Since 𝑥(5

˙ + ) − 𝑥(5
And as 𝜖 → 0, we have 𝑥(5 ˙ − ) = 0, the appropriate
initial conditions immediately after the impulse force are 𝑥(5+ ) = 0 and 𝑥(5
˙ +) =
1/2. This result can be confirmed using (4.14).
Chapter 5
Series solutions of
second-order linear
homogeneous differential
equations
We consider the second-order linear homogeneous differential equation for 𝑦 =

𝑦(𝑥):
𝑃 (𝑥)𝑦 ′′ + 𝑄(𝑥)𝑦 ′ + 𝑅(𝑥)𝑦 = 0, (5.1)
where 𝑃 (𝑥), 𝑄(𝑥) and 𝑅(𝑥) are polynomials or convergent power series (around
𝑥 = 𝑥0 ), with no common polynomial factors (that could be divided out). The
value 𝑥 = 𝑥0 is called an ordinary point of (5.1) if 𝑃 (𝑥0 ) ̸= 0, and is called a
singular point if 𝑃 (𝑥0 ) = 0. Singular points will later be further classified as
regular singular points and irregular singular points. Our goal is to find two
independent solutions of (5.1), valid in a neighborhood about 𝑥 = 𝑥0 .
5.1 Ordinary points

If 𝑥0 is an ordinary point of (5.1), then it is possible to determine two power
series solutions (i.e., Taylor series) for 𝑦 = 𝑦(𝑥) that converge in a neighborhood
of 𝑥 = 𝑥0 . We illustrate the method of solution by solving two examples.
Example: Find the general solution of 𝑦 ′′ + 𝑦 = 0.

view tutorial
By now, you should know that the general solution is 𝑦(𝑥) = 𝑎0 cos 𝑥 + 𝑎1 sin 𝑥,
with 𝑎0 and 𝑎1 constants. To find a power series solution about the point 𝑥 = 0,
we write
∞
∑︁
𝑦(𝑥) = 𝑎 𝑛 𝑥𝑛 ;
𝑛=0
61
62 CHAPTER 5. SERIES SOLUTIONS
and upon differentiating term-by-term

∞
∑︁
𝑦 ′ (𝑥) = 𝑛𝑎𝑛 𝑥𝑛−1 ,
𝑛=1
and
∞
∑︁
𝑦 ′′ (𝑥) = 𝑛(𝑛 − 1)𝑎𝑛 𝑥𝑛−2 .
𝑛=2
Substituting the power series for 𝑦 and its derivatives into the differential equa-
tion to be solved, we obtain
∞
∑︁ ∞
∑︁
𝑛−2
𝑛(𝑛 − 1)𝑎𝑛 𝑥 + 𝑎𝑛 𝑥𝑛 = 0. (5.2)
𝑛=2 𝑛=0
The power-series solution method requires combining the two sums on the left-
hand-side of (5.2) into a single power series in 𝑥. To shift the exponent of 𝑥𝑛−2
in the first sum upward by two to obtain 𝑥𝑛 , we need to shift the summation
index downward by two; that is,
∞
∑︁ ∞
∑︁
𝑛(𝑛 − 1)𝑎𝑛 𝑥𝑛−2 = (𝑛 + 2)(𝑛 + 1)𝑎𝑛+2 𝑥𝑛 .
𝑛=2 𝑛=0
We can then combine the two sums in (5.2) to obtain

∞ (︂
∑︁ )︂
(𝑛 + 2)(𝑛 + 1)𝑎𝑛+2 + 𝑎𝑛 𝑥𝑛 = 0. (5.3)
𝑛=0
For (5.3) to be satisfied, the coefficient of each power of 𝑥 must vanish separately.
(This can be proved by setting 𝑥 = 0 after successive differentiation.) We
therefore obtain the recurrence relation
𝑎𝑛
𝑎𝑛+2 = − , 𝑛 = 0, 1, 2, . . . .
(𝑛 + 2)(𝑛 + 1)
We observe that even and odd coefficients decouple. We thus obtain two inde-
pendent sequences starting with first term 𝑎0 or 𝑎1 . Developing these sequences,
we have for the sequence beginning with 𝑎0 :
𝑎0 ,
1
𝑎2 = − 𝑎0 ,
2
1 1
𝑎4 = − 𝑎2 = 𝑎0 ,
4·3 4·3·2
1 1
𝑎6 = − 𝑎4 = − 𝑎0 ;
6·5 6!
and the general coefficient in this sequence for 𝑛 = 0, 1, 2, . . . is
(−1)𝑛
𝑎2𝑛 = 𝑎0 .
(2𝑛)!
5.1. ORDINARY POINTS 63
Also, for the sequence beginning with 𝑎1 :
𝑎1 ,
1
𝑎3 = − 𝑎1 ,
3·2
1 1
𝑎5 = − 𝑎3 = 𝑎1 ,
5·4 5·4·3·2
1 1
𝑎7 = − 𝑎5 = − 𝑎1 ;
7·6 7!
and the general coefficient in this sequence for 𝑛 = 0, 1, 2, . . . is
(−1)𝑛
𝑎2𝑛+1 = 𝑎1 .
(2𝑛 + 1)!
Using the principle of superposition, the general solution is therefore

∞ ∞
∑︁ (−1)𝑛 2𝑛 ∑︁ (−1)𝑛 2𝑛+1
𝑦(𝑥) = 𝑎0 𝑥 + 𝑎1 𝑥
𝑛=0
(2𝑛)! 𝑛=0
(2𝑛 + 1)!
𝑥2 𝑥4 𝑥3 𝑥5
(︂ )︂ (︂ )︂
= 𝑎0 1 − + − . . . + 𝑎1 𝑥 − + − ...
2! 4! 3! 5!
= 𝑎0 cos 𝑥 + 𝑎1 sin 𝑥,
as expected.
In our next example, we will solve the Airy’s Equation. This differential
equation arises in the study of optics, fluid mechanics, and quantum mechanics.
Example: Find the general solution of 𝑦 ′′ − 𝑥𝑦 = 0.

view tutorial
With
∞
∑︁
𝑦(𝑥) = 𝑎 𝑛 𝑥𝑛 ,
𝑛=0
the differential equation becomes

∞
∑︁ ∞
∑︁
𝑛(𝑛 − 1)𝑎𝑛 𝑥𝑛−2 − 𝑎𝑛 𝑥𝑛+1 = 0. (5.4)
𝑛=2 𝑛=0
We shift the first sum to 𝑥𝑛+1 by shifting the exponent up by three, i.e.,
∞
∑︁ ∞
∑︁
𝑛(𝑛 − 1)𝑎𝑛 𝑥𝑛−2 = (𝑛 + 3)(𝑛 + 2)𝑎𝑛+3 𝑥𝑛+1 .
𝑛=2 𝑛=−1
When combining the two sums in (5.4), we separate out the extra 𝑛 = −1 term
in the first sum given by 2𝑎2 . Therefore, (5.4) becomes
∞ (︂
∑︁ )︂
2𝑎2 + (𝑛 + 3)(𝑛 + 2)𝑎𝑛+3 − 𝑎𝑛 𝑥𝑛+1 = 0. (5.5)
𝑛=0
Setting coefficients of powers of 𝑥 to zero, we first find 𝑎2 = 0, and then obtain

the recursion relation
1
𝑎𝑛+3 = 𝑎𝑛 . (5.6)
(𝑛 + 3)(𝑛 + 2)
Three sequences of coefficients—those starting with either 𝑎0 , 𝑎1 or 𝑎2 —decouple.
In particular the three sequences are
𝑎0 , 𝑎3 , 𝑎6 , 𝑎9 , . . . ;
𝑎1 , 𝑎4 , 𝑎7 , 𝑎10 , . . . ;
𝑎2 , 𝑎5 , 𝑎8 , 𝑎11 . . . .
Since 𝑎2 = 0, we find immediately for the last sequence
𝑎2 = 𝑎5 = 𝑎8 = 𝑎11 = · · · = 0.
We compute the first four nonzero terms in the power series with coefficients
corresponding to the first two sequences. Starting with 𝑎0 , we have
𝑎0 ,
1
𝑎3 = 𝑎0 ,
3·2
1
𝑎6 = 𝑎0 ,
6·5·3·2
1
𝑎9 = 𝑎0 ;
9·8·6·5·3·2
and starting with 𝑎1 ,
𝑎1 ,
1
𝑎4 = 𝑎1 ,
4·3
1
𝑎7 = 𝑎1 ,
7·6·4·3
1
𝑎10 = 𝑎1 .
10 · 9 · 7 · 6 · 4 · 3
The general solution for 𝑦 = 𝑦(𝑥), can therefore be written as
𝑥3 𝑥6 𝑥9 𝑥4 𝑥7 𝑥10
(︂ )︂ (︂ )︂
𝑦(𝑥) = 𝑎0 1 + + + + . . . + 𝑎1 𝑥 + + + + ...
6 180 12960 12 504 45360
= 𝑎0 𝑦0 (𝑥) + 𝑎1 𝑦1 (𝑥).
Suppose we would like to graph the solutions 𝑦 = 𝑦0 (𝑥) and 𝑦 = 𝑦1 (𝑥)
versus 𝑥 by solving the differential equation 𝑦 ′′ − 𝑥𝑦 = 0 numerically. What
initial conditions should we use? Clearly, 𝑦 = 𝑦0 (𝑥) solves the ode with initial
values 𝑦(0) = 1 and 𝑦 ′ (0) = 0, while 𝑦 = 𝑦1 (𝑥) solves the ode with initial values
𝑦(0) = 0 and 𝑦 ′ (0) = 1.
The numerical solutions, obtained using MATLAB, are shown in Fig. 5.1.
Note that the solutions oscillate for negative 𝑥 and grow exponentially for posi-
tive 𝑥. This can be understood by recalling that 𝑦 ′′ + 𝑦 = 0 has oscillatory sine
and cosine solutions and 𝑦 ′′ − 𝑦 = 0 has exponential hyperbolic sine and cosine
solutions.
5.2. REGULAR SINGULAR POINTS: CAUCHY-EULER EQUATIONS 65
Airy’s functions
2
1
y0
−1
−2
−10 −8 −6 −4 −2 0 2
1
y1
−1
−2
−10 −8 −6 −4 −2 0 2
x
Figure 5.1: Numerical solution of Airy’s equation.
5.2 Regular singular points: Cauchy-Euler equa-

tions
view tutorial
The value 𝑥 = 𝑥0 is called a regular singular point of the ode
(𝑥 − 𝑥0 )2 𝑦 ′′ + 𝑝(𝑥)(𝑥 − 𝑥0 )𝑦 ′ + 𝑞(𝑥)𝑦 = 0, (5.7)
if 𝑝(𝑥) and 𝑞(𝑥) have convergent Taylor series about 𝑥 = 𝑥0 , i.e., 𝑝(𝑥) and 𝑞(𝑥)
can be written as a power-series in (𝑥 − 𝑥0 ):
𝑝(𝑥) = 𝑝0 + 𝑝1 (𝑥 − 𝑥0 ) + 𝑝2 (𝑥 − 𝑥0 )2 + . . . ,
𝑞(𝑥) = 𝑞0 + 𝑞1 (𝑥 − 𝑥0 ) + 𝑞2 (𝑥 − 𝑥0 )2 + . . . ,
with 𝑝𝑛 and 𝑞𝑛 constants, and 𝑞0 ̸= 0 so that (𝑥 − 𝑥0 ) is not a common factor

of the coefficients. Any point 𝑥 = 𝑥0 that is not an ordinary point or a regular
singular point is called an irregular singular point. Many important differential
equations of physical interest have regular singular points, and their solutions
go by the generic name of special functions, with specific names associated
with now famous mathematicians like Bessel, Legendre, Hermite, Laguerre and
Chebyshev.
Here, we will only consider the simplest ode with a regular singular point at
𝑥 = 0. This ode is called a Cauchy-Euler equation, and has the form
𝑥2 𝑦 ′′ + 𝛼𝑥𝑦 ′ + 𝛽𝑦 = 0, (5.8)
with 𝛼 and 𝛽 constants. Note that (5.7) reduces to a Cauchy-Euler equation

(about 𝑥 = 𝑥0 ) when one considers only the leading-order term in the Taylor
series expansion of the functions 𝑝(𝑥) and 𝑞(𝑥). In fact, taking 𝑝(𝑥) = 𝑝0 and
𝑞(𝑥) = 𝑞0 and solving the associated Cauchy-Euler equation results in at least
one of the leading-order solutions to the more general ode (5.7). Often, this
is sufficient to obtain initial conditions for numerical solution of the full ode.
Students wishing to learn how to find the general solution of (5.7) can consult
Boyce & DiPrima.
An appropriate ansatz for (5.8) is 𝑦 = 𝑥𝑟 , when 𝑥 > 0 and 𝑦 = (−𝑥)𝑟 when
𝑥 < 0, (or more generally, 𝑦 = |𝑥|𝑟 for all 𝑥), with 𝑟 constant. After substitution
into (5.8), we obtain for both positive and negative 𝑥
𝑟(𝑟 − 1)|𝑥|𝑟 + 𝛼𝑟|𝑥|𝑟 + 𝛽|𝑥|𝑟 = 0,
and we observe that our ansatz is rewarded by cancelation of |𝑥|𝑟 . We thus
obtain the following quadratic equation for 𝑟:
𝑟2 + (𝛼 − 1)𝑟 + 𝛽 = 0, (5.9)
which can be solved using the quadratic formula. Three cases immediately
appear: (i) real distinct roots, (ii) complex conjugate roots, (iii) repeated roots.
Students may recall being in a similar situation when solving the second-order
linear homogeneous ode with constant coefficients. Indeed, it is possible to
directly transform the Cauchy-Euler equation into an equation with constant
coefficients so that our previous results can be used.
The idea is to change variables so that the power law ansatz 𝑦 = 𝑥𝑟 becomes
an exponential ansatz. For 𝑥 > 0, if we let 𝑥 = 𝑒𝜉 and 𝑦(𝑥) = 𝑌 (𝜉), then the
ansatz 𝑦(𝑥) = 𝑥𝑟 becomes the ansatz 𝑌 (𝜉) = 𝑒𝑟𝜉 , appropriate if 𝑌 (𝜉) satisfies
a constant coefficient ode. If 𝑥 < 0, then the appropriate transformation is
𝑥 = −𝑒𝜉 , since 𝑒𝜉 > 0. We need only consider 𝑥 > 0 here and subsequently
generalize our result by replacing 𝑥 everywhere by its absolute value.
We thus transform the differential equation (5.8) for 𝑦 = 𝑦(𝑥) into a differ-
ential equation for 𝑌 = 𝑌 (𝜉), using 𝑥 = 𝑒𝜉 , or equivalently, 𝜉 = ln 𝑥. By the
chain rule,
𝑑𝑦 𝑑𝑌 𝑑𝜉
=
𝑑𝑥 𝑑𝜉 𝑑𝑥
1 𝑑𝑌
=
𝑥 𝑑𝜉
𝑑𝑌
= 𝑒−𝜉 ,
𝑑𝜉
so that symbolically,
𝑑 𝑑
= 𝑒−𝜉 .
𝑑𝑥 𝑑𝜉
The second derivative transforms as
𝑑2 𝑦
(︂ )︂
−𝜉 𝑑 −𝜉 𝑑𝑌
= 𝑒 𝑒
𝑑𝑥2 𝑑𝜉 𝑑𝜉
(︂ 2 )︂
𝑑 𝑌 𝑑𝑌
= 𝑒−2𝜉 − .
𝑑𝜉 2 𝑑𝜉
5.2. REGULAR SINGULAR POINTS: CAUCHY-EULER EQUATIONS 67
Upon substitution of the derivatives of 𝑦 into (5.8), and using 𝑥 = 𝑒𝜉 , we obtain
𝑒2𝜉 𝑒−2𝜉 (𝑌 ′′ − 𝑌 ′ ) + 𝛼𝑒𝜉 𝑒−𝜉 𝑌 ′ + 𝛽𝑌 = 𝑌 ′′ + (𝛼 − 1)𝑌 ′ + 𝛽𝑌

(︀ )︀ (︀ )︀
= 0.
As expected, the ode for 𝑌 = 𝑌 (𝜉) has constant coefficients, and with 𝑌 = 𝑒𝑟𝜉 ,
the characteristic equation for 𝑟 is given by (5.9). We now directly transfer
previous results obtained for the constant coefficient second-order linear homo-
geneous ode.
5.2.1 Real, distinct roots

This simplest case needs no transformation. If (𝛼 − 1)2 − 4𝛽 > 0, then with 𝑟±
the real roots of (5.9), the general solution is
𝑦(𝑥) = 𝑐1 |𝑥|𝑟+ + 𝑐2 |𝑥|𝑟− .
5.2.2 Complex conjugate roots

If (𝛼 − 1)2 − 4𝛽 < 0, we can write the complex roots of (5.9) as 𝑟± = 𝜆 ± 𝑖𝜇.
Recall the general solution for 𝑌 = 𝑌 (𝜉) is given by
𝑌 (𝜉) = 𝑒𝜆𝜉 (𝐴 cos 𝜇𝜉 + 𝐵 sin 𝜇𝜉) ;
and upon transformation, and replacing 𝑥 by |𝑥|,
𝑦(𝑥) = |𝑥|𝜆 𝐴 cos (𝜇 ln |𝑥|) + 𝐵 sin (𝜇 ln |𝑥|) .

