Difference and Differential Equations: BA Econometrics and Operations Research

Difference and Differential equations
2013-2014
BA Econometrics and Operations Research
Module code EBB812A05
Author:
G.K. Immink
Contents
1 Difference equations 5
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.1 Asymptotic behaviour and stability . . . . . . . . . . . 13
1.1.2 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.2 Autonomous first order difference equations . . . . . . . . . . 19
1.2.1 Stability of periodic and equilibrium solutions of au-
tonomous first order difference equations . . . . . . . . 24
1.2.2 Bifurcations . . . . . . . . . . . . . . . . . . . . . . . . 32
1.2.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.3 Linear difference equations . . . . . . . . . . . . . . . . . . . . 38
1.3.1 Inhomogeneous linear difference equations . . . . . . . 44
1.3.2 First order linear difference equations . . . . . . . . . . 44
1.3.3 Stability of solutions of linear difference equations . . . 48
1.3.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 50
1.4 Systems of first order difference equations . . . . . . . . . . . 51
1.4.1 Homogeneous systems of first order linear difference
equations . . . . . . . . . . . . . . . . . . . . . . . . . 56
1.4.2 Inhomogeneous systems of first order linear difference
equations . . . . . . . . . . . . . . . . . . . . . . . . . 58
1.4.3 Systems of first order linear difference equations with
constant coefficients . . . . . . . . . . . . . . . . . . . . 58
1.4.4 Stability of solutions of systems of linear difference
equations with constant coefficients . . . . . . . . . . . 67
1.4.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 69
1.5 Autonomous systems of first order difference equations . . . . 71
1.5.1 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3
1.6 Higher order linear difference equations with constant coeffi-
cients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
1.6.1 The homogeneous equation . . . . . . . . . . . . . . . 87
1.6.2 Inhomogeneous higher order difference equations; the
annihilator method . . . . . . . . . . . . . . . . . . . . 90
1.6.3 Stability of solutions of higher order difference equations 94
1.6.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 97
2 Differential equations 101

2.1 First order scalar differential equations (supplement to S&B
§24.2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
2.1.1 Solving equations of type 1 . . . . . . . . . . . . . . . . 103
2.1.2 Separable equations (type 2) . . . . . . . . . . . . . . . 104
2.1.3 Homogeneous linear equations (type 3) . . . . . . . . . 106
2.1.4 Inhomogeneous linear equations (type 4) . . . . . . . . 107
2.1.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 110
2.2 Systems of first order differential equations . . . . . . . . . . . 111
2.2.1 Systems of first order linear differential equations . . . 114
2.2.2 Homogeneous linear vectorial differential equations with
constant coefficients . . . . . . . . . . . . . . . . . . . . 115
2.2.3 Inhomogeneous linear vectorial differential equations
with constant coefficients . . . . . . . . . . . . . . . . . 126
2.2.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 128
2.3 Stability of solutions of systems of first order differential equa-
tions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
2.3.1 Autonomous systems of nonlinear differential equations 133
2.3.2 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Appendix 143
A.1 Change of basis and Jordan normal form . . . . . . . . . . . . 143
A.2 Computation of J n . . . . . . . . . . . . . . . . . . . . . . . . 150
A.3 Definition and properties of eA . . . . . . . . . . . . . . . . . . 151
A.4 Vector and matrix functions . . . . . . . . . . . . . . . . . . . 153
The figures in these lecture notes were produced with the program Dynamics
of H. E. Nusse and J. A. Yorke.
Chapter 1
Difference equations
1.1 Introduction
Difference equations arise in discrete, dynamic economic models.
Example 1.1.1. One of the simplest examples is the classic cobweb model:
qtd = D(pt )
qts = S(pt−1 )
qtd = qts
qtd , qts , and pt denote demand, supply and price at time t, respectively. From
the above equations we deduce the following relationship between the price
at time t and that at time t − 1:
D(pt ) = S(pt−1 )
This is a first order difference equation, with independent variable t and

dependent variable p. A solution of the equation is a function p of t, satisfying
this relationship. If we consider discrete times, say t = 0, 1, ..., etc., then a
solution is a sequence of numbers: {p0 , p1 , ...}, such that D(p1 ) = S(p0 ),
D(p2 ) = S(p1 ), etc.. The tth element of the sequence represents the price
at time t (or in the period t). We can write the equation in the following,
equivalent, form:
D(pt+1 ) = S(pt )
5
or
D(pt+1 ) − S(pt ) = 0
If the demand function D is invertible, with inverse D−1 , then pt+1 can be
expressed explicitly in terms of pt :
pt+1 = D−1 ◦ S(pt )
If the price p0 at time t = 0 is known, the prices at subsequent times can be

computed successively: p1 = D−1 ◦S(p0 ), p2 = D−1 ◦S(p1 ), etc. The simplest
case is when demand and supply are linear, time-independent, functions of
p: D(p) = ap + b, S(p) = cp + d, where a < 0 and c > 0. In that case p
satisfies the linear difference equation
c d−b
pt+1 = pt +
a a
In mathematics, it is customary to place the independent variable in brackets
after the dependent variable: p(t) instead of pt .
Example 1.1.2. A neoclassical growth model in discrete time.
Yt = Q(Kt , Lt )
Kt+1 − Kt = sYt
Lt+1 − Lt = nLt
Here, Y , K and L denote production, capital and labour, respectively. The
rate of growth n of the labour force is a positive constant and the propensity
to save s is a number between 0 and 1. Q is a production function, assumed
homogeneous of degree 1. We can write the model in per capita form by
defining k := KL
, (capital per worker), and
Q(K, L) K
q(k) = = Q( , 1) = Q(k, 1)
L L
From the above equations we infer that k must satisfy the first order difference
equation
s 1
kt+1 = q(kt ) + kt (1.1)
1+n 1+n
6
Technical progress can be introduced into the model by replacing the first
equation with
Yt = Q(Kt , Et Lt )
Here, E is a parameter, representing the increase in labour efficiency due to
technological improvement. E is an increasing function of t, and E0 = 1.
K
Defining k := EL and
Q(K, EL) K
q(k) = = Q( , 1)
EL EL
we find that k satisfies the difference equation
Et s 1

kt+1 = q(kt ) + kt (1.2)
Et+1 1+n 1+n
Et+1 Et+1
If Et
is constant: Et
= 1 + g, with g > 0, we obtain
s 1
kt+1 = q(kt ) + kt (1.3)
(1 + n)(1 + g) (1 + n)(1 + g)
We begin by giving a general definition of a difference equation. A (scalar)

difference equation is an equation of the form
F (x, y(x), y(x + h), ..., y(x + kh)) = 0 (1.4)
Here, k is a natural number, the order of the difference equation, x is the

independent variable, which can be real or complex, and is often discrete. h is
a real or complex number: the difference or the shift, and F is a function of
k+2 variables, defined on a subset DF of Rk+2 or Ck+2 with the property that
(x, y1 , ..., yk+1 ) ∈ DF implies (x + jh, y1 , ..., yk+1 ) ∈ DF for all j ∈ {1, ..., k}.
y is the dependent variable, i.e. an unknown function of x that has to be
determined by solving the equation. The equation is autonomous if F
doesn’t depend explicitly on x, so that the equation can be written in the
form
F (y(x), y(x + h), ..., y(x + kh)) = 0
The equation in example 1.1.1 (the cobweb model) is autonomous, as we
assumed that D and S are functions of the price only and do not depend
explicitly on the time. Here, h = 1 and F (pt , pt+1 ) = D(pt+1 ) − S(pt ).
7
A difference equation is usually called linear if F is an affine function
of the last k + 1 variables, so a linear equation has the form
b(x) + a0 (x)y(x) + a1 (x)y(x + h) + ... + ak (x)y(x + kh) = 0
Frequently, the shift is normalized to either 1 or −1. This can be achieved

by the following simple transformation: choose a new independent variable
x
t=
h
and a new dependent variable u defined by
u(t) := y(x) = y(th)
Then we get
y(x + jh) = y((t + j)h) = u(t + j) j = 1, ..., k
Substituting this into (1.4), we obtain the equation
F (th, u(t), u(t + 1), ..., u(t + k)) = 0
Example 1.1.3. The equation
y(x + h) − y(x)
=3 (1.5)
h
is transformed into
u(t + 1) − u(t)
=3
h
by the change of variables
x
x 7→ t := , y(x) 7→ u(t) := y(x) = y(th)
h
All real-valued solutions of this equation have the form
u(t) = c + 3ht, where c ∈ R
Hence the solutions of (1.5) are given by the general expression
y(x) = c + 3x, where c ∈ R
8
The difference quotient on the left-hand side of (1.5) can be used to approx-
dy
imate dx . More generally, solutions of the difference equation
y(x + h) − y(x)
= f (x, y(x))
h
are used to approximate solutions of the differential equation
dy
= f (x, y(x))
dx
The simplest numerical method to “compute” solutions of differential equa-
tions, Euler’s method, is based on this idea.
From now on, we take the shift equal to 1 and work with a discrete variable
n ∈ Z. A solution of the difference equation
F (n, y(n), y(n + 1), ..., y(n + k)) = 0 (1.6)
is some function y, defined on an appropriate subset D of Z, such that (1.6)

holds for all n ∈ D. In this context, ‘appropriate subset’ refers to a set
D ⊂ Z with the property that n ∈ D implies n + j ∈ D for all j ∈ {1, ..., k}
and (n, y(n), ..., y(n + k)) ∈ DF . Usually, DF = Rk+2 or Ck+2 , and D = Z
or D is the set of all nonnegative integers, to be denoted by N. Often,
we note the equation as follows
F (n, y(n), y(n + 1), ..., y(n + k)) = 0, n∈D (1.7)
to indicate that the solutions we seek are defined on D. The following exam-
ple demonstrates the importance of this addition.
Example 1.1.4. The set of all solutions of the difference equation
y(n + 1) − ny(n) = 0, n∈N
consists of all functions y on N with the property that y(n) = 0 for all n ≥ 1.
y(0) can take on every real (or complex) value. However, the equation
y(n + 1) − ny(n) = 0, n ∈ {1, 2, ...} (1.8)
is satisfied by all sequences of the form y(n) = c(n − 1)!, c ∈ R or C.
9
Figure 1.2: 4-periodic solution of the equation y(n + 1) = 3.5y(n)(1 − y(n)),
n∈N
A function y on a discrete set D ⊂ Z or N generates the sequence {y(n)}n∈D .

In many textbooks and papers, the notation yn is used instead of y(n).
An equilibrium solution of (1.7) is a function y on D (or a sequence
{yn }n∈D ), that solves the equation and has the same value for all n ∈ D.
Thus, there exists a number c such that y(n) = c (yn = c) for all n ∈ D.
Such a solution is also called constant or stationary solution and is noted
y ≡ c. A sequence consisting entirely of zeroes is called a null sequence
and a null sequence satisfying the equation (1.7) is called a null solution
of (1.7). It is a particular case of an equilibrium solution. The equilibrium
solutions of (1.7) can be found by solving the equation F (n, c, ..., c) = 0 for c.
A periodic or cyclic solution is a solution y with the property that there
exists a positive integer p such that y(n + p) = y(n) for all n ∈ D. In that
case p is called a period of the solution. It is not unique, as any multiple of
p is again a period of the solution. The minimum period is referred to as
the period of the solution and here we will call a solution p-periodic if its
minimum period is p. An equilibrium solution can be viewed as a periodic
solution with period 1 (note that this is not true for difference equations with
a continuous variable). The graph of a solution of (1.7) consists of a discrete
set of points, cf. figs. 1.1 and 1.2. It is not uncommon to connect the distinct
points of the graph by line segments (cf. figs. 2.6 through 2.8).
y(n + 1) + ny(n) = (n − 1)(−1)n + 2(n + 1), n ∈ N
has no equilibrium solutions. Any 2-periodic solution must satisfy the system
consisting of the above equation and the equation y(n + 2) = y(n). For every
10
solution of the first equation we have
y(n + 2) = −(n + 1)y(n + 1) + n(−1)n+1 + 2(n + 2)
= −(n + 1)(−ny(n) + (n − 1)(−1)n + 2(n + 1)) − n(−1)n + 2(n + 2)
= n(n + 1)y(n) − (n2 + n − 1)((−1)n + 2)
Replacing y(n + 2) by y(n) on the left-hand side and rearranging terms, we

get
(n2 + n − 1)(y(n) − (−1)n − 2) = 0
As n2 + n − 1 6= 0 for all n ∈ N, this implies y(n) = (−1)n + 2. Hence the
equation has a unique 2-periodic solution: the sequence {(−1)n + 2}∞
n=0 .
The general solution of an equation is the set of all solutions. Even

though, in economic applications, one is usually interested in real-valued
solutions, the study of the set of all complex-valued solutions has certain
advantages. The real solutions obviously form a subset of the set of all
complex solutions. Thus, for example,
{y(n) = c(n − 1)!, n ∈ {1, 2, ...} : c ∈ C}, or {y = {c(n − 1)!}∞

n=1 : c ∈ C}
is the general (complex) solution of (1.8). In what follows we will usually

omit the braces. One specific (‘particular’) solution y is obtained by choosing
a specific (real or complex) value for the parameter c. There are, however,
many equations whose general solution cannot be written in a simple (so-
called “closed”) form.
We introduce a shift operator τ . This is a linear operator on the set of
all real- or complex-valued sequences on Z or N, defined by
τ y(n) := y(n + 1) ( or τ yn = yn+1 )
The inverse of this operator exists if we consider functions (sequences) on Z

and it is defined by
τ −1 y(n) = y(n − 1) ( or τ −1 yn = yn−1 )
Alternative notations for (1.6) that can be found in the literature are
F (n, yn , yn+1 , ..., yn+k ) = 0
and
F (n, y, τ y, ..., τ k y) = 0 (1.9)
11
A special case of (1.6) is the so-called recurrence relation. This is an equation
of the form
y(n + k) = f (n, y(n), y(n + 1), ..., y(n + k − 1)), n ∈ {n0 , n0 + 1, ...} (1.10)
Here, n0 is an integer and f is a function of k + 1 variables. For convenience,

let’s assume the domain of f is the entire space Ck+1 . In that case, the first
k elements y(n0 ), y(n0 + 1), ..., y(n0 + k − 1) of the sequence y can be taken to
be arbitrary complex numbers c1 , ..., ck and once they have been chosen, the
subsequent elements y(n0 + k), y(n0 + k + 1), etc., are determined by (1.10).
The numbers c1 , ..., ck are called the initial values of the solution and the
problem of finding a solution with given initial values c1 , ..., ck is an initial
value problem. Obviously, a solution of this type of equation is uniquely
determined by its k initial values c1 , ..., ck .
Example 1.1.6. The solution y of the equation
y(n + 2) = y(n)2 + 1, n∈N
with initial values: y(0) = 1, y(1) = 0, can be determined as follows: y(2) =

y(0)2 + 1 = 2, y(3) = y(1)2 + 1 = 1, etc.
Many equations arising in economic models, contain, besides the dependent

and independent variable(s), so-called parameters: additional quantities,
whose values can vary from case to case. In the case of n parameters, the
collection of equations obtained by varying these parameters, is called an n-
parameter family of equations. The solutions of an individual equation from
such a family obviously depend on the values of the parameters.
Example 1.1.7. The equations of the type
y(n + 1) = y(n)2 − a, n ∈ N
where a is allowed be any real number, form a 1-parameter family of nonlinear

difference equations. The equilibrium solutions are obtained by solving the
quadratic equation y 2 − y − a = 0. For each value of a 6= − 14 , the latter
equation has two complex solutions y1 and y2 , which are real when a > − 14 :
1 1√
y1,2 = ± 1 + 4a
2 2
12
and non-real when a < − 14 :
1 1 √
y1,2 = ± i −1 − 4a
2 2
If, in the above example, we restrict ourselves to real solutions, then the
equation has no equilibrium solutions for a < − 41 , a single equilibrium solu-
tion y ≡ 12 for a = − 14 and two equilibrium solutions for a > − 41 . At the
parameter-value a = − 41 , an abrupt change in the properties of the equation
occurs. Such qualitative changes, due to a change in a parameter are called
bifurcations. The value of the parameter(s) at which the bifurcation occurs
is called the bifurcation value of that parameter. There are many different
types of bifurcations, only a few of which will be reviewed here. The bifur-
cation in example 1.1.7 is a so-called ‘saddle-node bifurcation’. Bifurcations
can be represented in a bifurcation diagram, where certain characteristic
properties of the equation, such as the equilibrium values, are plotted versus
a parameter.
1.1.1 Asymptotic behaviour and stability

In most applications, the initial values of solutions of equations like (1.10),
are not known exactly and thus the subsequent values cannot be computed
with absolute accuracy either. Therefore, it is important to know how a
small change in the initial values affects the behaviour of the corresponding
solution. In many problems, our main concern is the long term behaviour
of solutions, i.e. their behaviour for large values of n. This is called the
asymptotic behaviour of the solution as n → ∞. There are many different
types of asymptotic behaviour. Some solutions approach a fixed limiting
value as n → ∞. More generally, a solution may approach a particular
sequence, in the sense that the difference between both sequences tends to
0. In that case, the two sequences are said to be asymptotically equal. If
one of the two sequences is periodic, the other one is called ‘asymptotically
periodic’. A solution may be unbounded, or it may remain bounded, but
display a very unpredictable, so-called chaotic behaviour.
While the asymptotic behaviour of a solution is a property of that specific
solution alone, the stability of a solution depends on the behaviour of other
solutions, namely those whose initial values are very close to those of the given
13
solution. A solution of the equation is said to be stable, if a small change
(‘perturbation’) of the initial values causes a (sufficiently) small change in
the entire solution, and is said to be asymptotically stable if, in the long
run, the effect of this change is no longer noticeable. Here are the precise
definitions:
Definition 1.1.8. A solution y of equation (1.10) is stable if, for every
positive number , there exists a positive number δ, such that every solution
ỹ, whose first k values do not differ more than δ from those of y, i.e.
|ỹ(n0 + j) − y(n0 + j)| ≤ δ for j = 0, ..., k − 1,
has the property that, for all n ≥ n0 ,
|ỹ(n) − y(n)| ≤
A stable solution y is called asymptotically stable or locally asymptoti-
cally stable, if there exists a positive number δ, such that, for every solution
ỹ with the property that
|ỹ(n0 + j) − y(n0 + j)| ≤ δ for j = 0, ..., k − 1
the following holds
lim (ỹ(n) − y(n)) = 0
n→∞
A stable solution y is globally asymptotically stable if
lim (ỹ(n) − y(n)) = 0
n→∞
for every solution ỹ. A stable solution that is not asymptotically stable, is
called neutrally stable. A solution of (1.10) that is not stable, is called
unstable.
If one is exclusively interested in real-valued solutions, then one only needs
to consider the effect of real-valued changes in the initial values, in order to
determine the stability of a solution. Cf. fig. 1.4 in §1.2. There, the null
solution (like all other solutions with initial value in the interval (−1, 2)) is
locally, but not globally asymptotically stable, whereas the solution y ≡ −1
is unstable.
Stability and instability are properties of solutions of equations and not
of individual sequences. Any particular sequence may be a stable solution of
one difference equation and an unstable solution of another equation, as is
demonstrated by the following example.
14
Example 1.1.9. The null sequence is a globally asymptotically stable solu-
tion of the equation
1
y(n + 1) − y(n) = 0, n ∈ N,
n+1
since, for every solution of this equation, the following relation holds
y(0)
y(n) =
n!
Hence limn→∞ y(n) = 0. On the other hand, the null sequence is an unstable
solution of
y(n + 1) − (n + 1)y(n) = 0, n ∈ N,
as, in this case, y(n) = n!y(0) for every solution of the equation, and thus
limn→∞ |y(n)| = ∞ when y(0) 6= 0. Finally, the null sequence is a neutrally
stable solution of the equation
y(n + 1) − y(n) = 0, n∈N
(Why?).
Example 1.1.10. Newton’s method is used to approximate zeroes of func-

tions by means of an iterative procedure. The difference en between the exact
value and the approximation, after n iterations, satisfies an equation of the
form
en+1 = φ(n)e2n , n ∈ N
The method converges if, for an appropriate choice of the starting value of the
algorithm, i.e. for sufficiently small e0 , this difference tends to 0 as n → ∞.
This happens precisely then, when the null solution of the above equation is
asymptotically stable. Obviously, this will depend on the function φ. Let’s
consider the simple case that φ ≡ 1 and replace en by y(n). Then the above
equation reads:
y(n + 1) = y(n)2 , n∈N (1.11)

We claim that the null solution of (1.11) is asymptotically stable. The null
solution is clearly stable, as |y(n+1)| = |y(n)|2 ≤ |y(n)|, provided |y(n)| ≤ 1.
Let > 0. An inductive argument shows that |y(n) − 0| ≤ , for all n ∈ N,
if |y(0) − 0| ≤ δ := min{1, }. It remains to prove the existence of a positive
15
number δ, such that limn→∞ y(n) = 0 for every solution y of (1.11), with the
property that |y(0)| ≤ δ. Let δ < 1. Another inductive argument shows that
|y(n)| ≤ δ n+1 for all n ∈ N, if |y(0)| ≤ δ. Suppose that |y(n)| ≤ δ n+1 for a
certain n ∈ N. Then
|y(n + 1)| = |y(n)|2 ≤ δ 2n+2 ≤ δ n+2
So the inequality is valid for n + 1 as well, and thus it is valid for all n ∈ N.
As limn→∞ δ n = 0, this implies that limn→∞ y(n) = 0.
The equation (1.11) has a second equilibrium solution, viz. the sequence
y1 (n) = 1 for all n ∈ N. This solution is unstable. This can be proved in
the following way. Let δ > 0 and consider the solution yδ with initial value
yδ (0) = 1 + δ. Thus, |yδ (0) − y1 (0)| = δ. A simple inductive argument shows
that
|yδ (n)| ≥ (1 + δ)n+1
for all n ∈ N. Hence it follows that limn→∞ yδ (n) = ∞. This implies that
|yδ (n) − y1 (n)| = |yδ (n) − 1| can take on arbitrarily large values. Since this
holds for any δ > 0, the condition for stability of the solution y1 is not
fulfilled. Note that the existence of a second equilibrium solution implies
that the null solution is not globally asymptotically stable!
Sometimes, by a change of variables, a difference equation can be simplified,
or reduced to an equation whose properties are already known. Thus, for
example, the change of variables: n = m + n0 , y(n) = y(m + n0 ) = z(m),
transforms the equation
F (n, y(n), ..., y(n + k)) = 0, n ∈ {n0 , n0 + 1, ...}
into
G(m, z(m), ..., z(m+k)) := F (m+n0 , y(m+n0 ), ..., y(m+n0 +k)) = 0, m ∈ N,
which is of the same type as the original equation, but now the independent
variable runs from 0 to ∞. This is why, from now on, we will usually take
n0 = 0. Another useful transformation is the following. Suppose that a
specific solution y∗ of equation (1.6) is known. Now substitute y = y∗ + z
into the equation. Then y is a solution of (1.6) iff the new dependent variable
z satisfies the equation
G(n, z(n), ..., z(n + k)) := F (n, z(n) + y∗ (n), ..., z(n + k) + y∗ (n + k)) = 0
A particular property of the new equation is that it possesses a null solution.
16
Lemma 1.1.11. The sequence y∗ is an asymptotically stable, neutrally sta-
ble, or unstable solution of the equation
y(n + k) = f (n, y(n), y(n + 1), ..., y(n + k − 1)), n∈N (1.12)
iff the null sequence is, respectively, an asymptotically stable, neutrally sta-
ble, or unstable solution of the equation
z(n + k) = g(n, z(n), z(n + 1), ..., z(n + k − 1)) :=
f (n, z(n) + y∗ (n), ..., z(n + k − 1) + y∗ (n + k − 1)) − y∗ (n + k), n∈N
(1.13)
Proof. Suppose, for example, that y∗ is a stable solution of (1.12). We then

have to show that the null sequence is a stable solution of (1.13). Let be any
positive number. Due to the stability of y ∗ , there exists a positive number δ
such that, if ỹ is a solution of (1.12) with the property that
|ỹ(n) − y∗ (n)| ≤ δ for n = 0, ..., k − 1, (1.14)
then we have, for all n ≥ 0,
|ỹ(n) − y∗ (n)| ≤
Now, let z̃ be a solution of (1.13) with the property that
|z̃(n) − 0| ≤ δ for n = 0, ..., k − 1
Then the sequence ỹ defined by: ỹ = z̃ + y∗ is a solution of (1.12) satisfying

(1.14), and thus |z̃(n)| = |ỹ(n) − y∗ (n)| ≤ for all n ≥ 0. This shows that
the null sequence is a stable solution of (1.13). The proofs of the converse
and the remaining statements are left to the reader.
Example 1.1.12. As we have seen, y∗ ≡ 1 is an unstable solution of the

equation y(n + 1) = f (y(n)) := y(n)2 . By lemma 1.1.11 this implies that
z ≡ 0 is an unstable solution of z(n + 1) = g(z(n)) := f (z(n) + 1) − 1 =
(z(n) + 1)2 − 1 = z(n)2 − 2z(n).
Now let y∗ be defined by y∗ (n) = n. We know that z ≡ 0 is an asymp-
totically stable solution of the equation z(n + 1) = g(z(n)) := z(n)2 . Again
by lemma 1.1.11, it follows that y∗ is an asymptotically stable solution of
y(n+1) = f (y(n)) := g(y(n)−y∗ (n))+y∗ (n+1) = y(n)2 −2ny(n)+n2 +n+1.
17
1.1.2 Exercises
1. Show that the solutions of equation (1.11) do not form a linear space.
2. Find all real, periodic solutions of the equation
y(n + 1) = y(n)2 , n∈N
3. Prove that all solutions of the equation

y(n + 2) − y(n) = 0, n∈N
are stable. Does this equation have any asymptotically stable solutions?
4. Determine the equilibrium solutions of the equation
y(n + 1) = αy(n)(1 − y(n)), n∈N
where α ∈ R.
5. Determine all equilibrium solutions of the equation
y(n + 2) = y(n) + 2hy(n + 1)(1 − y(n + 1)), n∈N
where h ∈ R.
6. Determine the equilibrium solutions of the equations (1.1) and (1.3) in
example 1.1.2, when Q is a Cobb-Douglas function of the form: Q(x, y) =
xα y 1−α , with 0 < α < 1.
y(n + 1) = 2y(n)2 , n∈N
and examine their stability.
y(n + 1) = 2y(n) + y(n)2 , n∈N
and examine their stability, using lemma 1.1.11.
9. Prove that the null solution of the equation
y(n + 1) = φ(n)y(n)2 , n∈N
is asymptotically stable if φ is bounded on N.
18
1.2 Autonomous first order difference equa-
tions
The most general first order difference equation of the type (1.10), with
n0 = 0, is
y(n + 1) = f (n, y(n)), n ∈ N
We restrict ourselves in this section to autonomous equations. These are of
the form
y(n + 1) = f (y(n)), n ∈ N (1.15)
(1.11) is an example of such an equation. Another example is the logistic
difference equation:
y(n + 1) = αy(n)(1 − y(n)) (1.16)
Here, α is a positive parameter. Equations of the type (1.15) play a role in

the determination of zeroes or fixed points of nonlinear functions by means
of an iterative procedure.
A subset G of the domain Df of f is said to be positive invariant if
f (G) ⊂ G, and invariant if f (G) = G.
From now on, we assume that Df itself is positive invariant. Then, for every
k ∈ N, the kth iterate of f , denoted by f k , is defined by
f 0 (x) = x, f k (x) = f (f k−1 (x)) if k ≥ 1
Thus,
f 1 (x) = f (x), f 2 (x) = f (f (x)), etc.
If y is a solution of (1.15), then, for every k ∈ N,
y(n + k) = f k (y(n))
The sequence y ≡ c is an equilibrium solution of (1.15) iff c is a fixed point

of f , i.e. iff f (c) = c. If y is a periodic solution of (1.15), of period p, then
y(0),..., y(p − 1) are fixed points of f p . Conversely, if c is a fixed point of f p ,
then, for every natural number k, f k (c) is a fixed point of f p as well (check
this), and thus the solution of (1.15) with initial value c is periodic of period
p: y(pn + k) = f pn+k (c) = f k (c) for k = 0, ..., p − 1 and all n ∈ N.
19
Example 1.2.1. f (x) = x2 (cf. (1.11)).
k
f k (x) = x2 for every k ∈ N. f has two fixed points: 0 and 1, both of which
are real. f 2 has four fixed points: the solutions of the 4th degree equation
x4 − x = 0
Besides the two (real) fixed points of f , this equation has two additional,
complex conjugated, non-real solutions:
2π 1 1 √
x1,2 = e± 3 i = − ± i 3
2 2
So, in this case, the difference equation (1.15) has two complex-valued peri-
odic solutions of period 2.
The positive orbit γ + (x0 ) of x0 ∈ Df subject to the map f is the set
γ + (x0 ) := {f n (x0 ), n ∈ N}
The positive (or forward) orbit of x0 , subject to the map f , is the set of all
elements of the sequence y with initial value x0 , that satisfies (1.15). We
will also call this the (positive) orbit of y. Every positive orbit is positive
invariant. The positive orbit of an equilibrium solution consists of one point,
and is also called an equilibrium point of the equation. The positive
orbit of a p-periodic solution is a set consisting of p elements, called a p-
periodic orbit. Different solutions of (1.15) may have the same positive orbit,
for instance, the positive orbit of both complex 2-periodic solutions of the
equation y(n + 1) = y(n)2 is the set {x1 , x2 } (cf. example 1.2.1). It is easily
seen that, in general, to each p-periodic orbit correspond exactly p p-periodic
solutions having this orbit (check this).
Example 1.2.2. In the case of the logistic difference equation, f (x) =
αx(1 − x) and
f 2 (x) = α2 x(1 − x)(1 − αx + αx2 )
In order to find the real-valued, periodic solutions of (1.16) with minimal
period 2, we begin by computing all real fixed points of f 2 , i.e. the solutions
of the equation
α2 x(1 − x)(1 − αx + αx2 ) = x
After eliminating the fixed points of f (cf. §1.1.2, exerc. 4), we find that the
remaining fixed points of f 2 , if there are any, should satisfy the equation
α2 x2 − α(α + 1)x + α + 1 = 0
20
If α > 3, this equation has two distinct, real-valued solutions x1 and x2 :
q
α+1± (α + 1)(α − 3)
x1,2 =
2α
Check that f (x1 ) = x2 and f (x2 ) = x1 . So, if α > 3, (1.16) has two real,
periodic solutions of minimal period 2, viz. the sequences x1 , x2 , x1 , x2 , ....
and x2 , x1 , x2 , x1 ..... Both have the same orbit: γ + (x1 ) = γ + (x2 ) = {x1 , x2 }
The following simple, graphical method serves to ‘determine’ the solution of

an equation of the type (1.15), with a given, real, initial value y0 (cf. figure
1.3). In an X − Y -coordinate system, draw the graph of the function f
and the line y = x. The first element of the sequence we are looking for
is y0 . The next element of the sequence, y(1), has the value f (y0 ), which
is found by intersecting the graph of f with the vertical line x = y0 . Now
draw a horizontal line through the point (y0 , f (y0 )) and determine the point
(y(1), y(1)), where it intersects the line y = x. The vertical line through
this point intersects the graph of f at (y(1), f (y(1))) = (y(1), y(2)), etc...
The sequence formed by the x-coordinates of the points of the graph of f ,
determined in this way, is the required solution of (1.15). This method can
also give us a first indication of the stability or instability of equilibrium
solutions of the equation. The equilibrium solutions themselves correspond
to the fixed points of f , i.e. the points where the graph of f intersects the
line y = x.
A point x ∈ Df is called a positive limit point or ω-limit point of x0 ,
if there exists a subsequence of {f n (x0 )}∞
n=0 converging to x. The positive
limit set or ω-limit set ω(x0 ) of x0 is the set of all positive limit points of
x0 .
Example 1.2.3. In the case that f (x) = x2 , x ∈ R, ω(x0 ) = {0} if x0 ∈

