Theoretical Background of Nonlinear System Stability and Control

Appendix A
Theoretical Background of Nonlinear System

Stability and Control
. . . Mr. Fourier had the opinion that the main purpose of

mathematics was public utility and explanation of natural
phenomena, but a philosopher like him should have known that
the sole purpose of science is the honor of the human spirit, and
that, as such, a matter of numbers is worth as much as a matter
of the world system.
Letter from Jacobi Legendre, July 2nd, 1830
Automatic control comprises a number of theoretical tools of mathematical charac-

teristics that enable to predict and apply its concepts to fulfill the objectives that are
directly attached to it. These tools are necessary for the synthesis of control laws on
a specific process and are utilized at various stages of the design. This is more so,
particularly during the modeling and identification of the parameters stage as well
as during the construction of control laws and during the verification of the stability
of the controlled system, just to mention a few. In fact, it is well-known that all con-
struction techniques of control laws or observation are narrowly linked to stability
considerations.
As a result, in the first part of this appendix, some definitions and basic concepts
of stability theory are recalled. The second part of the appendix is dedicated to the
presentation of some concepts and techniques of control theory.
Due to the numerous contributions in this area, in the past few years, we have
focused our interest only on points that are more directly related to our own work.
A.1 Stability of Systems
One of the tasks of the control engineer consists, very often, in the study of stability,
whether for the considered system, free from any control, or for the same system
when it is augmented with a particular control structure. At this stage, it might be
useful or even essential to ask what stability is. How do we define it? How to con-
ceptualize and formalize it? What are the criteria upon which one can conclude on
the stability of a system?
A. Choukchou-Braham et al., Analysis and Control of Underactuated Mechanical 93

Systems, DOI 10.1007/978-3-319-02636-7,
© Springer International Publishing Switzerland 2014
94 A Theoretical Background of Nonlinear System Stability and Control
A.1.1 What to Choose?
It is clear that drawing up an inventory as complete as possible of the forms of

stability that have appeared throughout the history of automatic control but also
of mechanics would be beyond the scope of this book. There will therefore not
be included in this presentation the stability method of Krasovskii [70], comparison
method, singular perturbations [37], the stability of the UUB (Uniformly Ultimately
Bounded) [15], the input–output stability [94], the input to state stability [71], the
stability of non-autonomous systems [2], contraction analysis [34], and descriptive
functions [70].
In addition, we will not be presenting the proofs of various results in this section.
We will assume that the conditions of existence and the uniqueness of solutions for
the considered systems of differential equations are verified everywhere.
From a notation point of view, we shall denote by x(t, t0 , x0 ) the solution at
time t with initial condition x0 at time t0 or by x(t, t0 , x0 , u) when the system is
controlled. In addition, for simplicity, we shall frequently use the notation x(t) or
even x, when the dependence on t0 , x0 or t is evident. Similarly, we shall consider
in the majority of cases, except in some cases, the initial time t0 = 0.
The class of systems considered will be those that can be put in the following
ordinary differential equation (ODE) form:
ẋ = f (x) (A.1)
where x ∈ Rn is the state vector and f : D → Rn a locally Lipschitzian function

and continuous on the subset D of Rn .
This type of systems is also called autonomous due to the absence of the temporal
term t in the function. Non-autonomous systems of the form ẋ = f (x, t) are not
considered in this work.
For the above equation (A.1), the point of the state space x = 0 is an equilibrium
point if it verifies
f (0) = 0 ∀t ≥ 0 (A.2)
Note that, by a change of coordinates, one can always bring the equilibrium point to
the origin.
A.1.2 The Lyapunov Stability Theory
The Lyapunov stability theory is considered as one of the cornerstones of automatic

control and stability for ordinary differential equations in general. The original the-
ory of Lyapunov dates back to 1892 and deals with the study of the behavior of
solution of differential equation for different initial conditions. One of its applica-
A.1 Stability of Systems 95
Fig. A.1 Intuitive illustration

of stability
tions that was contemplated at that time was the study of librations in astronomy.12
The emphasis is focused on the ordinary stability (i.e. stable but not asymptotically
stable), which we can represent as a robustness with respect to initial conditions,
and the asymptotic stability is only addressed in a corollary manner.
The automatic control community having inverted this preference, we will be
concentrating here on the concept of asymptotic stability rather than mere stability.
Note that there are many more complete presentations of Lyapunov stability in
many articles, for example [37, 49, 55, 62, 65, 66, 70, 87], which constitute the main
references of this part; this list also does not claim to be exhaustive.
A.1.2.1 Stability of Equilibrium Points
Roughly speaking, we say that a system is stable if when displaced slightly from its
equilibrium position, it tends to come back to its original position. On the other
hand, it is unstable if it tends to move away from its equilibrium position (see
Fig. A.1).
Mathematically speaking, this is translated into the following definitions:
Definition A.1 [37] The equilibrium point x = 0 is said to be:

• stable, if for every ε > 0, there exists η > 0 such that for every solution x(t) of
(A.1) we have

x(0) < η ⇒ x(t) < ε ∀t ≥ 0
• unstable, if it is not stable, that is, if for every ε > 0, there exists η > 0 such that
for every solution x(t) of (A.1) we have

x(0) < η ⇒ x(t) ≥ ε ∀t ≥ 0
• attractive, if there exists r > 0 such that for every solution x(t) of (A.1) we have

x(0) < r ⇒ lim x(t) = 0
t→∞
The basin of attraction of the origin is defined by the set B such that
x(0) ∈ B ⇒ lim x(t) = 0

t→∞
1 In astronomy, the librations are small oscillations of celestial bodies around their orbits.
2 The father of Alexander Michael Lyapunov was an astronomer.
Fig. A.2 Stability and

asymptotic stability of x̄
• globally attractive, if for every solution x(t) of (A.1) we have
lim x(t) = 0. In this case, B = Rn

t→∞
• asymptotically stable, if it is stable and attractive, and globally asymptotically

stable (GAS), if it is stable and globally attractive.
• exponentially stable, if there exist r > 0, M > 0 and α > 0 such that for every
solution x(t) of (A.1) we have

x(0) < r ⇒ x(t) ≤ M x(0)e−αt for all t ≥ 0
and globally exponentially stable (GES), if there exist M > 0 and α > 0 such that
for every solution x(t) of (A.1) we have

x(t) ≤ M x(0)e−αt for all t ≥ 0
Remark A.1
1. The difference between stable and asymptotically stable is that a small perturba-
tion on the initial state around a stable equilibrium point x̄ might lead to small
sustained oscillations, whereas these oscillations are dampened in time in the
case of asymptotically stable equilibrium point, see Fig. A.2 (U1 is the ball of
center 0 and radius ε and U2 is the ball of center 0 and radius η [46]).
2. For a linear system, all these definitions are equivalent (except for stable and
asymptotically stable). However, for a nonlinear system, stable does not imply
attractive, attractive does not imply stable, asymptotically stable does not imply
exponentially stable whereas exponentially stable implies asymptotically stable.
When the systems are represented by nonlinear differential equations, the verifi-
cation of stability is not trivial. On the contrary, for linear systems the verification
of stability is systematic and is determined as follows.
A.1.2.2 Stability of the Origin for a Linear System
Consider the linear system

ẋ = Ax (A.3)
where A is a square matrix of dimension n. Let λ1 , . . . , λs , be the distinct eigenval-
ues with algebraic multiplicity m(λi ) of the matrix A.
Theorem A.1 [37]

1. If ∃j Re(λj ) > 0 or if ∃k Re(λk ) = 0 and m(λk ) > 1 then x = 0 is unstable.
2. If ∀j Re(λj ) < 0 then x = 0 is exponentially (hence asymptotically) stable.
3. If Re(λj ) < 0 and if ∃k Re(λk ) = 0 and m(λk ) = 1 then x = 0 is stable but not
attractive.
Unfortunately, there does not exist an equivalent theorem to that of eigenvalues

for nonlinear systems. In some cases, one can characterize the stability of the origin
via the study of the linearized system.
A.1.2.3 Linear Approximation of a System
Consider a system of the form (A.1); we denote by
∂f
A= (x̄)
∂x
the Jacobian matrix of f evaluated at the equilibrium point x = x̄. The obtained sys-
tem will be of the form (A.3) and is called the linearization (or linear approximation)
of the nonlinear system (A.1).
Theorem A.2 [37]

1. If x = 0 is asymptotically stable for (A.3) then x = x̄ is asymptotically stable for
(A.1).
2. If x = 0 is unstable for (A.3) then x = x̄ is unstable for (A.1).
3. If x = 0 is stable but not asymptotically stable for (A.3) then we cannot conclude
on the stability of x = x̄ for (A.1).
Another criterion that allows to conclude on the stable behavior of the system for
both linear and nonlinear system is described next.
A.1.2.4 Lyapunov’s Direct Method
The principle of this method is a mathematical extension of the following physi-

cal phenomenon: if the total energy (of positive sign) of a mechanical or electrical
system is continuously decreasing then the system tends to reach a minimal energy
configuration. In other words, to conclude on the stability of a system, it suffices to
examine the variations of a certain scalar function called Lyapunov function without
having to solve the explicit solution of the system. This is precisely the strong point
of this method since the equation of motion of x(t) does not have to be computed
in order to characterize the evolution of the solution (the determination of explicit
solutions of nonlinear system is difficult and sometimes impossible).
Lyapunov Function Consider the system
ẋ = f (x) with f (0) = 0 (A.4)
x = 0 is an equilibrium point for (A.4) and D ⊂ Rn is a domain that contains x = 0.

