iii MAT217—X/1/00-02

APM214—Y/1/00-02

Contents

Page
Contents iii

Preface v

CHAPTER 1 Overview 1

CHAPTER 2 Linear Autonomous Dynamical Systems 11

CHAPTER 3 Linear Control Theory 72

CHAPTER 4 Autonomous Nonlinear Systems 114

CHAPTER 5 Advanced Topics 179

APPENDIX 193
 v MAT217—X/1
APM214—Y/1

Preface

In writing this study guide, we leant heavily on two books:

James T. Sandefur: Discrete Dynamical Systems: theory and

applications. Clarendon Press, Oxford, 1990. First Edition.

David G. Luenberger: Introduction to Dynamic Systems: the-

ory models and applications. John Willey, New York, 1979.
First Edition.

The following books were also useful in some places:

Dennis G. Zill and Michael R. Cullen: Differential equations.

PWS Kent, Boston, 1992. Third Edition.

K.J. Falconer: Fractal geometry: mathematical foundations

and applications.

This second—year module requires a background including first—year

mathematics, specifically calculus and linear algebra. Some expo-
sure to mathematical modelling is useful but not essential.

N.T. BISHOP F.E.S. BULLOCK

Revised by S.A. DE SWARDT
 1 MAT217—X/1
APM214—Y/1

CHAPTER 1

OVERVIEW

1. Introduction
First—year calculus courses can give the impression that all problems
involving differential equations are soluble in the sense that the solu-
tion can be expressed as a simple formula. In fact, this impression
is quite wrong : most differential equations do not have such so-
lutions. (The reason the impression arises, of course, is that the
exercises in a first calculus course are carefully chosen so as to have
simple solutions.) So, if faced with a problem that cannot, in the
above sense, be solved, what can one do (other than give up)? One
could use

(a) Numerical methods

(There are two modules on this topic in the Applied Mathematics
course.) Numerical methods form an important and useful tool, but
they do have limitations : all the coefficients in an equation must
have precise numerical values. This is fine if one wants a solution
to a specific problem, but is not very helpful for obtaining a general
understanding of the behaviour of the solutions. For this we turn to
 2

(b) Dynamical Systems theory

The use of dynamical systems theory avoids “not seeing the wood
for the trees”. In other words, we try and understand the overall
behaviour of a system, without getting bogged down in the details.
This is perhaps best illustrated by an example:

A woman, with a parachute, jumps out of an aeroplane. The equa-

tion describing her motion is

dv k
= g − v 1.9 , v > 0 (1)
dt m
where v is her downward speed, g is the acceleration due to gravity
(approximately 9.8ms−2 ), m is the mass of the woman and para-
chute, and k is a constant associated with the parachute. For safety,
the woman should hit the ground at not more than 5ms−1 . What
is the least possible value for the ratio k/m?

Solution 1 (which gets bogged down in detailed calculus)

At t = 0, v = 0, and suppose at t = T we have v = V . Then
integrating equation (1) gives
] v=V ] t=T
dv
k 1.9
= dt.
v=0 g− mv t=0

The integral on the right is easy (it equals T ), but the first integral
is difficult. Trying all the tricks learnt in MAT102, plus some oth-
ers, leads NOWHERE. So we resort to

Solution 2 (Dynamical systems approach)

The equation (1) has a so—called fixed point (i.e. a point where
 3 MAT217—X/1
APM214—Y/1

dv
= 0) at
dt
 mg 1/1.9
v = v∗ ≡ .
k
(The symbol “ ≡ ” means “defined to be”.)

For v > v∗ we have dv/dt < 0 and for v < v∗ we get dv/dt > 0.
Thus for this problem, whatever the initial value of v, as time evolves
v → v∗ . This is illustrated in the diagram below:

0 v* v

We call v∗ a stable fixed point. Thus v∗ is a good estimate of the

speed at which the woman hits the ground. Further, since we are
not given the time or height of the fall, it is in fact the only estimate
that can be made. The condition v∗ < 5ms−1 implies
mg
< 51.9 ,
k
and therefore
k 9.8
≥ 1.9 ≈ 0.46045 in S.I. units.
m 5

2. What is a dynamical system?

The essential property of a dynamical system is that it involves one
or more functions of time. For example,

x(t), (x(t), y(t)), x(t) = (x1 (t), ..., xk (t)).

 4

The dynamical system can be discrete or continuous, which means

that the time t is regarded as discrete or continuous. A continuous
dynamical system is usually described by a differential equation, or
a system of differential equations. For example,
dx
= cos x,
dt
or, in the case of a system of differential equations,
dx1
= x1 − x1 x2
dt
dx2 x1 x2
= x2 − .
dt 2
In a discrete dynamical system the time is regarded as a non—
negative integer (t = 0, 1, 2, 3, ...), and the evolution of the system is
usually described by a difference equation, or a system of difference
equations. For example, we might have

x(n + 1) = 2x(n),

or the system of difference equations

p(n + 1) = q(n)p(n)
q(n + 1) = 1 − p(n).

[Note that the recommended textbooks for this module use the
above notation, but other textbooks might write the difference equa-
tions as
xn+1 = 2xn
and

pn+1 = qn pn ,
qn+1 = 1 − pn .]
 5 MAT217—X/1
APM214—Y/1

In the module APM114—V (Mathematical modelling) some of the

following ideas of dynamical systems theory were introduced in the
context of a single differential equation or difference equation.

(a) Autonomous Systems

If the right—hand side of the differential or difference equation does
not involve time explicitly the system is called autonomous. For
example, the systems
dx
= x2 − x
dt
and
1
x(n + 1) = (x(n)2 − x(n))
2
are autonomous, but
dx
= x2 − t
dt
and
1
x(n + 1) = (x(n)2 − n)
2
are NOT autonomous. Much of the work in dynamical systems is
for autonomous systems because then the fixed points do not de-
pend on time.

(b) Fixed Points (also called singular or equilibrium points)

We call x∗ a fixed point if at x = x∗ we have

dx
=0
dt
or, in the case of a difference equation, if at x(n) = x∗ we have

x(n + 1) = x(n) = x∗ .
 6

As seen in the parachute example in section 1, fixed points can be

an extremely effective tool.

(c) Stability, or instability, of fixed points

A formal definition and precise methods for determining the nature
of fixed points are given later. For now, we just illustrate the ideas
by examples. Note that sometimes stable points are called attrac-
tors, or attracting points, and unstable points are called repellors, or
repelling points.

Examples

dx
1. = x2 − x.
dt

We solve the equation x2∗ − x∗ = 0 to get the fixed points. This

gives x∗ = 0 or x∗ = 1. If x < 0 then dx/dt > 0, and if 0 < x < 1
then dx/dt < 0. If x > 1 then dx/dt > 0. We therefore get the
following phase line for this system:

0 1

(See the module APM114—V for a discussion of phase lines.) The

point x∗ = 0 is a stable fixed point, and x∗ = 1 is an unstable fixed
point.

dx
2. = x2 .
dt

Solving the equation x2∗ = 0 gives x∗ = 0 as the only fixed point.

 7 MAT217—X/1
APM214—Y/1

We see that dx/dt > 0 if x < 0 or if x > 0, so we get the phase line
below:

0
This system is neither stable nor unstable. The reasons why will
be discussed later — we included this example to show some of the
subtleties that can arise in dynamical systems theory.
1
3. x(n + 1) = − x(n) + 6.
2
To find the fixed point(s) we substitute x∗ for x(n) and x(n + 1) in
the difference equation to get x∗ = − 12 x∗ + 6. Solving this equation
for x∗ gives x∗ = 4. Plotting x(n) against n, starting at x(0) = 2,
gives the following graph:
x(n)

5 *

4 * *
*
*
3

2*

n
0 1 2 3 4 5

(x(0) = 2, x(1) = 5, x(2) = 7/2, x(3) = 17/4, x(4) = 31/8, ...)

No matter what starting value is chosen for x(0), the same pattern
results. Thus the fixed point at x = 4 is stable.

3. Higher Order Differential and Difference Equations

Everything described so far has been for first order differential and
 8

difference equations. You may wonder whether the methods of dy-

namical systems theory can be applied to higher order equations.
The answer is “YES”, because one higher order equation is equiva-
lent to a system of first order equations. The following two examples
illustrate this:

1. Consider the third order differential equation

 2
d3 y dy
y 3 + + ey = 3.
dt dt
d2 y
Let x1 = y, x2 = dy
dt , and x3 = dt2 . Then we get an equivalent first
order system of equations
dx3 1  
= 3 − ex1 − x22
dt x1
dx2
= x3
dt
dx1
= x2 .
dt
2. Consider the difference equation

y(n + 1) = y(n)y(n − 1) − y(n)2 + a.

The order of a difference equation is the difference between the

largest time and the smallest time appearing in the equation. Here,
the times appearing are (n + 1), n and (n − 1), so the order is
(n + 1) − (n − 1) = 2. Thus the equation will be written as a
system of two first order difference equations. First, it is convenient
to rewrite the difference equation so that the smallest time value
appearing in it is n. We do this by replacing n − 1 by n, n by n + 1
and n + 1 by n + 2. Thus the difference equation can be written

y(n + 2) = y(n + 1)y(n) − y(n + 1)2 + a.

 9 MAT217—X/1
APM214—Y/1

Now let x1 (n) = y(n), x2 (n) = y(n + 1). Then the above difference
equation is equivalent to

x1 (n + 1) = x2 (n)
x2 (n + 1) = x2 (n)x1 (n) − x2 (n)2 + a.

There is more about this matter, and some more examples, in Lu-
enberger on pages 96 and 97.

4. Plan for the rest of the Module

The plan for the other chapters in this module is as follows: In
Chapter 2 we discuss linear autonomous dynamical systems. This
subject is important in its own right, and also because it forms the
basis for the material in Chapters 3 and 4.

Chapter 3 is about linear control theory. From a mathematical

point—of—view, in the discrete case, the system is linear but not
autonomous, the system equation being

x(n + 1) = Ax(n) + u(n)

where the vector u(n) is called the control. The idea is to use the
control u(n) to steer the system state vector x to some desired value.

Chapter 4 discusses non—linear autonomous dynamical systems by

linearising about the fixed points. Particular attention is paid to
systems of two first order equations and the construction of phase
portraits. We also investigate non—linear autonomous dynamical
systems using fully non—linear techniques. We describe the Lia-
punov theory of stability, and also limit cycles for continuous sys-
 10

tems and n—cycles for discrete systems.

In Chapter 5 we touch on some more advanced topics: chaos, bifur-

cation theory and fractals.

Throughout the module we will illustrate the theory with applica-

tions from physics, biology, economics and the social sciences.
 11 MAT217—X/1
APM214—Y/1

CHAPTER 2

LINEAR AUTONOMOUS
DYNAMICAL SYSTEMS

Linear Autonomous Dynamical systems are described by the equa-

tions
ẋ = Ax + f or x(n + 1) = Ax(n) + f (1)

where x and f are vectors and A is a matrix. Neither A nor f

vary with time. We consider in section 1 the simplest such systems,
namely those where only one dimension is involved.

1. One Dimensional Systems

In this case there is only one equation, so that A and f are just
numbers a and b. Then equations (1) become

ẋ = ax + f and x(n + 1) = ax(n) + f. (2)

There are standard methods for writing down a general solution of

(2), but here we proceed in a slightly different way, more in keeping
with the dynamical systems approach.

In the continuous case ẋ = ax + f , the fixed point x∗ is found by

 12

solving
f
ẋ = 0 ⇒ ax∗ + f = 0 ⇒ x∗ = − . (3a)
a
In the discrete case x (n + 1) = ax (n) + f , we replace x (n) and
x (n + 1) with x∗ yielding
f
x∗ = ax∗ + f ⇒ x∗ = . (3b)
1−a
Let
X(t) = x(t) − x∗ or X(n) = x(n) − x∗ . (4)

Then the continuous case in equation (2) becomes

ẋ (t) = Ẋ (t) = a (X (t) + x∗ ) + f

= aX (t) + ax∗ + f
= aX (t) ,

since x∗ = − fa .

Similarly, for the discrete case in equation (2), we have

x (n + 1) = X (n + 1) + x∗ = a (X (n) + x∗ ) + f.

Hence

X (n + 1) = aX (n) + x∗ (a − 1) + f
= aX (n) ,
f
since x∗ = 1−a .

Equation (2) therefore becomes

Ẋ = aX or X(n + 1) = aX(n). (5)

 13 MAT217—X/1
APM214—Y/1

We see that the substitution (4) has got rid of the constant term f .
Equations (5) can be solved immediately. For the continuous case
we have
Ẋ (t)
dt = a dt
X (t)
nX (t) = at + k (k is a constant)

so that
X (t) = Ceat (C = ek ).

From the initial condition X (0) it follows that

C = X (0)

so that the solution of Ẋ = aX is therefore given by

X (t) = X (0) eat . (6a)

For the discrete case in equation (5), it follows from the initial con-
dition X (0) that

X (1) = aX (0)
X (2) = aX (1) = a (aX (0)) = a2 X (0)
X (3) = aX (2) = a a2 X (0) = a3 X (0)
..
.

clearly yielding
X (n) = an X (0) . (6b)
Using (4), (3a) and (3b), the solutions of the original equation (2)
are therefore:
f f
x(t) = − + eat x(0) +
a a
 14

or (7)

f f
x(n) = + an x(0) − .
1−a 1−a

[The only problem that can arise with (7) is if a = 1 (discrete case),
or if a = 0 (continuous case). Then (2) is

ẋ = f or x(n + 1) = x(n) + f

which have solutions

x(t) = x(0) + f t and x(n) = x(0) + f n. (8)

In this situation there are no fixed points unless f = 0, in which

event every value of x is a fixed point. However, these cases are not
of much practical importance, so we will not dwell on them further.]

From (6a) and (6b), it is easy to see under what conditions this dy-
namical system has a stable fixed point. If |a| < 1 (discrete case),
or a < 0 (continuous case) then an → 0 as n → ∞ or eat → 0
as t → ∞, and therefore from equations (6a) and (6b) we see that
X → 0 so that x → x∗ and the fixed point is stable. On the other
hand, if |a| > 1 (discrete case) or a > 0 (continuous case) then
X = x − x∗ → ∞ and the fixed point is unstable.

To summarize the nature of the fixed point x∗ in the case of one

dimensional systems, we have the following table:
 15 MAT217—X/1
APM214—Y/1

discrete |a| < 1 stable

case |a| > 1 unstable
continuous a<0 stable
case a>0 unstable

We conclude this section with some examples to illustrate the above

theory.

Example 1
Consider the discrete system
1
x(n + 1) = x(n) + 2
2
with initial condition
x(0) = 2.
Then, replacing x (n + 1) and x (n) with the fixed point x∗ , we get
1
x∗ = x∗ + 2
2
and therefore
x∗ = 4.
Let X(n) = x(n) − 4. (This of course also means that X (n + 1) =
x (n + 1)−4.) In terms of the new variable X the difference equation
becomes
1
X (n + 1) + 4 = (X (n) + 4) + 2
2
clearly yielding
1
X(n + 1) = X(n)
2
with solution
1 n
X(n) = X(0).
2
 16

Transforming back to the original variable x gives

n
1
x(n) = (x(0) − 4) + 4
2
n n−1
1 1
= (−2) + 4 = − + 4.
2 2
Since (1/2)n → 0 as n → ∞ we see that x(n) → 4 as n → ∞ and
the fixed point x∗ = 4 is stable.
Plotting a graph we get:
x(n)

4 * * * *
*
3

2*

n
0 1 2 3 4 5

Example 2
Consider the discrete system

x(n + 1) = 2x(n) + 1

with initial condition

x(0) = 3.

Then
x∗ = 2x∗ + 1

and therefore
x∗ = −1.
 17 MAT217—X/1
APM214—Y/1

Let X(n) = x(n) + 1. In terms of the new variable X the difference

equation becomes
X(n + 1) = 2X(n)

with solution
X(n) = 2n X(0).

Transforming back to the original variable x gives

x(n) = 2n (x(0) + 1) − 1 = 2n+2 − 1.

Therefore x(n) → ∞ as n → ∞ and the fixed point is not stable.

Example 3
Consider the discrete system:
2
x(n + 1) = − x(n), x(0) = 3.
3
This is easy to solve. The solution is x(n) = 3(−2/3)n .

Since (−2/3)n → 0 as n → ∞ the system is stable.

Example 4
Consider the system
dx
= −2x + 1, x(0) = −1.
dt
Since the system is continuous, the fixed point is found by solving
dx
= 0 ⇒ 0 = −2x∗ + 1,
dt
which gives
1
x∗ = .
2
 18

Making the substitution

1
X (t) = x (t) − ,
2
the differential equation becomes
dX
= −2X,
dt
in terms of the new variable X which yields the solution

X (t) = X (0) e−2t .

Transforming back to the original variable x gives

1 1
x (t) − = x (0) − e−2t ,
2 2

which, together with the initial condition x (0) = −1, leads to the
solution
1 1
x (t) = + −1 − e−2t
2 2
1 3 −2t
= − e .
2 2
Example 5
Consider the system
dx 3
= x − 3, x(0) = 0.
dt 2
The solution is
3
x(t) = −2e 2 t + 2.
 19 MAT217—X/1
APM214—Y/1

2. Two Dimensional Systems

We now move on to the two—dimensional case, i.e. in equation (1)
x and f are 2—vectors and A is a 2 × 2 matrix. The theory will be
described for the continuous case (ẋ = Ax+ f ), and the correspond-
ing equations for the discrete case will be summarised in section 5.
The dynamical system
ẋ = Ax + f (9)

has a fixed point provided ẋ = 0, so the fixed point is given by

Ax + f = 0.

Therefore
x∗ = −A−1 f . (10)

[We assume that det(A) = 0.]

Then, as in the one—dimensional case, let

X(t) = x(t) − x∗ (11)

so that (9) becomes

Ẋ = AX. (12)

Put in words, the transformation (11) moves the fixed point from
−A−1 f to the origin. In order to understand the properties of equa-
tion (9), it is sufficient to understand the simpler equation (12).
Even so, the solution of (12) is rather more complicated than solv-
ing equation (5).
Since the solution of the one—dimensional equation (5) is X =
constant × exp(constant ×t), we try as a solution of (12):

X = Keλt . (13)
 20

[Note: this form of solution may appear to be just a lucky guess;

which, to some extent, is what it is.]
Substituting (13) into (12) gives

Kλeλt = AKeλt .

Therefore
AK = λK. (14)

Equation (14) may be familiar as the equation defining an eigenvec-

tor K, with eigenvalue λ, of a matrix A. Eigenvalues and eigen-
vectors are covered in APM113 and MAT211, but not in MAT103.
Thus students taking this module may not have come across eigen-
values and eigenvectors. The notes that follow will not assume any
prior knowledge of eigenvalue theory, and where necessary, standard
results will simply be stated.

Equation (14) may be rewritten:

(A − λI)K = 0 (15)

1 0
where I is the identity matrix . We require that K
0 1
should be non—trivial (i.e. K = 0), which implies that

det(A − λI) = 0. (16)

[WHY?: Because if det(A − λI) = 0, then (A − λI)−1 exists, and

multiplying (15) by (A − λI)−1 gives K = (A − λI)−1 .0 = 0, which
is not wanted.]
 21 MAT217—X/1
APM214—Y/1

Equation (16) is an equation that determines λ and can be regarded

as a generalisation of the one—dimensional case in which λ = a.
Writing
a b
A= (17)
c d

then equation (16) becomes

a−λ b
=0
c d−λ

i.e.
(a − λ)(d − λ) − bc = 0 (18)

which is a quadratic equation for λ. It can be rewritten as

λ2 − λ(a + d) + (ad − bc) = 0

or
λ2 − λσ + ∆ = 0 (19)

where σ = a + d is the sum of the diagonal components of A, and

is called the trace of A, and ∆ = ad − bc is det(A). The solution
of (19) is

σ+ σ2 − 4 σ− σ2 − 4
λ1 = , λ2 = . (20)
2 2
In general, the procedure for obtaining the solution of equation (12)
Kx
is as follows: First, we use (15) to find the eigenvector K(= ):
Ky

(A − λI)K = 0 (15)
 22

Now, because det(A − λI) = 0, (15) leads to two equations for

Kx , Ky that are multiples of each other:

aKx + bKy = 0
caKx + cbKy = 0

[For a 2 × 2 matrix det = 0 means that the two rows of the matrix
are multiples of each other.] Thus K is defined in direction only,
and we represent this by writing down the solution of (15) as cK for
some undetermined constant c. Of course, there are two eigenvalues
λ1 , λ2 so there are two corresponding eigenvectors

c1 K1 , c2 K2 .

Then the general solution of (12) is

X(t) = c1 K1 eλ1 t + c2 K2 eλ2 t . (21)

For any choice of X(0) we can find c1 and c2 such that

X(0) = c1 K1 + c2 K2 .

Further, these values of c1 , c2 are unique. This is because eigenvec-

tor theory shows that, if all the eigenvalues are distinct, then the
eigenvectors form a basis of the linear space. [The special case of
(16) having equal roots is considered later.] It is convenient to give
examples of the above general procedure under three sub—headings:

(A) σ 2 − 4∆ > 0: there are two distinct real solutions for λ.

(B) σ 2 − 4∆ < 0: there are two distinct complex solutions for λ.
(C) σ 2 − 4∆ = 0: “equal roots”.
 23 MAT217—X/1
APM214—Y/1

Case (A)
Once λ1 and λ2 are known, equation (15) is solved for K1 and K2 .

Example 1
Consider the system:
dx
= x + 3y
dt
dy
= 5x + 3y. (22)
dt
x 1 3
Let X = , then Ẋ = AX with A = . Thus equation
y 5 3
(19) becomes
λ2 − λ(4) − 12 = 0.
Thus λ1 = +6, λ2 = −2. Let K1 and K2 be the eigenvectors
corresponding to eigenvalues λ1 , λ2 respectively. Then equation
(15), for K1 is:
−5 3 0
K1 = . (23)
5 −3 0
Note that, although (23) looks like two equations, one is a multi-
ple of the other so there is, in effect, only one equation. As re-
marked earlier, this always happens because λ has been chosen so
that det(A − λI) = 0, and a zero determinant means, for a 2 × 2
matrix, that one row is a multiple of the other. Let, for example
a
K1 = .
b
Then (23) becomes

−5a + 3b = 0 ...(A)
5a − 3b = 0 ...(B)
 24

(Note that (B) = −1 × (A).) Choose for example, a = 3c1 , then

b = 5c1 , where c1 is an arbitrary constant. The solution of (23)
can therefore be written:

3
K1 = c1
5

for some constant c1 . Note that the choice of a and b are not
unique. We could also have chosen a = c1 (say). Then b = 53 c1 and
our vector K1 would be given by

1
K1 = c1 5
.
3

This however is just a constant multiple of the previous solution of

K1 :
1 1 5
c1 5 = c1 !
3
5 3

Similarly for K2 , equation (15) is

3 3
K2 = 0
5 5

so that
1
K2 = c2 .
−1

Thus from (13), and using the principle of superposition for linear
systems, the general solution of (22) is:

x 3 1
X= = c1 e6t + c2 e−2t . (24)
y 5 −1
 25 MAT217—X/1
APM214—Y/1

[Reminder: this means

x = 3c1 e6t + c2 e−2t

y = 5c1 e6t − c2 e−2t ]

Note that X → ∞ as t → ∞.

