Lecture Notes on Nonlinear Systems and Control

Spring Semester 2018

ETH Zurich

Peter Al Hokayem and Eduardo Gallestey

ABB Switzerland, Ltd.
Segelhof 1K
CH-5405, Baden-Dättwil
{peter.al-hokayem,eduardo.gallestey}@ch.abb.com
Contents

I Analysis 6
1 Introduction 7
1.1 Main Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.1 Nonlinear Models and Nonlinear Phenomena . . . . . . 8
1.2 Typical Nonlinearities . . . . . . . . . . . . . . . . . . . . . . 10
1.2.1 Memoryless Nonlinearities . . . . . . . . . . . . . . . . 10
1.2.2 Nonlinearities with Memory . . . . . . . . . . . . . . . 11
1.3 Examples of Nonlinear Systems . . . . . . . . . . . . . . . . . 12
1.3.1 Chemical Reactor . . . . . . . . . . . . . . . . . . . . . 12
1.3.2 Diode . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4 Second Order Systems . . . . . . . . . . . . . . . . . . . . . . 15
1.4.1 Qualitative Behavior of 2nd Order Systems Near Equi-
librium Points . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.2 Limit Cycles and Bifurcations . . . . . . . . . . . . . . 25
1.4.3 Bifurcations . . . . . . . . . . . . . . . . . . . . . . . 27
1.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2 Ordinary Differential Equations 33

2.1 Existence and Uniqueness . . . . . . . . . . . . . . . . . . . . 33
2.2 Continuity and Differentiability . . . . . . . . . . . . . . . . . 39
2.3 Differentiability and Sensitivity Equation . . . . . . . . . . . 41
2.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

3 Lyapunov Stability Theory 47

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2 Stability of Autonomous Systems . . . . . . . . . . . . . . . . 47
3.2.1 Lyapunov’s Direct Method . . . . . . . . . . . . . . . . 48
3.2.2 Lyapunov’s Indirect Method . . . . . . . . . . . . . . . 53

1
Nonlinear Systems and Control — Spring 2018

3.3 The Invariance Principle . . . . . . . . . . . . . . . . . . . . . 57

3.4 Instability Theorem . . . . . . . . . . . . . . . . . . . . . . . . 58
3.5 Comparison Functions . . . . . . . . . . . . . . . . . . . . . . 59
3.6 Stability of Nonautonomous Systems . . . . . . . . . . . . . . 61
3.7 Existence of Lyapunov Functions . . . . . . . . . . . . . . . . 66
3.8 Input-to-State Stability . . . . . . . . . . . . . . . . . . . . . . 67
3.9 Stability of Discrete-Time Systems . . . . . . . . . . . . . . . 70
3.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

4 Dissipative Systems 77
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.2 Dissipative Systems . . . . . . . . . . . . . . . . . . . . . . . . 79
4.3 Passive Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.3.1 Characterizations of Passivity for Linear Systems . . . 85
4.3.2 Stability of Passive Systems . . . . . . . . . . . . . . . 88
4.3.3 Passivity-Based Control . . . . . . . . . . . . . . . . . 90
4.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

II Control Design Techniques 95

5 Feedback Control Design 96
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.2 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.3 Integral Control . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.4 Gain Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

6 Feedback Linearization 110

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6.2 Input-Output Linearization . . . . . . . . . . . . . . . . . . . 113
6.2.1 Relationship to Linear Systems . . . . . . . . . . . . . 119
6.3 Full State Feedback Linearization . . . . . . . . . . . . . . . . 120
6.4 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.5 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

P. Al Hokayem & E. Gallestey 2 of 214

Nonlinear Systems and Control — Spring 2018

7 Sliding Mode Control 126

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
7.2 Sliding Mode Control . . . . . . . . . . . . . . . . . . . . . . . 128
7.3 Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

8 Optimal Control 136

8.1 Unconstrained Finite-Dimensional Optimization . . . . . . . . 136
8.2 Constrained Finite-Dimensional Optimization . . . . . . . . . 138
8.3 Pontryagin’s Minimum Principle . . . . . . . . . . . . . . . . . 142
8.4 Sufficient Conditions for Optimality . . . . . . . . . . . . . . . 147
8.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

9 State-Dependent Riccati Equation 153

9.1 State-Independent Riccati Equation Method . . . . . . . . . . 153
9.1.1 Results on Control Design for LTI Systems . . . . . . . 153
9.2 State-Dependent Riccati Equation Approach . . . . . . . . . . 155
9.2.1 H∞ control for nonlinear systems . . . . . . . . . . . . 164
9.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

10 Nonlinear Model Predictive Control 166

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
10.2 Main Theoretical Elements . . . . . . . . . . . . . . . . . . . 167
10.3 Case Study: Tank Level Reference Tracking . . . . . . . . . . 169
10.4 Chronology of Model Predictive Control . . . . . . . . . . . . 173
10.5 Stability of Model Predictive Controllers . . . . . . . . . . . . 174
10.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

11 State Estimation and Observers 182

11.1 Least Squares Estimation of Constant Vectors . . . . . . . . . 183
11.2 Weighted Least Squares Estimator . . . . . . . . . . . . . . . 187
11.3 Propagation of the State Estimate and its Uncertainty . . . . 187
11.4 Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
11.4.1 Extended Kalman Filter . . . . . . . . . . . . . . . . . 189
11.4.2 Moving Horizon Estimation . . . . . . . . . . . . . . . 191
11.5 High Gain Observers . . . . . . . . . . . . . . . . . . . . . . . 194
11.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

P. Al Hokayem & E. Gallestey 3 of 214

Nonlinear Systems and Control — Spring 2018

Sample Exam 1 200

Sample Exam 2 207

References 214

P. Al Hokayem & E. Gallestey 4 of 214

Nonlinear Systems and Control — Spring 2018

Preface
The objective of this course is to provide the students with an introduction
to nonlinear systems and the various methods of controlling them.
Part I of the course introduces the students to the notions of nonlinearities
and the various ways of analyzing existence and uniqueness of solutions to
ordinary differential equations, as well as understanding various notions of
stability and their characterizations.
Part II of the course arms the students with a variety of control methods
that are suitable for nonlinear systems and is designed in such a way as to
put the student in the position to deploy nonlinear control techniques in real
applications.
All chapters are combined with exercises that are geared towards attaining
better understanding of the pros and the cons of the different concepts.

P. Al Hokayem & E. Gallestey 5 of 214

 Part I

Analysis

6
Chapter 1

Introduction

1.1 Main Concepts

When engineers analyze and design nonlinear dynamical systems in elec-
trical circuits, mechanical systems, control systems, and other engineering
disciplines, they need to be able to use a wide range of nonlinear analysis
tools. Despite the fact that these tools have developed rapidly since the mid
1990s, nonlinear control is still largely a tough challenge.
In this course, we will present basic results for the analysis of nonlinear
systems, emphasizing the differences to linear systems, and we will introduce
the most important nonlinear feedback control tools with the goal of giving
an overview of the main possibilities available. Additionally, the lectures will
aim to give the context on which each of these tools are to be used. Table
1.1 contains an overview of the topics to be considered in this course.

Requirements Challenges Theoretical Results

Modeling and Simulation Well posedness ODE Theory
Bifurcations
Disturbance Rejection Sensors Observers
Stabilization Uncertainty Lyapunov
Tracking Nonlinearities Feedback Linearization
Sliding Mode
Economic Optimization Control Effort Optimal Control
Constraints Model Predictive Control

Table 1.1: Course Content

7
Nonlinear Systems and Control — Spring 2018

1.1.1 Nonlinear Models and Nonlinear Phenomena

We will deal with systems of the form:

ẋ1 = f1 (x1 , x2 , .., xn , u1 , u2 , . . . , um )

ẋ2 = f2 (x1 , x2 , .., xn , u1 , u2 , . . . , um )
...
ẋn = fn (x1 , x2 , .., xn , u1 , u2 , . . . , um )

where x ∈ Rn and u ∈ Rm .
Often, we will neglect the time-varying aspect. In the analysis phase,
external inputs are also often neglected, leaving system

ẋ = f (x). (1.1)

Working with an unforced state equation does not necessarily mean that the
input to the system is zero. It rather means that the input has been specified
as a given function of the state u = u(x).

Definition 1.1 A system is said to be autonomous or time invariant if the

function f does not depend explicitly on t, that is, ẋ = f (x).

Definition 1.2 A point x̄ is called equilibrium point of ẋ = f (x) if x(τ ) = x̄

for some τ implies x(t) = x̄ for t ≥ τ .

For an autonomous system the set of equilibrium points is equal to the

set of real solutions of the equation f (x) = 0.

• ẋ = x2 : isolated equilibrium point

• ẋ = sin(x): infinitely many equilibrium points

• ẋ = sin(1/x): infinitely many equilibrium points in a finite region

Linear Systems satisfy the following 2 properties:

1. Homogeneity: f (αx) = αf (x), ∀α ∈ R

2. Superposition: f (x + y) = f (x) + f (y), ∀x, y ∈ Rn

P. Al Hokayem & E. Gallestey 8 of 214

Nonlinear Systems and Control — Spring 2018

For example, consider the system given by the linear differential equation:
ẋ = Ax + Bu (1.2)
where x ∈ Rn , u ∈ Rm , A ∈ Rn×n , B ∈ Rn×m .
Then the solution is given by
Z t
At
x(t) = exp x0 + expA(t−τ ) Bu(τ )dτ. (1.3)
0

Note that the expression for x(t) is linear in the initial condition x0 and
in the control function u(·). Nonlinear systems are those systems that do not
satisfy these nice properties.
As we move from linear to nonlinear systems, we shall face a more difficult
situation. The superposition principle no longer holds, and analysis tools
necessarily involve more advanced mathematics. Most importantly, as the
superposition principle does not hold, we cannot assume that an analysis of
the behavior of the system either analytically or via simulation may be
scaled up or down to tell us about the behavior at large or small scales.
These must be checked separately.
The first step when analyzing a nonlinear system is usually to linearize it
about some nominal operating point and analyze the resulting linear model.
However, it is clear that linearization alone will not be sufficient. We must
develop tools for the analysis of nonlinear systems. There are two basic
limitation of linearization. First, since linearization is an approximation in
the neighborhood of an operating point, it can only predict the local behavior
of the nonlinear system in the vicinity of that point. Secondly, the dynamics
of a nonlinear system are much richer than the dynamics of a linear system.
There are essentially nonlinear phenomena that can take place only in the
presence of nonlinearity; hence they cannot be described or predicted by
linear models. The following are examples of nonlinear phenomena:
• Finite escape time: The state of an unstable linear system can go
to infinity as time approaches infinity. A nonlinear system’s state,
however, can go to infinity in finite time.
• Multiple isolated equilibrium points: A linear system can have
only one equilibrium point, and thus only one steady-state operating
point that attracts or repels the state of the system irrespective of the
initial state. A nonlinear system can have more than one equilibrium
point.

P. Al Hokayem & E. Gallestey 9 of 214

Nonlinear Systems and Control — Spring 2018

• Limit cycles: A linear system can have a stable oscillation if it has a

pair of eigenvalues on the imaginary axis. The amplitude of the oscilla-
tion will then depend on the initial conditions. A nonlinear system can
exhibit an oscillation of fixed amplitude and frequency which appears
independently of the initial conditions.

• Chaos: A nonlinear system can have a more complicated steady-state

behavior that is not equilibrium or periodic oscillation. Some of these
chaotic motions exhibit randomness, despite the deterministic nature
of the system.

• Multiple Modes of behaviour: A nonlinear system may exhibit

very different forms of behaviour depending on external parameter val-
ues, or may jump from one form of behaviour to another autonomously.
These behaviours cannot be observed in linear systems, where the com-
plete system behaviour is characterized by the eigenvalues of the system
matrix A.

1.2 Typical Nonlinearities

In the following subsections, various nonlinearities which commonly occur in
practice are presented.

1.2.1 Memoryless Nonlinearities

Most commonly found nonlinearities are:

• Relay, see Figure1.1. Relays appear when modelling mode changes.

• Saturation, see Figure 1.2. Saturations appear when modelling vari-

ables with hard limits, for instance actuators.

• Dead Zone, see Figure 1.3. Dead Zone appear in connection to actu-
ator or sensor sensitivity.

• Quantization, see Figure 1.4. Quantization is used to model discrete

valued variables, often actuators.

P. Al Hokayem & E. Gallestey 10 of 214

Nonlinear Systems and Control — Spring 2018

1
u
−1

Figure 1.1: Relay

y

k
u
δ

Figure 1.2: Saturation

This family of nonlinearities are called memoryless, zero memory or static

because the output of the nonlinearity at any instant of time is determined
uniquely by its input at that time instant; it does not depend on the history
of the input.

1.2.2 Nonlinearities with Memory

Quite frequently though, we encounter nonlinear elements whose input-output
characteristics have memory; that is, the output at any instant of time may
depend on the recent or even entire, history of the input.
In the case of hysteresis one is confronted with a situation where the
path forward is not the same as the path backward. This behavior is often
observed when dealing with defective actuators such as valves.

P. Al Hokayem & E. Gallestey 11 of 214

Nonlinear Systems and Control — Spring 2018

u
δ

Figure 1.3: Dead Zone

y

δ
u
δ/2

Figure 1.4: Quantization

1.3 Examples of Nonlinear Systems

In this section we present some examples of nonlinear systems which demon-
strate how nonlinearities may be present, and how they are then represented
in the model equations.

1.3.1 Chemical Reactor

This is an example of a strongly nonlinear system
q
[C˙a ] = ([Caf ] − [Ca ]) − r[Ca ]
V
The coefficient r is an exponential function of the temperature and the
reagent concentration.
E
r = K exp
RT

P. Al Hokayem & E. Gallestey 12 of 214

Nonlinear Systems and Control — Spring 2018

y
l+

s− s+
u

l−
Figure 1.5: Relay with hysteresis

Figure 1.6: Chemical Reactor

while the temperature T is given by

q
Ṫ = (Tf − T ) + Kr r[Ca ] + Kc (T − Tc )
V
The model has 2 states: the concentration [Ca ] of A and the temperature
T of the reaction vessel liquid. The manipulated variable is the jacket water
temperature Tc . Depending upon the problem formulation,the feed temper-
ature Tf and feed concentration [Caf ] can be considered either constant or
as a disturbance. At a jacket temperature of 305K, the reactor model has an
oscillatory response. The oscillations are characterized by reaction run-away
with a temperature spike. When the concentration drops to a lower value,
the reactor cools until the concentration builds and there is another run-away
reaction.

P. Al Hokayem & E. Gallestey 13 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.7: Chemical Reactor Phase Portrait

1.3.2 Diode
We assume a time invariant linear capacitor C, inductor L and resistor R.

R iC iR

vC C vR
E

Figure 1.8: Diode

The tunnel diode characteristic curve iR = h(vR ) is plotted in the Figure

1.9.
Choosing x1 = vC and x2 = iL and u = E we obtain the following
dynamical system
1
ẋ1 = (−h(x1 ) + x2 )
C
1
ẋ2 = (−x1 − Rx2 + E)
L

P. Al Hokayem & E. Gallestey 14 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.9: Diode Characteristic Curve

The equilibrium points of the system are determined by setting ẋ1 = ẋ2 =
0.
1
0 = (−h(x1 ) + x2 )
C
1
0 = (−x1 − Rx2 + E)
L
Therefore, the equilibrium points corresponds to the roots of the equation
E − x1
h(x1 ) =
R
The next figure shows graphically that, for certain values of E and R, this
equation has three isolated roots which correspond to three isolated equilib-
rium points of the system. The number of equilibrium points might change
as the values of E and R change. For example, if we increase E for the same
value of R, we will reach a point beyond which only Q3 will exist.
As we will see in the next chapter, the phase portrait in this case has two
stable equilibrium point and 1 unstable equilibrium point
The tunnel diode characteristic curve iR = h(vR ) is plotted in the next
figure.

1.4 Second Order Systems

Second-order autonomous systems occupy an important place in the study
of nonlinear systems because solution trajectories can be represented in the

P. Al Hokayem & E. Gallestey 15 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.10: Diode Equilibria

2D-plane. This allows for easy visualization of the qualitative behavior of

the system.
In the sequel we consider the following aspects of second order systems:

1. Behavior near equilibrium points

2. Nonlinear oscillations

3. Bifurcations

A second-order autonomous system is represented by two scalar differen-

tial equations

ẋ1 = f1 (x1 , x2 ), x1 (0) = x10

ẋ2 = f2 (x1 , x2 ), x2 (0) = x20

or in vector notation

ẋ = f (x), x(0) = x0 , x, x0 ∈ R2

The locus in the (x1 , x2 ) plane of the solution x(t) for all t ≥ 0 is a curve
that passes through the point x0 . This plane is usually called state plane or
phase plane. The vector f gives the tangent vector to the curve x(·).

Definition 1.3 We obtain a vector field diagram by assigning the vector

(f1 (x), f2 (x) to every point (x1 , x2 ) in a grid covering the plane.

P. Al Hokayem & E. Gallestey 16 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.11: Diode Phase Portrait

For example, iff (x) = (2x21 , x2 ), then at x = (1, 1), we draw an arrow point-
ing from (1, 1) to (1, 1) + (2, 1) = (3, 2).

Example 1.4 Pendulum without friction.

ẋ1 = x2
ẋ2 = −10 sin x1

See Figure 1.12.

Definition 2.2: The family of all trajectories of a dynamical system is called

the phase portrait.
As we will see later, the behavior of nonlinear systems near equilibrium points
is well described by the behavior of the linearization. We approximate the
nonlinear system by its linearization:
 
∂f1  ∂f1 
ẋ1 = ∂x  x 1 + ∂x2 
x2
1
x=x̄ x=x̄
∂f2  ∂f2 
ẋ2 = ∂x 1
 x1 + ∂x 2
 x2
x=x̄ x=x̄

P. Al Hokayem & E. Gallestey 17 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.12: Frictionless Pendulum.

1.4.1 Qualitative Behavior of 2nd Order Systems Near

Equilibrium Points
Consider the linear time-invariant system ẋ = Ax where A is a 2 × 2 real
matrix. The solution of the equation for a given state x0 is given by
x (t) = M exp (Jr t) M −1 x0 ,
where Jr is the real Jordan form of A and M is a real nonsingular matrix
such that M −1 AM = Jr . Depending on the properties of A, the real Jordan
form may take one of three forms:
     
λ1 0 λ k α −β
, ,
0 λ2 0 λ β α
where k is either 0 or 1. The first form corresponds to the case when the
eigenvalues λ1 and λ2 are real and distinct, the second form corresponds
to the case when the eigenvalues are real and equal, and the third form
corresponds to the case of complex eigenvalues λ1,2 = α ± jβ.

Real Distinct Eigenvalues: In this case λ1 and λ2 are different from zero
and M = [v1 , v2 ], where v1 and v2 are the real eigenvectors associated with
λ1 and λ2 . The change of coordinates z = M −1 x transforms the system into
two decoupled first-order differential equations,

P. Al Hokayem & E. Gallestey 18 of 214

Nonlinear Systems and Control — Spring 2018

ż1 = λ1 z1
ż2 = λ2 z2
whose solution, for a given initial state (z10 , z20 ), is given by
z1 = z10 eλ1 t
z2 = z20 eλ2 t .
Eliminating t between the two equations, we obtain

Figure 1.13: Stable Node.

λ /λ1
z2 = cz1 2
where
c = z20 / (z10 )λ2 /λ1 . (1.4)
There are then three cases:
• Stable node: Both eigenvalues are negative, Figure 1.13.
• Unstable node: Both eigenvalues are positive, Figure 1.14.
• Saddle point: Eigenvalues have different sign, Figure 1.15

P. Al Hokayem & E. Gallestey 19 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.14: Unstable Node.

Complex Eigenvalues: The change of coordinates z = M −1 x transforms

the system into the form

ż1 = αz1 − βz2

ż2 = βz1 + αz2

The solution of these equations is oscillatory and can be expressed more

conveniently in polar coordinates.
q
r = z12 + z22
 
−1 z2
θ = tan
z1
where we have two uncoupled first-order differential equation:

ṙ = αr
θ̇ = β

The solution for a given initial state (r0 , θ0 ) is given by

r (t) = r0 eαt
θ (t) = θ0 + βt

We now have three behaviors:

P. Al Hokayem & E. Gallestey 20 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.15: Saddle Point.

• Stable Focus, when α < 0, the spiral converges to the origin, see
Figure 1.16.
• Unstable Focus, when α > 0, it diverges away from the origin, see
Figure 1.17.
• Circle, when α = 0, the trajectory is a circle of radius r0 , see Figure
1.18.

Nonzero Multiple Eigenvalues

The change of coordinates z = M −1 x transforms the system into the form

ż1 = λz1 + kz2

ż2 = λz2
whose solution, for a given initial state (z10 , z20 ), is given by
z1 (t) = (z10 + kz20 t) eλt
z2 (t) = eλt z20 .
Eliminating t, we obtain the trajectory equation
  
z10 k z2
z1 = z2 + ln
z20 λ z20
Here there are only two cases:

P. Al Hokayem & E. Gallestey 21 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.16: Stable Focus.

Figure 1.17: Unstable Focus.

Figure 1.18: Circle.

P. Al Hokayem & E. Gallestey 22 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.19: As many Eigenvectors as Eigenvalues.

Figure 1.20: Fewer Eigenvectors than Eigenvalues.

• As many eigenvectors as eigenvalues: k = 0 (λ < 0, λ > 0), see Figure

1.19.

• Fewer eigenvectors than eigenvalues: k = 1 (λ < 0, λ > 0), see Figure

1.20.

One or more Eigenvalues are zero When one or both eigenvalues of A

are zero, the phase portrait is in some sense degenerate. Here, the matrix A
has a nontrivial null space. Any vector in the null space of A is an equilibrium
point for the system; that is, the system has an equilibrium subspace, rather
than an equilibrium point.
The dimension of the null space could be one or two; if it is two, the
matrix A will be the zero matrix. When the dimension of the null space is
one, the shape of the Jordan form of A will depend on the multiplicity of

P. Al Hokayem & E. Gallestey 23 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.21: One Zero Eigenvalue.

Figure 1.22: Zero Eigenvalues.

P. Al Hokayem & E. Gallestey 24 of 214

Nonlinear Systems and Control — Spring 2018

the zero eigenvalue. When λ1 = 0 and λ2 6= 0, the matrix M is given by

M = [v1 , v2 ] where v1 and v2 are the associated eigenvectors.
Here there are two cases:

• λ1 = 0 and (λ2 < 0, λ2 > 0), see Figure 1.21.

• λ1 = 0λ2 = 0, see Figure 1.22.

1.4.2 Limit Cycles and Bifurcations

Oscillations and Limit Cycles Oscillations are one of the most impor-
tant phenomena that occur in dynamical systems. A system

ẋ = f (x)

oscillates when it has a nontrivial periodic solution, i.e.

x (t + T ) = x (t) , ∀t ≥ 0

for some T > 0. The word nontrivial is used to exclude equilibrium point.
In a phase portrait an oscillation or periodic solution is given by a closed
trajectory, usually called a closed or periodic orbit. The trajectory resembles
a closed curve in the phase plane. The trivial example is a center.

Example 1.5 Van der Pol Oscillator

ẋ1 = x2


ẋ2 = −x1 + ε 1 − x21 x2

In the case ε = 0 we have a continuum of periodic solutions, while in the

case ε 6= 0 there is only one.

Definition 1.6 An isolated periodic orbit is called a limit cycle.

Figure 1.23 depicts a stable and an unstable limit cycle. Limit cycles are
themselves special cases of limit sets. However the study of general limit sets
is outside the scope of this course.

P. Al Hokayem & E. Gallestey 25 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.23: Limit Cycles.

Existence of Periodic Orbits Periodic orbits in the plane are special in

that they divide the plane into a region inside the orbit and a region outside it.
This makes it possible to obtain criteria for detecting the presence or absence
of periodic orbits for second-order systems, which have no generalizations to
higher order systems.

Theorem 1.7 Poincaré-Bendixson Criterion Consider the system on

the plane ẋ = f (x) and let M be a closed bounded subset of the plane such
that:
1. M contains no equilibrium points, or contains only one equilibrium
point such that the Jacobian matrix [df /dx] at this point has eigenvalues
with positive real parts.
2. Every trajectory starting in M remains in M for all future time.
Then M contains a periodic orbit of ẋ = f (x).

Sketch of Proof: A trajectory in M is bounded, and so it must either

converge to an equilibrium point or approach a periodic solution as time
tends to infinity. As there are no stable equilibrium points, it must converge
to a periodic orbit.

Theorem 1.8 Negative Pointcaré-Bendixson Criterion If, on a sim-

ply connected region D of the plane, the expression ∂f1 /∂x1 + ∂f2 /∂x2 is not
identically zero and does not change sign, then the system ẋ = f (x) has no
periodic orbits lying entirely in D.

P. Al Hokayem & E. Gallestey 26 of 214

Nonlinear Systems and Control — Spring 2018

Proof: On any orbit ẋ = f (x), we have

dx2 f2
= .
dx1 f1
Therefore, on any closed orbit γ, we have
I
f2 (x1 , x2 ) dx1 − f1 (x1 , x2 ) dx2 = 0
Γ

This implies, by Green’s theorem, that

ZZ  
∂f1 ∂f2
+ dx1 dx2 = 0
S ∂x1 ∂x2

where S is the interior of γ. Now, if

∂f1 ∂f2
+ 6= 0 on D,
∂x1 ∂x2
then we cannot find a region S ⊂ D such that the last equality holds. Hence,
there can be no closed orbits entirely in D.

1.4.3 Bifurcations
The qualitative behavior of a second-order system is determined by the pat-
tern of its equilibrium points and periodic orbits, as well as by their stability
properties. One issue of practical importance is whether the system main-
tains its qualitative behavior under infinitesimally small perturbations.
For a nonlinear autonomous dynamical system ẋ = f(x), the set of equi-
libria is given by {x : f (x) = 0}. The eigenvalues λ of ∂f  the Jacobian of
∂x x=x∗
∗
f at an equilibrium x determine the local stability properties for the system
for Re(λ) 6= 0. For Re(λ) = 0 further analysis is required.

Example 1.9 The systems

ẋ = −x3 , ẋ = x3

have the same linearization but quite different stability properties.

P. Al Hokayem & E. Gallestey 27 of 214

Nonlinear Systems and Control — Spring 2018

In general, for a dynamical system dependent on a parameter µ,

ẋ = f (x, µ)

the system is called structurally stable if small perturbations of the parameter

µ do not cause the equilibria to change their stability properties, and if there
are no additional equilibria created. Since the eigenvalues usually depend
continuously on µ, this will usually occur only when Re(λ(µ)) = 0.

Definition 1.10 The creation of a new set of equilibria is called a bifurca-

tion.

We are interested in perturbations that will change the equilibrium points

or periodic orbits of the system or change their stability properties. The
points in parameter space at which the system behavior change are called
bifurcation points. A bifurcation can include:

1. The creation of a stable-unstable pair

2. Vanishing of a stable-unstable pair

3. Pitchfork bifurcations

(a) A stable equilibrium becomes 2 stable and one unstable equilib-

rium
(b) An unstable equilibrium becomes 2 unstable and one stable equi-
librium

Example 1.11 We consider the system

ẋ1 = µ − x21
ẋ2 = −x2 .

√

1. If µ > 0 we have two equilibrium points Q1,2 = ± µ, 0 , of which one
is a saddle and the other one is a stable node.

2. If µ = 0 we have the degenerated case (0, 0). We have an equilibrium

point at the origin, where the Jacobian matrix has one zero eigenvalue.
This is the bifurcation point where saddle and stable node collide.

P. Al Hokayem & E. Gallestey 28 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.24: Bifurcations: phase portrait of the saddle-node bifurcation ex-

ample for µ > 0 (left), µ = 0 (center), µ < 0 (right).

3. If µ < 0 there are no equilibrium points anymore and both saddle and
stable node disappeared.

See Figure 1.24.

Theorem 1.12 Hopf Bifurcations Consider the system

ẋ = fµ (x1 , x2 ),
where µ is a parameter. Let (0, 0) be an equilibrium point for any value of
the parameter µ. Let µc be such that
 
df
eig (0, 0) = ±jβ.
dx
df d
If the real parts of the eigenvalues λ1,2 of dx
(0, 0) are such that dµ
Re (λ1,2 ) >0
and the origin is stable for µ = µc then
1. there is a µl such that the origin remains stable for µ ∈ [µl , µc ]
2. there is a µL such that the origin is unstable surrounded by a stable
limit cycle for µc < µ < µL .
Thus a Hopf Bifurcation is a bifurcation where a stable point bifurcates into
a stable limit cycle which surrounds an unstable point.

Example 1.13 We consider the system

ẋ1 = x2
ẋ2 = −x1 + (µ − x21 )x2

P. Al Hokayem & E. Gallestey 29 of 214

Nonlinear Systems and Control — Spring 2018

We realize that p
λ1,2 = 0.5µ ± 0.25µ2 − 1.
d
Thus, it is easily checked that, as for µ ∈ [−0.2; 0.2], the derivative dµ Re (λ1,2 .) =
0.5 > 0. The real parts become zero at µ = 0. In the next three figures we
depict the behavior of the system as µ goes from a negative to a positive value.

Figure 1.25: Hopf Bifurcation Example for µ = −1.

Example 1.14 Tunnel-Diode Circuit Revisiting Example 1.3.2 we see that

E and R are bifurcations parameters for the system.

Remark 1.15 A cascade of bifurcations is usually a precursor to chaotic

behavior. Chaotic behavior is observed when the effect of arbitrarily small
differences in initial conditions are sufficient to make long term prediction
of the trajectory of a system unpredictable. A further consideration of these
phenomena is outside the scope of this nonlinear control course, but the reader
is encouraged to read more on the subject of chaos of dynamical systems.

P. Al Hokayem & E. Gallestey 30 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.26: Hopf Bifurcation Example for µ = 0.

1.5 Exercises
1. Consider the nonlinear system
ẋ1 = µ − x1 (x21 − 1)
ẋ2 = −x2
What is the behavior of the nonlinear system as the parameter µ varies?
2. Consider the nonlinear system
ẋ1 = g(x2 ) + 4x1 x22
ẋ2 = h(x1 ) + 4x21 x2
Are there any limit cycles?
3. A pitchfork bifurcation example is given by
ẋ1 = µx1 − x31
ẋ2 = −x2
What is the behavior of the nonlinear system as the parameter µ varies?

P. Al Hokayem & E. Gallestey 31 of 214

Nonlinear Systems and Control — Spring 2018

Figure 1.27: Hopf Bifurcation Example for µ = 1.

4. Consider again the Tunnel Diode Circuit dynamics that are given by
1
ẋ1 = [−h(x1 ) + x2 ]
C
1
ẋ2 = [−x1 − Rx2 + u]
L
where we take C = 2, L = 5 and R = 1.5. Suppose that h(·) is given
by
h(x1 ) = 17.76x1 103.79x21 + 229.62x31 226.31x41 + 83.72x51

• Find the equilibrium points of the system for u = 1.2V ? What is

the nature of each equilibrium point?

P. Al Hokayem & E. Gallestey 32 of 214

Chapter 2

Ordinary Differential Equations

2.1 Existence and Uniqueness

Theorem 2.1 Local Existence and Uniqueness: Let f (t, x) be a piece-
wise continuous function in t and satisfy the Lipschitz condition

kf (t, x) − f (t, y)k ≤ L kx − yk

∀x, y ∈ B (x0 , r) = {x ∈ Rn | kx − x0 k ≤ r}, ∀t ∈ [t0 , t1 ]. Then there exists

some δ > 0 such that the state equation

ẋ = f (t, x)

with x (t0 ) = x0 has a unique solution over [t0 , t0 + δ].

Proof: We start by noting that if x (t) is a solution of

ẋ = f (t, x) , x (t0 ) = 0

then, by integration, we have

Z t
x (t) = x0 + f (s, x (s)) ds (2.1)
t0

It follows that the study of existence and uniqueness of the solution of the
differential equation is equivalent to the study of existence and uniqueness
of the solution of the integral equation above.

33
Nonlinear Systems and Control — Spring 2018

Let us introduce some convenient notation. Let

Z t
(P (x)(t) := x(t0 ) + f (s, x (s)) ds, t ∈ [t0 , t1 ],
t0

be a mapping of a continuous function

x : [t0 , t1 ] → Rn .

in to R.
Then, we realize that we can rewrite the Equation (2.1) as

x (t) = (P x) (t)

Note that (P x) (t) is continuous in t. A solution of x (t) = (P x) (t) can be

established by using the contraction mapping theorem, see Appendix 2.3.
Applying this result to our problem requires defining appropriate
1. Banach space X and

2. closed set S ⊂ X
such that P maps S into S and is a contraction over S. Let

X = C [t0 , t0 + δ]

with norm
kxkC = max kx (t)k
t∈[t0 ,t0 +δ]

and
S = {x ∈ X | kx − x0 kC ≤ r}
where r is the radius of the ball B and δ is a positive constant to be chosen.
We will restrict the choice of δ to satisfy δ ≤ t1 −t0 so that [t0 , t0 + δ] ⊂ [t0 , t1 ].
Note that kx (t)k denotes a norm on Rn , while kxkC denotes a norm on X .
Also, B is a ball in Rn , while S is a ball in X .
By definition, P maps X into X . To show that it maps S into S write
Z t Z t
(P x) (t) − x0 = f (s, x (s)) ds = [f (s, x (s)) − f (s, x0 ) + f (s, x0 )] ds
t0 t0

By piecewise continuity of f , we know that f (s, x0 ) is bounded on [t0 , t1 ].