Let V : D → R be a function that admits continuous partial derivatives. We de-
note
∂V (x)
n
∂V (x)
V̇ (x) = · f (x) = · fi (x)
∂x ∂xi
i=1
the derivative of the function V in the direction of the vector field f .
Definition A.2 The function V is a Lyapunov function for system (A.4 ) at x = 0

in D, if for all x ∈ D we have
• V (x) > 0 except at x = 0 where V (0) = 0
• V̇ (x) ≤ 0.
Theorem A.3 [37]

1. If there exists a Lyapunov function for (A.4) at x = 0 in a neighborhood D of 0,
then x = 0 is stable.
2. If, in addition, x = 0 ⇒ V̇ (x) < 0 then x = 0 is asymptotically stable.
3. If, in addition, D = Rn and V (x) → ∞ when x → ∞ then x = 0 is GAS.
Remark A.2
1. V̇ depends only on x, it is sometimes called the derivative of V along the system.
2. This derivative is also called Lie derivative and is denoted by Lf V .
3. To calculate V̇ , we do not require the knowledge of x but of ẋ, that is, of f (x).
Hence, for the same function V (x), V̇ is different for different systems.
4. For every solution x(t) of (A.4), we have dt d
V (x(t)) = V̇ (x(t)), consequently
if V̇ is negative, V decreases along the solution of (A.4) so that the trajectories
converge towards the minimum of V .
5. When V (x) → ∞ whenever x → ∞, V (x) is said to be radially unbounded.
6. V (x) is often a function that represents the energy or a certain form of energy of
the system.
7. From a geometric point of view, a Lyapunov function is seen as a bowl whose
minimum coincides with the equilibrium point. If this point is stable, then the
velocity vector ẋ (or f ), tangent to every trajectory will point towards the interior
of the bowl, see Fig. A.3 [46].
LaSalle’s Invariance Principle
Definition A.3 A set G ⊆ Rn is said to be positively invariant if every solution x(t)

such that x(0) ∈ G remains in G for all t ≥ 0.
Fig. A.3 Lyapunov

function V for vector fields f
If x̄ is an equilibrium point then {x̄} is positively invariant.
Theorem A.4 ([37] (Lyapunov–LaSalle)) Let V : D → R+ be a function having

continuous partial derivatives such that there exists l for which the region Dl defined
by V (x) < l is bounded V̇ (x) ≤ 0 for all x ∈ Dl . Let R = {x ∈ Dl : V̇ (x) = 0}
and let M be the largest positively invariant set that is included in R. Then, every
solution issued from Dl tends to M when t → ∞. In particular if {0} is the only
orbit contained in R then x = 0 is asymptotically stable and Dl is contained in its
basin of attraction.
Theorem A.5 [37] Let V : Rn → R+ be a function having continuous partial

derivatives. Suppose that V (x) is radially unbounded and that V̇ (x) ≤ 0 for all
x ∈ Rn . Let R = {x ∈ Rn : V̇ (x) = 0} and let M be the largest positively invariant
set that is included in R. Then, every solution tends to M when t → ∞. In particular
if {0} is the only orbit contained in R then x = 0 is GAS.
Remark A.3
1. The criteria for stability and asymptotic stability presented in Theorems A.3, A.4
and A.5 are easy to utilize. However, they do not give any information on how
to construct the Lyapunov function. In reality, there does not exist any general
method for the construction of Lyapunov functions except for some particular
classes of systems (namely for the class of linear systems).
2. The theorems given previously give sufficient conditions in the sense that if for a
certain Lyapunov function V , the conditions on V̇ are not satisfied, this does not
imply that the considered system is unstable (maybe with another function one
can demonstrate the stability of the system).
3. Contrary to Lyapunov functions which guarantee the stability of the equilibrium
points, there are functions, called Chetaev functions, that guarantee the instabil-
ity of the equilibrium points. Note that it is more difficult to demonstrate the
instability rather than stability (refer to [46] for more details).
In some cases, a dynamical system is represented, at a given instant of time t ≥ t0 ,

not by a single set of continuous differential equations, but by a family of contin-
uous subsystems together with a logic orchestrating the switching between these
subsystems: this is the class of switching systems.
In this book we have employed some controllers for this class of systems. Con-
sequently, in what follows, we shall present the stability criteria for these systems.
We shall present the controller design for this class of system for a later stage.
A.1.3 Stability of Switching Systems
Mathematically speaking, a switching system can be described by equations of the

form
ẋ = fp (x) (A.5)
where {fp : p ∈ P} is a family of functions sufficiently regular defined from Rn to
Rn and parameterized by a set of indices P.
For the system (A.5), the active subsystem at every instant of time is determined
by a sequence of switches of the form

σ = (t0 , p0 ), (t1 , p1 ), . . . , (tk , pk ), . . . (t0 ≤ t1 ≤ · · · ≤ tk )
σ is called the switching signal and can depend either on time or on the state or
both. Such systems are said to have variable structures or are called multi-models.
They represent a particularly simple class of hybrid systems [10, 81, 85].
Here, we shall assume that the origin is an equilibrium point that is common
for the individual subsystems fp (0) = 0. We shall also assume that the switching
is done without jumps and does not occur infinitely fast so that the Zeno phe-
nomenon is avoided. The reader who is interested in these properties can refer to
[6, 53, 63, 64].
The class of systems often considered in the literature are those for which the
individual systems are linear
ẋ = Ap x (A.6)
Just to mention a few, we cite the following references: [11, 26, 51, 50, 52, 54, 68,
75, 74, 90, 95, 96, 97]. On the other hand, there are only few works in the literature
for the class of nonlinear switching systems [9, 12, 16, 19, 47, 53, 98, 99].
At this stage one might ask the following question: given a switching system,
why do we need a theory of stability that is different from that of Lyapunov?
The main reason is that the stability of switching systems depends not only on
the different dynamics corresponding to several subsystems but also on the transition
law that governs the switchings. In effect, we have the case where two subsystems
are exponentially stable while the switching between the two subsystems drives the
trajectories to infinity.
In fact, it was shown in [12, 19, 47] that a necessary condition for the stability of
switching systems subjected to an arbitrary transition law is that all the individual
subsystems should be asymptotically stable, but this condition was not sufficient.
Nevertheless, it appears that when the switching between the subsystems is suffi-
ciently slow (so as to allow the transition period to settle down and to allow each
subsystems to be in steady state) then it is very likely that the global system will be
stable.
A.1.3.1 Common Lyapunov Function
It is clear that in the case where the family of systems (A.5) possesses a common
Lyapunov function V (x) such that ∇V (x)fp (x) < 0 for all x = 0 and all p ∈ P,
then the switching system is asymptotically stable for any transition signal σ [47].
Hence, a possibility for demonstrating the stability of switching systems consists in
finding a common Lyapunov function for all the individual subsystems of (A.5).
However, finding a Lyapunov function for a nonlinear system, even for a single
one is not simple. If, in addition, we allow the switchings between several subsys-
tems, the determination of such a function becomes much more difficult. It is also
the reason for which a non-classical theory of stability is necessary.
A.1.3.2 Multiple Lyapunov Functions
In the case where a common Lyapunov function cannot be determined, the idea is
to demonstrate the stability through several Lyapunov functions. One of the first
results of such procedure was developed by Peleties in [58, 59], then by Liberzon
[47], for the switching systems of the form (A.6).
Given N dynamical systems Σ1 , . . . , ΣN , and N pseudo Lyapunov functions
(Lyapunov-like functions) V1 , . . . , VN .
Definition A.4 [19] A pseudo Lyapunov function for system (A.5) is a function
Vi (x) with continuous partial derivatives defined on a domain Ωi ⊂ Rn , satisfying
the following conditions:
• Vi is positive definite: Vi (x) > 0 and Vi (0) = 0 for all x = 0.
• V̇ is semi negative definite: for x ∈ Ωi ,
∂Vi (x)
V̇i (x) = fi (x) ≤ 0 (A.7)
∂x
and Ωi is the set for which (A.7) holds true.

Theorem A.6 [19] Suppose that i Ωi = Rn . For i < j , let ti < tj be the tran-
sition instants for which σ (ti ) = σ (tj ) and suppose that there exists γ > 0 such
that
2
Vσ (tj ) x(tj +1 ) − Vσ (ti ) x(ti+1 ) ≤ −γ x(ti+1 ) . (A.8)
Then, the system (A.6) with fσ (t) (x) = Aσ (t) x and the transition function σ (t) is
GAS.
Fig. A.4 Energy profile of

the linear switching system
for N = 2
The condition (A.8) is illustrated by Fig. A.4.

The first generalization of this theorem to nonlinear systems is due to Branicky
[9, 10, 11, 12]
Theorem A.7 [9, 10] Given N switching systems of the form (A.5) and N pseudo
functions Vi in the region Ωi associated to each subsystem, and suppose
Lyapunov
that i Ωi = Rn and let σ (t) be the transition sequence that takes the value i when
x(t) ∈ Ωi . If in addition,

Vi x(ti,k ) ≤ Vi x(ti,k−1 ) (A.9)
− +
where ti,k is the kth time where fi is active, that is, σ (ti,k ) = σ (ti,k ) = i, then (A.5)
is stable in the sense of Lyapunov.
Figure A.5 illustrates the condition (A.9) (in dotted lines) [19]. A more general
result due to Ye [91, 92] concerns the utilization of weak Lyapunov functions for
which condition (A.9) is replaced by

Vi x(t) ≤ h Vi x(tj ) , t ∈ (tj , tj +1 ) (A.10)
where h : R+ → R+ is a continuous function with h(0) = 0 and tj is any transition

instant when the system i is activated.
In this case, it is no longer required that the Lyapunov functions be decreasing.
It suffices that they are bounded by a function that is zero at the origin. Hence, the
energy can grow in the intervals where the same system is activated but must be
decreasing at the end of these intervals, see Fig. A.5 (solid lines).
Liberzon in [47] extends these results by giving a condition on multiple Lya-
punov functions in order to demonstrate the global asymptotic stability.
Consider N subsystems of the form (A.5). When the subsystems of the family
(A.5) are assumed to be asymptotically stable, then there exists a family of Lya-
punov functions {Vp : p ∈ P} such that the value of Vp decreases on each interval
for which the pth subsystem is active.
Fig. A.5 Energy profile of a

nonlinear switching system
for N = 3
Fig. A.6 Energy profile of

the nonlinear switching
system for N = 2
If for each p, the value of Vp at the end of the interval where the system p is
active is higher than the value of Vp at the end of the following interval when the
system p is active (see Fig. A.6), then the system (A.5) is asymptotically stable.
Remark A.4
1. When N = 1, we obtain the classical results of stability. However, when N = ∞
the previous theorems are no longer valid.
2. These theorem are valid even when fp vary as a function of time.
3. These results can be extended by relaxing certain hypotheses, for example: the
individual subsystems can have different equilibrium points [53] or state jumps
during a switch [64].
Note that all the results of stability using multiple Lyapunov functions are con-
cerned with the decrease of these functions either at the beginning or at the end of
successive intervals where the same subsystem is active. Zhai in [98] has shown
that certain Lyapunov functions may not decrease at the beginning or at the end
of these intervals and yet decrease globally. His demonstration, which establishes
a complementary condition of stability to those that already exist, is based on the
Fig. A.7 Illustration of the

j
average values of Vi (x(Ti ))
evaluation of the average value of Lyapunov functions during the intervals where
the same subsystem is active.
Evidently, in the case where the subsystems are GAS, the result is practically
equivalent to the previous results. However, his conditions are given with respect to
the decrease of the average Lyapunov functions on the same intervals, see Fig. A.7.
Theorem A.8 [98] Suppose that the N subsystems of (A.5), associated to N radi-
ally unbounded Lyapunov functions are GAS. Define the average value of the Lya-
punov functions during the activation period for each subsystem as
2j
j Δ 1 ti 2j −1 j 2j
Vi x Ti = 2j 2j −1 2j −1
Vi x(τ ) dτ ti ≤ Ti ≤ ti (A.11)
ti − ti ti
Then, the switched system is GAS in the sense of Lyapunov if, for all i,
j +1 j j
Vi x T − Vi x T
i ≤ −Wi x T
i i (A.12)
holds for a positive definite continuous function Wi (x).
Additionally, this result is extended to the case when the subsystems are not sta-
ble under the condition that the Lyapunov functions are bounded. In this case, if the
average value of the Lyapunov functions decreases on the set of intervals associ-
ated to a subsystem i, then the switching system (A.5) is asymptotically stable, see
Fig. A.8 [98].
Remark A.5 More recently, a similar result to the above using the average value
of the derivative of the Lyapunov functions, rather than the average value of the
Lyapunov functions, for the stability analysis of linear switching systems has been
given by Michel in [54].
Recall that the stability is the first step in the study of a system in terms of its
performance evaluation. In fact, if a system is not stable (or not stable enough), it
is important to proceed to the stabilization of this system before looking to satisfy
other performances such as trajectory tracking, precision, control effort, perturba-
tion rejection, robustness, etc.
A.2 Control Theory 105
Fig. A.8 Illustration of the

decrease of energy in the
presence of unstable systems
A.1.4 Stabilization of a System
The problem of stabilization consists in maintaining the system near an equilibrium

point y ∗ . The aim is to construct stabilizing control laws such that y ∗ becomes an
asymptotically stable equilibrium point of the system under these control laws.
Remark A.6
1. The problem of trajectory tracking consists in maintaining the solution of the
system along a desired trajectory yd (t), t ≥ 0. The objective here is to find a
control law such that for every initial condition in a region D, the error between
the output and the desired output
e(t) = y(t) − yd (t)
tends to 0 when t → ∞. In addition, the state must remain bounded.

