Problem set 1

• Consider affine ODE defined as

ẏ = αy + b (0.1)

where b ∈ R and the associated homogeneous equation

ẏ = αy (0.2)

obtained from (0.1) by setting b = 0.

– (superposition principle) show that, if φ1 (t) and φ2 (t) are solutions of (0.2), so is
any linear combination of them λ1 φ1 (t) + λ2 φ2 (t);
– show that, if φ(t) is a solution of equation (0.2) and ψ(t) is a solution of (0.1), then
φ + ψ is also a solution of (0.1);
– setting α = −0.5 and b = −2, rewrite (0.1) in the new state variable Y := y + b/α,
showing that (0.1) turns into a linear ODE; from the general solution of the latter
go back to the general solution of the former; provide graphical representations of
particular solutions of (0.1) corresponding to initial conditions y 0 = −5, −4, −3, 0, 1;
– setting α = −0.5 and b = −2, perform qualitative analysis of (0.1) (that is, find
fixed points and highlight their stability properties by putting arrows in (y, ẏ) plain).

• Perform qualitative analysis of of the following ODE (that is, find fixed points, if any, and
highlight their stability properties by putting arrows in (y, ẏ) plain) choosing alternatively
α = −1, 0, 1

i) ẏ = −y + αy 3
ii) ẏ = αy + y 2
iii) ẏ = y 2 + α

• Definition of Lyapunov Stability

A fixed point y ∗ of an ODE is Lyapunov stable when, for any given l > 0, value d > 0
can be found such that the relation |y0 − y ∗ | < d implies |y(t) − y ∗ | < l for all t ≥ 0

Establish whether each of the following statements is true:

i) a fixed point y ∗ of an ODE is Lyapunov stable when, for any given l > 0, value
d > 0 can be found such that the relation |y0 − y ∗ | < l implies |y(t) − y ∗ | < d;
ii) a fixed point y ∗ of an ODE is not Lyapunov stable when there exist l > 0 and t > 0
such that |y(t) − y ∗ | > l even if, for all d > 0, the relation |y0 − y ∗ | < d holds;

1
 iii) if y ∗ is a fixed point of an ODE that is a Lyapunov stable, a distance l > 0 from
y ∗ does not exists such that, for all initial conditions y 0 , y(t) becomes far from y ∗
more than l at a certain point in time;
iv) if y ∗ is a fixed point of an ODE that is a Lyapunov stable, a distance l > 0 from y ∗
does not exists such that, for some initial conditions y 0 , y(t) becomes far from y ∗
more than l at a certain point in time;
v) a fixed point y ∗ of an ODE is Lyapunov stable when, for any l > 0, some initial
conditions y0 ̸= y ∗ can be found such that y(t) never becomes far from y ∗ more
than l > 0;
vi) a fixed point y ∗ of an ODE is Lyapunov stable when, for any given d > 0, the relation
|y0 − y ∗ | < d implies |y(t) − y ∗ | < l, for some small l > 0;

• Show that the fixed point of ẏ = −y + 3 is Lyapunov stable.

• In the rest of this file, the first chapter of Nonlinear Dynamics, by Medio and Lines is
printed. Solve exercises 1.3, 1.4, 1.6 at the end. For ODEs given in 1.8 try qualitative
analysis and sketch trajectories (in so called extended phase spaces) with different initial
conditions. The same for ODE in 1.10 by setting µ = 3.

• Fixed points of ODE ẏ = sy − y 2 depend on parameter s. Represents such fixed points

in the vertical axis of a cartesian plain where s is in abscissa. How stability varies with s?

2
 1
Statics and dynamics: some elementary concepts

Dynamics is the study of the movement through time of variables such as

heartbeat, temperature, species population, voltage, production, employ-
ment, prices and so forth.
This is often achieved by means of equations linking the values of vari-
ables at different, uniformly spaced instants of time, i.e., difference equa-
tions, or by systems relating the values of variables to their time derivatives,
i.e., ordinary differential equations. Dynamical phenomena can also be
investigated by other types of mathematical representations, such as par-
tial differential equations, lattice maps or cellular automata. In this book,
however, we shall concentrate on the study of systems of difference and
differential equations and their dynamical behaviour.
In the following chapters we shall occasionally use models drawn from
economics to illustrate the main concepts and methods. However, in general,
the mathematical properties of equations will be discussed independently of
their applications.

1.1 A static problem

To provide a first, broad idea of the problems posed by dynamic vis-à-vis
static analysis, we shall now introduce an elementary model that could be
labelled as ‘supply-demand-price interaction in a single market’. Our model
considers the quantities supplied and demanded of a single good, defined as
functions of a single variable, its price, p. In economic parlance, this would
be called partial analysis since the effect of prices and quantities determined
in the markets of all other goods is neglected. It is assumed that the demand
function D(p) is decreasing in p (the lower the price, the greater the amount
that people wish to buy), while the supply function S(p) is increasing in p
(the higher the price, the greater the amount that people wish to supply).