Let

P. Al Hokayem & E. Gallestey 34 of 214

Nonlinear Systems and Control — Spring 2018

h = max kf (t, x0 )k
t∈[t0 ,t1 ]

Using the Lipschitz condition and the fact that for each x ∈ S

kx (t) − x0 k ≤ r, ∀t ∈ [t0 , t0 + δ]

we obtain
Z t
k(P x) (t) − x0 k ≤ [kf (s, x (s)) − f (s, x0 )k + kf (s, x0 )k] ds
t
Z 0t
≤ [L kx (s) − x0 k + h] ds
t
Z 0t
≤ (Lr + h) ds
t0
≤ (t − t0 ) (Lr + h)
≤ δ (Lr + h)

and
kP x − x0 kC = max k(P x) (t) − x0 k ≤ δ (Lr + h)
t∈[t0 ,t0 +δ]

Hence, choosing δ ≤ r/ (Lr + h) ensures that P maps S into S.

To show that P is a contraction mapping over S, let x and y ∈ S and
consider
Z t


k(P x)(t) − (P y)(t)k ≤ [f (s, x(s)) − f (s, y(s))] ds
t0
Z t
≤ k[f (s, x(s)) − f (s, y(s))]k ds
t0
Z t
≤ L kx(s) − y(s)k ds
t0
Z t
≤ L kx − ykC ds
t0

Therefore,
kP x − P ykC < Lδ kx − ykC < ρ kx − ykC ,
for δ < ρ/L with ρ < 1. Thus, choosing ρ < 1 and δ ≤ ρ/L ensures that P is
a contraction mapping over S.

P. Al Hokayem & E. Gallestey 35 of 214

Nonlinear Systems and Control — Spring 2018

By the contraction mapping theorem, we can conclude that if δ is chosen to

satisfy
 
r ρ
δ ≤ min t1 − t0 , , , for ρ < 1,
Lr + h L
then the integral equation
Z t
x (t) = x0 + f (s, x (s)) ds
t0

will have a unique solution in S.

This is not the end of the proof though because we are interested in
establishing uniqueness of the solution among all continuous functions x (t),
that is, uniqueness in X . It turns out that any solution of
Z t
x (t) = x0 + f (s, x (s)) ds
t0

in X will lie in S. To see this, note that since x (t0 ) = x0 is inside the ball
B, any continuous solution x (t) must lie inside B for some interval of time.
Suppose that x (t) leaves the ball B and let t0 + µ be the first time x (t)
intersects the boundary of B. Then,
kx (t0 + µ) − x0 k = r
On the other hand, for all t ≤ t0 + µ,
Z t
kx (t) − x0 k ≤ [kf (s, x (s)) − f (s, x0 )k + kf (s, x0 )k] ds
t0
Z t
≤ [L kx (s) − x0 k + h] ds
t0
Z t
≤ (Lr + h) ds
t0

Therefore,
r
r = kx (t0 + µ) − x0 k ≤ (Lr + h) µ ⇒ µ ≥ ≥δ
Lr + h
Hence, the solution x (t) cannot leave the set B within the time interval
[t0 , t0 + δ], which implies that any solution in X lies in S. Consequently,
uniqueness of the solution in S implies uniqueness in X .

P. Al Hokayem & E. Gallestey 36 of 214

Nonlinear Systems and Control — Spring 2018

Remark 2.2 Some remarks about the Lipschitz condition:

1. Lipschitz is stronger than continuity

2. There exist continuous functions which are not Lipschitz. For example
1
f (x) = x 3

is not Lipschitz. In particular, the gradient becomes infinite at x=0.

Accordingly, the differential equation
1
ẋ = x 3 , x(0) = 0

does not have a unique solution. There exists an entire family of solu-
tions:   32
3 3
x(t) = 0 and x(t) = − (t − c) 2
2
for any c > 0.

3. The norm for which the Lipschitz condition holds is not relevant.
∂f
4. Locally bounded ∂x
⇒ locally Lipschitz

5. Lipschitz 6= differentiability. E.g.

 
x2
f (x) =
−sat (x1 + x2 )

is Lipschitz but not differentiable.

Theorem 2.3 Global Existence and Uniqueness: Let f (t, x) be piece-

wise continuous in t over the interval [t0 , t1 ] and globally Lipschitz in x. Then

ẋ = f (t, x) , x (t0 ) = x0

has a unique solution over [t0 , t1 ).

P. Al Hokayem & E. Gallestey 37 of 214

Nonlinear Systems and Control — Spring 2018

Example 2.4 Consider the function

0
f (x) = x2 ⇒ f (x) = 2x

The derivative is not bounded therefore f (x) is not globally Lipschitz. Accord-
ingly we cannot expect the solution to the differential equation ẋ = f (t, x)
to have a unique solution over any time interval. Indeed, the differential
equation:
ẋ = x2 , x(0) = c > 0
has the solution  −1
1
x(t) = −t
c
which has a vertical asymptote at t = 1c , which means that the solution ex-
hibits finite escape time.

Nevertheless, it may be that a function f (t, x) is not globally Lipschitz,

and yet the differential does have a solution which is defined over the entire
time interval. Consider the following example:

Example 2.5 Consider the function f (x) = −x3 . ∂f ∂x

is not bounded there-
fore the function f isn’t globally Lipschitz. However
s
x20
x (t) = sign (x0 )
1 + 2x20 (t − t0 )

is the solution of f (x) = −x3 for any x0 and t > t0 .

To develop this further, recall the mapping P from the proof of Theorem 2.1.
Then:

Theorem 2.6 Let f be piecewise continuous in t, locally Lipschitz in x for

x ∈ D. Let w be a compact subset of D. If x(t) = (P x)(t) ∈ w, ∀t > t0 , ie.
w is invariant, then the solution x(t) exists ∀t > t0 .

Example 2.7 Van der Pol Oscillator.

ẋ1 = x2


ẋ2 = −x1 + ε 1 − x21 x2

P. Al Hokayem & E. Gallestey 38 of 214

Nonlinear Systems and Control — Spring 2018

Then  
∂f 0 1
=
∂x −1 − 2εx1 x2 ε (1 − x21 )
The Jacobian ∂f∂x
is bounded in a bounded domain, thus f is locally Lipschitz
and the solution of the differential equation exists locally. However, global
existence has not been established due to unboundedness of ∂f∂x
over the entire
state space.

2.2 Continuity and Differentiability

For the solution of the state space equation

ẋ = f (t, x)

with x (t0 ) = x0 to be of interest in real applications, it must depend con-

tinuously on the initial state x0 , the initial time t0 , and the right-hand side
function f (t, x). Continuous dependence on the initial time t0 is obvious
from the integral expression
Z t
x (t) = x0 + f (s, x (s)) ds
t0

Let y (t) be a solution of the equation 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 time interval
and remain close to each other in that interval.
Let x (t, λ0 ) with nominal parameters λ0 be a solution of ẋ = f (t, x, λ0 )
defined on [t0 , t1 ], with x (t0 , λ0 ) = x0 . The solution is said to be depen-
dent continuously on λ if for any ε > 0, there is δ > 0 such that for all
λ in the ball {λ ∈ Rp | kλ − λ0 k < δ}, the equation ẋ = f (t, x, λ) has a
unique solution x (t, λ) defined on [t0 , t1 ], with x (t0 , λ) = x0 , and satisfies
kx (t, λ) − x (t, λ0 )k < ε for all t ∈ [t0 , t1 ].

Theorem 2.8 Let f (t, x, λ) be continuous in (t, x, λ) and locally Lipschitz

in x (uniformly in t and λ) on [t0 , t1 ] × D × {kλ − λ0 k ≤ c}, where D ⊂ Rn
is an open connected set.

P. Al Hokayem & E. Gallestey 39 of 214

Nonlinear Systems and Control — Spring 2018

Let y (t, y0 ) be a solution of ẋ = f (t, x, λ0 ) with y (t0 , y0 ) = y0 ∈ D.

Suppose y (t, λ0 ) is defined and belongs to D for all t ∈ [t0 , t1 ]. Then, given
ε > 0, there is δ > 0 such that if

kz0 − y0 k < δ, kλ − λ0 k < δ

Then there is a unique solution z (t, λ) of ẋ = f (t, x, λ) defined on [t0 , t1 ],

with z (t0 , λ) = z0 , and z (t, λ) satisfies

kz (t, λ) − y (t, λ0 )k < ε, ∀t ∈ [t0 , t1 ] .

Proof: By continuity of y (t, λ0 ) in t and the compactness of [t0 , t1 ], we know

that y (t, λ0 ) is bounded on [t0 , t1 ]. Define a tube U around the solution
y (t, λ0 ) by

Figure 2.1: Tube constructed around solution y(t, λ0 )

U = {(t, x) ∈ [t0 , t1 ] × Rn | kx − y (t, λ0 )k ≤ ε}

We can estimate
µ
ky (t) − z (t)k ≤ ky0 − z0 k exp [L (t − t0 )] + {exp [L (t − t0 ) − 1]}
L
where L is the Lipschitz constant and µ is a constant related to the size of
the perturbation of the parameters.


Remark 2.9 Note that the proof does not estimate the radius of the “tube”
in Figure 2.1.

P. Al Hokayem & E. Gallestey 40 of 214

Nonlinear Systems and Control — Spring 2018

2.3 Differentiability and Sensitivity Equa-

tion
The continuous differentiability of f with respect to x and λ implies the
additional property that the solution x (t, λ) is differentiable with respect to
λ near λ0 . To see this, write
Z t
x (t, λ) = x0 + f (s, x (s, λ) , λ) ds
t0

Taking partial derivative with respect to λ yields

Z t 
∂x ∂f ∂x ∂f
= + ds
∂λ t0 ∂x ∂λ ∂λ

Differentiating with respect to t, it can be seen that

∂x
(t0 , λ) = 0
∂λ
 
∂ ∂x ∂x
= A (t, λ) + B (t, λ)
∂t ∂λ ∂λ
where
∂f
A (t, λ) = (t, x (t, λ) , λ)
∂x
∂f
B (t, λ) = = (t, x (t, λ) , λ)
∂λ

Now we see that A (t, λ), B (t, λ) are continuous functions. Let

∂x 
S (t) =
∂λ λ=λ0

Then S (t) is the unique solution of the equation

Ṡ (t) = A (t, λ0 ) S (t) + B (t, λ0 )

S (0) = 0

P. Al Hokayem & E. Gallestey 41 of 214

Nonlinear Systems and Control — Spring 2018

The function S (t) is called the sensitivity function. Sensitivity functions

provide a first-order estimate of the effect of parameter variations on the
solutions of the differential equations.
They can also be used to approximate the solution when λ is sufficiently
close to it nominal value λ0 . Indeed, for small kλ − λ0 k, x (t, λ) can be
expanded in a Taylor series about the nominal solution x (t, λ0 ). Neglecting
the higher-order terms, the solution x (t, λ) can be approximated by

x (t, λ) ≈ x (t, λ0 ) + S (t) (λ − λ0 )

Example 2.10 Van der Pol Oscillator. Let us consider (again) the following
system.

ẋ1 = x2


ẋ2 = −x1 + ε 1 − x21 x2

First, we plot the nominal trajectories of the nonlinear system for ε = 1 and
the corresponding nominal phase portrait.

Figure 2.2: Oscillator Nominal Trajectories

P. Al Hokayem & E. Gallestey 42 of 214

Nonlinear Systems and Control — Spring 2018

Figure 2.3: Oscillator Nominal Phase Portrait

The expressions for the sensitivity equations are

 

δf  0 1
A (t, λ) = δx λ=λ0 = 2
 −1 − 2εx1 x 2 ε (1 − x1 )
 0
B (t, λ) = δf  =
δλ λ=λ0 (1 − x21 ) x2

The sensitivity equations can now be solved along the nominal trajectory.
The results can be inspected in Figure 2.4. Examining the shape of these
curves, we conclude that the solutions are more sensitive to parameter dis-
turbances at t = 4 and t = 7, which correspond to the points x = (2, 2)
and x = (1, −1) in the phase portrait’Figure 2.3. In Figure 2.5 we plot the
approximated trajectories together with the nominal phase portrait.

x(t, ε) = xnominal (t, ε0 ) + S(t)(ε − ε0 ), ε = 1.75

P. Al Hokayem & E. Gallestey 43 of 214

Nonlinear Systems and Control — Spring 2018

Figure 2.4: Oscillator Nominal Sensitivity

Appendix: Contraction Mappings and Fixed

Point Theorem
Let (S, kk) be a non-empty complete metric space (Banach). Let P : X → X
be a contraction mapping on S, i.e: there is a nonnegative real number ρ such
that for all x, y in S.

kP (x) − P (y)k ≤ ρ kx − yk , ∀x, y ∈ S, 0≤ρ<1

Then the map P admits one and only one fixed point x ∗ in S (this means
Px ∗ = x ∗ ). Furthermore, this fixed point can be found as follows: start
with an arbitrary element x 0 in S and define an iterative sequence by x[n] =
P (x[n − 1]) for n = 1, 2, 3, · · · This sequence converges, and its limit is x∗ .
When using the theorem in practice, the most difficult part is typically to
define S properly so that P actually maps elements from S to S, i.e. that
P x is always an element of S.

P. Al Hokayem & E. Gallestey 44 of 214

Nonlinear Systems and Control — Spring 2018

Figure 2.5: Phase Portrait after Perturbation

2.4 Exercises
1. Derive the sensitivity equations for the system

ẋ1 = tan−1 (ax1 ) − x1 x2

(2.2)
ẋ2 = bx21 − cx2

as the parameters a, b, c vary from their nominal values a0 = 1, b0 = 0,

and c0 = 1.

2. Determine whether the following differential equations have a unique

solution for t ∈ [0, ∞):

(a) ẋ(t) = 3x(t)

(b) ẋ(t) = ex(t) − 1 − x Hint: expand ex in a Taylor’s series
(c) ẋ(t) = e−|x(t)| − 1

ẋ1 (t) = −x31 (t)
(d)
ẋ2 (t) = x1 (t) − x2 (t) − x31 (t)

P. Al Hokayem & E. Gallestey 45 of 214

Nonlinear Systems and Control — Spring 2018

3. Consider the following Cauchy problem

( √
ẋ = x
.
x(0) = 1

Is there a unique solution to it? Why? If so, find its analytical expres-
sion.
What if the initial condition becomes x(0) = 0? Can you write down
at least two solutions to the problem? How many solution are there?
Motivate your answer.

4. The functions f1 : R → R and f2 : R → R are locally Lipschitz

continuous . Prove that the following functions are locally Lipschitz
continuous:

(a) f1 + f2 : x → f1 (x) + f2 (x)

(b) f1 · f2 :→ f1 (x)f2 (x)
(c) f1 o f2 : x → f2 (f1 (x))

5. Given a square matrix A ∈ Rn×n , consider the linear mapping P :

Rn 7→ Rn
P (x) = Ax .
If A is symmetric, under which conditions is P (x) a contraction from Rn
into itself? What about the general case, i.e. when A is not symmetric?
Hint: recall that a mapping P : Rn 7→ Rn is contractive if ||P (x) −
P (y)|| < ρ||x − y|| ∀x, y ∈ Rn with ρ < 1.

P. Al Hokayem & E. Gallestey 46 of 214

Chapter 3

Lyapunov Stability Theory

3.1 Introduction
In this lecture we consider the stability of equilibrium points of nonlinear
systems, both in continuous and discrete time. Lyapunov stability theory
is a standard tool and one of the most important tools in the analysis of
nonlinear systems. It may be utilized relatively easily to provide a strategy
for constructing stabilizing feedback controllers.

3.2 Stability of Autonomous Systems

Consider the nonlinear autonomous (no forcing input) system

ẋ = f (x) (3.1)

where f : D −→ Rn is a locally Lipschitz map from the domain D ⊆ Rn

to Rn . Suppose that the system (3.1) has an equilibrium point x̄ ∈ D, i.e.,
f (x̄) = 0. We would like to characterize if the equilibrium point x̄ is stable.
In the sequel, we assume that x̄ is the origin of state space. This can be
done without any loss of generality since we can always apply a change of
variables to ξ = x − x̄ to obtain

ξ˙ = f (ξ + x̄) , g(ξ) (3.2)

and then study the stability of the new system with respect to ξ = 0, the
origin.

47
Nonlinear Systems and Control — Spring 2018

Bδ
x0
B

Figure 3.1: Illustration of the trajectory of a stable system

Definition 3.1 The equilibrium x = 0 of (3.1) is

1. stable, if for each  > 0 there exists a δ > 0 such that

kx(t0 )k < δ ⇒ kx(t)k < , ∀t > t0 (3.3)

2. asymptotically stable if it is stable and in addition δ can be chosen

such that
kx(t0 )k < δ ⇒ lim kx(t)k = 0. (3.4)
t→∞

The following section provides characterizations of stability.

3.2.1 Lyapunov’s Direct Method

Let V : D −→ R be a continuously differentiable function defined on the
domain D ⊂ Rn that contains the origin. The rate of change of V along the
trajectories of (3.1) is given by
n
X ∂V d
d
V̇ (x) , V (x) = xi
dt ∂x i dt
i=1 (3.5)
 ∂V ∂V ∂V
 ∂V
= ∂x1 ∂x2 · · · ∂x ẋ = f (x)
n
∂x
The main idea of Lyapunov’s theory is that if V̇ (x) is negative along the
trajectories of the system, then V (x) will decrease as time goes forward.
Moreover, we do not really need to solve the nonlinear ODE (3.1) for every
initial condition; we only need some information about the drift f (x).

P. Al Hokayem & E. Gallestey 48 of 214

Nonlinear Systems and Control — Spring 2018

Figure 3.2: Lyapunov function in two states V = x21 + 1.5x22 . The level sets
are shown in the x1 x2 -plane.

Example 3.2 Consider the nonlinear system

   
f1 (x) −x1 + 2x21 x2
ẋ = f (x) = =
f2 (x) −x2

and the candidate Lyapunov function

V (x) = λ1 x21 + λ2 x22

with λ1 , λ2 > 0. If we plot the function V (x) for some choice of λ’s we obtain
the result in Figure 3.2. This function has a unique minimum over all the
state space at the origin. Moreover, V (x) → ∞ as kxk → ∞.
Calculate the derivative of V along the trajectories of the system

V̇ (x) = 2λ1 x1 (−x1 + 2x21 x2 ) + 2λ2 x2 (−x2 ) = −2λ1 x21 − 2λ2 x22 + 4λ1 x31 x2

If V̇ (x) is negative, V will decrease along the solution of ẋ = f (x).

We are now ready to state Lyapunov’s stability theorem.

P. Al Hokayem & E. Gallestey 49 of 214

Nonlinear Systems and Control — Spring 2018

Theorem 3.3 (Direct Method) Let the origin x = 0 ∈ D ⊂ Rn be an

equilibrium point for ẋ = f (x). Let V : D → R be a continuously differen-
tiable function such that

V (0) = 0 and V (x) > 0, ∀x ∈ D\ {0}

(3.6)
V̇ (x) ≤ 0, ∀x ∈ D

Then, x = 0 is stable. Moreover, if

V̇ (x) < 0, ∀x ∈ D\ {0}

then x = 0 is asymptotically stable

Remark 3.4 If V (x) > 0, ∀x ∈ D\ {0}, then V is called locally positive

definite. If V (x) ≥ 0, ∀x ∈ D\ {0}, then V is locally positive semi-definite.
If the conditions (3.6) are met, then V is called a Lyapunov function for the
system ẋ = f (x).

Proof: Given any ε > 0, choose r ∈ (0, ε] such that Br = {x ∈ Rn , kxk ≤ r} ⊂

D. Let α = min V (x). Choose β ∈ (0, α) and define Ωβ = {x ∈ Br , V (x) ≤ β}.
kxk=r

Ωβ
Bδ
Br

Figure 3.3: Various domains in the proof of Theorem 3.3

It holds that if x (0) ∈ Ωβ ⇒ x (t) ∈ Ωβ ∀t because

V̇ (x (t)) ≤ 0 ⇒ V (x (t)) ≤ V (x (0)) ≤ β

P. Al Hokayem & E. Gallestey 50 of 214

Nonlinear Systems and Control — Spring 2018

Further ∃δ > 0 such that kxk < δ ⇒ V (x) < β. Therefore, we have that

Bδ ⊂ Ωβ ⊂ Br

and furthermore

x (0) ∈ Bδ ⇒ x (0) ∈ Ωβ ⇒ x (t) ∈ Ωβ ⇒ x (t) ∈ Br

Finally, it follows that

kx (0)k < δ ⇒ kx (t)k < r ≤ ε, ∀t > 0

In order to show asymptotic stability, we need to to show that x(t) → 0 as

t → ∞. In this case, it turns out that it is sufficient to show that V (x(t)) → 0
as t → ∞. Since V is monotonically decreasing and bounded from below by
0, then
V (x) → c ≥ 0, as t → ∞
Finally, it can be further shown by contradiction that the limit c is actually
equal to 0, using the following estimate of the decay rate
Z t
V (t) ≤ V (0) + V̇ (x(τ ))dτ − γt
0

where −γ = maxd≤kxk≤r V̇ (x) is the slowest rate of decay of V (x) over a

compact set. γ exists since the function V̇ (x) is continuous.


Example 3.5 Recall Example 3.2. The derivative of the Lyapunov function
candidate was given by

V̇ (x) = −2λ1 x21 − 2λ2 x22 + 4λ1 x31 x2

For simplicity, assume that λ1 = λ2 = 1. Then

V̇ (x) = −2x22 − 2x21 g(x)

where g(x) , 1−2x1 x2 . Then the derivative of V is guaranteed to be negative

whenever g(x) > 0. The level sets of V, where V̇ < 0 will be invariant, or
equivalently when g(x) > 0, i.e., when x1 x2 < 1/2. So we conclude that the
origin is locally asymptotically stable.

P. Al Hokayem & E. Gallestey 51 of 214

Nonlinear Systems and Control — Spring 2018

F (x)

M
Ff (ẋ) = −δ ẋ

0 x

Figure 3.4: Mass spring system

Example 3.6 Consider a mass M connected to a spring, as shown in Figure

4.2, where x = 0 is defined as the equilibrium, the point where there is no
force exerted by the spring, i.e.,
F (x)x > 0, ∀x 6= 0, F (x) = 0 ⇔ x = 0
The dynamics of this system are given by
M ẍ = −F (x) − δ ẋ (3.7)
Let x1 , x, x2 , ẋ and M = 1, then
   
ẋ1 x2
= = f (x) (3.8)
ẋ2 −F (x1 ) − δx2
Consider the Lyapunov function candidate
Z x1
1
V (x) = F (s) ds + x22
0 2
 
Then ∂V∂x(x) = F (x1 ) , x2 and1
V̇ (x) = F (x1 ) x2 + x2 (−F (x1 ) − δx2 ) = −δx22 ≤ 0
Therefore,
the
system is stable but we cannot prove asymptotic stability be-
x1
cause V̇ = 0, ∀x1 . However, we shall see that using LaSalle’s in-
0
variance principle we can conclude that the system is in fact asymptotically
stable.
1
Here we use the Leibniz rule for differentiating integrals, i.e.,
 
b(φ)
Z b(φ)
Z
d   ∂ d d
 f (ξ, φ) dξ = f (ξ, φ)dξ + f (b(φ), φ) b(φ) − f (a(φ), φ) a(φ)
dφ ∂φ dφ dφ
a(φ) a(φ)

P. Al Hokayem & E. Gallestey 52 of 214

Nonlinear Systems and Control — Spring 2018

The result in Theorem 3.3 can be extended to become a global result.

Theorem 3.7 Let x = 0 be an equilibrium point of the system ẋ = f (x).

Let V : Rn −→ R be a continuously differentiable function such that

V (0) = 0 and V (x) > 0, ∀x 6= 0 (3.9)

kxk → ∞ ⇒ V (x) → ∞ (3.10)
V̇ (x) < 0, ∀x 6= 0 (3.11)

then the origin is globally asymptotically stable.

Remark 3.8 If the function V satisfies the condition (3.10), then it is said
to be radially unbounded.

Example 3.9 Consider the system

 
x2
ẋ =
−h(x1 ) − x2

where the function h is locally Lipschitz with h(0) = 0 and x1 h(x1 ) >
0, ∀x1 6= 0. Take the Lyapunov function candidate
  Zx1
1 T k k
V (x) = x x + h(s)ds
2 k 1
0

The function V is positive definite for all x ∈ R2 and is radially unbounded.

The derivative of V along the trajectories of the system is given by

V̇ (x) = −(1 − k)x22 − kx1 h(x1 ) < 0

Hence the derivative of V is negative definite for all x ∈ R2 , since 0 < k < 1
(otherwise V is not positive definite! - convince yourself). Therefore, the
origin is globally asymptotically stable.

3.2.2 Lyapunov’s Indirect Method

In this section we prove stability of the system by considering the properties
of the linearization of the system around the origin. Before proving the main
result, we require an intermediate result.

P. Al Hokayem & E. Gallestey 53 of 214

Nonlinear Systems and Control — Spring 2018

Definition 3.10 A matrix A ∈ Rn×n is called Hurwitz or asymptotically

stable, if and only if

Re(λi ) < 0, ∀i = 1, 2, · · · , n

where λi ’s are the eigenvalues of the matrix A.

Consider the system ẋ = Ax. We look for a quadratic function

V (x) = xT P x

where P = P T > 0. Then

V̇ (x) = ẋT P x + xT P ẋ = xT AT P + P A x = −xT Qx

If there exists Q = QT > 0 such that

AT P + P A = −Q,

then V is a Lyapunov function and x = 0 is globally stable. This equation

is called the Matrix Lyapunov equation.
We formulate this as a matrix problem: Given Q positive definite, sym-
metric, how can we find out if there exists P = P T > 0 satisfying the Matrix
Lyapunov equation.
The following result gives existence of a solution to the Lyapunov matrix
equation for any given Q.

Theorem 3.11 For A ∈ Rn×n the following statements are equivalent:

1. A is Hurwitz

2. For all Q = QT > 0 there exists a unique P = P T > 0 satisfying the

Lyapunov equation
AT P + P A = −Q

Proof: We make a constructive proof of 1. ⇒ 2. For a given Q = QT > 0,

consider the following candidate solution for P :
Z ∞
T
P = eA t QeAt dt
0

P. Al Hokayem & E. Gallestey 54 of 214

Nonlinear Systems and Control — Spring 2018

That P = P T > 0 follows from the properties of Q. Note that the integral
will converge if and only if A is a Hurwitz matrix. We now show that P
satisfies the matrix Lyapunov equation:
R∞h T T
i
AT P + P A = 0 AT eA t QeAt + eA t QeAt A dt
R ∞ d h AT t At i
= 0 dt
e Qe dt
∞
T 
= eA t QeAt  = −Q
0

Thus P satisfies the matrix Lyapunov equation. In order to show uniqueness,

assume that here exists another matrix P̄ = P̄ T > 0 that solves the Lyapunov
equation and P̄ 6= P . Then,

AT (P − P̄ ) + (P − P̄ )A = 0

From which it follows that

Tt  T  d h AT t i
0 = eA A (P − P̄ ) + (P − P̄ )A eAt = e (P − P̄ )eAt
dt
Therefore,
T
eA t (P − P̄ )eAt = a, ∀t
where a is some constant matrix. Now, this also holds for t = 0, i.e.,
T T
eA 0 (P − P̄ )eA0 = (P − P̄ ) = eA t (P − P̄ )eAt → 0, as t → ∞

Where the last limit follows from the fact that A is Hurwitz. Therefore,
P = P̄ .
The fact that 2. ⇒ 1. follows from taking V (x) = xT P x.

Theorem 3.11 has an interesting interpretation in terms of the energy avail-

able to a system. If we say that the energy dissipated at a particular point
in phase space x is given by q(x) = xT Qx - meaning that a trajectory pass-
ing through x is losing q(x)units of energy per unit time, then the equation
V (x) = xT P x, where P satisfies the matrix Lyapunov equation gives the
total amount of energy that the system will dissipate before reaching the
origin. Thus V (x) = xT P x measures the energy stored in the state x. We
shall revisit this concept in the next chapter.

P. Al Hokayem & E. Gallestey 55 of 214

Nonlinear Systems and Control — Spring 2018

Theorem 3.12 (Indirect Method) Let x = 0 be an equilibrium point for

ẋ = f (x) where f : D → Rn is a continuously differentiable and D is a
neighborhood of the origin. Let

∂f 
A= (3.12)
∂x x=0

then
1. The origin is asymptotically stable if Re (λi ) < 0 for all eigenvalues of
A

2. The origin is unstable if Re (λi ) > 0 for one or more of the eigenvalues
of A

Proof: If A is Hurwitz, then there exists P = P T > 0 so that V (x) = xT P x

is a Lyapunov function of the linearized system. Let us use V as a candidate
Lyapunov function for the nonlinear system

ẋ = f (x) = Ax + (f (x) − Ax) , Ax + g(x)

The derivative of V is given by

V̇ (x) = xT P f (x) + f (x)T P x = xT P Ax + xT AT P x + 2xT P g(x) (3.13)

= −xT Qx + 2xT P g(x) (3.14)

The function g(x) satisfies

kg(x)k2
→ 0 as kxk2 → 0
kxk2
Therefore, for any γ > 0, there exists r > 0 such that

kg(x)k2 < γ kxk2 , ∀ kxk2 < r

As such,
V̇ (x) < −xT Qx + 2γ kP k2 kxk22 , ∀ kxk2 < r
But we have the following fact xT Qx ≥ λmin (Q) kxk22 , where λmin indicates
the minimal eigenvalue of Q. Therefore,

V̇ (x) < −(λmin (Q) − 2γ kP k2 ) kxk22 , ∀ kxk2 < r

P. Al Hokayem & E. Gallestey 56 of 214

Nonlinear Systems and Control — Spring 2018

and choosing γ < λ2kP

min (Q)
k2
renders V̇ (x) negative definite (locally). Hence, the
origin of the nonlinear system is asymptotically stable. This proves point 1.
The proof of point 2 shall be omitted; however, it relies on the Instability
results in Section 3.4. More details can be found in [4, Ch. 4].

Theorem 3.12 does not say anything when Re (λi ) ≤ 0 ∀i with Re (λi ) = 0
for some i. In this case linearization fails to determine the stability of the
equilibrium point, and further analysis is necessary. The multi-dimensional
result which is relevant here is the Center Manifold Theorem. This theorem
is beyond the scope of this course.

Example 3.13 The system ẋ = ax3 , a > 0 has an unstable equilibrium point
at x = 0. The same system is asymptotically stable at the origin for a < 0.
In both cases the linearized system is given by ẋ = 0, and we cannot conclude
anything from the linearization or Lyapunov’s indirect method. However, if
we use Lyapunov’s direct method with V (x) = 41 x4 , then V̇ (x) = ax6 and if
a < 0 then the system is globally asymptotically stable.

3.3 The Invariance Principle

Definition 3.14 A domain D ⊆ Rn is called invariant for the system ẋ =
f (x), if
∀x (t0 ) ∈ D ⇒ x (t) ∈ D, ∀t ∈ R
A domain D ⊆ Rn is called positively invariant for the system ẋ = f (x), if

∀x (t0 ) ∈ D ⇒ x (t) ∈ D, ∀t ≥ t0

Theorem 3.15 (LaSalle’s Invariance Principle) Let Ω ⊂ D be a com-

pact set that is positively invariant with respect to ẋ = f (x). Let V : D → R
be a continuously differentiable function such that V̇ (x) ≤ 0 in Ω. Let E be
the set of all points in Ω where V̇ (x) = 0. Let M be the largest invariant set
in E. Then every solution starting in Ω approaches M as t → ∞.

Remark 3.16 Theorem 3.15 result can also be used to find limit cycles.

An important result that follows from Theorem 3.15 is the following corol-
lary.

P. Al Hokayem & E. Gallestey 57 of 214

Nonlinear Systems and Control — Spring 2018

Corollary 3.17 Let x = 0 ∈ D be an equilibrium point of the system

ẋ = f (x). Let V : D → R be a continuously differentiable positive defi-
nite function on the domain D, such that V̇ (x) ≤ 0, ∀x ∈ D. Let S , {x ∈
D | V̇ (x) = 0} and suppose that no solution can stay identically in S, other
than the trivial solution x(t) ≡ 0. Then, the origin is asymptotically stable.
Example 3.18 Consider again the mass-spring example 3.6. Apply LaSalles
invariance principle by noting that
V̇ (x) = 0 ⇔ x2 = 0
As such E = {x | x2 = 0}. Now, if x2 = 0, then from the dynamics we
obtain F (x1 ) = 0 and since F (x1 ) = 0 ⇔ x1 = 0, we conclude that M = {0}.
Thus, trajectories must converge to the origin, and we have proven that the
system is asymptotically stable.

3.4 Instability Theorem

Theorem 3.19 Let x = 0 be an equilibrium point of the system ẋ = f (x).
Let V : D → R be a continuously differentiable function such that
V (0) = 0, and V (x0 ) > 0
for some x0 with arbitrary small kx0 k. Define the set
U = {x ∈ Br | V (x) > 0}
and suppose that V̇ (x) > 0 in U . Then x = 0 is unstable.
Sketch of proof: First note that x0 belongs to the interior of U . Moreover,
we can show that the trajectory starting at x0 must leave the set U and that
the trajectory leaves the set U through the surface kxk2 = r. Since this can
happen for arbitrarily small kx0 k, thus the origin is unstable.

Note that the requirements on V (x) in Theorem 3.19 are not as strict as the
requirements on a Lyapunov function that proves stability.
Example 3.20 The set U for V (x) = 21 (x21 − x22 ) is shown in Figure 3.5.
Consider the system    
ẋ1 x1
=
ẋ2 −x2
then V̇ (x) = x21 + x22 > 0 for x ∈ U , proving that the system is unstable.