The above solution is the general solution because it has two arbi-
trary constants, which is correct as (21) is a system of two first—order
differential equations. The solution curves for various values of c1 ,
c2 are shown in Figure 1. The fixed point at (0, 0) is called a saddle
point.

Figure 1: Saddle point

Example 2
Consider the system:
dx
= 2x + 8y,
dt
 26

dy
= 4y, (25)
dt
so that clearly
2 8
A= .
0 4
Then (19) becomes
λ2 − 6λ + 8 = 0.
a
Solving this equation we get λ1 = 2, λ2 = 4. Let K1 = be
b
the eigenvector corresponding to the eigenvalue λ1 = 2. Then, from
equation (15) we need to solve

(A − 2I) K1 = 0,

i.e.
0 8 a 0
=
0 2 b 0
yielding the two equations

8b = 0
2b = 0.

This clearly gives b = 0, and since there is no restriction on the

unknown a, we can choose a = c1 , where c1 is an arbitrary constant.
Hence, K1 is given by

1
K1 = c1 .
0
 27 MAT217—X/1
APM214—Y/1

c
Next, let K2 = be the eigenvector corresponding to the eigen-
d
value λ2 = 4. Then, from (A − λ2 I) K1 = 0 (see equation (15)), we
get
−2 8 c 0
=
0 0 d 0
i.e.
−2c + 8d = 0 ⇒ 2c = 8d ⇒ c = 4d.

Again we have only one equation with two unknowns. (The second
equation is zero multiplied by the first one.) Clearly we can solve one
unknown in terms of the other, that is, let d = c2 , then c = 4d = 4c2
giving the vector
4
K2 = c2 ,
1
with c2 an arbitrary constant.

Therefore

x 1 4
X= = c1 e2t + c2 e4t . (26)
y 0 1

The solution curves for various values of c1 , c2 are shown in Figure

2. The fixed point at (0, 0) is called an unstable node. Clearly, from
(26), it follows that X → ∞ as t → ∞.
 28

Figure 2: Unstable node

Example 3
Consider the system:

dx
= −x − 2y + 4
dt

dy
= x − 4y + 2. (27)
dt
First, we need to find the location of the fixed point (x∗ , y∗ ) by
solving the equations:

0 = −x∗ − 2y∗ + 4,
0 = x∗ − 4y∗ + 2.
 29 MAT217—X/1
APM214—Y/1

These are easily solved to give x∗ = 2, y∗ = 1. Let X = (x − x∗ ,

y − y∗ ). Then
dX −1 −2
= X.
dt 1 −4

Equation (19) becomes

λ2 + 5λ + 6 = 0,

with solutions λ1 = −2, λ2 = −3. Substituting into equation (15)

we get
1 −2 2 −2
K1 = 0, K2 = 0.
1 2 1 −1

2 1
The solutions are K1 = c1 , K2 = c2 . Then
1 1

x − x∗ 2 1
X= = c1 e−2t + c2 e−3t ,
y − y∗ 1 1

or

x = 2c1 e−2t + c2 e−3t + 2

y = c1 e−2t + c2 e−3t + 1. (28)

The solution curves for various values of c1 , c2 are shown in Figure

3. The fixed point at (2, 1) is called a stable node. Also note from
(28) that (x, y) → (2, 1) as t → ∞.
 30

Figure 3: Stable node

Case (B)
In this case equation (19) has two distinct complex solutions for λ.
Writing β = 4 −σ 2 /2, the solutions are
σ σ
λ1 = + iβ, λ2 = − iβ. (29)
2 2
Note that λ1 and λ2 are complex conjugates of each other.

Example 1
Consider the system
dx
= 6x − y,
dt

dy
= 5x + 4y. (30)
dt
 31 MAT217—X/1
APM214—Y/1

Then
6 −1
A= ,
5 4
so that equation (19) becomes

λ2 − 10λ + 29 = 0.

The roots λ1 , λ2 are

λ1 = 5 + 2i, λ2 = 5 − 2i (31)

Thus equation (15) for K1 becomes

(1 − 2i) −1
K1 = 0. (32)
5 (−1 − 2i)

[Note that, as in case (A), the second row of the matrix is linearly
dependent on the first : just multiply row one by (1 + 2i) to see
this.] Solving (32) gives

1
K1 = c1 . (33)
1 − 2i

Similarly, we find the other eigenvector to be

1
K2 = c2 . (34)
1 + 2i

Note that if the matrix A is real the vector K1 is the complex

conjugate of K2 . It is simple to prove (but it won’t be done here)
that this is always the case. Thus the solution of (30) can be written

x 1 1
X= = c1 e(5+2i)t + c2 e(5−2i)t . (35)
y 1 − 2i 1 + 2i
 32

It would, however, be useful to be able to interpret (35) in terms of

real functions only. Using e2it = cos(2t) + i sin(2t), equation (35)
can be written:

x = e5t [(c1 + c2 ) cos 2t + (c1 i − c2 i) sin 2t],

y = e5t [(1 − 2i)(c1 cos 2t + ic1 sin 2t)
+(1 + 2i)(c2 cos 2t − ic2 sin 2t)]
= e5t [(c1 + c2 ) − 2(c1 i − c2 i)] cos 2t
+e5t [2(c1 + c2 ) + (c1 i − c2 i)] sin 2t. (36)

Let
C1 = c1 + c2 and C2 = c1 i − c2 i. (37)

Then

x cos 2t
X= = C1 e5t
y cos 2t + 2 sin 2t
sin 2t
+C2 e5t . (38)
−2 cos 2t − sin 2t

Since we want C1 , C2 to be real, (37) implies that c1 and c2 should

be complex conjugates of each other.
Equation (38) is then an interpretation of (35) in terms of real func-
tions.
There is a general and straightforward procedure for deriving the
general, real solution (38) from (31) (solution for λ) and (33) (solu-
tion for the first eigenvector K1 ). Using the notation K1 to mean
the complex conjugate of K1 , so K1 = K2 , and defining:
1
B1 = Re(K1 ) = (K1 + K1 )
2
 33 MAT217—X/1
APM214—Y/1
1
B2 = Im(K1 ) = (K1 − K1 ), (39)
2
(where the constant c1 multiplying K1 is taken as 1), then

X = [C1 (B1 cos βt − B2 sin βt) + C2 (B2 cos βt + B1 sin βt)]e(σ/2)t

(40)
The general proof of (40) is not difficult, but it involves some slightly
intricate complex number manipulation, and is omitted.

N.B. — In the above formula the eigenvector K1 must be worked

out for the eigenvalue λ1 which has positive imaginary part.

In example 1, K1 is given by (33), so that from

1 1 0
= +i ,
1 − 2i 1 −2

we have that
1 0
B1 = Re (K1 ) = , B2 = Im (K1 ) = .
1 −2

Also σ/2 = 5 and β = 2. Thus the solution is

1 0
X = e5t C1 cos 2t − sin 2t
1 −2
0 1
+e5t C2 cos 2t + sin 2t
−2 1

which, as expected, is the same as (38).

The solution curves are shown in Figure 4. The fixed point at (0,0)
is called an unstable focus.
 34

Figure 4: Unstable focus

Example 2
Consider the system

d x −1 2 x 2
= 1
+ . (41)
dt y − 2 −1 y 1

The fixed point (x∗ , y∗ ) is found by solving

−1 2 x∗ 2
1
=−
− 2 −1 y∗ 1

so that x∗ = 2, y∗ = 0. Then the eigenvalue equation (19) is

λ2 + 2λ + 2 = 0
 35 MAT217—X/1
APM214—Y/1

which has solutions λ1 = −1 + i, λ2 = −1 − i. Thus σ/2 = −1,

a
β = 1.Then K1 = is found from solving
b

−i 2 a 0
= ,
− 12 −i b 0
i.e.

−ia + 2b = 0 ...(1)
1
− a − ib = 0 ...(2)
2
i
From (1), choose a = 1 (say), then b = 2 so that

1
K1 = i
2

1 0
= +i 1
0 2

and therefore
1 0
B1 = and B2 = 1
.
0 2
Hence
x − x∗ 1 0
X = = e−t C1 cos t− 1
sin t
y − y∗ 0 2

0 1
+C2 1
cos t + sin t .
2 0
Thus

x = 2 + e−t (C1 cos t + C2 sin t)

1 1
y = e−t ( C1 sin t + C2 cos t). (42)
2 2
 36

The solution curves are shown in Figure 5. The fixed point at

(0, 0) is a stable focus.

Figure 5: Stable focus

Example 3
Consider the system

2 8
Ẋ = X. (43)
−1 −2

Equation (19) is:

λ2 + 4 = 0
so that
λ1 = 2i, λ2 = −2i.
The equation for K1 is:

(2 − 2i) 8
K1 = 0
−1 (−2 − 2i)
 37 MAT217—X/1
APM214—Y/1

which has a solution

2 + 2i 2 2
K1 = = +i
−1 −1 0
so that
2
B1 =
−1
and
2
B2 = .
0
Since σ = 0 the solution is
2 cos 2t − 2 sin 2t 2 cos 2t + 2 sin 2t
X = C1 + C2 . (44)
− cos 2t − sin 2t
The solution curves are shown in Figure 6. The fixed point (0,0) is
called a centre.

Figure 6: Centre
 38

Case (C)
Here we have “Equal roots”, so σ 2 − 4∆ = 0. This is a special case
which, for the purposes of this module, is not very important. So we
just briefly summarise the solutions. There are really two sub—cases:

C − 1:
Here
a 0
Ẋ = X (45)
0 a

x (t)
which, for X = , can be written as
y (t)

ẋ (t) a 0 x (t)
=
ẏ (t) 0 a y (t)

i.e.
ẋ (t) = ax (t) and ẏ = ay (t) .

These two equations clearly have the solutions

x (t) = c1 eat and y (t) = c2 eat

so that
x 1 0
X= = c1 eat + c2 eat . (46)
y 0 1
C − 2:
All other cases in which σ 2 − 4∆ = 0. It turns out that there is only
one eigenvector K. So one solution is

X = c1 Keλt (47)
 39 MAT217—X/1
APM214—Y/1

but the second solution still needs to be found. It is:

X = c2 (Kteλt + Peλt ). (48)

Substituting (48) into (12) yields

(AK − λK)teλt + (AP − λP − K)eλt = 0 (49)

so that
(A − λI)K = 0 (50)

and
(A − λI)P = K. (51)

Equation (50) is the usual eigenvector equation for K and (51) then
determines P.

Example
Consider the system

3 −18
Ẋ = X. (52)
2 −9

The eigenvalue equation is

λ2 + 6λ + 9 = 0

or
(λ + 3)2 = 0,

so
λ = −3.
 40

Then the equation for K is

6 −18
K=0
2 −6

3
which has a solution K = . Then from (51), the equation for
1
P is
6 −18 3
P=
2 −6 1
i.e.
2a − 6b = 1.
1
1 2
Now choose (say) b = 0, then a = 2 so that P = . Thus the
0
general solution of (52) is
1
3 3
X = c1 e−3t + c2 te−3t + 2 e−3t . (53)
1 1 0

3. Nature of fixed points : stability

The question of stability of a fixed point is rather more complicated
than for the case of one first—order equation. There are the follow-
ing possibilities determined by the parameters σ(= trace A) and
∆(= det(A)):

1. σ < 0, 0 < ∆ < σ 2 /4. This means that both eigenvalues λ1 , λ2

are distinct, real and negative. This type of fixed point is called a
stable node and was illustrated earlier in figure 3.

2. σ > 0, 0 < ∆ < σ 2 /4. This means that both eigenvalues λ1 , λ2

are distinct, real and positive. This is called an unstable node, and
 41 MAT217—X/1
APM214—Y/1

is illustrated in figure 2.

3. ∆ < 0. This implies that the eigenvalues λ1 , λ2 are distinct, real

and of opposite sign. The fixed point is then a saddle point, and is
illustrated in figure 1.

4. ∆ > σ 2 /4, σ > 0. The eigenvalues λ1 , λ2 are complex and with

positive real part. The fixed point is an unstable focus (also called
an unstable spiral -point) and is illustrated in figure 4.

5. ∆ > σ 2 /4, σ < 0. The eigenvalues λ1 , λ2 are complex and with

negative real part. The fixed point is a stable focus (also called a
stable spiral point) and is illustrated in figure 5.

6. σ = 0, ∆ > 0. The eigenvalues λ1 , λ2 are pure imaginary. The

fixed point is a centre and is illustrated in figure 6.

The above description of fixed points is summarised in the following

diagram:
 42

For stability we require

∆ > 0 and σ < 0.

4. Higher Dimensional Systems

We now investigate the general case for continuous systems

ẋ = Ax + f (54)

where the dimensionality is taken as m. As before, assuming A−1

exists,
ẋ = 0 at x∗ = −A−1 f . (55)

Let
X = x − x∗ (56)

then
Ẋ = AX. (57)

For simplicity we assume that the eigenvalues of A are distinct,

although they may be complex: λ1 , ..., λm . The eigenvectors Ki of
A are found by solving

(A − λi I)Ki = 0 (i = 1, ..., m). (58)

Then a general solution of (54) is:

m
x= (ci Ki eλi t ) − A−1 f . (59)
i=1

For a large system, carrying out the procedure to obtain (59) from
(54) is at least tedious if not impractical. So, is there any useful
 43 MAT217—X/1
APM214—Y/1

information that can be obtained from (54) without finding the com-
plete general solution? The stability of the solution about the fixed
point x∗ is often of great importance, and is determined entirely by
the eigenvalues λ1 , ..., λm . More precisely, it is obvious from (59)
that (54) is stable at x = x∗ if and only if

Re(λi ) < 0 (i = 1, ..., m). (60)

[Although it is not part of this module, we should mention that

there are powerful and efficient computer methods for calculating
the eigenvalues of a large matrix : it is horribly inefficient to form
det(A − λI) = 0 and then solve a polynomial equation in λ.]

Example 1
Consider the system
 
−4 1 1
 
Ẋ =  1 5 −1  X. (61)
0 1 −3

Thus det(A − λI) is

 
−4 − λ 1 1
 
det  1 5−λ −1  ,
0 1 −3 − λ

which is equal to

(−4 − λ) [(5 − λ) (−3 − λ) − (−1)] − 1 [(−3 − λ) − 1]

= (−4 − λ) λ2 − 2λ − 14 − (−λ − 4)
= (−4 − λ) λ2 − 2λ − 15
= (−4 − λ) (λ + 3) (λ − 5) .
 44

Thus the eigenvalues are —3, —4 and 5, and the general solution is

X = c1 K1 e−3t + c2 K2 e−4t + c3 K3 e5t (62)

where the eigenvectors K1 , K2 , K3 have not been calculated. Even

so, it is clear that the solution (62) is unstable (because of the term
e5t ), and is some form of 3—dimensional generalisation of a saddle
point.

Example 2
Consider the system
   
−1 1 −2 −1
   
ẋ =  1 −1 0  x+  3  . (63)
1 0 −1 2

First, the fixed point x∗ is the solution of 0 = Ax∗ +f , and, after

some linear algebra we find
 
−1
 
x∗ =  2  .
1

Then putting X = x − x∗ gives Ẋ = AX, and the eigenvalues are

found from 0 = det(A − λI):
 
−1 − λ 1 −2
 
det  1 −1 − λ 0 
1 0 −1 − λ
= (−1 − λ)[(−1 − λ)2 − 1] + 1(−(−2)(−1 − λ))
= −(λ + 1)3 + 1 + λ − 2 − 2λ
= −λ3 − 3λ2 − 4λ − 2.
 45 MAT217—X/1
APM214—Y/1

Thus we seek the solutions of

λ3 + 3λ2 + 4λ + 2 = 0.

One solution is λ = −1, so the above equation becomes

(λ + 1)(λ2 + 2λ + 2) = 0

which is straightforward to solve, yielding:

λ1 = −1,
λ2 = −1 + i,
λ3 = −1 − i.

The general solution is therefore

X = c1 K1 e−t + c2 K2 e(−1+i)t + c3 K3 e(−1−i)t . (64)

It is clear that the solution is stable, since all eigenvalues have neg-
ative real part. The terms e±it can be expressed in terms of sin t
and cos t, and therefore the solution has a component which oscil-
lates with frequency 1/2π. [Recall that sin ωt or cos ωt has period
T = 2π/ω and frequence ω/2π.] In some applications the frequency
(or frequencies) may be important.

5. Discrete systems
First we describe the general situation, and then investigate some
aspects of 2 − D systems in more detail.

The equation describing the system is

x(n + 1) = Ax(n) + f . (65)

 46

There is a fixed point x∗ given by solving

x∗ = Ax∗ + f .

Therefore
x∗ = (I − A)−1 f . (66)

[We assume that (I − A) has an inverse]. Then let X = x − x∗ so

that
X(n + 1) = AX(n). (67)
Of course, one can write down the solution of (67) as:

X(n) = An X(0). (68)

Unfortunately, (68) does not give much insight into how X(n) evolves
in (discrete) time. Rather, we again use eigenvalue theory. For sim-
plicity, we assume that A has m distinct eigenvalues λ1 , ..., λm , with
corresponding eigenvectors K1 , ..., Km . Then, as discussed earlier,
the eigenvectors form a basis of the linear space and for any X(0)
we can write
m
X(0) = ci Ki (69)
i=1
where the constants ci are unique. Putting (69) into (68) gives
m m m
X(n) = ci An Ki = ci λni Ki = λni ci Ki . (70)
i=1 i=1 i=1

From (70) it is clear that if |λi | < 1(i = 1, ..., m) then X(n) → 0
as n → ∞ for any given initial condition X(0). Such a system is
stable. On the other hand if any |λi | > 1 then, under certain initial
conditions, X(n) → ∞ as n → ∞ and the system is unstable.
 47 MAT217—X/1
APM214—Y/1

Thus the long term behaviour of a discrete dynamical system (65),

is determined by the eigenvalues of A. We illustrate this, and some
of the subtleties that can arise, by examples.

Example 1
Consider the system

x(n + 1) = x(n) + 2y(n), y(n + 1) = −x(n) + 4y(n) (71)

with
x(0) = 1, y(0) = 2.

Let
x (n)
X (n) = .
y (n)
Then equation (71) can be written

1 2
X (n + 1) = X (n)
−1 4

with
1
X (0) = .
2
We have calculated eigenvalues and eigenvectors for a 2 × 2 matrix
many times in this chapter, and we just state the result:

2 1
λ1 = 2, K1 = , λ2 = 3, K2 = .
1 1

We must now express X(0) in terms of K1 and K2 , i.e. we find

constants c1 and c2 such that

X(0) = c1 K1 + c2 K2 .
 48

This gives

1 = 2c1 + c2
2 = c1 + c2 .

Solving we get c1 = −1, c2 = 3. Thus the solution to the problem

is
2 1
X(n) = c1 K1 λn1 + c2 K2 λn2 = −1 2n + 3 3n .
1 1
Thus

x(n) = −2n+1 + 3n+1

y(n) = −2n + 3n+1 . (72)

Example 2
Consider the system

0.5 −0.25
X (n + 1) = X (n)
1 0.5

with
1
X (0) = . (73)
2
The eigenvalue equation is

λ2 − λ + 0.5 = 0

so that

λ1 = 0.5 + 0.5i,
λ2 = 0.5 − 0.5i.
 49 MAT217—X/1
APM214—Y/1

The eigenvectors are

i
K1 = ,
2
−i
K2 = .
2

Thus the general solution is

i −i
X(n) = c1 (0.5 + 0.5i)n + c2 (0.5 − 0.5i)n .
2 2

1
The constants c1 , c2 are determined from X(0) = .
2
We get
1 i (c1 − c2 )
=
2 2 (c1 + c2 )
which are solved to give

c1 = 0.5 − 0.5i
c2 = 0.5 + 0.5i.

Therefore

i
X(n) = (0.5 − 0.5i)(0.5 + 0.5i)n
2
−i
+ (0.5 + 0.5i) (0.5 − 0.5i)n . (74)
2

There is one important point to make about equation (74). The sec-
ond term is the complex conjugate of the first term, and therefore
their sum is real. This type of behaviour must always occur when
 50

the matrix A is real, because it is obvious that, if X(0) is real, then

so is X(n).

One can also interpret (74) in terms of sine and cosine functions.
Using the (r, θ) representation of complex numbers z = a + ib, i.e.

z = reiθ

where

r = a2 + b2
b
θ = tan−1 ,
a
we write:
1
(0.5 − 0.5i) = √ e−iπ/4 ,
2
1 iπ/4
(0.5 + 0.5i) = √ e .
2
Then (74) becomes
n+1
1 eiπ/2
X(n) = √ eiπ/4(n−1)
2 2
n+1
1 e−iπ/2
+ √ eiπ/4(1−n)
2 2
n+1
1 eiπ/4(n−1)+iπ/2 + eiπ/4(1−n)−iπ/2
= √
2 2eiπ/4(n−1) + 2eiπ/4(1−n)
n+1
1 2 cos(π/2 + (n − 1)π/4)
= √
2 4 cos((n − 1)π/4)
n−1
1 − sin((n − 1)π/4)
= √ , (75)
2 2 cos((n − 1)π/4)
 51 MAT217—X/1
APM214—Y/1

where the second last line of (75) was derived from

eit + e−it
cos t = .
2

Example 3
Consider the system

2 1 2
x (n + 1) = x (n) + . (76)
0 2 −1

Then the fixed point is given by

2 1 −2
x∗ = x∗ + .
0 2 −1

Therefore
1 1 −2
x∗ + = 0,
0 1 −1

and hence
1
x∗ = .
1

Let
X(n) = x(n) − x∗ .