(︀ )︀
5.2.3 Repeated roots
If (𝛼 − 1)2 − 4𝛽 = 0, there is one real root 𝑟 of (5.9). The general solution for
𝑌 is
𝑌 (𝜉) = 𝑒𝑟𝜉 (𝑐1 + 𝑐2 𝜉) ,
yielding
𝑦(𝑥) = |𝑥|𝑟 (𝑐1 + 𝑐2 ln |𝑥|) .
We now give examples illustrating these three cases.
Example: Solve 2𝑥2 𝑦 ′′ + 3𝑥𝑦 ′ − 𝑦 = 0 for 0 ≤ 𝑥 ≤ 1 with two-point

boundary condition 𝑦(0) = 0 and 𝑦(1) = 1.
Since 𝑥 > 0, we try 𝑦 = 𝑥𝑟 and obtain the characteristic equation
0 = 2𝑟(𝑟 − 1) + 3𝑟 − 1
= 2𝑟2 + 𝑟 − 1
= (2𝑟 − 1)(𝑟 + 1).
Since the characteristic equation has two real roots, the general solution is given
by
1
𝑦(𝑥) = 𝑐1 𝑥 2 + 𝑐2 𝑥−1 .
We now encounter for the first time two-point boundary conditions, which can
be used to determine the coefficients 𝑐1 and 𝑐2 . Since y(0)=0, we must have
𝑐2 = 0. Applying the remaining condition 𝑦(1) = 1, we obtain the unique
solution √
𝑦(𝑥) = 𝑥.
Note that 𝑥 = 0 is called a singular point of the ode since the general solution
is singular at 𝑥 = 0 when 𝑐2 ̸= 0. Our boundary condition imposes that 𝑦(𝑥) is
finite at 𝑥 = 0 removing the singular solution. Nevertheless, 𝑦 ′ remains singular
at 𝑥 = 0. Indeed, this is why we imposed a two-point boundary condition rather
than specifying the value of 𝑦 ′ (0) (which is infinite).
Example: Find the general solution of 𝑥2√𝑦 ′′ + 𝑥𝑦 ′ + 𝜋 2 𝑦 = 0 with two-

point boundary condition 𝑦(1) = 1 and 𝑦( 𝑒) = 1.
With the ansatz 𝑦 = 𝑥𝑟 , we obtain
0 = 𝑟(𝑟 − 1) + 𝑟 + 𝜋 2
= 𝑟2 + 𝜋2 ,
so that 𝑟 = ±𝑖𝜋. Therefore, with 𝜉 = ln 𝑥, we have 𝑌 (𝜉) = 𝐴 cos 𝜋𝜉 + 𝐵 sin 𝜋𝜉,

and the general solution for 𝑦(𝑥) is
𝑦(𝑥) = 𝐴 cos (𝜋 ln 𝑥) + 𝐵 sin (𝜋 ln 𝑥).
The first boundary

√ condition 𝑦(1) = 1 yields 𝐴 = 1. The second boundary
condition 𝑦( 𝑒) = 1 yields 𝐵 = 1.
Example: Find the general solution of 𝑥2 𝑦 ′′ + 5𝑥𝑦 ′ + 4𝑦 = 0 with two-

point boundary condition 𝑦(1) = 0 and 𝑦(𝑒) = 1.
With the ansatz 𝑦 = 𝑥𝑟 , we obtain
0 = 𝑟(𝑟 − 1) + 5𝑟 + 4
= 𝑟2 + 4𝑟 + 4
= (𝑟 + 2)2 ,
so that there is a repeated root 𝑟 = −2. With 𝜉 = ln 𝑥, we have 𝑌 (𝜉) =

(𝑐1 + 𝑐2 𝜉)𝑒−2𝜉 , so that the general solution is
𝑐1 + 𝑐2 ln 𝑥
𝑦(𝑥) = .
𝑥2
The first boundary condition 𝑦(1) = 0 yields 𝑐1 = 0. The second boundary
condition 𝑦(𝑒) = 1 yields 𝑐2 = 𝑒2 . The solution is therefore
𝑒2 ln 𝑥
𝑦(𝑥) = .
𝑥2
Chapter 6
Systems of first-order linear

equations
Systems of coupled linear equations can result, for example, from linear stability
analyses of nonlinear equations, and from normal mode analyses of coupled os-
cillators. Here, we will consider only the simplest case of a system of two coupled
first-order, linear, homogeneous equations with constant coefficients. These two
first-order equations are in fact equivalent to a single second-order equation, and
the methods of Chapter 3 could be used for solution. Nevertheless, viewing the
problem as a system of first-order equations introduces the important concept
of the phase space, and can easily be generalized to higher-order linear systems.
6.1 Determinants and the eigenvalue problem

We begin by reviewing some basic linear algebra. For the simplest 2 × 2 case,
let (︂ )︂ (︂ )︂
𝑎 𝑏 𝑥1
A= , x= , (6.1)
𝑐 𝑑 𝑥2
and consider the homogeneous equation
Ax = 0. (6.2)
When does there exist a nontrivial (not identically zero) solution for x? To
answer this question, we solve directly the system
𝑎𝑥1 + 𝑏𝑥2 = 0,
𝑐𝑥1 + 𝑑𝑥2 = 0.
Multiplying the first equation by 𝑑 and the second by 𝑏, and subtracting the
second equation from the first, results in
(𝑎𝑑 − 𝑏𝑐)𝑥1 = 0.
Similarly, multiplying the first equation by 𝑐 and the second by 𝑎, and subtract-
ing the first equation from the second, results in
(𝑎𝑑 − 𝑏𝑐)𝑥2 = 0.
69
70 CHAPTER 6. SYSTEMS OF EQUATIONS
Therefore, a nontrivial solution of (6.2) exists only if 𝑎𝑑 − 𝑏𝑐 = 0. If we define

the determinant of the 2 × 2 matrix A to be det A = 𝑎𝑑 − 𝑏𝑐, then we say that
a nontrivial solution to (6.2) exists provided det A = 0.
The same calculation may be repeated for a 3 × 3 matrix. If
⎛ ⎞ ⎛ ⎞
𝑎 𝑏 𝑐 𝑥1
A = ⎝ 𝑑 𝑒 𝑓 ⎠ , x = ⎝ 𝑥2 ⎠ ,
𝑔 ℎ 𝑖 𝑥3
then there exists a nontrivial solution to (6.2) provided det A = 0, where

det A = 𝑎(𝑒𝑖 − 𝑓 ℎ) − 𝑏(𝑑𝑖 − 𝑓 𝑔) + 𝑐(𝑑ℎ − 𝑒𝑔). The definition of the deter-
minant can be further generalized to any 𝑛 × 𝑛 matrix, and is typically taught
in a first course on linear algebra.
We now consider the eigenvalue problem. For A an 𝑛 × 𝑛 matrix and v an
𝑛 × 1 column vector, the eigenvalue problem solves the equation
Av = 𝜆v (6.3)
for eigenvalues 𝜆𝑖 and corresponding eigenvectors v𝑖 . We rewrite the eigenvalue

equation (6.3) as
(A − 𝜆I)v = 0, (6.4)
where I is the 𝑛×𝑛 identity matrix. A nontrivial solution of (6.4) exists provided
det (A − 𝜆I) = 0. (6.5)
Equation (6.5) is an 𝑛-th order polynomial equation in 𝜆, and is called the

characteristic equation of A. The characteristic equation can be solved for
the eigenvalues, and for each eigenvalue, a corresponding eigenvector can be
determined directly from (6.3).
We can demonstrate how this works for the 2 × 2 matrix A of (6.1). We
have
0 = det (A − 𝜆I)
⃒ ⃒
⃒ 𝑎−𝜆 𝑏 ⃒
= ⃒⃒ ⃒
𝑐 𝑑−𝜆 ⃒
= (𝑎 − 𝜆)(𝑑 − 𝜆) − 𝑏𝑐
= 𝜆2 − (𝑎 + 𝑑)𝜆 + (𝑎𝑑 − 𝑏𝑐).
This characteristic equation can be more generally written as
𝜆2 − TrA𝜆 + detA = 0, (6.6)
where TrA is the trace, or sum of the diagonal elements, of the matrix A. If
𝜆 is an eigenvalue of A, then the corresponding eigenvector v may be found by
solving (︂ )︂ (︂ )︂
𝑎−𝜆 𝑏 𝑣1
= 0,
𝑐 𝑑−𝜆 𝑣2
where the equation of the second row will always be a multiple of the equation
of the first row. The eigenvector v has arbitrary normalization, and we may
always choose for convenience 𝑣1 = 1.
In the next section, we will see several examples of an eigenvector analysis.
6.2. COUPLED FIRST-ORDER EQUATIONS 71
6.2 Two coupled first-order linear homogeneous

differential equations
With A a 2 × 2 constant matrix and x a 2 × 1 column vector, we now consider
the system of differential equations given by
ẋ = Ax. (6.7)
We will consider by example three cases separately: (i) eigenvalues of A are real
and there are two linearly independent eigenvectors; (ii) eigenvalues of A are
complex conjugates, and; (iii) A has only one linearly independent eigenvector.
These three cases are analogous to the cases considered previously when solving
the second-order, linear, constant-coefficient, homogeneous equation.
6.2.1 Two distinct real eigenvalues

We illustrate the solution method by example.
Example: Find the general solution of 𝑥˙ 1 = 𝑥1 + 𝑥2 , 𝑥˙ 2 = 4𝑥1 + 𝑥2 .

view tutorial
The equation to be solved may be rewritten in matrix form as
(︂ )︂ (︂ )︂ (︂ )︂
𝑑 𝑥1 1 1 𝑥1
= ,
𝑑𝑡 𝑥2 4 1 𝑥2
or using vector notation, written as (6.7). We take as our ansatz x(𝑡) = v𝑒𝜆𝑡 ,
where v and 𝜆 are independent of 𝑡. Upon substitution into (6.7), we obtain
𝜆v𝑒𝜆𝑡 = Av𝑒𝜆𝑡 ;
and upon cancelation of the exponential, we obtain the eigenvalue problem
Av = 𝜆v. (6.8)
Finding the characteristic equation using (6.6), we have
0 = det (A − 𝜆I)
= 𝜆2 − 2𝜆 − 3
= (𝜆 − 3)(𝜆 + 1).
Therefore, the two eigenvalues are 𝜆1 = 3 and 𝜆2 = −1.

To determine the corresponding eigenvectors, we substitute the eigenvalues
successively into
(A − 𝜆I)v = 0. (6.9)
We will write the corresponding eigenvectors v1 and v2 using the matrix nota-
tion (︂ )︂
(︀ )︀ 𝑣11 𝑣12
v1 v2 = ,
𝑣21 𝑣22
where the components of v1 and v2 are written with subscripts corresponding
to the first and second columns of a 2 × 2 matrix.
For 𝜆1 = 3, and unknown eigenvector v1 , we have from (6.9)

−2𝑣11 + 𝑣21 = 0,
4𝑣11 − 2𝑣21 = 0.
Clearly, the second equation is just the first equation multiplied by −2, so only
one equation is linearly independent. This will always be true, so for the 2 × 2
case we need only consider the first row of the matrix. The first eigenvector
therefore satisfies 𝑣21 = 2𝑣11 . Recall that an eigenvector is only unique up to
multiplication by a constant: we may therefore take 𝑣11 = 1 for convenience.
For 𝜆2 = −1, and eigenvector v2 = (𝑣12 , 𝑣22 )𝑇 , we have from (6.9)
2𝑣12 + 𝑣22 = 0,
so that 𝑣22 = −2𝑣12 . Here, we take 𝑣12 = 1.
Therefore, our eigenvalues and eigenvectors are given by
(︂ )︂ (︂ )︂
1 1
𝜆1 = 3, v1 = ; 𝜆2 = −1, v2 = .
2 −2
Using the principle of superposition, the general solution to the ode is therefore
x(𝑡) = 𝑐1 v1 𝑒𝜆1 𝑡 + 𝑐2 v2 𝑒𝜆2 𝑡 ,
or explicitly writing out the components,
𝑥1 (𝑡) = 𝑐1 𝑒3𝑡 + 𝑐2 𝑒−𝑡 ,
𝑥2 (𝑡) = 2𝑐1 𝑒3𝑡 − 2𝑐2 𝑒−𝑡 .
We can obtain a new perspective on the solution by drawing a phase-space
diagram, shown in Fig. 6.1, with “x-axis” 𝑥1 and “y-axis” 𝑥2 . Each curve corre-
sponds to a different initial condition, and represents the trajectory of a particle
for both positive and negative 𝑡 with velocity given by the differential equation.
The dark lines represent trajectories along the direction of the eigenvectors. If
𝑐2 = 0, the motion is along the eigenvector v1 with 𝑥2 = 2𝑥1 and motion is
away from the origin (arrows pointing out) since the eigenvalue 𝜆1 = 3 > 0.
If 𝑐1 = 0, the motion is along the eigenvector v2 with 𝑥2 = −2𝑥1 and motion
is towards the origin (arrows pointing in) since the eigenvalue 𝜆2 = −1 < 0.
When the eigenvalues are real and of opposite signs, the origin is called a saddle
point. Almost all trajectories (with the exception of those with initial conditions
exactly satisfying 𝑥2 (0) = −2𝑥1 (0)) eventually move away from the origin as 𝑡
increases.
The current example can also be solved by converting the system of two
first-order equations into a single second-order equation. Consider again the
system of equations
𝑥˙ 1 = 𝑥1 + 𝑥2 ,
𝑥˙ 2 = 4𝑥1 + 𝑥2 .
We differentiate the first equation and proceed to eliminate 𝑥2 as follows:
𝑥
¨1 = 𝑥˙ 1 + 𝑥˙ 2
= 𝑥˙ 1 + 4𝑥1 + 𝑥2
= 𝑥˙ 1 + 4𝑥1 + 𝑥˙ 1 − 𝑥1
= 2𝑥˙ 1 + 3𝑥1 .
phase space diagram

2
1.5
0.5
0
2
x
−0.5
−1
−1.5
−2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
x1
Figure 6.1: Phase space diagram for example with two real eigenvalues of oppo-
site sign.
Therefore, the equivalent second-order linear homogeneous equation is given by
¨1 − 2𝑥˙ 1 − 3𝑥1 = 0.
𝑥
If we had eliminated 𝑥1 instead, we would have found an identical equation for

𝑥2 :
¨2 − 2𝑥˙ 2 − 3𝑥2 = 0.
𝑥
The corresponding characteristic equation is 𝜆2 − 2𝜆 − 3 = 0, which is identical
to the characteristic equation of the matrix A. In general, a system of 𝑛 first-
order linear homogeneous equations can be converted into an equivalent 𝑛-th
order linear homogeneous equation. Numerical methods usually require the
conversion in reverse; that is, a conversion of an 𝑛-th order equation into a
system of 𝑛 first-order equations.
√ √
Example: Find the general solution of 𝑥˙ 1 = −3𝑥1 + 2𝑥2 , 𝑥˙ 2 = 2𝑥1 −
2𝑥2 .
The equations in matrix form are
(︂ )︂ (︂ √ )︂ (︂ )︂
𝑑 𝑥1 −3 2 𝑥1
= √ .
𝑑𝑡 𝑥2 2 −2 𝑥2
The ansatz x = v𝑒𝜆𝑡 leads to the eigenvalue problem Av = 𝜆v, with A the
matrix above. The eigenvalues are determined from
0 = det (A − 𝜆I)
= 𝜆2 + 5𝜆 + 4
= (𝜆 + 4)(𝜆 + 1).
phase space diagram

2
1.5
0.5
0
2
x
−0.5
−1
−1.5
−2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
x1
Figure 6.2: Phase space diagram for example with two real eigenvalues of same
sign.
Therefore, the eigenvalues of A are 𝜆1 = −4, 𝜆2 = −1. Proceeding to determine

the associated eigenvectors, for 𝜆1 = −4,
√
𝑣11 + 2𝑣21 = 0;
and for 𝜆2 = −1,

√
−2𝑣12 + 2𝑣22 = 0.
Taking the normalization 𝑣11 = 1 and 𝑣12 = 1, we obtain for the eigenvalues
and associated eigenvectors
(︂ )︂ (︂ )︂
𝜆1 = −4, v1 = √ 1 ; 𝜆2 = −1, v2 = √1 .
− 2/2 2
The general solution to the ode is therefore

(︂ )︂ (︂ )︂ (︂ )︂
𝑥1 √ 1 √1
= 𝑐1 𝑒−4𝑡 + 𝑐2 𝑒−𝑡 .
𝑥2 − 2/2 2
We show the phase space plot√in Fig. 6.2. If 𝑐2 = 0, the motion is along
the eigenvector v1 with 𝑥2 = −( 2/2)𝑥1 with eigenvalue √ 𝜆1 = −4 < 0. If
𝑐1 = 0, the motion is along the eigenvector v2 with 𝑥2 = 2𝑥1 with eigenvalue
𝜆2 = −1 < 0. When the eigenvalues are real and have the same sign, the origin
is called a node. A node may be attracting or repelling depending on whether
the eigenvalues are both negative (as is the case here) or positive. Observe that
the trajectories collapse onto the v2 eigenvector since 𝜆1 < 𝜆2 < 0 and decay is
more rapid along the v1 direction.
6.2.2 Complex conjugate eigenvalues

Example: Find the general solution of 𝑥˙ 1 = − 21 𝑥1 + 𝑥2 , 𝑥˙ 2 = −𝑥1 − 21 𝑥2 .
view tutorial
(︂ )︂ (︂ 1 )︂ (︂ )︂
𝑑 𝑥1 −2 1 𝑥1
= .
𝑑𝑡 𝑥2 −1 − 21 𝑥2
The ansatz x = v𝑒𝜆𝑡 leads to the equation
0 = det (A − 𝜆I)
5
= 𝜆2 + 𝜆 + .
4
Therefore, 𝜆 = −1/2±𝑖; and we observe that the eigenvalues occur as a complex
conjugate pair. We will denote the two eigenvalues as
1 ¯ = − 1 − 𝑖.
𝜆=− +𝑖 and 𝜆
2 2
¯ Therefore, the eigen-
Now, for A a real matrix, if Av = 𝜆v, then Av̄ = 𝜆v̄.
vectors also occur as a complex conjugate pair. The eigenvector v associated
with eigenvalue 𝜆 satisfies −𝑖𝑣1 + 𝑣2 = 0, and normalizing with 𝑣1 = 1, we have
(︂ )︂
1
v= .
𝑖
We have therefore determined two independent complex solutions to the ode,

that is,
¯
v𝑒𝜆𝑡 and v̄𝑒𝜆𝑡 ,
and we can form a linear combination of these two complex solutions to construct
two independent real solutions. Namely, if the complex functions 𝑧(𝑡) and 𝑧¯(𝑡)
are written as
𝑧(𝑡) = Re{𝑧(𝑡)} + 𝑖Im{𝑧(𝑡)}, 𝑧¯(𝑡) = Re{𝑧(𝑡)} − 𝑖Im{𝑧(𝑡)},
then two real functions can be constructed from the following linear combina-
tions of 𝑧 and 𝑧¯:
𝑧 + 𝑧¯ 𝑧 − 𝑧¯
= Re{𝑧(𝑡)} and = Im{𝑧(𝑡)}.
2 2𝑖
Thus the two real vector functions that can be constructed from our two complex
vector functions are
{︂(︂ )︂ }︂
1 1
Re{v𝑒𝜆𝑡 } = Re 𝑒(− 2 +𝑖)𝑡
𝑖
{︂(︂ )︂ }︂
1 1
= 𝑒− 2 𝑡 Re (cos 𝑡 + 𝑖 sin 𝑡)
𝑖
(︂ )︂
1 cos 𝑡
= 𝑒− 2 𝑡 ;
− sin 𝑡
phase space diagram