P. Al Hokayem & E. Gallestey 58 of 214

Nonlinear Systems and Control — Spring 2018

x2
x2 = x1
Br

U U
x1

x2 = −x1

Figure 3.5: Instability regions

3.5 Comparison Functions

Before we move on to the analysis of nonautonomous systems, we need to
introduce few definitions of classes of functions that will aid us in extending
the analysis we have done so far to the nonautonomous case.
Definition 3.21 A continuous function α1 : [0, a) → [0, ∞) belongs to class
K if it is strictly increasing and α1 (0) = 0. Moreover, α2 belongs to class
K∞ if a = ∞ and α2 (r) → ∞ as r → ∞.

α2 (r)

α1 (r)

r
0 a
Figure 3.6: Illustration of α1 ∈ K and α2 ∈ K∞ )

For example α(r) = tan−1 (r) belongs to class K but not to class K∞ , while
the functions α(r) = r2 and α(r) = min{r, r4 } belong to class K∞ (why?).

Definition 3.22 A continuous function β : [0, a) × [0, ∞) → [0, ∞) belongs

to class KL if for each fixed s, the mapping β(r, s) belongs to class K with

P. Al Hokayem & E. Gallestey 59 of 214

Nonlinear Systems and Control — Spring 2018

respect to r, and for each fixed r, β(r, s) is decreasing with respect to s and
β(r, s) → 0 as s → ∞.
r
The functions β(r, s) = r2 e−s and β(r, s) = sr+1
belong to class KL. The

r2e-s

4000

3000

2000

1000

0
100
0
50
5

r 0 10
s

Figure 3.7: Illustration of class KL function: β(r, s) = r2 e−s

following result is quite useful for analysis.

Lemma 3.23 Let α1 , α2 be class K functions on [0, a), a3 , a4 be class K∞

functions, and β be class KL function. Then,

• α1−1 is defined on [0, α1 (a)) and belongs to class K

• α3−1 is defined on [0, ∞) and belongs to class K∞

• α1 ◦ α2 belongs to class K

• α3 ◦ α4 belongs to class K∞

• γ(r, s) = α1 (β(α2 (r), s)) belongs to class KL

P. Al Hokayem & E. Gallestey 60 of 214

Nonlinear Systems and Control — Spring 2018

with the understanding that αi−1 denotes the inverse of αi and αi ◦ αj =

αi (αj (r)).
We can relate the foregoing definitions to our previous definitions of positive
definite functions.
Lemma 3.24 Let V : D → R be a continuous positive definite function
defined on the domain D ⊂ Rn that contains the origin. Let Br ⊂ D be a
ball of radius r > 0. Then, there exist class K functions α1 and α2 defined
on [0, r) such that
α1 (kxk) ≤ V (x) ≤ α2 (kxk)
Moreover, if D = Rn and V (x) is radially unbounded, then the last inequality
holds for α1 and α2 in class K∞ .
For example, if V (x) = xT P x with P being symmetric positive definite,
then it follows that

α1 (kxk) , λmin (P ) kxk22 ≤ V (x) ≤ λmax (P ) kxk22 , α2 (kxk)

3.6 Stability of Nonautonomous Systems

Consider the system
ẋ = f (t, x) (3.15)
where f : [0, ∞)×D → Rn is piecewise continuous in t and locally Lipschitz in
x on the domain [0, ∞)×D, and D ⊂ Rn contains the origin. We say that the
origin is an equilibrium point of (3.15) if f (t, 0) = 0, ∀t ≥ 0. Henceforth,
we shall assume that the origin is an equibrium point of the system (3.15),
since we can apply a similar analysis to the case of autonomous systems to
shift any equilibrium point to the origin.
We saw in the stability analysis of autonomous systems that the solution
depends on the difference (t − t0 ), where t0 is the initial time instant. For
nonautonomous systems, the solution generally depends on both t and t0 .
As such, we expect a modification of the definition of stability as seen next.
Definition 3.25 The equilibrium point x = 0 of the system (3.15) is
1. stable, if for each  > 0, there is a δ = δ(, t0 ) > 0 such that

kx(t0 )k < δ ⇒ kx(t)k < , ∀t ≥ t0 (3.16)

P. Al Hokayem & E. Gallestey 61 of 214

Nonlinear Systems and Control — Spring 2018

2. uniformly stable, if for each  > 0, there is a δ = δ() > 0 (independent

of t0 ) such that (3.16) holds.

3. unstable, if it is not stable

4. asymptotically stable if it is stable and there is a positive constant c =

c(t0 ) such that x(t) → 0 as t → ∞, for all kx(t0 )k < c

5. uniformly asymptotically stable, if it is uniformly stable and the con-

stant c is independent of t0 , i.e., there exists a time horizon T = T (η) >
0 such that

kx(t)k < η, ∀t ≥ t0 + T (η), ∀ kx(t0 )k < c (3.17)

6. globally uniformly asymptotically stable, if it is uniformly asymptoto-

icaly stable, δ() can be chosen to satisfy lim→∞ δ() = ∞, and, for
each pair of positive numbers η and c, there is T = T (η, c) > 0 such
that
kx(t)k < η, ∀t ≥ t0 + T (η, c), ∀ kx(t0 )k < c (3.18)

In terms of class K, K∞ and KL definitions above can be characterized as

follows: The origin of (3.15) is:

• Uniformly Stable (US) ⇔ there exist a function α ∈ K and a constant

c > 0 independent of t0 such that

kx(t)k ≤ α (kx(t0 )k) , ∀t ≥ t0 ≥ 0, ∀ kx(t0 )k < c (3.19)

• Uniformly asymptotically stable (UAS) ⇔ there exist a function β ∈

KL and a constant c > 0 independent of t0 such that

kx(t)k ≤ β (kx(t0 )k , t − t0 ) , ∀t ≥ t0 ≥ 0, ∀ kx(t0 )k < c (3.20)

• Globally UAS (GUAS) ⇔ there exist a function β ∈ KL such that

kx(t)k ≤ β (kx(t0 )k , t − t0 ) , ∀t ≥ t0 ≥ 0, ∀x(t0 ) (3.21)

• Exponentially stable if there exist constants c, k, λ > 0 such that

kx(t)k ≤ k kx(t0 )k e−λ(t−t0 ) , ∀t ≥ t0 ≥ 0, ∀ kx(t0 )k < c (3.22)

P. Al Hokayem & E. Gallestey 62 of 214

Nonlinear Systems and Control — Spring 2018

• Globally exponentially stable if the condition (3.22) is satisfied for any

initial state x(t0 ).

−→ Can you write down the corresponding notions for the autonomous sys-
tems?
Similarly to the autonomous systems case, we can write the Lyapunov
characterizations of stability for nonautonomous systems.

Theorem 3.26 Consider the system (3.15) and let x = 0 ∈ D ⊂ Rn be an

equilibrium point. Let V : [0, ∞) × D → R be a continuously differentiable
function such that

W1 (x) ≤ V (t, x) ≤ W2 (x) (3.23)

∂V ∂V
V̇ (t, x) = + f (t, x) ≤ 0 (3.24)
∂t ∂x
∀t ≥ 0 and ∀x ∈ D, where W1 and W2 are continuous positive definite
functions on D. Then, the origin is uniformly stable. If instead we have the
stronger conditions

W1 (x) ≤ V (t, x) ≤ W2 (x) (3.25)

∂V ∂V
V̇ (t, x) = + f (t, x) ≤ −W3 (x) (3.26)
∂t ∂x
with all the previous assumptions holding and W3 is continuous positive def-
inite on D, then the origin is UAS.
Moreover, if we choose some constants r and c such that Br = {kxk ≤
r} ⊂ D and c < minkxk=r W1 (x), then every trajectory starting in {x ∈
Br |W2 (x) ≤ c} satisfies the estimate

kxk ≤ β(kx0 k , t − t0 ) ∀t ≥ t0 ≥ 0 (3.27)

for some β ∈ KL. Finally, if D = Rn and W1 (x) is radially unbounded, then

the origin is GUAS.

Sketch of Proof: The proof is similar to that of Theorem 3.3, however now
we work with level sets that are formed by the functions Wi . Consider Figure
3.8. Choose some radius r > 0 and a constant c > 0, such that Br ⊂ D and
c < minkxk=r W1 (x) ⇒ S1 := {x ∈ Br |W1 (x) ≤ c} ⊂ interior(Br ). Define

P. Al Hokayem & E. Gallestey 63 of 214

Nonlinear Systems and Control — Spring 2018

W2 (x)

V (t, x)
D W1 (x)
Br
S1
c
Ωt,c

S2

Figure 3.8: A sketch of the various domains of Lyapunov stability for nonau-
tonomous systems (left) and the corresponding functions (right)

the set (time-independent) Ωt,c := {x ∈ Br |V (t, x) ≤ c}. It follows from the

relation (3.23) that
S2 := {x ∈ Br |W2 (x) ≤ c} ⊂ Ωt,c ⊂ S1 ⊂ Br ⊂ D
which shows along with (3.24) that any solution that starts in S2 remains in
S1 for all time beyond t0 . Moreover, we have that
V (t, x(t)) ≤ V (t0 , x(t0 )), ∀t ≥ t0 (3.28)
α1 (kxk) ≤ W1 (x) ≤ V (t, x) ≤ W2 (x) ≤ α2 (kxk) (3.29)
for some α1 , α2 ∈ K[0,r] , from which it follows that
kx(t)k ≤ α1−1 (α2 (kx(t0 )k)), ∀t0 (3.30)
and hence the origin is uniformly stable.
Under the stronger condition (3.26), we have the following differential
inequality
V̇ ≤ −W3 (x) ≤ −α3 (α1−1 (V (x))) (3.31)
where α3 ∈ K[0,r] and W3 (x) ≥ α3 (kxk). Using the comparison system
ẏ = −α3 (α1−1 (y)), y(t0 ) = V (t0 , x(t0 )) ≥ 0
we can show under some technical conditions that
V (t, x(t)) ≤ σ(V (t0 , x(t0 )), t − t0 ), ∀V (t0 , x(t0 )) ∈ [0, c] (3.32)
−1
⇒ kx(t)k ≤ α1 (σ(α2 (kx(t0 )k), t − t0 )) =: β(kx(t0 )k , t − t0 )
(3.33)

P. Al Hokayem & E. Gallestey 64 of 214

Nonlinear Systems and Control — Spring 2018

where σ ∈ KL[0,r]×[0,∞) . Hence the origin is UAS. Finally, if D = Rn , then

all αi ’s are defined on [0, ∞) from which we can deduce that the system is
GUAS.

The following two examples illustrate the utilization of the result of The-
orem 3.26.
Example 3.27 Consider the scalar system
ẋ = −(1 + g(t))x3
where g(t) is a continuous function with g(t) ≥ 0, ∀t. Take the time-invariant
2
candidate Lyapunov function V (x) = x2 . Then,

V̇ (x) = −(1 + g(t))x4 ≤ −x4 , ∀x ∈ R, t ≥ 0

Hence, the system is GUAS.
Example 3.28 Consider the second order system
ẋ1 = −x1 − g(t)x2
ẋ2 = x1 − x2
where g(t) is a continuously differentiable function, such that 0 ≤ g(t) ≤ k
and ġ(t) ≤ g(t), ∀t and for some positive constant k. We now consider the
time-varying Lyapunov function candidate V (t, x) = x21 + (1 + g(t))x22 . First,
we have that
α1 (kxk) , x21 + x22 ≤ V (t, x) ≤ x21 + (1 + k)x22 , α2 (kxk)
Hence, the function V is bounded above and below by class K∞ functions.
We now take the derivative of V along the trajectories of the system:
∂V ∂V
V̇ (t, x) = + ẋ
∂t ∂x
= x22 ġ(t) + 2x1 ẋ1 + 2(1 + g(t))x2 ẋ2
= x22 ġ(t) + 2x1 (−x1 − g(t)x2 ) + 2(1 + g(t))x2 (x1 − x2 )
= −2x21 + 2x1 x2 − (2 + 2g(t) − ġ(t))x22
 T   
2 2 x1 2 −1 x1
≤ −2x2 + 2x1 x2 − 2x2 = − <0
x2 −1 2 x2

P. Al Hokayem & E. Gallestey 65 of 214

Nonlinear Systems and Control — Spring 2018

 
2 −1
since Q = > 0. Hence, the system is GUAS.
−1 2
Moreover, since we have that
ρ1 kxk2 ≤ V (t, x) ≤ ρ2 kxk2 , and V̇ (t, x) ≤ −ρ3 kxk2 (3.34)

for a choice of ρ1 , 1, ρ2 , 1 + k, and ρ3 , λmin (Q), it follows that

ρ3
V̇ (t, x) ≤ − V (t, x)
ρ2
From which it follows that (using the so-called Comparison Lemma in the
Appendix)
ρ
− 3 (t−t )
V (t, x) ≤ V (t0 , x(t0 ))e ρ2 0 (3.35)
Using (3.34) again, we have that
1 1 ρ
− 3 (t−t )
kx(t)k2 ≤ V (t, x) ≤ V (t0 , x(t0 ))e ρ2 0
ρ1 ρ1
ρ2 ρ
2 − ρ3 (t−t0 )
≤ kx(t0 )k e 2
ρ1
q ρ
ρ2 − 2ρ3 (t−t0 )
⇒ kx(t)k ≤ ρ1
kx(t0 )k e 2 implying that the system is actually glob-
ally exponentially stable!

Remark 3.29 One can generalize the result of the last example on exponen-
tial stability if the bounds on V (t, x) and V̇ (t.x) satisfy the following
ρ1 kxkα ≤ V (t, x) ≤ ρ2 kxkα
V̇ (t, x) ≤ −ρ3 kxkα
for any positive constants ρ1 , ρ2 , ρ3 , α.

3.7 Existence of Lyapunov Functions

Having gone through Lyapunov methods of showing stability, a crucial ques-
tion remains regarding existence of a Lyapunov function: assume that x = 0
is an asymptotically stable equilibrium of the system (3.1), does a Lyapunov
function for the system exist? These types of results are often referred to
as converse theorems. Consider Autonomous systems for simplicity, we have
the following result.

P. Al Hokayem & E. Gallestey 66 of 214

Nonlinear Systems and Control — Spring 2018

Theorem 3.30 Let x = 0 be locally exponentially stable for the system ẋ =

f (x) on the domain D0 , {x ∈ Rn k kxk < r0 }. Then there exists a Lyapunov
function V : D → R+ for the system such that

c1 kxk2 ≤ V (x) ≤ c2 kxk2 (3.36)

2
V̇ (x) ≤ −c3 kxk (3.37)

∂V (x)

∂x ≤ c4 kxk (3.38)

where c1 , c2 , c3 , c4 > 0.

Sketch of Proof: Analogously to the construction of a solution to the matrix

Lyapunov equation, a Lyapunov Function is constructed by:
Z ∞
V (x) = φ(t, x)T φ(t, x)dt
0

where φ(t, x) is the solution to the system differential equations, defining the
trajectory starting at x at time t = 0. Due to local exponential stability, the
integral may be shown to converge locally. By bounding the rate of growth
of the integral away from 0, the properties may be proven.

The last theorem can be generalized for the non-autonomous systems case.

3.8 Input-to-State Stability

We end this chapter with a robustness-type result. Consider again the nonau-
tonomous system
ẋ = f (t, x, d) (3.39)
with the same assumptions as in the last section on t and x, and where d(t) is
a piecewise continuous function on Rm which is bounded for all t ≥ 0. Assume
that the unforced system ẋ = f (t, x, 0) enjoys some stability properties, what
can we say in the presence of the input d?

Definition 3.31 The system (3.39) is said to be input-to-state stable (ISS)

if there exists a class KL function β and a class K function γ such that for

P. Al Hokayem & E. Gallestey 67 of 214

Nonlinear Systems and Control — Spring 2018

any initial state x(t0 ) and any bounded input d(t), the solution x(t) exists for
all t ≥ t0 and
!
kx(t)k ≤ β (kx(t0 )k , t − t0 ) + γ sup kd(τ )k (3.40)
τ ∈[t0 ,t]

The function β allows us to determine the response of the system in terms

of the initial condition x0 and the corresponding overshoot. The function γ
determines the asymptotic behavior of the system with respect to a bounded
disturbance input d. This concept generalizes what we already know about
linear systems. More specifically, consider the linear system with A Hurwitz
ẋ = Ax + Bd, x(0) = x0 (3.41)
The solution is given by
Z t
At
x(t) = e x0 + eA(t−η) Bd(η)dη (3.42)
0

from which we can deduce the following inequality on the size of the state
over time
Z t
At
kx(t)k ≤ e kx0 k + e
A(t−η)
dηB
sup kd(τ )k
0 τ ∈[0,t]
(3.43)
, β(kx0 k , t) + γ sup kd(τ )k
τ ∈[0,t]

Therefore, assuming the matrix A is Hurwitz, the response due to the initial
condition asymptotically decays to 0 and we expect the state to asymptot-
ically converge to a ball that is proportional to the size of the input d. We
also have the following Lyapunov characterization of ISS.
Theorem 3.32 Let V : [0, ∞) × Rn → R be a continuously differentiable
function such that
α1 (kxk) ≤ V (t, x) ≤ α2 (kxk) (3.44)
∂V ∂V
+ f (t, x, d) ≤ −W (x), ∀ kxk ≥ ρ(kdk) > 0 (3.45)
∂t ∂x
for all (t, x, d) ∈ [0, ∞) × Rn × Rm , where α1 , α2 are class K∞ functions, ρ is
a class K function, and W is a continuous positive definite function on Rn .
Then, the system (3.39) is ISS with γ = α1−1 ◦ α2 ◦ ρ.

P. Al Hokayem & E. Gallestey 68 of 214

Nonlinear Systems and Control — Spring 2018

Note that if the system is autonomous, then the conditions of the result
above become necessary and sufficient.

Example 3.33 Consider the system

ẋ = f (x, u) = x + (x2 + 1)u (3.46)

We can simply design the control input u = x−2x

2 +1 that renders the closed-loop

system ẋ = −x which is globaly asymptotically stable. However, the question

remains if the resulting system is robust (in our context ISS). To see this,
assume that we have a small disturbance that comes in with the input, i.e.,
our input is actually u = x−2x
2 +1 + d. The resulting closed-loop system is given

by
ẋ = −x + (x2 + 1)d
This latter system can be shown to go unstable, for example if d ≡ 1 the
solution diverges to ∞ in finite time. Therefore, our feedback does not render
the system robust even to small disturbances! Alternatively, consider now the
new input u = x−2x
2 +1 − x, which gives the following closed-loop system under

the same type of disturbance as before

ẋ = −2x − x3 + (x2 + 1)d

Let us now show that this last system is actually ISS. Take the following
Lyapunov function candidate V = 12 x2 , then


V̇ = x −2x − x3 + (x2 + 1)d (3.47)
4
  4
 
x |d| x 2|d|
≤ −x2 − − x2 1 − − 1− (3.48)
2 |x| 2 |x|
x4
≤ −x2 − , ∀|x| > 2|d| (3.49)
2
Therefore, by Theorem 3.32, the closed-loop system is ISS.

Lemma 3.34 Consider the cascade system

ẋ1 = f1 (x1 , x2 ) (3.50)

ẋ2 = f2 (x2 , u) (3.51)

where the system (3.50) is ISS with respect to x2 , (3.51) is ISS with respect
to u. Then the origin of the cascade system (3.50)-(3.51) is ISS.

P. Al Hokayem & E. Gallestey 69 of 214

Nonlinear Systems and Control — Spring 2018

Sketch of Proof: Since each of the subsystems is ISS, then each admits an
ISS Lyapunov function, which can be shown to satisfy (after some rescaling)
the following inequalities

V̇1 (x1 , x2 ) ≤ −α1 (kx1 k) + ρ1 (kx2 k)

V̇2 (x1 , u) ≤ −2ρ1 (kx2 k) + ρ2 (kuk)

from which it follows that V := V1 + V2 is an ISS Lyapunov function that

satisfies
V̇ (x1 , x2 , u) ≤ −α1 (kx1 k) − ρ1 (kx2 k) + ρ2 (kuk)


Example 3.35 Consider the system

ẋ1 = −x1 − x2 + u1
ẋ2 = x1 − x2 + u2

and the quadratic function V (x1 , x2 ) = 21 (x21 + x22 ). The derivative of V along
the trajectories of the system is given by

V̇ = x1 (−x1 − x2 + u1 ) + x2 (x1 − x2 + u2 ) = −x21 − x22 + x1 u1 + x2 u2

≤ −(x21 + x22 )/2, ∀|xi | > 2|ui |

which shows that the system is ISS. Of course this is not surprising since we
know that the system is a stable linear one and hence is robust with respect
to bounded disturbances.

3.9 Stability of Discrete-Time Systems

One remaining question for this chapter is stability characterization of discrete-
time systems. Consider the discrete-time nonlinear system

x(k + 1) = f (x(k)) (3.52)

where f : D → Rn is a nonlinear map, and we use the shorthand notation of

indices k instead of the more precise one of kTs , with Ts being the sampling
period. We shall assume that the system has a equilibrium at the origin, i.e.,
f (0) = 0.

P. Al Hokayem & E. Gallestey 70 of 214

Nonlinear Systems and Control — Spring 2018

Remark 3.36 For the discrete-time system (3.52), the equilibrium point is
characterized by the fixed point condition x∗ = f (x∗ ).
Analogously to the continuous-time systems case, we can state the fol-
lowing global stability theorem.
Theorem 3.37 Let the origin x = 0 ∈ Rn be an equilibrium point for the
system (3.52). Let V : Rn → R be a continuous function such that
V (0) = 0 and V (x) > 0, ∀x 6= 0
∆V (x(k)) , V (x(k)) − V (x(k − 1)) < 0, ∀x(k) ∈ D (3.53)
kxk → ∞ ⇒ V (x) → ∞
then the origin is globally asymptotically stable.
Let us interpret the result in the previous theorem. For continuous time
systems, we require that the derivative of the Lyapunov function is negative
along the trajectories. For discrete-time systems, we require that the differ-
ence in the Lyapunov function is negative along the trajectories. Also, quite
importantly we do not require that V is continuously differentiable, but to
be only continuous.
If we now consider a linear discrete-time system given by
x(k + 1) = F x(k) (3.54)
where F ∈ Rn×n . The asymptotic stability of such a system is characterized
by the eigenvalues being strictly inside the unit circle in the complex plane.
Definition 3.38 The matrix F is called Schur or asymptotically stable, if
and only if
|λi | < 1, ∀i = 1, · · · , n
where λi ’s are the eigenvalues of the matrix F .

Theorem 3.39 Consider the linear discrete-time system (3.54), the follow-
ing conditions are equivalent:
1. The matrix F is Schur stable
2. Given any matrix Q = QT > 0 there exists a positive definite matrix
P = P T satisfying the discrete-time matrix Lyapunov equation
F T P F − P = −Q (3.55)

P. Al Hokayem & E. Gallestey 71 of 214

Nonlinear Systems and Control — Spring 2018

Proof: Let’s first show that 1. ⇒ 2. Let F be P∞SchurT stable and take any
T i i
matrix Q = Q > 0. Take the matrix P = i=0 (F ) QF , which is well
defined by the asymptotic stability of F , and P = P T > 0 by definition.
Now, substitute P into (3.55)
∞
! ∞
X X
T T T i i
F PF − P = F (F ) QF F − (F T )i QF i
i=0 i=0
∞
X ∞
X
= (F T )i QF i − (F T )i QF i = −Q
i=1 i=0

In order to show uniqueness, suppose that there is another matrix P̄ that

satisfies the Lyapunov equation. After some reccurssions, we can show that
if both P and P̄ satisfy the Lyapunov equation then

(F T )N (P − P̄ )F N = P − P̄

Letting N → ∞ yields the result.

In order to show that 2. ⇒ 1., consider the Lyapunov function V (x) =
T
x P x, and fix an initial state x(0). We have that (by applying the Lyapunov
equation recursively and summing up the steps)
N
X −1 N
X −1
V (x(N )) − V (x(0)) = − T
x(i) Qx(i) ≤ −λmin (Q) kx(i)k22
i=0 i=0

Therefore, the sequence [V (x(k))]k∈N is strictly decreasing and bounded from

below, hence it attains a non-negative limit. We can further show by contra-
diction that this limit is actually 0, or equivalently limi→∞ kx(i)k = 0, since
this holds for any choice of x(0), it follows that F is Schur stable.

Appendix: Comparison Lemma

Lemma 3.40 Consider the scalar system u̇ = f (t, u), u(t0 ) = u0 , where
f (t, u) is continuous in t and locally Liptschitz in u, ∀t ≥ 0 and ∀u ∈ D ⊂
R. Let [t0 , T ) be the maximal interval of existence of the solution u(t), and

P. Al Hokayem & E. Gallestey 72 of 214

Nonlinear Systems and Control — Spring 2018

suppose that u(t) ∈ D, ∀t ∈ [t0 , T ). Let v(t) be a continuously differentiable

function2 whose derivative satisfies

v̇(t) ≤ f (t, v(t)), v(t0 ) ≤ u0

with v(t) ∈ D, ∀t ∈ [t0 , T ). Then, the solution v(t) satisfies

v(t) ≤ u(t), ∀t ∈ [t0 , T )

2
This condition can be relaxed to simply having v(t) being continuous with an upper
right-hand derivative existing

P. Al Hokayem & E. Gallestey 73 of 214

Nonlinear Systems and Control — Spring 2018

3.10 Exercises
1. Show that the origin of the system
(
ẋ1 = x2
ẋ2 = −x31 − x32

is globally asymptotically stable, using a suitable Lyapunov function.

2. Show that the origin of the system

(
ẋ1 = x2
ẋ2 = x1 − sat(2x1 + x2 )

is locally but not globally asymptotically stable. Recall that the satu-
ration function is defined as
(
s, if |s| ≤ 1
sat(s) = .
sgn(s) , if |s| ≤ 1

3. Use a quadratic Lyapunov function in order to design an appropriate

(feedback) input function u(x) that renders the following system stable
(
ẋ1 = x2
.
ẋ2 = x21 + x1 x2 + u(x)

4. [Krasovskii’s Method] Consider the system

ẋ = f (x), f (0) = 0, x ∈ Rn

Assume that the Jackobian satisfies

   T
∂f ∂f
P (x) + (x) P ≤ −I, ∀x ∈ Rn
∂x ∂x

where P = P T > 0.
R1 ∂f (σx)
• Show using the representation f (x) = 0 ∂x
xdσ that

xT P f (x) + f (x)T P x ≤ −xT x, ∀x ∈ Rn

P. Al Hokayem & E. Gallestey 74 of 214

Nonlinear Systems and Control — Spring 2018

• Show that the function V (x) = f (x)T P f (x) is positive definite

and radially unbounded.
• Show that the origin is globally asymptotically stable.

5. Consider the nonlinear system

ẋ1 = h(t)x2 − g(t)x31

ẋ2 = −h(t)x1 − g(t)x32

where h(t) and g(t) are bounded continously differentiable functions,

with 0 ≤ g0 ≤ g(t).

• Is the equilibrium x = 0 uniformly asymptotically stable (UAS)?

Is it globally UAS?
• Is it (globally/locally) exponentially stable?

6. Consider the nonlinear system

(
ẋ1 = −x1 − e−2t x2
.
ẋ2 = x1 − x2

Determine whether the equilibrium point at 0 is stable or not?

Hint: use the following Lyapunov function V (x, t) = x21 + (1 + e−2t )x22

7. Consider the nonlinear system

ẋ1 = −φ(t)x1 + aφ(t)x2

ẋ2 = bφ(t)x1 − abφ(t)x2 − cψ(t)x32

where a, b, c > 0 are constant and 0 ≤ φ0 ≤ φ(t) and 0 ≤ ψ0 ≤ ψ(t),

∀t ≥ 0 are bounded functions.
Show that the origin is globally uniformly asymptotically stable. Is it
exponentially stable?

8. Consider the nonlinear system

ẋ1 = −x1 + x22

ẋ2 = −x2 + u

P. Al Hokayem & E. Gallestey 75 of 214

Nonlinear Systems and Control — Spring 2018

• Show that the unforced system (setting u ≡ 0) is globally asymp-

totically stable.
• Now, let u 6= 0 and show that the resulting system is input-to-state
stable (ISS).

Hint: For both parts use the candidate function V (x) = (1/2)x21 +
(1/4)ax42 for a proper choice of a > 0.

P. Al Hokayem & E. Gallestey 76 of 214

Chapter 4

Dissipative Systems

4.1 Introduction
We shall study a special class of systems called dissipative systems. Intu-
itively, we can think of systems interacting with the surrounding via some
input / output ports, exchanging power with the surrounding, storing some
energy and dissipating some (Figure 4.1).
u2 y2

u1 u3

y1
Σ y3

u4 y4

Figure 4.1: A system Σ interacting with the outside world via inputs and
outputs

For example, electrical systems exchange power with the surrounding via
an inner product between the current and voltage, i.e., Pe = v T i. Energy
in these systems may be stored in capacitors and/or inductors in the form
of voltage or current, respectively. Another example is mechanical systems;
these systems may by supplied with linear and/or rotational power, i.e.,
Pm = ω T T or Pm = v T F , where v is the velocity, F is the applied force,
ω is the rotational speed, and T is the applied torque. This supply may be

77
Nonlinear Systems and Control — Spring 2018

stored in the form of potential and/or kinetic energy. In what follows, we

shall provide a solid foundation for such supply and storage concepts that
allows us to describe systems from an input-output perspective.
Example 4.1 Consider again the mass-spring system from the lecture on
Lyapunov Stability Theory, as shown in Figure 4.2, with an extra input force
u. Define x1 = x and x2 = ẋ, the position and speed, respectively, let M = 1,
and assume that we measure the output y = ẋ = x2 . Then, we can write the
dynamical equations of the system as
     
ẋ1 x2 0
= + u, y = x2 (4.1)
ẋ2 −F (x1 ) − δx2 1
where F (x1 )x1 > 0, ∀x1 6= 0. How can we make a statement about the

F (x)
u
M
Ff (ẋ) = −δ ẋ

0 x

Figure 4.2: Mass spring system

stability of the system? We consider the energy that is stored in the system
Z x1
1
S (x) = F (ξ) dξ + x22 (4.2)
0 2
The derivative of S (x) along the trajectories of (4.1) is given by
Ṡ (x) = −δx22 + x2 u = −δy 2 + yu (4.3)
Assume that at t = 0, x1 = x2 = 0. Then
Z t Z t


S (t) = Ṡ (τ ) dτ = −δy 2 (τ ) + y(τ )u(τ ) dτ
0
Z0 t Z t
≤ y (τ ) u (τ ) dτ ≤ |y (τ )| |u (τ )| dτ
0 0

Therefore, we can see that if u and y are bounded signals (in some sense),
then S is also bounded. Due to the properties of the function S, we can
then limit the state. This reasoning helps us to go from an input-output
boundedness property to an internal state boundedness property.

P. Al Hokayem & E. Gallestey 78 of 214

Nonlinear Systems and Control — Spring 2018

4.2 Dissipative Systems

Consider the nonlinear state-space system, given by
ẋ = f (x, u)
(4.4)
y = h (x, u)
where x ∈ X ⊆ Rn , u ∈ U ⊆ Rm , and y ∈ Y ⊆ Rp . Associated with (4.4),
we have the following supply rate function
w(u, y) : U × Y → R (4.5)

Definition 4.2 The state-space system (4.4) is said to be dissipative with

respect to the supply rate w(u, y), if there exists a function S : X → R≥0 ,
called the storage function, such that ∀x0 ∈ X , ∀t1 > t0 , and all input
functions u the following dissipation inequality holds
Z t1
S (x (t1 )) ≤ S (x (t0 )) + w (u (t) , y (t)) dt (4.6)
t0

where x (t0 ) = x0 , and x (t1 ) is the state system (4.4) at time t1 resulting
from initial condition x0 and the input function u(t).

The dissipation inequality (4.6) expresses the concept that the stored en-
ergy S (x (t1 )) of the system (4.4) at any time t1 is at most equal to the
sum Rof the stored energy S (x (t0 )) present at the time t0 and the total en-
t
ergy t01 w (u (t) , y (t)) dt which is supplied externally during the time interval
[t0 , t1 ]. Hence, as the name suggests, dissipative systems cannot internally
create energy, but can rather either store it or dissipate it.
We shall study next a special type of storage functions.
Theorem 4.3 Consider the system (4.4) with a supply rate w. Then, it is
dissipative with respect to w if and only if the available storage function
 Z T 
Sa (x) = sup − w (u, y) dt , x(0) = x (4.7)
u(·),T ≥0 0

is well defined, i.e., Sa (x) < ∞, ∀x ∈ X . Moreover, if Sa (x) < ∞, ∀x ∈ X ,

then Sa is itself a storage function, and it provides a lower bound on all other
storage functions, i.e., for any other storage function S
Sa (x) ≤ S (x)

P. Al Hokayem & E. Gallestey 79 of 214

Nonlinear Systems and Control — Spring 2018

Proof: First note that Sa ≥ 0 (why?).

R t1 Suppose that Sa is finite. Compare
now Sa (x (t0 )) with Sa (x (t1 ))− t0 w (u (t) , y (t)) dt, for a given u : [t0 , t1 ] →
Rm and resulting state x (t1 ). Since Sa is given as theRsupremum over all u (.)
t
it immediately follows that Sa (x (t0 )) ≥ Sa (x (t1 ))− t01 w (u (t) , y (t)) dt and
thus Sa is a storage function, proving that the system (4.4) is dissipative with
respect to the supply rate w.
In order to show the converse, assume that (4.4) is dissipative with respect
to w. Then there exists a storage function S ≥ 0 such that for all u (.)
Z T
S (x (0)) + w (u (t) , y (t)) dt ≥ S (x (T )) ≥ 0
0

which shows that

 Z T 
S (x (0)) ≥ sup − w (u (t) , y (t)) dt = Sa (x (0))
u(·),T ≥0 0

proving finiteness of Sa , as well as Sa (x) ≤ S (x).