Then
2 1
X (n + 1) = X (n) .
0 2

The eigenvalue equation is

(λ − 2)2 = 0.
 52

There is therefore a repeated root at λ1 = 2. As in the continuous

system case, we describe the procedure for dealing with this situa-
tion, but do not prove it. Let K1 be the single eigenvector, [so that
1
here K1 = c1 ]. Then define P by
0
(A − λI)P = K1 . (76)

0
[Here P = c2 .] Then the general solution is
1

X(n) = c1 λn1 K1 + c2 (λn1 P + nλn−1

1 K1 ) (77)

so that, for this example,

1 0 1
X(n) = c1 2n + c2 2n + n2n−1 .
0 1 0
Thus
1 0 1 1
x(n) = c1 2n + c2 2n + n2n−1 + .
0 1 0 1
(78)
For discrete systems, the fixed points can be classified as follows:

(a) λ1 , λ2 are complex, and further λ2 = λ1 . In this case, as for

continuous systems, there is a spiralling effect — towards the
fixed point if |λ| < 1, (stable case), and away from it if |λ| > 1
(unstable case).

(b) λ1 , λ2 real and both greater than 1. The behaviour is analo-

gous to an unstable node.
 53 MAT217—X/1
APM214—Y/1

(c) λ1 , λ2 real and 0 < λ1 , λ2 < 1. The behaviour is analogous to

a stable node.

(d) λ1 , λ2 real and 0 < λ1 < 1 < λ2 . The behaviour is analogous

to a saddle point.

(e) One, or both λ1 , λ2 negative. There is no analogy with the

continuous situation. The solution keeps jumping from one
side of the fixed point to the other (as in section 1, example
3). If |λ1 | < 1 and |λ2 | < 1 then the behaviour is stable, and
if either |λ1 | > 1 or |λ2 | > 1 then it is unstable.

6. Some Applications

(i) Lanchester model of combat. [These models are named af-

ter F.W. Lanchester who investigated mathematical models
of combat situations during World War I.]

Suppose that two military forces are engaged in combat. Let

x(t) and y(t) be the strengths of the two forces in some ap-
propriate units. For example, the units could be number of
soldiers, or number of aircraft, or number of ships, depending
on the type of battle being fought. Strictly speaking x(t), y(t)
are integers, but it is convenient to regard them as continuous,
differentiable functions of time.

The key assumption in the model is that each force has a “hit-
ting strength”, which is proportional to its size; this “hitting
 54

strength” reduces the other force:

dx dy
= −αy, = −βx (79)
dt dt

where α and β are constants which depend upon the level of

technology, the level of training of the soldiers, etc. We now
investigate the consequences of the model.

Clearly, there is a fixed point at the origin. The eigenvalue

equation (16) is
λ2 − αβ = 0

which gives
λ1 = − aβ, λ2 = + aβ. (80)

The eigenvector equations are then

√
αβ −α K1x
√ =0
−β αβ K1y

and √
− αβ −α K2x
√ = 0.
−β − αβ K2y

These equations give

√ √
α − α
K1 = c1 √ , K2 = c2 √ . (81)
β β

Since λ1 , λ2 are of opposite signs, the fixed point is a sad-

dle point. We can therefore draw the following phase plane
diagram:
 55 MAT217—X/1
APM214—Y/1

Of course, there cannot be a negative number of soldiers, etc.,

so, on physical grounds, x(t) ≥ 0 and y(t) ≥ 0. Thus the only
part of the phase plane diagram that is relevant is the positive
quadrant:

In the sketch above you will see a dividing line between those
solutions which, as time increases, lead to a total elimination
 56

of x (above the line) and those which lead to a total elimina-

tion of y (below the line). Of course, if you start anywhere on
the line, both x and y will eventually be eliminated. Whether
you start on, above or below the line clearly depend on the
initial conditions x (0), y (0), so in order to derive the rela-
tionship between x (0) and y (0) which determines which force
will be eliminated, we calculate, for the solution
√ √
√ √
α − α
X (t) = c1 √ e− αβt + c2 √ e αβt ,
β β

the values of the constants c1 and c2 in terms of the initial

conditions x (0), y (0):
√
x (0) α (c1 − c2 )
= √ ,
y (0) β (c1 + c2 )

yielding
√ √
x (0) β + y (0) α
c1 = √
2 αβ
√ √
y (0) α − x (0) β
c2 = √ .
2 αβ

Since we want the relationship between x (0) and y (0) when

both x and y are eliminated as time increases, we solve

x (t) 0
lim = ,
t→∞ y (t) 0

that is
√ √
√ √
α − αβt − α αβt 0
c1 √ lim e + c2 √ lim e = .
β t→∞ β t→∞ 0
 57 MAT217—X/1
APM214—Y/1

This clearly gives

c2 = 0

yielding the desired relationship

y (0) = x (0) β/α.

We can therefore clearly summarize the outcome of the battle

in terms of the initial conditions as:

β
(a) If y(0) > x(0) α then x is totally eliminated.
β
(b) If y(0) < x(0) α then y is totally eliminated.
(c) In the case that the two sides are equally matched, y(0) =
x(0) β/α, then the two sides eliminate each other.

(ii) The cohort model of age distribution in a population

We investigate the growth of a population that is broken into
age groups. We let x1 (n), x2 (n), ..., xm (n) represent the num-
ber in the m age groups at the beginning of time period n.
Depending on the species, the time period could be anything
from a decade (perhaps appropriate for humans) to a day (e.g.
certain insects). Next, we introduce the birthrates, so that

x1 (n + 1) = b1 x1 (n) + b2 x2 (n) + ... + bm xm (n). (82)

The idea is that only the first cohort is increased due to birth,
and the birthrate depends very much on the age of the parents.
The other cohorts are not affected by birth but by the survival
rate from the previous cohort:

xi (n + 1) = Si−1 xi−1 (n). (i = 2, ..., m). (83)

 58

Of course 0 < Si < 1. We can then write the above equations

in matrix form
 
b1 b2 · · · · · · ··· · · · · · · bm
 
 S1 0 · · · · · · ··· 0 
 .. 
 
 0 S2 0 . 

x (n + 1) =  . .. 
. .. ..  x (n) .
 . 0 . . . 
 
 .. .. .. .. .. 
 . . . . . 
0 0 0 Sm−1 0
(84)
This model is often used to answer the question: What is
the age distribution of the population? To give the answer,
we need to know the dominant eigenvalue and correspond-
ing eigenvector. The term “dominant eigenvalue” means the
eigenvalue of greatest absolute magnitude. It can be shown
[Luenberger, 170—173], under reasonable assumptions about
the birthrates bi , that the dominant eigenvalue λ1 is positive,
real and simple (i.e. it is not a repeated root). Then if K1 is
the corresponding eigenvector, it is also real. Thus the solu-
tion is of the form
m
x(n) = c1 K1 λn1 + ci Ki λni
i=2
m n
λi
= λn1 c1 K1 + ci Ki . (85)
λ1
i=2

m
Now because λ1 > |λi | then as n → ∞, the i=2 term tends
to zero. Thus, as n → ∞,

x(n) → λn1 c1 K1 . (86)

 59 MAT217—X/1
APM214—Y/1

Thus, for large n, the population distribution amongst the

various age cohorts is given by K1 . Note that technically this
conclusion assumes that the initial population is such that
c1 = 0. However, even if c1 = 0 in the initial population, the
situation is not exactly described by equation (84), or indeed
by any other equation. In practice there is a certain amount of
randomness in the evolution of populations. This randomness
would ensure that c1 would become non—zero, and so equation
(86) still applies.

Example
Consider the fictional species S. Suppose that, measuring n
in years, the population can be divided into three age groups:
x1 , x2 , x3 . Let the birth rates be b1 = 0.5, b2 = 5, b3 = 3, and
let the survival rates be S1 = 0.5, S2 = 32 . Then equation
(84) is
 
0, 5 5 3
 
x(n + 1) =  0, 5 0 0  x(n). (85)
0 2/3 0

Then equation (16) is

2
(0.5 − λ)(−λ)2 − 5(0.5 × (−λ)) + 3 × 0.5 × = 0.
3

Therefore
1 1
−λ3 + λ2 + 2 λ + 1 = 0,
2 2
and hence
1
−(λ + )(λ + 1)(λ − 2) = 0. (86)
2
 60

The dominant eigenvalue is therefore λ1 = 2, and the cor-

responding eigenvector is found from
 
−1 12 5 3
 1 
 2 −2 0  K1 = 0. (87)
0 2/3 −2

Let  
K11
 
K1 =  K12  .
K13
Suppose that K13 = 1, then the last row of (87)
2
K1 − 2K13 = 0,
3 2
gives
K12 = 3,

so that from the middle row

1
K1 − 2K13 = 0,
2 2
we have
K11 = 12.

Thus  
12
 
K1 =  3  . (88)
1
and, for large n,    
x1 12
   
 x2  ∝  3  . (89)
x3 1
 61 MAT217—X/1
APM214—Y/1

(iii) Safety of buildings

Mechanical structures (bridges, buildings, etc.) are built from
materials having some elasticity, and are therefore dynamical
systems. To determine their response to forces such as wind,
earthquakes, etc. it is important to calculate the natural fre-
quencies of vibration. Consider a four storey building, as illus-
trated below, and assume that the elements of the frame are
inextensible and that the mass is concentrated in the floors.
The floors can be displaced laterally with respect to each other,
but the bending elasticity of the building frame then generates
restoring forces. Assuming that all frame members have the
same elasticity, the force vector f is related to the displace-
ment vector x by f = Ax, where A is the stiffness matrix. By
Newton’s laws, the force also satisfies −f = Mẍ(t), where M
is the mass matrix. Thus

ẍ(t) = −M−1 Ax = −Bx (90)

where
 
1 −1 0
 
A = a  −1 2 −1  ,
0 −1 2
 
1 0 0
 
M = m 0 1 0 . (91)
0 0 2

(m and B = M−1 A are constants associated with the material

of the building; m is mass and a is a coefficient of elasticity).
 62

First, equation (90) is not in the usual form, because it in-

volves derivatives of second—order (ẍ) rather than first order
(ẋ). There are two different, but ultimately equivalent, ways
of handling equation (90).

(a) We look for solutions of the form

x = Ke±iωt . (92)

Equation (92) is inspired by the solution of ẍ + ω2 x = 0,

x = e±iωt . Substituting (92) into (90) gives

−ω2 Ke±iωt = −Be±iωt K.

Therefore
ω 2 K = BK. (93)

Thus ω 2 is an eigenvalue, and K the corresponding eigen-

vector, of the matrix B. Now B = M−1 A, thus
 
1 −1 0
a  
B=  −1 2 −1  ,
m
0 − 12 1
 63 MAT217—X/1
APM214—Y/1

so that the eigenvalue equation is

0 = det B − w2 I
a a
m − w2 −m 0
= a 2a 2 a
−m m −w −m
a a 2
0 − 2m m −w
a 2a a a2
= − w2 − w2 − w2 −
m m m 2m2
a a a
+ − − w2
m m m
a 3 1
= (1 − λ) (2 − λ) (1 − λ) − − (1 − λ)
m 2
a 3 1
= (1 − λ) λ2 − 3λ + 2 − − 1
m 2
a 3 1
= (1 − λ) λ2 − 3λ −
m 2

m 2
where λ = aw .
Hence we solve
1
(1 − λ) λ2 − 3λ − =0 (94)
2

to find the solutions

λ1 = 1,
√
λ2 = (3 + 7)/2,
√
λ3 = (3 − 7)/2,

so that
a 1/2
ω1 = ,
m
 64

a 1/2
ω 2 = 1.68 ,
m
a 1/2
ω3 = 0.42 . (95)
m
Thus the lowest value of ω is ω 3 , and therefore the lowest
natural frequency (frequency ν = ω/2π) is
a 1/2
ν = 0.067 . (96)
m
Problems occur when the lowest frequency is small, mean-
ing in practice when ν ≤ 10 (cycles/second). Thus one
requires
m 10 2
≥ = 22277. (97)
a 0.067
The ratio elastic coefficient: mass must be high. The
analysis described above is very important in Engineer-
ing design. It was not applied to the design of the To-
coma Narrows bridge in the USA, and in 1940 that bridge
shook itself to pieces during a storm.

(b) An alternative way of deriving (95) is to write (90) as a

system of first order equations. Let
       
y1 x1 y4 ẋ1
 .   .   .   . 
 ..  =  ..  ,  ..  =  ..  . (98)
       
y3 x3 y6 x3

Then the system (90) is ẏ = Cy where

0 I
C= . (99)
−B 0
 65 MAT217—X/1
APM214—Y/1

It can be shown that, provided B is invertible (which it

is in this case), then the eigenvalues of C are

±i λ1 , ±i λ2 . ± i λ3 (100)

where λ1 , λ2 , λ3 are the eigenvalues of B. Thus the solu-

tion (92), with ω given by (95), is also found.

EXERCISES

1. For the following systems find the location of the singular

point, and its nature.
dx dy
(a) = x + 2y + 1, = 4x + 2y − 1.
dt dt
dx dy
(b) = 5x + y − 1, = −2x + 3y − 2.
dt dt
dx dy
(c) = −2x + y + 4, = −2x + 4y + 6.
dt dt

−1 2
(d) X(n + 1) = X(n).
−7 4

1 2 −4
(e) X(n + 1) = 1 1
X(n) + .
2 2 −1

2. Find the general solution of the following systems, and state

the location and nature of the singular point.
dx dy
(a) = 3x + 3y − 4, = −x − 3y + 5.
dt dt
 66

dX −2 3
(b) = X.
dt −3 5
dx dy
(c) = 6x − y − 5, = 5x + 2y − 3.
dt  dt 
−1 −1 0
dX  3 
(d) =  4 − 32 3  X.
dt 1 1
8 4 − 12

3. For the following systems, find the solution that satisfies the
given initial condition; state the location and nature of the
singular point.

dX −1 −2 3 3
(a) = X+ subject to X(0) = .
dt 3 4 3 4

dX 4 −5 6
(b) = X subject to X(0) = .
dt 5 −4 2
dx
(c) = −x + y, with x(0) = 2.
dt
dy
= x + 2y + z − 5, with y(0) = 3.
dt
dz
= 3y − z − 1, with z(0) = 4.
dt

4. Find the general solution of the following discrete dynamical

systems, and then find the particular solution for the given
initial condition. State the nature of the singular point.

−2 12 2
(a) X(n + 1) = X(n), with X(0) = .
5 2
 67 MAT217—X/1
APM214—Y/1

2 −5 5
(b) X(n + 1) = X(n), with X(0) = .
1 −2 1

(c) x(n + 1) = − 12 x(n) + 12 y(n),

y(n + 1) = − 12 x(n) + 34 y(n) with x(0) = y(0) = 3.

13 25 1
(d) X(n + 1) = X(n) +
−9 −17 0
2
with X(0) = .
2

5. Make the following simplifications in the cohort model of age

distribution:

— women have children between the ages of 13 and 38 in-

clusive;
— each woman has exactly one female child;
— each woman lives to be at least 39 years old.

Using a time interval of 13 years with these assumptions, and

considering only those females less than 39 years old gives the
following model:
    
x1 (k + 1) 0 α 1−α x1 (k)
    
 x2 (k + 1)  =  1 0 0   x2 (k) 
x3 (k + 1) 0 1 0 x3 (k)

where 0 ≤ α ≤ 1.

(a) Find the eigenvalues. Find the dominant eigenvalue and

the corresponding eigenvector.
 68

(b) Determine the growth rate.

(c) How quickly does the population move to equilibrium?

[The answer depends on the value of α.]

6 Assume that the population of a country is divided into two

distinct segments: rural and urban. The natural yearly growth
factors, due to procreation, in both segments are assumed to
be identical and equal to α (that is, the population at year
k + 1 would be α times the population at year k). The popu-
lation distribution, however, is modified by migration between
the rural and urban segments. The rate of this migration is
influenced by the need for a base of rural activity that is ad-
equate to support the total population of the country — the
optimal rural base being a given fraction γ of the total popu-
lation. The yearly level of migration itself, from rural to urban
areas, is proportional to the excess of rural population over the
optimal rural base.

If the rural and urban populations at year k are denoted by

r(k) and u(k), respectively, then the total population is r(k) +
u(k), the optimal rural base is γ[r(k) + u(k)], and thus, the
excess rural population is {r(k) − γ[r(k) + u(k)]}. A simple
dynamic model of the migration process, based on the above
assumptions, is then

r(k + 1) = αr(k) − β{r(k) − γ[r(k) + u(k)]}

u(k + 1) = αu(k) + β{r(k) − γ[r(k) + u(k)]}.
 69 MAT217—X/1
APM214—Y/1

In this model, the growth factor α is positive (and usually

greater than unity). The migration factor β is positive, and
is assumed to be less than α. The parameter γ is the ideal
fraction of the total population that would be rural in order to
support the total population. This parameter is a measure of
rural productivity. Each of these parameters might normally
change with time, but they are assumed constant for purposes
of this example.

The model can be easily put in the state vector form

x(k + 1) = Ax(k)

where
α − β (1 − γ) βγ
A=
β (1 − γ) α − βγ

and
r(k)
x(k) = .
u(k)

Show that the condition

α α
0 ≤ β ≤ min ,
1−γ γ

is necessary and sufficient in order to guarantee that both

urban and rural populations remain nonnegative given any
nonnegative initial conditions.
 70

7. Find the general solution of the following systems, and state

the location and nature of the singular point.

7 −1
(a) X (n + 1) = X (n)
5 3
6 1 1
(b) X (n + 1) = X (n) −
−2 4 2
5 −5
(c) X (n + 1) = X (n)
5 −3
(d) x (n + 1) = 4x(n − y (n) − 2
y (n + 1) = 9x (n) − 2y (n) − 6
(e) x (n + 1) = 3y (n) − 4
y (n + 1) = −3x (n) + 6y (n) − 4
(f) x (n + 1) = −5x (n) + 5y (n)
y (n + 1) = −5x (n) + 5y (n)

8. Find the nature of the singular point in the following systems.

 
2 −1 2
 
(a) X (n + 1) =  −1 2 0  X (n)
−1 0 2
 
3 5 1
 
(b) X (n + 1) =  −5 −5 4  X (n)
0 0 3
 
6 −4 0
 
(c) X (n + 1) =  1 1 2  X (n)
0 2 6
 71 MAT217—X/1
APM214—Y/1
 
2 0 0
 
(d) X (n + 1) =  2 3 −1  X (n)
0 1 1

9. Find the general solution of the following systems, and state

the location and nature of the singular point.

dx dy dz
(a) =z = −z =y
dt dt dt
   
−1 1 0 0
dX    
(b) = 1 2 1  X+  −4 
dt
0 3 −1 −2
 
1 1 4
dX  
(c) =  0 2 0 X
dt
1 1 1

10. Write the following as first—order systems and then determine

the nature of the singular point.

(a) x (n + 2) = 3x (n + 1) − 2x (n)

(b) x (n + 2) = 14 x (n)
d2 x
(c) dt2 + 4 dx
dt + 3 = 0

d2 x
(d) dt2 − 5 dx
dt + 6 = 0
 72

CHAPTER 3

LINEAR CONTROL THEORY

1. Introduction
It was not possible to change the state of the dynamical systems we
studied in Chapter 2 by means of an “outside” control — the changes
in state were entirely determined by the equations describing the dy-
namical system and the initial conditions. In many practical situa-
tions we do have some control over the system (for example, turning
a tap on or off to control the level of water in a tank, or a heater
on or off to control the temperature in a room). In this chapter we
consider some aspects of linear dynamical systems subject to linear
controls.

Mathematically such systems are described by the equation

ẋ(t) = Ax(t) + Bu(t)

for continuous systems, and

x(k + 1) = Ax(k) + Bu(k)

in the discrete case. If there are n state variables and m control

variables then, in both cases, x is an n × 1 column vector, u is an
 73 MAT217—X/1
APM214—Y/1

m × 1 column vector, A is an n × n matrix with constant entries,

and B is an n × m matrix with constant entries. We will call the
vector x the state vector of the system, or simply the state of the
system, and the vector u the control vector, or sometimes the input
vector to the system. In many cases some kind of measurement is
performed on the system, the results being given in the form of an
output vector y. We assume that y depends linearly on the state
vector x and input vector u. The equation for y for continuous
systems will then be

y(t) = Cx(t) + Du(t)

where y is an p×1 column vector, C is an p×n matrix with constant

entries, and D is an p × m matrix with constant entries.

For discrete systems the output equation is

y(k) = Cx(k) + Du(k).

Our aim in this chapter is to study two important properties of the

above control systems:

(i) Controllability
Here we want to know whether we can always choose an input u to
change the state vector x from any initial value to any final value.
Roughly, how well we can control the system. This is dealt with in
sections 3 and 4.

(ii) Observability
The question here is “If we know the output vector y and the input
u can we determine the state vector x?” In other words, can we use
 74

the output and input to observe the state of the system? Sections
5 and 6 are concerned with this question.

Lastly, in section 7 we investigate what happens when we connect

the output of the system in some way to the input. This is called
feedback.

First, however, we will give some examples of linear control systems

and show how to represent such systems by certain diagrams. This
visual aid will make it easier to understand and discuss the control
system concepts described above.

2. Dynamic Diagrams
These diagrams are built up from five elementary components shown
in figure 1:

Figure 1

The summer (a) adds whatever comes into it, instantaneously pro-
 75 MAT217—X/1
APM214—Y/1

ducing the sum. Any number of lines can come into a summer and
one line comes out. The transmission (b) multiplies what comes in
by the number written in the box. Diagram (c) is a splitter — it
connects an incoming line with any number of outgoing lines, each
having the value of the incoming line. The delay (d) is the basic
dynamic component of discrete—time systems. If x(k) comes in to
a delay diagram, then the value of x one time unit earlier goes out
— that is, x(k − 1) goes out. So if x(k + 1) goes in, x(k) goes out,
and if x(k − 1) goes in x(k − 2) goes out, and so on. The integrator
(e) is the basic dynamic component of continuous—time systems. If
ẋ goes in to the integrator then x goes out. The way in which these
diagrams are put together to form linear control systems is best de-
scribed by examples.

Example 1 (Pure Integrator)

The equation for this simple system is

ẋ(t) = u(t).

The output x(t) is the integral of the input function u(t). The dia-
gram is therefore:

Figure 2
 76

Example 2
Consider the control system
ẋ1 3 0 x1 1
= + u(t)
ẋ2 1 2 x2 0

y (t) = x2 (t) .

