Strogatz Notes PDF

Student Solutions Manual for
Nonlinear Dynamics and Chaos,

Second Edition
Mitchal Dichter
Boca Raton London New York
CRC Press is an imprint of the

Taylor & Francis Group, an informa business
A CHAPMAN & HALL BOOK
First published 2017 by Westview Press
Published 2018 by CRC Press

Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
CRC Press is an imprint of the Taylor & Francis Group, an informa business
Copyright © 2017 by Steven H. Strogatz and Taylor & Francis Group LLC
No claim to original U.S. Government works
This book contains information obtained from authentic and highly regarded sources. Reasonable efforts
have been made to publish reliable data and information, but the author and publisher cannot assume
responsibility for the validity of all materials or the consequences of their use. The authors and publishers
have attempted to trace the copyright holders of all material reproduced in this publication and apologize to
copyright holders if permission to publish in this form has not been obtained. If any copyright material has
not been acknowledged please write and let us know so we may rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter
invented, including photocopying, microfilming, and recording, or in any information storage or retrieval
system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, please access www.
copyright.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222
Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides
licenses and registration for a variety of users. For organizations that have been granted a photocopy license
by the CCC, a separate system of payment has been arranged.
Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used
only for identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com
and the CRC Press Web site at

http://www.crcpress.com
Every effort has been made to secure required permissions for all text, images, maps, and other art reprinted
in this volume.
A CIP catalog record for the print version of this book is available from the Library of Congress
ISBN 13: 978-0-8133-5054-7 (pbk)

CONTENTS
2 Flows on the Line 1

2.1 A Geometric Way of Thinking 1
2.2 Fixed Points and Stability 2
2.3 Population Growth 7
2.4 Linear Stability Analysis 9
2.5 Existence and Uniqueness 11
2.6 Impossibility of Oscillations 13
2.7 Potentials 13
2.8 Solving Equations on the Computer 14
3 Bifurcations 19
3.1 Saddle-Node Bifurcation 19
3.2 Transcritical Bifurcation 27
3.3 Laser Threshold 31
3.4 Pitchfork Bifurcation 33
3.5 Overdamped Bead on a Rotating Hoop 43
3.6 Imperfect Bifurcations and Catastrophes 45
3.7 Insect Outbreak 55
4 Flows on the Circle 65

4.1 Examples and Definitions 65
4.2 Uniform Oscillator 66
4.3 Nonuniform Oscillator 67
4.4 Overdamped Pendulum 75
4.5 Fireflies 77
4.6 Superconducting Josephson Junctions 80
5 Linear Systems 87
5.1 Definitions and Examples 87
5.2 Classification of Linear Systems 92
5.3 Love Affairs 101
6 Phase Plane 103

6.1 Phase Portraits 103
6.2 Existence, Uniqueness, and Topological Consequences 109
6.3 Fixed Points and Linearization 110
6.4 Rabbits versus Sheep 117
6.5 Conservative Systems 129
6.6 Reversible Systems 145
6.7 Pendulum 160
6.8 Index Theory 164
iii
iv CONTENTS
7 Limit Cycles 173

7.1 Examples 173
7.2 Ruling Out Closed Orbits 179
7.3 Poincaré-Bendixson Theorem 188
7.4 Liénard Systems 197
7.5 Relaxation Oscillations 198
7.6 Weakly Nonlinear Oscillators 203
8 Bifurcations Revisited 219

8.1 Saddle-Node, Transcritical, and Pitchfork Bifurcations 219
8.2 Hopf Bifurcations 226
8.3 Oscillating Chemical Reactions 237
8.4 Global Bifurcations of Cycles 241
8.5 Hysteresis in the Driven Pendulum and Josephson Junction 248
8.6 Coupled Oscillators and Quasiperiodicity 253
8.7 Poincaré Maps 267
9 Lorenz Equations 273

9.1 A Chaotic Waterwheel 273
9.2 Simple Properties of the Lorenz Equations 276
9.3 Chaos on a Strange Attractor 279
9.4 Lorenz Map 292
9.5 Exploring Parameter Space 292
9.6 Using Chaos to Send Secret Messages 303
10 One-Dimensional Maps 307

10.1 Fixed Points and Cobwebs 307
10.2 Logistic Map: Numerics 318
10.3 Logistic Map: Analysis 323
10.4 Periodic Windows 331
10.5 Liapunov Exponent 339
10.6 Universality and Experiments 342
10.7 Renormalization 352
11 Fractals 359
11.1 Countable and Uncountable Sets 359
11.2 Cantor Set 360
11.3 Dimension of Self-Similar Fractals 362
11.4 Box Dimension 366
11.5 Pointwise and Correlation Dimensions 369
12 Strange Attractors 371

12.1 The Simplest Examples 371
12.2 Hénon Map 381
12.3 Rössler System 387
12.4 Chemical Chaos and Attractor Reconstruction 389
12.5 Forced Double-Well Oscillator 391
2
Flows on the Line
2.1 A Geometric Way of Thinking

2.1.1
The fixed points of the flow ẋ = sin(x) occur when
ẋ = 0 ⇒ sin(x) = 0 ⇒ x = zπ z∈Z
2.1.3
a)
We can find the flow’s acceleration ẍ by first deriving an equation containing ẍ by taking the time derivative
of the differential equation.
d d
ẋ = sin(x) ⇒ ẍ = cos(x)ẋ
dt dt
We can obtain ẍ solely as a function of x by plugging in our previous equation for ẋ.
ẍ = cos(x) sin(x)
b)
We can find what values of x give the acceleration ẍ maximum positive values by using the trigonometric
identity
1
sin(2x) = sin(x) cos(x)
2
which can be used to rewrite ẍ as
1
ẍ = sin(2x)
2
which has maximums when Å ã
1
x= z+ π z∈Z
4
2.1.5
a)
A pendulum submerged in honey with the pendulum at the 12 o’clock position corresponding to x = 0 is
qualitatively similar to ẋ = sin(x). The force near the 12 o’clock position is small, is greatest at the 3 o’clock
position, and is again small at the 6 o’clock position.
b)
x = 0 and x = π being unstable and stable fixed points respectively is consistent with our intuitive under-
standing of gravity.
1
2 Chapter 2: Flows on the Line
2.2 Fixed Points and Stability

2.2.1
ẋ = 4x2 − 16
ẋ
−2 2
x(t) t ∈ [0, 2]
−2
2 x0 e16t + x0 − 2e16t + 2

x(t) =
−x0 e16t + x0 + 2e16t + 2
2.2.3
ẋ = x − x3
ẋ
−1 0 1
x(t) t ∈ [0, 1]
−1
±et
x(t) = q depending on the sign of the initial condition.
1
x20
+ e2t − 1
2.2.5
1
ẋ = 1 + cos(x)
2
ẋ
There are no fixed points, but the rate increase for x(t) does vary.
x(t)
å √ åå
√ tan x20
Ç Ç Ç
3t
x(t) = 2 arctan 3 tan arctan √ +
3 4
2.2.7
ẋ = ex − cos(x)
We can’t solve for the fixed points analytically, but we can find the fixed points approximately by looking
at the intersections of ex and cos(x), and determine the stability of the fixed points from which graph is
greater than the other nearby.
ẋ
x
0
We could also plot the graph using a computer.
ẋ
x
0
Å
ã
1
There are fixed points at x ≈ π − n , n ∈ N, and x = 0.
2
x(t) t ∈ [0, 4]
0
−1.57
−4.71
−7.85
Unable to find an analytic solution.
2.2.9
f (x) = x(x − 1)
ẋ
0 x
1
2.2.11
RC circuit
V0 Q 1 V0
Q̇ = − ⇒ Q̇ + Q= Q(0) = 0
R RC RC R
t
Multiply by an integrating factor e RC to both sides.
t 1 t V0 t d Ä t
ä V0 t
Q̇e RC + e RC Q = e RC ⇒ Qe RC = e RC
RC R dt R
Apply an indefinite integral to both sides with respect to t.

Z Z
d Ä t
ä V0 t
Qe RC dt = e RC dt
dt R
t t −t
Qe RC = V0 Ce RC + D ⇒ Q(t) = V0 C + De RC
And using the initial condition.

Ä −t ä
Q(0) = V0 C + D = 0 ⇒ D = −V0 C ⇒ Q(t) = V0 C 1 − e RC
2.2.13
Terminal velocity
The velocity v(t) of a skydiver falling follows the equation
mv̇ = mg − kv 2
with m the mass of the skydiver, g the acceleration due to gravity, and k > 0 the drag coefficient.
a) Ç… å
…
mg gk
v(t) = tanh t
k m
b) …
t→∞ mg
v(t) −−−→
k
c)
The terminal velocity should occur at a fixed point. The fixed point occurs when
…
2 mg
v̇ = 0 ⇒ mg − kv = 0 ⇒ v =
k
d)
31400ft − 2100ft ft
vavg = ≈ 253
116sec sec
e) Ç Ç… åå
m gk
s(t) = ln cosh t
k m
»
mg
which can be rewritten using V = k into
V2 g
s(t) = ln cosh t
g V
ft
Using s(116sec) = 29300ft and g = 32.2 sec 2 gives
ft
Ç Ç åå
V2 32.2 sec 2
29300ft = ft
ln cosh t
32.2 sec 2 V
ft
which can be solved numerically to give V ≈ 266 sec .
2.3 Population Growth 7
2.3 Population Growth

2.3.1
Logistic equation Å ã
N
Ṅ = rN 1 − N (0) = N0
K
a)
Solve by separating the variables and integrating using partial fractions.
rN 2 rKN − rN 2
Å ã
N r
KN − N 2

rN 1 − = rN − = =
K K K K
dN r dN r
KN − rN 2 ⇒

Ṅ = = = dt
dt K KN − N 2 K
1 1 1
2
= −
KN − N KN K(N − K)
Z Z Z
dN dN dN 1 1
= − = ln(N ) − ln(N − K)
KN − N 2 KN K(N − K) K K
Z
r r
= dt = t + C
K K
1 1 r
ln(N ) − ln(N − K) = t + C
K K K
ln(N ) − ln(N − K) = rt + KC
Å ã
N
ln = rt + KC
N −K
N −K
Å ã Å ã Å ã
N K
− ln = ln = ln 1 − = −rt − KC
N −K N N
K
1− = e−rt−KC = e−KC e−rt
N
K
1 − e−KC e−rt =
N
K
N (t) =
1 − e−KC e−rt
K K
N (0) = = N0 ⇒ e−KC = 1 −
1 − e−KC N0
K
N (t) = Ä ä
1 − 1 − NK0 e−rt
b)
1
Making the change of variables x = N.
Å ã
1 ẋ 1 1
N= ⇒ Ṅ = − 2 = r 1−
x x x xK
r
ẋ = − rx
K
1 1 1 KCe−rt + 1
x(t) = Ce−rt + ⇒ = Ce−rt + =
K N (t) K K
K
⇒ N (t) =
KCe−rt + 1
K
N (0) = N0 ⇒ N (t) = Ä ä
1 − 1 − NK0 e−rt
2.3.3
Tumor growth
Ṅ = −aN ln(bN )
a)
1
a can be interpreted as specifying how fast the tumor grows, and specifies the stable size of the tumor.
b
b)
Ṅ
N
0 1
b
t
N (t)
1
b
2.3.5
Dominance of the fittest
X(t)
Ẋ = aX Ẏ = bY x(t) = X0 , Y0 > 0 a>b>0
X(t) + Y (t)
a)
Show as t → ∞, x(t) → 1 by solving for X(t) and Y (t).
eat 1
X(t) = eat Y (t) = ebt ⇒ x(t) = =
eat + ebt 1 + e(b−a)t
1
a > b > 0 ⇒ 0 > b − a ⇒ x(t) → = 1 as t → ∞
1+0
b)
Show as t → ∞, x(t) → 1 by deriving an ODE for x(t).
ẊY − X Ẏ aXY − bXY

Å ã
X X
ẋ = = = (a − b) 1− = (a − b)x(1 − x)
(X + Y )2 (X + Y )2 X +Y X +Y
2.4 Linear Stability Analysis 9
This is the logistic equation

x
ẋ = rx 1 − r =a−b>0 K=1
K
X0
X0 , Y0 > 0 ⇒ 0 < x0 = < 1 ⇒ x(t) → 1 = K
X0 + Y0
and x(t) increases monotonically as t → ∞ from our previous analysis of the logistic equation.
2.4 Linear Stability Analysis

2.4.1
ẋ = x(1 − x)
dẋ
= 1 − 2x
dx
dẋ
(0) = 1 ⇒ unstable fixed point
dx
dẋ
(1) = −1 ⇒ stable fixed point
dx
2.4.3
ẋ = tan(x)
dẋ
= sec2 (x)
dx
dẋ
(πz) = 1 ⇒ unstable fixed point z ∈ Z
dx
π
There are also stable fixed-like points at πz + because ẋ is positive on the left and negative on the right.
2
However, these points aren’t true fixed points because tan(x) has infinite discontinuities at these points and
hence is not defined there.
2.4.5
2
ẋ = 1 − e−x
dẋ 2
= 2xe−x
dx
dẋ
(0) = 0 ⇒ inconclusive
dx
ẋ
The fixed point 0 is stable from the left and unstable from the right.
2.4.7
√ √
ẋ = ax − x3 = x( a − x)( a + x)
dẋ
= a − 3x2
dx
√
Assuming a ≥ 0 in order for a to be a real root.


 unstable :a<0
dẋ √


(− a) = −2a ⇒ see 0 root in this case : a = 0
dx 


 stable :0<a



 stable :a<0
dẋ 
(0) = a ⇒ inconclusive, but graph implies unstable :a=0
dx 


 unstable :0<a
√
Assuming a ≥ 0 in order for a to be a real root.


 unstable :a<0
dẋ √


( a) = −2a ⇒ see 0 root in this case :a=0
dx 


 stable :0<a
ẋ
0
2.5 Existence and Uniqueness 11
2.4.9
a)
±1
x(t) = q
1
x20
+ 2t
Depending on the sign of the initial condition
±1
lim x(t) → √ → 0
t→∞ 2t
b)
x(t) t ∈ [0, 10] x(t) = −x3 x(t) = −x
2.5 Existence and Uniqueness

2.5.1
ẋ = −xc x(t) ≥ 0 c∈R
a)
x = 0 is the only fixed point ⇐⇒ c 6= 0 and x(0) ≥ 0 ⇒ x(t) → 0 so x = 0 is stable from the right.
b)
1
x(t) = ((c − 1)t + x1−c
0 )
1−c
The particle will reach the origin in finite time if x0 = 0 or if x0 > 0 and c < 1 so that a t > 0 will make
(c − 1)t + x1−c
0 =0
If x0 = 1 then
1
x(t) = (1 + (c − 1)t) 1−c
And with c < 1 then

1
t∗ = >0 x(t∗ ) = 0
1−c
2.5.3
√
x0 rert
ẋ = rx + x3 r > 0 ⇒ x(t) = ± p 2
x0 (1 − e2rt ) + r
Which will blow up when the denominator vanishes at
r
ln(1 + x20
)
t=
2r
2.5.5
p
ẋ = |x| q x(0) = 0 p, q are coprime
x(t) = 0 is a solution, but we can find another solution through

Z Z
−p
|x| dx = dt
q
1 1− p
p |x|
q = t + C
1− q
ÅÅ ã ã 1p
p 1−
q
|x| = 1− (t + C)
q
x(0) = 0 ⇒ C = 0
ÅÅ ã ã 1p
p 1−
q
|x| = 1− t
q
a)
p p
p<q⇒ <1⇒0<1−
q q
Here there are an infinite number of solutions because we can specify that x(t) = 0 for an arbitrary amount
of time t0 and then switch to
ÅÅ ã ã 1p
p 1−
q
x(t) = 1− (t − t0 )
q
b)
p p
q<p⇒1< ⇒1− <0
q q
Here we only have x(t) = 0 for all time because the RHS of
ÅÅ ã ã 1p
p 1−
q
|x| = 1− t
q
would not always be nonnegative, which is inconsistent with the absolute value on the LHS.
2.7 Potentials 13
2.6 Impossibility of Oscillations

2.6.1
The harmonic oscillator does oscillate along the x-axis, but the position alone does not uniquely describe the
state of the system. The system is not uniquely determined unless both the position x(t) and the velocity
ẋ(t) are specified. The harmonic oscillator is a two-dimensional system so does not contradict the text by
the fact that solutions can oscillate.
2.7 Potentials
2.7.1
1 1
V (x) = − x2 + x3
2 3
V (x)
(0, 0) x
(1, −1/6)
2.7.3
V (x) = cos(x)
V (x)
(−2π, 1) (0, 1) (2π, 1)
x
(−π, 1) (π, 1)
The pattern of equilibrium points continues in both directions.

2.7.5
V (x) = cosh(x)
V (x)
(0, 1)
x
2.7.7
Assume for
dV
ẋ = f (x) = −
dx
there is an oscillating solution x(t) with period T 6= 0. Then
x(t) = x(t + T ) ⇒ V (x(t)) = V (x(t + T ))
but
dV
≤ 0 ⇒ x(t)
dt
is constant. In other words, the solution x(t) does not oscillate and never moves. Contradiction!
2.8 Solving Equations on the Computer
2.8.1
The slope is constant along horizontal lines because the equation for the slope ẋ explicitly depends on x but
not on t. Although t varies along the horizontal lines, x remains constant, so the slope remains constant.
2.8.3
a)
ẋ = −x, x(0) = 1 ⇒ x(t) = e−t x(1) = e−1 ≈ 0.3678794411
b)
∆t = 100 ⇒ x̂(1) = 0
∆t = 10−1 ⇒ x̂(1) ≈ 0.3486784401
∆t = 10−2 ⇒ x̂(1) ≈ 0.3660323413
∆t = 10−3 ⇒ x̂(1) ≈ 0.3676954248
∆t = 10−4 ⇒ x̂(1) ≈ 0.3678610464
c)
0.3
∆t
ln(E)
−10
ln(∆t)
−10
The slope is approximately 1 for small values of ∆t, which indicates that the error of the method is propor-
tional to ∆t.
ln(E) ln(C∆t) ln(C) + ln(∆t) ln(C) ln(∆t)

≈ = = + ≈ 0 + 1 = 1 for small ∆t
ln(∆t) ln(∆t) ln(∆t) ln(∆t) ln(∆t)
2.8.5
a)
ẋ = −x, x(0) = 1 ⇒ x(t) = e−t x(1) = e−1 ≈ 0.3678794411
b)
∆t = 100 ⇒ x̂(1) = 0.375
∆t = 10−1 ⇒ x̂(1) ≈ 0.367879774412498
∆t = 10−2 ⇒ x̂(1) ≈ 0.367879441202355
∆t = 10−3 ⇒ x̂(1) ≈ 0.367879441171446
∆t = 10−4 ⇒ x̂(1) ≈ 0.367879441171445
c)
0.0075
∆t
1
−10 ln(E)
ln(∆t)
−40
The slope is approximately 4 for small values of ∆t, which indicates that the error of the method is propor-
tional to (∆t)4 .
ln(E) ln(C(∆t)4 ) ln(C) + ln((∆t)4 ) ln(C) 4 ln(∆t)
≈ = = + ≈ 0 + 4 = 4 for small ∆t
ln(∆t) ln(∆t) ln(∆t) ln(∆t) ln(∆t)
2.8.7
Euler’s method
xn+1 = xn + ∆tf (xn )
a)
(∆t)2
x(t0 + ∆t) = x(t0 ) + ∆tẋ(t0 ) + ẍ(ξ) ξ ∈ (t0 , t0 + ∆t)
2
(∆t)2 0
= x(t0 ) + ∆tf (x(t0 )) + f (x(ξ))f (x(ξ))
2
b)
(∆t)2 0
|x(t1 ) − x1 | = |x(t0 + ∆t) − (x0 + ∆tf (x0 ))| = |f (x(ξ))f (x(ξ))| ≤ C(∆t)2
ß 0 2
|f (x(ξ))f (x(ξ))|
™
C = max : ξ ∈ (t0 , t0 + ∆t)
2
2.8.9
Runge-Kutta
k1 = ∆tf (xn )
k2 = ∆tf (xn + 21 k1 )
k3 = ∆tf (xn + 12 k2 )
k4 = ∆tf (xn + k3 )
xn+1 = xn + 16 (k1 + 2k2 + 2k3 + k4 )
a)
(∆t)2 d2 x (∆t)3 d3 x (∆t)4 d4 x (∆t)5 d5 x
x(t0 + ∆t) = x(t0 ) + ∆t dx
dt (t0 ) + 2 dt2 (t0 ) + 3! dt3 (t0 ) + 4! dt4 (t0 ) + 5! dt5 (ξ), ξ ∈ (t0 , t0 + ∆t)
(∆t)2 0 (∆t)3 00 (∆t)4 000 (∆t)5 0000
= x(t0 ) + ∆tf (t0 ) + 2 f (t0 ) + 3! f (t0 ) + 4! f (t0 ) + 5! f (ξ), ξ ∈ (t0 , t0 + ∆t)
Here we have to translate each derivative into f (x(t0 ))’s.
k1 = ∆tf (x0 )
k2 = ∆tf x0 + 12 k1

2 3 4 5
( 1 k1 ) ( 1 k1 ) ( 1 k1 ) ( 1 k1 )

= ∆t f (x0 ) + 21 k1 f (1) (x0 ) + 2 2! f (2) (x0 ) + 2 3! f (3) (x0 ) + 2 4! f (4) (x0 ) + 2 5! f (5) (ξ2 ) , ξ2 ∈
(x0 , x0 + 12 k1 )
And then we need to substitute all the k1 ’s as f (x0 )’s.
k3 = ∆tf x0 + 12 k2

2 3 4 5
( 1 k2 ) ( 1 k2 ) ( 1 k2 ) ( 1 k2 )

= ∆t f (x0 ) + 21 k2 f (1) (x0 ) + 2 2! f (2) (x0 ) + 2 3! f (3) (x0 ) + 2 4! f (4) (x0 ) + 2 5! f (5) (ξ3 ) , ξ3 ∈
(x0 , x0 + 12 k2 )

k4 = ∆tf x0 + k3
2 3 4 5

= ∆t f (x0 ) + k3 f (1) (x0 ) + (k2!3 ) f (2) (x0 ) + (k3!3 ) f (3) (x0 ) + (k4!3 ) f (4) (x0 ) + (k5!3 ) f (5) (ξ4 ) , ξ4 ∈ (x0 , x0 + k3 )

b) |x(t1 ) − x1 | = x(t0 + ∆t) − x0 + 16 (k1 + 2k2 + 2k3 + k4 )

And then we need to substitute the x(t0 + ∆t), k1 , k2 , k3 , and k4 as f (x0 )’s. This is a herculean task, but
at the end we will get a bound on the absolute value of the local truncation error, C(∆t)5 .
3
Bifurcations
3.1 Saddle-Node Bifurcation

3.1.1
ẋ = 1 + rx + x2
The fixed points are
√ 2 fixed points exist if |r| > 2

−r ± r2 − 4
x= 1 fixed point exists if r = ±2 ⇒ Saddle-node bifurcation occurs at r = ±2
2
0 fixed points exist if |r| < 2
To sketch the vector fields for different values of r, draw y = 1 + x2 and y = −rx. The intersections represent
the fixed points, and the flow is to the right and left if 1 + x2 > −rx and 1 + x2 < −rx respectively.
ẋ
r = −3
19
20 Chapter 3: Bifurcations
ẋ
r = −2
ẋ
r = −1
ẋ
r=0
x
ẋ
r=1
ẋ
r=2
x
ẋ
r=3
x
√
−r ± r2 − 4
Å ã
1
The bifurcation diagram is x = or equivalently r = − x + , which is easier to plot.
2 x
x
3.1.3
ẋ = r + x − ln(1 + x)
To sketch the vector fields for different values of r, draw y = r + x and y = ln(1 + x). The intersections
represent the fixed points, and the flow is to the right and left if r + x > ln(1 + x) and r + x < ln(1 + x)
respectively.
ẋ ẋ
r = −1 r=0
x x
ẋ
r=1
The bifurcation occurs when the curves y = r + x and y = ln(1 + x) have a tangential intersection. We can
find this r value analytically by solving the equations
d 1 d
r + x = ln(1 + x) and r+x =1= = ln(1 + x)
dx 1+x dx
1
1= ⇒x=0 r + x = ln(1 + x) −→ r = ln(1) = 0
1+x
Hence, a saddle-node bifurcation occurs at r = 0.
x
r
3.1.5
a)
ẋ = r2 − x2
ẋ ẋ
r = −1 r = −0.5
x x
ẋ ẋ
r=0 r = 0.5
x x
ẋ
r=1
x
A saddle-node bifurcation doesn’t occur. Instead there is more of a transcritical bifurcation, but not quite
since both of the fixed-point positions change with r.
b)
ẋ = r2 + x2
ẋ ẋ
r = −1 r = −0.5
x x
ẋ ẋ
r=0 r = 0.5
x x
ẋ x
r=1
x r
Again, a saddle-node bifurcation doesn’t occur. In this case there is almost nothing to plot on the bifurcation
diagram.
3.2 Transcritical Bifurcation

3.2.1
ẋ = rx + x2
r = −2 ẋ r = −1 ẋ
x x
r=0 ẋ r=1 ẋ
x x
r=2 ẋ
Transcritical bifurcation occurs at r = 0.

3.2.3
ẋ = x − rx(1 − x)
ẋ
r = −0.5
ẋ
r = −0.25
x
r=0 ẋ r = 0.5 ẋ
x x
Note: The positive fixed point goes to x = ∞ as r → 0− , disappears when r = 0, then reappears at x = −∞
when r becomes slightly positive.
r=1 ẋ r=2 ẋ
x x
r=3 ẋ
Transcritical bifurcation occurs at r = 1.

3.2.5
a)
1 k
−−
A+X )−−
−−−*
− 2X
k−1
k
X + B −−−−2−→ C
The k1 reaction has a net gain of one X at rate k1 [A][X], the k−1 reaction has a net loss of one X at rate
k−1 [X][X] = k−1 [X]2 , and the k2 reaction has a net loss of one X at rate k2 [B][X], making the rate of
˙ = ẋ the sum of the rates for positive gains and negative losses.
concentration change [X]
ẋ = k1 ax − k−1 x2 − k2 bx
= (k1 a − k2 b)x − k−1 x2
= c1 − c2 x2
So c1 = k1 a − k2 b and c2 = k−1 .
b)
dẋ
= k1 a − k2 b − 2k−1 x
dx
dẋ
= k1 a − k2 b < 0
dx ∗

x =0
⇒ k1 a < k2 b
This makes sense chemically because the amount of chemical X will always either increase or decrease to
0. Chemicals A and B are considered inexhaustible, so if the first reaction creates X faster than the second
reaction irreversibly destroys X(k2 b < k1 a), then the amount of X will grow unbounded. But if the first
reaction creates X slower than the second reaction irreversibly destroys X(k1 a < k2 b), then the amount of
X will approach 0.
3.3 Laser Threshold 31
3.3 Laser Threshold

3.3.1
ṅ = GnN − kn
Ṅ = −GnN − f N + p
a)
If Ṅ ≈ 0, then
0 ≈ −GnN − f N + p
p
N≈
Gn + f
Gpn
ṅ = GnN − kn ≈ − kn
Gn + f
b)
dṅ Gpf
= −k
dn (Gn + f )2

dṅ Gpf Gp
= −k = −k >0
dn f2 f

n∗ =0
kf
⇒p> = pc
Gp
c)
If we set G = k = f = 1,
pn
ṅ ≈ − n = 0 ⇒ n = 0 or n = p − 1
n+1
ṅ ṅ ṅ
p = 0.125 p=1 p=2
n n n
A transcritical bifurcation occurs at pc .

n
d)
ṅ = GnN − kn = (GN − k)n
Ṅ = −GnN − f N + p = −(Gn + f )N + p
The assumption that N almost isn’t changing relative to n is valid if |GN − k| |Gn + f )|. Then N will
initially change far more rapidly than n, but then stay put once it reaches an equilibrium with the current
n value. Now n will still change, but N will reach the equilibrium with the newer n value almost instantly,
keeping Ṅ ≈ 0.
3.4 Pitchfork Bifurcation

3.4.1
ẋ = rx + 4x3
ẋ ẋ ẋ
r = −2 r = −1 r=0
x x x
ẋ ẋ
r=1 r=2
x x
A subcritical pitchfork bifurcation occurs at r = 0.

x
r
3.4.3
ẋ = rx − 4x3
ẋ ẋ
r = −2 r = −1
x x
ẋ
r=0
ẋ ẋ
r=1 r=2
x x
A supercritical pitchfork bifurcation occurs at r = 0.

x
r
3.4.5
ẋ = r − 3x2
ẋ ẋ
r = −1 r = −0.5
x x
ẋ
r=0
ẋ ẋ
r = 0.5 r=1
x x
A saddle-node bifurcation occurs at r = 0.

x
3.4.7
2
ẋ = 5 − re−x
ẋ ẋ
r=3 r=4
x x
ẋ
r=5
x
ẋ ẋ
r=6 r=7
x x
A saddle-node bifurcation occurs at r = 5.

x
3.4.9
ẋ = x + tanh(rx)
ẋ ẋ
r = −3 r = −2
x x
ẋ ẋ
r = −1 r=0
x x
ẋ
r=1
A subcritical pitchfork bifurcation occurs at r = −1.

x
3.4.11
ẋ = rx − sin(x)
a)
r=0
ẋ
There are an infinite number of fixed points when r = 0 at every integer multiple of π. All even multiples of
π are stable, and all odd multiples of π are unstable.
b)
r = 1.25
ẋ
x
r=1
ẋ
For r ≥ 1, there is only one intersection between y = rx and y = sin(x), which is an unstable fixed point at
the origin.
c)
There are no bifurcations until r ≤ 1, and the first bifurcation is a subcritical pitchfork bifurcation at r = 1.
r = 0.75
ẋ
Note: To add more rigor to our bifurcation classification, we can expand in a Taylor series about x = 0,
which gives
x3 x5 x3
Å ã
ẋ = rx − sin(x) ≈ rx − x − + − · · · ≈ (r − 1)x +
3! 5! 3!
This is the normal form of a subcritical pitchfork bifurcation at r − 1 = 0.
There are more bifurcations as r continues to decrease toward zero. However, all these subsequent bi-
furcations are saddle-node bifurcations.
The first of these bifurcations occurs at r ≈ 0.1284 when the line y = rx becomes tangent to a hilltop
of the sine wave.
r = 0.25 ẋ
x
r ≈ 0.1284 ẋ
x
r = 0.05 ẋ
x
Saddle-node bifurcations occur repeatedly as the line y = rx becomes tangent to the next hilltop of the sine
wave, and an infinite number of hilltops implies an infinite number of saddle-node bifurcations.
d)
To find a formula for the r values of the saddle-node bifurcations, we can use the fact that y = rx and
y = sin(x) tangentially intersect at the bifurcation values.
d d
rx = sin(x) rx = r = cos(x) = sin(x) ⇒ x = tan(x) ⇒ r = cos(x)
dx dx
Therefore, the saddle-node bifurcation values are r = cos(x∗ ) where x∗ is a root of the equation x∗ = tan(x∗ ).
For an approximate formula when 0 < r 1, the saddle-node bifurcations occur approximately at the
hilltops of the sine wave, since y = rx intersects y = sin(x) tangentially and y = rx is almost flat. Therefore,
the bifurcations on the positive x-axis and negative x-axis respectively occur approximately at the local
minimums and local maximums of sin(x). These points are
±(4n + 1)π
ß ™
:n∈N
2
We have to skip the first local minimum and local maximum on the positive x-axis and negative x-axis
respectively because they are involved in the subcritical pitchfork bifurcation.
The approximate values of r at which the saddle-node bifurcations occur are the reciprocals of that set.
±2
ß ™
r∈ :n∈N
(4n + 1)π
So x and r in the first and second set respectively implies rx = ±1 and sin(x) = ±1, which will cancel out
in ẋ = rx − sin(x).
e)
For −∞ < r < 0, we again have saddle-node bifurcations occurring.
ẋ r = −0.05
x
ẋ r ≈ −0.2172
x
But there is no accompanying pitchfork bifurcation because after r < −0.2172 there is only one fixed point
left at x = 0.
f)
sin(x)
The bifurcation diagram is made of the line x = 0 and the curve r = plotted sideways. Hence, x can
x
take on any value from positive to negative infinity.
x
3.4.13
a)
ẋ = r − x − e−x
x
r
b)
ẋ = 1 − x − e−rx
x
3.4.15
dV −rx2 x4 x6
ẋ = rx + x3 − x5 = − ⇒V = − + +C
dx 2 4 6
The critical points occur when
dV
= −rx − x3 + x5
dx
= −x(r + x2 − x4 ) = 0
x1 = 0
p √
2 −1 ± 12 − 4(−1)r 1 ± 1 + 4r
x = =
2(−1) 2
√ √
1 + 1 + 4r 1 − 1 + 4r
x2 = x3 =
2 2
√ √
1 + 1 + 4r 1 − 1 + 4r
x4 = − x5 = −
2 2
and the local minima occur when
d2 V
= −r − 3x2 + 5x4 > 0
dx2
d2 V
(x1 ) = −r > 0 if r < 0
dx2
d2 V d2 V √ −1
(x2 ) = (x4 ) = 4r + 4r + 1 + 1 > 0 if <r
dx2 dx2 4
2 2 √
d V d V
2
(x3 ) = 2
(x5 ) = 4r − 4r + 1 + 1 > 0 if 0 < r
dx dx
−1
so the three candidate wells occur at x1 , x2 , and x4 with 4 < r < 0.
V (x0 ) = C, so for simplicity we’ll set C = 0, which leaves
V (x2 ) = V (x4 ) = V (x0 )

1 Ä 3
ä −3
−(4r + 1) 2 − 6r − 1 = 0 ⇒ r =
24 16
3.5 Overdamped Bead on a Rotating Hoop 43
−3
r= 16
C=0
−3
So rc = 16 and the three minima x1 , x2 , and x4 occur at the value of the integrating constant C.
3.5 Overdamped Bead on a Rotating Hoop

3.5.1
π
There shouldn’t be an equilibrium between 2 and π because the rotation would force the bead outward and
down, and gravity would force the bead down. The forces wouldn’t cancel, so the bead would move.
3.5.3

A Taylor series expansion of f (φ) = sin(φ) γ cos(φ) − 1 centered at φ = 0 gives
Å ã
1 2γ
φ3 + O φ5

(γ − 1)φ + −
6 3
3.5.5
a)
Starting with the dimensionless equation
d2 φ dφ
+ = f (φ)
dτ 2 dτ
t b
We already have Tslow = = from the text, but we don’t know what order of magnitude Tfaster is in
τ mg
relation to Tslow , so we’ll guess a new time scale τ = k ξ. The dimensionless equation is now
d2 φ dφ
1−2k 2
+ −k = f (φ)
dξ dξ
We still need to be able to solve the equation, meaning each of the terms must be of the same order or
negligible compared to all the other terms.
To have all the terms of the same order, we need 1−2k = −k = 1, which is impossible.
To have two terms of the same order and one term negligible, we need
1
1−2k = −k 1 ⇒ k = 1 or 1−2k = 1 −k ⇒ k = or −k = 1 1−2k ⇒ k = 0
2
1
Choosing k = 0 gives the equation we started with, and k = gives
2
d2 φ 1 dφ
2
+√ = f (φ)
dξ dξ
1
which is bad because 1 makes the √ term enormous. But the k = 1 choice looks good.

If we choose τ = ξ for the new time scale, then we can find Tfast in terms of m, g, r, ω, and b by
using
t b
Tslow = =
τ mg
from the text.
m2 gr b
Å ã
t t τ b τ b b mr
Tfast = = = = = =
ξ ξ t mg ξ mg mg b2 mg b
b)
Rescaling with the fast time scale
r d2 φ −b dφ rω 2
2 = − sin(φ) + sin(φ) cos(φ)
gTfast dξ 2 mgTfast dξ g
b2 d2 φ −b2 dφ rω 2
2 2
= 2 − sin(φ) + sin(φ) cos(φ)
m gr dτ m gr dτ g
b2 d2 φ b2 dφ rω 2
2 2
+ 2 = − sin(φ) + sin(φ) cos(φ)
m gr dτ m gr dτ g
d2 φ dφ m2 gr rω 2
Å ã
+ = − sin(φ) + sin(φ) cos(φ)
dτ 2 dτ b2 g
= f (φ)
Now the f (φ) is the negligible term.

c)
b
Tfast = = Tslow
mg
So if 1 then Tfast Tslow .
3.5.7
Å ã
N
Ṅ = rN 1 − N (0) = N0
K
a)
1
r∼ K ∼ population N0 ∼ population
time
b)
Å ã
N
Ṅ = rN 1 − N (0) = N0
K
N N0
x= K ẋ = rKx(1 − x) x(0) = = x0
K K
ẋ = rx(1 − x) x(0) = x0
dx
τ = rt r = rx(1 − x) x(0) = x0
dτ
dx
= x(1 − x) x(0) = x0
dτ
c)
Å ã
N
Ṅ = rN 1 − N (0) = N0
K
Å ã
N N0 N (0)
u= N0 u̇ = rN0 u 1 − u u(0) = =1
N0 K N0
Å ã
1 N0
u̇ = u 1 − u
r K
Å ã
du N0
τ = rt =u 1− u
dτ K
du
κ = rt = u (1 − κu) u(0) = 1
dτ
d)
The x nondimensionalization would be useful for investigating the effect of the initial condition relative to
carrying capacity on the solution, whereas the u nondimensionalization would be useful for studying the rate
of population growth relative to carrying capacity.
By the way, scaling in such a way that all solutions have the same initial condition is mathematically
valid but runs counter to the geometric approach we are using in this book. Our whole approach is based
on visualizing different solutions emanating from different initial conditions.
3.6 Imperfect Bifurcations and Catastrophes

3.6.1
ẋ = h + rx − x3
This x versus r bifurcation diagram

has a saddle-node bifurcation at a negative x value. Vertically shifting the cubic equation and then varying
r will achieve this.
ẋ ẋ
r=1 r=1
h=0 h = 0.25
x x
So h is positive.
3.6.3
ẋ = rx + ax2 − x3
a)
a = −1 x a=0 x
r r
a=1 x
b)
a
3 fi
xe
d
po
sa int
dd s
le-
no
de
transcritical
r
1 fixed point 3 fixed points
3.6.5
a)
The spring has spring constant k and rest length L0 , but only the force from the spring that is parallel, and
not perpendicular, with the wire will affect the bead.
Äp ä x
k(L − L0 ) = k x2 + a2 − L0 √
x + a2
2
Å ã
L0
= −kx 1 − √
x + a2
2
The force of gravity will also move the bead along the wire, but only the fraction of the force that is parallel
with the wire.
mg sin(θ)
At equilibrium, the force of the spring and the force from gravity have to be equal and opposite, which gives
Å ã
L0
mg sin(θ) = kx 1 − √
x2 + a2
b)
Å ã
L0
mg sin(θ) = kx 1 − √
x2 + a2
L0
mg sin(θ) = kx − kx √
x + a2
2
L0
mg sin(θ) − kx = −kx » 2
a xa2 + 1
−mg sin(θ) L0
+1= » 2
kx a xa2 + 1
mg sin(θ) L0
1− = »
kx a 1+ x2
a2
x L0 mg sin(θ)
u= R= h=
a a ka
h R
1− = √
u 1 + u2
c)
Both h and R are strictly positive quantities from the physical constants they’re made of. The following
graphs show the different types and number of possible intersections. In each graph, h shapes the discontin-
uous graph and R shapes the bell-shaped graph with u as the independent variable.
h=1
R = 0.5
u
h=1
R ≈ 2.8
h=1
R=5
For 0 < R < 1 there is only one intersection since the other curve asymptotes to height 1 from above. As
R increases, there is a point when the two curves just touch, making another fixed point in a saddle-node
bifurcation. As R increases further there are three fixed points.
d)
h R
1− =√
u 1 + u2
r =R−1
h r+1
1− =√
u 1 + u2
p p
u 1 + u2 − h 1 + u2 = (r + 1)u
p p
h 1 + u2 + (r + 1)u − u 1 + u2 = 0
Taylor series expansion

Å ã Å ã
1 2 4
1 2 4

h 1+ u +O u + (r + 1)u − u 1 + u + O u =0
2 2
h 1
h + u2 + ru + u − u − u3 = O u4

2 2
h 2 1 3
h + ru + u − u = O u4

2 2
h 2 1 3
h + ru + u − u ≈ 0
2 2
e)
The saddle-node bifurcations will occur approximately at the local maximum of the approximate equation
when the slopes of these equations are equal at an equal height.
1 3 h
u − ru = h + u2
2 2
h=1
r = 1.95
1 3 h
u − ru = h + u2
Å ã 2 2 Å ã
d 1 3 3 2 d h
u − ru = u − r = hu = h + u2
du 2 2 du 2
3 2
u − hu = r
2
Å ã
1 3 1 3 3 2 h
u − ru = u − u − hu u = h + u2
2 2 2 2
1 3 3 3 h
u − u + hu2 = h + u2
2 2 2
h 2 3
u −h=u
Å 2 ã
1 2
h u − 1 = u3
2
2u3
h= 2
u −2
3 3 2u3 u4 + 3u2
r = u2 − hu = u2 − 2 u=
2 2 u −2 2(1 − u2 )
f)
Here we use the same procedure as in part (e). The saddle-node bifurcations will occur at the local maximum
of the equation when the slopes of these equations are equal at an equal height.
h R
=√
1−
u 1 + u2
−Ru
Å ã Å ã
d h h d R
1− = 2 = 3 = √
du u u (1 + u2 ) 2 du 1 + u2
3
h(1 + u2 ) 2 = −Ru3
3
−h(1 + u2 ) 2
R=
u3
3
h R −h(1 + u2 ) 2 −h(1 + u2 )
1− = √ = √ =
u 1 + u2 u3 1 + u2 u3
u3 − hu2 = −h(1 + u2 )
h = −u3
3 3
−h(1 + u2 ) 2 −(−u3 )(1 + u2 ) 2 3
R= 3
= 3
= (1 + u2 ) 2
u u
The approximate result and the exact result agree in the limit of u → 0.
2u3 2u3
h = −u3 ≈ ≈ 2
0−2 u −2
3 3 04 + 3(0)2 u4 + 3u2
R = (1 + u2 ) 2 ≈ (1 + 02 ) 2 ≈ + 1 ≈ +1=r+1
2(1 − 02 ) 2(1 − u2 )
g)
32 23
h(u) = −u3 R(u) = 1 + u2 ⇒ r(u) = 1 + u2 −1
1 fixed point
e
nod
le-
dd
sa
r
3 fixed points
h)
3
h(u) = −u3 R(u) = 1 + u2 2
x
u=
a
x 3 mg sin(θ)
Å x 2 ã 32 L0
h=− = R= 1+ =
a ka a a
Solving for x and equating the two equations gives the conditions that the catastrophe will occur and the
position x of its occurrence. Since many of the parameters are fixed, we could find the catastrophe angle θ
in terms of the physical constants of the experiment.
3.6.7
a)
h = T arctanh(m) − Jnm
We can nondimensionalize to
h Jn
α + βm = arctanh(m) α= β= >0
T T
0<β≤1
α=1
α = 0.75
α=0
m
There is always a unique intersection point with this range of β.