Remark 4.4 Note that in linking dissipativity with the existence of the func-
tion Sa , we have removed attention from the satisfaction of the dissipation
inequality to the existence of the solution to an optimization problem.

Remark 4.5 The quantity Sa can be interpreted as the maximal energy

which can be extracted from the system (4.4) starting at an initial condition
x0 .

Consider the dissipation inequality in the limit where t1 → t0 . Then it

may be seen that satisfaction of the dissipation inequality is equivalent to
fulfilling the partial differential equation (assuming S is differentiable)

∂S(x)
Ṡ (x) = f (x, u) ≤ w (u, h (x, u)) , ∀x ∈ Rn , u ∈ Rm (4.8)
∂x
This version of the dissipation property is called the differential dissipation
inequality. Having this differential version, we can establish a connection to
what we have already seen in the lecture on Lyapunov stability theory. But
before we do that, let us first see what the rate at which the system (4.4)
dissipates energy is.

P. Al Hokayem & E. Gallestey 80 of 214

Nonlinear Systems and Control — Spring 2018

Definition 4.6 The function d : X × U → R is the dissipation rate of

the dissipative system (4.4) with supply rate w and storage function S, if
∀t0 , t1 ∈ R+ , x0 ∈ X , and u ∈ U, the following equality holds
Z t1
S(x(t0 )) + (w(t) + d(t)) dt = S(x(t1 )
t0

Of course, we would require that d is non-negative in order to obtain dissi-

pation!

Lemma 4.7 Let S be a continuously differentiable storage function for the

system (4.4) and assume that the supply rate w satisfies

w (0, y) ≤ 0, ∀y ∈ Y

Let the origin x = 0 be a strict local minimum of S (x). Then x = 0 is a

locally stable equilibrium for the unforced system ẋ = f (x, 0) and V (x) =
S (x) − S (0) ≥ 0 is a local Lyapunov function.

Proof: Consider the Lyapunov function candidate V (x) = S(x) − S(0),

which is positive definite (why?). Under u = 0, we have that

∂S(x)
V̇ (x) = Ṡ(x) = f (x, 0) ≤ w (0, y) ≤ 0
∂x


We can also show that the feedback interconnection of dissipative systems is

stable.
u1 y1
Σ1

y2 u2
Σ2

Figure 4.3: Feedback interconnection of dissipative systems

P. Al Hokayem & E. Gallestey 81 of 214

Nonlinear Systems and Control — Spring 2018

Lemma 4.8 Consider the two systems

(
ẋi = fi (xi , ui )
(Σi ) : (4.9)
yi = h(xi , ui )

connected in feedback as shown in Figure 4.3. Assume that both systems

are dissipative with respect to supply rates wi and positive definite storage
functions Si . Assume further that

w1 (u, y) + w2 (y, −u) = 0, ∀u, y

Then, the feedback system is stable.

Proof: Consider the Lyapunov function candidate V (x) = S1 (x1 ) + S2 (x2 ).

V̇ (x1 , x2 ) = Ṡ(x1 ) + Ṡ(x2 ) ≤ w1 (u1 , y1 ) + w2 (u2 , y2 )

= w1 (−y2 , y1 ) + w2 (y1 , y2 ) = 0

and the result follows.


Lemma 4.8 is an extremely powerful one and captures many of the stability
results in the frequency domain.

Example 4.9 Consider the RLC circuit shown in Figure 4.4. Define the
states x1 = vc and x2 = i, the input u = vin and the output y = x2 . The
state-space model is given by
   
0 1
0  
ẋ = Ax + Bu = C y = Hx = 0 1 x
1 R x + 1 u,
−L −L L

For simplicity, we shall take L = C = R = 1. The energy storage in the

system is captured by the inductor and the capacitor, i.e., the storage function
in the system is given by
1 1
S(x) = x21 + x22
2 2
which is positive definite, and the supply rate to the system is given by

w(u, y) = uy (4.10)

P. Al Hokayem & E. Gallestey 82 of 214

Nonlinear Systems and Control — Spring 2018

i R L

vin C vc

Figure 4.4: RLC electric circuit

which is the power injected (extracted) into (from) the system. Now,

Ṡ(x) = x1 ẋ1 + x2 ẋ2 = −x22 + uy ≤ w(u, y) (4.11)

Hence, the system is dissipative. Now, if we let u = 0, i.e., we short-circuit

the terminals, we obtain
Ṡ(x) = −x22 ≤ 0
and the energy that was initially stored in the capacitor and/or inductor is
cycled in the system and dissipated in the resistor. We can show in this
case, that using the so called strong Lyapunov function with u = 0 we get
asymptotic stability. Consider the Lyapunov function candidate
1 1
V (x) = x21 + x22 + x1 x2 (4.12)
2 2
where  ∈ (0, 1) (show that this choice of  makes V (x) positive definite!).
Taking the derivative of V along the trajectories of the closed-loop system,
we obtain

V̇ (x) = x1 ẋ1 + x2 ẋ2 + x1 ẋ2 + x2 ẋ1

= x1 x2 + x2 (−x1 − x2 ) + x1 (−x1 − x2 ) − x22
= −x21 − (1 + )x22 − x1 x2
2 1
≤ −x21 − (1 + )x22 + x21 + x22
  2 2
1 1
= − 1 −  x21 − +  x22 < 0
2 2

for our choice of . This way we have avoided using LaSalle’s invariance
principle, which we could have used instead to show asymptotic stability of
the system.

P. Al Hokayem & E. Gallestey 83 of 214

Nonlinear Systems and Control — Spring 2018

4.3 Passive Systems

Passive systems are a special subclass of dissipative systems, and they have
a special type of supply rate, given by

w(u, y) = uT y (4.13)

with the implicit condition that the number of inputs and outputs is the
same, i.e., u, y ∈ Rp . We can also differentiate among various types of passive
dynamical systems according to the following definition:

Definition 4.10 A state space system (4.4) is called

1. passive if it is dissipative with respect to the supply rate w (u, y) = uT y

2. lossless if Ṡ(x) = uT y

3. input-feedforward passive if it is dissipative with respect to the supply

rate w(u, y) = uT y − uT ϕ(u) for some function ϕ

4. input strictly passive if it is dissipative with respect to the supply rate

w(u, y) = uT y − uT ϕ(u) and uT ϕ(u) > 0, ∀u 6= 0

5. output feedback passive if it is dissipative with respect to the supply rate

w(u, y) = uT y − y T ρ(y) for some function ρ

6. output strictly passive if it is dissipative with respect to the supply rate

w(u, y) = uT y − y T ρ(y) for some function y T ρ(y) > 0, ∀y 6= 0

7. strictly passive if it is dissipative with respect to the supply rate w(u, y) =

uT y − ψ(x) for some positive definite function ψ

Example 4.11 Consider an integrator model given by

ẋ = u, y=x (4.14)

with the supply rate w(u, y) = uy. Take the storage function S(x) = 12 x2 .
The derivative Ṡ(x) = xẋ = uy, and hence the system is lossless.

P. Al Hokayem & E. Gallestey 84 of 214

Nonlinear Systems and Control — Spring 2018

Bode Diagram

40

30
1/(s+1)
20
1/s
Magnitude (dB)

10

−10

−20

−30

−40
0
Phase (deg)

−45

−90 −2 −1 0 1 2
10 10 10 10 10
Frequency (rad/s)

Figure 4.5: Bode plot of the two systems in Examples 4.11 and 4.12

Example 4.12 Now assume that instead of the pure integrator in Example
4.11 with transfer function 1s , we consider the low pass filter s+1
1
, which has
the state-space representation

ẋ = −x + u, y=x (4.15)

Consider the storage function S(x) = 21 x2 . The derivative Ṡ(x) = xẋ =

−x2 + uy. Hence the system is strictly dissipative, and globally asymptoti-
cally stable for u = 0 (convince yourself of the latter fact). This is one of
the main reasons that we would not implement ‘pure’ integrators in embed-
ded systems, but instead opt for low pass filters due to their inherent stable
behavior. Finally, notice from Figure 4.5 that both systems behave similarly
for high frequencies.

4.3.1 Characterizations of Passivity for Linear Systems

Passive systems are particularly interesting in the linear systems case, be-
cause we can get characterizations of passivity both in the frequency domain
and in the time domain. In the frequency domain, we think of transfer func-

P. Al Hokayem & E. Gallestey 85 of 214

Nonlinear Systems and Control — Spring 2018

tions and we can relate passivity to certain conditions being statisfied for the
transfer function.
Definition 4.13 A p × p proper transfer function matrix G(s) is called pos-
itive real if all the following conditions are satisfied:
1. the poles of all elements of G(s) have non-positive real part
2. for all real frequencies ω for which jω is not a pole of any element of
G(s), the matrix G(jω) + G(−jω)T is positive semi-definite
3. any pure imaginary pole jω of any element of G(s) is a simple pole
and the residue matrix lims→jw (s − jw)G(s) is positive semidefinite
Hermitian
G(s) is called strictly positive real if G(s − ) is positive real for some  > 0.

Remark 4.14 For p = 1, the second condition of Definition 4.13 reduces

to Re[G(jw)] ≥ 0, ∀w ∈ R. Moreover, this condition is satisfied only if the
relative degree of the transfer function G(s) is at most one.

The positive real property of transfer matrices can be equivalently charac-

terized as follows.
Lemma 4.15 Let G(s) be a proper rational p × p transfer function matrix.
Suppose that det(G(s) + G(−s)T ) is not equivalent to zero for all s. Then
G(s) is strictly positive real if and only if the following three conditions are
satisfied
1. G(s) is Hurwitz,
2. G(jω) + G(−jω)T is positive definite ∀ω ∈ R
3. either G(∞) + G(∞)T is positive definite or it is positive semidefinite
and limω→∞ ω 2 M T (G(jω)+G(−jω)T )M is positive definite for any full
rank p × (p − q) matrix M such that M T (G(∞) + G(∞)T )M = 0, where
q = rank(G(∞) + G(∞)T ).

Example 4.16 Recall the RLC circuit in Example 4.9. The transfer func-
tion (for R = L = C = 1) is given by
   
−1
  s −1 −1 0 s
G(s) = H(sI − A) B = 0 1 = 2 (4.16)
1 s+1 1 s +s+1

P. Al Hokayem & E. Gallestey 86 of 214

Nonlinear Systems and Control — Spring 2018

Let us check if it is positive real. Note: We can show that condition 2. of

Lemma 4.15 does not hold, and hence the transfer function is not strictly
positive real. √
The poles of G(s) are given by si = −1±j2
3
, i = 1, 2, thus G(s) is Hurwitz.
We also have that
w2
Re[G(jw)] = ≥ 0, ∀w ∈ R
(1 − w2 )2 + w2
Finally, we have no pure imaginary poles. Therefore, we can conclude that
the transfer function (4.16) is positive real.

One can also look at the state-space system directly and conclude that
the transfer function is actually positive real, as shown by the celebrated
KYP lemma below.
Lemma 4.17 (Kalman-Ykubovich-Popov) Consider the m × m trans-
fer function matrix G(s) = C(sI − A)−1 B + D, where the pair (A, B) is
controllable and the pair (A, C) is observable. G(s) is strictly positive real if
and only if there exist matrices P = P T > 0, L, and W , and  > 0 such that
the following equalities hold
P A + AT P = −LT L − P (4.17)
T T
PB = C − L W (4.18)
W T W = D + DT (4.19)

Remark 4.18 If  = 0 in Lemma 4.15, then the transfer function G(s) is

simply positive real.

Example 4.19 Consider again the RLC system in Example 4.9. Since there
is no direct feedthrough in the system,  D = 0 and by (4.19), W = 0.
  we have
0 0 p 0
As such, (4.18) implies P = ⇒P = , p > 0. Plugging P into
1 1 0 1
(4.17)
   
0 1−p 0
T
PA + A P = − ≤0⇔p=1⇒L = √ T
1−p 2 2
Therefore, again we reach the same conclusion that the system is passive.
However, we cannot satisfy the condition (4.17) with  > 0 and hence the
system is not strictly passive.

P. Al Hokayem & E. Gallestey 87 of 214

Nonlinear Systems and Control — Spring 2018

Finally, we are ready to state the connection/equivalence between positive

realness and passivity for Linear Time-Invariant (LTI).
Theorem 4.20 The LTI system

ẋ = Ax + Bu
y = Cx + Du

with the corresponding transfer matrix G(s) = C(sI − A)−1 B + D is

• passive if G(s) is positive real

• strictly passive, if G(s) is strictly positive real.

Proof: Consider the storage function S(x) = 21 xT P x. Then,
1
Ṡ(x) = xT P (Ax + Bu) = xT (P A + AT P )x + xT P Bu
2
1 T
= x (P A + AT P )x + xT C T u − xT LT W u
2
1 1
= xT (P A + AT P )x + y T u − uT (DT + D)u − xT LT W u
2 2
1 T 1
= x (P A + AT P )x + y T u − uT W T W u − xT LT W u
2 2
1 T T 1 T T
− x L Lx + x L Lx
2 2
1 T 1 1
= x (P A + AT P )x + y T u − (W u + Lx)T (W u + Lx) + xT LT Lx
2 2 2
1 T 1
≤ x (P A + AT P )x + y T u + xT LT Lx
2
 2

 − xT P x + y T u,  > 0, strictly passive
= 2
y T u,  = 0, passive

4.3.2 Stability of Passive Systems

Let us consider again the nonlinear system (4.4), where f : Rn × Rp → Rn
is locally Lipschitz, h : Rn × Rp → Rp is continuous, with f (0, 0) = 0 and
h(0, 0) = 0.

P. Al Hokayem & E. Gallestey 88 of 214

Nonlinear Systems and Control — Spring 2018

Lemma 4.21 Assume that the system (4.4) is passive with a positive definite
storage function S(x), then the origin of ẋ = f (x, 0) is stable.

Proof: Take V (x) = S(x) as Lyapunov function candidate. Then V̇ (x) ≤

uT y = 0, and stability follows.

We can further strengthen the previous Lemma, as follows.

Lemma 4.22 Assume that the system (4.4) is strictly passive with some
storage function S(x), then the origin of ẋ = f (x, 0) is asymptotically sta-
ble. Furthermore, if S(x) is radially unbounded, then the origin is globally
asymptotically stable.

Proof: Let V (x) = S(x) be a Lyapunov function candidate. Since the

system is strictly passive with a storage function V (x), it follows that (for
u = 0)
V̇ (x) ≤ −ψ(x) + uT y = −ψ(x)
Now consider any x ∈ Rn and let φ(t, x) be the solution to the differential
equation ẋ = f (x, 0), starting at x and time t = 0. As such, we have that
Z t
V (φ(t, x)) − V (x) ≤ − ψ(φ(τ, x))dτ ∀t ∈ [0, δ] (4.20)
0

for some positive constant δ. Since V (φ(t, x)) ≥ 0, then

Z t
V (x) ≥ ψ(φ(τ, x))dτ
0

Suppose now that there exists some x̄ 6= 0 such that V (x̄) = 0. This implies
that Z t
ψ(φ(τ, x̄))dτ = 0, ∀t ∈ [0, δ] ⇒ ψ(φ(τ, x̄)) ≡ 0 ⇒ x̄ = 0
0

which gives a contradiction. Hence, V (x) > 0 for all x 6= 0, i.e., positive
definite. Combining this with the V̇ (x) ≤ −ψ(x), yields asymptotic stability
of the origin. Finally, if V (x) is radially unbounded, we obtain asymptotic
stability.

P. Al Hokayem & E. Gallestey 89 of 214

Nonlinear Systems and Control — Spring 2018

u1 e1 y1
Σ1

y2 e2 u2
Σ2

Figure 4.6: Feedback interconnection of passive systems

We shall look at some of the stability properties of passive systems, when

connected in a feedback structure as shown in Figure 4.6.

Theorem 4.23 The feedback connection of two passive systems is passive.

Proof: Let Si (xi ) be the storage function of system Σi , i = 1, 2. Since both

systems are passive,we have that

V̇i (xi ) ≤ eTi yi

Using the feedback structure in Figure 4.6, we have that

Ṡ(x) = Ṡ1 (x1 ) + Ṡ2 (x2 ) ≤ eT1 y1 + eT2 y2 (4.21)

T T
= (u1 − y2 ) y1 + (u2 + y1 ) y2 = uT1 y1 + uT2 y2 T
=u y (4.22)

and the result follows. Note that if any of the two systems are memoryless,
i.e., yi = hi (ui ), then the corresponding storage function can be taken to be
0.

4.3.3 Passivity-Based Control

Having done some analysis on passive systems, we can now show a glimpse
of how this theory can be used to design the so-called passivity-based con-
trollers. Consider the dynamical system

ẋ = f (x, u)
(4.23)
y = h(x)

with the usual Lipschitz assumption on f and continuity of h. Moreover,

assume that f (0, 0) = 0 and h(0) = 0.

P. Al Hokayem & E. Gallestey 90 of 214

Nonlinear Systems and Control — Spring 2018

Definition 4.24 The system (4.23) is called zero-state observable, if no so-

lution of the unforced system ẋ = f (x, 0) can stay identically in the set
{h(x) = 0} other than the trivial solution x(t) ≡ 0.

Theorem 4.25 Assume that the system (4.23) is

1. passive with a radially unbounded positive definite storage function, and

2. is zero-state observable,
then the origin can be globally stabilized with a control law u = −φ(y), where
φ is any locally Lipschitz function such that φ(0) = 0, and y T φ(y) > 0,
∀y 6= 0.

Proof: Let V (x) be a storage function of the system and use it as a candidate
Lyapunov function for the closed-loop system with u = −φ(y). We have that

V̇ (x) ≤ uT y = −φ(y)T y ≤ 0

Hence, V̇ (x) ≤ 0, and V̇ (x) = 0 if and only if y = 0. By zero-state observ-

ability, the only solution that can stay in the set {y = h(x) = 0} is the trivial
solution x(t) ≡ 0, and we can conclude using LaSalle’s invariance principle
that the origin is globally asymptotically stable.

This last theorem is very useful when designing control laws for a large num-
ber of electrical and mechanical systems. Moreover, instead of starting with
systems for which the origin is open-loop stable, we can design control laws
that convert a nonpassive system into a passive one, a technique known as
feedback passivation.
Example 4.26 Consider the system
ẋ1 = x2
ẋ2 = x21 + u
y = x2

The open-loop system (u = 0) is unstable, and hence the system is not passive.
However, we can design the control law

u = −x22 − x31 + v

P. Al Hokayem & E. Gallestey 91 of 214

Nonlinear Systems and Control — Spring 2018

that yields the system passive with respect to the supply rate w = vy. Let
S(x) = 41 x41 + 12 x22 , then
Ṡ(x) = −x22 + vy
and the system is passive. Noting that v = 0 and y(t) ≡ 0 imply x(t) ≡ 0.
Therefore, all the conditions of Theorem 4.25 are satisfied, and accordingly
we can design a globally stabilizing control law, for example v = −kx2 or
v = −tan−1 (x2 ) and any k > 0.

For further reading on the subject, please see [4, 6, 7].

P. Al Hokayem & E. Gallestey 92 of 214

Nonlinear Systems and Control — Spring 2018

4.4 Exercises
1. The motion of a rigid n-link robot manipulator is described by the
following equations:

∂P (q) >
M (q)q̈ + C(q, q̇)q̇ + =u
∂q
where q ∈ Rn and q̇ ∈ Rn are the so-called generalized coordinates
of the system, u ∈ Rn represents the input (torques or forces), and
P (q) is the potential energy that is positive definite. The system above
satisfies the following skew-symmetry properties:

• The matrix L = Ṁ (q)−2C(q, q̇) is skew-symmetric, i.e. L> = −L;

• M (q)> = M (q) ≥ λI, ∀q ∈ Rn with λ > 0.

(a) Show that the system is passive from input u to output q̇, using
the storage function S(q, q̇) = 12 q̇ > M (q)q̇. Only for this point,
consider P (q) = 0.
(b) Now let us use the following control law u = −B q̇, where B =
>
B > > 0 is a damping matrix. Under the assumption that ∂P∂q(q)
has an isolated root at the origin, show that this control law is
asymptotically stabilizing using V (q, q̇) = S(q, q̇) + P (q) as a Lya-
punov fucntion.
>
(c) Now, consider the following control law u = ∂P∂q(q) − B q̇ − K(q −
q d ), where B = B T > 0, K = K T > 0, and q d is some de-
sired set point. Show that the equilibrium point (q, q̇) = (q d , 0)
is asymptotically stable using the Lyapunov function candidate
V = 21 q̇M q̇ + 12 (q − q d )> K(q − q d ) (Note that q̇d = 0).

2. Consider the system

ẋ1 = x2
ẋ2 = −h(x1 ) − ax2 + u
y = αx1 + x2

where 0 < α < a, yh(y) > 0 for all y 6= 0.

P. Al Hokayem & E. Gallestey 93 of 214

Nonlinear Systems and Control — Spring 2018

(a) Show that the system is passive by choosing an opportune storage

candidate function. Hint: In your candidate function, you need
to integrate the nonlinear function h(·).
(b) How would you render the system globally asymptotically stable
by using output feedback?

3. Euler equations for a rotating rigid spacecraft are given by

J1 ω̇1 = (J2 − J3 )ω2 ω3 + u1

J2 ω̇2 = (J3 − J1 )ω3 ω1 + u2
J3 ω̇3 = (J1 − J2 )ω1 ω2 + u3

where ω1 to ω3 are the components of the angular velocity vector along

the principal axes, u1 to u3 are the torque inputs applied about the
principal axes, and J1 to J3 are the principal moment of inertia.

(a) Show that the map from u = [u1 , u2 , u3 ]> to ω = [ω1 , ω2 , ω3 ]> is
lossless.
(b) Show that, for the control law u = −Kω, where K is a positive
definite matrix, the origin ω = 0 is globally asymptotically stable.

P. Al Hokayem & E. Gallestey 94 of 214

 Part II

Control Design Techniques

95
Chapter 5

Feedback Control Design

5.1 Introduction
Control design procedures require the use of feedback in order to steer the
state of the system towards some desired steady-state value, or a given output
towards some reference.

r e u y
CONTROL PLANT

Figure 5.1: Feedback Control

When dealing with nonlinear systems, one has a variety of design tech-
niques that can be used. We summarize in the table below some of the
dimensions that we shall consider in the design phase of the course.
In this lecture, we shall look into three control design techniques: Lin-
earization, integral control, and gain scheduling, all of which are based on
our knowledge of control design for linear systems.

5.2 Linearization
Based on the ideas we have already explored in Lyapunov’s indirect method,
we can consider designing a controller by linearizing the system about the de-

96
Nonlinear Systems and Control — Spring 2018

Requirements / Goals Tools Difficulties

Stabilization Linearization Nonlinearities
Tracking Integral Control Lack of measurements
Disturbance rejection Gain Scheduling Noise
Disturbance attenuation Robust control Uncertainties ...
Transient response ... Adaptive control
Feedback linearization
Sliding mode control ...

sired equilibrium point, and then designing a stabilizing linear state feedback
by pole placement, or other similar techniques.
Consider the nonlinear system
ẋ = f (x, u) (5.1)
where f (x, u) is continuously differentiable in a domain X × U ⊆ Rn × Rm
that contains the origin. Let (xss , uss ) ∈ X × U be the steady state point for
which
0 = f (xss , uss ) (5.2)
We can linearize the system (5.1) around the steady-state operation point
(xss , uss ) using the Taylor series expansion to obtain,
d
(x − xss ) = f (x, u)
dt  
∂f  ∂f 
≈ f (xss , uss ) + (x − xss ) + (u − uss )
∂x x=xss ,u=uss ∂u x=xss ,u=uss
= A(x − xss ) + B(u − uss ) (5.3)
 
where A = ∂f  and B = ∂f 
. Assume that (A, B) is
∂x x=xss ,u=uss ∂u x=xss ,u=uss
stabilizable, that is the eigenvalues that aren’t controllable have all negative
real parts. Then ∃K ∈ Rm×n such that (A − BK) has all eigenvalues in the
left hand side of the complex plane. If (A,B) is controllable, we can place all
the all the poles at the desired locations in the left-half complex plane. The
resulting feedback control law for the nonlinear system (5.1) is given by
u = uss − K (x − xss ) (5.4)
and the corresponding closed-loop system is given by
ẋ = f (x, uss − K (x − xss ))

P. Al Hokayem & E. Gallestey 97 of 214

Nonlinear Systems and Control — Spring 2018

which is asymptotically stable in some small neighborhood around (xss , uss ).

Example 5.1 Consider the pendulum shown in Figure 5.2 equation
θ̈ = −a sin θ − bθ̇ + cT (5.5)
where a = g/l, b = k/m ≥ 0, c = 1/ml2 > 0, and T is the torque applied
to the pendulum. We would like to stabilize the pendulum at some angle θss .

θ
m

Figure 5.2: Pendulum

For the pendulum to have an equilibrium at (θss , 0), the input torque must
have a steady state component Tss that satisfies
a
Tss = sin(θss ) (5.6)
c
Define the state variables as x1 = θ − θss , x2 = θ̇ and the control variable as
u = T − Tss . Accordingly, the state equations are given by
ẋ1 = x2
(5.7)
ẋ2 = −a[sin(x1 + θss ) − sin(θss )] − bx2 + cu
We linearize this system
  at the
 origin
 to obtain the system matrices A =
0 1 0
and B = . It is not difficult to show that the pair
−a cos(θss ) −b c  
(A, B) is controllable (show this). Take the feedback matrix K = k1 k2 ,
then (A − BK) is Hurwitz for any choice of the gains that satisfies 1
a cos(θss ) b
k1 > − , k2 > −
c c
1
you can drive this by looking at det(λI − (A − BK))) and enforcing that the sign of
all the coefficients of the characteristic polynomial is the same

P. Al Hokayem & E. Gallestey 98 of 214

Nonlinear Systems and Control — Spring 2018

Finally, the input torque that we would apply to the pendulum is given by
a
T = Tss − Kx = sin(θss ) − k1 (θ − θss ) − k2 θ̇
c
Assume now that we do not have access to the full state, but rather a
nonlinear function of the state, i.e., we have the nonlinear system

ẋ = f (x, u)
(5.8)
y = h(x)

We shall linearize the system (5.8) around the steady-state point (xss , uss ),
for which the following conditions are satisfied

f (xss , uss ) = 0, h(xss ) = 0 (5.9)

to obtain
d
(x − xss ) = A(x − xss ) + B(u − uss ) (5.10)
dt
y = C(x − xss ) (5.11)
  
where A = ∂f  , B = ∂f  , and C = ∂h  . If we
∂x x=xss ,u=uss ∂u x=xss ,u=uss ∂x x=xss
assume that (A, B) is stabilizable and (A, C) is detectable, then we can
design an observer-based linear dynamic output feedback controller
d
(x̂ − xss ) = A(x̂ − xss ) + B(u − uss ) + L(y − C(x̂ − xss ))
dt (5.12)
u − uss = −K(x̂ − xss )

such that the closed-loop system

      
d x − xss x − xss A − BK BK x − xss
= Acl = (5.13)
dt x − x̂ x − x̂ 0 A − LC x − x̂

is asymptotically (exponentially) stable, i.e., the matrix Acl is Hurwitz. Fi-

nally, the corresponding closed-loop nonlinear system is given by

ẋ = f (x, uss − K(x̂ − xss ))

x̂˙ = (A − BK − LC)(x̂ − xss ) + Lh(x)

P. Al Hokayem & E. Gallestey 99 of 214

Nonlinear Systems and Control — Spring 2018

5.3 Integral Control

The approach before is useful if there are no modeling errors. Integral control
is a method to handle some common uncertainties. We will present an inte-
gral control approach that ensures asymptotic regulation under all parameter
perturbations that do not destroy the stability of the closed-loop system.
Consider the nonlinear system

ẋ = f (x, u, w)
(5.14)
y = h (x, w)

where x ∈ Rn is the state, u ∈ Rp is the control input, y ∈ Rp is the output,

and w ∈ Rl is a vector of unknown constant parameters and disturbances.
We assume that f and h are continuously differentiable in x and u, and
continuous in w within some domain X × U × W ⊂ Rn × Rp × Rl . We would
like to design a control law that guarantees that y (t) → r as t → ∞, for some
constant reference r. We shall assume that there is a unique pair (xss , uss )
which depends on both r and w, such that

0 = f (xss (w, r), uss (w, r), w)

r = h (xss (w, r), w)

where xss is the desired equilibrium point and uss is the corresponding steady-
state control. As the name of the method suggests, we would like to introduce
an integral action with repsect to the error signal e = y − r. This goal is
achieved by introducing the extra state (multi-dimensional)

σ̇ = y − r = h(x, w) − r (5.15)

Accordingly, the augmented system is given by

ẋ = f (x, u, w)
(5.16)
σ̇ = h(x, w) − r

In general, it is quite difficult to find a globally stabilizing control law for

the system (5.16). As such, we shall proceed by seeking a locally stabilizing
control via the linearization technique we saw in the previous section.
We propose the linear control structure

u = −K1 x − K2 σ (5.17)

P. Al Hokayem & E. Gallestey 100 of 214

Nonlinear Systems and Control — Spring 2018

e σ u y
r −R CONTROL PLANT

measured signals

Figure 5.3: Integral Control

and require that the matrix gain K2 be nonsingular in order guarantee the
existence of a unique equilibrium (xss , uss , σss ) to the steady-state equations
0 = f (xss , uss , w)
0 = h(xss , w) − r
uss = −K1 xss − K2 σss
with the understanding that the equilibrium point depends on both the
known reference r and the unknown but constant parameters w.
We shall proceed by using the linearization design technique that we saw
in the previous section to obtain the system
ẋ = A(x − xss ) + B(u − uss )
(5.18)
σ̇ = C(x − xss )
 
 ∂f (x,u,w) 
where A = A(w, r) = ∂f (x,u,w)
∂x  , B = B(w, r) = ∂u  ,
 x=xss ,u=uss x=xss ,u=uss

and C = C(w, r) = ∂h(x,w)
∂x 
are dependent on the constant parameters w
x=xss
and the constant reference vector r. Now if we substitute the value steady-
state input uss into the linearized system (5.18), we obtain the following
augmented closed-loop system
 
˙ξ = (A − BK)ξ = A − BK1 −BK2 ξ (5.19)
C 0
 
x − xss
where ξ = . We have the following Lemma that guarantees the
σ − σss  
possibility of designing a stabilizing control matrix K = K1 K2 .
Lemma 5.2 Assume that the pair (A, B) is controllable and that
 
A B
rank =n+p (5.20)
C 0
Then, the pair (A, B) is controllable.

P. Al Hokayem & E. Gallestey 101 of 214

Nonlinear Systems and Control — Spring 2018

Proof: See the exercises.

Assuming that the conditions of Lemma 5.2 are satisfied, we can design
the state feedback integral controller as

u = −K1 x − K2 σ (5.21)
σ̇ = h(x, w) − r (5.22)
 
where K = K1 K2 is designed so that it renders (A − BK) Hurwitz for all
possible values of the uncertain vector w in some known set W (i.e., we have
robust stability).

Example 5.3 Consider again the Pendulum system in Example 5.1. This
time we don’t know the parameters in the system, but we have that a = g/l >
0, b = k/m ≥ 0, and c = 1/(ml2 ) > 0. We have the nonlinear system (using
the same definition for the state and input as before)

ẋ1 = x2
ẋ2 = −a sin(x1 + θss ) − bx2 + cu (5.23)
y = x1
with the equilibrium point
 
0 a
xss = , uss = sin(θss ) (5.24)
0 c
The linearization matrices are computed as
   
0 1 0  
A= , B= , C= 1 0 (5.25)
−a cos(θss ) −b c

Since the parameter c > 0 we can deduce that the pair (A, B) is controllable
(verify this). Moreover, the rank condition (5.20) is satisfied (verify this as
well). Hence the conditions of Lemma 5.2  are
 satisfied and we can design as
  1x
stabilizing controller u = − k1 k2 k3 x2 , where σ̇ = x1 . The resulting
σ
closed-loop matrix is given by
 
0 1 0
(A − BK) = −(a cos(θss ) + ck1 ) −b − ck1 −ck3  (5.26)
1 0 0

P. Al Hokayem & E. Gallestey 102 of 214

Nonlinear Systems and Control — Spring 2018

which is Hurwitz if the control gains satisfy

b + ck2 > 0, ck3 > 0, (b + ck2 )(a cos(θss ) + ck1 ) > ck3

These three inequalities have to be satisfied for all possible realization of the

a, b, and c.
system parameters
For example, a choice of k2 > 0, k3 > 0,
and k1 > max ac + kk23 max 1c would render the closed-loop matrix Hurwitz
(assuming that the maximum values are attained).

5.4 Gain Scheduling

So far we have looked at design methods that guarantee stability in some
neighborhood of the equilibrium point. However, this is very limiting as we
would like to stabilize the system over a large region of the state space about
several equilibrium points. This can be achieved by linearizing the system
about each equilibrium point and design a local controller for stability. Then,
we can online schedule the gains to go from one equilibrium point to another.
This design methodology which scales the previous design methods can be
described by the following steps:

1. Linearize the given nonlinear system about several equilibrium points,

which are parametrized by scheduling variables

2. Design a parametrized family of linear controllers to locally stabilize

the system around each of the equilibrium points

3. Construct a gain-scheduled controller

4. Check the performance of the gain-scheduled controller by simulating

the nonlinear closed-loop model

Remark 5.4 The region of attraction of one equilibrium point should con-
tain the neighboring equilibrium point, as shown in Figure 5.4. Moreover, we
should guarantee that the overshoot that happens whenever switching from
one steady-state point (xss,i ) to another (xss,i+1 ) does not drive the trajectory
of the system beyond the region of attraction of the latter steady state point
(xss,i+1 ).