Then ẋ1 = 3x1 + u. In that part of the diagram corresponding

to this equation we need an integrator for ẋ1 , a transmission box
for 3x1 and a summer for the “ + ”. The equation tells us these
components are connected as follows:

Figure 3
Similarly the diagram corresponding to the equation

ẋ2 = x1 + 2x2 is

Figure 4
 77 MAT217—X/1
APM214—Y/1

Putting these two diagrams together and noting that y = x2 gives

the following diagram for the whole system:

Figure 5
Example 3
The diagram for the one—dimensional discrete system

x(k + 1) = ax(k) + bu(k)

is

Figure 6
Example 4 (A very simple model for driving a car)
Suppose a car is driven along a straight road, its distance from an
initial point O being s(t) at time t. Assume the car is controlled
by the accelerator, providing a force of u1 (t) per unit mass, and the
 78

brake, which produces a retarding force of u2 (t) per unit mass. We

are interested in the car’s position given by

x1 (t) = s(t)

and speed
x2 (t) = ṡ(t).

Then
ẋ1 = x2

and by Newton’s 2nd law

ẋ2 = u1 − u2 .

In matrix notation:

ẋ1 0 1 x1 0 0 u1
= + .
ẋ2 0 0 x2 1 −1 u2

The dynamic diagram for this system is:

Figure 7
 79 MAT217—X/1
APM214—Y/1

Example 5 (Stick Balancing)

Balancing a stick on one’s hand can be formulated as a control prob-
lem as follows:

Suppose the stick has length L and the mass M of the stick is
concentrated at the top. Let the variables u, θ, and x be as shown
in the figure below:

Figure 8

It follows from Newton’s laws that the system is governed by the

equation
ü(t) cos θ(t) + Lθ̈(t) = g sin θ(t)

where g is the gravitational constant.

We also have
x(t) = u(t) + L sin θ(t).
 80

If we assume that the stick is nearly at rest close to the vertical

2
position, then we can put θ̇ = 0, cos θ = 1 and sin θ = θ, and the
above equations can be written
g
ẍ(t) = g[x(t) − u(t)].
L
Put
v(t) = ẋ(t),

then the matrix equation for the system is

ẋ (t) 0 1 x (t) 0
= g + g u(t).
v̇ (t) 0 v (t) −
L L
The dynamic diagram for the system is

Figure 9

Example 6 (Pendulum system)

Consider the pendulum system shown in the figure below:
 81 MAT217—X/1
APM214—Y/1

Figure 10

We take the angle θ to be the output of the system. The equation

for the system is
mL2 θ̈ = −mgL sin θ

or
g
θ̈ + sin θ = 0.
L
Define

x1 = θ
x2 = θ̇.

Then we get

ẋ1 = x2
g
ẋ2 = − sin x1 .
L
The output y is the angle θ so

y = x1 .
 82

If the angle θ is small we can put sin x1 = x1 and the matrix equa-
tions for the system are then

ẋ1 0 1 x1
= g
ẋ2 − 0 x2
L
x1
y = 1 0 .
x2

The dynamic diagram is

Figure 11

The system has no input.

Example 7
Consider a system consisting of an interconnected pair of water
tanks, as shown in figure 12:
 83 MAT217—X/1
APM214—Y/1

Figure 12

A1 and A2 represent the cross—sectional areas of the two tanks. The

single pump motor pumps water at a controlled rate of u litres/sec
into each tank. The outflow is proportional to the depth of water in
each tank (constant of proportionality µ). The flow in the intercon-
necting pipe is proportional to the difference of depths in the tanks
(constant of proportionality λ).

Let V1 , V2 denote the volume of water in each tank. Then

dV1 V2 V1 V1
= u − y1 + λ − , y1 = µ
dt A2 A1 A1
dV2 V1 V2 V2
= u − y2 + λ − , y2 = µ .
dt A1 A2 A2

Thus
dV1 V1 V2
= −(λ + µ) +λ + u,
dt A1 A2
dV2 V1 V2
= λ − (λ + µ) + u.
dt A1 A2
 84

The final combined outflow is

V1 V2
y=µ + .
A1 A2

The system equations in matrix form are therefore

 
λ+µ λ
 − A1 A2 
V̇1   V1 1
= 


 + u
V̇2  λ λ+µ  V2 1
−
A1 A2
µ µ V1
y = .
A1 A2 V2

We will return to some of these examples later.

3. Controllability for Discrete—Time Systems

We discussed the idea of controllability very briefly in the introduc-
tion. In this section we give a precise definition for discrete systems
and show how to determine if such a system is controllable.

Definition
The n—dimensional system

x(k + 1) = Ax(k) + Bu(k)

is said to be completely controllable if for x(0) = 0 and any given

vector w there exists a positive integer N and a sequence of inputs
u(0), u(1), u(2), ..., u(N − 1) such that this input sequence, applied
to the system, yields x(N ) = w.
 85 MAT217—X/1
APM214—Y/1

Thus if the system is completely controllable we can “steer” the zero

vector to any other vector we choose in a finite number of steps. The
choice of the zero vector as our initial vector in this definition is only
for convenience. We will show later that it follows that any vector
can be steered to any other in a finite number of steps — in fact in
n or fewer steps.

Here are some examples to illustrate this definition.

Example 8
Consider the system with dynamic diagram

Figure 13

The equations for this system are

x1 (k + 1) = ax1 (k)
x2 (k + 1) = x1 (k) + bx2 (k) + u(k).

It is clear from both the diagram and equations that the control u
cannot have any effect on the state variable x1 . Hence this system
 86

is obviously not completely controllable.

Example 9
Consider the one—dimensional discrete system with equation

x(k + 1) = x(k) + u(k).

To steer x(0) = 0 to any given number b in one step take N = 1 and

choose u(0) = b. Then

x(1) = x(0) + u(0) = b.

Hence this system is, by definition, completely controllable.

Of course, it is seldom as easy as it was in the above two exam-

ples to decide if a system is completely controllable. However, in
the next theorem we derive a fairly simple condition on the n × n
matrix A and the n × m matrix B that is equivalent to complete
controllability. You might need to revise the concept of the rank of a
matrix, which is discussed in the appendix, before working through
this theorem. We will also need the following lemma. We omit the
proof — if you are interested, the proof is on page 278 of Luenberger.

Lemma 1
Suppose A is an n × n matrix and B is an n × m matrix. Then, for
any integer N ≥ n ≥ 1, the rank of the matrix

[B, AB, ..., AN−1 B]

is equal to the rank of the matrix

[B, AB, ..., An−1 B].

 87 MAT217—X/1
APM214—Y/1

The square bracket [...] notation used here is explained in the ap-
pendix.

Theorem 1
The discrete—time system

x(k + 1) = Ax(k) + Bu(k)

is completely controllable if and only if the n × m controllability

matrix

M =[B, AB, A2 B, A3 B, ..., An−1 B]

has rank n.

(As usual, A is an n × n matrix and B is an n × m matrix.)

Proof
Suppose a sequence of inputs u(0), u(1), u(2), ..., u(N −1) is applied
to the system
x(k + 1) = Ax(k) + Bu(k)

with x(0) = 0. Then

x(1) = Ax(0) + Bu(0) = Bu(0),

x(2) = Ax(1) + Bu(1) = ABu(0) + Bu(1),
x(3) = Ax(2) + Bu(2) = A2 Bu(0) + ABu(1) + Bu(2),

and so on, giving finally

x(N ) = AN−1 Bu(0) + AN−2 Bu(1) + ... + Bu(N − 1).

 88

The expression on the right is simply a linear combination of the

columns of the matrix

K = B, AB, A2 B, ..., AN−1 B .

Thus all points of the n−dimensional state space can be reached if

and only if K has rank n. By lemma 1

rank K = rank M,

for N ≥ n, and hence the theorem is proved.

Remark
It follows from the above proof that N ≤ n. Thus we can reach any
vector from the zero vector in n or fewer steps if the system is com-
pletely controllable. In fact we can transfer the state between two
arbitrary vectors within n steps. To see this, suppose x(0) and x(n)
are specified arbitrarily. With zero input the system would move to
An x(0) at time n. Then the desired input sequence is the one that
would transfer the state from the zero vector to x(n) − An x(0) at
time n.

The following examples show how to apply this important theorem.

Example 10
Consider the control system in example 8. We have
a 0
A=
1 b
and
0
B= .
1
 89 MAT217—X/1
APM214—Y/1

Since n = 2, we have that the controllability matrix M is given by

M = [B, AB] which in this case is

0 0
M= .
1 b

The columns of M are linearly dependent, and therefore

rank M = 1 < 2.

Hence, as we concluded earlier, this system is not completely con-

trollable.

Example 11
Suppose we modify the system in example 10 by shifting the input
to the first stage rather than the second. Then we get the dynamic
diagram

Figure 14

and the equations for this modified system are

x1 (k + 1) = ax1 (k) + u(k)

x2 (k + 1) = x1 (k) + bx2 (k).
 90

Then
a 0
A=
1 b
and
1
B= .
0
The controllability matrix M is therefore
1 a
M = [B, AB] = .
0 1
Therefore rank M = 2, and hence the system is completely control-
lable.
Example 12
Consider the simple one—dimensional system in example 9. Then
A = 1 and B = 1. Hence

M = [B] = [1] .
Clearly rank M = 1, and therefore the system is completely con-
trollable.
Example 13
We investigate the discrete system with scalar control u and equa-
tion
x(k + 1) = Ax(k) + Bu(k) (1)
where
−2 2
A = ,
1 −1
1
B =
0
 91 MAT217—X/1
APM214—Y/1

and n = 2. Then
1 −2
M = [B, AB] = .
0 1
Clearly rank M = 2 = n, and so the system is completely control-
lable. Let us find a control sequence which transfers the zero vector
to the vector
1
.
2
We know we can do this in at most two steps. Let u(0), u(1) be a
control sequence such that

0
x(0) =
0
and
1
x(2) = .
2
Then from equation (1) it follows that

x(1) = Ax (0) + Bu(0)

x (2) = Ax (1) + Bu (1)
= A [Ax (0) + Bu (0)] + Bu (1)
= A2 x (0) + ABu (0) + Bu (1)
= ABu (0) + Bu (1)

since x (0) = 0.

This gives
1 −2 1
= u(0) + u(1).
2 1 0
 92

Therefore

u(1) − 2u(0) = 1
u(0) = 2.

Solving these equations gives u(0) = 2; u(1) = 5.

Is it possible to do the transfer in one step? We would require from

x (1) = Ax (0) + Bu (0) = Bu (0)

(since x (0) = 0) that

1 1
= u(0).
2 0
This is obviously not possible, and so the transfer cannot be accom-
plished in one step.

4. Controllability of Continuous— Time Systems

The definition of complete controllability for these systems is anal-
ogous to our definition for discrete systems:

Definition
The system
ẋ(t) = Ax(t) + Bu(t)
is said to be completely controllable if for x(0) = 0 and any given
state x1 there exists a finite number t1 and a piecewise continuous
input u(t), 0 ≤ t ≤ t1 , such that x(t1 ) = x1 .

(Piecewise continuous means that there are at most a finite number

of points of discontinuity of u in the interval [0, t1 ].)
 93 MAT217—X/1
APM214—Y/1

As with the discrete case there is a simple test for complete con-
trollability. The test is exactly the same as for the discrete case
but the proof is different. Some of the mathematical results used
will most likely be new to you, so you should study sections 3—5 of
the appendix before working through the next theorem, whose proof
uses the following lemma:

Lemma 2
Suppose A is a constant n × n matrix, B is a constant n × m matrix
and
rank B, ABA2 B, . . . , An−1 B = n.

Then for any t1 > 0 the matrix

t1
T
K= e−At BBT e−A t dt
0

has an inverse K−1 .

(Remember BT denotes the transpose of B.)

We omit the proof — to give it would be too much of a digression.

Theorem 2
The n dimensional continuous—time system

ẋ(t) = Ax(t) + Bu(t)

is completely controllable if and only if the controllability matrix

M = [B, AB, A2 B, ..., An−1 B]

 94

has rank n.
Proof
We first show that if rank M < n, then the system is not completely
controllable. From the results given in paragraph 5 in the appendix,
we get for any t1 and any input function u(t), 0 ≤ t ≤ t1 ,
t1
x(t1 ) = eA(t1 −t) Bu(t)dt (1)
0

Then, using the definition of eA(t1 −t) ,

t1
A2
x(t1 ) = I + A(t1 − t) + (t1 − t)2 ... Bu(t)dt
0 2!
or
t1 t1
x(t1 ) = B u(t)dt + AB (t1 − t)u(t)dt
0 0
t1
(t1 − t)2
+A2 B u (t) dt + ... .
0 2!
When evaluated, the integrals in the above expression are simply
constant m—dimensional column vectors. Therefore, the above ex-
pression shows that x(t1 ) is a linear combination of the columns of
B, AB A2 B, ... . By Lemma 1, if rank M < n, then even the
infinite set of vectors formed from the columns of B, AB, A2 B, ...
does not contain n linearly independent vectors. Therefore, there is
a vector x1 ∈ Rn that cannot be expressed as a linear combination
of any of these vectors, and so cannot be reached by any choice of
control.
Next we show that if rank M = n then the system is completely
controllable.
 95 MAT217—X/1
APM214—Y/1

We do this by explicitly constructing a control u which transfers the

zero vector to a given vector x1 . By Lemma 2 for any t1 > 0 the
matrix
t1
T
K= e−At BBT e−A t dt
0
has an inverse K−1 . Now, given x1 , select any t1 and let
T
u(t) = BT e−A t K−1 e−At1 x1 .

We verify that this input transfers the state from zero to x1 at time
t1 by substituting for u in equation (1):
t1
T
x(t1 ) = eA(t1 −t) BBT e−A t K−1 e−At1 x1 dt
0
= e At1
KK−1 e−At1 x1 = x1 .

Since x1 was chosen arbitrarily it follows from the definition that

the system is completely controllable.

Note that in the above proof we select any t1 > 0, so in fact the
state can be transferred to x1 in an arbitrarily short period of time.

All the examples used to illustrate controllability in discrete systems

can also be used for continuous systems — simply replace x(k + 1) by
ẋ in the equations and the delay symbol by the integrator symbol
in the dynamic diagrams. Let us explicitly calculate the control in
the continuous—time version of example 12.

Example 14
The system is 1—dimensional with equation

ẋ = x + u.
 96

As before, A = B = 1 and the controllability matrix M is

M = [1].

Since rank M = 1 this system is completely controllable. Let us

find a control u which transfers 0 to 1 at time t = 1. The matrix K
in this example is
t1 1
T 1 − e−2
K= e−At BB T e−A t dt = e−t (1)(1)e−t dt = .
0 0 2
Therefore
2
K −1 = .
1 − e−2
Hence
T 2e−(1+t)
u(t) = B T e−A t K −1 e−At1 x1 = .
1 − e−2
We can check that this control does work. We have x(0) = 0 and

dx 2e−(1+t)
=x+u=x+ .
dt 1 − e−2
Solving this 1st order linear differential equation for x(t) gives

e et − e−t
x(t) =
e2 − 1
and we see x(1) = 1.

We can now apply Theorem 2 to example 5 on stick balancing. We

know that it is possible to balance a stick on one’s hand so the
system should be completely controllable — let’s verify this.
 97 MAT217—X/1
APM214—Y/1

Example 15
The matrix form of the system equations is

ẋ 0 1 x 0
= g + g u.
v̇ 0 v −
L L
Therefore
0 1
A= g
0
L
and
0
B= g .
−
L
The controllability matrix M is then
 g 
0 −
M= g L .
− 0
L
Since rank M = 2 the system is controllable as we expected. But
you may be surprised by the following extension of the problem: Can
two sticks (with masses concentrated at their tops) placed side by
side a little distance apart on one hand be simultaneously balanced?
Let the suffixes 1 and 2 distinguish the various variables for the two
sticks. The same control is applied to both sticks, and there is no
interference between the sticks, so the system equations will be the
same as for one stick but repeated twice:

ẋ1 = v1
g
v̇1 = (x1 − u)
L1
ẋ2 = v2
g
v̇2 = (x2 − u).
L2
 98

In matrix form:
      
ẋ1 0 1 0 0 x1 0
   g    
 v̇1   L
   1 0 0 0 
 v1  
+ − Lg1 
 u.
 ẋ  =  0 0 0 1  x2   0 
 2      
g
v̇2 0 0 L2 0 v2 − Lg2

Calculating the controllability matrix M in the usual way gives

 2 
0 − Lg1 0 − Lg 2
 g2
1 
 −g 0 − 0 
 L1 L12 
M= g g2 .
 0 − 0 − 
 L 2 L 2
2 
g g2
− L2 0 − L2 0
2

If L1 = L2 then rank M = 4 and the system is completely con-

trollable, ie the sticks can be simultaneously balanced. However, if
L1 = L2 then rank M = 2, so the system is not completely control-
lable and the sticks cannot be balanced. It is not surprising that
two sticks of equal length cannot be balanced, since if they could it
would imply (because the sticks move independently) that the con-
trol used to balance a stick did not depend on the state of the stick!
I do find it quite surprising that two sticks of unequal length can be
simultaneously balanced — in fact if any finite number of sticks, no
two of equal length, are placed side by side on a (big enough) hand,
in theory they can all be balanced simultaneously. This is because
the form of the controllability matrix remains the same as that of
the above M, its dimension just gets larger. Note that the mass of
a stick plays no role in the analysis presented here.
 99 MAT217—X/1
APM214—Y/1

Example 16
Consider the water tanks system in example 7. Let us investigate
under what conditions this system is completely controllable. In
this case
1
B=
1
and  
λ+µ λ
 − A1 + A2 
 
AB = 

.

 λ λ+µ 
−
A1 A2
Hence the controllability matrix for this system is
 
λ+µ λ
 1 − A1 + A2 
 
M = [B, AB] =  
.

 λ λ+µ 
1 −
A1 A2

It is important to note that the values A1 and A2 are unknowns so

that the rank of the controllability matrix will clearly depend on
these values. Another thing to remember (see paragraph 2 of the
appendix):

If M is a n × n matrix then rank M =n if and only if det M = 0.

For the above matrix we have

λ λ+µ λ+µ λ
det M = − + −
A1 A2 A1 A2
(2λ + µ) (A2 − A1 )
= .
A1 A2
 100

Since λ and µ are positive, it follows that

det M = 0 ⇔ A1 = A2 .

Therefore the rank of M is 2 if and only if A1 = A2 .

Thus the system is completely controllable except when the tanks

have the same cross—sectional areas. For example, if A1 = A2 , and
the tanks are both empty to start with, then both tanks can be filled
to equal levels, even if one is much larger than the other.

5. Observability for Discrete—Time Systems

The definition and results for observability are very similar to those
for controllability. The definition is

Definition
The discrete—time system

x(k + 1) = Ax(k) + Bu(k)

y(k) = Cx(k) + Du(k)

is completely observable if there is a positive integer N such that

knowledge of the outputs y(0), y(1), y(2), ..., y(N − 1) when all in-
puts u(k) are zero is sufficient to determine the value of the initial
state x(0).

(As usual, A is n × n, B is n × m, C is p × n, and D is p × m.)

Note that once the initial state x(0) is known, all subsequent states
can be determined.
 101 MAT217—X/1
APM214—Y/1

Observability is important because inputs are often determined by

observation of outputs. If the outputs do not give full information
about the state vector, then it may not be possible to formulate a
suitable control scheme. Thus in general we should have the ability
to detect what the system is doing (observability) and to change the
system behaviour (controllability).
The following two examples serve to illustrate the above definition.

Example 17
Consider the system with dynamic diagram shown in figure 15 be-
low:

Figure 15

The system has the single output y which is not connected at all to
x1 , and therefore there is no way to infer the value of x1 from y. So
this system is not completely observable.

Example 18
Suppose we have a system with the same dynamic diagram as in
the previous example, except that the single output is taken from
x1 instead of x2 . The system is now completely observable. This
 102

is easily shown as follows: Suppose we know y(0) and y(1), that is,
x1 (0) and x1 (1). From Figure 15 we see that

x1 (1) = ax1 (0) + x2 (0).

But we know x1 (0) and x1 (1), so we can solve for x2 (0). Hence if we
know y(0) and y(1) we can determine x(0) and therefore the system
is completely observable.

The following theorem is analogous to Theorem 1. We will omit the

proof since it is similar to that of Theorem 1.

Theorem 3
The system

x(k + 1) = Ax(k) + Bu(k)

y(k) = Cx(k) + Du(k)

is completely observable if and only if the pn×n observability matrix

 
C
 
 CA 
 
 CA2 
S= 
 . 
 .. 
 
n−1
CA

has rank n.

Let us apply this theorem to examples 17 and 18:

 103 MAT217—X/1
APM214—Y/1

Example 19
The system equations for the dynamic diagram in figure 15 are:

x1 (k + 1) = ax1 (k) + x2 (k)

x2 (k + 1) = bx2 (k)
y(k) = x2 (k)

Then
a 1
A=
0 b
and
C= 0 1 .

For CA we get

a 1
CA = 0 1 = 0 b .
0 b

Therefore the observability matrix is

0 1
S= .
0 b

Since rank S = 1 the system is not completely observable.

Example 20
If we modify figure 15 as described in example 18 then the system
equations are

x1 (k + 1) = ax1 (k) + x2 (k)

x2 (k + 1) = bx2 (k)
y(k) = x1 (k).
 104

The matrix A is the same as in example 19 and for C we get

C= 1 0 .

Therefore
a 1
CA = 1 0 = a 1 ,
0 b
and the observability matrix S is

1 0
S= .
a 1

Hence rank S = 2 and the system is completely observable — as we

concluded earlier.

6. Observability for Continuous—Time Systems

For continuous—time systems the results are analogous to those for
discrete—time systems. We begin with the definition.

Definition
A system

ẋ = Ax(t) + Bu(t)
y(t) = Cx(t) + Du(t)

is completely observable if there is a t1 > 0 such that knowledge of

y(t), for all t, 0 ≤ t ≤ t1 , when u(t) = 0 for all t, is sufficient to
determine x(0).

The following theorem (we omit the proof) gives a condition for
complete observability for continuous systems:
 105 MAT217—X/1
APM214—Y/1

Theorem 4
The system in the above definition is completely observable if and
only if the observability matrix
 
C
 
 CA 
 
 CA2 
S= 
 . 
 .. 
 
CAn−1

has rank n.