2
1.5
0.5
0
2
x
−0.5
−1
−1.5
−2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
x1
Figure 6.3: Phase space diagram for example with complex conjugate eigenval-
ues.
and
{︂(︂
)︂ }︂
𝜆𝑡 − 21 𝑡 1
Im{v𝑒 } = 𝑒 Im (cos 𝑡 + 𝑖 sin 𝑡)
𝑖
(︂ )︂
1 sin 𝑡
= 𝑒− 2 𝑡 .
cos 𝑡
Taking a linear superposition of these two real solutions yields the general so-
lution to the ode, given by
(︂ (︂ )︂ (︂ )︂)︂
1 cos 𝑡 sin 𝑡
x = 𝑒− 2 𝑡 𝐴 +𝐵 .
− sin 𝑡 cos 𝑡
The corresponding phase space diagram is shown in Fig. 6.3. We say the
origin is a spiral point. If the real part of the complex eigenvalue is negative,
as is the case here, then solutions spiral into the origin. If the real part of the
eigenvalue is positive, then solutions spiral out of the origin.
The direction of the spiral—here, it is clockwise—can be determined using
a concept from physics. If a particle of unit mass moves along a phase space
trajectory, then the angular momentum of the particle about the origin is equal
to the cross product of the position and velocity vectors: L = x × ẋ. With
both the position and velocity vectors lying in the two-dimensional phase space
plane, the angular momentum vector is perpendicular to this plane. With
x = (𝑥1 , 𝑥2 , 0), ẋ = (𝑥˙ 1 , 𝑥˙ 2 , 0),
then
L = (0, 0, 𝐿), with 𝐿 = 𝑥1 𝑥˙ 2 − 𝑥2 𝑥˙ 1 .
By the right-hand-rule, a clockwise rotation corresponds to 𝐿 < 0, and a coun-
terclockwise rotation to 𝐿 > 0.
The computation of 𝐿 in our present example proceeds from the differential

equations as follows:
1 1
𝑥1 𝑥˙ 2 − 𝑥2 𝑥˙ 1 = 𝑥1 (−𝑥1 − 𝑥2 ) − 𝑥2 (− 𝑥1 + 𝑥2 )
2 2
= −(𝑥21 + 𝑥22 )
< 0.
And since 𝐿 < 0, the spiral rotation is clockwise, as shown in Fig. 6.3.
6.2.3 Repeated eigenvalues with one eigenvector

Example: Find the general solution of 𝑥˙ 1 = 𝑥1 − 𝑥2 , 𝑥˙ 2 = 𝑥1 + 3𝑥2 .
view tutorial
(︂ )︂ (︂ )︂ (︂ )︂
𝑑 𝑥1 1 −1 𝑥1
= . (6.10)
𝑑𝑡 𝑥2 1 3 𝑥2
The ansatz x = v𝑒𝜆𝑡 leads to the characteristic equation
0 = det (A − 𝜆I)
= 𝜆2 − 4𝜆 + 4
= (𝜆 − 2)2 .
Therefore, 𝜆 = 2 is a repeated eigenvalue. The associated eigenvector is found

from −𝑣1 − 𝑣2 = 0, or 𝑣2 = −𝑣1 ; and normalizing with 𝑣1 = 1, we have
(︂ )︂
1
𝜆 = 2, v = .
−1
We have thus found a single solution to the ode, given by

(︂ )︂
1
x1 (𝑡) = 𝑐1 𝑒2𝑡 ,
−1
and we need to find the missing second solution to be able to satisfy the initial
conditions. An ansatz of 𝑡 times the first solution is tempting, but will fail. Here,
we will cheat and find the missing second solution by solving the equivalent
second-order, homogeneous, constant-coefficient differential equation.
We already know that this second-order differential equation for 𝑥1 (𝑡) has a
characteristic equation with a degenerate eigenvalue given by 𝜆 = 2. Therefore,
the general solution for 𝑥1 is given by
𝑥1 (𝑡) = (𝑐1 + 𝑡𝑐2 )𝑒2𝑡 .
Since from the first differential equation, 𝑥2 = 𝑥1 − 𝑥˙ 1 , we compute
𝑥˙ 1 = 2𝑐1 + (1 + 2𝑡)𝑐2 𝑒2𝑡 ,

(︀ )︀
phase space diagram

2
1.5
0.5
0
2
x
−0.5
−1
−1.5
−2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
x1
Figure 6.4: Phase space diagram for example with only one eigenvector.
so that
𝑥2 = 𝑥1 − 𝑥˙ 1
= (𝑐1 + 𝑡𝑐2 )𝑒2𝑡 − 2𝑐1 + (1 + 2𝑡)𝑐2 𝑒2𝑡
(︀ )︀
= −𝑐1 𝑒2𝑡 + 𝑐2 (−1 − 𝑡)𝑒2𝑡 .
Combining our results for 𝑥1 and 𝑥2 , we have therefore found

(︂ )︂ (︂ )︂ [︂(︂ )︂ (︂ )︂ ]︂
𝑥1 1 0 1
= 𝑐1 𝑒2𝑡 + 𝑐2 + 𝑡 𝑒2𝑡 .
𝑥2 −1 −1 −1
Our missing linearly independent solution is thus determined to be

[︂(︂ )︂ (︂ )︂ ]︂
0 1
x(𝑡) = 𝑐2 + 𝑡 𝑒2𝑡 . (6.11)
−1 −1
The second term of (6.11) is just 𝑡 times the first solution; however, this is
not sufficient. Indeed, the correct ansatz to find the second solution directly is
given by
x = (w + 𝑡v) 𝑒𝜆𝑡 , (6.12)
where 𝜆 and v is the eigenvalue and eigenvector of the first solution, and w
is an unknown vector to be determined. To illustrate this direct method, we
substitute (6.12) into ẋ = Ax, assuming Av = 𝜆v . Canceling the exponential,
we obtain
v + 𝜆 (w + 𝑡v) = Aw + 𝜆𝑡v.
Further canceling the common term 𝜆𝑡v and rewriting yields
(A − 𝜆I) w = v. (6.13)
6.3. NORMAL MODES 79
If A has only a single linearly independent eigenvector v, then (6.13) can be

solved for w (otherwise, it cannot). Using A, 𝜆 and v of our present example,
(6.13) is the system of equations given by
(︂ )︂ (︂ )︂ (︂ )︂
−1 −1 𝑤1 1
= .
1 1 𝑤2 −1
The first and second equation are the same, so that 𝑤2 = −(𝑤1 + 1). Therefore,
(︂ )︂
𝑤1
w=
−(𝑤1 + 1)
(︂ )︂ (︂ )︂
1 0
= 𝑤1 + .
−1 −1
Notice that the first term repeats the first found solution, i.e., a constant times
the eigenvector, and the second term is new. We therefore take 𝑤1 = 0 and
obtain (︂ )︂
0
w= ,
−1
as before.
The phase space diagram for this ode is shown in Fig. 6.4. The dark line is
the single eigenvector v of the matrix A. When there is only a single eigenvector,
the origin is called an improper node.
There is a definite counterclockwise rotation to the phase space trajectories,
and this can be confirmed from the calculation
𝐿 = 𝑥1 𝑥˙ 2 − 𝑥2 𝑥˙ 1
= 𝑥1 (𝑥1 + 3𝑥2 ) − 𝑥2 (𝑥1 − 𝑥2 )
= 𝑥21 + 2𝑥1 𝑥2 + 𝑥22
= (𝑥1 + 𝑥2 )2
> 0.
6.3 Normal modes

view tutorial, Part 1 view tutorial, Part 2
We now consider an application of the eigenvector analysis to the coupled mass-
spring system shown in Fig. 6.5. The position variables 𝑥1 and 𝑥2 are measured
from the equilibrium positions of the masses. Hooke’s law states that the spring
force is linearly proportional to the extension length of the spring, measured
from equilibrium. By considering the extension of the spring and the sign of the
force, we write Newton’s law 𝐹 = 𝑚𝑎 separately for each mass:
𝑥1 = −𝑘𝑥1 − 𝐾(𝑥1 − 𝑥2 ),
𝑚¨
𝑥2 = −𝑘𝑥2 − 𝐾(𝑥2 − 𝑥1 ).
𝑚¨
Further rewriting by collecting terms proportional to 𝑥1 and 𝑥2 yields
𝑥1 = −(𝑘 + 𝐾)𝑥1 + 𝐾𝑥2 ,

𝑚¨
𝑥2 = 𝐾𝑥1 − (𝑘 + 𝐾)𝑥2 .
𝑚¨
Figure 6.5: Coupled harmonic oscillators.
The equations for the coupled mass-spring system form a system of two second-
order linear homogeneous odes. In matrix form, 𝑚ẍ = Ax, or explicitly,
𝑑2
(︂ )︂ (︂ )︂ (︂ )︂
𝑥1 −(𝑘 + 𝐾) 𝐾 𝑥1
𝑚 2 = . (6.14)
𝑑𝑡 𝑥2 𝐾 −(𝑘 + 𝐾) 𝑥2
In analogy to a system of first-order equations, we try the ansatz x = v𝑒𝜆𝑡 , and

upon substitution into (6.14) we obtain the eigenvalue problem Av = 𝑚𝜆2 v.
The values of 𝑚𝜆2 are determined by solving the characteristic equation
0 = det (A − 𝑚𝜆2 I)
⃒ ⃒
⃒ −(𝑘 + 𝐾) − 𝑚𝜆2 𝐾 ⃒
= ⃒⃒ ⃒
2 ⃒
𝐾 −(𝑘 + 𝐾) − 𝑚𝜆
= (𝑚𝜆2 + 𝑘 + 𝐾)2 − 𝐾 2 .
The solution for m𝜆2 is

𝑚𝜆2 = −𝑘 − 𝐾 ± 𝐾,
and the two eigenvalues are
𝜆21 = −𝑘/𝑚, 𝜆22 = −(𝑘 + 2𝐾)/𝑚.
Since 𝜆21 , 𝜆22 < 0, both values of 𝜆 are imaginary, and 𝑥1 (𝑡) and 𝑥2 (𝑡) oscillate
with angular frequencies 𝜔1 = |𝜆1 | and 𝜔2 = |𝜆2 |, where
√︀ √︀
𝜔1 = 𝑘/𝑚, 𝜔2 = (𝑘 + 2𝐾)/𝑚.
The positions of the oscillating masses in general contain time dependencies of

the form sin 𝜔1 𝑡, cos 𝜔1 𝑡, and sin 𝜔2 𝑡, cos 𝜔2 𝑡.
It is of further interest to determine the eigenvectors, or so-called normal
modes of oscillation, associated with the two distinct angular frequencies. With
specific initial conditions proportional to an eigenvector, the mass will oscillate
with a single frequency. The eigenvector associated with 𝑚𝜆21 satisfies
−𝐾𝑣11 + 𝐾𝑣12 = 0,
so
√︀ that 𝑣11 = 𝑣12 . The normal mode associated with the frequency 𝜔1 =
𝑘/𝑚 thus follows a motion where 𝑥1 = 𝑥2 . Referring to Fig. 6.5, during this
motion the center spring length does not change, and the two masses oscillate
as if the center spring was absent (which is why the frequency of oscillation is
independent of 𝐾).
6.3. NORMAL MODES 81
Next, we determine the eigenvector associated with 𝑚𝜆22 :
𝐾𝑣21 + 𝐾𝑣22 = 0,
√︀ that 𝑣21 = −𝑣22 . The normal mode associated with the frequency 𝜔2 =
so
(𝑘 + 2𝐾)/𝑚 thus follows a motion where 𝑥1 = −𝑥2 . Again referring to
Fig. 6.5, during this motion the two equal masses symmetrically push or pull
against each side of the middle spring.
A general solution for x(𝑡) can be constructed from the eigenvalues and
eigenvectors. Our ansatz was x = v𝑒𝜆𝑡 , and for each of two eigenvectors v,
we have a pair of complex conjugate values for 𝜆. Accordingly, we first apply
the principle of superposition to obtain four real solutions, and √︀ then apply the
principle
√︀ again to obtain the general solution. With 𝜔1 = 𝑘/𝑚 and 𝜔2 =
(𝑘 + 2𝐾)/𝑚, the general solution is given by
(︂ )︂ (︂ )︂ (︂ )︂
𝑥1 1 1
= (𝐴 cos 𝜔1 𝑡 + 𝐵 sin 𝜔1 𝑡) + (𝐶 cos 𝜔2 𝑡 + 𝐷 sin 𝜔2 𝑡) ,
𝑥2 1 −1
where the now real constants 𝐴, 𝐵, 𝐶, and 𝐷 can be determined from the four
independent initial conditions, 𝑥1 (0), 𝑥2 (0), 𝑥˙ 1 (0), and 𝑥˙ 2 (0).
Chapter 7
Nonlinear differential
equations and bifurcation
theory
Reference: Strogatz, Sections 2.2, 2.4, 3.1, 3.2, 3.4, 6.3, 6.4, 8.2
We now turn our attention to nonlinear differential equations. In particular, we

study how small changes in the parameters of a system can result in qualita-
tive changes in the dynamics. These qualitative changes in the dynamics are
called bifurcations. To understand bifurcations, we first need to understand the
concepts of fixed points and stability.
7.1 Fixed points and stability

7.1.1 One dimension
view tutorial
Consider the one-dimensional differential equation for 𝑥 = 𝑥(𝑡) given by
𝑥˙ = 𝑓 (𝑥). (7.1)
We say that 𝑥* is a fixed point, or equilibrium point, of (7.1) if 𝑓 (𝑥* ) = 0. At

the fixed point, 𝑥˙ = 0. The terminology fixed point is used since the solution to
(7.1) with initial condition 𝑥(0) = 𝑥* is 𝑥(𝑡) = 𝑥* for all time 𝑡.
A fixed point, however, can be stable or unstable. A fixed point is said to be
stable if a small perturbation of the solution from the fixed point decays in time;
it is said to be unstable if a small perturbation grows in time. We can determine
stability by a linear analysis. Let 𝑥 = 𝑥* + 𝜖(𝑡), where 𝜖 represents a small
perturbation of the solution from the fixed point 𝑥* . Because 𝑥* is a constant,
𝑥˙ = 𝜖;
˙ and because 𝑥* is a fixed point, 𝑓 (𝑥* ) = 0. Taylor series expanding about
𝜖 = 0, we have
𝜖˙ = 𝑓 (𝑥* + 𝜖)
= 𝑓 (𝑥* ) + 𝜖𝑓 ′ (𝑥* ) + . . .
= 𝜖𝑓 ′ (𝑥* ) + . . . .
83
84 CHAPTER 7. NONLINEAR DIFFERENTIAL EQUATIONS
The omitted terms in the Taylor series expansion are proportional to 𝜖2 , and
can be made negligible over a short time interval with respect to the kept term,
proportional to 𝜖, by taking 𝜖(0) sufficiently small. Therefore, at least over short
times, the differential equation to be considered, 𝜖˙ = 𝑓 ′ (𝑥* )𝜖, is linear and has
by now the familiar solution
′
𝜖(𝑡) = 𝜖(0)𝑒𝑓 (𝑥* )𝑡
.
The perturbation of the fixed point solution 𝑥(𝑡) = 𝑥* thus decays exponentially
if 𝑓 ′ (𝑥* ) < 0, and we say the fixed point is stable. If 𝑓 ′ (𝑥* ) > 0, the perturbation
grows exponentially and we say the fixed point is unstable. If 𝑓 ′ (𝑥* ) = 0, we
say the fixed point is marginally stable, and the next higher-order term in the
Taylor series expansion must be considered.
Example: Find all the fixed points of the logistic equation 𝑥˙ = 𝑥(1 − 𝑥)
and determine their stability.
There are two fixed points at which 𝑥˙ = 0, given by 𝑥* = 0 and 𝑥* = 1. Stability

of these equilibrium points may be determined by considering the derivative of
𝑓 (𝑥) = 𝑥(1 − 𝑥). We have 𝑓 ′ (𝑥) = 1 − 2𝑥. Therefore, 𝑓 ′ (0) = 1 > 0 so that
𝑥* = 0 is an unstable fixed point, and 𝑓 ′ (1) = −1 < 0 so that 𝑥* = 1 is a stable
fixed point. Indeed, we have previously found that all solutions approach the
stable fixed point asymptotically.
7.1.2 Two dimensions

view tutorial
The idea of fixed points and stability can be extended to higher-order systems
of odes. Here, we consider a two-dimensional system and will need to make use
of the two-dimensional Taylor series expansion of a function 𝐹 (𝑥, 𝑦) about the
origin. In general, the Taylor series of 𝐹 (𝑥, 𝑦) is given by
𝜕2𝐹 𝜕2𝐹 𝜕2𝐹

(︂ )︂
𝜕𝐹 𝜕𝐹 1
𝐹 (𝑥, 𝑦) = 𝐹 + 𝑥 +𝑦 + 𝑥2 2
+ 2𝑥𝑦 + 𝑦2 2 + ...,
𝜕𝑥 𝜕𝑦 2 𝜕𝑥 𝜕𝑥𝜕𝑦 𝜕𝑦
where the function 𝐹 and all of its partial derivatives on the right-hand-side are
evaluated at the origin. Note that the Taylor series is constructed so that all
partial derivatives of the left-hand-side match those of the right-hand-side at
the origin.
We now consider the two-dimensional system given by
𝑥˙ = 𝑓 (𝑥, 𝑦), 𝑦˙ = 𝑔(𝑥, 𝑦). (7.2)
The point (𝑥* , 𝑦* ) is said to be a fixed point of (7.2) if 𝑓 (𝑥* , 𝑦* ) = 0 and

𝑔(𝑥* , 𝑦* ) = 0. Again, the local stability of a fixed point can be determined by
a linear analysis. We let 𝑥(𝑡) = 𝑥* + 𝜖(𝑡) and 𝑦(𝑡) = 𝑦* + 𝛿(𝑡), where 𝜖 and 𝛿
are small independent perturbations from the fixed point. Making use of the
two dimensional Taylor series of 𝑓 (𝑥, 𝑦) and 𝑔(𝑥, 𝑦) about the fixed point, or
7.1. FIXED POINTS AND STABILITY 85
equivalently about (𝜖, 𝛿) = (0, 0), we have
𝜖˙ = 𝑓 (𝑥* + 𝜖, 𝑦* + 𝛿)
𝜕𝑓 𝜕𝑓
=𝑓 +𝜖 +𝛿 + ...
𝜕𝑥 𝜕𝑦
𝜕𝑓 𝜕𝑓
=𝜖 +𝛿 + ....
𝜕𝑥 𝜕𝑦
𝛿˙ = 𝑔(𝑥* + 𝜖, 𝑦* + 𝛿)
𝜕𝑔 𝜕𝑔
=𝑔+𝜖 +𝛿 + ...
𝜕𝑥 𝜕𝑦
𝜕𝑔 𝜕𝑔
=𝜖 +𝛿 + ...,
𝜕𝑥 𝜕𝑦
where in the Taylor series 𝑓 , 𝑔 and all their partial derivatives are evaluated at
the fixed point (𝑥* , 𝑦* ). Neglecting higher-order terms in the Taylor series, we
thus have a system of odes for the perturbation, given in matrix form as
(︂ )︂ (︃ 𝜕𝑓 𝜕𝑓 )︃ (︂ )︂
𝑑 𝜖 𝜕𝑥 𝜕𝑦 𝜖
= 𝜕𝑔 𝜕𝑔 . (7.3)
𝑑𝑡 𝛿 𝜕𝑥 𝜕𝑦
𝛿
The two-by-two matrix in (7.3) is called the Jacobian matrix at the fixed point.
An eigenvalue analysis of the Jacobian matrix will typically yield two eigenvalues
𝜆1 and 𝜆2 . These eigenvalues may be real and distinct, complex conjugate pairs,
or repeated. The fixed point is stable (all perturbations decay exponentially)
if both eigenvalues have negative real parts. The fixed point is unstable (some
perturbations grow exponentially) if at least one of the eigenvalues has a positive
real part. Fixed points can be further classified as stable or unstable nodes,
unstable saddle points, stable or unstable spiral points, or stable or unstable
improper nodes.
Example: Find all the fixed points of the nonlinear system 𝑥˙ = 𝑥(3 −
𝑥 − 2𝑦), 𝑦˙ = 𝑦(2 − 𝑥 − 𝑦), and determine their stability.
view tutorial
The fixed points are determined by solving
𝑓 (𝑥, 𝑦) = 𝑥(3 − 𝑥 − 2𝑦) = 0, 𝑔(𝑥, 𝑦) = 𝑦(2 − 𝑥 − 𝑦) = 0.
There are four fixed points (𝑥* , 𝑦* ): (0, 0), (0, 2), (3, 0) and (1, 1). The Jacobian
matrix is given by
(︃ )︃ (︂
𝜕𝑓 𝜕𝑓 )︂
𝜕𝑥 𝜕𝑦 3 − 2𝑥 − 2𝑦 −2𝑥
𝜕𝑔 𝜕𝑔 = .
𝜕𝑥 𝜕𝑦
−𝑦 2 − 𝑥 − 2𝑦
Stability of the fixed points may be considered in turn. With J* the Jacobian
matrix evaluated at the fixed point, we have
(︂ )︂
3 0
(𝑥* , 𝑦* ) = (0, 0) : J* = .
0 2
phase space diagram