P. Al Hokayem & E. Gallestey 103 of 214

Nonlinear Systems and Control — Spring 2018

xss,1
xss,2

xss,3 xss,4

Figure 5.4: Gain scheduling example

Consider the nonlinear system

ẋ=f (x, u, v, w) (5.27)
y = h (x, w) (5.28)
where x ∈ Rn is the state, u ∈ Rp is the input, v ∈ Rq is a known exogenous
input, w ∈ Rl is a vector of constant but unknown parameters. We would
like to design a state feedback controller that achieves the regulation of the
output y to a given reference vector r ∈ Rp . Let
 
v
ρ= (5.29)
r
be the vector of known exogenous quantities in the system. We shall use
αv
ρ as a scheduling variable. For each fixed value of ρ = α = , we use
αr
integral control to regulate the error e = y − r to zero in the presence of the
unknown vector w. Moreover, we rely on gain scheduling to achieve small
error whenever ρ is varying slowly over time.
In order to design the integral control, we assume (as we have seen before)
that there exists a unique pair (xss , uss ) that is continuously differentiable in
α and continuous in w, such that
0 = f (xss (α, w) , uss (α, w) , αv , w) (5.30)
αr = h (xss (α, w) , w) (5.31)

P. Al Hokayem & E. Gallestey 104 of 214

Nonlinear Systems and Control — Spring 2018

for all values of (α, w). Now, as in the previous section, we can design the
following integral controller

σ̇ = y − r (5.32)
u = −K1 x − K2 σ (5.33)

where K1 = K1 (α) and K2 = K2 (α), such that the matrix

 
A − BK1 −BK2
A= (5.34)
C 0

is Hurwitz for every (α, w), where


∂f (x, u, v, w) 
A = A(α, w) =  (5.35)
∂x
(x,u,v)=(xss ,uss ,αv )
∂f (x, u, v, w) 
B = B(α, w) =  (5.36)
∂u (x,u,v)=(xss ,uss ,αv )

∂h(x, w) 
C = C(α, w) = (5.37)
∂x (x,u,v)=(xss ,uss ,αv )

With such a design choice, we know that the closed-loop nonlinear system

ẋ = f (x, u, v, w)
σ̇ = y − r = h(x, w) − r
u = −K1 (α)x − K2 (α)σ

has a locally exponentially stable equilibrium point for a fixed ρ = α.

Remark 5.5 The fact that A is stable for every “frozen” ρ = α does not
guarantee stability of the closed loop system when ρ = α (t). However, there
are theoretical developments that relate the slow rate of change of ρ(t) to
the boundedness of the system response. Moreover, if ρ(t) converges asymp-
totically to some constant value, then the error in output tracking would
asymptotically converge to zero.

Remark 5.6 We can extent the method to the case when we don’t have full
state feedback, but rather we need to design an estimator of the state.

P. Al Hokayem & E. Gallestey 105 of 214

Nonlinear Systems and Control — Spring 2018

Example 5.7 Consider the second order system

   
ẋ1 tan(x1 ) + x2
ẋ = = = f (x, u)
ẋ2 x1 + u (5.38)
y = h(x) = x2
Assume that we measure the full state, and would like the output to track a
reference signal r. Now, for any fixed value of the reference r = α, we obtain
the following (unique) steady-state solution
 
− tan−1 (α)
xss (α) = , uss (α) = tan−1 (α) (5.39)
α
We use the following integral controller
σ̇ = y − r = y − α (5.40)
u = −K1 (α)x − K2 (α)σ (5.41)
The linearized system matrices of (5.38) are given by
     
∂f (x, u)  1 + α2 1 ∂f (x, u)  0
A = A(α) =  = , B =  =
∂x xss ,uss
1 0 ∂u xss ,uss
1

Therefore, we would like to design the gains K1 (α) and K2 (α) such that the
closed-loop matrix of the linearized system
 
A(α) − BK1 (α) −BK2 (α)
A(α) = (5.42)
C 0
is Hurwitz. The gains are designed as
 1

K1 (α) = (1 + α2 )(3 + α2 ) + 3 + 1+α2
3 + α2
1
K2 (α) = −
1 + α2
√
3
which would assign the closed-loop eigenvalues at −1, − 12 ±j 2
. The response
of the closed-loop system is shown in Figure 5.5.

Additional performance can be achieved if the derivative of the output y

can be reliably estimated. In those cases, one can extend the original system
with the equation
∂h dx ∂h
ẏ = = f
∂x dt ∂x
P. Al Hokayem & E. Gallestey 106 of 214
Nonlinear Systems and Control — Spring 2018

0.8
y
0.6 r

0.4

0.2

−0.2

−0.4

−0.6

−0.8
0 50 100 150 200

Figure 5.5: Response of integral control design

and apply the same procedures and ideas that have been discussed in this
chapter to the new vector of measurements given by [y; ẏ].
The result of applying this method to the previous example can be seen
in Figure 5.6. In the first plot, we see the linearization failing once we move
away from the linearisation point. In the second plot, we see the good effect
of adapting the linearization point using gain scheduling, but we see large
transients for some setpoints. The last plot shows that the large transients
disappear when the derivative of the observation is added to the original
system as new measurement.
For further reading on these design methods, see [4, Ch. 12].

P. Al Hokayem & E. Gallestey 107 of 214

Nonlinear Systems and Control — Spring 2018

Figure 5.6: Response of integral control design with added derivative action

5.5 Exercises
1. A magnetic suspension system is modelled as
ẋ1 = x2
k L0 ax23
ẋ2 = g − x2 −
m 2m(a + x1 )2
 
1 L0 ax2 x3
ẋ3 = −Rx3 + +u
L(x1 ) (a + x1 )2
where x1 = y is the elevation, x2 = ẏ, x3 = I the current, and u = V
is the input voltage. Use the following numerical data m = 0.1 kg,
k = 0.001 N/(m s), g = 9.81 m/s2 , a = 0.05 m, L0 = 0.01 H, L1 = 0.02
H, R = 1 Ω, and L(x1 ) = L1 + 1+L0x1 .
a

(a) Find the steady-state values Iss and Vss of I and V , respectively,
which are needed to keep the ball at a given position y = r.
(b) Is the equilibrium point obtained above stable ?
(c) Linearize the system and design a state feedback control law to
stabilize the ball at y = 0.05 m.

P. Al Hokayem & E. Gallestey 108 of 214

Nonlinear Systems and Control — Spring 2018

2. A simplified model of the low-frequency motion of a ship is given by

τ ψ̈ + ψ̇ = kσ

where ψ is the heading angle of the ship and σ is the rudder angle,
viewed as the control input. The time constant τ and the gain k de-
pends on the forward speed of the ship v, according to the expressions
τ = τ0 vv0 and k = k0 vv0 , where τ0 , k0 , and v0 are constants. Assuming
a constant forward speed, design a state feedback integral controller so
that ψ tracks a desired angle ψr .

3. Suppose we have a controllable system ẋ = Ax + Bu. We now augment

the system with an additional integral state σ̇ = Cx ∈ Rp , p the number
of inputs, and obtain the augmented system

ẋa = Aa xa + Ba u

where xa = [xT , σ T ]T and

   
A 0 B
Aa = , Ba =
C 0 0

Show that the augmented system is also controllable if and only if

 
A B
rank = n + p.
C 0

4. Consider the following non linear system for (x1 , x2 ) and measurements
y

ẋ1 = −x1 + x2
x3
ẋ2 = −β 2 − γx2 + u
3
y = x1

with β > 0, γ > 0 parameters. We want to output y to track a reference

signal r. Design, if possible, an integral controller that performes the
task and is robust with respect to parameter variation.

P. Al Hokayem & E. Gallestey 109 of 214

Chapter 6

Feedback Linearization

6.1 Introduction
Consider a class of single-input-single-output (SISO) nonlinear systems of
the form
ẋ = f (x) + g(x)u (6.1)
y = h(x) (6.2)
where x ∈ D ⊂ Rn , u, y ∈ R1 , f : D → Rn , g : D → Rn , and the domain D
contains the origin. 1
In this lecture, we shall answer the following two (tightly related) ques-
tions:
 
η
1. Does there exist a nonlinear change of variables z = = T (x), and
ξ
a control input u = α(x) + β(x)v that would transform the system
(6.1)-(6.2) into the following partially linear form?
η̇ = f0 (η, ξ)
ξ˙ = Aξ + Bv
y = Cξ

2. Does there exist a nonlinear change of variables z = T (x) and a control

input u = α(x) + β(x)v that would transform the system (6.1) into the
1
This condition is not a limitation to the theory of feedback linearization and the results
can be extended to the multiple-input-multiple-output (MIMO) case

110
Nonlinear Systems and Control — Spring 2018

following fully linear form?

ż = Ãz + B̃v

If the answer to question 2. is positive, then we say that the system (6.1) is
feedback linearizable. If the answer to question 1. is positive, then we say
that the system (6.1)-(6.2) is input-output linearizable. Both scenarios are
very attractive from the control design point of view, since in either case we
can rely on linear design techniques to render the closed-loop system stable.
Since we cannot expect every nonlinear system to possess such proper-
ties of feedback linearization, it is interesting to understand the structural
properties that the nonlinear system should possess that render it feedback
linearizable.
Before we dive into the theoretical developments of feedback linearization,
let us first look into two simple examples.

u ẋ = f (x) + g(x)u
v α(x) + β(x)v y
y = h(x)

Figure 6.1: General idea of feedback linearization

Example 6.1 Consider the example of the pendulum that we have seen in
previous lectures. The dynamics are given by

ẋ1 = x2
ẋ2 = −a [sin (x1 + δ) − sin δ] − bx2 + cu

If we choose the control

a v
u= [sin (x1 + δ) − sin δ] +
c c
we can cancel the nonlinear term a [sin (x1 + δ) − sin δ]. The resulting linear
system is given by

ẋ1 = x2
ẋ2 = −bx2 + v

P. Al Hokayem & E. Gallestey 111 of 214

Nonlinear Systems and Control — Spring 2018

As such, the stabilization problem for the nonlinear system has been reduced
to a stabilization problem for a controllable linear system. We can proceed to
design a stabilizing linear state feedback control
v = −k1 x1 − k2 x2
that renders the closed-loop system
ẋ1 = x2
ẋ2 = −k1 x1 − (k2 + b) x2
asymptotically stable. The overall state feedback control law comprises linear
and nonlinear parts
a 1
u= [sin (x1 + δ) − sin δ] − (k1 x1 + k2 x2 )
c c
Example 6.2 Consider the system
ẋ1 = a sin x2
ẋ2 = −x21 + u
We cannot simply choose u to cancel the nonlinear term a sin x2 . However,
let us first apply the following change of variables
z1 = x1
z2 = a sin x2 = ẋ1
Then, the new variables z1 and z2 satisfy
ż1 = z2
 z2 

ż2 = a cos x2 .ẋ2 = a cos sin−1 u − z12
a
and the nonlinearities can be canceled by the control
1
u = x21 + v
a cos x2
which is well defined for − π2 < x2 < π2 . The state equation in the new
coordinates can be found by inverting the transformation to express (x1 , x2 )
in the terms of (z1 , z2 ), that is,
x1 = z1
z 
−1 2
x2 = sin
a

P. Al Hokayem & E. Gallestey 112 of 214

Nonlinear Systems and Control — Spring 2018

which is well defined for −a < z2 < a. The transformed state equation is
given by
ż1 = z2
  z 

2
ż2 = a cos sin−1 −z12 + u
a
which is in the required form to use state feedback. Finally, the control input
that we would use is of the following form
1 1
u = x21 + (−k1 z1 − k2 z2 ) = x21 + (−k1 x1 − k2 a sin(x2 ))
a cos x2 a cos x2

6.2 Input-Output Linearization

Consider the single-input-single-output system (6.1)-(6.2), where the vector
fields f : D → Rn and g : D → Rn and the function h : D → R are sufficiently
smooth.
Definition 6.3 f is called smooth if f ∈ C ∞ , that is, f is continuous and
all its derivatives of all orders are continuous.
The goal is to derive conditions under which the input-output map can be
rendered linear. The idea is to take a sufficient number of derivatives of the
output, until the input appears.
We proceed by taking the derivative of y, which is given by
∂h
ẏ = [f (x) + g (x) u] = Lf h (x) + Lg h (x) u
∂x
where Lf h (x) , ∂h ∂x
f (x) is called the Lie Derivative of h with respect to
(along) f . If Lg h (x) = 0, then ẏ = Lf h (x), independent of u and we repeat
the differentiation process again. Calculating the second derivative of y,
denoted by y (2) , we obtain
∂ (Lf h)
y (2) = [f (x) + g (x) u] = L2f h (x) + Lg Lf h (x) u
∂x
Once again, if Lg Lf h (x) = 0, then y (2) = L2f h (x) is independent of u and we
repeat the process. Actually, we repeat the processes of taking derivatives of
the output until we see that h (x) satisfies
Lg Li−1
f h (x) = 0, i = 1, 2, ..., ρ − 1; Lg Lρ−1
f h (x) 6= 0

P. Al Hokayem & E. Gallestey 113 of 214

Nonlinear Systems and Control — Spring 2018

Therefore, u does not appear in the expressions of y, ẏ,. . . ,y (ρ−1) and appears
in the expression of y (ρ) with a nonzero coefficient, i.e.,

y (ρ) = Lρf h (x) + Lg Lρ−1

f h (x) u (6.3)

We can clearly see from (6.3), that the system is input-output linearizable,
since the state feedback control
1

u= ρ−1 −Lρf h (x) + v (6.4)
Lg Lf h (x)

reduces the input-output map (in some domain D0 ⊂ D) to

y (ρ) = v

which is a chain of ρ integrators. In this case, the integer ρ is called the

relative degree of the system.
Definition 6.4 The nonlinear system (6.1)-(6.2) is said to have a relative
degree ρ, 1 ≤ ρ ≤ n, in the region D0 ⊂ D if

Lg Lif h(x) = 0, i = 0, · · · , ρ − 2
(6.5)
Lg Lρ−1
f h(x) 6= 0

for all x ∈ D0 .

Example 6.5 Consider the controlled van der Pol equation

ẋ1 = x2


ẋ2 = −x1 + ε 1 − x21 x2 + u
with output y = x1 . Calculating the derivatives of the output, we obtain
ẏ = ẋ1 = x2
ÿ = ẋ2 = −x1 + ε (1 − x21 ) x2 + u

Hence, the system has relative degree two in R2 . For the output y = x1 + x22 ,
we have that 

ẏ = x2 + 2x2 −x1 + ε 1 − x21 x2 + u
and the system has relative degree one in D0 = {x ∈ R2 |x2 6= 0}. As such,
we can see that the procedure of input-output linearization is dependent on
the choice of the output map h(x).

P. Al Hokayem & E. Gallestey 114 of 214

Nonlinear Systems and Control — Spring 2018

Continuing with our derivations, we now let

 
φ1 (x)
 .. 
   .     
 φ (x) 
T1 (x)  n−ρ  φ (x) η
 
z = T (x) = − − − ,  − − −  ,  − − −  ,  − − − 
 h (x) 
T2 (x)   ψ (x) ξ
 . 
 .. 
ρ−1
Lf h (x)
(6.6)
where φ1 (x) to φn−ρ (x) are chosen such that
∂φi
g (x) = 0, 1≤i≤n−ρ (6.7)
∂x
which ensures that the η-dynamics

∂φ ∂φ 
η̇ = [f (x) + g (x) u] = f (x)
∂x ∂x x=T −1 (z)

are independent
  of u. Note also that our definition of T2 (x) in (6.6) results
y
 y (1) 
 
in ξ =  .. .
 . 
y (ρ−1)
Of course, it is crucial to know at this point if such functions φ exist in
order to define the transformation T ; the next result shows that.
Definition 6.6 A continuously differentiable transformation T with a con-
tinuously differential inverse is a called a diffeomorphism.

Theorem 6.7 Consider the system (6.1)-(6.2), and suppose that it has a
relative degree ρ ≤ n in D. If ρ = n, then for every x0 ∈ D, a neighborhood
N of x0 exists such that the map
 
h (x)
 Lf h(x) 
 
T (x) =  .. 
 . 
Ln−1
f h (x)

P. Al Hokayem & E. Gallestey 115 of 214

Nonlinear Systems and Control — Spring 2018

restricted to N is a diffeomorphism on N . If ρ < n, then for every x0 ∈ D,

a neighborhood N of x0 and smooth maps φ1 (x), · · · , φn−ρ (x) exist such that
the map such that the map
 
φ1 (x)
 .. 
 . 
 φ 
 n−ρ (x) 
 
T (x) =  − − − 
 h (x) 
 
 .. 
 . 
ρ−1
Lf h (x)

restricted to N , is a diffeomorphism on N .

We can now apply the change of variables z = T (x) to transform the

P. Al Hokayem & E. Gallestey 116 of 214

Nonlinear Systems and Control — Spring 2018

2
system (6.1)-(6.2). into

η̇=f0 (η, ξ) (6.8)

˙ c ξ + Bc γ (x) [u − α (x)]
ξ=A (6.9)
y=Cc ξ (6.10)

where ξ ∈ Rρ , η ∈ Rn−ρ , (AC , BC , CC ) is a canonical form representation of

a chain of ρ integrators, i.e.,
 
0 1 0 ··· 0  
0 0 1 · · · 0 0
   ..   
 . . . ..   
Ac =  ... .  , Bc =  .  , Cc = 1 0 · · · 0
  0
0 0 1
1
0 ··· 0 0
2
The η-dynamics can be derived as follows
∂φ ∂φ
η̇ = φ̇(x) = ẋ = (f (x) + g(x)u)
∂x ∂x 
∂φ ∂φ 
= f (x) = f (x) , f0 (η, ξ)
∂x ∂x x=T −1 (x)

The ξ-dynamics can be derived as follows

˙     (1)   
ξ1 y y ξ2
ξ˙2  d 
(1)   (2)   . 
˙ξ =    y  y   .. 
.=  . = . = 
 ..  dt  ..   ..   ξρ 
ξ˙ρ y (ρ−1) y (ρ) y (ρ)
" !#
Lρf h(x)
= Ac ξ + B c y (ρ)
= Ac ξ + Bc (Lg Lρ−1
f h(x)) (u − −
Lg Lρ−1
f h(x)

we can now define

Lρf h(x)
γ(x) , Lg Lρ−1
f h(x), α(x) , −
Lg Lρ−1
f h(x)

Finally, our choice of the transformation T2 (x) yields

y = h(x) = ξ1 = Cc ξ

P. Al Hokayem & E. Gallestey 117 of 214

Nonlinear Systems and Control — Spring 2018

and

∂φ 
f0 (η, ξ) = f (x) (6.11)
∂x x=T −1 (z)

γ (x) = Lg Lρ−1
f h (x) (6.12)
Lρf h (x)
α (x) = − (6.13)
Lg Lρ−1
f h (x)

R ξ˙ρ−1 = ξρ R ξ˙ρ−2 = ξρ−1 ξ˙1 = ξ2 R

v = ξ˙ρ y = ξ1

ξ η̇ = f0 (η, ξ)

Figure 6.2: Transformed system (6.8)-(6.10) and the input defined as in (6.4)

We have kept α and γ expressed in the original coordinates. These two

functions are uniquely determined in terms of f , g, and h and are independent
of the choice of φ. They can be expressed in the new coordinates by setting



α0 (η, ξ) = α T −1 (z) , γ0 (η, ξ) = γ T −1 (z) (6.14)
but now they are dependent on the choice of the functions φ. Regarding the
definition of the equilibrium point for the transformed system, assume that
x̄ is the open-loop equilibrium of the system (6.1), then
 T
η̄ = φ(x̄), ξ¯ = h(x̄) 0 · · · 0 (6.15)
The transformed system (6.8)-(6.10) is said to be in the normal form.
This form decomposes the system into an external part ξ and an internal
part η. The external part is linearized by the state feedback control
u = α (x) + β (x) v (6.16)
where β (x) = γ −1 (x), while the internal part is made unobservable by the
same control (see Figure 6.2). Setting ξ = 0 in the internal dynamics (6.8),
results in
η̇ = f0 (η, 0) (6.17)

P. Al Hokayem & E. Gallestey 118 of 214

Nonlinear Systems and Control — Spring 2018

which is called the zero dynamics of the system. If the zero dynamics of the
system are (globally) asymptotically stable, the system is called minimum
phase.
Finally, the linearized system may then be stabilized by the choice of an
appropriate state feedback:
v = −Kξ.

6.2.1 Relationship to Linear Systems

The notions of relative degree and minimum phase can also be found in linear
systems. Consider a linear system represented by the transfer function

bm sm + bm−1 sm−1 + ... + b0

G (s) = (6.18)
sn + an−1 sn−1 + ... + a0

where m < n and bm 6= 0. We can realize the transfer function in (6.18) in

state-space form as
ẋ = Ax + Bu, y = Cx (6.19)
where
 
0 1 0 ··· 0  
 0 0 1 ··· 0  0
 . .. ..  0
 . ... ...   
 . . .   
A= . ... ... , B =  ... 
 ..   
  0
 0 0 1 
1
−a0 −a1 · · · −an−1
 
C = b0 b1 ··· bm 0 · · · 0

This linear space model is a special case of

ẋ = f (x) + g (x) u, y = h (x)

where f (x) = Ax, g = B, and h (x) = Cx. We know that the relative degree
of the transfer function is given by n − m. However, let us verify that this is
the case. We take the derivative of the output, i.e.,

ẏ = C ẋ = CAx + CBu (6.20)

P. Al Hokayem & E. Gallestey 119 of 214

Nonlinear Systems and Control — Spring 2018

Now, if m = n − 1, then CB = bn−1 6= 0 and the system has relative degree

one. In general we have that

y (n−m) = CA(n−m) x + CA(n−m−1) Bu (6.21)

and we have the conditions that

CA(i−1) B = 0, ∀i = 1, 2, · · · , n − m − 1, and CA(n−m−1) B = bm 6= 0 (6.22)

and the relative degree of the system is n − m, i.e., the difference between the
order of the denominator and the numerator of the transfer function G(s).
Moreover, we know that a minimum phase system means that the zeros
of the transfer function G(s) lie in the open left-half complex plane. We can
show as well that there is a correspondence of the zeros of H(s) and the
normal form zero dynamics.

6.3 Full State Feedback Linearization

We know from the previous section, that the system (6.1) is feedback lin-
earizable if we can find a sufficiently smooth function h(x) in a domain D
such that the system (6.1)-(6.2) has relative degree n within some region
D0 ⊂ D. This is because the normal form would have no zero dynamics. In
fact, we can show that this result holds as an if and only if statement. As
such, it remains to show that such a function h(x) exists.
We begin with some definitions. For any two vector fields f and g on
D ⊂ Rn , the Lie bracket [f, g] is a third vector field that is defined as
∂g ∂f
[f, g](x) , f (x) − g(x) = Lf g(x) − Lg f (x) (6.23)
∂x ∂x
Taking Lie brackets can be repeated, as such we define the following notation:

ad0f g(x) = g(x)

adf g(x) = [f, g](x)
adkf g(x) = [f, adk−1
f g](x), k≥1

Example 6.8 Consider the two vector fields f (x) = Ax and g is constant.
Then, adf g(x) = [f, g](x) = −Ag, ad2f g(x) = [f, adf g](x) = −A(−Ag) =
A2 g, and adkf g = (−1)k Ak g.

P. Al Hokayem & E. Gallestey 120 of 214

Nonlinear Systems and Control — Spring 2018

Definition 6.9 (Distribution) Consider the vector fields f1 , · · · , fk on a

domain D ⊂ Rn . The distribution, denoted by

∆ = {f1 , f2 , · · · , fk } (6.24)

is the collection of all vectors spaces ∆(x) for x ∈ D, where

∆(x) = span{f1 (x), f2 (x), · · · , fk (x)}

is the subspace of Rn spanned by the vectors f1 (x), · · · fk (x).

Definition 6.10 A distribution ∆ is called involutive, if

g1 ∈ ∆, and g2 ∈ ∆ ⇒ [g1 , g2 ] ∈ ∆

And now we are ready to state the main result of this section.

Theorem 6.11 (Feedback Linearization) The SISO system (6.1) is feed-

back linearizable if and only if

1. the matrix M(x) = [g(x), adf g(x), · · · , adn−1

f g(x)] has full rank (= n)
for all x ∈ D0 , and

2. the distribution ∆ = span{g, adf g, · · · , adn−2

f g} is involutive on D0 .

Example 6.12 Consider the system

   
a sin(x2 ) 0
ẋ = f (x) + gu = + u
−x21 1

We compute the Lie bracket

 
∂f −a cos(x2 )
adf g = [f, g](x) = − g =
∂x 0

Accordingly, the matrix

 
0 −a cos(x2 )
M(x) = [g, adf g] =
1 0

has full rank for any x such that cos(x2 ) 6= 0. Moreover, the distribution
∆ = span{g} is involutive. Therefore, the conditions of Theorem (6.11) are

P. Al Hokayem & E. Gallestey 121 of 214

Nonlinear Systems and Control — Spring 2018

satisfied and the system is feedback linearizable in the domain D0 = {x ∈

R2 | cos(x2 ) 6= 0}. Let us now find the function h(x) for which the system
is feedback linearizable. h(x) should satisfy the following conditions
∂h ∂(Lf h)
g = 0, g 6= 0, h(0) = 0
∂x ∂x
∂h ∂h
Now, ∂x
g =
∂x2
= 0 which implies that h(x) is independent of x2 . Therefore,
∂h ∂(L h) ∂(L h) ∂h
Lf h(x) = ∂x1 a sin(x2 ). The condition ∂xf g = ∂xf2 = ∂x 1
a cos(x2 ) 6= 0 is
∂h
satisfied in D0 for any choice of h that satisfies ∂x1 6= 0. There are many
such choices for h, for example, h(x) = x1 or h(x) = x1 + x31 .

6.4 Stability
Consider again the input-output linearized system in normal form (6.8)-
(6.10), and assume that we have designed the feedback control law
u = α(x) + β(x)v
where β(x) = γ −1 (x), and v = −Kξ, such that (Ac − Bc K) is Hurwitz. The
resulting system reduces to the following dynamics
η̇ = f0 (η, ξ) (6.25)
ξ˙ = (Ac − Bc K)ξ (6.26)

Theorem 6.13 The origin of the system (6.25)-(6.25) is asymptotically sta-

ble, if the origin of the zero dynamics η̇ = f0 (η, 0) is asymptotically stable
(minimum phase).

Proof: The idea is to construct a special Lyapunov function. Since the

zero dynamics are asymptotically stable, there exists (by converse Lyapunov
theorem) a continuously differentiable function Vη (η) such that
∂Vη
f (η, 0) < −W (kηk)
∂η
in some neighborhood of η = 0, where W is a strictly increasing continuous
function with W (0) = 0. Let P = P T > 0 be the solution of the Lyapunov
equation.
P (Ac − Bc K) + (Ac − Bc K)T P = −I

P. Al Hokayem & E. Gallestey 122 of 214

Nonlinear Systems and Control — Spring 2018

Consider the function

p
V (η, ξ) = Vη (η) + k ξ T P ξ, k>0 (6.27)
The derivative of V is given by
∂Vη k
V̇ = f0 (η, ξ) + p ξ T [P (AC − BC K) + (AC − BC K)P ] ξ
∂η 2 ξ Pξ |
T {z }
−I
∂Vη ∂Vη k
= f (η, 0) + (f0 (η, ξ) − f0 (η, 0)) p ξT ξ
∂η ∂η T
2 ξ Pξ
≤ −W (kξk) + k1 (kξk) − kk2 (kξk)

for k1 , k2 > 03 . By choosing k large enough, we guarantee V̇ < 0, and the

result follows.

Remark 6.14 The result in the last theorem is local and does not extend to
the global setup, even if the zero dynamics are globally asymptotically sta-
ble. In order to make the result global, we have to impose more restrictive
requirements regarding the zero dynamics, namely the notion of input-to-
state stability of the zero dynamics (which is beyond the current scope of the
course).

6.5 Robustness
Feedback linearization is based on exact cancellation of the nonlinearities
in the system dynamics, which is practically very difficult to achieve. In a
realistic setup, we would have only approximations α̂, β̂ and T̂ (x) of the true
α, β, and T (x). The feedback control law has then the form
u = α̂(x) + β̂(x)v = α̂(x) − β̂(x)Kξ = α̂(x) − β̂(x)K T̂2 (x)
Accordingly, the closed-loop system of the normal form is given by
η̇ = f0 (η, ξ) (6.28)
h i
ξ˙ = Aξ + Bγ (x) (α̂(x) − α(x)) − β̂(x)K T̂2 (x) (6.29)
3
k1 > 0 follows from local continuity and differentiability of the function f0 , and k2 > 0
follows from P

P. Al Hokayem & E. Gallestey 123 of 214

Nonlinear Systems and Control — Spring 2018

Adding and subtracting BKξ to the ξ˙ equation we obtain

η̇ = f (η, ξ) (6.30)
ξ˙ = (A − BK) ξ + Bδ (z) (6.31)

where

δ (z) = γ(x)[(α̂(x) − α(x)) − (β̂(x) − β(x))KT

− β̂(x)K(T̂2 (x) − T (x))]|x=T −1 (z)

Hence, the local closed loop system differs from the nominal one by an ad-
ditive perturbation. Thus, we can show that locally the closed-loop system
remains stable, despite small uncertainties in the feedback linearization maps.
For more details on feedback linearization concepts and their extensions,
see [2].

P. Al Hokayem & E. Gallestey 124 of 214

Nonlinear Systems and Control — Spring 2018

6.6 Exercises
1. Consider the fourth order single input system

ẋ = f (x) + gu

with
   
x2 0
−a sin x1 − b(x1 − x3 ) 0
f (x) = 

,
 g=
0

x4
c(x1 − x3 ) d

where a, b, c and d are strictly positive constants. Give the transfor-

mation T = [T1 , . . . , T4 ] that transforms the system into the standard
form. Design a controller using feedback linearization to stabilize the
origin.

2. Consider the nonlinear system

ẋ1 = −x1 + x2 − x3
ẋ2 = −x1 x3 − x2 + u
ẋ3 = −x1 + u
y = x3

Is the system input-output linearizable? Is the system minimum phase?

3. Consider the following SISO nonlinear system

ẋ1 = x2 + sin x1
ẋ2 = x21 + γu
y = x1

where γ is a given scalar. Design a continuous feedback controller

via feedback linearization so that the output y tracks a given signal
v(t) = sin t.

P. Al Hokayem & E. Gallestey 125 of 214

Chapter 7

Sliding Mode Control

7.1 Introduction
Consider the second-order nonlinear system
ẋ1 = x2
(7.1)
ẋ2 = f (x) + g(x)u
where f (x) and g(x) ≥ g0 > 0 are unknown nonlinear functions and g0 is a
lower bound on the function g. The main goal is to design a feedback control
law u(x) that would guarantee stability in the presence of the uncertainty,
i.e., a robustly stabilizing control law.
Had we known the nonlinearities exactly, the control design problem
would be simple, i.e.,
1
u= (−f (x) − k1 x1 − k2 x2 ) (7.2)
g(x)
where any choice of the gains k1 , k2 > 0 guarantees stability of the closed-loop
system. However, since we do not know the nonlinear functions, our goal is
to design a potentially switched controller that guarantees robust stability of
the system.
For this reason, assume for a moment that we can directly manipulate
x2 , i.e., use x2 as an input to the first equation in (7.1), then we can pick
x2 = −ax1 with a > 0, and make the first equation exponentially stable. This
idea motivates us to define the so-called sliding surface or sliding manifold
σ = ax1 + x2 = 0 (7.3)

126
Nonlinear Systems and Control — Spring 2018

As such, if we can control the system to the manifold σ and keep it there,
the state trajectory would just asymptotically slide towards the origin. Dif-
ferentiating the function σ along the trajectories of the system yields
σ̇ = aẋ1 + ẋ2 = ax2 + f (x) + g(x)u
Assume further that the unknown functions satisfy the following bound
 
 ax2 + f (x) 
  2
 g(x)  ≤ ρ(x), ∀x ∈ R (7.4)

1 2
for some known function ρ(x). Taking V = 2
σ as a Lyapunov function
candidate, we have that
V̇ = σ σ̇ = σ(ax2 + f (x) + g(x)u) ≤ g(x) [|σ|ρ(x) + σu]
We can now design

1, σ>0
u = −(β0 + ρ(x))sgn(σ), β0 > 0, sgn(σ) = (7.5)
−1, σ<0
which yields
V̇ ≤ −g0 β0 |σ| (7.6)
√
We can now define the function W (σ) , 2V , which satisfies1
W (σ(t)) ≤ W (σ(0)) − β0 g0 t (7.7)
This shows that the trajectory of the closed-loop system (7.1) and (7.5)
reaches the surface σ in finite time, and once on this surface, it cannot leave
it, as seen by the inequality (7.6). The fact that the first phase is finished in
finite time is very important in order to claim asymptotic stability (why?).
We can now categorize the behavior of the closed-loop system under slid-
ing mode control into two phases:
Reaching Phase in which the designed input drives all trajectories that
start off the surface σ towards σ in finite time and keeps them there
Sliding Phase in which the motion of the system is confined to the surface
σ and can be described by some reduced order (stable) dynamics
Such a choice of surface σ is called a sliding surface and the corresponding
input is called sliding mode control.
1
The inequality (7.7) can be formally proven using the so-called comparison lemma and
the notion of Dini derivatives, which are beyond the scope of the course.