2. Note that the stabilization problem around an equilibrium point y ∗ is a special
case of the problem of trajectory tracking whereby
yd (t) = y ∗ , t ≥0
The control design techniques allowing to construct control laws for the stabiliza-
tion of systems are numerous and varied. In what follows, we are going to present
those that are most useful for the control of underactuated mechanical systems. The
main references where most of the results on this subject were borrowed from, in
the next section, are [32, 37, 41, 43, 44, 66, 67].
A.2 Control Theory
Given a physical system that we want to control and the system behavior we want to
obtain, designing a control amounts to construct control laws such that the system
subjected to these laws (the closed-loop system) presents the desired behavior.
Nonetheless, the control procedure is only possible if the system in question is

controllable. Otherwise, the uncontrollable modes would need to be stable [13]. For
more details, please refer to Appendix D.
The synthesis of control laws for nonlinear systems is difficult in general. There-
after, we propose some control design techniques for the class of nonlinear control
affine systems of the form
ẋ = f (x) + g(x)u, x ∈ Rn , u ∈ R (A.13)
The linearizability is a property that renders the systems more easy to control.
In addition, the control design techniques for linear systems are well-established
and largely developed. We can cite some examples such as pole placement control,
optimal control, and a frequency-based approach just to mention a few. For more
details on these subjects the interested reader can refer to [2, 17, 24, 35, 42, 56, 93].
This list is far from complete obviously.
Thus, it might be useful to highlight this linearizability property for nonlinear
systems too. In what follows, the most employed and well-known procedures are
briefly recalled.
A.2.1 Local Stabilization
Consider the system of the following form:
ẋ = f (x) + g(x)u, f (0, 0) = 0
In the presence of the control input u, the linear approximation around the equilib-
rium point is given by
ẋ = Ax + Bu (A.14)
where the matrices A and B are defined by
∂f ∂f
A= (0, 0), B= (0, 0).
∂x ∂u
The obtained form (A.14) justifies the utilization of linear control techniques men-
tioned above.
Unfortunately, the resulting linearized system is typically valid only around the
considered point so that the associated controller is valid only in a neighborhood of
this point. This leads to a local control only. In addition, determining the linearity
domain is not obvious.
In Appendix B, the reader will find more details of the limits of linearization and
the underlying dangers of destabilization.
Hence, even though this method is simple and practical, it is necessary to proceed
differently in order to increase the validity domain of the synthesized controllers.
To further benefit from the theory of linear control, there exists a control design
technique based on a change of coordinates and a state feedback allowing to render
the nonlinear dynamics equivalent to that of a linear dynamics: this is the so-called
feedback linearization.
A.2.2 Feedback Linearization
When we transform a system via a change of coordinates, some of its properties re-
main unchanged. For example, if a system is unstable then the transformed system is
also unstable. If a system is controllable, then the transformed system is also control-
lable. On the other hand, some systems might seem nonlinear in certain coordinates
while they can become linear in other coordinates and under certain feedback.
Thus, it is interesting, whenever possible, to analyze the dynamics of a system in
a transformed form that is easier to study.
In Appendix C, some results and concepts of differential geometry necessary for
the presentation of this approach are recalled.
Two procedures of linearization by feedback are possible: input–state lineariza-
tion and input–output linearization.
A.2.2.1 Input–State Linearization
The aim here is to transform the system of the form (A.13) via a diffeomorphism
z = ϕ(x) into a system of the form
ż1 = z2
ż2 = z3
..
. (A.15)
żn−1 = zn

żn = a ϕ −1 (z) + b ϕ −1 (z) u
This form is similar to the canonical form of Brunovsky or the canonical form of
controllability of linear systems.
If such transformation is possible, then for b(ϕ −1 (z)) = 0 the control
1
u= v − a ϕ −1 (z) (A.16)
b(ϕ −1 (z))
permits to linearize the system, which becomes
ż1 = z2 , ż2 = z3 , ..., żn−1 = zn , żn = v
where v is an external control.

One can then ask the following questions: Is it always possible to linearize a
system by feedback? When this is the case, how do we obtain the transformation
z = ϕ(x)?
The answer to these questions lies in the following theorem:
Theorem A.9 [33] The system (A.13) is input–state linearizable in a domain D if

and only if:
1. The rank of the controllability matrix Cfg = {g, adfg , . . . , adn−1
fg } is equal to n
for all x ∈ D.
fg } is involutive in D.
2. The distribution {g, adfg , . . . , adn−2
With regard to the diffeomorphism, when the conditions of linearization are sat-
isfied, then there exist several algorithms that permit to find the latter [14, 37, 65].
A.2.2.2 Input–Output Linearization
Consider the following nonlinear system:
ẋ = f (x) + g(x)u, x ∈ Rn , u, y ∈ R
(A.17)
y = h(x)
The idea is to generate linear equations between the output y and a certain input
v through a diffeomorphism z = φ(x) constituted of the output and its derivatives
with respect to time up to the order n − 1 when the relative degree r associated to
this system is equal to n:
φ1 (x) = h(x)
φ2 (x) = Lf h(x)
(A.18)
..
.
φn (x) = Ln−1
f h(x)
The system thus transformed is written as
ż1 = z2
ż2 = z3
..
. (A.19)
żn−1 = zn

żn = a φ −1 (z) + b φ −1 (z) u.
By choosing u of the form (A.16) and assuming that b(φ −1 (z)) = 0 for all z ∈ R n ,
the system becomes
ż1 = z2 , ż2 = z3 , ..., żn−1 = zn , żn = v.
Note that in this case, the form (A.19) is the same as in (A.15). In fact, when r = n
the two linearizations are equivalent. Hence, the conditions for applying the second
linearization will be the same as for the first.
For more details on these two linearizations, and for some useful examples, the
reader can refer to [32, 33, 37, 66].
Obviously, for a relative degree r < n, the system is no longer completely feed-
back linearizable. In this case, one can talk of a partial feedback linearization.
A.2.2.3 Partial Feedback Linearization
When r < n, it is only possible to partially linearize a system of the form (A.17)
through the diffeomorphism constituted partly by the output h(x) and its successive
derivatives up to order r − 1: z = φi (x) for 1 < i < r, and completed—by using
the theorem of Frobenius—by n − r other functions: η = φi (x) for r + 1 ≤ i ≤ n,
chosen in such a way that Lg φi = 0 for r + 1 ≤ i ≤ n. In the coordinates (z, η) the
equations of the system are given by
ż1 = z2
ż2 = z3
..
.
żr−1 = zr (A.20)
żr = a(z, η) + b(z, η)u
η̇ = q(z, η)
y = z1
This particular form is called the normal form.

If b(z, η) = 0, the input u can be chosen as
1
u= v − a(z, η) .
b(z, η)
In this case, the system takes the form
ż1 = z2
ż2 = z3
..
.
żr−1 = zr (A.21)
żr = v
η̇ = q(z, η)
Clearly, this system is composed of a linear subsystem of dimension r that is control-

lable by v—which is responsible for the input–output behavior—and of a nonlinear
subsystem of dimension n − r whose behavior is not affected by the control input.
It follows that the global behavior of the system depends on this internal dynamics
and that the verification of its stability is an essential step.
In [32], it was shown that the stability study of the internal dynamics can be
reduced to that of the zero dynamics. This is obtained when we apply a control u
that brings and maintains the output y to zero. In other words, the zero dynamics is
given by the system η̇ = q(0, η).
Remark A.7
1. When η̇ = q(0, η) is (locally) asymptotically stable then the associated system is
said to have (locally) minimum phase characteristic at the equilibrium point x̄.
2. When η̇ = q(0, η) is unstable then the associated system is said to be a non-
minimum phase system.
Even though the methods of linearization are useful for simplifying the study and
the control of nonlinear systems, they nevertheless present certain limitations. For
example, the lack of robustness in the presence of modeling errors, the verification
of certain conditions such as the involutivity, which, very often, is not verified by
many systems; even those belonging to the class of nonlinear control affine systems,
this is the case of UMSs. In addition, the state must be fully measured and accessi-
ble. Hence, the utilization of such techniques is confined to some classes of systems
only.
Therefore, one must find other linearization techniques that are applicable to a
wide range of systems without demanding restrictive and rigorous conditions as re-
quired by exact linearization. For example, approximative linearization techniques
allow the linearization of the systems up to certain order and neglect certain nonlin-
ear dynamics of high order. The authors that were interested in this technique are
[5, 28, 31, 36, 39, 73], just to mention a few of them.
A.2.2.4 Approximate Feedback Linearization
For certain systems the computation of the relative degree presents some singular-
ities in the neighborhood of the equilibrium point. For other systems the relative
degree is smaller than the order of the system. In this case, the condition of involu-
tivity is not verified.
The key idea of approximate linearization is to find an output function such that
the system approximatively verifies the former condition.
Several linearization algorithms are available; one can cite, for example, lin-
earization by the Jacobian, pseudo-linearization, Krener algorithm, Hunt and
Turi [39], the algorithm of Krener based on the theory of Poincaré [40], the al-
gorithm of Hauser, and that of Sastry and Kokotović [28]. A comparative study of
these algorithms applied to some examples can be found in [45].
In what follows, we shall recall briefly the algorithm of Hauser et al. [28]. Con-
sider the system of the form (A.17):
ẋ = f (x) + g(x)u
y = h(x)
Suppose that the relative degree associated to this system is equal to r < n. Con-
sequently, the system is not exactly feedback linearizable. In f and g, some terms
prevent the linearization to take place, in the sense that the relative degree in the
presence of these terms is smaller than n.
The idea is to neglect these terms so that we can achieve a complete relative
degree, called robust relative degree.
Definition A.5 [88] The robust relative degree of a regular output associated to
system (A.17) at 0 is the integer γ such that
γ −2
Lg h(0) = Lg Lf h(0) = · · · = Lg Lf h(0) = 0
γ −1
Lg L f h(0) = 0
In this case, we say that the system (A.17) is approximatively feedback lineariz-
able around the origin if there exists a regular output y = h(x) for which γ = n.
This transformation is possible via the following diffeomorphism z = φ(x):
z1 = h(x)
z2 = Lf h(x)
..
.
zn = Ln−1
f h(x)
Hence, during an approximate linearization, the nonlinear model (A.17) is simpli-

γ −2
fied by assuming that the functions Lg h(x) = Lg Lf h(x) = · · · = Lg Lf h(x),
preventing the definition of the classical relative degree and that cancel at 0, are
identically zero:
γ −2
Lg h(x) = Lg Lf h(x) = · · · = Lg Lf h(x) = 0
In this case, the system (A.17) is approximated by the following form:
ż1 = z2
ż2 = z3
..
. (A.22)
żn−1 = zn
−1
żn = Lnf h φ −1 (z) + Lg Ln−1
f h φ (z) u
which is of the canonical form of Brunovsky.

