Lecture notes 2013/2014

Dynamical Systems, 1
Mihaela Cistelecan, UT Cluj-Napoca

Motivation

Definition:
Process Modeling - addresses the following two related problems:
(1) - how to obtain the equations (= MODEL) describing certain phenomena / process
(2) - how to interpret the results of analyzing the phenomena within the
framework of the model.
We will focus only on the second part of the definition (often called the
theory of dynamical systems) - see the motivation below.
Process Modeling is rather used as a book title than a title for a subject
of study in Control Engineering Departments. This is because building
the mathematical model of a dynamical system (in particular, process) is
possible in the following two ways: 1). using the laws of physics or 2). using
different algorithms classified under the name of System Identification.
The first way of modeling is the mission of people with expertise in physics
(scientists). The second way of modeling is the responsibility of control
engineers.
For the following two reasons:
1). I do not have expertise in physics and, besides, there are no typical
rules to guide us in writing the physical laws;
2). System Identification is a subject you will study later because it
requires as a prerequisite the Systems Theory - subject that you will study
in the next semester;
the lecture notes will deal with problems concerning the Dynamical
systems not Process Modeling (in the sense of System Identification).
Still, we will use the keyword process model very often. More often
than process model we will use the more general concept of (dynamical)
system model.

We emphasize the fact that we assume the model as being given. We

then study the behavior of the given mathematical model and identify the
important phenomena it faces.
Several models for dynamical systems taken from the literature will be
provided.

Oportunity

Mathematical models are only approximations of reality. It is well-known

that the keyword dynamical systems is an alias for differential and difference equations. However, a dynamical system may be a model for a real
physical system (or process model) only if it fulfills certain conditions. You
will learn about these conditions (and more) from my lecture notes.
Which is the difference between the Dynamical Systems lecture and
the System Theory lecture ? - We have chosen to proceed by low dimensional nonlinear dynamical systems equations, and not linear high-dimensional
systems. Equations in dimension one and two constitute the majority of the
text. We explain the basic notions of stability and bifurcations of vector
fields using scalar and second-order autonomous equations.

How to learn and understand the concepts of

this lecture

Do not try to memorize anything from the notes I give you. Instead, try
to understand each new concept introduced by notes. I will be happy to
answer your questions (related to my notes) in the lab-classes.
Each of the given Home- and Lab- work assignment has as final goal the
writing of an essay, which will contribute to the final grade.
For the year 2009/ 2010 the final grade is computed as follows:
(oral + written exam) /2
or, for those presenting a case study (completed by individual analysis)
(oral + written exam + case study) / 3
The oral examination is based on the mini-projects I give you every two
or three weeks. You are required to submit your reports at the specified
deadline. If you miss the deadline your mark is automatically decreased
each week by one. A five-week delay for at least one report means you are
not qualified (eligible) to be examined in the first legal examination session.
There are only two legal examination sessions per year. I do not organize

special sessions or examinations, please do not pay for special examinations

!!!!!! That exams are not legal.

Dynamical systems

4.1

u/x/y - input/ state/ output representation (model)

For control engineers a dynamical system is often modeled by a finite number

of coupled first-order ordinary differential equations:
x 1 = f1 (t, x1 , ..., xn , u1 , ..., up )
...
x n = fn (t, x1 , ..., xn , u1 , ..., up )

(1)

where x i denotes the derivative of xi with respect to the time variable t, and
u1 , u2 , ...,un are specified input variables. We call the variables x1 , x2 , ...,
xn the state variables.
We usually use vector notation to write the above equations in a compact
form, as follows:
x = f (t, x, u); t R; x Rn ; u Rp

(2)

where:

x=

x1
x2
...
xn

u=

u1
u2
...
un

f (t, x, u) =

f1 (t, x, u)
f2 (t, x, u)
...
fn (t, x, u)

(3)

We call (2) the state equation, and refer to x as the state and u as the
input.
Sometimes, with (2) we associate the output equation
y = h(t, x, u); y Rq
Remark:

(4)

The output vector comprises variables of particular interest

in
(a). the analysis of the dynamical system, like variables which are physically measured or,
(b). variables which are required to behave in a specified manner
(a) and (b) will be investigated in the following semesters in the Systems
theory, Control Theory, Optimal / robust / adaptive control classes.
3

The state equation (2) together with the output equation (4) gives the
so-called STATE-SPACE MODEL of the dynamical process / system.
Mathematical models of finite-dimensional physical systems do not always
come in the form of a state-space model. However, more often than not,
we can model physical systems in this form by carefully choosing the state
variables.
We will often use the unforced state equation:
x = f (t, x)

(5)

Working with an unforced state equation does not necessarily mean that
the input to the system is zero. It could be that the input has been specified
as:
1. a given function of time, u = (t), or
2. a given feedback function of the state, u = (x), or
3. both, u = (t, x)
Substitution of u = in (2) eliminates u and yields an unforced state
equation.
The order of the dynamical system (3) is n.
We aims at studying differential equations with emphasis on some specific problems, such as:
1. special dynamics modeling some real, physical phenomena/ systems
/ plants/ benchmarks;
2. special dynamics that could help us find a good control (feedback)
law u = .

4.2

u/y - input/output models (representation)

In order to keep the exposition simple, we assume that the p = q = 1.

A mathematical model that describes a wide variety of physical nonlinear
systems is the n-th order differential equation:
dn y
dy dn1 y
dm u
...
=
g(t,
y,
,
u
...
)
,
dtn
dt dtn1
dtm

(6)

The order of the system is n.

Example: Pendulum equation: Using Newtons second law of motion we
can write the equation of motion in the tangential direction as:
ml = mgsin kl
where:
4

(7)

 l - denotes the length of the rod (the rod is rigid and has zero mass)
m - denotes the mass of the bob
g - the acceleration due to gravity
k - coefficient of friction
- the angle subtended by the rod and the vertical axis through the
pivot point
Equation (7) is the u/y unforced model of the pendulum (with u = 0).
The deviation (angle) is the output.
To obtain a state-space model of the pendulum, let us take the state
Then, the state equation is:
variables as x1 = and x2 = .
x 1 = x2
x 2 = gl sin x1

k
m x2

(8)

and the output equation is y = x1 .

If we can apply a torque T to the pendulum, this torque may be viewed
as a control input in the equation:
x 2 =

4.3

x 1 = x2
k
x1 m
x2 +

gl sin

1
T
ml2

(9)

equivalence u/y u/x/y

For a dynamical system the representations u/y and u/x/y are equivalent.
It is possible to transform one representantion into the other - this will be
studied in the Systems theory class.
Remark: Note that (7) and (8) describe the same dynamical system the pendulum.
-

4.4

Linear vs. Nonlinear Dynamical Systems

The definition of a linear system is based on the the superposition principle,

see.
http : //en.wikipedia.org/wiki/Superposition principle, which will be
studied in the Systems Theory class. Still, you may recognize a nonlinear
system when you see nonlinearities, for example: x2 , x1 x2 , sin(x), sign(x),
sat(x) (saturation), etc.

4.5

Autonomous vs. Nonautonomous Dynamical Systems

When the function f does not depend explicitly on t, that is:

x = f (x)

(10)

the system is called autonomous or time-invariant . The behavior of an

autonomous system is invariant to shifts in the time origin, since changing
the time variable from t to = t a does not change the right-hand side of
the state equation. If the system is not autonomous - see eq. (2) - then it
is called nonautonomous or time varying .
Note that sin(t) in f (x) denotes an non-autonomous system

4.6

Parameter-dependent dynamics, Bifurcations

Many times the model of a physical system depends on some parameters, :

x = f (x, ), = g(t, )

(11)

The bifurcation theory is the study of possible changes in the structure of

the orbits of a differential equation as parameters are varied. Bifurcation
behaviour of specific differential equations can be encapsulated in certain
pictures called bifurcation diagrams.

First order dynamical systems (given by first

order differential eq.)

The solutions of the differential equations/ systems are important because

they say everything about the systems behavior. Some differential equations
have analytical solutions. If an analytical solution cannot be found there
are two possibilities: 1. the equation is integrated using the Euler or RungeKutta method - what we get is the solution particularized by an initial state;
2. we make a qualitative analysis of the system/ equation dynamics basically
by studying the direction of change of the solution in different points.
In the following, we study:
1. Chapter 1 from:
Hirsch, Smale, Devaney - Differential Equations, Dynamical Systems
and An Introduction to Chaos
2. The Euler integration method from:
http : //en.wikipedia.org/wiki/Euler0 s method

Home-works

From the studied chapter Exercises: 1, 2, 3, 4, 5, 7, 10

Lab-works

1. Solve some ordinary differential equations using dsolve from Matlab.

2. Integrate the equations given in Chapter 1 using: 1). the Euler
method, 2) the function ode23 from Matlab, (Runge-Kutta method), 3) the
Runge-Kutta method explicitly written, see
http : //en.wikipedia.org/wiki/Runge Kutta method
3. Draw the slope field for the equations give in Chapter 1 using the
function quiver from Matlab.
4. Simulate the following first-order scalar difference equation:
xn+1 = xn (1 xn )

(12)

where is a positive real parameter

MATLAB

1. To solve the linear nonhomogeneous equation x = 5x + t3 type

>>eq=Dx=5*x+t3 and use the command dsolve as follows:
>>dsolve(eq)
To solve an initial value problem for the above equation, specify the
initial condition:
>> dsolve(eq,x(1)=1)
2. Numerical solutions: we will apply Eulers method to solve the initial
2
value problem x = 2tx + et , x(0) = 1, in the interval [0, 1.5], with step size
h = 0.1. First create the file called f1.m and type into it
function xp=f1(t,x)
xp=2 .*t .*x+ exp( t .* t);
Then, create another file called f2.m and type into it
t0=0;x0=1;
h=0.1;
n=15;
t=t0; x=x0; T=t; X=x;
for i=1:n
x=x+h*f1(t,x);
7

t=t+h; T=[T;t]; X=[X;x];

end
[T,X]
pause
plot(T,X)
Return to Matlab and type f2.

3. To draw the solution of the following equation

x = x(1 x)

(13)

for initial condition x(0) = 0.2 we proceed as follows. In a file called, say
f.m write the above equation as follows:
function xp=f(t,x)
xp=x - x .*x;
Then type in Matlab command window (or write in another file and
invoke the file name):
[t,x]=ode23(f,[0, 20],[0.2]);
plot(t,x)
We have taken the variable t from 0 to 20.
4. In order to draw the slope field for the system:
x = x (1 x);
use the Matlab code:
t1=0; t2=10; x1=-2; x2=2;
axis([t1 t2 x1 x2]);
[t,x]=meshgrid(t1:0.5:t2, x1:0.5:x2);
dt=t;
dx=2 *x .*(1-x);
quiver(t,x,dt,dx)

(14)

Lecture notes 2013/2014

Dynamical Systems, 2
Mihaela Cistelecan, UT Cluj-Napoca

Examples of dynamical systems

Example 1: A field-controled DC motor can be described by:

di

vf = Rf if + Lf dtf
va = c1 if + La didta + Ra ia
J d
dt = c2 if ia c3

(1)

The first equation is for the field circuit with vf , if , Rf , and Lf being its
voltage, current, resistance, and inductance. The variables va , ia , Ra , and
La are the corresponding variables for the armature circuit described by the
second equation. The third equation is a torque eq. for the shaft, with J as
the rotor inertia and c3 as damping coefficient. The term c1 if is the back
e.m.f induced in the armature circuit, and c2 if ia is the torque produced
by the interaction of the armature current with the field circuit flux. The
voltage va is held constant, and the voltage vf is the control input.
Example 2: The nonlinear dynamic equations for an m-link robot take
the form:
M (q)
q + C(q, q)
q + g(q) = u

(2)

where q is an m-dimensional vector of generalized coordinates representing

joint positions, u is an m-dimensional control (torque) input, M (q) is a
symmetric inertia matrix which is positive definite q Rm . The C(q, q)

accounts for centrifugal and Coriolis forces. The term g(q) accounts for
gravity forces.
Example 3: The logistic equation
p = p( p)

(3)

where and are constants is used in medicine and biology for simple
population and epiodemiological models.
9

Example 4: The eq.

x
+ bx + kx = sin(t)

(4)

where b, k, and are constants, models a simple electric circuit or the

motion of a damped spring with a periodic forcing term.
Example 5: The equation
x = kx
(5)
models growth or decay problems. At a certain time, the rate at which a
given quantity is changing is proportional to the amount existing at that
time.
Example 6: A special form of van der Pols equation is
x
+ (x2 1)x + x = cost

(6)

where , and are constants. This equation models an electric circuit

with a triode, the resistive properties of which change with the current.
Example 7: The Lorenz equations
x = (y x)
y = x y xz
z = z + xy

(7)

where , , > 0 are constants has important applications in meteorology.

Example 8: The Duffings equation
x
+ x x + x3 = cost

(8)

is used in mechanical engineering, for ex., in modeling the action of leaf

springs on trucks.
Example 9: The Newtons law of cooling or heating
T = k(T )

(9)

where k and are constants, models the cooling or heating of a body immersed in a medium of constant temperature . T is the unknown temperature function of the body and k is a constant depending on the body. The
rate at which the temperature of the body is changing is proportional to the
difference between the temperatures of the medium and the body.
Example 10: Competing species
In an isolated region of the Canadian Northwest Territories, a population
of white wolves, x, and one of silver foxes, y, compete for survival (For each
10

population, one unit represents 100 individuals). They have a common,

limited food supply, which consists mainly of mice. A model that describes
the dynamics of these competing species is given by the nonlinear system:
x = x x2 xy
y = 34 y y 2 12 xy

(10)

In other words, each population increases with its size, but decreases with
the square of its size and with the product of its own size and the size of the
other species. The proportionality factors were obtained statistically.

Second order (planar) autonomous systems

These notes complete the Chapter 1, pp. 19-56 from Khalil:

Nonlinear systems which is taught in class
A second order autonomous system is represented by two scalar differential equations
x 1 = f1 (x1 , x2 )
x 2 = f2 (x1 , x2 )

(11)

A linear second order autonomous system is represented by two scalar

differential equations:
x 1 = a11 x1 + a12 x2
(12)
x 2 = a21 x1 + a22 x2

2.1

Motivation for studying Linear, Autonomous, Unforced

Dynamics

We study in detail the linear, unforced, second order systems because of the
reasons given below.
1. Understanding the qualitative behavior of solutions of a
system x = f (x)
A geometrical study in the state/ phase plane of a dynamical system allows a qualitative study of the dynamics. This means that for the dynamical
system x = f (x), x = [x1 , x2 ]T , f (x) = [f1 (x), f2 (x)]T the flow (t, x(0))
can be studied from a geometric point of view. At each point of the plane,
f (x) gives a value of the derivative dx/dt which can be considered as the
slope of a line segment at that point. The collection of all such line segments
is called the direction field of the differential equation. The graph of the solution through x(0), represented in the phase/ state plane - the projection

11

of the curve through x(0), (t, (t, x(0))), on the plane (x1 , x2 ) - is called
phase / state trajectory or orbit. At each point through which t trajectory
passes, the trajectory is tangent to a line segment of the direction field of
the differential equation. We can assign to each point x the directed line
segment from x to x + f (x). The collection of all such vectors is called the
vector field f . To compensate for the loss of time parametrization in orbits
we insert arrows to indicate the direction in which (t, x(0)) is changing as t
increases. The flow of a differential equation is drawn as the collection of all
orbits together with the direction arrows - the resulting picture is called the
phase portrait of the differential equation. Certain distinguished orbits play
an important role in the qualitative theory of planar systems. The simplest
of such orbit is an equilibrium point. In planar systems there can be another
orbit of special interest, called a periodic orbit.
Equilibria play an important role in understanding the qualitative behavior of solutions.
Definition 1 : An equilibrium solution of equation x
= f (x, x)
is a constant function x that cancels the vector field f .
Definition 2: A point x
is called an equilibrium point (also critical point,
steady state solution) of x = f (x) if f (
x) = 0. For a second order system,
the equilibrium point x
satisfy the equations: fi (
x) = 0, i = 1, 2.
Example: To find the equilibrium solutions of the equation
x
= x + x2 x

(13)

solve the algebraic equation x2 x = 0 which yields the solutions x1 = 0 and

x2 = 1. The equilibrium solutions are thus the constant functions 1 (t) = 0
and 2 (t) = 1. In the phase plane they correspond to the points (0, 0) and
(1,0).
Usually, when we refer to the state space of the autonomous system
x = f (x), since the vector x(t) = [x1 (t), x2 (t), ...xn (t)]T is seen as a point
moving in the state space, an equilibrium point is seen as a fixed point
x
= [
x1 , x
2 , ...
xn ]T in space. When we refer to the time history of the
solutions x(t) = c (constant) of a differential system of equations we see
signals which are constant in time or a steady state solution of the system.
Unlike linear systems which have one equilibrium, nonlinear systems
can have more than one equilibrium or none at all. There is an important
difference between the equilibria of a linear system and those of a nonlinear
system. Whereas the former determine the global behavior of solutions, the
latter influence the qualitative picture only locally.
Definition 3: A solution (t, x(0)) of x = f (x) is called a periodic
12

solution of period T , with T > 0, if:

(t + T, x(0)) = (t, x(0))

(14)

for all t R. The minimal period T is that period with the property that
(t, x(0)) 6= x(0) for 0 < t < T . The orbit (t, x(0)) of a periodic solution
with period T is said to be a periodic orbit (also closed orbit) of period T .
A periodic orbit in the (x1 , x2 )-plane is a closed curve. It may be a center
or a limit cycle (only in the case of nonlinear systems). A periodic orbit
is called a limit cycle if there is a solution that tends to it from the interior
region, the exterior origin, or both, when t .
2. The Linear Control Problems reduce to Linear Systems A
linear control system (linear, autonomous, forced dynamics) in the form:
x = Ax + Bu,

x Rn

(15)

admits a control input in the form of a state feedback control, as follows:

u = Fx

(16)

Therefore, the dynamics of the control system is as follows:

x = (A + BF )x = Ad x

(17)

When the desired dynamics is given by Ad the matrix F is estimated

through simple matrix computations.
2. Nonlinear dynamics is usually linearized
Most nonlinear systems, in the neighborhood of equilibrium solutions
behave like the linear system that approximates them. In order to study
nonlinear systems we do the following: 1). shift the origin of the state space
to an equilibrium and, then, 2). linearize the new system around the origin.
See section 2- Nonlinear second-order dynamics.
3. Gradient and Hamiltonian systems
Many natural phenomena are modeled using Gradient and Hamiltonian
systems. A two-dimensional system of the form:
x = f (x, y)
y = g(x, y)

(18)

is called a:
Gradient system - if there is a real function G of variables x and y,
called gradient function that has continuous partial derivatives and satisfies
the relations:
f (x, y) = G
x (x, y)
(19)
G
g(x, y) = y (x, y)
13

These conditions imply that:

f
y

2G
yx

which means that:

g
x

2G
xy

f
g
=
y
x

(20)

(21)

To see if a given two dimensional system is a gradient system, we can either find a function G that satisfies conditions (19) or check whether relation
(21) is satisfied.
The geometric interpretation of a gradient system is that G is always
decreasing along the solutions of (18), and therefore it can be thought of as a
potential function of (18). A particle starting on the surface z = G(x1 , x2 )
moves downhill under the flow defined by (18).
Hamiltonian system - if there is a real function H of variables x and
y, called Hamiltonian function that has continuous partial derivatives and
satisfies the relations:
f (x, y) = H
y (x, y)
(22)
g(x, y) = H
x (x, y)
These conditions imply that
f
x

which means that

2H
xy ,

g
y

2H
yx

f
g
=
x
y

(23)

(24)

To see whether a given two-dimensional system is Hamiltonian, we can either

find a function H that satisfies conditions (22) or check whether relation (24)
is satisfied.