P. Al Hokayem & E. Gallestey 127 of 214

Nonlinear Systems and Control — Spring 2018

x2

x1

σ = x1 + ax2 = 0

Figure 7.1: Illustration of the sliding mode control in the plane

7.2 Sliding Mode Control

Let us now consider the more general scenario of a nonlinear system given
by
ẋ = f (x) + B(x)[G(x)E(x)u + δ(x, u)] (7.8)
where x ∈ Rn is the state, u ∈ Rm is the input, f , B, G, and E are sufficiently
smooth functions within some domain D ∈ Rn that contains the origin. We
assume that f , B, and E are known, while G and δ are uncertain, and that
δ(x, u) is sufficiently smooth in both arguments for (x, u) ∈ D × Rm . Assume
further that E(x) is invertible and that G(x) is a diagonal matrix with all
elements being bounded away from zero, i.e., gi (x) ≥ g0 > 0, for all x ∈ D.
Let f (0) = 0, so that if δ = 0 then the origin is an open-loop equilibrium
point.

Remark 7.1 The equation (7.8) shows that we have the scenario of matched
uncertainty, i.e., the input u and the uncertainty δ affect the dynamics via
the same B(x) matrix.

Let T : D → Rn be a diffeomorphism such that

 
∂T 0
B(x) = (7.9)
∂x Im×m

The dynamics of the new variables defined by the transformation T , i.e.,

P. Al Hokayem & E. Gallestey 128 of 214

Nonlinear Systems and Control — Spring 2018

   
η T1 (x)
z= = T (x) = are given by
ξ T2 (x)

∂T1 
η̇ = f (x) = fa (η, ξ) (7.10)
∂x x=T −1 (z)

∂T 2 
ξ˙ = f (x) + G(x)E(x)u + δ(x, u) = fb (η, ξ) + G(x)E(x)u + δ(x, u)
∂x x=T −1 (z)
(7.11)

Consider the sliding manifold

σ = ξ − φ(η) = 0 (7.12)

such that when we restrict the motion of the system to the manifold, the
reduced order model
η̇ = fa (η, φ(η)) (7.13)
is asymptotically stable to the origin. In order to solve for the function
φ(η) we just need to look for a stabilizing feedback in which ξ is taken as
the control input for the reduced dynamics (just as we have seen in the
Introduction section). The choice of the design method is open, and we can
use any of the design techniques we have learned in this course.
Assume that we have solved the stabilization problem for the system
(7.13) and obtained a stabilizing continuously differentiable function φ(η),
such that φ(0) = 0. We now need to design the input u that achieves the
reaching phase to the sliding manifold and maintains the trajectory on the
sliding manifold. The derivative of σ is given by
∂φ ∂φ
σ̇ = ξ˙ − η̇ = fb (η, ξ) + G(x)E(x)u + δ(x, u) − fa (η, ξ) (7.14)
∂η ∂η

Let Ĝ(x) be the nominal model of G(x) and consider the control input
   
−1 −1 ∂φ
u = E (x) −Ĝ (x) fb (η, ξ) − fa (η, ξ) + v (7.15)
∂η

where v is a free input that we would like to design. Substituting (7.15) into
(7.14), we obtain
σ̇ = G(x)v + ∆(x, v) (7.16)

P. Al Hokayem & E. Gallestey 129 of 214

Nonlinear Systems and Control — Spring 2018

where
∂φ
∆(x, v) = (I − G(x)Ĝ−1 (x))(fb (η, ξ) − fa (η, ξ))
∂η
    
−1 −1 ∂φ
+ δ x, E (x) −Ĝ (x) fb (η, ξ) − fa (η, ξ) + v
∂η

Assume that the ratio of the diagonal elements ∆i /gi satisfies

 
 ∆i (x, v) 
  m
 gi (x)  ≤ ρ(x) + κ0 kvk∞ , ∀(x, v) ∈ D × R , ∀i ∈ {1, 2, · · · , m}
(7.17)
where ρ(x) ≥ 0 is a known continuous function and κ0 ∈ [0, 1) is a known
constant.
Consider now the candidate Lyapunov function V = 12 σ T σ, whose deriva-
tive can be computed as

V̇ = σ T σ̇ = σ T (G(x)v + ∆(x, v))

m m   
X X  ∆i (x, v) 
= (σi gi (x)vi + σi ∆i (x, u)) ≤ gi (x) σi vi + |σi |  

i=1 i=1
g i (x)
m
X
≤ gi (x) (σi vi + |σi | (ρ(x) + κ0 kvk∞ ))
i=1

Choose the extra inputs vi as

vi = −β(x)sgn(σi ), i ∈ {1, 2, · · · , m} (7.18)

ρ(x)
β(x) ≥ + β0 , ∀x ∈ D (7.19)
1 − κ0
where β0 > 0. As such, we have that
m
X m
X
V̇ ≤ gi (x)(−β(x) + ρ(x) + κ0 β(x))|σi | ≤ − g0 β0 (1 − κ0 )|σi |
i=1 i=1

which ensures that all the trajectories that do not start on σ converge to the
manifold σ in finite time, and that those that start on σ stay on it.
Remark 7.2 We have used the sgn function for the design of the input v,
which poses theoretical questions in terms of the existence and uniqueness of

P. Al Hokayem & E. Gallestey 130 of 214

Nonlinear Systems and Control — Spring 2018

solutions of the closed-loop system of ODEs. However, this difficulty can be

alleviated via the using continuous approximations of the sgn function, for
example,

vi = −β(x)sat (σi /) , ∀i ∈ {1, 2, · · · , m}, and  > 0

The difference of these two choices is illustrated in Figure 7.2

sgn(y) sat(y)

1 1

y y
−
−1 −1

Figure 7.2: Illustrations of the sgn and sat functions

Remark 7.3 In practice, we would use a digital implementation of the any

control method. As such, it we may not be able to instantaneously switch the
control input v. This results in the so-called chattering phenomenon. The
effect of chattering may result in high heat losses and wearing out mechanical
systems. This effect can be alleviated by introducing some time and/or space
hysteresis bands that govern the switching commands whenever the trajectory
is close to the sliding surfaces.

Example 7.4 Consider the second-order system

ẋ1 = x2 + p1 x1 sin(x2 )
ẋ2 = p2 x22 + x1 + u

where the parameters p1 and p2 are unknown, but satisfy |p1 | ≤ a and |p2 | ≤ b,
where a and b are known. We take η = x1 and ξ = x2 , and we would like to
robustly stabilize the η = x1 dynamics to the origin by using x2 as an input.
To this end, we design x2 = −kx1 , where k > a, which yields

ẋ1 = −(k − p1 sin(−kx1 ))x1

P. Al Hokayem & E. Gallestey 131 of 214

Nonlinear Systems and Control — Spring 2018

which is a robustly stable system for our choice of k. Accordingly, define the
sliding manifold as
σ = x2 + kx1 = 0
The derivative of σ satisfies


σ̇ = ẋ2 + k ẋ1 = (x1 + kx2 ) + p2 x22 + kp1 x1 sin(x2 ) + u

We design the input u so as to cancel the known term, i.e.,

u = − (x1 + kx2 ) + v

For this choice of u, we have that

σ̇ = v + ∆(x)

where the uncertainty ∆(x) satisfies the bound

|∆(x)| ≤ bx22 + ka|x1 | = ρ(x)

Consider the Lyapunov function candidate V = 21 σ 2 , then

V̇ = σ σ̇ = σ(v + ∆(x))
≤ σv + |σ|ρ(x)

which proves finite time stability to the manifold σ, for any choice of v =
−β(x)sgn(s) (or some continuous approximation) with β(x) = ρ(x) + β0 ,
β0 > 0. Finally, the sliding mode controller is given by

u = − (x1 + kx2 ) − (bx22 + ka|x1 | + β0 )sgn(x2 + kx1 )

7.3 Tracking
Consider the system, comprised of a chain of integrators (SISO)

ẋ1 = x2
ẋ2 = x3
.. (7.20)
.
ẋn = f (x) + g(x)u + δ(x, u)
y = x1

P. Al Hokayem & E. Gallestey 132 of 214

Nonlinear Systems and Control — Spring 2018

where f (x) is a known function, while g(x) and δ(x) are uncertain functions
that satisfy, the first of which satisfies

g(x) ≥ g0 > 0

Remark 7.5 The system (7.20) may be the result of an input output feedback
linearization approach, i.e., the linearized dynamics in ξ that we have seen
in the lecture on feedback linearization.
We would like the output y to asymptotically track a given signal r(t),
where r(t) and its derivatives r(1) (t), · · · , r(n) (t) are bounded for all t, r(n) (t)
is piecewise continuous,
 and
 all these signals are available online.
x1 − r
 x2 − r(1) 
 
Let e =  ..  be the error vector, then we have the following
 . 
(n−1)
xn − r
error dynamics
ė1 = e2
ė2 = e3
.. (7.21)
.
ėn−1 = en
ėn = f (x) + g(x)u + δ(x) − r(n) (t)
Let us now assume that we can directly design en to stabilize the subsystem
ė1 = e2
ė2 = e3
.. (7.22)
.
ėn−1 = en
In this case, we can take any feedback of the form

en = −(k1 e1 + · · · + kn−1 en−1 )

where the coefficients are chosen so that the polynomial

sn−1 + kn−1 sn−2 + · · · + k1

P. Al Hokayem & E. Gallestey 133 of 214

Nonlinear Systems and Control — Spring 2018

is Hurwitz. Accordingly, we can now define the sliding manifold as

σ = k1 e1 + · · · + kn−1 en−1 + en = 0 (7.23)

The dynamics of σ are given by

σ̇ = k1 e2 + · · · + kn−1 en + f (x) + g(x)u + δ(x) − r(n) (t) (7.24)

We can now design the input

1

u=− k1 e2 + · · · + kn−1 en + f (x) − r(n) (t) + v
ĝ(x)

to yield
σ̇ = g(x)v + ∆(t, x, v)
where
 
g(x)

∆(t, x, v) = I − k1 e2 + · · · + kn−1 en + f (x) − r(n) (t)
ĝ(x)
 
1 (n)


+ δ x, − k1 e2 + · · · + kn−1 en + f (x) − r (t) + v
ĝ(x)

Assume that we have the following bound

 
 ∆(t, x, v) 
 
 g(x)  ≤ ρ(x) + κ0 |v|, κ0 ∈ [0, 1)

ρ(x)
then designing v = −β(x)sgn(σ) with β(x) ≥ 1−κ 0
+ β0 and β0 > 0 achieves
asymptotic tracking. You can show this using exactly the same argument as
in the previous section with a Lyapunov function candidate V = 12 σ 2 .

P. Al Hokayem & E. Gallestey 134 of 214

Nonlinear Systems and Control — Spring 2018

7.4 Exercises
1. Given the system:

ẋ1 = x2
ẋ2 = x21 + γu (7.25)
y = x1

with γ ∈ [1, 3] uncertain parameter, design a Sliding Mode Control that

tracks asymptotically a signal r(t) = sin t.

2. The motion of a rigid n-link robot manipulator is described by the

following equations:

H(q)q̈ + C(q, q̇)q̇ + G(q) = τ (7.26)

where q ∈ Rn and q̇ ∈ Rn are the so-called generalized coordinates

of the system, τ ∈ Rn represents the input (torques or forces). The
system above satisfies the following skew-symmetry properties:

• The matrix L = Ḣ(q)−2C(q, q̇) is skew-symmetric, i.e. L> = −L;

• H(q)> = H(q) ≥ λI, ∀q ∈ Rn with λ > 0.

Design a control τ that tracks the coordinates qd and q̇d , with the
following approaches:

(a) Design a feedback linearization control, assuming that the matri-

ces H, C and G are known exactly.
(b) Assume now that H, G and C are uncertain, and we only know
some nominal Ĥ, Ĉ and Ĝ. Design a SMC which accounts for this
mismatch and asymptotically stabilizes the system.
Hint: Take as sliding manifold σ = q̇ − q̇d + K (q − qd ) with
an accordingly defined K and leverage the positive definiteness of
H(q) by defining the Lyapunov function V = 12 σ > H(q)σ.

P. Al Hokayem & E. Gallestey 135 of 214

Chapter 8

Optimal Control

We consider first simple unconstrained optimization problems, and then

demonstrate how this may be generalized to handle constrained optimiza-
tion problems. With the observation that an optimal control problem is a
form of constrained optimization problem, variational methods are used to
derive an optimal controller, which embodies Pontryagin’s Minimum Princi-
ple. We provide necessary and sufficient conditions for usage in the context
of this important result.

8.1 Unconstrained Finite-Dimensional Opti-

mization
Consider a function
L:R→R
We want to find
min L(x)
x
Let us assume that L is sufficiently smooth, and consider the Taylor expan-
sion at the point x∗ :

∗ dL  ∗ 1 dL2 
L(x) = L(x ) +  (x − x ) +  (x − x∗ )2 + ..
dx x=x∗ 2 dx2 x=x∗
In order for the point x∗ to be a minima, we should have the following
condition holding (at least locally)
L(x∗ ) ≤ L(x), ∀x : |x − x∗ | < 

136
Nonlinear Systems and Control — Spring 2018

Then we have the following first-order necessary condition

dL 
 =0
dx x=x∗
and the second-order sufficient condition
dL2 
 >0
dx2 x=x∗
for optimality of x∗ . Note that these are only conditions for a local minimum.
Additional conditions are required to find the global minimum if the function
is non-convex. If we have a function with more than one variable, that is
L : Rn → R we have the following conditions
∂L ∂L ∂L 


∂x1 ∂x2
· · · ∂x n x=x∗
=0
and  
∂2L ∂2L 
 ∂x21
··· ∂x1 ∂xn 
∂ 2 L   .. ... .. 
= . .  >0
∂x2 x=x∗ 
∂2L
··· ∂2L 
∂xn ∂x1 ∂x2n x=x∗
i.e. is positive definite.

Example 8.1 Find the minimum of the function

L(x1 , x2 ) = 6x21 + 2x22 + 3x1 x2 − 8x1 + 3x2 − 4.
Necessary condition
  
∂L  12x1 + 3x2 − 8 
= =0
∂x x=x∗ 4x2 + 3x1 + 3 x=x∗
41 20
Solving these equations we find x∗ = ( 39 , − 13 ) and when we insert x∗ into
the Hessian Matrix we see that
  
∂ 2 L  12 3
= .
∂x2  ∗ x=x
3 4
the resulting matrix is positive definite. We conclude that the point x∗ is a
minimum.
The plot in Figure 8.1 shows the function for which we are finding the
minimum.

P. Al Hokayem & E. Gallestey 137 of 214

Nonlinear Systems and Control — Spring 2018

Figure 8.1: Convex Function

8.2 Constrained Finite-Dimensional Optimiza-

tion

Theorem 8.2 Consider the problem

min L(x)
x
(8.1)
s.t. f (x) = 0

with L : Rn → R and f : Rn → Rm . Then this problem is equivalent1 to

solving
min H(x, λ) := L(x) + λT f (x). (8.2)
x,λ

The function H(x, λ) is called the Lagrangian of the problem and the co-
efficients λ ∈ Rm are called Lagrangian multpliers.

Sketch of Proof: We shall split the proof into several steps:

1. The equality constraints f1 (x) = f2 (x) = · · · = fm (x) = 0 define a

feasible surface D in Rn , i.e., if x∗ is an optimum, it should belong to
D.
1
in the sense that first order necessary conditions are identical for both formulations

P. Al Hokayem & E. Gallestey 138 of 214

Nonlinear Systems and Control — Spring 2018

2. Assume that x∗ is a regular point of D, i.e., the partial derivatives

 
∂f1  ∂fm 
 ,··· , 
∂x x∗ ∂x x∗
are linearly independent.
3. Feasible directions/perturbations of the the optimal point x∗ belong to
the tangent space of the surface D at the point x∗ , denoted by Tx∗ D,
i.e., we can have the first-order approximation of the cost
∗ ∂L 
L(x) ≈ L(x ) +  (x − x∗ )
∂x x∗
perturbed using vectors d = x − x∗ ∈ Tx∗ D.
4. Necessary condition for optimality of the original problem can be stated
as
∂L 
 = 0, ∀d ∈ Tx∗ D
∂x x∗
or equivalently
∂L  ∂f1 
 ∗ d = 0, ∀d :  d = 0, ∀i = 1, · · · m
∂x x ∂x x∗
or equivalently
 
∂L  ∂f1 
 ∈ span  , ∀i = 1, · · · m
∂x x∗ ∂x x∗
or equivalently, ∃λ1 , · · · , λm ∈ R such that
∂L  ∂f1  ∂fm 
 + λ 1  + · · · + λ m  =0
∂x x∗ ∂x x∗ ∂x x∗

Now, let’s look at the necessary conditions for (8.2). Taking partial
derivaties and setting them equal to zero, we get
∂H  ∂L  ∂f 
 ∗ =0⇒  ∗ + λT 
∂x x ∂x x ∂x x∗
which is the same necessary condition as the original problem. More-
over,
∂H 
 = 0 ⇒ f (x∗ ) = 0
∂λ x∗
which is the equality constraint of the original problem.

P. Al Hokayem & E. Gallestey 139 of 214

Nonlinear Systems and Control — Spring 2018

Figure 8.2: Gradients of L and f must be colinear at the extrema

Example 8.3 Find the minimum of the function

L(x1 , x2 ) = x21 − 4x1 + x22
with the constraints
x1 + x2 = 0.
Indeed, the Hamiltonian is given by
H(x, λ) = x21 − 4x1 + x22 + λ(x1 + x2 ).
Thus, the expressions for
 
∂H  ∂H 
= 0, =0
∂x (x∗ ,λ∗ ) ∂λ (x∗ ,λ∗ )
are

∂H 
0= = 2x∗1 − 4 + λ∗
∂x1 (x∗ ,λ∗ )

∂H 
0= = 2x∗2 + λ∗
∂x2 (x∗ ,λ∗ )

∂H 
0=  = x∗1 + x∗2
∂λ (x∗ ,λ∗ )

P. Al Hokayem & E. Gallestey 140 of 214

Nonlinear Systems and Control — Spring 2018

Solving this system of linear equations we find the solution (x∗1 , x∗2 , λ∗ ) =
(1, −1, 2). This solution is only a candidate for solution. One must now test
whether these coordinates represented the solution we seek.
Example 8.4 Consider the following optimization problem
min L(x) = x21 + x22 + x23
x
s.t. f (x) = x1 x2 + x2 x3 + x3 x1 − 1 = 0
The Lagrangian is given by
H(x, λ) = x21 + x22 + x23 + λ(x1 x2 + x2 x3 + x3 x1 − 1)
The necessary conditions are given by
 ∗     ∗  
 2x1 + λx∗2 + λx∗3 2 λ λ x1 0
∂H  T
∗ ∗ ∗ 
 = 2x2 + λx1 + λx3 = λ 2 λ x2 = 0
   ∗ 
∂λ x∗
2x∗3 + λx∗1 + λx∗2 λ λ 2 x∗3 0
The
 equalities
 above can be satisfied, if x∗1 = x∗2 = x33 = 0 OR if the matrix
2 λ λ
λ 2 λ has no trivial null space, i.e., if
λ λ 2
 
2 λ λ
det λ 2 λ = 0 ⇒ 2λ3 − 6λ2 + 8 = 0
λ λ 2
which implies that the Lagrange multiplier can take values λ∗ = −1 or λ∗ = 2.
Let’s analyse
 all these cases using the final conditon of the necessary

conditions ∂H
∂λ 
= f (x∗ ) = 0.
x∗
If x∗1 = x∗2 = x33 = 0, then the condition f (x∗ ) = 0 cannot be satisifed,
therefore this is not an optimum point.
 If λ = 2,  then the null space is spanned by the following two vectors:
2 −1 −1 and 0 1 −1 , which means that we have the following two
potential solutions: (2x∗1 = −x∗2 , x∗2 = x∗3 ) OR (x∗1 = 0, x∗2 = −x∗3 ), both of
which do not yield a real-valued solution when substituted into the constraint
f (x∗ ) = 0.  
If λ = −1, then the null space is spanned by 1 1 1 , which means that
x∗1 = x∗2 = x∗3 , which
p when coupled with the constraint f (x∗ ) = 0, yields
x∗1 = x∗2 = x∗3 = 1/3 which is a candidate for an optimum. We still need
to check the sufficient conditions for this candidate solution to the problem.

P. Al Hokayem & E. Gallestey 141 of 214

Nonlinear Systems and Control — Spring 2018

Finally, note that The second-order sufficient condtion for optimality of

the problem (8.1) is given by

dT Hxx |(x∗ ,λ∗ ) d > 0, ∀d ∈ Tx∗ D

 P
m 
∂2L  2 
where Hxx |(x∗ ,λ∗ ) = ∂x2 
+ λi ∂∂xf21  or equivalently
x∗ i=1 x∗

T ∂fi 
d Hxx | (x∗ ,λ∗ ) d > 0, ∀d :  d = 0, ∀i = 1, · · · , m
∂x x∗

8.3 Pontryagin’s Minimum Principle

Consider the system
ẋ = f (x, u), x(0) = x0 (8.3)
with associated performance index
Z T
J[x0 , u(·)] = φ(x(T )) + L(x(t), u(t))dt (8.4)
0

and final state constraint

ψ(x(T )) = 0 (8.5)
The following terminology is customary:

1. J[x0 , u(·)] is called cost function.

2. φ(x(T )) is called end constraint penalty

3. L(x, u) is called running cost

4. H(x, u, λ) = L(x, u) + λf (x, u) is called the Hamiltonian.

The Optimal Control Problem is: Find the control function

u : [0, T ] 7→ Rm

whereas u is to sought in an appropriate function class, such that the per-

formance index is minimized and the final state constraint and the system
equations are satisfied.

P. Al Hokayem & E. Gallestey 142 of 214

Nonlinear Systems and Control — Spring 2018

Theorem 8.5 Solutions of the Optimal Control Problem also solve the fol-
lowing set of differential equations:
State Equation:ẋ = Hλ = f (x, u), (8.6)
∂L ∂f
Co-State Equation: − λ̇ = Hx = +λ (8.7)
∂x ∂x
∂L ∂f
Optimality Condition: 0 = Hu = +λ (8.8)
∂u ∂u
State initial condition: x(0) = x0 (8.9)
Co-state final condition: λ(T ) = (φx + ψx ν)|x(T ) (8.10)
where ν is the Lagrange multiplier corresponding to end condition given by
Equation 8.5.

Remark 8.6 The expression regarding u(.) in (8.8) is a special case of a

more general minimum condition. In general we must seek
u∗ = arg min H(x, u, λ)
for the optimal values of x(t) and λ(t). In other words, the Pontryagins Min-
imum Principle states that the Hamiltonian is minimized over all admissible
u for optimal values of the state and co-state.

Proof We use the Lagrange multipliers to eliminate the constraints. Since

the main constraints are now given by a dynamical system ẋ = f (x, u) it
is intuitive clear that they must hold for all t and thus the vector function
λ : [0, T ] → Rn is a function of time.
Using the notation H(x, u, λ) := L(x, u) + λf (x, u), the unconstrained
minimization problem can be written as
Z T
J(x, u, λ) = φ(x(T )) + ψ(x(T ))ν + [L(x, u) + λ(f (x, u) − ẋ)]dt
0
Z T
= φ(x(T )) + ψ(x(T ))ν + [H(x, u, λ) − λẋ)]dt.
0

The differentials are written as

δJ = [φx (x) + ψx (x)ν]δx)|x=x(T ) +

Z T
[Hx δx + Hu δu + Hλ δλ − λδ ẋ + δλẋ]dt + ψ(x)|x(T ) δν
0

P. Al Hokayem & E. Gallestey 143 of 214

Nonlinear Systems and Control — Spring 2018

Note that by integrating by parts:

Z T Z T
− λδ ẋ = −λδx|t=T + λδx|t=0 + λ̇δxdt
0 0

Furthermore, since x(0) = x(0) is constant, it holds that λδx|t=0 = 0. So we

can rewrite the previous expression as

δJ = [φx (x) + ψx (x)ν − λ]δx)|x=x(T ) +

Z T
[(Hx + λ̇)δx + (Hλ − ẋ)δλ + Hu δu]dt + ψ(x)|x(T ) δν
0

Now for the function u : [0, T ] → Rm to minimize the cost function, δJ must
be zero for any value of the differentials. Thus, all the expressions before the
differentials have to be zero for every t ∈ [0, T ]. This observation gives the
equations as required.

Remark 8.7 Just as in the static optimization case, where the zeros of the
derivatives represent candidates to be tested for extremum, the solutions of
the system described in Theorem 8.5 are to be seen as candidates to be the
optimal solution and their optimality must be tested for each particular case.
In other words, the Pontriaguin Maximum Principle delivers necessary, but
not sufficient, conditions for optimality.

Remark 8.8 The system of equations in Theorem 8.5 is a so called “Two

Point Boundary Value” problem. generally speaking, these problems are only
solvable by dedicated software implementing suitable numerical methods.

Remark 8.9 Special attention is needed in the case where Hu = const for all
u, in which case the optimal solution is found where the constraints are active.
This sort of solution is often called “Bang-Bang solution” to acknowledge for
the fact that they are often discontinuous and take values at the bounds.

Example 8.10 Consider the minimization problem

Z T
J[x0 , u(·)] = 0.5 u2 (t)dt
0

P. Al Hokayem & E. Gallestey 144 of 214

Nonlinear Systems and Control — Spring 2018

subject to

x˙1 = x2 , x1 (0) = x10

x˙2 = u, x2 (0) = x20

and with final constraint

ψ(x(T )) = x(T ) = 0.

Applying Theorem 8.5, we transform the system into

x˙1 = Hλ1 = x2 ; x1 (0) = x10

x˙2 = Hλ2 = u; x2 (0) = x20
λ˙1 = −Hx1 = 0; λ1 (T ) = ν1
λ˙2 = −Hx2 = −λ1 ; λ2 (T ) = ν2

where
H(x, u, λ) = 0.5u2 + λ1 x2 + λ2 u
Now we see that
Hu = u + λ2 = 0
Thus, we can solve the differential equations and see that

λ1 (t) = ν1
λ2 (t) = −ν1 (t − T ) + ν2
u(t) = ν1 (t − T ) − ν2

Placing these linear expressions in the dynamic equations for x and using
the initial conditions x(0) = x0 , we obtain a linear system of equations with
respect to (ν1 , ν2 ), which gives us the final parametrization of the control law
u. Figure 8.3 shows the result of applying this control law with T = 15.

Example 8.11 Consider the same problem as in Example 8.10, but with a
performance reflecting minimal time, i.e.
Z T
J[x0 , u(·)] = 1 dt
0

P. Al Hokayem & E. Gallestey 145 of 214

Nonlinear Systems and Control — Spring 2018

Figure 8.3: Trajectories for Minimal Energy Case. Arrival time T = 15.

and constrained input

−1 ≤ u(t) ≤ 1 ∀t ≥ 0.

Now
H(x, u, λ) = 1 + λ1 x2 + λ2 u.
We notice that Hu = λ2 , i.e. Hu does not depend on u and thus the extremum
is to be reached at the boundaries, i.e.

u∗ (t) = ±1, ∀t ≥ 0.

Using
u∗ = arg min H(u) = arg min λ2 (t)u(t)
u u

we see that
u∗ (t) = −signλ2 (t).
Figure 8.4 depicts the result of deploying this control. Note that the control
law is discontinuous. Further, we observe that the system now needs less
time to reach the origin than in the previous “Minimal Energy” example.

P. Al Hokayem & E. Gallestey 146 of 214

Nonlinear Systems and Control — Spring 2018

Figure 8.4: Trajectories for Minimal Time Case. Arrival time T ≈ 13.

Of course, this “success” is obtained at the cost of deploying more actuator

energy.

8.4 Sufficient Conditions for Optimality

The simplest sufficient condition for optimality is o course to establish unique-
ness of the solution that appears solving the first order necessary condition.
In applications it is often the case that this is possible by using application
specific knowledge.
A more generic and well understood sufficient condition follows directly
from the standard constrained optimization result and is can be stated as
follows.

Theorem 8.12 (Sufficiency Conditions I) Let us assume that (x∗ , λ∗ , u∗ )

fulfill the conditions in Theorem 8.5. Then, the condition

∂ 2H ∗ ∗ ∗
(x , u , λ ) > 0
∂u2

P. Al Hokayem & E. Gallestey 147 of 214

Nonlinear Systems and Control — Spring 2018

is sufficient for the optimality of u∗ .

Proof The results is proved by contradiction (i.e assuming that u∗ might

not be a local optimizer) and using the expression

∂ 2H ∗ ∗ ∗
H(x∗ , u, λ∗ ) ≈ H(x∗ , u∗ , λ∗ ) + (u − u∗ ) (x , u , λ )(u − u∗ )
∂u2
≥ H(x∗ , u∗ , λ∗ )

and this is a well known results from constrained optimization.

Further, useful results can be obtained using the concept of convexity.

Definition 8.13 A function f : Rn → R is called convex when it fulfills the

following property: for any given x and x∗ we have

f ((1 − t)x + tx∗ ) ≤ f (x)(1 − t) + tf (x∗ ), t ∈ [0, 1].

A smooth convex function f fulfills the following condition


∗ ∂f 
f (x) ≥ f (x ) + (x − x∗ ), ∀x, x∗ ∈ Rn .
∂x x=x∗

Before stating the next result we also note the following fact for the
Hamiltonian H = H(x, u, λ).

dH ∂H ∂H ∂u
= + .
dx ∂x ∂u ∂x
if u∗ is an optimal control policy candidate, then ∂H ∂u
= 0 and we establish
that
dH ∂H
(x, u∗ , λ) = (x, u∗ , λ).
dx ∂x

Theorem 8.14 (Sufficiency Conditions II) Let u*(t), and the correspond-
ing x∗ (t) and λ∗ (t) satisfy the minimum principle necessary condition for all
t ∈ [0, T ]. Then, u∗ is an optimal control if H(x, u∗ , λ) is convex in x and
φ(x) is convex in x.

P. Al Hokayem & E. Gallestey 148 of 214

Nonlinear Systems and Control — Spring 2018

Proof We want to prove that

J(x0 , u) − J(x0 , u∗ ) ≥ 0,

where x satisfies Equation 8.3. Now u∗ = arg min H and the convexity of H
give us

H(x, u, λ) ≥ H(x, u∗ , λ)

∗ ∗ ∂H 
≥ H(x , u , λ) + (x − x∗ )
∂x x∗ ,u∗ ,λ

Or equivalently

L(x, u) + λf (x, u) ≥ L(x∗ , u∗ ) + λf (x∗ , u∗ ) − λ̇(x − x∗ )

or

L(x, u) − L(x∗ , u∗ ) ≥ −λf (x, u) + λf (x∗ , u∗ ) − λ̇(x − x∗ )

= −λ[(ẋ − ẋ∗ ) − λ̇(x − x∗ )
d
= [−λ(x − x∗ )]
dt
Integrating, and using both the transversality condition and the fact that all
trajectories have the same initial condition x = x0 , we observe that
Z T
[L(x, u) − L(x∗ , u∗ , λ)]dt = −λ(T )(x(T ) − x∗ (T )) + λ(0))(x∗ (0) − x(0))
0 
∂φ 
= −  (x − x∗ )
∂x x∗

Then, since φ is convex, we know that


∗ ∂φ 
φ(x) − φ(x ) ≥ (x − x∗ )
∂x x=x∗

This implies

∂φ 
− (x − x∗ ) = −φ(x) + φ(x∗ ) + α, α ≥ 0.
∂x x=x∗

P. Al Hokayem & E. Gallestey 149 of 214

Nonlinear Systems and Control — Spring 2018

We now recall that

Z T
J[x0 , u(·)] = φ(x(T )) + L(x(t), u(t))dt.
0

Thus, after some manipulations, we obtain

J(x0 , u) − J(x, u∗ ) = α ≥ 0,

which proves the optimality of u∗ .

P. Al Hokayem & E. Gallestey 150 of 214

 2

We introduce the notation

h(t) = height at time t
Nonlinear Systems and Controlv(t) = velocity
— Spring 2018
m(t) = mass of spacecraft
u(t) = thrust at time t

8.5 We assume that

Exercises 0 ≤ a(t) ≤ 1
1. This
Let ussystem is modeled
consider by the following
a spacecraft equation
willing to land on the lunar surface. The
quantities h(t), v(t), m(t), and u(t)
ḣ(t) = v(t) represent the height over the sur-
face, the velocity, the mass
v̇(t) of
= the
−g +spacecraft and the vertical thrust
u(t)
m(t)
(control action), respectively.
ṁ(t) = −ku(t) is provided in the figure below.
A sketch

Fig. 1: Moon lander

The
We dynamics is described
want to minimize by of fuel used up, that is, to maximize the
the amount
amount remaining once we have landed, therefore find the optimal strategy u(t)
which minimize this term ḣ = v
! τ
u
J[u(.)]v̇== −g +
u(t)dt.
0 m
ṁ = −ku
Exercise 5
with k and g positive constants, and the control constrained to 0 ≤
[minimum time problem]
u(t) ≤ 1. the
Consider Thethelanding has to be soft, i.e. we want h(t∗ ) = 0, v(t∗ ) = 0
linear system
∗
at the landing time t . The initial conditions h(0), v(0), m(0) are fixed.
Use Pontryagin’s minimum x˙1 principle
= −x2 + to u1 find a candidate control u◦ (t)
x˙2 = x1 + u2
to minimize the fuel consumption
where the control should respect the constraint
Z t∗
J(u)
u21 +=
u22 = 1 u(t)dt .
0
we would like to reach the an arbitrary point starting from the pair (a, b) in
2. Let us consider
minimum time. a unidimensional optimal control problem with linear
dynamics and quadratics costs
Z 1 2
x u2 x2 (1)
min + dt +
|u(t)|≤10 0 2 2 2
s.t. ẋ = x + u, x(0) = 1 .