(−1, 1), ω(−1) = ω(1) = {1} and ω(x0 ) = ∅ for all other (real) values of x0 .
If {f n (x0 )}∞ +
n=0 is a periodic solution of (1.15), then ω(x0 ) = γ (x0 ). If
n
limn→∞ f (x0 ) = c, then ω(x0 ) = {c}, but the converse is not true. If
ω(x0 ) = {c} and γ + (x0 ) is unbounded, then limn→∞ f n (x0 ) does not exist
(cf. ex. 1.2.3; in such cases, it is sometimes said that ∞ is a positive limit
point of x0 ).
21
Figure 1.3: graphical determination of the first 8 elements of the solution of
the logistic difference equation (2.2), with α = 2.8, y(0) = 0.1.
The stable set of a fixed point c of f is the set of all x ∈ Df with the
property that limn→∞ f n (x) = c.
A closed, invariant set A is called an attracting set for f , if there exists
a neighbourhood U of A, such that the distance between f n (x) and A (i.e.
miny∈A |f n (x) − y|) tends to 0 as n → ∞, for all x ∈ U .
The basin of attraction of an attracting set A is the set of all x ∈ Df
such that the distance between f n (x) and A tends to 0 as n → ∞.
An attractor is an attracting set A, containing a “dense orbit”, i.e. there
exists an x ∈ A such that A = γ + (x).
The simplest example of an attractor is an asymptotically stable equi-
librium point c. If A = {c}, then A is closed and invariant (as f (c) = c).
The distance between f n (x) and A is |f n (x) − c|. By definition, there exists
a neighbourhood U of c, hence of A, such that limn→∞ f n (x) = c for all
x ∈ U . Moreover, γ + (c) = {c}, so A = γ + (c). The basin of attraction of {c}
coincides with its stable set.
Example 1.2.4. The basin of attraction of the fixed point 0, of the logistic
map
f (x) = αx(1 − x)
where α ∈ (0, 1), is the interval (1 − 1/α, 1/α) (cf. fig. 1.4).
Proof: Let x ∈ R and y(n) := f n (x), n ∈ N. We distinguish several cases.
22
Figure 1.4: solutions of the logistic difference equation, with α = 1/2, with
different initial values, varying from −2.2 to 2.5. Observe the asymptotic
behaviour of the different solutions. The basin of attraction of 0 is the interval
(−1, 2) (on the y-axis).
1. y(0) (= x) ≥ 1/α. This implies that αy(0) ≥ 1 and 1−y(0) ≤ 1−1/α < 0,
hence y(1) = αy(0)(1 − y(0)) ≤ 1 − 1/α. From here on, proceed as in case 2
below (starting from y(1) instead of y(0)).
2. y(0) ≤ 1 − 1/α. As an inductive argument shows, this implies that y(n) ≤
1−1/α for all n ∈ N. For, suppose that for some n ∈ N, y(n) ≤ 1−1/α (< 0),
then we have α(1−y(n)) ≥ 1, so y(n+1) = α(1−y(n))y(n) ≤ y(n) ≤ 1−1/α.
From 1. and 2. it follows that none of the points examined so far belongs to
the basin of attraction of 0, as in both cases y(n) ≤ 1 − 1/α for all n ≥ 1,
hence limn→∞ y(n) 6= 0. Thus, the basin of attraction of 0 must be contained
in the interval (1 − 1/α, 1/α).
3. 0 ≤ y(0) ≤ 1. If 0 ≤ y(n) ≤ 1 for some n ∈ N, then 0 ≤ 1−y(n) ≤ 1, and,
consequently, 0 ≤ y(n + 1) = αy(n)(1 − y(n)) ≤ αy(n). With the principle
of induction it follows that
0 ≤ y(n) ≤ αn y(0)
for all n ∈ N. As 0 < α < 1, limn→∞ αn = 0. With the aid of the “inclusion
theorem” we conclude that limn→∞ y(n) = 0 for all sequences y such that
0 ≤ y(0) ≤ 1. This implies that the interval [0, 1] is contained in the basin
23
of attraction of 0.
4. 1 < y(0) < 1/α. Then 1 − 1/α < 1 − y(0) < 0 and 0 < αy(0) < 1, hence
1 − 1/α < y(1) < 0. From here on, proceed as in case 5.
5. If 1 − 1/α < y(0) < 0, an inductive argument shows that 1 − 1/α <
y(n) < y(n + 1) < 0 for all n ∈ N. For, suppose that 1 − 1/α < y(n) < 0
for some n ∈ N. Then we have α(1 − y(n)) < 1, hence y(n) < y(n + 1) =
αy(n)(1 − y(n)) < 0. The sequence y is monotone increasing and bounded
above, hence it has a limit l ∈ (1 − 1/α, 0]. Letting n tend to ∞ on both
sides of the equation
y(n + 1) = αy(n)(1 − y(n))
we find that l = αl(1 − l). In the interval (1 − 1/α, 0] the latter equation has
the unique solution l = 0.
From 3., 4. and 5. we infer that limn→∞ y(n) = 0 for all sequences y with
1 − 1/α < y(0) < 1/α. Hence the interval (1 − 1/α, 1/α) is contained in the
basin of attraction of 0. As we have also proved that the basin of attraction of
0 is contained in the interval (1 − 1/α, 1/α), it must be equal to this interval.
1.2.1 Stability of periodic and equilibrium solutions of

autonomous first order difference equations
The following theorem provides a test to determine, in many cases, but not
all, the stability of equilibrium solutions of (1.15).
Theorem 1.2.5. Suppose c ∈ C is a fixed point of f and f is differentiable
at c.
(i) If |f 0 (c)| < 1, then y ≡ c is a (locally) asymptotically stable solution of
(1.15).
(ii) If |f 0 (c)| > 1, then y ≡ c is an unstable solution of (1.15).
Proof. (i) Suppose that |f 0 (c)| < 1. By the definition of derivative,
f (x) − f (c)
| lim |<1
x→c x−c
Let d ∈ (|f 0 (c)|, 1) and ρ = d − |f 0 (c)|. Then there exists a neighbourhood U
of c of the form |x − c| < δ, with δ > 0, such that, for all x ∈ U − {c},

f (x) − f (c) f (x) − f (c)
0
| | − |f (c)| ≤| − f 0 (c)| < ρ

x−c x−c
24
Hence
f (x) − f (c)
| | < |f 0 (c)| + ρ = d
x−c
Hence it follows that
|f (x) − c| = |f (x) − f (c)| ≤ d|x − c| < dδ < δ
for all x ∈ U , so that x ∈ U implies f (x) ∈ U . Using the principle of mathe-

matical induction, we conclude that f n (x) ∈ U for all n ∈ N. Moreover, we
have
|f n (x) − c| = |f (f n−1 (x)) − f (c)| ≤ d|f n−1 (x) − c|
for all x with the property that f n−1 (x) ∈ U and all n ∈ N. By means of
another inductive argument we find that, for all n ∈ N,
|f n (x) − c| ≤ dn |x − c|
Let > 0. As 0 < d < 1, |f n (x) − c| ≤ for all x ∈ U such that |x − c| ≤ .

Hence y ≡ c is stable. Moreover, limn→∞ dn = 0, so limn→∞ f n (x) = c. For
any solution ỹ of (1.15) with initial value y0 ∈ U , we have ỹ(n) = f n (y0 ) and
thus limn→∞ ỹ(n) = c. This implies that y ≡ c is an asymptotically stable
solution of (1.15).
(ii) Now suppose |f 0 (c)| > 1. Let d ∈ (1, |f 0 (c)|) and = |f 0 (c)| − d. Then
there exists a neighbourhood U of c of the form |x − c| < δ, with δ > 0, such
that
f (x) − f (c) f (x) − f (c)
0
| | − |f (c)| ≤ | − f 0 (c)| < ,

x−c x−c
hence
f (x) − f (c)
| | > |f 0 (c)| − = d
x−c
for all x ∈ U − {c}. It follows that
|f (x) − c| > d|x − c|
for all x ∈ U − {c}. We will prove the second statement of the theorem by
contradiction. Suppose that y ≡ c is a stable solution. Then there exists a
neighbourhood V of c, of the form |x − c| < δ 0 , with δ 0 > 0, such that any
solution ỹ with initial value y0 ∈ V “stays within U ”, i.e.
|ỹ(n) − c| < δ
25
for all n ∈ N. Then we have, for all n ∈ N,
|ỹ(n + 1) − c| = |f (ỹ(n)) − c| > d|ỹ(n) − c|
and thus, by the principle of mathematical induction,
|ỹ(n) − c| > dn |y0 − c|
As limn→∞ dn = ∞, this leads to a contradiction, and thus the solution y ≡ c

is unstable.
A fixed point c of a differentiable function f is called hyperbolic if |f 0 (c)| =

6
1. If |f 0 (c)| = 1, then, without additional information about the function f ,
nothing can be said about the stability of the equilibrium solution. Moreover,
the theorem is a strictly local result: in the case of an attracting fixed point,
it contains no information about the size of the basin of attraction.
y(n + 1) = 2(y(n) − n)2 + n + 1
is nonautonomous, but by the change of variable y(n) = z(n) + n it is trans-

formed into the autonomous equation
z(n + 1) = f (z(n)) := 2z(n)2
This equation has equilibrium solutions z ≡ 0 and z ≡ 1/2. From the facts
that f 0 (0) = 0 and f 0 ( 21 ) = 2 we deduce that the first is asymptotically stable,
whereas the second is unstable. The corresponding solutions of the original
equation are the sequences y(n) = n and y(n) = n + 21 . So, by lemma 1.1.11,
these are asymptotically stable and unstable, respectively.
Example 1.2.7. The functions f1 (x) = x − x3 and f2 (x) = x + x3 both have

the unique fixed point 0. As f10 (0) = f20 (0) = 1, both are nonhyperbolic and
Theorem 1.2.5 gives us no information about the stability of the null solution
of the equation y(n + 1) = fi (y(n)) for i = 1 or 2. Suppose we are interested
in real-valued solutions only. We have
|f1 (x)| = |x|(1 − x2 ) < |x| for all x ∈ (−1, 1) − {0}
Hence it follows that, if i = 1, for every solution y with initial value y(0) ∈
(−1, 1) − {0}, |y(n + 1)| < |y(n)|. Consequently, if ∈ (0, 1), then |y(0)| <
26
implies |y(n)| < for all n ∈ N, so y ≡ 0 is a stable solution of the equation
(here we can choose ‘δ 0 = ). Moreover, if y(0) ∈ (−1, 1) − {0}, the sequence
|y(n)| is monotone decreasing and thus it converges to a limit. Since this
limit must be a fixed point of |f1 |, it has to be 0 (check this). Thus we
find that all (real) solutions with initial value in (−1, 1) converge to 0, hence
the null solution of the equation y(n + 1) = y(n) − y(n)3 is asymptotically
stable. The null solution of the equation y(n + 1) = y(n) + y(n)3 , however, is
unstable. This can be seen as follows. For every solution y of the equation,
we have y(1) = y(0)(1 + y(0)2 ). By means of an inductive argument, one
easily proves that, for all n ∈ N,
|y(n)| ≥ |y(0)|(1 + y(0)2 )n
Hence it follows that limn→∞ |y(n)| = ∞ for any solution y 6≡ 0.
The following theorem is a particular case of Theorem 1.2.5.
Theorem 1.2.8. Let g be a continuous function on a (real or complex)

neighbourhood of 0, with the property that limx→0 g(x)
x
= 0, and let a ∈ C.
If |a| < 1, the null solution of the equation
y(n + 1) = ay(n) + g(y(n)) (1.17)
is asymptotically stable. If |a| > 1, the null solution of this equation is

unstable.
This theorem can be deduced from Theorem 1.2.5 as follows. The condition
limx→0 g(x)
x
= 0 implies that limx→0 g(x) = 0, and thus, in view of the con-
tinuity of g, g(0) = 0. Take c = 0 and define f by f (x) = g(x) + ax. Then
f (0) = 0. g is differentiable at 0, with derivative
g(x) − g(0) g(x)

g 0 (0) = lim = lim =0
x→0 x−0 x→0 x
f is differentiable at 0 too, and f 0 (0) = g 0 (0) + a = a. The statements of

Theorem 1.2.8 now follow immediately from Theorem 1.2.5.
Example 1.2.9.
y(n + 1) = αy(n)(1 − y(n)), n∈N
27
Here, g(x) = −αx2 and
|g(x)|
lim = lim |αx| = 0
x→0 |x| x→0
Application of Theorem 1.2.8 reproduces the known fact that the null solution
of the equation is asymptotically stable for α ∈ (0, 1) and unstable for α > 1.
Theorems 1.2.5 and 1.2.8 are based on a method that is frequently used in the
study of nonlinear difference equations, known as linearization. It consists
in approximating solutions of the nonlinear equation by solutions of a linear
equation. In the case of Theorem 1.2.8 that linear equation is
y(n + 1) = ay(n) (1.18)
As we will learn in §1.3.3, all solutions of this equation are asymptotically

stable when |a| < 1, stable when |a| ≤ 1 and unstable when |a| > 1. Thus, as
far as its stability is concerned, the null solution of (1.17) has, for |a| =
6 1, the
same properties as the null solution of the linear equation. (1.18) is referred
to as the linearized equation or the linearization of (1.17) at 0. Linearization
of equation (1.17) amounts to neglecting the influence of the function g in
(1.17) on the behaviour of solutions close to the null solution. This is not
always justified! In example 1.2.7, the linearized equation at 0 is the same
for both cases, viz. y(n + 1) = y(n). In that example, however, the influence
of the 3rd degree term in f on the stability of the null solution cannot be
neglected.
More generally, the equation (1.15) can be linearized at any fixed point c
where f is differentiable, i.e. where the graph of f has a tangent. Linearizing
now amounts to replacing f by its linear approximation, i.e. the equation of
that tangent. Thus, linearization of (1.15) at c yields the linear difference
equation
y(n + 1) = f (c) + f 0 (c)(y(n) − c) = c + f 0 (c)(y(n) − c)
Example 1.2.10. The linear approximation of f (x) = 2x2 at the fixed point
x = 1/2 is
f (x) ' f (1/2) + f 0 (1/2)(x − 1/2) = 1/2 + 2(x − 1/2)
28
Hence, linearization of the equation
y(n + 1) = 2y(n)2
at 1/2 yields the inhomogeneous, linear difference equation
y(n + 1) = 1/2 + 2(y(n) − 1/2)
Any p-periodic solution of (1.15) corresponds to p fixed points c, f (c), ...,

f p−1 (c) of f p , hence to p equilibrium solutions of the equation
z(n + 1) = f p (z(n)) (1.19)
If a p-periodic solution of (1.15) with initial value c is (asymptotically) stable,
then, obviously, the equilibrium solution z ≡ c of (1.19) is (asymptotically)
stable as well. If f is continuous at f k−1 (c) for k = 1, ..., p − 1, then
the converse is also true: (asymptotic) stability of the equilibrium solution
z ≡ c is a sufficient condition for the (asymptotic) stability of the p-periodic
solution. We give the proof of the last statement for the case of stability.
Suppose that z ≡ c is a stable solution of equation (1.19). Let > 0. The
continuity of f at f k−1 (c) for k = 1, ..., p − 1, implies that f k is continuous
at c for k = 1, ..., p − 1 (cf. §1.2.3, exercise 1). Hence there exists a positive
number δ such that |f k (x) − f k (c)| < for all x with the property that
|x − c| < δ and for k = 1, ..., p − 1. Furthermore, the stability of z ≡ c
implies the existence of a positive number δ0 , such that, for all n ∈ N,
|z(n) − c| < min{δ , }
for every solution of (1.19) with the property that |z(0) − c| < δ0 . Let y∗
be the p-periodic solution of (1.15) with initial value c and y an arbitrary
solution of (1.15) such that |y(0) − c| < δ0 . Then the sequence z defined by
z(n) = y(pn) is a solution of (1.19) and |z(0) − c| = |y(0) − c| < δ0 . Thus
|y(pn) − c| = |z(n) − c| < min{δ , } for all n ∈ N. Hence it follows that
|y(pn + k) − y∗ (pn + k)| = |f k (y(pn)) − f k (c)| <
for all n ∈ N and k = 0, ..., p − 1 and thus |y(n) − y∗ (n)| < for all n ∈ N.
This completes the proof of the stability of y∗ .
The following result can be deduced from Theorem 1.2.5 and the above
remarks.
29
Theorem 1.2.11. Let c ∈ C be a fixed point of f p for some positive integer
p and suppose that f p is differentiable at c.
(i) If f is continuous at f k−1 (c) for k = 1, ..., p−1, and |(f p )0 (c)| < 1, then the
(periodic) solution y of (1.15) with initial value c is (locally) asymptotically
stable.
(ii) If |(f p )0 (c)| > 1, then the (periodic) solution y of (1.15) with initial value
c is unstable.
Remark 1.2.12. If f is differentiable at f k (c) for k = 0, ..., p − 1, then f p

is differentiable at c, and its derivative is
(f p )0 (c) = f 0 (f p−1 (c))...f 0 (c)
Example 1.2.13. In example 1.2.1 we have: (f 2 )0 (x) = 4x3 , hence

2π
(f 2 )0 (x1,2 ) = 4(e± 3 i )3 = 4e±2πi = 4
According to Theorem 1.2.11, both periodic solutions are unstable.
Example 1.2.14. We can use Theorem 1.2.11 to examine the stability of

the 2-periodic solutions of the logistic difference equation, with α > 3. We
use the notation of example 1.2.2. We have
(f 2 )0 (x1 ) = f 0 (f (x1 ))f 0 (x1 ) = f 0 (x2 )f 0 (x1 ) = α2 (1 − 2x2 )(1 − 2x1 )
From the fact that x1 and x2 are solutions of the quadratic equation α2 x2 −
α(α + 1)x + α + 1 = 0, we deduce that
α2 (x1 + x2 ) = α(α + 1) and α2 x1 x2 = α + 1
(f 2 )0 (x1 ) = −α2 + 2α + 4
√ 1 to −∞
The function on the right-hand side decreases monotonically from
as α increases from 3 to ∞, and it equals −1 when α = 1 + 6. Thus,
according to Theorem 1.2.11,
√ both 2-periodic solutions are
√ asymptotically
stable when 3 < α < 1 + 6 and unstable when α > 1 + 6.
In the previous section we mentioned an asymptotically stable equilibrium

point as an example of an attractor. More generally, a “periodic attractor”
is an attractor consisting of a periodic orbit.
30
Lemma 1.2.15. If f is a continuous map, then the orbit of any asymptoti-
cally stable periodic solution of (1.15) is a (periodic) attractor.
Proof. We give the proof for the case of an asymptotically stable, 2-periodic
solution. Let c be the initial value of the solution and A := γ + (c) =
{c, f (c)}. A being a finite set, it is closed, and it is also invariant, as
f (A) = {f (c), f 2 (c)} = {f (c), c} = A. Due to the asymptotic stability
of the solution, there exists a neighbourhood U1 of c, of the form |x − c| < δ1 ,
with δ1 > 0, such that limn→∞ f n (x)−f n (c) = 0 for all x ∈ U1 . As f n (c) ∈ A
for all n ∈ N, this implies that the distance of f n (x) to A tends to 0 for all
x ∈ U1 . It remains to prove the existence of a neighbourhood U2 of f (c), of
the form |x − f (c)| < δ2 , with δ2 > 0, such that limn→∞ f n (x) − f n (f (c)) = 0
for all x ∈ U2 . Due to the continuity of f at f (c), there exists a positive
number δ2 , such that |f (x) − f (f (c))| < δ1 for all x with the property that
|x−f (c)| < δ2 . Consequently, f (x) ∈ U1 and thus limn→∞ f n (f (x))−f n (c) =
limn→∞ f n+1 (x)−f n (f 2 (c)) = limn→∞ f n+1 (x)−f n+1 (f (c)) = 0 for these val-
ues of x. Hence there exists a neighbourhood U := U1 ∪ U2 of A such that,
for all x ∈ U , the distance of f n (x) to A tends to 0 as n → ∞.
We conclude this section with two examples of very simple economic models
that give rise to nonlinear, autonomous difference equations.
Example 1.2.16. Inventory adjustment model

This is described by the equation
Xt − Xt−1 = βXt−1 (K − Xt−1 )
or, equivalently,
Xt+1 = (1 + βK)Xt − βXt2 ,
which is a variant of the logistic difference equation. Here, Xt denotes the
inventory level in the period t, K is the desired or maximal stock and β is a
positive constant.
Example 1.2.17. Cobweb model with adaptive expectations.

The cobweb model with adaptive expectations is described by the equations
qtd = D(pt )
qts = S(pet )
31
qtd = qts
pet = pet−1 + w(pt−1 − pet−1 ), 0 < w < 1
Here, qtd , qts , pt and pet denote the demand, supply, price and expected price
in the period t, respectively. We assume the demand to be a strictly de-
creasing function of the price, hence its inverse exists. Then pet satisfies the
autonomous, first order difference equation
pet+1 = (1 − w)pet + wD−1 ◦ S(pet ) (1.20)
1.2.2 Bifurcations
In this section we briefly discuss some examples of frequently occurring bi-
furcations. We already encountered one type of bifurcation in §1.2, in the
1-parameter-family of autonomous, first order difference equations
y(n + 1) = y(n)2 − a
where a is a real parameter. If a < − 41 , the equation has no real-valued

equilibrium solutions. If a > − 41 , however, it has 2 real-valued equilibrium
√ √
solutions: y ≡ y1 := 12 + 12 1 + 4a and y ≡ y2 := 21 − 12 1 + 4a. The
derivatives
√ = x2 − a at its fixed points are: f 0 (y1 ) =
of the function f (x) √
1 + 1 + 4a and f 0 (y2 ) = 1 − 1 + 4a. The first exceeds 1 for all a >
− 41 , the second one belongs to the interval (−1, 1) when − 14 < a < 43 . As
the parameter passes through the value − 14 , two real equilibrium solutions
appear, one of which is unstable, whereas the other one is asymptotically
stable. This is characteristic of the so-called saddle-node bifurcation.
A different type of bifurcation occurs in the 1-parameter family
y(n + 1) = ay(n) − y(n)2 (1.21)
where a ∈ R. The function f (x) = ax − x2 has two fixed points: 0 and

a − 1. f 0 (0) = a and f 0 (a − 1) = 2 − a. Thus, the null solution of (1.21)
is asymptotically stable when −1 < a < 1 and unstable when a > 1 (or
a < −1), whereas the equilibrium solution y ≡ a − 1 is asymptotically stable
when 1 < a < 3 and unstable when a < 1. At the parameter value a = 1, a
transfer of stability from one solution (y ≡ 0) to the other one (y ≡ a − 1)
32
Figure 1.5: Transcritical bifurcation in (2.7) at a = 1
Figure 1.6: Pitchfork bifurcation in (2.8) at a = 1
33
Figure 1.7: Bifurcation diagram of the logistic difference equation.
takes place. This is called a transcritical bifurcation (cf. fig. 1.5). A

third type of bifurcation occurs in the following family of equations
y(n + 1) = ay(n) − y(n)3 (1.22)
where a is a real parameter. If a < 1, 0 is the unique real fixed point of f (x) =
ax − x3 . f 0 (0) = a, so the null solution of (1.22) is asymptotically stable
when −1 < a < 1 and unstable when a > 1 (the case a = 1 was discussed
in example 1.2.7).√ If a > 1, there are two additional
√ real-valued equilibrium
solutions: y ≡ ± a − 1. From the fact that f 0 (± a − 1) = 3−2a we deduce
that both solutions are asymptotically stable when 1 < a < 2. At a = 1 a
so-called pitchfork bifurcation occurs (cf. fig. 1.6).
To illustrate a fourth type of bifurcation, we consider once more the
logistic difference equation (cf. the examples 1.2.2 and 1.2.4). The function
f (x) = αx(1 − x) has two fixed points: 0 and 1 − 1/α, when α 6= 1. f 0 (0) = α
and f 0 (1 − 1/α) = 2 − α. According to Theorem 1.2.5, the null solution of
(1.16) is asymptotically stable when 0 < α < 1 and unstable when α > 1.
The second equilibrium solution is asymptotically√stable when 1 < α < 3
and unstable when α > 3. When 3 < α < 1 + 6, the equation has two
asymptotically stable periodic solutions. At the parameter values α = 1 and
α = 3, the properties of the equation change significantly, in both cases it
undergoes a bifurcation. The bifurcation at α = 1 is of the same type as
in (1.21), this is a transcritical bifurcation. At α = 3 a so-called period-
doubling or flip-bifurcation occurs: an asymptotically stable equilibrium
solution gives way to two asymptotically stable 2-periodic solutions (having
the same orbit).
The above examples all concern bifurcations of equilibrium solutions.
Similar bifurcations occur in periodic solutions of period greater than 1. Bi-
furcations can be represented in a bifurcation diagram. Usually, the values
of the asymptotically stable equilibrium (or periodic) solutions are plotted
versus the parameter. Sometimes, unstable periodic solutions are indicated
by means of dotted lines. In order to make bifurcation diagrams, like the
ones shown in figs. 1.5 through 1.7, on a computer, one roughly proceeds as
follows. For “each” value of the parameter, first, a certain number of val-
ues of the solution with a given initial value x0 , the so-called “pre-iterates”)
is computed. Next, a number of subsequent entries of the same sequence
34
Figure 1.8: The solutions of (2.2) with α = 2.8, y(0) = 0.1 and y(0) = 0.2.
is computed and plotted. If the number of pre-iterates is sufficiently large,

the subsequent entries of the sequence will be close to the ω-limit set of
x0 (provided the latter is nonempty). In the case of a globally asymptoti-
cally stable equilibrium, all these entries will approximately coincide with the
equilibrium point. In the case of a globally asymptotically stable 2-periodic
solution they will approximately alternate between the two points of the pe-
riodic orbit, etc.. In the case of a locally, but not globally asymptotically
stable solution, the result will depend on the choice of initial value. Figure
1.7 shows a bifurcation diagram of (1.16), made with the computer program
Dynamics. The parameter α has been plotted on the horizontal axis. For a
great number of values of α in the interval (0, 4) the first 100 values of the
solution with a specific initial value have been computed, but not plotted,
and the next 200 values have been plotted. The result gives some idea of the
ω-limit set of the corresponding initial value as a function of α. The diagram
seems to indicate that, for α ∈ (1, 3), the solution has a limiting value (this is
the value 1−1/α of the asymptotically
√ stable equilibrium solution), for α > 3
up to a certain value (= 1 + 6), is asymptotically 2-periodic (approaches
the asymptotically stable 2-periodic solution), for even greater values of α
approaches a 4-periodic sequence, and so on.
The difference in asymptotic behaviour of the solutions of (1.16) for dif-
ferent values of the parameter is also illustrated by the figures 1.8 through
1.10. In each figure, two solutions with slightly different initial values have
been plotted versus n. (Successive points have been connected by line seg-
ments.) Compare the asymptotic behaviour of the solutions to figure 1.7 and
observe the effect of a change in the initial value in each of the three figures.
35
Figure 1.9: The solutions of (2.2) with α = 3.5, y(0) = 0.1 and y(0) = 0.2.
(Also compare this figure to Fig. 1.2.)
Figure 1.10: The solutions of (2.2) with α = 4, y(0) = 0.1 and y(0) = 0.11.
36
1.2.3 Exercises
1. Let f : R → R and c ∈ R and suppose that f is continuous at f k−1 (c)
for k = 1, ..., p. Use an inductive argument to prove that f k is continuous
at c for k = 1, ..., p.
√ √
2. Prove that the interval (− 2, 2) is the basin of attraction of the fixed
point 0 of the map f (x) = x − x3 , x ∈ R.
3. Find all fixed points of the map f (x) = x3 + 34 x, x ∈ R. Determine the

basin of attraction of the fixed point 0.
4. In example 1.1.2, let Q be a Cobb-Douglas function of the form: Q(x, y) =

xα y 1−α , with 0 < α < 1.
a. Linearize the equation (1.1) at the equilibrium point k 6= 0.

b. Prove that this equilibrium point is an attractor.
c. Take the domain of Q to be R2++ and prove that y ≡ k is a globally
asymptotically stable equilibrium point. (Hint: show that 0 < k < k
implies k < sq(k)+k
1+n
< k whereas k > k implies k < sq(k)+k
1+n
< k. Here,
q(k) = Q(k, 1).)
5. Determine the equilibrium solution of equation (1.15), where
f (x) = α sin x, x ∈ R, 0 < α ≤ 1
Linearize the equation at the equilibrium point. Examine the stability of

the equilibrium solution.
6. Find all real-valued equilibrium solutions of equation (1.15), where

αx
f (x) = , x ∈ R, α ∈ R
1 + x2
and examine the stability of these solutions.
7. Suppose that f is a continuous map from C to C and that z ≡ c is

an asymptotically stable solution of (1.19). Prove that the sequence
{f n (c)}∞
n=0 is an asymptotically stable periodic solution of (1.15), of period
p.
37
1.3 Linear difference equations
The difference equation (1.9) is linear if and only if F is an affine function of
y, τ y, ..., τ k y, thus a linear difference equation has the form
F (n, y, τ y, ..., τ k y) = b + a0 y + a1 τ y + ... + ak τ k y = 0 (1.23)
Here, a0 , a1 , ..., ak and b are functions of n, i.e. sequences. For every n ∈ N
(or Z) we have
b(n) + a0 (n)y(n) + a1 (n)y(n + 1) + ... + ak (n)y(n + k) = 0 (1.24)
We can write the equation in the more compact form
k
aj τ j y = −b
X
(1.25)
j=0
Here τ 0 is the identity operator, i.e. τ 0 y = y. The expression kj=0 aj τ j

P
defines a linear operator on the set of all sequences on Z or N. An operator

of this form is called a linear difference operator of order k, provided neither
a0 nor ak is a null sequence. We will often simply denote an operator of
this type by L. The equation is called homogeneous if the null sequence
y ≡ 0 is a solution. This is the case iff b ≡ 0. An equation that is not
homogeneous is called inhomogeneous. Thus, an inhomogeneous equation
never has a null solution. The set of all (real- or complex-valued) solutions
of a homogeneous linear difference equation is a linear space (over R
or C, respectively). This is an immediate consequence of the linearity of
the difference operator L := kj=0 aj τ j . For, suppose that the sequences yi ,
P
i = 1, ..., m, satisfy the homogeneous equation

k
X
Ly(n) = aj (n)y(n + j) = 0 (1.26)
j=0
Then we have, for arbitrary (real or complex) numbers c1 , ..., cm ,

L(c1 y1 + c2 y2 + ... + cm ym ) = c1 Ly1 + c2 Ly2 + ... + cm Lym
All terms on the right-hand side are null sequences, so the left-hand side is
a null sequence as well and this implies that c1 y1 + c2 y2 + ... + cm ym is again
a solution of (1.26). The dimension of the solution space (the kernel of
L) depends on the zeroes of the sequence ak , i.e. the values of n such that
ak (n) = 0.
38
Example 1.3.1. The solution space of the homogeneous, linear difference
equation
((−1)n − 1)y(n + 1) + y(n) = 0, n ∈ N
consists of the unique solution of this equation: the null solution, hence its
dimension is 0. On the other hand, the equation
(n + 1)y(n + 1) − y(n) = 0, n∈N
has the general solution

c
y(n) = , n ∈ N, c ∈ C,
n!
hence a 1-dimensional solution space.
From now on we assume that the sequence ak has no zeroes. Dividing each
term in (1.24) by ak (n), we obtain the equivalent equation
b(n) a0 (n) a1 (n) ak (n)
+ y(n) + y(n + 1) + ... + y(n + k) = 0
ak (n) ak (n) ak (n) ak (n)
It has the same form as (1.24), but now the coefficient of y(n + k) equals 1
for all n. Henceforth we consider linear difference equations of the form
y(n + k) + ak−1 (n)y(n + k − 1) + ... + a0 (n)y(n) = b(n), n ∈ N (1.27)
In fact, this equation is a special case of a recurrence relation (to see this, take
all terms except y(n + k) to the right-hand side). In particular, any solution
is determined by its first k values y(0), ..., y(k − 1) (the initial values), which
can be chosen arbitrarily. We begin by studying the homogeneous equation
y(n + k) + ak−1 (n)y(n + k − 1) + ... + a0 (n)y(n) = 0, n ∈ N (1.28)
If we take the first k entries of the sequence to be zero, then all subsequent
entries will be zero as well and we end up with the null solution.
Theorem 1.3.2. The (real- or complex-valued) solutions of the kth order

homogeneous linear difference equation (1.28) form a k-dimensional lin-
ear space (over R or C, respectively).
39
Proof. The k initial values y(0), ..., y(k − 1) form a k-dimensional vector.
If l ≤ k, we can choose l linearly independent such k-dimensional vectors
(y1 (0), ..., y1 (k−1)),...,(yl (0), ..., yl (k−1)) and these will determine l solutions
y1 ... yl of (1.28). These solutions are linearly independent. For, suppose
there exist numbers c1 , c2 , ..., cl with the property that c1 y1 (n) + c2 y2 (n) +
... + cl yl (n) = 0 for all n ∈ N, then, of course,
       
y1 (0) y2 (0) yl (0) 0
. . . .
       
       
       
c1 
 .  + c2 
  .  + ... + cl 
  . 
 = 
 . 
 (1.29)

 . 


 . 


 . 


