Lectures in Feedback Design For Multivariable Systems-Springer International Publishing (2017)
Lectures in Feedback Design For Multivariable Systems-Springer International Publishing (2017)
AlbertoIsidori
Lectures in
Feedback
Design for
Multivariable
Systems
Advanced Textbooks in Control and Signal
Processing
Series editors
Michael J. Grimble, Glasgow, UK
Michael A. Johnson, Kidlington, UK
More information about this series at http://www.springer.com/series/4045
Alberto Isidori
123
Alberto Isidori
Dipartimento di Ingegneria Informatica,
Automatica e Gestionale
Universit degli Studi di Roma
La Sapienza
Rome
Italy
ISSN 1439-2232
Advanced Textbooks in Control and Signal Processing
ISBN 978-3-319-42030-1 ISBN 978-3-319-42031-8 (eBook)
DOI 10.1007/978-3-319-42031-8
v
vi Series Editors Foreword
information transfer between the participating systems that makes the control
design challenging. Leaderfollower systems and the coordinated control design
problem are one continuing interest at todays international control conferences.
The insightful, reflective nature of the textbook assumes that the reader has
already studied the basics of control theory. A guide to the knowledge that is
assumed of the reader can be found in the two useful appendices.
The text is based on Prof. Isidoris considerable teaching experience, including
many years of lecturing at the Sapienza University in Rome, and at various inter-
national guest lecture courses. It provides a thoughtful and insightful exposition of
where we stand today in control theory and is a very welcome and valuable addition
to the Advanced Textbooks in Control and Signal Processing series.
Michael J. Grimble
Michael A. Johnson
Industrial Control Centre
University of Strathclyde
Glasgow, Scotland, UK
Preface
vii
viii Preface
Hangzhou. Encouragement and support from Francoise Lamnabhi and Hongye Su,
respectively, are gratefully acknowledged. In particular, while at Zhejiang
University, I have been able to establish a fruitful collaboration with Lei Wang, the
outcome of which is reflected in the material presented in Chaps. 10 and 11.
1 An Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Robust Stabilization of Linear Systems . . . . . . . . . . . . . . . . . . 2
1.3 Regulation and Tracking in Linear Systems. . . . . . . . . . . . . . . 6
1.4 From Regulation to Consensus . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Feedback Stabilization and State Observers
for Nonlinear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.6 Robust Stabilization of Nonlinear Systems . . . . . . . . . . . . . . . 12
1.7 Multi-input Multi-output Nonlinear Systems . . . . . . . . . . . . . . 13
1.8 Regulation and Tracking in Nonlinear Systems . . . . . . . . . . . . 17
ix
x Contents
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
Chapter 1
An Overview
1.1 Introduction
This is a book intended for readers, familiar with the fundamentals of linear system
and control theory, who are interested in learning methods for the design of feed-
back laws for (linear and nonlinear) multivariable systems, in the presence of model
uncertainties. One of the main purposes of the book is to offer a parallel presentation
for linear and nonlinear systems. Linear systems are dealt with in Chaps. 25, while
nonlinear systems are dealt with in Chaps. 612. Among the various design options
in the problem of handling model uncertainties, the focus of the book is on methods
that appealin one form or in anotherto the so-called Small-Gain Theorem. In
this respect, it should be stressed that, while some of such methods may require, for
their practical implementation, a high-gain feedback on selected measured vari-
ables, their effectiveness is proven anyway with the aid of the Small-Gain Theorem.
Methods of this kind lend themselves to a presentation that is pretty similar for
linear and nonlinear systems and this is the viewpoint adopted in the book. While
the target of the book are multi-input multi-output (MIMO) systems, for pedagog-
ical reasons in some cases (notably for nonlinear systems) the case of single-input
single-output (SISO) systems is handled first in detail. Two major design problems
are addressed (both in the presence of model uncertainties): asymptotic stabilization
in the large (that is, with a guaranteed region of attraction) of a given equilibrium
point and asymptotic rejection of the effect of exogenous (disturbance and/or com-
mands) inputs on selected regulated outputs. This second problem, via the design of
an internal model of the exogenous inputs, is reduced to the problem of asymptotic
stabilization of a closed invariant set.
The book presupposes, from the reader, some familiarity with basic concepts in
linear system theory (such as reachability, observability, minimality), elementary
control theory (transfer functions, feedback loop analysis), and with the basic con-
cepts associated with the notion of stability of an equilibrium in a nonlinear system.
Two Appendices collect some relevant background material in this respect, to the
purpose of fixing the appropriate notations and making the presentationto some
extentself-contained.
Chapters 2 and 3 of this book deal with the design of control laws by means of which
a linear system
x = Ax + Bu
(1.1)
y = Cx
x = (A + BF GC)x + Gy
u = F x
with F and G such that the matrices A + BF and, respectively, A GC have all
eigenvalues with negative real part, this method is not suitable in the presence model
uncertainties, because the design of the controller in question requires an accurate
knowledge of the matrices A, B, C that characterize the model (1.1).
There are various design methods by means of which stabilization in spite of
model uncertainties can be achieved, that dependamong other thingson the form
in which the model uncertainties are characterized. The methods described in these
two chapters are based on a classical result known as Small-Gain Theorem according
to which, if two stable systems are coupled to each other and the coupling is weak
(in a precise sense that will be defined) the resulting system is still stable. The reason
why these methods are preferred is that they find a natural counterpart in the same
design problem for nonlinear systems, as shown from Chap. 6 on.
The simpler example of a linear system subject to model uncertainties consists in
a system modeled by equations of the form (1.1), in which the state vector x, the input
vector u, and the output vector y have fixed dimensions but (some of) the entries of the
matrices A, B, C are known only within given (and possibly not negligible) tolerance
bounds. This case is usually referred to as the case of structured uncertainties. A
simple method by means of whichif the system has just one input and just one
outputthe effect of such uncertainties can be overcome is based on the following
elementary and well-known property. Suppose the transfer function
N(s)
T (s) =
D(s)
sN(s)
b = lim .
s D(s)
Then, it is well known from the theory of the root locus that if k is a real number
with the same sign as that of b, the n roots of polynomial
have the following property: as |k| , one (real) root diverges to while the
remaining n 1 roots converge to the roots of N(s) (that is to the zeros of T (s)).
Thus, if all zeros of T (s) have negative real part, the system can be made stable by
means of the simple control law
u = ky,
provided that |k| is sufficiently large. It is also well known that, if T (s) has nr zeros
(now with r > 1) all of which have negative real part, the same stability result can
be achieved by means of an output feedback controller characterized by the transfer
function
N (s)
Tc (s) = k
(1 + s)r1
in which N (s) is a monic polynomial of degree r 1 having all roots with negative
real part and > 0 is a small number. In fact, the law in question is classically
interpreted as the addition of r 1 new zeros having negative real part, which
results in a system with transfer function
N(s)N (s)
T (s) =
D(s)
whose n 1 zeros have negative real partto which the previous stability result
can be appliedand in the addition of r 1 poles which are real, negative, and
far away from the origin in the complex plane (so as to preserve the stability of
the closed-loop system), obtaining in this way a controller characterized by a proper
transfer function.
It is natural to guess that, if the coefficients of T (s) are subject to uncertainties but
these uncertainties range over closed and bounded sets, one ought to be able to find
one single (large) value of k and one single (small) value of , both not depending
on the uncertainties as such but only on the sets in which such uncertainties are
confined, for which the resulting closed-loop system is stable. The purpose of Chap. 2
is essentially to provide a formal proof of this result. This will be done considering,
for the controlled system, a state-space model as in (1.1). The state-space model
framework is particularly convenient because in this way a specific formalism and
appropriate concepts can be developed that prove very useful in the more general
4 1 An Overview
in which T1 (s), T2 (s), T3 (s) are accurately known while only some loose properties
of (s) are known. Typically, such properties are expressed as follows: all poles of
(s) have negative real part and a number is known such that
Suppose now that a system characterized by a transfer function of the form (1.2)
is controlled by a system having transfer function Tc (s). This yields the closed-loop
system depicted in Fig. 1.1 on the left. An elementary manipulation shows that such
closed-loop system coincides with the pure feedback interconnection (see Fig. 1.1
on the right) of a subsystem with input v and output z characterized by the transfer
function
T3 (s)T1 (s)Tc (s)
Tzv (s) =
1 T3 (s)T2 (s)T1 (s)Tc (s)
and a subsystem with input z and output v, characterized by the transfer function (s).
Observe that the various Ti (s)s appearing in Tzv (s) are by assumption accurately
known. Thus, it makes sense to consider the problem of designing Tc (s) in such a
way that the subsystem having transfer function Tzv (s) is stable. If this is the case,
the system of Fig. 1.1 is seen as the pure feedback interconnection of two stable
subsystems. If, in addition, for some number < (1/) the function Tzv (s) satisfies
the inequality
|Tzv (j)| for all R (1.3)
holds for some c0 < 1 and, by a well-known stability criterion, it can be concluded
that the closed-loop system of Fig. 1.1 is stable, actually regardless of what (s) is,
so long as its indicated loose properties hold.
It is seen in this way that a systematic, although conservative, method of achieving
stability in spite of unmodeled dynamics consists in looking at a system modeled by
equations of the form
x = Ax + B1 v + B2 u
z = C1 x + D11 v + D12 u
y = C2 x + D21 v
(see Fig. 1.2 and compare with Fig. 1.1), andregarding y as an output available for
measurement and u as an input available for controlseeking a controller yielding
a stable closed-loop system, having a transfer functionbetween input v and output
zsuch that a constraint of the form (1.3) holds. A design problem of this type,
which is known as the -suboptimal H control problem, is addressed in Chap. 3
for a general multi-input multi-output linear system. The solution of such problem
is based on a thorough analysis of the relations existing between the parameters that
characterize the state-space model of the system and the least upper bound of the
norm of its transfer function over the imaginary axis in the complex plane, which is
presented in the first section of the chapter. The design of a controller is presented in
the last section of the chapter, and is based on the feasibility analysis of appropriate
linear matrix inequalities.
In the previous section, we have seen that a method for dealing with unstructured
uncertainties consists in looking at a system such as the one described in Fig. 1.2 and
seeking a controller yielding a stable closed-loop system having a transfer function
between input v and output zthat satisfies a constraint of the form (1.3). The prob-
lem in question, motivated by a problem of robust stability, has also an independent
interest per se. In fact, regarding v as a vector of external disturbances, affecting the
behavior of the controlled plant, and z as a set of selected variables of special interest,
the design problem outlined above can be seen as the problem of finding a controller
thatwhile keeping the closed-loop stableenforces a prescribed attenuation of
the effect of the disturbance v on the variable of interest z.
In Chap. 4, the problem of designing a control so as to influence the effect of certain
external signals on some selected variables is studied in more detail. Specifically,
the problem is addressed of finding a control law that yields a closed-loop system in
which the influence of such external signals on the variables of interest ultimately
decays to zero. This requirement is substantially stronger than the requirement of
imposing a prescribed attenuation and its is not likely to be feasible in general. It can
be enforced, though, if a model of the external signals is available. This gives rise to
what is known as the problem of asymptotic tracking and/or disturbance rejection,
or more commonly the problem of output regulation. The problem in question can
be viewed as a generalized version of a classical problem in the elementary theory
of servomechanisms: the so-called set point control problem. In fact, this problem
consists in the design of a control ensuring that the output y(t) of the controlled plant
asymptotically tracks any prescribed constant reference yref (t), and does this in spite
of any possible constant disturbance d(t). In this case, the external signals are the
prescribed reference yref (t) and the disturbance d(t). Even if their (constant) value
is not available to the controller, a precise model for such signals is known. In fact,
= 0. It is
they are solution of the (trivial) differential equations y ref (t) = 0 and d(t)
well known that such problem is solved if the control guarantees that the tracking
error e(t) = y(t) yref (t) is subject to an integral action. It is also known that a
similar design goal can be achievedby means of similar design techniquesif the
external signals are oscillations of a fixed frequency but unknown amplitude.
The setup considered in Chap. 4 is that of a system modeled by equations of the
form
x = Ax + Bu + Pw
e = Ce x + Qe w (1.4)
y = Cx + Qw,
the first of which describes a plant with state x and control input u, subject to a set
of exogenous input variables w which includes disturbances (to be rejected) and/or
references (to be tracked). The second equation defines a set of regulated (or error)
variables e, the variables of interest that need to be protected from the effect of the
exogenous variables w, while the third equation defines a set of measured variables
1.3 Regulation and Tracking in Linear Systems 7
The problem is to design a controller, with input y and output u, yielding a stable
closed-loop system in which e(t) asymptotically decays to zero as t tends to .
Consistently with the approach pursued in the previous two chapters, it will be
required that the design goal in question be achieved in spite of model uncertainties,
be they structured or unstructured. In this setup, the results established in Chaps. 2
and 3 prove of fundamental importance. A particularly delicate robustness problem
arises when the model of the exogenous signal is uncertain, which is the case when
the parameters that characterize the matrix S in (1.5) are not known. This problem
can be handled by means of techniques borrowed from the theory of adaptive control,
as shown in Sect. 4.8.
Having learned, in Chap. 4, how to design a controller in such a way that the output
y(t) of a plant tracks a reference output yref (t) generated by an autonomous system
of the form
w = Sw
(1.6)
yref = Qw,
x k = Ak xk + Bk uk
(1.7)
yk = Ck xk
The challenge is to achieve the desired tracking goal with a limited exchange of
information, consisting only in the differences between the output of each individual
system and those of a limited number of other systems (that, with the idea of a spatial
distribution in mind, are usually called the neighbors), one of which could be, possibly
but not necessarily, the reference generator (1.6). Such restricted information pattern
renders the problem a little bit more challenging. In Chap. 5, the problem is addressed
under the assumption that the information available for the (decentralized) controller
of the kth system consists of a fixed linear combination the differences (yj yk ), of
the form
N
vk = akj (yj yk ) k = 1, . . . , N (1.8)
j=0
in which the akj s are elements of a N (N +1) sparse matrix and y0 = yref . Typically,
only few of the coefficients akj are nonzero and only those systems for which ak0 = 0
have access to yref (t).
If a problem of this kind is solved, the outputs yk (t) of all systems (1.7)the
followersasymptotically converge to a single function of time, the output yref (t)
of system (1.6)the leader. In this case, it is said that the outputs of all systems,
leader and followers altogether, achieve consensus. A more general version of such
design problem arises when a set of N system modeled by equations of the form (1.7)
is given, no leader is specified, and (decentralized) controllers are to be designed,
which receive information modeled as in (1.8),3 such that all outputs yk (t) achieve
consensus, i.e., asymptotically converge to a single function of time. A problem
of this kind is called a problem of leaderless coordination or, more commonly, a
consensus problem. This problem also will be addressed in Chap. 5. As expected,
instrumental in the solution of such problems is the theory developed in Chap. 4.
Chapter 6 deals with the problem of designing state feedback laws for single-input
single-output nonlinear systems modeled by equations of the form
x = f (x) + g(x)u
(1.9)
y = h(x).
Such systems are usually referred to as input-affine systems. The first part of the
chapter follows very closely the viewpoint presented in Sect. 2.1 of Chap. 2. An
integer r is defined, the relative degree of the system, that plays a role similar to
that of the difference between the number of poles and the number of zeros in the
transfer function of a linear system. On the basis of this concept and of some related
properties, a change of coordinates can be defined, by means of which the system
can be expressed in the form
z = f 0 (z, )
1 = 2
2 = 3
(1.10)
r1 = r
r = q(z, ) + b(z, )u
y = 1 ,
z = f 0 (z, ) (1.11)
in which
K = k0 k1 . . . kr1
is a vector of design parameters chosen in such a way that the roots of the polynomial
p() = r kr1 r1 k1 k0
have negative real part. In fact, system (1.10) subject to the feedback law (1.12) can
be seen as a stable linear system cascaded with an input-to-state stable system and
this, by known properties,5 is an asymptotically stable system (see Fig. 1.3). This
result, discussed in Sect. 6.4, is perhaps one of simplest results available that deal
with global stabilization via state feedback. It is observed, in this respect, that if the
system in question were a linear system, the dynamics of (1.11) would reduce to that
4 See Sect. B.2 of Appendix B for the definition of the concept of input-to-state stability and its
properties.
5 See Sect. B.3.
10 1 An Overview
Fig. 1.3 System (1.10) controlled by (1.12) and its equivalent form
with F0 a matrix whose eigenvalue coincide6 with the roots of the numerator of the
transfer function, i.e., with the zeros of the system. Thus, the assumption that the
dynamics of (1.11) are input-to-state stable can be seen as a nonlinear equivalent of
the assumption that the zeros of the system have negative real part, assumption that
is at the basis of the methods for robust stability presented in Chap. 2. In analogy,
the dynamics of (1.11) are called the zero dynamics of the nonlinear system (1.9).
If system (1.11) does not have the indicated input-to-state stability properties,
global stabilization of system (1.10) is still possible, under the assumption that: (i)
the map f 0 (z, ) only depends on z and on the first component 1 of , i.e.,
z = f 0 (z, 1 )
z = f 0 (z, 1 (z))
6 Under the assumption that the system is reachable and observable, see Sect. 2.1 for more details.
1.5 Feedback Stabilization and State Observers for Nonlinear Systems 11
attraction A that is guaranteed to contain the given set C . Clearly, if this is feasible,
a feedback law will result that depends on the choice of C . Such design goal can
be achieved, still under the assumption that the dynamics of (1.11) are input-to-state
stable, by means of a simple law of the form u = K in which the entries of the
matrix K are appropriate design parameters. A feedback law of this type is a partial-
state feedback law, identical to the one introduced in Sect. 2.4, and enjoys the same
robustness properties. All relevant results are presented in Sect. 6.5.
Having shown how, under suitable assumptions, a nonlinear system can be glob-
ally asymptotically (or only locally asymptotically, but with a guaranteed region
of attraction) stabilized by means of full state or partial-state feedback, in Chap. 7
the design of an asymptotic observer is considered. The existence of such observer
reposes on conditions that guarantee the existence of a change of variables leading
to a special observability normal form. Thus, also in the problem of asymptotic state
estimation, the passage to special forms plays an important role in the design. As
in the case of linear systems, the observer of a system modeled by equations of the
form
x = f (x, u)
(1.13)
y = h(x, u)
The difference with the case of linear systems is that the output injection vector G
needs to have a special form. In fact, G is chosen as
G = col(cn1 , 2 cn2 , . . . , n c0 ).
In Sects. 7.3 and 7.4, it is shown that, under reasonable hypotheses, there is a choice
of the design parameters c0 , c1 , . . . , cn1 and a number such that, if > , the
estimation error x(t) x (t) asymptotically decays to zero, for any value of x(0) x (0).
Since the parameter needs to be large enough, the observer in question is usually
referred to as a high-gain observer.
Finally, in the last section of the chapter, a nonlinear version of the separation
principle of linear system theory7 is discussed. Namely, assuming that the dynamics
of (1.13) can be stabilized by means of a state feedback of the form u = u (x), the
effect of a (dynamical) output feedback of the form
u = g
(u (x ))
in which g
(s) is a saturation function, that is a continuously differentiable function
that coincides with s when |s|
, is odd and monotonically increasing, and satisfies
lims g
(s) =
(1 + c) with c 1. This leads to a control scheme described as
in Fig. 1.4. In this way though, global asymptotic stability no longer can be assured
and only asymptotic stability with a guaranteed region of attraction can be obtained.
All details are presented in Sect. 7.5.
the method discussed in Chap. 2.8 In a relatively recent version of such methods, it
has been shown that all ingredients needed to implement the feedback law (1.12),
that is all components of as well the functions b(z, ) and q(z, ), can be robustly
estimated. The design of a robust observer that accomplishes this goal, known as
extended observer, is describedin the more general context of a multi-input multi-
output systemin Sect. 10.3, while its asymptotic properties as well as the resulting
separation principle are discussed in Sect. 10.4.
Chapter 8 deals with a nonlinear version of the Small-Gain Theorem, used in
Chap. 3 to address robust stabilization problems in the presence of unmodeled dynam-
ics. In a nutshell, the theorem in question says that the pure feedback interconnection
of two input-to-state stable systems is globally asymptotically stable if the compo-
sition of the respective gain functions is a contraction.9 This theorem can be used
as a tool to analyze the stability of interconnected nonlinear systems and, like in the
case of linear systems, to address problems of robust stability.
The methods for feedback stabilization described in Chap. 2, for linear systems, and in
Chap. 6, for nonlinear systems, are extended in Chaps. 911 to the case of multi-input
multi-output systems. To avoid duplications, only nonlinear systems are considered,
andfor convenienceonly the case of systems having the same number of inputs
and output channels is addressed. Extension of the design methods described in
Chap. 2 to the case of a (linear) system
x = Ax + Bu
(1.14)
y = Cx
having the same number m of inputs and output components is relatively straight-
forward in the special case in which there exists a set of integers r1 , r2 , . . . , rm (one
per each component of the output) such that, for each i = 1, . . . , m, the ith row ci of
C satisfies
ci B = ci AB = = ci Ari 2 B = 0, ci Ari 1 B = 0 (1.15)
cm Arm 1 B
with
det[J] = 0. (1.17)
Systems for which property (1.15)(1.17) holds are said to have a vector relative
degree {r1 , . . . , rm }, and it is known that the subclass of multivariable systems that
have such property coincides with the class of systems for which there exist a matrix
F and a nonsingular matrix G that render the transfer function matrix
purely diagonal.
A substantially more general (and hence more interesting) class of systems to deal
with is the class of systems whose transfer function matrix T (s) = C(sI A)1 B is
invertible, which in what follows are referred to as invertible systems.10 This class of
systems is more general because, as it is easy to check, if the properties (1.15)(1.17)
that hold the transfer function of the system is necessarily invertible, but the converse
is not true. This is shown in the following elementary example.
Example 1.1 Consider a linear system with two inputs and two outputs modeled by
the Eq. (1.14) in which
001 1 0
100
A = 0 0 2 , B = 1 0 , C = .
010
101 0 1
Since c1 B = 0 and c2 B = 0, it is seen that the two integers defined in (1.15) are
r1 = r2 = 1. However, the matrix
cB 10
J= 1 =
c2 B 10
10 Note that there is no ambiguity between left and right invertibility, because the matrix T (s) is
square.
1.7 Multi-input Multi-output Nonlinear Systems 15
is singular and hence condition (1.17) does not hold. This system, though, is invert-
ible, as it can be checked by looking at its transfer function matrix
2
1 s s s
T (s) = 3 .
s s2 s s2 s + 1 2s
The extension of the design methods presented in Chaps. 2 and 6 to the more
general class of multivariable systems that are invertible (but possibly do not possess
a vector relative degree) is not as straightforward as one might imagine. In fact, such
extension requires (notably in the case of nonlinear systems) a preliminary detailed
studythat will be conducted in Chap. 9of the consequences of the property of
invertibility. In Sects. 9.2 and 9.3, the concept of invertibility for a multivariable
nonlinear system, affine in the inputs, is discussed. The property of being invertible
is indirectly defined by means of a classical algorithm, known as structure algorithm,
that was originally conceived as a tool to characterizein a multivariable linear
systemcertain integers11 that can be associated with the limit of T (s) as s and
was then extended by various authors to nonlinear systems,12 as a tool to characterize
the property of when the input of a system is uniquely determined by its output (and
by its state). A by-product of paramount importance of the algorithm in question is
the possibility of defining new coordinates functions by means of which the system
is described in a special normal form (see Sect. 9.4). This normal form highlights
as the (simpler) normal form (1.10) of a single-input single-output system does
special relations between individual components of the state vector, and provides
the background for the development of feedback stabilizing laws. The algorithm in
question is also useful to obtain a classification of multivariable nonlinear systems,
on the basis of whichin order of increasing complexitythe stabilization problem
will be handled in Chaps. 10 and 11.
Example 1.2 The following elementary example shows how multivariable nonlin-
ear systems can be classified. Consider a two-input two-output system modeled by
equations of the form
z = f 0 (z, x)
x 1 = a1 (z, x) + b11 (z, x)u1 + b12 (z, x)u2
x 2 = x3 + (z, x)[a1 (z, x) + b11 (z, x)u1 + b12 (z, x)u2 ]
(1.18)
x 3 = a2 (z, x) + b21 (z, x)u1 + b22 (z, x)u2
y1 = x1
y2 = x2
This equation can be considered as a nonlinear equivalent of Eq. (1.16), written for
m = 2 and {r1 , r2 } = {1, 2}, and the system is said to have vector relative degree
{1, 2}.
If (z, x) is nonzero, while the first two components of x coincide with y1 and y2 ,
the third component x3 is a combination (with a state-dependent coefficients) of the
first derivatives of y1 and y2 , namely
Something special, though, occurs if (z, x) is a constant. If this is the case, in fact,
the system can be changed, via feedback, into a system whose inputoutput behavior
is the same as that of a linear system. This is achieved using the control
a1 (z, x) + v1
u = [B(z, x)]1
a2 (z, x) + v2
z = f 0 (z, x)
x 1 = v1
x 2 = x3 + (z, x)v1
x 3 = v2
y1 = x1
y2 = x2 .
The system is said to be inputoutput linearizable. Note also that, if = 0, the matrix
T (s) is invertible, but property (1.15)(1.17) fails to hold.
Finally, consider the (more general) case in which the multiplier (z, x) only
depends on z and x1 . If this is the case, the system is invertible, i.e., the input u(t)
can be uniquely recovered from y(t) and x(t). To see why this is the case, observe
1.7 Multi-input Multi-output Nonlinear Systems 17
that, while both y1(1) and y2(1) explicitly depend on u, the relations
13 In the equation that follows and in the consequences drawn immediately after, we take explicit
advantage of the fact that the multiplier (z, x1 ) does not depend on x2 .
14 The relevance of this special class of systems resides not so much in the fact that there is a feedback
law that can force a linear inputoutput behavior (law that is not likely to be used in practice since
a law that cancels the nonlinearities necessarily requires the exact knowledge of the system model
and hence is not suitable in the presence of model uncertainties) but, rather, in the special properties
of the normal forms of a system for which such law exists. In fact, the properties of such special
normal form can be fruitfully exploited in the design of robust output feedback stabilizing controls
for invertible multivariable systems.
18 1 An Overview
w = s(w)
x = f (w, x, u)
e= he (w, x)
y= h(w, x),
in such a way that all trajectories are bounded and e(t) asymptotically decays to zero
as time tends to . The analysis is cast in a form that follows, as much as possible,
the analysis carried out in Chap. 4. Prerequisite to this is an extension to nonlinear
systems of the notion of steady-state response, extension that is discussed in detail
in Sect. B.6 of Appendix B.
The design methods presented in the chapter only deal with the special case in
which y = e, i.e., the case in which measured output and regulated output coincides.
The more general case of systems in which y includes, in addition to e, an extra set of
measured variables yr , covered for linear systems in Sect. 4.4, is not covered here for
the simple reason that a full and satisfactory solution to such problem is not known
yet for nonlinear systems. In the case of systems having y = e, however, the design
methods covered in Sects. 4.6 and 4.7 can be extended to nonlinear systems, and the
resulting theory is presented in Chap. 12.
Part I
Linear Systems
Chapter 2
Stabilization of Minimum-Phase Linear
Systems
The purpose of this section is to show how single-input single-output linear systems
can be given, by means of a suitable change of coordinates in the state space, a
normal form of special interest, on which certain fundamental properties of the
system are highlighted, that plays a relevant role in the method for robust stabilization
discussed in the following sections.
The point of departure of the whole analysis is the notion of relative degree of
the system, which is formally defined as follows. Given a single-input single-output
system
x = Ax + Bu
(2.1)
y = Cx,
with A Rnn , B Rn1 , C R1n , consider the sequence of real numbers CB,
CAB, CA2 B, . . . , CAk B, . . . Let r denote the least integer for which CAr1 B = 0.
This integer is called the relative degree of system (2.1). In other words, r is the
integer uniquely characterized by the conditions
CB = CAB = . . . = CAr2 B = 0
(2.2)
CAr1 B = 0.
The relative degree of a system can be easily identified with an integer associated
with the transfer function of the system. In fact, consider the transfer function of
(2.1)
T (s) = C(sI A)1 B
1 1 1
(sI A)1 = I + 2 A + 3 A2 +
s s s
Springer International Publishing Switzerland 2017 21
A. Isidori, Lectures in Feedback Design for Multivariable Systems,
Advanced Textbooks in Control and Signal Processing,
DOI 10.1007/978-3-319-42031-8_2
22 2 Stabilization of Minimum-Phase Linear Systems
1 1 1
T (s) = CB + 2 CAB + 3 CA2 B +
s s s
Using the definition of r, it is seen that the expansion in question actually reduces to
1 1 1
T (s) = r
CAr1 B + r+1 CAr B + r+2 CAr+1 B +
s s s
1 1 1
= r [CAr1 B + CAr B + 2 CAr+1 B + ]
s s s
Thus
1 1
sr T (s) = CAr1 B + CAr B + 2 CAr+1 B +
s s
from which it is deduced that
Recall now that T (s) is a rational function of s, the ratio between a numerator
polynomial N(s) and a denominator polynomial D(s)
N(s)
T (s) = .