2.5
1.5
y
0.5
0
0 0.5 1 1.5 2 2.5 3 3.5
x
Figure 7.1: Phase space plot for two-dimensional nonlinear system.
The eigenvalues of J* are 𝜆 = 3, 2 so that the fixed point (0, 0) is an unstable

node. Next, (︂ )︂
−1 0
(𝑥* , 𝑦* ) = (0, 2) : J* = .
−2 −2
The eigenvalues of J* are 𝜆 = −1, −2 so that the fixed point (0, 2) is a stable
node. Next, (︂ )︂
−3 −6
(𝑥* , 𝑦* ) = (3, 0) : J* = .
0 −1
The eigenvalues of J* are 𝜆 = −3, −1 so that the fixed point (3, 0) is also a
stable node. Finally,
(︂ )︂
−1 −2
(𝑥* , 𝑦* ) = (1, 1) : J* = .
−1 −1
The characteristic
√ equation of J* is given by (−1 − 𝜆)2 − 2 = 0, so that 𝜆 =
−1 ± 2. Since one eigenvalue is negative and the other positive the fixed point
(1, 1) is an unstable saddle point. From our analysis of the fixed points, one can
expect that all solutions will asymptote to one of the stable fixed points (0, 2)
or (3, 0), depending on the initial conditions.
It is of interest to sketch the phase space diagram for this nonlinear system.
The eigenvectors associated with the unstable saddle point (1, 1) determine the
directions of the flow into and away from this fixed √ point. The eigenvector
associated with the positive eigenvalue 𝜆1 = −1 + 2 can determined from the
first equation of (J* − 𝜆1 I)v1 = 0, or
√
− 2𝑣11 − 2𝑣12 = 0,
√
so that 𝑣12 = −( √2/2)𝑣11 . The eigenvector
√ associated with the negative eigen-
value 𝜆1 = −1 − 2 satisfies 𝑣22 = ( 2/2)𝑣21 . The eigenvectors give the slope
7.2. ONE-DIMENSIONAL BIFURCATIONS 87
of the lines with origin at the fixed point for incoming (negative eigenvalue) and
outgoing (positive
√ eigenvalue) trajectories. The outgoing trajectories have
√ neg-
ative slope − 2/2 and the incoming trajectories have positive slope 2/2. A
rough sketch of the phase space diagram can be made by hand (as demonstrated
in class). Here, a computer generated plot obtained from numerical solution of
the nonlinear coupled odes is presented in Fig. 7.1. The curve starting from
the origin and at infinity, and terminating at the unstable saddle point is called
the separatrix. This curve separates the phase space into two regions: initial
conditions for which the solution asymptotes to the fixed point (0, 2), and initial
conditions for which the solution asymptotes to the fixed point (3, 0).
7.2 One-dimensional bifurcations

A bifurcation occurs in a nonlinear differential equation when a small change in
a parameter results in a qualitative change in the long-time solution. Examples
of bifurcations are when fixed points are created or destroyed, or change their
stability.
We now consider four classic bifurcations of one-dimensional nonlinear differ-
ential equations: saddle-node bifurcation, transcritical bifurcation, supercritical
pitchfork bifurcation, and subcritical pitchfork bifurcation. The corresponding
differential equation will be written as
𝑥˙ = 𝑓𝑟 (𝑥),
where the subscript 𝑟 represents a parameter that results in a bifurcation when

varied across zero. The simplest differential equations that exhibit these bifur-
cations are called the normal forms, and correspond to a local analysis (i.e.,
Taylor series expansion) of more general differential equations around the fixed
point, together with a possible rescaling of 𝑥.
7.2.1 Saddle-node bifurcation

view tutorial
The saddle-node bifurcation results in fixed points being created or destroyed.
The normal form for a saddle-node bifurcation is given by
𝑥˙ = 𝑟 + 𝑥2 .
√
The fixed points are 𝑥* = ± −𝑟. Clearly, two real fixed points exist when
𝑟 < 0 and no real fixed points exist when 𝑟 > 0. The stability of the fixed
points when 𝑟 < 0 are determined by the derivative of 𝑓 (𝑥) = 𝑟 + 𝑥2 , given by
𝑓 ′ (𝑥) = 2𝑥. Therefore, the negative fixed point is stable and the positive fixed
point is unstable.
Graphically, we can illustrate this bifurcation in two ways. First, in Fig. 7.2(a),
we plot 𝑥˙ versus 𝑥 for the three parameter values corresponding to 𝑟 < 0, 𝑟 = 0
and 𝑟 > 0. The values at which 𝑥˙ = 0 correspond to the fixed points, and
arrows are drawn indicating how the solution 𝑥(𝑡) evolves (to the right if 𝑥˙ > 0
and to the left if 𝑥˙ < 0). The stable fixed point is indicated by a filled circle
and the unstable fixed point by an open circle. Note that when 𝑟 = 0, solutions
converge to the origin from the left, but diverge from the origin on the right.
dx/dt
(a)
→ ← → x → →
r<0 r=0 r>0

x*
(b)
Figure 7.2: Saddlenode bifurcation. (a) 𝑥˙ versus 𝑥; (b) bifurcation diagram.
Second, in Fig. 7.2(b), we plot a bifurcation diagram illustrating the fixed point
𝑥* versus the bifurcation parameter 𝑟. The stable fixed point is denoted by a
solid line and the unstable fixed point by a dashed line. Note that the two fixed
points collide and annihilate at 𝑟 = 0, and there are no fixed points for 𝑟 > 0.
7.2.2 Transcritical bifurcation

view tutorial
A transcritical bifurcation occurs when there is an exchange of stabilities be-
tween two fixed points. The normal form for a transcritical bifurcation is given
by
𝑥˙ = 𝑟𝑥 − 𝑥2 .
The fixed points are 𝑥* = 0 and 𝑥* = 𝑟. The derivative of the right-hand-side is
𝑓 ′ (𝑥) = 𝑟 − 2𝑥, so that 𝑓 ′ (0) = 𝑟 and 𝑓 ′ (𝑟) = −𝑟. Therefore, for 𝑟 < 0, 𝑥* = 0
is stable and 𝑥* = 𝑟 is unstable, while for 𝑟 > 0, 𝑥* = 𝑟 is stable and 𝑥* = 0
is unstable. The two fixed points thus exchange stability as 𝑟 passes through
zero. The transcritical bifurcation is illustrated in Fig. 7.3.
7.2.3 Supercritical pitchfork bifurcation

view tutorial
The pitchfork bifurcations occur in physical models where fixed points appear
and disappear in pairs due to some intrinsic symmetry of the problem. Pitchfork
bifurcations can come in one of two types. In the supercritical bifurcation, a
pair of stable fixed points are created at the bifurcation (or critical) point and
dx/dt
(a)
← → ← x ← ← ← → ←
r<0 r=0 r>0

x
(b) *
Figure 7.3: Transcritical bifurcation. (a) 𝑥˙ versus 𝑥; (b) bifurcation diagram.
exist after (super) the bifurcation. In the subcritical bifurcation, a pair of

unstable fixed points are created at the bifurcation point and exist before (sub)
the bifurcation.
The normal form for the supercritical pitchfork bifurcation is given by
𝑥˙ = 𝑟𝑥 − 𝑥3 .
Note that the linear term results in exponential growth when 𝑟 > 0 and the
nonlinear
√ term stabilizes this growth. The fixed points are 𝑥* = 0 and 𝑥* =
± 𝑟, the latter fixed points existing only when √ 𝑟 > 0. The derivative of 𝑓 is
𝑓 ′ (𝑥) = 𝑟 − 3𝑥2 so that 𝑓 ′ (0) = 𝑟 and 𝑓 ′ (± 𝑟) = −2𝑟. Therefore, the fixed
point 𝑥√ * = 0 is stable for 𝑟 < 0 and unstable for 𝑟 > 0 while the fixed points
𝑥 = ± 𝑟 exist and are stable for 𝑟 > 0. Notice that the fixed point 𝑥* =√0
becomes unstable as 𝑟 crosses zero and two new stable fixed points 𝑥* = ± 𝑟
are born. The supercritical pitchfork bifurcation is illustrated in Fig. 7.4.
7.2.4 Subcritical pitchfork bifurcation

view tutorial
In the subcritical case, the cubic term is destabilizing. The normal form (to
order 𝑥3 ) is
𝑥˙ = 𝑟𝑥 + 𝑥3 .
√
The fixed points are 𝑥* = 0 and 𝑥* = ± −𝑟, the latter fixed points existing
only when 𝑟 ≤ 0. The √ derivative of the right-hand-side is 𝑓 ′ (𝑥) = 𝑟 + 3𝑥2 so
′ ′
that 𝑓 (0) = 𝑟 and 𝑓 (± −𝑟) = −2𝑟. Therefore, the fixed point 𝑥* = 0 is stable
(a) dx/dt
→ ← x → ← → ← → ←
r<0 r=0 r>0
x*
(b)
Figure 7.4: Supercritical pitchfork bifurcation. (a) 𝑥˙ versus 𝑥; (b) bifurcation

diagram.
√
for 𝑟 < 0 and unstable for 𝑟 > 0 while the fixed points 𝑥 = ± −𝑟 exist and are
unstable for 𝑟 < 0. There are no stable fixed points when 𝑟 > 0.
The absence of stable fixed points for 𝑟 > 0 indicates that the neglect of
terms of higher-order in 𝑥 than 𝑥3 in the normal form may be unwarranted.
Keeping to the intrinsic symmetry of the equations (only odd powers of 𝑥) we
can add a stabilizing nonlinear term proportional to 𝑥5 . The extended normal
form (to order 𝑥5 ) is
𝑥˙ = 𝑟𝑥 + 𝑥3 − 𝑥5 ,
and is somewhat more difficult to analyze. The fixed points are solutions of
𝑥(𝑟 + 𝑥2 − 𝑥4 ) = 0.
The fixed point 𝑥* = 0 exists for all 𝑟, and four additional fixed points can be
found from the solutions of the quadratic equation in 𝑥2 :
√︂
1 √
𝑥* = ± (1 ± 1 + 4𝑟).
2
These fixed points exist only if 𝑥* is real. √

Clearly, for the inner square-root to
be real, 𝑟 ≥ −1/4. Also observe that 1 − 1 + 4𝑟 becomes negative for 𝑟 > 0.
We thus have three intervals in 𝑟 to consider, and these regions and their fixed
x*
Figure 7.5: Subcritical pitchfork bifurcation.
points are
1
𝑟<− : 𝑥* = 0 (one fixed point);
4 √︂
1 1 √
− <𝑟<0: 𝑥* = 0, 𝑥* = ± (1 ± 1 + 4𝑟) (five fixed points);
4 2
√︂
1 √
𝑟>0: 𝑥* = 0, 𝑥* = ± (1 + 1 + 4𝑟) (three fixed points).
2
Stability is determined from 𝑓 ′ (𝑥) = 𝑟 + 3𝑥2 − 5𝑥4 . Now, 𝑓 ′ (0) = 𝑟 so 𝑥* = 0 is

stable for 𝑟 < 0 and unstable for 𝑟 > 0. The calculation for the other four roots
can be simplified by noting that 𝑥* satisfies 𝑟 + 𝑥2* − 𝑥4* = 0, or 𝑥4* = 𝑟 + 𝑥2* .
Therefore,
𝑓 ′ (𝑥* ) = 𝑟 + 3𝑥2* − 5𝑥4*

= 𝑟 + 3𝑥2* − 5(𝑟 + 𝑥2* )
= −4𝑟 − 2𝑥2*
= −2(2𝑟 + 𝑥2* ).
√
With 𝑥2* = 21 (1 ± 1 + 4𝑟), we have
√
(︂ )︂
1
𝑓 ′ (𝑥* ) = −2 2𝑟 + (1 ± 1 + 4𝑟)
2
(︀ √ )︀
= − (1 + 4𝑟) ± 1 + 4𝑟
√ (︀√ )︀
= − 1 + 4𝑟 1 + 4𝑟 ± 1 .
Clearly, the plus root is always stable since 𝑓 ′ (𝑥* ) < 0. The minus root exists
only for − 14 < 𝑟 < 0 and is unstable since 𝑓 ′ (𝑥* ) > 0. We summarize the
stability of the various fixed points:

1
𝑟<− : 𝑥* = 0 (stable);
4
1
− <𝑟<0: 𝑥* = 0, (stable)
4 √︂
1 √
𝑥* = ± (1 + 1 + 4𝑟) (stable);
2
√︂
1 √
𝑥* = ± (1 − 1 + 4𝑟) (unstable);
2
𝑟>0: 𝑥* = 0 (unstable)
√︂
1 √
𝑥* = ± (1 + 1 + 4𝑟) (stable).
2
The bifurcation diagram is shown in Fig. 7.5, and consists of a subcritical
pitchfork bifurcation at 𝑟 = 0 and two saddle-node bifurcations at 𝑟 = −1/4. We
can imagine what happens to 𝑥 as 𝑟 increases from negative values, supposing
there is some small noise in the system so that 𝑥 = 𝑥(𝑡) will diverge from
unstable fixed points. For 𝑟 < −1/4, the equilibrium value of 𝑥 is 𝑥* = 0. As
𝑟 increases into the range −1/4 < 𝑟 < 0, 𝑥 will remain at 𝑥* = 0. However, a
catastrophe occurs as soon as 𝑟 > 0. The 𝑥* = 0 fixed point becomes unstable
and the solution will jump up (or down) to the only remaining stable fixed point.
Such behavior is called a jump bifurcation. A similar catastrophe can happen
as 𝑟 decreases from positive values. In this case, the jump occurs as soon as
𝑟 < −1/4.
Since the stable equilibrium value of 𝑥 depends on whether we are increasing
or decreasing 𝑟, we say that the system exhibits hysteresis. The existence of
a subcritical pitchfork bifurcation can be very dangerous in engineering appli-
cations since a small change in a problem’s parameters can result in a large
change in the equilibrium state. Physically, this can correspond to a collapse of
a structure, or the failure of a component.
7.2.5 Application: a mathematical model of a fishery

view tutorial
We illustrate the utility of bifurcation theory by analyzing a simple model of a
fishery. We utilize the logistic equation (see S2.4.6) to model a fish population
in the absence of fishing. To model fishing, we assume that the government has
established fishing quotas so that at most a total of 𝐶 fish per year may be
caught. We assume that when the fish population is at the carrying capacity
of the environment, fisherman can catch nearly their full quota. When the fish
population drops to lower values, fish may be harder to find and the catch rate
may fall below 𝐶, eventually going to zero as the fish population diminishes.
Combining the logistic equation together with a simple model of fishing, we
propose the mathematical model
(︂ )︂
𝑑𝑁 𝑁 𝐶𝑁
= 𝑟𝑁 1 − − , (7.4)
𝑑𝑡 𝐾 𝐴+𝑁
where 𝑁 is the fish population size, 𝑡 is time, 𝑟 is the maximum potential growth
rate of the fish population, 𝐾 is the carrying capacity of the environment, 𝐶 is
x
*
0
α 2
(1+α) /4
c
Figure 7.6: Fishery bifurcation diagram.
the maximum rate at which fish can be caught, and 𝐴 is a constant satisfying
𝐴 < 𝐾 that is used to model the idea that fish become harder to catch when
scarce.
We nondimensionalize (7.4) using 𝑥 = 𝑁/𝐾, 𝜏 = 𝑟𝑡, 𝑐 = 𝐶/𝑟𝐾, 𝛼 = 𝐴/𝐾:
𝑑𝑥 𝑐𝑥
= 𝑥(1 − 𝑥) − . (7.5)
𝑑𝜏 𝛼+𝑥
Note that 0 ≤ 𝑥 ≤ 1, 𝑐 > 0 and 0 < 𝛼 < 1. We wish to qualitatively describe
the equilibrium solutions of (7.5) and the bifurcations that may occur as the
nondimensional catch rate 𝑐 increases from zero. Practically, a government
would like to issue each year as large a catch quota as possible without adversely
affecting the number of fish that may be caught in subsequent years.
The fixed points of (7.5) are 𝑥* = 0, valid for all 𝑐, and the solutions to
𝑥2 − (1 − 𝛼)𝑥 + (𝑐 − 𝛼) = 0, or
1 [︁ √︀ ]︁
𝑥* = (1 − 𝛼) ± (1 + 𝛼)2 − 4𝑐 . (7.6)
2
The fixed points given by (7.6) are real only if 𝑐 < 41 (1 + 𝛼)2 . Furthermore, the
minus root is greater than zero only if 𝑐 > 𝛼. We therefore need to consider
three intervals over which the following fixed points exist:
1 [︁ √︀ ]︁
0≤𝑐≤𝛼: 𝑥* = 0, 𝑥* = (1 − 𝛼) + (1 + 𝛼)2 − 4𝑐 ;
2
1 1 [︁ √︀ ]︁
𝛼<𝑐< (1 + 𝛼)2 : 𝑥* = 0, 𝑥* = (1 − 𝛼) ± (1 + 𝛼)2 − 4𝑐 ;
4 2
1
𝑐 > (1 + 𝛼)2 : 𝑥* = 0.
4
The stability of the fixed points can be determined with rigor analytically or
graphically. Here, we simply apply biological intuition together with knowledge
of the types of one dimensional bifurcations. An intuitive argument is made
simpler if we consider 𝑐 decreasing from large values. When 𝑐 is large, that
is 𝑐 > 41 (1 + 𝛼)2 , too many fish are being caught and our intuition suggests
that the fish population goes extinct. Therefore, in this interval, the single
fixed point 𝑥* = 0 must be stable. As 𝑐 decreases, a bifurcation occurs at
𝑐 = 41 (1 + 𝛼)2 introducing two additional fixed points at 𝑥* = (1 − 𝛼)/2. The
type of one dimensional bifurcation in which two fixed points are created as a
square root becomes real is a saddlenode bifurcation, and one of the fixed points
will be stable and the other unstable. Following these fixed points as 𝑐 → 0,
we observe that the plus root goes to one, which is the appropriate stable fixed
point when there is no fishing. We therefore conclude that the plus root is stable
and the minus root is unstable. As 𝑐 decreases further from this bifurcation,
the minus root collides with the fixed point 𝑥* = 0 at 𝑐 = 𝛼. This appears to
be a transcritical bifurcation and assuming an exchange of stability occurs, we
must have the fixed point 𝑥* = 0 becoming unstable for 𝑐 < 𝛼. The resulting
bifurcation diagram is shown in Fig. 7.6.
The purpose of simple mathematical models applied to complex ecological
problems is to offer some insight. Here, we have learned that overfishing (in
the model 𝑐 > 41 (1 + 𝛼)2 ) during one year can potentially result in a sudden
collapse of the fish catch in subsequent years, so that governments need to be
particularly cautious when contemplating increases in fishing quotas.
7.3 Two-dimensional bifurcations