1
2 Statics and dynamics: some elementary concepts

D,S

D( p) = S(p) D¢
S

p p

Fig. 1.1 The static partial equilibrium model

For example, in the simpler, linear case, we have:

D(p) = a − bp
(1.1)
S(p) = −m + sp
and a, b, m and s are positive constants. Only nonnegative values of these
variables are economically meaningful, thus we only consider D, S, p ≥ 0.
The economic equilibrium condition requires that the market of the
good clears, that is demand equals supply, namely:
D(p) = S(p) (1.2)
or
a − bp = −m + sp.

static solution Mathematically, the solution to our problem is the value

of the variable that solves (1.2) (in this particular case, a linear equation).
Solving (1.2) for p we find:
a+m
p̄ =
b+s
where p̄ is usually called the equilibrium price (see figure 1.1).1 We
call the problem static since no reference is made to time or, if you prefer,
1 The demand curve D � in figure 1.1 is provided to make the point that, with no further con-
straints on parameter values, the equilibrium price could imply negative equilibrium quantities
of supply and demand. To eliminate this possibility we further assume that 0 < m/s ≤ a/b, as
is the case for the demand curve D.
 1.2 A discrete-time dynamic problem 3

everything happens at the same time. Notice that, even though the static
model allows us to find the equilibrium price of the good, it tells us nothing
about what happens if the actual price is different from its equilibrium value.

1.2 A discrete-time dynamic problem

The introduction of dynamics into the model requires that we replace the
equilibrium condition (1.2) with some hypothesis concerning the behaviour
of the system off-equilibrium, i.e., when demand and supply are not equal.
For this purpose, we assume the most obvious mechanism of price adjust-
ment: over a certain interval of time, the price increases or decreases in
proportion to the excess of demand over supply, (D − S) (for short, excess
demand). Of course, excess demand can be a positive or a negative quan-
tity. Unless the adjustment is assumed to be instantaneous, prices must
now be dated and pn denotes the price of the good at time n, time being
measured at equal intervals of length h. Formally, we have

pn+h = pn + hθ[D(pn ) − S(pn )]. (1.3)

Since h is the period of time over which the adjustment takes place, θ can
be taken as a measure of the speed of price response to excess demand. For
simplicity, let us choose h = 1, θ = 1. Then we have, making use of the
demand and supply functions (1.1),

pn+1 = a + m + (1 − b − s)pn . (1.4)

In general, a solution of (1.4) is a function of time p(n) (with n taking

discrete, integer values) that satisfies (1.4).2

dynamic solution To obtain the full dynamic solution of (1.4), we begin

by setting α = a + m, β = (1 − b − s) to obtain

pn+1 = α + βpn . (1.5)

To solve (1.5), we first set it in a canonical form, with all time-referenced

terms of the variable on the left hand side (LHS), and all constants on the
right hand side (RHS), thus:

pn+1 − βpn = α. (1.6)

Then we proceed in steps as follows.

2 Weuse the forms pn and p(n) interchangeably, choosing the latter whenever we prefer to
emphasise that p is a function of n.
4 Statics and dynamics: some elementary concepts

step 1 We solve the homogeneous equation, which is formed by setting

the RHS of (1.6) equal to 0, namely:

pn+1 − βpn = 0. (1.7)

It is easy to see that a function of time p(n) satisfying (1.7) is p(n) = Cβ n ,

with C an arbitrary constant. Indeed, substituting in (1.7), we have

Cβ n+1 − βCβ n = Cβ n+1 − Cβ n+1 = 0.

step 2 We find a particular solution of (1.6), assuming that it has a

form similar to the RHS in the general form. Since the latter is a constant,
set p(n) = k, k a constant, and substitute it into (1.6), obtaining

k − βk = α

so that
α a+m
k= = = p̄ again!
1−β b+s
It follows that the p(n) = p̄ is a solution to (1.6) and the constant (or
stationary) solution of the dynamic problem is simply the solution of the
static problem of section 1.1.

step 3 Since (1.6) is linear, the sum of the homogeneous and the particular
solution is again a solution,3 called the general solution. This can be
written as
p(n) = p̄ + Cβ n . (1.8)

The arbitrary constant C can now be expressed in terms of the initial con-
dition. Putting p(0) ≡ p0 , and solving (1.8) for C we have

p0 = p̄ + Cβ 0 = p̄ + C

whence C = p0 − p̄, that is, the difference between the initial and equilibrium
values of p. The general solution can now be re-written as

p(n) = p̄ + (p0 − p̄)β n . (1.9)

Letting n take integer values 1, 2, . . ., from (1.9) we can generate a sequence

of values of p, a ‘history’ of that variable (and consequently, a history of
quantities demanded and supplied at the various prices), once its value at
any arbitrary instant of time is given. Notice that, since the function pn+1 =
3 This is called the superposition principle and is discussed in detail in chapter 2 section 2.1.
 1.2 A discrete-time dynamic problem 5

f (pn ) is invertible, i.e., the function f −1 is well defined, pn−1 = f −1 (pn )

also describes the past history of p.
The value of p at each instant of time is equal to the sum of the equilib-
rium value (the solution to the static problem which is also the particular,
stationary solution) and the initial disequilibrium (p0 − p̄), amplified or
dampened by a factor β n . There are therefore two basic cases:

(i) |β| > 1. Any nonzero deviation from equilibrium is amplified in time,
the equilibrium solution is unstable and as n → +∞, pn asymptotically
tends to +∞ or −∞.
(ii) |β| < 1. Any nonzero deviation is asymptotically reduced to zero,
pn → p̄ as n → +∞ and the equilibrium solution is consequently stable.