D(s)
Since the limit (2.3) is finite and (by definition of r) nonzero, it is concluded that r is
necessarily the difference between the degree of D(s) and the degree of N(s). This
motivates the terminology relative degree. Finally, note that, if T (s) is expressed
(as it is always possible) in the form
i=1
nr
(s zi )
T (s) = K ,
i=1 (s pi )
n
To determine the equations describing the system in the new coordinates, we take
the derivatives of z and with respect to time. For the former, no special structure is
found and we simply obtain
On the contrary, for the latter, a special structure can be displayed. In fact, observing
that i = CAi1 x, for i = 1, . . . , r, and using the defining properties (2.2), we obtain
and
24 2 Stabilization of Minimum-Phase Linear Systems
The equations thus found can be cast in a compact form. To this end, it is convenient
to introduce a special triplet of matrices, A Rrr , B Rr1 , C R1r , which
are defined as3
0 1 0 0 0
0 0 1 0 0
A =
, B =
, C = 1 0 0 0 . (2.7)
0 0 0 1 0
0 0 0 0 1
+ B(CA
= A r
x + CAr1 Bu)
(2.8)
y = Cx = 1 = C.
Note that the right-hand sides of (2.6) and (2.8) are still expressed in terms of the
original set of state variables x. To complete the change of coordinates, x should
be expressed as a (linear) function of the new state variables z and , that is as a
function of the form
x = M0 z + M1
Setting
A00 = T0 AM0 , A01 = T0 AM1 , B0 = T0 B,
z = A00 z + A01 + B0 u
+ B(A
= A 10 z + A11 + bu) (2.9)
y = C.
These equations characterize the so-called normal form of the Eq. (2.1) describing
the system. Note that the matrix B0 can be made equal to 0 if the option described
in Proposition 2.2 is used. If this is the case, the corresponding normal form is said
to be strict. On the contrary, the coefficient b, which is equal to the so-called high-
frequency gain of the system, is always nonzero.
In summary, by means of the change of variables indicated above, the original
system is transformed into a system described by matrices having the following
structure
A00 A01 B0
TAT 1 = , TB = , CT 1 = 0 C . (2.10)
BA10 A + BA11 Bb
One of the most relevant features of the normal form of the equations describing
the system is the possibility of establishing a relation between the zeros of the transfer
function of the system and certain submatrices appearing in (2.9).
We begin by observing that
A sI B
det
C 0
T (s) = . (2.11)
det(A sI)
from which, bearing in mind the fact that C(A sI)1 B is a scalar quantity, the
identity (2.11) immediately follows. Note that in this way we have identified simple
and appealing expressions for the numerator and denominator polynomial of T (s).
Next, we determine an expansion of the numerator polynomial.
from which, using the fact that the determinant of a product is the product of the
determinants and that
1 1
T 0 T 0 TT 0
det det = det =1
01 0 1 0 1
we obtain
A sI B TAT 1 sI TB
det = det .
C 0 CT 1 0
Furthermore, observe also that the left-hand side of (2.12) remains unchanged if
A is replaced by A + BF, regardless of how the matrix F R1n is chosen. This
derives from the expansion
A + BF sI B A sI B I 0
=
C 0 C 0 F1
and from the fact that the determinant of the right-hand factor in the product above
is simply equal to 1.
With these observations in mind, use for T the transformation that generates the
normal form (2.10), to arrive at
A00 sI A01 B0
A sI B 10 A + BA 11 sI
det = det BA
Bb
C 0
0 C 0
A00 sI A01 B0 B0
+ F
= det BA10 A + BA
11 sI Bb Bb .
0 C 0
The choice
1
F= A10 A11
b
yields 1
B0 b B0 A10 1b B0 A11
F= 10 BA 11 .
Bb BA
With this expansion in mind, we return to the formula (2.11), from which we
deduce that
(1)r CAr1 B det([A00 b1 B0 A10 ] sI)
T (s) = .
det(A sI)
Changing the signs of both matrices in the numerator and denominator yields the
final expression
det(sI [A00 b1 B0 A10 ])
T (s) = CAr1 B . (2.13)
det(sI A)
If the triplet A, B, C is a minimal realization of its transfer function, i.e., if the pair
(A, B) is reachable and the pair (A, C) is observable, the numerator and denominator
polynomial of this fraction cannot have common factors.5 Thus, we can conclude that
if the pair (A, B) is reachable and the pair (A, C) is observable, the n r eigenvalues
of the matrix
1
[A00 B0 A10 ].
b
can be identified with zeros of the transfer function T (s).
5 Otherwise,T (s) could be written as strictly proper rational function in which the denominator
is a polynomial of degree strictly less than n. This would imply the existence of a realization of
dimension strictly less than n, contradicting the minimality of A, B, C.
28 2 Stabilization of Minimum-Phase Linear Systems
Remark 2.1 Note that if, in the transformation T used to obtain the normal form,
the matrix T0 has been chosen so as to satisfy TB = 0, the structure of the normal
form (2.10) is simplified, and B0 = 0. In this case, the zeros of the transfer function
coincide with the eigenvalues of the matrix A00 .
x = A()x + B()u
(2.14)
y = C()x.
The theory described in what follows is based on the following basic hypothesis.
Assumption 2.1 A(), B(), C() are matrices of continuous functions of . For
every M, the pair (A(), B()) is reachable and the pair (A(), C()) is observ-
able. Moreover:
(i) The relative degree of the system is the same for all M.
(ii) The zeros of the transfer function C()(sI A())1 B() have negative real
part for all M.
In the classical theory of servomechanisms, systems whose transfer function has
zeros only in the (closed) left-half complex plane have been often referred to as
minimum-phase systems. This a terminology that dates back more or less to the
works of H.W. Bode.6 For convenience, we keep this terminology to express the
property that a system satisfies condition (ii) of Assumption 2.1, even though the
latter (which, as it will be seen, is instrumental in the proposed robust stabilization
strategy) excludes the occurrence of zeros on the imaginary axis. Thus, in what
follows, a (linear) system satisfying condition (ii) of Assumption 2.1 will be referred
to as a minimum-phase system.
Letting r denote the relative degree, system (2.14) can be put in strict normal
form, by means of a change of variables
z T0 ()
x = = x
T1 ()
in which
B,
in which A, C are the matrices introduced in (2.7). Moreover,
(i) b() = 0 for all M. By continuity, b() it is either positive or negative for
all M. In what follows, without loss of generality, it will be assumed that
(ii) The eigenvalues of A00 () have negative real part for all M. This being the
case, it is known from the converse Lyapunov Theorem7 that there exists a unique,
symmetric, and positive definite, matrix P(), of dimension (n r) (n r),
of continuous functions of such that
We want to prove that, under the standing assumptions, if k is large enough this
system is stable for all M. To this end, consider the positive definite n n matrix
P() 0
0 1
in which P() is the matrix defined in (2.16). If we are able to show that the matrix
8 The negative sign is a consequence of the standing hypothesis (2.15). If b() < 0, the sign must
be reversed.
2.3 The Case of Relative Degree 1 31
P() 0 A00 () A01 ()
Q() =
0 1 A10 () [A11 () b()k]
T
A00 () A01 () P() 0
+
A10 () [A11 () b()k] 0 1
is negative definite, then by the direct Lyapunov Theorem,9 we can assert that system
(2.19) has all eigenvalues with negative real part. A simple calculation shows that,
because of (2.16),
I [P()A01 () + AT10 ()]
Q() = (2.20)
[P()A01 () + A10 ()]
T T
2[A11 () b()k]
Thus, the conclusion is that Q() is negative definite for all M if (and only if)
Reverting to the notations of (2.20), we can say that Q() is negative definite, or
what is the samesystem (2.19) has all eigenvalues with negative real part, for all
M if
1
k> 2A11 () + [P()A01 () + AT10 ()] 2 for all M.
2b()
Now, set
2A () + [P()A () + AT ()] 2
k := max
11 01 10
.
M 2b()
This maximum exists because the functions are continuous functions of and M is
a compact set. Then we are able to conclude that if
k > k
The role of a large k is to render the term 2b()k sufficiently negative, so as: (i) to
overcome the uncertain term q0 (), and (ii) to overcome the effect of the uncertain
off-diagonal terms. Reverting to the Eq. (2.19) that describe the closed-loop system,
one may observe that the upper equation can be seen as a stable subsystem with state
z and input , while the lower equation can be seen as a subsystem with state and
input z. The role of a large k is: (i) to render the lower subsystem stable, and (ii) to
lower the effect of the coupling between the two subsystems. This second role, which
is usually referred to as a small-gain property, will be described and interpreted in
full generality in the next chapter.
Consider now the case of a system having higher relative degree r > 1. This sys-
tem can be artificially reduced to a system to which the stabilization procedure
described in the previous section is applicable, by means of a simple strategy. Let
the variable r of the normal form be replaced by a new state variable defined as
= r + a0 1 + a1 2 + + ar2 r1 (2.21)
in which a0 , a1 , . . . , ar2 are design parameters. With this change of variable (it is
only a change of variables, no control has been chosen yet !), a system is obtained
2.4 The Case of Higher Relative Degree: Partial State Feedback 33
Hence
A01 () = a01,1 1 + a01,2 2 + + a01,r r .
Since
r = (a0 1 + a1 2 + + ar2 r1 )
we see that
A01 () = [a01,1 a01,r a0 ]1 + [a01,2 a01,r a1 ]2 + + [a01,r1 a01,r ar2 ]r1 + a01,r
= F00 () + F01 ()
(2.22)
= F10 () + F11 () + b()u
with
0 1 0
A0 =
0 0 1 .
a0 a1 ar2
By Assumption 2.1, all eigenvalues of the submatrix A00 () have negative real
part for all . On the other hand, the characteristic polynomial of the submatrix A0 ,
which is a matrix in companion form, coincides with the polynomial
pa () = a0 + a1 + + ar2 r2 + r1 . (2.23)
The design parameters a0 , a1 , . . . , ar2 can be chosen in such a way that all eigen-
values of A0 have negative real part. If this is the case, we can conclude that all the
n 1 eigenvalues of F00 () have negative real part, for all .
Thus, system (2.22) can be seen as a system having relative degree 1 which satisfies
all the assumptions used in the previous section to obtain robust stability. In view of
this, it immediately follows that there exists a number k such that, if k > k , the
control law
u = k
u = k[a0 1 + a1 2 + + ar2 r1 + r ]
u = H. (2.24)
2.4 The Case of Higher Relative Degree: Partial State Feedback 35
It readily follows that, if T (s, ) = C()(sI A())1 B() is the transfer function
of (2.14), the transfer function of system (2.22), seen as a system with input u and
output , is equal to D(s)T (s, ). As confirmed by the state-space analysis and in
particular from the structure of F00 (), this new system has n 1 zeros, n r of
which coincide with the original zeros of (2.14), while the additional r 1 zeros
coincide with the roots of the polynomial (2.23). In other words, the indicated design
method can be interpreted as an addition of r 1 zeros having negative real part
(so as to lower the relative degree to the value 1 while keeping the property that all
zeros have negative real part) followed by high-gain output feedback on the resulting
output.
Example 2.1 Consider the problem of robustly stabilizing a rockets upright orien-
tation in the initial phase of the launch. The equation describing the motion on a
vertical plane is similar to those that describes the motion of an inverted pendulum
(see Fig. 2.1) and has the form11
d2 d
Jt 2
= mg sin() + cos()u.
dt dt
in which is the length of the pendulum, m is the mass concentrated at the tip of the
pendulum, Jt = J + m2 is the total moment of inertia, is a coefficient of rotational
viscous friction and u is a force applied at the base.
If the angle is sufficiently small, one can use the approximations sin()
and cos() 1 and obtain a linear model. Setting 1 = and 2 = , the equation
can be put in state space form as
1 = 2
2 = q1 1 + q2 2 + bu
in which
mg
q1 = , q2 = , b= .
Jt Jt Jt
Note that the system is unstable (because q1 is positive) and that, if is considered
as output, the system has relative degree 2, and hence is trivially minimum-phase.
According to the procedure described above, we pick (compare with (2.21))
= 2 + a0 1
1 = a0 1 +
= (q1 a0 q2 a02 )1 + (q2 + a0 ) + bu.
u = k,
It is known from the theory described above that, if k is sufficiently large, the
system is stable. To determine the minimal value of k as a function of the parameters,
we impose that the matrix
a0 1
(q1 a0 q2 a02 ) (q2 + a0 ) bk
(q2 + a0 ) bk < 0
(a0 )(q2 + a0 bk) (q1 a0 q2 a02 ) > 0,
2.4 The Case of Higher Relative Degree: Partial State Feedback 37
that is
q2 + a0 q1
k > max , .
b a0 b
Once the ranges of the parameters J, m, are specified, this expression can be used
to determine the design parameters a0 and k. Expressed in the original variable the
stabilizing control is
We have seen in the previous section that a system satisfying Assumption 2.1 can be
robustly stabilized by means of a feedback law which is a linear form in the states
1 , . . . , r that characterize its normal form. In general, the components of the state
are not directly available for feedback, nor they can be retrieved from the original
state x, since the transformation that defines in terms of x depends on the uncertain
parameter . We see now how this problem can be overcome, by designing a dynamic
controller that provides appropriate replacements for the components of in the
control law (2.24). Observing that these variables coincide, by definition, with the
measured output y and with its first r 1 derivatives with respect to time, it seems
reasonable to try to generate the latter by means of a dynamical system of the form
1 = 2 + cr1 (y 1 )
2 = 3 + 2 cr2 (y 1 )
(2.26)
r1 = r + c1 (y 1 )
r1
= r c (y ).
r 0 1
38 2 Stabilization of Minimum-Phase Linear Systems
In fact, if 1 (t) where identical to y(t), it would follow that i (t) coincides with
y(i1) (t), that is with i (t), for all i = 1, 2, . . . , r. In compact form, the system thus
defined can be rewritten as
= A + D G0 (y C ),
in which
cr1
cr2 0 0
0 2 0
G0 =
,
D =
,
c1
0 0 r
c0
= A + D G0 (y C )
(2.27)
u = H .
It will be shown in what follows that, if the parameters and c0 , . . . , cr1 which
characterize (2.27) are chosen appropriately, this dynamicoutput feedbackcontrol
law does actually robustly stabilize the system.
Controlling the system (assumed to be expressed in strict normal form) by means
of the control (2.27) yields a closed-loop system
ei = ri (i i ), i = 1, . . . , r.
e = r D1 ( ),
that is
= r D e.
The next step in the analysis is to determine the differential equations for the new
variables ei , for i = 1, . . . , r. Setting
2.5 The Case of Higher Relative Degree: Output Feedback 39
e = col(e1 , . . . , er ).
e = (A G0 C)e
+ B[A
10 ()z + A11 () + b()H ].
Replacing also with its expression in terms of and e we obtain, at the end, a
description of the closed-loop system in the form
The advantage of having the system written in this form is that we know that
the matrix F00 (), if H has been chosen as described in the earlier section, has
eigenvalues with negative real part for all . Hence, there is a positive definite
symmetric matrix P() such that
Moreover, it is readily seen that the characteristic polynomial of the matrix (AG
0 C)
coincides with the polynomial
pc () = c0 + c1 + + cr1 r1 + r . (2.28)
40 2 Stabilization of Minimum-Phase Linear Systems
Thus, the coefficients c0 , c1 , . . . , cr1 can be chosen in such a way that all eigenvalues
of (A G0 C)
have negative real part. If this is done, there exists a positive definite
symmetric matrix P such that
P( + (A G0 C)
A G0 C) T P = I.
This being the case, we proceed now to show that the direct criterion of Lyapunov
is fulfilled, for the positive definite matrix
P() 0
0 P
if the number is large enough. To this end, we need to check that the matrix
P() 0 F00 () F01 () r D
Q=
0 P F10 () (A G0 C)
+ F11 () r D
T
F00 () F01 () r D P() 0
+ + F11 () r D
F10 () (A G0 C) 0 P
in which
d(, ) = [P()F01 () r D + F10T
()P]
11 () r D + r D F11
q(, ) = [PF T
()P].
is positive definite. This is actually the case if is large enough. To check this claim,
assume, without loss of generality, that 1 and observe that in this case the
diagonal matrix
r D = diag( r+1 , . . . , 1 , 1),
is well-defined. To say that the quadratic form (2.29) is positive definite is to say
that, for any nonzero z Rr ,
Clearly, if > max{1, }, this inequality holds and the matrix (2.29) is positive
definite.
It is therefore concluded that if system (2.14) is controlled by (2.27), with H chosen
as indicated in the previous section, and > max{1, }, the resulting closed-loop
system has all eigenvalues with negative real part, for any M.
In summary, we have shown that the uncertain system (2.1), under Assumption 2.1,
can be robustly stabilized by means of a dynamic output-feedback control law of the
form
cr1 1 0 0 cr1
2 cr2 0 1 0 2 cr2
=
y
r1 + r1
c1 0 0 1 c1
r c0 0 0 0 r c0
u = k a0 a1 a2 ar2 1 ,
in which c0 , c1 , . . . , cr1 and, respectively, a0 , a1 , . . . , ar2 are such that the poly-
nomials (2.28) and (2.23) have negative real part, and and k are large (positive)
parameters.
Remark 2.4 Note the striking similarity of the arguments used to show the negative
definiteness of Q with those used in Sect. 2.3. The large value of is instrumental
in overcoming the effects of an additive term in the bottom-right block and of the
off-diagonal terms. Both these terms depend now on (this was not the case in
Sect. 2.3) but fortunately, if 1, such terms have bounds that are independent of
. Also in this case, we can interpret the resulting system as interconnection of a
stable subsystem with state x and input e, connected to a subsystem with state e and
input z. The role of a large is: (i) to render the lower subsystem stable, and (ii) to
lower the effect of the coupling between the two subsystems.
Example 2.2 Consider again the system of Example 2.1 and suppose that only the
variable is available for feedback. In this case, we use a control
u = k(a0 1 + 2 ),
(1 + 2a0 )s + a0
Tc (s) = k .
(s + 1)2
References
In this section,1 we analyze some properties of the forced response of a linear system
to piecewise continuous input functionsdefined on the time interval [0, )which
have the following property
T
lim u(t)2 dt < .
T 0
The space of all such functions, endowed with the so-called L2 norm, which is
defined as 21
u()L2 := u(t)2 dt ,
0
is denoted by UL2 . The main purpose of the analysis is to show that, if the system
is stable, the forced output response from the initial state x(0) = 0 has a similar
property, i.e.,
T
lim y(t)2 dt < . (3.1)
T 0
This makes it possible to compare the L2 norms of input and output functions and
define a concept of gain accordingly.
Consider a linear system described by equations of the form
x = Ax + Bu
(3.2)
y = Cx + Du
1 For an extended and more detailed coverage of the topics handled here in sections 3.1 through 3.5,
with state x Rn , input u Rm , output y Rp . Suppose the matrix A has all eigen-
values with negative real part. Let be a positive number. According to the direct
criterion of Lyapunov, the equation
PA + AT P = I (3.3)
d T V
[x (t)Px(t)] = x (t) = 2x T (t)P(Ax(t) + Bu(t))
dt x x=x(t)
= x T (t)(PA + AT P)x(t) + 2x T (t)PBu(t).
Adding and subtracting d 2 u(t)T u(t) to the right-hand side, and droppingfor
conveniencethe dependence on t, we obtain successively
V
(Ax + Bu) = x T (PA + AT P)x + d 2 uT u d 2 uT u + 2x T PBu
x
T
= x T x + d 2 uT u d 2 u d12 BT Px u d12 BT Px + 1 T
d2
x PBBT Px
x T x + d 2 uT u + 1 T
d2
x PBBT Px.
Subtracting and adding to the right-hand side the quantity yT y and using the
inequality
it is seen that
V
(Ax + Bu) x T x + d 2 uT u + 12 x T PBBT Px yT y + 2x T C T Cx + 2uT DT Du
x
d
T 1
= x I + 2 PBB P + 2C T C x + uT (d 2 I + 2DT D)u yT y.
T
d
Clearly, for any choice of > 0 there exists > 0 such that
I + 2C T C 2I.
Pick one such and let P be determined accordingly as a solution of (3.3). With P
fixed in this way, it is seen that, if the number d is sufficiently large, the inequality
1
PBBT P I
d2
3.1 The L2 Gain of a Stable Linear System 45
d 2 I + 2DT D 2 I.
Using the last two inequalities, it is concluded that the derivative of V (x(t)) along
the trajectories of (3.2) satisfies an inequality of the form
d V
V (x(t)) = (Ax(t) + Bu(t)) x(t)2 + 2 u(t)2 y(t)2 .
dt x x=x(t)
(3.4)
This inequality is called a dissipation inequality.2 Summarizing the discussion
up to this point, one can claim that for any stable linear system, given any positive
real number , it is always possible to find a positive definite symmetric matrix P and
a coefficient such that a dissipation inequality of the form (3.4), whichdropping
for convenience the dependence on tcan be simply written as
V (x)
(Ax + Bu) x2 + 2 u2 y2 , (3.5)
x
is satisfied.
The inequality thus established plays a fundamental role in characterizing a para-
meter, associated with a stable linear system, which is called the L2 gain. As a
matter of fact, suppose the input u() of (3.2) is a function in UL2 . Integration of the
inequality (3.4) on the interval [0, t] yields, for any initial state x(0),
t t
V (x(t)) V (x(0)) + 2 u( )2 d y( )2 d
0 0
2
V (x(0)) + 2
u( )2 d = V (x(0)) + 2 u()L2 ,
0
from which it is deduced that the response x(t) of the system is defined for all t
[0, ) and bounded. Now, suppose x(0) = 0 and observe that the previous inequality
yields
t t
V (x(t)) 2 u( )2 d y( )2 d
0 0
2 The concept of dissipation inequality was introduced by J.C. Willems in [13], to which the reader
is referred for more details.
46 3 The Small-Gain Theorem for Linear Systems
t t 2
y( )2 d 2 u( )2 dt 2 u()L2
0 0
for any t > 0 and therefore the property (3.1) holds. In particular,
2 2
y()L2 2 u()L2 ,
i.e.,
y()L2 u()L2 .
In summary, for any u() UL2 , the response of a stable linear system from the
initial state x(0) = 0 is defined for all t 0 and produces an output y() which has
the property (3.1). Moreover the ratio between the L2 norm of the output and the
L2 norm of the input is bounded by the number which appears in the dissipation
inequality (3.5).
Having seen that, in stable linear system, an input having finite L2 norm produces,
from the initial state x(0) = 0, an output response which also has a finite L2 norm,
suggests to look at the ratios between such norms, for all possible u() UL2 , and
to seek the least upper bound of such ratios. The quantity thus defined is called the
L2 gain of the (stable) linear system. Formally, the gain in question is defined as
follows: pick any u() UL2 and let y0,u () be the resulting response from the initial
state x(0) = 0; the L2 gain of the system is the quantity
y0,u ()L2
L2 gain = sup .
u()L2 =0 u()L2
With this definition in mind, return to the dissipation inequality (3.5). Suppose
that a system of the form (3.2) is given and that an inequality of the form (3.5) holds
for a positive definite matrix P. We have in particular (set u = 0)
V (x)
Ax x2 ,
x
from which it is seen that the system is stable. Since the system is stable, the L2 gain
can be defined and, as a consequence of the previous discussion, it is seen that
L2 gain . (3.6)
Thus, in summary, if an inequality of the form (3.5) holds for a positive definite matrix
P, the system is stable and its L2 gain is bounded from above by the number .
We will see in the next section that the fulfillment of an inequality of the form
(3.5) is equivalent to the fulfillment of a linear matrix inequality involving the system
data A, B, C, D.
3.2 An LMI Characterization of the L2 Gain 47
V
(Ax + Bu) x2 + 2 u2 Cx + Du2 (3.7)
x
DT D 2 I < 0 (3.8)
Conversely, suppose (3.8) and (3.9) hold for some . Then there exists > 0 such
that, for all satisfying 0 < < < , the inequality (3.7) holds for all x Rn
and all u Rm .
Proof Suppose that the inequality (3.7), that can be written as
2 I + DT D 0.
Since > , it follows that DT D < 2 I, which is precisely condition (3.8). More-
over, since > , (3.7) implies
in which
W = 2 I DT D
and
1 1 T
umax (x) = W N (x).
2
Hence, (3.11) holds if and only if the value of the form (3.12) at u = umax (x) is
nonpositive, that is if and only if
1
M(x) + N(x)W 1 N(x)T 0.
4
Using the expressions M(x), N(x), W , the latter reads as
provided that 1 > 0 is small enough. The left-hand sides of (3.8) and (3.13), which
are negative definite, by continuity remain negative definite if is replaced by any
satisfying 2 < < , provided that 2 > 0 is small enough. Take now =
min{1 , 2 }. The resulting form M(x) + N(x)u uT W u, in which W = 2 I DT D,
is non positive and (3.7) holds.
Remark 3.1 Note that the inequalities (3.8) and (3.9) take a substantially simpler
form in the case of a system in which D = 0 (i.e., systems with no direct feed-
through between input and output). In this case, in fact, (3.8) becomes irrelevant and
(3.9) reduces to
1
PA + AT P + C T C + 2 PBBT P < 0.
Lemma 3.2 Let be a fixed positive number. The inequality (3.8) holds and there
exists a positive definite symmetric matrix P satisfying (3.9) if and only if there exists
a positive definite symmetric matrix X satisfying
T
A X + XA XB C T
BT X I DT < 0. (3.14)
C D I
3.2 An LMI Characterization of the L2 Gain 49
1 T
I + D D
is negative definite, which is equivalent to condition (3.8), and so is its the Schurs
complement
1 T
1 1 1 1
AT X + XA + C T C XB + C T D I + DT D XB + C T D .
Functions having finite L2 norm may be seen as signals having finite energy3 over the
infinite time interval [0, ), and therefore the L2 gain can be given the interpretation
of (an upper bound of the) ratio between energies of output and input. Another similar
interpretation, in terms of energies associated with input and output, is possible,
which does not necessarily require the consideration of the case of finite energy over
the infinite time interval [0, ). Suppose the input is a periodic function of time,
with period T , i.e., that
for some piecewise continuous function u (t), defined on [0, T ). Also, suppose that,
for some suitable initial state x(0) = x , the state response x(t) of the system is defined
for all t [0, T ] and satisfies
x(T ) = x .
3 If,
in the actual physical system, the components of the input u(t) are voltages (or currents), the
quantity u(t)2 can be seen as instantaneous power, at time t, associated with such input and its
integral over a time interval [t0 , t1 ] as energy associated with such input over this time interval.
50 3 The Small-Gain Theorem for Linear Systems
Then, it is obvious that x(t) exists for all t 0, and is a periodic function, having
the same period T of the input, namely
Observe that the integrals on both sides of this inequality are independent of t0 ,
because the integrands are periodic functions having period T , and recall that the root
mean square value of any (possibly vector-valued) periodic function f (t) (which is
usually abbreviated as r.m.s. and characterizes the average power of the signal
represented by f (t)) is defined as
1 t0 +T 21
f ()r.m.s. = f ( )2 d .
T t0
In other words, the number that appears in the inequality (3.5) can be seen also
as an upper bound for the ratio between the r.m.s. value of the output and the r.m.s
value of the input, whenever a periodic input is producing (from an appropriate initial
state) a periodic (state and output) response.
Consider now the special case in which the input signal is a harmonic function of
time, i.e.,
u(t) = u0 cos(0 t)
It is known that, if the state x(0) is appropriately chosen,4 the output response of
the system coincides with the so-called steady-state response, which is the harmonic
function
yss (t) = Re[T (j0 )]u0 cos(0 t) Im[T (j0 )]u0 sin(0 t),
4 See Sect. A.5 in Appendix A. This is the case if x(0) = 1 , in which 1 is the first column of the
solution of the Sylvester equation (A.29).
3.3 The H Norm of a Transfer Function 51
in which
T (j) = C(jI A)1 B + D.
Recall that 2
0
u(t)2 dt = u0 2
0 0
and therefore 2
0
yss (t)2 dt = T (j0 )u0 2 .
0 0
In other words,
u()2r.m.s. = 1
2
u0 2
yss ()2r.m.s. = 1
2
T (j0 )u0 2 .
Thus, from the interpretation illustrated above one can conclude that, if the system
satisfies (3.5), then
i.e.,
T (j0 )u0 u0 ,
T (j0 ) .