2.2

Jordan normal forms

One could compute eAt from its power series given as follows:
eAt =

+
X

1 n n
1
A t = I + At + A2 t2 + ...
n!
2!
n=0

(25)

In general, it is impractical to sum series except for special coefficient

matrices. Using linear algebra the computation of eAt for any coefficient

14

matrix A can be reduced, by a linear change of coordinates, to one of the

canonical forms

1 0
0 2

0
0

(26)

where , 1 , 2 and 6= 0 are real numbers. These canonical forms are

called Jordan normal forms. The exponentials of these canonical forms are:

eAt =

2.3

e1 t 0
0
e2 t

eAt = et

1 t
0 t

eAt = et

cost
sint
sint cost
(27)

Linear Algebra helps solving systems of linear differential equations

Explanation that follows completes the information from the book chapter
2 by Khalil - Nonlinear systems. You may also read the book chapter
II by Hirsch, Smale, Devaney Differential Equations, Dynamical Systems
and an introduction to chaos.
For the linear dynamics x = Ax the solution results in the form x(t) =
eAt x(0). Lets assume that the matrix A is not in Jordan form. Consider
the new variables z so that z = M 1 x. Then, in the new coordinates the
linear system x = Ax becomes
z = M 1 AM z

(28)

The flow in the new coordinates is as follows:

x(t) = M z(t) = M eM

1 AM t

z(0) = (M eM

and thus
eAt = M eM
or, equivalently,
eM

1 AM t

1 AM t

1 AM t

M 1 )x(0)

(29)

M 1

(30)

= M 1 eAt M

(31)

The question now is to find the matrix M so that the new coefficient
matrix M 1 AM becomes one of the canonical matrices given above
We show in the following that M 1 AM = J (J-one of the three canonical
matrices) if M = [v1 v2 ] where v1 , v2 are the eigenvectors of the matrix A.
Case 1 , 2 R - real distinct eigenvalues
15

In this case the relations

Avi = i vi

(32)

i = 1, 2 are true. In order to understand the equality (32) please refer to a

mathematical textbook.
If we compute AM we get:
AM = [Av1 Av2 ] = [1 v1 2 v2 ] = M J
where

J=

1 0
0 2

(33)

(34)

If we multiply (33) by M 1 we obtain:

AM = J =

1 0
0 2

(35)

Case 1 = 2 = R - equal eigenvalues

In this case the following relation is true:
(A I)v2 = v1

(36)

In order to understand the equality (36) please refer to a mathematical

textbook. It follows:

AM = M

[Av1 Av2 ] = M

[v1 v1 +v2 ] = M

MJ = J =

0
0
(37)

Case = + i, 6= 0 - complex eigenvalues

In this case the following relation is true:
A(v1 + iv2 ) = ( + i)(v1 + iv2 )

(38)

In order to understand the equality (36) please refer to a mathematical

textbook. It follows:

M AM = M [Av1 Av2 ] = [v1 v2 v1 +v2 ] = M M J =

(39)
Remark . From x = M z, M = [v1 v2 ] it turns out that x1 and x2 are
linear combinations of z1 and z2 :
1

x1 = m11 z1 + m12 z2
x2 = m21 z1 + m22 z2
16

(40)

In the (z1 , z2 )-plane the axes x1 = 0 and x2 = 0 are estimated from (40) as
the new lines z1 = (m12 /m11 )z2 and z1 = (m22 /m21 )z2 .
The eigenvectors v1 and v2 in the (x1 , x2 )-plane are lines through the
points (0,0) and (m11 , m21 ) and respectively (0, 0) and (m12 , m22 ). The
eigenvectors v1 and v2 in the (z1 , z2 )-plane fall on z1 = 0 and z2 = 0 because after replacing [z1 , z2 ] in (40) by [1, 0] and [0, 1] we obtain [x1 , x2 ]T =
[m11 , m21 ]T = v1 and [x1 , x2 ]T = [m12 , m22 ]T = v2 .

2.4

u/x/y model u/y model

The u/y model for an unforced second-order system is usually given as

follows:
x
+ 20 x + 02 x = 0
(41)
where 0 is called undamped natural frequency and is called damping
ratio. The solution is:
x(t) = c1 e1 t + c2 e2 t
(42)
where c1 and c2 are constants that depend on the initial conditions and
q

1 , 2 = 0 j0 1 2

(43)

The equivalent u/x/y model, using the phase variables x = x1 , x = x2 ,

is as follows:
"

x 1
x 2

"

0
1
02 20

#"

x1
x2

"

=A

x1
x2

(44)

where the matrix A is not in Jordan form. Therefore, in order to draw the
flow or phase portrait for the system (41) we use the procedure given for
u/x/y models.
Remark 1: The slope of the phase trajectory depends on the system
parameters 0 and , as follows:
x 2
dx2
0 x1 + 2x2
=
= 0
x 1
dx1
x2

(45)

Remark 2: The advantage of having the phase plane representation for

1
a dynamical system is as follows. Since dx
1 = x2 it turns out
x 1 = dt and x
that:
R x (t)
R
(46)
dt = x12 dx1 x11(t0 ) x12 dx1 = tt0 d = t t0
where t t0 is the time required for translating the state x(t) from x(t0 ) to
x(t).
17

Nonlinear second-order (planar) systems

Most nonlinear systems, in the neighborhood of equilibrium solutions behave

like the linear system that approximates them. In order to study nonlinear
systems we do the following: 1). shift the origin of the state space to an
equilibrium and, then, 2). linearize the new system around 0.
The autonomous nonlinear system:
x = F (x),

x Rn

(47)

has the equilibrium solution x

given by F (x) = 0.
a). We shift the origin of the state space to x
, by a change of variables:
y =xx

(48)

In the new coordinates, y, system (47) takes the form:

y = G(y)

(49)

where G is the new form of the vector field , and the origin is an equilibrium, i.e. G(0) = 0.
Example 1: The nonlinear system
x 1 = (x1 3)(x2 1)
x 2 = (x1 + 2)(x2 + 5)

(50)

has (2, 1) as an isolated equilibrium. The change of variables that shifts

the origin of the frame to the equilibrium is:
y1 = x1 + 2
y2 = x2 1

(51)

y 1 = 5y2 + y1 y2
y 2 = 6y1 + y1 y2

(52)

The new system is as follows:

That is, G1 (y1 , y2 ) = 5y2 +y1 y2 , G2 (y1 , y2 ) = 6y1 +y1 y2 , and G1 (0, 0) =
G2 (0, 0) = 0. Thus, in (y1 , y2 )-coordinates, (0, 0) is the equilibrium corresponding to (2, 1) in (x1 , x2 )-coordinates.
b). We find the matrix A describing the linear approximation y = Ay
of the nonlinear system y = G(y) by computing the partial derivatives of
G with respect to the components of the vector y at the equilibrium point.
18

This comes from the Taylor series expansion about the origin. For example,
if G = (G1 , G2 ) and y = (y1 , y2 ), then:
A=

G
1
y1 (0, 0)
G2
y1 (0, 0)

G1
y2 (0, 0)
G2
y2 (0, 0)

(53)

Example 1, revisited For the system given by (50) the linearization is as

follows:
G
!
!
G1
1
0 5
y1 (0, 0)
y2 (0, 0)
A=
=
(54)
G2
G2
6 0
y1 (0, 0)
y2 (0, 0)
Definition 4: An equilibrium point of the system x = f (x) is called
hyperbolic (node, focus, saddle) or nonhyperbolic , respectively, if the corresponding equilibrium of the linearized system is hyperbolic (node, focus,
saddle) or nonhyperbolic.
If x
is an isolated equilibrium (no other equilibrium point in the neighborhood) of the nonlinear system x = f (x) which correspond to the equilibrium
(0, 0) of the linearized system x = Ax, then:
(a) if (0,0) is a hyperbolic equilibrium (node, focus, saddle) of the linear
system, the solutions of the nonlinear system near x
resemble the behavior
of the solutions of the linear system near (0,0); in this case the flow of the
linearized system is structurally stable and therefore a slight perturbation
due to nonlinearities does not affect the qualitative properties.
(b) if (0,0) is a nonhyperbolic equilibrium (center or at least one eigenvalue is zero) for the linear system, no conclusion can be drawn about the
behavior of the flow of the nonlinear system near x
; in this case the system
is not structurally stable and a slight perturbation which necessarily occurs
due to the nonlinear terms may break the configuration of a center into a
focus or it can keep it as it is - for this reason no general conclusion may be
drawn.
For details, see Chapter by Khalil - Nonlinear Systems.

Home-works

1. Determine the nature of the equilibria of the following linear systems:

x = x y
y = x + y

u = 2u 7s
s = 8u + 9s

= r
r =

(55)

2. Show that the following system

x 1 = ax1
x 1 = bx2
19

(56)

a, b R, has the geometric interpretation of a gradient system for the function:

1
1
F (x1 , x2 ) = ax12 bx22
(57)
2
2
Discussion on a, b.
3. For a linear system with A = [5 2; 0 1] and B = [1; 1] estimate
the state feedback matrix F so that the desired dynamics x = Ad x reaches
the equilibrium faster than x = Ax. Simulate the two dynamics x = Ax
and x = Ad x. Compare their dynamics.
4. Choose a matrix A with complex, stable, eigenvalues and solve the
previous problem.
5. Compute eAt for the matrices A given in the canonical forms (26)
using the formula (25) (Show that (27) is true).
Hint:
a) recognize the series of et in the matrix An when matrix A is diagonal
b) use the fact that if two matrices A and B commute, that is, AB=BA,
then e(A+B)t = eAt eBt
c).use the formula e(A+B)t = eAt eBt , for matrices:

A=

0 1
0 0

B=

0
0
1 0

(58)

and for n 2, An = B n = 0.
d). The power series expansions of eAt and eBt are:

eAt

= I + At =

1 t
0 1

eBt

1
0
t 1

(59)

e)

eAt eBt

e(A+B)t = e
f).
sin t =

0
1
1 0

1 t2 t
t
1

P+

n t2n+1
n=0 (1) (2n+1)!

cos t
sin t
sint t cos t

cos t =

P+

n t2n
n=0 (1) (2n)!

(60)

(61)

6. For the system given in (50) compute the matrix A for the rest of the
equilibrium points.
20

7. Assume that the dynamics of two competing species is described by

the system:
x = x x2 xy
(62)
y = 12 y 14 y 2 34 xy
Can the two species survive together ? (Hint: Study the nature of the
equilibrium points).
8. Assume that the dynamics of two competing species is described by
the system:
x = 4x x2 xy
y = 2y ay 2 bxy

(63)

where a and b are positive constants. For what values of a and b can the
two species survive together ?

Lab-works

1. For each of the examples from Section 1 choose some constants and do
numerical simulations.
2. Draw the phase portraits and indicate the direction of the flows of
the following linear system: x = Ax where A is chosen from the following
set:

(a) :

0 1
1 0

(b) :

0 1
1 0

(c) :

1 1
0 1

(d) :

0 1
0 0

(e) :
(64)

Explain the difference between (a) and (b).

3. Consider the linear system x 1 = x1 , x 2 = x2 . Transform this system
to new coordinates z = M 1 x for various choices of M below:

2 0
0 2

0 1
1 0

1 1
1 1

1 1
2 2

(65)

and sketch the phase portraits in these new coordinate systems.

4. Draw the phase portraits for the system x = Ax where:

1
0

0
0

0
0

0
1
!
0
0
!
1
0

1
0

1
0

0
1
21

0
1
!
0
0
!
1
0

1
0

1
0

0
1
!
0
0

(66)

1 1
0 1

5. Draw a) the direction fields and b). the flows (phase portrait) of the
systems:
(

x = y
(a)
y = 3y x
(
x = y
(a)
y = y 2x

x = y
(b)
y = 2y 2x
(
x = y
(a)
y = 2y + x

x = y
y = 4y x
(
x = y
(a)
y = x

(a)

(67)

Explain the equilibria types.

6. Draw the phase portraits of the linear system x 1 = x1 +x2 , x 2 = x2
for = 1, = 0, and = 1. Explain the equilibria types.
7. For a linear system, study the change of the nature of equilibria for
small perturbations on the matrix A. (Hint: use poly from Matlab)
8. For the system given in (56)-(57) draw the graphs of the surface
z = F (x1 , x2 ) for several choices of a and b. Study the motion of a particle on
the surface for the cases: a < 0, b < 0,a < 0, b < 0,a < 0, b > 0,a > 0, b > 0
and relate these cases to the types of the equilibrium of (56). (Hint: use
mesh from Matlab)
9. Simulate the system (50) (for a set of different initial conditions)
and its linear approximates. Draw the flow of the system (50) and of the
linearized systems near origin.
10. Consider the planar system :
x 1 = x2 + x1 (1 x21 x22 )
x2 = x1 + x2 (1 x21 x22 )

(68)

Draw the phase portrait. Transform the system using polar coordinates
(r, ): x1 = r cos , x2 = r sin , r2 = x21 + x22 . Draw the phase portrait.
Study the equilibrium points in both planes.
11. Simulate the following second-order scalar difference equation:
yn+1 = yn (1 yn1 )

(69)

where is a positive real parameter

12. Simulate the dynamical system which is described by the following
first-order planar difference equation:
x1n+1 = xn2
x2n+1 = xn2 (1 xn1 )

22

(70)

or, equivalently, by the following iteration of the map:

x1
x2

x2
x2 (1 x1 )

(71)

where will be chosen as a positive real parameter. Please try more than
one set of initial conditions.

Optional Assignments

1. For each of the following systems determine the equilibria and draw them
in the phase space:
x = x2 y 2
y = 2x 3y

r = r 2rp
p = 0.5r2 p

= sin r
r = r

(72)

2. From the chapter by Khalil - Nonlinear systems: Exercises: 1.7/pp.49,

1.13/52, 1.14/52, 1.17/54, 1.18/54, 1.22/55.

MATLAB

1. To draw a few solutions in the phase plane for the system

x = x(1 y)
y = 3y(x 1)

(73)

for initial conditions x(0), y(0) between 0 and 3, proceed as follows. In a

file called, say f.m write the above system as follows:
function xp=f(t,x)
xp=x;
y=x(2); x=x(1);
xp(1)=x - x .*y;
xp(2)=-3 .*y+ 3 .* x .* y ;
The flow can be produced with the following program,
>> for n=0:6
[t,x]=ode23(f,[0, 20],[n, n]);
plot(x(:,1),x(:,2),k), hold on
end
in which we have taken initial conditions of the form x(0) = y(0) = n
for n from 1 to 6. We have also taken the variable t from 0 to 20, to make
sure that the orbits are continued for a sufficiently long interval. The hold
23

on command is placed such that it tells Matlab to stop the drawing until all
values on n are taken.
2. For linear systems, use initial, as follows. Type the matrices A, B,
C, D in Matlab. For example:
>> A=[-2,-4;5,-6]; B=[1;2];C=[1 1]; D=0;
Then, use the functions ss, initial and plot as follows:
>> sys=ss(A,B,C,D);
>> x0=[10; -5]
>> t=0:0.1:20
>> sys=ss(A,B,C,D);
>> [Y,T,X] = initial(sys,x0, t)
>> subplot(2,1,1),plot(X(:,1),X(:,2))
>> subplot(2,1,2),plot(T,X(:,1),r,T,X(:,2),b)
3. In order to draw the vector field for the system:
x = y
y = y + x2 x

(74)

use the Matlab code:

x1=-1; x2=2; y1=-1; y2=2;
axis([x1 x2 y1 y2]);
[x,y]=meshgrid(x1:0.2:x2, y1:0.2:y2);
dx=y;
dy=y+ x. 2-x;
quiver(x,y,dx,dy)
4. To obtain the eigenvalues of a matrix A use the Matlab command:
>> eig(A)
To obtain both the eigenvalues and the eigenvalues of a matrix A use
the Matlab command:
>> [V,D]=eig(A)
5. Using mesh - Examples from MATLAB.
Produce a combination mesh and contour plot of the peaks surface:
>>[X,Y] = meshgrid(-3:.125:3);
>>Z = peaks(X,Y);
>> meshc(X,Y,Z);
>> axis([-33 -3 3 -10 5])
4. Use poly in order to get the coefficients of a polynomial associated
to an equation with imposed roots.

24

Lecture notes 2013/2014

Dynamical systems, 3
Mihaela Cistelecan, UT Cluj-Napoca
These notes were prepared using:
1. Hale, J.K., Kocak, H. Dynamics and Bifurcations. Springer-Verlag,
1991.
2. Khalil, H. Nonlinear Systems. Prentice Hall. 1996.

1
1.1

The Stability of the equilibrium points. Lyapunov stability

Motivation for studying

1) In physical systems, energy is a descriptor of the system dynamical behavior

The pendulum mathematical model:
x 1 = x2
x 2 = gl sin x1

k
x
m 2

(1)

has two equilibrium points: (0, 0) and (, 0). We study the stability of the
equilibrium point at the origin. The energy of the pendulum E(x) as the sum
of its potential and kinetic energies, with reference of the potential energy
chosen such that E(0) = 0, is:
Z x1
g
1
g
1
E(x) =
sin y dy + x22 = (1 cos x1 ) + x22
(2)
l
2
l
2
0
When friction is neglected (k = 0):
dE
g
g
g
= x 1 sin x1 + x2 x 2 = x2 sin x1 x2 sin x1 = 0
dt
l
l
l

25

(3)

Since for k = 0 the derivative dE

= 0, it follows that the system is conservadt
tive, that is there is no dissipation of energy. Hence, E(x)=const. during the
motion of the system or, dE
= 0 along the trajectories of the system. Since
dt
E(x) = c forms a closed contour around x = 0, for small c, we can arrive at
the conclusion that x = 0 is a stable equilibrium point.
When friction is accounted for (k > 0), the energy is as follows:
dE
g
k
= x 1 sin x1 + x2 x 2 = x22 0
dt
l
m

(4)

Therefore, the energy will dissipate during the motion of the system, that
is dE
0 along the trajectories of the system. Due to friction, E cannot
dt
remain constant indefinitely while the system is in motion. Hence, it keeps
decreasing until it eventually reaches zero, showing that the trajectory tends
to x = 0 as t tends to .
Thus, by examining the derivative of E along the trajectories of the system,
it is possible to determine the stability of the equilibrium point. In 1892,
Lyapunov showed that certain other functions could be used instead of energy to determine the stability of an equilibrium point.
2) When liniarization does not suffice
Consider the system of nonlinear diff eq. :
x 1 = x2 + x1 (x21 + x22 )
x 2 = x1 + x2 (x21 + x22 )

(5)

where is a scalar parameter. These equations are a nonlinear perturbation

of the harmonic oscillator but the perturbation does not affect the linear
terms. For all values of , the origin is an equilibrium point and the linearized
eq. at the origin is, of course, the harmonic oscillator:

0
1
x =
x
(6)
1 0
The eigenvalues of this system are i, which have zero real parts. To analyze
the behaviour of the nonlinear system (5) we compute the derivative of the
square of the distance of a solution from the origin:
d 2
(x + x22 ) = 2(x12 + x22 )2
dt 1
26

(7)

If < 0, then kx(t)k2 approaches zero monotonically as t +. Thus, the

equilibrium at the origin is asymptotically stable. However, if > 0, then
all solutions of (5) with initial data x(0) 6= 0 escape to ininity. Therefore, in
this case, the origin is unstable.