Example 21
Under what conditions is the system of water tanks in example 7
completely observable? In this case
µ µ
C=
A1 A2
and the observability matrix is
 µ µ 
 A1 A2 
C  
= .
CA  µ (λ (A1 − A2 ) − µA2 ) µ (λ (A2 − A1 ) − µA1 ) 
A21 A2 A1 A22
The determinant of the observability matrix is

C µ2 (A2 − A1 ) (µ + 2λ)
det = .
CA A21 A22

Of course λ > 0 and µ > 0 so the system is completely observable

if and only if A1 = A2 .
 106

7. Feedback
Control systems can be classified into two types: open loop and
closed loop or feedback control systems. In an open loop system the
control input follows a pre—set pattern without taking account of the
output or state variables, whereas in a feedback control system the
control input is obtained by the output being fed back in some way
to the input. Simple illustrations of open and closed loop systems
are provided by robots at traffic intersections which change at fixed
intervals of time, and those which measure traffic flow in some way
and change accordingly. Two advantages of feedback control are:

(ii) Feedback control is often simpler than open loop control. For
example, one could design a complicated open loop control for
a traffic robot which varied the green and red cycle lengths
according to the expected traffic flow at various hours of the
day, but feedback control would be simpler and more effective.

(ii) Feedback control can rapidly adjust to changes in the state

of the system. This means that feedback control can improve
the stability of the system, and it is this aspect of feedback
that we will discuss in this section.

Recall from Chapter 2 that the dynamical system ẋ = Ax is stable

at a fixed point if and only if the real part of the eigenvalues of A
are negative. Now consider the control system
ẋ = Ax + Bu.
Suppose that we apply linear feedback to this system, which means
each control variable is a linear combination of the state variables.
 107 MAT217—X/1
APM214—Y/1

Then
u = Cx

where C is a constant m × n matrix. Substituting for u in the

system equation we get

ẋ = (A + BC)x.

Therefore if we can find a matrix C such that the eigenvalues of

A + BC have negative real part then the system with feedback will
be stable. The following theorem (whose proof we omit) shows that
if the system without feedback is completely controllable, then we
can find such a feedback matrix C.

Theorem 5 (Eigenvalue placement theorem)

Suppose the control system

ẋ = Ax + Bu

is completely controllable and λ1 , λ2 , ..., λn are n complex numbers

such that any which are not purely real occur in conjugate pairs.
Then there exists a real matrix C such that the eigenvalues of
A + BC are the complex numbers λ1 , λ2 , ..., λn . In particular, we
can choose C such that all the eigenvalues have negative real part.

A very simple example to illustrate this theorem is

Example 22
Consider the one—dimensional system

ẋ(t) = u(t).
 108

This is certainly completely controllable. In this case A = 0, B = 1,

and A + BC = C. The eigenvalue of C is C, and we can choose C
real and C < 0. Then
ẋ = Cx
has the solution
x(t) = keCt .
If C < 0 then eCt → 0 as t → ∞, so the system with feedback
u = Cx is stable.

Example 23 (Stick balancing again)

Since the stick balancing system in examples 5 and 15 is completely
controllable we must be able to find a feedback matrix C such that
the closed loop system is stable. Let’s verify this. The matrices A
and B are
0 1
A = g ,
L 0
0
B = .
− Lg
Suppose that
C= c1 c2 .
Then
0 1
A + BC = g .
L (1 − c1 ) − Lg c2
As explained in Chapter 2, page 30, the eigenvalues of A + BC are
found by solving the equation det((A + BC) − λI) = 0. This is a
quadratic equation:
g g
λ2 + c2 λ − (1 − c1 ) = 0.
L L
 109 MAT217—X/1
APM214—Y/1

We can choose the solutions of this equation to be whatever we like

by taking appropriate values for c1 and c2 . For example, to make
both roots −1, choose c1 = 1 + Lg and c2 = 2Lg . For this value of C
the closed loop system will be stable.

Exercises for Chapter 3

1. Investigate the controllability of the system

ẋ = Ax + Bu

if u(t) is a scalar, and

−5 1 1
(a) A = , B= .
0 4 1
   
0 1 0 0
   
(b) A =  0 0 1 , B =  0 .
0 0 0 1
   
3 3 6 0
   
(c) A =  1 1 2 , B =  0 .
2 2 4 1

2. Consider the system

ẋ = Ax + Bu
y = Cx

where
2 −5
A =
−4 0
1
B =
−1
 110

and
C= 1 1 .

Is this system

(i) completely controllable?

(ii) completely observable?

3. For the discrete—time system

x(k + 1) = Ax(k) + bu(k)

where
2 1
A=
0 2

let

0
b1 = ,
1
1
b2 = .
0

(a) For each b, determine if the system is controllable.

(b) For each b that results in a completely controllable sys-
tem, find the shortest input sequence that drives the state
to zero if
2
x(0) = .
1
 111 MAT217—X/1
APM214—Y/1

4. Show that feedback does not destroy complete controllability.

Specifically, show that if the system

x(k + 1) = Ax(k) + Bu(k)

is completely controllable, then the system

x(k + 1) = Ax(k) + BCx(k) + Bu(k)

is also completely controllable.

(Hint: Rather than checking the rank condition, apply the
original definition of complete controllability.)

5. Suppose that the partitioned system

ẇ (t) A11 A12 w (t)
=
ẏ (t) A21 A22 y (t)

with output y(t) is completely observable. Show that the

combination of A11 as system and A21 as output matrix is
a completely observable pair.
(Hint: You may find it simpler to prove complete controlla-
bility of the transposed combination.)

6. Consider the two dynamic systems

S1 S2

ẋ1 = x2 + u ẋ3 = x3 + w
ẋ2 = −2x1 − 3x2 z = x3
y = αx1 + x2

S1 has state (x1 , x2 ), control u, and output y. S2 has state

x3 , control w and output z.
 112

(a) Determine whether each system is controllable, observ-

able. (Note: α is a parameter.)
(b) These two systems are connected in series, with w = y.
The resulting system is called S3 . Determine whether it
is controllable, observable.
(c) The systems are now connected in a feedback configura-
tion as shown in the figure below, to produce S4 .

Determine whether S4 is controllable, observable.

7. Two pendulums are connected by a spring as shown in the

diagram below:

The masses of the rods, which have unit length, can be ne-
glected. Equal and opposite control forces u(t) are applied to
the particles, which have mass m. For small oscillations (so
 113 MAT217—X/1
APM214—Y/1

that θ2i and higher powers can be neglected) the equations of

motion are

mθ̈1 = −mgθ1 − u − ka2 (θ1 − θ2 )

mθ̈2 = −mgθ2 + u + ka2 (θ1 − θ2 )

Take

x1 = θ1 + θ2
x2 = θ1 − θ2
x3 = ẋ1
x4 = ẋ2

as state variables. Is the system completely controllable?

 114

CHAPTER 4

AUTONOMOUS NONLINEAR

SYSTEMS

The introduction of nonlinearity into a system introduces a rich

variety of behaviours that are qualitatively different from the linear
case. For example,
zn+1 = zn2 + c

where zn and c are complex, leads to the amazing richness and

beauty of the Mandelbrot set (which is discussed further in Chapter
5). However, we start by using the results of linear theory, and we
will see that in most cases the behaviour of a nonlinear system in
the neighbourhood of a fixed point can be found from linear theory.
The crucial phrase in the previous sentence is “in the neighbour-
hood of a fixed point”: far from the fixed point the behaviour is
dominated by nonlinear effects and linear approximations are not
valid.

4.1 STABILITY
At this stage, we need to be rather more precise about the concept
of stability. A linear system has, in general, only one fixed point,
 115 MAT217—X/1
APM214—Y/1

and the whole space is affected by the stability (or not) of the fixed
point. In nonlinear systems there may be several fixed points, and
the stability (or not) about each fixed point is in general a concept
that is valid only locally. First, a trajectory means a path x(t)
or x(n) representing a solution of the system. The diagrams in
Chapter 2 showing the solution behaviour near a fixed point are
diagrams of trajectories. A fixed point x∗ is asymptotically stable
iff there exists a neighbourhood N of x∗ such that every trajectory
that starts in N approaches x∗ as time tends to infinity:

Continuous case: lim x(t) = x∗ . Discrete case: lim x(n) = x∗ .

t→∞ n→∞

Comment: The neighbourhood N represents the idea that the sta-

bility is local. While all trajectories in the space may not tend to
x∗ , there is a region around x∗ , i.e. N , in which the trajectories do
tend to x∗ .

Sometimes it is also useful to refer to the domain of asymptotic sta-

bility of a fixed point x∗ : this is simply the largest neighbourhood
N for which the definition of asymptotically stable holds.

Example. In linearised theory, a stable node and a stable focus

(figures 3 and 5 of Chapter 2) are both asymptotically stable.

The next definition that we need is that of an unstable fixed point.

Unfortunately the definition is a bit contorted, so study it carefully:
A fixed point x∗ is unstable in a neighbourhood N of x∗ iff there
exists a sequence xi of points, with limi→∞ xi = x∗ , such that every
trajectory, with an xi as starting point, contains points that are
outside N.
 116

Example. In linearized theory, an unstable node, an unstable

focus and a saddle point (figures 1, 2 and 4 of Chapter 2) are all
unstable.

Finally, we define marginal stability. A fixed point x∗ is stable iff

it is not unstable. Of course it is easy to see that

asymptotic stability ⇒ stability,

but the converse is not true. A fixed point x∗ is defined to be mar-

ginally stable iff it is neither unstable nor asymptotically stable.

Example. In linearised theory, a centre (figure 6 of Chapter 2) is

marginally stable.

The examples given above are all for linear systems and the domain
of stability is the whole space. A simple example illustrating a finite
domain of stability is:
dx
= x(x2 − 1).
dt
The singular points are at x = 0, x = −1, x = +1 and the phase
diagram for the system is:

Thus x = −1 and x = +1 are unstable, and x = 0 is stable with

domain of stability (−1, 1).
 117 MAT217—X/1
APM214—Y/1

4.2 LINEARISATION ABOUT SINGULAR POINTS

For clarity, we start in one dimension:

dx
= g(x) (1)
dt
and suppose that x = x∗ is a fixed point, i.e. g(x∗ ) = 0. Then by
Taylor’s theorem, for x near x∗ :

dx
≈ g(x∗ ) + (x − x∗ )g (x∗ ).
dt
Thus
dx
≈ (x − x∗ )g (x∗ ) . (2)
dt
In the language of the first part of Chapter 2, g (x∗ ) is equivalent
to the constant a, and it is clear that the fixed point x = x∗ is
asymptotically stable for g (x∗ ) < 0 and unstable for g (x∗ ) > 0.

For discrete systems in one dimension:

x(n + 1) = g(x(n)) (3)

and at a fixed point x = x∗ we have g(x∗ ) = x∗ . Then applying

Taylor’s theorem

x(n + 1) ≈ x∗ + (x(n) − x∗ )g (x∗ ).

Thus
x (n + 1) − x∗
≈ g (x∗ ), (4)
x (n) − x∗
and the system is asymptotically stable if |g (x∗ )| < 1 and unstable
if |g (x∗ )| > 1.
 118

Of course it may happen that g (x∗ ) is 0 (continuous case) or 1

(discrete case), and then linearisation cannot say whether x = x∗ is
stable or unstable.
Examples
1. The logistic equation of population growth is given by
dx
= ax(k − x) (5)
dt
where a and k are positive constants.
There are two fixed points, at x = 0 and x = k.
Then g(x) = ax (k − x) so that g (x∗ ) = ak − 2ax∗ , and therefore:

At x = 0 we have g (0) = ak > 0: unstable

At x = k we have g (k) = −ak < 0: asymptotically stable

2.
x(n + 1) = x(n)2 − 3x(n) + 4 (6)

This equation has fixed points given by solving

x∗ = x2∗ − 3x∗ + 4.

Thus
x2∗ − 4x∗ + 4 = 0

so that
x∗ = 2.
 119 MAT217—X/1
APM214—Y/1
d
Furthermore, g (x) = x2 − 3x + 4 so that g (x) = dx (x2 − 3x + 4) =
2x − 3, and at x = x∗ = 2 we have g (2) = 1. Thus linearisation
does not say whether x∗ = 2 is a stable or unstable fixed point.

In order to describe linearisation in higher dimensions it is first

necessary to introduce some notation. In the continuous case the
system is written as
dx
= g(x) (7)
dt
where x, g are vectors. For example, the system
dx dy
= x2 + y 2 − 6, = x2 − y (8)
dt dt
would have g1 = x2 + y 2 − 6, g2 = x2 − y. Then the Jacobian of
the system is  
∂g1 ∂g1
 ∂x ∂y 
 
A= 
.
 (9)
 ∂g2 ∂g2 
∂x ∂y
For the above example we have

2x 2y
A= . (10)
2x −1

The procedure for analysing the behaviour about the fixed points of
a nonlinear system is:
1. Find the fixed point(s).
2. For each fixed point evaluate the Jacobian.

It can be shown that the behaviour of the nonlinear system in the

neighbourhood of the fixed point is similar to that of the linear
 120

system
dx
= Ax (11)
dt

in the neighbourhood of the origin, provided the linear system is

not on the boundary between two qualitatively different types of
behaviour. For a system of two equations, and referring to the
diagram on page 41 of Chapter 2, this means that the singular point
behaves as shown in the diagram UNLESS
(a) ∆ = 0
or
(b) σ = 0, ∆ > 0
or
(c) σ 2 = 4∆.

For a system of three or more equations, there is no short—cut to

finding the behaviour of the linear system, and one has to calculate
the eigenvalues. If the real parts of all the eigenvalues are < 0, then
the nonlinear system is asymptotically stable in a neighbourhood of
the singular point. If the real part of at least one eigenvalue is > 0,
then the singular point is unstable, at least in some neighbourhood.
If at least one eigenvalue has real part exactly 0 and the others have
real part ≤ 0, then linear theory cannot say anything about the
behaviour of the singular point.

The above can be formulated as a precise theorem and proven. How-

ever the proof involves ( , δ) techniques in real analysis, and is omit-
ted.
 121 MAT217—X/1
APM214—Y/1

In the discrete case the dynamic system is written:

x(n + 1) = g(x(n)) (12)

and the Jacobian matrix A is defined as in the continuous case. A

fixed point x = x∗ is defined by the condition

x∗ = g(x∗ ). (13)

If the eigenvalues of A are denoted by λi , then the fixed point is lo-

cally (i.e. in a neighbourhood of x∗ ) asymptotically stable if |λi | < 1;
locally unstable if there exists an eigenvalue with |λi | > 1; and lin-
ear theory cannot say whether or not the fixed point is stable if the
largest value of |λi | is exactly 1.

Examples
1.

ẋ = x2 + y 2 − 6
ẏ = x2 − y (14)

The critical points are found by solving ẋ = 0 and ẏ = 0 simultane-

ously:

ẋ = 0 ⇒ x2 + y2 − 6 = 0 (1)
ẏ = 0 ⇒ x2 − y = 0 (2)

From (2) we have x2 = y which, substituted in (1) yields

y + y 2 − 6 = 0 ⇒ (y − 2) (y + 3) = 0,

i.e.
y = 2 or y = −3.
 122

We only consider real solutions of x and y and thus ignore the option
y = −3, since this would give complex values for x. From y = 2 we
√
get x = ± 2 and our two critical points are therefore given by
√ √
P1 = 2, 2 and P2 = − 2, 2 .

The Jacobian matrix is

2x 2y
A=
2x −1

and so at P1
√
2 2 4
A= √ .
2 2 −1
Thus
√
= −10 2 < 0
√
σ = 2 2−1

so, since < 0, the critical point is locally a saddle point.

At P2
√
−2 2 4
A= √ ,
−2 2 −1
from which
√
= 10 2
√
σ = −2 2 − 1 < 0

and
σ2 − 4 <0
 123 MAT217—X/1
APM214—Y/1

so that P2 is locally a stable focus.

2.
ẍ + x − x3 = 0 (15)

Writing the equation as a first order system with y = ẋ:

ẋ = y ẏ = x3 − x. (16)

Solving ẋ = 0, ẏ = 0 simultaneously we get:

ẋ = 0 ⇒ y = 0
ẏ = 0 ⇒ x3 − x = x (x − 1) (x + 1) = 0

yielding the critical points P1 = (0, 0), P2 = (1, 0) and P3 = (−1, 0).
The Jacobian matrix is

0 1
A= .
2x2−1 0

Thus
0 1
A1 = ,
−1 0

0 1
A2 = A3 = .
2 0
Now, det(A2 ) = det(A3 ) = −2 < 0, and therefore both P2 and P3
are saddle points. The matrix A1 has ∆ = 1, σ = 0. It is therefore a
focus, but since σ is on the borderline between stable and unstable,
it is not clear whether the behaviour near P1 is a stable focus, an
unstable focus, or a centre.
 124

3.

x(n + 1) = x(n)2 − 2y(n) + 3

y(n + 1) = x(n)y(n) − 7y(n) (17)

To find the singular points, we need to find values (x∗ , y∗ ) that

satisfy

x = x2 − 2y + 3, (i)
y = xy − 7y. (ii)

The equation (ii) is 0 = y(x − 8), and is satisfied by y = 0 or x = 8.

Substituting y = 0 into (i) gives x2 − x + 3 = 0, which has no real
solution. Substituting x = 8 in (i) gives y = 29 12 . Thus the only
singular point is
P1 = (x∗ , y∗ ) = (8, 29.5).

The Jacobian of the system is

2x −2
A= ,
y x−7

thus
16 −2
A1 =
29.5 1
at the singular point P1 . The eigenvalue equation has solutions
that are complex and |λ| > 1.

Thus the singular point P1 is unstable.

 125 MAT217—X/1
APM214—Y/1

Phase Plane Analysis

This is usually applied to a system of two continuous equations, and
is a means of obtaining the overall behaviour of the system from the
nature of the fixed points and other elementary considerations. It
is more difficult to apply the methods to discrete systems, and the
cases to be considered here will all be continuous.

It is useful to start by defining some terms. A trajectory is a path

in the (x, y) plane of a solution of the system. The diagrams in
Chapter 2 showing the solution behaviour near a fixed point are
diagrams of trajectories. An important fact about trajectories is
that at any non—singular point on the phase plane
dy dy/dt g1 (x, y)
= = (18)
dx dx/dt g2 (x, y)
has a definite value. Thus, at non—singular points the direction of
a trajectory is unique, and trajectories do not cross each other.

The x-isocline and y-isocline are curves in the (x, y) plane on which
dx/dt = 0 and dy/dt = 0, respectively. They are obtained by solving
the (algebraic) equations

g1 (x, y) = 0, g2 (x, y) = 0. (19)

The importance of the isoclines is that:

(a) singular points are at the intersection of an x—isocline and a

y—isocline;

(b) they divide the (x, y) plane into regions in which the direction
of a trajectory is known to within 90o ;
 126

(c) on the x—isocline dx/dt = 0, so trajectories crossing the x—

isocline must be parallel to the y—axis at the crossing point;
similarly, trajectories crossing the y—isocline must be parallel
to the x—axis at the crossing point.

Phase plane analysis usually proceeds as follows:

(a) Calculate and draw the isoclines.

(b) Find the location of the fixed points.

(c) Determine the nature of each fixed point by calculating its

Jacobian matrix.

(d) Sketch the direction of the trajectories in each region, bounded

by isoclines, of the (x, y) plane.

(e) Attempt to sketch trajectories throughout the whole (x, y)

plane.

These techniques are illustrated in some application models that are

discussed in the next section.

4.3 SOME APPLICATIONS

Application 1: Lotka—Volterra competition model
The model describes a competitive interaction between two species
that are competing for the food, water and other resources of an
ecosystem. The use of these resources by one species inhibits the
ability of the other species to use them, thereby reducing the growth
rate. The model aims to determine under what conditions two
competing species can coexist. Let x and y denote the populations
 127 MAT217—X/1
APM214—Y/1

of the two species, then the model is represented by the dynamical

system:
r1
ẋ = x(P1 − x − α12 y)
P1
r2
ẏ = y(P2 − y − α21 x) (20)
P2
Note that in the absence of one of the species, say y = 0, the equa-
tion for x is ẋ = (r1 /P1 )x(P1 − x), which is the standard logistic
equation for population growth; it has a stable fixed point at x = P1 .
Similarly, if x = 0 then y is described by the logistic equation. The
effect of competition is represented by the terms (−r1 /P1 )α12 xy and
(−r2 /P2 )α21 yx.

In this model it does not make sense to talk about negative popu-
lations, so we are only interested in solutions with x, y ≥ 0.

The isoclines are:

ẋ = 0: x = 0 (i) or P1 − x − α12 y = 0 (ii)
(21)
ẏ = 0: y = 0 (iii) or P2 − y − α21 x = 0 (iv)
The fixed points are therefore:

(0, 0) , (0, P2 ) , (P1 , 0) Q (22)

where Q is at the intersection of (ii) and (iv). There are four se-
parate cases to consider, depending on the values of the constants
P1 , P2 , α12 , α21 . The constants r1 , r2 do not appear in the iso-
cline equations and so do not affect the fixed points of the system:
they affect the rate at which the dynamic system evolves, but not
the end—point(s) to which it can evolve. The cases are presented
diagrammatically as:
 128

Figure 1

We will investigate case II here, i.e. the case where, as can clearly
be seen from the sketch, P1 > P2 /α21 and P2 > P1 /α12 . The other
cases are left to you to do as exercises.
 129 MAT217—X/1
APM214—Y/1

The Jacobian is:

 2r1 x r1 y −α12 r1 x 
r1 − − α12
 P1 P1 P1 
 
J= . (23)
 −r2 α21 y 2r2 y r2 α21 x 
r2 − −
P2 P2 P2
Thus at (0, 0) we get
r1 0
J=
0 r2
so that

= r1 r2 > 0
σ = r1 + r2
σ2 − 4 = (r1 + r2 )2 − 4r1 r2 = (r1 − r2 )2 > 0.

The origin is therefore always an unstable node.

At the singular point (0, P2 )

r1 (1 − α12 P2 /P1 ) 0
J= . (24)
−r2 α21 −r2

In this case, case II, P2 > P1 /α12 so the first term in the Jacobian
is negative. Thus, by setting
P2
β = 1 − α12
P1
and keeping in mind that β < 0 (from P2 > P1 /α12 ), we have

= −r1 r2 β > 0
σ = r1 β − r2 < 0
 130

and

σ2 − 4 = (r1 β − r2 )2 + 4r1 r2 β
= r12 β 2 − 2r1 r2 β + r22 + 4r1 r2 β
= r12 β 2 + 2r1 r2 β + r22
= (r1 β + r2 )2
> 0.

The singular point (0, P2 ) is therefore clearly a stable node.

Similarly, at the point (P1 , 0),

−r1 −r1 α12

J= (25)
0 r2 (1 − α21 P1 /P2 )

and P1 > P2 /α21 so the last term in the Jacobian is negative and
the singular point can also shown to be a stable node.