 . 

y1 (k−1) y2 (k−1) yl (k−1) 0
The linear independence of the vectors on the left-hand side of (1.29) implies
that c1 = c2 = ... = cl = 0. Hence the equation (1.28) has at least k
linearly independent solutions. Conversely, for any number l of linearly inde-
pendent solutions y1 , ..., yl of (1.28), the vectors (y1 (0), ..., y1 (k − 1)) through
(yl (0), ..., yl (k − 1)) are linearly independent. For, suppose there exist num-
bers c1 , c2 , ..., cl such that (1.29) holds. Define a sequence y by
y := c1 y1 + c2 y2 + ... + cl yl
Then y is a solution of (1.28) with the property that y(n) = 0 for 0 ≤ n ≤
k − 1 and thus y ≡ 0, i.e.
c1 y1 (n) + c2 y2 (n) + ... + cl yl (n) = 0 for all n ∈ N
The linear independence of the solutions y1 , ..., yl now implies that c1 = c2 =
... = cl = 0, and, consequently, the vectors
(y1 (0), ..., y1 (k − 1)), ..., (yl (0), ..., yl (k − 1)) are linearly independent. Since
there can be at most k linearly independent k-dimensional vectors, we con-
clude that the dimension of the solution space of (1.28) is at most k and
thus is exactly k.
We have proved the following lemma in the process.
Lemma 1.3.3. Let l ∈ N. l solutions y1 , ..., yl of the homogeneous difference

equation (1.28) are linearly independent (over R or C), if and only if the
vectors (y1 (0), ..., y1 (k − 1)), ..., (yl (0), ..., yl (k − 1)) are linearly independent
(over R or C, respectively).
40
It is of great importance to have a basis of the solution space of (1.28) at
our disposal. In that case, any solution of (1.28) can be written as a linear
combination of the basis elements. In view of lemma 1.3.3, it suffices to choose
k linearly independent vectors (y1 (0), ..., y1 (k − 1)),..., (yk (0), ..., yk (k − 1)).
To prove that k given solutions y1 , ..., yk of (1.28) constitute a basis of the
solution space, it suffices to prove the linear independence of the vectors
(y1 (0), ..., y1 (k − 1)),..., (yk (0), ..., yk (k − 1)). This is the case iff the matrix
having these vectors for its column vectors is nonsingular, i.e. iff

y1 (0) y2 (0) . . yk (0)

y1 (1) y2 (1) . . yk (1)

. . . . .

6= 0

. . . . .

. . . . .

y1 (k − 1) y2 (k − 1) . . yk (k − 1)
The above determinant is called the Casorati determinant. For all non-
negative integers n, Cy1 ,...,yk (n) is defined by

y1 (n) y2 (n) . . yk (n)

y1 (n + 1) y2 (n + 1) . . yk (n + 1)

. . . . .

Cy1 ,...,yk (n) =

. . . . .

. . . . .

y1 (n + k − 1) y2 (n + k − 1)
. . yk (n + k − 1)
Lemma 1.3.4. If the sequences y1 , ..., yk are solutions of the homogeneous

linear difference equation (1.28), then Cy1 ,...,yk is a solution of the first order,
linear difference equation
y(n + 1) = (−1)k a0 (n)y(n), n ∈ N (1.30)
Proof. We begin by proving the lemma for the case that k = 2. For every
nonnegative integer n we have

y (n + 1) y2 (n + 1)

Cy1 ,y2 (n + 1) = 1

y1 (n + 2) y2 (n + 2)

Furthermore,
yj (n + 2) = −a0 (n)yj (n) − a1 (n)yj (n + 1), j = 1, 2
41
Inserting this into the last row of the determinant, we get

y1 (n + 1) y2 (n + 1)

−a0 (n)y1 (n) − a1 (n)y1 (n + 1) −a0 (n)y2 (n) − a1 (n)y2 (n + 1)

The last row being the sum of the two row vectors (−a0 (n)y1 (n) −a0 (n)y2 (n))
and (−a1 (n)y1 (n + 1) − a1 (n)y2 (n + 1)), and the determinant of a matrix
being a linear function of its row vectors, this is equal to

y1 (n + 1) y2 (n + 1) y1 (n + 1) y2 (n + 1)
−a0 (n) − a1 (n)

y1 (n) y2 (n) y1 (n + 1) y2 (n + 1)

The second term vanishes for all n. Interchanging rows in the first determi-
nant, we find

y1 (n) y2 (n)
Cy1 ,y2 (n + 1) = a0 (n)
= a0 (n)Cy1 ,y2 (n)
y1 (n + 1) y2 (n + 1)

The proof for k > 2 is analogous. For every n ∈ N we have

y1 (n + 1) y2 (n + 1)

. . yk (n + 1)

y1 (n + 2) y2 (n + 2)

. . yk (n + 2)

. . . . .

Cy1 ,...,yk (n + 1) =

. . . . .

. . . . .

y1 (n + k) y2 (n + k)
. . yk (n + k)
If, in the last row of the determinant, we substitute:

k−1
X
yj (n + k) = − al (n)yj (n + l), j = 1, ..., k
l=0
and once more use the linearity of the determinant, we find

y1 (n + 1) y2 (n + 1) . . yk (n + 1)

y1 (n + 2) y2 (n + 2) . . yk (n + 2)

k−1
. . . . .
X

Cy1 ,...,yk (n + 1) = − al (n)
l=0 . . . . .

. . . . .

y1 (n + l) y2 (n + l) . . yk (n + l)
42
The determinants on the right-hand side are seen to vanish, except for l = 0,
hence

y1 (n + 1) y2 (n + 1) . . yk (n + 1)

y1 (n + 2) y2 (n + 2) . . yk (n + 2)

. . . . .

Cy1 ,...,yk (n + 1) = −a0 (n)
. . . . .

. . . . .

y1 (n) y2 (n) . . yk (n)
Interchanging rows in the determinant on the right-hand side, we obtain the

desired result.
The solutions y1 , ..., yk of (1.28) constitute a basis of the solution space if

and only if Cy1 ,...,yk (0) 6= 0. Due to (1.30), this condition is equivalent to:
Cy1 ,...,yk (n) 6= 0 for some n ∈ N, as Cy1 ,...,yk (0) = 0 implies that Cy1 ,...,yk (n) =
0 for all n ∈ N (in fact, this is a particular case of lemma 1.3.3). Thus we
have proved the following theorem.
Theorem 1.3.5. k solutions y1 , ..., yk of the kth order homogeneous linear

difference equation (1.28) constitute a basis of the solution space if and only
if Cy1 ,...,yk (n) 6= 0 for some n ∈ N.
Example 1.3.6. The sequences y1 and y2 defined by

π π
y1 (n) = sin n , y2 (n) = cos n , n∈N
2 2
are solutions of the homogeneous linear difference equation
y(n + 2) + y(n) = 0, n∈N
Moreover,
0 1

Cy1 ,y2 (0) = = −1

1 0

Thus, y1 and y2 constitute a basis of the solution space of the equation. The
Casorati-determinant Cy1 ,y2 (n) satisfies the first order difference equation
y(n + 1) = (−1)2 y(n) n ∈ N
hence Cy1 ,y2 (n) = −1 for all n ∈ N. (Check this by computing Cy1 ,y2 (n)
directly.)
43
1.3.1 Inhomogeneous linear difference equations
Next, we consider the inhomogeneous difference equation
k
X
Ly(n) := aj (n)y(n + j) = b(n), n ∈ N (1.31)
j=0
where b does not vanish identically (is not a null sequence). As the homo-
geneous equation Ly = 0 has other solutions besides the null solution, the
linear operator L is not invertible and equation (1.31) does not have a
unique solution. If y0 is a solution, i.e. if Ly0 = b and if ỹ is an arbitrary
solution of the homogeneous equation, so Lỹ = 0, then y0 + ỹ is another
solution of (1.31), as
L(y0 + ỹ) = Ly0 + Lỹ = b + 0 = b
Conversely, the difference of two solutions y1 and y2 of the inhomogeneous
equation (1.31) is a solution of the homogeneous equation, as
L(y1 − y2 ) = Ly1 − Ly2 = b − b = 0
Theorem 1.3.7. Let y0 be a particular solution of the inhomogeneous
linear difference equation (1.31). Every solution of (1.31) can be written
as the sum of y0 and a solution of the homogeneous equation. Conversely,
any sequence that can be written as the sum of y0 and a solution of the
homogeneous equation is a solution of (1.31).
1.3.2 First order linear difference equations

First, we consider the homogeneous, first order difference equation
y(n + 1) = a(n)y(n), n ∈ N (1.32)

This equation has a 1-dimensional solution space. If the initial value y(0)
is known, all subsequent values can be computed. It is easily seen that, for
n ≥ 1,
y(n) = a(n − 1)...a(0)y(0)
y(0) can be chosen arbitrarily. Therefore, the general solution of (1.32) has
the form
n−1
Y
y(0) = c, y(n) = c a(m) for n ≥ 1, c∈C
m=0
44
Example 1.3.8. The terms of a geometric progression with common ra-
tio r ∈ R or C satisfy the equation
y(n + 1) = ry(n), n∈N
Here a(n) equals the common ratio r for all n, and thus is a constant sequence.
For any solution y we have
y(n) = rn y(0), n∈N
and the general solution of the equation is
y(n) = crn , n ∈ N, c ∈ C
Example 1.3.9. For every solution of the equation
y(n + 1) = (n + 1)y(n), n∈N
the following relation holds
y(n) = n(n − 1)...1y(0) = n!y(0)
and the general solution is
y(n) = cn!, n ∈ N, c ∈ C
Next, we consider a very simple case of an inhomogeneous equation:
y(n + 1) − y(n) = b(n), n ∈ N (1.33)

Writing out the equation for the first n nonnegative integers, we have
y(1) − y(0) = b(0)

y(2) − y(1) = b(1)
. . . . .
y(n) − y(n − 1) = b(n − 1)
45
and summing the expressions on both sides of the equality signs, we get
y(n) − y(0) = b(0) + b(1) + ... + b(n − 1)
As y(0) can be chosen arbitrarily, the general solution of (1.33) is
n−1
X
y(0) = c, y(n) = c + b(m) for n ≥ 1, c∈C
m=0
Finally, we turn to the general case of a first order, inhomogeneous, linear

difference equation:
y(n + 1) = a(n)y(n) + b(n) (1.34)
If a solution y0 of the homogeneous equation is known, with the property
that y0 (n) 6= 0 for all n, then the inhomogeneous equation can be solved
by means of a method called ‘variation of constants’. Such a solution y0
exists iff a(n) 6= 0 for all n. Suppose that this condition is fulfilled. Then, for
any solution y of the inhomogeneous equation, the following relation holds:
y(n + 1) a(n)y(n) + b(n) a(n)y(n) b(n)
= = +
y0 (n + 1) y0 (n + 1) a(n)y0 (n) y0 (n + 1)
y(n)
This implies that the quotient z(n) := y0 (n)
satisfies the equation
b(n)
z(n + 1) = z(n) +
y0 (n + 1)
This is an equation of the form (1.33), where, on the right-hand side, b(n)
has been replaced by y0b(n)
(n+1)
. Hence we have, for n ≥ 1,
n−1
y(n) X b(m)
=c+ , c∈C
y0 (n) m=0 y0 (m + 1)
Consequently, the general solution of (1.34) is given by

n−1
X b(m)
y(0) = cy0 (0), y(n) = cy0 (n) + y0 (n) for n ≥ 1, c∈C
m=0 y0 (m + 1)
Alternatively, this can be written in the form

n−1
X
y(0) = cy0 (0), y(n) = cy0 (n) + a(n − 1)...a(m + 1)b(m) for n ≥ 1, c∈C
m=0
(1.35)
(1.35) represents the general solution of (1.34), even if y0 does have zeroes
(verify this).
46
Example 1.3.10. The general solution of the inhomogeneous, first order,
y(n + 1) − ay(n) = b(n), n∈N
where a is a real or complex constant, is

n−1
y(0) = c, y(n) = can + an−m−1 b(m) for n ≥ 1,
X
c∈C
m=0
Example 1.3.11. The general solution of the inhomogeneous, first order,

y(n + 1) − (n + 1)y(n) = 1, n∈N
is
n−1
X 1
y(0) = c, y(n) = cn! + n! for n ≥ 1, c∈C
m=0 (m + 1)!
(cf. example 1.3.9)
In some special cases, like in the following example, a particular solution can
be easily found.
Example 1.3.12. The inhomogeneous equation
y(n + 1) − 2y(n) = 3
has a constant (or equilibrium) solution y ≡ −3. The general solution of the
corresponding homogeneous equation is y(n) = c2n , with c ∈ C (cf. example
1.3.8). The general solution of the inhomogeneous equation is therefore
y(n) = −3 + c2n , n ∈ N, c∈C
Example 1.3.13. An arithmetic-geometric progression is a sequence

yn , n ∈ N, the terms of which satisfy the following relation
yn+1 = ryn + vrn+1 , n∈N
47
Here, r and v are constants: the common ratio and difference of the progres-
sion, respectively. The homogeneous equation has a solution yn0 := rn . Thus,
the general solution of the inhomogeneous equation is
n−1
y0 = c, yn = crn + rn
X
v for n ≥ 1, c ∈ C
m=0
So the general term of an arithmetic-geometric progression has the form
yn = (y0 + nv)rn , n ∈ N
1.3.3 Stability of solutions of linear difference equa-

tions
According to definition 1.1.8, a solution y of the kth order linear difference
equation (1.27) is stable iff, for every positive number , there exists a positive
number δ such that, for every solution ỹ, with the property that
|ỹ(n) − y(n)| ≤ δ for n = 0, ..., k − 1
we have
|ỹ(n) − y(n)| ≤ for all n ∈ N
Now, z := ỹ − y is a solution of the homogeneous equation and every solution
of the homogeneous equation can be written as the difference of some solution
ỹ of (1.27) and y. Hence we can state the above condition as follows: for
every positive number there exists a positive number δ, such that, for every
solution z of the homogeneous equation, with the property that
|z(n)| = |z(n) − 0| ≤ δ for n = 0, ..., k − 1
we have
|z(n) − 0| ≤ for all n ∈ N
But this is precisely the condition for stability of the null solution of the ho-
mogeneous equation. This shows that a solution of a linear difference equa-
tion is stable if and only if the null solution of the homogeneous equation
is stable. It is easily verified that the same is true for asymptotic stabil-
ity. Consequently, the stability properties are the same for all solutions:
48
all solutions are neutrally stable, or all are asymptotically stable, or all are
unstable. Moreover, every asymptotically stable solution is globally asymp-
totically stable. For, suppose that y is an asymptotically stable solution of
(1.27). Then z ≡ 0 is an asymptotically stable solution of the homogeneous
equation. Hence there exists a positive number δ, such that limn→∞ z(n) = 0
for every solution z of the homogeneous equation, with the property that
|z(n)| ≤ δ for n = 0, ..., k − 1. Now consider an arbitrary solution z 6≡ 0
of the homogeneous equation and let µ := max{|z(0)|, ..., |z(k − 1)|}. Ob-
viously, µ > 0 and the sequence z̃ defined by z̃(n) = µδ z(n) also satisfies
the homogeneous equation. Moreover, |z̃(n)| ≤ δ for n = 0, ..., k − 1, hence
limn→∞ z̃(n) = 0, and this implies that limn→∞ z(n) = 0. Consequently, all
solutions of the homogeneous equation tend to 0 as n → ∞ and it imme-
diately follows that limn→∞ ỹ(n) − y(n) = 0 for any solution ỹ of (1.27).
Summarizing, we have the following theorem.
Theorem 1.3.14. All solutions of the linear difference equation (1.27) are
neutrally stable, globally asymptotically stable, or unstable, iff the
null solution of the homogeneous equation is neutrally stable, asymptoti-
cally stable, or unstable, respectively.
Example 1.3.15. The solutions of the equation
y(n + 1) − ay(n) = b(n), n∈N (1.36)
where b is an arbitrary sequence, are stable when |a| ≤ 1, as in that case every
solution y of the homogeneous equation has the property that |y(n) − 0| =
|an y(0) − 0| = |an y(0)| ≤ |y(0)| for all n ∈ N (cf. example 1.3.8). The
solutions of (1.36) are globally asymptotically stable when |a| < 1, as in that
case every solution y of the homogeneous equation, has the property that
lim |an y(0)| = 0

lim |y(n) − 0| = n→∞
n→∞
which implies that the null solution of the homogeneous equation is asymp-
totically stable. The solutions of (1.36) are unstable when |a| > 1, as in that
case, for every solution y 6≡ 0, of the homogeneous equation, limn→∞ |y(n) −
0| = limn→∞ |y(n)| = ∞.
49
1.3.4 Exercises
1. Determine the solution of the following initial value problem
y(n + 1) − (n + 1)y(n) = 1, y(0) = 0

y(n)
and compute limn→∞ n!
. (Cf. Example 1.3.11.)
2. Prove that all solutions of the equation
y(n + 1) − (n + 1)y(n) = n2 , n∈N
are unstable.
3. Check whether or not the sequences y1 and y2 , defined by
y1 (n) = (−2)n , y2 (n) = 3n , n∈N
constitute a basis of the solution space of the equation
y(n + 2) − y(n + 1) − 6y(n) = 0, n∈N
4. a. Determine the general solution of the equation
y(n + 2) − y(n + 1) − 6y(n) = 2, n∈N
b. Compute the solution y with initial values y(0) = 1, y(1) = 0.

c. Prove that this solution is unstable.
5. Prove that the sequences y1 and y2 , defined by

n
X 1
y1 (n) = n!, y2 (n) = n! , n∈N
m=0 m!
form a basis of the solution space of the equation
y(n + 2) − (n + 3)y(n + 1) + (n + 1)y(n) = 0, n∈N
and determine Cy1 ,y2 (n) for arbitrary n ∈ N.
6. a. Find the equilibrium solutions of the equation
y(n + 2) − (n + 3)y(n + 1) + (n + 1)y(n) = 3, n∈N
50
b. Determine the general solution of the above equation.
c. Determine the solution y with initial values y(0) = y(1) = 1 and
examine its stability.
7. Determine the general solution of
y(n + 1) = αy(n)2 , n∈N
where α is a positive constant, by taking, on both sides, the logarithm.
1.4 Systems of first order difference equations

Many problems involve more than one, say k, dependent variables, simulta-
neously satisfying a certain number of, say m, relations. This is referred to
as a system of m equations in k unknowns. Here, we restrict ourselves to
systems of k first order difference equations in k unknowns, of the following
form
y1 (n + 1) = f1 (n, y1 (n), ..., yk (n))
y2 (n + 1) = f2 (n, y1 (n), ..., yk (n))
. .
(1.37)
. .
. .
yk (n + 1) = fk (n, y1 (n), ..., yk (n))
Defining, for every n, a k-dimensional vector y(n) with components y1 (n),...,yk (n),
and a k-dimensional vector function f with components f1 ,...,fk , we can write
the above system of equations in the form of a so-called vectorial difference
equation    
y1 (n + 1) f1 (n, y1 (n), ..., yk (n))
 y2 (n + 1)   f2 (n, y1 (n), ..., yk (n)) 
   
. .
   
   
 =  (1.38)

 . 


 . 

. .
   
   
yk (n + 1) fk (n, y1 (n), ..., yk (n))
or, briefly
y(n + 1) = f (n, y(n)) (1.39)
Its solutions are sequences of k-dimensional vectors, or k-dimensional vector
functions on a subset of Z or N.
51
The difference between (1.37) and (1.38) or (1.39) is mainly a matter of nota-
tion. In what follows we will not discriminate between systems of equations
and vectorial equations.
Example 1.4.1. The system of first order, linear difference equations
y1 (n + 1) = a11 (n)y1 (n) +...+ a1k (n)yk (n) +b1 (n)
y2 (n + 1) = a21 (n)y1 (n) +...+ a2k (n)yk (n) +b2 (n)
. . . .
(1.40)
. . . .
. . . .
yk (n + 1) = ak1 (n)y1 (n) +...+ akk (n)yk (n) +bk (n)
is equivalent to the linear vectorial difference equation.
      
y1 (n + 1) a11 (n) . . . a1k (n) y1 (n) b1 (n)

y
 2 (n + 1) 


a
 21 (n) . . . a 2k (n) 
 2y (n)   b2 (n)
  

. . . . . .  .  .
      
     
 =  +  (1.41)

 . 


 . . . . .  .   .
    

. . . . . .  .   .
      
   
yk (n + 1) ak1 (n) . . . akk (n) yk (n) bk (n)
or its more compact form
y(n + 1) = A(n)y(n) + b(n) (1.42)
Here, y(n + 1), y(n) and b(n) are k-dimensional vectors and A(n) is the k × k
matrix in (1.41).
Systems of first order difference equations are important for more than one
reason. In the first place, systems of difference equations emerge naturally in
mathematical models of various problems, involving several time-dependent
variables. Secondly, any kth order, scalar difference equation can be easily
transformed into a system of k first order equations. Here is one way to do
that. As a starting point we take the kth order equation
y(n + k) = f (n, y(n), y(n + 1), ..., y(n + k − 1)), n∈N (1.43)
We define k sequences y1 , ..., yk by
yj (n) = y(n + j − 1), j = 1, ..., k, n∈N
52
So
y1 (n) = y(n), y2 (n) = y(n + 1), ..., yk (n) = y(n + k − 1)
Then
y1 (n + 1) = y(n + 1) = y2 (n),
y2 (n + 1) = y(n + 2) = y3 (n),
........... . ............ . .....
........... . ............ . .....
yk−1 (n + 1) = y(n + k − 1) = yk (n)
If y is a solution of (1.43), we have
yk (n+1) = y(n+k) = f (n, y(n), y(n+1), ..., y(n+k−1)) = f (n, y1 (n), ..., yk (n))
Hence, the sequence of vectors or the vector function (y1 , ..., yk ), defined
by: (y1 , ..., yk )(n) = (y1 (n), ..., yk (n)), is a solution of the vectorial difference
equation    
y1 (n + 1) y2 (n)
 y2 (n + 1)  .
   
 
. .
   
   
 = 

 . 


 . 

. yk (n)
   
   
yk (n + 1) f (n, y1 (n), ..., yk (n))
Example 1.4.2. Suppose that y is a solution of the kth order, linear differ-
ence equation
y(n + k) = −ak−1 (n)y(n + k − 1) − . . . − a0 (n)y(n) − b(n) (1.44)
Then the sequence of vectors or vector function (y1 , ..., yk ), defined by: (y1 , ..., yk )(n) =
(y1 (n), ..., yk (n)) = (y(n), y(n+1), ..., y(n+k−1)) is a solution of the vectorial
difference equation
      
y1 (n+1) 0 1 . . 0 y1 (n) 0
 y2 (n+1)  0 0 1 . 0   y2 (n)   0 
     
 
. . . . . . . .
      
      
 =  + 

 . 


 . . . . . 

 .   . 
 

. . . . . . .   . 
      
   
yk (n+1) −a0 (n) . . . −ak−1 (n) yk (n) −b(n)
(1.45)
Conversely: if (y1 , ..., yk ) is a solution of (1.45), then y1 is a solution of the
kth order scalar equation (1.44) and, for j = 1, ..., k, yj (n) = y1 (n + j − 1).
53
Example 1.4.3. The second order, linear difference equation
y(n + 2) − (n + 3)y(n + 1) + (n + 1)y(n) = 3, n ∈ N
is transformed into a system of two first order, linear difference equations as

follows. Let
y1 (n) := y(n), y2 (n) := y(n + 1), n ∈ N
Then the new dependent variables y1 and y2 satisfy the system of equations
y1 (n + 1) = y(n + 1) = y2 (n)
y2 (n + 1) = y(n + 2) = (n + 3)y(n + 1) − (n + 1)y(n) + 3
= (n + 3)y2 (n) − (n + 1)y1 (n) + 3
This corresponds to the vectorial equation

! ! ! !
y1 (n + 1) 0 1 y1 (n) 0
= +
y2 (n + 1) −(n + 1) n + 3 y2 (n) 3
Example 1.4.4. Consider the third order, linear difference equation
y(n + 3) + ny(n) = 2n , n ∈ N
By the substitutions
y1 (n) := y(n), y2 (n) := y(n + 1), y3 (n) = y(n + 2), n ∈ N
it is transformed into the system
y1 (n + 1) = y(n + 1) = y2 (n)
y2 (n + 1) = y(n + 2) = y3 (n) ,
n n
y3 (n + 1) = y(n + 3) = −ny(n) + 2 = −ny1 (n) + 2
equivalent to the vectorial equation

      
y1 (n + 1) 0 1 0 y1 (n) 0
 y2 (n + 1)  =  0 0 1 
  y2 (n)  +  0 
      
y3 (n + 1) −n 0 0 y3 (n) 2n
54
Also in the context of systems of difference equations, the notions stability,
asymptotic stability and instability are used. They are defined analo-
gously to the case of a scalar difference equation. We give the definitions
for a system of first order difference equations. A solution (y1 , ..., yk ) of the
system (1.38) is called stable, iff, for every > 0 there exists a δ > 0, such
that, for every solution (ỹ1 , ..., ỹk ) with the property that
|ỹj (0) − yj (0)| ≤ δ for all j ∈ {1, ..., k}
we have
|ỹj (n) − yj (n)| ≤ for all j ∈ {1, ..., k} and all n ∈ N
The statement that the absolute value of each component of a vector v is

less than or equal to δ is equivalent to: kvk∞ ≤ δ. Hence, the definition of
stability can be rephrased as follows: a solution y of the vectorial equation
(1.39) is stable iff, for every > 0, there exists a δ > 0, such that every
solution ỹ with the property that kỹ(0) − y(0)k∞ ≤ δ, satisfies the condition
kỹ(n) − y(n)k∞ ≤ for all n ∈ N
As all norms on Ck are equivalent, the ∞-norm in the above definition can
be replaced with any other vector norm on Ck . (Cf. Simon & Blume, §29.4
and exercise 30.4. There, the ∞-norm is denoted by N0 .)
Definition 1.4.5. A solution y of the vectorial equation (1.39) is stable iff,

for a given norm k.k on Ck and for every > 0, there exists a δ > 0, such
that every solution ỹ with the property that kỹ(0) − y(0)k ≤ δ, satisfies the
condition
kỹ(n) − y(n)k ≤ for all n ≥ 0
A stable solution is called asymptotically stable, iff there exists a positive
number δ, such that, for every solution ỹ with the property that kỹ(0) −
y(0)k ≤ δ, we have
lim (ỹ(n) − y(n)) = 0
n→∞
A stable solution that is not asymptotically stable is called neutrally stable.

Every solution that is not stable is called unstable.
55
1.4.1 Homogeneous systems of first order linear differ-
ence equations
We consider a system of k first order, homogeneous, linear difference equa-
tions, in the form of the vectorial equation
y(n + 1) = A(n)y(n) (1.46)

where, for every n, A(n) is a k × k matrix. Analogously to the case k = 1
(cf. §1.3.2), for every vector solution y of (1.46), the following relation holds
y(n) = A(n − 1)A(n − 2)...A(0)y(0), n≥1
It should be noted, however, that, in general, the matrices A(n−1), A(n−2),
etc., do not commute and thus cannot be interchanged. One easily verifies
that y(0) can be chosen arbitrarily, i.e. for every k-dimensional vector c,
the sequence of vectors defined by
y(0) = c, y(n) = A(n − 1)A(n − 2)...A(0)c n ≥ 1, c ∈ Ck
satisfies the equation (1.46). Clearly, this is the general form of the solution.
Theorem 1.4.6. The solutions of the vectorial difference equation (1.46)

form a k-dimensional linear space. A collection of k solutions {y1 , ..., yk }
is a basis of the solution space, if and only if det(y1 (n), ..., yk (n)) (i.e. the de-
terminant of the matrix with column vectors y1 (n), ..., yk (n)) does not vanish
identically (i.e. is nonzero for at least one value of n).
Proof. The proof of the first statement is left to the reader. Let l ∈ N, let
y1 , ..., yl be solutions of (1.46) and c1 , ..., cl ∈ C. Then
c1 y1 (n) + ... + cl yl (n) = A(n − 1)...A(0)(c1 y1 (0) + ... + cl yl (0))
for all n ≥ 1. Hence it follows that c1 y1 (n) + ... + cl yl (n) = 0 for all n ∈
N iff c1 y1 (0) + ... + cl yl (0) = 0, and thus the solutions y1 , ..., yl of (1.46)
are linearly independent iff y1 (0), ..., yl (0) are linearly independent vectors
in Ck . Therefore, the dimension of the solution space equals that of Ck .
The solutions y1 , ..., yk constitute a basis of the solution space of (1.46) iff
y1 (0), ..., yk (0) are linearly independent vectors in Ck , or, equivalently, iff
det(y1 (0), ..., yk (0)) 6= 0. As
det(y1 (n), ..., yk (n)) = det A(n − 1)... det A(0) det(y1 (0), ..., yk (0))
56
for all n ≥ 1, det(y1 (n), ..., yk (n)) 6= 0 for at least one value of n iff
det(y1 (0), ..., yk (0)) 6= 0. This completes the proof of the theorem.
A k × k matrix function Y , having, for its column vectors, k linearly inde-
pendent solutions y 1 , ..., y k of the homogeneous vectorial difference equation
(1.46), is called a fundamental matrix of this equation. It is easily seen
that Y itself satisfies the equation, i.e.
Y (n + 1) = A(n)Y (n), n∈N
Moreover, Theorem 1.4.6 implies that det Y (n) 6= 0 for at least one value
of n. Conversely, the column vectors of a matrix function Y (n) with the
property that Y (n + 1) = A(n)Y (n) for all n ∈ N, also satisfy (1.46). If
det Y (n) 6= 0 for at least one value of n, then, according to Theorem 1.4.6,
these column vectors form a basis of the solution space of (1.46), hence Y
is a fundamental matrix of this equation. If Y is a fundamental matrix of
(1.46), with column vectors y 1 , ..., y k , then every solution y of (1.46) can be
written as a linear combination of y 1 , ..., y k :
c1
 
k
X  . 
y(n) = ci y i (n) = Y (n) 
 
.

 
i=1
ck
where c1 , ..., ck ∈ C. In solving initial value problems of the form
y(n + 1) = A(n)y(n), y(0) = y0
it is sometimes convenient to use the particular fundamental matrix Y of
(1.46) with the property that Y (0) = I. Then the solution is given by
y(n) = Y (n)y0
If equation (1.46) is derived from a scalar, kth order, homogeneous linear
difference equation, by means of the transformation described in the previous
section, then, for j ∈ {1, ..., k}, y j has the form
y j (n) = (yj (n), yj (n + 1), ..., yj (n + k − 1))
where yj is a solution of the scalar equation. Hence the determinant of

the fundamental matrix Y of (1.46) with column vectors y 1 , ..., y k in
that case is equal to the Casorati determinant Cy1 ,...,yk .
57
1.4.2 Inhomogeneous systems of first order linear dif-
ference equations
Next, we consider the inhomogeneous, first order, vectorial difference equa-
tion
y(n + 1) = A(n)y(n) + b(n), n ∈ N (1.47)

where A(n) again is a k × k matrix. Analogously to the case of a scalar
equation, we have the following theorem.
Theorem 1.4.7. Every solution of the inhomogeneous, linear, vectorial dif-

ference equation (1.47) can be written as the sum of a fixed (‘particular’)
solution y0 and a solution of the homogeneous equation (1.46). Conversely,
every sequence of vectors that can be written as the sum of a solution of
(1.47) and a solution of the homogeneous equation, is again a solution of
(1.47).
The general solution of (1.47) can be determined, once a particular so-

lution of (1.47) is known. Like in the scalar case, there exists a method of
‘variation of constants’ for systems of the form (1.47).
The following analogue of theorem 1.3.14 is easily proved (check this!)
Theorem 1.4.8. All solutions of the linear vectorial difference equation

(1.47) are neutrally stable, globally asymptotically stable, or unstable, iff
the null solution of the homogeneous vectorial equation (1.46) is neutrally
stable, asymptotically stable, or unstable, respectively.
1.4.3 Systems of first order linear difference equations

with constant coefficients
This section is concerned with systems of the form (1.40), where the sequences
aij (i ∈ {1, ..., k}, j ∈ {1, ..., k}) do not depend on n, so are constants. Again,
we write the system as a vectorial equation
y(n + 1) = Ay(n) + b(n) (1.48)

Here, A is a k × k matrix with constant entries (i.e. independent of n). The
homogeneous equation
58
y(n + 1) = Ay(n) (1.49)
has the general solution
y(n) = An c, c ∈ Ck
and, for every solution, the following relation holds:
y(n) = An y(0) (1.50)

If the k vectors c1 , ..., ck ∈ Ck are linearly independent, the sequences y1 , ..., yk
defined by
yj (n) = An cj , j = 1, ..., k, n ∈ N
form a basis of the solution space of (1.49). The matrix function Y with
column vectors y1 , ..., yk is a fundamental matrix of (1.49). Thus, if C is a
constant matrix with linearly independent column vectors, i.e. any nonsin-
gular matrix, then
Y (n) = An C, n ∈ N
is a fundamental matrix of (1.49). An obvious choice for C is the identity
matrix. Thus, the column vectors of the matrix An span the solution space.
An is the unique fundamental matrix of (1.49) with the property that Y (0) =
I. Another possibility is: C = S, where S is a matrix of (generalized)
eigenvectors, associated with a coordinate transformation converting A into
‘its’ Jordan normal form J, i.e.: S −1 AS = J. Thus,
An S = SJ n (1.51)
is another fundamental matrix of (1.49). If A has k linearly independent

eigenvectors s1 , ..., sk (this is not always the case!) with eigenvalues λ1 , ..., λk ,
respectively, we could choose S = (s1 , ..., sk ) and J = diag{λ1 , ..., λk }. In that
case, the column vectors of the fundamental matrix An S are:
λn1 s1 , ..., λnk sk
If, on the other hand, A has a single eigenvalue λ with geometric multiplicity 1
and generalized eigenvectors s1 , ..., sk , such that, for j = 2, ..., k, (A−λI)sj =
59
sj−1 , then J consists of a single Jordan block and the column vectors of An S
take the form
! ! !
n n n n−1 n n−1 n
λ s1 , λ s2 + λ s1 , ..., λn sk + λ sk−1 + ... + λn−k+1 s1
1 1 k−1
(1.52)
(Cf. the appendix, §A.1 and §A.2.)
Example 1.4.9. The vectorial difference equation