1<β
Now there are one, two, or three solutions depending on the values depending on α for this range of β.
b)
Jn
If h = 0 then α = 0 and the phase transition occurs when the slope of the line β = Tc = 1 ⇒ Tc = Jn.
3.7 Insect Outbreak

3.7.1
dx x x2
= rx 1 − −
dτ k 1 + x2
x2 −2rx
Å ã
d x 2x
rx 1 − − = +r− 2
dx k 1 + x2 k (x + 1)2
ã
x2
Å
d x
rx 1 − − =r>0

dx k 1 + x2

x=0
So x = 0 is always unstable.
3.7.3
Å ã
N
Ṅ = rN 1 − −H
K
a)
Å ã
Ṅ N N H
= 1− −
rK K K rK
N H
x= τ = rt h =
K rK
dx
= x(1 − x) − h
dτ
b)
dx dx dx
dτ dτ dτ
1 1
h=0 h= 4
h= 2
x x x
c)
1
A saddle-node bifurcation occurs at hc = 4 and x = 12 .
d)
For h < hc the fish are sustainably harvested so that their population never decreases or increases.
For hc < h the fish are hunted to extinction (assuming that the fish population can’t become negative).
3.7.5
a)
dg k3 g 2
ġ = = k1 s0 − k2 g + 2
dt k4 + g 2
g g2
k4 d k4 k1 s0 k2 k4 g k42
= − + 2
k3 dt k3 k3 k4 1 + kg 2
4
g k1 s0 k2 k4 k3
x= s= r= τ= t
k4 k3 k3 k4
dx x2
= s − rx +
dτ 1 + x2
b)
If s = 0
dx x2
= 0 − rx + =0
dτ 1 + x2
√
1 ± 1 − 4r2
x = 0,
2r
And there are three real fixed points, two of which are positive since the square root is at most one, if
r < rc = 12 .
c)
g(0) = 0 corresponds to x(0) = 0, and increasing the parameter s translates the graph vertically.
1
There is only one fixed point at x = 0 for 2 < r since the graph is strictly decreasing, and x will stay close
to the fixed point as s is increased.
dx dx dx
dτ dτ dτ
s=0 s=1 s=0

r=1 r=1 r=1
x x x
1
At r ≤ 2 one, two, and three fixed points are possible with two saddle-node bifurcations occurring as s
increases.
dx dx dx
dτ dτ dτ
1 1
s=0 s ≈ 32 s ≈ 16
r = 12 r = 12 r = 12
x x x
dx dx dx
dτ dτ dτ
1 1
s= 8 s ≈ 16 s=0
1
r= 2 r = 12 r = 12
x x x
dx dx dx
dτ dτ dτ
s=0 s=0 s=0

r = 12 r = 21 7
r = 16
x x x
dx dx dx
dτ dτ dτ
7 1 7
s ≈ 128 s= 16 s ≈ 128
7 7 7
r = 16 r= 16 r = 16
x x x
dx
dτ
s=0
7
r = 16
x
Starting with s = 0, increasing s, and then returning to s = 0:

1
If 2 < r, the system will go towards the x = 0 stable fixed point since that is the only fixed point.
If r = 12 , the system can go back to the x = 0 fixed point or the stable half of the semistable fixed point
depending on how fast s changes, but noise in a real system would most likely bump x to the unstable side
of the semistable fixed point and then the system would go towards the x = 0 fixed point.
If r < 12 , the system will almost surely be near the far right stable fixed point, despite s returning to its
original value.
d)
dx d dx
The saddle-node bifurcations will occur when dτ = 0 and dx dτ = 0, which gives us two equations to solve
for r and s in terms of x.
x2
Å ã
d 2x
s − rx + 2
= −r + =0
dx 1+x (1 + x2 )2
2x
r=
(1 + x2 )2
x2 2x x2
s − rx + = s − x + =0
1 + x2 (1 + x2 )2 1 + x2
x2 (1 − x2 )
s=
(x2 + 1)2
These are saddle-node bifurcations.

e)
s
1 fixed point
e
od
-n
le
ode
dd
saddle-n 3 fixed points r

sa
3.7.7
a)
g(p)
β n1
p
K
n
βpn β ppn β
lim = lim Kn pn = lim n = βH(p − K)
n→∞ K + pn
n n→∞ + n→∞ Kn +1
pn pn p
where H(p) is the Heaviside step function.

b)
β β
(i) β < δK − α ⇒ h(K) > β (ii) δK − α = ⇒ h(K) =
2 2
ṗ ṗ
p p
α α α+β
K
δ δ δ
(iii) δK − α < 0 ⇒ h(K) < 0

ṗ
α+β
δ
c)
We’ll assume α ≥ 0 and β > 0 due to being biological measurements, as well as δK > β from the problem
statement. If we start α = 0 and increase α, we’ll get a saddle-node bifurcation at p ≈ K when the line h(p)
touches the top corner of g(p). Then we get another saddle-node bifurcation when h(p) touches the bottom
corner of g(p).
α ≈ δK − β α ≈ δK
ṗ ṗ
p p
α α+β
K K
δ δ
In summary, the fixed points are

α
0 < α < δK − β ⇒ p≈ stable
δ
α
α ≈ δK − β ⇒ p ≈ , K stable, semistable
δ
α α+β
δK − β < α < δK ⇒ p ≈ , K, stable, unstable, stable
δ δ
α+β
α ≈ δK ⇒ p ≈ K, semistable, stable
δ
α+β
δK < α ⇒ p≈ stable
δ
and now we can plot the bifurcation diagram.
p
δK − β δK
d)
Hysteresis is possible for this system, in the path shown through parameter space below for example.
p
δK − β δK
Starting at α = 0, p = 0, the system stays at the lower fixed point until α increases past δK and the system
jumps to the upper fixed point. The system exhibits hysteresis because the system does not jump back to
the lower fixed point once α decreases past δK, despite the lower fixed point existing. Instead, we have to
decrease α past δK − β in order for the system to jump back to the lower fixed point.
Here is a similar path viewed on the phase line where we have combined the graph of both functions as
g(p) − h(p) with a very large n value, K = 3, δ = 1, and β = 2.
ṗ ṗ
1
α=0 α= 2
p p
ṗ ṗ
3
α=1 α= 2
p p
ṗ ṗ
5
α=2 α= 2
p p
ṗ ṗ
7
α=3 α= 2
p p
ṗ ṗ
5
α=3 α= 2
p p
ṗ ṗ
3
α=2 α= 2
p p
ṗ ṗ
1
α=1 α= 2
p p
ṗ
α=0
p
The system exhibits hysteresis when p starts at the left fixed point but moves to the right fixed point once
3 = δK < α. But then α must be decreased to α < δK − β = 1 instead of α < δK = 3 for p to go back to
the original left fixed point.
4
Flows on the Circle
4.1 Examples and Definitions

4.1.1
For what real values of a does the equation θ̇ = sin(aθ) give a well-defined vector field on the circle?
The value of a must ensure that f (θ) = sin(aθ) is 2π periodic.
f (θ) = f (θ + 2π)

sin(aθ) = sin a(θ + 2π) = sin(aθ + 2πa) ⇒ 2πa = 2πz, z ∈ Z ⇒ a ∈ Z
4.1.3
θ̇ = sin(2θ)
π 3π
θ̇ = 0 ⇒ sin(2θ) = 0 ⇒ θ = 0, , π, , 2π
2 2
θ̇
4.1.5
θ̇ = sin(θ) + cos(θ)
√ Å
pi
ã
3π 7π
θ̇ = 0 ⇒ sin(θ) + cos(θ) = 2 sin θ + =0⇒θ= ,
4 4 4
θ̇
65
66 Chapter 4: Flows on the Circle
4.1.7
θ̇ = sin(kθ) where k is a positive integer.

p
θ̇ = 0 ⇒ sin(kθ) = 0 ⇒ θ = π , p = 0, 1, 2, ...2k
k
Example: k = 3
θ̇
4.1.9
Exercise 2.6.2 and 2.7.7 does not carry over to vector fields on the circle because of the different interpreta-
tions of points. Put simply, the definition of periodic for vector fields on the line is fundamentally different
from vector fields on the circle.
A vector field on the line is an interval with each value in the interval corresponding to a value of x that is
interpreted as a unique point. Periodicity is defined as the particle returning to the same location after a
time T , which can’t occur.
A vector field on a circle with interval [0, 2π) interprets θ(t) = 1 and θ(t + T ) = 1 + 2π as the same
point. Periodicity is defined as the particle returning to the same location modulo 2π, which can occur. The
particle isn’t really where it started, but unlike vector fields on the line, the two locations are considered
identical.
4.2 Uniform Oscillator

4.2.1
Using common sense, we know the two bells will ring again in 12 seconds.
Using the method of Example 4.2.1
1 −1 1 1 −1
Å ã Å ã Å ã−1
1 1
− = − = = 12
T1 T2 3 4 12
4.2.3
Using the methods of this section
1 −1 1 −1
Å ã Å ã Å ã−1
1 1 11 12
− = − = = hours
T1 T2 1 12 12 11
So the hour and minute hands will be aligned again 1 hour, 5 minutes, and 27 seconds, to the nearest second,
later.
We also know that the hour and minute hands will be aligned again in 12 hours. During this time, the
minute hand will pass the hour hand 11 times. The time between passes is equal, so the minute hand will
1
lap the hour hand every 11 of the perimeter, which is 5 minutes and 27 seconds, to the nearest second.
4.3 Nonuniform Oscillator

4.3.1
∞
√ √
Z
dx
Tbottleneck = 2
x = r tan(θ) dx = r sec2 (θ)dθ
−∞ r + x
Z π2 √ Z π2 √
r sec2 (θ) r sec2 (θ)
= 2 dθ = dθ
−π 2
r + r tan (θ) −π
2
r(1 + tan2 (θ))
Z π2 Z π2
sec2 (θ) 1
= √ 2
dθ = √ dθ
−π 2
r sec (θ) −π 2
r
π
=√
r
4.3.3
θ̇ = µ sin(θ) − sin(2θ)
µ = −2 subcritical pitchfork bifurcation at θ = π

µ = 2 subcritical pitchfork bifurcation at θ = 0
θ̇ µ = −3
θ̇ µ = −2
θ̇ µ = −1
θ̇ µ=0
θ
θ̇ µ=1
θ̇ µ=2
θ̇ µ=3
4.3.5
θ̇ = µ + cos(θ) + cos(2θ)
µ = −2 saddle-node bifurcation at θ = 0
µ = 0 saddle-node bifurcation at θ = π
9
µ= 8saddle-node bifurcation at
»
5
θ = 2 arctan 3
»
5
θ = 2π − 2 arctan 3
θ̇ µ = −3
θ̇ µ = −2
θ̇ µ = −1
θ̇ µ=0
θ
9
θ̇ µ=
16
9
θ̇ µ=
8
θ̇ µ=2
4.3.7
sin(θ)
θ̇ =
µ + sin(θ)
π
µ = −1 saddle-node bifurcation at θ = 2
µ = 0 weird
The function hugs closer and closer to the θ̇ = 1 line and the two vertical asymptotes, forming right
angles there while simultaneously disappearing.
3π
µ = 1 saddle-node bifurcation at θ = 2
Note: Some of the fixed points on the circles are not actually fixed points but correspond to infinite disconti-
nuities. However, the discontinuities act like stable or unstable fixed points, since the point of discontinuity
is either attracting or repelling respectively.
θ̇ µ = −2
−3
θ̇ µ=
2
θ̇ µ = −1
θ
−1
θ̇ µ=
2
θ̇ µ=0
1
θ̇ µ=
2
θ
θ̇ µ=1
3
θ̇ µ=
2
θ̇ µ=2
4.3.9
a)
ẋ = r + x2 x = ra u t = rb τ
dx ra du du 2
= b = ra−b = r + x2 = r + (ra u) = r + r2a u2
dt r dτ dτ
b)
1
r2a u2 = ru2 ⇒ a =
2
du du 1 −1
ra−b =r ⇒a−b= −b=1⇒b=
dτ dτ 2 2
4.4 Overdamped Pendulum 75
4.4 Overdamped Pendulum

4.4.1
mL2 θ̈ + bθ̇ + mgL sin(θ) = Γ

d2 θ dθ
mL2 +b + mgL sin(θ) = Γ
dt2 dt
dθ 1 dθ d2 θ 1 d2 θ
t = Tτ = =
dt T dτ dt2 T 2 dτ 2
2 2
mL d θ b dθ
+ + mgL sin(θ) = Γ
T 2 dτ 2 T dτ
2
L d θ b dθ Γ
+ + sin(θ) =
gT 2 dτ 2 mgLT dτ mgL
dθ d2 θ
We want the dτ to be O(1) so that all the terms except the dτ 2 term are of the same order.
b b
=1⇒T =
mgLT mgL
m2 gL3 d2 θ dθ Γ
2 2
+ + sin(θ) =
b dτ dτ mgL
d2 θ d2 θ
Then we want the dτ 2 coefficient to be very small so the system relaxes quickly and we can neglect the dτ 2
term.
m2 gL3
1 ⇒ m2 gL3 b2
b2
So as long as the above condition is satisfied, the second-order differential equation can be well approximated
by the first-order differential equation after an initial transient.
4.4.3
The length of the simulation increases in each graph in order to compare the oscillations.
θ̇ γ = 100
The oscillations are almost indistinguishable due to the vertical scaling for 1 γ because the derivative is
approximately 100 all the time.
θ̇ γ=2
The oscillations look irregular as γ → 1 because the pendulum is slowing down significantly when going
through the horizontal portion of the swing.
θ̇ γ = 1.1
These oscillations continue to sharpen and lengthen as γ → 1, and the period will become infinity if γ = 1
because the pendulum never passes through the horizontal portion of the swing.
θ̇ γ=1
θ̇ γ = 0.9
Now the torque isn’t strong enough to continue the oscillations, so the pendulum comes to rest where the
torque balances the force of gravity.
4.5 Fireflies 77
θ̇ γ = 0.1
θ̇ γ=0
Now there is no torque and the pendulum settles to the bottom of the swing.
4.5 Fireflies
4.5.1

−π π
φ ≤φ≤

2 2
φ̇ = Ω − ω − Af (φ) f (φ) =
π − φ π 3π
≤φ≤

2 2
a)
f (φ)
b)
Entrainment occurs when there are still fixed points of the φ̇ equation. The maximum and minimum of f (φ)
are ± π2 , so
φ̇ = 0 = Ω − ω − Af (φ)
Ω − ω = Af (φ)
−π Ω−ω π
A≤ ≤ A
2 A 2
π
|Ω − ω| ≤ A
2
c)
The phase difference is simply the φ value of the stable fixed point.
φ̇ = 0 = Ω − ω − Af (φ∗ ) = Ω − ω − A(π − φ∗ )
ω−Ω π
φ∗ = +π |Ω − ω| ≤ A
A 2
d)
Z 2π Z 2π Z 3π
dφ dφ 2 dφ
Tdrift = = =
0 φ̇ 0 Ω − ω − Af (φ) −π Ω − ω − Af (φ)
2
Z π2 Z 3π
dφ 2 dφ
= +
−π Ω − ω − Af (φ) π Ω − ω − Af (φ)
2 2
Z π2 Z 3π
dφ 2 dφ
= +
−π Ω − ω − Aφ π Ω − ω − A(π − φ)
2 2
ò π2 ò 3π
−1
ï ï
1 2
= ln Ω − ω − Aφ + ln Ω − ω − A(π − φ)
A −π A π
2 2
4.5 Fireflies 79
òπ ò −π
−1
ï ï
2 1 2
= ln Ω − ω − Aφ + ln Ω − ω − Aφ
A −π A π
2 2
ï ò −π ï ò −π
1 2 1 2
= ln Ω − ω − Aφ + ln Ω − ω − Aφ
A π A π
2 2
ï ò −π
2 2
= ln Ω − ω − Aφ
A π
ã2
−π
Å Å Å ãã
2 π
= ln Ω − ω − A − ln Ω − ω − A
A 2 2
Å Å ã Å ãã
2 π π
= ln Ω − ω − A − ln Ω − ω + A
A 2 2
4.5.3
θ̇ = µ + sin(θ)
a)
θ̇
µ = 0.9
The system is periodic, so the unique globally attracting “rest state” and “threshold” are the stable and
unstable fixed points respectively.
Also, a large enough stimulus will shift the system past the unstable-fixed-point “threshold” and will only
return to the stable-fixed-point “rest state” after going through most of the 2π-length periodicity.
b)

cos θ(t)
t

cos θ(t)

cos θ(t)

cos θ(t)
4.6 Superconducting Josephson Junctions

4.6.1
a)
I
φ̇ = − sin(φ)
Ic
I
φ Ic = 1.1
t
I
φ Ic = 10
The vertical axis in the second case has been scaled for a reasonable size.
b) Å ã
~ ~ I
V (t) = φ̇ = − sin(φ)
2e 2e Ic
I
V (t) Ic = 1.1
I
V (t) Ic = 10
The vertical axis in the second case has been scaled for a reasonable size.
4.6.3
a)
I
φ̇ = − sin(φ)
Ic
dV I −I
− = − sin(φ) ⇒ V (φ) = φ − cos(φ)
dφ Ic Ic
V (φ) 6= V (φ + 2π)
So V (φ) is not a single-valued function on the circle.

b)
I
V (φ) Ic = 0.1
I
V (φ) Ic = 0.9
I
V (φ) Ic =1
φ
I
V (φ) Ic = 1.1
I
V (φ) Ic = 10
I
V (φ) Ic = 1.1
c)
Increasing I brings the potential well and hill closer together, which eventually collide and annihilate.
4.6.5
a)
Ib = Ia + IR
b)
The current running through the middle is Ia , which has to be the same current running through all sections
because they are in series. The current in a section is split between the current through the Josephson
junction Ic sin(φi ) and the current running through the resistor because they are in parallel. This gives
Vk
Ia = Ic sin(φk ) +
r
c)
Vk
Ia = Ic sin(φk ) +
r
Vk = rIa − rIc sin(φk )
~ I
V = φ̇ φ̇ = − sin(φ)
2e Ic
Å ã
Ia ~
Vk = rIc − sin(φk ) = φ˙k
Ic 2e
d)
Using Kirchhoff’s voltage law, the voltage from the top to the bottom of the circuit is the same going through
the Josephson junctions and the resistor.
N
X
Vi = VR
i=1
N
~ X
φ̇i = IR R = (Ib − Ia ) R
2e i=1
N
~ X
Ib = Ia + φ̇i
2eR i=1
N
Vk ~ X
= Ic sin(φk ) + + φ̇i
r 2eR i=1
N
~ ˙ ~ X
= Ic sin(φk ) + φk + φ̇i
2er 2eR i=1
e)
N
~ ˙ ~ X
Ib = Ic sin(φk ) + φk + φ̇i
2er 2eR i=1
N N N
X ~ X N~ X
N Ib = Ic sin(φi ) + φ̇i + φ̇i
i=1
2er i=1 2eR i=1
N N N
Ic X ~ X ~ X
Ib = sin(φi ) + φ̇i + φ̇i
N i=1
2N er i=1 2eR i=1
N
~ ˙ ~ X
= Ic sin(φk ) + φk + φ̇i
2er 2eR i=1
N N
Ic X ~ X ~ ˙
sin(φi ) + φ̇i = Ic sin(φk ) + φk
N i=1 2N er i=1 2er
N N
~ X ~ ˙ Ic X
φ̇i = Ic sin(φk ) + φk − sin(φi )
2N er i=1 2er N i=1
N N
~ X Nr N~ ˙ r X
φ̇i = Ic sin(φk ) + φk − Ic sin(φi )
2eR i=1 R 2eR R i=1
N
~ ˙ ~ X
Ib = Ic sin(φk ) + φk + φ̇i
2er 2eR i=1
N
~ ˙ Nr N~ ˙ r X
= Ic sin(φk ) + φk + Ic sin(φk ) + φk − Ic sin(φi )
2er R 2eR R i=1
Å ã Å ã 2
Nr 1 N r X
= 1+ Ic sin(φk ) + ~ + φ˙k − Ic sin(φj )
R 2er 2eR R j=1
Å ã Å ã 2
1 N Nr r X
~ + φ˙k = Ib − 1 + Ic sin(φk ) + Ic sin(φj )
2er 2eR R R j=1
2
~(R + N r) ˙ R + Nr r X
φk = Ib − Ic sin(φk ) + Ic sin(φj )
2erR R R j=1
2
~(R + N r) ˙ RIb R + Nr 1 X
φ k = − sin(φk ) + sin(φj )
2N er2 Ic N rIc Nr N j=1
2
dφk 1 X
= Ω + a sin(φk ) + sin(φj )
dτ N j=1
RIb −(R + N r) 2N er2 Ic
Ω= a= τ= t
N rIc Nr ~(R + N r)
5
Linear Systems
5.1 Definitions and Examples

5.1.1
Harmonic oscillator
ẋ = v v̇ = −ω 2 x
a)
Show that the orbits are given by ω 2 x2 + v 2 = C where C is any nonnegative constant.
Divide the first equation by the second equation.
ẋ v
= ⇒ −ω 2 xẋ = v v̇ ⇒ ω 2 xẋ + v v̇ = 0
v̇ −ω 2 x
Integrate
ω2 2 1 2
x + v = D ⇒ ω 2 x2 + v 2 = C
2 2
b)
Show that this is equivalent to conservation of energy.
The harmonic oscillator with mass m and spring constant k stores a constant amount of energy as the
sum of kinetic and potential energy.
Potential energy
−k 2
Z
PE = − kxdx = x +D
2
Kinetic energy
−m 2
KE = v
2
Total energy
−k 2 −m 2 −2 2D k
E = P E + KE = x +D+ v ⇒ E+ = x2 + v 2
2 2 m m m
If we set
−2 2D k
C= E+ ω2 =
m m m
then we get the same equation.
87
88 Chapter 5: Linear Systems
5.1.3
ẋ = −y ẏ = −x
Ñ é Ñ éÑ é
ẋ 0 −1 x
=
ẏ −1 0 y
5.1.5
ẋ = 0 ẏ =x+y
Ñ é Ñ éÑ é
ẋ 0 0 x
=
ẏ 1 1 y
5.1.7
ẋ = x ẏ = x + y
y
x
5.1.9
ẋ = −y ẏ = −x
a)
y
b)
xẋ − y ẏ = x(−y) − y(−x) = −xy + xy = 0

Z Z
xẋ − y ẏ dt = 0 dt
x2 − y 2 = C
c)
From the vector field, it looks like y = x and y = −x are the stable and unstable manifolds respectively. We
can check by plugging y = x and y = −x into the system.
y = x gives ẋ = −x and ẏ = −y, which has a stable fixed point at the origin.
y = −x gives ẋ = x and ẏ = y, which has an unstable fixed point at the origin.
d)
u=x+y v =x−y
u̇ = ẋ + ẏ = −y − x = −u
u̇ = −u ⇒ u(t) = Ce−t
u(0) = u0 ⇒ C = u0
u(t) = u0 e−t
v̇ = ẋ − ẏ = −y + x = v
v̇ = v ⇒ v(t) = Cet
v(0) = v0 ⇒ C = v0
v(t) = v0 et
e)
The stable and unstable manifold occurs when u = 0 and v = 0 respectively.
f)
u+v u0 e−t + v0 et (x0 + y0 )e−t + (x0 − y0 )et

x= = =
2 2 2
u−v u0 e−t − v0 et (x0 + y0 )e−t − (x0 − y0 )et
y= = =
2 2 2
5.1.11
a)
ẋ = y ẏ = −4x
ẍ = ẏ = −4x
y0
x(t) = x0 cos(2t) + sin(2t)
2
y(t) = −2x0 sin(2t) + y0 cos(2t)

δ < x0 ⇒ x(t), y(t) < 2x0 =
Liapunov stable
b)
ẋ = 2y ẏ = x
ẍ = 2ẏ = 2x
Ä√ ä √ Ä√ ä
x = x0 cosh 2t + 2y0 sinh 2t
x0 Ä √ ä Ä√ ä
y = √ sinh 2t + y0 cosh 2t
2
Trajectories go away from the origin

None of the above
c)
ẋ = 0 ẏ = x
x = x0 y = x0 t + y0

None of the above
d)
ẋ = 0 ẏ = −y
x = x0 y = y0 e−t

δ = x0 , y0 ≥ x(t), y(t) ≥ x0 , 0 for t ≥ 0
Liapunov stable
e)
ẋ = −x ẏ = −5y
x = x0 e−t y = y0 e−5t

δ = x0 , y0 ≥ x(t), y(t) for t ≥ 0
Asymptotically stable
f)
ẋ = x ẏ = y
x = x0 et y = y0 et

None of the above
5.1.13
Going back to multivariable calculus, a saddle point on a graph is neither a local minimum nor a local
maximum, but the first partial derivatives of x and y both equal zero at the saddle point. The typical
example
looks just like a horse saddle, and the level sets of graph make this pattern. The phase plane of a saddle
point for an ODE looks the same, so we call these fixed points saddle points.
5.2 Classification of Linear Systems

5.2.1
ẋ = 4x − y ẏ = 2x + y
a) Ñ é Ñ éÑ é
ẋ 4 −1 x
=
ẏ 2 1 y

4 − λ −1

= (4 − λ)(1 − λ) + 2 = λ2 − 5λ + 6 = (λ − 2)(λ − 3) = 0
1 − λ

2
λ1 = 2 λ2 = 3
Ñ é Ñ é Ñ é
4 − λ1 −1 2 −1 1
~v1 = ~v1 = ~0 ⇒ ~v1 =
2 1 − λ1 2 −1 2
Ñ é Ñ é Ñ é
4 − λ2 −1 1 −1 1
~v2 = ~v2 = ~0 ⇒ ~v2 =
2 1 − λ2 2 −2 1
b) Ñ é Ñ é Ñ é
x 1 1
= c1 eλ1 t~v1 + c2 eλ2 t~v2 = c1 e 2t
+ c2 e 3t
y 2 1
c)
Unstable node
d) Ñ é Ñ é Ñ é Ñ é
x(0) 3 1 1
= = c1 + c2 ⇒ c1 = 1 and c2 = 2
y(0) 4 2 1
Ñ é Ñ é Ñ é
x 2t
1 3t
2
=e +e
y 2 2
5.2.3
ẋ = y ẏ = −2x − 3y
Ñ é Ñ éÑ é
ẋ 0 1 x
=
ẏ −2 −3 y
Ñ é Ñ é
1 1
λ1 = −2 ~v1 = λ2 = −1 ~v2 =
−2 −1
Stable node
y
5.2.5
ẋ = 3x − 4y ẏ = x − y
Ñ é Ñ éÑ é
ẋ 3 −4 x
=
ẏ 1 −1 y
Ñ é
2
λ1 = 1 ~v1 =
1
Unstable degenerate node

y
5.2.7
ẋ = 5x + 2y ẏ = −17x − 5y
Ñ é Ñ éÑ é
ẋ 5 2 x
=
ẏ −17 −5 y
Ñ é Ñ é
5 + 3i 5 − 3i
λ1 = 3i ~v1 = λ2 = −3i ~v2 =
−17 −17
Center
5.2.9
ẋ = 4x − 3y ẏ = 8x − 6y
Ñ é Ñ éÑ é
ẋ 4 −3 x
=
ẏ 8 −6 y
Ñ é Ñ é
1 3
λ1 = −2 ~v1 = λ2 = 0 ~v2 =
2 4
Nonisolated fixed point

y
5.2.11
Ñ é Ñ éÑ é
ẋ a b x
=
ẏ 0 a y

a − λ b

= (a − λ)2 = 0 ⇒ λ = a
a − λ

0
Ñ é Ñ é Ñ é
a−λ b 0 b 1
~v = ~v = ~0 ⇒ ~v =
0 a−λ 0 0 0
ẏ = ay ⇒ y = c2 eat
ẋ = ax + by = ax + c2 beat
ẋ − ax = c2 beat
d −at
e−at ẋ − ae−at x = e x = c2 b
Z dt
e−at x = c2 b dt = c1 + c2 bt
x = c1 eat + c2 bteat
Ñ é Ñ é Ñ Ñ é Ñ éé
x at
1 b 0
= c1 e + c2 eat t +
y 0 0 1
Stable degenerate node
y
5.2.13
mẍ + bẋ + kx = 0 b>0
a)
−k b −k b
ẋ = y ẏ = ẍ = x − ẋ = x− y
m m m m
Ñ é Ñ éÑ é
ẋ 0 1 x
=
−k −b
ẏ m m y
b)
0 − λ 1

= −λ −b − λ + k = 1 (mλ2 + bλ + k) = 0
Å ã

−k
m −b
− λ
m m m
m
√
−b ± b2 − 4mk
λ=
2m
If the inside of the square root is positive, then the origin is stable because the square root can’t exceed
−b, and both roots will be nonnegative. If the inside of the square root is negative, then the −b term still
determines the stability. b > 0 ensures the origin is asymptotically stable.
y
m = 12 , b = 32 , k = 1 ⇒ b2 − 4mk > 0 ⇒ Stable node

m = 1, b = 1, k = 1 ⇒ b2 − 4mk < 0 ⇒ stable spiral

5.3 Love Affairs 101
m = 1, b = 2, k = 1 ⇒ b2 − 4mk = 0 ⇒ stable degenerate node

c)
Stable node ←→ overdamped
Stable degenerate node ←→ critically damped
Stable spiral ←→ underdamped
5.3 Love Affairs

5.3.1
See the “Answers to Selected Exercises” section in the textbook.
5.3.3
Ṙ = aJ J˙ = bR
Ñ é Ñ éÑ é
Ṙ 0 a R
=
J˙ b 0 J

0 − λ a

= λ2 − ab = 0
0 − λ

b
ab < 0 ⇒ center
0 < ab ⇒ saddle point
5.3.5
Ṙ = aR + bJ J˙ = bR + aJ
Ñ é Ñ éÑ é
Ṙ a b R
=
J˙ b a J

a − λ b

= (a − λ)2 − b2 = 0 ⇒ λ = a ± b
a − λ

b
a < b ⇒ saddle point

a < b < 0 ⇒ stable node
0 < b < a ⇒ unstable node
6
Phase Plane
6.1 Phase Portraits

6.1.1
ẋ = x − y ẏ = 1 − ex
(x, y) = (0, 0) is the fixed point.

y
103
104 Chapter 6: Phase Plane
6.1.3
ẋ = x(x − y) ẏ = y(2x − y)
x
6.1.5
ẋ = x(2 − x − y) ẏ = x − y
(x, y) = (0, 0) and (1, 1) are the fixed points.

y
x
6.1.7
ẋ = x + e−y ẏ = −y
−t
et−c2 et−c2 e
y = c2 e−t x = c1 e t − +
c2 c2
e−2 −1
If we pick c1 = 2 and c1 = 2 , then the growing exponentials destructively interfere with each other,
leaving the stable manifold.
et Ä −2e−t ä
y = 2e−t x= e −1
2
y
x
6.1.9
Dipole fixed point
ẋ = 2xy ẏ = y 2 − x2

y
x
6.1.11
Parrot
1 6
ẋ = y + y 2 ẏ = −x + y − xy + y 2
5 5
x
6.2 Existence, Uniqueness, and Topological Consequences 109
6.1.13
6.2 Existence, Uniqueness, and Topological Consequences

6.2.1
Trajectories in a phase portrait never intersect, despite trajectories that approach a stable fixed point. Tra-
jectories never actually reach the stable fixed point because they decay like an exponential function, never
vanishing but always decreasing.
However, less well-behaved ODEs such as having nonunique solutions can reach a fixed point in finite time.
6.3 Fixed Points and Linearization

6.3.1
ẋ = x − y ẏ = x2 − 4
(x, y) = (−2, −2) and (2, 2) are the fixed points.

y
6.3.3
ẋ = 1 + y − e−x ẏ = x3 − y

6.3.5
ẋ = sin(y) ẏ = cos(x)
1

(x, y) = n1 − 2 π, n2 π are the fixed points.
6.3.7
See graphs for Exercises 6.3.1, 6.3.3, and 6.3.5.
6.3.9
ẋ = y 3 − 4x ẏ = y 3 − y − 3x
a)
(x, y) = (−2, −2), (0, 0), and (2, 2) are the fixed points.
Ñ é
−4 3y 2
A=
−3 3y 2 − 1
Ñ é
−4 12
A(−2,−2) = ∆ = −8 ⇒ saddle point
−3 11
Ñ é
−4 0
A(0,0) = ∆ = 4 τ = −5 ⇒ unstable spiral
−3 −1
Ñ é
−4 12
A(2,2) = ∆ = −8 ⇒ saddle point
−3 11
b)
x = y ⇒ ẋ = x3 − 4x = y 3 − 4y = ẏ
So as long as the initial condition satisfies x(0) = y(0), then x(t) = y(t) since the two variables and equations
are indistinguishable.
c)
u=x−y
u̇ = ẋ − ẏ = y 3 − 4x − (y 3 − y − 3x) = −x + y = −u
u(t) = ce−t
|x(t) − y(t)| = |u(t)| = ce−t

lim |x(t) − y(t)| = lim ce−t → 0

t→∞ t→∞
d)
See part (e).
e)
y
x
6.3.11
1
ṙ = −r θ̇ =
ln(r)
a)
r(t) = r0 e−t
1 1 1 1
θ̇ = = = =
ln(r) ln(r0 e−t ) ln(r0 ) + ln(e−t ) ln(r0 ) − t
θ(t) = − ln | ln(r0 ) − t| + ln(ln(r0 )) + θ0
Å ã
ln(r0 )
= ln + θ0
| ln(r0 ) − t|
b)
lim r(t) = lim r0 e−t → 0 ⇒ stable

t→∞ t→∞
Å ã
ln(r0 )
lim |θ(t)| = lim ln + θ0 → ∞ ⇒ spiral
t→∞ t→∞ | ln(r0 ) − t|
c)
p y
r= x2 + y 2 θ = arctan
x
dp 2 x ẋ + y ẏ p
ṙ = x + y2 = p = −r = − x2 + y 2
dt x2 + y 2
d y xẏ − y ẋ 1 1 2
θ̇ = arctan = 2 = = =
x + y2 ln(x2 + y 2 )
p
dt x ln(r) ln x2 + y 2
xẋ + y ẏ p
p = − x2 + y 2 xẋ + y ẏ = −(x2 + y 2 )
2
x +y 2
xẏ − y ẋ 2 2(x2 + y 2 )
= xẏ − y ẋ =
x2 + y 2 ln(x2 + y 2 ) ln(x2 + y 2 )
2(x2 + y 2 )
x(xẋ + y ẏ) − y(xẏ − y ẋ) = (x2 + y 2 )ẋ = −x(x2 + y 2 ) − y
ln(x2 + y 2 )
2y
ẋ = −x −
ln(x2 + y2 )
2(x2 + y 2 )
y(xẋ + y ẏ) + x(xẏ − y ẋ) = (x2 + y 2 )ẏ = −y(x2 + y 2 ) + x
ln(x2 + y 2 )
2x
ẏ = −y +
ln(x2 + y 2 )
d)
4y 2
á ë
4xy 2
−1 −
(x + y ) ln2 (x2 + y 2 )
2 2 2 2 2 2 2
(x + y ) ln (a + y ) ln(a2 + y2 )
A=
2 4x2 −4xy
− 2 −1
+ y ) (x + y 2 ) ln2 (x2 + y 2 )
ln(x22
(x2 + y 2 ) ln2 (x2 + y 2 )
Ñ é
−1 0
A(0,0) = ẋ = −x ẏ = −y ⇒ stable star
0 −1
6.3.13
ẋ = −y − x3 ẏ = x
r 2 = x2 + y 2 2rṙ = 2xẋ + 2y ẏ
xẋ + y ẏ x(−y − x3 ) + yx −x4
ṙ = = p =p < 0 for (x, y) 6= (0, 0)
r x2 + y 2 x2 + y 2
Ñ é
−3x2 −1
A=
1 0
Ñ é
0 −1
A(0,0) = ∆ > 0 τ = 0 ⇒ center
1 0
6.3.15
ṙ = r(1 − r2 ) θ̇ = 1 − cos(θ)
(r, θ) = (0, 0) and (1, 0) are the fixed points.

(1,0) is attracting, but all trajectories except those starting on the positive x-axis have to traverse near the
unit circle before reaching the fixed point.
6.3.17
ẋ = xy − x2 y + y 3 ẏ = y 2 + x3 − xy 2

6.4 Rabbits versus Sheep

6.4.1
ẋ = x(3 − x − y) ẏ = y(2 − x − y)
(x, y) = (0, 0), (0, 2), and (3, 0) are the fixed points.
6.4.3
ẋ = x(3 − 2x − 2y) ẏ = y(2 − x − y)
(x, y) = (0, 0), (0, 2), and ( 23 , 0) are the fixed points.
6.4.5
N1
Å ã
Ṅ1 = r1 N1 1 − − b1 N1 N2 Ṅ2 = r2 N2 − b2 N1 N2
K1
N1
Å ã
dN1 dN2
= r1 N1 1 − − b1 N1 N2 = r2 N2 − b2 N1 N2
dt K1 dt
t = uτ N1 = vx N2 = wy
v dx vx w dy
Å ã
= r1 vx 1 − − b1 vwxy = r2 wy − b2 vwxy
u dτ K1 u dτ
vx b1 w r2
Å ã
1 dx 1 dy
=x 1− − xy = y − xy
r1 u dτ K1 r1 b2 uv dτ b2 v
1 1 b1 w r1 1 r1 r1
=1⇒u= =1⇒w= = =1⇒v=
r1 u r1 r1 b1 b2 uv b2 v b2
r1 x r2
Å ã
dx dy
=x 1− − xy = y − xy
dτ b2 K1 dτ r1
r2 r1
ρ= κ=
r1 b2 K1
dx dy
= x(1 − κx − y) = y(ρ − x)
dτ dτ
1
ρ=3 κ= 2
(x, y) = (0, 0), (0, 1), (2, 0), and (3, 0) are the fixed points.
1
ρ=3 κ= 3
(x, y) = (0, 0), (0, 1), and (3, 0) are the fixed points.
1
ρ=3 κ= 4
(x, y) = (0, 0), (0, 1), (3, 0), 3, 14 , and (4, 0) are the fixed points.

The top and bottom graphs have a different number of fixed points, with the middle graph being the
borderline case. Except by starting with an initial condition of zero population, the top two graphs grow
infinitely in y, and x decays to zero. Depending on the initial condition, the bottom graph can grow infinitely
in y, and x decays to zero, but it can also go towards a stable fixed point where x becomes finite and y
decays to zero.
6.4.7
n˙1 = G1 (N0 − α1 n1 − α2 n2 )n1 − K1 n1 n˙2 = G2 (N0 − α1 n1 − α2 n2 )n2 − K2 n2
a)
Ñ é
G1 (N0 − 2α1 n1 − α2 n2 ) − K1 −G1 α2 n2
A=
−G2 α1 n1 G2 (N0 − α1 n1 − 2α2 n2 ) − K2
Ñ é
G1 N0 − K1 0
A(0,0) =
0 G2 N0 − K2
λ1 = G1 N0 − K1 λ2 = G2 N0 − K2
So the origin is unstable if G1 N0 − K1 > 0 or G2 N0 − K2 > 0.

b)
Ä ä Ä ä
The other two fixed points that exist are 0, G2αN20G−K
2
2
and G1αN10G−K
1
1
,0 .
Ñ é
G1 K2 G1 K2
G2 − K1 −G1 N0 + G2
A 0,
G2 N0 −K2
=
α2 G2
0 −G2 N0 + K2
Ñ é
−G1 N0 + K1 0
A G1 N0 −K1
,0
=
G2 K 1 G2 K1
α1 G1
−G2 N0 + G1 G1 − K2
Since these are triangular matrices, the eigenvalues are on the diagonals. Notice that two of the eigenvalues,
G1 K2 G2 K1
G2 − K1 and G1 − K2 , will have opposite signs or both be zero, meaning at most one of the fixed points
can be stable.
The other pair of eigenvalues doesn’t have that relationship and can be positive, negative, or zero de-
pending on how big N0 is.
If you look closely, you’ll see that the eigenvalues of a fixed point and its coordinate are somewhat sim-
ply related. This relationship makes for some interesting graphs.
c)
N0 = 5 G1 = 1 G2 = 1 α1 = 1 α2 = 1 K1 = 7.5 K2 = 7.5
y
x
N0 = 5 G1 = 1 G2 = 1 α1 = 1 α2 = 1 K1 = 7.5 K2 = 2.5
y
N0 = 5 G1 = 1 G2 = 1 a1 = 1 a2 = 1 K1 = 2.5 K2 = 7.5
y
x
N0 = 5 G1 = 1 G2 = 1 α1 = 1 α2 = 1 K1 = 2.5 K2 = 2.5
y
There are four qualitatively different phase portraits. n1 and n2 decay to the origin, n1 axis, n2 axis, or an
infinite number of fixed points. No other phase portraits are possible because the nullclines are the axes and
parallel lines.
6.4.9
I˙ = I − αC Ċ = β(I − C − G) I, C, G ≥ 0, 1 < α < ∞, 1≤β<∞
a)
Ä ä
αG G
(x, y) = α−1 , α−1 is the fixed point.
Ñ é
1 −α
A=
β −β
Ñ é
1 −α
A αG G
=
α−1 , α−1
β −β
p
−(β − 1) ± (β − 1)2 − 4(α − 1)β
λ1,2 =
2
As long as β > 1, the real part is negative, making the fixed point stable. The sign of the expression inside
the square root determines the type of stable behavior.
(β−1)2
α> 4β + 1 ⇒ stable spiral
(β−1)2
α= 4β + 1 ⇒ stable star
(β−1)2
α< 4β + 1 ⇒ stable node
And the expression is purely imaginary when

β = 1 ⇒ Center
b)
Ä ä
αG0 G0 1
G = G0 + kI ⇒ (x, y) = α(1−k)−1 , α(1−k)−1 is the fixed point, and if k < 1 − α = kc , then it’s in the
positive quadrant.
Ñ é
1 −α
A=
β(1 − k) −β
Ñ é
1 −α
A αG0
,
G0
=
α(1−k)−1 α(1−k)−1
β(1 − k)
−β
»
−(β − 1) + (β − 1)2 − 4 α(1 − k) − 1 β
λ1 =
» 2

−(β − 1) − (β − 1)2 − 4 α(1 − k) − 1 β
λ2 =
2
If k > kc and β > 1, then λ2 < 0 < λ1 .