T H := sup T (j),
R
which is called the H norm of the matrix T (j). Observing that 0 in the above
inequality is arbitrary, it is concluded
T H . (3.18)
Therefore, a linear system that satisfies (3.5) is stable and the H norm of its frequency
response matrix is bounded from above by the number .
We have seen in the previous sections that, for a system of the form (3.2), if there exists
a number > 0 and a symmetric positive definite matrix X satisfying (3.14), then
there exists a number < and a positive definite quadratic form V (x) satisfying
(3.5) for some > 0. This, in view of the interpretation provided above, proves that
the fulfillment of (3.14) for some (with X positive definite) implies that
(i) the system is asymptotically stable,
(ii) its L2 gain is strictly less than ,
(iii) the H norm of its transfer function is strictly less than .
However, put in these terms, we have only learned that (3.5) implies both (ii) and
(iii) and we have not investigated yet whether converse implications might hold. In
this section, we complete the analysis, by showing that the two properties (ii) and
(iii) are, in fact, two different manifestations of the same property and both imply
(3.14).
This will be done by means of a circular proof involving another equivalent version
of the property that the number is an upper bound for the H norm of the transfer
function matrix of the system, which is very useful for practical purposes, since it
can be easily checked. More precisely, the fact that is an upper bound for the H
norm of the transfer function matrix of the system can be checked by looking at the
spectrum of a matrix of the form
A0 R0
H= , (3.19)
Q0 AT0
in which R0 and Q0 are symmetric matrices which, together with A0 , depend on the
matrices A, B, C, D which characterize the system and on the number .6
As a matter of fact the following result, known in the literature as Bounded Real
Lemma, holds.
Theorem 3.1 Consider the linear system (3.2) and let > 0 be a fixed number. The
following are equivalent:
(i) there exists < , > 0 and a symmetric positive definite matrix P such that
(3.7) holds for V (x) = x T Px,
(ii) all the eigenvalues of A have negative real part and the frequency response matrix
of the system T (j) = C(jI A)1 B + D satisfies
6 A matrix (of real numbers) with this structure is called an Hamiltonian matrix and has the property
that its spectrum is symmetric with respect to the imaginary axis (see Lemma A.5 in Appendix A).
3.4 The Bounded Real Lemma 53
(iii) all the eigenvalues of A have negative real part, the matrix W = 2 I DT D is
positive definite, and the Hamiltonian matrix
A + BW 1 DT C BW 1 BT
H= (3.21)
C C C DW D C A C T DW 1 BT
T T 1 T T
Proof We have already shown, in the previous sections, that if (i) holds, then (3.2)
is an asymptotically stable system, with a frequency response matrix satisfying
T H .
lim T (j) = D
it necessarily follows that Du < for all u with u = 1 and this implies 2 I >
DT D, i.e., the matrix W is positive definite.
Now observe that the Hamiltonian matrix (3.21) can be expressed in the form
H = L + MN
for
A 0 B
L= , M = ,
C T C AT C T D
N = W 1 DT C W 1 BT .
Observe now that the matrix L has no eigenvalues on the imaginary axis, because
its eigenvalues coincide with those of A and AT , and A is by hypothesis stable.
Thus (j0 I L) is nonsingular. Observe also that the vector u0 = Nx0 is nonzero
because otherwise x0 would be an eigenvector of L associated with an eigenvalue at
j0 , which is a contradiction. A simple manipulation yields
54 3 The Small-Gain Theorem for Linear Systems
where T (s) = C(sI A)1 B + D. In fact, it suffices to compute the transfer function
of
x = Lx + Mu
y = Nx
This implies
T (j0 )u0
T (j0 ) = sup
u0 =0 u0
F = (A + BW 1 DT C)T
Q = BW 1 BT
GGT = C T (I + DW 1 DT )C
and this matrix by hypothesis has no eigenvalues on the imaginary axis. Moreover, it
is also possible to show that the pair (F, GGT ) thus defined is stabilizable. In fact,
suppose that this is not the case. Then, there is a vector x = 0 such that
x T F I GGT = 0
3.4 The Bounded Real Lemma 55
Y F + F T Y Y GGT Y + Q = 0 , (3.25)
Y F + F T Y Y GGT Y + Q
= Y (AT + C T DW 1 BT ) + (A + BW 1 DT C)Y Y C T (I + DW 1 DT )CY BW 1 BT
= Y AT + AY [Y C T D B]W 1 [DT CY BT ] Y C T CY ,
z = AT z , (3.27)
U(z(t))
= 2zT (t)Y AT z(t) = zT (t)[Y AT + AY ]z(t) 0.
t
This inequality shows that the function V (z(t)) is non-decreasing, i.e., V (z(t))
V (z(0)) for any z(0) and any t 0. On the other hand, system (3.27) is by hypothesis
asymptotically stable, i.e., limt z(t) = 0. Therefore, necessarily, V (z(0)) 0, i.e.,
the matrix Y is negative semi-definite. From this, it is concluded that any solution
Y of (3.26), that is of the inequality
x = ax + bu
(3.29)
y = x + du
in which it is assumed that a > 0, so that the system is stable. The transfer function
of this system is
ds + (ad + b)
T (s) = .
s+a
This function has a pole at p = a and a zero at z = (ad + b)/d. Thus, bearing
in mind the possible Bode plots of a function having one pole and one zero (see
Fig. 3.1), it is seen that
ad + b ad + b
|d| < | | T H = |T (0)| = | |
a a
ad + b
|d| > | | T H = lim |T (j)| = |d|.
a
Thus,
ad + b
T H = max |d|, | | . (3.30)
a
Fig. 3.1 The Bode plots of the transfer function of (3.29). Left the case |T (0)| > |d|. Right the
case |T (0)| < |d|
3.4 The Bounded Real Lemma 57
The criterion in the Bounded Real Lemma says that the H norm of T (s) is strictly
less than if and only if 2 I DT D > 0, which in this case becomes
2 > d2 (3.31)
has no eigenvalues with zero real part. A simple calculation shows that the charac-
teristic polynomial of H is
2abd b2
p() = 2 a2 + + 2 .
d
2 2 d2
This polynomial has no roots with zero real part if and only if
2abd b2
a2 2 >0
d
2 2 d2
Using both (3.31) and (3.32), it is concluded that, according to the Bounded Real
Lemma, T H < if and only if
ad + b
> max |d|, | |
a
x i = Ai xi + Bi ui
(3.33)
yi = Ci xi + Di ui
dim(u2 ) = dim(y1 )
dim(u1 ) = dim(y2 ).
Suppose that the matrices D1 and D2 are such that the interconnections
u2 = y1
(3.34)
u1 = y2
u1 = C2 x2 + D2 u2
u2 = C1 x1 + D1 u1 ,
u1 D2 u2 = C2 x2
D1 u1 + u2 = C1 x1
Suppose now that both systems 1 and 2 are stable. As shown in Sect. 3.1, there
exist two positive definite matrices P1 , P2 , two positive numbers 1 , 2 and two real
numbers 1 , 2 such that 1 and 2 satisfy inequalities of the form (3.5), namely
Vi
(Ai xi + Bi ui ) i xi 2 + i2 ui 2 yi 2 (3.35)
xi
which, if a > 0, is positive definite. A simple calculation shows that, for any trajec-
tory (x1 (t), x2 (t)) of the pure feedback interconnection of 1 and 2 , the function
W (x1 (t), x2 (t)) satisfies
W W
x 1 + x 2 1 x1 2 a2 x1 2 + 12 u1 2 y1 2 + a22 u2 2 ay2 2
x1 x2
If
(1 + a22 )I 0
0, (3.36)
0 (12 a)I
we have
W W 1 I 0 x1
x 1 + x 2 1 x1 a2 x2 = x1T x2T .
x1 x2 0 a2 I x2
The quadratic form on the right-hand side is negative definite and hence, according
to the criterion of Lyapunov, the interconnected system is asymptotically stable.
Condition (3.36), on the other hand, can be fulfilled for some a > 0 if (and only if)
1 + a22 0, 12 a 0
i.e., if
1
12 a .
22
Theorem 3.2 Consider a pair of systems (3.33) and suppose that the matrix I
D2 D1 is nonsingular. Suppose (3.33) satisfy inequalities of the form (3.35), with
P1 , P2 positive definite and
1 2 1.
Remark 3.2 In the proof of the above statement, we have considered the general case
in which both component subsystems have an internal dynamics. To cover the special
case in which one of the two subsystem is a memoryless system, a slightly different
(actually simpler) argument is needed. Suppose the second component subsystem is
modeled as
y2 = D2 u2
V1
(A1 x1 + B1 u1 ) 1 x1 2 + 12 u1 2 y1 2
x1
1 x1 2 + (12 22 1)y1 2 .
Note also that, in view of the Bounded Real Lemma, the theorem above can be
rephrased in terms of H norms of the transfer functions of the two component
subsystems, as follows:
Corollary 3.1 Consider a pair of systems (3.33) and suppose that the matrix I
D2 D1 is nonsingular. Suppose both systems are asymptotically stable. Let
Ti (s) = Ci (sI Ai )1 Bi + Di
1
T1 H < .
T2 H
3.5 Small-Gain Theorem and Robust Stability 61
1
T1 H < 1 < .
T2 H
From the Bounded Real Lemma, it is seen thatfor both i = 1, 2there exists
i > 0, i < i and positive definite Pi such that (3.33) satisfy inequalities of the form
(3.35), with Vi (x) = xiT Pi xi . Moreover, 1 2 < 1 2 = 1. Hence, from Theorem 3.2
and Remark 3.2, the result follows.
The result just proven is known as the Small-Gain Theorem (of linear systems).
In a nutshell, it says that if both component systems are stable, a sufficient condition
for the stability of their (pure feedback) interconnection is that the product of the
H norms of the transfer functions of the two component subsystems is strictly less
than 1. This theorem is the point of departure for the study of robust stability via H
methods.
It should be stressed that the small gain condition (3.37) provided by this corol-
lary is only sufficient for stability of the interconnection and that the strict inequality
in (3.37) cannot be replaced, in general, by a loose inequality. Both these facts are
explained in the example which follows.
x 1 = ax1 + u1
y1 = x1
and
y2 = du2 .
Suppose a > 0 so that the first subsystem is stable. The stability of the interconnec-
tion can be trivially analyzed by direct computation. In fact, the interconnection is
modeled as
x 1 = ax1 + dx1
from which is it is seen that a necessary and sufficient condition for the interconnec-
tion to be stable is that d < a.
On the other hand, the Small-Gain Theorem yields a more conservative estimate.
In fact
1
T1 (s) = and T2 (s) = d
s+a
62 3 The Small-Gain Theorem for Linear Systems
and hence
1
T1 H = T1 (0) = and T2 H = |d|
a
The (sufficient) condition (3.37) becomes |d| < a, i.e., a < d < a.
This example also shows that the inequality in (3.37) has to be strict. In fact,
the condition T1 H T2 H = 1 would yield |d| = a, which is not admissible
because, if d = a the interconnection is not (asymptotically) stable (while it would
be asymptotically stable if d = a).
Example 3.3 It is worth observing that a special version of such sufficient condition
was implicit in the arguments used in Chap. 2 to prove asymptotic stability. Take, for
instance, the case studied in Sect. 2.3. System (2.19) can be seen as pure feedback
interconnection of a subsystem modeled by
= [A11 () b()k] + u2
(3.39)
y2 = .
The first of two such subsystems is a stable system, because A00 () has all eigen-
values in C . Let
T1 (s) = A10 ()[sI A00 ()]1 A01 ().
denote its transfer function. Its H norm depends on but, since ranges over a
compact set M, it is possible to find a number 1 > 0 such that
Note that this number 1 depends only on the data that characterize the controlled
system (2.17) and not on the coefficient k that characterizes the feedback law (2.18).
Consider now the second subsystem. If b()k A11 () > 0, this is a stable sys-
tem with transfer function
1
T2 (s) =
s + (b()k A11 ())
Clearly,
1
T2 H = T2 (0) = .
b()k A11 ()
3.5 Small-Gain Theorem and Robust Stability 63
It is seen from this that T2 H can be arbitrarily decreased by increasing the coef-
ficient k. In other words: a large value of the output feedback gain coefficient k in
(2.18) forces a small value of T2 H .
As a consequence of the Small-Gain Theorem, the interconnected system (namely,
system (2.19)) is stable if
1
<1
b()k A11 ()
In summary, the result established in Sect. 2.3 can be re-interpreted in the follow-
ing terms. Subsystem (3.38) is a stable system, with a transfer function whose H
norm has some fixed bound 1 (on which the control has no influence, though). By
increasing the value of the gain coefficient k in (2.18), the subsystem (3.39) can be
rendered stable, with a H norm that can be made arbitrarily small, in particular
smaller than the (fixed) number 1/1 . This makes the small-gain condition (3.37)
fulfilled and guarantees the (robust) stability of system (2.19).
We turn now to the discussion of the problem of robust stabilization. The problem
in question can be cast, in rather general terms, as follows. A plant with control
input u and measurement output y whose model is uncertain can be, from a rather
general viewpoint, thought of as the interconnection of a nominal system modeled
by equations of the form
x = Ax + B1 v + B2 u
z = C1 x + D11 v + D12 u (3.40)
y = C2 x + D21 v
in which the additional input v and the additional output z are seen as output
and, respectively, input of a system
x p = Ap xp + Bp z
(3.41)
v = Cp xp + Dp z
10 Note that, for the interconnection (3.40)(3.41) to be well-defined, the matrix I Dp D11 is
required to be invertible.
64 3 The Small-Gain Theorem for Linear Systems
which is indeed a special case of a system of the form (3.41). The interconnection
(3.40)(3.41) of a nominal model and of a dynamic perturbation is more general,
though, because it accommodates for perturbations which possibly include unmod-
eled dynamics (see example at the end of the section). In this setup, all modeling
uncertainties are confined in the model (3.41), including the dimension itself of xp .
Suppose now that, in (3.41), Ap is a Hurwitz matrix and that the transfer function
Tp (s) = Cp (sI Ap )1 Bp + Dp
has an H norm which is bounded by a known number p . That is, assume that, no
matter what the perturbations are, the perturbing system (3.41) is a stable system
satisfying
Tp H < p . (3.43)
for some p .
Let
x c = Ac xc + Bc y
(3.44)
u = Cc xc + Dc y
3.5 Small-Gain Theorem and Robust Stability 65
which is asymptotically stable and whose transfer function, between the input v and
output z has an H norm bounded by a number satisfying
p < 1. (3.45)
If this is the case, thanks to the Small-Gain Theorem, it can be concluded that the
controller (3.44) stabilizes any of the perturbed plants (3.40) and (3.41), so long as
the perturbation is such that (3.41) is asymptotically stable and the bound (3.43)
holds (Fig. 3.4).
In this way, the problem of robust stabilization is reduced to a problem of stabi-
lizing a nominal plant and to simultaneously enforce a bound on the H norm of its
transfer function.
Example 3.4 A simplified model describing the motion of a vertical takeoff and
landing (VTOL) aircraft in the vertical-lateral plane can be obtained in the following
way. Let y, h and denote, respectively, the horizontal and vertical position of the
center of mass C and the roll angle of the aircraft with respect to the horizon, as
in Fig. 3.5. The control inputs are the thrust directed out the bottom of the aircraft,
denoted by T , and the rolling moment produced by a couple of equal forces, denoted
by F, acting at the wingtips. Their direction is not perpendicular to the horizontal
body axis, but tilted by some fixed angle . Letting M denote the mass of the aircraft,
J the moment of inertia about the center of mass, the distance between the wingtips
and g the gravitational acceleration, it is seen that the motion of the aircraft is modeled
by the equations
M y = sin( )T + 2 cos( )F sin
M h = cos( )T + 2 sin( )F sin gM
J = 2F cos .
The purpose of the control input T is that of moving the center of mass of the airplane
in the vertical-lateral plane, while that of the input F is to control the airplanes
attitude, i.e., the roll angle .
In the hovering maneuver, the thrust T is expected to compensate exactly the
force of gravity. Thus, T can be expressed as T = gM + T in which T is a residual
input use to control the attitude. If the angle is sufficiently small one can use the
approximations sin( ) , cos( ) 1 and neglect the nonlinear terms, so as to
obtain a linear simplified model
M y = gM + 2F sin
M h = T
J = 2F cos .
In this simplified model, the motion in the vertical direction is totally decoupled, and
controlled by T . On the other hand, the motion in the lateral direction and the roll
angle are coupled, and controlled by F. We concentrate on the analysis of the latter.
The system with input F and output y is a four-dimensional system, having relative
degree 2. Setting 1 = , 2 = , 1 = y, 2 = y and u = F, it can be expressed, in
state-space form, as
3.5 Small-Gain Theorem and Robust Stability 67
1 = 2
2 = 2
J
(cos )u
1 = 2
2 = g1 + M2 (sin )u.
Note that this is not a strict normal form. It can be put in strict normal form, though,
changing 2 into
M cos
2 = 2 2 ,
J sin
as the reader can easily verify. The strict normal form in question is given by
M cos
1 0 1 1
= + J sin 2
2 2
Mg cos
J sin
0 0
(3.46)
1 01 1 0 0 2
= + + (sin )u.
2 00 2 g 1 1 M
It is seen from this that the zeros of the system are the roots of
Mg cos
N() = 2 ,
J sin
and hence the system is not minimum phase, because one zero has positive real part.
Therefore the (elementary) methods for robust stabilization described in the previous
section cannot be applied.
However, the approach presented in this section is applicable. To this end, observe
that the characteristic polynomial of the system is the polynomial D() = 4 and
consequently its transfer function has the form
2g
2
(sin )s2 (cos )
T (s) = M J
.
s4
From thisby known factsit is seen that an equivalent realization of (3.46) is given
by
0100 0
0 0 1 0 0
x =
0 0 0 1 x + 0 u := A0 + B0 u
(3.47)
0000 1
y = 2g
J
(cos ) 0 2
M
(sin ) 0 x := Cx.
The matrices A0 and B0 are not subject to perturbations. Thus, the only uncertain
parameters are the entries of the matrix C. We may write Cx in the form
68 3 The Small-Gain Theorem for Linear Systems
Cx = C0 x + Dp z
x = A0 x + B0 u
z = col(x1 , x3 )
y = C0 x + v,
Example 3.5 Consider a d.c. motor, in which the stator field is kept constant, and
the control is provided by the rotor voltage. The system in question can be charac-
terized by equations expressing the mechanical balance and the electrical balance.
The mechanical balance, in the hypothesis of viscous friction only (namely friction
torque purely proportional to the angular velocity) has the form
J + F = T
in which denotes the angular velocity of the motor shaft, J the inertia of the rotor,
F the viscous friction coefficient, and T the torque developed at the shaft. The torque
T is proportional to the rotor current I, namely,
T = km I.
The rotor current, in turn, is determined by the electrical balance of the rotor circuit.
This circuit, with reasonable approximation, can be modeled as in Fig. 3.6, in which
R is the resistance of the rotor winding, L is the inductance of the rotor winding, Rb
is the contact resistance of the brushes, Cs is a stray capacitance. The voltage V is the
control input and the voltage E is the so-called back electromotive force (e.m.f.)
which is proportional to the angular velocity of the motor shaft, namely
E = ke .
Rb (I + Is ) = V Vb
dVb
Cs = Is
dt
dI
L + RI = Vb E .
dt
These can be put in state-space form by setting
= col(Vb , I) , z = col(V , )
= F + Gz
I = H
in which
1 1
1
Cs Rb Cs 0
F = , G = Cs Rb k , H = 0 1 .
1 R e
0
L
L L
In this way, the rotor current I is seen as output of a system with input z. Note
that this system is stable, because characteristic polynomial of the matrix F
1 R 1 R
d() = + 2
+ + +
Cs Rb L LCs LCs Rb
70 3 The Small-Gain Theorem for Linear Systems
has roots with negative real part. A simple calculation shows that the two entries of
the transfer function of this system
T (s) = T1 (s) T2 (s)
1 (Cs Rb s + 1)ke
T1 (s) = , T2 (s) = .
(Cs Rb s + 1)(L s + R) + Rb (Cs Rb s + 1)(L s + R) + Rb
1 ke
T1 (s)
= , T2 (s)
= .
(L s + R) (L s + R)
and also that, if (L/R) 1, the functions can be further approximated, over a rea-
sonable range of frequencies, as
1 ke
T1 (s)
= , T2 (s) = .
R R
This shows that, neglecting the dynamics of the rotor circuit, the rotor current can
be approximately expressed as
1
I0
= (V ke ) = Kz,
R
in which K is the row vector
1 ke
K= .
R R
With this in mind, the (full) expression of the rotor current can be written as
I = I0 + v
= F + Gz
v = H Kz.
Replacing this expression into the equation of the mechanical balance, letting x1
denote the angular position of the rotor shaft (in which case it is seen that x 1 = ),
setting
x2 = , u=V
3.5 Small-Gain Theorem and Robust Stability 71
x 1 = x2
F km ke km km
x 2 = + x2 + v + u.
J JR J JR
In summary, letting y = x1 denote the measured output of the system, the (full) model
of the system in question can be seen as a system of the form (3.40), with
0 1 0 0
A= , B1 = , B2 =
F km ke km km
0 +
J JR J JR
00 0 1
C1 = , D11 = , D12 =
01 0 0
C2 = 1 0 , D21 = 0 ,
Motivated by the results discussed at the end of the previous section, we consider
now a design problem which goes under the name of problem of -suboptimal H
feedback design.12 Consider a linear system described by equations of the form
12 Inthis section, the exposition closely follows the approach of [7, 9, 10] to the problem of -
suboptimal H feedback design. For the numerical implementation of the design methods, see
also [8].
72 3 The Small-Gain Theorem for Linear Systems
x = Ax + B1 v + B2 u
z = C1 x + D11 v + D12 u (3.48)
y = C2 x + D21 v.
x c = Ac xc + Bc y
(3.49)
u = Cc xc + Dc y
which is asymptotically stable and whose transfer function, between the input v and
output z, has an H norm strictly less than .
The interest in such problem, in view of the results discussed earlier, is obvious. In
fact, if this problem is solved, the controller (3.49) robustly stabilizes any perturbed
system that can be seen as pure feedback interconnection of (3.48) and of an uncertain
system of the form (3.41), so long as the latter is a stable system having a transfer
function whose H norm is less than or equal to the inverse of . Of course, the
smaller is the value of for which the problem is solvable, the larger is the set of
perturbations against which robust stability can be achieved.
Rewrite system (3.50) as
x = A x + Bv
z = C x + Dv
where
A + B2 Dc C2 B2 Cc B1 + B2 Dc D21
A = , B=
Bc C2 Ac Bc D21
C = C1 + D12 Dc C2 D12 Cc , D = D11 + D12 Dc D21 .
In view of the Bounded Real Lemma, the closed-loop system has the desired
properties if and only if there exists a symmetric matrix X > 0 satisfying
X >0 (3.51)
A TX + X A X B C T
BT X I D T < 0. (3.52)
C D I
3.6 The Coupled LMIs Approach to the Problem 73
Thus, the problem is to try to find a quadruplet {Ac , Bc , Cc , Dc } such that (3.51) and
(3.52) hold for some symmetric X .
The basic inequality (3.52) will be now transformed as follows. Let
Set
A0 B1
A0 = , B0 = , C0 = C1 0 , (3.53)
00 0
AT0 X + X A0 X B0 CT0
T
(X ) = BT0 X I D11 , (3.54)
C0 D11 I
0 Ik 0 0
P= ,
B2T 0 0m2 m1 D12
T
0 Ik 0 0
Q= ,
C2 0 D21 0p2 p1
and
X 0 0
(X ) = 0 Im1 0 .
0 0 Ip1
Thus, the problem of -suboptimal H feedback design can be cast as the problem
of finding a symmetric matrix X > 0 and a matrix K such that (3.56) holds. Note
that this inequality, is not a linear matrix inequality in the unknowns X and K.
Rather, it is a bilinear matrix inequality.13 However, the problem of finding a matrix
X > 0 for which (3.56) is solved by some K can be cast in terms of linear matrix
inequalities. In this context, the following result is very useful.14
13 The inequality (3.56), for each fixed X is a linear matrix inequality in K and, for each fixed K
is a linear matrix inequality in X .
14 The proof of Lemma 3.3 can be found in [7].
74 3 The Small-Gain Theorem for Linear Systems
Im(WP ) = Ker(P)
Im(WQ ) = Ker(Q).
WQT WQ < 0
(3.58)
WPT WP < 0.
This lemma can be used to eliminate K from (3.56) and obtain existence conditions
depending only on X and on the parameters of the plant (3.48). Let WP (X ) be a
matrix whose columns span Ker(P (X )) and let WQ be a matrix whose columns
span Ker(Q). According to Lemma 3.3, there exists K for which (3.56) holds if and
only if
WQ T
(X )WQ < 0
(3.59)
WP (X ) (X )WP (X ) < 0.
T
The second of these two inequalities can be further manipulated observing that if WP
is a matrix whose columns span Ker(P), the columns of the matrix [ (X )]1 WP
span Ker(P (X )). Thus, having set
A0 X 1 + X 1 AT0 B0 X 1 CT0
(X ) = BT0 I D11
T , (3.60)
1
C0 X D11 I
It could therefore be concluded that (3.56) is solvable for some K if and only if the
matrix X satisfies the first of (3.59) and (3.61).
In view of this, the following (intermediate) conclusion can be drawn.
Proposition 3.1 There exists a k-dimensional controller that solves the problem of
-suboptimal H feedback design if and only if there exists a (n + k) (n + k)
symmetric matrix X > 0 satisfying the following set of linear matrix inequalities
15 Observe that, since Ker(P) and Ker(Q) are subspaces of Rm , then W and WQ are matrices having
P
m rows.
3.6 The Coupled LMIs Approach to the Problem 75
T
WQ (X )WQ < 0
(3.62)
T
WP (X )WP < 0,
The two inequalities (3.62) thus found can be further simplified. To this end, it is
convenient to partition X and X 1 as
S N 1 R M
X = , X = (3.63)
NT MT
in which R and S are n n and N and M are n k. With the partition (3.63), the
matrix (3.54) becomes
AT S + SA AT N SB1 C1T
N TA 0 N T B1 0
(X ) =
B1T S T
(3.64)
B1T N I D11
C1 0 D11 I
From the definition of Q, it is readily seen that a matrix WQ whose columns span
Ker(Q) can be expressed as
Z1 0
0 0
WQ = Z2 0
0 Ip1
in which
Z1
Z2
Then, an easy calculation shows that the first inequality in (3.62) can be rewritten as
T AT S + SA SB1 C1T Z1 0
Z1 Z2T 0 T Z
B1T S I D11 2 0 < 0.
0 0 I
C1 D11 I 0 I
in which
V1
V2
Then, an easy calculation shows that the second inequality in (3.62) can be rewritten
as
T T AR + RAT RC1T B1 V1 0
V1 V2 0
C1 R I D11 V2 0 < 0.
0 0 I T
B1 T
D11 I 0 I
In this way, the two inequalities (3.62) have been transformed into two inequalities
involving the submatrix S of X and the submatrix R of X 1 . To complete the
analysis, it remains to connect these matrices to each other. This is achieved via the
following lemma.
and 1
S N R
= (3.67)
NT Z
if and only if
rank(I SR) k (3.68)
3.6 The Coupled LMIs Approach to the Problem 77
and
S I
0. (3.69)
I R
Thus I SR = NM T and this implies (3.68), because N has k columns. Now, set
I R
T = .
0 MT
Using (3.70), the first of which implies MN T = I RS because S and R are sym-
metric, observe that
S I
T TX T = .
I R
because the matrix X is positive definite by assumption. This concludes the proof
of the necessity.
To prove sufficiency, let k k be the rank of I SR and let N,
M two n k
matrices of rank k such that
I SR = N M
T. (3.71)
N T R + Z M
T=0 (3.72)
N T R RN M
M T = (I RS)R R(I SR) = 0.
N T RL + RN = 0.
M
This shows that the matrix Z = (N T RL)T satisfies (3.72). The matrix Z is symmetric.
In fact, note that
N T R + Z M
0 = M( T ) = (I RS)R + M
Z M
T = R RSR + M
Z M
T
Z M
and hence M T is symmetric. This yields
Z M
0=M Z T M
TM Z Z T )M
T = M( T
is a solution of
S I I R .
= X (3.74)
N T 0 0MT
The symmetric matrix X thus found is invertible because, otherwise, the indepen-
dence of the rows of the matrix on the left-hand side of (3.74) would be contradicted.
It is easily seen that
R
X 1 = T . (3.75)
M
observe that
S I
T T X T = .