!
5 1
y(n + 1) = y(n), n∈N
−1 3
has a fundamental matrix of the form

!n ! !n !
5 1 1 0 4 1 1 0
Y1 (n) = =
−1 3 −1 1 0 4 1 1
Note that Y1 (0) = I. With the aid of §A.2 of the appendix, we find (cf. also
example A.1.6)
! ! ! !
1 0 4n n4n−1 1 0 n−1 4+n n
Y1 (n) = =4
−1 1 0 4n 1 1 −n 4−n
Verify that the column vectors of this matrix do indeed form a basis of the
solution space.
Another fundamental matrix is
! !n ! !
1 0 4 1 4n n4n−1 n−1 4 n
Y2 (n) = = =4
−1 1 0 4 −4n (4−n)4n−1 −4 4−n
Example 1.4.10. A population model (From the lecture notes: Linear

Algebra by W. Hesselink). A population is composed of three generations:
first-year, second-year and third-year individuals. Denoting the number of
ith-year individuals, present in the year n, by yi (n), the composition of the
population in that year can be represented by the population vector y(n) :=
(y1 (n), y2 (n), y3 (n)). In this model, it is assumed that the first-year and
second-year individuals have a constant survival rate of 50% and that all
third-year individuals die. Moreover, the birth rate, i.e. the average number
60
of offspring per individual per year, is assumed to be independent of time
and equal to 0, 23 and 1, respectively. The composition of this population
satisfies the following system of equations
3
y1 (n + 1) = y (n)
2 2
+ y3 (n)
1
y2 (n + 1) = y (n)
2 1
,
1
y3 (n + 1) = y (n)
2 2
equivalent to the vectorial difference equation

3
 
1 0 2
 1
y(n + 1) =  2 0 0  y(n), n∈N

0 12 0
The eigenvalues of the above matrix are the solutions of the equation
3 1 1
−λ3 + λ + = (1 − λ)(λ2 + λ + ) = 0
4 4 4
The matrix has a simple eigenvalue 1 and an eigenvalue − 12 with algebraic

multiplicity 2. The eigenspace associated with the eigenvalue 1 is spanned by
the vector (4, 2, 1). The eigenspace associated with the eigenvalue − 21 turns
out to be 1-dimensional as well and is spanned by the vector (1, −1, 1). We
can find a generalized eigenvector of degree 2, corresponding to the eigenvalue
− 21 , by solving the following system of equations
 1 3
   
1 x1 1
 21 2
1
  x2  =  −1 
0     
 2 2
1 1
0 2 2
x3 1
One of its solutions is    

x1 −2

x
 2 

=  0 
 
x3 2
 3
    −1
0 12
4 1 −2 1 0 0 4 1 −2
 1 1
A :=  2 0 0  =  2 −1 0   0 − 2 1   2 −1 0 
   

0 12 0 1 1 2 0 0 − 21 1 1 2
61
and, for all n ∈ N,
  n  
4 1 −2 1 0 0 2 4 2
n 1  1 
A =  2 −1 0   0 − 2 1   4 −10 4 
  
18
1 1 2 0 0 − 12 −3 3 6
Suppose we are interested in the long term development of a population

consisting of 100 first-year individuals in the year 0. Then we compute
    n  
100 4 1 −2 1 0 0 200
n 1  1 
A  0  =  2 −1 0   0 − 2 1  18  400 
   
0 
1 1 2   0 0 − 12 −300
  
4 1 −2 1 0 0 200
1 1  1 
= 2 −1 0   0 (− 2 )n n(− 2 )n−1  18  400 
 
1 1

2 
0 0 (− 21 )n −300
4 1 −2 2
= 9  2 −1 0   (6n + 4)(− 12 )n 
50   
1 n
1 1 2 −3(− 2
)
1 n
 
4 + (3n + 5)(− 2 )
= 100
9 

2 − (3n + 2)(− 12 )n 
1 + (3n − 1)(− 12 )n
Among other things we find that

 
4
100  
lim y(n) =  2  (1.53)
n→∞ 9
1
(1.53) is an example of the following theorem, which has applications in

demography.
Theorem 1.4.11. Let A > 0 be a primitive, irreducible k × k matrix, with

dominant eigenvalue λ1 and dominant eigenvector s1 . For every vectornorm
k.k on Ck and every solution y of the system (1.49), with positive initial
vector y(0), we have
y(n) s
lim = 1
n→∞ ky(n)k ks1 k
62
Proof. Let {s1 , ..., sk } be a basis of generalized eigenvectors of A, and S the
matrix with column vectors s1 ,...,sk , such that the corresponding change of
coordinates takes A to ‘its’ Jordan normal form J = S −1 AS. Every solution
y of (1.49) can be written in the form

λn1

0 . . 0
0 Jk2 (λ2 )n . . 0
 
 
y(n) = SJ n c =
 
S
 . . . . .  c,
 (1.54)

 . . . . . 

0 0 . . Jkr (λr )n
where c = (c1 , ..., ck ) ∈ Ck . J1 (λ1 ), Jk2 (λ2 ),..., Jkr (λr ) are the Jordan blocks
of J. A being a primitive matrix, with dominant eigenvalue λ1 , we have
λ1 > |λi | for i = 2, ..., r, and this implies that
lim λ−n n
1 Jki (λi ) = 0 (= the null matrix of order ki ) (1.55)
n→∞
for i = 2, ..., r. Moreover, λ−n n

1 J1 (λ1 ) = 1. From (1.54) and (1.55) we deduce
that  
1 0 . . 0
 0 0 . . 0
 

−n
 
lim λ1 y(n) = S  . . . . . c = c1 s 1 (1.56)
n→∞ 
 . . . . .
 

0 0 . . 0
Using the continuity of the norm (cf. Simon & Blume, exercise 30.4), we
obtain
lim λ−n −n
1 ky(n)k = lim kλ1 y(n)k = kc1 s1 k = |c1 |ks1 k (1.57)
n→∞ n→∞
Dividing (1.56) by (1.57) we find, for any solution y of (1.49) with y(0) 6= 0,
y(n) c1 s 1
lim = , provided c1 6= 0
n→∞ ky(n)k |c1 | ks1 k
It remains to be proved that c1 > 0 when y(0) >> 0, as that would imply
|c1 | = c1 . From (1.54) we deduce that c = S −1 y(0) and thus
k k
−1
y(0)j (S −1 )Tj1 = y(0) · (S −1 )∗ e1
X X
c1 = S1j y(0)j = (1.58)
j=1 j=1
63
(The scalar or inner product of two vectors x and y ∈ Ck is defined by:
x · y := kj=1 xj y j . For any matrix B, B ∗ is defined by: B ∗ = B T . Check
P
that, for a k × k matrix B, x · By = B ∗ x · y.) Furthermore, A∗ = AT and
AT (S −1 )∗ e1 = (S −1 A)∗ e1 = (JS −1 )∗ e1 = (S −1 )∗ J ∗ e1 = λ1 (S −1 )∗ e1 ,
so (S −1 )∗ e1 is an eigenvector of AT , with eigenvalue λ1 . Furthermore, it

is known that AT has a positive eigenvector s˜1 with eigenvalue λ1 (cf. the
lecture notes: Matrices, Graphs and Convexity). As the eigenspace of A
corresponding to the eigenvalue λ1 is 1-dimensional, there exists a complex
number c such that (S −1 )∗ e1 = cs˜1 . We are going to prove that c > 0. On
one hand, we have
(S −1 )∗ e1 · s1 = cs˜1 · s1
On the other hand,
(S −1 )∗ e1 · s1 = e1 · S −1 s1 = e1 · e1 = 1
This implies that

1
c= >0
s˜1 · s1
as both s1 and s˜1 are positive vectors. Consequently, (S −1 )∗ e1 >> 0. If
y(0) >> 0, (1.58) implies that c1 is the scalar product of 2 positive vectors,
so c1 > 0. This completes the proof of the theorem.
Remark 1.4.12. In the case discussed in example 1.4.10, the condition

y(0) >> 0 doesn’t hold. However, we do have y(4) >> 0 and this is sufficient
for the conclusion of theorem 1.4.11 to be true. (Check this!)
In most applications, one is interested in real-valued solutions only. If A ∈

Rk×k and y0 ∈ Rk , the solution of the initial value problem: y(n+1) = Ay(n),
y(0) = y0 , given by {An y0 }∞
n=0 , is real-valued. In this case, the column vectors
of the fundamental matrix An , form a real-valued basis of the solution space
of (1.49), but in general are not so easy to compute. The column vectors
of the fundamental matrix SJ n , where J is a Jordan normal form of A and
S −1 AS = J, are easier to compute, but will not all be real-valued if A has
eigenvalues with nonzero imaginary part. In that case, the following Theorem
can be used to find a simple real-valued basis.
64
Theorem 1.4.13. Let A ∈ Rk×k and suppose A has an eigenvalue λ with
nonzero imaginary part and eigenvector v. Then the sequences y1 and y2
defined by
y1 (n) = Re λn v and y2 (n) = Im λn v
are linearly independent solutions of (1.49).
Remark 1.4.14. By assumption, Av = λv. Taking complex conjugates, we

get Av = λv = λv. As A ∈ Rk×k , Av = Av = Av. Thus, A has an eigenvalue
n
λ with eigenvector v and {λn v}∞ ∞
n=0 and {λ v}n=0 are linearly independent
solutions of (1.49). Note that y1 (n) = 12 (λn v+λn v) and y2 (n) = 2i1 (λn v−λn v).
Example 1.4.15.
 
2 1 1
y(n + 1) = Ay(n) :=  0 1 1 
 y(n) (1.59)

−1 −1 0
 
1
A has an eigenvalue 1 with eigenvector s1 :=  −1 , an eigenvalue 1 − i
 
0
 
1
with eigenvector s2 :=  −i 

 and an eigenvalue 1 + i with eigenvector
−1
 
1
 i . Hence a basis of the solution space of (1.59) is formed by
s3 := 

−1
the equilibrium solution y1 ≡ s1 and the (complex conjugated)
√ −iπ/4 √ iπ/4sequences
{(1 − i)n s2 }∞
n=0 = {( 2e ) n
s } ∞
2 n=0 and {(1 + i) n
s 3 = {( 2e )n s3 }∞
n=0 .
The general term of the second solution can be written in the following form
−inπ/4
 
n
√ n e
{(1 − i) s2 = ( 2)  −ie−inπ/4 

e−inπ/4
By Theorem 1.4.13, a real-valued basis is formed by y1 and the sequences
  ∞   ∞

 cos(nπ/4) 
 
 − sin(nπ/4) 

n/2  n/2 
Re y2 = 2  − sin(nπ/4)  and Im y2 = 2  − cos(nπ/4) 
 
− cos(nπ/4) n=0
   
 
sin(nπ/4) 
n=0
65
A constant sequence y ≡ c, with c ∈ Ck , c 6= 0, is an equilibrium solution
of the homogeneous equation (1.49) iff Ac = c, i.e. iff c is an eigenvector of
A with eigenvalue 1. A periodic solution of (1.49), of period p, is a solution
y with the property that y(n + p) = y(n), so
Ap y(n) = y(n), n ∈ N
This implies that (1.49) has a p-periodic solution iff A has an eigenvalue λ
with the property that
λp = 1 and λr 6= 1 for r ∈ {1, ..., p − 1},
for example, λ = e2πi/p .

Analogously to the scalar case (i.e. k = 1, cf. example 1.3.10), the
inhomogeneous equation (1.48) (where b is not the null sequence), has a
particular solution y of the form
n−1
An−m−1 b(m) for n ≥ 1
X
y(0) = 0, y(n) = (1.60)
m=0
This can be verified by inserting the above expression for y into the equation:
n n−1
An−m b(m) − An−m b(m) = b(n) for n ≥ 1
X X
y(n + 1) − Ay(n) =
m=0 m=0
and
y(1) − Ay(0) = y(1) = A0 b(0) = b(0)
Hence the general solution of (1.48) is
n−1
y(0) = c, y(n) = An c + An−m−1 b(m) for n ≥ 1 c ∈ Ck
X
m=0
Suppose that b is periodic, of period p. We can find the periodic solutions of

period p of (1.48), by choosing an initial vector c such that, for all n ∈ N,
the following identity holds
n+p−1 n−1
An+p c + An+p−m−1 b(m) = An c + An−m−1 b(m)
X X
(1.61)
m=0 m=0
66
Due to the periodicity of b,
n+p−1 n+p−1 n−1
n+p−m−1 n+p−m−1
An−m−1 b(m)
X X X
A b(m) = A b(m − p) =
m=p m=p m=0
This implies that (1.61) is equivalent to

p−1
An (I − Ap )c = An Ap−m−1 b(m) for all n ∈ N
X
m=0
hence also to
p−1
(I − Ap )c = Ap−m−1 b(m)
X
m=0
(why?) This shows that (1.48) has a unique periodic solution of period p,
iff the matrix I − Ap is nonsingular. The initial vector of this solution is
p−1
y(0) = (I − Ap )−1 Ap−m−1 b(m)
X
m=0
1.4.4 Stability of solutions of systems of linear differ-

ence equations with constant coefficients
A solution of the vectorial difference equation
y(n + 1) = Ay(n) + b(n) (1.62)

where A is a constant k × k matrix, is stable, or asymptotically stable, if
the null solution of the homogeneous equation is stable, or asymptotically
stable, respectively and vice versa. For every solution y of the homogeneous
equation the following relation holds
y(n) = An y(0)
Thus, (asymptotic) stability of solutions of (1.62) can be expressed as follows

in terms of the properties of A. All solutions are stable if, for any positive
number , there exists a positive number δ, such that, for every vector x ∈ Ck
satisfying the condition kxk ≤ δ, we have kAn xk ≤ for all n ∈ N. All
solutions are (globally) asymptotically stable if, in addition, for all x ∈ Ck ,
limn→∞ An x = 0. Here, k.k is some vectornorm on Ck , for example, k.k1 (N1
in Simon & Blume), or k.k∞ (N0 in Simon & Blume).
67
Definition 1.4.16. Let A ∈ Ck×k . The spectrum σ(A) of A is the collection
of all eigenvalues of A. The spectral radius rσ (A) of A is defined by
rσ (A) = max{|λ| : λ ∈ σ(A)}
Without proof we state the following theorem.
Theorem 1.4.17. The solutions of (1.62) are stable if the following condi-
tions hold:
1) rσ (A) ≤ 1,
2) the algebraic and geometric multiplicities of every eigenvalue λ of A such
that |λ| = 1 are equal.
In all other cases the solutions are unstable. The solutions of (1.62) are
globally asymptotically stable iff rσ (A) < 1.
From this theorem two necessary (but not sufficient) conditions for stabil-
ity and asymptotic stability of the solutions of (1.62) can be deduced. Sup-
pose the Jordan normal form of the matrix A has diagonal entries λ1 , ..., λk
(the eigenvalues of A). If the solutions are asymptotically stable, then, ac-
cording to Theorem 1.4.17, |λi | < 1 for all i. As
k
Y k
X
det A = λi and tr A = λi
i=1 i=1
(this follows from lemma A.1.1 in the appendix), this implies that
| det A| < 1 and |tr A| < k
Analogously, stability of the solutions of (1.62) implies that
| det A| ≤ 1 and |tr A| ≤ k
The homogeneous equation y(n + 1) = Ay(n) is a simple example of an

autonomous system of the form y(n+1) = f (y(n)) (cf. §1.5), with f (x) = Ax.
Analogously to the scalar case, we define the stable set of the fixed point 0
(i.e. the null vector) of f to be the set of all vectors x ∈ Ck with the
property that limn→∞ An x = 0. Suppose x is an eigenvector of A, with
eigenvalue λ. Then An x = λn x and it is obvious that limn→∞ An x = 0 iff
|λ| < 1. With the aid of (1.52), it is easily shown that this also holds when x
is a generalized eigenvector with respect to λ, of degree greater than 1. Let
68
EAs denote the linear subspace of Ck spanned by all generalized eigenvectors
of A, corresponding to eigenvalues λ with the property that |λ| < 1, i.e.
EAs =
M
Eλ
λ∈σ(A):|λ|<1
where Eλ denotes the space of all generalized eigenvectors, corresponding to

the eigenvalue λ. We have the following result.
Theorem 1.4.18. (i) EAs is the stable set of the fixed point 0 of the linear
map f : x 7→ Ax.
(ii) If y and y∗ are solutions of the system (1.62), then limn→∞ y(n)−y∗ (n) = 0
iff y(0) − y∗ (0) ∈ EAs .
Example 1.4.19. A dynamic multiplier model involving two coun-

tries
Model equations:
Yi = Ci + Ii + Xi − Mi
Ci,t = ci Yi,t−1
Ii,t = Ii,0 + hi Yi,t−1
Mi,t = Mi,0 + mi Yi,t−1
Xi = Mj . i 6= j
Here, Yi , Ci , Ii , Mi and Xi denote production, consumption, investments,

import and export of country i, respectively. Ii,0 and Mi,0 are positive con-
stants. ci , hi and mi are numbers!between 0 and 1, such that ci +hi −mi > 0.
Y1
i = 1, 2. The vector Y := satisfies the following autonomous system
Y2
of linear, inhomogeneous difference equations
! !
c1 + h1 − m1 m2 I1,0 + M2,0 − M1,0
Yt = Yt−1 + (1.63)
m1 c2 + h2 − m2 I2,0 + M1,0 − M2,0
69
1.4.5 Exercises
1. Determine the general solution of the vectorial difference equation
!
2 −1
y(n + 1) = y(n), n∈N
1 4
!
1
and find the solution satisfying the initial condition: y(0) = .
1
!
1
2. Prove that is an eigenvector of the matrix
1+i
!
2 −1
A=
2 0
and determine the general, real-valued, solution of the equation
y(n + 1) = Ay(n), n ∈ N
3. Determine the general solution of the system

! !
5 1 1
y(n + 1) = y(n) + 4n , n∈N
−1 3 −1
4. Determine the solution of the initial value problem

   
1 0 2 1
y(n + 1) =  0 1 3  y(n), y(0) =  −1 
   
0 0 1 1

   
−1 −1 0 1
y(n + 1) =  0 −1 −1  y(n), y(0) =  2 
  

0 0 −1 −2
6. Determine a fundamental matrix of the system

 
1 0 1
y(n + 1) =  1 1 1  y(n), n∈N
 
1 0 1
70
7. Find the population vector y(n) in example 1.4.10, when the initial pop-
ulation is composed of 100 second-year individuals. Compute limn→∞ y(n).
Check whether there exists an initial vector y(0) such that, in the long
run, the population becomes extinct.
8. Determine a Jordan normal form and a basis of generalized eigenvectors

of the matrix in the model of example 1.4.10, when the birth rates
for the first-year, second-year and third-year individuals are 21 , 12 and
1, respectively. In this case we are dealing with a so-called Markov
matrix (i.e. a matrix with nonnegative entries, having the additional
property that the components of each column vector add up to 1).
Compute 3i=1 yi (n).
P
9. Examine the stability of the solutions of the system

1
− 13 1
   
2 3
1
y(n + 1) =  − 13 2
 y(n) +  2n  ,
0  n∈N
  
3
1 1
3
0 3
3n
10. a. Determine all equilibrium solutions of the linear difference equation

   
0 −1 0 1
y(n + 1) =  0

0 −1  y(n) +  2 
 
 , n = 0, 1, ... (1.64)
1 1 1 −1

b. Determine all real-valued solutions of the homogeneous part.
c. Determine
 the
 solutions of the equation in a. with initial vector
0
y(0) =  0 .
 
1
11. Find all periodic solutions of the system
!
−1 1
y(n + 1) = y(n)
a −1
where a is a real number, and examine their stability. (Hint: distinguish

different ranges of a.)
71
Figure 1.11: Orbits of various points (each of which is marked by a cross),
subject to the map f (x) = (x2 , αx2 (1 − x1 )), with α = 2.01; cf. example
1.5.6.
1.5 Autonomous systems of first order differ-

ence equations
In this section we extend the theory discussed in §1.2 to systems of first order
difference equations of the type:
y(n + 1) = f (y(n)), n∈N (1.65)
where f is a k-dimensional vector function on a subset Df of Rk or Ck , with

the property that f (Df ) ⊂ Df . The notions: kth iterate of f , (positive)
invariant set, (positive) orbit of an element of Df , or a solution of (1.65),
positive limit point or ω-limit point, stable set, attractor and basin of at-
traction are defined analogously to the scalar case. Like in the scalar case,
equilibrium solutions of (1.65) correspond to fixed points of f and p-periodic
solutions to fixed points of its pth iterate f p .
For systems of difference equations (with k ≥ 2) it is no longer feasible
to draw graphs of solutions as functions of “time”, since that would require
a k + 1-dimensional space. Instead, one can plot a single coordinate of the
solution vector versus time. If k = 2, orbits of different points may be drawn,
as shown in figs. 1.11 and 1.12. An equilibrium solution corresponds to a
single point, a p-periodic solution to p points (which are often connected by
line segments).
72
Figure 1.12: Orbit of the point ( 12 , − 21 ), subject to the map f (x) = (x2 , 1−x21 );
cf. example 1.5.7.
Example 1.5.1. The iterates of the function f : R2 −→ R2 , defined by

!
x2
f (x) := , x ∈ (R − {0})2 ,
x−1
1 x2
where x = (x1 , x2 ), are easily computed:
x−1 x−1 x−1

! ! !
2 1 x2 3 1 4 2
f (x) = , f (x) = , f (x) = ,
x−1
1 x−1
2 x−1
2 x1
x−1
!
5 2 x1
f (x) = , f 6 (x) = x
x1
!
1
f has a single fixed point, viz. . The only iterate to have more than
1
! !
3 −1 1
one fixed point is f , the additional fixed points being , and
−1 −1
!
−1
. The vectorial difference equation
1
! !
y1 (n + 1) y2 (n)
=
y2 (n + 1) y1 (n)−1 y2 (n)
73
!
1
has a unique equilibrium solution: y ≡ , no 2-periodic solutions and
1
three 3-periodic solutions having one and the same orbit. All other solutions
are 6-periodic.
The following result is a generalization of Theorem 1.2.5.
Theorem 1.5.2. Let c ∈ Ck be a fixed point of f and suppose f is differ-

entiable at c. Let A := Df (c) (the Jacobian matrix of f at c).
(i) If rσ (A) < 1, then y ≡ c is an asymptotically stable equilibrium solu-
tion of (1.65).
(ii) If rσ (A) > 1, then y ≡ c is an unstable equilibrium solution of (1.65).
In the case that c = 0, similarly to Theorem 1.2.8, the above theorem can be
stated as follows.
Theorem 1.5.3. Let A be a constant k × k matrix and g a k-dimensional

vector function, continuous at 0 (the origin of Rk or Ck ), with the property
that, for some (hence for any!) vectornorm k.k,
kg(x)k
lim =0 (1.66)
x→0 kxk
If rσ (A) < 1, then the null solution of the system of first order difference
equations
y(n + 1) = Ay(n) + g(y(n)) n ∈ N (1.67)
is asymptotically stable. If, on the other hand, rσ (A) > 1, then the null
solution of (1.67) is unstable.
Like in the scalar case, we say that the linear system
y(n + 1) = Ay(n)
is obtained by linearization of (1.67) at 0. More generally, (1.65) can be

linearized at any fixed point c where f is differentiable, and the linearized
equation is obtained from (1.65) by replacing the function f on the right-hand
side by its linear approximation at c, i.e.
y(n + 1) = f (c) + Df (c)(y(n) − c) = c + Df (c)(y(n) − c)
74
Example 1.5.4.
! ! !
y1 (n + 1) y1 (n) y2 (n)2
=A + , n∈N
y2 (n + 1) y2 (n) y1 (n)y2 (n)
where A is a 2 × 2 matrix. In this example

!
x22
g(x) = g(x1 , x2 ) =
x1 x2
For a vectornorm we choose the 1-norm. We have
kg(x)k1 = |x2 |2 + |x1 x2 | = |x2 |kxk1
Hence
kg(x)k1
lim = lim |x2 | = 0
x→0 kxk1 x→0
Example 1.5.5. We illustrate Theorem 1.5.2 with a discrete ‘predator-

prey model’. We assume that in a certain area, after n ‘units of time’ (e.g.
years, months, generations), there are y1 (n) prey animals and y2 (n) predators
present and that these numbers satisfy the following system of equations
y1 (n + 1) = ((1 + r) − ay1 (n) − by2 (n))y1 (n)

n∈N (1.68)
y2 (n + 1) = cy1 (n)y2 (n)
Here, a, b, c and r are positive constants. The number of prey that are eaten
during one unit of time (viz. by2 (n)y1 (n)) is proportional to both the number
of prey and the number of predators present, and the same applies to the
number of predators born during that period. Moreover, it is assumed in this
model that no predator lives more than 1 unit of time. The system (1.68)
has three equilibrium solutions, including the null solution and the solution
! !
1
y1 (n)
= 1
c for all n ∈ N (1.69)
y2 (n) b
(r − ac )
(check this). The Jacobian matrix of the vector function

!
(1 + r − ax1 − bx2 )x1
f (x) :=
cx1 x2
75
where x := (x1 , x2 ), is
!
1 + r − 2ax1 − bx2 −bx1
Df (x) =
cx2 cx1
Df (0, 0) is a diagonal matrix with diagonal entries 1+r and 0. From Theorem
1.5.3 it follows immediately that the null solution is unstable. In most cases
however, the more interesting question is, whether or not the equilibrium
solution (1.69) is (asymptotically) stable. In the given (real) context, this
solution exists, provided r > ac . The matrix A now reads
!
1 − ac − cb
A= 1
b
(cr − a) 1
The eigenvalues of this matrix are the solutions of the equation

a a
(λ − 1)(λ − 1 + ) + r − = 0
c c
For instance, let r = 38 and ac = 14 . In that case, the matrix has two complex
conjugated eigenvalues λ+ and λ− (so λ+ = λ− ). Thus we have
2a 7
|λ+ |2 = |λ− |2 = λ+ λ− = det A = 1 + r − =
c 8
This shows that rσ (A) < 1 and, with Theorem 1.5.3, we conclude that the
equilibrium solution under consideration is asymptotically stable. The
linearized system at the equilibrium point (1.69) is
! ! ! !
1
y1 (n + 1) c
1 − ac − cb y1 (n) − 1c
= 1 +
y2 (n + 1) b
(r − ac ) 1
b
(cr − a) 1 y2 (n) − 1b (r − ac )
Example 1.5.6. The delayed logistic equation.

This is the second order autonomous difference equation
u(n + 2) = αu(n + 1)(1 − u(n))
which, by a transformation of the form

!
u(n)
y(n) := (1.70)
u(n + 1)
76
Figure 1.13: The attractor of the equation in example 1.5.6, with α = 2.01.
The white area is the basin of attraction.
is converted into the system of first order equations y(n+1) = f (y(n)), where
!
x2
f (x) =
αx2 (1 − x1 )
If α 6= 1, f has two fixed points: the origin and (1 − 1/α, 1 − 1/α). The
Jacobian matrix of f is
!
0 1
Df (x) =
−αx2 α(1 − x1 )
Df (0, 0) has eigenvalues 0 and α, hence the null solution of the equation is
asymptotically stable when |α| < 1 and unstable when |α| > 1. The Jacobian
matrix at the second equilibrium point is
!
0 1
Df (1 − 1/α, 1 − 1/α) =
1−α 1
√
It has
√ eigenvalues λ = 1/2 ± 1/2 5 − 4α when α < 5/4 and λ = 1/2 ±
i/2 4α − 5 when α > 5/4. Both eigenvalues have absolute value < 1 iff 1 <
α < 2. For these values of α, the equilibrium solution y ≡ (1 − 1/α, 1 − 1/α)
is therefore asymptotically stable. At α = 2 a bifurcation occurs: for values
of α > 2, the system no longer has a point attractor, cf. fig. 1.13.
Example 1.5.7. Next, we consider the second order, autonomous difference

equation
u(n + 2) = a − u(n)2
77
By a transformation of the form (1.70), it is converted into the system of
first order equations y(n + 1) = f (y(n)), where
!
x2
f (x) =
a − x21
The iterates of f are easily computed:

! ! !
2 a − x21 3 a − x22 4 a − (a − x21 )2
f (x) = , f (x) = , f (x) =
a − x22 a − (a − x21 )2 a − (a − x22 )2
When a < −1/4, f has no real fixed points, when a > −1/4, √ it has two
real fixed points: c± =√(c±
1 , c±
2 ), where c +
1 = c +
2 = −1/2 + 1/2 1 + 4a, and
− − ±
c1 = c2 = −1/2 − 1/2 1 + 4a. The Jacobian matrix of f at c is
!
± 0 1
Df (c ) = ±
−2c1 0
qq √
For the eigenvalues we have: λ2 = −2c±
1 , so |λ| = 2|c±
1 | = | − 1 ± 1 + 4a|.
q √
The eigenvalues of Df (c− ) have absolute value 1 + 1 + 4a > 1 for all
a > −1/4, hence the equilibrium solution y ≡ c− is qunstable for all a > −1/4.
+
√
The eigenvalues of Df (c ) have absolute value | 1 + 4a − 1| and this is
less than 1 when −1/4 < a < 3/4 and greater than 1 for all a > 3/4. Thus the
equilibrium solution y ≡ c+ is asymptotically stable for all a ∈ (−1/4, 3/4)
and unstable for all a > 3/4. f 2 has two more fixed points, in addition to
c+ and c− , viz. p1 := (c− + + −
1 , c2 ) and p2 := (c1 , c2 ). These correspond to two
2
2-periodic solutions with orbit {p1 , p2 }. The
√ Jacobian matrix √ of f at p1 and
p2 is diagonal, with diagonal entries 1 − 1 + 4a and 1 + 1 + 4a. Hence it
follows that both 2-periodic solutions are unstable for all a > −1/4. Finally,
we look for 4-periodic solutions. To that end we try to solve the equation
! !
4 a − (a − x21 )2 x1
f (x) = =
a − (a − x22 )2 x2
This requires computing the zeroes of the 4th degree function a−(a−x2 )2 −x.
Since all fixed points of f and f 2 are fixed points of f 4 as well, all zeroes of
a − x2 − x are zeroes of a − (a − x2 )2 − x. This implies that a − x2 − x divides
a−(a−x2 )2 −x. Division yields: a−(a−x2 )2 −x = (a−x2 −x)(x 2
√ −x+1−a).
When a > 3/4, the second factor has 2 real zeroes: 1/2 ± 1/2 4a − 3. Hence
78
Figure 1.14: 4-periodic attractor of the map f (x) = (x2 , 1 − x21 ) and its basin
of attraction (the white region). Which points in the white region do not
belong to the basin of attraction? Also compare this figure to fig. 5.2
there are in total 16 real, periodic solutions of period 4. For 12 of these,

4 is the minimal period. Thus, there are 3 4-periodic orbits of the form
{c, f (c), f 2 (c), f 3 (c)}. The corresponding values of c are:
√ ! √ ! √ !
1+ 4a−3 1+ 4a−3 1+ 4a−3
c= √2 , c= 2
√ and c = √2
1+ 4a−3 −1+ 4a+1 −1− 4a+1
2 2 2
The Jacobian matrix of f 4 is a diagonal matrix, with diagonal entries 4x1 (a−
x21 ) and 4x2 (a − x22 ). At the first point, both entries are equal to 4 − 4a.
Hence the orbit of this point is a 4-periodic attractor for all values of a in
the interval (3/4, 5/4). In Figure 1.14 this attractor is shown, along with its
basin of attraction, for the case a = 1.
If a fixed point c of f has the property that none of the eigenvalues of Df (c)
has absolute value equal to 1, it is called a hyperbolic fixed point. For
a hyperbolic fixed point, Theorem 1.5.2 provides a necessary and sufficient
condition for stability of the equilibrium solution y ≡ c. Note that, in that
case, neutral stability does not occur. In the case of a non-hyperbolic fixed
point, linearization doesn’t give us any information about the stability of the
corresponding equilibrium solution of (1.65). In some situations, the so-called
direct method of Lyapunov may help us out.
Theorem 1.5.8. Let f be a continuous vector function on Rk or Ck , with

fixed point x0 . Suppose there exists a neighbourhood U of x0 and a continu-
ous function V : U → R with the property that V (x0 ) = 0 and V (x) > 0 for
79
all x ∈ U − {x0 }.
(i) If V (f (x)) ≤ V (x) for all x ∈ U such that f (x) ∈ U , then x0 is a stable
equilibrium point.
(ii) If V (f (x)) < V (x) for all x ∈ U − {x0 } such that f (x) ∈ U , then x0 is
an asymptotically stable equilibrium point.
(iii) If V (f (x)) > V (x) for all x ∈ U − {x0 } such that f (x) ∈ U , then x0 is
an unstable equilibrium point.
Proof. (i) Let be a sufficiently small positive number, so that the closed
“ball” B(x0 ; ) := {x ∈ Ck : kx − x0 k ≤ }, where k.k denotes a vectornorm
on Ck , is contained in U . Then V (x) > 0 for all x ∈ B(x0 ; ) − {x0 }. f is
continuous at x0 , hence there exists a δ1 ∈ (0, ) with the property that
kf (x) − f (x0 )k ≤ for all x ∈ B(x0 ; δ1 ) (1.71)
Let
m := min{V (x) : δ1 ≤ kx − x0 k ≤ } (1.72)
Then we have: m > 0 (why?). Since V is continuous and V (x0 ) = 0, there
exists a δ2 ∈ (0, δ1 ), such that
V (x) < m for all x ∈ B(x0 ; δ2 )
We want to prove that f n (x) ∈ B(x0 ; δ1 ) for every n ∈ N and every x ∈
B(x0 ; δ2 ), as this implies
kf n (x)−f n (x0 )k = kf n (x)−x0 k ≤ δ1 < for all x ∈ B(x0 ; δ2 ) and all n ∈ N
hence the stability of the equilibrium point x0 . We give a proof by contra-
diction: suppose there exist x ∈ B(x0 ; δ2 ) and n ∈ N such that f n (x) 6∈
B(x0 ; δ1 ). Let n0 denote the smallest value of n with this property. Thus,
n0 ≥ 1, as f 0 (x) = x ∈ B(x0 ; δ2 ) ⊂ B(x0 ; δ1 ). Now we have
f n0 −1 (x) ∈ B(x0 ; δ1 ), but f n0 (x) 6∈ B(x0 ; δ1 ) (1.73)
From (1.73) and (1.71) we deduce that
δ1 < kf n0 (x) − x0 k = kf (f n0 −1 (x)) − f (x0 )k ≤
In view of (1.72), this implies V (f n0 (x)) ≥ m . On the other hand, from the
properties of V and the fact that f n (x) ∈ B(x0 ; δ1 ) ⊂ U for all n ≤ n0 − 1,
it follows that
V (f n0 (x)) ≤ V (f n0 −1 (x)) ≤ ... ≤ V (x) < m
80
Thus we have produced a contradiction and hence f n (x) ∈ B(x0 ; δ1 ) for every
n ∈ N and every x ∈ B(x0 ; δ2 ).
(iii) Choose > 0 such that B := B(x0 ; ) ⊂ U . We will show that no solution
with initial vector x ∈ B − {x0 } “stays within B”. Again we give an indirect
proof. Suppose that f n (x) ∈ B for all n ∈ N. B being bounded and closed,
thus compact, there exists a subsequence of {f n (x)}∞ n=0 converging to an
nm ∞
element of B. Denote this subsequence by {f (x)}m=1 , where n1 < n2 < ...
and let x0 = limm→∞ f nm (x). Due to the continuity of V , the sequence
V (f nm (x)) converges to V (x0 ). The assumption that V (f (x)) > V (x) > 0
for all x ∈ U − {x0 }, implies that V (f n (x)) is an increasing sequence of
positive numbers, bounded above by the maximum of V on B. Hence the
sequence is convergent and
lim V (f n (x)) = V (x0 ) > 0
n→∞
Consequently,
lim V (f n+1 (x)) = V (x0 )
n→∞
On the other hand,
lim V (f n+1 (x)) = V (f (x0 )) > V (x0 )
n→∞
which contradicts the previous statement.