1.2

Stability - definitions

In this section we give considerations and theorems concerning n-dimensional

systems, not only second order systems.
Definition 1.3.1: An equilibrium point is stable if all solutions starting at
nearby points stay nearby; otherwise, it is unstable. The equilibrium point
is asymptotically stable if all solutions starting at nearby points not only stay
nearby, but also tend to the eq. point as time approaches infinity.
Definition 1.3.2: An equilibrium solution x of the system x = f (x) is
called stable if for any neighborhood U of x there is a neighborhood V of x
contained in U such that for any solution , with (0) in V , (t) belongs to
U for any t 0. The equilibrium x is called unstable if it is not stable. An
equilibrium solution x of system x = f (x) is called asymptotically stable
if it is stable and if there is a neighborhood V0 contained in V such that any
solution , with (0) in V0 , tends to x when t .
Linear systems
Theorem 1: If all the eigenvalues of the coefficient matrix A in the linear
system x = Ax have negative real parts, then its equilibrium point x = 0 is
asymptotically stable. Moreover, there are positive constants K and such
that:
kx(t)k = keAt x(0)k Ket kx(0)k

(8)

for all t 0, x(0) R2 .

If one of the eigenvalues of the coefficient matrix A has positive real part,
then the equilibrium point x = 0 is unstable.
Theorem 2: Let f be a continuous function. If all the eigenvalues of the
Jacobian matrix Df (
x) have negative real parts, then the equilibrium point
x of the differential equation x = f (x) is asymptotically stable.
27

1.3

Lyapunov stability

When linearization fails we may apply the Lyapunovs Method.This is a

method that helps us decide whether an equilibrium is stable, asymptotically
stable or unstable. It was given by Lyapunov in 1892. The idea is to find
a function whose properties determine the nature of equilibrium. Unfortunately, there is no algorithm for finding the desired function. Lyapunovs
method only ensures that if such a function exists, we can draw the right
conclusion. The right function should be guessed.
Consider the autonomous system x = f (x) where f : D Rn , D Rn .
Let V : D R be a continuously differentiable function defined in a domain
D Rn that contains the origin. The derivative of V along the trajectories
x = f (x), denoted by V (x), is given by:

n
n
f
(x)
1
X
X
V

V
V
V
... = V f (x)
... x
V (x) =
x i =
fi (x) = x
1
1
xi
xi
x
i=1
i=1
fn (x)
(9)
The derivative of V along the trajectories of a system is dependent on the
systems equations. Hence V (x) will be different for different systems. If
(t, x) is the solution of x = f (x) that starts at initial state x at time t = 0,
then:
d
V (x) = V ((t, x))|t=0
(10)
dt
Therefore, if V (x) is negative, V will decrease along the solution of
x = f (x).
3.1. Lyapunovs stability theorem: Let x = 0 be an equilibrium
point for x = f (x) and D Rn be domain containing x = 0. Let V : D R
be a continuously differentiable function, such that:
V (0) = 0 and V (x) > 0 in D {0}
V (x) 0 in D

(11)

Then, x = 0 is stable. Moreover, if V (x) < 0 in D {0} then x = 0 is

asymptotically stable. The function V is called the Lyapunov function.
Definition 1.3.3: The surface V (x) = c, for some c > 0, is called a
Lyapunov surface or a level surface .

28

The condition V 0 implies that when a trajectory crosses a Lyapunov

surface V (x) = c, it moves inside the set c = {x Rn |V (x) c} and can
never come out again. When V < 0, the trajectory moves from one Lyapunov
surface to an inner Lyapunov surface with a smaller c. As c decreases, the
Lyapunov surface V (x) = c shrinks to the origin, showing that the trajectory
approaches the origin as time progresses.
Positive / Negative definite / semidefinite function
Definition 1.3.4: A function V (x) satisfying the conditions
V (0) = 0
V (x) > 0 for x 6= 0
is said to be positive definite . If it satisfies the weaker condition V (x) 0
for x 6= 0) it is said to be positive semidefinite .
Definition 1.3.5: A function V (x) is said to be negative definite or
negative semidefinite if V (x) is positive definite or positive semidefinite,
respectively. If V (x) does not have a definite sign, it is said to be indefinite.
With this terminology, we can rephrase Lyapunovs theorem , to say
that: the origin is stable if there is a continuously differentiable
positive definite function V (x) so that V (x) is negative semidefinite,
and it is asymptotically stable if V (x) is negative definite.
A class of scalar functions V (x) for which sign definiteness
can be easily checked is the class of functions of the quadratic
form:
n X
n
X
V (x) = xT P x =
pij xi xj
(12)
i=1 j=1

where P is a real symmetric matrix.

In this case, V (x) is positive definite (positive semidef.) if and only if all the
eigenvalues of P are positive (nonnegative), which is true if and only if all
the leading principal minors of P are positive (all principal minors of P are
nonnegative).
If V (x) = xT P x is positive definite (positive semidef.) we say that the
matrix P is positive def. (positive semidef.) and write P > 0 (P 0).

29

Example: Linear system stabilityFor a linear autonomous system x = Ax

we can choose the Lyapunov function as V (x) = xT P x where P = P T . Then,
V (x) = xT (AT P + P A)x = xT Qx. If we can find P > 0 and Q 0 so that
AT P + P A = Q, then the system x = Ax is stable. If, Q > 0 then the
system is asymptotically stable.
Example: Control law designFor a linear control system x = Ax + Bu
we can use the Lyapunov function to design the control law u = F x (state
feedback control). We can choose the Lyapunov function as V (x) = xT P x
where P = P T . Then, V (x) = xT (AT P +P A+F T B T P +P BF )x = xT Qx.
If we can find the matrices F , P > 0 and Q 0 so that AT P +P A+F T B T P +
P BF = Q, then the system x = (A + BF )x is stable. If, Q > 0 then the
system is asymptotically stable.
Conclusion: Important
Lyapunovs theorem can be applied without solving the diff. eq. x =
f (x)
There is no systematic method for finding Lyapunov functions. In
some cases, there are natural Lyapunov functions like energy functions
in electrical or mechanical systems. In other cases, it is a matter of
trial and error.
The Lyapunovs Theorem conditions are only suffficient. Failure of a
Lyapunov function to satisfy the conditions for stability or asymptotic
stability does not mean that the equilibrium is not stable or asymptotically stable. It only means that such stability property cannot be
established by using this Lyapunov function... further investigation is
necessary.
The variable gradient method
Example: Pendulum,revisited From (4) we see that for the Lyapunov
function chosen as V (x) = E(x) the origin is stable but not asymptotically
stable. However, it is known that for the pendulum with k 6= 0 the origin
is asymptotically stable. We show in the following that by choosing the
Lyapunov function as follows the term x22 /2 is replaced by the more general
quadratic form xT P x/2, P > 0:

1
g
p11 p12
x1
V (x) = [ x1 x2 ]
+ (1 cos x1 )
(13)
p12 pp22
x2
2
l
30

For the quadratic form to be positive definite, the elements of the matrix
P must satisfy
2
p11 > 0, p22 > 0, p11 p22 p12
>0
(14)
The derivative V (x) is as follows:
g
g
k
k
V (x) = (1p22 )x2 sin x1 p12 x1 sin x1 +(p11 p12 )x1 x2 +(p12 p22 )x22
l
l
m
m
(15)

We can choose p11 , p12 and p22 such that V (x) < 0. Since the terms
x2 sinx1 and x1 x2 are sign indefinite, we will cancel them by taking p22 = 1
and p11 (k/m)p12 . With these choices, from (14), p12 must satisfy 0 < p12 <
(k/m). Let us take p12 = 0.5(k/m). Then, V (x) is given by:
1g k
1k 2
V (x) =
x1 sinx1
x
(16)
2lm
2m 2
The term x1 sinx1 > 0 for all 0 < |x1 | < . Taking D = {x R2 | |x| < }
we see that V (x) is positive definite and V (x) is negative definite over D.
Thus, we conclude that the origin is asymptotically stable.
The procedure that exploits the idea given in the above example (we
investigated the expression for V (x) and went back to choose the parameters
of V (x) so as to make V (x) negative) is known as the variable gradient method
. We briefly describe this procedure in the following.
Lets denote g(x) = V = (V /x)T . The derivative V (x) along the
trajectories of x = f (x) is given by:
V
V (x) =
f (x) = g T (x)f (x)
x

(17)

The idea now is to try to choose g(x) such that it would be the gradient
of a positive definite function V (x), and at the same time, V (x) would be
negative definite.
See the literature for further information...
3.4. Domain / region of attraction
The domain / region of attraction (or basin, or region of asymptotic stability) is defined as follows.
Let (t; x) be the solution of x = f (x) that starts at initial state x at
time t = 0. Then, the region of attraction is defined as the set of all points
x such that limt (t; x) = 0.

31

Finding the exact region of attraction analytically might be difficult or even

impossible. Still, a conservative approximation may be found using the Lyapunov function, that is: if there is a Lyapunov function that satisfies the conditions of asymptotic stability over a domain D, and if c = {x Rn |V (x)
c} is bounded and contained in D, then every trajectory starting in c remains in c and approaches the origin as t . Thus, c is an estimate
of the region of attraction. This estimate may be conservative - it may be
much smaller than the actual region of attraction.
Example: Consider the second-order dif. eq. :
z + 2az + z + z 3 = 0

(18)

where a is a constant satisfying 0 < a < 1, which is equivalent to the

system:
x 1 = x2
x 2 = x1 2ax2 x13

(19)

The origin is the only equilibrium point and

the eigenvalues of the liniarization at the origin are a i, where = 1 a2 . Consequently, it follows
that the origin is asymptotically stable. Let us now try to estimate the basin
of attraction of the origin by using a Lyapunov function. With the intent of
determining a quadratic Lyapunov function, we first put the linear part of
the vector field into Real Jordan Normal Form. If we use the new variables
y defined by:

1 0
0
1
1
y = P x, P =
P =
(20)
a
a 1
then the system becomes:
y1 = ay1 + y2
y2 = y1 ay2 1 y13

(21)

We now take the Lyapunov function as the quadratic function:

V (y1 , y2 ) =

1 2
(y + y22 )
2a 1

(22)

then a simple computation yields that:

1 3
V (y1 , y2 ) = (y12 + y22 )
y y2
a 1
32

(23)

Now, the main task is to determine the largest subset of R2 containing

the origin where V is positive definite. This is a difficult task, so let us
settle for the largest such disk. Observe that the level sets of the function
1 3
y12 + y22 are circles about the origin and the level sets of the function a
y1 y2

are similar to hyperbolas. From the symmetry of V , it is evident that the

radius r0 of the largest circle inside which V is positive definite must satisfy:
r02

1 4
r =0
a 0

(24)

that is, r0 = a.
Thus, every solution of (19) with initial value y(0) satisfying ky(0)k < r0
approaches the origin as t +.
Brute approximation of the region of attraction
You may want to transform the nonlinear system into an equivalent form
x = Ax + g(x) where x = Ax is the system linearization around the chosen
equilibrium point and g(x) is the nonlinear part of the system. Note that on
the domain of attraction:
V (x) = xT P x c

(25)

V (x) = xT Qx + 2xT P g(x)

(26)

the following is true:

where the first term is negative when Q > 0 and the second term is in general
indefinite.
Since for a matrix M the following inequalities are true:
min (M )kxk2 xT M x max (M )kxk2
min (M )kxkkyk xT M y max (M )kxkkyk

(27)

where
kxk2 = xT x
,
and the function g(x) satisfies:
kg(x)k
0 as
kxk
33

kxk 0

(28)

we may reason as follows:

> 0, r > 0 :

kg(x)k < kxk

kxk < r

V (x) < xT Qx + 2max (P )kxk2 ,

kxk < r

(29)

Since xT Qx min (Q)kxk2 we get:

V (x) < (min (Q) 2max (P ))kxk2 ,

kxk < r

(30)

Choosing < min (Q)/2max (P ) ensures that V (x) is negative definite.

Global asymptotical stability
The stability type of an equilibrium point is a local property. However,
in this section we deal with the following question: under what conditions
will the region of attraction be the whole space Rn . This will be the case
if we can show that for any initial state x, the trajectory (t; x) approaches
the origin as t , no matter how large kxk is. If an asymptotically stable
equilibrium point at the origin has this property, it is said to be globally
asymptotically stable .
Barbashin-Krasovskii Theorem Let x = 0 be an equilibrium point
for x = f (x). Let V : Rn R be a continuously differentiable function such
that:
V (0) = 0 and V (x) > 0 x 6= 0
kxk V (x) (V is radially unbounded)
V (x) < 0, x 6= 0

(31)

then x = 0 is globally asymptotically stable.

Example: Lets suppose we are given the system:
x 1 = x2
x 2 = h(x1 ) ax2
with yh(y) > 0 for all y 6= 0. The Lyapunov function:

Z x1
T ka2 ka
V (x) = x
x+
h(y)dy
ka 1
2
0

(32)

(33)

is positive definite for all x R2 and radially unbounded. The derivative:

34

V (x) = a(1 k)x22 akx1 h(x1 )

(34)

is negative definite for all x R2 since 0 < k < 1. Therefore, the origin
is globally asymptotically stable.
Remark: If the origin x = 0 is a GLOBALLY asymptotically stable equilibrium point of a system, then it must be the unique equilibrium point of the
system. For if there was another equilibrium point x
, the trajectory starting
at x would remain at x for all t 0; hence, it would not approach the origin
which contradicts the claim that the origin is globally asymptotically stable.
Therefore, global asymptotic stability is not studied for multiple equilibria
systems like the pendulum equation.

Home-works

1. For each of the following systems, use a quadratic Lyapunov function

candidate to show that the origin is asymptotically stable:
x 1
x 2
x 1
x 2

= x1 + x22
= x2
= x1 + x21 x2
= x2 + x1

x 1
x 2
x 1
x 2

= (x1 x2 )(x21 + x22 1)

= (x1 + x2 )(x21 + x22 1)
= x1 x2
= x1 x32

(35)

2. Using V (x) = x21 + x22 , study the stability of the origin of the system:
x 1 = x1 (k 2 x21 x22 ) + x2 (x21 + x22 + k 2 )
x 2 = x1 (k 2 + x21 + x22 ) + x2 (k 2 x21 x22 )

(36)

when: (a) k = 0 and (b) k 6= 0.

3. Consider the second-order linear control system x = Ax + Bu, A
R22 , B R21 where the coefficient matrix A has (a) real positive eigenvalues and (b) complex eigenvalues with positive real part. Investigate, using
the Lyapunov function, if there exists a linear control law u = Kx R able
to stabilize the control system.
4. Discuss the stability of the following systems, given in polar coordi-

35

nates:

(r
(r
(r
(r
(r

= r(r 4); = 1)
= ar2 + br3 ; = 1)
= r( r); = 1)
= ar2 ; = 1)
= ar(r 1)(r ); = 1)

(37)

where a, b are constants and (t) is time-variant.

5. Consider the following system of diff. eq.
x 1 = ax21 + bx1 x2 + cx22
x 2 = x2 + lx12 + mx1 x2 + nx22

(38)

where a, b, c, l, m and n are constants. Show that:

1. if a 6= 0 then the origin is unstable;
2. if a = 0, bl < 0, then the origin is asymptotically stable;
3. if a = 0, bl > 0, then the origin is unstable;
4. if a = b = 0, cl2 6= 0, then the origin is unstable;
5. if a = b = c = 0, then the origin is stable but not asymptotically;
6. if a = l = 0, then the origin is stable but not asymptotically.
Show that the origin of the second-order system y + y 3 = 0 is a stable
equilibrium point.
6. Show that the equilibrium point at the origin of the system:
x 1 = x1 x2 + ax31 + bx1 x2
x 2 = x2 + cx12 + dx21 x2

(39)

1. is asymptotically stable if either a+c < 0, or a+c = 0 and cd+bc2 < 0;

2. is unstable if either a + c > 0, or a + c = 0 and cd + bc2 > 0.
7. Consider the second order system:
x 1 = x2
x 2 = x1 x32

(40)

What is the stability type of the equilibrium point at the origin ?

8. Discuss the stability properties of the origin for the system:
x 1 = x2 + x1 x2 + ax1 x22
x 2 = x1 x21 + x22
for various values of a.
36

(41)

9. Discuss the stability and instability of the equilibrium points of the

system:
x 1 = x2 + x23
(42)
x 2 = x1 + x13

Lab-works

1. Draw the Lyapunov function for the systems given in the lecture notes, in
3-D as the function V (x1 , x2 ) and in 2-D as the function V (t).
2. Draw the phase portraits and the Lyapunov functions for the problems
given in the above section (Home-works).

Matlab

Study the Matlab functions lyap, contour, mesh.

Example: Given the nonlinear system

x 1 = x2
x 2 = x1 + (x21 1)x2

(43)

compute the domain of attraction of its equilibrium point (0, 0).

Method 1: We choose the Lyapunov function as the quadratic function
V (x1 (t), x2 (t)) = x(t)T P x(t),

P = PT > 0

P > 0,

and compute its time-derivative as follows:

dV
V =
= 2(xT Qx + p12 x31 x2 + p22 x21 x22 )
dt
where:

Q=

p12
0.5(p22 + p12 p11 )
0.5(p22 + p12 p11 )
p22 p12

(44)

(45)

Note that we should require p22 > 0 (P > 0). On the other hand, if
p12 = 0 it turns out that det(Q) = 0.52 (p22 p11 )2 0 which forbids
Q > 0.
We may compute the worst case scenario for V < 0 as follows:
x21 x22 < (x21 + x22 )(x21 + X22 ) = kxk2 kxk2 = r4
|x31 x2 |

<

x21 |x1 ||x2 |

<

(x21

x22 )

(x21
37

x22 )

(46)

(x21

x22 )

= kxk = r

where r2 = kxk2 = x21 + x22 . Thus, we should find a conservative (small)

approximation of the domain of attraction from:
xT Qx + (p12 + p22 )r4 < 0

(47)

min (Q) + (p12 + p22 )r2 < 0

(48)

or
Therefore:
r2 <

min (Q)
p12 + p22

(49)

is an approximation of the radius of the region of attraction.