First-order, discrete-time equations (where the order is determined as the

difference between the extreme time indices) can also have fluctuating be-
haviour, called improper oscillations,4 owing to the fact that if β < 0,
β n will be positive or negative according to whether n is even or odd. Thus
the sign of the adjusting component of the solution, the second term of the
RHS of (1.9), oscillates accordingly. Improper oscillations are dampened if
β > −1 and explosive if β < −1.
In figure 1.2 we have two representations of the motion of p through time.
In figure 1.2(a) we have a line defined by the solution equation (1.5), and the
bisector passing through the origin which satisfies the equation pn+1 = pn .
The intersection of the two lines corresponds to the constant, equilibrium
solution. To describe the off-equilibrium dynamics of p, we start on the
abscissa from an initial value p0 �= p̄. To find p1 , we move vertically to
the solution line and sidewise horizontally to the ordinate. To find p2 ,
we first reflect the value of p1 by moving horizontally to the bisector and
then vertically to the abscissa. From the point p1 , we repeat the procedure
proposed for p0 (up to the solution line, left to the ordinate), and so on and
so forth. The procedure can be simplified by omitting the intermediate step
and simply moving up to the solution line and sidewise to the bisector, up
again, and so on, as indicated in figure 1.2(a). It is obvious that for |β| < 1,
at each iteration of the procedure the initial deviation from equilibrium
is diminished again, see figure 1.2(b). For example, if β = 0.7, we have
β 2 = 0.49, β 3 = 0.34, . . . , β 10 ≈ 0.03, . . .) and the equilibrium solution is
approached asymptotically.
The reader will notice that stability of the system and the possibility
4 The term improper refers to the fact that in this case oscillations of variables have a ‘kinky’
form that does not properly describe the smoother ups and downs of real variables. We discuss
proper oscillations in chapter 3.
6 Statics and dynamics: some elementary concepts

pn+1 p

p0

p0 p1 p pn 1 2 3 4 5 6 7 time
(a) (b)

Fig. 1.2 Convergence to p̄ in the discrete-time partial equilibrium model

of oscillatory behaviour depends entirely on β and therefore on the two

parameters b and s, these latter denoting respectively the slopes of the
demand and supply curves. The other two parameters of the system, a and
m, determine α and consequently they affect only the equilibrium value
p̄. We can therefore completely describe the dynamic characteristics of the
solution (1.9) over the parameter space (b, s). The boundary between stable
and unstable behaviour is given by |β| = 1, and convergence to equilibrium
is guaranteed for
−1<β <1
2 > b + s > 0.
The assumptions on the demand and supply functions imply that b, s > 0.
Therefore, the stability condition is (b + s) < 2, the stability boundary is
the line (b + s) = 2, as represented in figure 1.3. Next, we define the curve
β = 1 − (b + s) = 0, separating the zone of monotonic behaviour from
that of improper oscillations, which is also represented in figure 1.3. Three
zones are labelled according to the different types of dynamic behaviour,
namely: convergent and monotonic; convergent and oscillatory; divergent
and oscillatory. Since b, s > 0, we never have the case β > 1, corresponding
to divergent, nonoscillatory behaviour.
If |β| > 1 any initial difference (p0 − p̄) is amplified at each step. In this
model, we can have |β| > 1 if and only if β < −1. Instability, then, is due
to overshooting. Any time the actual price is, say, too low and there is
positive excess demand, the adjustment mechanism generates a change in
the price in the ‘right’ direction (the price rises) but the change is too large.
 1.3 A continuous-time dynamic problem 7

s
Convergent, monotonic
2 Convergent, oscillatory
Divergent, oscillatory

1 2 b

Fig. 1.3 Parameter space for the discrete-time partial equilibrium model

After the correction, the new price is too high (negative excess demand) and
the discrepancy from equilibrium is larger than before. A second adjustment
follows, leading to another price that is far too low, and so on. We leave
further study of this case to the exercises at the end of this chapter.

1.3 A continuous-time dynamic problem

We now discuss our simple dynamical model in a continuous-time setting.
Let us consider, again, the price adjustment equation (1.3) (with θ = 1,
h > 0) and let us adopt the notation p(n) so that

p(n + h) = p(n) + h (D[p(n)] − S[p(n)]) .