The proof of (ii) is left to the reader (cf. also the proof of Theorem 1.5.11).
A continuous function V : U → R satisfying the conditions of part (i) of

Theorem 1.5.8, is called a Lyapunov function on U , centered at x0 , for
equation (1.65). A Lyapunov function satisfying the conditions of part (ii)
of Theorem 1.5.8 is called a strict Lyapunov function on U , centered at
x0 , for equation (1.65). Finding a suitable Lyapunov function often is not
an easy task. Sometimes, a natural candidate presents itself, like the energy
function in certain problems in physics. In a few simple cases, (the square
of) the distance of the fixed point to the origin is a Lyapunov function for
the equation. (For some economic applications of Lyapunov functions, cf. G.
Gandolfo: Economic Dynamics, Ch. 23.)
Example 1.5.9. The function f : R2 −→ R2 , defined by
!
x1 (1 − x22 )
f (x) =
x2 (1 − x21 )
81
has a fixed point at the origin. The matrix of the linearized system at (0, 0)
in this case is the identity matrix, so Theorem 1.5.2 doesn’t provide any
information about the stability of the null solution of (1.65). Now define
V (x) = kxk22 = x21 + x22
Then we have
V (f (x)) = x21 (1 − x22 )2 + x22 (1 − x21 )2 = x21 + x22 + x21 x22 (x21 + x22 − 4)
Hence it follows that V (f (x)) ≤ V (x), when x21 + x22 ≤ 4. Thus, V is a

Lyapunov function for (1.65) on the interior of the circle about the origin with
radius 2 (U := B(0; 2) = {(x1 , x2 ) ∈ R2 : x21 + x22 < 4), and, by Theorem
1.5.8, the null solution of this equation is stable. The null solution is not
asymptotically stable, as every solution with initial vector (c, 0), c ∈ R, again
is an equilibrium solution. So every neighbourhood of the origin contains
infinitely many equilibrium points.
Lyapunov’s method does not only provide us with a test for the stability of
equilibrium solutions (of periodic solutions). It can also be used to determine
the basin of attraction of an asymptotically stable equilibrium point, or a part
of that basin. We illustrate this with the following example.
Example 1.5.10. The function f : R2 −→ R2 , defined by
!
x2 (1 − 12 (x21 + x22 ))
f (x) =
x1 (1 − 12 (x21 + x22 ))
again has a fixed point at the origin. Df (0, 0) has eigenvalues 1 and −1, so
that Theorem 1.5.2 does not apply. We use the same function V as in the
previous example. In this case we have
1
V (f (x)) − V (x) = (x21 + x22 )2 (x21 + x22 − 4)
4
Hence V (f (x)) < V (x) for every x ∈ B(0; 2) − {0}. Thus, if y is a solution
of (1.65), with initial vector y(0) ∈ B(0; 2), then we have, for every n ∈ N,
ky(n)k2 < ... < ky(1)k2 < ky(0)k2 < 2
Hence it follows that limn→∞ ky(n)k2 exists. As y is a bounded sequence, it

has a subsequence {y(nm )}∞m=1 , with n1 < n2 < ..., converging to an element
82
x of B(0; 2). Due to the continuity of V (= the square of the 2-norm), we
have
lim ky(nm )k2 = n→∞
m→∞
lim ky(n)k2 = kxk2
and
lim ky(nm + 1)k2 = kxk2
m→∞
On the other hand, due to the continuity of f ,
lim y(nm + 1) = m→∞
m→∞
lim f (y(nm )) = f (x)
hence
lim ky(nm + 1)k2 = kf (x)k2
m→∞
Thus, V (f (x)) = V (x) and this implies x = (0, 0). We conclude that
limn→∞ ky(n)k2 = 0 for every solution with initial vector in B(0; 2), and
hence B(0; 2) is contained in the basin of attraction of (0, 0).
Theorem 1.5.11. Let f be a continuous vector function on Rk or Ck , with
fixed point x0 . Suppose there exists a positive invariant neighbourhood U of
x0 , such that, for every x ∈ U , the sequence {f n (x)}∞
n=0 is bounded and all
its accumulation points belong to U , and a continuous function V : U → R
with the property that V (x0 ) = 0, V (x) > 0 and V (f (x)) < V (x) for all
x ∈ U − {x0 }. Then U is contained in the basin of attraction of {x0 }.
Proof. Suppose that y is a solution of (1.65) with initial vector y(0) ∈
U − {x0 }. Then y is bounded and y(n) = f n (y(0)) ∈ U for all n ∈ N.
Now, 0 ≤ V (y(n + 1)) < V (y(n)) for all n ∈ N, so limn→∞ V (y(n)) exists.
Moreover, if {y(nm )}∞
m=1 with n1 < n2 < ..., is a convergent subsequence,
with limit x, then x ∈ U . Due to the continuity of V , limm→∞ V (y(nm )) =
V (x), hence limn→∞ V (y(n)) = V (x). From the continuity of f it follows
that limm→∞ f (y(nm )) = f (x), hence limm→∞ V (f (y(nm ))) = V (f (x)). On
the other hand, limm→∞ V (f (y(nm ))) = limm→∞ V ((y(nm + 1))) = V (x).
Thus, V (f (x)) = V (x) and this implies that x = x0 . The above argument
shows that every convergent subsequence of y has limit x0 and, since y is
bounded, the sequence as a whole converges to x0 .
Remark 1.5.12. If U is positive invariant and bounded, then, for every x ∈
U , the sequence {f n (x)}∞
n=0 is bounded. If U is positive invariant and closed,
then, for every x ∈ U , all accumulation points of the sequence {f n (x)}∞ n=0
belong to U . If U is positive invariant and compact, then, for every x ∈ U ,
both conditions on {f n (x)}∞n=0 in Theorem 1.5.11 are automatically fulfilled.
83
We conclude this section with two examples of simple economic models,
the dynamics of which is described by a system of nonlinear, autonomous
difference equations.
Example 1.5.13. Kaldor’s business-cycle model in discrete time
The model equations are
Yt+1 − Yt = α(I(Yt , Kt ) − S(Yt , Kt ))
Kt+1 − Kt = I(Yt , Kt ) − δKt
Here, Yt and Kt denote production and capital in the period t, respectively.
The investments I and the savings S are functions of Y and K. It is assumed
∂I
in this model that ∂Y is small, both for very small and very large values, and
∂S
larger for intermediate values of Y , whereas ∂Y is supposed relatively large
for very small and very large values of Y . Consequently, both I and S have
to be nonlinear functions of Y (otherwise all partial derivatives would be
constant).
Example 1.5.14. A nonlinear multiplier-accelerator model

The model is described by the following equations.
Ct = (1 − s)Yt−1 + sYt−2
It = a(Yt−1 − Yt−2 ) − b(Yt−1 − Yt−2 )3 + A
Yt = Ct + It
Here, A denotes autonomous investments, assumed independent of the time
t. Both the propensity to save s and the accelerator a are positive constants,
is a number between 0 and 1. Thus, the income Yt in this model satisfies
the second order difference equation
Yt+2 − Yt+1 = (a − s)(Yt+1 − Yt ) − b(Yt+1 − Yt )3 − (1 − )sYt + A, t ∈ N
Introducing a new variable Xt := Yt+1 − Yt , we can rewrite this equation
in the form of a system of first order equations as follows (this is a slightly
modified version of the transformation used in §1.4 and (1.70)):
Xt+1 = c1 Xt − c2 Xt3 − c3 Yt + β
Yt+1 = Xt + Yt
where c1 = a − s, c2 = b, c3 = (1 − )s, β = A.
84
1.5.1 Exercises
1. Deduce Theorem 1.5.3 from Theorem 1.5.2.
2. Prove that the system (1.68) has three unstable equilibrium solutions when
r > 2 ac .
3. Find the equilibrium solutions of the equation (1.65), where

!
x2 (1 − x21 )
f (x) = , x ∈ R2
x1 (1 − x22 )
4. Linearize the equation in example 1.5.6 at the equilibrium point (1 −

1/α, 1 − 1/α). Which scalar, second order equation is equivalent to the
linearized system?
5. Let f denote the map from R2 to R2 , defined by f (x1 , x2 ) = (x2 , αx2 (1 −

x1 )) (cf. example 1.5.6), with α > 1. Prove that the sequence {f n (x1 , x2 )}∞
n=0
diverges when x1 > 1 and 0 < x2 < 1. (Hint: begin by studying the orbits
of points in the third quadrant, next consider those of the points with
x1 < 1, x2 < 0 and, finally those of the set mentioned in the problem.
You may draw inspiration from figures 1.11 and 1.13.)
6. Determine all periodic solutions of period ≤ 2 of the equation (1.65),

where !
1 + x2 − x21
f (x) = , x ∈ R2
x1
and examine the stability of the equilibrium solutions.
7. Compute all real-valued, 4-periodic orbits of the equation in example 1.5.7

and examine the stability of the corresponding solutions.
8. Determine the equilibrium solutions of the equation (1.65), where

 
x2
1+x21
f (x) =  x1
, x ∈ R2
1+x22
85
9. Prove that the basin of attraction of the fixed point of the map in example
1.5.10 is B(0; 2).
10. The map f : R2 → R2 is defined by
f (x1 , x2 ) = (x21 − x22 , −2x1 x2 )
a. Determine all equilibrium solutions of the system y(n + 1) = f (y(n))

b. Prove that the basin of attraction of {(0, 0)} is the set {(x1 , x2 ) ∈ R2 :
x21 + x22 < 1}.
11. a. Convert the scalar difference equation
y(n + 2) = y(n + 1)2 y(n) − y(n)
into a system of first order difference equations.

b. Determine all equilibrium solutions of the system.
c. Prove that any solution starting from a point on one of the coordinate
axes is periodic, of period 4.
d. Using Lyapunov’s direct method, prove that the null solution is neu-
trally stable.
1.6 Higher order linear difference equations

with constant coefficients
The scalar, kth order, linear difference equation
y(n + k) + ak−1 y(n + k − 1) + ... + a0 y(n) + b(n) = 0 (1.74)

with constant coefficients a0 , ..., ak−1 , can be converted into a system of first
order equations, by means of the transformation described in §1.4. Let u
denote the k-dimensional vector, defined by
uj (n) = y(n + j − 1), j = 1, ..., k
Then the kth order equation (1.74) is equivalent to the first order vectorial
equation
86
 
  0
0 1 . . 0 
0 
0 0 1 . 0
   
.
   
   
u(n + 1) =  . . . . .  u(n) +   (1.75)
   . 

. . . . .   
.
   
−a0 . . . −ak−1
 
−b(n)
In what follows, the k × k matrix in (1.75) will be denoted by A. It is called

the companion matrix of the scalar equation (1.74). If u = (u1 , ..., uk )
is a solution of (1.75), then the sequence u1 is a solution of (1.74) and, for
j = 2, ..., k, uj = τ j−1 u1 . Conversely, if u1 is a solution of (1.74), then
(u1 , τ u1 , ..., τ k−1 u1 ) satisfies (1.75).
1.6.1 The homogeneous equation

The behaviour of the solutions of the homogeneous, vectorial equation asso-
ciated with (1.75), is, to a great extent, determined by the Jordan normal
form of A. We begin by computing the eigenvalues of A. These are the zeroes
of the characteristic polynomial of A, i.e. the solutions of the equation
det(A − λI) = 0
The above equation is known as the characteristic equation of (1.74).
Theorem 1.6.1. The eigenvalues of the companion matrix A of the kth

order, linear difference equation (1.74) are the solutions of the characteristic
equation
λk + ak−1 λk−1 + ... + a0 = 0
Proof. For every positive integer j ≤ k, we define a function Pj by

−λ 1 . . 0 0

0

−λ 1 . 0 0

. . . . . .

Pj (λ) =

. . . . . .

0 0 . . −λ 1

−ak−j −ak−j+1 . . −ak−2 −ak−1 −λ
87
Pj is a polynomial in λ of degree j. Pk (λ) = det(A − λI). We will use an
inductive argument to prove that
Pj (λ) = (−1)j (λj + ak−1 λj−1 + ... + ak−j ), j = 1, ..., k
For j = 1 we have: P1 (λ) = −ak−1 − λ. Now, let j > 1 and assume the above
equality holds for j − 1. Expanding the determinant along the first column,
we get:

−λ 1 . . 0 1 0 . . 0

0 −λ 1 . 0 −λ 1 . . 0

+(−1)j+1 .−ak−j

Pj (λ) = −λ . . . . . . . . . .

0 0 . . 1

0 0 . 1 0

−ak−j+1 . . −ak−2 −ak−1 −λ 0 . . −λ 1

The first determinant on the right-hand side equals Pj−1 (λ) and the second
one, being the determinant of a lower triangular matrix with diagonal entries
equal to 1, has the value 1. Thus, we find
Pj (λ) = −λPj−1 (λ) + (−1)j ak−j
Using the inductive hypothesis we obtain
Pj (λ) = −λ(−1)j−1 (λj−1 + ak−1 λj−2 + ...ak−j+1 ) + (−1)j ak−j

= (−1)j (λj + ak−1 λj−1 + ... + ak−j )
Hence, by the principle of mathematical induction, the equality holds for

j = 1, ..., k. In particular,
Pk (λ) = (−1)k (λk + ak−1 λk−1 + ... + a0 )
and this completes the proof of the theorem.
Equation (1.74) can be written in the more compact form
Ly = −b
where L denotes the kth order linear difference operator
L = τ k + ak−1 τ k−1 + ... + a0
88
(cf. §1.3). Denoting by λ1 , ..., λk the (not necessarily distinct) solutions of
the characteristic equation, we can write the kth degree polynomial λk +
ak−1 λk−1 + ... + a0 as a product of linear factors:
λk + ak−1 λk−1 + ... + a0 = (λ − λ1 )...(λ − λk )
The difference operator L can be factorized accordingly:
L = (τ − λ1 )...(τ − λk )
Suppose that λ is a solution of the characteristic equation of (1.74), i.e.

an eigenvalue of the companion matix A, on the right-hand side of (1.75).
Let x = (x1 , ..., xk ) be an eigenvector of A with eigenvalue λ. Then the
components of x satisfy the following relations
x2 = λx1 , ..., xk = λxk−1
Consequently, the corresponding eigenspace is always 1-dimensional! From

Theorem 1.6.1 and formula (1.51) it can be seen that every solution of the ho-
mogeneous equation associated with (1.74), is a linear combination of entries
of J n . If J consists of r Jordan blocks Jk1 (λ1 ),...,Jkr (λr ), then the solution
space

of the homogeneous

equation is spanned by the sequences {λnj }∞ n=0 ,
n n ∞ n n ∞
{ 1 λj }n=0 ,...,{ kj −1 λj }n=0 , j = 1, ..., r (cf. §A.1 and §A.2). Using the fact

n
that k
is a linear combination of 1, n, ..., nk , we obtain the following result.
Theorem 1.6.2. Suppose that the characteristic equation of (1.74) has r

solutions λj , each with multiplicity kj , j ∈ {1, ..., r}. Then the solution
space of the homogeneous equation associated with (1.74) has a basis
formed by the sequences
{λnj }∞ n ∞
n=0 , {nλj }n=0 , ..., {n
kj −1 n ∞
λj }n=0 , j = 1, ..., r
Example 1.6.3. Consider the second order, homogeneous, linear difference

equation
y(n + 2) − 4y(n + 1) + 4y(n) = 0, n ∈ N
Its characteristic equation
λ2 − 4λ + 4 = (λ − 2)2 = 0
89
has a double root λ = 2. Thus, the solution space is spanned by the se-
quences: {2n }∞ n ∞
n=0 and {n2 }n=0 . The general solution is
y(n) = (c1 + c2 n)2n , n ∈ N, c1 , c2 ∈ C
We can write the equation in a more compact form, as follows
(τ − 2)2 y = 0
(Verify this).
Example 1.6.4. The third order, homogeneous, linear difference equation
2 1
y(n + 3) − y(n + 1) + y(n) = 0, n∈N
3 3
has the following characteristic equation
2 1 1
λ3 − λ + = (λ + 1)(λ2 − λ + ) = 0
3 3 3
√ √
Its roots are λ1 = −1, λ2 = 61 (3 + i 3) and λ3 = 16 (3 − i 3). The last
two roots √are complex conjugated. The tangent of arg λ2 is the quotient of
Im λ2 = 61 3 and Re λ2 = 16 .3. Hence
1 π
arg λ2 = arctan √ =
3 6
The absolute value of λ2 and λ3 is the square root of the sum of the squares
of the real and imaginary parts, i.e.
s s
9 3 1 1√
|λ2 | = |λ3 | = + = = 3
36 36 3 3
(In fact, there is a simpler proof of the above identity : from the coefficient
of y(n) in the equation it follows that λ1 λ2 λ3 = − 31 . As λ1 = −1 and
λ2 λ3 = λ2 λ2 = |λ2 |2 = |λ3 |2 , we have |λ2 |2 = |λ3 |2 = 31 .) Every solution of
the equation is a linear combination of the sequences λn1 , λn2 and λn3 , i.e. of
n π n π
(−1)n , 3− 2 ein 6 and 3− 2 e−in 6
90
If one is solely interested in real-valued solutions, the last two sequences can
be replaced by the real and imaginary parts of λn2 , i.e. by
n π n π
3− 2 cos n and 3− 2 sin n ,
6 6
respectively. Thus, every real-valued solution of the equation can be writ-
ten as follows
n π n π
y(n) = c1 (−1)n + c2 3− 2 cos n + c3 3− 2 sin n , n ∈ N, c1 , c2 , c3 ∈ R
6 6
1.6.2 Inhomogeneous higher order difference equations;

the annihilator method
From (1.75) and (1.60) we deduce that the inhomogeneous equation (1.74)
has a particular solution y with initial values y(0) = ... = y(k − 1) = 0,
such that, for n ≥ 1, y(n) equals the first component of the vector
 
0
.
 
n−1  
n−m−1 
X  
A  . 

m=0 
 0 

−b(m)
i.e.
n−1
(An−m−1 )1k b(m) for n ≥ 1
X
y(0) = 0, y(n) = −
m=0
It is not easy, in general, to extract very precise information about the long
term behaviour of y from the above formula and therefore its use in practical
applications is rather limited. In some special cases, a particular solution of
(1.74) can be found by the so-called annihilator method. This is a system-
atic procedure to find particular solutions to certain types of inhomogeneous
difference (or differential) equations, using the technique of “undetermined
coefficients”. It can be applied in the case that the function b is itself a
solution of a homogeneous linear difference equation with constant
coefficients, that is, whenever b can be written as a linear combination of
91
sequences of the form {nk λn }∞
n=0 with k ∈ N and λ ∈ C\{0} (cf. Theorem
1.6.2). We begin by writing (1.74) in the form
Ly = −b
where L represents the linear difference operator τ k + k−1 j P

j=0 aj τ . Suppose
there exists an lth order difference operator with constant coefficients
l−1
L0 = τ l + bh τ h
X
h=0
such that
L0 b = 0
(L0 annihilates b.) Then Ly = −b implies
L0 Ly = −L0 b = 0
L0 L is a linear difference operator with constant coefficients, of order k + l:

l−1 k−1 l−1 k−1
L0 L = τ l+k + bh τ h+k + aj τ j+l + bh aj τ h+j
X X X X
h=0 j=0 h=0 j=0
Every solution of the kth order inhomogeneous difference equation Ly =

−b is a solution of the homogeneous equation L0 Ly = 0, of order k + l, but
the converse is not true!!! The set of all solutions of the inhomogeneous
equation Ly = −b, which is a k-dimensional affine space, is a strict subset
of the k+l-dimensional (linear) solution space of the homogeneous equation
L0 Ly = 0. The general solution of the latter equation can be derived from its
characteristic equation. This general solution contains k + l parameters. k of
these parameters can be set equal to 0, whereas the remaining l parameters
(the undetermined coefficients) can be computed so as to yield a particular
solution of (1.74). We illustrate this method with two examples.
Example 1.6.5.
y(n + 2) − 4y(n + 1) + 4y(n) = 3n n ∈ N
The right-hand side is of the form λn , with λ = 3, so it satisfies the first

order homogeneous linear difference equation
y(n + 1) − 3y(n) = 0
92
or, equivalently,
L0 y := (τ − 3)y = 0
The left-hand side can be written as follows
Ly := (τ 2 − 4τ + 4)y = (τ − 2)2 y
Hence, every solution of the inhomogeneous equation also satisfies the (third
order) homogeneous equation
L0 Ly = (τ − 3)(τ − 2)2 y = 0
with characteristic equation
(λ − 3)(λ − 2)2 = 0
This implies that y must have the following form
y(n) = c1 2n + c2 n2n + c3 3n ,
where c1 , c2 , c3 ∈ C (or R). Note that c1 2n + c2 n2n is the general solution

of the homogeneous equation Ly = 0. In order to find a particular solution
of the inhomogeneous equation, we set c1 = c2 = 0 and compute c3 by the
method of undetermined coefficients. Inserting y(n) = c3 3n into the equation,
we get
c3 3n+2 − 4c3 3n+1 + 4c3 3n = 3n for all n ∈ N
and hence we deduce that c3 = 1. Thus, the general solution of the inhomo-
geneous equation is
y(n) = (c1 + c2 n)2n + 3n , n ∈ N, c1 , c2 ∈ C
Example 1.6.6.
y(n + 1) − y(n) = n2 n ∈ N
The right-hand side can be written as: n2 1n and thus satisfies a third order,
homogeneous linear difference equation, with characteristic equation (λ −
1)3 = 0, viz.
L0 y := (τ − 1)3 y ≡ 0
93
(cf. Theorem 1.6.2). Hence, every solution of the inhomogeneous equation
also satisfies the 4th order homogeneous equation
L0 Ly := (τ − 1)4 y ≡ 0
and thus has the form
y(n) = a + bn + cn2 + dn3
The general solution of the homogeneous equation y(n + 1) − y(n) = 0 is:

y(n) = a, a ∈ C. We set a = 0 and seek a particular solution of the
inhomogeneous equation of the form: y(n) = bn + cn2 + dn3 . Inserting this
expression into the equation, we find
b + (2n + 1)c + (3n2 + 3n + 1)d = n2 for all n ∈ N,
or, equivalently,
b + c + d + (2c + 3d)n + 3dn2 = n2 for all n ∈ N
Hence we deduce: d = 31 , c = − 12 and b = 61 .
1.6.3 Stability of solutions of higher order difference

equations
The initial value problem for a kth order, scalar equation:
y(n+k)+ak−1 y(n+k −1)+...+a0 y(n)+b(n) = 0, y(0) = c1 , ..., y(k −1) = ck
where c1 , ..., ck are given real or complex numbers, is equivalent to a corre-

sponding initial value problem for a first order system:
u(n + 1) = Au(n) + (0, ..., −b(n)), u(0) = (c1 , ..., ck )
Comparing definitions 1.1.8 and 1.4.5, we immediately see that stability of

a solution y of the kth order, scalar difference equation (1.74) is equivalent
to stability of the corresponding solution u of the first order system (1.75).
For, suppose that y and ỹ are solutions of (1.74), with the property that
|ỹ(j) − y(j)| ≤ δ for j = 0, ..., k − 1
94
Then there are two corresponding solutions u and ũ of (1.75), such that
u(0) = (y(1), ..., y(k − 1)) and ũ(0) = (ỹ(1), ..., ỹ(k − 1))
hence
k ũ(0) − u(0) k∞ ≤ δ
and vice versa. From Theorem 1.4.17 we deduce the following result.
Theorem 1.6.7. Suppose that the characteristic equation of (1.74) has r

solutions λj , each with multiplicity kj , j ∈ {1, ..., r}. The solutions of (1.74)
are asymptotically stable iff |λj | < 1 for all j. The solutions are stable if
|λj | ≤ 1 for all j and, in addition, kj = 1 for all values of j such that |λj | = 1.
In all other cases the solutions are unstable.
Example 1.6.8. Samuelson’s multiplier-accelerator model.

This model is described by the following equations.
Ct = cYt−1
It = a(Yt−1 − Yt−2 ) + A
Yt = Ct + It
Here, A denotes autonomous investments, assumed independent of the time t,
c is a number between 0 and 1 and the accelerator a is a positive constant. In
this model, the income Yt satisfies the second order, linear difference equation
Yt+2 − (c + a)Yt+1 + aYt = A, t ∈ N (1.76)
This equation has a unique equilibrium solution
A
Yt = , t∈N
1−c
We are interested in the stability of the equilibrium solution and in deviations
from this equilibrium, so in the behaviour of non-equilibrium solutions close
to the equilibrium. The characteristic equation of (1.76) is
λ2 − (c + a)λ + a = 0 (1.77)
95
We distinguish 3 cases.
1). The discriminant D of (1.77) is negative. This is the case when
(c + a)2 − 4a = c2 + 2ca + a2 − 4a < 0
or, equivalently,
(a + c − 2)2 < 4(1 − c)
i.e.
√ √ √ √
(1 − 1 − c)2 = 2 − c − 2 1 − c < a < 2 − c + 2 1 − c = (1 + 1 − c)2
Then the characteristic equation (1.77) has two distinct, complex conjugated
solutions λ+ and λ− , given by
1 q
λ± = (c + a ± i 4a − (c + a)2 )
2
Furthermore, we have
|λ± |2 = λ+ λ− = a
The general, real-valued, solution of (1.76) has the form
A t
Yt = + a 2 (c1 cos φt + c2 sin φt), t ∈ N, c1 , c2 ∈ R
1−c
where
c+a
φ = arg λ+ = arccos √
2 a
The non-equilibrium solutions thus exhibit fluctuations, whose magnitude
increases with increasing a. All solutions are stable if and only if |a| ≤ 1
and asymptotically stable iff |a| < 1. In these cases, the non-equilibrium
solutions exhibit oscillations, or damped oscillations, respectively, about the
equilibrium.
2). D = (c + a)2 − 4a = 0, so
√ √
a = (1 + 1 − c)2 or a = (1 − 1 − c)2
The characteristic equation (1.77) has the unique solution

1 √
λ = (c + a) = a
2
96
with multiplicity 2. The general (real-valued) solution of (1.76) has the form
A t
Yt = + a 2 (c1 + c2 t), c1 , c2 ∈ R
1−c
The solutions are asymptotically stable iff |a| < 1. If, on the other hand,
|a| ≥ 1, all solutions are unstable.
3). D > 0, i.e.
√ √
a < (1 − 1 − c)2 or a > (1 + 1 − c)2
Then the characteristic equation (1.77) has two positive solutions λ1 and
λ2 , with the property that
λ1 + λ2 = c + a and λ1 λ2 = a
Hence we deduce that
(1 − λ1 )(1 − λ2 ) = 1 − λ1 − λ2 + λ1 λ2 = 1 − c > 0
Consequently, either λ1 and λ2 are both less than 1, or they are both greater
than 1. In the first case, all solutions are asymptotically stable, whereas
in the second case they are all unstable. The general (real-valued) solution
has the form
A
Yt = + c1 λt1 + c2 λt2 , c1 , c2 ∈ R
1−c
We conclude this section with a classic example from probability theory (cf.
W. Feller: Introduction to Probability Theory and its Applications).
Example 1.6.9. A gambler plays a series of identical games. In each game,
with probability p he gains one euro and with probability q = 1 − p he looses
one. He stops playing as soon as his capital (in euro’s) has attained a certain
value K, or has vanished entirely. Suppose we want to know in what way
the probability of the latter event depends on the gambler’s starting capital.
We denote the starting capital by the variable n (the number of euro’s) and
the probability that the gambler looses all his money by y(n). At the end
of one game, with probability p his capital has increased to n + 1 and with
probability 1 − p it has decreased to n − 1. Clearly, the following relation
holds
y(n) = py(n + 1) + (1 − p)y(n − 1), n = 1, 2, .....
97
This is a homogeneous, linear, second order difference equation, which can
be rewritten in the form
1 1
y(n + 2) − y(n + 1) + ( − 1)y(n) = 0, n = 0, 1, ..... (1.78)
p p
In addition, we have: y(0) = 1 and y(K) = 0. The last two conditions are
called boundary conditions. A similar argument leads to the following
inhomogeneous difference equation for the expected number of games, z(n),
assuming a starting capital of n euro’s:
z(n) = pz(n + 1) + (1 − p)z(n − 1) + 1, n = 1, 2, .....
or, equivalently,
1 1 1
z(n + 2) − z(n + 1) + ( − 1)z(n) + = 0, n = 0, 1, ..... (1.79)
p p p
with boundary conditions: z(0) = z(K) = 0.
1.6.4 Exercises
1. Find a homogeneous, linear difference equation, with constant coefficients,
that is satisfied by the sequence y(n) = 2−n sin n π4 (n ∈ N).
2. Find a homogeneous, linear difference equation, such that the sequence
y(n) = 2−n (n ∈ N) is a neutrally stable solution.
3. Find a linear difference equation, such that the sequence y(n) = 2−n sin n π4
(n ∈ N) is an unstable solution.
4. Find a linear difference equation, with constant coefficients, such that the
sequence y(n) = 5n (n ∈ N) is an asymptotically stable solution.
5. Set up a second order, linear difference equation, with constant coeffi-
cients, for the sum S(n) of the first n terms of an arithmetic progression,
with common difference v. Let S(0) = 0 and determine S(n) as the solu-
tion of an initial value problem.
6. Set up a second order, homogeneous, linear difference equation, with con-
stant coefficients, for the sum S(n) of the first n terms of a geometric
progression, with common ratio r. Let S(0) = 0 and determine S(n) as
the solution of an initial value problem.
98
7. Set up a second order, linear difference equation, with constant coeffi-
cients, for the sum S(n) of the first n terms of an arithmetic-geometric
progression, with common ratio r and common difference v (cf. Example
1.3.13). Let S(0) = 0 and determine S(n) as the solution of an initial
value problem.
y(n + 2) − 2y(n + 1) + 2y(n) = 3, y(0) = 5, y(1) = 6
9. a. Determine the solution of the initial value problem
2y(n + 2) + y(n + 1) − y(n) = 0, y(0) = −2, y(1) = 2
b. Convert the above initial value problem into an initial value problem
for a system of first order difference equations and solve it.
10. Determine the general solution of the equation
y(n + 2) − 4y(n + 1) + 3y(n) = 2n2 , n∈N
11. Find a solution of the equation
9y(n) − 6y(n − 1) + y(n − 2) = 3−n , n≥2
12. a. Determine the general solution of the equation
y(n + 2) − y(n + 1) + y(n) = (−1)n , n∈N
b. Convert the above equation into a system of first order equations and
determine the general solution of this system. Compare the result to
that of part a.
y(n + 2) − 4y(n + 1) + 4y(n) = 2n , n∈N
y(n + 2) − 4y(n + 1) + 4y(n) = 3n2 , n∈N
99
2 1
y(n + 3) − y(n + 1) + y(n) = n + 2n , n∈N
3 3
16. a. Determine the general solution of the difference equation
4y(n + 2) − 4y(n + 1) + y(n) = n, n = 0, 1, ...
and examine the stability of the solutions.