−1
f h(φ (z) =
Hence, if u is conveniently chosen (of type (A.16)) then, for Lg Ln−1
0), the model will be in the linear form and will be locally controllable,
ż1 = z2 , ż2 = z3 ..., żn−1 = zn , żn = v
This method is in many cases satisfactory but naturally the control engineer will
always try to improve it in order to increase its performances and its domain of
validity. This is how the theory of higher order approximations was introduced by
Krener [39] and Hauser [27].
A.2.2.5 Higher Order Approximations
The objective here is to maximize the order of terms to be neglected in order to

have better precision. Hence, fewer terms are neglected. Consequently, the higher
the order of the neglected residue is, the more effective the controller will be and
the larger its domain of validity will be [8].
Theorem A.10 [39] The nonlinear system (A.13) can be approximated by a state
feedback around an equilibrium point if and only if the distribution Δn−2 (f, g) is
involutive up to order3 p − 2 on E.
This means that there is a change of coordinates z = Ψ (x) such that, in the new
coordinates z, the dynamics (A.13) is given by
ż1 = z2
ż2 = z3
.. p p−1 (A.23)
. +OE (x) + OE (x)u
żn−1 = zn
żn = a(z) + b(z)u
with b(Ψ −1 (x)) = 0 in a neighborhood of E.
In other words, for the system (A.13), during a higher order approximation, the
γ −2
terms Lg h(x) = Lg Lf h(x), . . . , Lg Lf h(x) are no longer assumed to be zero but
will be taken into account in the model and consequently in the expression of the
control law.
3A distribution is involutive up to order p on E if ∀f, g ∈ Δ, [f, g] ∈ Δ + OE (πx).

The obtained model (A.23) is no longer fully linearizable, but it is at least easier
to control than the initial system (A.13).
Apart from these methods of linearization, there exist several other approaches
that are different from one another for the synthesis of control. The utilization of
one method over another will depend on the class of systems considered. Among
these methods, we shall be interested in three of them, namely: passivity approach,
backstepping, and sliding mode control.
A.2.3 Few Words on Passivity
The notion of passivity is essentially linked to the notion of the energy that is ac-
cumulated in the considered system and the energy brought by external sources to
the system [57, 67, 86]. The principal reference on the utilization of this concept of
passivity in automatic control is due to Popov [61]. The dissipativity, which is an
extension of this concept, is developed in the works of Willems [89].
Even though the concept of passivity is applicable to a large class of nonlinear
systems, we will restrict our attention, only to dynamics modeled by system (A.17).
A dissipative system is then defined as follows:
Definition A.6 [67] The system (A.17) is said to be dissipative if there exists a
function S(x) that is positive and such that S(0) = 0, and a function w(u, y) that is
locally integrable for all u, such that the following condition:
0

S(x) − S(x0 ) ≤ w u(τ ), y(τ ) dτ (A.24)
t
is satisfied over the interval [0, t].
This inequality expresses the fact that the energy stored in the system S(x) is at
most equal to the sum of energies initially stored and externally supplied. That is,
there is no creation of internal energy; only a dissipation of energy is possible.
If S(x) is differentiable, the expression (A.24) can be written as
Ṡ(x) ≤ w(u, y). (A.25)
One particularity form of w permits to define the passivity of a system.
Definition A.7 [67] The system (A.17) is said to be passive if it is dissipative and
if the function w is a bilinear function from the input to the output w(u, y) = uT y.
The passivity is a fundamental property of physical systems that is intimately

linked to the phenomenon of energy loss or dissipation. One can recognize the prin-
ciple similar to that of stability. In effect, the relation between passivity and stability
can be established by considering the storage function S(x) as a Lyapunov func-
tion V (x).
Remark A.8 Note that the definition of dissipativity and passivity does not require
that S(x) > 0 (it suffices that S(x) ≥ 0). Hence, in the presence of an unobservable
part, x = 0 can be unstable while the system is passive. For passivity to imply stabil-
ity, one must exclude a similar case. That is, one must verify that the unobservable
part is asymptotically stable. The reader should refer to [67] for a complete review
on the stability of passive systems and for some results on Lyapunov functions that
are semi positive definite.
A.2.4 Backstepping Technique
The backstepping is a recursive procedure for the construction of nonlinear control

laws and Lyapunov functions that guarantee the stability of the latter. This technique
is only applicable to a certain class of system which is said to be in strict feedback
form (lower triangular). A quick review of this control design approach is given
below, see [41] for more details.
Consider the problem of the stabilization of nonlinear systems in the following
triangular form:
ẋ1 = x2 + f1 (x1 )
ẋ2 = x3 + f2 (x1 , x2 )
..
.
(A.26)
ẋi = xi+1 + fi (x1 , x2 , . . . , xi )
..
.
ẋn = fn (x1 , , x2 , . . . , xn ) + u
The idea behind the backstepping technique is to consider the state x2 as a “vir-
tual control ” for x1 . Therefore, if it is possible to realize x2 = −x1 − f1 (x1 ), then
the state x1 will be stabilized. This can be verified by considering the Lyapunov
function V1 = 12 x12 . However, since x2 is not the real control for x1 , we make the
following change of variables:
z1 = x1
z2 = x2 − α1 (x1 )
with α1 (x1 ) = −x1 − f1 (x1 ). By introducing the Lyapunov function V1 (z1 ) = 12 z12 ,
we obtain
ż1 = −z1 + z2
∂α1
ż2 = x3 + f2 (x1 , x2 ) − x2 + f1 (x1 ) := x3 + f¯2 (z1 , z2 )
∂x1
V̇1 = −z12 + z1 z2
By proceeding recursively, we define the following variables:
z3 = x3 − α2 (z1 , z2 )
1
V2 = V1 + z22
2
In order to determine the expression of α2 (z1 , z2 ), one can observe that
ż2 = z3 + α2 (z1 , z2 ) + f¯2 (z1 , z2 )

V̇2 = −z12 + z2 z1 + z3 + α2 (z1 , z2 ) + f¯2 (z1 , z2 )
By choosing α2 (z1 , z2 ) = −z1 − z2 − f¯2 (z1 , z2 ) , we obtain
ż1 = −z1 + z2
ż2 = −z1 − z2 + z3
V̇2 = −z12 − z22 + z2 z3
Proceeding recursively, at step i, and defining
zi+1 = xi+1 − αi (z1 , . . . , zi )
1 2
i
Vi = zk
2
k=1
we obtain
żi = zi+1 + αi (z1 , . . . , zi ) + f¯i (z1 , . . . , zi )

i−1

V̇i = − zk2 + zi−1 zi + zi zi+1 + αi (z1 , . . . , zi ) + f¯i (z1 , . . . , zi )
k=1
By using the expression αi (z1 , . . . , zi ) = −zi−1 − zi − f¯i (z1 , . . . , zi ), we obtain
żi = −zi−1 − zi + zi+1

i−1
V̇i = − zk2 + zi zi−1
k=1
At step n, we obtain
żn = f¯n (z1 , . . . , zn ) + u
Choosing
u = αn (z1 , . . . , zn ) = −zn−1 − zn − f¯n (z1 , . . . , zn )
for the following Lyapunov function:
1 2
n
Vn = zk
2
k=1
it turns out that
żn = −zn−1 − zn

n
V̇n = − zk2
k=1
The stability of the system is proven by using simple quadratic Lyapunov func-
tions. One must also note that the dynamic obtained in the z coordinates is linear.
The advantage of the backstepping technique is its flexibility for the choice of the
stabilizing functions αi , which are simply chosen to eliminate all the nonlinearities
in order to render the function V̇i negative.
A.2.5 Sliding Mode Control
The theory of variable structure systems has been the subject of numerous studies
over the last 50 years. Initial works on this type of systems are those of Anosov
[1], Tzypkin [82]. and Emel’yanov [21]. These works have encountered a signifi-
cant revival in the late 1970s when Utkin introduced the theory of sliding modes
[83]. This control and observation technique received increasing interest because of
their relative ease of design, their strength vis-à-vis certain parametric uncertainties
and perturbations, and the wide range of their applications in varied fields such as
robotics, mechanics or power systems.
The principle of this technique is to force the system to reach and then to remain
on a given surface called sliding or switching surface (representing a set of static
relationships between the state variables). The resulting dynamic behavior, called
ideal sliding regime/mode, is completely determined by the parameters and equa-
tions defining the surface. The advantage of obtaining such behavior is twofold: on
one hand, there is a reduction of the system order, and on the other, the sliding mode
is insensitive to disturbances occurring in the same direction as the inputs.
The realization is done in two stages: a surface is determined so that the sliding
mode has the desired properties (not necessarily present in the original system), and
then a discontinuous control law is synthesized in order to make the surface invari-
ant (at least locally) and attractive. However, the introduction of this discontinuous
action, acting on the first derivative with respect to time of the sliding variable, does
not generate an ideal sliding mode. On average, the controlled variables can be con-
sidered as ideally moving on the surface. In reality, the movement is characterized
by high-frequency oscillations in the vicinity of the surface. This phenomenon is
known as chattering and is one of the major drawbacks of this technique. Further-
more, it may stimulate non-modeled dynamics and lead to instability [23].
The presentation of this theory and its applications would easily constitute an-
other book in itself. Therefore, in what follows, we swiftly present this technique
and we refer the reader to [8, 22, 23, 60, 72, 83] for an excellent presentation of
first order sliding modes and of the Fillipov theory for differential equations with
discontinuous second member as well as of the equivalent vector method of Utkin.
A.2.5.1 Sliding Modes of Order One
Even though the theory of sliding modes is applied to a large class of systems of the
form ẋ = f (x, u) [69], we shall restrict our attention to the class of single-output
control affine systems of the form
ẋ = f (x) + g(x)u (A.27)
where x = (x1 , . . . , xn )T belongs to χ , an open set of Rn , u is the input and f, g are

sufficiently differentiable functions. We define a sufficiently differentiable function
s : χ × R+ → R such that ∂x ∂s
is non-zero on χ . The set

S = x ∈ χ : s(t, x) = 0 (A.28)
is a submanifold of χ of dimension (n − 1), called the sliding surface. The function

s is called the sliding function.
Remark A.9 The systems studied here are governed by differential equations involv-
ing discontinuous terms. The classical theories do not allow to describe the behavior
of the solution in this case. One must, therefore, employ the theory of differential
inclusions [3] and the solutions in the Fillipov sense [22].
Definition A.8 [84] We say that there exists an ideal sliding mode on S if there
exists a finite time ts such that the solution of (A.27) satisfies s(t, x) = 0 for all
t ≥ ts .
The existence of the sliding mode is guaranteed by sufficient conditions: the slid-
ing surface must be locally attractive, which can be mathematically translated as
∂s ∂s
lim (f + gu) < 0 and lim (f + gu) > 0 (A.29)
s→0+ ∂x s→0− ∂x
This condition translates the fact that, in a neighborhood of the sliding surface, the
velocity vectors of the trajectories of the system must point towards this surface, see
Fig. A.9 [8].
Hence, once the surface is intersected, the trajectories stay in an ε-neighborhood
of S, and we say that the sliding mode is ideal if we have exactly s(t, x) = 0.
Fig. A.9 Attractivity of the

surface
The condition (A.29) is often written in the form
s ṡ < 0 (A.30)
and is called the attractivity condition.