In order to deal with the singular point Q, we could solve (ii) and
(iv), substitute into the Jacobian and then evaluate σ and ∆. How-
ever, the algebra involved in such a direct approach is rather lengthy
(and therefore it is easy to make a mistake). It is better to approach
the algebra with some cunning. At the point Q equations (ii) and
(iv) apply, so it is permissible to use them to simplify the diagonal
terms of the Jacobian. The result is:
 131 MAT217—X/1
APM214—Y/1

At Q,
 −r1 x∗ −α12 r1 x∗ 
 P1 P1 
 
J= . (26)
 −r α y −r2 y∗ 
2 21 ∗
P2 P2
Thus
r1 r2 x∗ y∗
∆ = (1 − α21 α12 ),
P1 P2
r1 x∗ r2 y∗
σ = − + .
P1 P2

At Q, both x∗ , y∗ are positive, so σ is clearly negative, and the sign

of ∆ will depend on the sign of (1−α21 α12 ). In case II, P2 > P1 /α12
and P1 > P2 /α21 , thus

P1 P2
P2 > > (27)
α12 α12 α21
so that clearly
α12 α21 > 1.

Thus ∆ < 0 and Q is a saddle point.

 132

The phase plane diagram is therefore:

In the above diagram we show:

(a) the direction of the trajectories on each isocline;

(b) the range of directions of the trajectories in each region bounded

by isoclines;

(c) the singular points and information about their nature.

Note from the sketch that a singular point occurs only where an x—
isocline intersects with a y—isocline. Since the y—axis is an x—isocline
and since the trajectories crossing the x—isocline are parallel to the
y—axis, the y—axis itself will be a trajectory. A similar argument
follows for the x—axis. The direction of the trajectories are away
from the unstable singular points, towards the stable points. Note
 133 MAT217—X/1
APM214—Y/1

also that the trajectories crossing the other x—isocline are crossing
it parallel to the y—axis and the trajectories crossing the y—isocline
are crossing it parallel to the x—axis.

From a diagram like that above there is usually only one way to
sketch the global, qualitative behaviour of the trajectories. In this
case:

Conclusion: Depending on the initial values of x and y, one or

other of the species dies out, and the other approaches the stable
value given by the logistic equation for a single population.

The other cases I, III and IV are left as exercises. Here we shall
simply state that stable coexistence is possible only in case I.
 134

Example
The above may be clearer if instead of using arbitrary constants r1 ,
P1 , α12 , r2 , P2 , α21 , we let the constants have definite numerical
values:

r1 = 0.1
P1 = 10000
α12 = 1
r2 = 0.2
P2 = 15000
α21 = 2.

The condition defining the case being considered here, case II, is
P2 > P1 /α12 and P1 > P2 /α21 , which is clearly satisfied. The fixed
points, and the values of the Jacobian at the fixed points, are:

0.1 0
1. (0, 0). From (23), J = .
0 0.2
Clearly this is an unstable node.

2. (0, P2 ). From (24),

0.1 (1 − 1.5) 0
J =
−0.2 × 2 −0.2
−0.05 0
= .
−0.4 −0.2

Thus ∆ = +0.01, σ = −0.25; thus σ 2 > 4∆ and the fixed

point is a stable node.
 135 MAT217—X/1
APM214—Y/1

−0.1 −0.1
3. (P1 , 0). From (25), J = .
0 −0.0666
Thus ∆ = 0.00666, σ = −0.1666; thus σ 2 > 4∆ and the fixed
point is a stable node.

4. Solving (21), (ii) and (iv) we obtain Q : (5000, 5000). Thus

from (26)
−0.05 −0.05
J= .
−0.13333 −0.06666

Thus ∆ = −0.0033333, σ = −0.116666; thus the fixed point is

a saddle point.

Application 2: A discrete predator—prey model

The model is represented by the dynamical system

x(n + 1) = x(n)(1 − a + by(n))

r
y(n + 1) = y(n)(1 + r − y(n) − cx(n)) (28)
P
where x(n) is the number of predators and y(n) is the number of
prey. If x(n) = 0, then y(n) obeys the usual logistic equation. For
x(n) > 0, the term cx(n) reduces the growth rate of the prey, so
allowing for the fact that they are being eaten by the predators. In
the absence of any prey, the growth rate of the predators is negative,
because they have nothing to eat. The term by(n) models the fact
that the growth rate of the predators depends on the size of their
food supply.

For discrete systems we analyse the behaviour near the fixed points,
but it is not possible to apply the other phase plane techniques
 136

that were described for continuous systems. The fixed points are
obtained by solving simultaneously:

x = x(1 − a + by) (i)

ry
y = y(1 + r − − cx) (ii) (29)
P
(i) has solutions: x = 0 or y = a/b; substituting these solutions into
(ii) leads to the fixed points:

x = 0, y = 0
x = 0, y = P
r a a
x = 1− ,y= . (30)
c bP b
The Jacobian is
1 − a + by bx
J= . (31)
−cy 1 + r − 2ry
P − cx

At this stage we should say a little about the values of the constants
a, b, c, r, P . They are all > 0, and the conditions that we will
impose are

1. bP > a (32)
What does this mean? If the prey population is at its (logis-
tic) equilibrium value P , and a small number of predators are
introduced, then the predators would have enough to eat and
their population would grow. It is clear that if this condition
is not satisfied then there is no possibility of a stable situation
involving both predators and prey.

2. r < 2 (33)
This is required so that the model consisting of prey only is
stable at y = P.
 137 MAT217—X/1
APM214—Y/1

We now calculate the Jacobian, and estimate its eigenvalues, at the

three critical points.

1−a 0
(i) x = 0, y = 0, J = . (34)
0 1+r
The eigenvalues are 1—a and 1 + r, so it is unstable and anal-
ogous to a saddle point.

1 − a + bP 0
(ii) x = 0, y = P , J = . (35)
−cP 1−r
Since bP > a, the diagonal elements of J, and therefore in this
case its eigenvalues, are λ1 = 1−a+bP > 1 and λ2 = 1−r with
|λ2 | = |1 − r| < 1. Thus the singular point is again unstable
and analogous to a saddle point.
 r a 
1 b 1−
r a a  c bP 
(iii) x = 1− ,y= ,J= 
 . (36)

c bP b ca ra
− 1−
b bP
It can be shown that both eigenvalues of J are less than 1,
although the algebra is rather intricate. For simplicity, let
ra
d = , and note that 0 < d < 2;
bP
a
e = ar 1 − and note that e > 0.
bP

Then the Jacobian can be rewritten as

 
be
 1 ac 
 
J= 
 ca 
− 1−d
b
 138

so that

= 1−d+e
σ = 2 − d.

The eigenvalues of J are (from equation (20) in Chapter 2):

√
2 − d ± 4 − 4d + d2 − 4 + 4d − e
√ 2
2 − d ± d2 − e
= ,
2
which is real for d2 − e ≥ 0. If d2 − e < 0, the eigenvalues are
complex and are given by
√
2 − d ± e − d2 i
.
2
For the real case d2 − e ≥ 0 we have
√ √
λ1 = 1 + 12 −d + d2 − e < 1 + 12 −d + d2 = 1.
√
Thus 0 < λ1 < 1. Also, λ2 = 1 + 12 −d − d2 − e < 1, and
also λ2 > 1 + 12 (−d − d) = 1 − d > −1. Thus −1 < λ2 < 1.

For the complex case d2 − e < 0,

2 √
d e − d2
|λ| = 1− +
2 2
e
= 1−d+ ,
4

so that, for stability we need |λ| < 1, hence,

e
− d < 0.
4
 139 MAT217—X/1
APM214—Y/1

In terms of the parameters a, b, r, an P , this means

bP < a + 4.

From equation (32) we now conclude that the condition

a < bP < a + 4

will ensure that this fixed point is asymptotically stable.

Example
We let the constants a, b, c, P , r have definite numerical values:

a = 0.5
b = 0.001
c = 0.01
P = 1000
r = 0.5.
 140

The conditions (32) and (33) are clearly satisfied. The fixed points
and corresponding Jacobians are:

0.5 0
1. (0, 0), J = .
0 1.5
The eigenvalues are 0.5 and 1.5, so the fixed point has one
eigenvalue greater than 1, and is therefore unstable and anal-
ogous to a saddle point.

1.5 0
2. (0, P ), J = .
−10 0.5
The eigenvalues are 1.5 and 0.5, so the fixed point is unstable
and analogous to a saddle point.

1 0.025
3. (25, 500), J = .
−5 0.75
Thus the eigenvalue equation is:

λ2 − 1.75λ + 0.875 = 0

which has solutions

λ = 0.875 ± 0.331i.
√
Thus |λ| = 0.8752 + 0.3312 ≈ 0.875; this is less than 1 so the
fixed point is stable, and analogous to a stable focus, because
of the imaginary part in λ.

Application 3: A model of the AIDS epidemic

The basic idea is to divide the population into two groups: healthy
but susceptible to the disease, x(t), and sick, y(t). Then the basic
 141 MAT217—X/1
APM214—Y/1

assumption is that the disease spreads at a rate proportional to xy.

A simple model would be:
dx dy
= −βxy, = βxy − γy (37)
dt dt
where γ represents the death rate, and β is the infection rate. The
system (37) is useful for a disease with a short time—scale, but in the
case of AIDS we also need to take into account population growth.
We get
dx
= αx − δ(x + y)2 − βxy
dt

dy
= βxy − γy. (38)
dt
When y = 0, this system reduces to the logistic equation. For
simplicity, we will scale the population to have a value of 1 at the
logistic population value x = P = α/δ; this implies α = δ. Thus
x and y are given as fractions of P . To translate into numbers of
people, we would multiply by P . For example P ≈ 50 million for
South Africa, and P ≈ 6 × 109 for the world. Then the system is
dx
= αx − α(x + y)2 − βxy (39)
dt

dy
= βxy − γy.
dt
In some places the algebra can be rather heavy, and we will some-
times use numerical values for
the constants α, β, γ:

α = 0.02
β = 0.5
γ = 0.2.
 142

The isocline diagram looks like:

The Jacobian is:

α − 2αx − βy − 2αy −2αy − βx − 2αx
J= . (40)
βy βx − γ

The singular points are therefore:

α 0
1. (0, 0), J = .
0 −γ
It is clearly a saddle point.

−α −2α − β
2. (1, 0), J = .
0 β−γ
If β < γ then J has two negative eigenvalues and the fixed
point is a stable node. If β > γ, which is the case considered
here, then ∆ < 0 and the fixed point is a saddle point.

3. To find the fixed point, we substitute x = γ/β into

αx − α(x + y)2 − βxy = 0.

 143 MAT217—X/1
APM214—Y/1

For the case considered here, this gives

y 2 + 10.8y − 0.24 = 0.

The positive solution is y = 0.022. So the fixed point is

(x∗ , y∗ ) = (0.4, 0.022). Thus J is
−0.0079 −0.217
J= .
0.011 0

Thus ∆ = 0.002, σ = −0.008, so that σ 2 < 4∆. The fixed

point is therefore a stable focus.

The phase plane diagram is therefore:

 144

Discussion
The values of α, β, γ have been chosen as follows. The con-
stant α = 0.02 represents a population that grows at 2% per
year, if resources are available. Putting β = 0.5 means that
the number of infected people increases by about 50% every
year. If γ = 0.2 then about 20% of infected people die every
year. In this case the stable value of the population is reduced
to (x∗ + y∗ ), i.e. 0.437 of what it could be in the absence of
AIDS. The crucial factor is γ/β, and the goal of public health
policy must be to increase γ/β. According to this model if
public education etc. can decrease β so that γ/β > 1, then
the AIDS epidemic will fade away. This can be understood as
saying that if the average infective, before he dies, passes the
disease on to less than, on average, one other person, then the
number of infectives will diminish to zero.
Of course this model is rather crude, but it does show the
power of applied dynamical systems theory to make predic-
tions about important matters in everyday life.

4.4 LIAPUNOV STABILITY THEORY

The method of determining stability by linearisation about a fixed
point is sometimes referred to as Liapunov’s first method. This
section is about Liapunov ’s second method, also called Liapunov ’s
direct method. It works explicitly with the nonlinear system and has
the advantage of being useful in marginal situations; and also that
it can be applied beyond a small region near the fixed point — it may
even determine global stability.
 145 MAT217—X/1
APM214—Y/1

The basic idea is to seek a function that continually decreases to-

wards a minimum as the system evolves (this idea comes from energy
in classical mechanics). Suppose that x∗ is an equilibrium point of
a given dynamical system. Then V (x) is a Liapunov function if:

(1) V (x) is defined in a domain Ω, with x∗ ∈ Ω.

(2) V (x) is real-valued and continuous and has continuous first

partial derivatives.

(3) V (x) has a unique minimum at x∗ , i.e. V (x∗ ) < V (x) for all
x ∈ Ω, x = x∗ .

(4) The value of V decreases along any trajectory in Ω. (41)

The key requirement is (4). Later it will be formulated more pre-

cisely, and in somewhat different ways for the discrete and continu-
ous cases. We will also show that if a Liapunov function exists, then
the equilibrium point x∗ is asymptotically stable.

The idea behind a Liapunov function is shown in the above figure.

The function V (x) has a minimum of zero at x = x∗ . A trajectory
 146

in the (x1 , x2 ) plane is shown: it is such that, when projected into

the surface V (x), its direction must be downhill.

One important point to remember about a Liapunov function is

that it is only a function of the coordinates (x1 , x2 , ..., xn ). It is not
defined by motion of the system. Rather, evolution of the system
means that the coordinates (x1 , x2 , ..., xn ) change, and in this way
V (x) changes. Also, note that Liapunov functions are not unique.

Example
Consider the system

x2 (k)
x1 (k + 1) = ,
1 + x1 (k)2 + x2 (k)2
x1 (k)
x2 (k + 1) = (42)
1 + x1 (k)2 + x2 (k)2

which has x = (0, 0) as an equilibrium point. Let us define the

function
V (x) = x21 + x22 . (43)

This V is defined everywhere, is real-valued and continuous, and

has a unique minimum at the equilibrium point. To determine how
V changes along a trajectory, consider any two vectors, x(k) and
x(k + 1), related by the system equations. We find

V (x (k + 1)) = x1 (k + 1)2 + x2 (k + 1)2

2 2
x2 (k) x1 (k)
= +
1 + x1 (k)2 + x2 (k)2 1 + x1 (k)2 + x2 (k)2

x2 (k)2 + x1 (k)2
= 2
1 + x1 (k)2 + x2 (k)2
 147 MAT217—X/1
APM214—Y/1

V (x (k))
= 2 ≤ V (x (k)) . (44)
1 + x1 (k)2 + x2 (k)2

Thus the function V obeys condition (4), and therefore V is a Lia-

punov function.
Liapunov theorem for continuous systems
We now focus on a continuous system
dx
(t) = g(x(t)) (45)
dt
together with a given equilibrium point x∗ . It is assumed that g
is continuous. The requirement (41-4), that along any trajectory V
is decreasing, is expressed as follows. Suppose x(t) is a trajectory.
Then V (x(t)) is a function of time and represents the values of V
along the trajectory. The requirement (41-4) is then

d
V̇ (x) ≡ (V (x (t))) ≤ 0 (46)
dt
with equality if and only if x = x∗ .
By the chain rule, this is
d ∂V dx1 ∂V dx2 ∂V dxn
(V (x (t))) = + +···+ ≤ 0. (47)
dt ∂x1 dt ∂x2 dt ∂xn dt
Now apply the system equation (45):

d ∂V ∂V ∂V
(V (x (t))) = g1 (x (t))+ g2 (x(t))+···+ gn (x (t)) ≤ 0
dt ∂x1 ∂x2 ∂xn
(48)
[Note: For those of you also taking MAT215 = APM212, this is
more conveniently written as
d
(V (x(t))) = (∇V ) · g ≤ 0 (49)
dt
 148

where ∇ is the gradient operator.]

Theorem Suppose that the dynamical system (45) has an equili-

brium point x∗ , that the function g(x) in (45) is continuous, and
that there exists a function V (x)[V : Ω → R, with the domain
Ω ⊆ Rn ] such that

(i) V (x) is continuous and has continuous first partial derivatives;

(ii) x∗ ∈ Ω;

(iii) V (x) has a unique minimum at x∗ ;

(iv) Ω is either Rn or bounded, and if it is bounded then V (x) is

constant on the boundary of Ω;

(v) V satisfies (48) [or equivalently (49)] with equality if and only
if x = x∗ .

Then for any trajectory that starts in Ω(x(t = 0) ∈ Ω),

lim x(t) = x∗ .
t→∞

[Alternatively, x∗ is asymptotically stable with Ω contained in the

domain of stability.]

Sketch of proof. From (v), V (x(t)) is monotonically decreasing

with t; further, by (iii) V is bounded below by V (x∗ ). Thus V∞ ≡
limt→∞ V (x(t)) exists. We now observe that V∞ = V (x∗ ), since only
at x = x∗ does dV (x(t))/dt = 0. Thus limt→∞ V (x(t)) = V (x∗ ).
Since V (x) has a unique minimum at x = x∗ , this implies

lim x(t) = x∗ .
t→∞
 149 MAT217—X/1
APM214—Y/1

Note that condition (iv) above was not used explicitly, but it is nec-
essary to ensure that a trajectory that starts in Ω cannot leave Ω.
Note also that condition (v) can be weakened to V̇ ≤ 0 with equality
only at x = x∗ and at isolated points on any trajectory; the proof
is somewhat intricate and is omitted.

Liapunov theorem for discrete systems

The dynamical system is:

x(n + 1) = g(x(n)) (50)

and x = x∗ is an equilibrium point. The requirement (41-4) is

expressed as follows. If at t = n the value of the Liapunov func-
tion is V (x(n)), then at the next time instant V = V (x(n + 1)) =
V (g(x(n))). Define

∆V (x) = V (g(x)) − V (x) (51)

then the condition that V is decreasing can be expressed as

∆V (x) ≤ 0 (52)

for all x ∈ Ω, with equality if and only if x = x∗ .

Theorem Suppose that the dynamical system (50) has an equi-

librium point x∗ , that the function g(x) in (50) is continuous, and
that there is a function V (x)[V : Ω → R, with the domain Ω ⊆ Rn ]
such that

(i) V (x) is continuous;

 150

(ii) x∗ ∈ Ω;

(iii) V (x) has a unique minimum at x = x∗ ;

(iv) ∆V , defined by (51), satisfies (50), with equality if and only

if x = x∗ .

Then there exists a domain Ω , with x∗ ∈ Ω ⊆ Ω, such that for any

trajectory that starts in Ω ,

lim x(n) = x∗
n→∞

[i.e. x∗ is asymptotically stable].

Proof. Omitted. The basic idea is the same as for the continuous
case, but because a trajectory can jump out of Ω the proof of the
theorem is rather complicated.

Liapunov theory can also be used to show that a fixed point is unsta-
ble. In the above theorems, if the condition (v) V̇ ≤ 0 (continuous
case) or (iv) ∆V ≤ 0 (discrete case) is replaced by (v) V̇ ≥ 0 (con-
tinuous case) or (iv) ∆V ≥ 0 (discrete case), then the fixed point
x = x∗ is unstable.

Examples

1. We have already shown that V (x) = x21 + x22 is a Liapunov

function for the system (42),
x2 (k)
x1 (k + 1) = ,
1 + x1 (k)2 + x2 (k)2
x1 (k)
x2 (k + 1) = .
1 + x1 (k)2 + x2 (k)2
 151 MAT217—X/1
APM214—Y/1

Thus the equilibrium point (0, 0) is asymptotically stable. Fur-

ther, this conclusion could not have been obtained by analysing
the Jacobian at (0,0), which is

0 1
J= . (53)
1 0

The eigenvalues are ±1, which is a borderline case, so lin-

ear theory provides no information about whether or not the
equilibrium point is stable.

2. ẋ = −x3 − xy2 + y (54)

ẏ = −y 3 − x2 y − x
V = x2 + y2 (55)
Clearly the system (54) has a fixed point at (0, 0), and V (x, y)
trivially satisfies the conditions (i) to (iv) of the continuous
Liapunov theorem. In this case (48) gives

2x(−x3 − xy 2 + y) + 2y(−y 3 − x2 y − x)
= −2x4 − 2x2 y 2 + 2xy − 2y4 − 2x2 y 2 − 2xy
= −2(x4 + 2x2 y 2 + y4 )
= −2(x2 + y 2 )2 . (56)

2
Since −2 x2 + y 2 ≤ 0 with equality only at x = y = 0,
the condition (v) is satisfied. Thus the fixed point (0, 0) is
asymptotically stable. In this case Ω is the whole space, so all
trajectories tend towards (0, 0). Note that linear analysis is of
no help. At (0, 0):
0 1
J= (57)
−1 0
 152

so that ∆ = 1, σ = 0 and the fixed point may be stable or

unstable.

If a Liapunov function is defined over a large region Ω, we can

say more than if it is defined only in a small region. In fact,
if the initial point x(0) has V (x(0)) = q, then the subsequent
trajectory never goes outside the region V (x) ≤ q [provided
this region is contained in Ω]. Therefore, the region Ω over
which the Liapunov function is defined, delineates a region Ω
in which system behaviour is related to the equilibrium point.
This information about extent of stability cannot be obtained
by linearization. For this reason a Liapunov analysis often fol-
lows a linearization analysis of a fixed point even if stability
has been established.

Suppose a Liapunov function can be found that does not change

along any trajectory: V̇ (x) ≡ 0, or ∆V (x) ≡ 0. Thus V is
constant along any trajectory and V is said to be a constant
of motion. We then know that any trajectory must lie on a
contour of the function V . In an n-dimensional space a con-
tour defines an (n − 1)-dimensional surface.

Example. Consider the system

ẋ1 = x2 , ẋ2 = −x1 . (58)

 153 MAT217—X/1
APM214—Y/1

Let V (x) = x21 + x22 . Then V̇ (x) = 2x1 x2 + 2x2 (−x1 ) = 0. Thus V
is a constant of the motion and the system trajectories are circles
centred at the origin.

In practice, we are given the equations describing a system and

have to try to construct a suitable Liapunov function. This can be
extremely difficult unless we use the context, or original motivation,
of the system equations. This idea is illustrated in two examples.

1. A simple swinging pendulum has an equilibrium point when

hanging straight down. Further, it is intuitively clear that,
assuming there is friction at the bearing, then the system will
approach the equilibrium point; i.e. it is stable. We assume
that the pendulum is of length R and has point mass M con-
centrated at the end. The position of the pendulum at any
time is described by the angle θ with the vertical. We assume
that the frictional force is proportional to the speed of the
pendulum.