All the one-dimensional bifurcations can also occur in two-dimensions along
one of the directions. In addition, a new type of bifurcation can also occur
in two-dimensions. Suppose there is some control parameter 𝜇. Furthermore,
suppose that for 𝜇 < 0, a two-dimensional system approaches a fixed point by
exponentially-damped oscillations. We know that the Jacobian matrix at the
fixed point with 𝜇 < 0 will have complex conjugate eigenvalues with negative real
parts. Now suppose that when 𝜇 > 0 the real parts of the eigenvalues become
positive so that the fixed point becomes unstable. This change in stability
of the fixed point is called a Hopf bifurcation. The Hopf bifurcation comes
in two types: supercritical Hopf bifurcation and subcritical Hopf bifurcation.
For the supercritical Hopf bifurcation, as 𝜇 increases slightly above zero, the
resulting oscillation around the now unstable fixed point is quickly stabilized at
small amplitude. This stable orbit is called a limit cycle. For the subcritical
Hopf bifurcation, as 𝜇 increases slightly above zero, the limit cycle immediately
jumps to large amplitude.
7.3.1 Supercritical Hopf bifurcation

A simple example of a supercritical Hopf bifurcation can be given in polar
coordinates:
𝑟˙ = 𝜇𝑟 − 𝑟3 ,
𝜃˙ = 𝜔 + 𝑏𝑟2 ,
7.3. TWO-DIMENSIONAL BIFURCATIONS 95
where 𝑥 = 𝑟 cos 𝜃 and 𝑦 = 𝑟 sin 𝜃. The parameter 𝜇 controls the stability of the
fixed point at the origin, the parameter 𝜔 is the frequency of oscillation near
the origin, and the parameter 𝑏 determines the dependence of the oscillation
frequency at larger amplitude oscillations. Although we include 𝑏 for generality,
our qualitative analysis of these equations will be independent of 𝑏.
The equation for the radius is of the form of the supercritical pitchfork
√
bifurcation. The fixed points are 𝑟* = 0 and 𝑟* = 𝜇 (note that 𝑟 > 0), and
the former fixed point is stable for 𝜇 < 0 and the latter is stable for 𝜇 > 0. The
transition of the eigenvalues of the Jacobian from negative real part to positive
real part can be seen if we transform these equations to cartesian coordinates.
We have using 𝑟2 = 𝑥2 + 𝑦 2 ,
˙ sin 𝜃
𝑥˙ = 𝑟˙ cos 𝜃 − 𝜃𝑟
= (𝜇𝑟 − 𝑟3 ) cos 𝜃 − (𝜔 + 𝑏𝑟2 )𝑟 sin 𝜃
= 𝜇𝑥 − (𝑥2 + 𝑦 2 )𝑥 − 𝜔𝑦 − 𝑏(𝑥2 + 𝑦 2 )𝑦
= 𝜇𝑥 − 𝜔𝑦 − (𝑥2 + 𝑦 2 )(𝑥 + 𝑏𝑦);
˙ cos 𝜃
𝑦˙ = 𝑟˙ sin 𝜃 + 𝜃𝑟
= (𝜇𝑟 − 𝑟3 ) sin 𝜃 + (𝜔 + 𝑏𝑟2 )𝑟 cos 𝜃
= 𝜇𝑦 − (𝑥2 + 𝑦 2 )𝑦 + 𝜔𝑥 + 𝑏(𝑥2 + 𝑦 2 )𝑥
= 𝜔𝑥 + 𝜇𝑦 − (𝑥2 + 𝑦 2 )(𝑦 − 𝑏𝑥).
The stability of the origin is determined by the Jacobian matrix evaluated at

the origin. The nonlinear terms in the equation vanish and the Jacobian matrix
at the origin is given by (︂ )︂
𝜇 −𝜔
𝐽= .
𝜔 𝜇
The eigenvalues are the solutions of (𝜇 − 𝜆)2 + 𝜔 2 = 0, or 𝜆 = 𝜇 ± 𝑖𝜔. As
𝜇 increases from negative to positive values, exponentially damped oscillations
change into exponentially growing oscillations. The nonlinear terms in the equa-
tions stabilize the growing oscillations into a limit cycle.
7.3.2 Subcritical Hopf bifurcation

The analogous example of a subcritical Hopf bifurcation is given by
𝑟˙ = 𝜇𝑟 + 𝑟3 − 𝑟5 ,
𝜃˙ = 𝜔 + 𝑏𝑟2 .
Here, the equation for the radius is of the form of the subcritical pitchfork
bifurcation. As 𝜇 increases from negative to positive values, the origin becomes
unstable and exponentially growing oscillations increase until the radius reaches
a stable fixed point far away from the origin. In practice, it may be difficult
to tell analytically if a Hopf bifurcation is supercritical or subcritical from the
equations of motion. Computational solution, however, can quickly distinguish
between the two types.
Chapter 8
Partial differential
equations
Differential equations containing partial derivatives with two or more inde-

pendent variables are called partial differential equations (pdes). These equa-
tions are of fundamental scientific interest but are substantially more difficult
to solve, both analytically and computationally, than odes. In this chapter, we
will derive two fundamental pdes and show how to solve them.
8.1 Derivation of the diffusion equation

view lecture
To derive the diffusion equation in one spacial dimension, we imagine a still
liquid in a long pipe of constant cross sectional area. A small quantity of dye
is placed in a cross section of the pipe and allowed to diffuse up and down the
pipe. The dye diffuses from regions of higher concentration to regions of lower
concentration.
We define 𝑢(𝑥, 𝑡) to be the concentration of the dye at position 𝑥 along the
pipe, and we wish to find the pde satisfied by 𝑢. It is useful to keep track of
the units of the various quantities involved in the derivation and we introduce
the bracket notation [𝑋] to mean the units of 𝑋. Relevant dimensional units
used in the derivation of the diffusion equation are mass 𝑚, length 𝑙, and time
𝑡. Assuming that the dye concentration is uniform in every cross section of the
pipe, the dimensions of concentration used here are [𝑢] = 𝑚/𝑙.
The mass of dye in the infinitesimal pipe volume located between position
𝑥1 and position 𝑥2 at time 𝑡, with 𝑥1 < 𝑥 < 𝑥2 , is given to order Δ𝑥 = 𝑥2 − 𝑥1
by
𝑀 = 𝑢(𝑥, 𝑡)Δ𝑥.
The mass of dye in this infinitesimal pipe volume changes by diffusion into and
out of the cross sectional ends situated at position 𝑥1 and 𝑥2 (Figure 8.1).
We assume the rate of diffusion is proportional to the concentration gradient,
a relationship known as Fick’s law of diffusion. Fick’s law of diffusion assumes
the mass flux 𝐽, with units [𝐽] = 𝑚/𝑡 across a cross section of the pipe is given
97
98 CHAPTER 8. PARTIAL DIFFERENTIAL EQUATIONS
J (x2)
J (x1)
X1 X2
Figure 8.1: Derivation of the diffusion equation.
by
𝐽 = −𝐷𝑢𝑥 , (8.1)
where the diffusion constant 𝐷 > 0 has units [𝐷] = 𝑙2 /𝑡, and we have used
the notation 𝑢𝑥 = 𝜕𝑢/𝜕𝑥. The mass flux is opposite in sign to the gradient of
concentration. The time rate of change in the mass of dye between 𝑥1 and 𝑥2
is given by the difference between the mass flux into and the mass flux out of
the infinitesimal cross sectional volume. When 𝑢𝑥 < 0, 𝐽 > 0 and the mass
flows into the volume at position 𝑥1 and out of the volume at position 𝑥2 . On
the other hand, when 𝑢𝑥 > 0, 𝐽 < 0 and the mass flows out of the volume at
position 𝑥1 and into the volume at position 𝑥2 . In both cases, the time rate of
change of the dye mass is given by
𝑑𝑀
= 𝐽(𝑥1 , 𝑡) − 𝐽(𝑥2 , 𝑡),
𝑑𝑡
or rewriting in terms of 𝑢(𝑥, 𝑡):
𝑢𝑡 (𝑥, 𝑡)Δ𝑥 = 𝐷 (𝑢𝑥 (𝑥2 , 𝑡) − 𝑢𝑥 (𝑥1 , 𝑡)) .
Dividing by Δ𝑥 and taking the limit Δ𝑥 → 0 results in the diffusion equation:
𝑢𝑡 = 𝐷𝑢𝑥𝑥 .
We note that the diffusion equation is identical to the heat conduction equation,
where 𝑢 is temperature, and the constant 𝐷 (commonly written as 𝜅) is the
thermal conductivity.
8.2 Derivation of the wave equation

view lecture
To derive the wave equation in one spacial dimension, we imagine an elastic
string that undergoes small amplitude transverse vibrations. We define 𝑢(𝑥, 𝑡)
to be the vertical displacement of the string from the 𝑥-axis at position 𝑥 and
time 𝑡, and we wish to find the pde satisfied by 𝑢. We define 𝜌 to be the
8.2. DERIVATION OF THE WAVE EQUATION 99
Figure 8.2: Derivation of the wave equation.
constant mass density of the string, 𝑇 the tension of the string, and 𝜃 the angle
between the string and the horizonal line. We consider an infinitesimal string
element located between 𝑥1 and 𝑥2 , with Δ𝑥 = 𝑥2 − 𝑥1 , as shown in Fig. 8.2.
The governing equations are Newton’s law of motion for the horizontal and
vertical acceleration of our infinitesimal string element, and we assume that the
string element only accelerates vertically. Therefore, the horizontal forces must
balance and we have
𝑇2 cos 𝜃2 = 𝑇1 cos 𝜃1 .
The vertical forces result in a vertical √
acceleration, and with√︀ 𝑢𝑡𝑡 the vertical
acceleration of the string element and 𝜌 Δ𝑥2 + Δ𝑢2 = 𝜌Δ𝑥 1 + 𝑢2𝑥 its mass,
where we have used 𝑢𝑥 = Δ𝑢/Δ𝑥, exact as Δ𝑥 → 0, we have
√︀
𝜌Δ𝑥 1 + 𝑢2𝑥 𝑢𝑡𝑡 = 𝑇2 sin 𝜃2 − 𝑇1 sin 𝜃1 .
We now make the assumption of small vibrations, that is Δ𝑢 ≪ Δ𝑥, or equiv-

alently 𝑢𝑥 ≪ 1. Note that [𝑢] = 𝑙 so that 𝑢𝑥 is dimensionless. With this
approximation, to leading-order in 𝑢𝑥 we have
cos 𝜃2 = cos 𝜃1 = 1,
sin 𝜃2 = 𝑢𝑥 (𝑥2 , 𝑡), sin 𝜃1 = 𝑢𝑥 (𝑥1 , 𝑡),

and √︀
1 + 𝑢2𝑥 = 1.
Therefore, to leading order 𝑇1 = 𝑇2 = 𝑇 , (i.e., the tension in the string is
approximately constant), and
𝜌Δ𝑥𝑢𝑡𝑡 = 𝑇 (𝑢𝑥 (𝑥2 , 𝑡) − 𝑢𝑥 (𝑥1 , 𝑡)) .

Dividing by Δ𝑥 and taking the limit Δ𝑥 → 0 results in the wave equation
𝑢𝑡𝑡 = 𝑐2 𝑢𝑥𝑥 ,
where 𝑐2 = 𝑇 /𝜌. Since [𝑇 ] = 𝑚𝑙/𝑡2 and [𝜌] = 𝑚/𝑙, we have [𝑐2 ] = 𝑙2 /𝑡2 so that
𝑐 has units of velocity.
8.3 Fourier series

view tutorial
Our solution of the diffusion and wave equations will require use of a Fourier
series. A periodic function 𝑓 (𝑥) with period 2𝐿, can be represented as a Fourier
series in the form
∞
𝑎0 ∑︁ (︁ 𝑛𝜋𝑥 𝑛𝜋𝑥 )︁
𝑓 (𝑥) = + 𝑎𝑛 cos + 𝑏𝑛 sin . (8.2)
2 𝑛=1
𝐿 𝐿
Determination of the coefficients 𝑎0 , 𝑎1 , 𝑎2 , . . . and 𝑏1 , 𝑏2 , 𝑏3 , . . . makes use of

orthogonality relations for sine and cosine. We first define the widely used
Kronecker delta 𝛿𝑛𝑚 as
{︂
1 if 𝑛 = 𝑚;
𝛿𝑛𝑚 =
0 otherwise.
The orthogonality relations for 𝑛 and 𝑚 positive integers are then given with
compact notation as the integration formulas
∫︁ 𝐿 (︁ 𝑚𝜋𝑥 )︁ (︁ 𝑛𝜋𝑥 )︁
cos cos 𝑑𝑥 = 𝐿𝛿𝑛𝑚 , (8.3)
−𝐿 𝐿 𝐿
∫︁ 𝐿 (︁ 𝑚𝜋𝑥 )︁ (︁ 𝑛𝜋𝑥 )︁
sin sin 𝑑𝑥 = 𝐿𝛿𝑛𝑚 , (8.4)
−𝐿 𝐿 𝐿
∫︁ 𝐿 (︁ 𝑚𝜋𝑥 )︁ (︁ 𝑛𝜋𝑥 )︁
cos sin 𝑑𝑥 = 0. (8.5)
−𝐿 𝐿 𝐿
We illustrate the integration technique used to obtain these results. To derive

(8.4), we assume that 𝑛 and 𝑚 are positive integers with 𝑛 ̸= 𝑚, and we make
use of the change of variables 𝜉 = 𝜋𝑥/𝐿:
∫︁ 𝐿 (︁ 𝑚𝜋𝑥 )︁ (︁ 𝑛𝜋𝑥 )︁
sin sin 𝑑𝑥
−𝐿 𝐿 𝐿
∫︁ 𝜋
𝐿
= sin (𝑚𝜉) sin (𝑛𝜉)𝑑𝜉
𝜋 −𝜋
𝐿 𝜋 [︀
∫︁
(︀ )︀ (︀ )︀]︀
= cos (𝑚 − 𝑛)𝜉 − cos (𝑚 + 𝑛)𝜉 𝑑𝜉
2𝜋 −𝜋
)︀ 𝜋
[︂ ]︂
𝐿 1 (︀ )︀ 1 (︀
= sin (𝑚 − 𝑛)𝜉 − sin (𝑚 + 𝑛)𝜉
2𝜋 𝑚 − 𝑛 𝑚+𝑛 −𝜋
= 0.
8.3. FOURIER SERIES 101
For 𝑚 = 𝑛, however,
𝐿
𝐿 𝜋
∫︁ (︁ 𝑛𝜋𝑥 )︁ ∫︁
2
sin 𝑑𝑥 = sin2 (𝑛𝜉)𝑑𝜉
−𝐿 𝐿 𝜋 −𝜋
𝐿 𝜋 (︀
∫︁
)︀
= 1 − cos (2𝑛𝜉) 𝑑𝜉
2𝜋 −𝜋
[︂ ]︂𝜋
𝐿 1
= 𝜉− sin 2𝑛𝜉
2𝜋 2𝑛 −𝜋
= 𝐿.
Integration formulas given by (8.3) and (8.5) can be similarly derived.