I R
Suppose x T X x < 0 for some x = 0. Using the fact that the rows of T are indepen-
dent, find y such that x = T y. This would make
S I
yT
y < 0,
I R
Remark 3.3 Note that condition (3.69) implies S > 0 and R > 0. This is the con-
sequence of the arguments used in the proof of the previous lemma, but can also
be proven directly as follows. Condition (3.69) implies that the two diagonal blocks
S and R are positive semidefinite. Consider the quadratic form associated with the
matrix on left-hand side of (3.69),
T T S I x
V (x, z) = x z = x T Sx + zT Rz + 2zT x.
I R z
Clearly, by choosing an appropriate c, the right hand side can be made strictly nega-
tive. Thus, V (x, z) is not positive semidefinite and this contradicts (3.69). The same
arguments are used to show that also S is nonsingular.
This lemma establishes a coupling condition between the two submatrices S and
R identified in the previous analysis that determines the positivity of the matrix X .
Using this lemma it is therefore possible to arrive at the following conclusion.
Theorem 3.3 Consider a plant of modeled by equations of the form (3.48). Let
V1 , V2 , Z1 , Z2 be matrices such that
Z1 V
Im = Ker C2 D21 , Im 1 = Ker B2T D12T
.
Z2 V2
The problem of -suboptimal H feedback design has a solution if and only if there
exist symmetric matrices S and R satisfying the following system of linear matrix
inequalities
80 3 The Small-Gain Theorem for Linear Systems
T T AT S + SA SB1 C1T Z1 0
Z1 Z2 0 T
B1T S I D11 Z2 0 < 0 (3.76)
0 0 I
C1 D11 I 0 I
T T AR + RAT RC1T B1 V1 0
V1 V2 0
C1 R I D11 V2 0 < 0 (3.77)
0 0 I
B1T T
D11 I 0 I
S I
0. (3.78)
I R
In particular, there exists a solution of dimension k if and only if there exist R and S
satisfying (3.76)(3.78) and, in addition,
The result above describes necessary and sufficient conditions for the existence
of a controller that solves the problem of -suboptimal H feedback design. For
the actual construction of a controller, one may proceed as follows. Assuming that S
and R are positive definite symmetric matrices satisfying the system of linear matrix
inequalities (3.76)(3.78), construct a matrix X as indicated in the proof of Lemma
3.4, that is, find two n k matrices N and M such that
I SR = NM T
By construction, the matrix in question is positive definite which satisfies (3.62) and
their equivalent versions (3.59). Thus, according to Lemma 3.3, there exists a matrix
K satisfying (3.56). This is a linear matrix inequality in the only unknown K. The
solution of such inequality provides the required controller as
Ac Bc
K= .
Cc Dc
References
1. B.D.O. Anderson, J.B. Moore, Optimal Control: Linear Quadratic Methods (Prentice Hall,
Englewoof Cliffs, 1990)
2. T. Basar, P. Bernhard, H Optimal Control and Related Minimaz Design Problems (Birkuser,
Boston, 1990)
3. S.P. Boyd, L. El Ghaoui, E. Feron, V. Balakrishnan, Linear Matrix Inequalities in System and
Control Theory. Studies in Applied and Numerical Mathematics (SIAM, Philadelphia, 1994)
4. C.A. Desoer, M. Vidyasagar, Feedback Systems: Input-Output Properties (Academic Press,
New York, 1975)
5. J. Doyle, B. Francis, A. Tannenbaum, Feedback Control Theory (MacMillan, New York, 1992)
6. B. Francis, A Course in H Control Theory (Springer, Berlin, 1987)
7. P. Gahinet, P. Apkarian, A linear matrix inequality approach to H control. Int. J. Robust
Nonlinear Control 4, 421448 (1994)
8. P. Gahinet, A. Nemirovski, A.J. Laub, M. Chilali, LMI Control Toolbox: For Use with MATLAB
(The Mathworks Inc., Massachusetts, 1995)
9. C. Scherer, S. Weiland, Linear matrix inequalities in control, in The Control Handbook: Control
Systems Advanced Methods, Chap. 24, ed. by W.S. Levine (CRC Press, Boca Raton, 2011)
10. C.W. Scherer, P. Gahinet, M. Chilali, Multiobjective output-feedback control via LMI opti-
mization. IEEE Trans. Autom. Control 42, 896911 (1997)
11. A.J. Van der Schaft, L2 Gain and Passivity Techniques in Nonlinear Control (Springer, London,
1996)
12. J.G. VanAntwerp, R.D. Braatz, A tutorial on linear and bilinear matrix inequalities. J. Process
Control 10, 363385 (2000)
13. J. Willems, Dissipative dynamical systems. Arch. Ration. Mech. Anal. 45, 321393 (1972)
Chapter 4
Regulation and Tracking in Linear Systems
In the last part of the previous chapter, we have considered a plant modeled as
in (3.48), andregarding y as an output available for measurement and u as an
input available for controlwe have studied the problem of finding a controller of
the form (3.49) yielding a stable closed-loop system having a transfer function
between input v and output zwhose H norm does not exceed a given number
. The problem in question was motivated by the interest in solving a problem of
robust stability, but it has an independent interest per se. In fact, regarding v as a
vector of external disturbances, affecting the behavior of the controlled plant, and
z as set of variables of special interest, the problem considered in Sect. 3.6 can be
regarded as the problem of finding a controller thatwhile keeping the closed-loop
stableenforces a prescribed attenuation of the effect of the disturbance v on the
variable of interest z, the attenuation being expressed in terms of the H norm (or
of the L2 gain, if desired) of the resulting system.
In the present chapter, we continue to study a problem of this kind, i.e., the control
of a plant affected by external disturbances, from which certain variables of interest
have to be protected, but with some differences. Specifically, while in Sect. 3.6 we
have considered the case in which the influence of the disturbances on the variables of
interest had to be attenuated by a given factor, we consider in this section the case in
which the influence of the disturbances on the variables of interest should ultimately
vanish. This is indeed a much stronger requirement and it is unlikely that it might be
enforced in general. It can be enforced, though, if the disturbances happen to belong
to a specific (well-defined) family of signals. This gives rise to a specific setup, known
as problem of asymptotic disturbance rejection and/or tracking, or more commonly
problem of output regulation, that will be explained in more detail hereafter.
For consistency with the notations currently used in the literature dealing with
the specific problem addressed in this chapter, the controlled plant (compare with
x = Ax + Bu + Pw
e = Ce x + Qe w (4.1)
y = Cx + Qw.
The first equation of this system describes a plant with state x Rn and control
input u Rm , subject to a set of exogenous input variables w Rnw which includes
disturbances (to be rejected) and/or references (to be tracked). The second equation
defines a set of regulated (or error) variables e Rp , which are expressed as a linear
combination of the plant state x and of the exogenous input w.1 The third equation
defines a set of measured variables y Rq , which are assumed to be available for
feedback, and are also expressed as a linear combination of the plant state x and of
the exogenous input w.
The control action to (4.1) is to be provided by a feedback controller which
processes the measured information y and generates the appropriate control input.
In general, the controller in question is a system modeled by equations of the form
x c = Ac xc + Bc y
(4.2)
u = Cc xc + Dc y,
with state xc Rnc , which yields a closed-loop system modeled by equations of the
form
x = (A + BDc C)x + BCc xc + (P + BDc Q)w
x c = Bc Cx + Ac xc + Bc Qw (4.3)
e = Ce x + Qe w.
This system is seen as a system with input w and output e. The purpose of the
control is to make sure that the closed-loop system be asymptotically stable and that
the regulated variable e, viewed as a function of time, asymptotically decays to zero
as time tends to , for every possible initial state and for every possible exogenous
input in a prescribed family of functions of time. This requirement is also known
as property of output regulation. For the sake of mathematical simplicity, and also
because in this way a large number of relevant practical situations can be covered, it
is assumed that the family of the exogenous inputs w() which affect the plant, and
for which asymptotic decay of the regulated variable is to be achieved, is the family
of all functions of time which is the solution of a homogeneous linear differential
equation
w = Sw (4.4)
1 If some components of w are (external) commands that certain variables of interest are required to
track, then some of the components of e can be seen as tracking errors, that is differences between
the actual values of those variables of interest and their expected reference values. Overall, the
components of e can simply be seen as variables on which the effect of w is expected to vanish
asymptotically.
4.1 The Problem of Asymptotic Tracking and Disturbance Rejection 85
with state w Rnw , for all possible initial conditions w(0). This system, which is
viewed as a mathematical model of a generator for all possible exogenous input
functions, is called the exosystem (Fig. 4.1). In this chapter, we will discuss this
design problem at various levels of generality.
Note that, without loss of generality, in the analysis of this problem it can be
assumed that all the eigenvalues of S have nonnegative real part. In fact, if there were
eigenvalues having negative real part, system (4.4) could be split (after a similarity
transformation) as
w s = Ss ws
w u = Su wu
For expository reasons we consider first the (nonrealistic) case in which the full state
x of the plant and the full state w of the exosystem are available for measurement,
i.e., the case in which the measured variable y in (4.1) is y = col(x, w). This is called
the case of full information. We also assume that all system parameters are known
exactly.
In this setup, we consider the special version of the controller (4.2) in which
u = Dc y, that we rewrite for convenience as
86 4 Regulation and Tracking in Linear Systems
u = Fx + Lw. (4.5)
Proposition 4.1 The problem of output regulation in the case of full information has
a solution if and only if
(i) the matrix pair (A, B) is stabilizable
(ii) there exists a solution pair (, ) of the linear matrix equations
S = A + B + P
(4.6)
0 = Ce + Qe .
Suppose the pair of matrices F, L is such the controller (4.5) solves the problem
of output regulation. Then, the matrix A + BF has all eigenvalues with negative real
part and, hence, the matrix pair (A, B) must be stabilizable.
The eigenvalues of system (4.7) are those of (A + BF) and those of S. The former
are all in C while none of the latter, in view of the standing assumption described
at the end of the previous section, is in C . Hence this system possesses a stable
eigenspace and a center eigenspace. The former can be described as
0
V s = Im ,
I
while the latter, which is complementary to the stable eigenspace, can be described
as
I
V c = Im ,
Setting
= L + F
If the controller (4.5) solves the problem of output regulation, we must have
ess (t) = 0 and this can only occur if satisfies Ce + Qe = 0. In terms of state
trajectories, this is equivalent to say that the steady-state component of any state
trajectory must be contained in the kernel of the map e = Qe w + Ce x. We have
shown in this way that the solution of the Sylvester equation (4.8) necessarily
satisfies the second equation of (4.6). In summary, if the pair of matrices F, L is such
that the controller (4.5) solves the problem of output regulation, Eq. (4.6) must have
a solution pair (, ).
[Sufficiency] Suppose that (A, B) is stabilizable and pick F such that the eigen-
values of A + BF have negative real part. Suppose Eq. (4.6) have a solution (, )
and pick L as
L = F,
u = w + F(x w).
x = x w
and, bearing in mind the first equation of (4.6) and the choice of L, observe that, in
the new coordinates,
x = (A + BF)x .
Hence limt x (t) = 0, because the eigenvalues of A + BF have negative real part.
In the new coordinates,
lim e(t) = 0
t
88 4 Regulation and Tracking in Linear Systems
Equation (4.6) are called the linear regulator equations or, also, the Franciss
equations.2 The following result is useful to determine the existence of solutions.3
Lemma 4.1 The Francis equation (4.6) have a solution for any (P, Qe ) if and only
if
A I B
rank = # rows (S). (4.9)
Ce 0
If this is the case and the matrix on the left-hand side of (4.9) is square (i.e., the
control input u and the regulated output e have the same number of components),
the solution (, ) is unique.
The condition (4.9) is usually referred to as the nonresonance condition. Note
that the condition is necessary and sufficient if the existence of a solution for any
(P, Qe ) is sought. Otherwise, it is simply a sufficient condition. Note also that such
condition requires m p, i.e., that the number of components of the control input u
be larger than or equal to the number of components of the regulated output.
Example 4.1 As an example of what the nonresonance condition (4.9) means and
why it plays a crucial role in the solution of (4.6), consider the case in which dim(e) =
dim(u) = 1. Observe that the first two equations of (4.1), together with (4.4), can be
written in the form of a composite system with input u and output e as
w S 0 w 0
= + u
x PA x B
(4.10)
w
e = Qe Ce .
x
Bearing in mind the conditions (2.2) used to identify the value of the relative degree
of a system, let r be such that Ce Ak B = 0 for all k < r 1 and Ce Ar1 B = 0. A
simple calculation shows that, for any k 0,
k
S 0 0 0
= .
PA B Ak B
Therefore,
S 0 k 0
Qe Ce =0
PA B
for k = 1, . . . , r 2 and
2 See [1].
3 See Appendix A.2.
4.2 The Case of Full Information and Francis Equations 89
S 0 r1 0
Qe Ce = Ce Ar1 B.
PA B
Thus, system (4.10), viewed as system with input u and output e has relative
degree r and can be brought to (strict) normal form, by means of a suitable change
of variables. In order to identify such change of variables, let
1 = Qe w + Ce x
i+1 = i
i+1 = i = Ri w + Ce Ai1 x
= Ri Sw + Ce Ai1 (Ax + Bu + Pw) = (Ri S + Ce Ai1 P)w + Ce Ai1 x,
which has the required form with Ri+1 = (Ri S + Ce Ai1 P).
Then, set (compare with (2.4))
Ce R1
Ce A R2
T1 =
,
R=
,
Ce Ar1 Rr
w = Sw
z = T0 Ax + T0 Pw
1 = 2
r1 = r
r = Ce Ar x + bu + (Rr S + Ce Ar1 P)w
e = 1 ,
in which
b = Ce Ar1 B = 0.
Then, it can be concluded that the (strict) normal form of such system has the fol-
lowing expression
w = Sw
z = A00 z + A01 + P0 w
1 = 2
(4.11)
r1 = r
r = A10 z + A11 + bu + P1 w
e = 1 ,
in which the four matrices A00 , A01 , A10 , A11 are precisely the matrices that charac-
terize the (strict) normal form of a system defined as
x = Ax + Bu
e = Ce x,
and P0 , P1 are suitable matrices. Hence, as shown in Sect. 2.1, the matrix A00 is a
matrix whose eigenvalues coincide with the zeros of the transfer function
The normal form (4.11) can be used to determine a solution of Francis equations.
To this end, let be partitioned as
0
=
1
consistently with the partition of x , and let 1,i denote the ith row of 1 .
Then, it is immediate to check that Franciss equations are rewritten as
4.2 The Case of Full Information and Francis Equations 91
0 S = A00 0 + A01 1 + P0
1,1 S = 1,2
(4.13)
1,r1 S = 1,r
1,r S = A10 0 + A11 1 + b + P1
0 = 1,1 .
The last equation, along with the second, third, and rth, yield 1,i = 0 for all i,
i.e.,
1 = 0.
0 S = A00 0 + P0
(4.14)
0 = A10 0 + b + P1 .
The first of these equations is a Sylvester equation in 0 , that has a unique solution
if and only if the none of the eigenvalues of S is an eigenvalue of the matrix A00 . Since
the eigenvalues of A00 are the zeros of (4.12), this equation has a unique solution
if and only if none of the eigenvalues of S is a zero of Te (s). Entering the solution
0 of this equation into the second one yields an equation that, since b = 0, can be
solved for , yielding
1
= [A10 0 + P1 ].
b
In summary, it can be concluded that Francis equations have a unique solution if
and only if none of the eigenvalues of S coincides with a zero of the transfer function
Te (s).
Of course the condition thus found must be consistent with the condition resulting
from Lemma 4.1, specialized to the present context in which m = p = 1. In this
case, the condition (4.9) becomes
A I B
det = 0 (S).
Ce 0
Now, it is known (see Sect. 2.1) that the roots of the polynomial
A I B
det
Ce 0
coincide with the zeros of the transfer function Te (s). Thus, the condition indicated
in Lemma 4.1 is the condition that none of the eigenvalues of S is a zero of Te (s),
which is precisely the condition obtained from the construction above.
92 4 Regulation and Tracking in Linear Systems
Consider now the case in which the feedback law is provided by a controller that does
not have access to the full state x of the plant and the full state w of the exosystem.
We assume that the controller has access to the regulated output e and, possibly, to
an additional supplementary set of independent measured variables. In other words,
we assume that the vector
y = Cx + Qw
in which e is the regulated output and yr Rpr . As opposite to the case considered
in the previous section, this is usually referred to as the case of measurement feed-
back. In this setup the controlled plant, together with the exosystem, is modeled by
equations of the form
w = Sw
x = Ax + Bu + Pw
(4.15)
e = Ce x + Qe w
yr = Cr x + Qr w.
The control is provided by a generic dynamical system with input y and output u,
modeled as in (4.2). Proceeding as in the first part of the proof of Proposition 4.1, it
is easy to deduce the following necessary conditions for the solution of the problem
of output regulation.
Proposition 4.2 The problem of output regulation in the case of measurement feed-
back has a solution only if
(i) the triplet {A, B, C} is stabilizable and detectable
(ii) there exists a solution pair (, ) of the Francis equation (4.6).
If a controller of the form (4.2) solves the problem of output regulation, all the
eigenvalues of the matrix
A + BDc C BCc
Bc C Ac
4.3 The Case of Measurement Feedback: Steady-State Analysis 93
have negative real part. Hence, the triplet {A, B, C} must be stabilizable and
detectable.
Since by assumption S has all eigenvalues on the imaginary axis, system (4.16)
possesses two complementary invariant subspaces: a stable eigenspace and a center
eigenspace. The latter, in particular, can be expressed as
I
V c = I
c
Setting
= Dc C + Cc c + Dc Q
x c = Ac xc + Bce e + Bcr yr
u = Cc xc + Dce e + Dcr yr .
Bearing in mind the fact that, if the controller solves the problem of output reg-
ulation, the matrix must satisfy the second of (4.6), the equation above reduces
to
A + BDcr Cr BCc
P + BDcr Qr
S= + ,
c Bcr Cr Ac c Bcr Qr
c S = Ac c + Bcr (Cr + Qr )
(4.20)
= Cc c + Dcr (Cr + Qr ),
and the state xc (t) of the controller evolves (see again (4.18)) as
The first equation of (4.20) expresses precisely the property that c w(t) is the steady-
state response of the controller, when the latter is forced by a steady-state input of
the form (4.21)(4.22). The second equation of (4.20), in turn, shows that the output
of the controller, in steady state, is a function of the form
uss (t) = Cc xc,ss (t) + Dce ess (t) + Dcr yr,ss (t)
= Cc c w(t) + Dcr (Cr + Qr )w(t) = w(t).
The latter, as predicted by Francis equations, is a control able to force in the con-
trolled plant a steady-state trajectory of the form xss (t) = w(t) and consequently to
keep ess (t) identically zero. The property thus described is usually referred to as the
internal model property: any controller that solves the problem of output regulation
necessarily embeds a model of the feedforward inputs needed to keep e(t) identically
zero.
The possibility of constructing a controller that solves the problem in the case of
measurement feedback reposes on the following preliminary result. Let
() = s0 + s1 + + sd1 d1 + d
and T
G0 = 0 0 0 1 .
With and G constructed in this way, consider now the system obtained by letting
the regulated output of (4.1) drive a postprocessor characterized by the equations
= + Ge. (4.28)
Since the minimal polynomial of the matrix in (4.28) coincides with the minimal
polynomial of the matrix S that characterizes the exosystem, system (4.28) is usually
called an internal model (of the exosystem). Note that, in the coordinates = T ,
with T such that (4.25) hold, system (4.28) can be seen as a bench of p identical
subsystems of the form
i = S0 i + G0 ei
= + Ge
= As + Bs y + Js (4.29)
u = Cs + Ds y + Hs
4 The arguments uses hereafter are essentially the same as those used in [24].
4.4 The Case of Measurement Feedback: Construction of a Controller 97
Fig. 4.3 The plant, augmented with the exosystem, controlled by (4.29)
The first step in proving that a controller of the form (4.29) can solve the problem
of output regulation consists in the analysis of the properties of stabilizability and
detectability of a systemwith state (x, ), input u and output ya defined as follows
x A 0 x B
= + u
GCe 0
(4.30)
C0 x
ya = ,
0 I
that will be referred to as the augmented system. As a matter of fact, it turns out that
the system in question has the following important property.
Lemma 4.2 The augmented system (4.30) is stabilizable and detectable if and only
if
(i) the triplet {A, B, C} is stabilizable and detectable
(ii) the nonresonance condition
A I B
rank =n+p (S). (4.31)
Ce 0
holds.
Proof To check detectability we look at the linear independence of the columns of
the matrix
98 4 Regulation and Tracking in Linear Systems
A I 0
GCe I
C 0
0 I
for any having nonnegative real part. Taking linear combinations of rows (and
using the fact that the rows of Ce are part of the rows of C), this matrix can be easily
reduced to a matrix of the form
A I 0
0 0
C 0
0 I
from which it is seen that the columns are linearly independent if and only if so are
those of the submatrix
A I
.
C
for any having nonnegative real part. Taking linear combinations of columns, this
matrix is initially reduced to a matrix of the form
A I 0 0 0 0 B
0 0 I 0 0 0
0 0 0 I 0 0
0 0 0 0 I 0
Ce ()I 0
After permutation of rows and columns, one finally obtains a matrix of the form
A I B 0 0
Ce 0 ()I 0
0 0 0 I
Thus, it is concluded that system (4.30) is stabilizable if and only if so is the pair
(A, B) and the nonresonance condition (4.31) holds.
If the assumptions of this lemma hold, the system (4.30) is stabilizable by means
of a (dynamic) feedback. In other words, there exists matrices As , Bs , Cs , Ds , Js , Hs ,
such that the closed-loop system obtained controlling (4.30) by means of a system
of the form
= As + Bs ya
+ Js ya
(4.32)
u = Cs + Ds ya
+ Hs ya
in which ya
and ya
Remark 4.1 A simple expression of such stabilizer can be found in this way. By
Lemma 4.2 the augmented system (4.30) is stabilizable. Hence, there exist matrices
L Rmn and M Rmpd such that the system
x A 0 B x
= + LM
GCe 0
is stable. Moreover, since the pair (A, C) is detectable, there is a matrix N such that
(A NC) has all eigenvalues in C . Let the augmented system (4.30) be controlled
by
100 4 Regulation and Tracking in Linear Systems
= A + N(ya
C ) + B(L + Mya
)
(4.33)
u = L + Mya
With this result in mind, consider, for the solution of the problem of out-
put regulation, a candidate controller of the form (4.29), in which the matrices
As , Bs , Cs , Ds , Js , Hs , are chosen in such a way that (4.32) stabilizes system (4.30).
This yields a closed-loop system of the form
w S 0 0 0 w
x P + BDs Q A + BDs C BHs BCs x
= . (4.34)
GQe GCe 0
Bs Q Bs C Js As
has all eigenvalues with negative real part. Hence, the closed-loop system possesses a
stable eigenspace and a center eigenspace. The latter, in particular, can be expressed
as
I
x
V = Im
c
(4.36)
S = + G(Ce x + Qe ). (4.38)
This is a key property, from which it will be deduced that the proposed controller
solves the problem of output regulation. In fact, the following result holds.
Lemma 4.3 If is the matrix defined in (4.23) and G is the matrix defined in (4.24),
Eq. (4.38) implies
4.4 The Case of Measurement Feedback: Construction of a Controller 101
Ce x + Qe = 0. (4.39)
,1 S = ,2
,2 S = ,3
,d1 S = ,d
,d S = s0 ,1 s1 ,2 sd1 ,d + Ce x + Qe .
,i = ,1 S i1 ,
= Cs + Ds (Cx + Q) + Hs ,
We consider in this section the case in which the plant is affected by structured
uncertainties, that is the case in which the coefficient matrices that characterize the
model of the plant depend on a vector of uncertain parameters, as in
w = Sw
x = A()x + B()u + P()w
(4.41)
e = Ce ()x + Qe ()w
yr = Cr ()x + Qr ()w
We show in what follows how the results discussed in the previous section can be
enhanced to obtain a robust controller. First of all, observe that if a robust controller
exists, this controller must solve the problem of output regulation for each of .
Hence, for each of such values, the necessary conditions for existence of a controller
determined in the earlier sections must hold: the triplet {A(), B(), C()} must be
stabilizable and detectable and, above all, the -dependent Francis equations
It should be stressed, though, that a result such as that of the Lemma 4.2
cannot be invoked anymore. In fact, the (necessary) assumption that the triplet
5 Itis worth observing that, since by assumption the matrix S is not affected by parameter uncer-
tainties, so is its minimal polynomial () and consequently so is the matrix defined in (4.23).
4.5 Robust Output Regulation 103
{A(), B(), C()} is stabilizable and detectable for every no longer guaran-
tees the existence of a robust stabilizer for (4.43). For example, a stabilizer having
the structure (4.33) cannot be used, because the latter presumes a precise knowledge
of the matrices A, B, C that characterize the model of the plant, and this is no longer
the case when such matrices depend on a vector of uncertain parameters.
We will show later in the chapter how (and under what assumptions) such a robust
stabilizer may be found. In the preset section, we take this as an hypothesis, i.e., we
suppose that there exists a dynamical system, modeled as in (4.32), that robustly
stabilizes (4.43). To say that the plant (4.43), controlled by (4.32), is robustly stable
is the same thing as to say that the matrix
A() + B()Ds C() B()Hs B()Cs
GCe () 0 (4.44)
Bs C() Js As
= + Ge
= As + Bs y + Js (4.45)
u = Cs + Ds y + Hs .
The resulting closed-loop system is an autonomous linear system having the form
w S 0 0 0 w
x P() + B()Ds Q() A + B()Ds C() B()Hs B()Cs x
= .
GQe () GCe () 0
Bs Q() Bs C() Js As
(4.46)
Since the matrix (4.44) is Hurwitz for every value of and S has all eigenvalues on
the imaginary axis, the closed-loop system possesses two complementary invariant
subspaces: a stable eigenspace and a center eigenspace. The latter, in particular, has
the expression
I
x ()
V c = Im ()
()
in which (x (), (), ()) is the (unique) solution of the Sylvester equation
(compare with (4.37))
104 4 Regulation and Tracking in Linear Systems
x A() + B()Ds C() B()Hs B()Cs x
S = GCe () 0
Bs C() Js As
P() + B()Ds Q()
+ GQe () .
Bs Q()
From this Sylvester equation, one deduces (compare with (4.38)) that
Ce ()x () + Qe () = 0.
In steady state xss (t) = x ()w(t) and, in view of equation above, we conclude that
ess (t) = 0. We summarize the discussion as follows.
Proposition 4.4 Let be a matrix of the form (4.23) and G a matrix of the form
(4.24). Suppose the system (4.43) is robustly stabilized by a stabilizer of the form
(4.32). Then the problem of robust output regulation is solvable, in particular by
means of a controller of the form (4.45).
Remark 4.3 One might be puzzled by the absence of the nonresonance condition
in the statement of the proposition above. It turns out, though, that this condition
is implied by the assumption of robust stabilizability of the augmented system. As
a matter of fact, in this statement it is assumed that there exists a stabilizer that
stabilizes the augmented plant (4.43) for every . As a trivial consequence, the latter
is stabilizable and detectable for every . This being the case, it is seen from Lemma
4.2 that, if the system (4.43) is robustly stabilized by a stabilizer of the form (4.32),
necessarily the triplet {A(), B(), C()} is stabilizable and detectable for every
and the nonresonance condition must hold for every . We stress also that the
nonresonance condition, which we have seen is necessary, also implies the existence
of a solution of Franciss equations (4.42). This is an immediate consequence of
Lemma 4.1.
We discuss in this section the design of regulators in the special case of a plant in
which the number of regulated outputs is equal to the number of control inputs,
and no additional measurements outputs are available. Of course, the design of a
regulator could be achieved by following the general construction described in the
previous section, but in this special case alternative (and somewhat simpler) design
procedures are available, which will be described in what follows.
4.6 The Special Case in Which m = p and pr = 0 105
Immediate consequences of the assumption m = p are the fact that the nonreso-
nance condition can be rewritten as
A I B
det = 0 (S), (4.47)
Ce 0
and the fact that, if this is the case, the solution , of Franciss equations (4.6) is
unique. If, in addition, pr = 0, the construction described above can be simplified
and an alternative structure of the controller is possible.
If pr = 0, in fact, the structure of the controller (4.29) becomes
= + Ge
= As + Bs e + Js (4.48)
u = Cs + Ds e + Hs .
Js = Bs
Hs = Ds
= + Ge
(4.49)
e = + e,
= As + Bs e
(4.50)
u = Cs + Ds e
i.e., if the latter is stabilizable and detectable. In this respect, the following result is
useful.