The control u is constructed such that the trajectories of the system are brought
towards the sliding surface and are then maintained in a neighborhood of the latter.
u is a variable control law defined as follows:

+
u (x) if s(t, x) > 0,
u= u+ = u− (A.31)
u− (x) if s(t, x) < 0,
with u+ and u− being continuous functions. It must be noted that it is this discontin-
uous characteristic of the control law that permits to obtain a convergence in finite
time on the surface as well as the properties of robustness with respect to certain
perturbations.
An example of a classical control by sliding mode that ensures the convergence
towards the surface s = 0 in finite time is as follows: if for the nonlinear system
(A.13) of relative degree r, we have |Lg Lr−1
f | > K > 0, Lf < M < ∞ then there
r
exists λ > 0 such that the control [4]

u = − sign Lg Lr−1
f λ sign(s) (A.32)
ensures the convergence of s to 0 in finite time.
Remark A.10 Often, we assume that Lg h is positive. In this case, it is sufficient to

take
u = −λ sign(s). (A.33)
Whenever that is not the case, it is more accurate to consider the expression (A.32).
A.2.5.2 Convergence in Finite Time
When on one hand, the control is chosen of the form (A.32) or simply of the form
(A.33), and on the other hand, the previous conditions for the boundedness of certain
functions are verified then the convergence in finite time is ensured. We shall try to
demonstrate this result through an example.
Example A.1 [4] Consider the following simple example:
ẋ = b + u
(A.34)
u = −λ sign(x − xd )
with xd the desired state, s = x − xd the sliding surface and λ > |b| + sup|ẋd |, then
x converges to xd in finite time and remains on the surface x = xd .
Proof
s = x − xd
ṡ = b − λ sign(s) − ẋd
s2
Consider the Lyapunov function: V = 2. In this case, we have

V̇ = s b − λ sign(s) − ẋd
if λ > |b| + sup|ẋd | then V̇ < 0.

Hence, the convergence is demonstrated. It now remains to show that the conver-
gence is achieved in finite time.
Since V̇ < 0, there exists a constant K > 0 such that V̇ < −K|s|.
2 √ √ √
Now V = s2 ⇒ |s| = 2V therefore V̇ < − 2K V ,
√ √
Set K1 = − 2K ⇒ V̇ < −K1 V , let us take √ the worst case where the maxi-
mum convergence time is the limit case: V̇ = −K1 V
The solution of this equation gives
V −1/2 dV = −K1 dt
2V 1/2 = −K1 t + V0

−K1 t + V0 2
V (t) =
2
V0
The time from which V (t) = 0 corresponds to t = K1 , which is finite.
A.2.5.3 The Chattering Phenomenon
In practice, an ideal sliding mode does not exist since it will imply that the control
can switch with an infinite frequency. There then occurs the problem of chattering
which means that we no longer have s(t, x) = 0 but s(t, x) < Δ from t > t0 where
t0 is the convergence time and Δ a constant representing the maximum variations
along the ideal trajectory s = 0.
This maximum depends on the “slew rate ”of the components intervening in the
injection of the input u in the system, on wear, and on the sensitivity of actuator
Fig. A.10 Chattering

phenomenon
Fig. A.11 Saturation

function
Fig. A.12 Sigmoid function
noise in the case of an analog control hence limiting the variation of speed between
u+ and u− , see Fig. A.10. In the discrete case, the switching speed is limited by
the data measurement which is in turn constrained by the sampling period and the
computation time [8].
This phenomenon constitutes a non-negligible disadvantage because even if it is
possible to filter the output of the process, it is susceptible to excite high-frequency
modes that were not taken into account in the model of the system. This can degrade
the performances and even lead to instability [29].
The chattering also implies important mechanical requirements at the actuator
level. This can cause rapid wear and tear of the actuators as well as non-negligible
energy loss at the power circuits level. Numerous studies have been carried out
in order to reduce this phenomenon. One of them consists in replacing the sign
function by saturation functions Fig. A.11, or by sigmoid functions such as tan(s)
or arctan(s), Fig. A.12 [8].
Nevertheless, it has been proven that to overcome this chattering phenomenon the
best solution is to consider higher order sliding modes such as the twisting algorithm
or the super twisting [25, 60].
A.3 Summary 121
Fig. A.13 Architecture of

multi controllers
A.2.6 Control Design Technique Based on the Switching Between

Several Controllers
The control design techniques based on the switching between several controllers
have been the subject of intensive applications these last few years. The importance
of such methods comes from the existence of systems that are not stabilizable by a
single controller. In effect, a large range of dynamical systems is modeled by a fam-
ily of continuous subsystems and a logic rule orchestrating the switching between
these subsystems, see Fig. A.13
Based on this, switching systems appear as a rigorous concept for studying com-
plexes systems, even if their theoretical properties are still the subject of intensive
research.
A.3 Summary
This appendix has been devoted to the presentation of the theoretical aspects on sta-
bility and control of nonlinear and switching systems. There is no general method-
ology for the design of controller for nonlinear systems as opposed to controller
design for linear systems. Depending on the class of nonlinear systems under study,
some approaches are better suited than others. In addition, we have attempted to
explain in a simple way the principle of some nonlinear control design techniques
that fall within the scope of this book with the aim of using some of them for the
stabilization of underactuated mechanical systems.
Appendix B
Limits of Linearization and Dangers
of Destabilization
A common practice of automatic community is to assume that a system can be

described by a set of differential equations around an operating point as follows:
ẋ = Ax + Bu
(B.1)
y = Cx
Assuming that (B.1) describes the system behavior, we can then exploit linear con-
trol design procedures, where powerful analysis and design tools are available. How-
ever, nonlinear system behaviors can be more complex than what can be represented
by an equivalent linear model.
Neglecting such behaviors, unpredictable instability may arise and may cause
performance degradation. Moreover, the obtained linear system is valid only around
the considered operating point. Hence, it can describe the system only in the neigh-
borhood of this point. On the other hand, some phenomena such as Coulomb fric-
tion, backlash, and hysteresis, called hard nonlinearities, cannot be captured by lin-
ear equations. Therefore, these nonlinearities are neglected.
Additional nonlinear phenomena include finite escape time, multiple equilibrium
points, limit cycles, and chaos. A more complete description of these phenomena
and others is given in [18, 37].
To illustrate the impact of the loss of information due to linearization, let us
consider the following examples [20]:
Example B.1 Several equilibrium points:

ẋ = −x + x 2
(B.2)
x(t = 0) = x0
After linearization around x(t) = 0, the obtained dynamic and associated solution
are given by
ẋ = −x(t)
(B.3)
x(t) = x0 e(−t)

Systems, DOI 10.1007/978-3-319-02636-7,
124 B Limits of Linearization and Dangers of Destabilization
Fig. B.1 Responses of a

nonlinear system for several
initial conditions
Equation (B.3) indicates that for any initial condition x0 , the solution exponen-
tially converges towards the equilibrium point.
However, according to (B.2), the nonlinear system possesses a second equilib-
rium point at x(t) = 1.
The impact of negligence of this point can be illustrated by calculating the solu-
tion of the nonlinear system:
x0 e−t
x(t) = (B.4)
1 + x0 (e−t − 1)
according to (B.4), note that:

For x0 < 1, the solution tends to 0 when t → ∞ like for the linear case.
For x0 > 1, the solution explodes to infinity in finite time, see Fig. B.1.
Example B.2 The linearized system is not controllable.

Consider the following unicycle robot model (see [7] for other models):
⎡ ⎤ ⎡ ⎤
x˙1 cos x3 0
⎣ x˙2 ⎦ = ⎣ sin x3 0 ⎦ u1 (B.5)
u2
x˙3 0 1
Clearly, (B.5) is controllable while the linearized system around the point
x3 (t) = 0 given by
⎡ ⎤ ⎡ ⎤
x˙1 1 0
⎣ x˙2 ⎦ = ⎣ 0 0 ⎦ u1 (B.6)
u2
x˙3 0 1
is not controllable for x2 (t)!
Other examples of performance degradation are given in [20].

B Limits of Linearization and Dangers of Destabilization 125
On the other hand, the use of linear controller can sometimes lead to destabiliza-
tion; for example, the consequence of the peaking phenomenon on a linear system
can lead to the system instability [78, 80].
To illustrate this concept let us consider the partially linear coupled system de-
scribed by the dynamics:
ẋ = f (x, y)
(B.7)
ẏ = Ay + By
Let us make the following assumptions:

(b1) The pair (A, B) is supposed controllable.
(b2) The nonlinear function f is differentiable to first order with respect to the time.
(b3) The origin is an GAS equilibrium point for the zero dynamic ẋ = f (x, 0).
According to (B.7), and assumption (b3), it seems rather clear intuitively that a
linear controller can be designed to lead the dynamics y(t) to 0 exponentially such
that the zero dynamic of the nonlinear system is GAS. However, this strategy can
lead the nonlinear dynamics to instability and its trajectory may escape to infinity in
finite time. For example consider the following system:
Example B.3 Finite escape time:

−(1 + y2 ) 3
ẋ = x (B.8)
2
y˙1 = y2
y˙2 = u
From (B.8), one can verify that the assumption (b3) is satisfied.
By designing a linear controller as follows:
u = −a 2 y1 − 2ay2 (B.9)
multiple eigenvalues for the closed system result at −a.