It might be possible to find combinations of matrices P and Q which
shows a bigger domain of attractions than others. In order to find the best
matrices P and Q we must reason as follows:
By denoting the eigenvalues of the matrix Q: 1 = and 2 = + ,
, > 0, from (45) we get:
2 + = p22
( + ) = p12 (p22 p12 ) 0.52 (p22 + p12 p11 )2
or, equivalently:
2 p22 + p12 (p22 p12 ) 0.52 (p22 + p12 p11 )2 = 0
Since:

(50)

p22 1

2
2
2
= p22
4p12 (p22 p12 ) + (p22 + p12 p11 )2
1,2 =

we require a positive great minimum eigenvalue by = 0. In addition, from

2
the positivity of P and Q we should require: p22 > 0, p11 > 0, p11 p22 > p12
,
p12 > 0, p22 > p12 ,
p12 (p22 p12 ) 0.52 (p22 + p12 p11 )2 > 0
We may choose:

Q=

1 0
0 1

; P =

38

3 1
1 2

(51)

10

25 20

10

1.5

20

0.5

18

1
2
0
2

2
2

10
8

10

1.5

15
20 25

12

10

0
2

0.5

14

16

10

10

10

15
0

22
20

25

15

30

24

6
1

Figure 1: The Lyapunov function and its level sets

Using the following Matlab code:

[x, y] = meshgrid(2 : 0.1 : 2, 2 : 0.1 : 2);

z = 3 x. x + 2 y. y + 2 x. y;
subplot(1, 2, 1), mesh(x, y, z)
subplot(1, 2, 2), [c, h] = contour(x, y, z); clabel(c, h), colorbar
we show the Lyapunov function and its level sets.
Remark: The Matlab function contour shows the level sets of the Lyapunov function.
With Q and P chosen as in (51) we get:
r2 < 1/3 r < 0.5774
However, if we use the following Matlab code:
[x, y] = meshgrid(.9 : 0.1 : .9, .9 : 0.1 : .9)
z = (x. x + y. y) 2 x. x. y. y 1 x. x. x. y;
subplot(2, 2, 1), mesh(x, y, z)
39

0.5

0.5

0
1

0.3

0.3

0
1

0.70.8

0.5

0.5

0.4
0.8 0.6

0.2
0

0.3
0.2
0.1

0.5

0.6
0.5
0.4

0.3

0.4

0.6

0.5

0.4 0.05.6

0.4

0.2

0.3

0.5 0
.

0.8
0.6

0.2

2
1

0.

0.1

0.2

0.4
0.5
0.6
0

0
0.06.5 .4

0.4

0.7
0.6
0.5
0.4
0.3

0
0..87

1
0.6 0.8
0.4

0.5

0.6
2

0.2

0.5

Figure 2: -V for x1 , x2 [0.9, 0.9] and x1 , x2 [0.8, 0.8]

subplot(2, 2, 2), [c, h] = contour(x, y, z); clabel(c, h), colorbar

[x, y] = meshgrid(.8 : 0.1 : .8, .8 : 0.1 : .8)
z = (x. x + y. y) 2 x. x. y. y 1 x. x. x. y;
subplot(2, 2, 3), mesh(x, y, z)
subplot(2, 2, 4), [c, h] = contour(x, y, z); clabel(c, h), colorbar

we get Fig. 2. Note that for x1 , x2 [0.8, 0.8] V < 0 (since V > 0)
but for x1 , x2 [0.9,0.9] this is not true anymore.
Thus, the stability is
guaranteed for r < 0.8 2 = 1.1314 but not for r < 0.9 2 = 1.2728.

Method 2: The system (43) may be also written as follows:

0
1
0
x = Ax + g(x) =
x+
1 1
x21 x2
where
A=

f
|(0,0)
x
40

(52)

8
6
4
2
0
2
4
6
8
4

Figure 3: Phase portrait

is the Jacobian of the system in the equilibrium point (0, 0). Note that:
q
q
2
2
2
2
kg(x1 , x2 )k = 0 + (x1 x2 ) < (x1 + x2 ) x21 + x22 = kxk3
Using the Matlab lyap function for Q chosen as the identity matrix:
>> P = lyap([0, 1; 1, 1]0 , [1, 0; 0, 1])
we get:
P =[ 1.5000 , 0.5000 ; 0.5000 , 1.0000 ]
>> eig(P )
ans =
1.8090, 0.6910
Thus, using (30):
= r2 <

min (Q)
1
=
2 max (P )
2 1.8090
r < 0.5257

We show a phase portrait in Fig 3.

41

35

30

25

V(t)

20

15

10

10

Figure 4: The time history of the energy for different initial conditions

We also show the time history of V (t) for P = [3 1; 1 2] in Figure 4 for

different initial conditions inside and outside the attraction domain. Note
that when the initial condition is inside the domain of attraction the energy
decreases monotonically.
In Figure 5 we show the difference between the time history of V (t) for
P = [3 1; 1 2] and P = [1 0; 0 1] for an initial condition contained in the
domain of attraction.
The Matlab code used was based on the sequence:
[t,x]=ode23(vpol11,[0,10], [1,1]), P=[3 1; 1 2]; V=[];
for i=1:length(t)
V=[V x(i,:) * P * x(i,:)];
end plot(t,V,b), hold on

42

4.5
4
3.5
3

V(t)

2.5
2
1.5
1
0.5
0

10

Figure 5: The time history of the energy versus the time history of the state
vector norm for the same initial condition

43

Lecture notes 2013/ 2014

Dynamical systems, 4
Mihaela Cistelecan, UT Cluj-Napoca
These notes were prepared using:
1. Hale, J.K., Kocak, H. Dynamics and Bifurcations. Springer-Verlag,
1991.
2. Hirsch, W., Smale, S, Devaney, R. Differential Equations, Dynamical
Systems and An Introduction to Chaos. elsevier, 2004
3. Diacu, F. An Introduction to Differential Equations: Order and Chaos.
W.H. Freeman and Company, 2000.
4. Khalil, H. Nonlinear Systems. Prentice Hall. 1996.
5. http://www.soe.ucsc.edu/classes/ams146/Spring05/Lectures.html

Bifurcations in linear and nonlinear systems. Topological Equivalence. Structural

stability.

The reasons for studying bifurcations and structural stability in dynamical

systems are:
1. In the case of a dynamical system with perturbed / uncertain parameters it may be that case that for a combination of parameter values the
equilibrium is nonhyperbolic. In such a case the system dynamics shows
different types of equilibrium for different parameter values.
2. When the parameters of a system are time-varying either the number
of the equilibria or the equilibrium types may change.
3. In the nonlinear systems we need to examine how a particular nonlinear
term of a vector field affects the flow near a nonhyperbolic equilibrium point.
Bifurcations: Let
x = F (, x),
44

x Rn

(1)

be a dynamic system depending on k parameters = [1 , ..., k ]T . The study

of changes in the qualitative structure of the flow as parameters are varied
is called bifurcation theory.
Sink and Source At this point we think it is useful to introduce
two new keywords: sink and source. We will refer to a stable node/focus as
a sink and to an unstable node/ focus as a source.
Definition: A linear system whose eigenvalues have negative real parts is
called a hyperbolic sink ; a linear system whose eigenvalues have positive real
parts is called a hyperbolic source ; and, when one eigenvalue is negative and
the other is positive the linear system is said to be a hyperbolic saddle .
Remark: These new terms refer to the topological viewpoint - there are
only three cases and they are determined solely by the signs of the real parts
of the eigenvalues: sink, source, saddle. It means that a stable focus and a
node are topologically equivalent.
Structural stability (stable orbit structure) - Let x = F (, x) be a
vector field that depends on k parameters = [1 , ..., k ]T . For a fixed value
the vector field x = F (,
x) is called structurally stable if there is an
= ,
x) for all
> 0 such that x = F (, x) is topologically equivalent to x = F (,

values of satisfying k k < .

Thus, at a given parameter value, a differential equation is said to have
stable orbit structure if the qualitative structure of the flow does not change
for sufficiently small variations of the parameter. A parameter value for
which the flow does not have stable orbit structure is called a bifurcation
value , and the equation is said to be a bifurcation point.

1.1

Motivation

A. Linear Systems - Structural stability and bifurcation Since mathematical models are approximations of the reality, solving the associated
differential equations may not suffice for understanding the phenomenon it
describes; we might need additional information.
Example 1: If observations of experiments suggest as a model of a physical
phenomenon a certain eq., say
x + 2.1x + 0.76x = 0

(2)

then we might need to determine the behavior of the solutions for all equations of the type
x + bx + cx = 0
(3)
45

where b and c belong to some intervals that contain the coefficients in (2),
say b in (1.9, 2.3) and c in (0.6, 0.9). Since all measurements are approximate,
the value 2.1 could in truth be 2.05 or 2.17. If the solutions of all equations
given by (3), for b and c in the chosen intervals behave like the solution of
(2), then the results are trustworthy because the qualitative structure of the
phase-plane portrait remains the same. We then say that (3) is structurally
stable . In the phase plane, the system:
x = y
y = cx by

(4)

is structurally stable around the values b0 and c0 if the nature of equilibrium

remains unchanged for all b and c near b0 and c0 , respectively. At the values
b0 or c0 at which the nature of the equilibrium changes, we say that the
solution encounters a bifurcation .
Discussion: If we assume that c 6= 0 the equilibrium (0, 0) of the system
(4) is a:
(1) source (unstable node / unstable focus) if and only if b < 0 and c > 0;
if additionally b2 4c < 0 then the equilibrium is a spiral source (unstable
focus);
(2) sink (stable node / stable focus) if and only if b > 0 and c > 0; if
additionally b2 4c < 0 then the equilibrium is a spiral sink (stable focus);
(3) saddle if and only if c < 0;
(4) center if and only if b = 0 and c > 0.
This means that models involving linear second-order equations with coefficients belonging to the interior of structurally stable regions of the parameter plane are robust and can be trusted. Models whose coefficients come
close to the bifurcation lines must be regarded more carefully.
B. Nonlinear Systems and bifurcations - Change in the number
of equilibrium points or the type of equilibria
B1. Scalar systems For a scalar differential equation x = f (x), x R,
the equilibrium points and the sign of the function f (x) between equilibria
determine the number of orbits and the direction of the flow on the orbits
(the orbit structure of the differential equation or the qualitative structure of
the flow).
Example 2: Saddle-node bifurcation
x = c + x2 = F (c, x)
46

(5)

where c is a parameter.
Discussion The flows of system (30) depend
parameter c, as follows:
on the
1). c < 0 - the equilibrium
points are c
and
c
and the orbits are

given by the intervals (, c), ( c, c) and ( c, +);

2). c = 0 - the equilibrium point is 0 and the orbits are (, 0) and
(0, +);
3). c > 0 - there is no equilibrium point, the only orbit is (, +).
If the parameter c is varied, as long as c < 0, the number and the direction
of the orbits remain the same;
the only change is the shifting of the location
of the equilibrium points c. Similarly, for all c > 0 there is only one
orbit and its direction is from left to right. However, if c = 0, regardless of
how small an amount c is varied, the number of orbits changes: there are
two equilibria for any c < 0 and none for c > 0.
In this case, the point c = 0 is a bifurcation point.
B2. Planar systems
Example 3: Damped pendulum with constant torque M (the case where
the linearized vector field has one zero and one negative eigenvalue):
+ + sin = M

(6)

If the rod of the pendulum is rigid, then one can interpret the application of
the torque as pushing the pendulum with constant force that is perpendicular
to the rod.
By choosing y1 = and y2 = the phase plane model is:
y1 = y2
y2 = y2 sin y1 + M

(7)

Since sin y1 is periodic with period 2, we confine our discussion to the

values of y1 in the interval [, ].
The equilibrium points of (7) are given by y2 = 0 and sin y1 = M .
Discussion: For M > 1 there are no equilibrium points;when M = 1
there is a single equilibrium point at (/2, 0); for M < 1 there are two equilibrium points. It appears that the equilibrium point at (/2, 0) undergoes
a bifurcation at the parameter value M=1.
Example 4: Van der Pols oscillator (the case where the linearized vector
field has two imaginary eigenvalues):
y (2 y 2 )y + y = 0
47

(8)

is equivalent to the planar system:

x 1 = x2
x 2 = x1 + 2x2 x21 x2

(9)

The eigenvalues
of the linearization of (9) about the equilibrium point at
the origin are i 1 2 . For < 0 the origin is asymptotically stable
because the real parts of the eigenvalues are negative. At = 0 the origin is
still stable, but for > 0 the real parts of the eigenvalues become positive
and thus the origin is unstable.

1.2

Linear equivalence vs. Topological equivalence

The eigenvalues of a matrix depend continuously on its parameters

1). If the equilibrium point x = 0 of x = Ax is hyperbolic (node, focus,
saddle point), then the equilibrium point x = 0 of x = (A + A)x will be
of the same type for sufficiently small perturbations. When A has multiple
nonzero real eigenvalues, infinitesimally small perturbations could result in
a pair of complex eigenvalues. Hence a stable node would either remain a
stable node or become a stable focus.
2). For a nonhyperbolic equilibrium the type of equilibrium may change
under small perturbations. For example, consider the following perturbation
of the real Jordan form in the case of a center:
"

1
1

where is a perturbation parameter. If > 0 the equilibrium point of

the perturbed system is an unstable focus; if < 0 the eq. point of the
perturbed system is a stable focus. This is true no matter how small is, as
long as it is different from zero.
In the course of transforming matrices into their Jordan Normal Forms
we have investigated the question of topological equivalence in a limited
context by considering only invertible linear maps. The problem was stated
as follows: if there is an invertible 2 2 matrix P such that A = P 1 BP
then the flows of x = Ax and x = Bx are related by P eAt = eBt P . In other
words, if matrices A and B are similar (conjugate) then the flows of x = Ax
and x = Bx are linear equivalent .
48

Unfortunately, the linear equivalence is a bit too restrictive for comparing

the qualitative features of the linear systems flows. For example, the two
linear systems x = Ix and x = 2Ix should be considered qualitatively
equivalent, yet they are not linearly equivalent (the matrices I and 2I
have different eigenvalues).
Therefore, the concept of topological equivalence was introduced. Topological equivalence - requires (the existence of ) a nonlinear transformation
(homeomorphism) that maps the orbits of x = f (x) onto the orbits of
x = g(x) and preserves the sense of direction of time.
The stability type of a hyperbolic equilibrium point is preserved under
arbitrary but small nonlinear perturbations - from Taylor expansion of the
function f (x) (see again the previous lecture):
Theorem (Grobman-Hartman) If x Rn is a hyperbolic equilibrium
point of x = f (x), x Rn , then there is a neighborhood of x in which f
is topologically equivalent to the linear vector field x = Df (
x)x (Df - the
Jacobian of f).
Definition: Linear systems with hyperbolic equilibrium types - Any two
linear systems with hyperbolic equilibrium points are topologically equivalent
if they have an equal number of eigenvalues with positive real parts and an
equal number with negative real parts.
To study bifurcations and structural stability the following theorem is of
paramount importance.
The Implicit Function Theorem Assume that f : Rn Rm Rn is
continuously differentiable at each point (x, y) of an open set S Rn Rm .
Let (x0 , y0 ) be a point in S for which f (x0 , y0 ) = 0 and for which the Jacobian
matrix [f /x](x0 , y0 ) is nonsingular. Then there exist neighborhoods U
Rn of x0 and V Rm of y0 such that for each y V the equation f (x, y) = 0
has a unique solution x U . Moreover, this solution can be given as x = g(y)
where g is continuously differentiable at y = y0 .

1.3

Bifurcations and structural stability in Planar Linear Systems

In linear systems there exists only one equilibrium point. Up to a translation

the equilibrium point can be considered the origin of the state space. The
phase portrait of a hyperbolic linear system does not change topologically if
the entries of its coefficient matrix are varied by a small amount. This means

49

that a linear system may face bifurcations only in the following two cases:
a) The equilibrium of the linear system is nonhyperbolic and the system
(the coefficient matrix) undergoes small perturbations.
b) The linear system is parameter time-varying and under these variations
there exists a parameter configuration that produce a modification of the
hyperbolic equilibrium into a nonhyperbolic equilibrium. For example, we
may be given a linear system that contains parameters and the parameters
enter into the system as continuous functions. In this case it is unavoidable
to hit one of the nonhyperbolic systems if there are parameter values at which
the linear systems belong to two different hyperbolic equivalence classes.
Topological classifications of orbits
In the following we give two theorems (without proof) taken from literature concerning the topological classification of hyperbolic/ nonhyperbolic
systems.
Theorem T1: Hyperbolic linear systems:- Suppose that the eigenvalues of two matrices A and B have nonzero real parts. Then the two linear
systems x = Ax and x = Bx are topologically equivalent if and only if A
and B have the same number of eigenvalues with negative (and hence positive) real parts. Consequently, up to topological equivalence, there are three
distinct equivalence classes of hyperbolic planar linear systems with, for ex.,
the following representatives:

1
0

1
0

1
0

0
1!
0
1 !
0
1

two negative eigenvalues;

two positive eigenvalues;

(10)

one positive and one negative eigenvalue.

Theorem T2: If a coefficient matrix A has at least one eigenvalue with zero
real part, then the planar linear system x = Ax is topologically equivalent to
precisely one of the following five linear systems with the indicated coefficient
matrices:

50

0
0

1
0

1
0

0
0

0
1

0
the zero matrix;
0 !
0
one negative and one zero eigenvalue;
0!
0
one positive and one zero eigenvalue;
0 !
1
two zero eigenvalues but one eigenvector;
0 !
1
two purely imaginary eigenvalues.
0

(11)

Sketch of the proof for Theorem 1, first matrix The idea of the classification is to find the transformation function that transform a dynamics
x = Ax into its topologically equivalent dynamics x = Bx. This problem
is equivalent to the following one: for each linear system, find an ellipse encircling the origin such that each orbit, except the equilibrium point at the
origin, crosses the ellipse in the same direction and only at one point. This
is also equivalent to finding (or proving the existence) of two positive definite
symmetric matrices CA and CB such that:
AT CA + CA A = I; B T CB + CB B = I

(12)

because, if x(t) is the solution of x = Ax, then

i
d h T
x (t)CA x(t) = xT (t)[AT CA + CA A]x(t) = xT (t)x(t)
dt

(13)

Similarly,
i
d h T
x (t)CB x(t) = xT (t)[B T CB + CB B]x(t) = xT (t)x(t)
dt

(14)

Therefore, any nonequilibrium solution x(t) of x = Ax crosses the level sets

of xT CA x inward. Similarly, any nonequilibrium solution x(t) of x = Bx
crosses the level sets of xT CB x inward, too.
Remark: A real symmetric matrix C, (C T = C) is said to be positive
definite if the quadratic form xT Cx > 0 for all x 6= 0. An important geometric
property of a positive definite quadratic form on the plane is that the level
set {x : xT Cx = k}, for any k > 0, is an ellipse encircling the origin.