b. Convert the equation into a first order vectorial difference equation and
determine, with the aid! of (a), the solution of the vectorial equation
1
with initial vector .
1
17. a. Determine all equilibrium solutions of (1.78) and (1.79) in example

1.6.9.
b. Determine the general solution of (1.78) and the particular solution
satisfying the boundary conditions: y(0) = 1 and y(K) = 0.
c. Find a particular solution of (1.79) for p 6= 1/2.
d. Find a particular solution of (1.79) for p = 1/2.
e. Determine the solution of (1.79) satisfying the boundary conditions:
z(0) = z(K) = 0.
18. Consider the multiplier-accelerator model, described by the following equa-
tions:
Ct = c1 Yt−1 + c2 Yt−2
It = a1 (Yt−1 − Yt−2 ) + a2 (Yt−2 − Yt−3 ) + A(1 + g)t
Yt = Ct + It
Here, c1 , c2 , a1 , a2 , A and g are positive numbers.
a. Show that the income Yt satisfies a third order difference equation.
b. Under what condition does this equation have a solution of the form
Yt = Y0 (1 + g)t ?
c. Under what condition is the solution mentioned in b. relevant from an
economic point of view?
d. Prove that all solutions are unstable when a2 > 1.
100
Chapter 2
Differential equations
An ordinary differential equation, to be denoted by ODE, is an equation of

the form
F (t, y(t), y (1) (t), ..., y (k) (t)) = 0 (2.1)
Here, k is a positive integer: the order of the ODE, t a real- or complex-
valued, independent variable, F is a function of k + 2 variables, defined on
a subset D of Rk+2 or Ck+2 . y is an unknown function of t (the dependent
variable) and y (j) denotes the jth order derivative of y with respect to t (we
usually write y 0 and y 00 instead of y (1) and y (2) , other authors use ẏ and ÿ,
respectively). A given function y, defined on a subset T of R or C, is a
solution of (2.1) if, for every t ∈ T , we have (t, y(t), y (1) (t), ..., y (k) (t)) ∈ D
and F (t, y(t), y (1) (t), ..., y (k) (t)) = 0.
In many applications, the independent variable t is the time. From now on,
t will be a real variable and T a (finite or infinite) interval of R. Often, the
t-dependence of y is omitted and (2.1) is written in the form
F (t, y, y (1) , ..., y (k) ) = 0
The ODE is called autonomous when F doesn’t depend explicitly on t, so
that the equation can be written in the form
F (y(t), y (1) (t), ..., y (k) (t)) = 0 or F (y, y (1) , ..., y (k) ) = 0
F can be a scalar or a complex-valued function. In this section we will only
consider the first case.
Example 2.0.10. The logistic differential equation
y 0 = y(a − by)
101
is an autonomous ODE, as opposed to, for instance, the equation
y0 = t − y2
Both equations can be written in the form (2.1). In the first case, F (t, y, y 0 ) =
y 0 − ay + by 2 , in the second one F (t, y, y 0 ) = y 0 − t + y 2 . Both equations are
examples of non-linear, first order ODE.
The ODE is called linear when F is an affine function of y, y (1) ,...,y (k) .
So a linear ODE has the form
b(t) + a0 (t)y(t) + a1 (t)y (1) (t) + ... + ak (t)y (k) (t) = 0
2.1 First order scalar differential equations

(supplement to S&B §24.2)
In a first order differential equation only the unknown, say y, itself and its
first derivative y 0 can occur. In this section, we discuss a number of standard
methods to solve simple, first order differential equations.
Let f, g and G be continuous functions. We consider the following 4 types
of differential equations:
1. y 0 = g
2. y 0 = f G(y) (separable differential equation)
3. y 0 = f y (homogeneous linear differential equation)
4. y 0 = f y + g (inhomogeneous linear differential equation if g 6≡ 0)
Obviously, type 1 is a special case of type 4 and type 3 is both a special

case of type 2, with G(y) = y, and of type 4, with g ≡ 0.
Example 2.1.1. A neoclassical growth model in continuous time:
Y (t) = Q(K(t), L(t))
K 0 (t) = sY (t)
L0 (t) = nL(t)
102
Y , K and L denote production, capital and labour, respectively, As in the
discrete case, n > 0, 0 < s < 1 and the production function Q is assumed
homogeneous of degree 1. Again defining k := K
L
, (capital per worker), and
Q(K, L) K
q(k) = = Q( , 1) = Q(k, 1)
L L
we find that k must satisfy the first order separable differential equation (type
2):
k 0 (t) = sq(k(t)) − nk(t) (2.2)
Here, we can choose, for example, f ≡ 1 and G(k) = sq(k) − nk.
2.1.1 Solving equations of type 1

A solution of the differential equation of type 1 :
y 0 (t) = g(t), t ∈ T (2.3)

where T is an interval of R, is a continuously differentiable function y with
the property that y 0 (t) = g(t) for all t ∈ T (if T is not an open interval, y
should have a continuous right- or left-derivative at the end-points). Let y
be a solution of (2.3) and t0 ∈ T . Then
Z t Z t
0
y(t) = y (s) ds + y(t0 ) = g(s) ds + y(t0 )
t0 t0
Conversely, for any t0 ∈ T and C ∈ R or C,

Z t
y(t) = g(s) ds + C
t0
is a solution of (2.3). Solving an equation of type 1 is equivalent to finding

a primitive function
R
of the right-hand side. In other words, a solution is
given by y(t) = g(t) dt, and if no initial condition is given, the solution is
determined up to a additive constant.
Remark 2.1.2. If we drop the assumption that g is continuous, but allow

it to be discontinuous at a finite number of points, then the above solution
is only valid on the intervals where g is continuous. On each interval the
additive constant may have a different value!
103
Example 2.1.3. The general solution of the differential equation
y 0 (t) = 4e2t
is Z
y(t) = 4e2t dt = 2e2t + C.
Here, C is an arbitrary (real- or complex-valued) constant. If we require that

y(0) = 1, then substituting t = 0 in the general solution gives 1 = y(0) =
2 + C, so C = −1 and the solution to the initial value problem is given by
y(t) = 2e2t − 1. If y(0) = y0 , then the solution is y(t) = 2e2t + y0 − 2, since
y0 = y(0) = 2 + C.
2
Example 2.1.4. The differential equation y 0 (t) = 6tet has the general so-
2 2
lution y(t) = 6tet dt = 3et + C, C constant. The unique solution with the
R
2
property that y(1) = 7 is given by y(t) = 3et + 7 − 3e.
Example 2.1.5. The general solution of the differential equation
y 0 (t) = (t − 1)−1 + t−2 , t ∈ [2, ∞)
is
y(t) = log(t − 1) − t−1 + C, C constant.
The solution with initial value y(2) = 1.5 is y(t) = log(t − 1) − t−1 + 2.
2.1.2 Separable equations (type 2)

The notation
y 0 = f G(y) (2.4)
is shorthand for: y 0 (t) = f (t)G(y(t)) for all t in the domain of y. We assume
that G(y) 6= 0 for all y and define p(y) = 1/(G(y)). Multiplying both sides
with p(y) and integrating them we get
Z Z
0
p(y(t))y (t) dt = f (t) dt.
Thus, if P is a primitive of p, F a primitive of f and y a solution of (2.4),

then we have P (y(t)) = F (t) + C, where C ∈ R or C. Conversely, any
continuously differentiable function y satisfying P (y(t)) = F (t) + C for some
104
(real- or complex-valued) constant C is a solution of (2.4). This can be
verified by differentiating both sides of the equation P (y(t)) = F (t) + C. C
can be determined from the initial condition: C = P (y(t0 )) − F (t0 ). This
solution method is called the method of separation of variables.
Remark 2.1.6. Here we have assumed that G(y) 6= 0, but, more generally,
the solution is valid for all t for which y(t) is not a zero of G. If G(k) = 0
then the constant function y ≡ k is a solution of (2.4) and, conversely, if
y ≡ k is a solution of (2.4), then G(k) = 0 (check this!).
Example 2.1.7. Consider the initial value problem

y 0 (t) = 8t(y(t))2 , y(0) = 1
Here, we can choose f (t) = 8t and G(y) = y 2 . Note that y ≡ 0 satisfies the
equation, but not the initial condition. We begin by applying the method
of
R 0
separation of variables, without worrying about the zeroes of G. We get
y (t)/y(t) dt = 8t dt, from which we derive that −1/y(t) = 4t2 + C, hence
2
R
y(t) = −1/(4t2 + C). The condition y(0) = 1 then yields C = −1. As the
function y : [0, 1/2) −→ R defined by y(t) = −1/(4t2 − 1) is C 1 and different
from 0 for all t ∈ [0, 1/2), G(y(t)) is never zero and y is the solution of the
initial value problem.

2y(t)y 0 (t) = 4t, y(0) = 3
the method of separation of variables, we get 2y(t)y 0 (t) dt =
R
Applying
2 2
R
4t dt hence
√ y(t) = 2t + C. The initial condition implies that C√= 9, so
y(t) = ± 2t + 9. Obviously, y : [0, ∞) −→ R defined by y(t) = 2t2 + 9
2
is the (unique) solution to the initial value problem.
Example 2.1.9. Consider the initial value problem for the logistic differen-
tial equation
y 0 (t) = a(1 − y(t)/k)y(t), y(0) = y0 , (2.5)
where the parameters a and k are positive and y0 ∈ R. Check that the
constant functions y ≡ 0 and y ≡ k satisfy the equation. Applying the
method of separation of variables, we get
Z
y0 Z
dt = a dt.
(1 − y(t)/k)y(t)
105
From example 24.11 in Simon and Blume (pp 644-645) we conclude that
y(t) = ky0 /(y0 + (k − y0 )e−at ).
Note that this expression also represents the solution of the initial value
problem when y0 = 0 or y0 = k, even though 0 and k are zeroes of “G” (the
right-hand side of (2.5)). So it represents the solution to the initial value
problem for all values of y0 . (Determine the domain of the solution y with
initial value y0 < 0.)
2.1.3 Homogeneous linear equations (type 3)

Next, we consider the homogeneous linear differential equation
y0 = f y (2.6)
or, equivalently, y 0 (t) = f (t)y(t). As this is a special case of type 2, we
can apply the method of separation of variables, providedR y(t) 6= 0 for all t.
This yields y 0 (t)/y(t) dt = f (t)dt, hence log |(y(t)| = f (t)dt. If F is a
R R
primitive function of f we get

R
f (t)dt
|(y(t)| = e = eF (t)+C ,
where C is an arbitrary real-valued constant. Thus, if y is a real-valued
solution, we have y(t) = DeF (t) , with D = eC = y(0) if y(0) > 0 and
D = −eC = y(0) if y(0) < 0. Note that D = 0 corresponds to y ≡ 0: the
unique solution with the property that y(0) = 0.
y 0 = 3y, y(0) = 6
Separation of variables gives y 0 (t)/y(t) dt = 3 dt, hence log |y(t)| = 3t+C,
R R
and y(t) = ±eC e3t = De3t . We determine D from 6 = y(0) = D, so the

solution of the initial value problem is y(t) = 6e3t .
Example 2.1.11. More generally, consider the initial value problem
y 0 = ky, y(0) = a,
where a and k are positive R 0
numbers. Using the same method as in the
previous example, we get y (t)/y(t) dt = k dt, hence log |y(t)| = kt + C,
R
and y(t) = ±eC ekt = Dekt . The initial condition then implies that D = a so
the solution is y(t) = aekt .
106
Example 2.1.12. Determine the general real-valued solution of
y 0 (t) = (At + B)y(t)
If y(t) 6= 0 for all t, separation of variables gives y 0 (t)/y(t) dt = (At+B) dt,

R R
1 2 1 2
so log |y(t)| = 21 At2 +Bt+C, and y(t) = ±eC e 2 At +Bt = De 2 At +Bt . A priori,
D can have any value except 0 (D 6= 0 implies y(t) 6= 0 for all t). However,
D = 0 corresponds to y ≡ 0 which obviously satisfies the equation as well.
1 2
Hence the general real-valued solution is given by: y(t) = Ce 2 At +Bt , C ∈ R.
2.1.4 Inhomogeneous linear equations (type 4)

Finally, we consider the equation:
y0 = f y + g (2.7)
or y 0 (t) = f (t)y(t) + g(t) for all t in the domain of y. We assume that g 6≡ 0.

Similarly to the case of inhomogeneous linear difference equations, in order
to solve (2.7), all we need is one particular solution of (2.7) and the general
solution of the homogeneous equation (2.6).
Theorem 2.1.13. Let y0 be a particular solution of the inhomogeneous lin-

ear differential equation (2.7).
(i) Every solution of (2.7) can be written as the sum of y0 and a solution of
the homogeneous equation (2.6).
(ii) Conversely, any function that can be written as the sum of y0 and a
solution of the homogeneous equation is a solution of (2.7).
Proof. (i) Suppose that y1 is another solution of (2.7), then y10 (t) − y00 (t) =
f (t)y1 (t) + g(t) − (f (t)y0 (t) + g(t)) = f (t)(y1 (t) − y0 (t)). So the function
z := y1 − y0 solves the associated homogeneous differential equation z 0 = f z
and y1 = y0 + z.
(ii) Conversely, if z is a solution of (2.6) and y1 = y0 + z, then y10 (t) =
y00 (t) + z 0 (t) = f (t)y0 (t) + g(t) + f (t)z(t) = f (t)y1 (t) + g(t).
In simple cases, a particular solution can be found with some guess-work, or

by means of an annihilator-method, similar to the one discussed in §1.6.2.
107
y 0 (t) = −4ty(t) + 6t, y(0) = 3
It is not difficult to see that this equation has the constant solution y0 ≡ 32 :
y00 (t) = 0 = −4ty(t) + 6t = −4t 23 + 6t = 0. According to example 2.1.12,
2
the general solution of the homogeneous equation is yh (t) = Ce−2t , so the
2
general solution of the inhomogeneous equation is y(t) = 32 + Ce−2t . The
initial condition implies that C = 32 , hence the solution of the initial value
2
problem is y(t) = 32 (1 + e−2t ).
y 0 = 2ty(t) + (3 − 2t2 )t2 , y(0) = 4
This equation has a solution y0 (t) = t3 : y00 (t) = 3t2 = 2t4 + 3t2 − 2t4 . Ac-
cording to example 2.1.12, the general solution of the homogeneous equation
2
is yh (t) = Cet , so the general solution of the inhomogeneous equation is
2
y(t) = t3 + Cet and from the initial condition we deduce that C = 4, hence
2
the solution of the initial value problem is y(t) = t3 + 4et .
Variation of constants.
A very general method for solving (2.7) is by means of “variation of con-
stants”, similarly to the case of a first order, inhomogeneous linear difference
equation (cf. §1.3.2). Note, however, that it doens’t necessarily lead to a
simple expression for the solutions.
Let F be a primitive of f . Then the general solution of the homogeneous
linear differential equation y 0 = f y, is y(t) = CeF (t) , where C is a (real- or
complex-valued) constant. We try to find a solution of the inhomogeneous
equation (2.7) by allowing C to vary, i.e. to depend on t. Substituting
y(t) = C(t)eF (t) into (2.7) we get a differential equation for C:
y 0 (t) = C 0 (t)eF (t) + C(t)eF (t) F 0 (t) = f (t)C(t)eF (t) + g(t)
As F 0 (t) = f (t), this is equivalent to
C 0 (t)eF (t) = g(t) ⇔ C 0 (t) = e−F (t) g(t)
which has the general solution C(t) = e−F (t) g(t) dt. Hence the general
R
solution of (2.7) can be written in the form

Z
y(t) = e F (t)
e−F (τ ) g(τ )dτ
108
Moreover, the solution of (2.7) with initial value y(t0 ) = y0 can be represented
as follows Z t
y(t) = eF (t) {e−F (t0 ) y0 + e−F (τ ) g(τ )dτ } (2.8)
t0
Remark 2.1.16. In the above expressions eF (t) can be replaced by any non-
vanishing solution of the homogeneous equation (2.6). (Verify this.)
y 0 (t) = −3y(t) + 8t, y(0) = 2
We begin by solving the associated homogeneous equation y 0 (t) = −3y(t).

As we have seen in example 2.1.11 the general solution is ae−3t . By (2.8),
the solution to the initial value problem is given by the expression
Z t
y(t) = e−3t (2 + e3τ 8τ dτ )
0
Using partial integration, we get 0t e3τ 8τ dτ = 38 te3t − 89 e3t + 89 , so y(t) =

R
8
3
t − 89 + 26
9
e−3t solves the initial value problem.
y 0 (t) = 2Aty(t) + 4Bt, y(0) = C,
for given nonzero real numbers A, B and C.

As we have seen in example 2.1.12, the general solution of the associated
2
homogeneous differential equation y 0 (t) = 2Aty(t) is aeAt . By (2.8), the
solution to the initial value problem is given by the expression
Z t
2 2
y(t) = eAt (C + e−Aτ 4Bτ dτ )
0
2 2
Using the substitution method, we find 0t e−Aτ 4Bτ dτ = −2 B (e−At − 1).
R
A
2
So y(t) = −2 B A
+ (C + 2 B
A
)eAt solves the initial value problem. (In fact, it
is easily seen that the equation admits the constant solution y ≡ −2 B A
, cf.
example 2.1.14.)
109
2.1.5 Exercises
1. Determine the solutions of the following differential equations:
a. y 0 (t) = 3y(t), t ∈ [0, ∞), with initial condition y(0) = 5.

b. y 0 (t) = −8ty(t)2 , t ∈ [1, ∞), with initial condition y(1) = 1.
c. y 0 (t) = −2ty(t), t ∈ [0, ∞), such that y(1) = 3.
d. y 0 (t) = −2ty(t) + 6t, t ∈ [0, ∞), with initial condition y(0) = 4.
2. Determine the general solution of the following differential equations:
a. y 0 (t) = −4ty(t), t ∈ [0, ∞).

b. y 0 (t) = −4y(t) + 10et , t ∈ [0, ∞).
3. Consider a lake with a volume of 9.108 m3 of water, containing an amount

of chemical waste. The waste is distributed evenly at every moment.
Let C(t) denote the concentration of chemical waste at time t, where
0 ≤ C(t) ≤ 1 for all t. The waste enters the lake by a polluted river P; the
concentration of the chemical waste in the river P has the constant value
0.2. At time t = 0 the concentration in the lake, C(0), equals 0.05. Every
second 180m3 of polluted water enters the lake through river P, and at
the same time 180m3 leaves the lake through another river, N.
a. Explain that the above process can be modelled by the following initial
value problem:
C 0 (t) = −2.10−7 C(t) + 4.10−8 , C(0) = 0.05
b. Solve the initial value problem.

c. Determine limt→∞ C(t) from the solution found in b, and explain the
result.
d. At a certain moment a new inlet into the lake is realized. From that
moment on every second 90m3 clean water is pumped into the lake.
The outlet N is widened and every second it removes 270m3 water
from the lake.
e. Write down the differential equation describing the change in concen-
tration in the new situation.
110
f. What will be the concentration of the chemical waste in the lake on
the long run? (Underpin your answer)
4. Consider the neoclassical growth model in Example 2.1.1 and let Q be a

Cobb-Douglas function of the form Q(K, L) = K α L1−α , with 0 < α < 1.
K
a. Derive a first order differential equation for the variable k := L
.
b. Show that this equation has a nonvanishing equilibrium solution k ≡ k̄.
c. Use separation of variables to derive the following expression for the
general solution of this equation
s 1
k(t) = ( + Cen(α−1)t ) 1−α
n
where C is an “arbitrary” constant (explain the quotation marks).
2.2 Systems of first order differential equa-

tions
Every scalar, kth order ODE can be converted into a system of k first order
ODE’s, by means of a transformation resembling the one we discussed in the
case of a kth order difference equation. Consider the general ODE
F (t, y, y (1) , ..., y (k) ) = 0 (2.9)
We introduce new dependent variables y1 through yk , defined by
y1 := y, y2 := y (1) , ...., yk := y (k−1) (2.10)
Now, (2.9) is equivalent to the following system of first order ODE’s:
y10 (t) = y2 (t),

............................ . .....
............................ . .....
0
yk−1 (t) = yk (t),
0
F (t, y1 (t), y2 (t), ...yk (t), yk (t)) = 0
111
Example 2.2.1. Consider the second order, nonlinear ODE
y 00 (t) + y(t) − y(t)3 = 0 (2.11)
Define new dependent variables y1 and y2 , by
y1 := y, y2 := y 0
If y is a solution of (2.11), then y1 and y2 satisfy the following system of first

order, linear ODE’s
y10 (t) = y 0 (t) = y2 (t)

y20 (t) = y 00 (t) = y1 (t)3 − y1 (t)
Conversely, if two functions y1 and y2 satisfy the system of ODE’s
y10 (t) = y2 (t)

y20 (t) = y1 (t)3 − y1 (t)
then y1 is a solution of the second order equation (2.11) en y2 (obviously!) is

the first order derivative of y1 .
Example 2.2.2. Consider the second order, inhomogeneous, linear ODE

√
y 00 (t) + 3ty 0 (t) − ty(t) = et , t ∈ [0, ∞) (2.12)
Defining y1 and y2 , by
y1 := y, y2 := y 0
we obtain the following system of first order, linear ODE’s
y10 (t) = y 0 (t) = y√

2 (t)
y20 (t) = y 00 (t) = ty1 (t) − 3ty2 (t) + et
which is equivalent to (2.12) in the same sense as in the previous example.
Example 2.2.3. Now consider the general case of a kth order linear ODE:
ak (t)y (k) (t) + ak−1 (t)y (k−1) (t) + ... + a0 (t)y(t) = b(t), t ∈ T, (2.13)
112
where it is assumed that ak (t) 6= 0 for all t ∈ T . Introducing new dependent
variables y1 through yk , as defined in (2.10), we obtain the following system
of first order, linear ODE’s
y10 (t) = y2 (t)

y20 (t) = y3 (t)
...................
...................
ak−1 (t)
yk0 (t) = − aak0 (t) y (t)... −
(t) 1 ak (t) k
y (t) + b(t)
ak (t)
This system of first order differential equations can also be written in the
form of a first order, linear, vectorial ODE:
   
0 1 0 . . 0 0
  
y1 (t) y1 (t)
0 0 1 . . 0 0
    

 y2 (t) 
 


 y2 (t)   
  
d

 .

  . . . . . .  .   . 
= +
    
dt 
 . 


 . . . . . . 
 .  
  . 

. . . . . .
    

 .

 


 .  
 . 

yk (t) − aak0 (t)
(t)
. . . . − ak−1 (t)
ak (t)
yk (t) b(t)
ak (t)
(2.14)
If the function y is a solution of the kth order, linear ODE (2.13), then the
vector function (y1 , ..., yk ), defined by (2.10), is a solution of the vectorial
ODE (2.14) and, conversely, if the vector function (y1 , ..., yk ) satisfies (2.14),
(j−1)
then y1 is a solution of the kth order ODE (2.13) and yj = y1 for j =
2, ..., k.
From now on, we consider systems of k first order ODE’s, of the form
y10 (t) = f1 (t, y1 (t), ..., yk (t))

y20 (t) = f2 (t, y1 (t), ..., yk (t))
.......................... , t ∈ [t0 , ∞)
..........................
yk0 (t) = fk (t, y1 (t), ..., yk (t))
where f1 , ...fk are (real- or complex-valued) functions of k + 1 variables,

defined on a set of the form [t0 , ∞) × D, with D ⊂ Rk or D ⊂ Ck . This
system is equivalent to the vectorial ODE:
y 0 (t) = f (t, y(t)), t ∈ [t0 , ∞) (2.15)
113
Theorem 2.2.4 (Existence and uniqueness of solutions). Let D ⊂ Ck ,
y0 ∈ D and f : [t0 , ∞)×D → Ck be continuous on a neighbourhood of (t0 , y0 ).
Then there exists a positive number δ and a C 1 -function y : (t0 − δ, t0 + δ) →
Ck such that y 0 (t) = f (t, y(t)) and y(t0 ) = y0 . Moreover, if f is C 1 on a
neighbourhood of (t0 , y0 ), then this solution is unique.
2.2.1 Systems of first order linear differential equa-

tions
The system of k first order, linear ODE’s
y10 (t) = a11 (t)y1 (t) + a12 (t)y2 (t) + ... + a1k (t)yk (t) + b1 (t)
y20 (t) = a21 (t)y1 (t) + a22 (t)y2 (t) + ... + a2k (t)yk (t) + b2 (t)
......................................................................... (2.16)
.........................................................................
yk0 (t) = ak1 (t)y1 (t) + ak2 (t)y2 (t) + ... + akk (t)yk (t) + bk (t)
can also be written in vectorial form:
      
y1 (t) a11 (t) a12 (t) . . . a1k (t) y1 (t) b1 (t)

 y2 (t) 


 a21 (t) a22 (t) . . . a2k (t) 
 y2 (t)  
  b2 (t)] 

d . . . . . . . . .
      
      
  =   + 
dt 
 . 


 . . . . . . 
 .  
  . 

. . . . . . . . .
      
      
yk (t) ak1 (t) ak2 (t) . . . akk (t) yk (t) bk (t)
or, briefly,
y 0 (t) = A(t)y(t) + b(t)
Here, y and b denote k-dimensional vector functions and A is a k × k matrix
function. In what follows we will not discriminate between systems of ODE’s
and vectorial ODE’s. (2.16) is called a system of coupled ODE’s, if there
exists at least one pair (i, j), with i 6= j, such that the function aij does
not vanish identically. In that case, the behaviour of yi depends on that of
yj . When the ODE’s are not coupled, the corresponding matrix A(t) is a
diagonal matrix for every value of t. The vectorial ODE
y 0 (t) = A(t)y(t) + b(t), t ∈ T
is called homogeneous iff b ≡ 0 (i.e. b(t) = 0 for all t ∈ T ). A homogeneous,
linear, vectorial ODE has the solution y ≡ 0, and every linear vectorial
114
ODE that is satisfied by y ≡ 0 is homogeneous (check this). Thus, an
inhomogeneous, linear vectorial ODE never has a null solution. We will
restrict ourselves to the very simple case that the coefficients of the system,
or, equivalently, the entries of the matrix function A, do not depend on time,
thus are constants.
2.2.2 Homogeneous linear vectorial differential equa-

tions with constant coefficients
We begin by discussing the homogeneous case, i.e. the equation
y 0 (t) = Ay(t), t ∈ T
where A ∈ Ck×k .
Example 2.2.5. The first order ODE
y 0 (t) = ay(t), t ∈ [t0 , ∞)
where a ∈ C and t0 ∈ R, has the general solution
{y(t) = ceat , c ∈ C}
(or {y(t) = ceat , c ∈ R} if a ∈ R and we only consider real-valued solutions.)

In order to find the solution of the initial value problem
y 0 (t) = ay(t), y(t0 ) = y0
where y0 is a given real (or complex) number, we have to choose the parameter
c such that ceat0 = y0 . Hence, this solution is
y(t) = ea(t−t0 ) y0
Example 2.2.6. If A is a k × k diagonal matrix, with diagonal entries

λ1 , ..., λk , then the vectorial ODE
y 0 (t) = Ay(t)
115
is equivalent to a system of uncoupled, scalar ODE’s in the coordinates
y1 , ..., yk :
yi0 (t) = λi yi (t), i = 1, ..., k
The general solution is given by
yi (t) = ci eλi t , i = 1, ..., k
where c1 , ..., ck are arbitrary, real or complex numbers. In vectorial form:


c1 eλ1 t
  
c1
c2 eλ2 t c2
   
   
diag{eλ1 t , ..., eλk t } 
   
y(t) = 
 . 
 =  . 


 . 


 . 

ck eλk t ck
Since, in this particular case,
etA = diag{eλ1 t , ..., eλk t }
(cf. §A.3, (45) for the definition of etA ), the general solution can be written
as follows
{y(t) = etA c, c ∈ Ck } (2.17)
We will prove that the general solution of the vectorial ODE y 0 (t) = Ay(t) is
represented by (2.17), even if A is not a diagonal matrix. For this purpose,
we need the following lemma.
Lemma 2.2.7. For every k-dimensional vector c ∈ Ck , the vector function

y defined by: y(t) = etA c is a solution of the vectorial ODE
y 0 (t) = Ay(t) (2.18)
From the above lemma it is easily deduced that

d tA
e = AetA
dt
(Verify this.) Thus, the matrix function etA satisfies the matrix ODE
Y 0 (t) = AY (t)
116
d tA −tA
(e e ) = AetA e−tA + etA .(−A)e−tA
dt
As, for every value of t, the matrix etA commutes with A, the right-hand side
vanishes identically, so we have
d tA −tA
(e e ) = 0
dt
This implies that the matrix function etA e−tA does not depend on t, hence
etA e−tA = e0A e−0A = I
Thus we have shown that, for every value of t, the matrix etA is invertible,
with inverse e−tA . This property is a special case of Theorem A.3.1 in the
appendix.
Theorem 2.2.8. (i) The general solution of the vectorial ODE (2.18) has
the form
{y(t) = etA c, c ∈ Ck } (2.19)
(ii) The solution of the initial value problem
y 0 (t) = Ay(t), y(t0 ) = y0
where y0 ∈ Ck , is
y(t) = e(t−t0 )A y0
Proof. (i) According to lemma 2.2.7, every vector function of the form (2.19)
is a solution of equation (2.18). It remains to be proved that all solutions are
of this form. So suppose that y is a k-dimensional vector function, satisfying
(2.18). We have to show the existence of a vector c ∈ Ck , such that y = etA c.
Noting that
d −tA
(e y(t)) = −Ae−tA y(t) + e−tA y 0 (t) = −Ae−tA y(t) + e−tA Ay(t) = 0
dt
we conclude that the vector function e−tA y(t) is a constant (vector), which
we denote by c. Then we have
etA e−tA y(t) = y(t) = etA c
117
(ii) In order to find the solution of the initial value problem, we need to
choose the vector c in such a manner that
y(t0 ) = et0 A c = y0
Left multiplication with the matrix e−t0 A yields
e−t0 A y(t0 ) = c = e−t0 A y0
Inserting this into the general expression for y, we find
y(t) = etA e−t0 A y0 = e(t−t0 )A y0
(The last equality is due to the fact that the matrices tA and t0 A commute,
cf. Theorem A.3.1.)
Theorem 2.2.9. The solutions of the k-dimensional vectorial ODE (2.18)

form a k-dimensional linear space. The vector functions yi , i = 1, ..., k,
defined by
yi (t) = etA ci , ci ∈ Ck
constitute a basis of the solution space if and only if the vectors c1 , ..., ck
constitute a basis of Ck .
Proof. etA is a nonsingular matrix for every value of t. Hence, the linear
Pk tA Pk
combination of vector functions: i=1 αi yi = e i=1 αi ci , with αi ∈ C,
vanishes identically (i.e. equals the null vector for all values of t) if and only
if
k
X
αi ci = 0
i=1
i.e. precisely then, when the vectors c1 , ..., ck are linearly dependent. Fur-
thermore, according to the previous theorem, every solution y of (2.18) has
the form y = etA c, with c ∈ Ck . If the vectors c1 , ..., ck form a basis of Ck ,
then c is a linear combination of c1 , ..., ck and thus y is a linear combination
of the vector functions etA c1 , ..., etA ck . Hence the statements of the theorem
follow easily.
A basis of the solution space of (2.18) can be obtained by choosing for

c1 , ..., ck the standard basis of Ck , i.e.
yi (t) = etA ei , i = 1, ..., k
118
The solution yi is the ith column vector of the matrix function etA . If A
has k linearly independent eigenvectors s1 , ..., sk , with eigenvalues λ1 , ..., λk ,
respectively, then also the vector functions ỹi (t) := etA si , i = 1, ..., k, span
the solution space. Since, for every value of t, si is an eigenvector of the
matrix etA , with eigenvalue etλi , equation (2.18) in this case has a second
basis of solutions, of the form
ỹi (t) = eλi t si , i = 1, ..., k
The advantage of this basis over the previous one is, that the t-dependence
of the individual basis elements is immediately clear, without the need to
compute the entries of etA . The first basis is sometimes convenient in solving
initial value problems.
Example 2.2.10. The system of first order, homogeneous ODE’s
y10 (t) = 4y1 (t) + y2 (t)
y20 (t) = −2y1 (t) + y2 (t)

can be written in the form of the 2-dimensional, vectorial ODE:
!
dy 4 1
= y (2.20)
dt −2 1
The matrix A in this case has an eigenvector (1, −2), with eigenvalue 2, and
an eigenvector (1, −1), with eigenvalue 3. The vector functions e2t (1, −2)
and e3t (1, −1) span the solution space of (2.20) and the general solution is
! ! ! ! !
2t 1 3t 1 e2t e3t c1 c1
y(t) = c1 e +c2 e = , ∈ C2
−2 −1 −2e2t −e3t c2 c2
So, y1 (t) = c1 e2t + c2 e3t and y2 (t) = −2c1 e2t − c2 e3t .