Dividing this equation throughout by h, we obtain

p(n + h) − p(n)
= D[p(n)] − S[p(n)]
h
whence, taking the limit of the LHS as h → 0, and recalling the definition
of a derivative, we can write
dp(n)
= D[p(n)] − S[p(n)].
dn
Taking the interval h to zero is tantamount to postulating that time is a
continuous variable. To signal that time is being modelled differently we
substitute the time variable n ∈ Z with t ∈ R and denote the value of p at
time t simply by p, using the extended form p(t) when we want to emphasise
that price is a function of time. We also make use of the efficient Newtonian
8 Statics and dynamics: some elementary concepts

notation dx(t)/dt = ẋ to write the price adjustment mechanism as

dp
= ṗ = D(p) − S(p) = (a + m) − (b + s)p. (1.10)
dt
Equation (1.10) is an ordinary differential equation relating the values of
the variable p at a given time t to its first derivative with respect to time
at the same moment. It is ordinary because the solution p(t) is a function
of a single independent variable, time. Partial differential equations, whose
solutions are functions of more than one independent variable, will not be
treated in this book, and when we refer to differential equations we mean
ordinary differential equations.

dynamic solution The dynamic problem is once again that of finding a

function of time p(t) such that (1.10) is satisfied for an arbitrary initial
condition p(0) ≡ p0 .
As in the discrete-time case, we begin by setting the equation in canonical
form, with all terms involving the variable or its time derivatives on the LHS,
and all constants or functions of time (if they exist) on the RHS, thus

ṗ + (b + s)p = a + m. (1.11)

Then we proceed in steps as follows.

step 1 We solve the homogeneous equation, formed by setting the RHS of

(1.11) equal to 0, and obtain

ṗ + (b + s)p = 0 or ṗ = −(b + s)p. (1.12)

If we now integrate (1.12) by separating variables, we have

� �
dp
= −(b + s) dt
p
whence
ln p(t) = −(b + s)t + A

where A is an arbitrary integration constant. Taking now the antilogarithm

of both sides and setting eA = C, we obtain

p(t) = Ce−(b+s)t .
 1.3 A continuous-time dynamic problem 9

step 2 We look for a particular solution to the nonhomogeneous equation

(1.11). The RHS is a constant so we try p = k, k a constant and conse-
quently ṗ = 0. Therefore, we have
ṗ = 0 = (a + m) − (b + s)k
whence
a+m
k= = p̄.
b+s
Once again the solution to the static problem turns out to be a special
(stationary) solution to the corresponding dynamic problem.

step 3 Since (1.12) is linear, the general solution can be found by summing
the particular solution and the solution to the homogeneous equation, thus
p(t) = p̄ + Ce−(b+s)t .
Solving for C in terms of the initial condition, we find
p(0) ≡ p0 = p̄ + C and C = (p0 − p̄).
Finally, the complete solution to (1.10) in terms of time, parameters, initial
and equilibrium values is
p(t) = p̄ + (p0 − p̄)e−(b+s)t . (1.13)
As in the discrete case, the solution (1.13) can be interpreted as the sum
of the equilibrium value and the initial deviation of the price variable from
equilibrium, amplified or dampened by the term e−(b+s)t . Notice that in
the continuous-time case, a solution to a differential equation ṗ = f (p)
always determines both the future and the past history of the variable p,
independently of whether the function f is invertible or not. In general, we
can have two main cases, namely:
(i) (b + s) > 0 Deviations from equilibrium tend asymptotically to zero as
t → +∞.
(ii) (b + s) < 0 Deviations become indefinitely large as t → +∞ (or, equiv-
alently, deviations tend to zero as t → −∞).
Given the assumptions on the demand and supply functions, and therefore
on b and s, the explosive case is excluded for this model. If the initial price
is below its equilibrium value, the adjustment process ensures that the price
increases towards it, if the initial price is above equilibrium, the price de-
clines to it. (There can be no overshooting in the continuous-time case.) In
a manner analogous to the procedure for difference equations, the equilibria
10 Statics and dynamics: some elementary concepts

.
p p
p

30

20

10

0
p p 5 10 15
(a) (b) time

Fig. 1.4 The continuous-time partial equilibrium model

of differential equations can be determined graphically in the plane (p, ṗ)

as suggested in figure 1.4(a). Equilibria are found at points of intersection
of the line defined by (1.10) and the abscissa, where ṗ = 0. Convergence
to equilibrium from an initial value different from the equilibrium value is
shown in figure 1.4(b).
Is convergence likely for more general economic models of price adjust-
ment, where other goods and income as well as substitution effects are taken
into consideration? A comprehensive discussion of these and other related
microeconomic issues is out of the question in this book. However, in the
appendixes to chapter 3, which are devoted to a more systematic study of
stability in economic models, we shall take up again the question of conver-
gence to or divergence from economic equilibrium.
We would like to emphasise once again the difference between the discrete-
time and the continuous-time formulation of a seemingly identical problem,
represented by the two equations

pn+1 − pn = (a + m) − (b + s)pn (1.4)

ṗ = (a + m) − (b + s)p. (1.10)

Whereas in the latter case (b+s) > 0 is a sufficient (and necessary) condition
for convergence to equilibrium, stability of (1.4) requires that 0 < (b+s) < 2,
a tighter condition.
This simple fact should make the reader aware that a naive translation of
a model from discrete to continuous time or vice versa may have unsuspected
consequences for the dynamical behaviour of the solutions.
 1.4 Flows and maps 11