Linear analysis tools can be used to find the exact solution y2 (t) given by
y2 = −a 2 te−at (B.10)
and from this solution it appears that the dynamic |y2 (t)| rises to a peak, then con-
verges exponentially to 0. By computation, we can show that the peak time is t = a1 .
From (B.10), we can conclude that for important values of a, y converges faster
towards 0. Hence, from assumption (b3), it seems that important values of a allow
a fast stabilization of the nonlinear system.
Nevertheless, in [38], it was shown that this is not true. Indeed, if we substitute
(B.10) in (B.8). By integrating, the resulting expression is given by
x02
x 2 (t) = (B.11)
1 + x02 (t + (1 + at)e−at − 1)
126 B Limits of Linearization and Dangers of Destabilization
The peaking phenomenon destabilization effect is now apparent when we replace

the values of x0 , a and t in (B.11). For example, for a = 10 and x02 = 2.176, the
response x 2 (t ∼
= 0.5) becomes unbounded and we have escape to infinity in finite
time.
Other examples and discussion of this phenomenon are in [38, 78, 80].
Appendix C
A Little Differential Geometry
This section is devoted to the definition of some concepts and basic tools of dif-
ferential geometry introduced in nonlinear automatic control theory, since the early
1970s, by Eliott, Lobry, Hermann, Krener, Brockett, and others.
Diffeomorphism A diffeomorphism is a nonlinear change of coordinates z =

Φ(x) where Φ is a vectorial function
⎛ ⎞
Φ1 (x1 , . . . , xn )
⎜ Φ2 (x1 , . . . , xn ) ⎟
⎜ ⎟
Φ(x) = ⎜ .. ⎟
⎝ . ⎠
Φn (x1 , . . . , xn )
with the following properties:
• Φ(x) is a bijective application
• Φ(x) and Φ −1 are differentiable applications.
If these properties are verified for all x ∈ Rn then Φ is a global diffeomorphism.
Otherwise, Φ is a local diffeomorphism.
Proposition C.1 If the Jacobian matrix of Φ, evaluated at the point x = x0 is

nonsingular, then Φ(x) is a local diffeomorphism.
Lie Derivative and Bracket Let f and g be two vector fields on an open Ω of
Rn with all continuous partial derivatives, and denote by ∂f ∂g
∂x and ∂x the Jacobian
matrices.
The Lie derivative of g along f is the vector field
∂g
Lf g = f.
∂x
The Lie bracket of f and g is the vector field
[f, g] = Lf g − Lg f.

Systems, DOI 10.1007/978-3-319-02636-7,
128 C A Little Differential Geometry
We define also the vector fields
adff g = [f, g]

adkf g = f, adk−1
f g , k = 2, 3, . . .
Distribution, Involutivity and Complete Integrability

• A distribution Δ on a manifold M assigns to each point x ∈ M a subspace of the
tangent space T .
• A set of vectors {g1 , . . . , gm } in Ω is said involutive if for all the i and j the
bracket [gi , gj ] is a linear combination of vectors g1 , . . . , gm , that is, there exist
functions αijk defined in Ω such that

k=m
[gi , gj ] = αijk gk .
k=1
That is, if for all the f and g in Δ, [f, g] belongs to Δ (Δ is closed by Lie
bracket).
• A set of linearly independent vectors {g1 , . . . , gm } is a complete integrable set, if
the system of n − m partial derivative equations
∂h ∂h
g1 = 0, ..., gn−m = 0
∂x ∂x
admits a solution h : Ω → Rn such that ∂h

∂x = 0
Theorem C.1 ([70] (Frobenius)) A set of linearly independent vectors {g1 , . . . , gm }

is involutive if and only if it is completely integrable.
For proof and examples see [37].
Relative Degree The relative degree associated with the system
ẋ = f (x) + g(x)u
(C.1)
y = h(x)
in a region Ω ⊂ Rn is given by the integer γ such that

γ −2
Lg h(x) = Lg Lf h(x) = · · · = Lg Lf h(x) = 0
γ −1
Lg L f h(x) = 0
for all x ∈ Ω.
Appendix D
Controllability of Continuous Systems
One of the main goals of automatic control is to establish control laws so that a
system evolves according to a predetermined objective. This requires controllability
of the system. Intuitively, the controllability means that we can bring a system from
one state to another by means of a control. Conversely, non-controllability implies
that some states are unreachable for any control.
D.1 Controllability of Linear Systems
For controlled linear systems
ẋ = Ax + Bu (D.1)
y = Cx (D.2)
where An×n is the state matrix, x ∈ Rn is the vector states, Bn×m the control matrix,
u controls belonging to a set of admissible controls U, Cp×n the output matrix and
y ∈ Rp the system outputs.
Definition D.1 [35] The system (D.1) is controllable if for each couple (x0 , xd ) of
Rn there exist a finite time T and a control u defined on [0, T ] that brings the system
from an initial state x(0) = x0 to a desired state x(T ) = xd .
D.1.1 Kalman Controllability Criterion
An algebraic characterization of linear systems controllability, due to Kalman, is

given as follows:

Systems, DOI 10.1007/978-3-319-02636-7,
130 D Controllability of Continuous Systems
Fig. D.1 Static state

feedback stabilization
Theorem D.1 The linear system (D.1) is controllable if and only if the rank of its
controllability matrix

C = B AB . . . An−1 B (D.3)
is equal to n. We say that the pair (A, B) is controllable.
Details and proofs can be found in [17].

For controllable linear systems, we seek to design a controller that makes the
origin asymptotically stable. Several approaches are available, for example, we can
design control laws by state feedback.
D.1.2 State Feedback Stabilization
A linear state feedback or controller for (D.1) is a control law
u(t) = −Kx(t) (D.4)
where Km×n is a gain matrix.

When the value of u(t) at t depends only on x(t) then the feedback is called
static feedback, Fig. D.1.
The gain matrix can be computed in several ways, for example by pole place-
ment.
Pole Placement Design When a system is controllable, the pole placement prin-
ciple consists of determining a control law u = −Kx such that σ (A − BK) = σd ,
where σ is the spectrum of (A − BK) and σd is the desired spectrum.
The difficulty of this approach lies in the determination of the spectrum since
there is no general methodology for doing so. This method offers the possibility to
place the closed-loop poles anywhere in the negative half-plane, regardless of the
open-loop poles location. As a result, the response time can be controlled. However,
if the poles are placed too far into the negative half-plane, the values of the gain K
are very large and can cause saturation problems and can lead to instability.
Remark D.1 The control law u is designed assuming that the state vector x is avail-
able. This assumption is not always true. Sometimes, some states are not available,
D.2 Controllability Concepts for Nonlinear Systems 131
because it is either difficult or impossible to physically measure these states or it

is too expensive. In this case, we proceed to a reconstruction of the missing states
using observers. However, throughout this book we are interested in the problem of
control under the assumption that states are measurable.
D.2 Controllability Concepts for Nonlinear Systems
The notion of controllability which seems simple and intuitive for linear systems
is rather complicated for nonlinear systems where several definitions of the latter
exist. The first results on nonlinear system controllability are due to Sussmann and
Jurdjevic [79], Lobry [48], Hermann and Krener [30], Sussmann [76, 77] and for a
nice presentation see also Nijmeijer and Van der Schaft [55].
A nonlinear system is generally represented by
ẋ = f (x, u)
(D.5)
y = h(x)
where x ∈ M ⊂ Rn , u ∈ Rm , y ∈ Rp , and f, h are C ∞ .
Definition D.2 Let U be a subset of M and let (x0 , xd ) ∈ U . We say that xd is

U -accessible from x0 which we denote by xd Au x0 , if there exist a measurable and
bounded control u and a finite time T , such that the solution x(t) of (D.5), for
t ∈ [0, T ], satisfies
x(0) = x0 , x(T ) = xd and x ∈ U for t ∈ [0, T ]
we denote by A(x0 ) the set of points in M accessible from x0 :
A(x0 ) = {x ∈ M/ xAM x0 } (D.6)
Definition D.3 The system (D.5) is controllable at x0 if A(x0 ) = M and is control-

lable if A(x0 ) = M for all x ∈ M.
When a system is controllable at x0 , it may be necessary to cover a considerable

distance or time for reaching a point near x0 . This leads us to introduce a local
version of the concept of controllability.
Definition D.4 The system (D.5) is said locally controllable at x0 , if for all neigh-
borhood U of x0 , Au (x0 ) is a neighborhood of x0 , where
Au (x0 ) = {x ∈ U/ xAu x0 } (D.7)
and it is said to be locally controllable if it is locally controllable for all x ∈ M.

Fig. D.2 Relationship between the nonlinear controllability concepts
We can weaken the controllability notion:
Definition D.5 The system is said to be weakly controllable at x0 , if W A(x0 ) = M,

it is said to be weakly controllable if it is weakly controllable for all x ∈ M.
Remark D.2 W Au is the smallest set containing U-accessible pairs (that is,

x W Au x if and only if there exist x 0 , . . . , x k , such that x 0 = x , x k = x and let
x i Au x i−1 or x i−1 Au x i for i = 1, . . . , k).
The concept of weak controllability is a global concept which does not reflect
the behavior of a system in the neighborhood of a point. Therefore, it is necessary
to introduce the concept of weak local controllability
Definition D.6 The system (D.5) is said to be locally weakly controllable at x0 if

for all neighborhood U of x0 , W Au (x0 ) is a neighborhood of x0 and it is said to be
locally weakly controllable if it is for all x ∈ M.
Different notions of controllability are illustrated by the diagram of Fig. D.2.