Ñ é Ñ é
β+λ1 β+λ2
β(1−k) β(1−k)
v1 = v2 =
1 1
The attracting eigendirection is v1 , which has positive slope and is present in the positive quadrant. The
economy will follow this line, with I and C always increasing.
c) Å √ √ ã
α−1± (α−1)2 −4α2 G0 k α−1± (α−1)2 −4α2 G0 k
G = G0 + kI 2 ⇒ (x, y) = 2αk , 2α2 k are the fixed points.
(α−1)2
If G0 < 4α2 k then there are two fixed points in the positive quadrant.
(α−1)2
If G0 = 4α2 k then there is only one fixed point in the positive quadrant.
(α−1)2
If G0 > 4α2 k then there are no real fixed points.
α = 2 β = 2 k = 0.05 G0 = 0.75
I
α = 2 β = 2 k = 0.01 G0 = 1.25
I
α = 2 β = 2 k = 0.01 G0 = 1.75
In all cases, the trajectories can result in the national income I going negative, but the amount of consumer
spending C is still positive. Also, it looks like the government spending lots of money is a good thing. With
too little spending the economy can get stuck at a fixed point, but a lot of government spending with the
right initial condition will result in the national income growing continually, with consumer spending not
increasing much. This is the unstable manifold of the saddle point.
6.4.11
ẋ = rxz ẏ = ryz ż = −rxz − ryz
a)
x+y+z =1 (The total sum of leftists, centrists, and rightists should add up to 1.)
ẋ + ẏ + ż = rxz + ryz − rxz − ryz = 0
So the sum never changes, also known as invariant.

b)
(x, y, z) = (x, y, 0) and (0, 0, z) are the fixed points.
á ë
rz 0 rx
A= 0 rz ry
−rz −rz −rx − ry
á ë
0 0 rx
A(x,y,0) = 0 0 ry
0 0 −rx − ry
λ1 = 0 λ2 = 0 λ3 = −rx − ry
á ë
rz 0 0
A(0,0,z) = 0 rz 0 λ1 = 0 λ2 = 0 λ3 = rz
−rz −rz 0
c)
For r > 0, a purely centrist government z = 1 is unstable. If there are any leftists or rightists, then the
centrists are converted to each side with the limiting number of leftists and rightists determined by the initial
values.
For r < 0, a purely centrist government z = 1 is stable. If there are no centrists z = 0, then the leftists and
rightists will stay at their initial value because neither side can ever convert opposite members.
6.5 Conservative Systems

6.5.1
ẍ = x3 − x
a)
ẋ = y ẏ = x3 − x
(x, y) = (−1, 0), (0, 0), and (1, 0) are the fixed points.
b)
Z
1 1 2 1 2 1 4
E = ẋ2 + −(x3 − x) dx = ẋ + x − x
2 2 2 4
c)
y
6.5.3
ẍ = a − ex
Z
1 2 1 2
E= ẋ + −(a − ex ) dx = ẋ − ax + ex
2 2
ẋ = y ẏ = a − ex
a<0
y
a=0
y
x
a>0
6.5.5
ẍ = (x − a)(x2 − a)
√
(x, ẋ) = (a, 0) and (± a, 0) are the fixed points.
ẋ = y ẏ = (x − a)(x2 − a)
Ñ é
0 1
A=
−2ax − a + 3x2 0
Ñ é
0 1 p
A(a,0) = λ1,2 = ± a2 − a
a2 − a 0
Ñ é
0 1 »
3
A(√a,0) = 3
λ1,2 = ± −2a 2 + 2a
−2a 2 + 2a 0
Ñ é
0 1 »
3
A(−√a,0) = 3
λ1,2 = ± 2a 2 + 2a
2a 2 + 2a 0
The eigenvalues tell us that the fixed points, if they are real, are saddle points. The only ones to be unsure
about are when a = −1, 0, 1, which causes some eigenvalues to be zero.
We can determine what happens at these troublesome values of a by finding a conserved quantity E and
classifying the critical points.
1 4 a 3 a 2
Z
1 2 1
E = ẋ − (x − a)(x2 − a) dx = y 2 − x + x − x + a2 x
2 2 4 3 2

Exx Exy −3x2 + 2ax − a y − x3 + ax2 − ax + a2

=
Eyx Eyy y − x3 + ax2 − ax + a2

1
a = −1

Exx Exy −4 0

= = −4 < 0 ⇒ saddle point
Eyx Eyy

0 1
(1,0)
a=0

Exx Exy 0 0

= = 0 < 0 ⇒ inconclusive
Eyx Eyy

0 1
(0,0)
a=1

Exx Exy −2 0

Eyx Eyy

0 1
(1,0)

Exx Exy −6 4

Eyx Eyy

4 1
(−1,0)
So everything was a saddle point, except possibly (x, y) = (0, 0) when a = 0, but we can figure it out by
looking at the conserved quantity.
1 2 1 4
E= y − x
2 4
which is most definitely a saddle point since the x-axis cross section is concave down and the y-axis cross
section is concave up.
6.5.7
d2 u 1
+ u = α + u2 u=
dθ2 r
a)
du
=v
dθ
dv
= α + u2 − u
dθ
b)
Ä √ ä Ä √ ä
1+ 1−4α 1− 1−4α
(u, v) = 2 ,0 and 2 ,0 are the fixed points.
c)
Ñ é
0 1
A=
2u − 1 0
Ñ é
0 1
A √ =
1+ 1−4α
,0
√
2 1 − 4α 0
√
∆ = − 1 − 4α τ = 0 ⇒ Saddle point
Ñ é
0 1
AÄ 1−√1−4α ä = √
,0
2 − 1 − 4α 0
√
∆ = 1 − 4α τ = 0 ⇒ linear center
This is also a nonlinear center by Theorem 6.5.1. The only difficult condition to check is that the fixed point
is a local minimum of a conserved quantity.

Z
1 2 1 1
E = ẋ + −(α + x2 − x) dx = ẋ2 − αx − x3 + x2
2 2 3 2
∇E = hα + x2 − x, ẋi
Ç √ å
1 − 1 − 4α
∇E , 0 = (0, 0)
2

Exx Exẋ −2 ẋ − α − x2 + x

=
Eẋx Eẋẋ ẋ − α − x2 + x

1
∇E = hα + x2 − x, ẋi
Ç √ å
1 − 1 − 4α
∇E , 0 = (0, 0)
2

Exx Exẋ

= −2 < 0
Eẋx Eẋẋ Ä 1−√1−4α ä

2 ,0
Ç √ å
1− 1 − 4α
⇒ , 0 is a local minimum
2
d)
If the fixed point corresponds to a circular orbit, then the radius is constant.
√
1 1 − 1 − 4α 2
=u= ⇒r= √
r 2 1 − 1 − 4α
which is a constant.
6.5.9
If H(x, p) is a conserved quantity, then H is constant in time.
d ∂H ∂H
H(x, p) = ẋ + ṗ = ṗẋ − ẋṗ = 0
dt ∂x ∂p
So Ḣ = 0 and H is a conserved quantity.
6.5.11
ẋ = y ẏ = −by + x − x3 0<b1
Adding the small amount of damping to the double-well oscillator of Example 6.5.2 changes the nonlinear
centers to stable spirals and destroys the homoclinic orbits of the saddle point.
Using a bit of intuition, we know all the trajectories will go to either the left or right stable spiral, with the
exception of the stable manifold of the saddle point. The two trajectories leading to the saddle point form
the separatrices for the basins of attraction of each spiral.
6.5.13
ẍ + x + x3 = 0
a)
By checking that the isolated fixed point (x, ẋ) = (0, 0) is the minimum of a conserved quantity

Z
1 1 1
E = ẋ2 − (−x − x3 ) dx = ẋ2 + x2 + x4
2 2 2 4

Exx Exẋ 1 + 3x2 ẋ + x + x3 1 0

= = =1>0
Eẋx Eẋẋ ẋ + x + x3

1 0 1
(0,0) (0,0)
So Exx (0, 0) = 1 > 0 means the fixed point is a local minimum and therefore a nonlinear center of the
system.
b)
The origin will still be a nonlinear center for any value of because the origin is always a local minimum
of E and therefore will be surrounded by closed trajectories. However, we don’t know how far away those
closed trajectories exist for any value of , or even if the value of matters at all.
To find out, we can look at the level sets of E for arbitrary .
1 2 1 2 4
E= ẋ + x + x
2 2 4
4E = 2ẋ2 + 2x2 + 2 x4
= 2ẋ2 + (2 x4 + 2x2 + 1) − 1
= 2ẋ2 + (x2 + 1)2 − 1
4E + 1 = 2ẋ2 + (x2 + 1)2
This equation looks a lot like, but not quite due to the x exponent, the equations for the conic sections
(hyperbola, parabola, ellipse, circle). Even more interesting is that by varying we can get all the different
conic-like sections as level sets.
> 0 ⇒ ellipse-like trajectories
= 0 ⇒ (x, ẋ) = (0, 0)
−1
4E < < 0 ⇒ hyperbola-like trajectories
Ä ä Ä √ ä
−1
= 4E ⇒ (x, ẋ) = √±1
−
, 0 = ±2 E, 0
−1
< 4E ⇒ hyperbola-like trajectories
Next we can look at the graphs to see what these conic-like sections actually look like.
Here > 0 and all trajectories loop around the origin no matter how far away we are.
ẋ
x
<0
ẋ
−1
By picking different initial conditions we specify different values of E and can transition from <<0
4E
−1 −1
to = to < .
4E 4E
We can see that the origin is still a nonlinear center, but the trajectories are no longer closed after go-
ing past the heteroclinic trajectories.
6.5.15
a)
mrφ̈ = −mg sin(φ) + mrω 2 sin(φ) cos(φ)

1 d2 φ g
= sin(φ) cos(φ) −
ω 2 dt2 rω 2
rω 2
τ = ωt γ=
g
2
d φ
Å ã
1
= sin(φ) cos(φ) −
dτ 2 γ
b)
Ä ä Ä Ä ä ä
φ, dφ
dτ = (zπ, 0) with z ∈ Z and ± arccos 1
γ , 0 if
1
γ < 1 are the fixed points.
1
γ >1
dφ
dτ
φ
1
γ =1
dφ
dτ
1
0< γ <1
dφ
dτ
c)
The phase portraits imply that the bead either oscillates in position as in one of the closed trajectories, or
the bead keeps making loops around the entire hoop as in the upper and lower periodic trajectories.
6.5.17
mrφ̈ = −mg sin(φ) + mrω 2 sin(φ) cos(φ)

g
φ̈ = ω 2 sin(φ) cos(φ) − 2
rω
g
Z
1 2
E = φ̇ − ω 2 sin(φ) cos(φ) − 2 dφ
2 rω
Å ã
1 2 1 2 2g
= φ̇ + ω cos(φ) cos(φ) − 2
2 2 rω
1 2 1 2 g
= φ̇ + ω cos2 (φ) − cos(φ)
2 2 r
We define the bottom of the hoop as zero potential energy. Then the potential energy of the bead is mass
times gravity times the vertical distance that the bead has traveled.
PE = mgr(1 − cos(φ))
There are two kinetic energies to calculate. One is from the velocity that the bead is moving in the plane of
the hoop. We can calculate that using the rate of change of angle φ̇ and the radius r.
1 1
KEtranslational = mv 2 = m(rφ̇)2
2 2
The other is from the bead’s rotation around the vertical rotation axis of the hoop, which depends on the
hoop’s rotation rate and the bead’s horizontal distance from the rotation axis.
1 1
KErotational = mv 2 = m(r sin(φ)ω)2
2 2
Then we add these together to get the total kinetic energy.
1 1 1
KE = mv 2 = m(rφ̇)2 + m(r sin(φ)ω)2
2 2 2
which leads us to the bead’s total energy.
1 1
Ebead = PE + KE = mgr(1 − cos(φ)) + m(rφ̇)2 + m(r sin(φ)ω)2
2 2
If we check whether Ėbead = 0, we find that

Ėbead = mrφ̇ g sin(φ) + rφ̈ + rω 2 sin(φ) cos(φ)
g
= mrφ̇ g sin(φ) + rω 2 sin(φ) cos(φ) − 2 + rω 2 sin(φ) cos(φ)
rω
= 2mr2 ω 2 sin(φ) cos(φ)φ̇ 6= 0
So the bead’s total energy is not constant. This actually makes sense because the hoop transfers energy
to the bead when it moves away from the bottom. This hoop has a constant rotation rate ω, so whatever
is rotating the hoop has to put in more energy to maintain the constant rotation rate when the bead gets
farther away from the rotation axis and take away energy when the bead gets closer to the rotation axis.
As for what the real conserved quantity is actually conserving,
1 2 1 2 g
φ̇ + ω cos2 (φ) − cos(φ)
E=
2 2 r
1 1 2 2
Emr = m(rφ̇) + mr ω cos2 (φ) − mgr cos(φ)
2 2
2 2
1
= KEtranslational + mr2 ω 2 1 − sin2 (φ) + PE − mgr

2
2 1 2 2 1 2
Emr − mr ω + mgr = KEtranslational − m r sin(φ)ω + PE
2 2
2 1 2 2
Emr − mr ω + mgr = KEtranslational − KErotational + PE
2
If we define a new conserved quantity by lumping all the constants together on the other side (keeping in
mind there is a new zero for potential energy), we get
E = PE + KEtranslational − KErotational
This equation makes sense. The bead has an initial amount of energy in the rotating frame, and the hoop-
rotating energy source transfers energy to and from the bead’s total energy in the rotating frame. A conserved
quantity of energy is what the bead starts with in the rotating frame minus any energy the hoop-rotating
energy source transfers to the bead.
The conserved quantity is H, the “Hamiltonian” for the system. When the potential energy V is a function
only of the coordinates for a stationary coordinate system, H is the total energy. But for problems with
moving constraints (as here), H is not the total energy. The force moving the hoop is doing work on it, and
that work (= energy) needs to be included in the energy accounting. Another way to handle the problem
is to go into a rotating frame with the spinning hoop; then you get a fictitious centrifugal force and its
associated potential energy. This lets you pretend the system has stationary coordinates, and in that frame
the conserved quantity represents the total “energy” of the system.
6.5.19
Ṙ = aR − bRF Ḟ = −cF + dRF a, b, c, d > 0
a)
a represents the growth rate of the rabbits.
b represents the probability for a rabbit to be eaten when encountering a fox, assuming rabbits can only die
by being eaten.
c represents the death rate for foxes due to illness, accidents, etc.
d represents the rate at which rabbits can be converted into foxes when eaten.
This model is unrealistic in that the rabbits have no carrying capacity in the absence of foxes, the foxes
would all die if there were no rabbits (meaning there is nothing else for the foxes to eat), the foxes might
mate seasonally, both species could have a threshold population above and below which the species would
grow and die off respectively, etc.
b)
dR dF
= aR − bRF = −cF + dRF
dt ã dt
b d
Å Å ã
1 dR dF
=R 1− F = cF R−1
a dt a dt c
d b
x = R y = F τ = at
c a
dx dy
= x(1 − y) a = cy(x − 1)
dτ dτ
c
µ=
a
dx dy
= x(1 − y) = µy(x − 1)
dτ dτ
c)
Believe it or not, we can use the spirit of solving exact equations and separable equations from an introductory
ordinary differential equations course to find the conserved quantity.
d
E(x, y) =Ex x0 + Ey y 0
dτ
=Ex x(1 − y) + Ey µy(x − 1) = 0
1−y x−1
dy + µ dx = 0
y x
1−y 1−x
dy = µ dx
y x
Z Z
1 1
− 1 dy = µ − 1 dx
y x
ln |y| − y = µ(ln |x| − x) + C
µx − y − µ ln |x| − ln |y| = C
This expression equals a constant whose value is determined by a point (x0 , y0 ) on the curve, so the LHS must
be the conserved quantity. (We can also drop the absolute value signs since we only consider nonnegative
populations of rabbits and foxes.)
E(x, y) = µx − y − µ ln(x) − ln(y)
d)
(x, y) = (0, 0) and (1, 1) are the fixed points.
We can show that the fixed point (1, 1) is a nonlinear center since we know the conserved quantity along with
the theorem in the section. In fact, everything in the first quadrant is a closed orbit around (1, 1) except
for trajectories starting on the axes, which go to infinity on that axis. This can be verified by showing that
(1, 1) is a global minimum and each level set of E is bounded in the first quadrant. x − ln(x) and y − ln(y)

grow unbounded near their axes and far away from their axes.
6.6 Reversible Systems

6.6.1
We can prove the system is reversible by verifying that f (x, y) is odd in y and g(x, y) is even in y.
ẋ = y(1 − x2 ) = f (x, y) f (x, −y) = −y(1 − x2 ) = −f (x, y)
ẏ = 1 − y 2 = g(x, y) g(x, −y) = 1 − (−y)2 = 1 − y 2 = g(x, y)
So the system is reversible.
(x, y) = (1, 1), (1, −1), (−1, 1), and (−1, −1) are the fixed points.
x
6.6.3
a)
ẋ = sin(y) ẏ = sin(x)
We can prove the system is reversible by verifying that f (x, y) is odd in y and g(x, y) is even in y.
f (x, y) = sin(y) f (x, −y) = sin(−y) = − sin(y) = −f (x, y)
g(x, y) = sin(x) g(x, −y) = sin(x) = g(x, y)
So the system is reversible.

b)
(x, y) = (z1 π, z2 π) are the fixed points. z1 , z2 ∈ Z
Ñ é
0 cos(y)
A=
cos(x) 0
The eigenvalues are λ1,2 = ±1, which is a saddle point if z1 and z2 have the same parity.
The eigenvalues are λ1,2 = ±i, which is a center if z1 and z2 have the opposite parity.
c)
If the lines y = ±x are invariant, then the derivative will be zero.
d
y=x⇒ (y − x) = ẏ − ẋ = sin(x) − sin(y)
dt
= sin(x) − sin(x) = 0
d
y = −x ⇒ (y + x) = ẏ + ẋ = sin(x) + sin(y) = sin(x) + sin(−x)
dt
= sin(x) − sin(x) = 0
Hence the trajectories that start on y = ±x stay on that line forever.

d)
6.6.5
ẍ + f (ẋ) + g(x) = 0 f is even and f, g are smooth.
a)
d2 x
Å ã
dx
ẍ(t) + f ẋ(t) + g x(t) = 2 (t) + f (t) + g x(t) = 0
dt dt
2
1 d x
Å ã
1 dx
t → −t ⇒ 2 2
(−t) + f (−t) + g x(−t)
(−1) dt −1 dt
2
d x
Å ã
dx
2
(−t) + f − (−t) + g x(−t)
dt dt
2
d x
Å ã
dx
(−t) + f (−t) + g x(−t) = 0
dt2 dt
So x(−t) is also a solution if x(t) is a solution.

b)
The system transforms into
ẋ = y ẏ = −f (y) − g(x)
Since the system is reversible, the trajectories for y < 0 are reflections of trajectories with y > 0 with the
arrows reversed.
Now supposing we have a trajectory with y > 0 leading into a stable node or spiral, we know that the
trajectory has to always travel towards the fixed point sufficiently close to the fixed point. However, the
twin trajectory with y < 0 will necessarily travel away from the fixed point because the arrows are reversed.
This is a contradiction because the trajectory has to travel toward the fixed point whether or not y is positive
or negative.
So a reversible system will never contain stable nodes or spirals.
6.6.7
ẍ + xẋ + x = 0
This system’s symmetry is across the ẋ-axis. As a consequence, we’ll check that f (x, y) and g(x, y) are even
and odd in x respectively.
ẋ = y = f (x, y) f (−x, y) = y = f (x, y)
ẏ = −xy − x = g(x, y) g(−x, y) = xy + x = −g(x, y)

6.6.9
N
dφk 1 X
= Ω + a sin(φk ) + sin(φj ) for k = 1, 2
dτ N j=1
a)
π
θ k = φk −
2
N
dφk π 1 X π
= Ω + a sin θk + + sin θj +
dτ 2 N j=1 2
N
dθk 1 X
= Ω + a cos(θk ) + cos(θj )
dτ N j=1
Applying the transformation
θk → −θk τ → −τ
N
−1 dθk 1 X
= Ω + a cos(−θk ) + cos(−θj )
−1 dτ N j=1
N
dθk 1 X
= Ω + a cos(θk ) + cos(θj )
dτ N j=1
results in the same equations, so the system is reversible.

b)
N
dθk 1 X
= 0 = Ω + a cos(θk ) + cos(θj )
dτ N j=1
1
Ω + a cos(θk ) + cos(θ1 ) + cos(θ2 ) = 0
2
2a cos(θk ) + cos(θ1 ) + cos(θ2 ) = −2Ω
(2a + 1) cos(θ1 ) + cos(θ2 ) = −2Ω
cos(θ1 ) + (2a + 1) cos(θ2 ) = −2Ω
(2a + 1)2 cos(θ1 ) + (2a + 1) cos(θ2 ) = −2(2a + 1)Ω
− cos(θ1 ) − (2a + 1) cos(θ2 ) = 2Ω
(2a + 1)2 − 1 cos(θ1 ) = 2Ω − 2(2a + 1)Ω

(4a2 + 4a + 1 − 1) cos(θ1 ) = (2 − 4a − 2)Ω
4a(a + 1) cos(θ1 ) = −4aΩ

−Ω
cos(θ1 ) =
a+1

Ω
which has two solutions if < 1,
a + 1

Ω
one solution if = 1,
a + 1

Ω
and zero solutions if > 1.
a + 1
(2a + 1) cos(θ1 ) + cos(θ2 ) = −2Ω
cos(θ1 ) + (2a + 1) cos(θ2 ) = −2Ω
− cos(θ2 ) − (2a + 1) cos(θ1 ) = 2Ω
(2a + 1)2 cos(θ2 ) + (2a + 1) cos(θ1 ) = −2(2a + 1)Ω

..
.
−Ω
cos(θ2 ) =
a+1
which has the same conditions on the number of solutions.

Ω
There are four (θ1 , θ2 ) fixed points if < 1.
a + 1

Ω
There is one (θ1 , θ2 ) fixed point if = 1.
a + 1

Ω
There are zero (θ1 , θ2 ) fixed points if > 1.
a + 1
c)
Ω=0
θ2
θ1
Ω=1
θ2
θ1
Ω = 1.75
θ2
θ1
Ω=2
θ2
θ1
Ω = 2.5
θ2
θ1
Ω=3
θ2
θ1
6.6.11
φ̇ = cos2 (φ) + A sin2 (φ) sin(θ)

θ̇ = cot(φ) cos(θ)
a)
t → −t θ → −θ
−1 dθ 1 dφ
= cos2 (φ) + A sin2 (φ) sin(−θ)

= cot(φ) cos(−θ)
−1 dt −1 dt
dθ dφ
= cos2 (φ) + A sin2 (φ) − sin(θ)

= cot(φ) cos(−θ) −
dt dt
dφ
= cos2 (φ) + A sin2 (φ) sin(θ)

dt
t → −t φ → −φ
1 dθ −1 dφ
= cos2 (−φ) + A sin2 (−φ) sin(θ)

= cot(−φ) cos(θ)
−1 dt −1 dt
dθ dφ 2
− = − cot(φ) cos(θ) = cos2 (φ) + A − sin(φ) sin(θ)
dt dt
dθ dφ
= cos2 (φ) + A sin2 (φ) sin(θ)

= cot(φ) cos(θ)
dt dt
b)
A = −1
θ
A=0
θ
A=1
θ
c)
From the phase portraits, it seems that almost all trajectories approach one of two stable fixed points when
A = −1 or A = 0. This means that the object settles down to a fixed orientation in the shear flow. In
contrast, when A = 1, the trajectories trace out neutrally stable periodic orbits starting from almost all
initial conditions. This means the object tumbles periodically in the shear flow without ever converging to
a fixed orientation.
6.7 Pendulum
6.7.1
θ̈ + bθ̇ + sin θ = 0 b>0
ẋ = y ẏ = − sin(x) − by
(x, y) = (zπ, 0) are the fixed points, z ∈ Z.

Ñ é
0 1
A=
− cos(x) −b
6.7 Pendulum 161
Ñ é
0 1
A(2zπ,0) =
−1 −b
∆ = 1, τ = −b < 0 ⇒ stable fixed point

√
−b ± b2 − 4
λ1,2 =
2
0 < b < 2 ⇒ stable spiral
b = 2 ⇒ stable degenerate node
2 < b ⇒ stable node
Ñ é
0 1
A =
(2z+1)π,0
1 −b
∆ = −1, τ = −b < 0 ⇒ saddle point
b=1
x
b=2
b=4
6.7.3

θ̈ + 1 + a cos(θ) θ̇ + sin θ = 0 a≥0

ẋ = y ẏ = − sin(x) − 1 + a cos(x) y
(x, y) = (zπ, 0) are the fixed points, z ∈ Z.

Ñ é
0 1
A=
− cos(x) + ay sin(x) −1 − a cos(x)
6.7 Pendulum 163
Ñ é
0 1
A(2zπ,0) =
−1 −1 − a
∆ = 1, τ = −1 − a < 0 ⇒ stable fixed point
Ñ é
0 1
A =
(2z+1)π,0
1 a−1
∆ = −1 ⇒ saddle point
So as expected, the low point of the swing is stable and the very top of the swing is unstable for this damped
pendulum.
6.7.5
1 2

α 2π 1 + 16 α Quadrature
0
18 π 2π 2π
1
18 π 6.29514 6.29517
2
18 π 6.3310 6.3313
3
18 π 6.3908 6.3925
4
18 π 6.4746 6.4801
5
18 π 6.5822 6.5960
6
18 π 6.7138 6.7430
7
18 π 6.8693 6.9250
8
18 π 7.0488 7.1471
9
18 π 7.2521 7.4163
10
18 π 7.4794 7.7423
11
18 π 7.7306 8.1389
12
18 π 8.0058 8.6261
13
18 π 8.3048 9.2352
14
18 π 8.6278 10.0182
15
18 π 8.9747 11.0723
16
18 π 9.3455 12.6135
17
18 π 9.7403 15.3270
18
18 π 10.1590 82.0530
6.8 Index Theory

6.8.1
a)
Stable spiral
y
3
4 2
1,9
8
2
x 7
5 1,9
3
6
4 5
6 8
7
b)
Unstable spiral
y
3
4 2
5 4
6
x 3
5 1,9
7
2
8 1,9
6 8
7
c)
Center
y
3
4 2
5
6 4
x
5 1,9 7 3
8 2
1,9
6 8
7
d)
Star
y
3
2
4 3
4 2
1,9
x
5 1,9
5
6 8
8 7
6
7
e)
Degenerate node
y
The vector field on this circle is pretty finicky near the point when the trajectory turns around and starts
going the opposite direction. Due to this, we’ll be graphing the path of the vector heads instead of the
individual vectors themselves since they really bunch around θ = π and the side directly opposite.
y
And the degenerate node has an index of +1 as expected.

6.8.3
ẋ = y − x ẏ = x2
y
Despite looking complicated, this curve has an index of 0 because the arrows never point down so can’t make
a full rotation.
6.8.5
ẋ = xy ẏ = x + y
y
4 3
5 1,4,13
2,3
6
2 5,12
x
7 1,13
12 6,11
8
8,9
9 11 7,10
10
The arrows almost make a full counterclockwise rotation but then turn around, so the index is 0.
6.8.7
ẋ = x(4 − y − x2 ) ẏ = y(x − 1)
(x, y) = (0, 0), (2, 0), (1, 3), and (−2, 0) are the fixed points.
Ñ é
4 − y − 3x2 −x
A=
y x−1
Ñ é
4 0
A(0,0) =
0 −1
λ1 = 4 λ2 = −1 ⇒ saddle point
Ñ é Ñ é
1 0
~v1 = ~v2 =
0 1
Ñ é
−8 −2
A(2,0) =
0 1
λ1 = −8 λ2 = 1 ⇒ saddle point
Ñ é Ñ é
1 −2
~v1 = ~v2 =
0 9
Ñ é
−2 −1
A(1,3) =
3 0
√
λ1,2 = −1 ± 2i ⇒ stable spiral
Ñ é
−8 2
A(−2,0) =
0 −3
λ1 = −8 λ2 = −3 ⇒ stable node
Also notice from the equations that x(0) = 0 ⇒ x(t) = 0 and y(0) = 0 ⇒ y(t) = 0. This means the x-axis
and y-axis are invariant. This means a closed orbit can’t cross the axes.
A closed orbit must enclose a fixed point, and the only candidate is the stable spiral. However, this is
also impossible since the unstable manifold of the saddle point goes into the stable spiral. We didn’t actually
prove this but use numerical results as evidence.
An intuitive argument is that the upper branch of the unstable manifold of the saddle point (2,0) must
approach the spiral. It has to go somewhere. If it does not go to the stable spiral, where else could it go?
i) Both the x and y axes are invariant, so the unstable manifold can’t get out of the first quadrant. Said
another way, both of the axes contain trajectories and trajectories cannot cross by the uniqueness theorem
for solutions of ODEs.
ii) The unstable manifold can’t approach the saddle point at (0,0) since the only trajectories that do lie
on the y-axis, which is the stable manifold of the saddle point at (0,0).
iii) Conceivably the upper branch of the unstable manifold could escape to infinity. However, the flow
in the first quadrant pushes the unstable manifold upward and to the left until it crosses the vertical line
x = 1. Then the unstable manifold moves left because ẋ < 0 when x lies to the right of the nullcline given
the parabola 4 − y − x2 = 0 and moves upward because ẏ > 0 when x > 1, which is the case initially since
the unstable manifold starts at (2,0). Once the manifold crosses the line x = 1 it has to move downward.
And after that, there’s no way for the trajectory to escape out to infinity.
Clearly this is only a plausibility argument, not a proof. To really prove that the unstable manifold cannot
escape to infinity, one could construct a trapping region, a technique discussed in Section 7.3.
6.8.9
False. A counter example is
ṙ = r(r − 1)(r − 3)
θ̇ = r2 − 4
There is only one fixed point at the origin. The inner closed orbit is attracting and θ̇ < 0, so rotation is in
the clockwise direction. The outer closed orbit is repelling with θ̇ > 0, so rotation is in the counterclockwise
direction.
6.8.11
a)
k=1
p y
ż = z = x + iy = reiθ r= x2 + y 2 tan(θ) =
x
ż = z̄ = x − iy = re−iθ
k=2
ż = z 2 = (x + iy)2 = x2 − y 2 + 2ixy
2
= reiθ = r2 e2iθ
ż = (z̄)2 = (x − iy)2 = x2 − y 2 − 2ixy

2
= re−iθ = r2 e−2iθ
k=3
ż = z 3 = (x + iy)3 = x3 − 3xy 2 + i(3x2 y − y 3 )

3
= reiθ = r3 e3iθ
ż = (z̄)3 = (x − iy)3 = x3 + i(y 3 − 3x2 y) − 3xy 2

3
= re−iθ = r3 e−3iθ
bc)
It’s not so easy to see that the origin is the only stable fixed point in the Cartesian coordinates, but in the
polar coordinates it’s clear that the only way to satisfy ż = 0 is r = 0 in both cases, which is the origin.
The index, also not easy to see from the Cartesian coordinates, is ±k in the z k and (z̄)k cases respec-
tively. The arrows will do k rotations on a loop enclosing the origin when θ goes through a 2π rotation due
to the k in e±kiθ , with +kiθ causing counterclockwise and −kiθ causing clockwise rotations.
6.8.13
a)
ẏ g(x, y)
Å ã Å ã
φ = arctan = arctan
ẋ f (x, y)
d g(x,y)
g(x, y) d g
Å ã Å ã
d d dt f (x,y) 1
φ= arctan = Ä g(x,y) ä2 = 2
dt dt f (x, y) 1 + f (x,y) dt f 1 + fg 2
f dg − gdf 1 f dg − gdf
dφ = 2 2 =
f g
1 + f2 f 2 + g2
b)
f dg − gdf
I I
1 1
IC = dφ =
2π C 2π C f 2 + g2
7
Limit Cycles
7.1 Examples
7.1.1
ṙ = r3 − 4r θ̇ = 1
y
173
174 Chapter 7: Limit Cycles
7.1.3
ṙ = r(1 − r2 )(4 − r2 ) θ̇ = 2 − r2
y
7.1.5
ṙ = r(1 − r2 ) θ̇ = 1
d
ẋ = r cos(θ)
dt
= ṙ cos(θ) − r sin(θ)θ̇
= r(1 − r2 ) cos(θ) − r sin(θ)
= (1 − r2 )r cos(θ) − r sin(θ)
= (1 − x2 − y 2 )x − y
= x − y − x(x2 + y 2 )
7.1 Examples 175
d
ẏ = r sin(θ)
dt
= ṙ sin(θ) + r cos(θ)θ̇
= r(1 − r2 ) sin(θ) + r cos(θ)
= (1 − r2 )r sin(θ) + r cos(θ)
= (1 − x2 − y 2 )y + x
= x + y − y(x2 + y 2 )
7.1.7
ṙ = r(4 − r2 ) θ̇ = 1
The initial condition of (r, θ) = (0.1, 0) in Cartesian coordinates is (x, y) = (0.1, 0). The system will approach
the r = 2 limit cycle, which would make x(t) = 2 cos(t) the limiting waveform for any initial condition except
the origin.
We can see this in the following graph.

y
7.1.9
a)
Since the dog and the duck are moving at the same speed, they will both move the same amount in the time
it takes θ to change by ∆θ. Thus the problem can be done using ∆θ instead of ∆t in the derivation.
During a small amount of time ∆t, the length of the segment R decreases from the dog swimming a length
of ∆θ and increases from the duck moving along the perimeter of the circle also by a length of ∆θ. We’ll
eventually be taking a limit, so we can make our lives easier with some approximations.
The line segment along the perimeter of the circle is approximately the length of the arc that the line segment
spans, and the angle bordering φ is approximately a right angle. Using these approximations, the law of
cosines, and a trick for taking a limit, we can arrive at a differential equation for R.
∆R ≈ −∆θ + c − R
… π
= −∆θ + (∆θ)2 + R2 − 2∆θR cos φ + −R
2
(∆θ)2 + R2 − 2∆θR cos φ + π2 − R2

= −∆θ + »
(∆θ)2 + R2 − 2∆θR cos φ + π2 + R

(∆θ)2 + 2∆θR sin(φ)

= −∆θ + p
(∆θ)2 + R2 + 2∆θR sin(φ) + R
∆R ∆θ + 2R sin(φ)
= −1 + p
∆θ (∆θ) + R2 + 2∆θR sin(φ) + R
2
∆R ∆θ + 2R sin(φ)
lim = −1 + p
∆θ→0 ∆θ (∆θ) + R2 + 2∆θR sin(φ) + R
2
2R sin(φ)
R0 = −1 + √ = −1 + sin(φ)
R2 + R
For the φ0 equation we can use some basic geometry.
7.1 Examples 177
Next we have to remember that 4acb is similar to 4dce (angles subtended by the same arc are equal),
meaning the ratio between ab and de is equal to the ratio of ac and ce.
Approximating ab as ab
Ù = ∆θ and de as de
Ù
2 cos(φ) − R
ab : de = ac : cd ⇒ de
Ù≈ ∆θ
R
Next we have to remember that the arc made by the inscribed angle φ is half of the arc of the central angle
(inscribed angle theorem).
a
φ
2φ
Now putting it all together, the change in the central angle 2∆φ is from an increase of arc de
Ù and a decrease
of arc ∆θ.
Ù − ∆θ ≈ 2 cos(φ) − R 2 cos(φ) − 2R
2∆φ = de ∆θ − ∆θ = ∆θ
R R
∆φ
R ≈ cos(φ) − R
∆θ
∆φ
lim R = Rφ0 = cos(φ) − R
∆θ→0 ∆θ
As for whether or not the dog catches the duck, the answer is no. The proof is difficult enough that it is the
subject of a published paper. Please take a look at Applications of Center Manifolds by Keith Promislow if
you’re interested.
b)
Repeating the derivation now with the speed of the dog as k∆θ
R0 = −k + sin(φ)
Rφ0 = cos(φ) − R
c)
Assuming there is a limit cycle
R0 = −k + sin(φ) = 0 ⇒ φ = arcsin(k)
p
Rφ0 = cos(φ) − R = 0 ⇒ R = cos(φ) = 1 − k 2
√
1 3
So for k = 2 the dog will endlessly chase the duck while asymptotically approaching a circle of radius 2 as
a limit cycle.
7.2 Ruling Out Closed Orbits

7.2.1
V = x2 + y 2
y
x
7.2.3
V = ex sin(y)
y
7.2.5
Let ẋ = f (x, y) and ẏ = g(x, y) be a smooth vector field defined on the phase plane.
a)
If this is a gradient system, then
Å ã
∂V ∂V
−∇V = − ,− = (f (x, y), g(x, y))
∂x ∂y
∂V ∂V ∂f ∂2V ∂2V ∂g
− = f and − =g⇒ =− =− =
∂x ∂y ∂y ∂y∂x ∂x∂y ∂x
b)
∂f ∂g
Yes. If f and g are smooth on R2 and ∂y = ∂x , the system is a gradient system.
f : R2 → R being smooth and the fundamental theorem of calculus together imply there exists V : R2 → R
R
such that V (x, y) is continuous and V (x, y) = f (x, y)dx.
f (x, y)dx = ∂f
R ∂g
Leibniz integral rule ⇒ ∂V ∂
R R
∂y = ∂y ∂y (x, y)dx = ∂x (x, y)dx = g
∂V ∂V
Hence there exists V (x, y) such that ∂x = f and ∂y = g.
7.2.7
ẋ = y + 2xy ẏ = x + x2 − y 2
a)
∂ ∂
x + x2 − y 2

y + 2xy = 0 + 2y = 2y = 0 + 0 + 2y =
∂x ∂y
b)
∂V
− = ẋ = y + 2xy
∂x Z
⇒ −V = y + 2xy dx = xy + x2 y + g(y)
∂V
− = ẏ = x + x2 − y 2
∂y
∂
xy + x2 y + g(y) = x + x2 + g 0 (y)

=
∂y
1
⇒ g 0 (y) = −y 2 ⇒ g(y) = − y 3 + C (set to 0 for convenience)
3
2 1 3
V (x, y) = −xy − x y + y
3
c)
y
7.2.9
Recall that a gradient system will have
hẋ, ẏi = hf (x, y), g(x, y)i = h−Vx , −Vy i ⇒ fy = −Vxy = −Vyx = gx
a)
ẋ = y + x2 y ẏ = −x + 2xy
∂ ∂
fy = (y + x2 y) = 1 + x2 6= −1 + 2y = (−x + 2xy) = gx
∂y ∂x
So this is not a gradient system.
b)
ẋ = 2x ẏ = 8y
∂ ∂
fy = (2x) = 0 = (8y) = gx
∂y ∂x
So this is a gradient system with V (x, y) = −x2 − 4y 2 .
y y
x x
c)
2 2 2 2
ẋ = −2xex +y ẏ = −2yex +y
∂ 2 2
2 2 ∂ 2 2

fy = − 2xex +y = −4xyex +y = − 2yex +y = gx
∂y ∂x
2
+y 2
So this is a gradient system with V (x, y) = ex .
y y
x x
7.2.11
V = ax2 + 2bxy + cy 2
The shape is paraboloid-like, but the middle term can make the origin into a saddle point for the surface.
This problem is exactly like a multivariable calculus problem to find an absolute minimum, so we will be
using the second derivative test.