Lemma 4.4 Let be such that G is a Hurwitz matrix. Then, the augmented
system (4.51) is stabilizable and detectable if and only if
(i) the triplet {A, B, Ce } is stabilizable and detectable
(ii) the nonresonance condition (4.47) holds.
are independent for all having nonnegative real part. Taking linear combinations
of rows, we transform the latter into
A I 0
0 G I ,
Ce
and we observe that, if is such that the matrix G is Hurwitz, the augmented
system (4.51) is detectable if and only if so is the pair (A, Ce ).
6 Note that the filter (4.49) is an invertible system, the inverse being given by
= ( G ) + Ge
e = + e .
Hence, if is chosen in such that a way G is Hurwitz, the inverse of (4.49) is a stable system.
4.6 The Special Case in Which m = p and pr = 0 107
in which , G have the form (4.23), (4.24), is such that a G is Hurwitz and
As , Bs , Cs , Ds are such that (4.50) stabilizes the augmented plant (4.51).
The controller (4.52) is, as observed, the cascade of two subsystems. It would be
nice to check whether these subsystems could be swapped, i.e., whether the same
result could be obtained by means of a controller consisting of a preprocessing filter
of the form
= + Gu
(4.53)
u = + u
= As + Bs e
(4.54)
u = Cs + Ds e
On the basis of our earlier discussions, it can be claimed that the problem of output
regulation is solved if the matrix
A + BDs Ce B BCs
GDs Ce GCs (4.56)
Bs Ce 0 As
is Hurwitz and the associated center eigenspace of (4.55) is contained in the kernel
of the error map e = Ce x + Qe w.
The matrix (4.56) is Hurwitz if (and only if) the stabilizer (4.54) stabilizes the
augmented plant
x A B x B
= + u
0 G
(4.57)
x
e = Ce 0 .
In what follows the conditions under which this is possible are discussed.
Lemma 4.5 Let and G be defined as in (4.23) and (4.24). Let be such that the
matrix G is Hurwitz. Then, the pair (, ) is observable.
Proof Since G is Hurwitz, the pair (, ) is by definition detectable. But has
by assumption all eigenvalues with nonnegative real part. In this case, detectability
(of the pair (, )) is equivalent to observability.
Lemma 4.6 Let be defined as in (4.23). Let be a (p dp) matrix such that the
pair (, ) is observable. Then, there exists a nonsingular matrix T such that
= I 0 0 T
and
T = T .
d1
is invertible. In view of the second of (4.27), a simple calculation shows that this
matrix renders the two identities in the lemma fulfilled.
Lemma 4.7 Let be such that G is a Hurwitz matrix. Then, the augmented
system (4.57) is stabilizable and detectable if and only if
4.6 The Special Case in Which m = p and pr = 0 109
are linearly independent for each having nonnegative real part if and only if the
pair (A, B) is stabilizable. To check detectabililty, we need to look at the linear
independence of the columns of the matrix
A I B
0 I . (4.58)
Ce 0
By Lemma 4.5, the pair (, ) is observable. Let T be a matrix that renders the two
identities in Lemma 4.6 fulfilled, and have the matrix (4.58) replaced by the matrix
I 00 A I B A I B T 1
0 T 0 0 I I 01 = 0 I ,
0T
00 I Ce 0 Ce 0
From this, the check of detectability condition proceeds essentially as in the check of
stabilizability condition in Lemma 4.2. Taking linear combinations and permutation
of columns and rows, one ends up with a matrix of the form
A I B 0
Ce 0 0
0 ()I 0
0 0 I
110 4 Regulation and Tracking in Linear Systems
If the matrix (4.56) is Hurwitz, the center eigenspace of (4.55) can be expressed
as
I
x
V = Im
c
Lemma 4.8 Let be defined as in (4.23). Let be a (p dp) matrix such that the
pair (, ) is observable. Then, given any matrix , there exists a matrix such
that
S =
(4.59)
= .
Proof Let T be a matrix such that the two identities in Lemma 4.6 hold. Change
(4.59) in
(T )S = T T 1 (T ) = (T )
= T 1 (T ) = I 0 0 (T ),
S d1
x = , = , = 0
is a solution of the Sylvester equation above, as a matter of fact, the unique solution
of that equation. Thus, in particular, the steady state of the closed-loop system (4.55)
4.6 The Special Case in Which m = p and pr = 0 111
is such that xss (t) = w(t) with obeying the second equation of (4.6). As a con-
sequence, ess (t) = 0 and the problem of output regulation is solved. We summarize
the discussion as follows.
in which , G have the form (4.23), (4.24), is such that a G is Hurwitz and
As , Bs , Cs , Ds are such that
= As + Bs e
(4.61)
u = Cs + Ds e.
x =
S =
S = As + Bs
= Cs + Ds
where , is the unique solution of Franciss equations (4.6). In the second mode
of control, on the other hand, in steady state we have
in which
x =
S =
=
with , the unique solution of Franciss equations (4.6). In particular, in the second
mode of control, in steady state the stabilizer (4.61) is at rest.
In both modes of control the internal model has an identical structure, that of the
system
= + Gu
(4.62)
y = + u ,
in which is the matrix defined in (4.23). In the discussion above, the matrix G
has been taken as in (4.24), and the matrix was any matrix rendering G a
Hurwitz matrix. However, it is easy to check that one can reverse the roles of G and
. In fact, using the state transformation
= T
with T defined as in the proof of Lemma 4.6 it is seen that an equivalent realization
of (4.62) is
= + G
u
(4.63)
y = + u ,
= TG and
in which G
= I 0 0 (4.64)
(here all blocks are p p, with p = m by hypothesis). Thus, one can design the
internal model picking as in (4.23), as in (4.64) and then choosing a G that
makes G a Hurwitz matrix.
Finally, note that if we let F denote the Hurwitz matrix F = G , the internal
model can be put in the form
= F + Gy
(4.65)
y = + u .
The two modes of control lend themselves to solve also a problem of robust
regulation. From the entire discussion above, in fact, one can arrive at the following
conclusion.
Proposition 4.7 Let be a matrix of the form (4.23), G a matrix of the form (4.24),
and a matrix such that G is Hurwitz (alternatively: a matrix of the form
(4.64) and G a matrix such that G is Hurwitz). If system (4.51) is robustly
stabilized by a stabilizer of the form (4.50), the problem of robust output regulation
is solvable, in particular by means of a controller of the form (4.52). If system (4.57)
4.6 The Special Case in Which m = p and pr = 0 113
is robustly stabilized by a stabilizer of the form (4.61), the problem of robust output
regulation is solvable, in particular by means of a controller of the form (4.60).
w = Sw
x = A()x + B()u + P()w (4.66)
e = C()x + Q()w.
Let the system be controlled by a controller of the form (4.60). We know from
the previous analysis that the controller in question solves the problem of output
regulation if there exists a stabilizer of the form (4.61) that stabilizes the augmented
system
x A() B() x B()
= + u
0 G
(4.67)
x
e = C() 0 .
Suppose the triplet in question has a well-defined relative degree r between control
input u and regulated output e, independent of . It is known from Chap. 2 that a
single-input single-output system having well-defined relative degree and all zeros
with negative real part can be robustly stabilized by (dynamic) output feedback.
Thus, we seek assumptions ensuring that the augmented system so defined has a well-
defined relative degree and all zeros with negative real part. An easy calculation shows
that, if the triplet {A(), B(), C()} has relative degree r,8 then the augmented
system (4.67) still has relative degree r, between the input u and the output e. In fact,
for all k r 1
A() B() k
C() 0 = C()Ak () 0
0
from which it is seen that the relative degree is r, with high-frequency gain
A() B() r1 B()
C() 0 = C()Ar1 ()B().
0 G
no matter what is. Thus, the zeros of the system are the roots of the polynomial
A() I 0 B()
det 0 G I G = 0.
C() 0 0
Thus, the n + d r zeros of the augmented plant are given by the n r zeros of
the triplet {A(), B(), C()} and by the d eigenvalues of the matrix G . If,
as indicated above, the matrix is chosen in such a way that the matrix G
is Hurwitz (as it is always possible), we see that the zeros of the augmented system
have negative real part if so are the zeros of the triplet {A(), B(), C()}.
We summarize the conclusion as follows.
Proposition 4.8 Consider an uncertain system of the form (4.66). Suppose that the
triplet {A(), B(), C()} has a well-defined relative degree and all its n r zeros
have negative real part, for every value of . Let be a matrix of the form (4.23),
G a matrix of the form (4.24), and a matrix such that G is Hurwitz. Then,
the problem of robust output regulation is solvable, in particular by means of a
controller of the form (4.60), in which (4.61) is a robust stabilizer of the augmented
system (4.67).
Example 4.2 To make this result more explicit, it is useful to examine in more detail
the case of a system having relative degree 1. As a by-product, additional insight
in the design procedure is gained, that proves to be useful in similar contexts in the
next sections. Consider the case in which the triplet {A(), B(), C()} has relative
degree 1, i.e., is such that C()B() = 0. Then, as shown in Example 4.1, system
4.7 The Case of SISO Systems 115
w S 0 w 0
= + u
x P() A() x B()
w
e = Q() C()
x
has relative degree 1. It can be put in normal form, obtaining a system (see again
Example 4.1)
w = Sw
z = A00 ()z + a01 () + P0 ()w
(4.68)
= A10 ()z + a11 () + b()u + P1 ()w
e = ,
in which it is assumed that b() > 0. As shown in Example 4.1, since by hypothesis
the eigenvalues of A00 () have negative real part, the Sylvester equation9
has a solution 0 () and therefore the regulator equations (4.6) have (unique) solu-
tion
0 () 1
() = , () = [A10 ()0 () + P1 ()].
0 b()
w = Sw
z = A00 ()z + a01 ()
= A10 ()z + a11 () + b()[u ()w].
Let now this system be controlled by a preprocessor of the form (4.53), with such
that the matrix F = G is Hurwitz and u provided by a stabilizer (4.54). As
shown in Lemma 4.8, there always exists a matrix () such that (see (4.59) in this
respect)
()S = ()
(4.70)
() = ().
Using these identities and changing into = ()w yields the following
(augmented) system
w = Sw
z = A00 ()z + a01 ()
(4.71)
= + Gu
= A10 ()z + a11 () + b()[ + u ],
9 Since the parameters of the equation are -dependent so is expected to be its solution.
116 4 Regulation and Tracking in Linear Systems
in which
z A00 () 0 a01 ()
x = , A() = 0 Gk . (4.72)
A10 () b() a11 () b()k
Since the eigenvalues of A00 () and those of F have negative real part, the converse
Lyapunov Theorem says that there exists a (n 1 + d) (n 1 + d) positive-definite
symmetric matrix Z() such that
T
A00 () 0 A00 () 0
Z() + Z() < 0.
b()
1
GA10 () F b()
1
GA10 () F
is such that
P()A() + A T ()P() < 0 (4.75)
4.7 The Case of SISO Systems 117
if k is large enough.10 From this, using the direct Theorem of Lyapunov, we conclude
that, if k is large enough, the matrix (4.74) has all eigenvalues with negative real part.
If this is the case, the lower subsystem of (4.71) is robustly stabilized and the problem
of output regulation is robustly solved.
The remarkable feature of the controller discussed in the previous section is the
ability of securing asymptotic decay of the regulated output e(t) in spite of parameter
uncertainties.
Thus, control schemes consisting of an internal model and of a robust stabilizer
efficiently address the problem of rejecting all exogenous inputs generated by the
exosystem. In this sense, they generalize the classical way in which integral-control-
based schemes cope with constant but unknown disturbances, even in the presence
of parameter uncertainties. There still is a limitation, though, in these schemes: the
necessity of a precise model of the exosystem. As a matter of fact, the controller
considered above contains a pair of matrices , whose construction (see above)
requires the knowledge of the precise values of the coefficients of the minimal poly-
nomial of S. This limitation is not sensed in a problem of set point control, where the
uncertain exogenous input is constant and thus obeys a trivial, parameter independent,
differential equation, but becomes immediately evident in the problem of rejecting,
e.g., a sinusoidal disturbance of unknown amplitude and phase. A robust controller is
able to cope with uncertainties on amplitude and phase of the exogenous sinusoidal
signal, but the frequency at which the internal model oscillates must exactly match
the frequency of the exogenous signal: any mismatch in such frequencies results in
a nonzero steady-state error.
In what follows we show how this limitation can be removed, by automatically
tuning the natural frequencies of the robust controller. For the sake of simplicity,
we limit ourselves to sketch here the main philosophy of the design method.11
Consider again the single-input single-output system (4.66) for which we have
learned how to design a robust regulator but suppose, now, that the model of the
exosystem that generates the disturbance w depends on a vector of uncertain para-
meters, ranging on a prescribed compact set Q, as in
w = S()w. (4.76)
We retain the assumption that the exosystem is neutrally stable, in which case S()
can only have eigenvalues on the imaginary axis (with simple multiplicity in the
denote the minimal polynomial of S() and assume that the coefficients s0 (), s1 (),
. . . , sd1 () are continuous functions of . Following the design procedure illustrated
in the previous sections, consider a pair of matrices , G defined as
0 1 0 0 0
0 0 1 0 0
=
,
G=
0 0 0 1 0
s0 () s1 () s2 () sd1 () 1
= + Gu
(4.77)
u = + u ,
with u provided by a robust stabilizer of the form (4.61), would solve the problem of
output regulation. We have also seen that such a stabilizer exists if is any matrix
that renders G a Hurwitz matrix and if, in addition, all n r zeros of the
triplet {A(), B(), C()} have negative real part for all .
Note that in this context the choice of is arbitrary, so long as the matrix G
is Hurwitz. In what follows, we choose this matrix as follows. Let F be a fixed d d
matrix
0 1 0 0
0 0 1 0
F = (4.78)
0 0 0 1
a0 a1 a2 ad1
p() = a0 + a1 + + ad1 d1 + d ,
having all roots with negative real part, let G be defined as above, i.e.,
T
G = 0 0 0 1 , (4.79)
4.8 Internal Model Adaptation 119
(note that we have added the subscript to stress the dependence of such on
the vector of possibly uncertain parameters). This choice is clearly such that
G = F
is a Hurwitz matrix. Hence, if were known, this choice would be admissible. With
this choice, the preprocessing internal model (4.77) can be rewritten as
= F + G( + u )
(4.81)
u = + u .
Essentially, what we have done is to shift the uncertain data from the matrix
to the vector . The realization (4.81) of the internal model, though, lends itself to
the implementation of some easy (and standard) adaptive control techniques.
If were known, the controller (4.81), with u provided by a robust stabilizer of
the form (4.61) would be a robust controller. In case is not known, one may wish
to replace the vector with an estimate , to be tuned by means of an appropriate
adaptation law.
We illustrate how this can be achieved in the simple situation in which the system
has relative degree 1. To facilitate the analysis, we assume that the controlled plant
has been initially put in normal form (see Example 4.1), which in the present case
will be12
w = S()w
z = A00 ()z + a01 () + P0 (, )w
(4.82)
= A10 ()z + a11 () + b()u + P1 (, )w
e = .
By assumption, the n 1 eigenvalues of the matrix A00 () have negative real part
for all .
Consider a tunable preprocessing internal model of the form
= F + G( + u )
(4.83)
u = + u
12 Itis seen from the construction in Example 4.1 that, in the normal form (4.11), the matrices P0
and P1 are found by means of transformations involving A, B, C, P and also S. Thus, if the former
are functions of and the latter is a function of , so are expected to be P0 and P1 .
120 4 Regulation and Tracking in Linear Systems
w = S()w
z = A00 ()z + a01 () + P0 (, )w
= A10 ()z + a11 () + b()[ + u ] + P1 (, )w
= F + G[ + u ]
e = .
w = S()w
z = A00 ()z + a01 () + P0 (, )w
= + Gu + G
= A10 ()z + a11 () + b()[ + u ] + b() + P1 (, )w
e = .
We know from the analysis in Example 4.2 that, if were zero, the choice of a
stabilizing control
u = k (4.84)
(with k > 0 and large) would solve the problem of robust output regulation. Let u
be chosen in this way and consider, for the resulting closed-loop system, a change
of coordinates (see again Example 4.2)
z = z 0 (, )w, = (, )w
in which 0 (, ) is a solution of
and (, ) satisfies
(, )S() = (, )
(, ) = (, ),
in which
1
(, ) = [A10 ()0 () + P1 (, )].
b()
w = S()w
z = A00 ()z + a01 ()
= Gk + G
= A10 ()z + a11 () + b()[ k ] + b()
4.8 Internal Model Adaptation 121
(note that we have not modified the terms for reasons that will become clear in
a moment).
The dynamics of w is now completely decoupled, so that we can concentrate on
the lower subsystem, that can be put in the form (compare with (4.72))
in which
z A00 () 0 a01 () 0
x = , A(, ) = 0 Gk , B() = G .
A10 () b() a11 () b()k b()
It is known from Example 4.2 thatsince the eigenvalues of A00 () and those of
F = G have negative real parta matrix such as A(, ) is Hurwitz,
provided that k is large enough. Specifically, let x be changed in
x = T ()x
x = A(,
)x + B() , (4.86)
in which A(, ) = T ()A(, )T 1 (), and
0
B() = T ()B() = 0 b().
1
It has been shown in Example 4.2 that there exists a positive-definite matrix
Z() 0
P() =
0 1
P()A(, ) + A T (, )P() < 0 (4.87)
if k > k .13
Consider now the positive-definite quadratic form
13 Recall, in this respect, that both the uncertain vectors and range on compact sets.
122 4 Regulation and Tracking in Linear Systems
and compute the derivative of this function along the trajectories of system (4.86).
Letting Q(, ) denote the negative definite matrix on the left-hand side of (4.87)
and observing that
x T P()B() = b(),
this yields
U = x T [P()A(, ) + A T (, )P()]x + 2x T P()B()
+ 2b() T
= x T Q(, )x + 2 b() + 2b() T
= x T Q(, )x + 2b() [ + T ].
T = T .
The function U cannot be made negative definite, but it can be made negative
semidefinite, by simply taking
T =
so as to obtain
x , ) = x T Q(, )x 0.
U(
x , ) = 0
U( x = 0.
w = S()w
x = A()x + B()u + P()w
e = C()x + Q()w.
Suppose the system has relative degree 1 and, without loss of generality, C()B() >
0. Suppose the n 1 zeros of the triplet {A(), B(), C()} have negative real part,
for every value of . Then, the problem of robust output regulation is solved by a
controller of the form
= F + G( ke)
u = ke
in which F, G are matrices of the form (4.78) and (4.79), k > 0 is a large number
and is provided by the adaptation law
T = e.
The extension to systems having higher relative degree, which relies upon argu-
ments similar to those presented in Sect. 2.5, is relatively straightforward, but it will
not be covered here.
Example 4.3 A classical control problem arising in the steel industry is the control
of the steel thickness in a rolling mill. As shown schematically in Fig. 4.6, a strip of
steel of thickness H goes in on one side, and a thinner strip of steel of thickness h
comes out on the other side. The exit thickness h is determined from the balance of
two forces: a force proportional to the difference between the incoming and outgoing
thicknesses
FH = W (H h)
and a force proportional to the gap between the rolls, that can be expressed as
Fs = M(h s)
Work roll
W
H h
Backup roll
124 4 Regulation and Tracking in Linear Systems
Rolls that are not perfectly round can be thought of as rolls of variable radius and
this radiusif the rolls rotate at constant speedis a periodically varying function
of time, the period being equal to the time needed for the roll to perform a complete
revolution. Thus, the term d that models the perturbation of the nominal gap h s is
a periodic function of time. The period of such function is not fixed though, because
it depends on the rotation speed of the rolls.
Balancing the two forces FH and Fsd yields
1
h= (Ms + WH + Md).
M +W
The purpose of the design is to control the thickness h. This, if href is the prescribed
reference value for h, yields a tracking error defined as
1
e= (Ms + WH + Md) href .
M +W
The unloaded screw position s is, in turn, proportional to the angular position of
the shaft of a servomotor which, neglecting friction and mechanical losses, can be
modeled as
s = bu
e = Cx + Qw
in which
C = c1 0 , Q = q1 q2 q3 0
and w R4 satisfies
4.8 Internal Model Adaptation 125
0 0 0 0
0 0 0 0
w =
0
:= S w.
0 0
0 0 0
w = S w
1 = 2
2 = c1 bu + QS2 w
e = 1 .
This system has relative degree 2. Thus, according to the method described in
Sect. 2.4, we define a new variable as
= 2 + a0 1 = e + a0 e (4.88)
in which a0 > 0, and set z = 1 to obtain a system which, viewed as a system with
input u and output , has relative degree 1 and one zero with negative real part
w = S w
z = a0 z +
= a02 z + a0 + c1 bu + QS2 w.
The theory developed above can be used to design a control law, driven by the
regulated variable , that will steer (t) to 0 as t . This suffices to steer also
the actual tracking error e(t) to zero. In fact, in view of (4.88), it is seen that e(t) is
the output of a stable one-dimensional linear system
e = a0 e +
() = 3 + 2 ,
which has a fixed root at = 0. Thus, it is natural to seek a setting in which only two
parameters are adapted (those that correspond to the uncertain roots in j). This
can be achieved in this way. Pick
0 H2 0
F= , G= , = 1 2,
G2 F2 G2
with a0 and a1 both positive (so that F2 is Hurwitz) and 2, such that
0 1
F2 + G2 2, = .
2 0
P2 F2 + F2T P2 = I
would solve the problem of output regulation. Since is not know, has to be
replaced by a vector of tunable parameters. Such vector, though, needs not to be a
(1 3) vector because the first component of , being equal to 1, is not uncertain.
Accordingly, in the above controller, this vector is replaced by a vector of the form
= 1 2
satisfies
0 0
QF + F T Q = 0,
0 I
and use LaSalles invariance principle. The proof of (ii) is achieved by direct substitution. Property
(iii) is a consequence of (i) and of (ii), which says that all eigenvalues of F + G have zero real
part. Property (iv) follows from Lemma 4.8.
4.8 Internal Model Adaptation 127
law is chosen for 2 , is identical to the one presented above, and will not repeated
here. We limit ourself to conclude with the complete model of the controller which,
in view of all of the above, reads as
1 = H2 2
2 = F2 2 + G2 (2 2 k(e + a0 e))
2T = 2 (e + a0 e)
u = 1 + 2 2 k(e + a0 e).
In Sect. 4.7, appealing to the results presented in Chap. 2, we have shown how robust
regulation can be achieved in the special case m = p = 1, under the assumption that
the triplet {A(), B(), C()} has a well-defined relative degree and all its n r
zeros have negative real part, for every value of . In this section, we discuss how
robust regulation can be achieved in a more general setting, appealing to the method
for robust stabilization presented in Sects. 3.5 and 3.6.16
For consistency with the notation used in the context of robust stabilization via
H methods, we denote the controlled plant as
w = Sw
x = Ax + B1 v + B2 u + Pw
z= C1 x + D11 v + D12 u + Q1 w (4.89)
y= C2 x + D21 v + Q2 w
e= Ce x + De1 v + Qe w,
in which
Ce = EC2 , De1 = ED21 , Qe = EQ2
with
E = Ip 0 .
According to the theory presented in Sects. 4.4 and 4.5, we consider a controller
that has the standard structure of a postprocessing internal model
= + Ge, (4.90)
in which , G have the form (4.23) and (4.24), cascaded with a robust stabilizer.
The purpose of such stabilizer is to solve the problem of -suboptimal H feed-
back design for an augmented plant defined as
16 Theapproach in this section essentially follows the approach of [6]. See also [7, 8] for further
reading.
128 4 Regulation and Tracking in Linear Systems
x A 0 x B1 B2
= + v+ u
GCe GDe1 0
x
z = C1 0 + D11 v + D12 u (4.91)
C2 0 x D21
ya = + v.
0 I 0
If this is the case, in fact, on the basis of the theory of robust stabilization via H
methods, one can claim that the problem of output regulation is solved, robustly with
respect to dynamic perturbations that can be expressed as
v = P(s)z,
in which P(s) is the transfer function of a stable uncertain system, satisfying
P
H <
1/ .
For convenience, let system (4.91) be rewritten as
x a = Aa xa + Ba1 v + Ba2 u
z = Ca1 x + Da11 v + Da12 u (4.92)
ya = Ca2 x + Da21 v,
in which
A 0 B1 B2
Aa = , Ba1 = , Ba2 = ,
GCe GDe1 0
Ca1 = C1 0 , Da11 = D11 , Da12 = D12 ,
C2 0 D21
Ca2 = , Da21 = .
0 I 0
Theorem 4.1 Consider a plant of modeled by equations of the form (4.92). Let
Va1 , Va2 , Za1 , Za2 be matrices such that
Z V T T
Im a1 = Ker Ca2 Da21 , Im a1 = Ker Ba2 Da12 .
Za2 Va2
The problem of -suboptimal H feedback design has a solution if and only if there
exist symmetric matrices Sa and Ra satisfying the following system of linear matrix
inequalities
4.9 Robust Regulation via H Methods 129
AT S + S A S B C T Z 0
T
Za1 T
Za2 0 a a T a a a a1 Ta1 a1
Ba1 Sa I Da11 Za2 0 < 0 (4.93)
0 0 I
Ca1 Da11 I 0 I
T T Aa Ra + Ra ATa Ra Ca1
T
Ba1 Va1 0
Va1 Va2 0
Ca1 Ra I Da11 Va2 0 < 0 (4.94)
0 0 I T
Ba1 T
Da11 I 0 I
Sa I
0. (4.95)
I Ra
In particular, there exists a solution of dimension k if and only if there exist Ra and
Sa satisfying (4.93)(4.95) and, in addition,
rank(I Ra Sa ) k. (4.96)
In view of the special structure of the matrices that characterize (4.92), the con-
ditions above can be somewhat simplified. Observe that the inequality (4.93) can be
rewritten as
T
Za1 Sa (Aa Za1 +Ba1 Za2 )+(Aa Za1 +Ba1 Za2 )T Sa Za1 Za2
TZ
a2
T C T +Z T DT
Za1 a1 a2 a11 < 0.
Ca1 Za1 +Da11 Za2 I
(4.97)
The kernel of the matrix
C2 0 D21
Ca2 Da21 =
0 I 0
Thus, if Sa is partitioned as
S
Sa = ,
130 4 Regulation and Tracking in Linear Systems
T
Moreover, Za2 Za2 = Z2T Z2 and
Corollary 4.1 Consider the problem of -suboptimal H feedback design for the
augmented plant (4.91). Suppose there exists positive-definite symmetric matrices S
and Ra satisfying the system of linear matrix equations (4.98), (4.99) and
4.9 Robust Regulation via H Methods 131
S In 0
In > 0. (4.100)
Ra
0
Y22 0 0 I
one obtains
0 0 0 0
S I 0 S I 0
TT a T =
0
.
I Ra I Ra11 Ra12
T
0 0 Ra12 Ra22
By (4.100), the matrix on the right is positive semidefinite and this implies (4.95).
Thus, (4.98)(4.100) altogether imply (4.93)(4.95).
Finally, observe that
S Y12 0
Ra Sa = Ra =
T
Y12 Y22 I
rank(I Sa Ra ) n,
132 4 Regulation and Tracking in Linear Systems
from which it is concluded that the problem can be solved by a controller dimension
not exceeding n.
Remark 4.4 It is worth stressing the fact that the dimension of the robust stabilizer
does not exceed the dimension n of the controlled plant, and this despite of the fact
that the robust stabilizer is designed for an augmented plant of dimension n + dp.
This is essentially due to the structure of the augmented plant and in particular to the
fact that the component of the state of such augmented plant is not affected by the
exogenous input v and is directly available for feedback.
= Ac + Bc ya
+ Js ya
u = Cc + Dc ya
+ Hs ya
in which ya
and ya
that solves the problem of output regulation for the perturbed plant (4.89), robustly
with respect to dynamic perturbations that can be expressed as v = P(s)z, in which
P(s) is the transfer function of a stable uncertain system, satisfying
P
H < 1/ .
References
1. B.A. Francis, The linear multivariable regulator problem. SIAM J. Control Optim. 14, 486505
(1976)
2. B. Francis, O.A. Sebakhy, W.M. Wonham, Synthesis of multivariable regulators: the internal
model principle. Appl. Math. Optim. 1, 6486 (1974)
3. E.J. Davison, The robust control of a servomechanism problem for linear time-invariant multi-
variable systems. IEEE Trans. Autom. Control 21, 2534 (1976)
4. B.A. Francis, M.W. Wonham, The internal model principle of control theory. Automatica 12,
457465 (1977)
5. A. Serrani, A. Isidori, L. Marconi, Semiglobal nonlinear output regulation with adaptive internal
model. IEEE Trans. Autom. Control, AC 46, 11781194 (2001)
6. H. Kroglu, C.W. Scherer, An LMI approach to H synthesis subject to almost asymptotic
regulation constraints. Syst. Control Lett. 57, 300308 (2008)
7. J. Abedor, K. Nagpal, P.P. Khargonekar, K. Poolla, Robust regulation in the presence of norm-
bounded uncertainty. IEEE Trans. Autom. Control 40, 147153 (1995)
8. H. Kroglu, C.W. Scherer, Scheduled control for robust attenuation of non-stationary sinusoidal
disturbances with measurable frequencies. Automatica 47, 504514 (2011)
Chapter 5
Coordination and Consensus of Linear
Systems
In this chapter, we study control problems that can be seen as extensions of the
problem of asymptotic tracking discussed in the previous chapter and, in fact, can be
addressed by means of appropriate enhancements of the design techniques presented
therein. One of such problems is the so-called leaderfollower coordination problem,
which can be cast in the following terms.