From Newton’s laws of motion we have:

M Rθ̈ = −M g sin θ − M kθ̇. (59)
 154

Expressing (59) as a first-order system:

g k
θ̇ = ω, ω̇ = − sin θ − ω. (60)
R R
There is a fixed point at (θ = 0, ω = 0).

The question now is: How do we find a Liapunov function

for the system (60)? We want something that we expect will
decrease with time. The physical situation is that energy is
being dissipated by friction, and thus we try the mechanical
energy of the system. [It was this type of situation that led
to the development of the Liapunov theorem.] V = kinetic
energy + potential energy, so
1
V (θ, ω) = M R2 ω2 + M gR(1 − cos θ). (61)
2
It is clear that, without the motivation of energy, it would be
difficult to guess the form (61) from (60).

We now show that (61) is a Liapunov function. It is continuous

and so are its first partial derivatives, and it has a unique
minimum at (θ = 0, ω = 0). Note that the range of θ can be
taken as −π < θ ≤ π. We compute V̇ :

V̇ (θ, ω) = M R2 ωω̇ + M gR sin θ.θ̇

g kω 2
= M R2 − ω sin θ − + M gR ω sin θ
R R
= −kM Rω2 ≤ 0. (62)

Thus V̇ ≤ 0 with equality whenever ω = 0. This includes the

fixed point (θ = 0, ω = 0) as well as points that are isolated on
 155 MAT217—X/1
APM214—Y/1

any trajectory. Thus the fixed point is asymptotically stable.

2. Suppose a dog is chasing a rabbit. The rabbit runs a straight

course along the x-axis at constant velocity R. The dog runs
at a constant velocity D, but in such a way as to always point
directly toward the rabbit. Let xr , yr and xd , yd denote the x
and y coordinates of the rabbit and dog respectively. Then

ẋr = R, ẏr = yr = 0. (63)

The fact that the velocity of the dog is D means that

ẋ2d + ẏd2 = D2 . (64)

The velocity vector of the dog always pointing toward the

rabbit means that:

ẋd = −k(xd − xr ) (65)

ẏd = −kyd . (66)
 156

The constant k is found by calculating ẋ2d + ẏd2 from (65) and

(66), and then equating it to D2 from (64). We find

D
k= 1 (67)
2 2
(xd − xr ) + yd2

so that (65) and (66) become:

(xd − xr ) D
ẋd = −
(xd − xr )2 + yd2
yd D
ẏd = − . (68)
(xd − xr )2 + yd2

The system is more meaningful when expressed in relative co-

ordinates, representing the difference in position of the rabbit
and dog. Define

x = xd − xr , y = yd (69)

then
ẋd = ẋ + R, ẏd = ẏ (70)

and (68) becomes:

xD −yD
ẋ = − − R, ẏ = . (71)
x2 + y2 x2 + y 2

We examine the system (71). Will the dog catch the rabbit?
This is equivalent to asking whether a trajectory of (71) with
arbitrary initial condition x(0), y(0) will eventually go to the
origin, where the relative coordinates are zero. Since (71)
 157 MAT217—X/1
APM214—Y/1

is undefined at the origin and the origin is therefore not an

equilibrium point, we cannot apply Liapunov’s theorem to this
problem. However, it is still useful to seek a Liapunov function
for (71). The dog is trying to catch the rabbit by minimising
the distance between them. It is therefore natural to try the
distance (or distance squared) from the origin as a Liapunov
function:
V (x, y) = x2 + y. (72)

Now
d
V̇ (x, y) = V (x(t), y(t))
dt
xD 2y (−yD)
= 2x − −R +
x2 + y2 x2 + y2
= −2D x2 + y 2 − 2Rx. (73)

If D > R it easily follows that V̇ (x, y) < 0 everywhere except

at x = y = 0 where V̇ (x, y) = 0. Thus V (x, y) decreases along
any trajectory if D > R. Hence if the dog runs faster than the
rabbit, he will catch it.

(Note that conditions (i) to (iv) are also satisfied. Thus V (x, y)
is a Liapunov function.)

Attractors
So far we have talked about equilibrium and stability in terms of
one isolated point. While this is (usually) the case for linear sys-
tems, non-linearity permits more interesting possibilities. We start
by generalising the concept of an equilibrium point to that of an
 158

invariant set:

A set G is an invariant set (for a dynamical system) if whenever a

point x on a system trajectory is in G, the trajectory remains in G.

An equilibrium point is perhaps the simplest example of an invariant

set. Also, if a system has several equilibrium points then the union
of these points is an invariant set. Here are some more substantial
examples.

1. ẋ = y + x(1 − x2 − y 2 ), ẏ = −x + y(1 − x2 − y2 ) (74)

The origin is an equilibrium point, and it can easily be shown
that the system has no other equilibrium points. However,
once the system is on the unit circle x2 + y 2 = 1, it will stay
there. To see this, observe that at any point on the unit circle
the direction of a trajectory is given by ẏ/ẋ = −x/y, which
is the direction of a tangent to the unit circle. Thus the unit
circle is an invariant set of the system (74).

2. x(n + 1) = 3.2x(n) − 2.2x(n)2 (75)

The system has equilibrium points at x = 0 and 1. Writing
(75) as x(n + 1) = g(x(n)), we see that g(x) = 3.2x − 2.2x2 ,
so that g (x) = 3.2 − 4.4x. Thus |g (x)| > 1 at both equi-
librium points and they are therefore unstable. However, if
we calculate the evolution of this system on a computer we
find that it settles down to a stable 2-cycle:

x (n) = . . . , 0.746; 1.163; 0.746; 1.163; 0.746; . . . . (76)

 159 MAT217—X/1
APM214—Y/1

Thus the system has an invariant set G = {0.746; 1.163}.

Having generalised the concept of an equilibrium point to that of an

invariant set, we now generalise the concept of stability. Although
the phrase “asymptotically stable” is sometimes used to describe an
invariant set, it is more usual to talk about an attractor.

An invariant set G is an attractor if there exists > 0 such that

if |x(0) − G| < then limt→∞ |x(t) − G| = 0 (continuous case), or
limn→∞ |x(n) − G| = 0 (discrete case). Note that |x − G| means the
greatest lower bound of the distance between x and every point in
G, i.e. the shortest distance between x and G.

In general, invariant sets and attractors may have an extremely com-

plicated structure. However, here we will restrict attention to cases
where the attractor is a closed curve, i.e. analogous to a circle, and
is then called a limit cycle; or consists of a finite number (n) of
points and is called an n-cycle. Examples 1 and 2 above illustrated
these two cases.

We now investigate conditions under which an invariant set is an

attractor.

First we state a general result, which is an extension of the Liapunov

theory of the previous section.

Theorem Consider the dynamical system (50) [or (45)]. Let V (x)
be a scalar function with continuous first partial derivatives and let
 160

s > 0. Let Ωs denote the region where V (x) < s. Assume that Ωs is
bounded and that ∆V (x) ≤ 0 [or V̇ (x) ≤ 0 in continuous time]. Let
S be the set of points within Ωs where ∆V (x) = 0 [or V̇ (x) = 0],
and let G be the largest invariant set within S i.e. the union of all
the invariant sets within S. Then every trajectory in Ωs tends to G
as time increases.

Proof Omitted.

Comment. Not only does the theorem show that G is an attractor,

but it also gives a lower estimate of the domain (Ωs ) of attraction.

A special case of the above result is the Poincar é-Bendixson theo-

rem which applies to 2-dimensional continuous systems: If a region
R of the phase plane contains no fixed points, and trajectories enter
but never leave R, then R contains a limit cycle. Again the proof is
omitted.

This implies that the only attractors in 2-dimensional continuous

systems are fixed points and limit cycles. Attractors with a more
complex structure are found in higher dimensional continuous sys-
tems, and in discrete systems.

Example. Consider (74) again —

ẋ = y + x[1 − x2 − y2 ], ẏ = −x + y[1 − x2 − y 2 ].

Define
V (x, y) = (1 − x2 − y 2 )2 . (77)
 161 MAT217—X/1
APM214—Y/1

Then

V̇ (x, y) = −2(1 − x2 − y 2 )(2xẋ + 2y ẏ)

= −4(x2 + y 2 )(1 − x2 − y 2 )2 . (78)

The function V (x, y) has the graph shown:

V̇ (x, y) ≤ 0 everywhere, with equality at x = y = 0 and on the

unit circle. V has a maximum at x = y = 0 and therefore by a
simple linearization argument, as in the previous chapter, this is an
unstable equilibrium point. Let Ω1 be the region where V (x) < 1
[then 0 ∈ / Ω1 ]. In applying the theorem the set S is the unit circle.
We have already seen that the unit circle is an invariant set; thus
every trajectory in Ω1 tends towards the limit cycle consisting of
the unit circle.
 162

For an n-cycle in a discrete system we can use linear methods to de-

termine stability. We illustrate this for a 2-cycle in a 1-dimensional
system
x(n + 1) = g(x(n)) (79)

and suppose that (a1 , a2 ) form a 2-cycle:

a1 = g(a2 ), a2 = g(a1 ). (80)

Let b be near a1 . Let c = g(b), so c is near a2 , and let d = g(c), so

that d is near a1 . Then the 2-cycle is an attractor if d is nearer to
a1 than b, i.e. if
|d − a1 | < |b − a1 |. (81)

[Technically we should also check that if e = g(d), then e is nearer

to a2 than c is. However, it turns out that this is not necessary
because the condition (80) is symmetric in a1 , a2 .]
 163 MAT217—X/1
APM214—Y/1

Using Taylor’s theorem,

c = g(b) = g((b − a1 ) + a1 ) = g(a1 ) + (b − a1 )g (a1 )

= a2 + (b − a1 )g (a1 ). (82)

Similarly,
d = g(c) = a1 + (c − a2 )g (a2 ). (83)
Substituting for (c − a2 ) from (82) into (83) gives

d = a1 + (b − a1 )g (a1 )g (a2 )
(d − a1 )
∴ = g (a1 ) g (a2 ) . (84)
(b − a1 )
Thus the 2-cycle is an attractor if

|g (a1 )g (a2 )| < 1. (85)

 164

We will not prove the condition for an n-cycle to be an attractor,

but it is the natural generalisation of (85):

|g (a1 )g (a2 ) · · · g (an )| < 1. (86)

Example. Consider (75) again:

x(n + 1) = 3.2x(n) − 2.2x(n)2 .

We will show algebraically that the system has a 2-cycle, and that
it is an attractor. Having a cycle (a1 , a2 ) means:

a2 = g(a1 ), a1 = g(a2 ) ∴ a1 = g(g(a1 )). (87)

We therefore solve
0 = g(g(x)) − x (88)

with
g(x) = 3.2x − 2.2x2 . (89)

Thus (88) is:

0 = g(3.2x − 2.2x2 ) − x
∴ 0 = 3.2(3.2x − 2.2x2 ) − 2.2(3.2x − 2.2x2 )2 − x. (90)

Now x = g(g(x)) is also solved by solutions of x = g(x), so we

already know that x = 0 and x = 1 are solutions of (90). Thus we
can divide (90) by x(x − 1). Doing all the algebra we get
11
(−121x2 + 231x − 105) = 0. (91)
125
Thus using the quadratic formula
√
21 ± 21
x= = 0.746 or 1.163. (92)
22
 165 MAT217—X/1
APM214—Y/1

Thus the 2-cycle is, as stated earlier (0.746; 1.163). To investigate

stability note that
g (x) = 3.2 − 4.4x (93)
Thus
|g (0.746).g (1.163)| = 0.158 (94)
which is less than 1 and the 2-cycle is therefore an attractor.

Comment. The dynamical system (75) represents logistic popu-

lation growth, with a large growth factor (3.2). As such it could
be used to describe the population of various species of insect. The
oscillations in population size predicted by the model are observed
in nature.

Exercises

1. Consider the cases I, III and IV of the Lotka—Volterra com-

petition model. For each case, find the nature of the singular
points and sketch the phase plane diagrams with the trajecto-
ries.

2. For the AIDS epidemic model consider the case γ/β > 1. Find
the nature of the singular points and sketch the phase plane
diagram with the trajectories.

3. Classify (if possible) each critical point of the given plane au-
tonomous system as a stable node, an unstable node, a stable
spiral point, an unstable spiral point, or a saddle point.

(a) ẋ = 1 − 2xy
ẏ = 2xy − y
 166

(b) ẋ = y − x2 + 2
ẏ = x2 − xy
(c) ẋ = −3x + y2 + 2
ẏ = x2 − y 2
(d) ẋ = −2xy
ẏ = y − x + xy − y3
(e) ẋ = x2 − y 2 − 1
ẏ = 2y
(f) ẋ = 2x − y 2
ẏ = −y + xy
(g) ẋ = xy − 3y − 4
ẏ = y 2 − x2
(h) ẋ = x(1 − x2 − 2y 2 )
ẏ = y(3 − x2 − 3y 2 )
(i) ẋ = −2x + y + 10
y
ẏ = 2x − y − 15 y+5 .

4. Classify (if possible) each critical point of the given second—

order differential equation as a stable node, an unstable node,
a stable spiral point, an unstable spiral point, or a saddle.

(a) θ̈ = (cos θ − 0.5) sin θ, |θ| < π.

(b) ẍ + ẋ(1 − x3 ) − x2 = 0.

5. Consider the nonlinear dynamical system

a(n + 1) = [1.8 − 0.8a(n) − 0.2b(n)]a(n)

b(n + 1) = [1.8 − 0.8b(n) − 0.2a(n)]b(n).
 167 MAT217—X/1
APM214—Y/1

(a) Find the equilibrium vectors (or points) for this system,
and for each equilibrium vector, determine if it is asymp-
totically stable or unstable.
(b) Pick a point (a(0), b(0)) close to each equilibrium vector
and compute (a(1), b(1)), (a(2), b(2)) and (a(3), b(3))
using this nonlinear system to see if (a(k), b(k)) seems to
agree with your results in part (a).

6. Repeat Problem 5 for the nonlinear dynamical system

a(n + 1) = [1.4 − 0.4a(n) − 1.2b(n)]a(n)

b(n + 1) = [1.4 − 0.4b(n) − 1.2a(n)]b(n).

7. For the system

ẋ(t) = sin[x(t) + y(t)]

ẏ(t) = ex(t) − 1

determine all the equilibrium points, and classify each equilib-

rium point as asymptotically stable or unstable.

8. The van der Pol equation. The equation

ẍ + [x2 − 1]ẋ + x = 0

arises in the study of vacuum tubes. Show that if < 0, the

origin is asymptotically stable and that for > 0 it is unstable.

9. In many fishery science models, the rate at which a species is

caught is assumed to be directly proportional to its abundance.
 168

If both predator and prey are being exploited in this manner,

the Lotka-Volterra differential equations take the form

ẋ = −ax + bxy − 1x

ẏ = −cxy + dy − 2 y,

where 1 and 2 are positive constants.

(a) When 2 < d, show that there is a new critical point in

the first quadrant that is a centre.
(b) Volterra’s principle states that a moderate amount of
exploitation increases the average number of prey and de-
creases the average number of predators. Is this fisheries
model consistent with Volterra’s principle?

10. Suppose that the damping force of free, damped motion is

proportional to the square of the velocity, and the differential
equation is
β k
ẍ = − ẋ|ẋ| − x.
m m

(a) Write the second-order differential equation as a plane

autonomous system and find all critical points.
(b) The system is called overdamped when (0,0) is a stable
node and is called underdamped when (0,0) is a stable
spiral point. Physical considerations suggest that (0,0)
must be an asymptotically stable critical point. Show
that the system is necessarily underdamped.
d
[Hint: dy y|y| = 2|y|.]
 169 MAT217—X/1
APM214—Y/1

11. The plane autonomous system

y
ẋ = α x−x
1+y
y
ẏ = − x−y+β
1+y

arises in a model for the growth of microorganisms in a chemo-

stat, a simple laboratory device in which a nutrient from a
supply source flows into a growth chamber. In the system, x
denotes the concentration of the microorganisms in the growth
chamber, y denotes the concentration of nutrients, and α and
β are positive constants that can be adjusted by the experi-
menter. Show that when α > 1 and β(α − 1) > 1, the system
has a unique critical point (x∗ , y∗ ) in the first quadrant, and
demonstrate that this critical point is a stable node.

12. Consider the dynamical system

a(n + 1) = [1.4 − 0, 4a(n) − 1.2(n)]a(n)

b(n + 1) = [1.4 − 0.4b(n) − 1.2a(n)]b(n).

Use phase plane analysis in the first quadrant to analyze the

solutions to this system. Do this by constructing the lines

a = 0 and 0.4 − 0.4a = 1.2b = 0

and drawing arrows to the left or right according to the solu-

tion to an appropriate inequality, and then constructing the
lines
b = 0 and 0.4 − 0.4b − 1.2a = 0
 170

and drawing arrows up or down. Finally, combine the two

constructions to get a phase plane analysis of the behaviour
of solutions.

13. Use phase plane analysis to analyze the solutions to the dy-
namical system
2
a(n + 1) = [2 − a(n) − b(n)]a(n)
3
2
b(n + 1) = [2 − b(n) − a(n)]b(n)
3
in the first quadrant.

14. Use phase plane analysis to analyze the solutions to the dy-
namical system

a(n + 1) − a(n) = 1.2[1 − b(n)]a(n)

b(n + 1) − b(n) = 0.5[1 − 5b(n) + a(n)]b(n)

for the predator-prey model. Notice that the solutions oscillate

about the equilibrium vector
4
A= .
1

15. No Marginal Stability. Consider the system

ẋ1 = x1 a2 − x21 − x22 + x2 a2 + x21 + x22

ẋ2 = −x1 a2 + x21 + x22 + x2 a2 − x21 − x22

and the function V (x) = x21 + x22 . Show that the system is sta-
ble for a = 0 and unstable for a = 0. Thus, the transition from
(asymptotic) stability to instability does not pass through a
point of marginal stability.
 171 MAT217—X/1
APM214—Y/1

16. Prove that the origin is asymptotically stable for each of the
systems below using Liapunov’s direct method. [In parts (a)
and (b) find a suitable Liapunov function. In part (c) try the
suggested function.]

(a) ẋ = y
ẏ = −x3
(b) ẋ = −x3 − y2
ẏ = xy − y3
(c) ẋ = y(1 − x)
ẏ = −x(1 − y)
V = −x − log(1 − x) − y − log(1 − y)

17. The van der Pol Equation. The equation

ẍ + [x2 − 1]ẋ + x = 0

arises in the study of vacuum tubes. Show that if > 0 there

is a limit cycle that is approached by all other trajectories.
[Hint: The origin is unstable, and Liapunov’s method will
show that trajectories a long way from the origin move towards
the origin. Now use the Poincaré-Bendixson theorem.]

18. Consider the dynamical system

A(n + 1) = 3, 2A(n) − 3, 2A2 (n).

The numbers a = 0 and a = 0.6875 are equilibrium values for

this dynamical system.
 172

(a) Write down the equation, f (f (a)) − a = 0, factor out the

terms a and a − 0.6875, then use the quadratic formula
to find the 2-cycle for this equation.
(b) Show that this 2-cycle is an attractor.

19. Consider the dynamical system

A(n + 1) = 3.2A(n) − 1.6A2 (n).

(a) Approximate the numbers a1 and a2 that form a 2-cycle

for this dynamical system, by letting ao = 1 and comput-
ing A(1), A(2), and so forth, until you repeat the same
two numbers. Round off each A(k) value to 3 decimal
places.
(b) Show that this 2-cycle is an attractor by computing
f (a1 )f (a2 ).

20. Determine whether the origin is a stable equilibrium point for

each of the following systems:

(a) x1 (k + 1) = 2x1 (k) + x2 (k)2

x2 (k + 1) = x1 (k) + x2 (k)
(b) x1 (k + 1) = 1 − ex1 (k)x2 (k)
x2 (k + 1) = x1 (k) + 2x2 (k).

21. Given the system

ẋ1 = x2
ẋ2 = x3
ẋ3 = −(x1 + cx2 )3 − bx3
 173 MAT217—X/1
APM214—Y/1

consider the function

b 1 1 b2 1
V (x1 , x2 , x3 ) = x41 + (x1 +cx2 )4 − x41 + x22 +bx2 x3 + x23 .
4 4c 4c 2 2
(a) What condition on b > 0, c > 0 insures that V (x1 , x2 , x3 ) >
0 for (x1 , x2 , x3 ) = 0?
(b) Show that V̇ has the form
2
3 3
V̇ (x1 , x2 , x3 ) = γx22 x1 + cx2 + x21
2 4

and determine the constant γ.

(c) Is V always a Liapunov function if b and c satisfy the
conditions of part (a)?
(d) Is the origin asymptotically stable under these condi-
tions?

22. Consider the dynamical system

A(n + 1) = 3.1A(n) − 3.1A2 (n).

(a) Show that a1 = 0.76456652 and a2 = 0.558014125 form

a 2-cycle by showing that, if A(0) = a1 , then A(1) = a2 ,
and A(2) = a1 .
(b) Show that this 2-cycle is stable by showing that

|f (a1 )f (a2 )| < 1.

23. Consider the dynamical system

A(n + 1) = 3.3A(n) − 3.3A2 (n).

 174

(a) Find a 2-cycle for this dynamical system.

(b) Show that this 2-cycle is stable by showing that

|f (a1 )f (a2 )| < 1.

24. Consider the dynamical system

A(n + 1) = 3.6A(n) − 3.6A2 (n).

(a) Find a 2-cycle for this system.

(b) Show that this 2-cycle is unstable by showing that

|f (a1 )f (a2 )| > 1.

25. Consider the dynamical system

A(n + 1) = 3.5A(n) − 2.5A2 (n).

(a) Show that a1 = 0.701237896 is one point in a 4-cycle by

finding a2 = f (a1 ), a3 = f (a2 ), and a4 = f (a3 ), and
showing that a1 = f (a4 ).
(b) Show that this 4-cycle is stable by showing that

|f (a1 )f (a2 )f (a3 )f (a4 )| < 1.

26. The rod of a pendulum is attached to a horizontally movable

joint and rotates at a constant angular speed ω. The resulting
differential equation for θ (the angle that the pendulum makes
with the vertical) is:

d2 θ dθ
m 2
= ω2 m sin θ cos θ − mg sin θ − β .
dt dt
 175 MAT217—X/1
APM214—Y/1

(a) If ω 2 < g/ , show that (θ = 0, dθ/dt = 0) is a stable fixed

point and is the only critical point with −π < θ < π.
(b) If ω 2 > g/ , show that (θ = 0, dθ/dt = 0) is unstable and
that there are two additional stable critical points with
−π < θ < π.

27.(a) State and prove Liapunov’s theorem for a continuous-time dy-

namical system.