2 2
= 4 ac − b2

Vxx Vyy − Vxy = (2a)(2c) − (2b)
(0,0) (0,0)
The second derivative test states that the origin is a local extrema if ac − b2 > 0 and a local minimum if
Vx (0, 0) = Vy (0, 0) = 0 for the origin to be a critical point and for Vxx (0, 0) = 2a > 0. (We could have used
Vyy instead of Vxx but we simply had to use one of them.)
Hence, V (x, y) is positive definite if and only if a > 0 and ac − b2 > 0.
7.2.13
Å ã Å ã
N1 N2
Ṅ1 = r1 N1 1 − − b1 N1 N2 Ṅ2 = r2 N2 1 − − b2 N1 N2
K1 K2
1
Using Dulac’s criterion with weighting function g(N1 , N2 ) =
N1 N2

∇ · g(Ṅ1 , Ṅ2 )hN1 , N2 i
≠ Å ã Å ã ∑
1 1 N1 1 1 N2
= ∇ · r1 − − b1 , r2 − − b2
N2 K1 N2 N1 K2 N1
r1 1 r2 1
=− − <0
K1 N2 K2 N1
The last step holds since all the parameters and variables are strictly positive. Hence, by Dulac’s criterion
there are no periodic orbits contained entirely in the first quadrant.
7.2.15
ẋ = x(2 − x − y) ẏ = y(4x − x2 − 3)
a)
(x, y) =(0,0), (1,1), (2,0), and (3,-1) are the fixed points.
Ñ é
2 − 2x − y −x
A=
y(4 − 2x) 4x − x2 − 3
Ñ é
2 0
A(0,0) =
0 −3
λ1 = 2 λ2 = −3 ⇒ saddle point
Ñ é
−1 −1
A(1,1) =
2 0
√
−1 ± i 7
λ1,2 = ⇒ stable spiral
2
Ñ é
−2 −2
A(2,0) =
0 1
λ1 = −2 λ2 = 1 ⇒ saddle point
Ñ é
−3 −3
A(3,−1) =
2 0
√
−3 ± i 15
λ1,2 = ⇒ stable spiral
2
b)
y
x
7.2.17
We’ll prove this by contradiction by assuming that there are two or more closed orbits in R. Next we assume
all the conditions of Dulac’s criterion, except that R is not simply connected and has exactly one hole in it.
There are two possible properties for the closed orbits. Either a closed orbit does not encircle the hole
or the closed orbit does encircle the hole.
R R
We can actually rule out the former case immediately by trimming the R into a region R̃ that is simply
connected.
R̃
Then applying Dulac’s criterion rules out the existence of the closed orbit.
As for when the closed orbit encircles the hole, zero or one closed orbit are possible inside R, but two
or more leads to a contradiction.
If there are two closed orbits encircling the hole, then one closed orbit must encircle the inner closed orbit,
and both closed orbits are the boundary to a region R̂ they enclose.
R̂
We can then split the region R̂ into two simply connected regions and apply Green’s theorem to each.
R̂1
R̂2
Z Z Z
0 6= ∇ · (g ẋ) dA = ∇ · (g ẋ) dA + ∇ · (g ẋ) dA
R̂
IR̂1 I R̂2
= g ẋ · n d` + g ẋ · n d`
I∂ R̂1 ∂ R̂2
= g ẋ · n d` = 0
∂ R̂
(The orientation of the closed orbits does not matter since the the normal vector is perpendicular regardless
if the closed orbit is traversed backwards or forwards.)
And here is our contradiction of 0 6= 0. All the steps are correct, which leaves the assumption at the
very beginning that this region exists at all, meaning there can’t be two closed orbits encircling the hole in
R. More than two closed orbits are also ruled out by the same reasoning.
Hence the conditions of Dulac’s criterion with R as a region topologically equivalent to an annulus instead
of a simply connected region implies that there is at most one closed orbit in the region R.
7.2.19
Ṙ = −R + AS + kSe−S Ṡ = −S + AR + kRe−R k, AS , AR > 0
a)
The −R and −S terms signify that the more love Rhett and Scarlett feel for the other, the faster their love
for the other decreases.
The AS and AR are the baseline amounts of love one feels for the other. For instance, if Scarlett’s love
for Rhett is fixed at 0, then Rhett will still love Scarlett an amount AS , and similarly for Scarlett when the
roles are reversed.
The kSe−S and kRe−R are 0 for R, S = 0 and approximately 0 when R, S are large, with the maxi-
mum somewhere in between. The function having a maximum implies too little love. Too little love for each
partner has effectively no effect, and there is an optimal amount of love for each partner that will cause the
greatest increase in each partner’s love.
b)
Making normal vectors pointing inward on the boundary of the first quadrant
and then dot product with the vector field on the boundary

h1, 0i · hṘ, Ṡi = h1, 0i · h−R + AS + kSe−S , −S + AR + kRe−R i

R=0 R=0
−S
= h1, 0i · hAS + kSe , −S + AR i
= AS + kSe−S > 0

h0, 1i · hṘ, Ṡi = h0, 1i · h−R + AS + kSe−S , −S + AR + kRe−R i

S=0 S=0
−R
= h0, 1i · h−R + AS , AR + kRe i
= AR + kRe−R > 0
meaning the vector field always points inward on the positive axes. Thus the trajectory can never escape
the first quadrant.
c)
Picking g(R, S) = 1
∇ · hṘ, Ṡi = ∇ · h−R + AS + kSe−S , −S + AR + kRe−R i
= (−1) + (−1) = −2
then by Dulac’s criterion there are no periodic solutions, and not just in the first quadrant.
d)
AS = 1.2 AR = 1 k = 15 R(0) = S(0) = 0
7.3 Poincaré-Bendixson Theorem

7.3.1
ẋ = x − y − x(x2 + 5y 2 ) ẏ = x + y − y(x2 + y 2 )
a)
Classify the fixed point at the origin.
Ñ é
1 −1
A=
1 1
(1 − λ)2 + 1 = 0 ⇒ λ2 − 2λ + 2 = (λ − (1 + i)) (λ − (1 − i)) = 0
λ1 = 1 + i λ2 = 1 − i ⇒ unstable spiral
b)
Rewrite the system in polar coordinates using
xẏ − y ẋ
rṙ = xẋ + y ẏ θ̇ =
r2
xẋ = x2 − xy − x2 (x2 + 5y 2 ) y ẏ = xy + y 2 − y 2 (x2 + y 2 )

xẋ + y ẏ = x2 − xy − x2 (x2 + 5y 2 ) + xy + y 2 − y 2 (x2 + y 2 )
= x2 + y 2 − x2 (x2 + 5y 2 ) − y 2 (x2 + y 2 )
= x2 + y 2 − x4 − y 4 − 6x2 y 2
= r2 − r4 cos4 (θ) − r4 sin4 (θ) − 6r4 cos2 (θ) sin2 (θ)
= r2 − r4 cos4 (θ) + r4 cos2 (θ) sin2 (θ) − r4 sin4 (θ) + r4 cos2 (θ) sin2 (θ) − 4r4 cos2 (θ) sin2 (θ)

= r2 − r4 cos2 (θ) cos2 (θ) + sin2 (θ) − r4 sin2 (θ) sin2 (θ) + cos2 (θ) − 4r4 cos2 (θ) sin2 (θ)

= r2 − r4 cos2 (θ) − r4 sin2 (θ) − 4r4 cos2 (θ) sin2 (θ)
= r2 − r4 cos2 (θ) + sin2 (θ) − 4r4 cos2 (θ) sin2 (θ)

= r2 − r4 − 4r4 cos2 (θ) sin2 (θ)

2
= r2 − r4 − r4 (2 cos(θ) sin(θ))
= r2 − r4 − r4 sin2 (2θ)
rṙ = xẋ + y ẏ = r2 − r4 − r4 sin2 (2θ) ⇒ ṙ = r − r3 − r3 sin2 (2θ) = r 1 − r2 − r2 sin2 (2θ)

xẏ = x x + y − y(x2 + y 2 ) = x2 + xy − xy(x2 + y 2 )

y ẋ = y x − y − x(x2 + 5y 2 ) = xy − y 2 − xy(x2 + 5y 2 )

xẏ − y ẋ = x2 + y 2 − xy(x2 + y 2 ) − xy − y 2 − xy(x2 + 5y 2 )

= x2 + y 2 + xy(−x2 − y 2 + x2 + 5y 2 )
= x2 + y 2 + 4xy 3
= r2 + 4r4 cos(θ) sin3 (θ)
xẏ − y ẋ
θ̇ = = 1 + 4r2 cos(θ) sin3 (θ)
r2
c)
Determine the circle of maximum radius, r1 , centered on the origin such that all trajectories have a radially
outward component on it.
This is equivalent to 0 ≤ ṙ for all θ.
0 ≤ r 1 − r2 − r2 sin2 (2θ) ⇒ 0 ≤ 1 − r2 − r2 sin2 (2θ)

1
⇒0≤r≤
1 + sin2 (2θ)
1
⇒ 0 ≤ r ≤ √ = r1
2
d)
Determine the circle of minimum radius, r2 , centered on the origin such that all trajectories have a radially
inward component on it.

1
r 1 − r − r sin (2θ) ≤ 0 ⇒ 1 − r2 − r2 sin2 (2θ) ≤ 0
2 2 2

⇒ ≤ r ⇒ r2 = 1 ≤ r
1 + sin2 (2θ)
e)
1
We know that there is at least one limit cycle in the region √ ≤ r ≤ 1 by the Poincaré-Bendixson theorem.
2
The trapping region is a closed and bounded subset of the plane.
The vector field is continuously differentiable, easily seen in the Cartesian representation, on an open set
containing the trapping region.
The trapping region does not contain any fixed points.
1
ṙ = 0 ⇒ r 1 − r2 − r2 sin2 (2θ) = 0 ⇒ r = 0 or r2 =

1 + sin2 (2θ)
−1
θ̇ = 0 ⇒ 1 + 4r2 cos(θ) sin3 (θ) = 0 ⇒ r2 =
4 cos(θ) sin3 (θ)
1 −1
= ⇒ 4 cos(θ) sin3 (θ) = −1 − sin2 (2θ)
1 + sin2 (2θ) 4 cos(θ) sin3 (θ)
⇒ sin2 (2θ) (4 cos(θ) sin(θ) + 1) = −1
⇒ sin2 (2θ) (2 sin(2θ) + 1) = −1
⇒ 2 sin3 (2θ) + sin2 (2θ) + 1 = 0
⇒ (sin(2θ) + 1) 2 sin2 (2θ) − sin(2θ) + 1 = 0

3π
The only real solution is sin(2θ) = −1 ⇒ θ = 2 ⇒ x = 0 ⇒ y = 0.
7.3.3
ẋ = x − y − x3 ẏ = x + y − y 3
The only fixed point of the system is the origin, and linearization predicts the origin is an unstable spiral.
We’ll make a trapping region by converting to polar coordinates.
r 2 = x2 + y 2
2rṙ = 2xẋ + 2y ẏ
x(x − y − x3 ) + y(x + y − y 3 )
ṙ =
r
x2 + y 2 − x4 − y 4
=
r
r2 − r4 cos4 (θ) − r4 sin4 (θ)
=
r
= r − r3 cos4 (θ) − r3 sin4 (θ)
= r − r3 cos4 (θ) + sin4 (θ)

We can bound the θ term with 1

2 ≤ cos4 (θ) + sin4 (θ) ≤ 1, which gives
r3
r − r3 ≤ ṙ ≤ r −
2
So a suitable trapping region is bounded by a circle of radius less than 1, which has a positive radial com-
√
ponent, and a circle of radius greater than 2, which has a negative radial component.
Then by the Poincaré-Bendixson theorem, there is at least one periodic solution in the trapping region.
7.3.5
ẋ = −x − y + x(x2 + 2y 2 ) ẏ = x − y + y(x2 + 2y 2 )
This system has a fixed point at (x, y) = (0, 0), which is best seen using a graph of the nullclines.
y
x
Ñ é
3x2 + 2y 2 − 1 4xy − 1
A=
2xy + 1 x2 + 6y 2 − 1
Ñ é
−1 −1
A(0,0) =
1 −1
λ1,2 = −1 ± i ⇒ stable spiral
The Poincaré-Bendixson theorem can still be applied here, despite the fixed point being unstable. In this
case the trapping region will be more of a repulsion region, as the vector field will point outward on the
boundaries. Thus there will be at least one unstable limit cycle in the region. (You can see this by trans-
forming t → −t, which runs time backwards for the system. Then we’re showing existence for the usual
stable limit cycle.)
Next we’ll transform the system into polar coordinates to find the boundaries of the trapping region.
r 2 = x2 + y 2
2rṙ = 2xẋ + 2y ẏ
xẋ + y ẏ
ṙ =
r
x − x − y + x(x2 + 2y 2 ) + y x − y + y(x2 + 2y 2 )

=
r
x4 + 3x2 y 2 − x2 + 2y 4 − y 2
=
r
(x2 + y 2 )(x2 + 2y 2 − 1)
=
r
r2 (r2 + y 2 − 1)
=
r
= r r + r2 sin2 (θ) − 1
2

From here we can upper and lower bound the θ term to bound ṙ.
r(r2 − 1) ≤ ṙ ≤ r(2r2 − 1)
So a suitable trapping region is bounded by a circle of radius less than √1 , which has a negative radial
2
component, and a circle of radius greater than 1, which has a positive radial component.
Then by the Poincaré-Bendixson theorem, there is at least one periodic solution in the trapping region.
7.3.7
1
ẋ = y + ax(1 − 2b − r2 ) ẏ = −x + ay(1 − r2 ) 0<a≤1 0≤b<
2
a)
xẋ + y ẏ
ṙ =
r
x y + ax(1 − 2b − r2 ) + y − x + ay(1 − r2 )

=
r
−2abx − ar x − ar y + ax2 + ay 2
2 2 2 2 2
=
r
a − 2bx2 − r4 + r2
=
r
a − 2br2 cos2 (θ) − r4 + r2

=
r
= ar 1 − r − 2b cos2 (θ)
2

y
θ = tan
x
xẏ − ẋy
θ̇ =
x2 + y 2
x − x + ay(1 − r2 ) − y + ax(1 − 2b − r2 ) y

=
r2
2abxy − x2 − y 2
=
r2
2abr sin(θ) cos(θ) − r2
2
=
r2
= 2ab sin(2θ) − 1
b)
A fixed point can occur only if ṙ = θ̇ = 0 or r = ṙ = 0 is true. The ranges of a and b and the θ̇ equation
make the latter impossible, leaving the origin as the only fixed point.
We can make an annular trapping region using ar(1 − r2 ) ≤ ṙ ≤ ar(1 − 2b − r2 ), for which 1 < r ⇒ ṙ < 0,
√
and 1 − 2b < r ⇒ ṙ > 0 will have all vectors pointing inward on the boundary. Therefore, by the Poincaré-
Bendixson theorem, there is at least one limit cycle in the trapping region.
We also know that all limit cycles have the same period because θ̇ = 2ab sin(2θ) − 1 is independent of
r, meaning the rotation rate is the same at every θ value no matter which limit cycle is being traversed.
c)
Setting b = 0 simplifies the equations to
ṙ = ar 1 − r2

θ̇ = −1
for which r < 1 ⇒ ṙ > 0 and 1 < r ⇒ ṙ > 0, making r = 1 the only limit cycle.
7.3.9
ṙ = r(1 − r2 ) + µr cos(θ) θ̇ = 1
a)
dr dr dθ dr
= = = r(1 − r2 ) + µr cos(θ)
dt dθ dt dθ
r(θ) = 1 + µr1 (θ) + O µ2

dr
= µr10 (θ)
dθ
ṙ = r(1 − r2 ) + µr cos(θ)
2
µr10 (θ) = 1 + µr1 (θ) 1 − 1 + µr1 (θ) + µ 1 + µr1 (θ) cos(θ)
r10 = −2r1 + cos(θ)

1 2
r1 = sin(t) + cos(θ) + Ce−2t
5 5
The exponential term eventually dies off, leaving only the periodic trajectory.
Å ã
1 2 1 2
sin(θ) + cos(θ) + O µ2

r1 = sin(θ) + cos(θ) r(θ) = 1 + µ
5 5 5 5
b)
Å ã
dr 1 2
=µ cos(θ) − sin(θ) = 0
dθ 5 5
Å ã Å ã
1 2 1 1
cos(θ) − sin(θ) = 0 ⇒ θ = arctan , arctan +π
5 5 2 2
Å Å ãã Å ã
1 1 1 2 2 µ
+ O µ2 = 1 + √ + O µ2

r arctan =1+µ √ + √
2 5 5 5 5 5
1 −1 2 −2
Å Å ã ã Å ã
1 µ
+ O µ2 = 1 − √ + O µ2

r arctan +π =1+µ √ + √
2 5 5 5 5 5
p µ µ p
1−µ<1− √ <r <1+ √ < 1+µ for µ 1
5 5
c)
√ √
The dashed circles are the error bounds r = 1 − µ and r = 1 + µ, the thick solid line is the numerical
solution, and the thin solid line is the series approximation solution.
µ = 1 − 2−5 y µ = 1 − 2−4 y
x x
µ = 2−1 y µ = 2−2 y
x x
The numerical solution and series approximation solution become indistinguishable on the graphs, but the
upper and lower bounds continue to restrict the closed orbit to a more and more circular path as µ decreases.
µ = 2−3 y µ = 2−4 y
x x
As for how the maximum error in r depends on µ, some values have been computed below.
µ Error µ Error
1/8 0.0016312 1/32 9.8704e-05
2/8 0.0068165 2/32 3.9907e-04
3/8 0.016030 3/32 9.0765e-04
4/8 0.029790 4/32 0.0016312
5/8 0.048636 5/32 0.0025766
6/8 0.073110 6/32 0.0037510
7/8 0.10371 7/32 0.0051618
Using only the right side as the data to a quadratic fit gives
(1.1368e-01)µ2 + (−1.4908e-03)µ + (4.1271e-05)
The µ2 coefficient is dominant in the polynomial, giving us a little reassurance that the error really is O µ2

for µ 1.
7.3.11
ṙ = r(1 − r2 ) r2 sin2 (θ) + (r2 cos2 (θ) − 1)2 θ̇ = r2 sin2 (θ) + (r2 cos2 (θ) − 1)2

a)
y
x
7.4 Liénard Systems 197
b)
x
The above graph has the qualitative behavior of the system, but the real graph has greatly exaggerated
pauses.
The trajectory here starts at (r, θ) = (3, 0), which is (x, y) = (3, 0), and quickly asymptotes to the cir-
cle r = 1. The trajectory then follows the circumference for all time, but what shows in the x versus t
graph is that the trajectory is spending more and more time near the fixed points (x, y) = (±1, 0) with each
traversal because r → 1 as t → ∞, which makes θ̇ ≈ 0 near the fixed points.
7.4 Liénard Systems

7.4.1
ẍ + µ(x2 − 1) + tanh(x) = 0
Applying Liénard’s theorem

(1)
f (x) = µ(x2 − 1) g(x) = tanh(x)
which are both continuously differentiable for all x.

(2)
g(−x) = tanh(−x) = − tanh(x) = −g(x)
g(x) is odd.
(3)
g(x) > 0 for x > 0
(4)
f (−x) = µ (−x)2 − 1 = µ(x2 − 1) = f (x)

f (x) is even.
(5)
x x
x3
Z Z ã Å
2 1
F (x) = f (u)du = µ(u − 1)du = µ − x = µ(x3 − 3x)
0 0 3 3
1 √ √
= µx(x + 3)(x − 3)
3
√
F (x) has exactly one positive root at x = 3.
√
F (x) is negative for 0 < x < 3.
√
F (x) is positive and nondecreasing for 3 < x.
and F (x) → ∞ as x → ∞.
All the properties of Liénard’s theorem are satisfied; therefore the system has a unique stable limit cy-
cle surrounding the origin.
7.5 Relaxation Oscillations

7.5.1
For the van der Pol oscillator with µ >> 1, show that the positive branch of the cubic nullcline begins at
xA = 2 and ends at xB = 1.
The cubic nullcline is from

1 3
ẋ = µ[y − F (x)] F (x) = x −x
3
y
(xA , yA )
(xB , yB )
xB occurs at the local minimum

F 0 (x) = x2 − 1 = 0 ⇒ x = ±1 ⇒ xB = 1
xA occurs at the intersection of the tangent line of the local maximum with F (x)
2
F (−1) = yA = 3 = F (x) ⇒ x = −1, 2 ⇒ xA = 2
7.5.3
ẍ + k(x2 − 4)ẋ + x = 1
We can do a substitution to transform the equation into the form of Example 7.5.1.
ẍ + k(x2 − 4)ẋ + x = 1 → z̈ + k (z + 1)2 − 4 ż + (z + 1) = 1

x→z+1
Å Å ãã
d 1
0 = z̈ + k (z + 1)2 − 4 ż + z = (z + 1)3 − 4z

ż + k
dt 3
1
F (z) = (z + 1)3 − 4z
3
w = ż + kF (z) ẇ = z̈ + k (z + 1)2 − 4 ż = −z

w z
y= ż = k y − F (z) ẏ = −
k k
Now that we transformed the equation into something similar to Example 7.5.1, we can perform the integral
over one branch, then multiply by 2 for the approximate period.
Z tB
T ≈2 dt
tA
z dy dz dz
− = ≈ F 0 (z) = (z + 1)2 − 4
k dt dt dt
−k (z + 1)2 − 4
dt ≈ dz
z
Next we have to find the turning point of the limit cycle for the integral bounds.
F 0 (z) = (z + 1)2 − 4 = z 2 + 2z − 3 = (z + 3)(z − 1)
zA = 1
28
zB 6= −3 and F (−3) = = F (zB ) ⇒ zB = 3
3
And now we can compute the integral for the period.
Z tB Z 1
(z + 1)2 − 4
T ≈2 dt = −2k dz = 2k ln 8 − ln(27)
tA 3 z
7.5.5

ẍ + µ |x| − 1 ẋ + x = 0
Define
x
ẋ = µ y − F (x) ẏ = −
µ
 
− x2 − x

x≤0 d x − 1 x ≤ 0

2
F (x) = ⇒ F (x) = = |x| − 1
 x2 − x 0 < x dx −x − 1 0<x
 
2
Verifying these two systems are the same,
d d
ẋ = ẍ = µ y − F (x) = µ ẏ − µF 0 (x)ẋ = µẏ − µ |x| − 1 ẋ

dt Å dt ã
x
=µ − − µ |x| − 1 ẋ = −x − µ |x| − 1 ẋ
µ

ẍ = −x − µ |x| − 1 ẋ ⇒ ẍ + µ |x| − 1 ẋ + x = 0
The period is approximately twice the traversal time of the right branch.
y
Z tB
T ≈2 dt
tA
dy x dy
ẏ = = − ⇒ dt = −µ
dt µ x
x2 p
0 ≤ x y = F (x) = − − x ⇒ x = 1 ± 1 + 2y
2
And here we need to pick the correct side of the parabola to integrate.
Z − 21 Z − 21 √ √
dy dy
T ≈ −2µ = −2µ √ = 2µ 2 − ln 1 + 2
1
2
x 1
2
1 + 1 + 2y
7.5.7
u̇ = b(v − u)(α + u2 ) − u v̇ = c − u b1 α1 8αb < 1
ab)
The nullclines are Å ã
1
u=c v=u +1
b(α + u2 )
u
c1 c c2
The vector field is similar to Example 7.5.1 as shown in the graph. Therefore there should be a stable limit
cycle as long as the vertical nullcline lies between the local maximum and local minimum, but we also have
to check that the intersection is an unstable fixed point.
Assuming 8αb 1, the local maximum and minimum occur when

ã!
(α + u2 ) − 2u2
Å
d d 1
v= u + 1 = +1=0
du du b(α + u2 ) b(α + u2 )2
(α + u2 ) − 2u2 + b(α + u2 )2 = 0
bu4 + (2αb − 1)u2 + α2 b + α = 0

p
2 −(2αb − 1) ± (2αb − 1)2 − 4(b)(α2 b + α)
u =
2b
√

8αb 2
1 − 2αb ± 1 − 8αb 1 − 2αb ± 1 − 2 + O (αb)
= =
2b 2b
1 − 2αb ± (1 − 4αb) 1 − 3αb
≈ = 2α,
2b b
We can determine stability by linearizing.

Ñ é
−αb − 3bu2 + 2buv − 1 b(α + u2 )
A=
−1 0
∆ = b(α + c2 ) > 0 τ = −αb − 3bu2 + 2buv − 1
∂ u̇
The stability of the fixed point is determined by the sign on τ , which is the upper-left element. ∂u < 0
∂ u̇
implies the point is stable, and ∂u > 0 implies it is unstable. We can actually see from the graph that the
unstable point occurs only between the local maximum and local minimum.
v
The u̇ has a positive increase to the rightward direction, either starting positive and becoming more positive
above the nullcline, or starting negative and increasing towards zero below the nullcline.
∂ u̇
Therefore ∂u > 0 when c lies on the middle branch. Thus the fixed point is unstable and there is a
stable limit cycle when
1 − 3αb
2α ≈ c1 < c < c2 ≈
b
c)
v
The fixed point is stable when c < c1 and all trajectories eventually approach the fixed point, but the route
can be very long depending on the initial condition. Consequently a bit of noise in the system can have
varying amounts of effect depending on which way the noise nudges the system.
7.6 Weakly Nonlinear Oscillators

7.6.1
√
sin(t 1 − 2 )
x(t, ) = √
et 1 − 2
2 ∂ 2 x

∂x
= x(t, 0) + + 0<ξ<
∂ (t,0) 2 ∂2 (t,ξ)
√
sin(t 1 − 2 )

x(t, 0) = √ = sin(t)
et 1 − 2 (t,0)
√ √ √
∂x te−t sin( 1 − 2 t) e−t sin( 1 − 2 t) te−t cos( 1 − 2 t)
=− √ + −
∂ 1 − 2 (1 − 2 )3/2 1 − 2

∂x
= −t sin(t)
∂ (t,0)
x(t, ) = sin(t) − t sin(t) + O 2

7.6.3
ẍ + x = x(0) = 1 ẋ(0) = 0
a)
First we solve the homogeneous equation.
ẍ + x = 0
x(t) = c1 sin(t) + c2 cos(t)
Next we can use the method of undetermined coefficients to deduce the nonhomogeneous solution.
ẍ + x =
x = At2 + Bt + C ẋ = 2At + B ẍ = 2A
2A + At2 + Bt + C = ⇒ x =
Now applying the initial conditions
x = + c1 sin(t) + c2 cos(t)
x(0) = + c2 = 1 ⇒ c2 = 1 −
ẋ(0) = c1 = 0
x = + (1 − ) cos(t)
b)
x(t, ) = x0 (t, ) + x1 (t, ) + 2 x2 (t, ) + O 3

ẍ + x = → (ẍ0 + ẍ1 + 2 ẍ2 ) + (x0 + x1 + 2 x2 ) =
(ẍ0 + x0 ) + (ẍ1 + x1 − 1) + 2 (ẍ2 + x2 ) = 0
Now we can solve for each of the order terms separately by using the method of undetermined coefficients
and applying the initial conditions
ẍ0 + x0 = 0 ⇒ x0 = cos(t)
ẍ2 + x2 = 0 ⇒ x2 = cos(t)
ẍ1 + x1 − 1 = 0 ⇒ x1 = 1
Now we can plug these all back into the starting equation
x(t, ) = x0 (t, ) + x1 (t, ) + 2 x2 (t, ) + O 3

≈ cos(t) + + 2 cos(t) = + (1 + 2 ) cos(t)

c)
The perturbation does not contain any secular terms because there is a single time scale in this problem.
The governing equation describes an undamped harmonic oscillator with a constant drive. The most general
formula for the solution is x(t) = A cos(ωt + φ) + d, with appropriately chosen constants to satisfy the initial
conditions and forcing term. The solution can also be interpreted by a shift.
ẍ + x = x(0) = 1 ẋ(0) = 0
z = x − → z̈ + z = 0 z(0) = 1 − ż(0) = 0
which now describes an undamped harmonic oscillator with a zero drive, but in either case the only time
scale is the natural period of the harmonic oscillator.
7.6.5
h(x, ẋ) = xẋ2 x(0) = a ẋ(0) = 0
Z 2π
dr 1
= hh sin(θ)i = h r cos(θ), −r sin(θ) sin(θ)dθ
dT 2π 0
Z 2π
1
= r3 cos(θ) sin3 (θ)dθ = 0 ⇒ r(T ) = r0
2π 0
Z 2π
dφ 1
r = hh cos(θ)i = h r cos(θ), −r sin(θ) cos(θ)dθ
dT 2π 0
Z 2π
1 r3
= r3 cos2 (θ) sin2 (θ)dθ =
2π 0 8
2 2 2
dφ r r r
= = 0 ⇒ φ(T ) = 0 T + φ0
dT 8 8 8
r02
r = r0 + O() ω = 1 + φ0 = 1 + + O 2

8
The amplitude of the closed orbit can be anything and the closed orbit is approximately circular.
»
r(0) = x(0)2 + ẋ(0)2 = a r(T ) = a
!
ẋ(0) a2
φ(0) = arctan −τ =0−0=0 φ(T ) = T
x(0) 8
Å 2 ã
a
x0 = r(T ) cos φ(T ) + τ = a cos T +τ
8
Å 2 ã Å 2 ã
a a
x(t, ) = a cos T + τ + O() = a cos + 1 t + O()
8 8
7.6.7
h(x, ẋ) = (x4 − 1)ẋ x(0) = a ẋ(0) = 0
Z 2π
dr 1
dT 2π 0
Z 2π
1 1
= −r(r4 cos4 (θ) − 1) sin2 (θ)dθ = r(8 − r4 )
2π 0 16
1 T
84 e 2
⇒ r(T ) = 1
(C + e2T ) 4
Z 2π
dφ 1
dT 2π 0
Z 2π
1
= −r(r4 cos4 (θ) − 1) sin(θ) cos(θ)dθ = 0
2π 0
dφ
= 0 ⇒ φ(T ) = φ0
dT
dr 1 1
= r(8 − r4 ) = 0 ⇒ r = 0, 8 4
dT 16
d 1 1 d 1
r(8 − r4 ) = r(8 − r4 ) 1 = −2

dr 16 r=0 2 dr 16 r=8 4
1
The origin is an unstable fixed point, and there is a stable limit cycle at r = 8 4 .
1 T
84 e 2
+ O() ω = 1 + φ0 = 1 + O 2

r(T ) = 1
(C + e2T ) 4
r(T )
1
84
C = 1000
T
10
Next, we solve the initial value problem.

1 T
» 84 e 2
r(0) = x(0)2 + ẋ(0)2 = a r(T ) = 41
8
a4 − 1 + e2T
!
ẋ(0)
φ(0) = arctan −τ =0−0=0 φ(T ) = 0
x(0)
1 T
84 e 2
x0 = r(T ) cos φ(T ) + τ = 41 cos(τ )
8
a4 − 1 + e2T
1 T 1 t
84 e 2 84 e 2
x(t, ) = 14 cos(τ ) + O() = 14 cos(t) + O()
8 8
a4 − 1 + e2T a4 − 1 + e2t
7.6.9
h(x, ẋ) = (x2 − 1)ẋ3 x(0) = a ẋ(0) = 0
Z 2π
dr 1
dT 2π 0
Z 2π
1 1 3
= −r3 (r2 cos2 (θ) − 1) sin4 (θ)dθ = r (6 − r2 )
2π 0 16
6 9
⇒ 2 ln(r) − 2 − ln(6 − r2 ) = T + C
r 2
Z 2π
dφ 1
dT 2π 0
Z 2π
1
= −r3 (r2 cos2 (θ) − 1) sin3 (θ) cos(θ)dθ = 0
2π 0
dφ
= 0 ⇒ φ(T ) = φ0
dT
dr 1 3 √
= r (6 − r2 ) = 0 ⇒ r = 0, 6
dT 16
d 1 3 d 1 3 9
r (6 − r2 ) =0 r (6 − r2 ) √ = −

dr 16 r=0 dr 16 r= 6 2
dr
dT
√
The origin is an unstable fixed point from the right, and there is a stable limit cycle at r = 6.
6 9
− ln(6 − r2 ) = T + C ω = 1 + φ0 = 1 + O 2

2 ln(r) −
r2 2
r(T )
√
6
C = −27
T
10
We can’t solve for the general solution with an initial condition, but we can solve for the limit cycle.
√
x0 = r(T ) cos φ(T ) + τ = 6 cos(τ )
√ √
x(t, ) = 6 cos(τ ) + O() = 6 cos(t) + O()
7.6.11
We start off by taking the expansion out one more term.
x(t, ) = x0 (τ, T ) + x1 (τ, T ) + 2 x2 (τ, T ) + O 3

d
ẋ = x0 (τ, T ) + x1 (τ, T ) + 2 x2 (τ, T ) + O 3
dt

= ∂τ x0 + ∂T x0 + ∂τ x1 + ∂T x1 + 2 ∂τ x2 + ∂T x2 + O 3

= ∂τ x0 + ∂T x0 + ∂τ x1 + 2 ∂T x1 + ∂τ x2 + O 3

d
ẍ = ∂τ x0 + ∂T x0 + ∂τ x1 + 2 ∂T x1 + ∂τ x2 + O 3
dt

= ∂τ τ x0 + ∂τ T x0 + ∂T τ x0 + ∂T T x0 + ∂τ τ x1 + ∂τ T x1

+ 2 ∂T τ x1 + ∂T T x1 + ∂τ τ x2 + ∂τ T x2 + O 3

= ∂τ τ x0 + ∂τ T x0 + ∂T τ x0 + ∂τ τ x1

+ 2 ∂T T x0 + ∂τ T x1 + ∂T τ x1 + ∂τ τ x2 + O 3

= ∂τ τ x0 + 2∂T τ x0 + ∂τ τ x1 + 2 ∂T T x0 + 2∂T τ x1 + ∂τ τ x2 + O 3

Then we plug these into the van der Pol oscillator and collecting terms.
ẍ + x + x2 − 1 ẋ = 0

O 1 : ∂τ τ x0 + x0 = 0
O : ∂τ τ x1 + x1 = −2∂T τ x0 − x20 − 1 ∂τ x0

O 2 : ∂τ τ x2 + x2 = −∂T T x0 − 2∂T τ x1 − x20 − 1 ∂T x0 + ∂τ x1 − 2x0 x1 ∂T x0

The O is Equation (39) from the text.
ï ò
1 1
∂τ τ x1 + x1 = −2r0 + r − r3 sin(τ + φ) + [−2rφ0 ] cos(τ + φ) − r3 sin 3(τ + φ)

4 4
The first two terms on the RHS are constrained to be zero as in Example 7.6.2, so the differential equation
simplifies to
1
∂τ τ x1 + x1 = − r3 sin 3(τ + φ)

4
with general solution
1 3
x1 = r (T ) sin 3(τ + φ) + A(T ) cos(τ ) + B(T ) sin(τ )
32
and applying the initial conditions gives
1
x1 = sin(3τ ) + A(T ) cos(τ ) + B(T ) sin(τ )
4
Unfortunately we don’t know much about A(T ) and B(T ) right away, but looking at the O 2 term for the

x(t, ) expansion we can see that the cos(τ ) and sin(τ ) attached to A(T ) and B(T ) will create at least one
secular term.
The work required to solve explicitly for A(T ) and B(T ) is not worth the trouble, but we can derive

differential equations for A(T ) and B(T ), just as in the O terms, in order to cancel out all the secular
terms in the O 2 terms.

1
sin(3τ ) resulting from the higher harmonic term − 14 r3 sin 3(τ + φ)

So all in all the leftover bit of x1 = 4
has hardly any effect at all. However, for the practically minded we could use these results to make a more
accurate prediction of the error of our approximate solution.
7.6.13
ẍ + x + x3 = 0 0 < 1 x(0) = a ẋ(0) = 0

a)
Since there’s no ẋ term, it’s quite straightforward to find the conserved energy equation
x2 x4
Z
1 1
E = ẋ2 + x + x3 dx = ẋ2 +

+
2 2 2 4
which we’ll use later.
Looking at our initial conditions x(0) = a and ẋ(0) = 0, we see that the t = 0 corresponds to the turning-
around point for the oscillation. The amplitude of the oscillation is therefore a.
An obvious relationship for the period of the limit cycle is

T T
Z T Z Z
T 4 4
T = dt ⇒ = dt ⇒ T = 4 dt
0 4 0 0
but we can change this into an integral in terms of x.

T
Z Z 0
4 dx
T =4 dt = 4
0 a ẋ
We can solve for ẋ in the conservation of energy equation for this particular limit cycle and substitute it into
the above integral.
1 2 x2 (0) x4 (0) a2 a4
E= ẋ (0) + + ⇒E= +
2… 2 4 2 4
a 4 x 4
ẋ = ± a2 + − x2 −
2 2
Since x is decreasing from a to 0 on this quarter of the cycle, ẋ < 0 and we pick the negative square root.
Z 0 Z 0 Z a
dx dx dx
T =4 =4 » =4 »
a ẋ a 4 x 4 4 4
a − a2 + − x2 −
2
0 a2 + − x2 − x
a
2 2 2
We can simplify this a bit with a substitution of x = ay, then factor.

Z a Z 1
dx ady
T =4 » 4 4
=4 » 4 4 4
0 a x
a2 + 2 − x2 − 2 0 a2 + a2 − a2 y 2 − a 2y
Z 1 Z 1
dy 4 dy
=4 » 2 2 y4
= » 2
»
a a
a

0 1+ 2 −y − 22 1+ 2 0 1 − y 1 − k2 y2
2
2
4 − a2
T () = » K(k) k2 = 2
1 + a2
2
1 + a2
where K(k) is the complete elliptic integral of the first kind.

b)
The power series is a tad easier to do if we go back a few steps and expand from there.
Z 1 Z 1
dy dy
T () = 4 » 2 2 y4
=4 q
a
a
Ä 2 2 4
ä
0 1 + 2 − y2 − 2 0 1 − y 2 + a2 − a 2y
ä2
a2 y 4 a2 y 4
Ä 2 ä Ä 2
a a
2 − 2 −
Z 1 3
1 2 2 2 3

=4 p − 32 + 52 + O dy
1−y 2
0 2 1 − y2 8 1 − y2
Z 1 4 2
a2 1 − y 4 4

4 2 3a 1 − y 3

= p − 32 + 25 + O dy
1−y 2
0 1 − y2 8 1 − y2
2 4
3a2 57a4 2
Å ã
3a π 57a π 2
+ O 3 = 2π 1 − + O 3

= 2π − + +
4 128 8 256
To check against Equation 7.6.57, we have to convert angular frequency ω to period T .
3
ω = 1 + a2 + O 2

8
3 2 2
Ç å
3a2
Å ã
2π 2π
T = ≈ = 2π 1 − + a + ···
ω 1 + 38 a2 8 8
3a2
Å ã
+ O 2

= 2π 1 −
8
So the two methods have agreement in at least the first two terms.
7.6.15
a)
The pendulum equation can be changed into
ẍ + sin(x) = 0 x(0) = a ẋ(0) = 0

1
sin(x) = x − x3 + O(x5 )
6
1
ẍ + sin(x) = 0 → ẍ + x − x3 = O(x5 ) ≈ 0
6
1
ẍ + x + h(x, ẋ) = 0 = h(x, ẋ) = −x3
6
Z 2π
dr 1
dT 2π 0
Z 2π
1
= −r3 cos4 (θ)dθ = 0 ⇒ r(T ) = r0
2π 0
Z 2π
dφ 1
r = hh cos(θ)i = h r cos(θ), sin(θ) cos(θ)dθ
dT 2π 0
Z 2π
1 3
= r3 sin3 (θ) cos(θ)dθ = r3
2π 0 8
dφ 3 2 3 2
= r ⇒ φ(T ) = r0 + φ0
dT 8 8
»
r(0) = x(0)2 + ẋ(0)2 = a r(T ) = a
3 1
ω = 1 + φ0 = 1 + r02 + O 2 ≈ 1 − a2

8 16
b)
The exact period from Exercise 6.7.4 is
Å ã
1 2 4

T = 2π 1 + a + O a
16
We can compare the exact result to our approximation by first converting the angular frequency ω to period
T , and then converting the expression for T into an infinite geometric series since α 1.
1 2 2 1 2 4
Ç Å ã Å ã å
2π 2π 1 2
T = ≈ 1 2 = 2π 1 + 16 a + 16 a + a + ···
ω 1 − 16 a 16
Å ã
1
= 2π 1 + a2 + O a4

16
And the two derivations do agree!
7.6.17

ẍ + 1 + γ + cos(2t) sin(x) = 0 0<1
a)

ẍ + 1 + γ + cos(2t) x ≈ 0

h(x, ẋ) = γ + cos(2t) x ⇒ h(θ) = r γ + cos 2(θ − φ) cos(θ)
Z 2π
1 1
r0 = h(θ) sin(θ)dθ = r sin(2φ)
2π 0 4
Z 2π Å ã Å ã
0 1 1 1 0 1 1
rφ = h(θ) cos(θ)dθ = r γ + cos(2φ) ⇒ φ = γ + cos(2φ)
2π 0 2 2 2 2
b)
Due to the r present in the r0 equation, r0 φ0 is a good approximation for r(0) ≈ 0. Thus φ should reach
equilibrium quickly enough relative to r that we can use the φ stable fixed point, assuming it exists, in the
r0 equation and solve.
Å ã
0 1 1 1
φ = γ + cos(2φ) = 0 ⇒ cos(2φ) = −2γ |γ| ≤
2 2 2
The graphs below illustrate the phase line for γ positive, and γ negative is similar.
φ̇ φ̇ φ̇
1 1 1
0≤γ< 2 γ= 2 2 <γ
φ φ φ
1
We won’t say much about |γ| = 2 because convergence to the fixed point is slower, so our assumptions may
no longer hold.
We can now approximate the solution to the r0 equation when r ≈ 0.