For control affine nonlinear systems described by
ẋ = f (x) + g(x)u
(D.8)
y = h(x)
the controllability rank condition is defined as
Definition D.7 System (D.8) satisfies the rank condition if the rank of the nonlinear
controllability matrix

Cfg = g(x) adf g(x) ad2f g(x) . . . adn−1 f g(x) (D.9)
is equal to n for all x.
Theorem D.2 [30] If the system (D.8) satisfies the rank condition then it is locally
weakly controllable.
This theorem highlights the advantage of weak local controllability compared to

previous forms of controllability since verification of such a concept is reduced to a
simple algebraic criterion.
References 133
References
1. D.V. Anosov, On stability of equilibrium points of relay systems. J. Autom. Remote Control
2,135–149 (1959)
2. P.J. Antsalkis, A.N. Michel, Linear Systems (McGraw-Hill, New York, 1997)
3. J-P. Aubin, H. Frankowska, Set-Valued Analysis (Birkhäuser, Basel, 1990)
4. J.P. Barbot, Systèmes à structure variable. Technical report, École Nationale Supérieure
d’Electronique et de Ses Applications, ENSEA, France, 2009
5. N. Bedrossian, Approximate feedback linearisation: the cart pole example, in Proc. IEEE Int.
Conf. on Robotics and Automation (1992), pp. 1987–1992
6. I. Belgacem, Automates hybrides Zénon: Exécution-simulation. Master at Aboubekr Belkaid
University, Tlemcen, Algeria, 2009
7. A.M. Bloch, M. Reyhanoglu, N.H. McClamroch, Control and stabilization of nonholonomic
dynamic systems. IEEE Trans. Autom. Control 37(11),1746–1757 (1992)
8. T. Boukhobza, Contribution aux formes d’observabilité pour les observateurs à modes glis-
sants. Ph.D. thesis, Paris Sud Centre d’Orsay University, France, 1997
9. M.S. Branicky, Stability of switched and hybrid systems, in Proc. 33rd IEEE Conf. on Decision
and Control, USA (1994), pp. 3498–3503
10. M.S. Branicky, Studies in hybrid systems: modeling, analysis, and control. Ph.D. thesis, De-
partement of Electrical and Computer Engineering, Massachusetts Institute of Technology,
1995
11. M.S. Branicky, Stability of hybrid systems: State of the art, in Proc. 36th IEEE Conf. on
Decision and Control, USA (1997), pp. 120–125
12. M.S. Branicky, Multiple Lyapunov functions and other analysis tools for switched and hybrid
systems. IEEE Trans. Autom. Control 43(4), 475–482 (1998)
13. R.W. Brockett, Asymptotic Stability and Feedback Stabilization (Birkhäuser, Basel, 1983)
14. K. Busawon, Lecture notes in control systems engineering. Technical report, Northumbria
University, Newcastle, United Kingdom, 2008
15. S. Canudas, G. Bastin, Theory of Robot Control (Springer, Berlin, 1996)
16. G. Chesi, Y.S. Hung, Robust commutation time for switching polynomial systems, in Proc.
46th IEEE Conf. on Decision and Control, USA (2007)
17. B. D’Andréa Novel, M.C. De-Lara, Commande Linéaire des Systèmes Dynamiques (École des
Mines Press, 2000)
18. H. Dang Vu, C. Delcarte, Bifurcation et Chaos (Ellipses edition, 2000)
19. R. DeCarlo, M. Branicky, S. Petterssond, B. Lennartson, Perspectives and results on the sta-
bility and stabilizability of hybrid systems. Proc. IEEE 88, 1069–1082 (2000)
20. W.E. Dixon, A. Behal, D. Dawson, S. Nagarkatti, Nonlinear Control of Engineering Systems
(Birkhäusser, Basel, 2003)
21. S.V. Emel’yanov, Variable Structure Control Systems (Nauka, Moscow, 1967)
22. A.G. Fillipov, Differential Equations with Discontinuous Right Hand-Sides. Mathematics and
its Applications (Kluwer Academic, Dordrecht, 1983)
23. T. Floquet, Contributions à la commande par modes glissants d’ordre supérieur. Ph.D. thesis,
University of Sciences and Technology of Lille, France, 2000
24. G.F. Franklin, J.D. Powell, A. Emami-Naeini. Feedback Control of Dynamic Systems
(Prentice-Hall, Englewood Cliffs, 2002)
25. L. Fridman, A. Levant, Sliding Modes of Higher Order as a Natural Phenomenon in Control
Theory (Springer, Berlin, 1996)
26. L. Gurvits, R. Shorten, O. Mason, On the stability of switched positive linear systems. IEEE
Trans. Autom. Control 52, 1099–1103 (2007)
27. J. Hauser, Nonlinear control via uniform system approximation. Syst. Control Lett. 17, 145–
154 (1991)
28. J. Hauser, S. Sastry, P. Kokotović, Nonlinear control via approximate input-output lineariza-
tion. IEEE Trans. Autom. Control 37(3), 392–398 (1992)
29. B. Heck, Sliding mode control for singulary perturbed systems. Int. J. Control 53(4) (1991)
30. R. Hermann, A.J. Krener, Nonlinear controllability and observability. IEEE Trans. Autom.
Control 22, 728–740 (1977)
31. L.R. Hunt, R. Su, G. Meyer Gurvits, R. Shorten, O. Mason, Global transformations of nonlin-
ear systems. IEEE Trans. Autom. Control 28(1), 24–31 (1983)
32. A. Isidori, Nonlinear Control Systems (Springer, Berlin, 1995)
33. B. Jakubczyk, W. Respondek, On linearisation of control systems. Rep. Sci. Pol. Acad. 28(9),
517–522 (1980)
34. J. Jouffroy, Stabilité et systèmes non linéaire: réflexion sur l’analyse de contraction. Ph.D.
thesis, Savoie University, France, 2002
35. T. Kailath, Linear Systems (Prentice-Hall, Englewood Cliffs, 1981)
36. W. Kang, Approximate linearisation of nonlinear control systems, in Proc. IEEE Conf. on
Decision and Control (1993), pp. 2766–1771
37. H.K. Khalil, Nonlinear Systems (Prentice-Hall, Englewood Cliffs, 2002)
38. P. Kokotović, The joy of feedback: Nonlinear and adaptive. IEEE Control Syst. Mag. 12(3),
7–17 (1992)
39. A.J. Krener, Approximate linearisation by state feedback and coordinate change. Syst. Control
Lett. 5, 181–185 (1984)
40. A.J. Krener, M. Hubbard, S. Karaban, B. Maag, Poincaré’s Linearisation Method Applied to
the Design of Nonlinear Compensators (Springer, Berlin, 1991)
41. M. Kristić, P. Kokotović, Nonlinear and Adaptive Control Design (Wiley, New York, 1995)
42. H. Kwakernak, R. Sivan, Linear Optimal Control Systems (Library of Congress, 1972)
43. H.G. Kwatny, G.L. Blankenship, Nonlinear Control and Analytical Mechanics: A Computa-
tional Approach (Birkhäuser, Basel, 2000)
44. F. Lamnabhi Lagarrique, P. Rouchon, Commandes Non Linéaires (Lavoisier, Paris, 2003)
45. M. Latfaoui, Linéaisation approximative par feedback. Master at Aboubekr Belkaid Univer-
sity, Tlemcen, Algeria, 2002
46. J. Lévine, Analyse et commande des systèmes non linéaires. Technical report, École des Mines
de Paris, France, 2004
47. D. Liberzon, A.S. Morse, Basic problems in stability and design of switched systems. IEEE
Control Syst. Mag. 19(5), 59–70 (1999)
48. C. Lobry, Contrôlabilité des Systèmes Non Linéaires (CNRS, Paris, 1981)
49. D.G. Luenberger, Introduction to Dynamic Systems (Wiley, New York, 1979)
50. N. Manamanni, Aperçu rapide sur les systèmes hybrides continus. Technical report, University
of Reims, France, 2007
51. M. Margaliot, M.S. Branicky, Nice reachability for planar bilinear control systems with appli-
cations to planar linear switched systems. IEEE Trans. Autom. Control (May 2008)
52. O. Mason, R. Shorten, A conjecture on the existence of common quadratic Lyapunov functions
for positive linear systems, in Proc. American Control Conf., USA (2003), pp. 4469–4470
53. S. Mastellone, D.M. Stipanović, M.W. Spong, Stability and convergence for systems with
switching equilibria, in Proc. 46th IEEE Conf. on Decision and Control, USA (2007),
pp. 4013–4020
54. A.N. Michel, L. Hou, Stability results involving time-averaged Lyapunov function derivatives.
J. Nonlinear Anal., Hybrid Syst. 3(1), 51–64 (2009)
55. H. Nijmeijer, A. van der Schaft, Nonlinear Dynamical Control Systems (Springer, Berlin,
1990)
56. K. Ogata, Modern Control Engineering (Prentice-Hall, Englewood Cliffs, 1997)
57. R. Ortega, A. Loria, P. Nicklasson, H. Sira-Ramirez, Passivity-Based Control of Euler La-
grange Systems (Springer, Berlin, 1998)
References 135
58. P. Peleties, Modeling and design of interacting continuous-time/discrete event systems. Ph.D.
thesis, Electrical Engineering, Purdue Univ., West Lafayette, IN, 1992
59. P. Peleties, R.A. DeCarlo, Asymptotic stability of m-switched systems using Lyapunov-like
functions, in Proc. American Control Conf., USA (1991), pp. 1679–1684
60. W. Perruquetti, J.P. Barbot, Sliding Mode Control in Engineering (Taylor and Francis, London,
2002)
61. V.M. Popov, Absolute stability of nonlinear control systems of automatic control. J. Autom.
Remote Control 22, 857–875 (1962)
62. J.P. Richard, Mathématiques pour les Systèmes Dynamiques (Lavoisier, Paris, 2002)
63. H. Saadaoui, Contribution à la synthèse d’observateurs non linéaires pour des classes de sys-
tèmes dynamiques hybrides. Ph.D. thesis, Cergy Pontoise University, France, 2007
64. R.G. Sanfelice, A.R. Teel, R. Sepulchre, A hybrid systems approach to trajectory tracking con-
trol for juggling systems, in Proc. 46th IEEE Conf. on Decision and Control, USA (2007), pp.
5282–5287
65. T. Sari, C. Lobry, Contrôle Non Linéaire et Applications (Hermann, Paris, 2005)
66. S.S. Sastry, Nonlinear Systems: Analysis, Stability and Control (Springer, Berlin, 1999)
67. R. Sepulchre, M. Janković, P. Kokotović, Contructive Nonlinear Control (Springer, Berlin,
1997)
68. C. Shen, Q. Wei, F. Shumin, On exponential stability of switched systems with delay: multiple
Lyapunov functions approach, in Proc. Chineese Control Conf., China (2007), pp. 664–668
69. H. Sira Ramirez, Sliding regimes in general nonlinear systems: a relative degree approach. Int.
J. Control 50(4), 1487–1506 (1989)
70. J. Slotine, W. Li, Nonlinear Systems Analysis (Prentice-Hall, Englewood Cliffs, 1993)
71. E.D. Sontag, Y. Wang, On characterizations of the input-to-state stability property. Syst. Con-
trol Lett. 24, 351–359 (1995)
72. S.K. Spurgeon, C. Edwards, Sliding Mode Control: Theory and Applications (Taylor and Fran-
cis, London, 1983)
73. T. Sugie, K. Fujimoto, Control of inverted pendulum systems based on approximate linearisa-
tion, in Proc. IEEE Conf. on Decision and Control (1994), pp. 1647–1649
74. Z. Sun, S.S. Ge, Analysis and synthesis of switched linear control systems. Automatica 41,
181–195 (2005)
75. Z. Sun, S.S. Ge, Switched Linear Systems (Springer, Berlin, 2005)
76. H.J. Sussmann, Lie Brackets, Real Analyticity and Geometric Control (Birkhäuser, Basel,
1983)
77. H.J. Sussmann, A general theorem on local controllability. SIAM J. Control Optim. 25(1),
158–194 (1987)
78. H.J. Sussmann, Limitations on stabilizability of globally minimum phase systems. IEEE
Trans. Autom. Control 35(1), 117–119 (1990)
79. H.J. Sussmann, V. Jurdjevic, Controllability of nonlinear systems. J. Differ. Equ. 12(2), 95–
116 (1991)
80. H.J. Sussmann, P. Kokotović, The peaking phenomenon and the global stabilization of nonlin-
ear systems. IEEE Trans. Autom. Control 36(4), 424–440 (1991)
81. C.J. Tomlin, J. Lygeros, S. Sastry, Introduction to Dynamic Systems (Springer, Berlin, 2003)
82. Y.Z. Tzypkin, Theory of Control of Relay Systems (Gostekhizdat, Moscow, 1955)
83. V.I. Utkin, Variable structure systems with sliding mode. IEEE Trans. Autom. Control 22(2),
212–222 (1977)
84. V.I. Utkin, Sliding Modes in Control Optimization (Springer, Berlin, 1992)
85. A. van der Schaft, H. Schumacher, An Introduction to Hybrid Dynamical Systems (Springer,
Berlin, 2002)
86. A. van der Shaft, L2 Gain and Passivity Techniques in Nonlinear Control (Springer, Berlin,
2000)
87. M. Vidyasagar, Nonlinear Systems Analysis (Prentice-Hall, Englewood Cliffs, 1993)