51

The trace-determinant plane

From the topological equivalence point of view we may classify the dynamics of a planar linear system using only two parameters, as follows.
For a matrix:

!
a b
A=
(15)
c d
we know that the eigenvalues are the roots of the characteristic equation
which can be written:
2 (a + d) + (ad bc) = 0

(16)

2 trace(A) + det(A) = 0

(17)

or, equivalently:
The eigenvalues are given by:
q
1
= (tr(A) (tr(A))2 4det(A))
2

(18)

+ + = tr(A), + = det(A)

(19)

Note that:
Thus the point (tr(A), det(A)) defines the type of the equilibrium point,
as follow:
- if tr(A)2 4det(A) < 0 then the equilibrium is either a focus (tr(A) 6= 0)
or a center (tr(A) = 0); if tr(A) < 0 the focus is stable, otherwise it is
unstable;
- if tr(A)2 4det(A) 0 the equilibrium is a node; the node stability
depends on the sign of tr(A);
- if det(A) < 0 the equilibrium is a saddle.
Remarks: 1). The trace-determinant plane is a two-dimensional representation of a four dimensional space, since 2 2 matrices are determined by
four parameters, the entries of the matrix. Thus, there are infinitely many
different matrices corresponding to each point in the trace-determinant plane.
2). We can think of the trace-determinant plane as the analog of the bifurcation diagram for planar linear systems. A one-parameter family of linear
systems corresponds to a curve in the trace-determinant plane. When this
curve crosses the trace-axis, the positive determinant axis, or the parabola
tr(A)2 4det(A) = 0, the phase portrait of the linear system undergoes

52

a bifurcation, that is a major change occurs in the geometry of the phase

portrait.
Example: Study the bifurcations of the following planar linear system:
x 1 = x2 , x 2 = x1 + 2x2 .

trace
det
dynamics

|
|
|
|
|

<0

++
++
source

=0

+
0
bif.

>0

++

saddle

(20)

Discussion: Since tr A = 2 = 1 + 2 and det A = = 1 2 we

conclude that the bifurcation point is = 0 (one zero eigenvalue), for > 0
the equilibrium is a saddle (one positive eigenvalue, one negative eigenvalue),
for < 0 the equilibrium is a source (both eigenvalues with positive real
part).
Unfolding of the non-hyperbolic equilibria - optional reading
Nonhyperbolic equilibria are the only points where a bifurcation is possible. In the following, an analysis of the possible changes in the system
dynamics is given, taken into account the change in the system parameters.
Let A be a nonhyperbolic 2 2 matrix in Jordan Normal Form. Observe that, for any 6= 0 and || small, the matrix A + I is hyperbolic.
Under arbitrarily small perturbations a nonhyperbolic system can be made
hyperbolic.
The method of studying bifurcations in nonhyperbolic linear systems is
based on inserting a minimum number of parameters in any given nonhyperbolic system in such a way that as the parameters are varied we can reach,
up to topological equivalence, any linear system in a small neighborhood of
the nonhyperbolic linear system. The resulting parameter-dependent linear
system is called an unfolding of the nonhyperbolic linear system.
We study the bifurcations of nonhyperbolic systems in Jordan Normal
Form. Excluding the zero matrix, there are three such matrices:

0
0 0

0
0

0 1
0 0

(21)

where 6= 0 and 6= 0. Furthermore, without loss of generality we shall

take = 1 and = 1.

53

For a 2 2 matrix A all the relevant information regarding eigenvalues is

contained in its characteristic polynomial:
det(A I) = 2 (tr A) + det A

(22)

where:
tr A = a11 + a22 = 1 + 2 ; det A = a11 a22 a21 a12 = 1 2

(23)

In the following the matrix 2 2 A0 is viewed as a point in R4 by ordering

the entries of this matrix into a four-vector.
Case 1: Unfolding

!
1 0
A0 =
(24)
0 0
In this case, there is a neighborhood of A0 in R4 such that in this neighborhood the matrices that are topologically equivalent to the nonhyperbolic
matrix A0 satisfy the conditions tr A < 0 and det A = 0. The set of such
matrices is a three-dimensional surface S in R4 containing the matrix (24)
because a22 det A|A0 = 1 6= 0 (see IFT). All the matrices on one side of
the surface S are topologically equivalent. We construct a curve of matrices
which crosses this surface transversally at the point A0 . In other words, we
define a matrix A() depending on a scalar parameter such that A(0) = A0 .
As the parameter is varied from negative to positive values, one crosses the
surface S from the side with (tr A() < 0, det A() < 0) to the side with
(tr A() < 0, det A() > 0). Observe that on the first side, the matrices have
two real negative eigenvalues, and on the other side, they have one negative
and one positive eigenvalue. One such one-parameter matrix is:

A() =

1 0
0

(25)

In the first case the matrices have two real negative eigenvalues, and in
the other case they have one negative and one positive eigenvalue.
It follows that in the case of a nonhyperbolic equilibrium in the form (24)
the system x = Ax may switch from a stable node equilibrium to a saddle
equilibrium or vice-versa when passing through (24).
Case 2: Unfolding

!
0 1
A0 =
(26)
1 0
54

A similar reasoning as above gives a one-parameter matrix as follows:

A() =

1
1

(27)

It means that for the nonhyperbolic equilibrium (26) the system x = Ax may
switch from a stable focus equilibrium to an unstable focus equilibrium or
vice-versa when passing through (26).
Case 3: Unfolding

!
0 1
A0 =
(28)
0 0
Since in this case the matrix A0 has tr A = 0 and det A = 0 the A0 unfolding
depends on two parameters. It was proved that bifurcations are given by the
two following two-parameters matrices:

A(1 , 2 ) =

1
1
2 1

, A(1 , 2 ) =

0 1
1 2

(29)

As we vary the parameters 1 and 2 we can reach matrices whose eigenvalues

are either both negative/positive or one is negative and the other positive.

1.4

Bifurcations in NONLINEAR SCALAR Systems

We study the system x = F (, x), Rk , x R.

Relevant graphs:
(t, x(t)) - system solution
(x(t), F (, x) for a fixed - we study the equilibrium points and their
stability
(i , x) - the bifurcation diagram for a fixed , we are interested in the
variation of i
Definition An equilibrium point x R of x = f (x), x R, is called a
hyperbolic equilibrium if f 0 (
x) 6= 0. If f 0 (
x) = 0, then x is called a nonhyperbolic or degenerate equilibrium point.
Elementary bifurcations in scalar systems

55

1. Saddle-node bifurcation
x = + x2 = F (, x)

(30)

x = x + x2 = F (, x)

(31)

x = + x x3 = F (, x)

(32)

x = x x3 = F (, x)

(33)

x = c + dx x3 = F (c, d, x)

(34)

2. Transcritical bifurcation

3. Hysteresis
4. Pitchfork bifurcation

5. Fold or cusp
A numerical procedure to draw the bifurcation diagram for a
scalar one-parameter system
A graphical method for depicting some of the important dynamical features in equations x = F (, x) depending on the parameter R is the
bifurcation diagram. This method consists of drawing curves on the (, x)plane, where the curves depict the equilibrium points for each value of the
parameter. More specifically, a point (0 , x0 ) lies on one of these curves if and
only if F (0 , x0 ) = 0. For instance, the bifurcation diagram of the saddlenode bifurcation x = + x2 is the parabola = x2 . Also, to represent the
stability types of these equilibria, we label stable equilibria with solid curves
and unstable equilibria with dotted curves.
For a scalar system x = F (, x) depending on a scalar parameter, we can
parameterize the curve F ((t), x(t)) = 0 by t, satisfying (0) = 0 , x(0) = x0
and F (0 , x0 ) = 0. To represent this parameterized curve as an orbit of a
differential equation, we differentiate the eq. F ((t), x(t)) = 0 with respect
to t:
F
F
(, x) +
(, x)x = 0
(35)

x
x)
To satisfy this identity (,
must be a constant multiple of the vector
F
F
( x (, x), (, x)). If we choose the parametrization of the curve ((t), x(t))
in such a way that this constant is one, then we arrive at the following system:
= F
(, x)
x
F
x = (, x)
56

(36)

satisfying the initial conditions (0) = 0 , x(0) = x0 . The orbits of

this system are investigated on the (, x)-plane. A constructive numerical
procedure for obtaining the zero set of F (, x) is as follows:
Step 1. Fix = 0 and use the scalar differential equation x = F (0 , x)
to find the set E0 of zeros of F (0 , x).
Step 2. For each x0 E0 , compute numerically the solution of the
differential eq. (36) with initial value (0) = 0 and x(0) = x0 , in both
forward and reverse time (using negative step size) directions.
Step 3. Repeat Step 1 as necessary for other values of the parameter
because there may be some components of the curves of zeros of F (, x)
that never intersect the vertical line = 0 .
Thus, the bifurcation diagram is composed on branches that are constructed numerically from segments around the equilibrium points chosen at
Step 1. From a study of the sign of we may decide on the sign of F
and
x
then, on the stability of branches.
How to study equilibria using the Implicit Function Theorem
Since the STABILITY of an equilibrium point x is a local property of
the flow near the equilibrium, it is reasonable to expect that under certain
conditions the stability properties of x can be determined from the linear
approximation, that is, the derivative
f 0 (x) =

d
f (x)
dx

of the function f near x. The linear differential equation x = f 0 (

x)x is called
the linear variational equation or the linearization of the vector field x = f (x)
about its equilibrium point x. A theorem proves that when f 0 (x) 6= 0 the
stability type of the equilibrium point x of x = f (x) is the same as the
stability type of the equilibrium point at the origin of its linearized vector
field.
While for a hyperbolic equilibrium the stability of the equilibrium point
depends only on the first-order term in the Taylor expansion, for a nonhyperbolic equilibrium the stability of the equilibrium point depends on higherorder terms in the Taylor expansion of the function f (x) about x (or its
translation about 0).

57

The Implicit Function Theorem (IFT) for Bifurcation theory in scalar

systems
Let = [1 , ..., k ] be a vector in Rk . Suppose that F : Rk R R,
(, x) F (, x), is a continuous function satisfying:
F (0, 0) = 0 and

F
(0, 0)
x

6= 0

(37)

Then there are constants > 0 and > 0 and a continuous function
: { : kk < } R
such that
(0) = 0 and F (, ()) = 0 f or kk <

(38)

Moreover, if there is a (0 , x0 ) Rk R such that k0 k < and |0 | < ,

and satisfies the equation F (0 , x0 ) = 0, then x0 = (0 ).
To apply the IFT in a specific situation, you may need to transform variables in such a way that the origin in the (, x)-space becomes zero of the
given function. For ex., if G : Rk R R, (, x) G(0 , x0 ) = 0, then
we can set F (, x) = G(0 + , x0 + x) and F (0, 0) = 0.
The IFT can be used to study equilibria in the following context. Let
x = F (, x) be a differential equation depending on k parameters =
[1 , ..., k ]T . If x = 0 is a hyperbolic equilibrium point of the differential
equation x = F (, x) at = 0, then the conditions of the IFT are satisfied. This guarantees that the equation F (, x) = 0 may be solved locally
for x = () as a function of the parameters = [1 , ..., k ]T . Furthermore
F
(, ()) 6= 0 for sufficiently small. Thus, the qualitative structure of
x
the flow does not change near x = 0. Consequently, there are no bifurcations
in the neighborhood of x = 0 for sufficiently small values of the parameters.
Briefly - if F (0, 0) = 0 and F
(0, 0) 6= 0 we should always study
x
the sign of F
(,
())
and
draw
conclusions
about the stability of
x
x = F (, x) about (0, 0).
Example: The system x = c x = F (c, x) is insensitive to the variation
of the parameter c R. By drawing the phase portraits (x, F (c, x)) for different values of c, or by an analytic study of the system we find a unique
stable equilibrium point no matter what value was chosen for c.

58

What if (0, 0) is a nonhyperbolic equilibrium ? Then, the parameter may change everything ... Therefore, we have to investigate the sign
of the remaining terms of the Taylor expansion.
Remember that the Taylor expansion of function F (, x) about (+, x+
) is as follows:
F ( + , x + ) = F (, x) + F
+ F
+
x

1 2F 2
1 2F 2
1 2F
+ 2 2 + + 2 x + higher order terms
2 x2

(39)

Example Study the bifurcation of the scalar system x = F (, x) where:

F (, x) = x x3 (the supercritical Pitchfork bifurcation )
Equilibrium points are on F (, x) = 0.
It turns out we have the following cases: 1). x = 0 and 2). x = if > 0
In the bifurcation diagram (r, x) we may decide on stable / unstable
branches as follows: Compute f 0 (x) = F
. If f 0 (x) 6= 0 then, we may
x
decide on the stability around the echilibrium by studying the sign of
f 0 (x). Since f 0 (x) = 3x2 , in the equilibrium points we get:
1. f 0 (0) = stable branch for < 0 and unstable branch for
>0

2. f 0 ( ) = 2 stable branch for > 0 and unstable branch

for < 0
In all, four branches in the bifurcation diagram.
Discussion: The system has three equilibria and a stable orbit structure
for all > 0. At = 0 the system is at a bifurcation point (the equilibria
come together at the origin). For all < 0 the system has stable orbit structure with one stable equilibrium point. Thus, x = 0 is always an equilibrium
point. However, as the parameter passes through the bifurcation value
= 0, the equilibrium at the origin loses its stability by giving it up to two
new stable equilibria which bifurcate from the origin.
Remember that we assumed k parameters i in the system !
However, it has been mathematically shown that if the first term
of the Taylor expansion of the vector field f at x is linear, quadratic,
or cubic, then under rather general perturbations the bifurcation
59

of equilibria near x is essentially the same as in the systems given

above.
We study the following three cases of local perturbations near equilibria:
Hyperbolic equilibria
Equilibria with quadratic degeneracy
Equilibria with cubic degeneracy
Hyperbolic equilibria:- Consider the perturbed system x = F (, x),
satisfying:
F (0, x) = f (x) and F
(0, 0) = f 0 (0) 6= 0
(40)
x
where f (0) = 0 and f 0 (0) 6= 0. We investigate the existence of equilibria of
the perturbed eq. x = F (, x). If F (, 0) 6= 0, then the origin will no longer
be an equilibrium. However, from (40) and the fact that f (0) = 0 we have:
F (0, 0) = 0 and

F
(0, 0)
x

= f 0 (0) 6= 0

(41)

Hence, the IFT implies that there exists a function () with (0) = 0
such that x = () and F (, ()) = 0. In order to establish the stability
behavior of the equilibrium () we compute the sign of the derivative:
F
(, ())
x

(42)

From (40) and the fact that (0) = 0, we have F

(0, (0)) = f 0 (0) 6= 0 therex
0
fore, the sign of (42) is the same as that of f (0). Therefore, the stability type
of the equilibrium () of the perturbed equation x = F (, x) is the same as
the stability type of the equilibrium 0 of the unperturbed equation x = f (x).
Therefore, the flow near a hyperbolic equilibrium point is insensitive to small
perturbations of the vector field.
Equilibria with quadratic degeneracy:- Consider the perturbed system x = F (, x), Rk , satisfying:
F (0, x) = f (x)

F
(0, 0)
x

=0

2F
(0, 0)
2x

= f 00 (0) 6= 0

(43)

where f (0) = 0, f 0 (0) = 0 but f 00 (0) 6= 0. These conditions together with

f (0) = 0 imply that the Taylor expansion of F about the origin has the
following form:
F (, x) = a() + b()x + c()
60

x2
+ G(, x)
2

(44)

with a(0) = 0, b(0) = 0, c(0) = f 00 (0) 6= 0, and |G(, x)| < |x|2 .
It has been proved mathematically that the qualitative structure of the flow
of the perturbed equation x = F (, x) is determined from a single function
of the parameters , namely the function () corresponding to the extreme
value of F (, x). The extreme points of F correspond to the solutions x of the
equation F
(, x) = 0. The number of equilibrium points of eq. x = F (, x)
x
depends upon the extreme value () = F (, ()) of the function F , as
follows:
()f 00 (0) < 0 there are two hyperbolic equilibria near the origin
() = 0 implies that there is a nonhyperbolic eq. at the origin
()f 00 (0) > 0 there are no eq. points about the origin
Thus, even though there may be k components of the vector parameter , the
bifurcation behaviour of the perturbed eq. x = F (, x) depends on a single
parameter - the function () corresponding to the extreme value of F (, x).
Example: Study the orbit structure of the dynamical system x = 1 +
2 x + x2 , depending on two small parameters - = [1 2 ]T . The function
() for this example is as follows: (1 , 2 ) = 1 (1/4)22 (corresponds to
the minimum value of the function F (, x)). Thus, the bifurcations occur as
we cross the curve 1 = 22 /4 in the (1 , 2 )-plane. There are two equilibrium
points if 1 < 22 /4 and none if 1 > 22 /4. The same discussion results if we
study the roots of the equations F = 0. Why ?
Equilibria with cubic degeneracy:- Consider the perturbed system
x = F (, x), satisfying:
F (0, x) = f (x)
2F
(0, 0) = 0
2x

F
(0, 0)
x

3F
(0, 0)
3x

=0
= f 000 (0) 6= 0

(45)

where f (0) = 0, f 0 (0) = 0, f 00 (0) = 0 but f 000 (0) 6= 0.

It has been mathematically proven that the two-parameter case is representative of what can happen in the general case near the equilibrium point
x = 0 of f satisfying the conditions above when f is subjected to perturbations that are small together with their derivatives up through order three.
The Taylor expansion about origin is for a two-parameter perturbation
F (, x) as follows:
61

x2
x3
+ d() + G(, x)
(46)
2
6
where a(0) = b(0) = c(0) = 0, d(0) = f 000 (0) 6= 0 and |G(, x)| < |x|3 .
Moreover, it has been proven that an arbitrary two-parameter perturbation
of the cubic degeneracy, under certain reasonable conditions, can be reduced
to the special two-parameter perturbation on the following form:
F (, x) = a() + b()x + c()

F (, x) = 1 + 2 x + c()

x2
x3
+ d() + G(, x)
2
6

(47)

We find the bifurcation values of the parameters, that is, the values of =
(1 , 2 ) for which the function F (, x) has multiple zeroes. To accomplish
this we must solve the following two eq:
2

0 = F (, x) = 1 + 2 x + c() x2 + d() x6 + G(, x)

2
0 = F
(, x) = 2 + c()x + d() x2 + G
(, x)
x
x

(48)

These eq. may be viewed as two equations defining 1 and 2 parametrically

in terms of x. To obtain the solutions locally, we can use the method of
Gaussian elimination in conjunction with repeated applications of the IFT.
The local solutions 1 (x) and 2 (x) of (48) near the origin are given by:
1 (x) = 31 f 000 (0)x3 + ...
2 (x) = 12 f 000 (0)x2 + ...