1 1 1 p 1
r0 = r sin(2φ) ≈ r sin arccos(−2γ) = r 1 − 4γ 2 |γ| <
4 4 4 2
1
This equation has an unstable fixed point at r = 0 and therefore γ < 2 = γc is the critical value.
c)
Now solving for r is relatively straightforward.
dr 1 p 1p 1
r0 = = r 1 − 4γ 2 ⇒ r(T ) = r0 ekT k= 1 − 4γ 2 |γ| <
dT 4 4 2
d)
For γc < |γ| we can see that φ0 > 0, but we can still make a solvable differential equation.
dr dr dT r0
= = 0
dφ dT dφ φ
1
r sin(2φ) r sin(2φ)
= 1 4 1 =
γ + cos(2φ) 2γ + cos(2φ)
Z Z2 2
dr sin(2φ)
= dφ
r 2γ + cos(2φ)
1
ln(r) = − ln 2γ + cos(2φ) + C
2
C
r(φ) = p
2γ + cos(2φ)
r(φ) is bounded, so trajectories that start near the origin stay near the origin and are closed orbits.
e)
As for the physical interpretation, let’s start with a pendulum. A frictionless pendulum will continue its
oscillation with the same period forever; but imagine that we have the power to raise or lower the mass at
the end of the rod whenever we want, which effectively changes the length of the rod. If we move the mass
up (shorten the rod) at the bottom of the swing, and move the mass farther away (lengthen the rod) at the
top of the swing, then the pendulum will swing higher each time.
The rationale is from conservation of angular momentum. Decreasing the rod length at the bottom of
the swing, which is when the mass is rotating the fastest, makes the mass swing faster in order to maintain
the same rotational inertia. Therefore we can put energy into the system this way. Also, increasing the rod
length at the top of the swing does not add or remove energy from the system because the mass has zero
velocity at the top of the swing, and all the energy is stored as potential energy.
So repeating in this fashion, which is twice the frequency of the pendulum since one-half of the pendu-
lum swing contains one full up-down-up cycle, will increase the oscillation amplitude more and more.
This reasoning also works for a regular swing with flexible chains because the leg-pumping motion causes
the chain to bend at the swinger’s handholds, effectively raising the swinger’s center of mass and changing
the length of the chain. (Consequently, pumping your legs on a swing with rigid rods instead of chains won’t
work because the swinger can’t bend the rods.) The differential equation for this problem doesn’t have an
instantaneous up-down motion, but instead has a continuously changing rod length.
g
θ̈ + sin(θ) = 0 ←→ ẍ + 1 + γ + cos(2t) sin(x) = 0
L
and the value of |γ| affects when the length of the rod is changed. Only some values of |γ| will change the
length of the rod at the correct values of φ to impart more energy into the system.
7.6.19
ẍ + x + x3 = 0 0 < 1 x(0) = a ẋ(0) = 0
a)
d2 d2 x
Å ã Å ã
d d d dx
= ω2 2

τ = ωt ⇒ ẍ = x τ (t) = x = ω
dt dt dt dt dτ dτ
d2 x 3
ẍ + x + x3 = 0 → ω 2 2
+ x(τ ) + x(τ ) = ω 2 x00 + x + x = 0
dτ
b)
x(τ, ) = x0 (τ ) + x1 (τ ) + 2 x2 (τ ) + O 3

ω = 1 + ω1 + 2 ω2 + O 3

ω 2 x00 + x + x3 = 0
2 00
1 + ω1 + 2 ω2 + O 3 x0 + x001 + 2 x002 + O 3
3
+ x0 + x1 + 2 x2 + O 3 + x0 + x1 + 2 x2 + O 3 =0

1 + 2ω1 + 22 ω2 + 2 ω12 + O 3 x000 + x001 + 2 x002 + O 3

3
+ x0 + x1 + 2 x2 + O 3 + x0 + x1 + 2 x2 + O 3 =0

There’s no need to expand everything out. We only need the O 1 and O terms, which we can pick out.
O 1 : x000 + x0 = 0

O : x001 + x1 = −2ω1 x000 − x30

c)
The x0 term has to match the initial conditions of the expand function x(t, ) since the series is supposed to
hold for all values, and all the other terms in the series consequently have zero as their initial conditions.
x(0) = a ẋ(0) = 0 ⇒ x0 (0) = a x00 (0) = 0 xk (0) = 0 x0k (0) = 0 for all k > 0
d)
x000 + x0 = 0 x0 (0) = a x00 (0) = 0 ⇒ x0 = a cos(τ )
e)
x001 + x1 = −2ω1 x000 − x30
= 2aω1 cos(τ ) − a3 cos3 (τ )

Triple angle formula : cos(3x) = 4 cos3 (x) − 3 cos(x)
1
= 2aω1 cos(τ ) − a3 cos(3τ ) + 3 cos(τ )

Å ã4
3 3 1
= 2aω1 − a cos(τ ) − a3 cos(3τ )
4 4
3 3
No secular terms ⇒ 2aω1 − a3 = 0 ⇒ ω1 = a2
4 3
f)
1
x001 + x1 = − a3 cos(3τ ) x1 (0) = 0 x01 (0) = 0
4
1 3
a cos(3τ ) − a3 cos(τ )

x1 =
32
7.6.21
ẍ + (x2 − 1)ẋ + x = 0 x(0) = a ẋ(0) = 0
x(τ, ) = x0 (τ ) + x1 (τ ) + 2 x2 (τ ) + O 3 ω = 1 + ω1 + 2 ω2 + O 3

ω 2 x00 + (x2 − 1)ωx0 + x = 0
O 1 : x000 + x0 = 0 x0 (0) = a x00 (0) = 0 x0 = a cos(τ )

O : x001 + x1 = −2ω1 x000 + 1 − x20 x00 x1 (0) = 0 x01 (0) = 0

= 2ω1 a cos(τ ) − 1 − a2 cos2 (τ ) a sin(τ )

Å ã
1 1
= 2ω1 a cos(τ ) − a 1 − a2 sin(τ ) + a3 sin(3τ )
4 4
No secular terms ⇒ ω1 = 0 a = 2
1
x0 = 2 cos(τ ) x1 =3 sin(τ ) − sin(3τ )
4
O : x2 + x2 = − ω1 + 2ω2 x000 − 2ω1 x001 + 1 − x20 x01 + ω1 x00 − 2x0 x1 x00
2
00 2

= −2ω2 x000 + 1 − x20 x01 − 2x0 x1 x00

3
1 − 4 cos2 (τ ) cos(τ ) − cos(3τ )

= 4ω2 cos(τ ) +
4

+ 2 cos(τ ) 3 sin(τ ) − sin(3τ ) sin(τ )
Å ã
1
= 4ω2 + cos(τ ) + · · ·
4
1
No secular terms ⇒ ω2 = −
16
1 2
+ O 3

ω =1−
16
7.6.23
ẍ − xẋ + x = 0 x(0) = a ẋ(0) = 0
x(τ, ) = x0 (τ ) + x1 (τ ) + 2 x2 (τ ) + O 3 ω = 1 + ω1 + 2 ω2 + O 3

ω 2 x00 + ωxx0 + x = 0
O 1 : x000 + x0 = 0 x00 (0) = 0

x0 (0) = a x0 = a cos(τ )
O : x1 + x1 = −2ω1 x0 + x0 x00
00
x1 (0) = 0 x01 (0) = 0
= −2ω1 a cos(τ ) − a2 cos(τ ) sin(τ )

1
− 2a2 sin(τ ) + a2 sin(2τ ) − 6aω1 τ sin(τ )

x1 =
6
No secular terms ⇒ ω1 = 0
1 2
x1 = a − 2 sin(τ ) + sin(2τ )
6
O 2 : x002 + x2 = x1 x00 + x0 x01 + ω1 x0 x00 − 2ω2 x000 − 2ω1 x001 − w12 x000

= x1 x00 + x0 x01 − 2ω2 x000

1
= − a3 − 2 sin(τ ) + sin(2τ ) sin(τ )

6
1
+ a3 cos(τ ) − 2 cos(τ ) − 2 cos(2τ ) + 2ω2 a cos(τ )

6
1
(−3a3 + 24aω2 ) cos(τ ) − 4a3 cos(2τ ) − a3 cos(3τ )

=
12
1
No secular terms ⇒ −3a3 + 24aω2 = 0 ⇒ ω2 = − a2
8
1
ω(a) = 1 − 2 a2 + O 3

8
7.6.25
ẍ + x + h(x, ẋ, t) = 0

x(t) = r(t) cos t + φ(t) ẋ(t) = −r(t) sin t + φ(t)
a)
We can derive a differential equation for r(t) and φ(t) by comparing the derivative of the x(t) definition to
the ẋ(t) definition.
d d
x(t) = r(t) cos t + φ(t) = ṙ cos(t + φ) − r(1 + φ̇) sin(t + φ)
dt dt
ẋ(t) = −r sin(t + φ)
−r sin(t + φ) = ṙ cos(t + φ) − r(1 + φ̇) sin(t + φ)
0 = ṙ cos(t + φ) − rφ̇ sin(t + φ)
We can derive another differential equation for r(t) and φ(t) by substituting the derivative of the ẋ(t)
definition into the main differential equation.
d d
ẋ(t) = ẍ = − r(t) sin t + φ(t) = −ṙ sin(t + φ) − r(1 + φ̇) cos(t + φ)
dt dt
0 = ẍ + x + h(x, ẋ, t)
= −ṙ sin(t + φ) − r(1 + φ̇) cos(t + φ) + r cos(t + φ) + h
= −ṙ sin(t + φ) − rφ̇ cos(t + φ) + h
Now we have two linear equations of ṙ(t) and φ̇(t) and can solve for each explicitly.
ṙ cos(t + φ) − rφ̇ sin(t + φ) = 0
ṙ sin(t + φ) + rφ̇ cos(t + φ) = h
ṙ cos2 (t + φ) − rφ̇ sin(t + φ) cos(t + φ) = 0
ṙ sin2 (t + φ) + rφ̇ cos(t + φ) sin(t + φ) = h sin(t + φ)
ṙ cos2 (t + φ) + ṙ sin2 (t + φ) = ṙ = h sin(t + φ)
−ṙ cos(t + φ) sin(t + φ) + rφ̇ sin2 (t + φ) = 0
ṙ sin(t + φ) cos(t + φ) + rφ̇ cos2 (t + φ) = h
rφ̇ sin2 (t + φ) + rφ̇ cos2 (t + φ) = rφ̇ = h cos(t + φ)

b) Z t+π
1
hri(t) = r̄(t) = r(τ ) dτ
2π t−π
We can use Leibniz’s integral rule here, assuming that ṙ(t) is continuous for τ ∈ [t − π, t + π].
Z t+π !
dhri d 1
= r(τ ) dτ
dt dt 2π t−π
! ! Z !
t+π
1 d d ∂
= r(t + π) (t + π) − r(t − π) (t − π) + r(τ ) dτ
2π dt dt t−π ∂t
Z t+π !
1
= r(t + π)(1) − r(t − π)(1) + 0 dτ
2π t−π
! Z t+π ≠ ∑
1 1 d dr
= r(t + π) − r(t − π) = r(τ ) dτ =
2π 2π t−π dτ dt
c)
≠ ∑
dhri dr

= = ṙ(t) = h(x, ẋ, t) sin(t + φ)
dt dt

= h r(t) cos(t + φ), −r(t) sin(t + φ), t sin(t + φ)
d)
We can intuitively see that these equations

r = r̄ + O φ = φ̄ + O
are true because the variable will be the average value over one oscillation plus the rate of change over that
cycle.

r = r̄ + hṙi = r̄ + h sin(t + φ) = r̄ + O

φ = φ̄ + hφ̇i = φ̄ + h cos(t + φ) = φ̄ + O
r̄
The other two equations follow easily from here.
dr̄

= hṙi = h r(t) cos(t + φ), −r(t) sin(t + φ), t sin(t + φ)
dt

= h r̄ cos(t + φ̄), −r̄ sin(t + φ̄), t sin(t + φ̄) + O
= h r̄ cos(t + φ̄), −r̄ sin(t + φ̄), t sin(t + φ̄) + O 2

dφ̄

r̄ = hrφ̇i = h r(t) cos(t + φ), −r(t) sin(t + φ), t cos(t + φ)
dt

= h r̄ cos(t + φ̄), −r̄ sin(t + φ̄), t cos(t + φ̄) + O
= h r̄ cos(t + φ̄), −r̄ sin(t + φ̄), t cos(t + φ̄) + O 2

8
Bifurcations Revisited
8.1 Saddle-Node, Transcritical, and Pitchfork Bifurcations

8.1.1
a)
ẋ = µx − x2 ẏ = −y
y y y
x x x
µ<0 µ=0 µ>0
b)
ẋ = µx + x3 ẏ = −y
y y y
x x x
µ<0 µ=0 µ>0
219
220 Chapter 8: Bifurcations Revisited
8.1.3
ẋ = µx − x2 ẏ = −y
(x, y) = (0, 0) and (µ, 0) are the fixed points.

Ñ é
µ − 2x 0
A=
0 −1
Ñ é
µ 0
A(0,0) =
0 −1
λ1 = µ λ2 = −1 ⇒ stable for µ < 0
λ1 = µ → 0 as µ → 0−
Ñ é
−µ 0
A(µ,0) =
0 −1
λ1 = −µ λ2 = −1 ⇒ stable for µ > 0
λ1 = −µ → 0 as µ → 0+
One fixed point is always at the origin and the other fixed point moves along the x-axis. The stable fixed
point moves along the negative half of the x-axis while the origin is an unstable fixed point. The two fixed
points combine at the origin with eigenvalues λ1 = 0 and λ2 = −1, then separate into a stable fixed point
at the origin and an unstable fixed point that moves along the positive half of the x-axis.
The fixed points exchange stability, but for the stable fixed point we have λ1 = µ → 0, as µ → 0 regardless
of which side of 0 µ is on.
8.1.5
True. The nullclines always intersect tangentially at a zero-eigenvalue bifurcation in two dimensions.
For a system with (x∗ , y ∗ ) as a fixed point
ẋ = f (x, y) ẏ = g(x, y)
f (x∗ , y ∗ ) = 0 g(x∗ , y ∗ ) = 0
Ñ é Ñ é
fx fy ∇f (x∗ , y ∗ )
A(x∗ ,y∗ ) = =
gx gy ∇g(x∗ , y ∗ )
With a zero eigenvalue, the two rows of the A matrix evaluated at the fixed point must be multiples of each
other, which are the gradient of f (x, y) and the gradient of g(x, y). The gradient vectors are orthogonal to
the level curves, which includes the nullclines. Therefore the tangent lines of the nullclines at the fixed point
are parallel because their normal vectors are parallel.
8.1.7
x
ẋ = y − ax ẏ = −by +
1+x
First find the fixed points, which are the intersections of the nullclines.
x 1 1
Å ã
y = ax y= ⇒ (x, y) = (0, 0) and − 1, − a
b(1 + x) ab b
Next find the linearization at the fixed points.

Ñ é
−a 1
A=
1
(1+x)2 −b
(x, y) = (0, 0) ⇒ ∆ = ab − 1 τ = −(a + b)

1
− 1, 1b − a ⇒ ∆ = ab − (ab)2 τ = −(a + b)

(x, y) = ab
Looking more closely at this system, we see that the fixed point goes to infinity when either a or b is
zero. Also notice that there is only one fixed point when ab = 1, and there are two fixed points when ab 6= 1,
a 6= 0, and b 6= 0. There are several cases for the stability of the two fixed points.
ab > 1 and a + b > 0 ⇒ The origin is a stable node and the other fixed point is a saddle point.
0 < ab < 1 and a + b > 0 ⇒ The origin is a saddle point and the other fixed point is a stable node.
ab > 1 and a + b < 0 ⇒ The origin is an unstable node and the other fixed point is a saddle point.
0 < ab < 1 and a + b < 0 ⇒ The origin is a saddle point and the other fixed point is an unstable node.
ab < 0 ⇒ The origin and the other fixed point are saddle points. This last case completes all possible cases.
1
From this we can conclude that transcritical bifurcations occur along the boundary a = in parameter
b
space.
8.1.9
ẍ + bẋ − kx + x3 = 0
First we transform this into a 2-dimensional system.
ẋ = y ẏ = kx − by − x3
Next we find the fixed points, which are the intersections of the nullclines.
kx − x3 √
y=0 y= ⇒ (x, y) = (0, 0) and (± k, 0)
b
Now we find the linearization at the fixed points.
Ñ é
0 1
A=
k − 3x2 −b
√
Ñ é
0 1 −b ± b2 + 4k
A(0,0) = λ1,2 =
k −b 2
√
Ñ é
0 1 −b ± b2 − 8k
A(±√k,0) = λ1,2 =
−2k −b 2
The second pair of fixed points comes into existence and splits off from the origin when k > 0, and the
stability of all three fixed points is determined by the sign of b. If b < 0 and k become positive, then
there is a subcritical pitchfork bifurcation. If b > 0 and k become positive, there is a supercritical pitchfork
bifurcation. When b = 0 there is a bifurcation when k becomes positive too. The origin changes from a
center to a saddle point, and the two new fixed points are centers.
Note that the new fixed points can be spirals or nodes depending on k and b, but we only care about
the stability for the stability diagram.
k
2 centers
1 saddle
1 saddle 1 saddle
2 unstable 2 stable
subcritical pitchfork supercritical pitchfork b

1 center
1 unstable 1 stable
8.1.11
u̇ = a(1 − u) − uv 2 v̇ = uv 2 − (a + k)v a, k > 0
The fixed points of the system are (u, v) = (1, 0), which always exists, and
Ç p p å
a ± a2 − 4(a + k)2 a ∓ a2 − 4(a + k)2
(u, v) = ,
2a 2(a + k)
if the inside of the square root is positive.
We definitely have a saddle-node bifurcation when the inside of the square root is negative with zero fixed-
points, the inside of the square root is zero with one fixed point, and the inside of the square root is positive
with two fixed points.
The inside of the square root is zero when

√
2 2 a a2
a − 4(a + k) = 0 ⇒ (a + k) = ⇒ k = −a ±
4 2
and thus saddle-node bifurcations occur on this curve in a − k parameter space.
8.1.13
ṅ = GnN − kn Ṅ = −GnN − f N + p G, k, f > 0
a)
dn dN
= GnN − kn = −GnN − f N + p
dt dt
G dn G2 nN Gn G dN G2 nN f GN pG
= − = − − + 2
k 2 dt k2 k k 2 dt k2 k2 k
Gn GN f pG
τ = kt x = y= a= b= 2
k k k k
dx dy
= x(y − 1) = −xy − ay + b
dτ dτ
b) (x, y) = 0, ab and (b − a, 1) are the fixed points.

Ñ é
y−1 x
A=
−y −x − a
Ñ é
b
a −1 0 b
A(0, b ) = ∆=a−b τ= −1−a
a
− ab −a a
Ñ é
0 b−a
A(b−a,1) = ∆=b−a τ = −b
−1 −b
c)
y a=2 b=1 y a=2 b=2 y a=2 b=3
x x x
d)
b
0, ab unstable

a
(b − a, 1) stable
b=
n
io
at
rc
fu
bi
al
ic
rit
sc
an
tr
0, ab stable

(b − a, 1) unstable
The picture only includes positive values of a due to the sign restrictions of the nondimensionalized variables.
8.1.15
The power of true believers
ṅA = (p + nA )nAB − nA nB ṅB = nB nAB − (p + nA )nB nAB = 1 − (p + nA ) − nB
a)
The first equation is the rate of change of the A-believers, which is the sum of gaining A-believers and
losing A-believers. Gain comes from AB-believers converting at a rate proportional to the fraction of A-
believers, and loss comes from A-believers converting to AB-believers at a rate proportional to the fraction
of B-believers.
The second equation is the rate of change of the B-believers, which is the sum of gaining B-believers
and losing B-believers. Gain comes from AB-believers converting at a rate proportional to the fraction of
B-believers, and loss comes from B-believers converting to AB-believers at a rate proportional to the fraction
of A-believers.
While there is a fixed fraction of unconvertable A-believers, p, there is no analogous group for the B-
believers. So the total fraction of A-believers is (p + nA ), but the total number of B-believers is nB .
An equation for ṅAB is unnecessary because nAB can be found by conservation of believers.
b)
nB
p = 0.1
p = 0.15
p = 0.2
p = 0.5
nA
An initial condition of p = 0.15 successfully converts everyone to A-believers, but an initial condition of
p = 0.1 is unsuccessful. Then there should be a 0.1 < pc < 0.15 that is a critical point for the change in
behavior.
c)
ṅB = 0 ⇒ nB = 0 or nB = −2nA − 2p + 1
ṅA = 0 and nB = 0 ⇒ nA = 1 − p
p
−4p + 1 ± 4p2 − 8p + 1
ṅA = 0 and nB = −2nA − 2p + 1 ⇒ nA =
6
Then the fixed points in terms of p are
Ç p p å
−4p + 1 ± 4p2 − 8p + 1 4p − 1 ∓ 4p2 − 8p + 1
(1 − p, 0) , − 2p + 1
6 3
The first fixed point is always there, but the other two fixed points will collide when
√
2 3
4p − 8p + 1 = 0 ⇒ p = 1 ±
2
√
3
out of which we have pc = 1 − 2 to be in the [0,1] interval.
The last two fixed points are complex if pc < p, and therefore to check if the remaining fixed point (1 − p, 0)
is globally attracting we only need to look at its linearization.
Ñ é Ñ é
d d
dnA ṅA dnA ṅB −2nA − 2nB − 2p + 1 −2nA − 2p
A= =
d d
dnB ṅA dnB ṅB −2nB −2nA − 2nB − p + 1
Ñ é
d d
dnA ṅA dnA ṅB −1 −1
A(1−p,0) = = λ1 = −1 λ2 = p − 1
d d
dnB ṅA dnB ṅB 0 p−1
The eigenvalues are always negative. Therefore (1 − p, 0) is a stable node and when pc < p all trajectories
will converge to it.
d)
A saddle-node bifurcation occurs at p = pc as two of the fixed points approach, collide, and then disappear.
8.2 Hopf Bifurcations

8.2.1
ẍ + µ(x2 − 1)ẋ + x = a
First convert the ODE into a system.
ẋ = y ẏ = a − µ(x2 − 1)y − x
Now we solve for fixed points, which gives (x, y) = (a, 0), and then compute the Jacobian.
Ñ é
0 1
A=
−2µxy − 1 −µ(x2 − 1)
Ñ é
0 1
A(a,0) = ∆ = 1 τ = −µ(a2 − 1)
−1 −µ(a2 − 1)
Hopf bifurcations occur when the trace changes from negative to zero to positive, and vice versa. The trace
changes sign when crossing the lines µ = 0 and a = ±1 in parameter space.
8.2.3
ẋ = −y + µx + xy 2 ẏ = x + µy − x2
y y
µ = −0.04 µ = −0.02
x x
y y
µ=0 µ = 0.02
x x
There’s an unstable limit cycle for µ negative, and the limit cycle constricts towards the origin as µ increases
to zero. (Only a portion of the spiral trajectory is shown for µ = −0.02 because the individual spiral
trajectories are very densely packed.) The limit cycle disappears when µ = 0 with a very slight unstable
spiral now at the origin. The unstable spiral’s radial growth increases as µ increases past zero. This is a
subcritical Hopf bifurcation.
8.2.5
ẋ = y + µx ẏ = −x + µy − x2 y
y y
µ = −0.1 µ = 0.1
x x
Supercritical Hopf bifurcation
8.2.7
ẋ = µx + y − x2 ẏ = −x + µy + 2x2
These graphs have bounds of (x, y) ∈ [−0.4, 0.8] × [−0.4, 0.8].

µ = −0.06 y µ = 0.06 y
x x
This is a subcritical Hopf bifurcation. For µ < 0 the origin is stable and there is an unstable limit cycle, but
for µ > 0 the origin is unstable and there is no limit cycle.
8.2.9
y x
Å ã Å ã
ẋ = x b − x − ẏ = y − ay x, y ≥ 0 a, b > 0
1+x 1+x
a)
The x and y nullclines are
x
x = 0 y = (1 + x)(b − x) y=0 y=
a(1 + x)
The fixed points are (x, y) = (0, 0), (b, 0), and the intersection of the curves, which is difficult to find because
we need to solve a cubic equation.
Next, we check stability. Ñ é

y −x
b− (x+1)2 − 2x 1+x
A=
y x
(x+1)2 1+x − 2ay
Ñ é
b 0
A(0,0) = b > 0 ⇒ unstable
0 0
Ñ é
−b
−b 1+b b
A(b,0) = λ1 = −b < 0 < = λ2 ⇒ unstable
0 b 1+b
1+b
As for the third fixed point, there can’t be any fixed point creating bifurcations since there is only one
intersection. There could be a Hopf bifurcation though.
b)
We’re pretty certain from the nullcline graph above that there is always a fixed point in the positive quad-
rant. Being careful though, we can note that the x nullcline always has a positive y intercept and the y
nullcline always starts at the origin. Also, the x nullcline has an x intercept at b, and the y nullcline has
1
a horizontal asymptote of a. Therefore the x and y nullcline will always have an intersection at a strictly
positive coordinate.
We should also check that the strictly positive fixed point is unique because of the y nullcline intersect-
ing on different sides of the x nullcline maximum. The easiest way to do this is to set the x and y nullclines
equal and set up the cubic equation we would need to solve to find all the roots.
x
(1 + x)(b − x) = y =
a(1 + x)
ax3 + (2a − ab)x2 + (1 + a − 2ab)x − ab = (x − x1 )(x − x2 )(x − x3 ) = 0

From this we know that x1 x2 x3 = ab > 0, so we have to have exactly one strictly positive intersection, or
exactly three strictly positive intersections. We can’t have three intersections graphically, which means the
strictly positive fixed point is unique.
c)
A Hopf bifurcation can only occur if the trace of the linearized system is zero. So we’ll plug in the strictly
positive fixed point (x∗ , y ∗ ) and find conditions on a and b.
y x
τ =b− − 2x + − 2ay
(x + 1)2 1+x
y∗ x∗
τ |(x∗ ,y∗ ) = b − ∗ − 2x∗ + − 2ay ∗
(x + 1) 2 1 + x∗
x∗
y ∗ = (1 + x∗ )(b − x∗ ) y∗ =
a(1 + x∗ )
∗ ∗
(1 + x )(b − x ) ∗ x∗ 2x∗
τ |(x∗ ,y∗ ) = b − − 2x + −
(x∗ + 1)2 1 + x∗ 1 + x∗
∗ ∗ ∗
b−x x 2x
=b− − 2x∗ + −
1 + x∗ 1 + x∗ 1 + x∗
b b−2
= b − 2x∗ − ∗
= 0 ⇒ x∗ =
1+x 2
x∗
(1 + x∗ )(b − x∗ ) =
a(1 + x∗ )
b−2
x∗ 2
a= =
(1 + x∗ )2 (b − x∗ ) b−2 2 b−2

1+ 2 b− 2
b−2
2 4(b − 2)
= = = ac
b 2 b+2 b2 (b + 2)

2 2
We should also check that the determinant is positive here to show there is a Hopf bifurcation.
Ñ ∗
é
−x∗
b − (x∗y+1)2 − 2x∗ 1+x∗
A(x∗ ,y∗ ) =
y∗ x∗ ∗
(x∗ +1)2 1+x∗ − 2ay
x∗ b−2
y ∗ = (1 + x∗ )(b − x∗ ) y∗ = x∗ = ⇒ b = 2x∗ + 2
a(1 + x∗ ) 2
Ñ é
b−x∗ −x∗
2x∗ + 2 − 1+x∗ − 2x∗ 1+x∗
A(x∗ ,y∗ ) =
b−x∗ x∗ x ∗
1+x∗ 1+x∗ − 2 1+x ∗

Ñ é Ñ é
b−x∗ −x∗ x∗ +2 −x∗
2− 1+x∗ 1+x∗ 2− 1+x∗ 1+x∗
= =
b−x∗ −x∗ x∗ +2 −x∗
1+x∗ 1+x∗ 1+x∗ 1+x∗
Ñ é Ñ é
2(1+x∗ )−(x∗ +2) −x∗ x∗ −x∗
1+x∗ 1+x∗ 1+x∗ 1+x∗
= =
x∗ +2 −x∗ x∗ +2 −x∗
1+x∗ 1+x∗ 1+x∗ 1+x∗
x∗ −x∗ −x∗ x∗ + 2 2x∗

∆= ∗ ∗
− ∗ ∗
= >0
1+x 1+x 1+x 1+x (1 + x∗ )2
The last inequality follows from x∗ being strictly positive.
Hence we have a whole curve of Hopf bifurcations in (a, b) parameter space.

d)
1
a = 0.08 b = 4 a= b=4
y y 12
x x
y a = 0.1 b = 4
x
8.2.11
ẍ + µẋ + x − x3 = 0
a)
ẋ = y ẏ = −µy − x + x3
Ñ é
0 1
A=
−1 + 3x2 −µ
Ñ é
0 1
A(0,0) = ∆ = 1 τ = −µ
−1 −µ
A negative µ close to zero implies an unstable spiral.
A positive µ close to zero implies a stable spiral.
b)
µ<0 y µ=0 y
x x
µ>0 y
The origin becomes a center when µ = 0, which is a degenerate case of a Hopf bifurcation because there are
an infinite number of closed orbits instead of a single closed orbit.
8.2.13
ẋ = y + µx ẏ = −x + µy − x2 y µ=0
ω = −1 f (x, y) = 0 g(x, y) = −x2 y
fx = 0 fy = 0 gx = −2xy gy = −x2
fxx = 0 fxy = 0 fyy = 0 gxx = −2y gxy = −2x gyy = 0
fxxx = 0 fxyy = 0 gxxy = −2 gyyy = 0
16a = fxxx + fxyy + gxxy + gyyy

1h i
+ fxy (fxx + fyy ) − gxy (gxx + gyy ) − fxx gxx + fyy gyy
ω
−1
16a = −2 ⇒ a = <0
8
Hence this is a supercritical Hopf bifurcation, and numerical simulations agree.
8.2.15
ẋ = µx + y − x2 ẏ = −x + µy + 2x2 µ=0
ω = −1 f (x, y) = −x2 g(x, y) = 2x2
fx = −2x fy = 0 gx = 4x gy = 0
fxx = −2 fxy = 0 fyy = 0 gxx = 4 gxy = 0 gyy = 0
fxxx = 0 fxyy = 0 gxxy = 0 gyyy = 0
16a = fxxx + fxyy + gxxy + gyyy

1h i
+ fxy (fxx + fyy ) − gxy (gxx + gyy ) − fxx gxx + fyy gyy
ω
−1
16a = −8 ⇒ a = <0
2
Hence this is a supercritical Hopf bifurcation, and numerical simulations agree.
8.2.17
−y1 + x1
ẋ1 = −x1 + F (I − bx2 − gy1 ) ẏ1 =
T
−y2 + x2
ẋ2 = −x2 + F (I − bx1 − gy2 ) ẏ2 =
T
1
F (x) =
1 + e−x
a)
Assuming that x∗1 = y1∗ = x∗2 = y2∗ = u, substituting into the equations gives
−u + u
ẋ1 = −u + F (I − bu − gu) ẏ1 = =0
T
−u + u
ẋ2 = −u + F (I − bu − gu) ẏ2 = =0
T
The ẏ1 and ẏ2 are satisfied, and what’s left are the ẋ1 and ẋ2 equations.
0 = −u + F (I − bu − gu)
1
u=
1 + e−(I−bu−gu)
The LHS and RHS are zero and positive respectively when u = 0. Also, the LHS and RHS are strictly
increasing and decreasing respectively, with the LHS greater than the RHS maximum at u = 0 within a
positive finite u. Therefore the curves intersect and this symmetric solution is unique.
b)
 
∂ ẋ1 ∂ ẋ1 ∂ ẋ1 ∂ ẋ1
 ∂x1 ∂y1 ∂x2 ∂y2 
 ∂ ẏ1 ∂ ẏ1 ∂ ẏ1 ∂ ẏ1 
 ∂x1 ∂y1 ∂x2 ∂y2 

A= 
 ∂ ẋ2
 ∂x ∂ ẋ2 ∂ ẋ2 ∂ ẋ2 
∂y1 ∂x2 ∂y2 

 1
∂ ẏ2 ∂ ẏ2 ∂ ẏ2 ∂ ẏ2
∂x1 ∂y1 ∂x2 ∂y2
 
−ge−I+bx2 +gy1 −be−I+gx2 +gy1
−1 (1+e−I+bx2 +gy1 )2 (1+e−I+bx2 +gy1 )2
0
 
1
− T1 0 0
 
T
 
=
 −be−I+bx1 +gy2 −I+bx1 +gy2 

−ge
 (1+e−I+bx1 +gy2 )2 0 −1 (1+e−I+bx 1 +gy 2
2) 


1 1
0 0 T −T
e−I+bu+gu 1 + e−I+bu+gu 1
= −
(1 + e−I+bu+gu )2 (1 + e−I+bu+gu )2 (1 + e−I+bu+gu )2
1 1
= −
1 + e−I+bu+gu (1 + e−I+bu+gu )2
= F (I − bu − gu) − F 2 (I − bu − gu)
= u − u2
 
−1 −g(u − u2 ) −b(u − u2 ) 0
 
1
− T1 0 0
 
T
 
A(u,u,u,u) = 
−b(u − u2 ) 0 −1 2 
−g(u − u )

 
1
0 0 T − T1
Now to prove the determinant law

a1,1 a1,2 b 0

a2,1 a2,2 0 0

b 0 a1,1 a1,2

0 0 a2,1 a2,2
(This is in no way a general proof but suffices for our purposes.) Expanding along the last column

a1,1 a1,2 b a1,1 a1,2 b

−a1,2 a2,1 a2,2 0 + a2,2 a2,1 a2,2 0

0 0 a2,1

b 0 a1,1

And then the bottom row

a1,1 a1,2 a b a a1,2

−a1,2 a2,1 + a2,2 b 1,2 + a1,1 a2,2 1,1
a2,1 a2,2 a2,2 0 a2,1 a2,2

a1,1 a1,2 a a

− (a2,2 b)2 + a1,1 a2,2 1,1 1,2

−a1,2 a2,1
a2,1 a2,2 a2,1 a2,2

a1,1 a1,2

(a1,1 a2,2 − a1,2 a2,1 )
− (a2,2 b)2
a2,1 a2,2

(a1,1 a2,2 − a1,2 a2,1 )2 − (a2,2 b)2

(a1,1 a2,2 − a1,2 a2,1 ) + a2,2 b (a1,1 a2,2 − a1,2 a2,1 ) − a2,2 b

(a1,1 + b)a2,2 − a1,2 a2,1 (a1,1 − b)a2,2 − a1,2 a2,1

a1,1 + b a1,2 a1,1 − b a1,2

a2,1 a2,2 a2,1 a2,2

Therefore
A B

= A + B A − B
B A

We can use this to compute the eigenvalues of the 4 × 4 matrix.

−1 − λ −g(u − u2 ) −b(u − u2 ) 0

1 1
− − λ 0 0

T T

=0
−b(u − u2 ) 0 −1 − λ −g(u − u2
)

1 1
0 0 T − T − λ

−1 − b(u − u2 ) − λ −g(u − u2 ) −1 + b(u − u2 ) − λ −g(u − u2 )

=0
1 1 1 1
− − λ − − λ

T T T T

c)
Looking at the A + B matrix, we can use that the determinant of a matrix is the product of the eigenvalues
and the trace of a matrix is the sum of the eigenvalues.

−1 − b(u − u2 ) −g(u − u2 )

1 1
−T

T

1 1
Å ã
2
− − g(u − u2 )

∆ = − 1 − b(u − u ) −
T T
2
1 + (b + g)(u − u )
= = λ1 λ2
T
Going back to the definition of u
1
u = F (I − bu − gu) =
1 + e−(I−bu−gu)
We also know that 0 < u < 1 because 0 < F |u=0 < 1 and F is strictly decreasing. Therefore the intersection
occurs for 0 < u < 1. Hence u − u2 > 0 and
1 + (b + g)(u − u2 )
0< = λ1 λ2
T
From this we know that λ1 and λ2 are both positive or both negative.
Then from the trace
1
τ = −1 − b(u − u2 ) −
T
2 1
0 > −1 − b(u − u ) − = λ1 + λ2
T
Therefore λ1 and λ2 are both negative.

d)
However, for the A − B matrix
1 1 1 + (g − b)(u − u2 )
Å ã
2
− − g(u − u2 )

∆ = − 1 + b(u − u ) − =
T T T
1
τ = −1 + b(u − u2 ) −
T
The determinant is positive or negative if g > b or b > g respectively, and the trace can be positive or
negative depending on the relative sizes of b and T .
From these facts about the eigenvalues, we can cause either a pitchfork bifurcation or a Hopf bifurcation at
(u, u, u, u) by varying g and T respectively. If we start with a negative trace, then by changing g, the system
will undergo a pitchfork bifurcation at (u, u, u, u) when the determinant switches sign. If we start with a
positive determinant, changing T will cause the system to undergo a Hopf bifurcation at (u, u, u, u) when
the trace switches sign.
e)
For small T , the variables settle to the stable fixed point (u, u, u, u).
z
Increasing T causes the stable limit cycle to appear.

z
And increasing T further enlarges the size of the limit cycle.

z
We can’t plot the limit cycle since it’s 4-dimensional, but we can plot the individual variables in time. We
used z for the horizontal label because all four of the variables x1 , y2 , x3 , and y4 have indistinguishable
graphs. So not only does the system tend toward a symmetric fixed point, the variables even follow the limit
cycle symmetrically.
8.3 Oscillating Chemical Reactions

8.3.1
ẋ = 1 − (b + 1)x + ax2 y ẏ = bx − ax2 y a, b > 0 x, y ≥ 0
a)
x, y = 1, ab is the fixed point.

Ñ é
−(b + 1) + 2axy ax2
A=
b − 2axy −ax2
Ñ é
b−1 a ∆=a
A b
=
1, a
−b −a τ = b − (1 + a)
The determinant is positive since a is positive, and the trace can vary greatly depending on b and a. Therefore
all we really know is that the fixed point can’t be a saddle point.
b)
Finding a trapping region implies that we’re looking for a limit cycle and b > 1 + a for an unstable fixed
point. A trapezoid with a circle around the fixed point will suffice.
y
1
2−
b(b + 1) b+1
a
1 b(b + 1)
Å ã
y =− x−2+ +
b+1 a
x
1
b+1
First, we should make sure that the fixed point is always inside the trapezoid. This is actually very easy
with the choice of trapezoid.
1 1 b b(b + 1)
x: <1<2− y:0< <
b+1 b+1 a a
Next, to confirm that the vector field points inward on the trapezoid edges, we’ll have to check that the
vector field points inward on all the edges (starting with the vertical edge and going clockwise).
ã2
1 1 ay
Å
h1, 0i · hẋ, ẏi = 1 − (b + 1) +a y= >0

x= b+11
b+1 b+1 (b + 1)2
b(b + 1)
h0, −1i · hẋ, ẏi b(b+1) = bx − ax2

= bx x(b + 1) − 1 > 0

y= a a

h−1, −1i · hẋ, ẏi = 1 − (b + 1)x + ax2 y − (bx − ax2 y) = −1 + x > 0

1
x>2− b+1

h0, 1i · hẋ, ẏi = bx − ax2 (0) = bx > 0

y=0
Lastly, since the linearization predicts an unstable fixed point, we can always draw a sufficiently small circle
enclosing the fixed point with the vector field pointing outward. Therefore we have our trapping region.
c)
The Hopf bifurcation has to occur at b = 1 + a = bc where the sign of the trace changes and there are no
other fixed points.
d)
The limit cycle exists for b > bc since we were able to construct a trapping region and apply the Poincaré-
Bendixson theorem. (However, we should keep in mind that there could be more than one limit cycle inside
the trapping region, but a stable limit cycle will appear after the Hopf bifurcation.)
e)
b=1+a+µ 0<µ1
Ñ é Ñ é
b−1 a a+µ a
A b
= =
1, a
−b −a −(1 + a + µ) −a
p
µ−
µ2 − 4a √ √
λ1,2 = ≈i a⇒ω= a
2
2π
Therefore the period of the limit cycle is approximately √ for 0 < µ 1.
a
8.3.3
4xy y
Å ã
ẋ = a − x − ẏ = bx 1 − a, b, x, y > 0 b 1
1 + x2 1 + x2
The solution for the problem done in the text is applicable here as well. The only portion that needs to be
redone is estimating the period of the limit cycle since b bc now.
y
a = 10 b 1
B
C
A
D
x
The entire period of the relaxation oscillation is the time spent on all four branches, but the time spent
on the BC and DA branches is negligible compared to the time spent on the AB and CD branches, since
ẏ ẋ, so we can neglect the fast branches.
Z tB Z tC Z tD Z tA Z tB Z tD
T = dt + dt + dt + dt ≈ dt + dt
tA tB tC tD tA tC
Next we change the integrals from dt to dx.

Z tB Z xB Z xB Z xB
dy dt dy 1 dy 1
dt = dx = dx = Ä ä dx
tA xA dx dy xA dx ẏ xA dx y
bx 1 − 1+x 2
From here we can make a substitution from the ẋ nullcline equation, making the integrand a function of x
only.
(a − x)(1 + x2 ) y a−x dy ax2 − a − 2x3
y= ⇒ = =
4x 1 + x2 4x dx 4x2
Z tB Z xB Z xB
dy 1 ax2 − a − 2x3 1
dt = ä dx ≈ dx
bx 1 − a−x
Ä 2
xA dx bx 1 − 1+x2 4x
y
tA xA 4x
We had to go this very roundabout way because we couldn’t use ẋ for the approximation, our first of which
is in the previous step. We know that the trajectory is close to the x nullcline, so we approximated the (x, y)
coordinates as exactly on the x nullcline. The problem is by definition ẋ = 0 on the x nullcline, which would
make an undefined integral.
Z tB yB xB xB
1 1 1
Z Z Z
dt = dx = 4xy dx = dx
tA yA ẋ xA a−x− 1+x2 xA 0
Now back to integrating.

Z tB xB xB
ax2 − a − 2x3 ax2 − a − 2x3
Z Z
dt ≈ dx = dx
tA xA 4bx3 1 − a−x
4x xA 4bx − bx (a − x)
3 2
xB
(3a2 − 125)x ln(5x − a) − 5a(2x2 + 5) + 125x ln(x)
=
25abx

xA
Next we can use a bit of basic calculus to find the turnaround points xA , xB , xC , and xD from the local
minimum and maximum of the x nullcline. Unfortunately, this requires solving a cubic equation, and plugging
the results back in gets rather messy. We don’t show the details here, but the approximate period would be
Z tB Z tD
T ≈ dt + dt
tA tC
xB
(3a2 − 125)x ln(5x − a) − 5a(2x2 + 5) + 125x ln(x)
=
25abx

xA
xD
(3a2 − 125)x ln(5x − a) − 5a(2x2 + 5) + 125x ln(x)
+
25abx

xC
8.4 Global Bifurcations of Cycles

8.4.1
ṙ = r(1 − r2 ) θ̇ = µ − sin(θ)
x, y ∈ [−1.25, 1.25] t ∈ [0, 100]
µ = 1.1
x
µ = 1.01
x
8.4.3
ẋ = µx + y − x2 ẏ = −x + µy + 2x2
Homoclinic bifurcation at µ ≈ 0.066.
These graphs have bounds of (x, y) ∈ [−0.4, 0.8] × [−0.4, 0.8].