Suppose a set of linear systems is given, modeled by equations of the form
x k = Ak xk + Bk uk
(5.1)
yk = Ck xk
with state xk Rnk , control input uk Rm and controlled (and also measured) output
yk Rp , for k = 1, . . . , N.1 The number N of such systems is large. The problem
is to design a control law under which the output yk of each system asymptotically
tracks the output y0 of a single autonomous system
x 0 = A0 x0
(5.2)
y0 = C0 x0 .
In this context, system (5.2) is called the leader and each of the N systems (5.1)
is called a follower. Altogether, the N systems (5.1) and (5.2) are often called the
agents.
The theory presented in the previous section could provide an elementary approach
to the solution of such problem. If fact, it would suffice to define a tracking error
ek = yk y0 (5.3)
1 Note that all such systems have the same number of input and output components.
and to solve a set of N problems of output regulation for N systems, each one of
which is modeled by equations of the form
x 0 = A0 x0
x k = Ak xk + Bk uk
ek = Ck xk C0 x0 ,
where the role of exogenous input is now taken by the state of x0 of the leader. This
approach will end up with the design of a set of N controllers, each one of which is
modeled by equations of the form
vk = ek .
The problem with this approach, though, is that each controller is supposed to
be driven by the corresponding tracking error ek , i.e., each controller needs to have
access to the (reference) output of the leader. If the number N of agents is very large,
this entails an excessively large exchange of information, which could be prohibitive
if the followers are spatially distributed over a very large (geographical) region. The
challenge is to achieve the desired coordination (between the output of each follower
and the output of the leader) with a limited exchange of information, consisting only
in the differences between the output of each agent and those of a limited number of
other agents (one of which could be, possibly but not necessarily, the leader).
Such (more restricted) information pattern indeed makes the problem more chal-
lenging. In what follows we will consider the case in which the input vk to each local
controller is a fixed linear combination of the differences between the outputs of the
agents, namely a function of the form
N
vk = akj (yj yk ) k = 1, . . . , N (5.5)
j=0
in which the akj s are elements of a N (N +1) sparse matrix. Note that the coefficient
ak0 is the weight by means of which the difference (yk y0 ), that is the k-th tracking
error ek , is weighted in the linear combination vk , which characterizes the information
provided to the control of the kth follower. In this chapter, we address the case in
which only a few of such coefficients are nonzero, which means that only a few of
the local controllers have access to the output of the leader. A problem of this kind
is called a problem of leaderfollowers coordination.
If the problem in question is solved (and in the next sections we will discuss how
such problem can be solved), eventually all outputs of the followers asymptotically
5.1 Control of a Network of Systems 137
converge to a single function of time, the output of the leader. In this case, it is said
that the outputs of all agents, leader and followers altogether, achieve consensus. A
more general version of the problem is when a set of N agents modeled by equations
of the form (5.1) is given, no leader is specified, and a control law is sought, similar
to the one described above (that is consisting of a set of N local controllers of the
form (5.4) with vk a fixed linear combination of the differences between the outputs
of the agents) under which the outputs of all the agents achieve consensus, i.e.,
asymptotically converge to a single function of time. A problem of this kind is called
a problem of leaderless coordination or, more commonly, a consensus problem.2
As specified in the previous section, the local controllers (5.4) are driven by a fixed
linear combination of the differences between the outputs of the agents. This pattern
of communication between individual agents is conveniently described by means of
a few concepts borrowed from the theory of graphs. A communication graph3 is a
triplet G = {V , E , A} in which:
V is a set of N nodes (or vertices), V = {1 , 2 , . . . , N }, one for each of the N
agents in the set.
E V V is a set of edges that models the interconnection between nodes,
according to the following convention: (k , j ) belongs to E if there is a flow of
information from node j to node k .
A is a N N matrix whose entry akj represents a weight in the flow of information
from node j to node k . The matrix A is called the adjacency matrix of the graph
and its elements are nonnegative.
It is assumed that there are no self-loops, i.e., that (k , k ) / E or, what is the
same, akk = 0 for all k = 1, . . . , N. The graph is undirected if akj = ajk , for all (k, j)
(with k = j). Otherwise, the graph is said to be directed. In what follows, we will
consider the (more general) case of directed graphs. The set of neighbors of node k
is the set Nk = {j V : akj = 0}, i.e., the set of nodes from which there is a direct
flow of information to node k . A path from node j to node k is a sequence of r
distinct nodes {1 , . . . , r } with 1 = j and r = k such that (i+1 , i ) E (or,
what is the same, a sequence {1 , . . . , r } such that, for each i, i is a neighbor of
i+1 ).
2 Further motivations for the interest of an information pattern such as the one described by (5.5)
can be found in [16] and [20].
3 Background material on graphs can be found, e.g., in [7, 8]. All objects defined below are assumed
1 2
3 4
Definition 5.1 A graph G is connected if there is a node such that, for any other
node k V \ {}, there is a path from to k .
In other words, a graph is connected if there is a node from which information
can propagate to all other nodes along paths. This node is sometimes called a root.4
In the leaderless coordination problem, we will consider the case in which the
information available for control purpose at the kth agent has the form
N
vk = akj (yj yk ) k = 1, . . . , N (5.6)
j=1
viewing vk as a weighted sum of the relative values of the measurements taken from
agent k and its neighbors.
Letting L denote the matrix defined by
kj = akj for k = j
kk = Ni=1 aki
The matrix L is called the Laplacian matrix of the graph. Observe that, by definition,
the diagonal entries of L are nonnegative, the off-diagonal elements are nonpositive
and that, for each row, the sum of all elements on this row is zero. Letting 1N denote
the all-ones N-vector
4 Note
that in a connected graph there may be more than one node with such property. See the
example in Fig. 5.1.
5.2 Communication Graphs 139
1N = col(1, 1, . . . , 1),
L 1N = 0.
It is seen from the above that the matrix L is singular or, what is the same, that
= 0 is always an eigenvalue of L. Such eigenvalue is referred to as the trivial
eigenvalue of L. Let the other (possibly nonzero) N 1 eigenvalues of L be denoted
as 2 (L), . . . , N (L). The real parts of these eigenvalues play an important role with
respect to the property of connectivity.
Theorem 5.1 A graph is connected if and only if its Laplacian matrix L has only one
trivial eigenvalue 1 = 0 and all other eigenvalues 2 (L), . . . , N (L) have positive
real parts.
N
N
vk = akj [yj y0 (yk y0 )] = akj (ej ek ) k = 1, . . . , N,
j=0 j=0
and this, in view of the definition of the Laplacian matrix L (associated with the
(N + 1) (N + 1) adjacency matrix that characterizes the expression (5.5)) and of
the fact that e0 = 0, can be rewritten as (compare with (5.7))
N
vk = kj ej k = 1, . . . , N. (5.8)
j=1
In what follows, consistently with a basic assumption in the theory of output reg-
ulation, it is assumed that the matrix A0 which characterizes the leader (5.2) has a
purely diagonal Jordan form and all eigenvalues with zero real part. As far as the
140 5 Coordination and Consensus of Linear Systems
For the purpose of using the theory of output regulation, we check first the exis-
tence of a solution of the Francis equations. A standard calculation shows that if
k,0 is a solution of the equation
(solution which exists because the spectra of A0 and Ak,00 are disjoint), the solution
of Francis equation is given by
k,0 1
k = , k = [Ak,10 k,0 + Ak,11 C0 C0 A0 ].
0 bk
z k = Ak,00 zk + Ak,01 ek
(5.11)
e k = Ak,10 zk + Ak,11 ek + bk [uk k x0 ].
With the results discussed in Sect. 4.7 in mind, we consider for each agent a
preprocessing internal model of the form
k = k + Guk
(5.12)
uk = k + u k
in which and G are as in (4.23) and (4.24), where we take p = 1 and pick
s0 , s1 , . . . , sd1 as the coefficients of the minimal polynomial of A0 , and where is
such that G is a Hurwitz matrix. As shown in Sect. 4.6 there always exists a
matrix k such that (see (4.59) in this respect)
5.3 LeaderFollower Coordination 141
k A0 = k
k = k .
z k = Ak,00 zk + Ak,01 ek
k = k + Guk (5.13)
e k = Ak,10 zk + Ak,11 ek + bk [ k + u k ],
and obtain, after simple manipulations, a system having the following structure
k = Fk,00 k + Fk,01 ek
(5.14)
e k = Fk,10 k + Fk,11 ek + bk u k ,
in which
Ak,00 0
Fk,00 = .
b1k GAk,10 G
u k = gek
with a sufficiently large value of g, the state of (5.14) would asymptotically converge
to 0. In particular ek would converge to 0 and this would imply that the desired
tracking goal is achieved. This control mode, though, in general is not feasible,
as anticipated in Sect. 5.1, because this control mode requires that, for each k =
1, . . . , N, the controller of agent k has access to the tracking error ek . This would entail
an excessively large exchange of information between the leader and the followers.
Fortunately, the desired control goal can still be achieved if u k is taken as (compare
with (5.5))
N
u k = g akj (yj yk ) k = 1, . . . , N (5.15)
j=0
in which g is a suitable gain parameter and the akj are the entries of the adjacency
matrix A of a connected graph modeling the pattern of communication between leader
142 5 Coordination and Consensus of Linear Systems
and followers. To see why this is the case, let us stack all systems (5.14) together, to
obtain a system modeled as
= F00 + F01 e
(5.16)
e = F10 + F11 e + Bu
in which
= col(1 , . . . , N )
e = col(e1 , . . . , eN )
u = col(u1 , . . . , u N )
and
F1,00 0 0 F1,01 0 0
0 F2,00 0 0 F2,01 0
F00 =
F01 =
0 0 FN,00 0 0 FN,01
F1,10 0 0 F1,11 0 0
0 F2,10 0 0 F2,11 0
F10 =
F11 =
0 0 FN,10 0 0 FN,11
b1 0 0
0 b2 0
B=
.
0 0 bN
Note that the vector u , whose entries are defined in (5.15), can be expressed as
(see (5.8) in this respect)
u = gL22 e
in which the matrix BL22 is fixed and g is still a free design parameter.
We discuss now the properties of the matrix BL22 . Bearing in mind the fact that
B is a purely diagonal matrix with positive entries, consider the (N + 1) (N + 1)
matrix A in which the element a kj is defined as
a 0j = 0 j = 0, 1, . . . , N
a kj = bk akj , j = 0, 1, . . . , N k = 1, . . . , N.
5.3 LeaderFollower Coordination 143
Clearly, the matrix A has been obtained from the adjacency matrix A, associated with
communication graph G that characterizes the communication protocol between
leader and followers, by multiplying all elements on the kth row (for k = 1, . . . , N)
of A by the positive coefficient bk . The matrix A can be regarded as an adjacency
matrix of a communication graph, that will be denoted as G. Since all coefficients
bk are nonzero, it is readily seen that if the graph G is connected, so is the graph
G.5 Bearing in mind the definition of the Laplacian matrix of a graph, observe that
the matrix BL22 coincides with the lower-right (N N) submatrix of the Laplacian
matrix L of the graph G. By virtue of Theorem 5.1, if the graph G is connected, the
matrix L has one trivial eigenvalue at 0 while all other eigenvalues have positive real
part. But, in the present case, the first row of L has all zero entries and hence the
eigenvalues of BL22 coincide with the N nontrivial eigenvalues of L. In this way, it
is concluded that, if the graph G is connected, the matrix BL22 has all eigenvalues
with positive real part.
Return now to the system (5.17) in which we know that, by assumption (and also
by construction), all eigenvalues of F00 have negative real part and that, if the graph
G is connected, the negative of the matrix BL22 has all eigenvalues with negative
real part as well. Let P0 be a positive definite symmetric solution of the Lyapunov
equation
P0 F00 + F00
T
P0 = I
P1 (BL22 ) + (BL22 )T P1 = I.
Set
P0 0
P=
0 P1
5 Infact, if the graph G possesses a root from which information can propagate to all other nodes
along paths, the same is true for the graph G, because the nonzero entries of A coincide with the
nonzero entries of A.
144 5 Coordination and Consensus of Linear Systems
gI + P1 F11 + F11
T
P1 + [P1 F10 + F01
T
P0 ][P0 F01 + F10
T
P1 ] ,
for all k = 1, . . . , N.
The construction described above can easily be extended to the case in which
the agents have a relative degree higher than 1. Without loss of generality, we can
assume that all agents in (5.1) have the same relative degree r. In fact, it is always
possible to achieve such property by adding a suitable number of integrators on the
input channel of each agent. This being the case, the model of each agent can be
expressed in (strict) normal form as (compare with (2.9))
zk = Ak,00 zk + Ak,01 k
k + B[A
k = A k,10 zk + Ak,11 k + bk uk ] (5.18)
yk = Ck
in which
k = col(k1 , k2 , . . . , kr ).
can be seen as a system having relative degree 1 between input uk and output k ,
modeled in normal form as (compare with (2.22))
5.3 LeaderFollower Coordination 145
k = Fk,00 k + Fk,10 k
(5.20)
k = Fk,10 k + Fk,11 k + bk uk .
Moreover, if the coefficients a0 , a1 , . . . , ar2 are such that the polynomial (2.23) has
all roots in C , all eigenvalues of Fk,00 are in C .
Recall that, by definition, ki (t) coincides with the derivative of order (i 1) of
yk (t) with respect to time (for i = 1, . . . , r). Thus, the function k (t), in terms of the
original output yk (t), can be expressed as
For consistency, also the output y0 of the leader (5.2) is replaced by variable having
a similar structure, namely
which yields
The systems defined in this way satisfy the assumptions of Proposition 5.1. Thus,
it can be claimed that if each agent is controlled by a controller of the form (5.12),
with parameters chosen as specified in the proposition, with u k given by
N
u k = g akj (j k ) k = 1, . . . , N, (5.21)
j=0
and the communication graph is connected, there is g such that, if g > g , the
(modified) outputs of the followers achieve consensus with the (modified) output of
the leader, namely
lim (k (t) 0 (t)) = 0
t
k (t) = e(r1)
k (t) + a0 ek (t) + a1 e(1) (r2)
k (t) + ar2 ek (t), (5.22)
146 5 Coordination and Consensus of Linear Systems
Since the a0 , a1 , . . . , ar2 are such that the polynomial (2.23) has all roots in C ,
this is a stable system, driven by an input k (t) that asymptotically decays to 0. Thus
e k (t) asymptotically decays to zero and so does the tracking error ek (t).
In conclusion, if each agent is controlled by a controller of the form (5.12), with
parameters chosen as specified in the Proposition 5.1, with u k given by (5.21), and
the communication graph is connected, for large enough values of g the tracking
errors (5.3) asymptotically decay to zero, i.e., outputs of the followers (5.1) achieve
consensus with the output of the leader (5.2).
The communication protocol (5.21) presumes the availability of the variables k s.
If this is not the case, one could replace them by quantities of the form
k = kr + a0 k1 + a1 k2 + ar2 k,r1 ,
in which the ci s are such that the polynomial p() = r1 +cr1 r2 + +c2 +c0
has all eigenvalues in C and is a design parameter. Arguments similar to those
used in Sect. 2.5, which are not repeated here, can be used to show that if is large
enough the desired consensus goal is achieved.
5.4 Consensus in a Homogeneous Network: Preliminaries 147
In this and in the following sections, we will consider the problem of achieving
consensus among the outputs of a set of N agents, modeled as in (5.1).6 The input
uk to each of such systems is provided by a local controller modeled as in (5.4), in
which vk represents (relative) information received from other agents, and has the
form (5.6), consisting in a weighted linear combination of the relative values of the
outputs of the neighboring agents. The weights akj are the entries of the adjacency
matrix A that characterizes the communication pattern among agents. It is said that
the set of (controlled) agents achieve consensus, if for some function of time, that
will be denoted by ycons (t), it occurs that
for all k = 1, . . . , N.
There is no leader in this setting, thus the consensus output ycons (t) is not the
output of a separate autonomous system. Note, in this respect, that if all outputs of
the (controlled) agents coincide, vk = 0, and hence the control uk generated by the
controller (5.4) is the output of the autonomous system
As a consequence, the output yk (t) of the kth agent is necessarily the output of the
autonomous system
x k = Ak xk + Bk Cc,k xc,k
x c,k = Ac,k xc,k
yk = Ck xk .
In this and in the following section we will study the simpler case in which all
agents are identical, i.e.,
x k = Axk + Buk
for all k = 1, . . . , N (5.24)
yk = Cxk ,
in which A Rnn , B Rnm , C Rpn are fixed matrices. In this case the network
of controlled agents is said to be homogeneous. In the last section of the chapter we
will discuss the general case in which the agents may be different, which is called
the case of a heterogeneous network.
We will consider two modes of control. A static control mode, in which the control
uk is given by
uk = Kvk (5.25)
with K Rmp a matrix of feedback gains to be found, and a dynamic control mode,
in which the control uk is the output of a dynamical system
k = Ac k + Bc vk
(5.26)
uk = Cc k
while the second mode of control results in a closed-loop system modeled by a set
of equations of the form
0
xj
N
x k A BCc xk
= kj C0 i = 1, . . . , N. (5.28)
k 0 Ac k Bc j
j=1
Both these sets have the same structure, that of a set of systems of the form
N
x k = F x k kj GH x j i = 1, . . . , N. (5.29)
j=1
x k = xk , F = A, G = BK, H = C,
x = col(x1 , x 2 , . . . , x N )
and using the Kroeneker product of matrices,7 the set (5.29) of systems can be
expressed as
x = (IN F)x (L GH)x. (5.30)
x = (T 1 In )[(IN F) (L GH)](T In )x
= (T 1 In )[(T F) (LT GH)]x
= [(IN F) (L GH)]x .
x N x 1
(Footnote 7 continued)
a11 B a12 B a1n B
a21 B a22 B a2n B
AB=
.
am1 B am2 B amn B
8 We use in what follows the property (AB)(C D) = (AC BD), whichin particularimplies
(T 1 In )1 = (T In ).
150 5 Coordination and Consensus of Linear Systems
x 1
x =
x
in which
x 2 x 1
x = .
x N x 1
Then, it is easily checked that the system thus obtained has a block-triangular struc-
ture, of the following form
x 1 = F x 1 (L 12 GH)x
(5.33)
x = [(I(N1) F) (L 22 GH)]x .
and choose R in such a way that is a triangular matrix.10 If this is the case, the entries
on the diagonal of coincide with the nontrivial eigenvalues 2 (L), . . . , N (L) of
L. The similarity transformation R In changes the matrix (5.34) in
(I(N1) F) ( GH).
Since the matrix is diagonal, the matrix in question is block-diagonal, the diagonal
blocks being n n matrices of the form
F i (L)GH, i = 2, . . . , N. (5.35)
Thus, it can be claimed that the eigenvalues of the matrix (5.34) have negative real
part if and only if, for each i = 2, . . . , N, the eigenvalues of the matrix (5.35) have
negative real part. This yields the conclusion summarized as follows.
Proposition 5.2 Consider the controlled system (5.30). Suppose that for each i =
2, . . . , N, the eigenvalues of the matrix (5.35) have negative real part. Then, the
states x k (t) of (5.29) achieve consensus.
In the next section we will discuss how the property indicated in this proposition
can be fulfilled, separately for the static and for the dynamic control mode. Note also
that, in both cases, the output yk of the kth agent (5.24) can be expressed as a linear
function of the state x k of (5.29). In fact, in both cases
yk = H x k .
Thus, if the conditions of Proposition 5.2 hold, the outputs yk of (5.24) achieve
consensus, as desired.
One may wonder how the consensus trajectory looks like. To this end, it is con-
venient to recall the following basic result.
z1 = A11 z1 + A12 z2
(5.36)
z2 = A22 z2 .
with z1 Rn1 and z2 Rn2 . Suppose that all the eigenvalues of A22 have negative
real part. Then, there exists numbers M > 0, > 0 and a matrix R Rn1 (n1 +n2 )
such that, for all (z1 (0), z2 (0)) Rn1 Rn2
z1 (t) eA11 t z
Met
z2 (0)
,
in which
z1 (0)
z = R .
z2 (0)
This lemma can be used to estimate the asymptotic behavior of the response x 1 (t)
in system (5.33). In fact, because of this lemma, it is easily seen that, if all the
eigenvalues of the matrix (5.34) have negative real part, there exist numbers M > 0,
> 0 and a matrix R Rn N n such that, for all x(0), the state x 1 (t) satisfies
x1 (t) e Ft x
Met
x (0)
.
In other words, if all the eigenvalues of the matrix (5.34) have negative real part,
the states of the individual subsystems of (5.30) reach consensus along a particular
solution of
x = F x ,
the solution resulting from an initial state x which is a (fixed) linear function of the
initial states x 1 (0), . . . , x N (0) of the individual subsystems.
In the original consensus problem in static control mode, x k = xk and the matrix
F coincides with the matrix A that describes the dynamics of the agents. Thus, if the
hypotheses of the previous proposition are fulfilled, we conclude that
for some x of the form x = Rx(0), with x(0) = col(x1 (0), . . . , xN (0)). In particular,
the outputs of the individual systems reach consensus along the function
If the hypotheses of the previous proposition are fulfilled, the states xk (t) of the
individual subsystems reach consensus along a particular solution of
x k A BCc xk
= .
k 0 Ac k
If, in addition, the matrix Ac is a Hurwitz matrix, the upper component xk (t) of this
solution satisfies (again, use Lemma 5.2)
xk (t) eAt xk
Ket
k (0)
,
for some x . Hence, it can be concluded that the states xk of the individual agents
still satisfy a inequality of the form
for some x of the form x = Rx(0), with x(0) = col(x1 (0), . . . , x N (0)) and x k (0) =
col(xk (0), k (0)). Once again, the outputs of the individual agents reach consensus
along a function
ycons (t) = CeAt x .
5.4 Consensus in a Homogeneous Network: Preliminaries 153
Remark 5.1 Note that the consensus output is determined by the set of initial states
of all the individual agents (and controllers). In fact, the state x is determined by
all such initial states. In other words, there is not a unique consensus trajectory, but
rather one consensus trajectory for each choice of initial states.
In the static control mode, we consider the special case of agents (5.24) in which
n = m and having B = I (or, equivalently, having a nonsingular B). In this case the
matrix (5.35) reduces to a matrix of the form
A i (L)KC, i = 2, . . . , N.
Set
K = PC T .
Ai = A i (L)PC T C
has all the eigenvalues with negative real part, it suffices to check that12
and hence it is concluded that the choice K = PC T solves the problem. For conve-
nience, we summarize this result a follows.13
11 The existence of such P > 0 is guaranteed by the assumption that the pair (A, C) is observable.
Proposition 5.3 Consider a set of agents of the form (5.24), with B = I. Suppose
the pair (A, C) is observable. Suppose the communication graph is connected. Let
the control be given by
N
uk = PC T akj (yi yk ) k = 1, . . . , N,
j=1
with P the unique positive definite symmetric solution of the algebraic Riccati equa-
tion (5.37). Then, output consensus is reached.
In the dynamic control mode, the matrix (5.35) reduces to a matrix of the form
A BCc
i (L)Bc C Ac
To place the eigenvalues of this matrix in C , the idea is to choose for Ac , Bc the
structure of an observer-based stabilizer, with
Ac = A + GC BCc , Bc = G,
and G such that A + GC has all eigenvalues in C (which is possible under the
assumption that (A, C) is observable).
For the design of the matrix Cc , consider the the family if algebraic Riccati equa-
tions
P A + AT P P BBT P + I = 0 (5.39)
in which > 0.14 It is known15 that if the matrix A has eigenvalues with non-positive
real parts and the pair (A, B) is reachable, this equation has a unique positive definite
symmetric solution P . The matrix P depends continuously on and lim0 P = 0.
This being the case, set Cc = BT P .
In summary, the triplet {Ac , Bc , Cc } is chosen as
Ac = A + GC BBT P
Bc = G
Cc = BT P
with sufficiently small . Note that, since lim0 P = 0, for small the eigenvalues
of Ac have negative real part, as desired. In what follows, it is shown that a controller
defined in this way is capable of solving the problem of output consensus for the
homogeneous network of systems (5.24)
Proposition 5.4 Consider a set of agents of the form (5.24). Suppose that A has
eigenvalues with non-positive real parts, that the pair (A, B) is reachable and the
pair (A, C) is observable. Suppose the communication graph is connected. Let the
control be given by
k = (A + GC BBT P )k G N
a (y y )
j=1 kj j k
k = 1, . . . , N,
uk = (BT P )k
with G such that (A + GC) C and P the unique positive definite symmetric
solution of the algebraic Riccati equation (5.39). Then, there is a number such
that, if 0 < < , output consensus is reached.
and pick
P 0
P=
0 cQ
2
i x P BBT P i P BBT P x x P BBT P x + |i |2 P BBT P
2
and
c[(i 1) QBBT P x + (i 1)x P BBT Q ]
2 x P BBT P x + 2c2
|i 1|2 QBBT Q.
x (P Ai + Ai P )x
x
2 c
2 + 2 |i |2
BBT
P
2
2
2c2
+ |i 1|2
QB
2
2 + 2c|i 1|2
QBBT
2 .
Set now
2 2
= (max |i |2 )
BBT
, = (max |i 1|2 )
QB
2 ,
i i
= 2(max |i 1|2 )
QBBT
.
i
Then,
x (PAi + Ai P)x
x
2 (c
P
2 c2 c
P
)
2 .
Picking c =
P
:= c() and setting
() = c() ( + + )c2 ()
yields
x (PAi + Ai P)x
x
2 ()
2 .
and, as a consequence, all eigenvalues of (5.38) have negative real part. This, in
view of the results discussed in the earlier section, shows that the proposed control
structure solves the consensus problem.
The analysis carried out in the previous sections can be easily extended, so as to
handle also the problem of consensus in a heterogeneous network. The (simple) idea
is to use the result of Proposition 5.3 to induce consensus among the outputs in a set
of N identical reference generators, modeledfor k = 1, . . . , Nas
w k = Swk + Kvk
(5.40)
k = Qwk ,
N
vk = akj (j k ), (5.41)
j=1
and then to design a (local) regulator, so as to force the output yk (t) of the kth agent
in the set (5.1) to asymptotically track the output k (t) of the corresponding local
reference generator (5.40).17
This yields a two-step design procedure. The first step can be simply completed
by exploiting the result of Proposition 5.3, as shown below in more detail. The second
step consists in considering, for each of the (nonidentical) agents (5.1), a tracking
error ek defined as
ek = yk k
17 The approach described in this section is motivated by the works of [18] and [19]. In particular,
the work [18] shows that the approach outline above is in some sense necessary for the solution of
the problem in question.
158 5 Coordination and Consensus of Linear Systems
k = Fk k + Gk ek
uk = Hk k
x k = Ak xk + Bk uk
(5.42)
ek = Ck xk Qwk .
SP + PS T 2PQT QP + aI = 0. (5.43)
Then, it follows from Proposition 5.3 that, if the matrix K in (5.40) is chosen as
K = PQT , the outputs of the set (5.40) of local reference generators controlled by
(5.41) reach consensus. In particular, as shown at the end of Sect. 5.4, the consensus
is reached along a function of the form
for all k = 1, . . . , N.
Then, we proceed with the second step of the design. In the analysis of the (local)
regulation problem thus defined, a little extra care is need, to correctly characterize
the exosystem (which, as in any regulation problem, is supposed to generate the
exogenous input: in this case, the input wk in (5.42)). In fact, the various wk s are not
independent of each other. Thus, the correct approach to the problem is to consider
as exosystemthe entire set local reference generators (5.40) controlled by (5.41).
Setting
w = col(w1 , w2 , . . . , wN ),
w = [(IN S) (L KQ)]w.