1.4 Flows and maps

To move from the elementary ideas and examples considered so far to a

more general and systematic treatment of the matter, we need an appro-
priate mathematical framework, which we introduce in this section. When
necessary, the most important ideas and methods will be discussed in greater
detail in the following chapters. For the sake of presentation, we shall begin
with continuous-time systems of differential equations, which typically take
the canonical form

dx
= ẋ = f (x) (1.14)
dt

where f is a function with domain U , an open subset of Rm , and range

Rm (denoted by f : U → Rm ). The vector5 x = (x1 , x2 , . . . , xm )T denotes
the physical variables to be studied, or some appropriate transformations
of them; t ∈ R indicates time. The variables xi are sometimes called ‘de-
pendent variables’ whereas t is called the ‘independent variable’.
Equation (1.14) is called autonomous when the function f does not
depend on time directly, but only through the state variable x. In this
book we shall be mainly concerned with this type of equation, but in our
discussions of stability in chapters 3 and 4 we shall have something to say
about nonautonomous equations as well.
The space Rm , or an appropriate subspace of dependent variables — that
is, variables whose values specify the state of the system — is referred to
as the state space. It is also known as the phase space or, sometimes,
the configuration space, but we will use only the first term. Although
for most of the problems encountered in this book the state space is the
Euclidean space, we occasionally discuss dynamical systems different from
Rm , such as the unit circle. The circle is a one-dimensional object embedded
in a two-dimensional Euclidean space. It is perhaps the simplest example
of a kind of set called manifold. Roughly speaking, a manifold is a set
which locally looks like a piece of Rm . A more precise definition is deferred
to appendix C of chapter 3, p. 98.
In simple, low-dimensional graphical representations of the state space
the direction of motion through time is usually indicated by arrows pointing
to the future. The enlarged space in which the time variable is explicitly

5 Recallthat the transposition operator, or transpose, designated by T , when applied to a

row vector, returns a column vector and vice versa. When applied to a matrix, the operator
interchanges rows and columns. Unless otherwise indicated, vectors are column vectors.
12 Statics and dynamics: some elementary concepts

considered is called the space of motions. Schematically, we have

R × Rm = R1+m
↓ ↓ ↓
time state space of
space motions
The function f defining the differential equation (1.14) is also called a
vector field, because it assigns to each point x ∈ U a velocity vector f (x).
A solution of (1.14) is often written as a function x(t), where x : I → Rm
and I is an interval of R. If we want to specifically emphasise the solution
that, at the initial time t0 , passes through the initial point x0 , we can write
x(t; t0 , x0 ), where x(t0 ; t0 , x0 ) = x(t0 ) = x0 . We follow the practice of setting
t0 = 0 when dealing with autonomous systems whose dynamical properties
do not depend on the choice of initial time.
remark 1.1 In applications, we sometimes encounter differential equations
of the form
dm x � dx dm−1 x �
= F x, , . . . , x∈R (1.15)
dtm dt dtm−1
where dk x/dtk denotes the kth derivative of x with respect to time. Equa-
tion (1.15) is an autonomous, ordinary differential equation of order m,
where m is the highest order of differentiation with respect to time appear-
ing in the equation. It can always be put into the canonical form (1.14) by
introducing an appropriate number of auxiliary variables. Specifically, put
dk x
= zk+1 , 0≤k ≤m−1
dtk
(where, for k = 0, dk x/dtk = x) so that
żk = zk+1 , 1≤k ≤m−1
żm = F (z1 , . . . , zm ).

If we now denote by z ∈ Rm the m-dimensional vector (z1 , . . . , zm )T we can

write:
ż = f (z)
where f (z) = [z2 , . . . , zm , F (z1 , . . . , zm )]T .
We can also think of solutions of differential equations in a different man-
ner which is now prevalent in dynamical system theory and will be very
helpful for understanding some of the concepts discussed in the following
chapters.
 1.4 Flows and maps 13

If we denote by φt (x) = φ(t, x) the state in Rm reached by the system

at time t starting from x, then the totality of solutions of (1.14) can be
represented by a one-parameter family of maps6 φt : U → Rm satisfying
�
d �
[φ(t, x)] �� = f [φ(τ, x)]
dt t=τ

for all x ∈ U and for all τ ∈ I for which the solution is defined.
The family of maps φt (x) = φ(t, x) is called the flow (or the flow map)
generated by the vector field f . If f is continuously differentiable (that
is, if all the functions in the vector are continuously differentiable), then
for any point x0 in the domain U there exists a δ(x0 ) > 0 such that the
solution φ(t, x0 ) through that point exists and is unique for |t| < δ. The
existence and uniqueness result is local in time in the sense that δ need not
extend to (plus or minus) infinity and certain vector fields have solutions
that ‘explode’ in finite time (see exercise 1.8(c) at the end of the chapter).
When the solution of a system of differential equations ẋ = f (x) is not
defined for all time, a new system ẋ = g(x) can be determined which has
the same forward and backward orbits in the state space and such that each
orbit is defined for all time. If ψ(t, x) is the flow generated by the vector field
g, the relation between ψ and the flow φ generated by f is the following:

ψ(t, x) = φ[τ (t, x), x] x∈U

and
τ :R×U →R

is a time-reparametrisation function monotonically increasing in t for all

x ∈ U.

example Suppose we have a system

ẋ = f (x) (1.16)

with f : Rm → Rm , a continuously differentiable function with flow φ(t, x)

defined on a maximal time interval −∞ < a < 0 < b < +∞. Then the
6 The terms map or mapping indicate a function. In this case, we speak of y = f (x) as the
image of x under the map f . If f is invertible, we can define the inverse function f −1 , that
is, the function satisfying f −1 [f (x)] = x for all x in the domain of f and f [f −1 (y)] = y for all
y in the domain of f −1 . Even if f is not invertible, the notation f −1 (y) makes sense: it is the
set of pre-images of y, that is, all points x such that f (x) = y. The terms map, mapping
are especially common in the theory of dynamical systems where iterates of a map are used to
describe the evolution of a variable in time.
14 Statics and dynamics: some elementary concepts

y y

time

x
(a) (b) x

Fig. 1.5 A damped oscillator in R2 : (a) space of motions; (b) state space

differential equation ẋ = g(x) with g(x) : Rm → Rm and

f (x)
g(x) =
1 + �f (x)�
(where � · � denotes Euclidean norm), defines a dynamical system whose
forward and backward orbits are the same as those of (1.16) but whose
solutions are defined for all time.7

The set of points {φ(t, x0 ) | t ∈ I} defines an orbit of (1.14), starting from

a given point x0 . It is a solution curve in the state space, parametrised by
time. The set {[t, φ(t, x0 )] | t ∈ I} is a trajectory of (1.14) and it evolves
in the space of motions. However, in applications, the terms orbit and
trajectory are often used as synonyms. A simple example of a trajectory in
the space of motions R × R2 and the corresponding orbit in the state space
R2 is given in figure 1.5. Clearly the orbit is obtained by projecting the
trajectory onto the state space.
The flows generated by vector fields form a very important subset of a
more general class of maps, characterised by the following definition.
definition 1.1 A flow is a map φ : I ⊂ R × X → X where X is a metric
space, that is, a space endowed with a distance function, and φ has the
following properties
(a) φ(0, x) = x for every x ∈ X (identity axiom);
(b) φ(t + s, x) = φ[s, φ(t, x)] = φ[t, φ(s, x)] = φ(s + t, x), that is, time-
translated solutions remain solutions;
7 For details see Bhatia and Szegö (1970), p. 78; Robinson (1999), pp. 146–7.
 1.4 Flows and maps 15

(c) for fixed t, φt is a homeomorphism on X.

Alternatively, and equivalently, a flow may be defined as a one-parameter

family of maps φt : X → X such that the properties (a)–(c) above hold for
all t, s ∈ R.

remark 1.2 A distance on a space X (or, a metric on X) is a function

X × X → R+ satisfying the following properties for all x, y ∈ X:

(1) d(x, y) ≥ 0 and d(x, y) = 0 if and only if x = y;

(2) d(x, y) = d(y, x) (symmetry);
(3) d(x, y) ≤ d(x, z) + d(z, y) (triangle inequality).

Notice that there also exist notions of distance which are perfectly mean-
ingful but do not satisfy the definition above and therefore do not define a
metric, for example:

the distance between a point and a set A;

d(x, A) = inf d(x, y).

y∈A

the distance between two sets A and B

d(A, B) = inf inf d(x, y).

x∈A y∈B

Neither of these cases satisfies property (1) in remark 1.2. However, there
exists a ‘true’ distance between two sets which is a metric in the space of
nonempty, compact sets, i.e., the Hausdorff distance.8

In this book we are mainly concerned with applications for which φ is a

flow generated by a system of differential equations and the state space is
an Euclidean space or, sometimes, a manifold. However, some concepts and
results in later chapters of the book will be formulated more generally in
terms of flows on metric spaces.
Consider now a system of nonautonomous differential equations such that

ẋ = f (t, x) (1.17)

where f : R × U → Rm , and assume that a unique solution exists for all

(t0 , x0 ) ∈ R × U . Then we can represent solutions of (1.17) by means of
a flow φ : R × X → X, where X ⊂ (R × Rm ). This suggests that a
8 See, for example, Edgar (1990), pp. 65–6.
16 Statics and dynamics: some elementary concepts