P. Al Hokayem & E. Gallestey 151 of 214

Nonlinear Systems and Control — Spring 2018

Use the Pontryagin’s minimum principle to find a candidate control

u◦ (t). Is such candidate optimal? [Hint: is the Hamiltonian convex? ]

3. This exercise is taylored to show that the Pontryagin’s minimum prin-

ciple only provides a necessary condition, in general cases when the
Hamiltonian is non convex.
Consider the following
Z 1
x2
min − dt
u(t)∈[−1,1] 0 2
s.t. ẋ = u, x(0) = 0 .

Do p◦ (t) = 0, x◦ (t) = 0 and u◦ (t) = 0 satisfy Pontryagin’s minimum

principle equations? Is u◦ (t) = 0 optimal? If not, by just looking and
the problem, can you come up with an optimal control u∗ (t)?

P. Al Hokayem & E. Gallestey 152 of 214

Chapter 9

State-Dependent Riccati
Equation

9.1 State-Independent Riccati Equation Method

9.1.1 Results on Control Design for LTI Systems
The Linear Quadratic Regulator (LQR)

ẋ = Ax + Bu (9.1)

and the performance criteria

Z ∞
J[x0 , u(·)] = [xT Qx + uT Ru]dt, . . . Q ≥ 0, R > 0 (9.2)
0

Problem: Calculate function u : [0, ∞] 7→ Rp such that J[u] is minimized.

The LQR controller has the following form

u(t) = −R−1 B T P x(t) (9.3)

where P ∈ Rn×n is given by the positive (symmetric) semi definite solution

of
0 = P A + AT P + Q − P BR−1 B T P. (9.4)
This equation is called Riccati equation. It is solvable if and only if the pair
(A, B) is controllable and (Q, A) is detectable.
Note that

153
Nonlinear Systems and Control — Spring 2018

1. LQR assumes full knowledge of the state x.

2. The pair (A, B) is given by “the process dynamics” and can not be
modified at this stage
3. The pair (Q, R) make the controller design parameters. Large Q pe-
nalizes transients of x, large R penalizes usage of control action u.

The Linear Quadratic Gaussian Regulator (LQG) In LQR we as-

sumed that the whole state is available for control at all times (see formula
for control action above). This is unrealistic as the very least there is always
measurement noise.
One possible generalization is to look at

ẋ = Ax + Bu + w (9.5)
y(t) = Cx + v (9.6)

where v, w are stochastic processes called measurement and process noise

respectively. For simplicity in the explanation, we assume these processes to
be white noise (ie zero mean, uncorrelated, Gaussian distribution).
Crucially, now only y(t) is available for control. It turns out that for
linear systems a separation principle holds
1. First, calculate xe (t) estimate the full state x(t) using the available
information
2. Secondly, apply the LQR controller, using the estimation xe (t) replac-
ing the true (but unknown!) state x(t).

Observer Design (Kalman Filter) The estimation xe (t) is calculated

by integrating in real time the following ODE

ẋe = Axe + Bu + L(y − Cxe ) (9.7)

With the following matrices calculated offline:

L = P C T Re−1 (9.8)
0 = AP + P AT − P C T Re−1 CP + Qe , P ≥ 0 (9.9)
Qe = E(wwT ) (9.10)
Re = E(vv T ) (9.11)

P. Al Hokayem & E. Gallestey 154 of 214

Nonlinear Systems and Control — Spring 2018

The Riccati equation above has its origin in the minimization of the cost
functional Z 0
J[xe (·)] = [(xe − x)(xe − x)T ]dt
−∞

9.2 State-Dependent Riccati Equation Approach

The LQR/LQG method is extremely powerful and widely used in applica-
tions, where linearisations of the nonlinear process representations are valid
over large operating areas. How can this framework be extended beyond that
obvious case?
The State-Dependent Riccati Equation (SDRE) strategy provides an ef-
fective algorithm for synthesizing nonlinear feedback controls by allowing for
nonlinearities in the system states, while offering design flexibility through
state-dependent weighting matrices.
The method entails factorization of the nonlinear dynamics into the state
vector and its product with a matrix-valued function that depends on the
state itself. In doing so, the SDRE algorithm brings the nonlinear system
to a non-unique linear structure having matrices with state-dependent co-
efficients. The method includes minimizing a nonlinear performance index
having a quadratic-like structure. An algebraic Riccati equation (ARE), as
given by the SDC matrices, is then solved on-line to give the suboptimum
control law.
The coefficients of this Riccati equation vary with the given point in state
space. The algorithm thus involves solving, at a given point in the state
space, an algebraic state-dependent Riccati equation. The non-uniqueness
of the factorization creates extra degrees of freedom, which can be used to
enhance controller performance.

Extended Linearization of a Nonlinear System Consider the deter-

ministic, infinite-horizon nonlinear optimal regulation (stabilization) prob-
lem, where the system is full- state observable, autonomous, nonlinear in the
state, and affine in the input, represented in the form

ẋ = f (x) + B(x)u(t) (9.12)

x(0) = x0 (9.13)

P. Al Hokayem & E. Gallestey 155 of 214

Nonlinear Systems and Control — Spring 2018

where x ∈ Rn is the state vector, u ∈ Rm is the input vector, with smooth

functions f and B of appropriate domains such that
• B(x) 6= 0 for all x
• f (0) = 0.
Extended linearisation is the process of factorizing a nonlinear system into a
linear-like structure which contains SDC matrices. Under the fairly generic
assumptions
f (0) = 0, f ∈ C1 (Rn )
a continuous nonlinear matrix-valued function

A : Rn 7→ Rn×n

always exists such that

f (x) = A(x)x (9.14)
where A : Rn 7→ Rn×n is found by algebraic factorization and is clearly
non-unique when n > 1.
After extended linearization of the input-affine nonlinear system (9.12)
becomes

ẋ = A(x)x + B(x)u(t)
x(0) = x0 .

which has a linear structure with state dependent matrices A(x) and B(x).
• Note that these parameterizations are not unique for n > 1. For in-
stance, if A(x)x = f (x), then (A(x) + E(x))x = f (x) for any E(x) such
that E(x)x = 0.
• We also note that given A1 (x)x = f (x) and A2 (x)x = f (x), then for
any α ∈ R

A(x, α)x = [αA1 (x) + (1 − α)A2 (x)]x = αf (x) + (1 − α)f (x) = f (x)

is also a valid parametrization.

Remark 9.1 In general, one needs to answer the question about the optimal
choice of α for the given application at hand. In principle, α can also be
made a function of x, delivering additional degree of freedom for design.

P. Al Hokayem & E. Gallestey 156 of 214

Nonlinear Systems and Control — Spring 2018

Pointwise Hurwitz A Matrix does not imply stability We could

naively think that if we find a controller u = K(x) such that the closed
loop matrix
Acl (x) = A(x) − B(x)K(x)
is point-wise Hurtwitz then we would have a good design for stabilization of
the nonlinear system given by (9.15). However, this condition is not sufficient
for closed loop stability as the following example shows.
Example 9.2 Let us consider the system in form of (9.15)
 p 
  − √1 2 2 1 + x21 + x22 )  
x˙1 1+ x +x x
= 1 2  1
,  ≥ 0.
x˙2 0 − √1 2 2 x2
1+ x1 +x2 )

We note that the eigenvalues of the matrix, given simply by the matrix
components a11 and a22 are negative for any value of x. Moreover, we also
notice that for  = 0 the eigenvalues are both equal to −1, with the effect that
the state x2 will converge to zero rapidly. This will make the term a12 also
small, and eventually lead to x1 to also converge to zero, see Figure 9.1.
However, if  = 1 then at least for some initial conditions, x1 may grow
fast and make the term a22 converge to zero, which will lead to x2 becoming
a constant. Since also a11 = a22 will be small, the term a12 will be dominant
in the dynamics of x1 and will lead x1 to grow to infinite, see Figure 9.2.

SDRE Method Formulation For the system

ẋ = A(x)x + B(x)u(t) (9.15)

x(0) = x0 . (9.16)

We shall look at minimization of the infinite-time performance criterion

Z ∞
J[x0 , u(·)] = [xT Q(x)x + uT R(x)u]dt, . . . Q(x) ≥ 0, R(x) > 0. (9.17)
0

Note, that the state and input weighting matrices are also assumed state
dependent.
Under the specified conditions, we apply point-wise the LQR method for
(A(x), B(x)),(Q(x), R(x)) generating a control law

u(t) = −K(x)x = R(x)−1 B(xT )P (x)x(t), K : Rn 7→ Rp×n

P. Al Hokayem & E. Gallestey 157 of 214

Nonlinear Systems and Control — Spring 2018

Figure 9.1: System behavior with  = 0

where P : Rn 7→ Rn×n satisfies

P (x)A(x) + A(x)T P (x) − P (x)B(x)R(x)−1 B(x)T P (x) + Q(x) = 0.

By applying this method one expects to obtain the good properties of the
LQR control design, namely, that

• the control law minimizes the cost given by (9.17)

• regulates the system to the origin limt→∞ x(t) = 0.

Main Stability Results Below we establish how close we are to that

”wish”. Indeed, we have the following theorem.

Theorem 9.3 (Mracek & Cloutier, 1998) Let us assume the following con-
ditions hold.

1. The matrix valued functions A(x), B(x), Q(x), R(x) ∈ C 1 (Rn ).

P. Al Hokayem & E. Gallestey 158 of 214

Nonlinear Systems and Control — Spring 2018

Figure 9.2: System behavior with  = 1

1
2. The pairs (A(x), B(x)) and (A(x), Q 2 (x)) are pointwise stabilizable,
respectively, detectable, state dependent parameterizations of the non-
linear system for all x ∈ Rn .

Then, the nonlinear multivariable system given by (9.15) is rendered locally

asymptotically stable by the control law

u(x) = −K(x)x = R−1 (x)B T (x)P (x)x, K : Rn 7→ Rp×n (9.18)

where P (x) ∈ Rn×n is the unique, symmetric, positive-definite solution of the

algebraic State-Dependent Riccati Equation

P (x)A(x) + AT (x)P (x) − P (x)B(x)R−1 (x)B T (x)P (x) + Q(x) = 0. (9.19)

Proof: The dynamics is given by the pointwise Hurwitz matrix

ẋ = [A(x) − B(x)R−1 (x)B T (x)P (x)]x = Acl (x)x.

P. Al Hokayem & E. Gallestey 159 of 214

Nonlinear Systems and Control — Spring 2018

Under the assumptions of the theorem one can show that

Acl (x) = Acl (0) + φ(x), lim φ(x) = 0.

x→0

which means that the linear term is dominant near the origin. 

Remark 9.4 Note that global stability has not been established, this is a local
result. In general, as we saw in Example 9.2, even when

Acl (x) = A(x) − B(x)K(x)

is Hurtwitz for all x, global stability can not be guaranteed.

Remark 9.5 One can prove though that if Acl (x) Hurtwitz and symmetric
for all x, then global stability holds. The proof is simply obtained by showing
that under these conditions V (x) = xT x is a Lyapunov function for system
(9.15).

Theorem 9.6 Under conditions of Theorem 9.3, the SDRE nonlinear feed-
back solution and its associated state and costate trajectories satisfy the first
necessary condition for optimality of the nonlinear optimal regulator problem

u(x) = arg min H(x, λ, u), λ = P (x)x

where

H(x, λ, u) = 0.5xT Q(x)x + 0.5uT R(x)u + λT [A(x)x + B(x)u]

Proof: The Hamiltonian of this problem is given by

H(x, λ, u) = 0.5xT Q(x)x + 0.5uT R(x)u + λT [A(x)x + B(x)u].

Thus, it is clear that

Hu = R(x)u + B T (x)λ
This implies that for any choice of λ

u(x) = −R−1 (x)B T (x)λ =⇒ Hu = 0.

In particular, the choice λ = P (x)x renders a controller given by (9.18).



P. Al Hokayem & E. Gallestey 160 of 214

Nonlinear Systems and Control — Spring 2018

When using the SDRE method one observes that as the state converges
to zero, the solution also converges to the optimal solution given by the Pon-
triaguin Maximum Principle. This observation is supported by the following
result.
Theorem 9.7 Assume that all state dependent matrices are bounded func-
tions along the trajectories. Then, under the conditions of Theorem 9.3 the
Pontriaguin second necessary condition for optimality.

λ̇ = −Hx (x, λ, u)

is satisfied approximately by
λ = P (x)x
at a quadratic rate along the trajectory. Here P (x) denotes the solution of
Equation 9.19.
Proof: The proof of this statement regarding λ(·) is quite technical and is
obtained as follows. First, we observe

λ = P (x)x =⇒ λ̇ = Ṗ (x)x + P (x)ẋ.

Thus, we need it to hold:

−Hx (x, λ, u) ≈ Ṗ (x)x + P (x)ẋ.

Equating these expressions, grouping linear and quadratic terms and using
(9.19), one obtains that the residuals are all quadratic functions of x. The-
orem 9.3 proves that these quadratic terms in x decay quadratically as x
converges to the origin.


Example 9.8 . Steer to x = (d, 0) the following system.

ẋ1 = x2
ẋ2 = −a sin(x1 ) − bx2 + cu(t).

Indeed,
   
0 1 0
A(x) = B(x) =
−a sin(x1 − d)/(x1 − d) −b c

P. Al Hokayem & E. Gallestey 161 of 214

Nonlinear Systems and Control — Spring 2018

Figure 9.3: SDRE versus LQR

We choose  
1 + (x1 − d)2 0
Q(x) = , R = 1.
0 1 + x22
The choice of Q(x) ensures larger control actions for large deviations from
the equilibrium.
The magenta trajectory in Figure 9.3 is obtained using LQR on the stan-
dard linearization of the original system with
 
1 0
Q(x) = , R = 1.
0 1

Note how the state is brought faster to the equilibrium in the SDRE case.

P. Al Hokayem & E. Gallestey 162 of 214

Nonlinear Systems and Control — Spring 2018

Output Feedback and SDRE Approach One can extend these same
ideas to the case where the state is not available for control. As expected
the method reduces to use the matrix functions A(·), B(·), C(·) to calculate
pointwise the corresponding Kalman gain L described in (9.8). Clearly, since
x is not available, one must evaluate the matrix functions at the current
estimate xe , which is calculated following (9.7).
In Figure 9.3 the line in black depicts the result of this strategy for the
previous example, but using only the position x1 for calculation the control.
One observes that the controller is able to bring the system to the desired
position.
We also observe though that the transient is now much longer. The
length of this transient being dependent, among other issues, on the quality
of the initial estimate and on the measurement noise. Figure 9.4 shows the
transients of x and xe . Note that the controller has only xe to plan its actions.

Figure 9.4: Estimate and true state using the SDRE approach

P. Al Hokayem & E. Gallestey 163 of 214

Nonlinear Systems and Control — Spring 2018

9.2.1 H∞ control for nonlinear systems

H∞ control is a systematic method for generating robust controllers for lin-
ear systems that implies designing a controller that minimizes the effect on
a performance related variable z = Cz x + Du u of the worst possible distur-
bance w, see Figure 9.5. In complete analogy to LQR and LQG controllers,
the SDRE method is similarly deployed to implement H∞ controllers for
nonlinear systems.
H∞ control creates a beautiful mathematical apparat related to game
theory that unfortunately we will not touch further due to time and space
constraints. The interested reader can refer to [1, 3, 5] for further information.

w(t) z(t)
P
u(t) y(t)

Figure 9.5: H∞ control components

P. Al Hokayem & E. Gallestey 164 of 214

Nonlinear Systems and Control — Spring 2018

9.3 Exercises
1. Verify that the Hamiltonian-Jacobi equation for a linear system with
quadratic costs reduces to the Riccati equation.

2. Consider the controlled dynamical system

ẋ = f (x) + b(x)u ,

with x, u ∈ R, f (x) = x − x3 and b(x) = 1 .

(a) Design, if possible, a SDRE controller us (x) to minimize

Z ∞
J= x2 (s) + u2 (s)ds
0

(b) Study the stability of the closed loop system. Is there any a priori
guarantee you can exploit?
(c) Recall that the previous optimal control problem can be solved
looking at the infinite horizon Hamilton-Jacobi equation. Write
the equation and find an explicit solution uo (x) for the controller.
(d) Are us (x) and uo (x) the same or different? Did you expect this?
Motivate your answer.

3. Given the following system

ẋ = x2 + u

(a) Determine its unique factorization.

(b) RDesign, if possible, a SDRE controller us (x) to minimize J =
∞ 2
0
x (s) + u2 (s)ds, with Q(x) = 1 and R(x) = 1.
(c) Check that
λ̇ = Hu (x, λ, u)
is satisfied approximately by λ = P (x)x.

P. Al Hokayem & E. Gallestey 165 of 214

Chapter 10

Nonlinear Model Predictive

Control

10.1 Introduction
In the first section we looked at the following simple system.
ẋ1 = x2
ẋ2 = u
y = x1
The goal is keep the output at a given setpoint ysp using the control action
u(·), which is bounded −M ≤ u ≤ M . Figure 10.1 shows the result of
designing a control using LQR plus saturation at M , i.e.
u = sat(−Kx, M )
We observe that as the the value of M decreases this simple strategy fails
to keep the system stable. The reason is that although the controller does
know about the system dynamics, the controller design ignores the constraint,
being over optimistic about its ability to slow down the system once the speed
is high.
Intuitively, it is clear that better results would be achieved if we made the
controller see not only the dynamics but also the barrier. Model Predictive
Control (MPC) offers that framework. Indeed, Figure 10.2 shows the result
of applying MPC to this problem. We observe that stability is not lost
as M decreases. Model Predictive Control has largely conquered industrial
applications by means of being both systematic and intuitive.

166
Nonlinear Systems and Control — Spring 2018

Figure 10.1: LQR plus saturation for double integrator

10.2 Main Theoretical Elements

Model predictive control is nowadays probably the most popular method for
handling disturbances and forecast changes. The main ingredients of the
method are

1. Plant model
ẋ = f (t, x, u), x(0) = x0

2. Constraints
g(t, x, u) = 0
gi (t, x, u) ≤ 0

3. Objective function
Z t+T
J[x0 , u(·)] = f0 (τ, x(τ ), u(τ ))dτ + F (x(t + T ), u(t + T ))
t

Resembling a chess game as played by a computer algorithm, the method

works in iterations comprising the following steps.

P. Al Hokayem & E. Gallestey 167 of 214

Nonlinear Systems and Control — Spring 2018

Figure 10.2: MPC for double integrator with constraints

1. Evaluate position (=measurement) and estimate system state

2. Calculate the effect on the plant of (possible) sequences of actuator

moves

3. Select the best control sequence (mathematical algorithm, optimiza-

tion)

4. Implement the first move (new actuator set-point)

5. Restart after the opponents move (plant/process reaction to actuator

move)

Remark 10.1

1. The model is used to predict the system behavior into the future.

2. The method requires solution of optimization problem at every sampling

time.

P. Al Hokayem & E. Gallestey 168 of 214

Nonlinear Systems and Control — Spring 2018

Figure 10.3: Model Predictive Control Trajectories

3. Additional constraints on the actuators can be added. For instance,

that the actuators are to remain constant during the last few steps.

4. Normally linear or quadratic cost functions are used. These functions

represent trade off among deviations from setpoints, actuator action
costs, economical considerations, etc.

5. For nonlinear systems, one has to deal with risks like loss of convexity,
local minima, increased computational time, etc. Still, the optimization
problems to be solved online are highly sparse. This allows for efficiency
gains of several orders of magnitudes.

6. Useful functionality is added when mathematical models in the “Mixed

Logical Dynamical Systems” framework are used. In this case, logi-
cal constraints and states can be included in the mathematical model,
fact that is extremely useful when dealing with plant wide planning and
scheduling problems.

10.3 Case Study: Tank Level Reference Track-

ing

P. Al Hokayem & E. Gallestey 169 of 214

Nonlinear Systems and Control — Spring 2018

Figure 10.4: Tank Schema

Problem Statement
The process consists of a tank with a measurable inflow disturbance and a
controllable outflow in form of a pump. Constraints are set on the tank level
and the pump capacity. A level reference target is given.

Model Variables

1. Inputs of the model are the known but not controllable tank inflow and
the controlled valve opening variation.

2. States are the tank level and the current valve opening:

3. Outputs are the tank level, and the outflow

Model Equations

P. Al Hokayem & E. Gallestey 170 of 214

Nonlinear Systems and Control — Spring 2018

State dynamics
volume(t + 1) = volume(t) + fin (t) − fout (t)

fout (t) = α level(t) u(t)

level(t) = volume(t)/area

u(t + 1) = u(t) + du(t)

Outputs
ylevel (t) = level(t)
yout (t) = fout (t)
Inputs
• fin : inflow, measured but not controlled
• u: valve opening, controlled via its increment du(t) = u(t + 1) − u(t).
Cost Function
T
X
J[x0 , u(·)] = q [ylevel (t) − yref (t)]2 + r du(t)2
t=0

Problem Constraints
1. Level within min/max
2. Outflow within min/max
3. Max valve variation per step

Model Representation Industry has created dedicated software tools

aimed to facilitate the tasks of modelling and control design. As a rule,
the engineer is provided tools for graphical representation of the model. Fig-
ure 10.5 shows the tank model equations depicted in a typical industrial
software.
There are also standard displays for fast visualization of the plant dy-
namics. In Figure 10.6 we see such standard representation. The graphic
is a grid of graphs, with as many rows as outputs and as many columns as
inputs. Each column represents the response of each output to a step in a
given input.

P. Al Hokayem & E. Gallestey 171 of 214

Nonlinear Systems and Control — Spring 2018

Figure 10.5: Tank Model in Graphical Package

Results Discussion
Figure 10.7 shows optimization results obtained with a receding horizon of
T=20 steps. Note how the Model Predictive Controller is able to nicely bring
the level to the desired setpoint.
On the other hand, Figure 10.8 shows the same optimization obtained with
a receding horizon of T = 2 steps. In this case, the performance considerably
deteriorates: stability is lost!
In the last example we look at the problem is controlling 3 interconnected
tanks. The problem to solve is the same as before, namely, to keep the tank
levels at given setpoints. The problem is now not just larger, but also more
complicated due to the interaction among the tanks. We design 2 controllers,
one that knows about the true plant dynamics, and one that treats the tanks
as if they were independent.
Figure 10.10 shows the step response of this model. Note how the effect
of each actuators propagates in the system from one tank to the next.
In Figure 10.11 we represent the closed loop responses of 2 different con-
trollers:

• one that knows about the interconnection among the tanks and can
better predict the system behavior and

P. Al Hokayem & E. Gallestey 172 of 214

Nonlinear Systems and Control — Spring 2018

Figure 10.6: Tank Model Step Response

• one that treats each tanks as independent entities, where the valve is
used to control the level.

As expected we observe that the multivariable controller is able to solve

the problem in a much more efficient fashion than the ”single input single
output” one, showing the benefits of the model predictive control approach
over the idea of cascaded SISO controllers.

10.4 Chronology of Model Predictive Control

Below we find some major milestones in the journey leading to the current
MPC approach:

1. 1970’s: Step response models, quadratic cost function, ad hoc treat-

ment of constraints

2. 1980’s: linear state space models, quadratic cost function, linear con-
straints on inputs and output

P. Al Hokayem & E. Gallestey 173 of 214

Nonlinear Systems and Control — Spring 2018

Figure 10.7: Tank Trajectories for T=20

3. 1990’s: constraint handling: hard, soft, ranked

4. 2000’s: full blown nonlinear MPC

10.5 Stability of Model Predictive Controllers

When obtaining stability results for MPC based controllers, one or several
of the following assumptions are made
1. Terminal equality constraints
2. Terminal cost function
3. Terminal constraint set

P. Al Hokayem & E. Gallestey 174 of 214

Nonlinear Systems and Control — Spring 2018

Figure 10.8: Tank Example Trajectories for T=2

4. Dual mode control (infinite horizon): begin with NMPC with a termi-
nal constraint set, switch then to a stabilizing linear controller when
the region of attraction of the linear controller is reached.

In all these cases, the idea of the proofs is to convert the problem cost function
into a Lyapunov function for the closed loop system.
Let us consider at least one case in details. For that, we introduce the
MPC problem for discrete time systems. Note that in practice, this is the
form that is actually used.
Consider the time invariant system

1. Plant model

x(k + 1) = f (x(k), u(k)), x(0) = x0

P. Al Hokayem & E. Gallestey 175 of 214

Nonlinear Systems and Control — Spring 2018

Figure 10.9: Three Interconnected Tanks

2. Constraints
g(k, x, u) = 0, k = 0 : ∞
gi (k, x, u) ≤ 0, k = 0 : ∞

3. Objective function
k+N
X
J[x0 , u(·)][= L(x(l), u(l))
l=k

The optimal control is a function

u? (·) = arg min J[x0 , u(·)], u? (l), l =k :k+N

Theorem 10.2 Consider an MPC algorithm for the discrete time plant,
where x = 0 is an equilibrium point for u = 0, i.e. f (0, 0) = 0.
Let us assume that

P. Al Hokayem & E. Gallestey 176 of 214

Nonlinear Systems and Control — Spring 2018

Figure 10.10: Step Response in the Interconnected Tanks Case

• The problem contains a terminal constraint x(k + N ) = 0

• The function L in the cost function is positive definite in both argu-

ments.

Then, if the optimization problem is feasible at time k, then the coordinate

origin is a stable equilibrium point.

Proof. We use the Lyapunov result on stability of discrete time systems

introduced in the Lyapunov stability lecture. Indeed, consider the function

V (x) = J ? (x),

where J ? denotes the performance index evaluated at the optimal trajectory.

We note that:

P. Al Hokayem & E. Gallestey 177 of 214

Nonlinear Systems and Control — Spring 2018

• V (0) = 0

• V (x) is positive definite.

• V (x(k + 1)) − V (x(k)) < 0. The later is seen by noting the following
argument. Let
u?k (l), l = k : k + N
be the optimal control sequence at time k. Then, at time k + 1, it is
clear that the control sequence u(l), l = k + 1 : N + 1, given by

u(l) = u?k (l), l =k+1:N

u(N + 1) = 0

generates a feasible albeit suboptimal trajectory for the plant. Then,

we observe

V (x(k+1))−V (x(k)) < J(x(k+1), u(·))−V (x(k)) = −L(x(k), u? (k)) < 0

which proves the theorem.

Remark 10.3 Stability can be lost when receding horizon is too short, see
Figure 10.12.

Remark 10.4 Stability can also be lost when the full state is not available
for control and an observer must be used. More on that topic in the Observers
Lecture.

P. Al Hokayem & E. Gallestey 178 of 214

Nonlinear Systems and Control — Spring 2018

10.6 Exercises
1. Let us consider the following standard MPC problem :
P −1 T

min J = N i=0 xi Qxi + uTi Rui + xTN P xN
s.t. xk+1 = Axk + Buk , x0 is given
kuk k∞ ≤ γ, ∀k ∈ [0, . . . , N − 1].

(a) Rewrite the cost, dynamics and input constraints as a convex

quadratic program in the variable U = [uT0 ; . . . ; uTN −1 ].
(b) Consider A = −1, B = 1, x0 = 10, Q = 1, R = 1, P = 2, N = 2
and γ = 6. Compute analytically the solution of the quadratic
program.

2. Consider the nonlinear system

xk+1 = −x3k + u.

Design a nonlinear MPC for such a system in Matlab. Close the loop
and, starting from an initial point x0 = 2, plot the state evolution in
time.
Hint: Use the Matlab function fmincon to solve the nonlinear opti-
mization problem

P. Al Hokayem & E. Gallestey 179 of 214

Nonlinear Systems and Control — Spring 2018

Figure 10.11: MPC and SISO Responses in the Interconnected Tanks Case

P. Al Hokayem & E. Gallestey 180 of 214

Nonlinear Systems and Control — Spring 2018

Figure 10.12: Double Integrator looses stability lost for short horizon

P. Al Hokayem & E. Gallestey 181 of 214

Chapter 11

State Estimation and Observers

In practice no perfect observation of the system state is available, either be-

cause it is costly, technically unfeasible or because the measurements quality
is low. In this case state feedback control laws,
u(t) = u(x(t)), t≥0
as derived in previous lectures is often impractical. There is a need for a
systematic approach for the evaluation or estimation of the system state
using the information available.
One natural approach is to compute an estimate x̂ of the state x and
apply the feedback:
u(t) = u(x̂(t)), t ≥ 0
The idea that a stabilizing controller can consist of a state estimator plus (es-
timated) state feedback is called the separation principle. For linear systems
this is a valid approach. Indeed, given a linear time invariant system
ẋ = Ax + Bu
y = Cx
Consider the observer
x̂˙ = Ax̂ + Bu + L (y − C x̂)
= (A − LC) x̂ + Bu + Ly
Let e = x − x̃, then
ė = (Ax + Bu) − (A − LC) x̃ − Bu − LCx
= (A − LC) e

182
Nonlinear Systems and Control — Spring 2018

Let
u = −K x̃ = −K(x − e)
Then     
ẋ A − BK BK x
=
ė 0 A − LC e
Thus if A − BK and A − LC are stable matrices, the resulting closed loop
system is also stable.
Unfortunately, for nonlinear systems this approach does not work in gen-
eral. The problem is that generally speaking it is not possible to estimate
the error dynamics – i.e. the dynamics of the difference between the actual
state and the estimated state.
There are several approaches to state estimation that may be applied:
1. Extended Kalman Filter (EKF): Extension of linear Kalman Flter
2. Recursive Prediction Error (RPE): Filter based on the sensitivity equa-
tion
3. Unscented Kalman Filter (UKF): Mix of Monte-Carlo with Kalman
Filter
4. Moving Horizon Estimation (MHE)
5. High Gain Observers
In the sequel we present techniques 1, 4 and 5 for estimating the state of
a nonlinear system.

11.1 Least Squares Estimation of Constant

Vectors
We consider the process model
y = Cx + w, x ∈ Rn , y, w ∈ Rp , n ≥ p.
with w denotes white noise.
The goal is to compute the best estimate x̂ of x using proces measurements
y. Given x̂ we can give an estimate ŷ of the output by
ŷ = C x̂

P. Al Hokayem & E. Gallestey 183 of 214

Nonlinear Systems and Control — Spring 2018

Then we define the residual as

y = ŷ − y
We want to obtain the optimal estimate in the sense that the quadratic
function
J[x̂] = 0.5Ty y
It is easily derived that this optimal estimate is given by
x̂ = (C T C)−1 C T y
Where the matrix inversion is made in the sense of the so called “pseudo
inverse”. Averaging effect takes place when the dimension of the output is
larger than the dimension of the state.
Example 11.1 Consider a process modelled by the simple linear relationship
 
−5 1
y = Cx + w; C =
0 1
The magnitude y is measured directly and is affected by white noise w. The
actual value of x and needs to be reconstructed from the data series
Y = [y1 ; y2 , . . . , yN ] = [Cx + w1 ; Cx2 + w2 ; . . . ; Cx + wN ]
The optimal estimate x̂ is then given by
x̂ = KY, K = (H T H)−1 H T ; H = [C; C; . . . ; C];
Figure 11.1 shows how the estimate x̂ approaches the true value x = [5; 5] as
N , the number of measurements, grows.
Our second example shows how to use the results above in the context of
dynamical system parameter estimation.
Example 11.2 Consider the equations of a simple heat exchanger. Steam at
temperature Tsi enters the system and passes its energy to a gas that enters
the system at temperature Tgi . The fluids are separated by a metal inter-
face, which has temperature Tm . At the output, the steam and the gas have
temperatures
Ṫg = αg (Tm − Tg ) + f g(Tgi − Tg ) (11.1)
Ṫm = αg (Tg − Tm ) + αs (Ts − Tm ) (11.2)
Ṫs = αs (Tm − Ts ) + f s(Tsi − Ts ) (11.3)

P. Al Hokayem & E. Gallestey 184 of 214

Nonlinear Systems and Control — Spring 2018

Figure 11.1: Static Estimation. Note how accuracy improves as N grows

Our task is to estimate the values of αg and αs from the measurements

Tgi ,Tg ,Tsi ,Ts and Tm .
We proceed by realizing that for any given point in time we can create a
linear relationship
y = Cx + w
where  
Ṫg − f g(Tgi − Tg )
y= Ṫm 
Ṫs − f s(Tsi − Ts )
and  
(Tm − Tg ) 0
C =  (Tg − Tm ) (Ts − Tm ) 
0 (Tm − Tm )
 
αg
x=
αs

P. Al Hokayem & E. Gallestey 185 of 214

Nonlinear Systems and Control — Spring 2018

Figure 11.2: Parametric Static Estimation. Note how accuracy improves as

N grows

Note that the derivatives Ṫg , Ṫm , Ṫs in y need to be reconstructed from the
temperature measurements. Due to measurement noise, one must expect that
these “reconstructed” derivatives are noisy measurements of the true deriva-
tives.
Figure 11.2 shows how the estimates αge , αse approach the true values
[3; 1] as N , the number of measurements, grows.

Remark 11.3 The static method is designed to introduce averaging effects

over the whole data set y. Thus, in practice one needs methods to manage
which data is presented to the algorithm so that results remain meaningful.
For instance, if the process has large steady state phases interrupted by periods

P. Al Hokayem & E. Gallestey 186 of 214

Nonlinear Systems and Control — Spring 2018

of transient behavior, then one must make sure that the steady state phases
are not presented to the algorithm in the same data set because otherwise the
result will be the average of two different steady state modes, which is not
what is desired.
For example, the results in Figure 11.2 were obtained by resetting the time
arrays after the change of parameters, i.e. at times t = 50, 100. Without this
manipulation the observer does not deliver meaningful results.

11.2 Weighted Least Squares Estimator

The measurement error statistical properties can change from one point to
another thus it makes sense to give different weight to different points while
seeking the optimum. More importantly , as discussed in the previous section,
when designing an observer there is always need for methods to fine tune the
significance of different values in the method trade off’s between model and
noise. For instance, the designer may want to make sure that old values are
considered less important than recent ones.
If the weights are distributed according to the equation

J[x̂] = 0.5Ty y , y = N −1 (ŷ − y).