To determine the coefficient 𝑎𝑛 , we multiply both sides of (8.2) by cos (𝑛𝜋𝑥/𝐿)
with 𝑛 a nonnegative integer, and change the dummy summation variable from
𝑛 to 𝑚. Integrating over 𝑥 from −𝐿 to 𝐿 and assuming that the integration can
be done term by term in the infinite sum, we obtain
𝐿
𝑎0 𝐿
∫︁ ∫︁
𝑛𝜋𝑥 𝑛𝜋𝑥
𝑓 (𝑥) cos 𝑑𝑥 = cos 𝑑𝑥
−𝐿 𝐿 2 −𝐿 𝐿
∞
{︃ ∫︁ }︃
𝐿 ∫︁ 𝐿
∑︁ 𝑛𝜋𝑥 𝑚𝜋𝑥 𝑛𝜋𝑥 𝑚𝜋𝑥
+ 𝑎𝑚 cos cos 𝑑𝑥 + 𝑏𝑚 cos sin 𝑑𝑥 .
𝑚=1 −𝐿 𝐿 𝐿 −𝐿 𝐿 𝐿
If 𝑛 = 0, then the second and third integrals on the right-hand-side are zero
and the first integral is 2𝐿 so that the right-hand-side becomes 𝐿𝑎0 . If 𝑛 is
a positive integer, then the first and third integrals on the right-hand-side are
zero, and the second integral is 𝐿𝛿𝑛𝑚 . For this case, we have
∫︁ 𝐿 ∞
𝑛𝜋𝑥 ∑︁
𝑓 (𝑥) cos 𝑑𝑥 = 𝐿𝑎𝑚 𝛿𝑛𝑚
−𝐿 𝐿 𝑚=1
= 𝐿𝑎𝑛 ,
where all the terms in the summation except 𝑚 = 𝑛 are zero by virtue of the
Kronecker delta. We therefore obtain for 𝑛 = 0, 1, 2, . . .
∫︁ 𝐿
1 𝑛𝜋𝑥
𝑎𝑛 = 𝑓 (𝑥) cos 𝑑𝑥. (8.6)
𝐿 −𝐿 𝐿
To determine the coefficients 𝑏1 , 𝑏2 , 𝑏3 , . . . , we multiply both sides of (8.2) by

sin (𝑛𝜋𝑥/𝐿), with 𝑛 a positive integer, and again change the dummy summation
variable from 𝑛 to 𝑚. Integrating, we obtain
𝐿
𝑎0 𝐿
∫︁ ∫︁
𝑛𝜋𝑥 𝑛𝜋𝑥
𝑓 (𝑥) sin 𝑑𝑥 = sin 𝑑𝑥
−𝐿 𝐿 2 −𝐿 𝐿
∞
{︃ ∫︁ 𝐿 ∫︁ 𝐿 }︃
∑︁ 𝑛𝜋𝑥 𝑚𝜋𝑥 𝑛𝜋𝑥 𝑚𝜋𝑥
+ 𝑎𝑚 sin cos 𝑑𝑥 + 𝑏𝑚 sin sin 𝑑𝑥 .
𝑚=1 −𝐿 𝐿 𝐿 −𝐿 𝐿 𝐿
Here, the first and second integrals on the right-hand-side are zero, and the
third integral is 𝐿𝛿𝑛𝑚 so that
∫︁ 𝐿 ∞
𝑛𝜋𝑥 ∑︁
𝑓 (𝑥) sin 𝑑𝑥 = 𝐿𝑏𝑚 𝛿𝑛𝑚
−𝐿 𝐿 𝑚=1
= 𝐿𝑏𝑛 .
Hence, for 𝑛 = 1, 2, 3, . . . ,
∫︁ 𝐿
1 𝑛𝜋𝑥
𝑏𝑛 = 𝑓 (𝑥) sin 𝑑𝑥. (8.7)
𝐿 −𝐿 𝐿
Our results for the Fourier series of a function 𝑓 (𝑥) with period 2𝐿 are thus
given by (8.2), (8.6) and (8.7).
8.4 Fourier cosine and sine series

view tutorial
The Fourier series simplifies if 𝑓 (𝑥) is an even function such that 𝑓 (−𝑥) = 𝑓 (𝑥),
or an odd function such that 𝑓 (−𝑥) = −𝑓 (𝑥). Use will be made of the following
facts. The function cos (𝑛𝜋𝑥/𝐿) is an even function and sin (𝑛𝜋𝑥/𝐿) is an odd
function. The product of two even functions is an even function. The product
of two odd functions is an even function. The product of an even and an odd
function is an odd function. And if 𝑔(𝑥) is an even function, then
∫︁ 𝐿 ∫︁ 𝐿
𝑔(𝑥)𝑑𝑥 = 2 𝑔(𝑥)𝑑𝑥;
−𝐿 0
and if 𝑔(𝑥) is an odd function, then

∫︁ 𝐿
𝑔(𝑥)𝑑𝑥 = 0.
−𝐿
We examine in turn the Fourier series for an even or an odd function. First,
if 𝑓 (𝑥) is even, then from (8.6) and (8.7) and our facts about even and odd
functions,
2 𝐿
∫︁
𝑛𝜋𝑥
𝑎𝑛 = 𝑓 (𝑥) cos 𝑑𝑥,
𝐿 0 𝐿 (8.8)
𝑏𝑛 = 0.
The Fourier series for an even function with period 2𝐿 is thus given by the
Fourier cosine series
∞
𝑎0 ∑︁ 𝑛𝜋𝑥
𝑓 (𝑥) = + 𝑎𝑛 cos , 𝑓 (𝑥) even. (8.9)
2 𝑛=1
𝐿
Second, if 𝑓 (𝑥) is odd, then

𝑎𝑛 = 0,
∫︁ 𝐿 (8.10)
2 𝑛𝜋𝑥
𝑏𝑛 = 𝑓 (𝑥) sin 𝑑𝑥;
𝐿 0 𝐿
8.4. FOURIER COSINE AND SINE SERIES 103
−1
−2*pi −pi 0 pi 2*pi
Figure 8.3: The even triangle function.
and the Fourier series for an odd function with period 2𝐿 is given by the Fourier
sine series
∞
∑︁ 𝑛𝜋𝑥
𝑓 (𝑥) = 𝑏𝑛 sin , 𝑓 (𝑥) odd. (8.11)
𝑛=1
𝐿
Examples of Fourier series computed numerically can be obtained using the

Java applet found at http://www.falstad.com/fourier. Here, we demonstrate an
analytical example.
Example: Determine the Fourier cosine series of the even triangle

function represented by Fig. 8.3.
view tutorial
The triangle function depicted in Fig. 8.3 is an even function of 𝑥 with period
2𝜋 (i.e., 𝐿 = 𝜋). Its definition on 0 < 𝑥 < 𝜋 is given by
2𝑥
𝑓 (𝑥) = 1 − .
𝜋
Because 𝑓 (𝑥) is even, it can be represented by the Fourier cosine series given by
(8.8) and (8.9). The coefficient 𝑎0 is
2 𝜋
∫︁
𝑎0 = 𝑓 (𝑥)𝑑𝑥
𝜋 0
2 𝜋
∫︁ (︂ )︂
2𝑥
= 1− 𝑑𝑥
𝜋 0 𝜋
]︂𝜋
𝑥2
[︂
2
= 𝑥−
𝜋 𝜋 0
= 0.
The coefficients for 𝑛 > 0 are

2 𝜋
∫︁
𝑎𝑛 = 𝑓 (𝑥) cos (𝑛𝑥)𝑑𝑥
𝜋 0
2 𝜋
∫︁ (︂ )︂
2𝑥
= 1− cos (𝑛𝑥)𝑑𝑥
𝜋 0 𝜋
∫︁ 𝜋 ∫︁ 𝜋
2 4
= cos (𝑛𝑥)𝑑𝑥 − 2 𝑥 cos (𝑛𝑥)𝑑𝑥
𝜋 0 𝜋
{︂[︁ 0 ]︁𝜋 1 ∫︁ 𝜋 }︂
2 ]︀𝜋 4 𝑥
= sin(𝑛𝑥) 0 − 2 sin (𝑛𝑥) − sin (𝑛𝑥)𝑑𝑥
𝑛𝜋 𝜋 𝑛 0 𝑛 0
∫︁ 𝜋
4
= sin (𝑛𝑥)𝑑𝑥
𝑛𝜋 2 0
4 ]︀𝜋
= − 2 2 cos (𝑛𝑥) 0
𝑛 𝜋
4 (︀ )︀
= 2 2 1 − cos (𝑛𝜋) .
𝑛 𝜋
Since {︂
−1, if 𝑛 odd;
cos (𝑛𝜋) =
1, if 𝑛 even;
we have
8/(𝑛2 𝜋 2 ), if 𝑛 odd;
{︂
𝑎𝑛 =
0, if 𝑛 even.
The Fourier cosine series for the triangle function is therefore given by
(︂ )︂
8 cos 3𝑥 cos 5𝑥
𝑓 (𝑥) = 2 cos 𝑥 + + + ... .
𝜋 32 52
Convergence of this series is rapid. As an interesting aside, evaluation of this
series at 𝑥 = 0, using 𝑓 (0) = 1, yields an infinite series for 𝜋 2 /8:
𝜋2 1 1
= 1 + 2 + 2 + ....
8 3 5
With Fourier series now included in our applied mathematics toolbox, we are
ready to solve the diffusion and wave equations in bounded domains.
8.5 Solution of the diffusion equation

8.5.1 Homogeneous boundary conditions
view lecture
view lecture
We consider one dimensional diffusion in a pipe of length 𝐿, and solve the
diffusion equation for the concentration 𝑢(𝑥, 𝑡),
𝑢𝑡 = 𝐷𝑢𝑥𝑥 , 0 ≤ 𝑥 ≤ 𝐿, 𝑡 > 0. (8.12)
Both initial and boundary conditions are required for a unique solution. That
is, we assume the initial concentration distribution in the pipe is given by
𝑢(𝑥, 0) = 𝑓 (𝑥), 0 ≤ 𝑥 ≤ 𝐿. (8.13)
8.5. SOLUTION OF THE DIFFUSION EQUATION 105
Furthermore, we assume that boundary conditions are given at the ends of

the pipes. When the concentration value is specified at the boundaries, the
boundary conditions are called Dirichlet boundary conditions. As the simplest
example, we assume here homogeneous Dirichlet boundary conditions, that is
zero concentration of dye at the ends of the pipe, which could occur if the ends
of the pipe open up into large reservoirs of clear solution,
𝑢(0, 𝑡) = 0, 𝑢(𝐿, 𝑡) = 0, 𝑡 > 0. (8.14)
We will later also discuss inhomogeneous Dirichlet boundary conditions and

homogeneous Neumann boundary conditions, for which the derivative of the
concentration is specified to be zero at the boundaries. Note that if 𝑓 (𝑥) is
identically zero, then the trivial solution 𝑢(𝑥, 𝑡) = 0 satisfies the differential
equation and the initial and boundary conditions and is therefore the unique
solution of the problem. In what follows, we will assume that 𝑓 (𝑥) is not iden-
tically zero so that we need to find a solution different than the trivial solution.
The solution method we use is called separation of variables. We assume
that 𝑢(𝑥, 𝑡) can be written as a product of two other functions, one dependent
only on position 𝑥 and the other dependent only on time 𝑡. That is, we make
the ansatz
𝑢(𝑥, 𝑡) = 𝑋(𝑥)𝑇 (𝑡). (8.15)
Whether this ansatz will succeed depends on whether the solution indeed has
this form. Substituting (8.15) into (8.12), we obtain
𝑋𝑇 ′ = 𝐷𝑋 ′′ 𝑇,
which we rewrite by separating the 𝑥 and 𝑡 dependence to opposite sides of the

equation:
𝑋 ′′ 1 𝑇′
= .
𝑋 𝐷 𝑇
The left hand side of this equation is independent of 𝑡 and the right hand side is
independent of 𝑥. Both sides of this equation are therefore independent of both
𝑥 and 𝑡 and equal to a constant. Introducing −𝜆 as the separation constant, we
have
𝑋 ′′ 1 𝑇′
= = −𝜆,
𝑋 𝐷 𝑇
and we obtain the two ordinary differential equations
𝑋 ′′ + 𝜆𝑋 = 0, 𝑇 ′ + 𝜆𝐷𝑇 = 0. (8.16)
Because of the boundary conditions, we must first consider the equation for
𝑋(𝑥). To solve, we need to determine the boundary conditions at 𝑥 = 0 and
𝑥 = 𝐿. Now, from (8.14) and (8.15),
𝑢(0, 𝑡) = 𝑋(0)𝑇 (𝑡) = 0, 𝑡 > 0.
Since 𝑇 (𝑡) is not identically zero for all 𝑡 (which would result in the trivial
solution for 𝑢), we must have 𝑋(0) = 0. Similarly, the boundary condition at
𝑥 = 𝐿 requires 𝑋(𝐿) = 0. We therefore consider the two-point boundary value
problem
𝑋 ′′ + 𝜆𝑋 = 0, 𝑋(0) = 𝑋(𝐿) = 0. (8.17)
The equation given by (8.17) is called an ode eigenvalue problem. The allowed
values of 𝜆 and the corresponding functions 𝑋(𝑥) are called the eigenvalues and
eigenfunctions of the differential equation. Since the form of the general solution
of the ode depends on the sign of 𝜆, we consider in turn the cases 𝜆 > 0, 𝜆 < 0
and 𝜆 = 0. For 𝜆 > 0, we write 𝜆 = 𝜇2 and determine the general solution of
𝑋 ′′ + 𝜇2 𝑋 = 0
to be
𝑋(𝑥) = 𝐴 cos 𝜇𝑥 + 𝐵 sin 𝜇𝑥.
Applying the boundary condition at 𝑥 = 0, we find 𝐴 = 0. The boundary
condition at 𝑥 = 𝐿 then yields
𝐵 sin 𝜇𝐿 = 0.
The solution 𝐵 = 0 results in the trivial solution for 𝑢 and can be ruled out.
Therefore, we must have
sin 𝜇𝐿 = 0,
which is an equation for the eigenvalue 𝜇. The solutions are
𝜇 = 𝑛𝜋/𝐿,
where 𝑛 is an integer. We have thus determined the eigenvalues 𝜆 = 𝜇2 > 0 to

be
2
𝜆𝑛 = (𝑛𝜋/𝐿) , 𝑛 = 1, 2, 3, . . . , (8.18)
with corresponding eigenfunctions
𝑋𝑛 = sin (𝑛𝜋𝑥/𝐿). (8.19)
For 𝜆 < 0, we write 𝜆 = −𝜇2 and determine the general solution of
𝑋 ′′ − 𝜇2 𝑋 = 0
to be
𝑋(𝑥) = 𝐴 cosh 𝜇𝑥 + 𝐵 sinh 𝜇𝑥,
where we have previously introduced the hyperbolic sine and cosine functions
in S3.4.1. Applying the boundary condition at 𝑥 = 0, we find 𝐴 = 0. The
boundary condition at 𝑥 = 𝐿 then yields
𝐵 sinh 𝜇𝐿 = 0,
which for 𝜇 ̸= 0 has only the solution 𝐵 = 0. Therefore, there is no nontrivial

solution for 𝑢 with 𝜆 < 0. Finally, for 𝜆 = 0, we have
𝑋 ′′ = 0,
with general solution

𝑋(𝑥) = 𝐴 + 𝐵𝑥.
The boundary condition at 𝑥 = 0 and 𝑥 = 𝐿 yields 𝐴 = 𝐵 = 0 so again there is
no nontrivial solution for 𝑢 with 𝜆 = 0.
We now turn to the equation for 𝑇 (𝑡). The equation corresponding to the
eigenvalue 𝜆𝑛 , using (8.18), is given by
𝑇 ′ + 𝑛2 𝜋 2 𝐷/𝐿2 𝑇 = 0,
(︀ )︀
which has solution proportional to

2
𝜋 2 𝐷𝑡/𝐿2
𝑇𝑛 = 𝑒−𝑛 . (8.20)
Therefore, multiplying the solutions given by (8.19) and (8.20), we conclude

that the functions
2
𝑢𝑛 (𝑥, 𝑡) = sin (𝑛𝜋𝑥/𝐿)𝑒−𝑛 (8.21)
satisfy the pde given by (8.12) and the boundary conditions given by (8.14) for
every positive integer 𝑛.
The principle of linear superposition for homogeneous linear differential
equations then states that the general solution to (8.12) and (8.14) is given
by
∞
∑︁
𝑢(𝑥, 𝑡) = 𝑏𝑛 𝑢𝑛 (𝑥, 𝑡)
𝑛=1
∞
(8.22)
−𝑛2 𝜋 2 𝐷𝑡/𝐿2
∑︁
= 𝑏𝑛 sin (𝑛𝜋𝑥/𝐿)𝑒 .
𝑛=1
The final solution step is to satisfy the initial conditions given by (8.13). At
𝑡 = 0, we have
∞
∑︁
𝑓 (𝑥) = 𝑏𝑛 sin (𝑛𝜋𝑥/𝐿). (8.23)
𝑛=1
We immediately recognize (8.23) as a Fourier sine series (8.11) for an odd func-
tion 𝑓 (𝑥) with period 2𝐿. Equation (8.23) is a periodic extension of our original
𝑓 (𝑥) defined on 0 ≤ 𝑥 ≤ 𝐿, and is an odd function because of the boundary
condition 𝑓 (0) = 0. From our solution for the coefficients of a Fourier sine series
(8.10), we determine
2 𝐿
∫︁
𝑛𝜋𝑥
𝑏𝑛 = 𝑓 (𝑥) sin 𝑑𝑥. (8.24)
𝐿 0 𝐿
Thus the solution to the diffusion equation with homogeneous Dirichlet bound-
ary conditions defined by (8.12), (8.13) and (8.14) is given by (8.22) with the
𝑏𝑛 coefficients computed from (8.24).
Example: Determine the concentration of a dye in a pipe of length 𝐿,

where the dye has unit mass and is initially concentrated at the center
of the pipe, and the ends of the pipe are held at zero concentration
The governing equation for concentration is the diffusion equation. We model

the initial concentration of the dye by a delta-function centered at 𝑥 = 𝐿/2,
that is, 𝑢(𝑥, 0) = 𝑓 (𝑥) = 𝛿(𝑥 − 𝐿/2). Therefore, from (8.24),
2 𝐿
∫︁
𝐿 𝑛𝜋𝑥
𝑏𝑛 = 𝛿(𝑥 − ) sin 𝑑𝑥
𝐿 0 2 𝐿
2
= sin (𝑛𝜋/2)
𝐿
⎧
⎨ 2/𝐿 if 𝑛 = 1, 5, 9, . . . ;
= −2/𝐿 if 𝑛 = 3, 7, 11, . . . ;
0 if 𝑛 = 2, 4, 6, . . . .
⎩
With 𝑏𝑛 determined, the solution for 𝑢(𝑥, 𝑡) given by (8.22) can be written as
∞ (︂ )︂
2 ∑︁ 𝑛 (2𝑛 + 1)𝜋𝑥 −(2𝑛+1)2 𝜋2 𝐷𝑡/𝐿2
𝑢(𝑥, 𝑡) = (−1) sin 𝑒 .
𝐿 𝑛=0 𝐿
When 𝑡 ≫ 𝐿2 /𝐷, the leading-order term in the series is a good approximation

and is given by
2 2 2
𝑢(𝑥, 𝑡) ≈ sin (𝜋𝑥/𝐿)𝑒−𝜋 𝐷𝑡/𝐿 .
𝐿
8.5.2 Inhomogeneous boundary conditions

view lecture
Consider a diffusion problem where one end of the pipe has dye of concentration
held constant at 𝐶1 and the other held constant at 𝐶2 , which could occur if the
ends of the pipe had large reservoirs of fluid with different concentrations of
dye. With 𝑢(𝑥, 𝑡) the concentration of dye, the boundary conditions are given
by
𝑢(0, 𝑡) = 𝐶1 , 𝑢(𝐿, 𝑡) = 𝐶2 , 𝑡 > 0.
The concentration 𝑢(𝑥, 𝑡) satisfies the diffusion equation with diffusivity 𝐷:
𝑢𝑡 = 𝐷𝑢𝑥𝑥 .
If we try to solve this problem directly using separation of variables, we will

run into trouble. Applying the inhomogeneous boundary condition at 𝑥 = 0
directly to the ansatz 𝑢(𝑥, 𝑡) = 𝑋(𝑥)𝑇 (𝑡) results in
𝑢(0, 𝑡) = 𝑋(0)𝑇 (𝑡) = 𝐶1 ;
so that
𝑋(0) = 𝐶1 /𝑇 (𝑡).
However, our separation of variables ansatz assumes 𝑋(𝑥) to be independent of
𝑡! We therefore say that inhomogeneous boundary conditions are not separable.
The proper way to solve a problem with inhomogeneous boundary conditions
is to transform it into another problem with homogeneous boundary conditions.
As 𝑡 → ∞, we assume that a stationary concentration distribution 𝑣(𝑥) will
attain, independent of 𝑡. Since 𝑣(𝑥) must satisfy the diffusion equation, we have
𝑣 ′′ (𝑥) = 0, 0 ≤ 𝑥 ≤ 𝐿,
with general solution