It is known from the analysis of Sect. 5.4 that the system thus defined can be
transformed, by similarity, into a system in block-triangular form. More precisely,
changing w into (T 1 In0 )w, with T defined as in (5.32), the system can be changed
into a system of the form
w 1 = Sw1 (L 12 KQ)w
(5.45)
w = [(I(N1) S) (L 22 KQ)]w ,
where
w2 w1
w = .
wN w1
Moreover, thanks to the results of the analysis in Sect. 5.5, it is known that if K is
taken as indicated above, i.e., K = PQT with P solution of (5.43), the matrix
[(I(N1) S) (L 22 KQ)]
w 1 = Sw1 + Rw
w = S w .
160 5 Coordination and Consensus of Linear Systems
wk = w1 + Hk w
The design of the regulator can be accomplished along lines similar (but not quite
identical) to those followed in the construction described in Sect. 4.6. Specifically,
consider the case in which m = p and assume thatfor each kthere is a solution
(k , k ) of the Francis equations
k S = Ak k + Bk k
(5.47)
0 = Ck k Q.
Let and G be as in (4.23) and (4.24), where we pick s0 , s1 , . . . , sd1 as the coeffi-
cients of the minimal polynomial of S, and where is such that G is a Hurwitz
matrix. Then, it is known from Sect. 4.6 that there exists a matrix k satisfying
k S = k
(5.48)
k = k .
k = As,k k + Bs,k ek
k = k + G(Cs,k k + Ds,k ek ) (5.49)
uk = k + (Cs,k k + Ds,k ek ).
To check that such controller is able to achieve the desired goal, consider the
interconnection of (5.46) and (5.49), and change the variables as
x k = xk k w1
k = k k w1 .
5.6 Consensus in a Heterogeneous Network 161
This, using (5.47) and (5.48), after a few standard manipulation yields
w 1 = Sw1 + Rw
w = S w
x k = (Ak + Bk Ds,k Ck )xk + Bk k + Bk Cs,k k (Bk Ds,k QHk + k R)w (5.50)
k = k + GCs,k k + GDs,k Ck x k (GDs,k QHk + k R)w
k = As,k k + Bs,k Ck x k Bs,k QHk w
and
ek = Ck x k QHk w .
Note that the four bottom equations of (5.50) are not affected by w1 . They can be
organized as a block-triangular system of the form
cw ccccS 0 0 0 cw
x k Ak + Bk Ds,k Ck Bk Bk Cs,k x k
.
k = GDs,k Ck GCs,k k
k Bs,k Ck 0 As,k k
It is known from the analysis in Sect. 4.6 that, if (Ak , Bk ) is stabilizable, if (Ak , Ck )
is detectable, if the nonresonance condition
Ak I Bk
det = 0 (S) (5.51)
Ck 0
has all eigenvalues in C . Let this be the case. The matrix S , on the other hand,
has all eigenvalues in C because of the choice K = PQT in (5.40). Hence, it is
concluded that the system abovewhich is a block-triangular systemis stable. In
particular, both x k (t) and w (t) tend to 0 as t , and this yields
As a consequence
lim [yk (t) k (t)] = 0.
t
18 Note that the latter guarantees the existence a (unique) the solution pair of (5.47).
162 5 Coordination and Consensus of Linear Systems
This, since all k (t)s achieve consensus, shows that under the proposed control
scheme also all yk (t)s achieve consensus, in particular along a function of the form
(5.44).
To summarize, we conclude that the consensus problem for a heterogeneous net-
work can be solved by means of a controller having the following structure
N
w k = Swk + PQT j=1 akj (j k )
k = Qwk
k = k + G(Cs,k k + Ds,k [yk k ])
k = As,k k + Bs,k [yk k ]
uk = k + (Cs,k k + Ds,k [yk k ]).
The assumptions under which the parameters of this controller can be chosen so
as to induce output consensus are that the communication graph is connected, the
pair (S, Q) is observable, the pairs (Ak , Bk ) are stabilizable, the pairs (Ak , Ck ) are
detectable and the nonresonance conditions (5.51) hold. Note also that an exten-
sion that covers the cases in which the models of the agents (5.1) are affected by
uncertainties is possible, appealing to the methods for the design of robust regulators
discussed in Sects. 4.7 and 4.9. Details are left to the reader.
References
x = f (x) + g(x)u
(6.1)
y = h(x)
1 By smooth map we mean a C map, i.e., a map for which partial derivatives of any order are
defined and continuous. In what follows, f (x) and g(x) will be sometimes regarded as smooth vector
fields of Rn .
2 In fact, the right-hand side of the upper equation for each x is an affine function of u.
3 Some of the topics presented in this chapter are covered in various textbooks on nonlinear systems,
such as [16]. The approach here follows that of [2]. For further reading, see also [1416].
dx dx
= = f (x(t)) + g(x(t))u(t) .
dt x dt x
where
f (x ) = f (x) g (x ) =
g(x) x ) = [h(x)]x= 1 (x) .
h(
x x= 1 (x ) x x= 1 (x )
Sometimes, a transformation possessing properties (i) and (ii) and defined for all
x is difficult to find. Thus, in some cases one rather looks at transformations defined
only in a neighborhood of a given point. A transformation of this type is called a local
diffeomorphism. A sufficient condition for a map () to be a local diffeomorphism
is described below.
Lemma 6.1 Suppose () is a smooth map defined on some subset U of Rn . Suppose
the jacobian matrix of is nonsingular at a point x = x U. Then, on a suitable
neighborhood U U of x , () defines a local diffeomorphism.
Having understood how coordinates are transformed, we proceed now with the
definition of an integer that plays a role identical to that of the relative degree defined
in Sect. 2.1 for a linear system. The nonlinear system (6.1) is said to have relative
degree r at a point x if5 :
(i) Lg Lfk h(x) = 0 for all x in a neighborhood of x and all k < r 1
(ii) Lg Lfr1 h(x ) = 0.
4 The first of the two properties is clearly needed in order to have the possibility of reversing
the transformation and recovering the original state vector as x = 1 (x ), while the second one
guarantees that the description of the system in the new coordinates is still a smooth one.
5 Let be real-valued function and f a vector field, both defined on a subset U of Rn . The function
n
Lf (x) = fi (x) := f (x).
xi x
i=1
6.1 Relative Degree and Local Normal Forms 169
Note that the concept thus introduced is a local concept, namely r may depend
on the specific point x where the functions Lg Lfk h(x) are evaluated. The value of r
may be different at different points of Rn and there may be points where a relative
degree cannot be defined. This occurs when the first function of the sequence
which is not identically zero (in a neighborhood of x ) is zero exactly at the point
x = x . However, since f (x), g(x), h(x) are smooth, the set of points where a relative
degree can be defined is an open and dense subset of Rn .
We have seen in Sect. 2.1 that the relative degree of a linear system can be inter-
preted as a simple property of the transfer function, namely the difference between
the degrees of the denominator and the numerator polynomials of such function.
However, an alternative interpretation is also possible, that does not appeal to the
notion of transfer function and also holds for a nonlinear system of the form (6.1).
Assume the system at time t = 0 is in the state x(0) = x and let us calculate the value
of the output y(t) and of its derivatives with respect to time y(k) (t), for k = 1, 2, . . .,
at t = 0. We obtain
y(0) = h(x(0)) = h(x )
and
h dx h
y(1) (t) = = [f (x(t)) + g(x(t))u(t)] = Lf h(x(t)) + Lg h(x(t))u(t).
x dt x
At time t = 0,
y(1) (0) = Lf h(x ) + Lg h(x )u(0),
from which it is seen that, if r = 1, the value y(1) (0) is an affine function of u(0).
Otherwise, suppose r is larger than 1. If |t| is small, x(t) remains in a neighborhood
of x and hence Lg h(x(t)) = 0 for all such t. As a consequence
(Footnote 5 continued)
This function is sometimes called the derivative of along f . If g is another vector field, the notation
Lg Lf (x) stands for the derivative of the real-valued function Lf along g and the notation Lfk (x)
stands for the derivative of the real-valued function Lfk1 along f .
170 6 Stabilization of Nonlinear Systems via State Feedback
This yields
Lf h dx Lf h
y(2) (t) = = [f (x(t)) + g(x(t))u(t)] = Lf2 h(x(t)) + Lg Lf h(x(t))u(t).
x dt x
At time t = 0,
y(2) (0) = Lf2 h(x ) + Lg Lf h(x )u(0),
from which it is seen that, if r = 2, the value y(2) (0) is an affine function of u(0).
Otherwise, if r larger than 2, for all t near t = 0 we have Lg Lf h(x(t)) = 0 and
y(k) (t) = Lfk h(x(t)) for all k < r and all t near t = 0
y(r) (0) = Lfr h(x ) + Lg Lfr1 h(x )u(0).
Thus, the integer r is exactly equal to the number of times one has to differentiate
the output y(t) at time t = 0 in order to have the value u(0) of the input explicitly
appearing.
The calculations above suggest that the functions h(x), Lf h(x), . . . , Lfr1 h(x) must
have a special importance. As a matter of fact, such functions can be used in order
to define, at least partially, a local coordinates transformation around x . This fact is
based on the following property, which extends Proposition 2.1 to the case of systems
of the form (6.1).6
Proposition 6.1 The differentials7 of the r functions h(x), Lf h(x), . . . , Lfr1 h(x) are
linearly independent at x = x .
Proposition 6.1 shows that necessarily r n and that the r functions h(x),
Lf h(x), . . . , Lfr1 h(x) qualify as a partial set of new coordinate functions around
the point x . If r = n these functions define a full change of coordinates. Otherwise,
if r < n the change of coordinates can be completed by picking n r additional
functions, yielding a full change of coordinates that can be seen as equivalent of the
change of coordinates introduced in Proposition 2.2.
Proposition 6.2 Suppose the system has relative degree r at x . Then r n.
If r is strictly less than n, it is always possible to find n r more functions
1 (x), . . . , nr (x) such that the mapping
row vector
d(x) = := .
x1 x2 xn x
6.1 Relative Degree and Local Normal Forms 171
1 (x)
...
nr (x)
(x) =
h(x)
Lf h(x)
...
Lfr1 h(x)
The description of the system in the new coordinates is found very easily. Set
z1 1 (x) 1 h(x)
z2 2 (x) 2 Lf h(x)
z=
= , = =
and
x = col(z1 , . . . , znr , 1 , . . . , r ) := (x).
d1 h dx
= = Lf h(x(t)) = 2 (t)
dt x d t
dr1 (Lfr2 h) dx
= = Lfr1 h(x(t)) = r (t).
dt x dt
while for r we obtain
dr
= Lfr h(x(t)) + Lg Lfr1 h(x(t))u(t).
dt
On the right-hand side of this equation, x must be replaced by its expression as a
function of x , which will be written as x = 1 (z, ). Thus, setting
dr
= q(z(t), (t)) + b(z(t), (t))u(t).
dt
dzi i
= [f (x(t)) + g(x(t))u(t)] = Lf i (x(t)) + Lg i (x(t))u(t) = Lf i (x(t)).
dt x
Setting
Lf 1 ( 1 (z, ))
f0 (z, ) =
1
Lf nr ( (z, ))
In addition to these equations, one has to specify how the output of the system is
related to the new state variables. Being y = h(x), it is immediately seen that
y = 1 . (6.3)
The equations thus introduced are said to be in strict normal form. They are
useful in understanding how certain control problems can be solved. The equations in
question can be given a compact expression if we use the three matrices A Rr Rr ,
B Rr R and C R Rr defined in (2.7). With the aid of such matrices, the Eqs.
(6.2) and (6.3) can be rewritten as
6.1 Relative Degree and Local Normal Forms 173
z = f0 (z, )
+ B[q(z,
= A ) + b(z, )u] (6.4)
y = C,
We address, in this section, the problem of deriving the global version of the coor-
dinates transformation and normal form introduced in Sect. 6.1. The exposition is
limited to the introduction of a few basic concepts involved and to the main results.
A detailed analysis and the appropriate differential-geometric background can be
found in the literature.8
Consider again a single-input single-output nonlinear system described by equa-
tions of the form (6.1), in which f (x) and g(x) are smooth vector fields, and h(x)
is a smooth function, defined on Rn . Assume also that f (0) = 0 and h(0) = 0. This
system is said to have uniform relative degree r if it has relative degree r at each
x Rn .
If system (6.1) has uniform relative degree r, the r differentials
Lfr h(x) 1
(x) = (x) =
Lg Lfr1 h(x) Lg Lfr1 h(x)
are complete.11
Then Z is connected. Suppose Z is diffeormorphic to Rnr . Then, the smooth
mapping
1 : Z Rr Rn
(6.6)
(z, (1 , . . . , r )) r1 r1
2
1r (z),
in which t (x) denotes the flow of the vector field , has a globally defined smooth
inverse
(z, (1 , . . . , r )) = (x) (6.7)
11 If is a vector field of Rn , the flow of is the map t (x), in which t R and x Rn , defined by
the following properties: 0 (x) = x and
dt (x)
= (t (x)).
dt
In other words, t (x) is the value at time t R of the integral curve of the o.d.e. x = (x) passing
through x = 0 at time t = 0. The vector field is said to be complete if t (x) is defined for all
(t, x) R Rn .
6.2 Global Normal Forms 175
r 2 1
z = h(x) L r2
h(x)
L r1
h(x)
(x)
f f
i = Lfi1
h(x) 1 i r.
The globally defined diffeomorphism (6.7) changes system (6.1) into a system
described by equations of the form
z = f0 (z, )
+ B[q(z,
= A ) + b(z, )u] (6.8)
y = C
B,
in which A, C are matrices of the form (2.7) and
= col(1 , . . . , r )
q(z, ) = Lfr h( 1 (z, (1 , . . . , r )))
b(z, ) = Lg Lfr1 h( 1 (z, (1 , . . . , r ))).
If, and only if, the vector fields (6.5) are such that
[i , j ] = 0 for all 1 i, j r,
then the globally defined diffeomorphism (6.7) changes system (6.1) into a system
described by equations of the form
z = f0 (z, 1 )
+ B[q(z,
= A ) + b(z, )u] (6.9)
y = C.
Remark 6.3 Note that, if r < n, the submanifold Z is the largest (with respect to
inclusion) smooth submanifold of h1 (0) with the property that, at each x Z , there
is u (x) such that the vector field
is tangent to Z . Actually, for each x Z there is only one value of u (x) that renders
this condition satisfied, which is
Lfr h(x)
u (x) = . (6.10)
Lg Lfr1 h(x)
x = f (x) (6.11)
176 6 Stabilization of Nonlinear Systems via State Feedback
Remark 6.4 Note that (6.9) differs from (6.8) in the fact that while the map f0 (, ) in
(6.8) possibly depends on all components 1 , . . . , r of , in (6.8) the map in question
only depends on its first component 1 .
In this section, we introduce and discuss an important concept, that in many instances
plays a role exactly similar to that of the zeros of the transfer function in a linear
system.12 In order to motivate this concept, consider again for a moment a linear
system in (strict) normal form (2.9). Pick any z Rnr , set = 0, and note that
there is a (unique) input function u () such that the output response y(t) of (2.9)
from the initial state (z(0), (0)) = (z , ) = (z , 0) is identically zero for all t R.
This is the function
1
u (t) = A10 eA00 t z .
b
A simple calculation, in fact, shows that the pair
is a solution pair of (2.9) when u(t) = u (t), and this solution passes through (z , 0)
at time t = 0. The output y(t) associated with such pair is clearly identically zero. It
is seen in this way that the dynamics of the autonomous system
z = A00 z
characterize the forced state behavior of (2.9), whenever input and initial conditions
are chosen in such a way that the output is identically zero. Bearing in mind the fact
that the eigenvalues of A00 are the zeros of the transfer function,13 this characterization
could be pushed a little bit further. In fact, if z is an eigenvector of A00 , associated
with an eigenvalue , it is seen that the forced output response of the system to the
input
1
u (t) = A10 et z ,
b
from the initial condition (z(0), (0)) = (z, 0), is identically zero. In other words,
as it is well known, for each zero of the transfer function one can find an input and
an initial condition yielding an output which is identically zero. In this section, we
extend such interpretation to the case of a nonlinear system of the form (6.1).
Consider again system (6.1), assume that f (0) = 0 and h(0) = 0, suppose that the
system has relative degree r at x = 0 and consider its normal form (6.4) in which,
in view of Remark 6.2, f0 (0, 0) = 0 and q(0, 0) = 0. For such system we study the
problem of finding all pairs consisting of an initial state x and of an input function
u(), for which the corresponding output y(t) of the system is identically zero for all
t in a neighborhood of t = 0.14
Recalling that in the normal form y(t) = 1 (t), it is seen that if y(t) = 0 for all t,
then
1 (t) = 2 (t) = = r (t) = 0,
that is (t) = 0 for all t. Thus, when the output of the system is identically zero
its state is constrained to evolve in such a way that also (t) is identically zero. In
addition, the input u(t) must necessarily be the unique solution of the equation
(recall that b(z(t), 0) = 0 if z(t) is close to 0). As far as the variable z(t) is concerned,
it is seen that
z (t) = f0 (z(t), 0).
From this analysis we deduce the following facts. If the output y(t) is identically
zero, then necessarily the initial state of the system must be such that (0) = 0,
whereas z(0) = z can be arbitrary. According to the value of z , the input must be
q(z(t), 0)
u(t) =
b(z(t), 0)
The dynamics of (6.12) characterize the forced state behavior of the system when
input and initial conditions are chosen in such a way as to constrain the output to
remain identically zero. These dynamics, which are rather important in many of the
subsequent developments, are called the zero dynamics of the system.
Remark 6.5 Note the striking analogy with the corresponding properties of a linear
system. Note also that the same analysis could be performed in the original coor-
14 Ifwe deal with a locally defined normal form, all functions of time are to be seen as functions
defined in a neighborhood of t = 0. Otherwise, if the normal form is globally defined, such functions
are defined for all t for which the solution of (6.12) is defined.
178 6 Stabilization of Nonlinear Systems via State Feedback
dinates. In fact, bearing in mind the results presented in Sect. 6.2, it is seen that if
y(t) = 0 for all t R, then x(t) Z . Moreover,
u(t) = u (x(t))
where u (x) is the function (6.10), and x(t) is a trajectory of the autonomous system
(6.11).
As we will see in the sequel of this Chapter, the asymptotic properties of the
system identified by the upper equation in (6.8) play a role of paramount importance
in the design of stabilizing feedback laws. In this context, the properties expressed
by the following definition are relevant.15
Definition 6.1 Consider a system of the form (6.1), with f (0) = 0 and h(0) = 0.
Suppose the system has uniform relative degree r and possesses a globally defined
normal. The system is globally minimum-phase if the equilibrium z = 0 of the zero
dynamics
z = f0 (z, 0) (6.13)
z = f0 (z, ), (6.14)
15 In the second part of the definition, of the property of input-of-state stability is invoked. For a
summary of the main characterizations of such property and a number of related results, see Sect.
B.2 of Appendix B. For further reading about the property of input-to-state stability, see references
[10, 11, 14] of Appendix B. For further reading about the notion of a minimum-phase system, see
also [8, 9].
16 See Sect. B.2 in Appendix B.
17 In what follows, property (6.16) will be sometime expressed in the equivalent form: V (x) is
and
V0
f (z, ) (
z
) for all (z, ) Rnr Rr such that
z
(
).
z
(6.17)
Of course, this criterion, written for = 0, includes as a special case the criterion of
Lyapunov to determine when a system is globally minimum-phase.
In what follows, it will be useful to consider the special case of systems that are
globally minimum-phase and the convergence to the equilibrium of the trajectories of
(6.13) is locally exponential andalsothe special case of systems that are strongly
minimum-phase and the functions () and (, ) of the estimate (6.15), for some
d > 0 and some
> 0, M > 0, > 0, are bounded as
(r)
r
for |r| d. (6.18)
(r, t) Met r
Remark 6.6 It is seen from the analysis above that a linear system is globally
minimum-phase if the system z = A00 z is asymptotically stable, i.e., if it is minimum-
phase in the sense defined in Sect. 2.2.18 Of course, this property automatically
implies that the linear system
z = A00 z + A01
is input-to-state stable and hence the system is strongly and also locally exponentially
minimum-phase.
A typical setting in which normal forms are useful is the derivation of systematic
methods for stabilization in the large of certain classes of nonlinear system, even in
the presence of parameter uncertainties. We begin this analysis with the observation
that, if the system is strongly minimum-phase, it is quite easy to design a globally
stabilizing state feedback law. To be precise, consider again a system in normal form
18 This
analogy is the motivation for the introduction of the term minimum-phase to indicate the
asymptotic properties considered in Definition 6.1.
180 6 Stabilization of Nonlinear Systems via State Feedback
(6.8), which we assume to be globally defined, and assume that the system is strongly
minimum-phase, i.e., assume that f0 (0, 0) = 0 and that
z = f0 (z, ),
viewed as a system with input and state z, is input-to-state stable. Bearing in mind
the fact that the coefficient b(z, ) is nowhere zero, consider the feedback law
1 ),
u= (q(z, ) + K (6.19)
b(z, )
r
=
K k i Lfi1 h(x)
i=1
1
r
u(x) = Lfr h(x) + k i Lfi1 h(x) , (6.21)
Lg Lfr1 h(x) i=1
6.4 Stabilization via Full State Feedback 181
This feedback strategy, although very intuitive and elementary, is not useful in a
practical context because it relies upon exact cancelation of certain nonlinear func-
tion and, as such, possibly non-robust. Uncertainties in q(z, ) and b(z, ) would
make this strategy unapplicable. Moreover, the implementation of such control law
requires the availability, for feedback purposes, of the full state (z, ) of the system,
a condition that might be hard to ensure. Thus, motivated by these considerations,
we readdress the problem in what follows, by seeking feedback laws depending on
fewer measurements (hopefully only on the measured output y) and possibly robust
with respect to model uncertainties. Of course, in return, some price has to be paid.
We conduct this analysis in the next section and in the following Chapters. For
the time being we conclude by showing how, in the context of full state feedback,
the assumption that the system is strongly minimum-phase can be weakened. This is
possible, to some extent, if the normal form of the system has the special structure
(6.9). However, the expression of the control law becomes more involved.
The design of such feedback law is based on a recursive procedure, known as
backstepping, by means of which it is possible to construct, for a system having
the special structure (6.9), a state feedback stabilizing law as well as a Lyapunov
function.19 The procedure in question reposes on the following results.
z = f (z, )
(6.22)
= u
19 For additional reading on such design procedure, as well as on its use in problems of adaptive
control, see [3, 10].
20 To check that this is always possible, observe that the difference
V
u(z, ) = p(z, ). (6.24)
z
Then, the equilibrium (z, ) = (0, 0) of (6.22) controlled by (6.24) is globally asymp-
totically stable, with Lyapunov function
1
W (z, ) = V (z) + 2 .
2
Proof By assumption, V (z) is positive definite and proper, which implies that the
function W (z, ) in the lemma is positive definite and proper as well. Moreover,
V
f (z, 0) (
z
) z Rnr
z
W W V V
W = f (z, ) + u= f (z, 0) + p(z, ) + u.
z z z
W ( z ) 2 (z, ) Rnr R.
The quantity on the right-hand side is negative for all nonzero (z, ) and this proves
the lemma.
In the next lemma (which contains Lemma 6.2 as a particular case), this result is
extended by showing that to the purpose of stabilizing the equilibrium (z, ) = (0, 0)
of system (6.22), it suffices to assume that the equilibrium z = 0 of
z = f (z, ),
z = f (z, (z) + )
(6.25)
= f (z, (z) + ) + u,
z
6.4 Stabilization via Full State Feedback 183
Pick now
u= f (z, (z) + ) + u
z
z = f (z, (z) + )
= u .
This system has the same structure as that of system (6.22), andby construction
satisfies the assumptions of Lemma 6.2. Thus, this system can be globally stabilized
by means of a control u having the structure of the control indicated in this lemma.
This yields the following consequence.
Lemma 6.3 Consider a system described by equations of the form (6.22), in which
(z, ) Rnr R, and f (0, 0) = 0. Suppose there exists a smooth real-valued func-
tion : Rnr R, with (0) = 0, such that the equilibrium z = 0 of
z = f (z, (z))
Set
V
u (z, ) = f (z, (z) + ) p (z, ) (6.26)
z z
Then, the equilibrium (z, ) = (0, 0) of (6.25) controlled by (6.26) is globally asymp-
totically stable, with Lyapunov function
1
W (z, ) = V (z) + 2 .
2
As a consequence, the equilibrium (z, ) = (0, 0) of (6.22) controlled by
V
u(z, ) = f (z, ) + (z) p (z, (z))
z z
1
W (z, ) = V (z) + ( (z))2 .
2
The function (z), which is seen as a control imposed on the upper subsystem
of (6.22), is usually called a virtual control. The property indicated in Lemma 6.3
184 6 Stabilization of Nonlinear Systems via State Feedback
can be used repeatedly, to address the problem of stabilizing a system of the form
(6.9). In the first iteration, beginning from a virtual control 1 (z) that stabilizes the
equilibrium z = 0 of
z = f0 (z, 1 (z)),
using this lemma one finds a virtual control 2 (z, 1 ) that stabilizes the equilibrium
(z, 1 ) = (0, 0) of
z = f0 (z, 1 )
1 = 2 (z, 1 ).
Then, using the lemma again, one finds a virtual control 3 (z, 1 , 2 ) that stabilizes
the equilibrium (z, 1 , 2 ) = (0, 0, 0) of
z = f0 (z, 1 )
1 = 2
2 = 3 (z, 1 ).
Proposition 6.5 Consider a system of the form (6.9), in which f0 (0, 0) = 0. Suppose
there exists a smooth real-valued function 1 : Rnr R, with 1 (0) = 0, such that
the equilibrium z = 0 of
z = f0 (z, 1 (z))
is globally asymptotically stable. Then, there exists a smooth state feedback law
u = u(z, 1 , . . . , r )
Of course, a special case in which the result of this proposition holds is when
z = f0 (z, 0) has a globally asymptotically stable equilibrium at z = 0, i.e., when
the system is globally minimum-phase. In this case, in fact, the assumption of the
proposition holds with 1 (z) = 0.
Remark 6.7 Note that, while the feedback law (6.19) has a very simple expression,
the feedback law derived above cannot be easily expressed in closed form. Rather, it
can only be derived by means of a recursive procedure. The actual expression of the
law in question also requires the explicit knowledge of the function V (z). In return,
the stabilization method just described does not require the upper subsystem of the
form (6.9) to be input-to-state stable, as assumed in the case of the law (6.19), but
only relies upon the assumption that the subsystem in question is stabilizable, by
means of an appropriate virtual control 1 (z).
6.4 Stabilization via Full State Feedback 185
z = f0 (z, 1 )
1 = q1 (z, 1 ) + b1 (z, 1 )2
2 = q2 (z, 1 , 2 ) + b2 (z, 1 , 2 )3
r = qr (z, 1 , . . . , r ) + br (z, 1 , . . . , r )u.
In the previous section, we have described how certain classes of nonlinear systems
(namely systems that, after a globally defined change of coordinates can be expressed
in the form (6.8) or (6.9)), under suitable hypotheses on the subsystem z = f0 (z, )
(input-to-state stability, or global stabilizability in the special case of a system of
the form (6.9)) can be globally stabilized via full state feedback. In this section, we
show how similar results can be obtained using a partial state feedback. Of course, if
limited state information are available, we expect that weaker results will be obtained.
We consider first the case of a system having relative degree 1, which in normal
form is written as
z = f0 (z, )
= q(z, ) + b(z, )u (6.27)
y=
f0 (0, 0) = 0
q(0, 0) = 0.
u = ky (6.28)
186 6 Stabilization of Nonlinear Systems via State Feedback
z = f0 (z, )
(6.29)
= q(z, ) b(z, )k.
Set21
x = col(z, )
in which
f0 (z, )
Fk (x) = .
q(z, ) b(z, )k
Proposition 6.6 Consider system (6.1), with f (0) = 0 and h(0) = 0. Suppose the
system has uniform relative degree 1 and possesses a globally defined normal form.
Suppose the system is globally minimum-phase. Let the control be provided by the out-
put feedback u = ky so that a closed-loop system modeled as in (6.30) is obtained.
Then, for every choice of a compact set C and of a number > 0, there is a number
k and a finite time T such that, if k k , all trajectories of the closed-loop system
with initial condition x(0) C remain bounded and satisfy
x(t)
< for all t T .
Proof Consider, for system (6.30), the candidate Lyapunov function22
1
W (x) = V (z) + 2
2
which is positive definite and proper. For any real number a > 0, let a denote the
sublevel set of W (x)
a = {x Rn : W (x) a}
and let
B = {x Rn :
x
< }
denote the (open) ball radius . Assume, without loss of generality that C is such that
B C . Since W (x) is positive definite and proper, there exist numbers 0 < d < c
such that
d B C c .
21 With a mild abuse of notation, we use here x instead of x , to denote the vector of coordinates that
characterize the normal form of the system.