µ = 0.060 y µ = 0.072 y
x x
In each case, the gray trajectory starts near the origin. The left graph traps the trajectory in a stable limit
cycle, but the the stable limit cycle eventually makes a homoclinic orbit with the saddle point and then
disappears in a homoclinic bifurcation. The right graph is after the homoclinic bifurcation.
8.4.5
ẍ + x + bx3 + k ẋ − ax − F cos(t) = 0

0 < << 1 b, k, F > 0
h(x, ẋ) = bx3 + k ẋ − ax − F cos(t) = bx3 + k ẋ − ax − F cos(θ − φ)h r cos(θ), −r sin(θ)

= br3 cos3 (θ) − kr sin(θ) − ar cos(θ) − F cos(θ − φ)
Z 2π
dr 1
dT 2π 0

− kr + F sin(φ)
=
2
Z 2π
dφ 1
dT 2π 0
− 4ar − 3br3 + 4F cos(φ)

=
8
3

dφ − 4ar − 3br + 4F cos(φ)
=
dT 8r
8.4.7
x0 = (r0 , rφ0 ) g(rφ) ≡ 1
The reason for the (r0 , rφ0 ) instead of (r0 , φ0 ) is due to using the polar coordinate unit vectors. x0 = (r0 , rφ0 ) =
r0 r̂ + rφ0 φ̂ and the polar coordinate unit vectors, unlike Cartesian coordinates, are dependent on the angle
of the point.
− 4ar − 3br3 + 4F cos(φ)

0 − kr + F sin(φ) 0
r = φ =
2 8r
1 ∂ 1 ∂
∇ · x0 = (rr0 ) + (rφ0 )
r ∂r r ∂φ
1 ∂ kr2 + F r sin(φ) 1 ∂ 4ar − 3br3 + 4F cos(φ)
Å ã Å ã
=− −
r ∂r 2 r ∂φ 8
1 2kr + F sin(φ) 1 −4F sin(φ)
Å ã Å ã
=− − = −k
r 2 r 8
Thus, the value of ∇ · (gx0 ) < 0 since k > 0 and there are no closed orbits for the averaged system by Dulac’s
criterion.
8.4.9
− 4ar − 3br3 + 4F cos(φ)

0 − kr + F sin(φ) 0
r = φ =
2 8r
a)

0 − kr + F sin(φ) kr
r = =0⇒− = sin φ
2 F
k2 r2
sin2 (φ) + cos2 (φ) = 1 ⇒ cos2 (φ) = 1 − 2
F
3

− 4ar − 3br + 4F cos(φ)
φ0 = = 0 ⇒ 4ar − 3br3 + 4F cos(φ) = 0
8r
3br3 − 4ar = 4F cos(φ)
k2 r2
Å ã
(3br − 4ar) = 16F cos (φ) = 16F 1 − 2 = 16F 2 − 16k 2 r2
3 2 2 2 2
F
16k 2 r2 + (3br3 − 4ar)2 = 16F 2
" ã2 #
3
Å
r2 k2 + br2 − a = F2
4
b)
a
b=0
r
c)
Increasing b slightly positive, and then a bit more gives the following two graphs.
a 0 < b < bc a bc < b
r r
The tip of the b = 0 graph continually moves up until the range is eventually overlapped. Now there are
three different r values in a range of a values where the graph doubles back.
We can solve for a in terms of r somewhat easily and then find when the solutions overlap.
…
3 2 F2
a = br ± − k2
4 r2
The graph will double back when there is zero slope with respect to r.
da 3 F2
= br ± p =0
dr 2 r2 (F − kr)(F + kr)
We throw out the plus from the ± sign because we need the terms to cancel.
da 3 F2
= br − p =0
dr 2 r2 (F − kr)(F + kr)
We also know that the second derivative will equal zero at b = bc when the graph is cubic-shaped and about
to develop a local minima.
d2 a 3 2F 4 − 3F 2 k 2 r2
2
= b− 3 = 0
dr 2 r3 (F − kr)(F + kr) 2
We can find the conditions under which there is a real value of r as a solution to these two equations. The
result is √
32 3k 3
b≥ = bc
27F 2
d)
r
b
a
8.4.11
4
k=1 b= 3 F =2
a)
ẍ + x + bx3 + k ẋ − ax − F cos(t) = 0

0 < << 1 b, k, F > 0
Averaged system
− 4ar − 3br3 + 4F cos(φ)

0 − kr + F sin(φ) 0
r = φ =
2 8r
r a = −1
φ
r a=0
r a=1
r a=2
φ
b)
r a = 2.8
c)
k = 1, b = 43 , F = 2, = 0.01
The simulation is run from t = 0 to t = 1200. a starts at −1, linearly increases to 5 at t = 300, then linearly
decreases back to −1 at t = 600.
t
8.5 Hysteresis in the Driven Pendulum and Josephson Junction

8.5.1
1
f (I) =
ln(I − Ic )
Taking the first few derivatives
!
d 1 −1
= 2
dI ln(I − Ic ) (I − Ic ) ln(I − Ic )
!
d2 1 2 1
= 3 + 2
dI 2 ln(I − Ic ) (I − Ic )2 ln(I − Ic ) (I − Ic )2 ln(I − Ic )
!
d3 1 −6 −6 −2
= 4 + 3 + 2
dI 2 ln(I − Ic ) (I − Ic )3 ln(I − Ic ) (I − Ic )3 ln(I − Ic ) (I − Ic ) ln(I − Ic )
3
we can rewrite them all in terms of f (I) = f .
−f 2
f (1) =
I − Ic
2f 3 f2 2f 3 + f 2
f (2) = + =
(I − Ic )2 (I − Ic )2 (I − Ic )2
−6f 4 −6f 3 −2f 2 −6f 4 − 6f 3 − 2f 2
f (3) = + + =
(I − Ic )3 (I − Ic ) 3 (I − Ic ) 3 (I − Ic )3
Maybe we can rewrite all the derivatives in terms of f .
Pn+1 p
p=2 ap f
f (n) =
(I − Ic ) n
Pn+1 Pn+1
(n+1)
ap pf p−1 (1)
p=2 p=2 ap f
p
f = f − n
(I − Ic )n (I − Ic )n+1
Pn+1 p−1 Pn+1 p
p=2 ap pf −f 2 p=2 ap f
Å ã
= − n
(I − Ic ) n I − Ic (I − Ic )n+1
Pn+1 p+1 Pn+1 p
− p=2 ap pf p=2 ap f
= − n
(I − Ic )n+1 (I − Ic )n+1
− p=2 ap pf p+1 − n n+1
Pn+1 p Pn+2 p
p=2 ap f p=2 ãp f
P
= =
(I − Ic )n+1 (I − Ic ) n+1
The coefficients are a bit confusing to figure out, but we have a pattern. Now we just need to make sure that
there is at least one term in the numerator for all derivatives in case everything cancels to zero somehow.
Looking a little closer at the recurrence relation, the coefficient of the highest power of f in the numerator is
Pn+1 p
(n) p=2 ap f
f = an+1 = (−1)n n!
(I − Ic )n
So the highest power of f in f (n) has an infinite limit at Ic . (In fact, all terms have an infinite limit as long
as the coefficient doesn’t cancel out.) Therefore f (I) has infinite derivatives of all orders at Ic .
8.5.3
N
Å ã
Ṅ = rN 1 −
K(t)
a)
K(t) is positive and smooth, and T -periodic in t implies K(t) has a maximum and minimum Kmax and
Kmin on the cylindrical phase space. Notice that N > Kmax ⇒ Ṅ < 0 and N < Kmin ⇒ Ṅ > 0 from the
differential equation.

Next, we’ll define a Poincaré map P N (t) = N (t + T ) as successive iterates of trajectories around the
cylinder. Just as in the text, we know P (N ) is continuous and monotonic, and there must be a fixed point
of the Poincaré map P (N ∗ ) = N ∗ corresponding to a limit cycle of the differential equation.
P (N )
P (Nj )
P (Ni )
Ni N∗ Nj
We also know that Kmin < N ∗ < Kmax since N = Kmin and N = Kmax form a trapping region for the stable
limit orbit.
b)
We don’t know the stable limit cycle is unique from part (a) since there could be multiple crossings. We’ll
need to prove it.
First, we do a change of variables.
1 ẋ N r 1
Å ã Å ã
x= ⇒ − 2 = Ṅ = rN 1 − = 1−
N x K(t) x xK(t)
r
⇒ ẋ + rx =
K(t)
This is a first-order linear ODE, which can be solved with an integrating factor.
r
ẋ + rx =
K(t)
rert
ert ẋ + rert x =
K(t)
d rt rert
e x =
dt K(t)
rert
Z
ert x = dt + C
K(t)
!
rert
Z
−rt
x(t) = e dt + C
K(t)
We know the integral has a unique solution because K(t) being positive and smooth is well-behaved, and
we know that all solutions converge to this unique solution.
!
rert rert
Z Z
−rt
lim x(t) = lim e dt + C −→ e−rt dt
t→∞ t→∞ K(t) K(t)
Therefore the stable limit cycle is unique.
8.5.5
θ̈ + αθ̇|θ̇| + sin(θ) = F α, F > 0
a)
θ̇ = ν ν̇ = F − αν|ν| − sin(θ)

The fixed points are θ, ν = arcsin(F ), 0 and π − arcsin(F ), 0 F < 1
Ñ é
0 1
A=
− cos(θ) −2α|ν|
Ñ é
1
0 1 λ1,2 = ±i(1 − F 2 ) 4
A = √
arcsin(F ),0
− 1 − F2 0 center
Ñ é
1
0 1 λ1,2 = ±(1 − F 2 ) 4
A = √
π−arcsin(F ),0
1 − F2 0 saddle point
Now we aren’t sure if the predicted center is really a center. A quick numerical simulation will show it’s
actually a stable spiral. Now we look for a suitable Liapunov function. A little bit of experimentation with
an “energy” relative to the fixed point gives
1 2 p
V (θ, ν) = ν − F θ − cos(θ) + F arcsin(F ) + 1 − F 2
2
V̇ = ν ν̇ − F θ̇ + sin(θ)θ̇ = ν F − αν|ν| − sin(θ) − F ν + sin(θ)ν = −αν 2 |ν|

Our V is not strictly Liapunov because V̇ (θ, 0) = 0, meaning V̇ is not negative on the line ν = 0. However,
this suffices to prove that the fixed point is a stable spiral. That’s because the trajectories always move
inward to lower and lower contours of V , except at instants when they cross the line ν = 0. At each such
instant they move tangentially to the V -contour, but an instant later they resume moving inward. Also, the
V contours approach ellipses as we approach the fixed point, so the geometry of this situation is clear.
Thus the Liapunov “energy” V is almost always decreasing (but never increasing) towards the fixed point
after each intersection with ν = 0, and the “energy” of the fixed point is zero. Therefore the fixed point is a
stable spiral.
b)
We’ll construct a trapping region to prove the existence of the stable limit cycle.
ν̇ = F − αν|ν| − sin(θ) ≥ 0
F − sin(θ) ≥ F − 1 ≥ αν|ν|
…
F −1
ν= ⇒ ν̇ ≥ 0
α
ν̇ = F − αν|ν| − sin(θ) ≤ 0
F − sin(θ) ≤ F + 1 ≤ αν|ν|
…
F +1
ν= ⇒ ν̇ ≤ 0
α
There are no fixed points in this region. Therefore there must be at least one stable limit cycle within
… …
F −1 F +1
0 ≤ θ < 2π ≤ν≤
α α
Next, we can use the total energy of the system after each cycle to prove there is a unique stable limit
cycle within the bounds. Even though the system is losing energy to friction and gaining energy from the
externally applied torque, the system will repeat the same velocity at the same angle each cycle. Therefore
the total energy at that position will always be the same.
The total energy for a pendulum is given by
1 2
E= ν + 1 − cos(θ)
2
Now we can repeat the reasoning in Section 8.5, “Uniqueness of the Limit Cycle,” with some small changes.
E is different, and also y ⇒ ν and φ → θ. Starting with Equation (7),
dE d 1 2 dν
= ν + 1 − cos(θ) = ν + sin(θ)
dθ dθ 2 dθ
dν ν̇ F − αν|ν| − sin(θ)
= =
dθ θ̇ ν
Hence
dE F − αν|ν| − sin(θ)
=ν + sin(θ) = F − αν|ν|
dθ ν
Going back to the change in energy integral
Z 2π Z 2π Z 2π
dE 2πF
0= dθ = F − αν|ν| dθ ⇒ ν(θ)ν(θ)dθ =
0 dθ 0 0 α
To prove uniqueness, suppose there were two distinct limit cycles. Following the notation and reasoning in
Section 8.5 of the text, denote these by νU (θ) and νL (θ). Hence νU (θ) > νL (θ) implies
Z 2π
Z 2π
νU (θ)νU (θ)dθ > νL (θ)νL (θ)dθ
0 0
which is a contradiction. Therefore the limit cycle is unique.

c)
1 2
u(θ) = ν t(θ)
2
du 1 d 2 d dν dt 1
= ν t(θ) = ν ν t(θ) = ν = ν ν̇ = ν̇ = θ̈
dθ 2 dθ dθ dt dθ ν
d)
ν > 0 −→ θ̇ > 0
θ̈ + αθ̇|θ̇| + sin(θ) = F −→ θ̈ + αθ̇2 + sin(θ) = F

du 1 2
= θ̈ u= ν
dθ 2
du
θ̈ + αθ̇2 + sin(θ) = F −→ + 2αu + sin(θ) = F
dθ
This is an integrating factor problem with solution
F 1 2α
u(θ) = + cos(θ) − sin(θ) + Ce−2αθ
2α 1 + 4α2 1 + 4α2
And the limit cycle is when the transient decays to zero, leaving
F 1 2α
u(θ) = + cos(θ) − sin(θ)
2α 1 + 4α2 1 + 4α2
e)
We can find F at the homoclinic bifurcation by constraining u(θ) to go through one of the fixed points,
which has to be the saddle point because homoclinic bifurcations occur with saddles.
The saddle point is

(θ, ν) = π − arcsin(F ), 0 −→ (θ, u) = π − arcsin(F ), 0
F 1 2α
u(θ) = + cos(θ) − sin(θ)
2α 1 + 4α2 1 + 4α2
F 1 Ä p ä 2α
0= + 2
− 1 − F2 − F
2α 1 + 4α 1 + 4α2
2α F −1 p
2
F− = 1 − F2
1 + 4α 2α 1 + 4α2
p
4α2 F − F (1 + 4α2 ) = −2α 1 − F 2
p
− F = −2α 1 − F 2
F 2 = 4α2 (1 − F 2 ) = 4α2 − 4α2 F 2
F 2 (1 − 4α2 ) = 4α2
2α
F = Fc (α) = √
1 + 4α2
And that’s where the homoclinic bifurcation occurs.
8.6 Coupled Oscillators and Quasiperiodicity

8.6.1
θ̇1 = ω1 + sin(θ1 ) cos(θ2 ) θ̇2 = ω2 + sin(θ2 ) cos(θ1 ) ω1 , ω2 ≥ 0
a)
Let φ1 = θ1 − θ2 , which corresponds to the phase difference of the two oscillators.
φ̇1 = θ̇1 − θ̇2 = ω1 + sin(θ1 ) cos(θ2 ) − ω2 − sin(θ2 ) cos(θ1 )
= ω1 − ω2 + sin(θ1 − θ2 ) = ω1 − ω2 + sin(φ1 )
The fixed points correspond to phase-locked solutions of the original system; i.e., periodic trajectories on the
torus.
φ˙1 = ω1 − ω2 + sin(φ1 ) = 0
φ1 = 2π − arcsin(ω1 − ω2 ) π + arcsin(ω1 − ω2 ) |ω1 − ω2 | ≤ 1
A graphical approach for ω1 − ω2 ≥ 0 (and similarly for ω1 − ω2 ≤ 0) shows
φ̇1
ω1 − ω2
φ1
φ̇1
ω1 − ω2
φ1
φ̇1
ω1 − ω2
φ1
φ̇1
φ1
ω1 − ω2
Next, we’ll look at φ2 = θ1 + θ2 , the fixed points of which correspond to periodic trajectories on the torus
running in opposite directions, in contrast to the fixed points of φ1 . (θ̇1 + θ̇2 = 0 ⇒ θ̇1 = −θ̇2 )
φ̇2 = θ̇1 + θ̇2 = ω1 + sin(θ1 ) cos(θ2 ) + ω2 + sin(θ2 ) cos(θ1 )
= ω1 + ω2 + sin(θ1 + θ2 ) = ω1 + ω2 + sin(φ2 )
The fixed points are
φ˙2 = ω1 + ω2 + sin(φ2 ) = 0
φ2 = 2π − arcsin(ω1 + ω2 ) π + arcsin(ω1 + ω2 ) ω1 + ω2 ≤ 1
A graphical approach for ω1 − ω2 ≥ 0 is the same as previously, except this time there is no ω1 + ω2 ≤ 0
case to consider because ω1 , ω2 ≥ 0.
Now we have to go through all the possible cases for ω1 and ω2 and the existence of fixed points in the
(φ1 , φ2 ) system and see what happens in the (θ1 , θ2 ) system. (Actually we only show graphs from about half
the regions due to symmetry.)
1 1
θ2 ω1 = 4 ω2 = 4
θ1
Fixed points
π 13π 5π 17π 7π 7π 11π 11π 13π π 17π 5π 19π 19π

Å ã Å ã Å ã Å ã Å ã Å ã Å ã
(θ1 , θ2 ) = , , , , , , , , , , , , ,
12 12 12 12 12 12 12 12 12 12 12 12 12 12
1
θ2 ω1 = 2 ω2 = 2
θ1
θ2 ω1 = 1 ω2 = 1
θ1
Stable and unstable limit cycles are present.

1 1
θ2 ω1 = 2 ω2 = 2
θ1
Fixed points
π 5π 3π 3π 5π π 7π 7π
Å ã Å ã Å ã Å ã
(θ1 , θ2 ) = , , , , , , ,
4 4 4 4 4 4 4 4
In this phase portrait, the fixed points are about to disappear in an infinite-period bifurcation if we view the
phase portrait as on a torus. In their places will be limit cycles.
θ2 ω1 = 1 ω2 = 2
θ1
θ2 ω1 = 0 ω2 = 0
θ1
Fixed points
0π 0π 0π 2π 0π 4π 1π 1π 1π 3π 2π 0π
Å ã Å ã Å ã Å ã Å ã Å ã
(θ1 , θ2 ) = , , , , , , , , , , , ,
2 2 2 2 2 2 2 2 2 2 2 2
2π 2π 2π 4π 3π 1π 3π 3π 4π 0π 4π 2π 4π 4π
Å ã Å ã Å ã Å ã Å ã Å ã Å ã
, , , , , , , , , , , , ,
2 2 2 2 2 2 2 2 2 2 2 2 2 2
θ2 ω1 = 0 ω2 = 1
θ1
Fixed points
0π 3π 2π π 4π 3π
Å ã Å ã Å ã
(θ1 , θ2 ) = , , , , ,
2 2 2 2 2 2
b)
We have saddle-node bifurcation of cycles occurring along the |ω1 + ω2 | = 1 lines and infinite-period bifur-
cations occurring along the ω1 − ω2 = 1 line.
c)
1
=
2|
ω2
ω
−
1
|ω
s
No fixed points
e
cl
cy
1
No limit cycles
=
of
2|
n
ω
io
−
at
rc
1
|ω
fu
bi
de
o
-n
le
dd
Sa
Limit cycles
ω1
es
cl
+
cy
ω2
of
=
n
1
io
at
In
rc
fin
fu
ite
bi
-p
e
od
er
No fixed points
-n
io
d
le
No limit cycles
dd
bi
fu
Sa
Fixed
rc
at
points
io
ω1
n
The oscillator death region is the lower-left triangle. Once we cross the infinite-period bifurcation ω1 +ω2 = 1
line, the limit cycles are broken by fixed points and the oscillations stop.
8.6.3
This problem is actually quite common and there are multiple proofs. Here is one that doesn’t require much
prerequisite knowledge.
ω1
Without loss of generality, assume that q = (0, 0), as we can translate to any point and ω = < 1
ω
since exchanging ω1 with ω2 is an equivalent problem. Also, we only need to prove that the irrational orbit
is dense in {(θ1 , θ2 ) : θ2 = 0} (the bottom of the phase space) since a dense set at a cross section of the torus
will also be dense everywhere when translated by the irrational flow. So now we have a somewhat simpler
problem to work on.
We’ll be using Dirichlet’s approximation theorem, which implies that for every irrational number ω there
exist integers a and b such that
a 1
ω − < 2

b b
a
The orbits that start at (θ1 , θ2 ) = (0, 0) with slope intersect the bottom of the phase space at
b
1 2 b−1
ß ™
0, , , · · ·
b b b
1
So the rational orbit passes within distance of every point at the bottom of the phase space.
2b
Now applying the implication of Dirichlet’s approximation theorem

a n 1
n ω − < 2 < n = 0, 1, 2, . . . (b − 1)

b b b
1 n
As long as 0 ≤ n ≤ b − 1, the nth intersection of the irrational orbit is within distance of and therefore
b b
1
the first b − 1 iterations of the irrational orbit are within distance of every point at the bottom of the
b
phase space. Hence the irrational orbit can be made arbitrarily close to any point at the bottom of the phase
space, and consequently any point on the entire torus, by picking a large enough value for b.
8.6.5
a) b) c)
y y y
x x x
2 5
ω=3 ω= 3 ω= 3
d) e) f)
y y y
x x x
√ √
1+ 5
ω= 2 ω=π ω= 2
8.6.7
h2 h
mr̈ = −k θ̇ = m, k, h > 0
mr3 mr2
a)
r(t) = r0 ⇒ r̈ = 0
ã 13
h2 h2 h2
Å
mr̈ = − k −→ 0 = − k ⇒ r0 =
mr 3 mr03 mk
ã2 Å 2 ã 13
h h mk 3 k
Å
θ̇ = = = = ωθ
mr02 m h2 mh
b)
First, we’ll transform the r̈ equation to a system and then linearize about (r, ṙ) = (r0 , 0).
h2 k
x=r y = ṙ ẋ = y ẏ = 2 3
−
m x m
Ñ é Ñ é
0 1 0 1
A= A(r0 ,0) =
−3h2 −3h2
m 2 x4 0 m2 r04
0
Then the frequency of small radial oscillations is

3h2 √ h √
ωr = 2 4 = 3 2 = 3ωθ
m r0 mr0
c)
The winding number
ωr ωθ √
= = 3
ωθ ωθ
is irrational and hence the small radial oscillations are quasiperiodic.
d)
The linearization from part (c) predicts a center, but we didn’t actually check that it’s a center. It really is a
center and we can find a conserved quantity; or, if we look closely, we see that the r equation is a nonlinear
spring with no damping and no r = r0 rest position. The solution to r(t) is periodic for any initial conditions.
The periodicity of r(t) implies that θ̇ is periodic. The θ̇ equation is also strictly positive. Both these
facts together imply that θ(t) has a constant finite period, which is determined by the r(t) initial conditions.
The period can vary for different r(t) initial conditions, but the period is constant for each choice of r(t)
initial conditions. More precisely,

r(0), ṙ(0) = (a, b) θ(Ta,b ) − θ(0) = 2π ⇒ θ(t + Ta,b ) − θ t) = 2π

r(0), ṙ(0) = (α, β) θ(Tα,β ) − θ(0) = 2π ⇒ θ(t + Tα,β ) − θ t) = 2π
However, if (a, b) 6= (α, β), then Ta,b may or may not equal Tα,β because they correspond to different initial
conditions.
Now consider the geometric argument. We just proved that ωr and consequently ωθ are constant for any
ωr
amplitude of radial oscillation. Then the winding number is also constant. The winding number is
ωθ
either a rational or irrational number that corresponds to periodic and quasiperiodic orbits on a torus with
dimensions corresponding to the r(t) and θ(t) periods respectively. There is no chaos.
e)
Two masses are connected by a string of fixed length. The first mass plays the role of the particle; it moves
on a frictionless, horizontal “air table.” It is connected to the second mass by a string that passes through a
hole in the center of the table. This second mass hangs below the table, bobbing up and down and supplying
the constant force of its weight. This mechanical system obeys the equations given in the text, after some
rescaling.
8.6.9
θ̇1 = ω + H(θ2 − θ1 )
θ̇2 = ω + H(θ1 − θ2 )
θ̇1 = ω + H(θ2 − θ1 ) + H(θ3 − θ1 )
θ̇2 = ω + H(θ1 − θ2 ) + H(θ3 − θ2 )
θ̇3 = ω + H(θ1 − θ3 ) + H(θ2 − θ3 )
a)
φ̇ = θ̇1 − θ̇2 = H(θ2 − θ1 ) − H(θ1 − θ2 ) = H(−φ) − H(φ)
φ̇ = θ̇1 − θ̇2 = H(θ2 − θ1 ) + H(θ3 − θ1 ) − H(θ1 − θ2 ) − H(θ3 − θ2 )
= H(−φ) + H(−φ − ψ) − H(φ) − H(−ψ)
ψ̇ = θ̇2 − θ̇3 = H(θ1 − θ2 ) + H(θ3 − θ2 ) − H(θ1 − θ3 ) − H(θ2 − θ3 )
= H(φ) + H(−ψ) − H(φ + ψ) − H(ψ)
b)
H(x) = a sin(x)
For the two-frog system,
φ̇ = H(−φ) − H(φ) = a sin(−φ) − a sin(φ) = −2a sin(φ) = 0 ⇒ φ = 0, π

dφ̇ dφ̇ dφ̇
= −2a cos(φ) = −2a = 2a
dφ dφ dφ

φ=0 φ=π
If a < 0 then the φ = π antiphase synchronization state will be stable.

For the three-frog system,
φ̇ = H(−φ) + H(−φ − ψ) − H(φ) − H(−ψ)
= a sin(−φ) + a sin(−φ − ψ) − a sin(φ) − a sin(−ψ)
= −2a sin(φ) + a sin(ψ) − a sin(φ + ψ)
ψ̇ = H(φ) + H(−ψ) − H(φ + ψ) − H(ψ)
= a sin(φ) + a sin(−ψ) − a sin(φ + ψ) − a sin(ψ)
= −2a sin(ψ) + a sin(φ) − a sin(φ + ψ)
φ̇(0, π) = ψ̇(0, π) = 0
2π 2π 2π 2π
Å ã Å ã
φ̇ , = ψ̇ , =0
3 3 3 3
The two experimentally stable states are fixed points according to the model.
Next we linearize Ñ é
−2a cos(φ) − a cos(φ + ψ) a cos(ψ) − a cos(φ + ψ)
A=
a cos(φ) − a cos(φ + ψ) −2a cos(ψ) − a cos(φ + ψ)
Ñ é
−a 0 ∆ = −3a2 τ = 2a
A(0,π) =
2a 3a ∆ < 0 ⇒ saddle point
Ñ é
3a 9a2
2 0 ∆= 4 τ = 3a
A( 2π , 2π ) =
3 3
0 3a
2 a<0⇒τ <0 stable
So the experimental fixed points exist, but the pair of frogs in unison and the third out of phase are unstable.
c)
H(x) = a sin(x) + b sin(2x) a = −5 b = 1
φ
2π
t
0
φ
d)
H(x) = a sin(x) + b sin(2x) + c cos(x) a = −5 b = 1 c = 0.1
2π
t
0
So the fixed points and their stability remained even after adding a small even component to H.
8.7 Poincaré Maps

8.7.1
r1
dr
Z
= 2π
r0 r(1 − r2 )
1 1 A B C
= = + +
r(1 − r2 ) r(1 + r)(1 − r) r 1+r 1−r
1 = A(1 + r)(1 − r) + Br(1 − r) + Cr(1 + r)
= A(1 − r2 ) + B(−r2 + r) + C(r2 + r)
= (−A − B + C)r2 + (B + C)r + A

−1 1
⇒A=1 B= C=
2 2
r1 r1 Z r1 Å
dr 1 1 1 1 1 1
Z Z Å ã ã
= − + dr = − − dr
r0 r(1 − r2 ) r r 2(1 + r) 2(1 − r) r0 r 2(r + 1) 2(r − 1)
ï 0 òr1
1 1
= ln |r| − ln |r + 1| − ln |r − 1|
2 2 r0
1 1 1 1
Å ã Å ã
= ln |r1 | − ln |r1 + 1| − ln |r1 − 1| − ln |r0 | − ln |r0 + 1| − ln |r0 − 1|
2 2 2 2
p
r r r r2 − 1 1 r2 (r2 − 1)
1 0 1
= ln p 2 − ln p 2 = ln p 02 = ln 1 0 = 2π

r1 − 1 r0 − 1 r0 r1 − 1 2 r02 (r12 − 1)
We can drop the absolute value because r = 1 is a limit cycle, so (r1 − 1) and (r0 − 1) have to have the same
sign.
r12 (r02 − 1)
= e4π
r02 (r12 − 1)
r12 (r02 − 1) = r02 (r12 − 1)e4π
r12 (r02 − 1) − r02 (r12 − 1)e4π = 0
r12 (r02 − 1 − r02 e4π ) + r02 e4π = 0

r02 e4π 1 1
r12 = = =
−r02 + 1 + r02 e4π −e−4π + r0−2 e−4π + 1 1 + e−4π (r0−2 − 1)
1
r1 = »
1 + e−4π (r0−2 − 1)
1
rn+1 = P (rn ) = »
1+ e−4π (rn−2 − 1)
−4π −3
d e r
P (r) = p
dr 1+e −4π (r−2 − 1)
d
P (1) = e−4π
dr
8.7.3

+A 0 < t <
 T
2
ẋ + x = F (t) F (t) =
−A T
<t<T

2
a)
ẋ + x = F (t)
d t
et ẋ + et x = e x = et F (t)
dt
Z T Z T
d t
e x dt = et F (t) dt
0 dt 0
h iT Z T2 Z T
t t
ex = e (+A) dt + et (−A) dt
0 0 T
2
T T
eT x(T ) − x(0) = A e 2 − 1 − A eT − e 2

T T
x(T ) = e−T x(0) + A e− 2 − e−T − A 1 − e− 2

T 2
= e−T x0 − A 1 − e− 2
b)
If we assume there is a T -periodic solution and solve
T 2
x(T ) = x(0) = x0 = e−T x0 − A 1 − e− 2
T 2
x0 − e−T x0 = x0 1 − e−T = −A 1 − e− 2

T 2 T 2
1 − e− 2 1 − e− 2
x0 = −A = −A
1 − e−T T
1 + e− 2 1 − e− 2
T
T
1 − e− 2 T
Å ã
= −A T = −A tanh
1+e 2− 4
then we arrive at the result.
c)
T 2 2
lim x(T ) = lim e−T x0 − A 1 − e− 2 = x0 − A 1 − 1 = x0
T →0 T →0
T 2 2
lim x(T ) = lim e−T x0 − A 1 − e− 2 = (0)x0 − A 1 − 0 = A
T →∞ T →∞
These results are quite plausible. As the period goes to zero, F (t) has almost no time to do anything and
x(t) has no time to move anywhere. If x(0) = x0 then x(T ) ≈ x0 and T → ∞. Then x(0) = x(T ) = x0 and
the solution doesn’t go anywhere.
As the period becomes infinite, F (t) is essentially not periodic. The equation and solution become
ẋ + x = F (t) = A x(t) = (x0 − A)e−t + A

d)
P (x)
T =2 A=3
The fixed point is stable, and not just for these choices of parameters. P (x) is a line.
T 2
P (x) = e−T x − A 1 − e− 2 = mx + b
and the slope m is always between 0 and 1 since T > 0 and m = e−T, guaranteeing the existence of a unique
globally stable fixed point.
8.7.5
θ̇ + sin(θ) = sin(t)
ṫ = 1 θ̇ = sin(t) − sin(θ)
The trajectories here can be wrapped between −π and π, but the modulo operation is removed to more
easily follow the paths of the trajectories.
θ
2π
θ̂1
3π 3π

2 , 2
π
θ1
π π

θ2 2, 2
t
0
θ̂2
θ̂3
3π π

2 ,−2
−π
θ3
−2π
The above trajectories were computed numerically, but the important point analytically is when these
trajectories are increasing and decreasing. From the intervals of increase and decrease, we can determine
where the trajectories start and end relative to the nullclines.
The Poincaré map P (θ), while we don’t know it explicitly, maps the interval [θ2 , θ3 ] into [θ̂2 , θ̂3 ] continuously.
Therefore there has to be at least one θ∗ ∈ [θ2 , θ3 ] such that P (θ∗ ) = θ∗ , which represents a periodic orbit.
The same argument can be applied to [θ1 , θ2 ], meaning there has to be at least one θ∗ ∈ [θ1 , θ2 ] such that
P (θ∗ ) = θ∗ , which represents a periodic orbit.
As for stability, there can’t be just a single stable limit cycle in the interval [θ1 , θ2 ] because the interval
expanded after one mapping. However, there could be unstable limit cycles surrounding a stable limit cycle,
or maybe even an interval of unstable periodic solutions since we haven’t explicitly ruled out that possibility.
We do know for sure that there is at least one unstable limit cycle in the interval.
The [θ2 , θ3 ] interval on the other hand contracts after one mapping. Therefore we can’t have just a single
unstable periodic orbit in [θ2 , θ3 ]. There has to be at least one stable periodic orbit in order to account for
the contracting interval.
8.7.7
θ̇ + sin(θ) = sin(t)
The trajectories here can be wrapped between −π and π, but the modulo operation is removed to more
easily follow the paths of the trajectories.
θ
2π
t
0
−π
−2π
And the graph clearly indicates the existence of one stable and one unstable periodic solution.
8.7.9
ṙ = r − r2 θ̇ = 1
a)
We can solve for r(t) by separating variables and then integrating using partial fraction decomposition and
θ(t) by direct integration.
e2π r0
r(t) = θ(t) = t + θ0
(e2π − 1)r0 + 1
The positive x-axis as the surface of section corresponds to θ = 0 and r positive.
e2π rn
rn+1 = P (rn ) =
(e2π − 1)rn + 1
b)
Fixed point
e2π r
P (r) = =r⇒r=1
(e2π − 1)r + 1
Stability
e2π
P 0 (r) = 2 P 0 (1) = e−2π < 1 ⇒ stable
(e2π − 1)r + 1
c)
We need to compute the linearized Poincaré map at the fixed point, but we already did that in part (b).
Therefore e−2π is the characteristic multiplier for the periodic orbit.
8.7.11
From Example 8.7.4, we know that the linearization of the in-phase fixed point of Equation 8.7.1
N
1 X
φ̇i = Ω + a sin(φi ) + sin(φj )
N j=1
predicts a center.
Going back to Section 6.6, a higher-order system ẋ = f (x) is reversible if it is invariant under the transfor-
mation
t → −t x → R(x) where R2 (x) = x
The simplest choice, R(x) = −x, doesn’t work, but R(x) = π − x does.
R2 (x) = R(π − x) = π − (π − x) = x
dφi d(π − φi ) −dφi dφi
−→ = =
dt −dt −dt dt
sin(π − φi ) = sin(φi )
N N
1 X 1 X
Ω + a sin(π − φi ) + sin(π − φj ) = Ω + a sin(φi ) + sin(φj )
N j=1 N j=1
Then by Theorem 6.6.1, the linear center is a nonlinear center; i.e., the in-phase state is not attracting.
9
Lorenz Equations
9.1 A Chaotic Waterwheel

9.1.1
a)
The moment of inertia of a ring is the mass times radius squared.
Iwater = M r2
b)
The rate of change of water is the rate at which water flows in minus the rate at which water flows out.
Ṁ = Qtotal − KM
c)
First, we make a differential equation for Iwater
I˙water = Ṁ r2 = Qtotal r2 − KM r2 = Qtotal r2 − KI
which has solution
Qr2 −Kt Qr2 Qr2

Å ã
Iwater (t) = Iwater (0) − e + lim Iwater (t) =
K K t→∞ K
9.1.3
ν πgrw
ȧ1 = ωb1 − Ka1 ḃ1 = −ωa1 + q1 − Kb1 ω̇ = − ω + a1
I I
ẋ = σ(y − x) ẏ = rL x − xz − y ż = xy − bz
We distinguish between the r in the waterwheel equations and Lorenz equations with a w and L subscript
respectively.
ω = αx a1 = βy b1 = γz + φ t = ξτ
ν πgrw 1 dx ν πgrw
ω̇ = − ω + a1 → α = − αx + βy
I I ξ dτ I I
1 dy
ȧ1 = ωb1 − Ka1 → β = αx(γz + φ) − Kβy
ξ dτ
1 dz
ḃ1 = −ωa1 + q1 − Kb1 → γ = −αxβy + q1 − K(γz + φ)
ξ dτ
273
274 Chapter 9: Lorenz Equations
dx νξ πgrw βξ
=− x+ y = σ(y − x)
dτ I αI
dy αξγ αφξ
= xz + x − ξKy = rL x − xz − y
dτ β β
dz αβξ ξ
=− xy + (q1 − Kφ) − ξKz = xy − bz
dτ γ γ
νξ πgrw βξ
σ= =
I αI
αφξ αξγ
rL = −1= 1 = ξK
β β
αβξ ξ
1=− 0 = (q1 − Kφ) b = ξK
γ γ
1 q1
⇒ξ= φ=
K K
ν πgrw β
σ= =
IK αIK
αq1 αγ
rL = −1=
βK 2 βK
αβ
1=− b=1
γK
πgrw β α πgrw
⇒α= =
ν β ν
πgrw q1 πgrw γ
rL = 2
−1=
νK νK
πgrw β 2
1=−
γνK
νK 1 πgrw
⇒γ=− =−
πgrw γ νK
πgrw β 2
Å ã
1=
νK
νK
⇒β=± α = ±K Choose +
πgrw
νK
β= α=K
πgrw
πgrw K νK νK q1 1
ω= x a1 = y b1 = − z+ t= τ
ν πgrw πgrw K K
ν πgrw q1
σ= rL = b=1
IK νK 2
9.1 A Chaotic Waterwheel 275
9.1.5
We start off the problem with the same equations, but now Q(θ) has pn sin(nθ) terms in the Fourier series.
Previously we used only the qn cos(nθ) because we assumed Q(θ) was an even function.
∂m ∂m
= Q − Km − ω
∂t ∂θ
∞
X
m(θ, t) = an (t) sin(nθ) + bn (t) cos(nθ)
n=0
∞
X
Q(θ) = pn sin(nθ) + qn cos(nθ)
n=0
If we go back through the derivation

" ∞ # ∞
∂ X X
an (t) sin(nθ) + bn (t) cos(nθ) = pn sin(nθ) + qn cos(nθ)
∂t n=0 n=0
∞ ∞
X ∂ X
−K an (t) sin(nθ) + bn (t) cos(nθ) − ω an (t) sin(nθ) + bn (t) cos(nθ)
n=0
∂θ n=0
∞
X ∞
X
ȧn (t) sin(nθ) + ḃn (t) cos(nθ) = pn sin(nθ) + qn cos(nθ)
n=0 n=0
∞
X ∞
X

−K an (t) sin(nθ) + bn (t) cos(nθ) − ω n an (t) cos(nθ) − bn (t) sin(nθ)
n=0 n=0
The ω̇ equation stays the same because Q(θ) wasn’t involved in the derivation.
−νω + πgra1
ω̇ =
I
Equating terms gives
ȧn = nωbn − Kan + pn
ḃn = −nωan − Kbn + qn
And all equations for n = 1 are
ȧ1 = ωb1 − Ka1 + p1
ḃ1 = −ωa1 − Kb1 + q1

−νω + πgra1
ω̇ =
I
We have a new term p1 in the ȧ1 equation.
Now solving for the fixed points,
0 = ωb1 − Ka1 + p1
0 = −ωa1 − Kb1 + q1
−νω + πgra1
0= ⇒ 0 = −νω + πgra1
I
Next we eliminate b1 from the first two equations, solve for a1 , then substitute into the third equation.
0 = Kωb1 − K 2 a1 + Kp1
0 = −ω 2 a1 − Kωb1 + ωq1
0 = (Kωb1 − K 2 a1 + Kp1 ) + (−ω 2 a1 − Kωb1 + ωq1 )
= −K 2 a1 + Kp1 − ω 2 a1 + ωq1 )
Kp1 + ωq1
a1 =
K 2 + ω2
Kp1 + ωq1
0 = −νω + πgra1 = −νω + πgr
K 2 + ω2
= −νω 3 − K 2 νω + πgrKp1 + πgrq1 ω
0 = νω 3 + (νK 2 − πgrq1 )ω − πgrKp1
= ω(νω 2 + νK 2 − πgrq1 ) − πgrKp1
Looking at this, we can see the same solutions as before.

…
πgrq1
p1 = 0 ⇒ 0 = ω(νω 2 + νK 2 − πgrq1 ) ⇒ ω = 0, ± − K2
ν
This creates a pitchfork bifurcation. However, for p1 6= 0 along with proper values for the other coefficients,
the pitchfork bifurcation is ruined and replaced with an imperfect bifurcation similar to that in Figure 3.6.3
in the text.
9.2 Simple Properties of the Lorenz Equations

9.2.1
a)
ẋ = σ(y − x)
ẏ = rx − xz − y
ż = xy − bz
» »
C± = (± b(r − 1), ± b(r − 1), r − 1)
á ë á ë
−σ σ 0 −σ σ 0
p
r−z −1 −x = 1 −1 ∓ b(r − 1)
p p
y x −b ± b(r − 1) ± b(r − 1) −b
C±
á ë
−σ − λ σ 0
p
det 1 −1 − λ ∓ b(r − 1)
p p
± b(r − 1) ± b(r − 1) −b − λ
9.2 Simple Properties of the Lorenz Equations 277
= (−σ − λ)((−1 − λ)(−b − λ) + b(r − 1)) − σ((−b − λ) + b(r − 1))
= −(σ + λ)((1 + λ)(b + λ) + b(r − 1)) + σ(b + λ − b(r − 1))
= −(σ + λ)(b + (1 + b)λ + λ2 + br − b) + σ(b − br + b) + σλ
= −(σ + λ)(λ2 + (1 + b)λ + br) + σ(2b − br) + σλ
= −σλ2 − σ(1 + b)λ − σbr − λ3 − (1 + b)λ2 − brλ + 2σb − bσr + σλ
= −λ3 − (σ + 1 + b)λ2 − b(r + σ)λ − 2bσ(r − 1) = 0
= λ3 + (σ + 1 + b)λ2 + b(r + σ)λ + 2bσ(r − 1) = 0
b)
Assuming a root of the form λ = iw
λ3 + (σ + 1 + b)λ2 + b(r + σ)λ + 2bσ(r − 1) = 0
(iw)3 + (σ + 1 + b)(iw)2 + b(r + σ)(iw) + 2bσ(r − 1) = 0
−iw3 − (σ + 1 + b)w2 + ib(r + σ)w + 2bσ(r − 1) = 0
Both the real and imaginary parts have to separately equal 0.