𝑣(𝑥) = 𝐴 + 𝐵𝑥.
Since 𝑣(𝑥) must satisfy the same boundary conditions of 𝑢(𝑥, 𝑡), we have 𝑣(0) =
𝐶1 and 𝑣(𝐿) = 𝐶2 , and we determine 𝐴 = 𝐶1 and 𝐵 = (𝐶2 − 𝐶1 )/𝐿.
We now express 𝑢(𝑥, 𝑡) as the sum of the known asymptotic stationary con-
centration distribution 𝑣(𝑥) and an unknown transient concentration distribu-
tion 𝑤(𝑥, 𝑡):
𝑢(𝑥, 𝑡) = 𝑣(𝑥) + 𝑤(𝑥, 𝑡).
Substituting into the diffusion equation, we obtain
𝜕 𝜕2
(𝑣(𝑥) + 𝑤(𝑥, 𝑡)) = 𝐷 2 (𝑣(𝑥) + 𝑤(𝑥, 𝑡))
𝜕𝑡 𝜕𝑥
or
𝑤𝑡 = 𝐷𝑤𝑥𝑥 ,
since 𝑣𝑡 = 0 and 𝑣𝑥𝑥 = 0. The boundary conditions satisfied by 𝑤 are
𝑤(0, 𝑡) = 𝑢(0, 𝑡) − 𝑣(0) = 0,

𝑤(𝐿, 𝑡) = 𝑢(𝐿, 𝑡) − 𝑣(𝐿) = 0,
so that 𝑤 is observed to satisfy homogeneous boundary conditions. If the initial

conditions are given by 𝑢(𝑥, 0) = 𝑓 (𝑥), then the initial conditions for 𝑤 are
𝑤(𝑥, 0) = 𝑢(𝑥, 0) − 𝑣(𝑥)

= 𝑓 (𝑥) − 𝑣(𝑥).
The resulting equations may then be solved for 𝑤(𝑥, 𝑡) using the technique for
homogeneous boundary conditions, and 𝑢(𝑥, 𝑡) subsequently determined.
8.5.3 Pipe with closed ends

view lecture
view lecture
There is no diffusion of dye through the ends of a sealed pipe. Accordingly,
the mass flux of dye through the pipe ends, given by (8.1), is zero so that the
boundary conditions on the dye concentration 𝑢(𝑥, 𝑡) becomes
𝑢𝑥 (0, 𝑡) = 0, 𝑢𝑥 (𝐿, 𝑡) = 0, 𝑡 > 0, (8.25)
which are known as homogeneous Neumann boundary conditions. Again, we

apply the method of separation of variables and as before, we obtain the two
ordinary differential equations given by (8.16). Considering first the equation
for 𝑋(𝑥), the appropriate boundary conditions are now on the first derivative
of 𝑋(𝑥), and we must solve
𝑋 ′′ + 𝜆𝑋 = 0, 𝑋 ′ (0) = 𝑋 ′ (𝐿) = 0. (8.26)
Again, we consider in turn the cases 𝜆 > 0, 𝜆 < 0 and 𝜆 = 0. For 𝜆 > 0, we
write 𝜆 = 𝜇2 and determine the general solution of (8.26) to be
𝑋(𝑥) = 𝐴 cos 𝜇𝑥 + 𝐵 sin 𝜇𝑥,
so that taking the derivative
𝑋 ′ (𝑥) = −𝜇𝐴 sin 𝜇𝑥 + 𝜇𝐵 cos 𝜇𝑥.

Applying the boundary condition 𝑋 ′ (0) = 0, we find 𝐵 = 0. The boundary

condition at 𝑥 = 𝐿 then yields
−𝜇𝐴 sin 𝜇𝐿 = 0.
The solution 𝐴 = 0 results in the trivial solution for 𝑢 and can be ruled out.
Therefore, we must have
sin 𝜇𝐿 = 0,
with solutions
𝜇 = 𝑛𝜋/𝐿,
where 𝑛 is an integer. We have thus determined the eigenvalues 𝜆 = 𝜇2 > 0 to
be
2
𝜆𝑛 = (𝑛𝜋/𝐿) , 𝑛 = 1, 2, 3, . . . , (8.27)
𝑋𝑛 = cos (𝑛𝜋𝑥/𝐿). (8.28)
For 𝜆 < 0, we write 𝜆 = −𝜇2 and determine the general solution of (8.26) to be
𝑋(𝑥) = 𝐴 cosh 𝜇𝑥 + 𝐵 sinh 𝜇𝑥,
𝑋 ′ (𝑥) = 𝜇𝐴 sinh 𝜇𝑥 + 𝜇𝐵 cosh 𝜇𝑥.
Applying the boundary condition 𝑋 ′ (0) = 0 yields 𝐵 = 0. The boundary

condition 𝑋 ′ (𝐿) = 0 then yields
𝜇𝐴 sinh 𝜇𝐿 = 0,
which for 𝜇 ̸= 0 has only the solution 𝐴 = 0. Therefore, there is no nontrivial

solution for 𝑢 with 𝜆 < 0. Finally, for 𝜆 = 0, the general solution of (8.26) is
𝑋(𝑥) = 𝐴 + 𝐵𝑥,

𝑋 ′ (𝑥) = 𝐵.
The boundary condition 𝑋 ′ (0) = 0 yields 𝐵 = 0; 𝑋 ′ (𝐿) = 0 is then trivially
satisfied. Therefore, we have an additional eigenvalue and eigenfunction given
by
𝜆0 = 0, 𝑋0 (𝑥) = 1,
which can be seen as extending the formula obtained for eigenvalues and eigen-
vectors for positive 𝜆 given by (8.27) and (8.28) to 𝑛 = 0.
We now turn to the equation for 𝑇 (𝑡). The equation corresponding to eigen-
value 𝜆𝑛 , using (8.27), is given by
𝑇 ′ + 𝑛2 𝜋 2 𝐷/𝐿2 𝑇 = 0,
(︀ )︀
which has solution proportional to

2
𝑇𝑛 = 𝑒−𝑛 , (8.29)
valid for 𝑛 = 0, 1, 2, . . . . Therefore, multiplying the solutions given by (8.28)

and (8.29), we conclude that the functions
2
𝑢𝑛 (𝑥, 𝑡) = cos (𝑛𝜋𝑥/𝐿)𝑒−𝑛 (8.30)
satisfy the pde given by (8.12) and the boundary conditions given by (8.25) for
every nonnegative integer 𝑛.
The principle of linear superposition then yields the general solution as
∞
∑︁
𝑢(𝑥, 𝑡) = 𝑐𝑛 𝑢𝑛 (𝑥, 𝑡)
𝑛=0
∞
(8.31)
𝑎0 ∑︁ 2 2 2
= + 𝑎𝑛 cos (𝑛𝜋𝑥/𝐿)𝑒−𝑛 𝜋 𝐷𝑡/𝐿 ,
2 𝑛=1
where we have redefined the constants so that 𝑐0 = 𝑎0 /2 and 𝑐𝑛 = 𝑎𝑛 , 𝑛 =

1, 2, 3, . . . . The final solution step is to satisfy the initial conditions given by
(8.13). At 𝑡 = 0, we have
∞
𝑎0 ∑︁
𝑓 (𝑥) = + 𝑎𝑛 cos (𝑛𝜋𝑥/𝐿), (8.32)
2 𝑛=1
which we recognize as a Fourier cosine series (8.9) for an even function 𝑓 (𝑥)
with period 2𝐿. We have obtained a cosine series for the periodic extension of
𝑓 (𝑥) because of the boundary condition 𝑓 ′ (0) = 0, which is satisfied by an even
function. From our solution (8.8) for the coefficients of a Fourier cosine series,
we determine
2 𝐿
∫︁
𝑛𝜋𝑥
𝑎𝑛 = 𝑓 (𝑥) cos 𝑑𝑥. (8.33)
𝐿 0 𝐿
Thus the solution to the diffusion equation with homogenous Neumann bound-
ary conditions defined by (8.12), (8.13) and (8.25) is given by (8.31) with the
coefficients computed from (8.33).
Example: Determine the concentration of a dye in a pipe of length

𝐿, where the dye has unit mass and is initially concentrated at the
center of the pipe, and the ends of the pipe are sealed
Again we model the initial concentration of the dye by a delta-function centered
at 𝑥 = 𝐿/2. From (8.33),
2 𝐿
∫︁
𝐿 𝑛𝜋𝑥
𝑎𝑛 = 𝛿(𝑥 − ) cos 𝑑𝑥
𝐿 0 2 𝐿
2
= cos (𝑛𝜋/2)
⎧𝐿
⎨ 2/𝐿 if 𝑛 = 0, 4, 8, . . . ;
= −2/𝐿 if 𝑛 = 2, 6, 10, . . . ;
0 if 𝑛 = 1, 3, 5, . . . .
⎩
The first two terms in the series for 𝑢(𝑥, 𝑡) are given by
2 [︁ 2 2
]︁
𝑢(𝑥, 𝑡) = 1/2 − cos (2𝜋𝑥/𝐿)𝑒−4𝜋 𝐷𝑡/𝐿 + . . . .
𝐿
Notice that as 𝑡 → ∞, 𝑢(𝑥, 𝑡) → 1/𝐿: the dye mass is conserved in the pipe
(since the pipe ends are sealed) and eventually diffused uniformly throughout
the pipe of length 𝐿.
8.6 Solution of the wave equation

view lecture
8.6.1 Plucked string
view lecture
We assume an elastic string with fixed ends is plucked like a guitar string. The
governing equation for 𝑢(𝑥, 𝑡), the position of the string from its equilibrium
position, is the wave equation
𝑢𝑡𝑡 = 𝑐2 𝑢𝑥𝑥 , (8.34)
with 𝑐2 = 𝑇 /𝜌 and with boundary conditions at the string ends located at 𝑥 = 0

and 𝐿 given by
𝑢(0, 𝑡) = 0, 𝑢(𝐿, 𝑡) = 0. (8.35)
Since the wave equation is second-order in time, initial conditions are required
for both the displacement of the string due to the plucking and the initial velocity
of the displacement. We assume
𝑢(𝑥, 0) = 𝑓 (𝑥), 𝑢𝑡 (𝑥, 0) = 0, 0 ≤ 𝑥 ≤ 𝐿. (8.36)
Again we use the method of separation of variables and try the ansatz
𝑢(𝑥, 𝑡) = 𝑋(𝑥)𝑇 (𝑡). (8.37)
Substitution of our ansatz (8.37) into the wave equation (8.34) and separating
variables results in
𝑋 ′′ 1 𝑇 ′′
= 2 = −𝜆,
𝑋 𝑐 𝑇
yielding the two ordinary differential equations
𝑋 ′′ + 𝜆𝑋 = 0, 𝑇 ′′ + 𝜆𝑐2 𝑇 = 0. (8.38)
We solve first the equation for 𝑋(𝑥). The appropriate boundary conditions for
𝑋 are given by
𝑋(0) = 0, 𝑋(𝐿) = 0, (8.39)
and we have solved this equation for 𝑋(𝑥) previously in S8.5.1 (see (8.17)).
A nontrivial solution exists only when 𝜆 > 0, and our previously determined
solution was
2
𝜆𝑛 = (𝑛𝜋/𝐿) , 𝑛 = 1, 2, 3, . . . , (8.40)
𝑋𝑛 = sin (𝑛𝜋𝑥/𝐿). (8.41)
With 𝜆𝑛 specified, the 𝑇 equation then becomes

𝑛 2 𝜋 2 𝑐2
𝑇𝑛′′ + 𝑇𝑛 = 0,
𝐿2
8.6. SOLUTION OF THE WAVE EQUATION 113
with general solution given by

𝑛𝜋𝑐𝑡 𝑛𝜋𝑐𝑡
𝑇𝑛 (𝑡) = 𝐴 cos + 𝐵 sin . (8.42)
𝐿 𝐿
The second of the initial conditions given by (8.36) implies
𝑢𝑡 (𝑥, 0) = 𝑋(𝑥)𝑇 ′ (0) = 0,
which can be satisfied only if 𝑇 ′ (0) = 0. Applying this boundary condition to

(8.42), we find 𝐵 = 0. Combining our solution for 𝑋𝑛 (𝑥), (8.41), and 𝑇𝑛 (𝑡), we
have determined that
𝑛𝜋𝑥 𝑛𝜋𝑐𝑡
𝑢𝑛 (𝑥, 𝑡) = sin cos , 𝑛 = 1, 2, 3, . . .
𝐿 𝐿
satisfies the wave equation, the boundary conditions at the string ends, and the
assumption of zero initial string velocity. Linear superposition of these solutions
results in the general solution for 𝑢(𝑥, 𝑡) of the form
∞
∑︁ 𝑛𝜋𝑥 𝑛𝜋𝑐𝑡
𝑢(𝑥, 𝑡) = 𝑏𝑛 sin cos . (8.43)
𝑛=1
𝐿 𝐿
The remaining condition to satisfy is the initial displacement of the string, the
first equation of (8.36). We have
∞
∑︁
𝑓 (𝑥) = 𝑏𝑛 sin (𝑛𝜋𝑥/𝐿),
𝑛=1
which is observed to be a Fourier sine series (8.11) for an odd function with
period 2𝐿. Therefore, the coefficients 𝑏𝑛 are given by (8.10),
2 𝐿
∫︁
𝑛𝜋𝑥
𝑏𝑛 = 𝑓 (𝑥) sin 𝑑𝑥, 𝑛 = 1, 2, 3, . . . . (8.44)
𝐿 0 𝐿
Our solution to the wave equation with plucked string is thus given by (8.43)
and (8.44). Notice that the solution is time periodic with period 2𝐿/𝑐. The
corresponding fundamental frequency is the reciprocal of the period and is given
by 𝑓 = 𝑐/2𝐿. From our derivation of the wave equation in S8.2, the velocity 𝑐
is related to the density of the string 𝜌 and tension of the string 𝑇 by 𝑐2 = 𝑇 /𝜌.
Therefore, the fundamental frequency (pitch) of our “guitar string” increases (is
raised) with increasing tension, decreasing string density, and decreasing string
length. Indeed, these are exactly the parameters used to construct, tune and
play a guitar.
The wave nature of our solution and the physical significance of the velocity
𝑐 can be made more transparent if we make use of the trigonometric identity
1 (︀ )︀
sin 𝑥 cos 𝑦 = sin (𝑥 + 𝑦) + sin (𝑥 − 𝑦) .
2
With this identity, our solution (8.43) can be rewritten as
∞ (︂ )︂
1 ∑︁ 𝑛𝜋(𝑥 + 𝑐𝑡) 𝑛𝜋(𝑥 − 𝑐𝑡)
𝑢(𝑥, 𝑡) = 𝑏𝑛 sin + sin . (8.45)
2 𝑛=1 𝐿 𝐿
The first and second sine functions can be interpreted as a traveling wave moving
to the left or the right with velocity 𝑐. This can be seen by incrementing time,
𝑡 → 𝑡 + 𝛿, and observing that the value of the first sine function is unchanged
provided the position is shifted by 𝑥 → 𝑥 − 𝑐𝛿, and the second sine function is
unchanged provided 𝑥 → 𝑥 + 𝑐𝛿. Two waves travelling in opposite directions
with equal amplitude results in a standing wave.
8.6.2 Hammered string

view lecture
In contrast to a guitar string that is plucked, a piano string is hammered. The
appropriate initial conditions for a piano string would be
𝑢(𝑥, 0) = 0, 𝑢𝑡 (𝑥, 0) = 𝑔(𝑥), 0 ≤ 𝑥 ≤ 𝐿. (8.46)
Our solution proceeds as previously, except that now the homogeneous initial
condition on 𝑇 (𝑡) is 𝑇 (0) = 0, so that 𝐴 = 0 in (8.42). The wave equation
solution is therefore
∞
∑︁ 𝑛𝜋𝑥 𝑛𝜋𝑐𝑡
𝑢(𝑥, 𝑡) = 𝑏𝑛 sin sin . (8.47)
𝑛=1
𝐿 𝐿
Imposition of initial conditions then yields

∞
𝜋𝑐 ∑︁ 𝑛𝜋𝑥
𝑔(𝑥) = 𝑛𝑏𝑛 sin .
𝐿 𝑛=1 𝐿
The coefficient of the Fourier sine series for 𝑔(𝑥) is seen to be 𝑛𝜋𝑐𝑏𝑛 /𝐿, and we
have
2 𝐿
∫︁
𝑛𝜋𝑐𝑏𝑛 𝑛𝜋𝑥
= 𝑔(𝑥) sin 𝑑𝑥,
𝐿 𝐿 0 𝐿
or ∫︁ 𝐿
2 𝑛𝜋𝑥
𝑏𝑛 = 𝑔(𝑥) sin 𝑑𝑥.
𝑛𝜋𝑐 0 𝐿
8.6.3 General initial conditions

view lecture
If the initial conditions on 𝑢(𝑥, 𝑡) are generalized to
𝑢(𝑥, 0) = 𝑓 (𝑥), 𝑢𝑡 (𝑥, 0) = 𝑔(𝑥), 0 ≤ 𝑥 ≤ 𝐿, (8.48)
then the solution to the wave equation can be determined using the principle of
linear superposition. Suppose 𝑣(𝑥, 𝑡) is the solution to the wave equation with
initial condition (8.36) and 𝑤(𝑥, 𝑡) is the solution to the wave equation with
initial conditions (8.46). Then we have
𝑢(𝑥, 𝑡) = 𝑣(𝑥, 𝑡) + 𝑤(𝑥, 𝑡),
since 𝑢(𝑥, 𝑡) satisfies the wave equation, the boundary conditions, and the initial
conditions given by (8.48).
8.7. THE LAPLACE EQUATION 115
Figure 8.4: Dirichlet problem for the Laplace equation in a rectangle.
8.7 The Laplace equation