(49)

These equations are the parametric representation of a cusp in the (1 , 2 )plane which near the origin approximately coincides with the curve
9
32 = f 000 (0)21
8

(50)

Therefore, the bifurcations are determined only by the constant and the
linear terms of the Taylor expansion of the vector field. The term c()x2 has
little influence on the bifurcations of (47) because this term does not enter
into the first approximation to the cusp in the (1 , 2 )-plane.
Example: Fold or Cusp
x = c + dx x3 = F (c, d, x)

(51)

The vector field (51) is the most general perturbation of the function
x with lower order terms (because the term involving x2 can always be
3

62

eliminated by an appropriate translation of variable. At bifurcation points,

a differential equation must have a multiple equilibrium point, that is,
F (c, d, x) = 0 and

F (c, d, x)
x

=0

(52)

It follows:
c + dx x3 = 0 and d 3x2 = 0

(53)

The objective is to determine all values of c and d for which these two
equations can have some common solution x. It follows that
d = 3x2 and c = 2x3

(54)

If we eliminate x from these two equations, we obtain the following equation

of the cusp:
4d3 = 27c2

(55)

Elementary bifurcations in planar systems

Saddle-node bifurcation

Pitchfork bifurcation

x 1 = + x21
x 2 = x2

(56)

x 1 = x1 x31
x 2 = x2

(57)

Vertical bifurcation
x 1 = x1 + x2
x 2 = x1 + x2
Poincare-Andronov-Hopf bifurcation
x 1 = x2 + x1 ( x21 x22 )
x 2 = x1 + x2 ( x21 x22 )

(58)

(59)

Homoclinic or saddle-loop bifurcation

x 1 = x2
x 2 = x1 + x2 x21

(60)

In order to study the dynamics of these systems we proceed as follows:

1. Find the equilibrium points and study their dependence on parameters;
2. Find the system Jacobian for all equilibrium points;
3. Study how parameters influence the eigenvalues for all Jacobians; if
the eigenvalues do not have a zero real part the liniarization approximates
the non-linear system nearby the equilibrium point, else see the next lecture.
63

Home-works

1. Prove that the system (4) is structurally stable in the whole parameter
plane except for the b-axis and the positive part of the c-axis.
2. Using Lemma 1 verify that
a). for x = x the equilibrium point at 0 is asymptotically stable;
b). for x = x2 the equilibrium point at 0 is unstable;
c). for x = x x3 the equilibrium points at -1 and 1 are asymptotically
stable, and 0 is unstable.
3. Prove that () = is the function whose existence is guaranteed
by the IFT for the function F (, x) = + (1 + )x + x2 .
4. Apply the IFT to show that there is a unique solution of the equation
+ (1 )y + y 3 = 0 near (, y) = (1, 1).
5. Show that the conditions of the IFT are satisfied for the following
functions: a). + sin x, b). 1/2 + cos(/6 + x), c). sin + tan x.
6. Consider an RLC electric circuit which has a resistor, an inductor and
a capacitor all connected in series. The circuit is characterized by q, the total
charge in the plates of the capacitor and by i, the intensity of the current,
which are connected by the relation q = i:
q +

R
1
q +
q=0
L
LC

(61)

Study the type of the equilibrium and prove that the charge in plates will
eventually tend to zero.
7. The model describing the motion of water in a pipe is
k
(a) x = m
x

2g
x
L

(62)

2g
x
L

(63)

if we took friction into consideration, and

k
(b) x = m
x

if frictions are neglected. Study the structural stability in cases (a) and (b).
8. Prove the topological equivalence for the systems x = A1 x, x = A2 x,
x = A3 x, x R2 , where:

A1 =

1 0
0 2

; A2 =

and 1 < 0, 2 < 0, > 0, 6= 0, > 0

64

; A3 =

1
0

(64)

9. Establish the topological equivalence of the flows of the following two

linear systems and compute the positive definite matrices that define the
transformation of their orbits:

x =

1 0
0 2

x; x =

1 2
2 1

x;

(65)

10. For the following matrices, discuss the transition from one topological
equivalence class to another as the scalar parameter is varied:

(a)

, (b)

1
2
2 + 2

(66)

11. Let A() be a 2 2 real matrix depending on a real parameter , for

0 1 with A(0) = A0 and A(1) = A1 . For the following choices of A0
and A1 , must there be any bifurcations as the parameter is varied ?

1 0
0 1

(a) A0 =

(b) A0 =

(c) A0 =

1 0
0 1

1 0
0 2

A1 =

A1 =

A1 =

1 0
0 1

1 0
0 1

(67)

1 1
0 1

(68)
!

(69)

12. Consider the following vector fields depending on two parameters:

x = 1 + 2 x + x2

(70)

where = (1 , 2 ) are two small parameters. Show that:

a. the function () corresponds to the minimum value of the function
F (, x) = 1 + 2 x + x2 and this is given by:
1
(1 , 2 ) = 1 22
4
b. the bifurcations occur as we cross the curve 1 = 22 /4 in the (1 , 2 )plane: there are two equilibrium points if 1 < 22 /4, and none if 1 > 22 /4.
13. Consider the following perturbation of the vector field f (x) = x2
given by :
x = 2 + 2ax + x2
(71)
65

where is a scalar parameter and a is a given constant. Show that:

a.
() = 2 (1 a2 )

(72)

b. The bifurcation takes place at = 0 and the flow is different depending

on whether |a| < 1 or |a| > 1.

Lab-works

1. For (30)-(33) draw the bifurcation diagrams in the plane (

x, ). Follow
the following procedure: 1). Draw the graph (x, F (, x)) for different values
of 0 ; 2). read the approximate values of x so that F (0 , x)
= 0; 3). use
the Matlab function f zero in order to find the equilibrium values x so that
F (0 , x) = 0 for the chosen 0 and construct the set E0 , or equivalently the
set of the initial states (0 , x0 ) = (0 , x) for (36); 4). integrate (36) using the
Matlab function ode23 from the determined initial states E0 , from t0 < 0
to tf > 0 - take small values, for example t0 = 0.1 and tf = 0.1.
2. Show that the following two linear systems are topologically equivalent
but they are not linearly equivalent.

x =

2 0
0 2

x; x =

2 1
0 2

x;

(73)

Draw the phase portraits of these two linear systems and compare.
3. Consider the following five coefficient matrices A of the linear systems
x = Ax:

!
1 0
1 1
1 0
,
,
0
1
0 1
0 2
!
!
(74)
0
2
1 1
1 3
0 2
(a) For each A above, determine a symmetric matrix CA satisfying the
matrix equation AT CA + CA A = I.
(b) Draw the level set {x : xT CA x = 1} for each of your CA .
4. Obtain the bifurcation curves of the following one and two-parameter
perturbations of the cubic vector field f (x) = x3 and sketch some representative phase portraits:
x = 1 + 1 + 21 x x3
x = 1 + 22 + 2 x x3
66

(75)

Lecture notes 2013/2015

Dynamical systems, 5
Mihaela Cistelecan, UT Cluj-Napoca
These notes were prepared using:
1. Hale, J.K., Kocak, H. Dynamics and Bifurcations. Springer-Verlag,
1991.

The Stability and Bifurcations of nonhyperbolic equilibria of planar systems

Since:
1). the stability is critical (the linearization fails to determine the stability
properties) when the equilibrium point is nonhyperbolic
2). the parameter varying systems undergo bifurcations when their dynamics encounters a nonhyperbolic equilibrium
Lets assume a planar non-linear system with eigenvalues 1 , 2 so that
1 2 = 0. Therefore the Jacobians determinant is zero. In this case any
perturbation from the nonlinear terms (HOT from linearization) matters.
An example - lets investigate what is going on in a small vicinity of the
equilibrium point (0, 0) for the following system
"

x =

0 0
0 1

#"

x1
x2

"

x1 x2
x22

(1)

If we re-write (1) by introducing the nonlinear terms as perturbations (|x1 |

0, |x2 | 0) into the Jacobian matrix we get:
"

x =

0 + x2
0
0
1 + x2

67

#"

x1
x2

(2)

Thus, the determinant of the above matrix is not zero anymore and the
stability of the equilibrium (0,0) should be thoroughly investigated.
Therefore, a further study of the nonhyperbolic equilibria is necessary.
We study two cases of nonhyperbolic equilibrium, as follows:
The linearized vector field has one zero and one negative eigenvalue
The linearized vector field has two imaginary eigenvalues
In these two cases there is a chance that the system might be stable when
taken into account the high order terms which were removed when we performed linearization. The case when the linearized vector field has one zero
and one positive eigenvalue leads to an unstable system.
Thus, for both types of nonhyperbolic equilibria we study the following
two problems:
Local stability about the origin of the state plane of the system x =
f (x)
Bifurcations of the system x = F (, x)
We prove that the local dynamics and bifurcations of planar systems can
be determined from those of an appropriate scalar differential equation.

1.1

The linearized vector field has one zero and one

negative eigenvalue

a. The study of the LOCAL STABILITY in a neighborhood of the

origin
Let f be a continuous and differentiable function:
f : R2 R2 ; x f (x)

(3)

f (0) = 0; Df (0) = 0;

(4)

satisfying:

where Df is the Jacobian of the function f .

We assume the planar system as follows:
x 1 = f1 (x1 , x2 )
x 2 = x2 + f2 (x1 , x2 )
68

(5)

To bring the linear part of this system to the forefront, we write (5 ) in vector
notation, as follows:
"

x =

0 0
0 1

x + f (x)

(6)

Note that the linear part of the vector field about the equilibrium
point at the origin is in Jordan Normal Form with eigenvalues 0
and 1.
In applications, the linearization of a vector field with one zero and one
negative eigenvalue may not always come in normal form (5); however, such
a vector field can always be put into this form with a linear change of coordinates.
Idea: In order to study the stability of the origin of eq. (5) we reason as
follows:
If x(t) = (x1 (t), x2 (t)) is a solution of eq. (5) with initial values close
to zero, then the variation of x1 (t) occurs more slowly than the variation of
x2 (t). Thus, it is reasonable to consider x1 as constant in the second equation,
and consider x2 in the first equation as the zero (x1 ) of x2 + f2 (x1 , x2 ).
As a result, the dynamics of eq. (5) in a neighborhood of the origin should
be determined by the scalar differential equation x 1 = f1 (x1 , (x1 )).
Therefore, the following theorem was proven to be true:
Theorem 1.1.1.: Let (x1 ) be a continuous function, defined as follows:
: {x1 : |x1 | < } {x2 : |x2 | < }; > 0
(x1 ) + f2 (x1 , (x1 )) = 0

(0) = 0; x
(0) = 0;
1

(7)

Then the equilibrium x = (

x1 , x2 ) = (0, 0) of eq. (5) is stable (resp., asymptotically stable, unstable) if and only if the equilibrium x
1 = 0 of the scalar
differential equation:
x 1 = f1 (x1 , (x1 ))
(8)
is stable (resp., asymptotically stable, unstable).
Thus, we identify the following three steps we should consider in the
stability analysis:
69

1 We look for the Taylor expansion of (x1 ) about the origin so that (0) = 0
3
4
and 0 (0) = 0
1 + a2 x1 + a3 x1 + . ...
(x1 ) = a1 x2
2 We replace this expansion into the equation (x1 ) + f2 (x1 , (x1 )) = 0
and find the unknown coefficients of the expansion.
3 Using f1 (x1 , (x1 )) in the expression of x 1 - eq. (9) - we draw conclusions about the stability of the second order system.
Example: Consider the system:
x 1 =ax31+x1 x2
x 2 = x2 + x2 2 +x1 x2x3

(9)

where a is a given real constant. The function (x1 ) should be a solution of

(x1 ) + ((x1 ))2 + x1 (x1 ) x31 = 0

(10)

Substituting a Taylor series for (x1 ) and equating the coefficients of like
powers of x1 , we obtain:
(x1 ) = x31 + O(|x41 |)

(11)

f1 (x1 , (x1 )) = ax31 x41 + O(|x51 |)

(12)

and
It follows that the equilibrium x = (
x1 , x2 ) = (0, 0) is asymptotically stable
if a < 0 and unstable if a 0.
b. The study of bifurcations
Let F be a continuous and k-differentiable function, depending on parameters :
F : Rm R2 R2 ; (, x) F (, x)
(13)
satisfying:
F (0, x) = f (x); f (0) = 0; Df (0) = 0;

(14)

Below, we consider the system of differential equations:

x 1 = F1 (, x1 , x2 )
x 2 = x2 + F2 (, x1 , x2 )
70

(15)

where F = (F1 , F2 ), and determine the nature of the bifurcations of equilibria near the origin for small values of the parameter .
Lemma 1.1.1. If F is a continuous function (C k ) , k 1 then there are
constants 0 > 0, > 0 and a C k function
: { : kk < 0 } {x1 : |x1 | < } {x2 : |x2 | < }
satisfying
(0, 0) = 0,

(0, 0)
x1

=0

(16)
(17)

with kk
< 0 , a point (
such that, for each ,
x1 , x2 ), with |
x1 | < and
x1 ) and
|
x2 | < , is an equilibrium point of eq. (15) if and only if x2 = (,

F1 (, x1 , (, x1 ) = 0
Definition: Let be as in Lemma 1.1.1. Then, the function G(, x1 )
defined by
G(, x1 ) = F1 (, x1 , (, x1 ))
(18)
is called the bifurcation function and the equation
G(, x1 ) = 0

(19)

is referred to as the bifurcation equation of the system (15).

It is obvious that the equilibrium points of the scalar differential equation:
x 1 = G(, x1 )

(20)

are in one-to-one correspondence with the equilibrium points of the planar

system (15). It turns out that the stability types of the corresponding equilibria of eq. (15) and (20) agree as well.
If the equilibrium points of the scalar differential equation (20) are hyperbolic, the following theorem establish the stability types of the corresponding
equilibria of the planar system (15).
x1 ))
Theorem 1.1.1.: Let (, x1 ) be as in Lemma 1.1.1 and x = (
x1 , (,

is an equilibrium point of eq. (15) with kk < 0 , |

x1 | < , and |(, x1 )| < .
Then 0 and can be chosen small enough so that
(i) x is a hyperbolic stable node if

x1 )
G(,
x1

71

<0

(ii) x is a saddle point if

x1 )
G(,
x1

> 0.

When the equilibrium points of (20) are not hyperbolic we use the following Theorem:
Theorem 1.1.2.: Let (, x1 ) be defined as in Lemma 1.1.1 and, for
let x = (
x1 )) be an equilibrium point of eq. (15).
a given small ,
x1 , (,
Then the equilibrium point x is stable (respectively, asymptotically stable,
unstable) if and only if the equilibrium x1 of the scalar differential equation
x1 )
x 1 = G(,

(21)

is stable (respectively, asymptotically stable, unstable).

Example: Damped pendulum with torque. Consider the following secondorder differential equation describing the motion of a planar pendulum in the
presence of damping and constant torque M:
+ + sin = M

(22)

If the rod of the pendulum is rigid, then one can interpret the application of the torque as pushing the pendulum with constant force that is
perpendicular to the rod.
By choosing y1 = and y2 = the phase plane model is:
y1 = y2
y2 = y2 sin y1 + M

(23)

Since sin y1 is periodic with period 2, we confine our discussion to the

values of y1 in the interval [, ].
The equilibrium points of (23) are given by y2 = 0 and sin y1 = M .
Discussion: For M > 1 there are no equilibrium points;when M = 1
there is a single equilibrium point at (/2, 0); for M < 1 there are two equilibrium points. It appears that the equilibrium point at (/2, 0) undergoes
a bifurcation at the parameter value M=1.
To analyze the behavior of the orbits of eq. (23) near (M = 1; (x1 , x2 ) =
(/2, 0)), let us use the translation of variables:
M = 1 + , y1 =

72

+ z1 , y2 = z2

(24)

and transform eq. (23)to the system:

z1 = z2
z2 = z2 cos z1 + 1 +

(25)

so that for = 0 the origin is an equilibrium point. The coefficient matrix

of the linearization of eq. (25) at the origin is :
"

0 1
0 1

To put this matrix into Jordan Normal Form, observe that its eigenvalues
are 0 and -1 with the corresponding eigenvectors (1, 0) and (1, 1). Therefore, if we make the transformation of variables
"

z = N x, N =

1 1
0 1

(26)

then, eq. (25) becomes:

"

x =

0 0
0 1

"

x+N

0
cos(x1 x2 ) + 1 +

(27)

or equivalently,
x 1 = cos(x1 x2 ) + 1 +
x 2 = x2 cos(x1 x2 ) + 1 +

(28)

The system (28) is now in the normal form (15).

In order to compute the bifurcation function we need to determine the
function (, x1 ) satisfying:
(, x1 ) cos(x1 (, x1 )) + 1 + = 0

(29)

Since we are interested in local bifurcations, it suffices to determine a

few terms of the Taylor series about (, x1 ) = (0, 0). Therefore, consider the
power series of given by:
(, x1 ) = c10 + c20 2 + c11 x1 + c02 x21 + O((|| + |x1 |)3 )

73

(30)

where the coefficients cij are to be determined if we substitute this expression

into (29) and use the Taylor series of the cosine. A somewhat laborious
calculation yields:
1
1
(, x1 ) = + 2 x1 + x21 + O((|| + |x1 |)3 )
2
2
Therefore, the bifurcations are given by:
1
1
x 1 = + 2 x1 + x21 + O((|| + |x1 |)3 )
2
2

(31)

(32)

Discussion: We analyze the bifurcations of equilibria of (32) as follows.

If = 0 the system (28) undergoes a saddle-node bifurcation; when < 0
there are two equilibria, one an asymptotically stable node and the other a
saddle, and there are no equilibria when > 0.
c. Center Manifolds (optional reading)
In order to determine fine structures of flows and bifurcations near an
equilibrium point at which the matrix of linear approximation has one zero
and one negative eigenvalue we apply the theory of center manifold. It has
been proven that there is some invariant curve - local center manifold
- tangent to the line containing the eigenvectors corresponding to the zero
eigenvalue of the linearized vector field. Since the other eigenvalue is negative, all orbits starting near the origin approach this invariant curve. The
qualitative behavior of the local flow on the plane can then be determined
from the flow of an appropriate scalar diff. eq. on the center manifold.
To utilize the theory of center manifolds in applications, we now show how
to compute them in specific equations. For practical purposes, it suffices to
have a procedure for the computation of several terms of the Taylor series of
a function h(x1 ) defining a center manifold. If we differentiate with respect
to t the defining equation x2 (t) = h(x1 (t)), then we obtain:
x 2 =

h
(x1 )x 1
x1

(33)

If we now substitute the expressions for x 1 and x 2 from (5) then a solution
of the partial differential equation
74

h(x1 ) + f2 (x1 , h(x1 )) =

h
(x1 )f1 (x1 , h(x1 ))
x1

(34)

subject to the initial values:

h(0) = 0;

h
(0)
x1

=0

(35)

yields a center manifold defined by h(x1 ). The partial differential equation

(34) cannot, of course, be solved for h in most cases. Therefore, we opt for a
sufficiently accurate approximate solution to facilitate a local analysis of the
flow. To accomplish this, we expand h(x1 ) into a power series in the variable
x1 as follows:
h(x1 ) = c2 x21 + c3 x13 + O(|x41 |)

(36)

where the coefficients ci are to be determined. We now substitute this

series into the partial diff. eq. (34) and determine ci by equating the coefficients of like terms. Although center manifolds are not unique, the power
series method above always yields a unique Taylor series, if it exists. It turns
out that all center manifolds have the same Taylor expansions near the origin.
Example: Consider the system:
y = yz
z = z + ay 2

(37)

The center manifold equation is:

h
(yh(y)) + h(y) ay 2 = 0
y
where
h(0) =

h
(0) = 0
y

(38)

(39)

We try the center manifold in the form: h(y) = h2 y 2 + O(|y|3 ) and calculate
h2 by matching coefficients of y 2 in (38). We get h2 = a. Therefore the
reduced system is:
y = ay 3 + O(|y|4 )

75

(40)

Therefore, the origin of (37) is asymptotically stable if a < 0 and unstable if

a > 0.
The center manifold concept can be extended to parameter varying systems... some other time....
end of optional reading

1.2

The linearized vector field has two imaginary eigenvalues

When the eigenvalues of the linearized vector field at an equilibrium point

are purely imaginary, the local dynamics about the equilibrium point cannot
be determined by the linear approximation. Depending on the nonlinear
terms, the equilibrium can be unstable, stable or even asymptotically stable.
Consequently, we need to investigate the effects of the nonlinear terms in
each particular situation.
The systems we consider in this section are in the following form:
"

x =

0
1
1 0

x + f (x)