88. D.A. Voytsekhovsky, R.M. Hirschorn, Stabilization of single-input nonlinear systems using
higher order compensating sliding mode control, in Proc. 45th IEEE Conf. on Decision and
Control and the European Control Conf. (2005), pp. 566–571
89. J.C. Willems, Dissipative dynamical systems. Rational mechanics and analysis 45, 321–393
(1972)
90. X. Xu, G. Zhai, Practical stability and stabilization of hybrid and switched systems. IEEE
91. H. Ye, A.N. Michel, L. Hou, Stability analysis of discontinuous dynamical systems with ap-
plication, in 13th IFAC Congress, USA (1996), pp. 461–466
92. H. Ye, A.N. Michel, L. Hou, Stability theory for hybrid dynamical systems. IEEE Trans. Au-
tom. Control 43(4), 461–474 (1998)
93. S.H. Zak, Systems and Control (Oxford University Press, London, 2003)
94. G. Zames, On the input-output stability of nonlinear time-varying feedback systems. IEEE
95. G. Zhai, B. Hu, K. Yasuda, A.N. Michel, Stability analysis of switched systems with stable
and unstable subsystems: an average dwell time approach, in Proc. American Control Conf.,
USA (2000), pp. 200–204
96. G. Zhai, D. Liu, J. Imae, T. Kobayashi, Lie algebraic stability analysis for switched systems
with continuous-time and discrete-time subsystems. IEEE Trans. Circ. Syst. 53(2), 152–156
(2006)
97. G. Zhai, H. Lin, X. Xu, A.N. Michel, Stability analysis and design of switched of normal
systems, in Proc. 43rd IEEE Conf. on Decision and Control, USA (2004), pp. 3253–3258
98. G. Zhai, I. Matsune, J. Imae, T. Kobayashi, A note on multiple Lyapunov functions and sta-
bility condition for switched and hybrid systems, in Proc. 16th IEEE Int. Conf. on Control
Application, Singapore (2007), pp. 1189–1199
99. J. Zhao, D.J. Hill, Dissipativity theory for switched systems. IEEE Trans. Autom. Control
53(4), 941–953 (2008)
Index
A E
Acrobot, 8, 9, 18, 27, 29, 42, 47, 64, 65 Energy, 9, 113
Approximation kinetic, 16
high order, 83, 85, 112 potential, 16, 28, 29
linear, 97, 106 Euler–Lagrange equation, 15, 17, 19, 27,
65
B Equilibrium, 94
Backstepping, 10, 11, 35, 40, 44, 49, 55, 64,
114 F
Ball and beam, 8, 18, 27, 31, 43, 48, 64, 83, Finite time convergence, 84, 118
85 Forwarding, 10, 11, 44, 49
BIBS, 56 Frobenius theorem, 109, 128
Brockett, 8, 10, 21, 22, 71 Function
Brunovsky form, 107, 112 Chetaev, 99
Lipscitzian, 94
C Lyapunov, 64, 97
Centrifugal term, 16 common, 101
CFD, 10, 35, 40, 55, 65 multiples, 74, 101
Chaos, 123 weak, 102
Chattering, 88, 119
Christoffel symbols, 16 G
Classification, 10, 35, 43, 55 Generalized momentum, 27, 50
class I, 47
class II, 47, 48 H
Complete integrability, 128 Hurwitz, 84
Constraint
holonomic, 19, 27 I
non-holonomic, 18, 27, 37 Inertial wheel pendulum, 9, 30, 65
Controllability, 8, 17, 21–23, 40, 108, 129, Inverted pendulum, 8, 9, 18, 28, 42, 48, 49, 65,
131 67
Coriolis, 16 Involutivity, 128
D J
Degree of complexity, 36, 40 Jacubezyk–Respondek theorem, 108
Diffeomorphism, 108, 109, 127
Distribution, 128 K
Kalman criterion, 129

Systems, DOI 10.1007/978-3-319-02636-7,
138 Index
L equilibrium, 95, 96
Lagrange vector of multipliers, 19 exponential, 96
LaSalle invariance principle, 98 global, 96
Legendre switching system, x, 100
normal form, 17 Stabilization, 8, 10, 11
transformation, 16 feedback, 130
Lie local, x, 106
bracket, 127 of a system, x, 105
derivative, 127 Structural properties, 8, 10, 19, 35, 50
Limit cycles, 123 Structure
Linearization chain, 10, 35, 39–41, 55
approximate, 82, 83, 110 isolated vertex, 10, 35, 39, 40, 43, 55, 82
feedback, 23, 40, 56, 107 tree, 10, 35, 39, 40, 42, 55, 64, 78
input–output, 108 Switch, 9, 71
input–state, 107 Switching control, 10, 71
limits, 123 Symmetry, 26
partial, 23, 43, 66, 68, 81, 109 kinetic, 27, 47, 49
collocated, 24 System
coupled inputs, 25 aeronautic, 18
non-collocated, 24 aeronautical, 7, 26, 27
LQR, 9, 71 autonomous, 94
chain, 20
N Lagrangian, 18, 26
Normal form, 109 marine, 8, 18
feedforward, 11, 48, 50, 53 mechanical, 7, 27
non-triangular, 11, 53 fully actuated, 21
non-triangular quadratic, 50 underactuated, 10, 27, 35
strict feedback, 11, 45–48, 50, 53, 66 minimum phase, 110
mobile, 8
O
non-autonomous, 94
Observability, 17
non-minimum phase, 21, 110
P nonlinear, 15, 19, 66, 96, 106
Passivity, x, 8, 9, 113 spatial, 7
Peaking phenomenon, 56, 125 switching, 121
Pendubot, 8, 9, 18, 27, 29, 42, 48, 64, 65, 67
Pendular systems, 64 T
Pole placement, 9, 71, 80, 130 Tora, 8, 9, 18, 27, 29, 42, 47, 64, 65, 67
Trajectory tracking, 10, 57, 105
R Triangular form, 114
Reduction, 49 Two order sliding mode, 88
Relative degree, 8, 21, 40, 41, 43, 108, 109
robust, 111 U
Robot Under water vehicles, 19
flexible, 8, 18 Underactuated mechanical systems, 55
joint elasticity, 41, 64 Unicycle, 18, 20, 124
walking biped, 10
wheeled, 18 V
Variable
S external, 27, 47, 49
Satellites, 8, 18, 27 shape, 27, 47, 49, 50, 90
Sliding mass on cart, 41, 64 VTOL, 9
Sliding mode, 10, 116
Stability, 93, 94 Z
asymptotic, 96 Zero dynamics, 17, 56, 110

Theoretical Background of Nonlinear System Stability and Control

Uploaded by

Copyright:

Available Formats

Theoretical Background of Nonlinear System Stability and Control

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Theoretical Background of Nonlinear System Stability and Control

Uploaded by

Copyright:

Available Formats

Appendix A

Theoretical Background of Nonlinear System

. . . Mr. Fourier had the opinion that the main purpose of

Automatic control comprises a number of theoretical tools of mathematical charac-

A.1 Stability of Systems

A. Choukchou-Braham et al., Analysis and Control of Underactuated Mechanical 93

A.1.1 What to Choose?

It is clear that drawing up an inventory as complete as possible of the forms of

where x ∈ Rn is the state vector and f : D → Rn a locally Lipschitzian function

A.1.2 The Lyapunov Stability Theory

The Lyapunov stability theory is considered as one of the cornerstones of automatic

Fig. A.1 Intuitive illustration

A.1.2.1 Stability of Equilibrium Points

Definition A.1 [37] The equilibrium point x = 0 is said to be:

x(0) ∈ B ⇒ lim x(t) = 0

Fig. A.2 Stability and

• globally attractive, if for every solution x(t) of (A.1) we have

lim x(t) = 0. In this case, B = Rn

• asymptotically stable, if it is stable and attractive, and globally asymptotically

A.1.2.2 Stability of the Origin for a Linear System

Consider the linear system

Theorem A.1 [37]

Unfortunately, there does not exist an equivalent theorem to that of eigenvalues

A.1.2.3 Linear Approximation of a System

Consider a system of the form (A.1); we denote by

Theorem A.2 [37]

A.1.2.4 Lyapunov’s Direct Method

The principle of this method is a mathematical extension of the following physi-

Lyapunov Function Consider the system

ẋ = f (x) with f (0) = 0 (A.4)

x = 0 is an equilibrium point for (A.4) and D ⊂ Rn is a domain that contains x = 0.

Definition A.2 The function V is a Lyapunov function for system (A.4 ) at x = 0

Theorem A.3 [37]

LaSalle’s Invariance Principle

Definition A.3 A set G ⊆ Rn is said to be positively invariant if every solution x(t)

Fig. A.3 Lyapunov

If x̄ is an equilibrium point then {x̄} is positively invariant.

Theorem A.4 ([37] (Lyapunov–LaSalle)) Let V : D → R+ be a function having

Theorem A.5 [37] Let V : Rn → R+ be a function having continuous partial

In some cases, a dynamical system is represented, at a given instant of time t ≥ t0 ,

A.1.3 Stability of Switching Systems

Mathematically speaking, a switching system can be described by equations of the

A.1.3.1 Common Lyapunov Function

A.1.3.2 Multiple Lyapunov Functions

Fig. A.4 Energy profile of

The condition (A.8) is illustrated by Fig. A.4.

where h : R+ → R+ is a continuous function with h(0) = 0 and tj is any transition

Fig. A.5 Energy profile of a

Fig. A.6 Energy profile of

Fig. A.7 Illustration of the

holds for a positive definite continuous function Wi (x).

Fig. A.8 Illustration of the

A.1.4 Stabilization of a System

The problem of stabilization consists in maintaining the system near an equilibrium

e(t) = y(t) − yd (t)

tends to 0 when t → ∞. In addition, the state must remain bounded.

A.2 Control Theory

Nonetheless, the control procedure is only possible if the system in question is

ẋ = f (x) + g(x)u, x ∈ Rn , u ∈ R (A.13)