2 2bσ(r − 1)
−(σ + 1 + b)w + 2bσ(r − 1) = 0 ⇒ w = ±
(σ + 1 + b)
»
−iw3 + ib(r + σ)w = −iw(w2 − b(r + σ)) = 0 ⇒ w = 0, ± b(r + σ)
Equating the two roots for w and solving for r

2bσ(r − 1) »
± = ± b(r + σ)
(σ + 1 + b)
2bσ(r − 1)
= b(r + σ)
(σ + 1 + b)
2bσ(r − 1) = b(r + σ)(σ + 1 + b)
2bσr − 2bσ = br(σ + 1 + b) + bσ(σ + 1 + b)
2bσr − br(σ + 1 + b) = bσ(σ + 1 + b) + 2bσ
r(2bσ − b(σ + 1 + b)) = bσ(σ + 1 + b) + 2bσ
rb(σ − 1 − b) = bσ(σ + 3 + b)
(σ + 3 + b)
r=σ
(σ − 1 − b)
We need b + 1 < σ in order for the derivation to work. We solved for
(σ + 3 + b) σ
1 < r < rH = σ = (σ + 3 + b)
(σ − 1 − b) (σ − (1 + b))
σ and b are positive, so 1 < (σ + 3 + b), and as long as 0 < σ − (1 + b) < σ ⇒ 1 + b < σ, we ensure that
1 < r.
c)
»
w= b(r + σ)
(λ + iw)(λ − iw)(λ − λ3 )
(λ2 + w2 )(λ − λ3 )
(λ2 + b(r + σ))(λ − λ3 )
λ3 − λ3 λ2 + . . .
λ3 + (σ + 1 + b)λ2 + b(r + σ)λ + 2bσ(r − 1) = λ3 − λ3 λ2 + . . .
λ3 = −(σ + 1 + b)
9.2.3
ẋ = σ(y − x)
ẏ = rx − xz − y
ż = xy − bz
V = x2 + y 2 + (z − r − σ)2
V̇ = 2xẋ + 2y ẏ + 2(z − r − σ)ż
= 2xσ(y − x) + 2y(rx − xz − y) + 2(z − r − σ)(xy − bz)

1
V̇ = σxy − σx2 + rxy − xyz − y 2 + xyz − bz 2 − (r + σ)xy + b(r + σ)z
2
= −σx2 − y 2 − bz 2 + b(r + σ)z
= −σx2 − y 2 − b(z 2 − (r + σ)z)

(r + σ)2 (r + σ)2
Å ã
2 2 2
= −σx − y − b z − (r + σ)z + −
4 4
2 2
r+σ (r + σ)
Å ã
= −σx2 − y 2 − b z − −
2 4
r+σ 2 (r + σ)2
= −σx2 − y 2 − b z − +b
2 4
We need V̇ < 0 for the trajectories to point inward on the surface.
1 r + σ 2 (r + σ)2
V̇ = −σx2 − y 2 − b z − +b <0
2 2 4
(r + σ)2 r + σ 2
b < σx2 + y 2 + b z −
4 2
4σ 2 4 2 4 r + σ 2
1< x + y + z −
b(r + σ)2 b(r + σ)2 (r + σ)2 2
r + σ
This last inequality describes all of R3 minus an ellipsoid with center 0, 0, , with the maximum
2
b(r + σ)2 b(r + σ)2 (r + σ)2
distance from the center in x, y, and z as , , and respectively.
4σ 4 4
Thus we can enclose this ellipsoid with a sphere centered at (0, 0, r + σ) by picking a sufficiently large
radius. However, the radius is a bit annoying to calculate because the center of the ellipsoid is not the center
of the sphere.
9.2.5
For b = 1 we have the following stability diagram for the Lorenz equations.
r
30
Origin saddle point
Hopf bifurcation
C + and C − saddle points σ(σ + b + 3) σ(σ + 4)
r= =
σ−b−1 σ−2
Origin saddle point
C + and C − stable nodes

encircled by saddle cycles
Pitchfork bifurcation at r = 1
σ
Origin stable node
30
9.3 Chaos on a Strange Attractor

9.3.1
a)
The system is periodic, but it is not chaotic because it’s not sensitive to initial conditions, which is one of
the requirements for chaos.
b)
In fact, the error never changes.
θ̇1 = w1 ⇒ θ1 (t) = w1 t + C1 , θ1 (0) = C1
θ̇2 = w2 ⇒ θ2 (t) = w2 t + C2 , θ2 (0) = C2
Adding error to the initial conditions makes a blob around the initial conditions, but the blob moves together
and doesn’t distort.
The system can be interpreted as movement of a dot on a sphere. w1 controls how fast the dot moves
in the latitudinal direction and w2 how fast the dot moves in the longitudinal direction.
~ + ~δ(t)) − θ(t)||
||(θ(t) ~ = ||~δ(t)|| = ||(δ1 , δ2 )|| = ||(δ1 , δ2 )||e0t
So the largest Liapunov exponent is 0.
9.3.3
With the parameters r = 22, σ = 10, and b = 38 , some of the trajectories can be chaotic for a little while but
eventually stay in the middle of one of the pieces. (Trajectories can also go directly to the middle of a piece
with no chaos, but those are not shown.) The two simulations were run until t = 100.
Initial condition (x, y, z) = (−58.26, −3.3144, 12.221)
x(t)
y(t)
t
This other trajectory has an initial condition (x, y, z) = (−20, 0, −20), but the simulation had to be run for
much, much longer. We show only the final graph due to the simulation length.
x
9.3.5
With the parameters r = 100, σ = 10, and b = 83 , the trajectories go to one of two stable limit cycles.
Initial condition (x, y, z) = (10, 10, 10)
x(t)
t
y(t)
t
x
The other limit cycle is symmetric and starts at initial condition (x, y, z) = (−10, −10, 10).
x
9.3.7
With the parameters r = 400, σ = 10, and b = 83 , trajectories go towards a single stable limit cycle.
Initial condition (x, y, z) = (40, −40, 200)
x(t)
y(t)
t
x
9.3.9
The following graph was made by running an arbitrary initial condition until the trajectory was on the
Lorenz attractor at (x, y, z) = (0.8961, 1.5605, 11.246). Then the trajectory and a perturbed trajectory of
(x, y, z) = (0.8961, 1.5605, 11.246+0.001) were simulated until t = 25 and the error computed in the following
graph. Then we used a linear fit of the error between t ∈ [1, 10] so as to avoid the initial transient and leveling
off of the error. The slope is an estimate of the largest Liapunov exponent, λ = 0.8623 ≈ 0.906. The latter
is the “real” value calculated with more advanced methods.
ln ||δ(t)||
t
9.4 Lorenz Map

9.4.1
zn+1
50
25
zn
25 50
9.5 Exploring Parameter Space

9.5.1
With the parameters r = 166.3, σ = 10, and b = 83 , the simulation was run until t = 30, showing intermittent
chaos.
Initial condition (x, y, z) = (10, −10, 0)

x(t)
y(t)
This graph is for t ∈ [20, 30], and you can see the mostly periodic behavior of the grouped lines and the
trajectory splitting off at some point.
z
x
9.5.3
8
Parameters 145 < r < 166 σ = 10 b = 3
This range of r is a period doubling region for the Lorenz system.

All these inequaliti es are approximate.
166.07 < r
No limit cycle, intermittent chaos
r = 166.07
Single stable symmetric limit cycle born
r = 154.4
Single stable symmetric limit cycle splits into two asymmetric limit cycles
r = 148.2
Each of two stable asymmetric limit cycles split, creating a total of four asymmetric limit cycles with
double the period
145 < r < 148.2
An infinite number of period double bifurcations occurs in this region
r = 145
Chaos
A sampling of r values
167.1 166 160 155 154 148.5 148 147.5 145
in this order were simulated to show changes in behavior.
Unless otherwise noted, each simulation was run until t = 10 after reaching the limit cycle.
x(t) r = 167.1
t
y(t)
x(t) r = 166
y(t)
t
x(t) r = 160
y(t)
x(t) r = 155
t
y(t)
x(t) r = 154
y(t)
t
x(t) r = 148.5
y(t)
x(t) r = 148
t
y(t)
x(t) r = 147.5
y(t)
t
x(t) r = 145
y(t)
z z z z
x x x x
r = 167.1 r = 166 r = 160 r = 155

z z z z
x x x x
r = 154 r = 148.5 r = 148 r = 147.5
Initial condition (x, y, z) = (10, −10, 0) t ∈ [0, 50]

z
r = 145
9.5.5
ẋ = σ(y − x) ẏ = rx − xz − y ż = xy − bz
a)
t 1
X = x Y = 2 σy Z = σ(2 z − 1) τ = = √
r
1 dX Y X
Å ã
ẋ = 2 = σ(y − x) = σ 2 −
dτ σ
1 dY X X 1 Z Y
Å ã
ẏ = 3 = rx − xz − y = r − + 1 − 2
σ dτ 2 σ σ
1 1 dZ X Y 1 Z
Å ã Å ã
ż = 2 +1 = xy − bz = − b + 1
σ dτ 2 σ 2 σ
1 dX Y X
Å ã
=σ 2 −
2 dτ σ
dX
⇒ = Y − σX
dτ
1 dY X X 1 Z Y
Å ã
=r − +1 − 2
3 σ dτ 2 σ σ
dY
⇒ = r2 σX − XZ − σX − Y
dτ
= σX − XZ − σX − Y
= −XZ − Y
1 1 dZ X Y 1 Z
Å ã Å ã
+1 = −b 2 +1
2 σ dτ 2 σ σ
dZ Z
Å ã
⇒ (1 + σ) = XY − bσ +1
dτ σ
dX
lim =Y
→0 dτ
dY
lim = −XZ
→0 dτ
dZ dZ
lim (1 + σ) = = XY
→0 dτ dτ
b)
(Y 2 + Z 2 )0 = 2Y Y 0 + 2ZZ 0 = 2Y (−XZ) + 2Z(XY )
= −2XY Z + 2XY Z = 0
(X 2 − 2Z)0 = 2XX 0 − 2Z 0 = 2XY − 2XY = 0
c)
Z Z Z
V̇ = ∇ · f dV = ∇ · hX 0 , Y 0 , Z 0 i dV = ∇ · hY, −XZ, XY i dV
V V
Z
= 0 dV = 0
V
So the set of equations has a rate of change of zero on a volume, otherwise known as volume preserving.
d)
As discussed in the text, the Rayleigh number r is a dimensionless ratio. Its numerator includes the effects
driving the system. Its denominator includes the two sources of energy dissipation (for the waterwheel, these
were the leakage of water through the holes and the viscous damping due to the brake). The limit of infinite
r means that the dissipation terms are tending to zero, in a relative sense. And without dissipation, the
system would be expected to act conservatively.
e)
Initial condition (x, y, z) = (10, −10, 0) t ∈ [0, 10]
x(t)
y(t)
z(t)
The behavior here is a stable limit cycle. Numerical solutions of the Lorenz equations for large r also show
a single symmetric limit cycle, so the Lorenz equations and this approximate system are in agreement.
9.6 Using Chaos to Send Secret Messages

9.6.1
a)
From the text
1 2
Å ã
V̇ = −e22 − 4be23 < −k 2
e + 2e3 = −kV
2 2
We can solve the inequality for each component separately.

1
− e22 < − ke22 ⇒ k < 2
2
− 4be23 < −2ke23 ⇒ k < 2b
Therefore, as long as k < min{2, 2b}, the inequality will hold. Integration then yields
0 ≤ V (t) ≤ V0 e−kt
b)
1 2 p kt
e2 ≤ V < V0 e−kt ⇒ e2 (t) < 2V0 e− 2
2
Similarly, …
−kt V0 − kt
2e23 ≤ V < V0 e ⇒ e3 (t) < e 2
2
c)
First take the ė1 equation, rearrange it, bound it,
p kt
ė1 = σ(e2 − e1 ) ⇒ ė1 + σe1 = σe2 < σ 2V0 e− 2
and then use an integrating factor.

√
2σ 2V0 − kt
e1 (t) = Ce−σt − e 2
k − 2σ
So all components of e(t) decay exponentially fast.
9.6.3
a)
Initial conditions t ∈ [0, 3]
(x, y, z) = (15, 5, −10)
(xr , yr , zr ) = (x, −10, 20)
y(t)
y
yr
t
b)
z
(y, z)
(yr , zr )
9.6.5
The idea here is that the receiver and transmitter will still synchronize even though we aren’t transmitting
x(t) but s(t) = x(t) + m(t) instead. As long as m(t) has a small amplitude relative to x(t) on the strange at-
tractor, then s(t) ≈ x(t) and the receiver xr (t) should still synchronize to the transmitter’s x(t) pretty closely.
In all the following simulations, we start the transmitter and receiver variables at (x, y, z) = (xr , yr , zr ) =
(7.6200, 9.6101, 48.3368), which is already close to the Lorenz attractor since we assume the receiver and
transmitter will synchronize. Then we add a sine wave A sin(2πf t) (A = 0.1 and f is varied) to xr and
simulate for ten periods of the sine wave. Lastly, we simulate for ten more periods of the sine wave, record
E(f ) = max|m(t) − m̂(t)| for this time interval, and graph the results from 1 Hz to 20,000 Hz.
8
Parameters r = 60 σ = 10 b = 3
log(E)
10 102 103 104 105 log(f )
10−1
10−2
10−3
10−4
10−5
10−6
10−7
10−8
The error is rather erratic at low frequencies, but then at about 1000 Hz the error just keeps going down
and down.
10
One-Dimensional Maps
10.1 Fixed Points and Cobwebs

10.1.1
√
xn+1 = xn
The fixed points occur when
√
x= x
x2 = x
x(x − 1) = 0
So the fixed points are x∗ = 0, 1.
As for stability,
√
f (x) = x
1
f 0 (x) = √
2 x
|f 0 (0)| = ∞ ⇒ unstable
1
f 0 (1) = ⇒ stable
2
Here is some numerical and graphical confirmation as well.
x0 0 0.5 1 2
x1 0 0.707107 1 1.41421
x2 0 0.840896 1 1.18921
x3 0 0.917004 1 1.09051
x4 0 0.957603 1 1.04427
x5 0 0.978572 1 1.0219
x6 0 0.989228 1 1.01089
x7 0 0.994599 1 1.00543
x8 0 0.997296 1 1.00271
x9 0 0.998647 1 1.00135
307
308 Chapter 10: One-Dimensional Maps
xn+1
xn
10.1.3
xn+1 = exp(xn )
x = ex
and there are none.

As for stability, there are no fixed points, but the sequence should always be increasing.
x0 -3
x1 0.0497871
x2 1.05105
x3 2.86065
x4 17.4728
xn+1
xn
10.1.5
xn+1 = cot(xn )

x = cot(x)
So the fixed points are x∗ = −1, 1.
As for stability,
f (x) = cot(x)
f 0 (x) = − csc2 (x)
f 0 (1) ≈ −1.41 ⇒ unstable
f 0 (−1) ≈ −1.41 ⇒ unstable

x0 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2

x1 0.45766 -0.07091 -0.64209 -1.83049 ? 1.83049 0.64209 0.07091 -0.45766
x2 2.03031 -14.07777 -1.33725 0.26569 -0.26569 1.33725 14.07777 -2.03031
x3 -0.49485 -0.23788 3.67479 -3.67479 0.23788 0.49485
x4 -1.85312 -4.12414 1.69430 -1.69430 4.12414 1.85312
x5 0.29007 -0.66703 -0.12414 0.12414 0.66703 -0.29007
x6 3.35023 -1.26996 -8.01428 8.01428 1.26996 -3.35023
x7 4.72318 -0.31026 0.31026 -4.72318
x8 -0.01079 -3.11906 3.11906 0.01079
x9 -92.66893 44.37344 -44.37344 92.66893
xn+1
xn
10.1.7
xn+1 = sinh(xn )

x = sinh(x)
So the fixed point is x∗ = 0.
As for stability,
f (x) = sinh(x)
f 0 (x) = cosh(x)
f 0 (0) = 1 ⇒ inconclusive, but the graph shows it’s unstable
x0 -0.50000 0.50000
x1 -0.52110 0.52110
x2 -0.54500 0.54500
x3 -0.57238 0.57238
x4 -0.60415 0.60415
x5 -0.64158 0.64158
x6 -0.68651 0.68651
x7 -0.74173 0.74173
x8 -0.81163 0.81163
x9 -0.90372 0.90372
x10 -1.03186 1.03186
x11 -1.22497 1.22497
x12 -1.55515 1.55515
x13 -2.26231 2.26231
x14 -4.75059 4.75059
x15 -57.82220 57.82220
xn+1
xn
10.1.9
2xn
xn+1 =
1 + xn
2x
x=
1+x
x(1 + x) = 2x
x(x − 1) = 0
So the fixed points are x∗ = 0, 1.
As for stability,
2x
f (x) =
1+x
0 2(1 − x) − 2x(−1) 2
f (x) = =
(1 + x)2 (1 + x)2
f 0 (0) = 2 ⇒ unstable
1
f 0 (1) = ⇒ stable
2
x0 -4 -0.01 0.01 1.5 4

x1 2.666667 -0.020202 0.019802 1.200000 1.600000
x2 1.454545 -0.041237 0.038835 1.090909 1.230769
x3 1.185185 -0.086022 0.074766 1.043478 1.103448
x4 1.084746 -0.188235 0.139130 1.021277 1.049180
x5 1.040650 -0.463768 0.244275 1.010526 1.024000
x6 1.019920 -1.729730 0.392638 1.005236 1.011858
x7 1.009862 4.740741 0.563877 1.002611 1.005894
x8 1.004907 1.651613 0.721127 1.001304 1.002938
x9 1.002447 1.245742 0.837971 1.000651 1.001467
xn+1
xn
xn+1
xn
10.1.11
xn+1 = 3xn − x3n
a)
√ √
x = 3x − x3 ⇒ 0 = 2x − x3 = x(x + 2)(x − 2)
√
So the fixed points are x = 0, ± 2.
b)
xn+1
xn
c)
xn+1
xn
d)
There exists a 2-cycle xn = ±2 ⇒ xn+1 = 3(±2) − (±2)3 = ∓2 and 2 → −2 → 2 → · · · , which is the reason
for the change in behavior from x0 = 1.9 to x0 = 2.1.
xn+1
xn
If −2 < xn < 2 then −2 < xn+1 = 3xn − x3n < 2 simply because the graph fits in the box made by the
2-cycle and thus every point maps to a point in the box.
As for starting outside the 2-cycle box,
|xn | = 2 + >0
|x1 | = |3x0 − x30 | = |x0 ||3 − x20 | = |x0 ||3 − (2 + )2 |
= |x0 ||3 − (4 + 2 + 2 )| = |x0 || − 1 − 2 − 2 )|
= |x0 |(1 + 2 + 2 )
|xn | = |x0 |(1 + 2 + 2 )n ⇒ lim |xn | = ∞

n→∞
So the iteration map will always be increasing in magnitude and spiraling away from the 2-cycle.
10.1.13
g(xn )
xn+1 = f (xn ) = xn −
g 0 (xn )
g 0 (x) g(x)g 00 (x)
f 0 (x) = 1 − 0 +
g (x) (g 0 (x))2
If g 0 (x∗ ) 6= 0, then along with g(x∗ ) = 0,
g 00 (x)
f 0 (x) = 1 − 1 + 0 =0
(g 0 (x))2
is implied. The multiplier is zero and x∗ is a superstable fixed point.
10.2 Logistic Map: Numerics

10.2.1
a)
1 < xn
xn+1 = rxn (1 − xn ) < 0
xn+2 = rxn+1 (1 − xn+1 ) < 0

..
.
x∞ < 0
But that’s not enough to show that the sequence diverges since it might have some convergent negative
value.
0<
xn = 1 +
xn+1 = r(1 + )(1 − (1 + )) = −r(1 + )
xn+1 − xn = −r(1 + ) − (1 + ) = −(r + )(1 + ) < −r
So the sequence decreases by at least r every iteration. Since r 6= 0,
lim xn → −∞
n→∞
b)
It’s a good idea to have x0 ∈ [0, 1], otherwise x1 < 0 and the sequence would diverge.
r ∈ [0, 4] is necessary to ensure that the sequence is trapped within [0,1] if the sequence starts in that
interval. r < 0 would have the next iterate map to a negative value; and 4 < r would have the next iterate
able to map outside [0,1] around the maximum of the parabola, but
x0 ∈ [0, 1] and r ∈ [0, 4] ⇒ rxn (1 − xn ) ∈ [0, 1]
f (xn ) f (xn )
1
r<0 4<r
xn xn
10.2.3, 10.2.5, and 10.2.7

Good black-and-white orbit diagrams of Exercises 10.2.3, 10.2.5, and 10.2.7 can be made with the following
pseudocode.
begin function orbit_diagrammer()

#create the bounds of the orbit diagram
r_min = ???
r_max = ???
x_min = ???
x_max = ???
#what is the r sample resolution
r_step = ???
#how many sample points per r value
x_samples = ???
#loops through all sampled r values

for r from r_min to r_max by increment r_step
#initial x value to start iterations
x = random number between x_min and x_max
#loop to pass through transient behavior until the orbits are reached
for n = 1 to 1000
#the iteration map
x = f(r,x)
end for
#loop to calculate orbits
for n = 1 to x_samples
x = f(r,x)
draw (r,x) coordinate
end for
end for
end function orbit_diagrammer()
However, the following orbit diagrams use gray scale to show how often each pixel region was visited in the
simulations.
10.2.3
xn+1 = rxn (1 − xn )
1.5
−0.5
−2 r 4
0
2.95 3 r 4
1.5
−0.5 −0.95
−2 r −1
10.2.5
xn+1 = e−rxn
e
1
e
0
1 e r 8
−
e
Look up the Lambert W function for those who are intrigued by all the e’s.
10.2.7
xn+1 = r tan(xn )
10
−10
−2 r 2
10.3 Logistic Map: Analysis

10.3.1
x = f (x) = rx(1 − x) = rx − rx2

r−1
rx2 + (1 − r)x = x(rx + 1 − r) = 0 ⇒ x∗ = 0,
r
f 0 (x) = r(1 − 2x)
f 0 (0) = r = 0
x∗ = 0 is a superstable fixed point when r = 0

r−1 r−1
Å ã Å ã
f0 =r 1−2 = 0 ⇒ r = 0, 2
r r
r−1
x∗ = is a superstable fixed point when r = 2
r
x∗ = 0 is superstable if r = 0 because x1 = 0 no matter what x0 is. The other fixed point is superstable
when r = 2 but not when r = 0 because x1 = 0 as before.
10.3.3
rxn
xn+1 =
1 + x2n
rx
x=
1 + x2
x(1 + x2 ) = rx
x(1 + x2 ) − rx = x(1 − r + x2 ) = 0
√
x = 0, ± 1 − r
So there are three fixed points if r ≤ 1 and one fixed point if 1 < r.
rx
f (x) =
1 + x2
r(1 + x2 ) − 2rx2 r(1 − x2 )
f 0 (x) = =
(1 + x2 )2 (1 + x2 )2
√ 2
f 0 (0) = r f 0 (± 1 − r) = − 1
r
√
r < 1 ⇒ x = 0 is stable, and x∗ = ± 1 − r are unstable
∗
r = 1 ⇒ f 0 (0) = 1 is inconclusive

xn xn
|xn+1 | = < = |xn | ⇒ x∗ = 0 is stable
1 + x2n 1
1 < r ⇒ x∗ = 0 is stable
As for the existence of periodic orbits, we already proved that the sequence |xn | is monotonically decreasing
when r ≤ 1 and xn 6= 0, which rules out any periodic orbits.
It turns out that there are no periodic orbits for 1 < r either, and never any chaos. We can prove this
by showing that the sequence |xn | is monotonically increasing, monotonically decreasing, or constant for any
positive r value.

rx
|x| <
= r|x|
1 + x 1 + x2
2
|x|(1 + x2 ) < r|x|

√
|x|(1 − r + x2 ) < 0 ⇒ x2 < r − 1 ⇒ |x| < r−1

rx r|x|
1 + x2 = 1 + x2 < |x|

r|x| < |x|(1 + x2 )

√
0 < |x|(1 − r + x2 ) ⇒ r − 1 < x2 ⇒ r − 1 < |x|
So the last bunch of equations shows that the sequence |xn | is monotonically increasing or monotonically
decreasing, and constant only if the sequence starts at a fixed point. Essentially the iterations always travel
away from the origin, always travel towards the origin, or stay at a fixed point.
10.3.5
xn+1 = rxn (1 − xn ) yn+1 = yn2 + c xn = ayn + b
xn+1 = ayn+1 + b = a(yn2 + c) + b
xn+1 = r(ayn + b)(1 − ayn − b)
a(yn2 + c) + b = r(ayn + b)(1 − ayn − b)
ayn2 + ac + b = −a2 ryn2 + ar(1 − 2b)yn + rb(1 − b)

1 1 r(2 − r)
a=− b= c=
r 2 4
10.3.7
xn+1 = 10xn (mod 1)
a)
xn+1
1
xn
1
b)
x = 0, 0.1̄, 0.2̄, 0.3̄, 0.4̄, 0.5̄, 0.6̄, 0.7̄, and 0.8̄ are the fixed points. It looks like we’re missing x = 0.9̄, but that’s
actually a fancy way to write x = 1, which will get mapped to x = 0 after the first iteration.
c)
Any x ∈ I with a p long repeating decimal expansion is a p-cycle of the map.
To find the stability of the p-cycles, we take the derivative of
d p d p
f (x) = 10 x (mod 1) = 10p > 1
dx dx
so they’re all unstable.

d)
Any orbit starting at an irrational number x0 will be aperiodic, since the decimal expansion of an irrational
number never repeats. There are an infinite number of irrational numbers in [0,1], so there are an infinite
number of aperiodic orbits.
e)
The definition of the Liapunov exponent is
( n−1 ) ( n−1 )
1 X 0 1X
λ = lim ln f (xi ) = lim
ln(10) = ln(10) > 0
n→∞ n i=0 n→∞ n i=0
The Liapunov exponent is positive, so the decimal shift map has sensitive dependence on initial conditions.
10.3.9
xn+1 = 2xn (mod 1) in decimal ≡ 10xn (mod 1) in binary
We’ll be using binary representation in this problem, as the binary shift map applies easily to binary numbers.
a)
xn+1
1
xn
1
b)
x = 0 is the only fixed point. It looks like x = 0.1̄ is also a fixed point, but that’s actually a fancy way to
write x = 1, which will get mapped to x = 0 after the first iteration.
c)
Any x ∈ I with a p long repeating decimal expansion is a p-cycle of the map.
To find the stability of the p-cycles, we take the derivative of
d p d p
f (x) = 10 x (mod 1) = 10p > 1 2p > 1 in decimal
dx dx
so they’re all unstable.

d)
Any orbit starting at an irrational number x0 will be aperiodic, since the binary expansion of an irrational
number never repeats. There are an infinite number of irrational numbers in [0,1], so there are an infinite
number of aperiodic orbits.
e)
The definition of the Liapunov exponent is
( n−1 ) ( n−1 )
1 X 0 1X
λ = lim ln f (xi ) = lim
ln(10) = ln(10) > 0
n→∞ n i=0 n→∞ n i=0
ln(2) > 0 in decimal
The Liapunov exponent is positive, so the decimal shift map has sensitive dependence on initial conditions.
This one is a bit of a trick. Start the construction as zero. Then append all the number sequences of length
one. Then append all number sequences of length two, then of length three, then four, etc.
x=0
0, 1
x = 0.01
00, 01, 10, 11
x = 0.0100011011
x = 0.0100011011000001010011100101110111
..
.
This number x contains all numbers in [0,1] with terminating binary expansions. Therefore this number will
be arbitrarily close to any particular number in [0,1] within a finite number of iterations of the binary shift
map. We need only find a terminating binary number within distance of the number of interest and apply
the binary shift map to x until the terminating binary number is at the front, and then a little bit of error
attached to the end.
Note: It is necessary to append enough zeros onto the terminating binary representation in order to achieve
the tolerance. For example, if we want to approximate 0.1101 within 0.0001 tolerance, then a suitable
number would be 0.1101 . . . and we could iterate the map on x until 0.1101 is at the front. But if we want
a tolerance of 0.0000001, a suitable number would be 0.1101000 . . . and we could iterate the map on x until
0.1101000 is at the front.
10.3.11
xn+1 = f (xn ) = −(1 + r)xn − x2n − 2x3n
a)
f 0 (x) = −(1 + r) − 2x − 6x2
f 0 (0) = −(1 + r)
− 1 < −(1 + r) < 1 ⇒ 0 < r < 2
0 < r < 2 ⇒ stable r < 0 or 2 < r ⇒ unstable
b)
A flip bifurcation will occur at x = 0 and r = 0 if the stability changes on either side of the r value.
f (0) = 0
r < 0 ⇒ f 0 (0) = −(1 + r) < −1 ⇒ unstable
r = 0 ⇒ f 0 (0) = −(1 + r) = −1 ⇒ inconclusive
0 < r ⇒ f 0 (0) = −(1 + r) > −1 ⇒ stable
Hence at x = 0 and r = 0 a flip bifurcation occurs.

c)
The following graph shows the plot at r = −0.5.
xn+1
xn
The origin is a fixed point of the original map, but the other two intersections are the 2-cycle points. The
slope of the 2-cycle intersections with the diagonal have a magnitude greater than one, meaning the 2-cycle
is unstable. Trajectories will be drawn towards the origin or repelled to infinity if the initial conditions start
between the 2-cycle points or the outside respectively. Also, increasing r towards 0 will move the 2-cycle
points closer to the origin until all three coalesce.
d)
The long-term behavior for orbits that start near x∗ = 0 diverges for r > 0. For −2 < r < 0, the origin
x∗ = 0 is stable, but then there is a series of period-doubling bifurcations leading to chaos for r < −2, and
the periodic orbits and chaos disappear shortly before r = −4.
1.25
−0.25
−4 r 2
1.125
−0.125
−4 r −3.5
10.3.13
0
2.95 3 r 4
a)
1
The curves being referred to are r, f k

2, r , a graph of which appears below for k = 1, 2, 3, 4, 5. (Keep in
mind that the curves are extended backwards to r values that don’t correspond to periodic orbits.)
0
2.95 3 r 4
k=1 k=2 k=3 k=4 k=5
The reason these curves are darker in the high-resolution orbit diagram is that the slope of the map is zero
1
when x = for all the iterated maps.
2
d k
f (x) = f 0 f k−1 (x) f 0 f k−2 (x) f 0 f k−3 (x) · · · f 0 f (x) f 0 (x)

dxÅ ã Å ã
1 d k 1
f0 =0⇒ f =0
2 dx 2
1
So all the points near x = are mapped to approximately the same value after an iteration of the map since
2
the graph is flat there. Hence those regions are darker because many iterations are required to escape the
region once a trajectory passes nearby.
b)
1 1 1
The corner of the “big wedge” occurs when f 3 = f4 = f5

2, r 2, r 2, r , using the hint and the graphs
of part (a). However, the first equality is enough to solve for the big wedge r value. Now being a bit clever,
Å ã Å ã Å Å ã ã
1 1 1
u = f3 , r = f4 , r = f f3 , r , r = f (u, r)
2 2 2
The r value at the big wedge must be a fixed point of the original function, which we’ve called u. There are
1 1
two fixed points of the logistic map, 0 and 1 − , and the vertical coordinate definitely isn’t 0, so u = 1 − .
r r
This means
Å ã
1 1
f3 ,r = 1 −
2 r
1 4 1 5 1 1 1 8
1 − r + r − r − r6 + r7 − r =0
4 16 16 32 256
Normally you would use a computer to estimate the roots and pick the one that fits, but luckily this has
some nice roots.
(r − 2)4 (r + 2)(r3 − 2r2 − 4r − 8) = 0
from which the cubic formula yields our root of interest. (The other roots are too far off or imaginary.)
2 8 √ − 1 2 √ 1
r= + 19 + 297 3 + 19 + 297 3 = 3.67857...
3 3 3
10.4 Periodic Windows

10.4.1
xn+1 = rexp(xn ) r > 0

a)
xn+1
r = 0.2
xn
1 xn+1
r= ≈ 0.367
e
xn
xn+1
r = 0.5
xn
b)
1
A tangent bifurcation does occur at r = e as is evident in the graphs. We can verify by confirming that the
graph of the iteration function lies tangent to the diagonal line.
1 d x
r= x = 1 ⇒ rex = x re = 1
e dx
c)
xn
1
r = 0.368 >
e
xn
1
r = 0.366 <
e
Here you can see the ghost of the tangent bifurcation. The first graph takes quite a while for the trajectory
to iterate through, but the second graph still has the unstable and stable fixed points in existence.
Note: The graphs look continuous, but that really is an artifact of so many points near the ghost of the
tangent bifurcation and connecting the dots.
10.4.3
A superstable 3-cycle
The iteration map
xn+1 = f (xn ) = 1 − rx2n
will have a superstable 3-cycle if

d
f (f (f (x))) = 0
dx
and
x = f (f (f (x)))
for some x
d
f (f (f (x))) = f 0 (f (f (x)))f 0 (f (x))f 0 (x)
dx
= (−2rf (f (x)))(−2rf (x))(−2rx)
= −8r3 f (f (x))f (x)x = 0
⇒ x = 0 or f (x) = 0 or f (f (x)) = 0
Now we need to check if any of these potential x values are 3-cycles, but we only need to check one of them
because it’s a cycle. The cycle, if it exists, will consist of the points
. . . → xn → xn+1 → xn+2 → xn+3 = xn → . . .
where f (f (xn )) = 0, f (xn+1 ) = 0, and xn+2 = 0. So let’s just pick the easiest equation to use, which is
f (f (x)) = 0.
The x satisfying f (f (x)) = 0 should be a point of the 3-cycle, so
f (f (f (x))) = x and f (f (x)) = 0
⇒ f (f (f (x))) = f (0) = 1 = x
Now we plug x = 1 into f (f (x)) = 0 to derive an equation for r.
f (f (1)) = f (1 − r) = 1 − r(1 − r)2 = 0
So as long as r satisfies
1 − r(1 − r)2 = 0
there will exist a superstable 3-cycle or the iteration map.

10.4.5
A large sampling of initial values was iterated in the logistic map 300 times, but only the last 100 iterations
were plotted in both cases. Clearly something happened after increasing r slightly.
xn+1
r = 3.8495
xn
10.4.7
a)
1 √
x0 = r =1+ 5
2 √
r 1+ 5
x1 = rx0 (1 − x0 ) = = ≈ 0.8 R
4 4
r2 r 4r2 − r3 1
x2 = rx1 (1 − x1 ) = 1− = =
4 4 16 2
√ 1
r > 1 + 5 ⇒ x2 < L
2
b)
RLRR
10.4.9
8
Parameters r = 148.5 σ = 10 b = 3
Initial condition (x, y, z) = (15, 5, −10)

8
Parameters r = 147.5 σ = 10 b = 3
Initial condition (x, y, z) = (15, 5, −10)

10.4.11
xn+1 = axn x1 = a > 0
0 < a < e−e xn tends to a stable 2-cycle

∗
e−e < a < 1 xn → x∗ where x∗ is the unique root of x∗ = ax
1
1 < a < ee xn tends to the smaller root of x = ax
1
a > ee xn → ∞
10.5 Liapunov Exponent

10.5.1
xn+1 = rxn
The explicit formula in terms of the initial condition is not hard to find, and is
xn = r n x0
Now we can add in the error term and compute the Liapunov component exactly.
1 rn δ0

1 δn
λ = = ln = ln |r|
n δ0 n δ0
So the Liapunov exponent λ = ln |r|.
10.5.3
The origin is globally stable for r < 1, by cobwebbing.
xn+1
r = 0.8
xn
There is an interval of marginally stable fixed points for r = 1.

xn+1
r=1
xn
10.5.5
0
0 r 2
0
1 r 2
10.5.7
Suppose x∗ is a fixed point of f undergoing a period-doubling bifurcation. Then f (x∗ ) = −1. Regarding x∗
as x0 , we see x1 = x2 = x3 = · · · = x∗ since x∗ is fixed. Hence,
n−1 n−1
1 X 0 1 X 0 ∗
λ = lim ln f (xi ) = lim ln f (x ) n terms
n→∞ n n→∞ n
i=0 i=0
1
= lim n ln f 0 (x∗ ) = lim ln f 0 (x∗ ) = ln f 0 (x∗ ) = ln − 1

n→∞ n n→∞
= ln(1) = 0
This shows λ = 0 at r = r1 for the logistic map.

Next, consider where period-2 bifurcates to period-4: At r = r2 , we have period-2 points p, q with f (p) = q,
0 0
f (q) = p, f 2 (p) = −1, and f 2 (q) = −1, so f 0 f (p) f 0 (q) = −1; i.e., f 0 (p)f 0 (q) = −1.

Consider the sequence {an } in the definition of the Liapunov exponent:

n−1
!
1 Y
0

lim an = λ an = ln f (xi )
n→∞ n i=0
1
a1 = ln f 0 (x0 ) = ln f 0 (p)
1
1 1
a2 = ln f 0 (x0 )f 0 (x1 ) = ln f 0 (p)f 0 (q)

2 2
1
= ln | − 1| = 0 Since f (p)f 0 (q) = −1 when r = r2
0

2
1
a3 = ln f 0 (x0 )f 0 (x1 )f 0 (x2 )
3
1 1
= ln f 0 (p)f 0 (q)f 0 (p) = ln (−1)f 0 (p)
3 3
1 0
= ln f (p)
3
In general 
1
ln f 0 (p)

a2k+1 =

2k+1
a2k = 0

and as k → ∞, this means a2k+1 → 0 and therefore an → 0 as n → ∞ ⇒ λ = 0 at r = r2 .
The same idea works at higher rm .

10.6 Universality and Experiments

10.6.1
a)
(1.1, 1.1)
(−0.1, −0.1) r
See the solution to Exercise 10.2.3 for the psuedocode.

b)
Zooming in on the orbit diagram further and further with many more iterations gives the approximate ri
values.
r1 ≈ 0.71994 r2 ≈ 0.83326 r3 ≈ 0.85861
r4 ≈ 0.86408 r5 ≈ 0.86526 r6 ≈ 0.86551
c)
r2 − r1
≈ 4.47022
r3 − r2
r3 − r2
≈ 4.63437
r4 − r3
r4 − r3
≈ 4.63559
r5 − r4
r5 − r4
≈ 4.72
r6 − r5
The first few numbers look like they’re converging to Feigenbaum δ = 4.669201609 . . ., except the last one
because five significant figures are not enough since the ri values are getting so close together.
10.6.3
a)
Logistic map with r = 3.6275575
xn+1
xn
Sine map with r = 0.8811406

xn+1
xn
b)
The iteration pattern for both cycles is RLRRR.
10.6.5
a)
Logistic map with r = 3.39057065 5-cycle
xn+1
xn
If you look very carefully you can see that this one doesn’t complete. See part (b).
xn+1
xn

xn+1
xn

xn+1
xn

xn+1
xn

xn+1
xn
b)
The logistic map with r = 3.39057065 was supposed to have a 5-cycle when we plotted the cobweb, but the
numerical simulation didn’t complete.
xn+1
xn
The cobweb wasn’t truncated in this graph. Instead the trajectory got stuck in the following pattern.
xn+1
xn
The trajectory alternates from x ≈ 0.454357 to x ≈ 0.840579, which turns out to be a 2-cycle of the map.
0
The derivative evaluates to approximately f f (x) = −0.71503 at the two alternating x coordinates, so
this is a stable 2-cycle. The 5-cycle passes very close by this stable 2-cycle and was trapped due to roundoff
error.
10.7 Renormalization
10.7.1
a)
g(x) = 1 + c2 x2
x
g(x) = αg 2
α
2c22 x2
1 + c2 x2 = α(1 + c2 ) + + O(x3 )
α
2c22
1 = α(1 + c2 ) c2 =
α

(1, 0)



Ä √ −1−√3 ä
⇒ (α, c2 ) = −1 − 3, 2 ≈ (−2.73205, −1.36603)

√ √

Ä ä
−1 + 3, −1+2 3 ≈ (0.732051, 0.366025)


We want the second solution since that has α closest to Feigenbaum α = −2.5...
b)
g(x) = 1 + c2 x2 + c4 x4
x
g(x) = αg 2
α
2(c22 + 2c2 c4 )x2
1 + c2 x2 + c4 x4 = α(1 + c2 + c4 ) +
α
(c + 6c2 c4 + 2c2 c4 + c24 )x4
3 2
+ 2 + O(x5 )
α3
2(c22 + 2c2 c4 ) (c32 + 6c22 c4 + 2c2 c4 + c24 )
1 = α(1 + c2 + c4 ) c2 = c4 =
α α3
We can use the first and second equations to solve for c2 and c4 in terms of α.
2 α 1 α
c2 = −2 + − c4 = 1 − +
α 2 α 2
Then we can substitute into the third equation to obtain a polynomial in α.
128
4α6 + 3α5 − 60α4 − 104α3 + 168α2 + 240α − 384 + =0
α
The root that is closest to the Feigenbaum α = −2.534... with corresponding c2 and c4 is
α = −2.53403... c2 = −1.52224... c4 = 0.12761...
10.7.3
x
for all x. Then replace x with µx to get

Suppose g(x) = αg g α
Å ã Å Å ãã
x x
g = αg g
µ µα
Multiply both sides by µ to get Å ã Å Å ãã
x x
µg = µαg g (1)
µ µα
Let Å ã
x
h(x) = µg (2)
µ
We want to show that h satisfies the same equation as g with the same α.
From Equation (1), we get (after using µα = αµ)

Å ã Å Å ãã Å Å ãã
x x x
h(x) = µg = µαg g = αµg g (3)
µ µα µα
x

Compare αh h α , which gives
x Å Å ãã
x
αh h = αh µg inner argument equality from Equation (2)
α αµ
Ñ Ä äé
x
µg αµ
= αµg again from Equation (2)
µ
Ñ Ä äé
x
µg αµ
= αµg cancel µ in fraction
µ
= h(x) from Equation (3)
So h satisfies the functional equation with the same α.
10.7.5
a)
From Example 10.7.1, we know that for f (x, r) = r − x2 we have R0 = 0 and R1 = 1. Thus
f (x, R0 ) = −x2 f (x, r1 ) = 1 − x2
Hence
x Åh
x i2
ã ñ Å h x i2 ã2 ô
αf f ,1 ,1 = α f ,1 =α 1− 1−
α α α
2x2 x4 2x2 x4
ï Å ãò
=α 1− 1− 2 + 4 = − 3
α α α α
b)
Matching O(x2 ) terms
x 2x2 2
f (x, r0 ) = −x2 αf 2 ,1 = + O(x4 ) ⇒ −1 =
α α α
and hence α = −2.
10.7.7
Redoing Exercise 10.7.1 part (a)
x
g(x) = αg g
α
4c2 x4
1 + c1 x4 = α(1 + c1 ) + 13 + O(x5 )
α
2
4c
1 = α(1 + c1 ) c1 = 31 ⇒ α = −1.83509 . . . c1 = −1.54493 . . .
α
We chose the α and c1 negative roots as found in the quadratic case.
Redoing Exercise 10.7.1 part (b)

x
g(x) = αg g
α
4(c21 + 2c1 c2 )x4 2(3c31 + 14c21 c2 + 2c1 c2 + 4c22 )x8

1 + c1 x4 + c2 x8 = α(1 + c1 + c2 ) + + + O(x9 )
α3 α7
4(c21 + 2c1 c2 )
1 = α(1 + c1 + c2 ) c1 =
α3
2(3c1 + 14c1 c2 + 2c1 c2 + 4c22 )
3 2
c2 =
α7
Solving for α in the first equation gives
1
α=
1 + c1 + c2
Solving for c1 and c2 in the second and third equations gives
2 α3 1 α3
c1 = −2 + − c2 = 1 − +
α 4 α 4
And then plug in to obtain a polynomial in α.
8
256 + 768α2 + 308α5 + 26α8 + 4α10 + α13 = 6144 + 2048α2 + 1216α3 + 1248α5 + 248α8 + 11α11

α
which has relevant root and corresponding coefficients
α = −1.732354430 . . . c1 = −1.85478 . . . c2 = 0.277528 . . .
Redoing Exercise 10.7.5 part (a)

f (x) = r − x4
Clearly R0 = 0 since f (0) = 0 when r = 0. Now we can find R1 by demanding f (f (0)) = 0, which has
relevant root r = 1. So R1 = 1 as in the quadratic case.
Redoing Exercise 10.7.5 part (b), we match the leading order terms in
x
f (x, R0 ) = αf 2 , R1
α
4
4x 2
−x4 = + O(x5 ) ⇒ α = −2 3 = −1.58740 . . .
α3
Numerically solving
n

f 2 (x, Rn ) =0

x=0
gives
R0 = 0 R1 = 1 R2 = 1.14571 . . . R3 = 1.16527 . . .
R4 = 1.16794 . . . R5 = 1.16831 . . .
And using these to estimate δ

R4 − R3
δ≈ ≈ 7.30235
R5 − R4
For the fourth-degree case, Briggs (1991) finds numerically that the true δ = 7.2846862171 . . . α =
−1.6903029714 . . ., which is close to our calculation.
10.7.9
Following Exercise 10.7.8, we rescale x by a factor of α and observe that the second iterate f 2 resembles f
near the bifurcation point x = 0, r = 0. This gives
x
g(x) = αg g
α
The boundary conditions are g(0) = 1 and g 0 (0) = 1 because of this tangency condition at the bifurcation.
g(x)
a)
Try
x
g(x) =
1 + ax
Then
x x xα
αg g = Å ã=
α ax
ax 2ax + α
1+ α 1+
(1+ ax
α )α
which has the same form as g(x) and equals g(x) when
x xα x
g(x) = = = αg g ⇒α=2
1 + ax 2ax + α α
which works for all a.

b)
α = 2 is (almost) obvious because the steps to find g 2 are twice as long as those for g, as discussed in Exercise
10.7.8 part (c). Otherwise g and g 2 have the same dynamics—namely, slow passage through a bottleneck.
10.7.11
p
µk = −2 + 6 + µk−1 = f (µk−1 )
The graph and cobweb of f are

µk
µk = µk−1
f (µk−1 )
µk−1
−2 µ∗
If we start from µ1 = 0, we see µk → µ∗ > 0, where

√
∗
p −3 + 17
∗
µ = −2 + 6+ µ∗ ⇒µ =
2
as in the text.
11
Fractals
11.1 Countable and Uncountable Sets

11.1.1
The same argument is not valid because the constructed r is not necessarily rational. What if we pick the
order of our list to be such that r = √1 when we’re done?
2
1
√ = 0.707106 . . .
2
x1 = 0.6x1,2 x1,3 x1,4 x1,5 x1,6 . . .
x2 = 0.x2,1 9x2,3 x2,4 x2,5 x2,6 . . .
x3 = 0.x3,1 x3,2 6x3,4 x3,5 x3,6 . . .
x4 = 0.x4,1 x4,2 x4,3 0x4,5 x4,6 . . .
x5 = 0.x5,1 x5,2 x5,3 x5,4 9x5,6 . . .
x6 = 0.x6,1 x6,2 x6,3 x6,4 x6,5 5 . . .