In order to solve the corresponding initial value problem, with initial values:
y1 (0) = 1 and y2 (0) = 0, i.e. with initial vector y(0) = (1, 0), we let t = 0
in the above expressions for y1 and y2 . This yields the following system of
equations for the parameters c1 and c2 :
c1 + c2 = 1
−2c1 − c2 = 0
119
which has the solution c1 = −1, c2 = 2. Thus, the solution of the initial
value problem, in vectorial form, is:
! !
2t 1 3t 1
y(t) = −e + 2e
−2 −1
We can summarize the above procedure for solving the initial value problem
as follows. First, write the initial vector (1, 0) as a linear combination of the
eigenvectors (1, −2) and (1, −1) of A (and thus of etA ):
! ! !
1 1 1
= −1 +2
0 −2 −1
Second, multiply each vector on the right-hand side with the corresponding
eigenvalue of etA .
A slightly different approach is by using the second part of Theorem 2.2.8:
!
tA 1
y(t) = e
0
The matrix function etA can be computed with the aid of the eigenvectors
and eigenvalues found above (cf. §A.3):
! ! !−1 !
tA 1 1 e2t 0 1 1 −e2t + 2e3t −e2t + e3t
e = =
−2 −1 0 e3t −2 −1 2e2t − 2e3t 2e2t − e3t
Hence we deduce that

! ! !
−e2t + 2e3t −e2t + e3t 1 −e2t + 2e3t
y(t) = =
2e2t − 2e3t 2e2t − e3t 0 2e2t − 2e3t
Any matrix function, whose column vectors form a basis of the solution space
of (2.18), is called a fundamental matrix of (2.18). Every fundamental
matrix is itself a (matrix) solution of the equation. For, if Y is a fundamental
matrix of (2.18), with column vectors y1 , ..., yk , then
d
Y (t) = (y10 (t)...yk0 (t)) = (Ay1 (t)...Ayk (t)) = A(y1 (t)...yk (t)) = AY (t)
dt
120
One example of a fundamental matrix of (2.18) is the matrix function with
column vectors yi (t) = etA ei , i = 1, ..., k, i.e. the matrix function etA . If the
vectors c1 , ..., ck form a basis of Ck , then the matrix function Y defined by
Y (t) = etA C
where C denotes the matrix with column vectors c1 , ..., ck , is a fundamental

matrix of (2.18). Obviously, C is a nonsingular k × k matrix. Conversely,
every matrix function of the form etA C, where C is nonsingular, is a fun-
damental matrix of (2.18). The equation in the preceding example has a
fundamental matrix
!
tA −e2t + 2e3t −e2t + e3t
Y1 (t) := e =
2e2t − 2e3t 2e2t − e3t
A second fundamental matrix is

! ! ! !
tA 1 1 1 1 e2t 0 e2t e3t
Y2 (t) := e = =
−2 −1 −2 −1 0 e3t −2e2t −e3t
If Y is a fundamental matrix of (2.18), with column vectors y1 , ..., yk , then

every solution y of this equation can be written as follows
 
c1
.
 
k
X 



y= ci yi = Y 
 . 

i=1 
 . 

ck
where c1 , ..., ck ∈ C. If we take Y to be the matrix function etA , we recover

the familiar expression (2.19) for the general solution of (2.18).
Example 2.2.11. We seek the solution of the initial value problem
! !
0 −5 3 1
y (t) = y(t), y(0) =
−15 7 1
The matrix A in this example has two complex conjugated eigenvalues: λ1 =

1 + 3i and λ2 = λ1 = 1 − 3i. An eigenvector with eigenvalue 1 + 3i is, for
instance, (1, 2 + i) and thus (1, 2 − i) is an eigenvector with eigenvalue 1 − 3i
(in general, if A is a real matrix and x an eigenvector with eigenvalue λ, so
121
that Ax = λx, then Ax = Ax = λx = λx). The initial vector should be a
linear combination of these eigenvectors:
! ! ! ! !
1 1 1 1 1 α
=α +β =
1 2+i 2−i 2+i 2−i β
Hence ! !−1 ! !
α 1 1 1 1 1+i
= =
β 2+i 2−i 1 2 1−i
It now follows that
! !
1 1 1
y(t) = (1
+ i)e (1+3i)t
+ 1 (1 − i)e(1−3i)t
2 2 + i!! 2 2−i
1+i
= Re e(1+3i)t
!
1 + 3i !
t 1 t 1
= e cos 3t − e sin 3t
1 3
Alternatively, we can use the second part of Theorem 2.2.8 to compute this
solution:
! ! !−1 !
1 1 e(1+3i)t 0 1 1 1
y(t) = etA y(0) = (1−3i)t
2+i 2−i 0 e 2+i 2−i 1
After some computations we find

! !
1 e3it (1 + i) + e−3it (1 − i) cos 3t − sin 3t
y(t) = et =e t
2 e (1 + 3i) + e−3it (1 − 3i)
3it
cos 3t − 3 sin 3t
(Verify that this vector function does indeed satisfy all requirements.) Note
that it is unnecessary, and somewhat akward, to compute the matrix function
etA .
If A is nondiagonalizable, there is no basis of Ck consisting of eigenvec-
tors of A, but there always exists a basis {s1 , ..., sk } of generalized eigen-
vectors. The solution space of (2.18) then again is spanned by the vector
functions
yi (t) = etA si , i = 1, ..., k
However, only for those values of i for which si is a genuine eigenvector, do
we have yi (t) = eλi t si .
122
Example 2.2.12. The vectorial ODE
!
0 0 1
y (t) = y(t)
0 0
is equivalent to the system
y10 (t) = y2 (t)
y20 (t) = 0
This has the general solution
y2 (t) = c2 , y1 (t) = c2 t + c1 , c1 , c2 ∈ C
In vectorial form: ! !
1 t
y(t) = c1 + c2
0 1
The matrix A has an eigenvector s1 := (1, 0), with eigenvalue 0, and a gen-
eralized eigenvector s2 := (1, 1), of degree 2, with respect to this eigenvalue.
The matrix function etA is easily computed by means of (45) in §A.3, since
An vanishes for all n > 1, so
! !
tA 0 1 1 t
e =I +t =
0 0 0 1
The column vectors of this fundamental matrix form a basis of the solution
space of the equation. Another basis consists of the vector functions
! !
tA 1 tA 1+t
e s1 = and e s2 =
0 1
Verify that both bases yield the general solution determined above.
In the general case, the matrix function etA can be computed as follows.
Suppose that, by a change of coordinates, A is transformed into a Jordan
normal form:
S −1 AS = J = Λ + N
where Λ = diag{λ1 , ..., λk } and N is a nilpotent matrix of a particular form,
which commutes with Λ (cf. §A.1). Then, due to Theorem A.3.1,
k−1
X nt Nn
S −1 etA S = etJ = etΛ etN = diag{eλ1 t , ..., eλk t } (2.21)
n=0 n!
123
In particular, for a single k × k Jordan block, with eigenvalue λ, we have
t2 tk−1
 
1 t 2!
. (k−1)!
tk−2
 

 0 1 t . (k−2)!


etJk (λ) = eλt  (2.22)
 
 . . . . . 


 . . . . t 

. . . . 1
Example 2.2.13. The matrix
!
5 1
A=
−1 3
is nondiagonalizable. In example A.1.6, in the appendix, it is shown that
! ! ! !
5 1 1 0 4 1 1 0
=
−1 3 −1 1 0 4 1 1
Hence the equation y 0 (t) = Ay(t) has a fundamental matrix of the form
!
! t 4 1 ! ! ! !
tA 1 0 0 4 1 0 4t 1 0 1 t 1 0
e = e =e
−1 1 1 1 −1 1 0 1 1 1
Another fundamental matrix is
! ! ! !
tA 1 0 4t 1 0 1 t 4t 1 t
e =e =e
−1 1 −1 1 0 1 −1 1 − t
The general solution of the equation has the form
! ! !
4t 1 t c1 4t c1 + c2 t
y(t) = e =e , c1 , c2 ∈ C
−1 1 − t c2 c2 − c1 − c2 t
The corresponding initial value problem, with initial vector
!
1
y(0) =
2
has the following solution
! ! ! ! ! !
tA 1 4t 1 0 1 t 1 0 1 4t 1 + 3t
y(t) = e =e =e
2 −1 1 0 1 1 1 2 2 − 3t
(Verify that this vector function has all the required properties.)
124
In cases where A ∈ Rk×k has eigenvalues with nonzero imaginary part and
where one is exclusively interested in real-valued solutions, it is inconvenient
to use the fundamental matrix SetJ . In such cases, the following theorem
can be used to construct a real-valued basis of the solution space of (2.18).
Theorem 2.2.14. Suppose A ∈ Rk×k has an eigenvalue λ with nonzero
imaginary part and eigenvector v. Then the functions y1 and y2 defined by
y1 (t) = Re etλ v and y2 (t) = Im etλ v
are linearly independent solutions of (2.18).
Example 2.2.15. Consider once more the equation in example 2.2.11. The
matrix A in this example has an eigenvector (1, 2 + i) with eigenvalue 1 + 3i.
With the aid of Theorem 2.2.14, we obtain the following real-valued basis of
the solution space of (2.18):
( !) !
(1+3i)t 1 t cos 3t
y1 (t) = Re e =e ,
2+i 2 cos 3t − sin 3t
( !) !
(1+3i)t 1 t sin 3t
y2 (t) = Im e =e
2+i 2 sin 3t + cos 3t
We can solve the initial value problem in example 2.2.11 by writing y as a
linear combination y = αy1 + βy2 of y1 and y2 and solving α and β from the
equation
! ! !
1 1 0
y(0) = = αy1 (0) + βy2 (0) = α +β
1 2 1
(Show that this yields the solution found in example 2.2.11).
Example 2.2.16. The matrix
 
−1 1 0
A :=  −1 1 −1 
 
0 1 −1
 
1
has an eigenvector s1 =  0  with eigenvalue −1, an eigenvector s2 =
 
−1
   
1 1
 1 + i  with eigenvalue i and (thus) an eigenvector s3 =  1 − i  with
   
1 1
125
eigenvector −i. Hence the solution space of the equation y 0 (t) = Ay(t) has
a basis formed by e−t s1 , eit s2 and e−it s3 . By Theorem 2.2.14, the solution
space has a real-valued basis {y1 , y2 , y3 }, where
      
1 
 1 
 cos t
−t  it 
y1 (t) = e  0  , y2 (t) = Re e  1 + i  =  cos t − sin t 
   
−1 
1 
cos t
and     

 1 
 sin t
y3 (t) = Im eit 
 1 + i  =  sin t + cos t 
  
 

1 
sin t
2.2.3 Inhomogeneous linear vectorial differential equa-

tions with constant coefficients
Next, we consider the inhomogeneous system of first order ODE’s, or the
inhomogeneous vectorial ODE
y 0 (t) = Ay(t) + b(t), t ∈ T
where A denotes a constant, k × k matrix and b a k-dimensional vector

function, which is supposed continuous on an interval T of the form T =
[t0 , t1 ] or T = [t0 , ∞). If a vector function y satisfies this equation, then
d
dt
(e−tA y(t)) = −Ae−tA y(t) + e−tA y 0 (t) = −Ae−tA y(t) + e−tA (Ay(t) + b(t))
= −Ae−tA y(t) + Ae−tA y(t) + e−tA b(t) = e−tA b(t)
The right-hand side being continuous on T , both sides are integrable and,
for every t ∈ T , we have
Z t Z t
d −τ A
(e y(τ ))dτ = e−τ A b(τ )dτ
t0 dτ t0
This yields the identity

Z t
e−tA y(t) − e−t0 A y(t0 ) = e−τ A b(τ )dτ
t0
126
Multiplying both sides from the left with etA , we obtain
Z t
y(t) = e(t−t0 )A y(t0 ) + etA e−τ A b(τ )dτ, t ∈ T
t0
k
Thus, for every vector y0 ∈ C , the initial value problem
y 0 (t) = Ay(t) + b(t), y(t0 ) = y0
has the (unique) solution
Z t
y(t) = e (t−t0 )A
y0 + e tA
e−τ A b(τ )dτ, t ∈ T
t0
Conversely, it is easily seen that, for every k-dimensional vector c ∈ Ck , the

vector function y defined by
Z t
y(t) = etA c + etA e−τ A b(τ )dτ, t ∈ T
t0
satisfies the equation (check this). Hence, the above expression represents
the general solution of the inhomogeneous equation. Note that the second
term on the right-hand side is the solution of the initial value problem with
initial vector 0 (the nullvector of Ck ). Setting b ≡ 0 we retrieve the general
solution of the homogeneous equation (cf. Theorem 2.2.8).
Example 2.2.17. The solution of the initial value problem
! ! !
0 4 1 0 1
y (t) = y(t) + , y(0) =
−2 1 −2et 0
is given by ! !
1 t Z
0
tA
y(t) = e + e(t−τ )A dτ
0 0 −2eτ
With the aid of example 2.2.10, we find
! ! !
1 1 e2t 0 −1
y(t) = +
−2 −1 0 e3t 2
!Z ! ! !
1 1 t e2(t−τ ) 0 −1 −1 0
+ dτ
−2 −1 0 0 e3(t−τ )
2 1 −2eτ
After some computations, we obtain
!
e3t + e2t − et
y(t) =
−e3t − 2e2t + 3et
(Verify this.)
127
2.2.4 Exercises
1. Determine the matrix function etA in the following cases:
! ! !
1 0 0 1 0 1
A= , A= , A=
1 2 1 0 −1 0
2. Determine the general solution of the system

y10 = 6y1 − 3y2
y20 = −2y1 + y2
3. Determine a fundamental matrix of the vectorial ODE

!
0 1 4
y (t) = y(t)
3 2

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

! !
0 1 1 1
y (t) = y(t), y(0) =
4 −2 0

! !
0 1 −5 1
y (t) = y(t), y(0) =
2 −5 0
7. Determine a real-valued basis of the solution space of the vectorial ODE

!
0 1 1
y (t) = y(t)
−2 −1
8. A vectorial ODE of the form y 0 (t) = Ay(t) has a fundamental matrix

!
4e6t −et
Y (t) =
e6t et
Compute A and etA .
128
9. Prove that the vector functions y1 and y2 , defined by
! !
e−t e−2t
y1 (t) = , y2 (t) = , t∈R
3e−t −4e−2t
are linearly independent. Find out whether or not there exists a homoge-
neous, linear, vectorial ODE, such that both y1 and y2 are solutions.
10. Determine a fundamental matrix and the general solution of the vectorial
ODE !
5 −2
y 0 (t) = y(t)
4 −1
11. a. Convert the second order ODE
y 00 (t) + 4y 0 (t) + 5y(t) = 0
into a first order vectorial ODE.

b. Determine a real-valued basis of the solution space of the vectorial
ODE.
c. From the result in b. derive a real-valued basis of the solution space of
the original, second order ODE.
12. Determine the general solution of the vectorial ODE

 
2 0 0
0
y (t) =  0 1 0 

 y(t)
0 1 1

   
3 −1 1 1
0
y (t) =  2

0 1  y(t), y(0) =  −1 
 

1 −1 2 2
14. Determine the general solution of the vectorial ODE: y 0 (t) = Ay(t), where
 
0 1 0
A= 0

0 1 

2 −5 4
129
and find the solution of the initial value problem
et
   
1
y 0 (t) = Ay(t) +  2et  , y(0) =  1 
   
4et 1
2.3 Stability of solutions of systems of first

order differential equations
We begin by presenting the definitions of stability and asymptotic stability
for the general case of a first order vectorial ODE, of the form
y 0 (t) = f (t, y(t)), t ∈ [t0 , ∞) (2.23)
where f is an Rk or Ck -valued function, defined on a set of the form [t0 , ∞)×

D, with D ⊂ Rk or D ⊂ Ck .
Definition 2.3.1. A solution y of (2.23), defined on [t0 , ∞), is called stable

if, for a given vectornorm k.k, for every > 0 there exists a δ > 0, such that
every solution ỹ, defined on an interval [t0 , t1 ], with t1 > t0 , and satisfying
the condition kỹ(t0 ) − y(t0 )k ≤ δ, has the following properties:
ỹ is a solution on [t0 , ∞), and
kỹ(t) − y(t)k ≤ for all t ∈ [t0 , ∞)
A stable solution y is called (locally) asymptotically stable, if there exists

a positive number δ, such that, for every solution ỹ, with the property that
kỹ(t0 ) − y(t0 )k ≤ δ, we have
lim (ỹ(t) − y(t)) = 0

t→∞
A stable solution is called globally asymptotically stable, if every solution

ỹ, defined on an interval [t0 , t1 ], with t1 > t0 , is a solution on [t0 , ∞), and
limt→∞ (ỹ(t) − y(t)) = 0.
A stable solution that is not asymptotically stable is called neutrally stable.
A solution that is not stable is called unstable.
If f is an affine (vector) function of y, then (2.23) takes the form of a linear

system of first order ODE’s, or a linear, vectorial ODE:
130
y 0 (t) = A(t)y(t) + b(t), t ∈ [t0 , ∞) (2.24)
Here, A and b are a k×k matrix function and a k-dimensional vector function
on [t0 , ∞), respectively. Analogously to the case of (a system of) linear
difference equations, we have the following theorem.
Theorem 2.3.2. All solutions of the linear, vectorial ODE (2.24) are neu-
trally stable, globally asymptotically stable, or unstable, if and only if the
null solution of the homogeneous equation is neutrally stable, asymp-
totically stable, or unstable, respectively.
Again, we consider the special case of a system of first order, linear ODE’s
with constant coefficients, i.e. an equation of the form
y 0 (t) = Ay(t) + b(t), t ∈ [t0 , ∞) (2.25)

where A is a constant k × k matrix. From now on, we take t0 = 0. For every
solution of the homogeneous equation
y 0 (t) = Ay(t), t ∈ [0, ∞) (2.26)

the following relation holds
y(t) = etA y(0)
Similarly to systems of difference equations with constant coefficients, it is

again the Jordan normal form of the matrix A that determines the stability
of the null solution of (2.26). Suppose that S is a nonsingular matrix, such
that S −1 AS is in Jordan normal form: S −1 AS = J = Λ + N , where Λ is
a diagonal matrix with diagonal entries λ1 , ..., λk (the eigenvalues of A) and
N a nilpotent matrix, commuting with Λ. For every solution y of (2.26) we
now have (cf. also (2.21)),
k−1
X l t Nl
!
−1
tΛ tN
y(t) = Se e S y(0) = Se tΛ
S −1 y(0) (2.27)
l=0 l!
The matrix etΛ is a diagonal matrix, with diagonal entries etλ1 , ..., etλk and
the sum, in parentheses, on the right-hand side of (2.27) is a nilpotent, upper
triangular matrix, whose entries are polynomials in t, of degree not exceeding
k − 1.
131
Theorem 2.3.3. The null solution of equation (2.26) is asymptotically
stable if and only if the real parts of all eigenvalues of A are negative. The
null solution is stable if A has no eigenvalues with positive real part, and, in
addition, the algebraic and geometric multiplicities of every purely imaginary
eigenvalue (i.e. with real part = 0) are equal. In all other cases, the null
solution is unstable.
From the above theorem the following necessary (but not sufficient!) con-
ditions for stability and asymptotic stability of the null solution of (2.26) (and
hence of all solutions of (2.25)) can be derived.
Lemma 2.3.4. If the null solution of (2.26) is stable, then
Re tr A ≤ 0
If the null solution is asymptotically stable, then
Re tr A < 0
The next theorem provides sufficient conditions for asymptotic stability of

the null solution of (2.26).
Theorem 2.3.5. Suppose that the matrix A has a dominant negative

diagonal, i.e.:
(i) A has a dominant diagonal:
X
|Aij | < |Aii |, for i = 1, ..., k
j6=i
(ii) Aii < 0 for i = 1, ..., k.

Then the null solution of equation (2.26) is asymptotically stable.
Proof. We intend to show that all eigenvalues of a matrix with a dominant

negative diagonal have negative real parts. By Theorem 2.3.3, this implies
the asymptotic stability of the null solution of (2.26). For every λ ∈ C we
have
|Aii − λ|2 = (Aii − Re λ)2 + (Im λ)2 ≥ (Aii − Re λ)2 (2.28)
Due to (ii), Aii = −|Aii | for i = 1, ..., k, hence
(Aii − Re λ)2 = (−|Aii | − Re λ)2 = (|Aii | + Re λ)2 (2.29)
132
Suppose that Re λ ≥ 0. Then |Aii | + Re λ ≥ |Aii | and, with (2.28) and
(2.29), it follows that
|Aii − λ| ≥ |Aii | (2.30)
Now consider the matrix B := A − λI. We have: Bii = Aii − λ and Bij = Aij
if j 6= i. From (2.30) and assumption (i) we deduce
X X
|Bij | = |Aij | < |Aii | ≤ |Aii − λ| = |Bii |
j6=i j6=i
This shows that A − λI has a dominant diagonal and, therefore, must be

nonsingular. Thus, λ cannot be an eigenvalue of A and we conclude that all
eigenvalues of A have negative real parts.
2.3.1 Autonomous systems of nonlinear differential equa-

tions
To conclude this chapter, we discuss the generalization of Theorem 2.3.3 to
systems of first order ODE’s of the form
y 0 (t) = f (y(t)), t ∈ [0, ∞) (2.31)
This is an autonomous system. The homogeneous, linear system with con-

stant coefficients, y 0 (t) = Ay(t), is a particular case of (2.31), where f (x) =
Ax.
Theorem 2.3.6. Let c ∈ Ck be a zero of f and suppose that f is differen-

tiable at c (sufficient condition: all first order partial derivatives of f exist
and are continuous on a neighbourhood of c). Let A := Df (c). The solution
y ≡ c of equation (2.31) is (locally) asymptotically stable if all eigen-
values of A have negative real parts. If at least one eigenvalue of A has a
positive real part, then the solution y ≡ c is unstable.
The following theorem is the continuous analogue of Theorem 1.5.3.
Theorem 2.3.7. Let A be a constant k × k matrix and g a k-dimensional

vector function, continuous at 0 (the origin of Rk or Ck ), with the property
that, for some (hence for any) given vectornorm k.k,
kg(y)k
lim =0 (2.32)
y→0 kyk
133
If all eigenvalues of A have negative real parts, then the null solution of the
equation
y 0 (t) = Ay(t) + g(y(t)), t ∈ [0, ∞) (2.33)
is asymptotically stable. If, on the other hand, A has at least one eigen-
value with positive real part, then the null solution is unstable.
Like in the case of systems of first order difference equations, Theorems

2.3.6 and 2.3.7 are based on linearization of the system of ODE’s at the
equilibrium point under consideration. The linearized system is obtained
by replacing f with its linear approximation at c, thus:
y 0 (t) = f (c) + Df (c)(y(t) − c) = Df (c)(y(t) − c)
Example 2.3.8. Replacing, in the predator-prey model of example 1.5.5,

the discrete variable n by a continuous time-variable t and the ‘difference’
yi (n + 1) − yi (n) by the derivative yi0 (t) (i = 1, 2), we obtain a continuous
predator-prey model, described by the following system of ODE’s
y10 (t) = (r − ay1 (t) − by2 (t))y1 (t)

t ∈ [0, ∞) (2.34)
y20 (t) = (cy1 (t) − 1)y2 (t)
Here, y1 (t) and y2 (t) denote the number of prey and the number of predators
at time t, respectively. a, b, c and r again are positive constants. The
Jacobian matrix of the vector function
!
(r − ax1 − bx2 )x1
f (x) :=
(cx1 − 1)x2
at any point x := (x1 , x2 ), is given by

!
r − 2ax1 − bx2 −bx1
Df (x) =
cx2 cx1 − 1
Df (0, 0) is a diagonal matrix, with diagonal entries r and −1. By Theorem

2.3.6, due to the fact that r > 0, the null solution of (2.34) is unstable.
The linearized system at (0, 0) is
!
0 r 0
y (t) = y(t)
0 −1
134
Like its discrete analogue, the system (2.34) has two more equilibrium solu-
tions, one of which is the following:
! !
1
y1 (t)
= 1
c for all t ∈ [0, ∞)
y2 (t) b
(r − ac )
Let’s assume that r > ac . The matrix A in that case is

!
− ac − cb
A= 1
b
(cr − a) 0
The eigenvalues of A are the solutions of the equation

a a
λ2 + λ + r − = 0
c c
If for r and ac we choose the same values as in the discrete example, i.e.
r = 38 and ac = 14 , then the discriminant of the above quadratic equation
is negative, hence the matrix A has two complex conjugated eigenvalues λ+
and λ− . These eigenvalues have identical real parts, and
1
2Re λ+ = λ+ + λ− = tr A = −
4
According to Theorem 2.3.6, the equilibrium solution under consideration is
asymptotically stable.
Example 2.3.9 (The pendulum). A classic example of a nonlinear ODE,

possessing both a stable and an unstable equilibrium solution, is the equation
of motion of a frictionless pendulum:
y 00 (t) + sin y(t) = 0, t ∈ [0, ∞) (2.35)

Here, y denotes the angle, in radians, that the pendulum makes with the
vertical. The equation has two equilibrium solutions: y1 ≡ 0 (or y1 ≡ 2nπ,
n ∈ Z) and y2 ≡ π (or y2 ≡ (2n + 1)π, n ∈ Z). From ‘experience’, we
expect the first equilibrium to be stable, but not asymptotically stable:
if, at time t = 0, the pendulum is released from an initial position close
to, but different from 0 (with initial velocity = 0), it will, in the absence of
friction, start a periodic movement about the equilibrium position 0 (with
maximal deviation from 0 equal to the initial one). The second equilibrium
135
is unstable: an arbitrarily small disturbance suffices to send the pendulum
swinging downward.
This second order ODE can be converted into a system of two first order
ODE’s, by means of the following transformation. Define new (dependent)
variables y1 and y2 by
y1 := y, y2 := y 0 (2.36)
If y is a solution of (2.35), then y1 and y2 satisfy the system of equations
y10 (t) = y2 (t)

, t ∈ [0, ∞)
y20 (t) = y 00 (t) = − sin y1 (t)
Hence, the vector y := (y1 , y2 ) satisfies the vectorial differential equation
! ! ! !
d y1 (t) 0 1 y1 (t) 0
= + , t ∈ [0, ∞)
dt y2 (t) −1 0 y2 (t) y1 (t) − sin y1 (t)
(2.37)
This is of the form (2.33), with g(y) = (0, y1 − sin y1 ). Now we have
kg(y)k1 |y1 − sin y1 | |y1 − sin y1 |
lim = lim ≤ lim = 0,
y→0 kyk1 (y1 ,y2 )→(0,0) |y1 | + |y2 | (y1 ,y2 )→(0,0) |y1 |
and thus condition (2.32) holds. The eigenvalues of the matrix A are the
solutions of the equation λ2 +1 = 0, so the numbers i and −i. Unfortunately,
however, both eigenvalues are purely imaginary, so that Theorem 2.3.7 does
not apply. This was to be expected, as the theorem provides a sufficient
condition for asymptotic stability, but not for neutral stability of the
null solution. Next, we consider the equilibrium solution y ≡ π of (2.35). It
corresponds to the solution
! !
y1 π
≡
y2 0
of the vectorial ODE (2.37) (check this). Introducing new (dependent) vari-
ables u1 and u2 , defined by
u1 := y1 − π, u2 := y2
we have ! !
d u1 (t) u2 (t)
= , t ∈ [0, ∞)
dt u2 (t) − sin(u1 (t) + π)
136
or, equivalently,
! ! ! !
d u1 (t) 0 1 u1 (t) 0
= + , t ∈ [0, ∞)
dt u2 (t) 1 0 u2 (t) sin u1 (t) − u1 (t)
(2.38)
This is of the form (2.33) again. The matrix A in this case has eigenvalues 1
and −1. By Theorem 2.3.7, the null solution of (2.38), and thus the solution
y ≡ (π, 0) of (2.37), is unstable.
We conclude the example with a version of (2.35) that takes into account the
friction experienced by the pendulum:
y 00 (t) + wy 0 (t) + sin y(t) = 0, t ∈ [0, ∞) (2.39)

Here, w a positive constant, the frictional coefficient. By means of the trans-
formation (2.36), the equation is converted into the vectorial differential equa-
tion
! !
d y1 (t) y2 (t)
=
dt y2 (t) −sin y1 (t)−wy2 (t)
! ! !
0 1 y1 (t) 0
= + , t ∈ [0, ∞)
−1 −w y2 (t) y1 (t)−sin y1 (t)
(2.40)
The eigenvalues of the 2 × 2 matrix on the right-hand side of (2.40) are the
solutions of the equation
λ2 + wλ + 1 = 0
For every positive value of w, both eigenvalues have negative real part (the
proof of this assertion is left to the reader, cf. exercise 10 below). With
the aid of Theorem 2.3.7, we conclude that the null solution of (2.40) is
asymptotically stable.
Similarly to the case of difference equations, stability or instability of solu-

tions of systems of differential equations can sometimes be established by
means of Lyapunov’s direct method.
Theorem 2.3.10. Let f be a differentiable function from Rk to Rk , vanish-

ing at x0 . Suppose there exists a neighbourhood U of x0 and a C 1 -function
V : U → R, with the property that V (x0 ) = 0 and V (x) > 0 for all
137
x ∈ U − {x0 }.
(i) If V̇ (x) := DV (x)f (x) = kj=1 ∂x
∂V
(x)fj (x) ≤ 0 for all x ∈ U , then y ≡ x0
P
j
is a stable equilibrium solution of (2.31).
(ii) If V̇ (x) < 0 for all x ∈ U − {x0 }, then x0 is an asymptotically stable
equilibrium solution. Moreover, any solution y of (2.31), with the property
that there exists r > 0 such that, for all t ∈ [0, ∞), y(t) ∈ B(x0 , r) ⊂ U ,
converges to x0 as t → ∞.
(iii) If V̇ (x) > 0 for all x ∈ U − {x0 }, then x0 is an unstable equilibrium
solution.
A function V satisfying the conditions of part (i) of Theorem 2.3.10, is called a
Lyapunov function on U , centered at x0 , for equation (2.31). A Lyapunov
function, satisfying the conditions of part (ii) of Theorem 2.3.10, is called a
strict Lyapunov function on U , centered at x0 , for equation (2.31).
Remark 2.3.11. If f is a continuous function from Rk to Rk , then any
solution y of equation (2.31) is a C 1 -function. Consequently, if V : U → R
is a C 1 -function, then so is V ◦ y and
d
V (y(t)) = DV (y(t))y 0 (t) = DV (y(t))f (y(t)) = V̇ (y(t))
dt
The condition that f should be differentiable ensures the existence and
uniqueness of solutions with a given initial value. It is a sufficient condi-
tion and is often replaced by a weaker, so-called Lipschitz condition.
If x0 is an asymptotically stable equilibrium for equation (2.31), the set of
all x in the domain of f , with the property that the solution y of (2.31),
with initial value y(0) = x, converges to x0 as t → ∞, is called the basin of
attraction of x0 .
Example 2.3.12. We can apply Theorem 2.3.10 to the example of the pen-
dulum, with or without friction, i.e. equation (2.40) with w ≥ 0. We define:
U = {x ∈ R2 : |x1 | < 2π} and
1
V (x) = x22 + 1 − cos x1
2
(the energy of the pendulum). It is easily seen that the right-hand side ≥ 0
for all x ∈ R2 and = 0 iff x2 = 0 and x1 = 0 mod 2π, hence V (x) > 0 for all
x ∈ U with the property that x2 6= 0. Furthermore,
∂V ∂V
= sin x1 , = x2 ,
∂x1 ∂x2
138
V̇ (x) = −wx22 ≤ 0 for all x ∈ U
According to the first part of Theorem 2.3.10, the null solution of (2.40) is
stable, also in the frictionless case, i.e. when w = 0. However, this particular
Lyapunov function doesn’t allow us to conclude that the null solution is
asymptotically stable when w > 0, as V̇ (x) = 0 whenever x2 = 0, so for all
points on the x1 -axis, and hence (ii) doesn’t hold.
2.3.2 Exercises
1. a. Determine the general solution of the system of ODE’s
!
0 1 1
y (t) = y(t) t ∈ [0, ∞)
−2 −2
b. Compute, for every solution y of this system, limt→∞ y(t).

c. Examine the stability of the solutions.
2. Examine the stability of the solutions of the following systems of ODE’s:

a.
! !
0 −1 −1 t
y (t) = y(t) + , t ∈ [0, ∞)
2 −1 e−t
b.
0 2 1 0
 
 −2 0 0 1 
y 0 (t) =   y(t), t ∈ [0, ∞)
 
 0 0 0 2 
0 0 −2 0
3. Determine all equilibrium solutions of the system
y10 (t) = (1 − 2y1 (t) − y2 (t))y1 (t)

, t ∈ [0, ∞)
y20 (t) = (1 − y1 (t) − 2y2 (t))y2 (t)
139
y10 (t) = (1 − y1 (t) − y2 (t))y1 (t)
, t ∈ [0, ∞)
y20 (t) = (2y1 (t) − 1)y2 (t)
y10 (t) = 2y1 (t) − y12 (t) − y1 (t)y2 (t)
, t ∈ [0, ∞)
y20 (t) = y1 (t)y2 (t) − y2 (t)
6. a. Determine the equilibrium solutions of the second order ODE
y 00 (t) + y 0 (t) + y(t)y 0 (t)2 + y(t)2 − 1 = 0, t ∈ [0, ∞)
b. Convert the equation into a system of first order ODE’s.

c. Determine the equilibrium solutions of this system and examine their
stability.
7. a. Determine all real-valued equilibrium solutions of the ODE
y 00 (t) + y 0 (t) + sin(2y(t)) = 0, t ∈ [0, ∞)
b. Convert the equation into a system of first order ODE’s.

c. Determine the real-valued equilibrium solutions of this system and ex-
amine their stability.
8. Prove that the system (2.34) has an asymptotically stable equilibrium
a2
solution when r > ac + 4c 2.
9. Linearize equation (2.34) at the equilibrium point ( 1c , 1b (r − ac )).