Making the derivative of this function equal to zero, we derive the following
formula for optimal estimate

x̂ = (C T S −1 C)−1 C T S −1 y; S = N −T N −1 .

This formula is used exactly as in the previous section. Care is of course

needed in the choice of the weights N .

11.3 Propagation of the State Estimate and

its Uncertainty
The state of a dynamic system changes with time and with its statistical
properties. State estimators must take this fact into account. If the initial
conditions, the inputs and the dynamics were perfectly known then it would
be enough to integrate forward the system equations to have a good estimate
of the system state. Unfortunately, this is never the case and uncertainty will

P. Al Hokayem & E. Gallestey 187 of 214

Nonlinear Systems and Control — Spring 2018

always play a role. In this section we must learn to propagate the statistical
properties of the state, so that we can use them when calculating the optimal
estimator.
Let us assume we have a linear dynamical system

ẋ = Ax + Bu + Lw

Let us replace deterministic values by the “means”. Then

E[ẋ(t)] = AE[x(t)] + BE[u(t)] + LE[w(t)]

Via simple manipulations this implies

ṁ(t) = Am(t) + Bu(t)

Ṗ (t) = AP + P AT + LQLT .

where

1. m(t) = E[x()]

2. P (t) = E[(x(t) − m(t))T (x(t) − m(t))]

3. Q is covariance of w.

Note how the uncertainty (covariance) of the mean of the state will always
grow due to the driving force LQLT .

11.4 Kalman Filter

A recursive optimal filter propagates the conditional probability density func-
tion from one sampling time to the next, incorporating measurements and
statistical of the measurements in the estimate calculation.
The Kalman Filter consists of the following steps

1. State Estimate Propagation

2. State Covariance Propagation

3. Filter Gain Calculation

4. State Estimate Update using the newest measurements

P. Al Hokayem & E. Gallestey 188 of 214

Nonlinear Systems and Control — Spring 2018

5. State Covariance Update using the newest measurements

Steps 1 and 2 were considered before. Step 3 (calculation of Filter gain)
is made using ideas of the first part of this lecture namely weighted least
squares. The result is
K = Pk C T R−1 .
where R is the covariance of the measurement noise. The formula shows a
tradeoff between the measurement noise statistics R and the quality of our
estimation as given by P .
The Filter gain K is used in the Steps 4 and 5 as follows.
1. x̂k = xk + K(y − C ŷ)

2. Pk = [Pk−1 + Ck Rk−1 Ck ]−1 .

11.4.1 Extended Kalman Filter

The system is now nonlinear and modifications to the schema above are
needed. They are
• Steps 1 is made with the true nonlinear equations

• Steps 2 to 5 are made with linearization of the linear system at the

current estimate x̂.
A summary of the formulae is given in Figure 11.3.
The same formalism can be used to estimate parameters. Indeed, one can
augment the natural system states with states representing the parameters.
The dynamic equation for the parameter vector is chosen to be zero, i.e.
ṗ = 0 + wp , where wp has covariance Qpp . Substituting these expressions in
the standard EKF formulate produces the algorithm represented in Figure
11.4.

Example 11.4 We consider the same heat exchanger, but now we are pre-
pared to confront the realistic case where all measurements are noisy and thus
calculating the derivatives accurately is not realistic. Also, we should drop the
assumption that we know when the heat exchanger changes the operational
modes. Under these conditions, the static estimation of parameters can not
be used because both we would not have the elements to construct the matrix
C used in Example 2.

P. Al Hokayem & E. Gallestey 189 of 214

Nonlinear Systems and Control — Spring 2018

Formulae Meaning
At time tk , given x̂(tk ), y(tk ), P (tk ), Q, R Data at each step
for τ ∈ [tk , t−
k+1 ] Step forward in time for es-
timates and covariances
x̂˙ = f (x̂)
∂f ∂f T
Ṗ = (x̂(tk ))P + P (x̂(tk ))
∂x ∂x

Estimation error
e(tk+1 ) = y(tk+1 ) − h(x̂(t−
k+1 ))

Estimation error covari-

ance
∂h ∂h T
A(tk+1 ) = R + (x(t−
k+1 ))P (t− −
k+1 ∂x (x(tk+1 ))
)
∂x

Kalman gain

∂h T
K(tk+1 ) = P (t−
k+1 ) (x(t−
k+1 ))[A(tk+1 )]
−1
∂x

˙ k+1 ) = x̂(t− ) + K(tk+1 )e(tk+1 )

x̂(t k+1
Final update
P (tk+1 ) = P (t− T
k+1 ) − K(tk+1 )A(tk+1 )K (tk+1 )

Figure 11.3: Extended Kalman Filter Formulae

In this more realistic case we resort to use the Extended Klaman Filter,
whereas we will estimate the states Tg ,Tm ,Ts and the parameters αg and αs .
The equations remain the same, i.e. Equations 11.1. The important non
trivial step is to create dynamic equations for the parameters. In this case it
is enough

ṗg = 0
ṗs = 0

and  
c1 0
Qpp = .
0 c2

P. Al Hokayem & E. Gallestey 190 of 214

Nonlinear Systems and Control — Spring 2018

ŷ(tk+1 ) = h(x̂(t− −
k+1 ), u(tk+1 ), p̂(tk+1 ))
ẋ = f (x, u, p) + w1 e(tk+1 ) = y(tk+1 ) − ŷ(tk+1 )
ṗ = 0 + w2
y = h(x, p) + η ∂f ∂f T ∂f T ∂f T
Q = E[w1 w1T ] Ṗxx = P xx + P xx + P + P xp +Q
∂xT ∂x ∂pT xp ∂p
Qpp = E[w2 w2T ] ∂f ∂f
Ṗxp = T
Pxp + T Ppp
R = E[ηη T ] ∂x ∂p
Ṗpp = Qpp

T T
∂h ∂h
K(tk+1 ) = (Pxx (t−
k+1 ) + Pxp (t−
k+1 ) )A−1 (tk+1 )
∂x ∂p
T
T − ∂h ∂h T −1
L(tk+1 ) = (Pxp (tk+1 ) + Ppp (t−
k+1 ) )A (tk+1 )
∂x ∂p

x̂(t+ −
k+1 ) = x̂(tk+1 ) + K(tk+1 )e(tk+1 )
p̂(tk+1 ) = p̂(tk ) + L(tk+1 )e(tk+1 )

Figure 11.4: Augmented State Extended Kalman Filter Formulae

Qxp = 02×3 .
Note that now there is no need for manipulating the arrays: the algorithm
is recursive and reacts correctly to operational mode changes.

11.4.2 Moving Horizon Estimation

In analogy to Model Predictive Control, one can think of the observation
problem as one of optimization under a finite horizon. Variables for opti-
mization would be in this case the system states, and parameters if parame-
ter estimation is also required. This approach has the advantage that Model
Predictive Control formalism can be reused to deal with this task.
In concrete terms, the ”attack plan” is to collect

1. the history of actuators moves and measurements uM (τ ), yM (τ ), τ ∈

[t − M : t]

P. Al Hokayem & E. Gallestey 191 of 214

Nonlinear Systems and Control — Spring 2018

Figure 11.5: Extended Kalman Filter Results for Heat Exchanger

2. the plant model, including the constraints

with which the observer calculates the trajectory of states x(τ ), τ ∈ [t−M : t]
that better “explains” the given history and the model.
In a somewhat simplified notation the problem being solved is one of
minimization of the function

J[x(t − M ), . . . , x(0)] =
−1
X
{[x(t + 1) − f (x(t), uM (t))]2Qx + [yM (t) − h(x(t), uM (t))]2Qu }
t=−M

Tuning parameters are qx , qy > 0, which denote the degree of trust on the
existent measurements (so called measurement noise level) and in the model

P. Al Hokayem & E. Gallestey 192 of 214

Nonlinear Systems and Control — Spring 2018

(process noise level). For large estimation horizon M , this method converges
towards a Kalman Filter.

Remark 11.5 Moving Horizon Estimation allows keeping the problem con-
straints explicitly, which might be important in application for the sake of
eliminating ambiguity.

Remark 11.6 (Findeisen, 2003): Stability results of MPC and MHE can be
obtained assuming that

1. MHE can be made to converge very fast: given error bound is satisfied
after a given number of iterations

2. MPC has a continuous value function: NMPC robust to small pertur-

bations.

Neither of these facts are obvious to prove for a given system, but they show
in what direction ones needs to work to guarantee stability.

Industrial implementations of this method hide complexity from the user.

All is required it is to attach to the model those variables containing the
current values of the measurement and actuators. Industrial grade systems
automatically:

• collect the data needed, i.e. the values of actuators, measurements and
model parameters,

• create the associated mathematical programming problem,

• calculate the estimation of the states as required,

• pass these values to the controller for calculation of the optimal moves.

This capability can also be used in stand-alone fashion for construction

of model based soft sensors or even simply for noise filtering purposes. One
example of this usage is the implementation of material balancing appli-
cations, where the values of certain magnitudes in the process (eg, metal
content and/or reagent usage) are reconstructed by looking at magnitudes
measured at other points of the circuit and a (simple) mathematical model
of the installation.

P. Al Hokayem & E. Gallestey 193 of 214

Nonlinear Systems and Control — Spring 2018

11.5 High Gain Observers

Another approach to estimate the state is to consider a High Gain Observer.
Consider the system
ẋ1 = x2
ẋ2 = φ(x, u)
y = x1
Suppose that the state feedback u = γ(x) is a locally Lipschitz state feedback
control that stabilizes the origin x = 0 of the closed loop sytem:

ẋ1 = x2
ẋ2 = φ(x, γ(x))
y = x1

To implement this feedback control using only y, use the following estimation
scheme:
x̂˙ 1 = x̂2 + h1 (y − x̂1 )
x̂˙ 2 = φ0 (x̂, γ(x̂)) + h1 (y − x̂1 )
Where φ0 (x, u) is a nominal model for φ(x, u). The estimation error x̃ = x− x̂
then satisfies
x̃˙ 1 = −h1 x̃1 + x̃2
x̃˙ 2 = −h2 x̃1 + δ(x, x̂)
where
δ(x, x̂) = φ(x, γ(x̂)) − φ0 (x̂, γ(x̂)).
We want (h1 ,h2 ) to be such that x̃(t) → 0, and if δ ≡ 0, this is easily
achieved by making the matrix
 
−h1 1
A0 =
−h2 0

to be stable. For δ 6= 0, we design (h1 ,h2 ) in such a way that the transfer
function
G0 : δ 7→ x̃
is as small as possible, i.e. that the gain of the system is as small as possible.
In this case:  
1 1
G0 = 2 ,
s + h1 s + h2 s + h1

P. Al Hokayem & E. Gallestey 194 of 214

Nonlinear Systems and Control — Spring 2018

which can be made arbitrarily small if h2 >> h1 >> 1. In particular if

α1 α2
h1 = , h2 = 2 , ε << 1,
ε ε
Then
lim kG0 (s)k = 0
ε→0

Unfortunately there is a problem, the (2,1) term of exp(A0 t) grows as h1 ,

h2 grow. This gives a peaking phenomenon. This means that the estimates
of x̃ will become arbitrarily large for some period of time. This will make
u = γ(x̂) large, and may drive the system state out of the region of attraction
– the region where the error dynamics are stable.
This situation may be overcome by saturating the control action. Since
the peaking phenomena is of short duration, we assume that the state esti-
mate will converge to the system state faster than the system diverges, and
then the estimated state feedback will be sufficient to stabilize the system.

Example 11.7 Let us consider the system:


 ẋ1 = x2
ẋ2 = x32 + u

y = x1

A stabilizing state feedback is given by:

u = γ(x) = −x32 − x1 − x2

The output feedback controller is then given by:



 x̂˙ 1 = x̂2 + 2ε (y − x̂
1 )
x̂˙ = −x̂1 − x̂2 + ε12 (y − x̂1 )
 2
u = sat (−x̂23 − x̂1 − x̂2 )

The following plots compare the behaviour of the system with state feed-
back, with output feedback and with saturated output feedback for value of ε
equal to 0.2 and 0.1. In Figure 11.6 and Figure 11.7 we see how for smaller
values of ε, the transient is faster but larger, leading to earlier instabilities of
the unsaturated system, but faster convergence of the estimated state to the
actual state for the saturated system.

P. Al Hokayem & E. Gallestey 195 of 214

Nonlinear Systems and Control — Spring 2018

In particular, note how for the case  = 0.1 depicted Figure 11.7, the
brief transient is sufficient to drive the controller without saturation into an
unstable region (the simulation stops around 0.9s), while in the case with
saturation, the controller performs almost as well as the state feedback con-
troller.
Further, as ε → 0 the region of attraction under output feedback ap-
proaches the region of attraction under state feedback. This holds whenever
γ(x) is a globally bounded stabilizing function. Indeed, as ε → 0 the time in
saturation becomes smaller. Eventually this time is so small that the system
state does not move during this time.
In general we have the following result:
Theorem 11.8 Consider the SISO system

ẋ = AC x + BC φ(x, u)
y = CC x
Assume that we have a stabilizing state feedback controller:

v̇ = Γ(v, x)
u = γ(v, x)
with g and G globally bounded. Consider then the output feedback controller

 v̇ = Γ(v, x̂)
x̂˙ = AC x̂ + Bφ0 (x̂, u) + H (y − Cc x̂)

u = γ(v, x̂)
where  
α1/
ε
 α2/ 2 
 ε 
H= ..  ∈ Rρ , (r the relative degree)
 . 
αρ/ ρ
ε
and ε > 0, and α1 , α2 , . . . αρ are such that the polynomial
sρ + α1 sρ−1 + . . . + αρ = 0
has all its roots in the complex left half plane. Then ∃ ε0 > 0 such that for
all ε < ε0 , the closed loop system is exponentially stable.
Remark 11.9 In this formulation we have neglected the unobserved zero
dynamics. If they were to exist, then we need to assume that the original
system is minimum phase. Then the result will hold.

P. Al Hokayem & E. Gallestey 196 of 214

Nonlinear Systems and Control — Spring 2018

11.6 Exercises
1. Observability tells us whether it is possible to design a state estimator
for an autonomous system

xk+1 = (A + BK) xk , x0 = unknown

yk = Cxk .

with neither process nor measurement noise.

(a) How many observations yk are necessary to determine the initial

state of the system?
(b) What is the worst-case number of observations necessary to de-
termine the initial state, assuming that the system is observable?

2. Consider the linear stochastic system

xk+1 = (A + BK) xk , x0 = unknown

yk = Cxk + wk ,

where the measurements yk are corrupted by independent identically

distributed (i.d.d.) measurement noise wk Laplace distributed. Derive
a moving horizon estimator as the maximum likelihood estimator of
the state sequence (x0 , . . . , xN ) given the measurements (y0 , . . . , yN ).

3. Given the system:

x+
1 = (1 − Ts )x1 + Ts x2 + Ts w
x+ 3
2 = x2 − Ts x1 + Ts w
y = sin (x1 ) + 2x2 + v

with Ts = 1, unknown initial state x0 , independent identically dis-

tributed process noise w ∼ N (0, 0.22 ) and v ∼ N (0, 0.12 ).
Write a Matlab program to design a discretized Kalman Filter which
estimates the state x from output measurements yk , available at every
time step tk , and simulate its closed loop behavior.

P. Al Hokayem & E. Gallestey 197 of 214

Nonlinear Systems and Control — Spring 2018

Figure 11.6: High Gain Observer Plots for  = 0.2

P. Al Hokayem & E. Gallestey 198 of 214

Nonlinear Systems and Control — Spring 2018

Figure 11.7: High Gain Observer Plots for  = 0.1

P. Al Hokayem & E. Gallestey 199 of 214

Sample Exam 1

Problem 1: 2D Systems Existence & Uniqueness for ODEs

Consider the following system
(
ẋ1 = tan−1 (ax1 ) − x1 x2
ẋ2 = bx21 − cx2

Show that the solution to the system exists and is unique. Hint: Use the Jaco-
d
bian and remember that for any function g(x), we have that dx tan−1 (g(x)) =
d
g(x)
dx
1+g(x)2
.

Solution: The Jacobian of the system’s vector field f (x) is given by

" #  
f1 (x) f1 (x) a
∂f (x) x1 x2 1+x21
− x2 −x1
= f2 (x) f2 (x) =
∂x x1 x2
2bx1 −c

from which it follows that

∂f (x)
2
∂x ≤ max{a + |x2 | + |x1 |, 2b|x1 | + c} (11.4)
∞

which is bounded on any compact set R2 . Hence we have local existence and
uniqueness of the solution.

200
Nonlinear Systems and Control — Spring 2018

Problem 2: Lyapunov Stability

Consider the second order system

ẋ1 = −x1 + x1 x2
ẋ2 = −x2

1. Compute the equilibrium point of the system. (4pts)

2. Show using a quadratic Lyapunov function that the obtained equilib-

rium point is asymptotically stable. Is the obtained result local or
global? (6pts)

Solution:

1. Setting the derivative to 0 in both equations, we obtain

0 = −x1 + x1 x2 ⇒ x1 (x2 − 1) = 0
0 = −x2 ⇒ x2 = 0

From which it follows that the equilibrium point is the origin (0, 0).

2. Take V (x) = 12 (x21 + x22 ). Then,

V̇ (x) = x1 ẋ1 + x2 ẋ2 = −x21 + x21 x2 − x22

= −x21 (1 − x2 ) − x22

Therefore, V̇ (x) ≤ 0, if 1 − x2 > 0, i.e., ∀x2 < 1. Hence the system

is locally asymptotically stable. To be precise, we need the condition
1 − x2 ≥ , with  > 0 ⇒ V̇ (x) ≤ −x21 − x22 .

P. Al Hokayem & E. Gallestey 201 of 214

Nonlinear Systems and Control — Spring 2018

Problem 3: Sliding Mode Control

Consider the system (
ẋ1 = x2 + ax1 sin(x1 )
ẋ2 = bx1 x2 + u
where the parameters a and b belong to the intervals [−1 2] and [0 1], respec-
tively. Design using slide mode control a continuous and globally stabilizing
state feedback controller.
Hint: Use the sliding surface s = x2 + kx1 , and design the parameter k so
that you can achieve stability of the reduced system on this manifold.

Solution: On the manifold s = x2 + kx1 , we have the following system


ẋ1 = −kx1 + ax1 sin(x1 ) = −x1 (k − a sin(x1 ))
ẋ2 = − b x22
k
which is stable as long as k > 2. This guarantees that the sliding phase is
asympotically stable. It remains to show that we can design a control input
that renders convergence to the sliding surface in finite time. Consider the
following Lyapunov function
1
V = s2
2
Taking the derivative of V along the trajectory of s, we obtain
V̇ = sṡ = s(ẋ2 + k ẋ1 )
= s(bx1 x2 + u + kx2 + kax1 sin(x1 ))
= s(bx1 x2 + u + kx2 + kax1 sin(x1 )) + su
≤ |s||bx1 x2 + u + kx2 + kax1 sin(x1 )| + su
≤ |s|(|x1 x2 | + k|x2 | + 2k|x1 |) + su
now designing
u = −µ · sgn(s), µ > β + |x1 x2 | + k|x2 | + 2k|x1 |
with β > 0, yields
V̇ ≤ −β|s|
which implies that the designed controller forces the system towards the
sliding surface in finite time.

P. Al Hokayem & E. Gallestey 202 of 214

Nonlinear Systems and Control — Spring 2018

Problem 4: Feedback Linearization

Consider the system
(
ẋ1 = x2
ẋ2 = a sin(x1 ) − b cos(x1 )u

where a and b are some known positive constants.

1. Is the system feedback linearizable?
2. If so, design using feedback linearization a state feedback controller to
stabilize the system at x1 = θ, for a given θ ∈ [0 π/2).

Solution:
1. We can rewrite the equations as
ẋ = f (x) + g(x)u
   
x2 0
where f (x) = and g(x) = .
a sin(x1 ) −b cos(x1 )
We need to check the two conditions for full state feedback linearization:
 
G(x) = g(x), adf g(x)
Now,
∂g ∂f
adf g(x) = [f, g](x) = f (x) − g(x)
 ∂x ∂x   
0 0 x2 0 1 0
= −
b sin(x1 ) 0 a sin(x1 ) a cos(x1 ) 0 −b cos(x1 )
 
b cos(x1 )
=
bx2 sin(x1 )
Therefore,  
0 b cos(x1 )
G(x) =
−b cos(x1 ) bx2 sin(x1 )
And, det(G(x)) = b2 cos(x1 )2 6= 0, ∀x1 ∈ [0 π/2).
We also have that the distribution ∆ = {g(x)} being involutive. As
such, the system is feedback linearizable.

P. Al Hokayem & E. Gallestey 203 of 214

Nonlinear Systems and Control — Spring 2018

2. Consider the feedback control law

1
u= (a sin(x1 ) + k1 (x1 − θ) + k1 x2 )
b cos(x1 )

Then the closed-loop system is given by

(
ẋ1 = x2
ẋ2 = −k1 (x1 − θ) − k1 x2
 
0 1
which has the system matrix A = which is Hurwitz for all
−k1 −k2
k1 , k2 > 0.

Problem 5: Optimal Control

Consider a boat anchored at a river bank at time t=0. The river current
velocity is given by v = (v1 (x2 ), 0), i.e. it is parallel to the shore. The boat
has a motor u = (u1 , u2 ) that can be used to change the boat direction and
speed, always acting with total effort equal to 1, i.e.,

kuk22 = u21 + u22 = 1

The dynamics are given by

ẋ1 = v1 (x2 ) + u1
ẋ2 = u2

with the initial condition x(0) = (x1 (0), x2 (0)) = (0, 0).

1. Assuming that the river current speed v = (v1 (x2 ), 0) = (x2 , 0), find an
optimal control u(t) that moves the boat to the maximum distance in
the x1 -direction in a fixed time T .

2. How would you solve the problem if the condition is added that the
boat must be at the river bank at time 2T , i.e., x2 (2T ) = 0?

P. Al Hokayem & E. Gallestey 204 of 214

Nonlinear Systems and Control — Spring 2018

Solution:

1. The cost to minimize is

Z T
J(x(0), u) = φ(x(T )) + L(x, u)dt = −x1 (T )
0

subject to the boat dynamic equations and the motor effort constraint.
The Hamiltonian is given by

H = 0 + λ1 (x2 + u1 ) + λ2 u2

The adjoint equations are given by

∂H
λ̇1 = − =0
∂x1
∂H
λ̇2 = − = −λ1
∂x2
with the boundary conditions
∂φ
λ1 (T ) = = −1
∂x1
∂φ
λ2 (T ) = − =0
∂x2
The solution to the adjoint equations is given by

λ1 (t) = −1
λ2 (t) = t − T

for t ∈ [0, T ]. The optimal control law u∗ (t) is found by mininizing the
Hamiltonian function subject to the input constraints, i.e.,

min H = min min H

kuk22 =1 2 2
kuk2 =1 kuk2 =1

which yields
" #  
λ1 (t) √1
λ − kλ(t)k 2
u∗ (t) = − = 2
=  T −t ) 
1+(t−T
kλk2 λ2 (t)
− kλ(t)k √
2 1+(t−T )2

P. Al Hokayem & E. Gallestey 205 of 214

Nonlinear Systems and Control — Spring 2018

2. By the principle of optimality, the optimal trajectory will be one where

the Solution a is applied till t = T , followed by the symmetric one, i.e.,
 
√ 1 2
1+(t−T )
u∗ (t) = −  √ T −t 
1+(t−T ) 2

for t ∈ (T, 2t].

P. Al Hokayem & E. Gallestey 206 of 214

Sample Exam 2

Problem 1: ODEs and Planar Systems

Consider the nonlinear system

ẋ1 = ax1 − x1 x2
ẋ2 = bx21 − cx2

with a, b, and c being positive constants.

1. Determine all equilibrium points of the system.

2. Compute the Jacobian of the system.

3. Determine the type of each equilibrium point.

Solution:

0 = ax1 − x1 x2 ⇒ x∗1 = 0 or x∗2 = a

p
0 = bx21 − cx2 ⇒ x∗1 = ± cx∗2 /b

combining the two equations we obtain the following equilibria

p
x∗ = (0, 0), x∗ = (± cx∗2 /b, a)

The Jacobian is computed as

 
a − x2 −x1
J|x∗ = |∗
2bx1 −c x

Then, we have that  

a 0
Jx∗ =(0,0) =
0 −c

207
Nonlinear Systems and Control — Spring 2018

which is a saddle point. And for the other two equilibria

 p 
√ 0
√ ∓ ca/b
Jx∗ =±( cx∗ /b,a) =
2 ±2 cab −c
which has the characteristic polynomial
√
2 c2 − 8ac
−c ±
λ + cλ + 2ac = 0 ⇒ λ =
2
which gives a stable focus for c < 8a and a stable node if c > 8a.

Problem 2: Lyapunov Stability

Consider the system


ẋ = −a In + S(x) + xxT x
where x ∈ Rn , a is a posivitve constant, In is the n × n identity matrix, S(x)
is a skew symmetric matrix that depends on x, i.e. S(x)T = −S(x), ∀x.
1. Show that the origin is globally asymptotically stable.
2. Show further that the origin is globally exponentially stable.
Solution:
1. Take a quadratic Lyapunov function V (x) = 21 xT x, then



V̇ (x) = xT −a In + S(x) + xxT x
= −a kxk22 − axT S(x)x −axT xxT x
| {z }
=0

= −a kxk22 − a kxk42
hence GAS.
2. We also have that 21 kxk22 ≤ V (x) ≤ 1
2
kxk22 , and V̇ (x) ≤ −a kxk22 from
which it follows that
V̇ (x) ≤ −a kxk22 ≤ −2aV (x) ⇒ V (x(t)) ≤ e−2at V (x(0))
where the last inequality follows from the comparison lemma. Finally,
we have that
1 1
kx(t)k22 ≤ V (x(t)) ≤ e−2at V (x(0)) ≤ e−2at kx(0)k22 ⇒ kx(t)k2 ≤ e−at kx(0)k2
2 2
and exponential stability follows.

P. Al Hokayem & E. Gallestey 208 of 214

Nonlinear Systems and Control — Spring 2018

Problem 3: Radial Unboundedness of Lyapunov Functions

Consider the following positive definite function
(x1 + x2 )2
V (x) = + (x1 − x2 )2
(x1 + x2 )2 + 1

When checking radial unboundedness of the function V (x), we may think

that it is sufficient to check what happens to V (x) as kxk → ∞ along the
principal axes. However, this is not true!

1. Show that V (x) is positive definite.

2. Show that V (x) → ∞ as kxk → ∞ along the principal axes/lines

x1 ≡ 0 or x2 ≡ 0.

3. Show that V (x) is not radially unbounded?

4. Why do we need radial unboundedness in Lyapunov’s theory?

Solution:
1. We can show that V (x) is positive definite via the following argument:

(x1 + x2 )2 + (x1 − x2 )2 + (x1 − x2 )2 (x1 + x2 )2

V (x) =
(x1 + x2 )2 + 1
x2 + x22 + (x1 − x2 )2 (x1 + x2 )2 x21 + x22
= 1 ≥ >0
(x1 + x2 )2 + 1 (x1 + x2 )2 + 1

2. Along x1 = 0, we have that

x22
V (0, x2 ) = + x22 ≥ x22 → ∞, as |x2 | → ∞
x22 + 1
The same holds for the case of x2 = 0.

3. No, because for x1 = x2 = ξ, we have that

4ξ 2
V (x) = → 1, as |ξ| → ∞
4ξ 2 + 1

4. For showing global asymptotic stability

P. Al Hokayem & E. Gallestey 209 of 214

Nonlinear Systems and Control — Spring 2018

Problem 4: Sliding Mode Control

Consider the nonlinear system
ẋ1 = x2 + x1 e−|x2 |
ẋ2 = ψ(x1 , x2 ) + u

where the function ψ is unknown and satisfies |ψ(x1 , x2 )| ≤ kxk4 , ∀x ∈ R2 .

1. Design as state feedback sliding mode controller to globally stabilize the
system to the origin. You need to prove that you controller is actually
stabilizing using a proper Lyapunov function.
2. Is the resulting controller continuous? If not, can you make it continu-
ous? Hint: remember the discussion in the lecture regarding alternative
choices to the sign function.

Solution:
1. Consider the manifold s = x2 − kx1 . On the manifold, we have that
ẋ1 = −kx1 + x1 e−|x2 | = −(k − e−|x2 | )x1
which is exponentially asymptotically stable for any choice of k > 1. It
remains to show that we can drive the system towards the mainfold in
finite time. First, we have that
ṡ = ẋ2 − k ẋ1 = ψ(x1 , x2 ) + u − kx2 − kx1 e−|x2 |
Consider the Lyapunov function candidate V = 21 s2 , then

V̇ = sṡ = s(ψ(x1 , x2 ) + u − kx2 − kx1 e−|x2 | ) ≤ |s| kxk4 + s(u − kx2 − kx1 e−|x2 | )

taking u = kx2 + kx1 e−|x2 | − (kxk4 + β)sgn(s) we obtain

V̇ ≤ −β|s|
which implies that the system using the sliding mode control above
converges to the manifold in finite time.
2. In order to make the sliding mode controller continuous, we can use
the following alternative input
s
u = kx2 + kx1 e−|x2 | − (kxk2 + β)sat , >0 (11.5)

P. Al Hokayem & E. Gallestey 210 of 214
Nonlinear Systems and Control — Spring 2018

Problem 5: Feedback Linearization

Consider the nonlinear system

ẋ1 = −x1 − x2 + ψ(x2 )

ẋ2 = x1 + ψ(x2 ) + u
y = x2

where the nonlinear function ψ is continuously differentiable and satisfies:

dψ(0) dψ(x2 )
ψ(0) = 0, = 1, 0< < 1, ∀x2 6= 0
dx2 dx2
1. Is the system input-output linearizable? What is the relative degree?
2. Is it full state feedback linearizable? (Hint: you need to check some
conditions involving Lie brackets)

Solution:
1.
ẏ = ẋ2 = x1 + ψ(x2 ) + u
hence the system has relative degree ρ = 1 and is input-output lineariz-
able.
2. We can rewrite the equations as

ẋ = f (x) + g(x)u
   
−x1 − x2 + ψ(x2 ) 0
where f (x) = and g(x) = .
x1 + ψ(x2 ) 1
We need to check the two conditions for full state feedback linearization:
 
G(x) = g(x), adf g(x)

Now,
∂g ∂f
adf g(x) = [f, g](x) = f (x) − g(x)
" ∂x # ∂x " #
−1 −1 + dψ(xdx2
2)
0 1 − dψ(x2 )
dx2
=− dψ(x2 ) = dψ(x2 )
1 dx2
1 − dx2

P. Al Hokayem & E. Gallestey 211 of 214

Nonlinear Systems and Control — Spring 2018

Therefore, " #
0 1 − dψ(x
dx2
2)

G(x) =
1 − dψ(x
dx2
2)

dψ(x2 )
And, det(G(x)) = dx2
− 1 6= 0, ∀x2 6= 0.
We also have that the distribution ∆ = {g(x)} being involutive. As
such, the system is feedback linearizable.

Problem 6: Stability of Model Predictive Control Designs

Consider
1. Discrete time, time invariant system

x(k + 1) = f (x(k), u(k)), f (0, 0) = 0, x(0) = x0

2. Objective function
k+N
X
J[x0 , u(·)][= L(x(l), u(l))
l=k

Consider an MPC algorithm for this system, where

• The problem contains a terminal constraint x(k + N ) = 0

• The function L in the cost function is positive definite in both argu-

ments.
Prove that if the optimization problem is feasible at time k, then the coor-
dinate origin is a stable equilibrium point.

Solution: Clearly, the origin is an equilibrium point of this discrete time

system because f (0, 0) = 0. In order to prove stability of the origin, we use
the Lyapunov result on stability of discrete time systems introduced in the
Lyapunov stability lecture. Indeed, consider the function

V (x) = J ? (x),

where J ? denotes the performance index evaluated at the optimal trajectory.

We note that:

P. Al Hokayem & E. Gallestey 212 of 214

Nonlinear Systems and Control — Spring 2018

• V (0) = 0

• V (x) is positive definite.

• V (x(k + 1)) − V (x(k)) < 0. The later is seen by noting the following
argument. Let
u?k (l), l = k : k + N
be the optimal control sequence at time k. Then, at time k + 1, it is
clear that the control sequence u(l), l = k + 1 : N + 1, given by

u(l) = u?k (l), l =k+1:N

u(N + 1) = 0

generates a feasible albeit suboptimal trajectory for the plant. Then,

we observe

V (x(k+1))−V (x(k)) < J(x(k+1), u(·))−V (x(k)) = −L(x(k), u? (k)) < 0

which proves the theorem.

P. Al Hokayem & E. Gallestey 213 of 214

Bibliography

[1] J. W. Helton and M. R. James. A general framework for extending h∞

control to nonlinear systems. SIAM, 1998.

[2] A. Isidori. Nonlinear control systems. Springer-Verlag, New York, 3rd

edition, 1995.

[3] James Cloutier et al. Nonlinear regulation and nonlinear h∞ control via
the state dependent riccati equation technique. Proceedings of First Inter-
national Conference on Nonlinear Problems in Aviation and Aerospace,
Florida, May,1996.

[4] H. K. Khalil. Nonlinear systems. Prentice hall, 3rd edition, 2002.

[5] C. Mracek. Control designs for the nonlinear benchmark problem via the
SDRE method. International Journal of Robust and Nonlinear Control,
(8):401–433, 1998.

[6] A. van der Schaft. L2 -Gain and Passivity Techniques in Nonlinear Con-
trol. Springer-Verlag, London, 2000.

[7] J. Willems. Dissipative dynamical systems - Part I: General theory.

Archive for Rational Mechanics and Analysis, (45):321–351, 1972.

214