..
.
One possible r = 0.707106 . . . = √1 , which is irrational and thus outside the set xi . So r may not be in the
2
list, but sometimes it’s not supposed to be in the list anyway.
11.1.3
We can prove that the irrational numbers are uncountable by contradiction. We know the real numbers are
uncountable, and taking away the countable set of rational numbers leaves the irrational numbers. Therefore
the irrational numbers must be uncountable since they were made by starting with an uncountable set and
then removing a countable set.
11.1.5
The bijective function f : N → Z × Z × Z is very difficult to create, but we can convince ourselves the
mapping exists by visualizing the points in Cartesian space.
The very first point is the origin. The next set of lattice points lies on the surface of the side length 2
cube centered at the origin, which will have 33 − 1 = 26 points since the middle is taken out. The next set
of lattice points lies on the side length 5 cube centered at the origin, which will have 53 − 33 = 98 points
359
360 Chapter 11: Fractals
since the interior side length 3 cube is taken out. The next set of lattice points will lie on the side length 7
cube, and so on and so forth.
The point (no pun intended) is that the lattice points can be labeled first by cube size and then arbitrarily
on the surface of the cube by a finite natural number. Hence, the set of lattice points in 3-dimensional space
is countable.
11.1.7
xn+1 = 2xn (mod 1)
The problem becomes a little more intuitive if we convert to binary.
bn+1 = 10bn (mod 1)
No matter what value starts the sequence, the first iteration maps into the interval [0,1] and stays there
forever due to the modulo operation. The periodic orbits are the rational numbers in the interval [0,1] since
the decimal eventually repeats or terminates, which is a countable set. The aperiodic orbits of the map are
the irrational numbers, which is an uncountable set.
11.2 Cantor Set

11.2.1
The Cantor set has measure zero. The first iteration removes one interval of length 13 , the second iteration
removes two intervals of length 19 , the third iteration removes four intervals of length 1
27 , etc.
The interval we start with has length one, and the length of the set after the iterated removals is
Å ã Å ã
1 1 1 1 2 4
1− + 2 + 4 + ··· = 1 − 1 + + + ···
3 9 27 3 3 9
∞ Å ã
1 X 2
=1−
3 n=0 3
1 1 11
=1− 2 =1− 1
31− 3 33
1
=1− 3=1−1=0
3
Therefore the measure of the set left over after completing the infinite number of iterated removals (the
Cantor set) is zero.
11.2 Cantor Set 361
11.2.3
The plan for this problem is to upper bound the measure of this countable subset of the real numbers with
another set that we can shrink to zero measure.
First, this set of numbers is countable, so we can list them as {x1 , x2 , x3 , x4 , . . .}.
We can make an interval centered around every xn with length 21−n where > 0.

Å ã

In = xn − n , xn + n
2 2
This set of intervals is countable, and the union of this countable number of intervals contains {x1 , x2 , x3 , x4 , . . .}
and then some. Therefore the measure of the countable set of intervals is an upper bound for the measure
of our countable subset.
∞ ∞
!
[ X
21−n = 2

0 ≤ µ {x1 , x2 , x3 , x4 , . . .} < µ In =
n=1 n=1
We never explicitly chose , so we can scale the size of In and consequently the upper bound for the measure
of {x1 , x2 , x3 , x4 , . . .} to as small as we want, meaning that the measure of {x1 , x2 , x3 , x4 , . . .} has to be 0.
11.2.5
a)
We’ll attack this by pretending we know the ternary representation and multiplying by 3.
[0.5]10 = [0.t1 t2 t3 . . .]3
[3]10 · [0.5]10 = [10]3 · [0.t1 t2 t3 . . .]3
[1.5]10 = [t1 .t2 t3 . . .]3
[1]10 + [0.5]10 = [t1 ]3 + [0.t2 t3 . . .]3
From this we know that we can replace t1 with 1, and then if we iterate we can replace t2 with 1 and t3 with
1 and so on and so forth. Therefore the value decimal 0.5 is 0.1̄ in ternary.
b)
The Cantor set is all ternary numbers in [0,1] that do not contain any 1’s in their ternary representation.
This satisfies every point in the Cantor set corresponding to a point in [0,1], but not vice versa since any
value with a 1 in its ternary representation is not mapped into the Cantor set.
However, we can use binary numbers to uniquely pair every point in [0,1] with a point in the Cantor
set, and vice versa. Basically we read the digit in the binary number and pick the left or right third for that
subinterval if the digit is 0 or 1 respectively. Then we repeat for the next digit and so on, and the Cantor
set point is the left endpoint of the subinterval we end in. For example, the binary number 0.110101 directs
us to pick right right left right left right and then the left endpoint of that subinterval, which is
ï ò ï ò ï ò ï ò ï ò
2 8 24 25 74 75 222 223 222
[0, 1] −→ , 1 −→ , 1 −→ , −→ , −→ , −→
R 3 R 9 L 27 27 R 81 81 L 243 243 L 243
c)
Endpoints in the Cantor set are ternary numbers made entirely of 0’s and 2’s that either eventually reach
endlessly repeating 2’s or terminate, which is the same as endlessly repeating 0’s. Hence, a number in the
Cantor set that is not an endpoint must not terminate and must not eventually become endlessly repeating
2’s, an example of which is 0.02 in ternary.
11.3 Dimension of Self-Similar Fractals

11.3.1
a)
Taking out the middle half of the [0,1] interval set leaves two intervals that will each contain the middle-halves
Cantor set scaled by a factor of four. Then the similarity dimension is
ln(2) ln(2) 1
2 = 4d ⇒ d = = =
ln(4) 2 ln(2) 2
b)
The first iteration removes one interval of length 21 , the second iteration removes two intervals of length 18 ,
1
the third iteration removes four intervals of length 32 , etc.
The interval we start with has length one, so the the length of the set after the iterated removals is
Å ã Å ã
1 1 1 1 2 4
1− + 2 + 4 + ··· = 1 − 1+ + + ···
2 8 32 2 4 16
Å ã
1 1 1
=1− 1 + + + ···
2 2 4
∞ Å ã
1 X 1
=1−
2 n=0 2
1 1 11
=1− =1− 1
2 1 − 12 22
1
=1− 2=1−1=0
2
Hence, the measure of the middle-halves Cantor set is zero.
You can also see this intuitively by taking away the left half of the remaining interval every time instead of
the middle half of each remaining interval. [0, 1] → [0, 0.5] → [0, 0.25] → [0, 0.125] → · · · → [0, 2−n ], so the
length keeps getting closer and closer to 0.
11.3.3
a)
One iteration of the even-sevenths Cantor set leaves four intervals that will each contain the even-sevenths
Cantor set scaled by a factor of seven. Then the similarity dimension is
ln(4)
4 = 7d ⇒ d =
ln(7)
b)
n+1
For odd n, one iteration of the even-nths Cantor set leaves intervals that will each contain the even-nths
2
Cantor set scaled by a factor of n. Then the similarity dimension is
n+1 ln(n + 1) − ln(2)

= nd ⇒ d =
2 ln(n)
11.3.5
One iteration of this no-8’s Cantor set will remove the interval [0.8,0.9) from [0,1] and leave nine intervals
that will each contain this no-8’s Cantor set scaled by a factor of ten. Then the similarity dimension is
ln(9)
9 = 10d ⇒ d =
ln(10)
11.3.7
a)
S0 S1
b)
S2 S3
c)
4
The arc length of S0 = 3 and the arc length are multiplied by 3 each iteration.
∞
Then the arc length of S∞ = 3 34 = ∞.
d)
√
3
S0 has side length one with area 4 and every iteration adds a new equilateral triangle onto each side that
1
is 3 the length of the side.
Å ãn
1
Side length of Sn =
3
n
Number of sides of Sn = 3 (4)
√ Å ã2
3 1
Area added from Sn−1 to Sn = (Number of sides of Sn−1 ) (Side length of Sn−1 )
4 3
√ Ç Å ãn−1 å2
n−1 3 1 1
= 3 (4)
4 3 3
√ ÅÅ ãn ã2
n−1 3 1
= 3 (4)
4 3
√ √ Å ã
3 4 n−1
3 3 3 4 n
= =
4 (32n ) 16 9
If we sum the area of S0 and the area added after each iteration we get
√ √ ∞ Å ã √ √ √ √
3 3 3X 4 n 3 3 3 49 3 3 349
+ = + = +
4 16 n=1 9 4 16 1 − 49 4 16 9 5
√ Å ã √ Å ã
3 349 3 3
= 1+ = 1+
4 495 4 5
√ √
38 2 3
= =
4 5 5
e)
The Koch snowflake is made of three Koch curves. Each Koch curve requires four copies of itself scaled by a
factor of three in order to recreate itself. Therefore the Koch snowflake can be made by taking four complete
copies of itself and scaling each one by a factor of three. Then each scaled Koch snowflake is cut into thirds
and placed onto the full-size Koch snowflake.
Hence, the similarity dimension of the Koch snowflake is
ln(4)
4 = 3d ⇒ d =
ln(3)
11.3.9
At each iteration of the Menger sponge construction, the number of copies is m = 27 − 6 − 1 = 20 (six smaller
cubes for the faces, and one smaller cube from the center) and each copy is scaled by a factor of r = 3 in
each dimension. Therefore the similarity dimension is
ln(m) ln(20)
d= =
ln(r) ln(3)
As for the Menger hypersponge in N dimensions, each iteration of the construction makes 3N smaller N -
dimensional cube regions, and then we delete some. To find how many we delete, we need a pattern for the
construction of the N -dimensional fractal. The pattern turns out to be straightforward.
For the first construction iteration of the 2-dimensional Sierpinski carpet, label the squares of the 3x3
grid as
(1,3) (2,3) (3,3)
(1,2) (2,2) (3,2)
(1,1) (2,1) (3,1)
and then we delete the square with at least two 2’s in the coordinate, which is the middle square. Then we
descend into the remaining eight squares, each with a side length reduced by a factor of three.
For the first construction iteration of the 3-dimensional Menger sponge, label the cubes of the 3x3x3 grid as
(a, b, c) with a, b, c ∈ {1, 2, 3}. Then delete all the cubes with at least two 2’s in the coordinate, which makes
one cube for each surface and the middle cube. Then we descend into the remaining 33 − 6 − 1 = 20 cubes,
each with a side length reduced by a factor of three.
For the first construction iteration of the 4-dimensional Menger hypersponge, label the hypercubes of the
3x3x3x3 grid as (a, b, c, d) with a, b, c, d ∈ {1, 2, 3} and then delete all the hypercubes with at least two 2’s in
the coordinate. Combinatorially counting the number of hypercubes with less than two 2’s in the coordinate
gives the same answer and has an easier formula.
Ç å Ç å
4 3 4 4
2 + 2 = 48
1 0
The number of ways to choose one coordinate out of four to place a 2 with 1’s or 3’s in the remaining three
slots plus the number of ways to place zero 2’s with 1’s or 3’s in the remaining four slots.
Hence, for the 4-dimensional Menger hypersponge, there are m = 48 copies after each iteration and each
copy is scaled by a factor of r = 3. Therefore the similarity dimension is
ln(m) ln(48)
d= =
ln(r) ln(3)
For the N -dimensional Menger hypersponge, the similarity dimension is
ln N 2N −1 + 2N

ln(m)
d= =
ln(r) ln(3)
Just to be safe, let’s check the N -dimensional formula against the results for the Sierpinski carpet
ln (2)(22−1 ) + 22

ln(8)
N =2⇒d=
ln(3)
=
ln(3)
X
and Menger sponge
ln (3)(23−1 ) + 23

ln(20)
N =3⇒d=
ln(3)
=
ln(3)
X
11.4 Box Dimension

11.4.1
Iterations of the Koch snowflake
n=0
n=1
11.4 Box Dimension 367
n=2
n=3
1
If the box has side length = √ , the number of boxes needed is N () = 3 (4n ).
3n 2

ln N () ln (3 (4n ))
d = lim 1 = lim Ä √ ä
→0 ln( ) n→∞ ln 3n 2
ln (4n ) + ln(3) ln(4)

= lim √ =
n→∞ ln(3n ) + ln( 2) ln(3)
11.4.3
Upon each iteration, the box to be descended into is divided into twenty new boxes with the side length
reduced by a factor of three. From this we can calculate
Å ãn
1
= N () = 20n
3
and the box dimension is

ln N () ln(20n ) ln(20)
d = lim = lim =
→0 ln 1

n→∞ ln(3n ) ln(3)

11.4.5
For the 4-dimensional Menger hypersponge, each iteration of the construction descends into 48 smaller
4-dimensional hypercubes with the side length reduced by a factor of three. (See Exercise 11.3.9 for the
derivation.) From this we can calculate
Å ãn
1
= N () = 48n
3

ln N () ln(48n ) ln(48)
d = lim 1
= lim =
→0 ln n→∞ ln(3n ) ln(3)

11.4.7
a)
0 1
S0
S1
S2
S3
S4
b)
1 1 1
We will take the covering interval to be 4 for S1 requiring 3 boxes, 42 for S2 requiring 32 boxes, 43 for S3
requiring 33 boxes, and so on and so forth. From this we can calculate
Å ãn
1
= N () = 3n
4

ln (N () ln(3n ) ln(3)
d = lim 1
= lim n
=
→0 ln n→∞ ln(4 ) ln(4)
c)
S∞ is self-similar. However, its self-similar pieces are not scaled by the same amount, unlike most of the
other fractals we’ve seen.
11.4.9
Upon each iteration we descended into p2 − m2 new boxes with the side length reduced by a factor of p.
From this we can calculate Å ãn
1
= N () = p2 − m2
p
ln (N () ln(p2 − m2 )
d = lim 1
= lim
→0 ln n→∞ ln(p)
11.5 Pointwise and Correlation Dimensions 369
11.5 Pointwise and Correlation Dimensions

11.5.1
This is a challenging problem, and consequently the topic of several publications. See Grassberger and
Procaccia (1983) for guidance.
12
Strange Attractors
12.1 The Simplest Examples

12.1.1
xn+1 = axn yn+1 = byn
The equations are uncoupled, which makes things much simpler. The only thing that matters for conver-
gence is the magnitude of the coefficients a and b. Less than 1, equal to 1, or greater than 1 will converge,
remain a constant distance from or diverge from 0 along that coordinate respectively. The sign of a and b
will determine whether or not the behavior is alternating.
Below are the types of graphs that can occur. Not every case is explicitly shown because some graphs
look the same in forwards and backwards time for different a and b values.
y y y
x x x
y y y
x x x
371
372 Chapter 12: Strange Attractors
12.1.3
12.1.5
a)
B(x, y) = (0.a2 a3 a4 ..., 0.a1 b1 b2 b3 ...) To describe the dynamics more transparently, associate the symbol
...b3 b2 b1 .a1 a2 a3 ... with (x, y) simply by placing x and y back-to-back. Then B(x, y) = ...b3 b2 b1 a1 .a2 a3 ... in
this notation. In other words, B just shifts the binary point one place to the right.
b)
In the notation above, ...1010.1010... and ...0101.0101... are the only period-2 points. They correspond to
( 23 , 31 ) and ( 13 , 23 ).
(1,1)
1 2

3, 3
2 1

3, 3
(0,0)
c)
The points with periodic orbits are when x is a repeating rational number and the digits of y are in the
reverse order of x.
x = .a1 a2 a3 . . . an y = .an . . . a3 a2 a1
(x, y) = a1 a2 a3 . . . an .a1 a2 a3 . . . an
There is a countable number of repeating binary numbers, so there is a countable number of periodic orbits
for the baker’s map.
d)
Pick x = irrational and y = anything. The orbit will never repeat because irrationals never repeat. There
is an uncountable number of irrational numbers, so there is an uncountable number of aperiodic orbits.
e)
There are dense orbits. The plan will be to fill out a grid of points where the coordinates are integer multiples
of 20 , then 2−1 , then 2−2 , then 2−3 ,
and so on.
Next, we’ll need some notation. p and q are sequences of 0’s and 1’s. A point in the unit square can
be represented by (0.p, 0.q) for some p and q. q̃ is the reverse sequence of q.
Now take all the points in the first square and find their binary representation as (0.p11 , 0.q11 ), (0.p12 , 0.q21 ),
(0.p13 , 0.q31 ), and (0.p14 , 0.q41 ), and set
x = 0.q̃11 p11 q̃21 p12 q̃31 p13 q̃41 p14
Now take all the points in the second square and find their binary representation as
(0.p21 , 0.q12 ), (0.p22 , 0.q22 ), . . . , (0.p29 , 0.q92 )
Append to x with the same pattern as before.
x = 0.q̃11 p11 . . . q̃41 p14 q̃12 p21 . . . q̃92 p29
and so on.
Set y = 0 and we have a dense orbit. Applying the baker’s map enough times
(x, y) = 0.q̃11 p11 . . . q̃41 p14 q̃12 p21 . . . q̃92 p29 . . .
B n (x, y) = . . . q̃ij .pji . . . = (0.pji . . . , 0.qij . . .)
will get to any grid point with some junk on the end of the binary representation. The distance between
grid points goes to zero, so this orbit can get arbitrarily close to any point in the unit square within a finite
number of steps. The orbit it dense.
12.1.7
First, we need to make a function definition for the map.
The square will have corners (±1, ±1) to make scaling and shifting easier.
c d
a b
Scale the square by a factor of w in the x direction and h in the y direction.
c d
a b
Instead of making the horseshoe shape, we’re going to make two copies. The top and bottom copy correspond
to −1 ≤ x < 0 and 0 ≤ x ≤ 1 respectively. We’re also going to flip the top piece in both directions before
shifting up.
b a
d c
c d
a b
Next, we shift both the pieces to the left.

b a
d c
c d
a b
And finally chop off the excess.
b a
d c
c d
a b
The series of steps




 x0 = wx x00 = −x0 x000 = x00 − sx


 −1 ≤ x < 0
 y 0 = hy y 00 = −y 0 + sy y 000 = y 00


F (x, y) =
x0 = wx x00 = x0 x000 = x00 − sx



0≤x≤1




 y 0 = hy y 00 = y 0 − s

y 000 = y 00
y
Putting all the steps together 



 −wx − sx



 −1 ≤ x < 0
 −hy + sy

F (x, y) =
wx − sx



0≤x≤1




 hy − s

y
But this is actually missing the chopping-off step. We can restrict the mapping to lie inside the square
instead and not worry about the original. (This is assuming sy isn’t too big to chop off anything in the
vertical direction.)
xn+1 = ±wxn − sx yn+1 = ±hyn ∓ sy
−1 ≤ xn+1 = ±wxn − sx ≤ 1
−1 ≤ wxn − sx ≤ 1 − 1 ≤ wxn + sx ≤ 1
−1 + sx 1 + sx −1 − sx 1 − sx
≤ xn ≤ ≤ xn ≤
w w w w
It looks like the only points that survived from the original square are in two vertical strips.
Next we can find the inverse map explicitly.

±xn+1 + sx ±yn+1 + sy
xn = yn =
w h
1 2
We picked w = 5 h = 5 sx = 2 sy = 5 so that the map came out symmetrically.
a)
Applying the map and then the inverse while remembering to chop off the excess gives
F F −1
b)
F2 F −2
c)
Continuing gives
F3 F −3
F4 F −4
Note: F 3 and F 4 are different, but the detail is too small to see. The color was also changed to black to
show detail.
If we continue, remembering to chop off the excess at every application of F and F −1 to the square, we’ll get
the points that never get chopped off. The invariant set of the map is the Cartesian product of the Cantor
set with itself, also known as Cantor dust.
Intuitively, this occurs because F and F −1 look just like each other but rotated one-quarter turn anti-
clockwise. The intersection of these two sets is Cantor dust.
12.1.9
xn+1 = xn + yn+1 = xn + yn + k sin(xn ) yn+1 = yn + k sin(xn )
a) Ñ é
1 + k cos(x) 1
J=
k cos(x) 1

det(J) = 1 + k cos(x) 1 − 1 k cos(x) = 1
b)
k = 0 ⇒ xn+1 = xn + yn+1 yn+1 = yn ⇒ xn = x0 + ny0
(x0 , y0 )
Å
31π
ã 2π
,6
32
Å ã
π 13π
,
12 9
Å ã
3π
1,
7
Å ã
1 1
,
4 4 0
The orbits are periodic if y0 is a rational multiple of π, otherwise the orbits are quasiperiodic.
c)
k = 0.5
x
d)
k=1
x
e)
k=2
12.2 Hénon Map

12.2.1
T 000 (x, y) = (y, x) T 00 (x, y) = (bx, y) T 0 (x, y) = (x, 1 + y − ax2 )
T 000 (T 00 (T 0 (x, y))) = T 000 (T 00 (x, 1 + y − ax2 )) = T 000 (bx, 1 + y − ax2 )
= (1 + y − ax2 , bx)
12.2.3
a)
a = 0.25 b = 0.5
y y0 y 00 y 000
x x0 x00 x000
(a) (b) (c) (d)
12.2.5
Ñ é
−2ax∗ 1
J=
b 0
(−2ax∗ − λ)(0 − λ) − b = 0
λ2 + 2ax∗ λ − b = 0
−2ax∗ ± (2ax∗ )2 − 4(−b)
p
λ=
2(1)
»
= −ax∗ ± (ax∗ )2 + b
12.2.7
xn+2 = bxn + 1 − a(yn + 1 − ax2n )2 yn+2 = b(yn + 1 − ax2n )
x = bx + 1 − a(y + 1 − ax2 )2 y = b(y + 1 − ax2 )
We can solve for y in the right-hand equation by rearranging a bit.
b(ax2 − 1)
y=
b−1
We can then substitute for y in the left-hand equation.

ã2
y 2 b(ax2 − 1)
Å
a
x = bx + 1 − a(y + 1 − ax2 )2 = bx + 1 − a = bx + 1 −
b b2 b−1
Moving everything to one side and simplifying gives an equation without y.
0 = (b − 1)3 x + (b − 1)2 − a(ax2 − 1)2

The roots of this equation will be the x coordinates of the 2-cycle. However, two of the four roots to this
equation will also be fixed points of the Hénon map x = y + 1 − ax2 = bx + 1 − ax2 and we can divide them
out.
(b − 1)3 x + (b − 1)2 − a(ax2 − 1)2
0= = a2 x2 + a(b − 1)x − a + (b − 1)2
bx + 1 − ax2 − x
which has solutions
p p
1−b± 4a − 3(b − 1)2 b 1 − b ∓ 4a − 3(b − 1)2
x= ⇒y= a 6= 0
2a 2a
We have to throw out the a = 0 case since it’s not a 2-cycle.
The coordinates must be real for the 2-cycle to exist, meaning
3
4a − 3(b − 1)2 > 0 ⇒ a > a1 = (1 − b)2
4
To determine stability, we’ll have to linearize and find when all |λ| < 1.
Ñ é Ñ é
dxn+2 dxn+2 2 2 2
dxn dyn b + 4a x(y + 1 − ax ) −2a(y + 1 − ax )
J= =
dyn+2 dyn+2
dxn dyn −2abx b
Ñ é
4a2 x∗ y ∗ ∗
b+ b − 2ay
b
J(x∗ ,y∗ ) =
−2abx∗ b
where we simplified using the y equation for the 2-cycle.
2a2 x∗ y ∗ + b2 ± 2 a2 x∗ y ∗ (a2 x∗ y ∗ + b2 )
p
λ± =
b
2(a2 x∗ y ∗ ) + b2 ± 2 (a2 x∗ y ∗ )2 + b2 (a2 x∗ y ∗ )
p
=
b
We can use the equations from finding the 2-cycle to simplify.
b a(x∗ )2 − 1 abx∗ a2 (x∗ )2 − a

a 2 x∗ y ∗ = a 2 x∗ =
b−1 b−1
0 = a2 (x∗ )2 + a(b − 1)x∗ − a + (b − 1)2 a2 (x∗ )2 − a = −(b − 1) ax∗ + (b − 1)

a2 x∗ y ∗ = −abx∗ ax∗ + (b − 1) = −b a2 (x∗ )2 + a(b − 1)x∗

0 = a2 (x∗ )2 + a(b − 1)x∗ − a + (b − 1)2 a2 (x∗ )2 + a(b − 1)x∗ = a − (b − 1)2
a2 x∗ y ∗ = −b a − (b − 1)2 = b (b − 1)2 − a

2(a2 x∗ y ∗ ) + b2 ± 2 (a2 x∗ y ∗ )2 + b2 (a2 x∗ y ∗ )

p
λ± =
b »
2
2
2
2b (b − 1) − a + b ± 2 b2 (b − 1)2 − a + b3 (b − 1)2 − a
=
q b
2
= 2 (b − 1)2 − a + b ± 2 (b − 1)2 − a + b (b − 1)2 − a

Finally, we look at |λ± | < 1 to determine stability. This is a bit of a headache because we have to worry
about complex eigenvalues. Eventually we obtain the following by assuming a positive argument to the
square root, a negative argument to the square root, and the inequality from the existence of the 2-cycle.
1 2 3
a< (5b − 6b + 5) −1<b<1 a> (b − 1)2
4 4
The intersection of all three inequalities is plotted below.
a
12.2.9
a)
L1 (t) = (−1.33 + 2.65t, 0.42 − 0.287t) t ∈ [0, 1]
L2 (t) = (1.32 − 0.075t, 0.133 − 0.273t) t ∈ [0, 1]
L3 (t) = (1.245 − 2.305t, −0.14 − 0.36t) t ∈ [0, 1]
L4 (t) = (−1.06 − 0.27t, −0.5 + 0.92t) t ∈ [0, 1]
H1 (t) = 0.42 − 0.287t + 1 − a(−1.33 + 2.65t)2 , b(−1.33 + 2.65t) t ∈ [0, 1]

H2 (t) = 0.133 − 0.273t + 1 − a(1.32 − 0.075t)2 , b(1.32 − 0.075t) t ∈ [0, 1]

H3 (t) = − 0.14 − 0.36t + 1 − a(1.245 − 2.305t)2 , b(1.245 − 2.305t) t ∈ [0, 1]

H4 (t) = − 0.5 + 0.92t + 1 − a(−1.06 − 0.27t)2 , b(−1.06 − 0.27t) t ∈ [0, 1]

a = 1.4 b = 0.5
y
b)
For this part, it’s easier to define the sides of the quadrilateral in Cartesian coordinates.
y1 = −0.108302x + 0.275958
y2 = 3.64x − 4.6718
y3 = 0.156182x − 0.334447
y4 = −3.40741x − 4.11185
We need to ensure that the top and bottom lines L1 and L3 bound the y value of the transformation for
every given x value in the transformation.
H1 : x = 0.42 − 0.287t + 1 − a(−1.33 + 2.65t)2
y1 = −0.108302(0.42 − 0.287t + 1 − 1.4(−1.33 + 2.65t)2 ) + 0.275958
= 1.06477t2 − 1.03771t + 0.390375
y3 = 0.156182(0.42 − 0.287t + 1 − 1.4(−1.33 + 2.65t)2 ) − 0.334447
= −1.5355t2 + 1.49647t − 0.499447
y3 ≤ 0.3(−1.33 + 2.65t) ≤ y1
y3 − 0.3(−1.33 + 2.65t) ≤ 0 ≤ y1 − 0.3(−1.33 + 2.65t)
− 1.5355t2 + 0.70147t − 0.100447 ≤ 0 ≤ 1.06477t2 − 1.83271t + 0.789375
We then would have to find the local minimum or maximum of these parabolas to determine if the inequality
is satisfied, keeping in mind that t ∈ [0, 1] for the parabolas. We would have to do this for the H2 , H3 , and
H4 curves as well.
Similarly, we need to ensure that the left and right lines, L4 and L2 respectively, bound the x value of the
transformation for every given y value in the transformation.
Verification of the other inequalities is left to the reader.

12.2.11
For b = 0, the map reduces to a one-dimensional map after one iteration.
x1 = y0 + 1 − ax2n y1 = (0)x0 = 0
xn+1 = yn + 1 − ax2n = 1 − ax2n yn+1 = 0 for n 6= 0
From then on, yn = 0, making xn+1 effectively a function of xn only.
12.2.13
A cursory investigation of the area preserving the Hénon map doesn’t reveal much. Some iterations of a
square centered at the origin are graphed below.
12.2.15
xn+1 = 1 + yn − a|xn |
yn+1 = bxn
Ñ é
±a 1
J= det(J) = −b
b 0
So the Lozi map contracts areas if |det(J)| = | − b| < 1 ⇒ −1 < b < 1.
12.2.17
xn+2 = 1 + yn+1 − a|xn+1 | yn+2 = bxn+1

xn+2 = 1 + bxn+1 − a1 + yn − a|xn | yn+2 = b 1 + yn − a|xn |

x = 1 + bx − a1 + y − a|x| y = b 1 + y − a|x|

12.3 Rössler System 387
2-cycle points
1−a−b b(1 + a − b)
Å ã
(x, y) = , a>1−b
a2 + (b − 1)2 a2 + (b − 1)2
1+a−b b(1 − a − b)
Å ã
= , a>b−1
a2 + (b − 1)2 a2 + (b − 1)2
but along with the condition that −1 < b < 1, only a > 1 − b must be satisfied for the fixed points to exist.
Linearization Ñ é
b + a2 sgn x(1 + y − a|x|) −a sgn

A=
−ab sgn(x) b
Ñ é
b − a2 −a
AÄ 1−a−b b(1+a−b)
ä=
,
a2 +(b−1)2 a2 +(b−1)2 ab b
Ñ é
b − a2 a
AÄ 1+a−b b(1−a−b)
ä=
,
a2 +(b−1)2 a2 +(b−1)2 −ab b
Plugging in the coordinates and solving for |λ± | < 1 along with the other conditions gives the stable region.
√
−a2 + 2b ± a a2 − 4b
|λ± | = <1 −1<b<1 a>1−b
2
a
(1, 2)
(0, 1)
b
(1, 0)
12.3 Rössler System

12.3.1
a)
Finding when the Hopf bifurcation occurs is most easily done with an orbit diagram.
b=2 c=4
0 a 0.3342 0.4
0.3736
This graph is made by numerically integrating the system until the trajectory reaches the attractor; then
the positive values of x are recorded every time y changes sign for each a.
The first branching occurs at a ≈ 0.3342, which is the Hopf bifurcation, and the next branching occurs
at a ≈ 0.3736, which is the period-doubling bifurcation.
b)
a = 0.3342 b = 2 c = 4
Initial condition (x, y, z) = (−5, 0, 0) t ∈ [0, 50]
z(t)
y y
x x
12.4 Chemical Chaos and Attractor Reconstruction 389
a = 0.3736 b = 2 c = 4
Initial condition (x, y, z) = (−5, 0, 0) t ∈ [0, 50]
z(t)
y y
x x
12.3.3
The Rössler system lacks the symmetry of the Lorenz system. The Lorenz equations are invariant under

the transformation (x, y, z) → (−x, −y, z), so if x(t), y(t), z(t) is a solution to the Lorenz system, then

−x(t), −y(t), z(t) is also a solution.
12.4 Chemical Chaos and Attractor Reconstruction

12.4.1
The general equation of an ellipse is
Ax2 + Bxy + Cy 2 + Dx + Ey + F = 0
such that A, B, C, D, E, and F are constants and (x, y) is a set of points. The ellipse is a non-degenerate
real ellipse if B 2 − 4AC < 0 and
B D
A

2 2

C B2 C E = C∆ < 0

2
D E
2 2 F
Otherwise, if C∆ > 0 then the ellipse is imaginary, and if ∆ = 0 then the ellipse is a point.

Hence, if (x, y) = sin(t), sin(t + τ ) is an ellipse, then we should be able to find constants A, B, C, D, E,
and F such that
A sin2 (t) + B sin(t) sin(t + τ ) + C sin2 (t + τ ) + D sin(t) + E sin(t + τ ) + F = 0
Simplifying using symbolic mathematics software gives a big mess.
A sin2 (t) − A cos2 (t) + A − B cos(τ ) cos2 (t) + B cos(τ ) sin2 (t)
+ 2B sin(τ ) sin(t) cos(t) + B cos(τ ) − C sin2 (τ ) sin2 (t) − C cos2 (τ ) cos2 (t)
+ C sin2 (τ ) cos2 (t) + C cos2 (τ ) sin2 (t) + 4C sin(τ ) cos(τ ) sin(t) cos(t)
+ C + 2D sin(t) + 2E sin(τ ) cos(t) + 2E cos(τ ) sin(t) + 2F = 0
But we know that all the like t terms must cancel, giving
sin2 (t) : A + B cos(τ ) − C sin2 (τ ) + C cos2 (τ ) = 0

sin(t) cos(t) : 2B sin(τ ) + 4C sin(τ ) cos(τ ) = 0
2
cos (t) : −A − B cos(τ ) − C cos2 (τ ) + C sin2 (τ ) = 0
sin(t) : 2D + 2E cos(τ ) = 0
cos(t) : 2E sin(τ ) = 0
1 : A + B cos(τ ) + C + 2F = 0
which is really a homogeneous matrix equation with solution

   
A 1
   
B  −2 cos(τ )
   
   
C   1
   

 = 
   
D  0 
   
   
E   0 
   
F − sin2 (τ )
Therefore the trajectory in Figure 12.4.5 is an ellipse.

12.4.3
Following is the delayed trajectory. The initial conditions were (x, y, z) = (−5, 0, 0), and the simulation was
run until t = 1000, with the initial transient discarded.
a = 0.4 b = 2 c = 4 τ = 1.5
Trajectories circle around the middle of the attractor in about 6 units of time, so the delay is about one-
quarter of this period.
12.5 Forced Double-Well Oscillator

12.5.1
ẋ = y ẏ = −δy + x − x3 + F cos(ωt) ω=1 F =0
The following simulations have initial conditions of x, y ∈ [−2.5, 2.5] in a 900 × 900 pixel image. After
simulating for 200π units of time (100 cycles)—which is a bit silly since F = 0 means there is no drive—
points that were within 0.1 of the −1 and +1 basins with a speed less than 0.1 were considered to have
settled in that basin. A black pixel corresponds to that initial condition settling into the −1 basin, and
white corresponds to the +1 basin.
δ = 0.25
δ = 0.125
δ = 0.0625
The basins become thinner as the damping decreases, and they appear to wrap around each other more and
more as δ tends to zero. This will make for sensitive dependence on initial conditions, which means that
accurate prediction of which basin an initial condition will settle into is virtually impossible unless the initial
condition is practically settled into one of the basins already.
12.5.3
ẋ = y ẏ = −ky − x3 + B cos(t) k = 0.1 B = 12
(x, y) = (2, 2)
t ∈ [0, 100] (10, 10) t ∈ [0, 7000] (10, 10)
y(t) y(t)
(−10, −10) x(t) (−10, −10) x(t)
There is also a limit cycle for this system.

(x, y) = (1, 0)
t ∈ [250, 500] (10, 10) t ∈ [0, 7000] (10, 10)
y(t) y(t)
(−10, −10) x(t) (−10, −10) x(t)

12.5.5
θ̈ + bθ̇ − sin(θ) + F cos(t) b = 0.2 F = 2
Note: The θ value is taken modulo 2π.

a)
(θ, θ̇) = (5, 10) (0.5, 0) on the left and right respectively
t ∈ [0, 200] (7.5, 7.5) t ∈ [0, 200] (7.5, 7.5)
θ̇(t) θ̇(t)
(−7.5, −7.5) θ(t) (−7.5, −7.5) θ(t)
The left and right Poincaré section settles down to approximately (5.807,2.0338) and (5.806,-0.61238) re-
spectively.
These fixed points correspond to period motion of the pendulum, shown below.
(5.807,2.0338)
θ̇(t)
θ
−2π 2π
(5.806,-0.61238)
θ̇(t)
θ
−2π 2π
Note: The jumps in the middle are due to the modulo 2π operation.
b)
The following simulations have initial conditions of θ, θ̇ ∈ [−2.5, 2.5] in a 900x900 pixel image. A point was
determined to be in a basin if, when strobed at an integer multiple of 2π, the θ and θ̇ were within 0.1 and
had a speed less than 0.1 of the fixed point of the Poincaré map (5.807,2.0338) and (5.806,-0.61238). A black
pixel corresponds to that initial condition settling into the (5.806,-0.61238) basin, and white corresponds to
the (5.807,2.0338) basin.
You can see the fractal aspect of the basins at the top of the image.

Strogatz Notes PDF

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

Strogatz Notes PDF

Uploaded by

Copyright:

Available Formats

Student Solutions Manual for

Nonlinear Dynamics and Chaos,

Boca Raton London New York

CRC Press is an imprint of the

Published 2018 by CRC Press

No claim to original U.S. Government works

Visit the Taylor & Francis Web site at

and the CRC Press Web site at

ISBN 13: 978-0-8133-5054-7 (pbk)

2 Flows on the Line 1

4 Flows on the Circle 65

6 Phase Plane 103

7 Limit Cycles 173

8 Bifurcations Revisited 219

9 Lorenz Equations 273

10 One-Dimensional Maps 307

12 Strange Attractors 371

2.1 A Geometric Way of Thinking

2.2 Fixed Points and Stability

We could also plot the graph using a computer.

Unable to find an analytic solution.

Apply an indefinite integral to both sides with respect to t.

And using the initial condition.

2.3 Population Growth

ẊY − X Ẏ aXY − bXY

This is the logistic equation

2.4 Linear Stability Analysis

Depending on the sign of the initial condition

x(t) t ∈ [0, 10] x(t) = −x3 x(t) = −x

2.5 Existence and Uniqueness

ẋ = −xc x(t) ≥ 0 c∈R

And with c < 1 then

x(t) = 0 is a solution, but we can find another solution through

2.6 Impossibility of Oscillations

The pattern of equilibrium points continues in both directions.

x(t) = x(t + T ) ⇒ V (x(t)) = V (x(t + T ))

2.8 Solving Equations on the Computer

ln(E) ln(C∆t) ln(C) + ln(∆t) ln(C) ln(∆t)

3.1 Saddle-Node Bifurcation

The fixed points are

√ 2 fixed points exist if |r| > 2

3.2 Transcritical Bifurcation

Transcritical bifurcation occurs at r = 0.

Transcritical bifurcation occurs at r = 1.

= (k1 a − k2 b)x − k−1 x2

3.3 Laser Threshold

p = 0.125 p=1 p=2

A transcritical bifurcation occurs at pc .

ṅ = GnN − kn = (GN − k)n

3.4 Pitchfork Bifurcation

A subcritical pitchfork bifurcation occurs at r = 0.

A supercritical pitchfork bifurcation occurs at r = 0.

A saddle-node bifurcation occurs at r = 0.

A saddle-node bifurcation occurs at r = 5.

A subcritical pitchfork bifurcation occurs at r = −1.

V (x0 ) = C, so for simplicity we’ll set C = 0, which leaves

V (x2 ) = V (x4 ) = V (x0 )

3.5 Overdamped Bead on a Rotating Hoop

from the text.

Now the f (φ) is the negligible term.

3.6 Imperfect Bifurcations and Catastrophes

This x versus r bifurcation diagram

Taylor series expansion

Now the f (φ) is the negligible term.

To find out, we can look at the level sets of E for arbitrary .

= 2ẋ2 + (2 x4 + 2x2 + 1) − 1

= 2ẋ2 + (x2 + 1)2 − 1

4E + 1 = 2ẋ2 + (x2 + 1)2