(41)

Notice that the linear part of the vector field about the equilibrium point at the origin is in Jordan normal form with eigenvalues
i.
When the vector field is subjected to small perturbations, the original
equilibrium point persists, and there can be no new equilibria in the neighborhood. However, if the eigenvalues of the linearized system move away
from the imaginary axis, one expects the equilibrium point to change its stability type. This change is typically marked by the appearance of a small
periodic orbit encircling the equilibrium point.
In order to investigate the dynamics of the system (41) in a sufficiently
small neighborhood of the origin we introduce polar coordinates (43).
Example: Consider the planar system:
x 1 = x2 + ax1 (x21 + x22 )
x 2 = x1 + ax2 (x12 + x22 )
76

(42)

where a is a given real number. Regardless of the value of the constant

a, the origin is an equilibrium point and the eigenvalues of the linearization
at the origin are i. If we introduce polar coordinates (r, ) defined by:
x1 = rcos , x2 = rsin

(43)

r = ar3
= 1

(44)

then (42) becomes:

Since > 0, the orbits spiral monotonically in around the origin. Therefore, the stability type of the origin of (42) is the same as that of the equilibrium point r = 0 of the radial equation r = ar3 . Therefore, r = 0 is
asymptotically stable if a < 0, stable at a = 0, and unstable for a > 0.
In the above case the dynamics of r(t) and (t) was not coupled, but what
if we deal with the case where the system (41) becomes:
r = R(r, )
= 1 + (r, )

(45)

where, for r 6= 0:
R(r, ) = f1 (rcos, rsin)cos f2 (rcos, rsin)sin
(r, ) = 1r (f1 (rcos, rsin)sin + f2 (rcos, rsin)cos)

(46)

When r is sufficiently small, the discussion of the orbits of eq. (41) can
be reduced to the discussion of solutions of a 2-periodic scalar differential
equation. If we can eliminate t in (45) we obtain an equation for r as a
function of through the differential equation:
dr
R(r, )
=
d
1 + (r, )

(47)

which, from (46), is a 2-periodic function and satisfies R(0, ) = 0. The

solutions of (47) give the orbits of (41).
We can recover the solutions of (41) as a function of time from the solutions of (47) by the following steps below:

77

1. Fix r0 and find the solution r(, r0 ) of eq. (47) satisfying the initial
value r(0, r0 ) = r0 . The orbit of (41) through the point x(0) = (r0 , 0) is then
given by:
(x(0)) = {(x1 , x2 ) : x1 = r(, r0 )cos , x2 = r(, r0 )sin , 0 < }
(48)
2. Find the solution (t) of the initial-value problem
= 1 + (r(, r0 ), ); (0) = 0;

(49)

3. The solution x(t) of eq. (41) through the point x(0) = (r0 , 0) is then
given by:
x1 (t) = r((t), r0 )cos (t)
(50)
x2 (t) = r((t), r0 )sin (t)
Why can we eliminate t ? - Since |(0, )| < 1 for all , and (r, ) is
continuous, we can choose > 0 such that 1 + (r, ) > 0 for all and
|r| < . Consequently, in a neighborhood of the origin, we have > 0. The
implication of this is that the orbits of eq. (45) spiral monotonically in
around the origin.
Concluding - using polar coordinates, we capture the dynamics
of (41) in the neighborhood of the equilibrium point in terms of
the dynamics of an appropriate nonautonomous scalar differential
equation with periodic coefficients.
So, what if, for example:

or, even worse:

dr
= r3 + sin()
d

(51)

dr
= cos()r3 + sin()
(52)
d
How could we study the stability of r = 0 in these cases ?
Note that (52) is worse than (51) because of the periodic coefficient that
multiplies the variable r. We may apply a transformation and replace the
periodic coefficient by a constant one - see below.
Then, the following should be performed:
78

1). The periodic coefficients from the equation dr/d should be transformed into constant coefficients, using the transformations from Appendix
3.
2). The Poincare map for dr/d should computed. The derivative of the
Poincare map decide on the stability of the scalar nonautonomous system
dr/d which is equivalent to the stability of the original planar system.
STABILITY analysis
Lemma: There is a bounded neighborhood U of the origin in R2 such
that each periodic orbit of eq. (41) lying in U encircles the origin; also, if
x(0) = (r0 , 0) with r0 > 0, then the solution r(, r0 ) of eq. (47) satisfying
the initial value r(0, r0 ) = r0 is 2-periodic in . Conversely, if r(, r0 ) is a
2-periodic solution of (47), then the orbit (x(0)) with x(0) = (r0 , 0) of the
planar system (41) is a periodic orbit. The minimal period T of such a is
the first value of t for which the solution (t) of eq:
= 1 + (r(, r0 ), ), (0) = 0

(53)

(T ) = 2

(54)

satisfies:
Example: Consider the damped oscillator:
y + y 2 y + y = 0

(55)

x 1 = x2
x 2 = x1 x21 x2

(56)

or, equivalently the system:

Using polar coordinates we get:

dr
1
1
= cos2 sin2 r3 +O(r4 ) = (1cos(4))r3 +O(r4 )
= (1cos(4))r3
d
8
8
(57)
We need a transformation to convert the scalar equation dr/d to the eq.
d/d for which the lowest order term in the Taylor expansion of the vector
field has a nonzero constant coefficient.
In order to make several terms of the Taylor series independent of we
make the following transformation:
r = + a()3
79

(58)

where a() is a 2-periodic function to be determined. In the new variable

the system becomes:
d
1
1
= (1 cos(4) + 8a0 ())3 + O(4 )
= (1 cos(4) + 8a0 ())3 (59)
d
8
8
To make the coefficient of the 3 term independent of we choose:
1
a0 () = cos(4)
8
Therefore,

d
1
= 3 + O(4 )
(60)
d
8
From (60) it turns out that = 0 is asymptotically stable, hence r = 0 is
asymptotically stable. The Poincare map is:
Y

1
1
(0 ) = 0 30 + O(40 )
= 0 30
8
8

Remark: In order to compute the Poincare map approximate d/d '

Q
d/dt then integrate (60) in order to find (0 ) = (t = 1, t0 = 0, (0)),
where (t, t0 , (0)) is the solution of (60):
Y

(0 ) = 0 +

1.3

Z t=1
t=0

1
1
1
( 03 + O(04 ))dt ' 0 03 + O(04 )
= 0 03
8
8
8

Orbitally stable

In the following we give the definitions for orbitally stable.

We consider a periodic orbit of:
x 1 = x2 + f1 (x1 , x2 )
x 2 = x1 + f2 (x1 , x2 )

(61)

as a closed curve and neglect its time parametrization. In this context,

we define the distance of a point x R2 to a periodic orbit , denoted by
dist(
x, ) as follows:
dist(
x, ) min{k
x xk, f or all x }

(62)

Definition: A periodic orbit of the planar system (61) is said to be

orbitally stable if, for any > 0, there is a > 0 such that dist(x0 , ) <
80

implies that dist((t, x0 ), ) < for all t 0. The periodic orbit is said
to be orbitally unstable if it is not orbitally stable.
Definition: A periodic orbit of (61) is said to be orbitally asymptotically stable if it is orbitally stable and, in addition, there is a b0 such
that dist(x0 , ) < b implies that dist((t, x0 ), ) 0 as t + that is,
(x0 ) .
BIFURCATIONS - Poincare-Andronov-Hopf Bifurcation
In order to study the bifurcations in the neighborhood of a nonhyperbolic
equilibrium point with nonzero purely imaginary eigenvalues the PoincareAndronov-Hopf Theorem is applied. This theorem may be found in the
literature. We only give a simplified version and some important remarks
in the following. Under small perturbations of the vector field, an equilibrium point may persists. However, if the stability type of the equilibrium
changes when subjected to perturbations, then this change is usually accompanied with either the appearance or disappearance of a small periodic orbit
encircling the equilibrium point.
Example: Consider the special planar system of the form:
x 1 = x2 + F (, r2 )x1
x 2 = x1 + F (, r2 )x2

(63)

where is a scalar parameter, r2 = x21 + x22 and F (0, 0) = 0 so that

the origin is an isolated equilibrium point. In polar coordinates, this system
becomes:
r = F (, r2 )r
(64)
= 1
The existence and stability properties of periodic solutions of eq. (63)
are the same as the corresponding equilibrium points (and their stability
types) of the scalar differential eq. r = F (, r2 )r. Indeed, if either r = 0
or F (, a2 ) = 0, then (a cost t, a sint t) is a 2- periodic solution of eq.
(63) with amplitude a.The bifurcation diagram for periodic solutions of (63)
is simply a plot of the solutions of F (, a2 ) = 0 in the (, a)-plane.
In the following we study several specific forms for F (, r).
1 F (, r2 ) = - we obtain a linear perturbation of the linear harmonic
oscillator. There is no nontrivial periodic orbit except at = 0, at
81

which case there is one periodic orbit for each amplitude a. All periodic
orbits are orbitally stable.
2 F (, r2 ) = r2 - there is a unique nontrivial
periodic orbit if > 0 for

a particular value of a; namely a = . The periodic orbit is orbitally

asymptotically stable. Because the bifurcation curve emanates from
the origin to the right, the bifurcation is called supercritical.
3 F (, r2 ) = (r2 c)2 + c2 + with c > 0 a fixed constant: there are
two nontrivial periodic orbits, one orbitally unstable and the other
orbitally asymptotically stable, for c2 < < 0 with amplitudes c
[( + c2 )1/2 ]1/2 . The two periodic orbits coalesce and disappear as
decreases through c2 . There is only one periodic orbit for > 0
and it is orbitally asymptotically stable. Because the bifurcation curve
emanates from the origin to the left, the bifurcation is called subcritical.
Example: - Van der Pols oscillator:
x 1 = x2
x 2 = x1 + 2x2 x21 x2

(65)

The eigenvalues ofthe linearization of (65) about the equilibrium point

at the origin are i 1 2 . For < 0 the origin is asymptotically stable
because the real parts of the eigenvalues are negative. At = 0, the origin
is still asymptotically stable. For > 0 the real parts of the eigenvalues
become positive and thus the origin is unstable. Using polar coordinates
we may show that, as the eigenvalues cross the imaginary axis, the origin
gives up its stability to a periodic orbit. Moreover, for each small > 0
there is a unique nontrivial periodic orbit near the origin which is orbitally
asymptotically stable.
After some computations (in the manner given in the previous section) it
turns out:

2rsin2 r3 cos2 sin2

dr
=
(66)
d
1 + 2sincos r2 cos3 sin
The Taylor series for || and r small is as follows (see http : //en.wikipedia.org/wiki/Rationa
dr
d

= [ cos(2) + O(2 )]r 81 [1 cos(4) + O()]r3 + O(r4 )

=

= [ cos(2)]r 18 [1 cos(4)]r3
82

(67)

which is a periodic scalar equation.

Now, we should apply the transformation theory of periodic scalar equationsto make the coefficients of the powers of r up through r3 independent
of . After the following transformations (we will not give the explanation
of the first transformation here, see literature for further explanations):
r = e(/2)sin(2)
+ a()3
da
= 18 cos(4)
d

(68)

we get:
d
1
1
= ( + O(2 )) ( + O())3 + O(4 )
= () ( )3
d
8
8
Q
The Poincare map of (69) is:
1
1
(0 ) = 0 + ( + O(2 ))0 ( + O())03
= 0 + 0 ( )30
8
8

(69)

(70)

To find the periodic solutions of (69) near origin, we need to

locate the fixed points of its Poincare map, that is to solve the eq.:
(0 ) = 0
For (70) this requires:
1
1
( + O(2 ))0 ( + O())30 + O(40 )
= 0 30 = 0
8
8
The solution 0 = 0 of this eq. corresponds to the equilibrium point at
the origin. The other solution satisfies the eq.:
1
1
+ O(2 ) ( + O())20 + O(30 )
= ( )20 = 0
8
8
and represents the bifurcation curve in the (, 0 )-plane. This curve has the
property that there are periodic orbits of the equation with approximate
amplitude 0 if and only if (, 0 ) lie on this curve.

83

1.4

Appendix 1 - Nonautonomous scalar systems with

periodic coefficients

For a nonautonomous system

f : R R R x = f (t, x)

(71)

we denote the solution through x0 at t0 by (t, t0 , x0 ) with (t0 , t0 , x0 ) = x0 .

Therefore, for such systems we must consider the trajectories in R R. The
trajectories are uniquely defined by (t0 , x0 ). When the function f (t, x) is
periodic in t,the qualitative behavior of solutions is very similar to the
autonomous case if we replace equilibrium points with periodic solutions.
Eq. (71) is equivalent with the autonomous system:
= 1
x = f (, x)

(72)

Remark: - In order to find the solution of this kind of equations you may
either apply the variation of constants formula (http : //en.wikipedia.org/wiki/
M ethod of variation of parameters) or may recur to Mathematica / Matlab / Maple .
Thus, the main problems here are
how to establish the existence of periodic solutions
how to investigate the stability of periodic solutions
Answer - We use Poincare maps !
Consider the differential equation, where f (t, x) is 1 periodic in t:
x = f (t, x) with f (t + 1, x) = f (t, x)

(73)

For simplicity of notation, we assume the period to be 1; if not, we can

always rescale t to make it so.
Lemma: A solution (t, t0 , x0 ) of a 1-periodic differential equation x =
f (t, x) is 1-periodic if and only if (t0 + 1, t0 , x0 ) = x0 .
Definition: A periodic solution (t, t0 , x0 ) of (73) is said to be stable if,
for any given > 0, there is a > 0 (depending only on and not on t0 ) such
that for any y0 for which |y0 x0 | < the solution (t, t0 , y0 ) satisfies the
inequality |(t, t0 , y0 ) (t, t0 , x0 )| < for all t t0 . The periodic solution
84

(t, t0 , x0 ) is said to be unstable if it is not stable.

Definition: A periodic solution (t, t0 , x0 ) of (73) is said to be asymptotically stable if it is stable and, in addition, there is an r > 0, independent
of t0 , such that |(t, t0 , y0 ) (t, t0 , x0 )| 0 as t + for all y0 satisfying
|y0 x0 | < r.

1.5

Appendix 2 - Poincare Map

Definition: The Poincare map of a 1-periodic differential eq. is the scalar

mapping
: R R x0 (1, 0, x0 )

(74)

Thus, the Poincare map takes the initial value x0 at t0 = 0 to the value of
the solution (t, 0, x0 ) at t=1.
The k th iterate of x0 under is given by
k (x0 ) = (k, 0, x0 )

(75)

Useful properties of the Poincare map:

the Poincare map is monotone
the Poincare map is differentiable with nonnegative derivative
0 for all x0

Q0

(x0 )

is asymptotically stable if 0 (x0 ) < 1 and unstable if 0 (x0 ) > 1

In order to understand the above properties and how the Poincare
map measures some properties of the continuous systems please
be aware of the following facts:
1 A continuous-time system is given by one of the following dynamics:
x = f (x); x = f (t, x); x = f (t, x, u(t))

(76)

A discrete-time system is a map (see it as an iteration) written in

the following form:
x(k + 1) = f (x(k)); x(k + 1) = f (k, x(k)); x(k + 1) = f (k, x(k), u(k))
(77)
85

or, equivalently,
x(t + 1) = f (x(t)); x(t + 1) = f (t, x(t)); x(t + 1) = f (t, x(t), u(t))
(78)
where t should be understood as a time instant.
2 For an unforced scalar linear discrete time system x(k + 1) = ax(k) the
stability analysis is based on the following recurrence:
x(k + 1) = ax(k) = a2 x(k 1) = ....ak+1 x(0)

(79)

Thus, if |a| < 1, x(k + 1) 0 and x = 0 is stable, otherwise the

sequence x(k) diverges as time (k-instant) increases.
3 The asymptotic behavior of the solution (t, 0, x0 ) can be studied by looking at the following sequence of functions on the interval [0, 1]:
: [0, 1] R k = 0, 1, 2, ... def ined by

(80)

0 (t) = (t, 0, x0 )
1 (t) = (t + 1, 0, x0 )
...
k (t) = (t + k, 0, x0 )

(81)

0 t 1. See the dynamics given by the iterations (81) as the

sequence:
0 (t) 1 (t) 2 (t)....
From math - the uniqueness of solutions and the periodicity in the
differential equation imply that this sequence of functions is monotone,
that is, if 1 (0) 0 (0) = x0 , then k+1 (t) k (t) (or, if 1 (0)
0 (0) = x0 then k+1 (t) k (t)).
4 In order to intuitively understand the relation between 0 (x0 ) and stability (avoiding the linear variational equation) please reason as follows:

0 (x0 ) = x
(1, 0, x0 ) which may be understand as a small deviation of
0
the solution due to a small variation of the initial condition. Thus, we
are interested in how the distance between (t, 0, x0 ) and (t, 0, x0 + )

86

- |(t, 0, x0 + ) (t, 0, x0 )| - behaves in time, at t = 0, t = 1, .... From

the reasoning for discrete-time systems given above, we may write:
|(t, 0, x0 + ) (t, 0, x0 )| = a|(t 1, 0, x0 + ) (t 1, 0, x0 )| (82)
where
|a| > 1 if the difference increases in time
|a| < 1 if the difference decreases in time
Now, if we replace t = 1 in (82) we get
|(1, 0, x0 +)(1, 0, x0 )| = a|(0, 0, x0 +)(0, 0, x0 )| = a|x0 +x0 |
(83)
or

0 (x0 ) =
(1, 0, x0 ) = a
(84)
x0
and from above, a > 1 stands for an unstable periodic solution and
a < 1 stands for a stable periodic solution.
Theorem: If a solution (t, 0, x0 ) of a 1-periodic differential eq. (73) is
bounded for t 0, then there is a 1-periodic solution (t) of eq (73) such
that:
(t + k, 0, x0 ) (t) as the integer k +

(85)

monotonically and uniformly for 0 t 1. Similarly, if (t, 0, x0 ) is

bounded for t 0, then there is a 1-periodic solution (t) of (73) such that
(t k, 0, x0 ) (t) as the integer k +

(86)

monotonically and uniformly for 0 t 1.

A restatement of the previous Lemma:
A point x0 is the initial value of a 1-periodic solution of (73) if
and only if x0 is a fixed point of the Poincare map, that is, (x0 ) = x0
In addition The stability properties of a 1-periodic solution of eq. (73) are the
same as the stability properties of the corresponding fixed point of
the Poincare map
Example: Consider the 1-periodic differential equation:
87

x = x + cos 2t 2sin 2t

(87)

Any solution of this linear equation satisfying x(0) = x0 is given by:

(t, 0, x0 ) = et (x0 1) + cos 2t

(88)

It turns out that there is a unique periodic solution (x0 = 1) and it is

asymptotically stable. This observation is also reflected in the Poincare map:
(x0 ) = e1 (x0 1) + 1

(89)

there is a unique fixed point of this map at x0 = 1:

(1) = 1

(90)

which corresponds to the periodic (t, 0, 1) = cos 2t. Since

0 (1) = e1 < 1

(91)

the fixed point, hence the corresponding periodic solution, ia asymptotically

stable.
It appears that we need an explicit formula for the DERIVATIVE of the Poincare map which is difficult to compute unless the
general solution is available. To circumvent this difficulty, in the following we derive a formula (theorem) for the derivative of the Poincare map
in terms of only the 1-periodic solution and the vector field f (t, x):
Lemma: Let (t, 0, x0 ) be the solution of a 1-periodic equation
x = f (t, x) with (0, 0, x0 ) = x0 and be the Poincare map. Then
the derivative of the Poincare map is given by
0

Z 1
f

(x0 ) = exp(

(t, (t, 0, x0 ))dt)

(92)
x
Theorem: Let (t, 0, x0 ) be a 1-periodic solution of the 1-periodic eq.
x = f (t, x). Define:
Z 1
f
a0 =
(t, (t, 0, x0 ))dt
(93)
0 x
Then,
0

1 (t, 0, x0 ) is asymptotically stable if a0 < 0, and

88

2 (t, 0, x0 ) is unstable if a0 > 0

Example: Consider the eq.
x = x3 + sin3 2t + 2cos 2t

(94)

and its 1-periodic solution (t, 0, 0) = sin2t. It turns out that:

a0 =

Z 1
0

3sin2 2tdt = 3/2 < 0

(95)

therefore, the 1-periodic solution is asymptotically stable.