nonautonomous system ẋ = f (t, x) can be transformed into an equivalent

autonomous system by introducing an arbitrary variable θ = t and writing
θ̇ = 1
(1.18)
ẋ = f (θ, x).
Notice that, by definition, the extended autonomous system (1.18) has no
equilibrium point in X. However, if the original, nonautonomous system
(1.17) has a uniformly stable (uniformly, asymptotically stable) equilibrium
point, then for the extended autonomous system (1.17), the t-axis is a stable
(asymptotically stable) invariant set. The precise meaning of (asymptotic,
uniform) stability will be discussed in chapters 3 and 4.
Solutions of system (1.14) can be written in either the simpler form x(t),
x : I → Rm , or φt (x) : U → Rm , or again φ(t, x), φ : I × U → Rm ,
depending on what aspect of solutions one wants to emphasise. The notation
φt (x) is especially suitable for discussing discrete-time maps derived from
continuous-time systems.
If time t is allowed to take only uniformly sampled, discrete values, sep-
arated by a fixed interval τ , from a continuous-time flow we can derive a
discrete-time map (a difference equation)
xn+τ = G(xn ) (1.19)
where G = φτ . Certain properties of continuous-time dynamical systems
are preserved by this transformation and can be studied by considering the
discrete-time systems derived from them. If the unit of measure of time is
chosen so that τ = 1, we have the canonical form
xn+1 = G(xn ). (1.20)
Let the symbol ◦ denote the composition of functions, so that, f ◦ g(x)
means f [g(x)]. Then we write
xn = G(xn−1 ) = G ◦ G(xn−2 ) = . . . = G ◦ G ◦ . . . ◦ G(x0 ) = Gn (x0 )
where Gn is the composition of G with itself n times, or the nth iteration
of G, with n ∈ Z+ . If G is invertible and G−1 is a well defined function, Gn
with n ∈ Z− denotes the nth iterate of G−1 . (Note that Gn (x) is not the
nth power of G(x).) Thus, iterates of the map G (or G−1 ) can be used to
determine the value of the variable x at time n, when the initial condition
x0 is fixed.9
9 For autonomous difference equations whose solutions do not depend on the choice of the initial
time, in a manner analogous to our practice for autonomous differential equations, we take the
initial time as zero.
 1.4 Flows and maps 17

remark 1.3 There exists another way of deriving a discrete-time map from
a continuous-time dynamical system, called Poincaré map, which de-
scribes the sequence of positions of a system generated by the intersections
of an orbit in continuous time and a given space with a lower dimension,
called surface of section. Clearly, in this case the time intervals between
different pairs of states of the systems need not be equal. Poincaré maps
are a powerful method of investigation of dynamical systems and we shall
make some use of them in chapter 4, when we discuss periodic solutions and
in chapters 7 and 8.

Of course, there exist problems that are conceived from the beginning
as discrete dynamical systems (difference equations). In fact, there are
difference equations that cannot be derived from differential equations. In
particular, this is true of noninvertible maps which have been extensively
used in recent years in the study of dynamical problems in many applied
fields. Intuitively, the reason why a noninvertible map cannot be a flow map
(derived from a differential equation as explained above) is that such a map
uniquely determines the dynamics in one time direction only whereas, under
standard assumptions, solutions of a differential equation always determine
the dynamics in both directions uniquely.

remark 1.4 Orbits of differential equations are continuous curves, while

orbits of maps are discrete sets of points. This has a number of important
consequences, the most important of which can be appreciated intuitively.
If the solution of an autonomous system of differential equations through a
point is unique, two solution curves cannot intersect one another in the state
space. It follows that, for continuous-time dynamical systems of dimension
one and two, the orbit structure must be drastically constrained. In the
former, simpler case, we can only have fixed points and orbits leading to (or
away from) them; in the two-dimensional case, nothing more complex than
periodic orbits can occur. For maps the situation is different. It remains true
that the orbit starting from a given point in space is uniquely determined in
the direction defined by the map. However, since discrete-time orbits, so to
speak, can ‘jump around’, even simple, one-dimensional nonlinear maps can
generate very complicated orbits, as we shall see in the following chapters.

Generalising the simple examples discussed in sections 1.2 and 1.3 above,
the stationary, equilibrium solutions of multi-dimensional dynamical sys-
tems in both continuous and discrete time can be identified by solving sys-
tems of equations.
In the former case, setting ẋ = 0 in (1.14) the set of equilibrium or
18 Statics and dynamics: some elementary concepts

fixed points is defined by

E = {x̄|f (x̄) = 0}
that is, the set of values of x such that its rate of change in time is nil.
Analogously, in the discrete-time case,
xn+1 = G(xn )
we have
E = {x̄|x̄ − G(x̄) = 0}
that is, the set of values of x that are mapped to themselves by G. Because
the functions f and G are generally nonlinear, there are no ready-made
procedures to find the equilibrium solutions exactly, although geometrical
and numerical techniques often give us all the qualitative information we
need. Notice that linear systems typically have a unique solution, whereas
nonlinear systems typically have either no solutions, or a finite number of
them. It follows that only nonlinear systems may describe the interesting
phenomenon of (finite) multiple equilibria.
For a system of autonomous, differential equations like (1.14), a general
solution φ(t, x) can seldom be written in a closed form, i.e., as a combi-
nation of known elementary functions (powers, exponentials, logarithms,
sines, cosines, etc.). Unfortunately, closed-form solutions are available only
for special cases, namely: systems of linear differential equations; one-
dimensional differential equations (i.e., those for which m = 1); certain
rather special classes of nonlinear differential equations of order greater than
one (or systems of equations with m > 1). The generality of nonlinear sys-
tems which are studied in applications escapes full analytical investigation,
that is to say, an exact mathematical description of solution orbits cannot
be found. Analogous difficulties arise when dynamical systems are repre-
sented by means of nonlinear maps. In this case, too, closed-form solutions
are generally available only for linear systems.
The importance of this point should not be exaggerated. On the one
hand, even when a closed-form solution exists, it may not be very use-
ful. A handbook of mathematical formulae will typically have a hundred
pages of integrals for specific functions, so that a given nonlinear model
may indeed have a solution. However, that solution may not provide much
intuition, nor much information if the solution function is not a common,
well known function. On the other hand, in many practical cases we are
not especially interested in determining (or approximating) exact individual
solutions, but we want to establish the qualitative properties of an ensemble
 Exercises 19

of orbits starting from certain practically relevant sets of initial conditions.