The diffusion equation in two spatial dimensions is
𝑢𝑡 = 𝐷(𝑢𝑥𝑥 + 𝑢𝑦𝑦 ).
The steady-state solution, approached asymptotically in time, has 𝑢𝑡 = 0 so
that the steady-state solution 𝑢 = 𝑢(𝑥, 𝑦) satisfies the two-dimensional Laplace
equation
𝑢𝑥𝑥 + 𝑢𝑦𝑦 = 0. (8.49)
We will consider the mathematical problem of solving the two dimensional
Laplace equation inside a rectangular or a circular boundary. The value of
𝑢(𝑥, 𝑦) will be specified on the boundaries, defining this problem to be of Dirich-
let type.
8.7.1 Dirichlet problem for a rectangle

view lecture
We consider the Laplace equation (8.49) for the interior of a rectangle 0 < 𝑥 < 𝑎,
0 < 𝑦 < 𝑏, (see Fig. 8.4), with boundary conditions
𝑢(𝑥, 0) = 0, 𝑢(𝑥, 𝑏) = 0, 0 < 𝑥 < 𝑎;
𝑢(0, 𝑦) = 0, 𝑢(𝑎, 𝑦) = 𝑓 (𝑦), 0 ≤ 𝑦 ≤ 𝑏.
More general boundary conditions can be solved by linear superposition of so-
lutions.
We take our usual ansatz
𝑢(𝑥, 𝑦) = 𝑋(𝑥)𝑌 (𝑦),
and find after substitution into (8.49),
𝑋 ′′ 𝑌 ′′
=− = 𝜆,
𝑋 𝑌
with 𝜆 the separation constant. We thus obtain the two ordinary differential
equations
𝑋 ′′ − 𝜆𝑋 = 0, 𝑌 ′′ + 𝜆𝑌 = 0.
The homogeneous boundary conditions are 𝑋(0) = 0, 𝑌 (0) = 0 and 𝑌 (𝑏) = 0.
We have already solved the equation for 𝑌 (𝑦) in S8.5.1, and the solution yields
the eigenvalues
(︁ 𝑛𝜋 )︁2
𝜆𝑛 = , 𝑛 = 1, 2, 3, . . . ,
𝑏
𝑛𝜋𝑦
𝑌𝑛 (𝑦) = sin .
𝑏
The remaining 𝑋 equation and homogeneous boundary condition is therefore
𝑛2 𝜋 2
𝑋 ′′ − 𝑋 = 0, 𝑋(0) = 0,
𝑏2
and the solution is the hyperbolic sine function
𝑛𝜋𝑥
𝑋𝑛 (𝑥) = sinh ,
𝑏
times a constant. Writing 𝑢𝑛 = 𝑋𝑛 𝑌𝑛 , multiplying by a constant and summing
over 𝑛, yields the general solution
∞
∑︁ 𝑛𝜋𝑥 𝑛𝜋𝑦
𝑢(𝑥, 𝑡) = 𝑐𝑛 sinh sin .
𝑛=0
𝑏 𝑏
The remaining inhomogeneous boundary condition 𝑢(𝑎, 𝑦) = 𝑓 (𝑦) results in

∞
∑︁ 𝑛𝜋𝑎 𝑛𝜋𝑦
𝑓 (𝑦) = 𝑐𝑛 sinh sin ,
𝑛=0
𝑏 𝑏
which we recognize as a Fourier sine series for an odd function with period 2𝑏,
and coefficient 𝑐𝑛 sinh (𝑛𝜋𝑎/𝑏). The solution for the coefficient is given by
∫︁ 𝑏
𝑛𝜋𝑎 2 𝑛𝜋𝑦
𝑐𝑛 sinh = 𝑓 (𝑦) sin 𝑑𝑥.
𝑏 𝑏 0 𝑏
8.7.2 Dirichlet problem for a circle

view lecture
view lecture
The Laplace equation is commonly written symbolically as
∇2 𝑢 = 0, (8.50)
where ∇2 is called the Laplacian, sometimes denoted as Δ. The Laplacian can

be written in various coordinate systems, and the choice of coordinate systems
usually depends on the geometry of the boundaries. Indeed, the Laplace equa-
tion is known to be separable in 13 different coordinate systems! We have solved
the Laplace equation in two dimensions, with boundary conditions specified on
a rectangle, using
𝜕2 𝜕2
∇2 = + .
𝜕𝑥2 𝜕𝑦 2
Here we consider boundary conditions specified on a circle, and write the Lapla-
cian in polar coordinates by changing variables from cartesian coordinates. Polar
coordinates is defined by the transformation (𝑟, 𝜃) → (𝑥, 𝑦):
𝑥 = 𝑟 cos 𝜃, 𝑦 = 𝑟 sin 𝜃; (8.51)
and the chain rule gives for the partial derivatives

𝜕𝑢 𝜕𝑢 𝜕𝑥 𝜕𝑢 𝜕𝑦 𝜕𝑢 𝜕𝑢 𝜕𝑥 𝜕𝑢 𝜕𝑦
= + , = + . (8.52)
𝜕𝑟 𝜕𝑥 𝜕𝑟 𝜕𝑦 𝜕𝑟 𝜕𝜃 𝜕𝑥 𝜕𝜃 𝜕𝑦 𝜕𝜃
After taking the partial derivatives of 𝑥 and 𝑦 using (8.51), we can write the
transformation (8.52) in matrix form as
(︂ )︂ (︂ )︂ (︂ )︂
𝜕𝑢/𝜕𝑟 cos 𝜃 sin 𝜃 𝜕𝑢/𝜕𝑥
= . (8.53)
𝜕𝑢/𝜕𝜃 −𝑟 sin 𝜃 𝑟 cos 𝜃 𝜕𝑢/𝜕𝑦
Inversion of (8.53) can be determined from the following result, commonly

proved in a linear algebra class. If
(︂ )︂
𝑎 𝑏
A= , det A ̸= 0,
𝑐 𝑑
then (︂ )︂
1 𝑑 −𝑏
A−1 = .
det A −𝑐 𝑎
Therefore, since the determinant of the 2 × 2 matrix in (8.53) is 𝑟, we have
(︂ )︂ (︂ )︂ (︂ )︂
𝜕𝑢/𝜕𝑥 cos 𝜃 − sin 𝜃/𝑟 𝜕𝑢/𝜕𝑟
= . (8.54)
𝜕𝑢/𝜕𝑦 sin 𝜃 cos 𝜃/𝑟 𝜕𝑢/𝜕𝜃
Rewriting (8.54) in operator form, we have
𝜕 𝜕 sin 𝜃 𝜕 𝜕 𝜕 cos 𝜃 𝜕
= cos 𝜃 − , = sin 𝜃 + . (8.55)
𝜕𝑥 𝜕𝑟 𝑟 𝜕𝜃 𝜕𝑦 𝜕𝑟 𝑟 𝜕𝜃
To find the Laplacian in polar coordinates with minimum algebra, we combine
(8.55) using complex variables as
(︂ )︂
𝜕 𝜕 𝑖𝜃 𝜕 𝑖 𝜕
+𝑖 =𝑒 + , (8.56)
𝜕𝑥 𝜕𝑦 𝜕𝑟 𝑟 𝜕𝜃
so that the Laplacian may be found by multiplying both sides of (8.56) by its
complex conjugate, taking care with the computation of the derivatives on the
right-hand-side:
𝜕2 𝜕2
(︂ )︂ (︂ )︂
𝑖𝜃 𝜕 𝑖 𝜕 −𝑖𝜃 𝜕 𝑖 𝜕
+ = 𝑒 + 𝑒 −
𝜕𝑥2 𝜕𝑦 2 𝜕𝑟 𝑟 𝜕𝜃 𝜕𝑟 𝑟 𝜕𝜃
𝜕2 1 𝜕 1 𝜕2
= 2+ + .
𝜕𝑟 𝑟 𝜕𝑟 𝑟2 𝜕𝜃2
We have therefore determined that the Laplacian in polar coordinates is given
by
𝜕2 1 𝜕 1 𝜕2
∇2 = 2 + + 2 2. (8.57)
𝜕𝑟 𝑟 𝜕𝑟 𝑟 𝜕𝜃
We now consider the solution of the Laplace equation in a circle with radius
𝑟 < 𝑎 subject to the boundary condition
𝑢(𝑎, 𝜃) = 𝑓 (𝜃), 0 ≤ 𝜃 ≤ 2𝜋. (8.58)
An additional boundary condition due to the use of polar coordinates is that

𝑢(𝑟, 𝜃) is periodic in 𝜃 with period 2𝜋. Furthermore, we will also assume that
𝑢(𝑟, 𝜃) is finite within the circle.
The method of separation of variables takes as our ansatz
𝑢(𝑟, 𝜃) = 𝑅(𝑟)Θ(𝜃),
and substitution into the Laplace equation (8.50) using (8.57) yields
1 1
𝑅′′ Θ + 𝑅′ Θ + 2 𝑅Θ′′ = 0,
𝑟 𝑟
or
𝑅′′ 𝑅′ Θ′′
𝑟2 +𝑟 =− = 𝜆,
𝑅 𝑅 Θ
where 𝜆 is the separation constant. We thus obtain the two ordinary differential
equations
𝑟2 𝑅′′ + 𝑟𝑅′ − 𝜆𝑅 = 0, Θ′′ + 𝜆Θ = 0.
The Θ equation is solved assuming periodic boundary conditions with period
2𝜋. If 𝜆 < 0, then no periodic solution exists. If 𝜆 = 0, then Θ can be constant.
If 𝜆 = 𝜇2 > 0, then
Θ(𝜃) = 𝐴 cos 𝜇𝜃 + 𝐵 sin 𝜇𝜃,
and the requirement that Θ is periodic with period 2𝜋 forces 𝜇 to be an integer.
Therefore,
𝜆𝑛 = 𝑛2 , 𝑛 = 0, 1, 2, . . . ,
Θ𝑛 (𝜃) = 𝐴𝑛 cos 𝑛𝜃 + 𝐵𝑛 sin 𝑛𝜃.
The 𝑅 equation for each eigenvalue 𝜆𝑛 then becomes
𝑟2 𝑅′′ + 𝑟𝑅′ − 𝑛2 𝑅 = 0, (8.59)
which is an Euler equation. With the ansatz 𝑅 = 𝑟𝑠 , (8.59) reduces to the

algebraic equation 𝑠(𝑠 − 1) + 𝑠 − 𝑛2 = 0, or 𝑠2 = 𝑛2 . Therefore, 𝑠 = ±𝑛, and
there are two real solutions when 𝑛 > 0 and degenerate solutions when 𝑛 = 0.
When 𝑛 > 0, the solution for R(r) is
𝑅𝑛 (𝑟) = 𝐴𝑟𝑛 + 𝐵𝑟−𝑛 .
The requirement that 𝑢(𝑟, 𝜃) is finite in the circle forces 𝐵 = 0 since 𝑟−𝑛 becomes
unbounded as 𝑟 → 0. When 𝑛 = 0, the solution for 𝑅(𝑟) is
𝑅𝑛 (𝑟) = 𝐴 + 𝐵 ln 𝑟,
and again finite 𝑢 in the circle forces 𝐵 = 0. Therefore, the solution for 𝑛 =
0, 1, 2, . . . is 𝑅𝑛 = 𝑟𝑛 . Thus the general solution for 𝑢(𝑟, 𝜃) may be written as
∞
𝐴0 ∑︁ 𝑛
𝑢(𝑟, 𝜃) = + 𝑟 (𝐴𝑛 cos 𝑛𝜃 + 𝐵𝑛 sin 𝑛𝜃), (8.60)
2 𝑛=1
where we have separated out the 𝑛 = 0 solution to write our solution in a form
similar to the standard Fourier series given by (8.2). The remaining boundary
condition (8.58) specifies the values of 𝑢 on the circle of radius 𝑎, and imposition
of this boundary condition results in
∞
𝐴0 ∑︁ 𝑛
𝑓 (𝜃) = + 𝑎 (𝐴𝑛 cos 𝑛𝜃 + 𝐵𝑛 sin 𝑛𝜃). (8.61)
2 𝑛=1
Equation (8.61) is a Fourier series for the periodic function 𝑓 (𝜃) with period 2𝜋,
i.e., 𝐿 = 𝜋 in (8.2). The Fourier coefficients 𝑎𝑛 𝐴𝑛 and 𝑎𝑛 𝐵𝑛 are therefore given
by (8.6) and (8.7) to be
1 2𝜋
∫︁
𝑎𝑛 𝐴𝑛 = 𝑓 (𝜑) cos 𝑛𝜑𝑑𝜑, 𝑛 = 0, 1, 2, . . . ;
𝜋 0
1 2𝜋
∫︁
𝑎𝑛 𝐵𝑛 = 𝑓 (𝜑) sin 𝑛𝜑𝑑𝜑, 𝑛 = 1, 2, 3, . . . . (8.62)
𝜋 0
A remarkable fact is that the infinite series solution for 𝑢(𝑟, 𝜃) can be summed
explicitly. Substituting (8.62) into (8.60), we obtain
∞ (︁ )︁𝑛
∫︁ 2𝜋 [︃ ]︃
1 ∑︁ 𝑟
𝑢(𝑟, 𝜃) = 𝑑𝜑𝑓 (𝜑) 1 + 2 (cos 𝑛𝜃 cos 𝑛𝜑 + sin 𝑛𝜃 sin 𝑛𝜑)
2𝜋 0 𝑛=1
𝑎
∞ (︁ )︁𝑛
∫︁ 2𝜋 [︃ ]︃
1 ∑︁ 𝑟
= 𝑑𝜑𝑓 (𝜑) 1 + 2 cos 𝑛(𝜃 − 𝜑) .
2𝜋 0 𝑛=1
𝑎
We can sum the infinite series by writing 2∑︀cos 𝑛(𝜃 − 𝜑) = 𝑒𝑖𝑛(𝜃−𝜑) + 𝑒−𝑖𝑛(𝜃−𝜑) ,
∞
and using the sum of the geometric series 𝑛=1 𝑧 𝑛 = 𝑧/(1 − 𝑧) to obtain
∞ (︁ )︁𝑛 ∞ (︂ )︂𝑛 ∑︁ ∞ (︂ )︂𝑛
∑︁ 𝑟 ∑︁ 𝑟𝑒𝑖(𝜃−𝜑) 𝑟𝑒−𝑖(𝜃−𝜑)
1+2 cos 𝑛(𝜃 − 𝜑) = 1 + +
𝑛=1
𝑎 𝑛=1
𝑎 𝑛=1
𝑎
𝑟𝑒𝑖(𝜃−𝜑)
(︂ )︂
=1+ + c.c.
𝑎 − 𝑟𝑒𝑖(𝜃−𝜑)
𝑎2 − 𝑟2
= 2 .
𝑎 − 2𝑎𝑟 cos (𝜃 − 𝜑) + 𝑟2
Therefore,
2𝜋
𝑎2 − 𝑟2
∫︁
𝑓 (𝜑)
𝑢(𝑟, 𝜃) = 𝑑𝜑,
2𝜋 0 𝑎2 − 2𝑎𝑟 cos (𝜃 − 𝜑) + 𝑟2
an integral result for 𝑢(𝑟, 𝜃) known as Poisson’s formula. As a trivial example,

consider the solution for 𝑢(𝑟, 𝜃) if 𝑓 (𝜃) = 𝐹 , a constant. Clearly, 𝑢(𝑟, 𝜃) = 𝐹
satisfies both the Laplace equation and the boundary conditions so must be the
solution. You can verify that 𝑢(𝑟, 𝜃) = 𝐹 is indeed the solution by showing that
∫︁ 2𝜋
𝑑𝜑 2𝜋
= 2 .
0 𝑎2 − 2𝑎𝑟 cos (𝜃 − 𝜑) + 𝑟2 𝑎 − 𝑟2

Introduction To Differential Equations: Jeffrey R. Chasnov

Uploaded by

Copyright:

Available Formats

Introduction To Differential Equations: Jeffrey R. Chasnov

Uploaded by

Document Information

Original Title

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

Introduction To Differential Equations: Jeffrey R. Chasnov

Uploaded by

Copyright:

Available Formats

Introduction to Differential

Lecture notes for MATH 2351/2352

The Hong Kong University of

0 A short mathematical review 1

3 Second-order odes, constant coefficients 29

4 The Laplace transform 47

7 Nonlinear differential equations 83

8 Partial differential equations 97

A basic understanding of calculus is required to undertake a study of differential

0.1 The trigonometric functions

0.2 The exponential function and the natural

𝑒𝑥 𝑒𝑦 = 𝑒𝑥+𝑦 , 𝑒𝑥 /𝑒𝑦 = 𝑒𝑥−𝑦 , (𝑒𝑥 )𝑝 = 𝑒𝑝𝑥 .

The corresponding rules for the logarithmic function are

0.3 Definition of the derivative

0.4 Differentiating a combination of functions

Similarly, the derivative of the difference is

0.4.2 The product rule

0.4.3 The quotient rule

0.4.4 The chain rule

0.5 Differentiating elementary functions

0.5.2 Trigonometric functions

(sin 𝑥)′ = cos 𝑥, (cos 𝑥)′ = − sin 𝑥.

(sin 𝑥)′′ = − sin 𝑥, (cos 𝑥)′′ = − cos 𝑥.

0.5.3 Exponential and natural logarithm functions

0.6 Definition of the integral

Also, if 𝑎 < 𝑏 < 𝑐, then

0.7 The fundamental theorem of calculus

Unfortunately, finding antiderivatives is much harder than finding derivatives,

to choose the same ℎ in both limits. With 𝐹 ′ (𝑥) = 𝑓 (𝑥), we have

0.8 Definite and indefinite integrals

where F(x) is the antiderivative of 𝑓 (𝑥).

0.9 Indefinite integrals of elementary functions

Trigonometric functions can also be integrated:

Easily proved identities are an addition rule:

and multiplication by a constant:

This permits integration of functions such as

This integration formula is usually implemented by letting 𝑦 = 𝑔(𝑥). Then one

0.11 Integration by parts

Commonly, the above integral is done by writing

0.12 Taylor series

𝑓 (𝑥, 𝑦) = 𝑓 (𝑎, 𝑏) + 𝑓𝑥 (𝑎, 𝑏)(𝑥 − 𝑎) + 𝑓𝑦 (𝑎, 𝑏)(𝑦 − 𝑏)

0.13 Complex numbers

Using 𝑧 and 𝑧¯, we have

and we define the absolute value of 𝑧, also called the modulus of 𝑧, by

𝑧 + 𝑤 = (𝑥 + 𝑠) + 𝑖(𝑦 + 𝑡); 𝑧 − 𝑤 = (𝑥 − 𝑠) + 𝑖(𝑦 − 𝑡);

𝑧𝑤 = (𝑥𝑠 − 𝑦𝑡) − 𝑖(𝑥𝑡 + 𝑦𝑠)

𝑒𝑖(𝑥+𝑦) = 𝑒𝑖𝑥 𝑒𝑖𝑦 .

cos(𝑥 + 𝑦) + 𝑖 sin(𝑥 + 𝑦) = (cos 𝑥 + 𝑖 sin 𝑥)(cos 𝑦 + 𝑖 sin 𝑦)

De Moivre’s Theorem derives from 𝑒𝑖𝑛𝜃 = (𝑒𝑖𝜃 )𝑛 , yielding the identity

cos(𝑛𝜃) + 𝑖 sin(𝑛𝜃) = (cos 𝜃 + 𝑖 sin 𝜃)𝑛 .

For example, if 𝑛 = 2, we derive

cos 2𝜃 + 𝑖 sin 2𝜃 = (cos 𝜃 + 𝑖 sin 𝜃)2

A differential equation is an equation for a function that relates the values of

1.1 The simplest type of differential equation

2.1 The Euler method

2.2 Separable equations

The integral on the left can be transformed by substituting 𝑢 = 𝑦(𝑥), 𝑑𝑢 =

dy/dx + y/2 = 3/2

and checking the initial conditions, 𝑦(0) = 3 − 𝑒0 = 2. Therefore, our solution