(b) By finding a suitable Liapunov function show that the origin

is a stable fixed point for the dynamical system:

ẋ = yx2 − x
ẏ = −x3 − y

What is the domain of stability?

28.(a) A farmer’s crops are eaten by one species of insect, I1 , who

are in turn eaten by another species, I2 . The populations are
described by the dynamical system:

dI1
= (0.5 − 0.01I2 − 0.0005I1 )I1
dt

dI2
= (−0.5 + 0.001I1 )I2 .
dt

Determine the location of the equilibrium point for which both

I1 and I2 are non-zero.

Determine the nature of the equilibrium point.

 176

(b) The farmer now decides to try to reduce the pest population
I1 by spraying his crops with insecticide. The effect of this is
to amend the dynamical system equations to:
dI1
= ((0.5 − k) − 0.01I2 − 0.0005I1 )I1
dt
dI2
= ((−0.5 − k) + 0.001I1 )I2 .
dt

For small k, how is the value of I1 at the equilibrium point

affected? Was spraying with insecticide a good idea?

29(a) Given a nonlinear discrete dynamical system in one variable,

x,
x(n + 1) = g(x(n))

that forms a 2-cycle

a1 = g(a2 ), a2 = g(a1 ),

state and prove the condition for the 2-cycle to be an attractor.

(b) Given that (0.746 and 1.163) form a 2-cycle of

x(n + 1) = 3.2x(n) − 2.2x(n)2

determine whether or not the 2-cycle is stable.

30. The competing species model for species x, y is:

dx x 0, 1
= 0.1x 1 − − y
dt 10000 10000
dy y 0, 2x
= 0.2y 1 − −
dt 15000 15000
 177 MAT217—X/1
APM214—Y/1

Restrict attention to x, y ≥ 0 and sketch the phase plane dia-

gram. Find the location and nature of all singular points. Is
coexistence of the species x and y possible?

31. Show that V (x1 , x2 ) = x21 + x22 is a Liapunov function at the

origin for each of the following systems:

(a) ẋ1 = −x2 − x31 , ẋ2 = x1 − x32 ;

(b) ẋ1 = −x31 + x2 sin x1 , ẋ2 = −x2 − x21 x2 − x1 sin x1 ;
(c) ẋ1 = −x1 − 2x22 , ẋ2 = 2x1 x2 − x32 ;
(d) ẋ1 = −x1 sin2 x1 , ẋ2 = −x2 − x52 ;
(e) ẋ1 = −(1 − x2 )x1 , ẋ2 = −(1 − x1 )x2 .

Find domains of stability at the origin for each of these sys-

tems.

32. Find domains of stability for the following systems by using

an appropriate Liapunov function:

(a) ẋ1 = x2 − x1 (1 − x21 − x22 )(x21 + x22 + 1)

ẋ2 = −x1 − x2 (1 − x21 − x22 )(x21 + x22 + 1);
(b) ẋ1 = x2 , ẋ2 = −x2 + x32 − x51 .

33. Use V (x1 , x2 ) = (x1 /a)2 + (x2 /b)2 to show that the system

ẋ1 = x1 (x1 − a), ẋ2 = x2 (x2 − b). a, b > 0,

has an asymptotically stable origin. Show that all trajectories

tend to the origin as t → ∞ in the region
x21 x22
+ 2 < 1.
a2 b
 178

34. Given the system

ẋ1 = x2 , ẋ2 = x2 − x31

show that a positive definite function of the form

V (x1 , x2 ) = ax41 + bx21 + cx1 x2 + dx22

can be chosen such that V̇ (x1 , x2 ) is also positive definite.

Hence deduce that the origin is unstable.

35. Show that the origin of the system

ẋ1 = x22 − x21 , ẋ2 = 2x1 x2

is unstable by using

V (x1 , x2 ) = 3x1 x22 − x31 .

36. Show that the fixed point at the origin of the system

ẋ1 = x41 , ẋ2 = 2x21 x22 − x22

is unstable by using the function

V (x1 , x2 ) = αx1 + βx2

for a suitable choice of the constants α and β.

 179 MAT217—X/1
APM214—Y/1

CHAPTER 5

ADVANCED TOPICS

In this final Chapter we will look at some recent developments in

Dynamical Systems Theory. This will be done mainly by example,
i.e. we will write down explicit dynamical systems and investigate
certain aspects of their behaviour. In this way we will illustrate the
following complex features that can arise in nonlinear dynamical
systems:

• bifurcation

• chaos

• fractals.

Bifurcation theory
We often think that the world is continuous in the sense that a small
change in input causes a small change in the output. But this is not
always the case, as exemplified by the phrase “the straw that broke
the camel’s back”. Bifurcation theory is the study of that breaking
point, that is, it is the study of the point at which the qualitative
behaviour of a system changes.
 180

Complex behaviour for nonlinear dynamical systems

While you may not be familiar with the term “bifurcation”, you are
certainly familiar with the concept. Consider a vertical rectangular
metal bar. Suppose you start putting weights on the top of the
bar. At first nothing happens, but after you put enough weight
on the bar, the bar bends to one side or the other. There is some
weight ω such that if the total weight is less than ω the bar has
one equilibrium value (it is straight), but if the weight exceeds ω
then the bar has three equilibrium values (bent to the left and bent
to the right which are stable equilibriums, and straight which is an
unstable equilibrium). The weight ω is a bifurcation value in that
as the weight on the bar goes from less than ω to greater than ω,
one equilibrium point divides (or bifurcates) into three equilibrium
values.

Another example of bifurcation is the velocity ν of a rocket. If

its velocity is below a fixed velocity ν e (the escape velocity), it will
fall back to earth; if its velocity exceeds ν e , it will break free of the
earth’s gravitational force and go into outer space.

The key to bifurcation theory is that we have one unknown input

(or parameter) in our dynamical system (the weight on the rod,
the velocity of the rocket). For some values of the parameter we
have one number m1 of equilibrium values, but for another value of
the parameter we have a different number of equilibrium values, say
m2 , that is, the qualitative behaviour of the system we are studying
changes as the parameter changes.
 181 MAT217—X/1
APM214—Y/1

We look at an example of bifurcation involving a 1—dimensional dy-

namic system with 1 parameter. We will see that the situation
is described in a 1 + 1 = 2—dimensional diagram. In general the
bifurcation properties of a d—dimensional dynamic system with p
parameters are described in a (d + p)—dimensional space.

Example
Consider
dx
= g(x) = (1 + x)(1 − a + x2 ). (1)
dt
First, we find all the fixed points:
√
x∗ = −1 and, for a ≥ 1, x∗ = ± a − 1. (2)

Next, we determine stability of the fixed points. To do this we need

to know g (x) :

g (x) = (1 − a + x2 ) + (1 + x)2x = 1 − a + 2x + 3x2 . (3)

Fixed points:

(i) x∗ = −1, g (x∗ ) = 1 − a − 2 + 3 = 2− a. Stable for a > 2,

unstable for a < 2.
√ √
(ii) x∗ = + a − 1, g (x∗ ) = 1 − a + 2 a − 1 + 3(a − 1) = 2(a −
√
1) + 2 a − 1. Thus g (x∗ ) > 0 for a > 1 and the fixed point
is unstable.
√
(iii) x∗ = − a − 1,
√ √ √ 
g (x∗ ) = 2(a − 1) − 2 a − 1 = 2 a − 1 a − 1 − 1 . This
is negative (thus stable) for 1 < a < 2, and positive (thus
unstable) for a > 2.
 182

We can therefore draw the bifurcation diagram:

In this case there are points of bifurcation at a = 1 and at a = 2.

Chaos
From a mathematical point of view there are three ingredients nec-
essary for a dynamical system to exhibit chaotic behaviour. We
will not give formal definitions of these ingredients, but the essen-
tial idea is to have a combination of attraction and repulsion. More
precisely we need:

1. An attractor G (as defined in the previous Chapter).

2. The attractor must be transitive. This means that if x(0) is

close to G then in the future x(t), or x(n), will get arbitrarily
close to every point of G. For example, if G is a stable limit
 183 MAT217—X/1
APM214—Y/1

cycle, or a stable n—cycle, then it is transitive. But if G

consists of two distinct fixed points then it is not transitive.

3. The crucial condition for chaotic behaviour is that the system

should have sensitive dependence on the initial values. That
is, for every point x(0) there are points y(0) arbitrarily close
to x(0), such that at some times in the future x(t) and y(t)
(or x(n) and y(n)) are not close to each other. An example
of such a system is

x(n + 1) = 2x(n) (4)

which has solution

x(n) = 2n x(0). (5)

Given initial values, say x0 , y0 , we find

|x(n) − y(n)| = |2n xo − 2n y0 | = 2n |x0 − y0 | → ∞ (6)

no matter how close the initial values x0 and y0 are to each

other. Thus this system has sensitive dependence on initial
values. But notice that it does not have an attractor G, and
it is not chaotic.

The following two examples illustrate the concept of chaos in dy-

namical systems theory.

1. We have already mentioned the logistic equation of population

growth. In discrete form it can be written as

x(n + 1) = x(n) + rx(n)(1 − x(n)/N ). (7)

 184

The dynamic system (7) has fixed points at 0 and N . We

have seen that the fixed point at 0 is unstable (for r > 0),
and that the fixed point at N is stable for 0 < r < 2. At
the end of the last Chapter we saw that for r > 2 the fixed
point at N becomes unstable, and instead the system develops
stable n−cycles. [We showed that if r = 2.2 there is a stable
2—cycle.]
Now what happens as r is further increased? First, it is con-
venient to re—write (7) as
x(n + 1) = λx(n)(1 − x(n)). (8)

This is simply a re—scaling and is brought about by writing

x(n) = ay(n) for suitable a. It turns out that λ = 1 + r. In
terms of λ the system does not have a stable (non—zero) fixed
point for λ ≥ 3; instead it may have stable n−cycles. For
√ √
3 < λ < 6 + 1 there is a stable 2—cycle, but for λ > 6 + 1
there is no stable 2—cycle; instead there may be a stable 4—
cycle. As λ increases we find that the order n of the n−cycle
must become larger and larger to ensure stability — n−cycles
with low n exist, but they are unstable. Finally, at λ = 3.57,
there are n−cycles (with n arbitrarily large) but they are all
unstable.
For 3.57 < λ < 4, the dynamic system (8) is chaotic.

2. Consider thermal convection of a horizontal layer of fluid heated

from below: the warm fluid may rise owing to its buoyancy and
circulate in cylindrical rolls. Under certain conditions these
cells are a series of parallel rotating cylindrical rolls; see the
figure on the next page.
 185 MAT217—X/1
APM214—Y/1

Lorenz used the continuity equation and Navier—Stokes equa-

tions from fluid dynamics, together with the heat conduction
equation to describe the behaviour of one of these rolls. A se-
ries of approximations and simplifications lead to the Lorenz
equations

ẋ = σ(y − x)
ẏ = rx − y − xz
ż = xy − bz. (9)

The term x represents the rate of rotation of the cylinder,

z represents the deviation from a linear vertical temperature
gradient, y corresponds to the difference in temperature at
opposite sides of the cylinder and σ, b and r are constants.
The non—linearity in the second and third equations results
from the non—linearity of the equations of flow.
With σ = 10, b = 83 , r = 28 (values often chosen for study)
the trajectories are concentrated onto an attractor of a highly
complex form. This Lorenz attractor consists of two “discs”
each made up of spiralling trajectories. Certain trajectories
leave each of the discs almost perpendicularly and flow into
the other disc. If a trajectory x(t) is computed, the following
behaviour is typical. As t increases, x(t) circles around one of
 186

the discs a number of times and then “flips” over to the other
disc. After a few loops round this second disc, it flips back to
the original disc. This pattern continues, with an apparently
random number of circuits before leaving each disc.

The above figure shows a view of the Lorenz attractor for σ = 10,
b = 83 , r = 28. Note the spiralling round the two discs and the
“jumps” from one disc to the other. Points that are initially close
together soon have completely different patterns of residence in the
two discs of the attractor. Thus the system exhibits sensitive de-
pendence on initial values, and is chaotic.

The Lorenz equations (9) describe a part of the Physics of the

Earth’s atmosphere. Since the Lorenz system is chaotic it follows
that so is the Earth’s atmosphere. This leads to an important prac-
tical conclusion: precise long—term weather prediction is impossible.
 187 MAT217—X/1
APM214—Y/1

Fractals
Consider the dynamic system

z(n + 1) = z(n)2 (10)

where z is complex, z = x + iy. The system could also be written

as
x(n + 1) = x(n)2 − y(n)2 , y(n + 1) = 2x(n)y(n),

but the behaviour is clearer in terms of complex z. The system (10)

has fixed points at
z∗ = 0, z∗ = 1 (11)

and it is straightforward to check that z∗ = 0 is stable and z∗ = 1

is unstable. There is also an invariant set at |z| = 1, because if
|z(n)| = 1 then |z(n + 1)| = 1; this invariant set is not an attractor,
but is unstable. It forms the boundary of the domain of stability
of the fixed point z∗ = 0. If |z(0)| < 1 then z(n) → 0; if |z(0)| > 1
then z(n) → ∞.

We now amend the dynamic system (10) to

z(n + 1) = z(n)2 + c, (12)

where c is a constant complex number, and again we are interested

in the domain of stability. The question can be formulated in two
different ways.

(a) For fixed c, find the domain of stability, i.e. find those points
z(0) such that |z(n)| is bounded for all n > 0. The boundary
of this domain of stability is called the Julia set. If c = 0,
the Julia set is the unit circle |z| = 1, and for small c one
 188

would expect that the Julia set would be a deformation of the

unit circle. In is indeed close to the unit circle, but whereas the
unit circle is a smooth curve, the Julia set for non—zero c is not
smooth. It has a very complex structure, and a magnification
of any region has approximately the same shape as the original.
It is a fractal. The definition of a fractal is beyond the scope
of this course, but the idea is illustrated in figure on page 13.

(b) For fixed z(0) (usually taken as z(0) = 0), find those values
of c for which |z(n)| is bounded for all n > 0. This is the
Mandelbrot set M . It is a bifurcation diagram in the sense
that as you cross the boundary of M you go from bounded to
unbounded solutions, or vice versa.
Often pictures of the Mandelbrot set, and of Julia sets, are
given in colour. From their popularity it is clear that these
pictures are having an influence on modern art and design.
The definition of the Julia and Mandelbrot sets is a yes/no
issue — either a point is in the set, or it isn’t. So how is colour
introduced into what is essentially a black/white picture? In
practice diagrams of the Julia and Mandelbrot sets are calcu-
lated by computer. Starting with z(0), we apply (12) repeat-
edly to find z(1), z(2), ..., z(N ), where N is some (previously
decided) maximum. If |z(n)| ≤ a, n = 0, ..., N , where a is
a constant usually taken as 2, then the trajectory starting at
z(0) is regarded as bounded, so that c ∈ M , or z(0) ∈ interior
of the Julia set. Otherwise the trajectory may be regarded as
escaping to infinity. In this case, let k be the first iterate such
that |z(k)| > a. The larger the value of k, the closer z(0) is to
the boundary. It is this value of k that is used to determine
 189 MAT217—X/1
APM214—Y/1

the colour in diagrams of the Mandelbrot and Julia sets.

Pictures of the Mandelbrot set suggest that it has a highly

complicated form — see the figure on the next page. It has
certain obvious features: a main cardioid to which a series of
prominent circular “buds” are attached. Each of these buds is
surrounded by further buds, and so on. However, this is not
all. In addition, fine, branched “hairs” grow outwards from
the buds, and these hairs carry miniature copies of the entire
Mandelbrot set along their length. It is easy to miss these
hairs in computer pictures. However, accurate pictures sug-
gest that M is a connected set, a fact that has been confirmed
mathematically.
 190

Julia sets for c at various points in the Mandelbrot

set.

Exercises

1. A simple fish harvesting model uses the logistic equation,

amended by assuming that a proportion b of the population is

 191 MAT217—X/1
APM214—Y/1

harvested in each time period:

 
x (n)
x(n + 1) = x(n) + 0.5 1 − x(n) − bx(n)
N
where N is the population in the absence of harvesting. Deter-
mine the fixed points and their nature (i.e. stable or unstable)
and draw a bifurcation diagram.

2. A discrete 2—dimensional dynamic system is described in terms

of polar coordinates (r, θ) :
1
θn+1 = θ2n , rn+1 = rn2

with θ ∈ [0, 2π). If θ2n ≥ 2π, say θ2n = ψ n + k(2π) with

ψ n ∈ [0, 2π) and k an integer, then we put θn+1 = ψ n . Show
that the unit circle r = 1 is an attractor. Show also that
the system is not transitive, thereby implying that it is not
chaotic.

3. Consider the dynamic system

dx
= 4x − x2 − b.
dt
Determine the fixed points and their nature and draw a bifur-
cation diagram.

4(a) A model of harvesting a fixed amount b of a population (that

would otherwise obey the logistic equation) is:

A(n + 1) = 1.8A(n) − 0.8A(n)2 − b.

For b > 0 find the fixed points of the system, and then draw
the bifurcation diagram. What limit is imposed on the amount
of harvesting?
 192

(b) State the dynamic equation that is used for generating the
Mandelbrot set, and then define the Mandelbrot set.
 193 MAT217—X/1
APM214—Y/1

APPENDIX

1. Notation
If A is an n × p matrix and B is an n × q matrix then [A, B] denotes
the n × (p + q) matrix obtained by placing B alongside and to the
right of A. For example, if

1 3
A=
5 4

and
4
B=
2
then
1 3 4
[A, B] = .
5 4 2
This notation is extended in an obvious way to more than two ma-
trices.

Similarly if A is an p × m matrix and B is a q × m matrix then

A
denotes the (p + q) × m matrix obtained by placing A above
B
B. Obviously, this notation can also be extended to more than two
matrices.
 194

2. Rank of a Matrix
The set of all column m—vectors
 
k1
 
 k2 
 
 k3 
 
 .. 
 . 
 
km
is denoted by Rm . (The vectors can also be written in row form, but
in this chapter we often think of the columns of a matrix as vectors,
so for us it is more convenient to use column vectors.) As explained
in module MAT103, we can multiply these vectors by scalars and
add them, so forming linear combinations of vectors. Thus if

x1 , x2 , x3 , ..., xn

are scalars and

a1 , a2 , a3 , a4 , ..., an
are n column m—vectors, we can form the linear combination

x1 a1 + x2 a2 + x3 a3 + ... + xn an .

If a set of x s, not all of them zero, can be found such that

x1 a1 + x2 a2 + x3 a3 + ... + xn an = 0

then the set of vectors {a1 , a2 , a3 , ..., an } is said to be linearly de-

pendent. Otherwise we say they are linearly independent.

Our interest in the concept of linear independence lies in the follow-

ing fact:
 195 MAT217—X/1
APM214—Y/1

If S is a subset of Rn then every vector in Rn can be expressed as a

linear combination of vectors in S if and only if S contains a subset
of n linearly independent vectors.

Now suppose we have a n × m matrix A. The columns of A can

be thought of as vectors in Rn . The rank of A is defined as the
maximum number of linearly independent columns of A. (It can
be shown that this is also equal to the maximum number of linearly
independent rows of A, where each row is thought of as a vector in
Rm .) In the proof of Theorem 1 we use the following result:

If A is an n × m matrix, then every vector in Rn can be expressed

as a linear combination of columns of A if and only if rank A = n.

It is important to be able to calculate the rank of a matrix. The

rank can be calculated by first reducing the matrix to row—echelon
form (this process is explained in the module MAT103) and then
counting the number of non—zero rows in the row—echelon matrix.
This number will be the rank of the matrix. For example, the
row—echelon form of the matrix

1 2
A=
4 8

is
1 2
0 0

and therefore rank A = 1.

 196

Another result which is sometimes useful is:

If A is an n × n matrix then rank A = n if and only if det A = 0.

3. Derivatives and Integrals of Matrices

The calculation of derivatives and integrals of matrices is very sim-
ple — you merely differentiate or integrate each entry in the matrix
using the calculus rules you learned in MAT101 and MAT102. For
example, if
1 t3 3
C= 2
, D =
t t sin t
then
d 0 d 3t3 0
C= , D= ,
dt 2t dt 1 cos t
while  
t1  
 1dt  t1
t1  0   
Cdt = 

=
 


0  t1  1 3
t2 dt t
3 1
0
and
 t1 t1   
t3 dt 3dt 1 4
  t 3t1
t1  0 0  
 4 1 

Ddt = 

=
  .
0  t1 t1  1 2 
tdt sin t dt t 1 − cos t1
0 0
2 1
4. Exponential Function
The exponential function ex , where x ∈ R, can be written as a power
series:
x2 x3 xk
ex = 1 + x + + + ... + +...
2! 3! k!
 197 MAT217—X/1
APM214—Y/1

which converges for all x ∈ R. If A is a square matrix, we define

A2 A3 A4 Ak
eA = I + A + + + + ... + + ...
2! 3! 4! k!
It can be shown that this series converges for any square matrix A.
If A is an n × n matrix then eA will also be an n × n matrix. In
this module you will never have to calculate the value of eA , all you
need to know is the above definition, and the following results:

If t ∈ R and A is a constant matrix, then

d At
e = AeAt = eAt A.
dt
This follows easily from the definition of eAt . We have

d At d A2 t2 A3 t3
e = I + At + + +...
dt dt 2! 3!
A3 t2 A4 t3
= A + A2 t + + + ...
2! 3!
A2 t2 A2 t2
= A I + At + + . . . or I + At + + ... A
2! 2!
= AeAt or eAt A.

It can also be shown that

(i) If A is a square matrix then eA e−A = e−A eA = I, and there-

fore the inverse of eA is e−A .

(ii) If s, t ∈ R then eAt eAs = eAs eAt = eA(s+t) .

 198

5. Solution of the Equation ẋ = Ax + Bu.

Multiply both sides of the equation ẋ = Ax + Bu on the left by
e−At . This gives

e−At ẋ(t) = e−At Ax(t) + e−At Bu(t)

or
e−At ẋ(t) − e−At Ax(t) = e−At Bu(t).

But
d −At
e x(t) = e−At ẋ(t) − e−At Ax(t)
dt
so we get
d −At
e x(t) = e−At Bu(t).
dt
Therefore
t
e−At x(t) = e−As Bu(s)ds.
a

Multiply both sides of this equation on the left by eAt to get

t t
x(t) = eAt e−As Bu(s)ds = eA(t−s) Bu(s)ds.
a a

If we assume that x(0) = 0 then we can take a = 0 and we get

t
x(t) = eA(t−s) Bu(s)ds.
0