These properties can often be investigated effectively by a combination of
mathematical, geometrical, statistical and numerical methods. Much of
what follows is dedicated precisely to the study of some of those methods.
Before turning to this goal, however, we review in chapter 2 the class of
dynamical systems which is best understood: linear systems. Dynamical lin-
ear systems in both continuous and discrete time are not terribly interesting
per se because their behaviour is morphologically rather limited and they
cannot be used effectively to represent cyclical or complex dynamics. How-
ever, linear theory is an extremely useful tool in the analysis of nonlinear
systems. For example, it can be employed to investigate qualitatively their
local behaviour, e.g., their behaviour in a neighbourhood of a single point
or of a periodic orbit. This is particularly important in stability analysis
(chapters 3 and 4) and in the study of (local) bifurcations (chapter 5).

Exercises
1.1 Consider the discrete-time partial equilibrium model summarised in
(1.6) given the parameter values a = 10, b = 0.2, m = 2, s = 0.1.
Write the general solution given the initial values p0 = 20 and p0 =
100. Calculate the values for the price at time periods 0, 1, 2, 4,
10, 100 starting from each of the above initial values and sketch the
trajectories for time periods 0–10.
1.2 State a parameter configuration for the discrete-time partial equi-
librium model that implies β < 0. Describe the dynamics implied
by that choice. Using these parameter values and a = 10, m = 2,
sketch the dynamics in the space (pn , pn+1 ). Draw the bisector line
and from the chosen initial condition, iterate 3 or 4 times. Show the
direction of movement with arrows.
1.3 If we define the parameters as in exercise 1.1 (b = 0.2, s = 0.1,
a = 10, m = 2), the continuous-time, partial equilibrium model of
(1.11) gives the constant exponent of the solution as b + s = 0.3. Let
this be case 1. If s = 0.6, b+s = 0.8. Let this be case 2. Calculate the
solution values for case 1 and case 2 at periods t = 0, 1, 2, 4.67, 10, 100
starting from the initial condition p0 = 20. Comment on the speed
of the adjustment process. Note the different integer values of t for
which equilibrium in Case 2 is approximated using a precision of 1
decimal point, 2 decimal points, 3 decimal points.
1.4 Suppose that the good under consideration is a ‘Giffen’ good (for
which dD/dp > 0 and therefore b < 0). It is unlikely, but possible
20 Statics and dynamics: some elementary concepts

that b + s < 0. Sketch the differential equation (1.10) under that hy-
pothesis in the (p, ṗ) plane, note the equilibrium point and comment
on the adjustment process.
1.5 Convert these higher-order differential equations to systems of first-
order differential equations and write the resulting systems in matrix
form:
(a) ẍ + x = 1
3
(b) ddt3x + 0.4ẍ − 2x = 0
4
(c) ddt4x + 4ẍ − 0.5ẋ − x = 11.
1.6 Convert the following higher-order system of differential equations
into a system of first-order differential equations
ẍ + x = 1
ÿ − ẏ − y = −1.
1.7 Higher-order difference equations and systems can also be converted
to first-order systems using auxiliary variables. A kth-order equation
xn+k = G(xn+k−1 , . . . , xn ) can be converted by setting
xn = zn(1)
(1)
zn+1 = xn+1 = zn(2)
(2)
zn+1 = xn+2 = zn(3)
.. .. ..
. . .
(k)
zn+1 = xn+k = G(xn+k−1 , . . . , xn ) = G(zn(k) , . . . , zn(1) ).
Convert the following difference equations into systems of first-order
difference equations and write them in matrix form
(a) xn+2 − axn+1 + bxn = 1
(b) 0.5xn+3 + 2xn+1 − 0.1xn = 2.
1.8 Use integration techniques to find exact solutions to the following
differential equations and sketch trajectories where possible, assum-
ing an initial value of x(0) = 1
(a) ẋ = 2x
(b) ẋ = x12
(c) ẋ = x2 .
1.9 Use the technique described in the example in section 1.4 to find
a function g, defined over all time and such that ẋ = g(x) has the
same backward and forward orbits in the state space as ẋ = x2 .
 Exercises 21

1.10 Write the exact solution of the following differential equation (Hint:
rewrite the equation as dx/dt = µx(1 − x) and integrate, separating
variables) and discuss the dynamics of x
ẋ = µx(1 − x) x ∈ [0, 1].