10. Determine all equilibrium solutions of the second order ODE (2.39) and
examine their stability.
11. a. Prove that the system of real-valued differential equations
dy1
dt
= −y1 + y22
dy2
dt
= y1 y2 − y23
has an infinite number of equilibrium solutions.
140
b. Prove that none of these equilibrium solutions is asymptotically stable.
c. Use the function V (x1 , x2 ) = x21 + x22 to prove that the null solution is
stable.
12. a. Determine the equilibrium solutions of the system of real-valued differ-
ential equations
y10 (t) = −2y2 (t)2
y20 (t) = y1 (t)y2 (t)
b. Linearize the system around the solution (y1 , y2 ) ≡ (1, 0) and examine
the stability of all nonzero equilibrium solutions of the equation in a.
c. Use a Lyapunov function of the form V (y1 , y2 ) = y12 + by22 to examine
the stability of the null solution of the nonlinear system of differential
equations.
d. Does the system have any asymptotically stable equilibrium solutions?
13. Examine the stability of the null solution of the system of real-valued
differential equations
dy1
dt
= ay2 + by1 (y12 + y22 )
dy2
dt
= −ay1 + by2 (y12 + y22 )
where a, b ∈ R.
14. a. Convert the real-valued equation
y 00 (t) + µy 0 (t) + y(t)3 = 0 (µ > 0)
into a system of first order ODE’s.
b. Use the function V (x1 , x2 ) = 21 x41 + x22 to examine the stability of the
null solution of this system.
15. a. Convert the real-valued equation
y 00 (t) + y 0 (t)y(t)2 + y(t)3 − a2 y(t) = 0
into a system of first order differential equations.
b. Determine the equilibrium solutions of this system and examine their
stability in the case that a ∈ R, a 6= 0.
c. Examine the stability of the null solution of the system in the case that
a = 0, using Lyapunov’s direct method. Try a Lyapunov function of
the form V (x1 , x2 ) = x41 + bx22 .
141
142
Appendix
A.1 Change of basis and Jordan normal form

Any k × k matrix A with real entries represents a linear map L : Rk → Rk .
The column vectors of A are the coordinate vectors of the images of the
basis vectors e1 , ..., ek , i.e. of Lei , i = 1, ..., k. The matrix that represents
L thus depends on the choice of a basis. A change of basis will change the
matrix representing a given linear map. ! Thus,! for instance, with respect to
1 0
the standard basis {e1 , e2 } := { , }, the rotation over 90◦ in R2
0 1
is represented by the matrix
!
0 −1
1 0
since
Le1 = 0.e1 + 1.e2 and Le2 = −1.e1 + 0.e2
! !
1 1
whereas, w.r.t. the basis {e01 , e02 } := { , }, it is represented by
1 −1
!
0 1
−1 0
as
Le01 = 0.e01 + −1.e02 and Le02 = 1.e01 + 0.e02
A change of basis of Rk , usually called a coordinate transformation, or
change of coordinates, is a bijective linear map, represented by a nonsingular
k × k matrix. Conversely, every nonsingular matrix with real entries cor-
responds to a change of coordinates in Rk . Likewise, there is a one-to-one
correspondence between changes of coordinates in Ck and nonsingular k × k
143
matrices with complex entries. In what follows, a ‘coordinate transforma-
tion’ refers to a coordinate transformation of Ck . If {e1 , ..., ek } denotes
the standard basis of Ck and {s1 , ..., sk } any other basis, then the change of
coordinates transforming the first basis into the second, is, w.r.t. the stan-
dard basis, represented by the matrix S with column vectors s1 , ..., sk . A
linear map L : Ck → Ck , which, w.r.t. the standard basis, is represented by
a matrix A, in the new coordinate system is represented by
S −1 AS
Matrices that are related through a change of coordinates, are, in a certain

sense, equivalent, and are called similar matrices. Several important proper-
ties of square matrices, such as determinant, trace, etc., are invariant under
coordinate transformations.
Lemma A.1.1. Similar square matrices A and B possess the same charac-
teristic polynomial.
Proof. Suppose that A and B ∈ Ck×k are similar matrices. Then there
exists a nonsingular matrix S ∈ Ck×k , such that B = S −1 AS. The charac-
teristic polynomial pB of B is given by
pB (λ) = det(B − λI) = det(S −1 AS − λI) =
= det(S −1 (A − λI)S) = det S −1 det(A − λI) det S

As det S −1 = (det S)−1 , this implies that pB (λ) = det(A − λI).
Many problems, related to linear mappings, can be simplified by an appro-

priate choice of coordinates, such that the action of a specific, linear map
is as simple as possible. In terms of matrices, this amounts to choosing,
from the similarity class of matrices representing this map, an element with
the simplest possible form, such as a triangular matrix, or, even better, a
diagonal matrix. Suppose, for instance, that a coordinate transformation,
represented by the nonsingular matrix S, converts a given matrix A into a
diagonal matrix Λ, i.e. S −1 AS = Λ. Then we have, for every nonnegative
integer n,
An = SΛn S −1
and the right-hand side of this identity can be easily computed, for arbitrary
large values of n.
144
A k × k matrix A is diagonalizable if and only if it has exactly k
linearly independent eigenvectors v 1 , ..., v k . These eigenvectors form a
basis of Ck and the action of A on these basis vectors simply is multiplication
with a scalar: Av i = λi v i , i = 1, ..., k, where λi is the eigenvalue of A,
corresponding to the eigenvector v i . Let S be the matrix with column vectors
v 1 , ..., v k , then
AS = (Av 1 , ..., Av k ) = (λ1 v 1 , ..., λk v k ) = (v 1 , ..., v k ) diag{λ1 , ..., λk }
Hence S −1 AS = diag{λ1 , ..., λk } (i.e. the diagonal matrix with diagonal
entries λ1 ,...,λk ). A k × k matrix with k distinct eigenvalues, always has
k linearly independent eigenvectors and thus is diagonizable. Furthermore,
every hermitian and, in particular, every real, symmetric matrix is diago-
nizable. (A k × k matrix A is hermitian iff A∗ := AT = A.) Not all matrices
are diagonizable. But every matrix can be brought to a (lower or upper)
triangular form (such as the Schur normal form). The simplest triangular
matrix, that is similar to a given matrix A, is the so-called Jordan normal
form, or Jordan canonical form, of A. If A is diagonizable, then, naturally,
its Jordan normal form is diagonal. A nondiagonizable matrix has at least
one multiple eigenvalue, whose eigenspace has dimension (the geometric
multiplicity) less than the algebraic multiplicity of the eigenvalue. To give
some idea of how such a matrix can be brought to a Jordan normal form,
we begin by considering the case that the k × k matrix A has a single eigen-
value λ, with algebraic multiplicity k and geometric multiplicity l < k. Note
that, if A has an eigenvalue λ, with geometric multiplicity l, then the matrix
B := A−λI has an eigenvalue 0, with geometric multiplicity l and vice versa.
A matrix with a single eigenvalue 0 is nilpotent.
Definition A.1.2. A square matrix B is nilpotent if there exists a positive
integer n such that B n = 0.
Lemma A.1.3. A k × k matrix B is nilpotent, if and only if B has a single
eigenvalue equal to 0. Every nilpotent k × k matrix B has the property that
B k = 0.
To begin with, we will try to bring the nilpotent matrix B = A − λI to the
simplest possible upper triangular form, by means of a coordinate transfor-
mation, i.e. we seek a basis of Ck , on which the action of B is as simple as
possible. If we succeed, the same transformation can be used for A, as
S −1 BS = S −1 AS − λI
145
If the geometric multiplicity of the eigenvalue 0 of B is l, there are l linearly
independent eigenvectors, with eigenvalue 0. We have to complement these
with k − l linearly independent vectors, to a basis of Ck . Note that every
eigenvector of B is an eigenvector of B 2 , B 3 , etc., but the converse is not
true, in general. Consider, for instance, the nilpotent matrix
 
0 1 0
 0 0 1 
B= 
0 0 0
The matrix  
0 0 1
B2 =  0 0 0 
 
0 0 0
has an eigenvector (0, 1, 0), but this is not an eigenvector of B, as B(0, 1, 0)
= (1, 0, 0). (0, 1, 0) is a so-called generalized eigenvector of B, of degree
2, with respect to the eigenvalue 0. At the same time, it is a generalized
eigenvector of A w.r.t. the eigenvalue λ.
Definition A.1.4. A generalized eigenvector of degree j w.r.t. the

eigenvalue λ of a square matrix A, is a vector x with the property that
(A − λI)j x = 0, but (A − λI)j−1 x 6= 0
The generalized eigenspace of an eigenvalue is the linear space, spanned

by all generalized eigenvectors w.r.t. that eigenvalue.
An ordinary eigenvector is a generalized eigenvector of degree 1. If x is a

generalized eigenvector of degree j, w.r.t. the eigenvalue λ of A and y is a
generalized eigenvector of degree h < j, then x+y is a generalized eigenvector
of degree j, for (A − λI)j (x + y) = 0 and
(A − λI)j−1 (x + y) = (A − λI)j−1 x 6= 0
If λ 6= 0, then also Ax is a generalized eigenvector of degree j, w.r.t. λ, for
(A − λI)j Ax = (A − λI)j+1 x + λ(A − λI)j x = 0
whereas
(A − λI)j−1 Ax = (A − λI)j x + λ(A − λI)j−1 x = λ(A − λI)j−1 x 6= 0
146
If λ = 0, then Ax is a generalized eigenvector of degree j − 1 w.r.t. λ (check
this).
We can now try to supplement the eigenvectors of B = A − λI with
eigenvectors of B 2 and, if necessary, of higher order powers of B, i.e. with
generalized eigenvectors of B (hence of A), until we obtain a basis of Ck .
According to the following theorem, this can always be done.
Theorem A.1.5. Every square matrix A ∈ Ck×k has k linearly indepen-

dent, generalized eigenvectors.
We will give the proof for the very special case that A has a single eigenvalue
λ, and, moreover, B k−1 = (A − λI)k−1 6= 0. Thus, there exists a vector x,
such that B k−1 x 6= 0. Now define
xj := B k−j x, j = 1, ..., k
Note that B is nilpotent, hence, by lemma A.1.3, B k = 0. For every j ∈

{1, ..., k}, B j xj = B k x = 0 and B j−1 xj = B k−1 x 6= 0, i.e. xj is an eigenvector
of B j , hence a generalized eigenvector of degree j of B (and of A as well).
It remains to prove the linear independence of these vectors. Suppose there
exist numbers c1 , ..., ck , not all equal to zero, such that kj=1 cj xj = 0. Let
P
l ≤ k be the largest number such that cl 6= 0. Then we have
c1 x1 + ... + cl xl = c1 B k−1 x + ... + cl B k−l x = 0
B l−1 (c1 B k−1 x + ... + cl B k−l x) = c1 B k+l−2 x + ... + cl B k−1 x = cl B k−1 x = 0
The last identity implies that cl = 0, in contradiction with the definition

of l. Thus we conclude that cj = 0 for all j and, consequently, the vectors
x1 , ..., xk are linearly independent.
It is not hard to see what the matrices, representing the linear mappings
corresponding to A and B, w.r.t. the basis {x1 , ..., xk } defined above, look
like. For j = 2, ..., k we have
Bxj = B k−j+1 x = xj−1
and Bx1 = B k x = 0. Hence it follows that B(x1 ...xk ) = (x1 ...xk )Nk , where
Nk is defined by
147
 
0 1 0 . . 0
0 0 1 . . 0
 
 
 
Nk = 
 . . . . . . 
 (41)

 0 0 0 . . 1 

0 0 0 . . 0
The matrix Nk is an upper triangular matrix of a very special type: it is
the Jordan normal form of B. The Jordan normal form of A is the matrix
Jk (λ) := λIk + Nk , i.e.
 
λ 1 0 . . 0
0 λ 1 . . 0
 
 
 
Jk (λ) = 
 . . . . . . 
 (42)

 0 0 0 . . 1 

0 0 0 . . λ
A matrix of the above form is called a Jordan block of order k. J1 (λ) is
defined as the 1 × 1 matrix with entry λ.
In general, the Jordan normal form of any particular matrix represent-
ing a linear map is a block matrix, whose diagonal blocks are Jordan
blocks, whereas all other blocks are null matrices. For the Jordan normal
form to be uniquely defined, an additional condition on the order in which
the different Jordan blocks are arranged, is needed. Here, we will consider
‘the’ Jordan normal form as being defined up to a permutation of the Jor-
dan blocks. Obviously, the Jordan normal form of a diagonizable matrix is
diagonal, and all its Jordan blocks are of order 1. In order to find a transfor-
mation that converts the matrix A into its Jordan normal form, one proceeds
as follows. First, determine the eigenvalues of A and, for every eigenvalue as
many linearly independent eigenvectors as possible. The maximal number of
linearly independent eigenvectors, corresponding to a particular eigenvalue
λ, equals the dimension of the eigenspace, i.e. the geometric multiplicity
of λ. If the geometric multiplicity m is less than the algebraic multiplicity
l, there are l − m linearly independent generalized eigenvectors of degree
> 1 w.r.t. λ to be determined. This is done by seeking successively, for
every previously determined generalized eigenvector x0 (note that there is
always at least one eigenvector!), as many linearly independent solutions of
the equation
(A − λI)x = x0
148
as possible. If x0 is a generalized eigenvector of degree g, then
(A − λI)g+1 x = (A − λI)g x0 = 0
and
(A − λI)g x = (A − λI)g−1 x0 6= 0
hence x is a generalized eigenvector of degree g + 1.
Example A.1.6. Consider the matrix
!
5 1
A=
−1 3
Its characteristic polynomial is
det(A − λI) = (5 − λ)(3 − λ) + 1 = (λ − 4)2
The eigenspace, associated with the eigenvalue 4, consists of the solutions of
the vectorial equation
! !
1 1 0
(A − 4I)x = x=
−1 −1 0
It is 1-dimensional and is spanned by, for example, x1 := (1, −1). The
generalized eigenvectors of degree 2 are solutions of the equation
(A − 4I)2 x = (A − 4I){(A − 4I)x} = 0
that is, the vectors x with the property that (A − 4I)x is an eigenvector.
This implies that (A − 4I)x = cx1 for some c ∈ C. For our purpose, we seek
a solution of (A − 4I)x = x1 , i.e.
! !
1 1 1
x=
−1 −1 −1
There are infinitely many possibilities, for example, (0, 1) or ( 12 , 12 )T . We

choose the basisvectors x1 and x2 := (0, 1). The matrix S, representing the
corresponding coordinate transformation, is the matrix with column vectors
x1 and x2 , and the Jordan normal form of A is
! ! ! !
1 0 5 1 1 0 4 1
S −1 AS = =
1 1 −1 3 −1 1 0 4
149
We conclude with some general remarks, concerning the Jordan normal form.
Let J be a k × k matrix in Jordan normal form. J is a block matrix, with the
property that all off-diagonal blocks are null matrices. A matrix with this
property is called a block-diagonal matrix. In the case considered here,
the diagonal entries are Jordan blocks, so matrices of the form (42). We will
denote them by Jki (λi ), i = 1, .., r, where r is a positive integer: 1 ≤ r ≤ k,
and ri=1 ki = k. Each Jordan block can be written in the form
P
Jki (λi ) = λi Iki + Nki , i = 1, ..., r
Here, Nki is a nilpotent ki × ki matrix, of the form (41). Obviously, the

matrices λi Iki and Nki commute for every i. Hence it immediately follows
that the block-diagonal matrix Λ, with diagonal entries λi Iki , i = 1, ..., r,
commutes with the block-diagonal matrix N , with diagonal entries Nki , i =
1, ..., r. Λ is an ordinary diagonal matrix and N is a nilpotent matrix, with
the property that Nij = 0 for all pairs (i, j) such that j 6= i + 1. Thus, every
matrix J in Jordan normal form can be written as the sum of a diagonal
matrix Λ and a nilpotent matrix N that commutes with Λ.
A.2 Computation of J n
As a Jordan matrix J is a block-diagonal matrix, with diagonal entries
Jk1 (λ1 ), ..., Jkr (λr ), J n too is block-diagonal, with diagonal entries Jk1 (λ1 )n ,
..., Jkr (λr )n . We are going to examine in detail a single block, of the form
Jk (λ)n = (λIk + Nk )n
We will use the following lemma.
Lemma A.2.1. Suppose that A and B are two commuting k × k matrices.

Then, for every positive integer n, the following identity holds
n
!
n n
An−m B m
X
(A + B) =
m=0 m
Proof. We prove the lemma by means of an inductive argument. If n = 1,

the statement reads
! !
1 1 0 1 0 1
A+B = AB + AB
0 1
150
which is obviously true. Now suppose the statement holds for a certain
positive integer n. Then we have

Pn n
(A + B)n+1 = (A + B) An−m B m
m=0 m
n n n
B + nm=0 m
n−m+1 m
An−m B m+1
P P
= m=0 m A
Pn n!
B + n+1
n+1−m m n! n+1−m m
P
= m=0 m!(n−m)! A m=1 (m−1)!(n+1−m)! A B
n+1 n n! n+1−m m
+ m=1 m!(n+1−m)! (n + 1 − m + m)A B + B n+1
P
= A
Pn+1 (n+1)! n+1−m m Pn+1 n+1 n+1−m m
= m=0 m!(n+1−m)! A B = m=0 m
A B
Hence, by the principle of mathematical induction, the identity holds for all
positive integers.
Noting that (Nk )k = 0 and using lemma A.2.1, we find

min{n,k−1} !
n n n−m
(Nk )m
X
Jk (λ) = λ (43)
m=0 m
It remains to compute the matrix (Nk )m for m ≤ k − 1. To begin with,
we observe that (Nk )ij = 1 if j = i + 1, and (Nk )ij = 0, otherwise. Hence
it follows that (Nk )m
P
ij = j1 ,...,jm−1 (Nk )ij1 (Nk )j1 j2 ...(Nk )jm−1 j = 1 if j1 =
i + 1, j2 = j1 + 1, ..., j = jm−1 + 1, i.e. if j = i + m and (Nk )m ij = 0
m
otherwise. Thus, (Nk ) is an upper triangular matrix with ones on the mth
upper diagonal and zeroes everywhere else. Hence we deduce, with the aid of
formula (43), that Jk (λ)n is an upper triangular matrix, with diagonal entries
λn , while, for
m ≤ min{n, k − 1}, the entries on the mth upper diagonal are
n
equal to m λn−m . All other entries vanish. For example: if n ≥ 3, we have

n n
λn nλn−1 λn−2 λn−3
 
2 3
n

λn nλn−1 λn−2

 0
J4 (λ)n = 

2 
 0 0 λn nλ n−1
 

0 0 0 λn
A.3 Definition and properties of eA

It can be proved that, for every matrix A ∈ Ck×k , the sequence of partial
sums
N
X An
SN := (44)
n=0 n!
151
converges as N → ∞, i.e. limN →∞ (SN )ij exists for every i and j ∈ {1, ..., k}.
The limit of (44) is noted eA , thus
∞
An
eA :=
X
(45)
n=0 n!
If S ∈ Ck×k is nonsingular, then

∞ ∞
−1 A −1 An (S −1 AS)n −1
= eS AS
X X
S e S=S S=
n=0 n! n=0 n!
From lemma A.2.1 we deduce the following theorem.
Theorem A.3.1. If two square matrices A and B commute, then
eA+B = eA eB
Proof. By definition,
∞
(A + B)n
eA+B =
X
n=0 n!
Applying lemma A.2.1, we find
∞ n ∞ X n
!
1 X n 1
eA+B An−m B m = An−m B m
X X
=
n=0 n! m=0 m n=0 m=0 (n − m)!m!
Setting n = m + l and interchanging summations, we obtain

∞ ∞
Al X Bm
eA+B = = eA eB
X
l=0 l! m=0 m!
The following example demonstrates the importance of the condition that A

and B commute.
! !
0 1 1 0
Example A.3.2. Let A = and B = . A and B do not
0 0 0 0
commute, for !
0 −1
AB − BA =
0 0
152
Now, we have !
A 1 1
e =I +A=
0 1
and
∞ P∞
Bn
! !
1
B
X
n=0 n! 0 e 0
e = = =
n=0 n!
0 1 0 1
Hence ! !
A B e 1 B A e e
e e = and e e =
0 1 0 1
Furthermore,
! ! ! !
1 1 1 −1 1 0 1 1
A+B = =
0 0 0 1 0 0 0 1
Consequently,
!
! 1 0 !
A+B 1 −1 0 0 1 1
e = e
0 1 0 1
and thus
! ! ! !
A+B 1 −1 e 0 1 1 e e−1
e = =
0 1 0 1 0 1 0 1
A.4 Vector and matrix functions

An m × n matrix function of a variable x can be viewed as an m × n matrix,
whose entries are functions of x. An m-dimensional vector function is an
m × 1 matrix function.
Addition and multiplication of matrix functions are defined in an obvious
manner: if A and B are m × n matrix functions of x, then (A + B)(x) =
A(x) + B(x) and if A is an l × m and B an m × n matrix function, then
AB(x) = A(x)B(x).
153
A matrix function is continuous, or differentiable, iff all its entries are
continuous or differentiable, respectively. The derivative dA
dx
or A0 of a differ-
entiable matrix function A is the matrix function whose entries are obtained
by differentiating the entries of A, so
A0ij (x) = (Aij )0 (x)
The integral ab A(x)dx over [a, b] of a continuous matrix function A, is the

R
matrix function whose entry at the position (i, j) is given by

Z b
Aij (x)dx
a
The computational rules for integration and differentiation are easily deduced
from those for ‘ordinary’ (i.e. scalar) functions. Thus, for example, the
derivative of the sum of two differentiable m × n matrix functions A and B
equals the sum of the derivatives:
d dA dB
(A + B) = +
dx dx dx
One should be well aware of the fact that matrix multiplication is noncom-
mutative. If A and B are a differentiable l × m and a differentiable m × n
matrix function, respectively, then the (i, j)th entry of (AB)0 is given by
m m
d X
(AB)0ij = (A0ih Bhj + Aih Bhj
0
) = (A0 B + AB 0 )ij
X
( Aih Bhj ) =
dx h=1 h=1
So, similarly to the product rule for scalar functions, we have
(AB)0 = A0 B + AB 0 ,
However, in general, the factors in each of the terms on the right-hand side
can not be interchanged. For instance, we have
d 2 dA dA
A = A+A ,
dx dx dx
but, in general,
d 2 dA
A 6= 2A ,
dx dx
154
as the following example shows.
!
1 0
A(x) =
x 2
In this case, ! !
d 2 d 1 0 0 0
A (x) = =
dx dx 3x 4 3 0
However,
! ! ! !
dA 1 0 0 0 0 0 dA 0 0
A = = and A=
dx x 2 1 0 2 0 dx 1 0
Exercise. Show that, for every nonsingular k × k matrix function A with

the property that A and A−1 are differentiable, the following identity holds
d −1 dA
A = −A−1 A−1
dx dx
Give an example of a matrix function A such that
d −1 dA
A 6= −A−2
dx dx
155
Index
annihilator method, 91 inhomogeneous, 38

arithmetic-geometric progression, 47 linear, 8, 38
asymptotic behaviour, 13 linearized, 28
asymptotically stable, 14, 24, 30, 49, logistic, 19
55, 58, 67, 73, 74, 94, 130– scalar, 7
134, 138 differential equation
attracting set, 22 autonomous, 101
attractor, 22, 71 linear, 102
autonomous difference equation, 7 logistic, 101
autonomous differential equation, 101 dominant diagonal, 132
dynamic multiplier model, 69
basin of attraction, 22, 23, 71, 77, 79,
138 equilibrium point, 20
bifurcation, 13, 32 equilibrium solution, 10, 65
flip-, 34 Euler’s method, 9
period-doubling, 34
pitchfork-, 33 fixed point, 19
saddle-node, 13, 32 flip-bifurcation, 34
transcritical, 34 fundamental matrix, 57, 59, 120
bifurcation diagram, 13, 34, 35
boundary conditions, 97 general solution, 11
business-cycle model, 83 generalized eigenspace, 146
generalized eigenvector, 146
Casorati determinant, 41, 57 geometric multiplicity, 148
cobweb model, 5, 31 graphical solution of difference eqs.,
coordinate transformation, 143 21, 22
cyclic solution, 10
homogeneous difference equation, 89
difference equation, 7 homogeneous linear difference equa-
autonomous, 7 tion, 38
homogeneous, 38 hyperbolic, 26, 78
156
inhomogeneous difference equation, 44, periodic attractor, 31, 79
90 periodic solution, 65
inhomogeneous linear difference equa- pitchfork bifurcation, 33, 34
tion, 38 positive invariant, 19
initial value problem, 12, 94 positive limit point, 21, 71
initial values, 39 positive limit set, 21
invariant set, 19, 71 pre-iterates, 34
iterate, 19, 71 predator-prey model
continuous, 134
Jordan block, 148 discrete, 74
Jordan normal form, 148
recurrence relation, 12
linear difference equation, 8, 38
linear difference operator, 38 saddle-node bifurcation, 32
linear differential equation, 102 saddle-node-bifurcation, 13
linearization, 28, 74, 134 Samuelson’s multiplier-accelerator model,
linearized difference equation, 28 94
logistic difference equation, 19, 30, 34 separation of variables, 105
logistic differential equation, 101 shift operator, 11
logistic map, 22 solution space, 38, 39, 41, 56, 89
Lyapunov function, 81, 138 spectral radius, 67
stable, 14, 49, 55, 67, 94, 130, 132,
multiplier-accelerator model 138
Samuelson’s, 94 asymptotically, 24, 30, 49, 55, 58,
neoclassical growth model in contin- 67, 73, 74, 94, 130–134, 138
uous time, 102 neutrally, 55, 58, 130, 131
neoclassical growth model in discrete stable set, 22, 68, 71
time, 6 stationary solution, 10
neutrally stable, 14, 55, 58, 130, 131 transcritical bifurcation, 33, 34
Newton’s method, 15
non-linear multiplier-accelerator model,unstable, 14, 24, 30, 49, 55, 58, 67,
84 73, 74, 94, 130–134, 138
null solution, 10
variation of constants (difference equa-
orbit, 20, 71 tions), 46
variation of constants (differential equa-
parameters, 12
tions), 108
period, 10
period-doubling bifurcation, 34
157

Difference and Differential Equations: BA Econometrics and Operations Research

Uploaded by

Copyright:

Available Formats

Difference and Differential Equations: BA Econometrics and Operations Research

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Difference and Differential Equations: BA Econometrics and Operations Research

Uploaded by

Copyright:

Available Formats

Difference and Differential equations

BA Econometrics and Operations Research

Module code EBB812A05

2 Differential equations 101

This is a first order difference equation, with independent variable t and

pt+1 = D−1 ◦ S(pt )

If the price p0 at time t = 0 is known, the prices at subsequent times can be

Example 1.1.2. A neoclassical growth model in discrete time.

We begin by giving a general definition of a difference equation. A (scalar)

F (x, y(x), y(x + h), ..., y(x + kh)) = 0 (1.4)

Here, k is a natural number, the order of the difference equation, x is the

b(x) + a0 (x)y(x) + a1 (x)y(x + h) + ... + ak (x)y(x + kh) = 0

Frequently, the shift is normalized to either 1 or −1. This can be achieved

u(t) := y(x) = y(th)

y(x + jh) = y((t + j)h) = u(t + j) j = 1, ..., k

Substituting this into (1.4), we obtain the equation

F (th, u(t), u(t + 1), ..., u(t + k)) = 0

Example 1.1.3. The equation

u(t) = c + 3ht, where c ∈ R

Hence the solutions of (1.5) are given by the general expression

y(x) = c + 3x, where c ∈ R

F (n, y(n), y(n + 1), ..., y(n + k)) = 0 (1.6)

is some function y, defined on an appropriate subset D of Z, such that (1.6)

F (n, y(n), y(n + 1), ..., y(n + k)) = 0, n∈D (1.7)

Example 1.1.4. The set of all solutions of the difference equation

y(n + 1) − ny(n) = 0, n∈N

y(n + 1) − ny(n) = 0, n ∈ {1, 2, ...} (1.8)

is satisfied by all sequences of the form y(n) = c(n − 1)!, c ∈ R or C.

A function y on a discrete set D ⊂ Z or N generates the sequence {y(n)}n∈D .

Example 1.1.5. The equation

y(n + 1) + ny(n) = (n − 1)(−1)n + 2(n + 1), n ∈ N

Replacing y(n + 2) by y(n) on the left-hand side and rearranging terms, we

The general solution of an equation is the set of all solutions. Even

{y(n) = c(n − 1)!, n ∈ {1, 2, ...} : c ∈ C}, or {y = {c(n − 1)!}∞

is the general (complex) solution of (1.8). In what follows we will usually

τ y(n) := y(n + 1) ( or τ yn = yn+1 )

The inverse of this operator exists if we consider functions (sequences) on Z

τ −1 y(n) = y(n − 1) ( or τ −1 yn = yn−1 )

F (n, yn , yn+1 , ..., yn+k ) = 0

Here, n0 is an integer and f is a function of k + 1 variables. For convenience,

Example 1.1.6. The solution y of the equation

y(n + 2) = y(n)2 + 1, n∈N

with initial values: y(0) = 1, y(1) = 0, can be determined as follows: y(2) =

Many equations arising in economic models, contain, besides the dependent

where a is allowed be any real number, form a 1-parameter family of nonlinear

1.1.1 Asymptotic behaviour and stability

y(n + 1) − y(n) = 0, n∈N

Example 1.1.10. Newton’s method is used to approximate zeroes of func-

y(n + 1) = y(n)2 , n∈N (1.11)

Proof. Suppose, for example, that y∗ is a stable solution of (1.12). We then

|ỹ(n) − y∗ (n)| ≤ δ for n = 0, ..., k − 1, (1.14)

then we have, for all n ≥ 0,

Now, let z̃ be a solution of (1.13) with the property that

|z̃(n) − 0| ≤ δ for n = 0, ..., k − 1

Then the sequence ỹ defined by: ỹ = z̃ + y∗ is a solution of (1.12) satisfying

Example 1.1.12. As we have seen, y∗ ≡ 1 is an unstable solution of the

3. Prove that all solutions of the equation

y(n + 1) = αy(n)(1 − y(n)) (1.16)

Here, α is a positive parameter. Equations of the type (1.15) play a role in

Let > 0. As 0 < d < 1, |f n (x) − c| ≤ for all x ∈ U such that |x − c| ≤ .