DefinitionA periodic solution (t, 0, x0 ) is called hyperbolic if 0 (x0 ) 6=
1.
Example: Consider the eq.
x = x2 2sin 2t cos2 2t

(96)

Its periodic solution (t, 0, 1) = cos 2t. This periodic solution corresponds to the fixed point of the Poincare map at x0 = 1. Also:
0

a0

(1) = e

Z 1

= exp(

2cos2tdt) = 1

(97)

therefore, the periodic solution is not hyperbolic and thus its stability type
cannot be determined from the liniarization of the Poincare map.

Appendix 3 - Transformation for the periodic coefficients

The idea is to determine an appropriate transformation of the variable r(, )

dr
so that the 2-periodic differential equation d
is transformed to one with
constant coefficients. Once this is accomplished , the approximate form of the
Poincare map, and thus the corresponding bifurcation curve, can be written
down rather easily.
Lets write:
X
dr +
=
Rk (, )rk
d k=1

where the coefficients Rk (, ) are 2-periodic functions in .

89

(98)

Next, we introduce the new variable defined by:

r =+

+
X

bk (, )k

(99)

k=1

where b1 (0, ) = 0 and bk (, ) are 2-periodic functions to be determined.

In new variable the differential eq. (98) has the form:
X
d +
bk
=
[ck (, )
(, )]k
d k=1

(100)

Now choose the functions bk so that

bk
(, ) = ck (, ) ck ()

where ck is the aaverage of ck given by:

1 Z 2
ck () =
ck (, s)ds
2 0
Thus, (98) formally reduces to:
X
d +
=
ck (, )k
d k=1

(101)

(102)

(103)

Lemma: The formal power series (103) contains only the odd powers of
, that is
X
d +
=
c2k+1 ()2k+1
(104)
d k=1
In most applications we need to compute only the first few bk .

Home-works

1. For each of the following systems, use a quadratic Lyapunov function

candidate to show that the origin is asymptotically stable:
x 1 = x1 + x22
x 2 = x2
x 1 = x1 + x21 x2
x 2 = x2 + x1

x 1 = (x1 x2 )(x21 + x22 1)

x 2 = (x1 + x2 )(x21 + x22 1)
x 1 = x1 x2
x 2 = x1 x32
90

(105)

2. Using V (x) = x21 + x22 , study stability of the origin of the system:
x 1 = x1 (k 2 x21 x22 ) + x2 (x21 + x22 + k 2 )
x 2 = x1 (k 2 + x21 + x22 ) + x2 (k 2 x21 x22 )

(106)

when: (a) k = 0 and (b) k 6= 0.

3. Consider the second-order linear control system x = Ax + Bu, A
22
R , B R21 where the coefficient matrix A has (a) real positive eigenvalues and (b) complex eigenvalues with positive real part. Investigate, using
the Lyapunov function, if there exists a linear control law u = Kx R able
to stabilize the control system.
4. Discuss the stability and bifurcations of the following systems, given
in polar coordinates:
(r = r(r 4); = 1)
(r = ar2 + br3 ; = 1)
(r = r( r); = 1)
(r = ar2 ; = 1)
(r = ar(r 1)(r ); = 1)

(107)

where a, b are constants and (t) is time-variant.

5. Consider the following system of diff. eq.
x 1 = ax21 + bx1 x2 + cx22
x 2 = x2 + lx12 + mx1 x2 + nx22

(108)

where a, b, c, l, m and n are constants. Show that:

1. if a 6= 0 then the origin is unstable;
2. if a = 0, bl < 0, then the origin is asymptotically stable;
3. if a = 0, bl > 0, then the origin is unstable;
4. if a = b = 0, cl2 6= 0, then the origin is unstable;
5. if a = b = c = 0, then the origin is stable but not asymptotically;
6. if a = l = 0, then the origin is stable but not asymptotically.
Show that the origin of the second-order system y + y 3 = 0 is a stable
equilibrium point.
6. Show that the equilibrium point at the origin of the system:
x 1 = x1 x2 + ax31 + bx1 x2
x 2 = x2 + cx21 + dx21 x2
91

(109)

1. is asymptotically stable if either a+c < 0, or a+c = 0 and cd+bc2 < 0;

2. is unstable if either a + c > 0, or a + c = 0 and cd + bc2 > 0.
7. Consider the second order system:
x 1 = x2
x 2 = x1 x32

(110)

What is the stability type of the equilibrium point at the origin ?

8. Discuss the stability properties of the origin for the system:
x 1 = x2 + x1 x2 + ax1 x22
x 2 = x1 x21 + x22

(111)

for various values of a.

9. Discuss the stability and instability of the equilibrium points of the
system:
x 1 = x2 + x23
(112)
x 2 = x1 + x13

Lab-works

1. Use dsolve from Matlab to find the solutions of the nonautonomous

scalar systems given in this document.
2. Draw the phase portraits for the problems given in the above section
(Home-works).

92

Lecture notes 2013/2014

Dynamical systems, 6
Mihaela Cistelecan, UT Cluj-Napoca
These notes were prepared entirely from
Khalil, H. Nonlinear Systems. Prentice Hall. 1996.

Fundamental properties

For the state equation x = f (t, x) to be a useful mathematical model for a

physical system, its solution should have the following properties:
existence
uniqueness
continuous dependence on initial conditions
continuous dependence on parameters
The meaning of these properties is as follows. In experimenting with a
physical system, we expect that starting the experiment from a given initial
state at time t0 , the system will move and its state will be defined in the
future time t > t0 . Moreover, with a deterministic system, we expect that
if we could repeat the experiment exactly, we would get exactly the same
motion and the same state at t > t0 . For the mathematical model to predict
the future state of the system from its current state at t0 , the initial value
problem
x = f (t, x) with x(t0 ) = x0
must have a unique solution. This is a question of existence and uniqueness
that can be ensured by imposing some constraints on the right-hand side
function f (t, x). The key constraint is the Lipschitz condition - see below.
93

In addition, an essential factor in the validity of any mathematical model

is the continuous dependence of its solutions on the data of the problem.
This refer to the fact that arbitrary small errors in the data will not result
in large errors in the solution obtained by the model. If f is differentiable
with respect to its parameters, then the solution will be differentiable with
respect to these parameters. In this context, we may study the effect of small
parameter variations on the system, using the sensitivity equations.

1.1

Existence and uniqueness

Note that solution x(t) of :

x = f (t, x) with x(t0 ) = x0
is:
x(t) = x0 +

Z t
t0

f (s, x(s))ds

(1)
(2)

Thus, the study of existence and uniqueness of the solution of the differential
eq. (1) is equivalent to the study of existence and uniqueness of the solution
of the integral equation (2).
However, a differential equation with a given initial condition might have
several solutions. For example, the scalar equation
x = x1/3 with x(0) = 0

(3)

has a solution x(t) = (2t/3)3/2 . This solution is not unique, since x(t) 0
is another solution. Noting that the right-hand side of (3) is continuous in
x, it is clear that continuity of f (t, x) in its arguments is not sufficient to
ensure uniqueness of the solution. Extra conditions must be imposed on the
function f .
Theorem - Local existence and uniqueness Let f (t, x) be piecewise
continuous in t and satisfy the Lipschitz condition
kf (t, x) f (t, y)k Lkx yk

(4)

x, y B = {x Rn |kx x0 k r}, t [t0 , t1 ]. Then, there exists some

> 0 such that the state equation:
x = f (t, x) with x(t0 ) = x0
94

(5)

has a unique solution over [t0 , t0 + ].

A function satisfying (4) is said to be Lispchitz in x, and the positive
constant L is called a Lipschitz constant. We also use the words locally Lipschitz and globally Lipschitz to indicate the domain over which the Lipschitz
condition holds.
We explain locally / globally Lipschitz only for a function f (x) depending
only on the x variable.
A function f (x) is said to be locally Lipschitz on a domain D Rn if
each point of D has a neighborhood D0 such that f satisfies the Lipschitz
condition for all points in D0 with some Lipschitz constant L. A locally
Lipschitz function on a domain D is not necessarily Lipschitz on D, since
the Lipschitz condition may not hold uniformly (with the same Lipschitz
constant L) for all points in D.
A function f (x) is said to be globally Lipschitz if it is Lipschitz on Rn .
When f : R R the Lipschitz condition can be written as:
|f (y) f (x)|
L
|y x|

(6)

which implies that on a plot of f (x) versus x, a straight line joining any
two points of f (x) cannot have a slope whose absolute value is greater than
L. Therefore, any function f (x) that has infinite slope at some point is not
locally Lipschitz at that point. For ex., any discontinuous function is not
locally Lipschitz at that point.

1.2

Locally Lipschitz property

Lemma: Let f : [a, b] D Rm be continuous for some domain D Rn .

Suppose [f /x] exists and is continuous on [a, b]D. If, for a convex subset
W D, there is a constant L 0 such that

(t, x) L
x

(7)

kf (t, x) f (t, y)k Lkx yk

(8)

on [a, b] W , then

for all t [a, b], x W and y W .

95

Remark: Due to equivalence of norms, the choice of a norm on Rn does

not affect the Lipschitz property of a function. It only affects the value of
the Lipschitz constant.
In the following we show by two examples how to use the above theorem.
Example: The function:
"

f (x) =

x1 + x1 x2
x2 x1 x2

(9)

is continuously differentiable on R2 . On any compact subset of R2 , f is

Lipschitz. Suppose we are interested in calculation a Lipschitz constant over
the convex set:
W = {x R2 | |x1 | a1 , |x2 | a2 }
The Jacobian matrix [f /x] is given by
f
[ ]=
x

"

1 + x2
x1
x2
1 x1

Using k k for vectors in R2 and the induced matrix norm for matrices,
we have:

f
x

= max{| 1 + x2 | + |x1 |, |x2 | + |1 x1 |}

All points satisfy

| 1 + x2 | + |x1 | 1 + a2 + a1
|x2 | + |1 x1 | a2 + 1 + a1
Hence

f
x

1 + a1 + a2

and a Lipschitz constant can be taken as L = 1 + a1 + a2 .

Example: Consider the function
"

f (x) =

x2
sat(x1 + x2 )

(10)

where the saturation function sat() is defined by:

sat(y) =

1 f or

y < 1
f or |y| 1
f or y > 1

96

(11)

The function f is not continuously differentiable on R2 . Let us check its

Lipschitz property by examining f (x) f (y). Using k k2 for vectors in R2
and the fact that the saturation function sat() satisfies
|sat() sat()| | |

(12)

we obtain
kf (x) f (y)k22 (x2 y2 )2 + (x1 + x2 y1 y2 )2
(x1 y1 )2 + 2(x1 y1 )(x2 y2 ) + 2(x2 y2 )2

(13)

Using the inequality

"

a2 + 2ab + 2b2 = [a b]

1 1
1 2

#"

a
b

"

max

1 1
1 2

#!

k[a b]k22

(14)

we conclude that:
kf (x) f (y)k2 1.618kx yk2

x, y R2

(15)

Here we have used xT P x max (P )xT x for all x Rn where P is

symmetric and positive definite and max () is the maximum eigenvalue of
the matrix.
The local Lipschitz property of a function is basically a smoothness requirement. It is implied by continuous differentiability. Except for discontinuous nonlinearities, which are idealizations of physical phenomena, it is
reasonable to expect models of physical systems to have locally Lipschitz
right-hand side functions.

1.3

Globally Lipschitz property

The theorem given in the previous section guarantees existence and uniqueness only over an interval [t0 , t0 + ] where may be very small. This means
that the phenomenon that the trajectory (solution) may have a finite escape
time is not taken into account.
For example, consider the scalar system:
x = x2 with x(0) = 1
97

The function f (x) = x2 is locally Lipschitz for all x R. Hence, it is

Lipschitz on any compact subset of R. The unique solution
x(t) =

1
t1

exists over [0, 1). As t 1, x(t) leaves any compact set.

A global Lipschitz condition would take care on the finite escape time,
as shown by the theorem below.
Theorem: (Global existence and uniqueness) Suppose f (t, x) is piecewise continuous in t and satisfies :
kf (t, x) f (t, y)k Lkx yk
kf (t, x0 )k h
x, y Rn , t [t0 , t1 ]. Then, the state equation
x = f (t, x) with x(t0 ) = x0
has a unique solution over [t0 , t1 ].
The global Lipschitz property is restrictive. One can easily construct
smooth meaningful examples which do not have the global Lipschitz property,
but do have unique global solutions, which is an indication of the conservative
nature of this theorem.
Example: Consider the scalar system
x = x3 = f (x)
The function f (x) does not satisfy a global Lipschitz condition since the
Jacobian f /x = 3x2 is not globally bounded. Nevertheless, for any
initial state x(t0 ) = x0 , the equation has the unique solution:
v
u
x(t) = sign(x0 )t

x20
1 + 2x20 (t t0 )

which is well defined for all t t0 .

However, the global Lipschitz condition may be imposed on a linear system, as follows. Consider the linear system:
x = A(t)x + g(t) = f (t, x)
98

where A() and g() are piecewise continuous functions of t. Over any
finite interval of time [t0 , t1 ], the elements of A(t) and g(t) are bounded.
Hence kA(t)k a and kg(t)k b, where kgk can be any norm on Rn and
kAk is the induced matrix norm. The conditions of the above given theorem
are fulfilled:
1 kf (t, x)f (t, y)k = kA(t)(xy)k kA(t)kkxyk akxyk, x, y Rn ,
t [t0 , t1 ]
2 kf (t, x0 )k = kA(t)x0 + g(t)k akx0 k + b h for each finite x0 , t [t0 , t1 ]

1.4

Continuous dependence on initial conditions and

parameters

For the solution of (1) to be of any interest, it must depend continuously

on the initial state x0 , the initial time t0 and the right-hand side function
f (x, t).
Let y(t) be a solution of (1) that starts at y(t0 ) = y0 and is defined on
the compact time interval [t0 , t1 ]. The solution depends continuously on y0
if solutions starting at nearby points are defined on the same interval and
remain close to each other in this interval
Another important problem to study is how f = f (t, x, ) depends on
some constant parameters . This problem is related to the case when the
constant parameters represent physical parameters of the system and the
study of perturbation of these parameters accounts for modeling errors or
changes in the parameter values due to aging. Let x(t, 0 ) be a solution of
x = f (t, x, 0 ) defined on [t0 , t1 ], with x(t0 , 0 ) = x0 . The solution is said
to depend continuously on if for any > 0, there is > 0 such that for
all in the ball { Rp |k 0 k < } , the equation x = f (t, x, ) has
a unique solution x(t, ) defined on [t0 , t1 ] with x(t0 , ) = x0 , and satisfies
kx(t, ) x(t, 0 )k < for all t [t0 , t1 ].
More about this subject in the next section.

1.5

Differentiability of solutions and sensitivity equations

Suppose f (t, x, ) is continuous in (t, x, ) and has continuous first partial

derivatives with respect to x and for all (t, x, ) [t0 , t1 ] Rn Rp . Let
99

0 be a nominal value of , and suppose the nominal state equation

x = f (t, x, 0 ) with x(t0 ) = x0

(16)

has a unique solution x(t, 0 ) over [t0 , t1 ]. A theorem proves that for all
sufficiently close to 0 , that is, k 0 k sufficiently small, the state eq:
x = f (t, x, ) with x(t0 ) = x0

(17)

has a unique solution x(t, ) over [t0 , t1 ]:

x(t, ) = x0 +

Z t
t0

f (s, x(s, ), )ds

(18)

that is close to the nominal solution x(t, 0 ).

In this context the sensitivity function S(t) is defined, to provide first
order estimates of the effect of parameter variations on solutions. That is:
x(t, ) x(t, 0 ) + S(t)( 0 )

(19)

An approach for calculating S(t) is to solve for the nominal solution and
the sensitivity function simultaneously, by numerically solving the system:
x =

x = f (t, x, ),
x(t0 ) = x0
f (t,x,)
+ ( ), x (t0 ) = 0

( f (t,x,)
)x
x

(20)

The sensitivity function S(t) can be calculated by solving (20) at the

nominal value = 0 .
Example: Consider the system:
x 1 = x2
= f1 (x1 , x2 )
x 2 = c sin x1 (a + b cos x1 )x2 = f2 (x1 , x2 )

(21)

and suppose the parameters a, b, c have the nominal values:

a0 = 1 b0 = 0 c0 = 1
The nominal system is given by:
x 1 = x2
x 2 = sin x1 x2
The Jacobian matrices [f /x] and [f /] are given by:
100

(22)

f
=
x
h
f
=

"

0
1
c cos x1 + bx2 sin x1 (a + b cos x1 )

f
a

f
b

f
c

"

(23)

0
0
0
x2 x2 cos x1 sin x1

(24)

Evaluate these Jacobian matrices at the nominal parameters to obtain:

"

f
|nominal =
x
f
|nominal =

"

0
1
cos x1 1

(25)

0
0
0
x2 x2 cos x1 sin x1

(26)

To solve for S(t) numerically, we solve (20) at the nominal values of the
parameters. Let:
"

S=

x3 x5 x7
x4 x6 x8

"

x1
a
x2
a

x1
b
x2
b

x1
c
x2
c

|nominal

(27)

Then, (20) is given:

x 1
x 2
x 3
x 4
x 5
x 6
x 7
x 8

1.6

= x2
= sin x1 x2
= x4
= x3 cos x1 x4 x2
= x6
= x5 cos x1 x6 x2 cos x1
= x8
= x7 cos x1 x8 sin x1

x1 (0) = x10
x2 (0) = x20
x3 (0) = 0
x4 (0) = 0
x5 (0) = 0
x6 (0) = 0
x7 (0) = 0
x8 (0) = 0

(28)

Homework

1. For each of the following functions find whether (a) f is locally Lipschitz
and (b) f is globally Lipschitz:
"

f (x) =

x1 + sgn(x2 )
x2

"

; f (x) =

x1 + sat(x2 )
x1 + sin x2
101

x3 sat(x1 + x2 )

; f (x) =
x22

x1

2. Derive the sensitivity equations for the system:

x 1 = tan1 (ax1 ) x1 x2
x 2 = bx21 cx2
as the parameters a, b, c vary from their nominal values a0 = 1, b0 = 0, c0 = 1.

102