22 The arguments used in this proof are essentially those originally proposed in [11] and frequently
W V
W (x) := Fk (x) = f0 (z, ) + q(z, ) b(z, )k 2
x z
is negative at each point of Sdc . To this end, proceed as follows. Consider the compact
set
S0 = {x Sdc : = 0}.
At each point of S0
V
W (x) = f0 (z, 0) (
z
)
x
Since minxS0
z
> 0, there is a number a > 0 such that
W (x) a x S0 .
V
M = max { f0 (z, ) + q(z, )} m = min {b(z, ) 2 }
xS z xS
and observe that m > 0 because b(z, ) > 0 and cannot vanish at any point of S .23
Thus, since k > 0, we obtain
W (x) M km x S .
fact, let x be a point of Sdc for which = 0. Then x S0 , which implies x S . Any of such x
23 In
cannot be in S .
188 6 Stabilization of Nonlinear Systems via State Feedback
as anticipated.
This being the case, suppose the initial condition x(0) of (6.30) is in Sdc . It follows
from known arguments24 that x(t) c for all t 0 and, at some time
T 2(c d)/a,
x(T ) is on the boundary of the set d .25 On the boundary of d the derivative of
W (x(t)) with respect to time is negative and hence the trajectory enters the set d and
remains there for all t T . Since all x d are such that
x
< , this completes
the proof.
This proposition shows that, no matter how large the set C of initial conditions is
chosen and no matter how small a target set B is chosen, there is a value k of the
gain in (6.28) and a finite time T such that, if the actual gain parameter k is larger
than or equal to k , all trajectories of the closed-loop system with origin in C are
bounded and for all t T remain in the set B . This property is commonly referred to
by saying that the control law (6.28) is able to semiglobally and practically stabilize
the point (z, ) = (0, 0). The term practical (as opposite to asymptotic) is meant
to stress the fact that the convergence is not to a point, but rather to a neighborhood
of that point, that can be chosen arbitrarily small, while the term semiglobal (as
opposite to global) is meant to stress the fact that the convergence to the target set
is not for all initial conditions, but rather for a compact set of initial conditions, that
can be chosen arbitrarily large.26
The result presented in Proposition 6.6 can be seen as nonlinear analogue of the
stabilization result presented in Sect. 2.3 of Chap. 2. The standing assumption in
both cases is that the system is minimum-phase and the stability result is obtained
via high-gain output feedback. In the case of the nonlinear system, the minimal value
k of the feedback gain k is determined by the choice of the set C and by the value
of . In particular, as it is clear from the proof of Proposition 6.6, k increases as C
increases (in the sense of set inclusion) and also increases as decreases.
To obtain asymptotic stability, either a nonlinear control law u = (y) is needed
or, if one insists in using a linear law u = ky, extra assumptions are necessary. The
first option shall be briefly covered in Sect. 8.2 of Chap. 8.27 For the time being, we
address here the second option.
Proposition 6.7 Consider system (6.1), with f (0) = 0 and h(0) = 0. Suppose the
system has relative degree 1 and possesses a globally defined normal form. Sup-
pose the system is globallyand also locally exponentiallyminimum-phase. Let
the control be provided by the output feedback u = ky. Then, for every choice of
a compact set C there is a number k such that, if k k , the equilibrium x = 0 of
the resulting closed-loop system is asymptotically (and locally exponentially) stable,
with a domain of attraction A that contains the set C .
in which
f0
g(z)
F0 = (0, 0) and lim = 0. (6.33)
z z0
z
If the zero dynamics are locally exponentially stable, all eigenvalues of F0 have
negative real part and there exists a positive-definite solution P of
1
U(x) = zT Pz + 2
2
yielding
U(x) = 2zT Pf0 (z, ) + q(z, ) b(z, )k 2 .
Write f0 (z, ) as
1
z B
g(z)
z
2
P
Set
S = {(z, ) :
z
< , | | < }.
Since the function [f0 (z, ) f0 (z, 0)] is a continuously differentiable function that
vanish at = 0, there is a number M1 such that
190 6 Stabilization of Nonlinear Systems via State Feedback
Finally, since b(z, ) is positive and nowhere zero, there is a number b0 such that
Putting all these inequalities together, one finds that, for all for all (z, ) S
U(x)
z
2 + (M1 + N1 )
z
| | (kb0 N2 )| |2 .
then
1
U(x)
x
2 .
2
This shows that there is a number k2 such that, if k k2 , the function U(x) is negative
definite for all (z, ) S . Pick now any (nontrivial) sublevel set d of U(x) entirely
contained in the set S and let r be such that Br d . Then, the argument above
shows that, for all k k2 , the equilibrium x = 0 is asymptotically stable with a
domain of attraction that contains Br . Note that the set Br depends only by the value
of and on the matrix P, with both and P determined only by the function f0 (z, 0).
In particular, the set Br does not depend on the value chosen for k.
Pick < r and any C . By Proposition 6.6, we know that there is a number k1
such that, if k k1 , all trajectories with initial condition in C in finite time enter
the (closure of) the set B , and hence enter the region of attraction of x = 0. Setting
k = max{k1 , k2 }, it can be concluded that for all k k the equilibrium x = 0 of
(6.30) is asymptotically (in fact, locally exponentially) stable, with a domain of
attraction that contains C .
The case of a system having relative degree r > 1 can be reduced to the case
discussed above by means of a technique which is reminiscent of the technique
introduced in Sect. 2.4. Suppose the normal form
z = f0 (z, 1 , . . . , r1 , r )
1 = 2
(6.34)
r1 = r
r = q(z, 1 , . . . , r1 , r ) + b(z, 1 , . . . , r1 , r )u.
6.5 Stabilization via Partial State Feedback 191
is globally defined and let the variable r be replaced by a new state variable defined
as (compare with (2.21))
= r + a0 1 + a1 2 + + ar2 r1
This system, with regarded as output, has a structure which is identical to that
of system (6.27). In fact, if we set
= col(z, 1 , . . . , r1 ) Rn1
y =
), y as
and define f0 (, ), q (, ), b(,
r1
f0 (z, 1 , . . . , r1 , i=1 ai1 i + )
2
f0 (, ) =
r1
r1
i=1 ai1 i +
r1
q (, ) = a0 2 + a1 3 + + ar2 ( i=1 a + )
r1 i1 i
+ q(z, 1 , . . . , r1 , i=1 ai1 i + )
) = b(z, 1 , . . . , r1 , r1 ai1 i + ),
b(, i=1
= f0 (, )
)u
= q (, ) + b(, (6.35)
y = ,
From this viewpoint, it is trivial to check that if the functions f0 (z, ), q(z, ), and
b(z, ) satisfy
f0 (0, 0) = 0
q(0, 0) = 0
b(z, ) > 0 for all (z, )
then also
f0 (0, 0) = 0
q (0, 0) = 0
) > 0
b(, for all (, ).
= f0 (, 0)
u = k = k(a0 1 + a1 2 + + ar2 r1 + r )
6.5 Stabilization via Partial State Feedback 193
and bearing in mind the fact that i = Lfi1 h(x) for i = 1, . . . , r, one can conclude
the following stabilization result.
Proposition 6.8 Consider system (6.1), with f (0) = 0 and h(0) = 0. Suppose the
system has uniform relative degree r and possesses a globally defined normal form.
Suppose the system is strongly minimum-phase. Let the control be provided by a
feedback of the form
r
u = k ai Lfi1 h(x) . (6.38)
i=1
with the ai s such that the polynomial (6.37) is Hurwitz and ar1 = 1. Then, for
every choice of a compact set C and of a number > 0, there is a number k and a
finite time T such that, if k k , all trajectories of the closed-loop system with initial
condition x(0) C remain bounded and satisfy
x(t)
< for all t T . If the system
is stronglyand also locally exponentiallyminimum-phase, for every choice of a
compact set C there is a number k such that, if k k , the equilibrium x = 0 of
the resulting closed-loop system is asymptotically (and locally exponentially) stable,
with a domain of attraction A that contains the set C .
Remark 6.9 Note that in Proposition 6.8 we have used the assumption that the system
is strongly minimum-phase while in Propositions 6.6 and 6.7 we had used the weaker
assumption that the system is globally minimum-phase. This is due to the fact that
we are considering here the general case in which the normal form of the system has
the structure (6.8), where the dynamics of the upper subsystem
z = f0 (z, )
are (possibly) affected by all the components of . If the normal form of the system
had the special structure (6.9), in which the dynamics of z are affected only by 1 ,
one might have weakened the assumption, requiring only the system to be globally
minimum-phase. Details are not covered here and can be found in the literature.28
If r = 1 the difference between the structure (6.8) and (6.9) ceases to exists and this
explains the weaker assumption used in Propositions 6.6 and 6.7.
Remark 6.10 It is worth comparing the feedback laws (6.21) and (6.38) which, in
the coordinates of the normal form, read as
1 r
u= q(z, ) + ki i
b(z, ) i=1
and, respectively,
r
u = k ai i .
i=1
r
r = ki i ,
i=1
while if the latter is used such cancelation does not takes place and
r
r = q(z, ) b(z, )k ai i . (6.39)
i=1
The first law is indeed sensitive to uncertainties in b(z, ) and q(z, ) while the
second is not. If the second law is used, the presence of the nonlinear terms in (6.39)
is made negligible by taking a large value of the gain parameter k. Such simpler
and more robust feedback law, though, is not expected to yield global stability, but
only practical stability with a guaranteed region of attraction. If asymptotic stability
is sought, the additional assumption that the equilibrium z = 0 of the zero dynamics
is locally exponentially stable is needed.
The stabilizing feedback laws discussed in this section are, to some extent, nonlin-
ear counterparts of the feedback law considered in Sect. 2.4. Such laws presume the
availability of all the components 1 , . . . , r of the partial state that characterizes
the normal form (6.8). One might wonder, at this point, whether procedures similar
to those described in Sect. 2.5 could be set up, and whether the goal of achieving
asymptotic stability could be obtained using, instead of , an appropriate replace-
ment generated by a dynamical systems driven only by the measured output y. As
a matter of fact, this is actually feasible. The extension of the method presented in
Sect. 2.5 and the associated convergence proof involve some subtleties and require
appropriate care. Since one of the purposes of the book is to discuss problems of
stabilization also for multi-input multi-output systems, in order to reduce duplica-
tions we defer the discussion of this topic to Chap. 10, where the case of multi-input
multi-output systems is addressed. In that Chapter, it will be shown not only how a
suitable replacement of the partial state can be generated, but also how a robust
version of the feedback law (6.21) can be obtained, if semiglobal stabilization is
sought.
Example 6.1 In this example it is shown that, if the relative degree of the system
is higher than 1 and the normal form does not have the special structure (6.9), the
assumption that system is globally minimum phase may not suffice to guarantee the
existence of a feedback law that stabilizes the equilibrium x = 0 with an arbitrarily
large region of attraction. Consider the system having relative degree 2
6.6 Examples and Counterexamples 195
z = z + z2 1 + (z, 2 )
1 = 2
2 = u
and suppose
1
(z, 2 ) = (z 2 )2 z.
4
This system is globally minimum phase. However, it is not strongly minimum phase.
Set = z1 and observe that
= + 2 + (z, 2 )1 + z2 .
At each point of S
1 1 1
(z, 2 )1 + z2 = (z 2 )2 + z2 (z 2 )2 + z2 = (z2 + 22 ) 0.
4 2 2
Therefore, at each point of the set S,
+ 2 2.
This shows that if (0) = z(0)1 (0) 2, then z(t)1 (t) 2 for all times for which
the trajectory is defined. As a consequence, it is concluded thatno matter how the
control u is chosentrajectories with initial conditions satisfying z(0)1 (0) 2 can
never converge to the equilibrium point (z, 1 , 2 ) = (0, 0, 0).
z = z + z2
= u.
z = z + z2
= k.
is locally exponentially stable, with a domain of attraction that contains the set C .
This fact is not in contradiction with the conclusions of the Example B.3. In fact, as
196 6 Stabilization of Nonlinear Systems via State Feedback
shown in that example, for each z0 > 1, the solution would escape to infinity in finite
time if
(t) = exp(kt)0 1 for all t [0, tmax (z0 )) (6.40)
where tmax (z0 ) = ln(z0 ) ln(z0 1). Now, for any choice of (z0 , 0 ) it is always
possible to find a value of k such that (t) decreases fast enough so that the estimate
(6.40) is violated. Thus, the arguments used in the example to conclude that the
system has finite escape time are no longer applicable. From the example we learn,
though, that even if by choosing a suitable k the system can be made asymptotically
stable, with a domain of attraction that contains the set C , there are initial conditions
outside C from which the trajectory escapes to infinity in finite time.
Example 6.3 This example is meant to show that, in a system like (6.29), asymptotic
stability of the equilibrium z = 0 of z = f0 (z, 0) is not sufficient to guarantee, even
for large k > 0, local asymptotic stability of the equilibrium (z, ) = (0, 0). Consider
the system
z = z3 +
= z + u.
This system is globally minimum phase but not globally and also locally exponen-
tially minimum phase. Set u = k to obtain the system (compare with (6.29))
z = z3 +
= z k. (6.41)
No matter how k is chosen, the equilibrium (z, ) = (0, 0) cannot be locally asymp-
totically stable. In fact, the linear approximation at this equilibrium
z 0 1 z
=
1 k
has a characteristic polynomial 2 + k 1 with one root having positive real part
for any k.
This is not in contradiction, though, with Proposition 6.6. In fact, system (6.41),
for k > 0, has three equilibria
(z, ) = (0,
0)
(z, ) = (1/ )
k, 1/ k 3
(z, ) = (1/ k, 1/ k 3 ).
The former of these is unstable, but the other two are locally asymptotically stable.
The unstable equilibrium, which is a saddle point, has a stable manifold M s (the
set of all points x with the property that the integral curve satisfying x(0) = x is
defined for all t > 0 and limt x(t) = 0) and an unstable manifold M u (the set of
all points x with the property that the integral curve satisfying x(0) = x is defined
6.6 Examples and Counterexamples 197
for all t < 0 and limt x(t) = 0). Trajectories with initial conditions that are not
in M s asymptotically converge, as t , to either one of the two locally stable
equilibria. From the characterization above, it is seen that, for any , there is a number
k such that, if k > k , both the stable equilibria are in B . Thus, all trajectories enter
in finite time the target set B (but, with the exception of those with origin in M s ,
do not converge to x = 0).
x 1 = x1 x13 + x2
x 2 = x2 + x3
(6.42)
x 3 = x1 + x22 + u
y = x2 .
This system has relative degree 2. Its normal form is obtained setting
1 = x2 , 2 = x2 + x3 , z = x1
and reads as
z = z z 3 + 1
1 = 2
2 = z + 12 + u
y = 1 .
z = z z3 + 1
1 = 2 2z 2z3 + 21
2 = z + (1 2z)2 + u.
The subsystem consisting of the two upper equations, viewed as a system with
state (z, 1 ) and control 2 , can be stabilized by means of a virtual control
W1
2 (z, 1 ) = (2z 2z3 + 21 ) 1 = 31 + z + 2z3 .
z
198 6 Stabilization of Nonlinear Systems via State Feedback
2 = 2 2 (z, 1 ) = 2 + 31 z 2z3
z = z z3 + 1
1 = z 1 + 2
2 = a(z, 1 , 2 ) + u
in which
W2
u(z, 1 , 2 ) = a(z, 1 , 2 ) 2 = a(z, 1 , 2 ) 2 1 ,
1
1 2
W3 (z, 1 , 2 ) = (z + 12 + 22 ).
2
Reversing all changes of coordinates used in the above construction, one may find a
stabilizing law u = (x) expressed in the original coordinates (x1 , x2 , x2 ).
References
1. H. Nijmeijer, A. van der Schaft, Nonlinear Dynamical Control Systems (Springer, New York,
1990)
2. A. Isidori, Nonlinear Control System, 3rd edn. (Springer, London, 1995)
3. M. Kristic, I. Kanellakopoulos, P. Kokotovic, Nonlinear Adaptive Control Design (Wiley, New
York, 1995)
4. R. Marino, P. Tomei, Nonlinear Control Design: Adaptive, Robust (Prentice Hall, New York,
1995)
5. H. Khalil, Nonlinear Systems, 3rd edn. (Prentice Hall, Upper Saddle River, 2002)
References 199
In this chapter we discuss the design of observers for nonlinear systems modeled by
equations of the form
x = f (x, u)
(7.1)
y = h(x, u)
: Rn Rn
x z
that carries system (7.1) into a system described by equations of the form
z 1 = f1 (z 1 , z 2 , u)
z 2 = f2 (z 1 , z 2 , z 3 , u)
(7.2)
z n1 = fn1 (z 1 , z 2 , . . . , z n , u)
z n = fn (z 1 , z 2 , . . . , z n , u)
1 , u)
y = h(z
1 High-gain observers have been considered by various authors in the literature. Here we closely
follow the approach of Gauthier and Kupca, who have thoroughly investigated the design of high-
gain observers in [1].
Springer International Publishing Switzerland 2017 201
A. Isidori, Lectures in Feedback Design for Multivariable Systems,
Advanced Textbooks in Control and Signal Processing,
DOI 10.1007/978-3-319-42031-8_7
202 7 Nonlinear Observers and Separation Principle
h fi
= 0, and = 0, for all i = 1, . . . , n 1 (7.3)
z 1 z i+1
for all z Rn , and all u Rm . Equations having this structure and such properties
are said to be in GauthierKupcas uniform observability canonical form.
Remark 7.1 The term uniform here is meant to stress the fact that the possibility of
observing the state is not influenced by the actual input u() affecting the system. In
fact, in contrast with the case of a linear system, in which the property of observability
is independent of the input, this might not be the case for a nonlinear system, as it will
also appear from the analysis below. The use of the term observability canonical
form can be motivated by various arguments. To begin with, consideras a special
casea single-output linear system
x = Ax + Bu
y = C x + Du,
z 1 C CAx + CBu z 2 + CBu
z 2 CA CA2 x + CABu z 3 + CABu
= x = =
n2 n1
z n1 CA CA x + CAn2 Bu z n + CAn2 Bu
z n CAn1 CAn x + CAn1 Bu CAn T 1 z + CAn1 Bu
y = z 1 + Du
which clearly have the form (7.2) and in which the properties (7.3) hold.
For a general nonlinear system, suppose equations of the form (7.2), with prop-
erties (7.3), hold. Pick an input u(), fix an initial condition z(0) and let z(t)
and y(t) denote the corresponding state and output trajectories. Define a function
F1 (z 1 , u, y) as
1 , u).
F1 (z 1 , u, y) = y h(z
7.1 The Observability Canonical Form 203
F1
= 0 (z 1 , u, y).
z 1
Therefore, by the implicit function theorem, Eq. (7.4)that can be seen as defin-
ing z 1 (t) as implicit function of (u(t), y(t))can be solved for z 1 (t), at least in a
neighborhood of a point {z 1 (0), u(0), y(0)} satisfying
1 (0), u(0)).
y(0) = h(z
for all (u, y) in a neighborhood of (u(0), y(0)) and z 1 (t) can be expressed as
F2 (z 2 , z 1 , z 1 , u) = z 1 f1 (z 1 , z 2 , u).
F2
= 0 (z 2 , z 1 , z 1 , u).
z 2
F2 (H2 (z 1 , z 1 , u), z 1 , z 1 , u) = 0
for all (z 1 , z 1 , u) in a neighborhood of (z 1 (0), z 1 (0), u(0)) and z 2 (t) can be expressed
as
z 2 (t) = H2 (z 1 (t), z 1 (t), u(t)). (7.7)
In view of the previous calculation, which has shown that z 1 (t) can be expressed as
in (7.5), define a function H 1 (u, u (1) , y, y (1) ) as
H1 (1) H1 (1)
H 1 (u, u (1) , y, y (1) ) = u + y
u y
z 2 (t) = H2 (H1 (u(t), y(t)), H 1 (u(t), u (1) (t), y(t), y (1) (t)), u(t)).
We review in what follows how the existence of canonical forms of this kind
can be checked and the canonical form itself can be constructed. We begin with the
description of a set of necessary conditions for the existence of such canonical form.
Consider again system (7.1), suppose that f (0, 0) = 0, h(0, 0) = 0, and define
recursivelya sequence of real-valued functions i (x, u) as follows
i1
1 (x, u) := h(x, u), i (x, u) := f (x, u), (7.8)
x
for i = 1, . . . , n. Finally, with each of such i (x, u)s, associate the subspace
i
Ki (x, u) = Ker .
x (x,u)
dimKi (x, u) = n i.
Ki (x, u) = independent of u.
In other words, condition (i) says that the distribution Di (u) has constant dimension
n i. Condition (ii) says that, for each x, the subspace Ki (x, u) is always the same,
regardless of what u Rm is.3
Proposition 7.1 System (7.1) is globally diffeomorphic to a system in Gauthier
Kupcas observability canonical form only if its canonical flag is uniform.
Proof (Sketch of) Suppose a system is already in observability canonical form and
compute its canonical flag. A simple calculation shows that the each of the functions
i (x, u) is a function of the form
i (z 1 , . . . , z i , u),
and
i
= 0, for all z 1 , . . . , z i , u.
z i
2 See[1].
3 Condition (i) is a regularity condition. Condition expresses the independence of Ki (x, u) on
the parameter u.
206 7 Nonlinear Observers and Separation Principle
This shows that the canonical flag of a system in observability canonical form is
uniform. This property is not altered by a diffeomorphism and hence the condition
in question is a necessary condition for an observability canonical form to exist.
Remark 7.2 The necessary condition thus identified is also sufficient for the exis-
tence of a local diffeomorphism carrying system (7.1) into a system in observability
canonical form.4
Proposition 7.2 Consider the nonlinear system (7.1) and define a map
: Rn Rn
x z = (x)
as
1 (x, 0)
2 (x, 0)
(x) =
.
n (x, 0)
Suppose that:
Proof By assumption,
i
Ker
x (x,u)
n (T (z), 0) = z.
Then
n T
=I
x x=T (z)
u=0 z
T
i
Ker , z Rn .
z j x x=T
u=0
(z)
T
i
Ker , j > i, z Rn , u Rm . (7.9)
z j x x=T
u=u
(z)
Suppose the map z = (x) is used to change coordinates in (7.1) and consider the
system in the new coordinates
z = f(z, u)
u),
y = h(z,
where
u) = h(T (z), u),
h(z, f(z, u) =
n
f (T (z), u).
x x=T
u=0
(z)
Define
u), i1
1 (z, u) := h(z, i (z, u) := f (z, u) for 2 = 1, . . . , n.
z
which implies
1 (z, u)
..
i (z, u) := . = i (T (z), u).
i (z, u)
208 7 Nonlinear Observers and Separation Principle
1 (z, u) 1 T
= =0
z j x x=T (z)
u=u z j
h 1
= = 0.
z 1 z
2 (z, u) 2 T
= =0
z j x x=T (z)
u=u z j
for all j > 2 which means that 2 (z, u) only depends on z 1 , z 2 . Looking at the form
of 2 (z, u), we deduce that f1 (z, u) only depends on z 1 , z 2 . Moreover, an easy check
shows that
2 (z, u) h f1
= 0 0
z z 1 z 2
and hence
f1
= 0
z 2
In this section we specialize the results discussed in the previous section to the case in
which the model of the system is input-affine, i.e., the system is modeled by equations
of the form
7.2 The Case of Input-Affine Systems 209
x = f (x) + g(x)u
(7.10)
y = h(x).
It is easy to check that, in a system of the form (7.10), the functions i (x, u) defined
by (7.8) have the following expressions:
1 (x, u) = h(x)
2 (x, u) = L f h(x) + L g h(x)u
3 (x, u) = L 2f h(x) + [L g L f h(x) + L f L g h(x)]u + L 2g h(x)u 2
4 (x, u) = L 3f h(x) + [L g L 2f h(x) + L f L g L f h(x) + L 2f L g h(x)]u+
+[L 2g L f h(x) + L g L f L g h(x) + L f L 2g h(x)]u 2 + L 3g h(x)u 3
5 (x, u) =
Hence
h(x)
L f h(x)
n (x, 0) =
:= (x).
L n1
f h(x)
It is known from Proposition 7.2 that if the canonical flag of (7.10) is uniform
and if n (x, 0) is a global diffeomorphism, the system is transformable into uniform
observability canonical form. The form in question is
z = f(z) + g(z)u
y = h(z)
in which
(x) (x)
f(z) = f (x) ,
g(z) = g(x) ,
x x= 1 (z) x x= 1 (z)
h(z) = h( 1 (z)).
By construction,
h(z) = z1.
Moreover, bearing in mind the special structure of (x), it is easy to check that f(z)
has the following form:
z2
z3
f (z) = .
zn
f n (z 1 , . . . , z n )
210 7 Nonlinear Observers and Separation Principle
Finally, since we know that the ith entry of f(z) + g(z)u can only depend on
z 1 , z 2 , . . . , z i+1 , we deduce that g(z)
must necessarily be of the form
g 1 (z 1 , z 2 )
g 2 (z 1 , z 2 , z 3 )
g(z)
= .
g n1 (z 1 , z 2 , . . . , z n )
g n (z 1 , z 2 , . . . , z n )
1 (z, u) = z 1
2 (z, u) = z 2 + g 1 (z 1 , z 2 )u
g 1
1+ u=0
z 2
in which case
1 (z, u)
1 0 0 0
z = .
2 (z, u) 0 0 0
z
At this value of (z, u) the subspace K2 (z, u) has dimension n 1 and not n 2 as
prescribed. Hence the uniformity conditions are violated. We conclude that
g 1
=0
z 2
Proceeding in this way, it can be concluded that the uniform observability canon-
ical form of an input-affine system has the following structure:
z 1 = z 2 + g 1 (z 1 )u
z 2 = z 3 + g 2 (z 1 , z 2 )u
z n1 = z n + g n1 (z 1 , z 2 , . . . , z n1 )u
z n = fn (z 1 , z 2 , . . . , z n ) + g n (z 1 , z 2 , . . . , z n )u
y = z1.
In this section, we describe how to design a global asymptotic state observer for
a system in GauthierKupcas observability canonical form. Letting z i denote the
vector
z i = col(z 1 , . . . , z i )
the canonical form in question can be rewritten in more concise form as5
z 1 = f 1 (z 1 , z 2 , u)
z 2 = f 2 (z 2 , z 3 , u)
(7.11)
z n1 = f n1 (z n1 , z n , u)
z n = f n (z n , u)
y = h(z 1 , u).
The construction described below reposes on the following two additional tech-
nical assumptions:
(i) each of the maps f i (z i , z i+1 , u), for i = 1, . . . , n 1, is globally Lipschitz with
respect to z i , uniformly in z i+1 and u, and the map f n (z n , u) is globally Lipschitz
with respect to z n , uniformly in u.
(ii) there exist two real numbers , , with 0 < < , such that
h f
i
, and , for all i = 1, . . . , n 1
z 1 z i+1
Remark 7.3 Note that the properties (i) and (ii) can be assumed without loss of
generality if it is knowna priorithat z(t) remains in a compact set C .
5 For convenience, we drop the tilde above h() and the f i ()s.
212 7 Nonlinear Observers and Separation Principle
The observer for (7.11) consists of a copy of the dynamics of (7.11) corrected by
an innovation term proportional to the difference between the output of (7.11) and
the output of the copy. More precisely, the observer in question is a system of the
form
z 1 = f 1 (z 1 , z 2 , u) + cn1 (y h(z 1 , u))
z 2 = f 2 (z 2 , z 3 , u) + 2 cn2 (y h(z 1 , u))
(7.12)
z n1 = f n1 (z n1 , z n , u) + n1 c1 (y h(z 1 , u))
z n = f n (z n , u) + n c0 (y h(z 1 , u)),
i = z i z i , i = 1, 2, . . . , n,
an estimate of this error can be obtained as follows. Observe that, using the mean
value theorem, one can write
f i (z i , z i+1 , u) f i (z i , z i+1 , u) = f i (z i (t), z i+1 (t), u(t)) f i (z i (t), z i+1 (t), u(t))
+ f i (z i (t), z i+1 (t), u(t)) f i (z i (t), z i+1 (t), u(t))
fi
= (z (t), i (t), u(t)) i+1 (t)
z i+1 i
+ f i (z i (t), z i+1 (t), u(t)) f i (z i (t), z i+1 (t), u(t))
in which i (t) is a number in the interval [z i+1 (t), z i+1 (t)]. Note also that
h
y h(z 1 , u) = h(z 1 (t), u(t)) h(z 1 (t), u(t)) = (0 (t), u(t)) 1
z 1
h
g1 (t) = (0 (t), u(t))
z 1
(7.13)
fi
gi+1 (t) = (z (t), i (t), u(t)) for i = 1, . . . , n 1,
z i+1 i
i = f i (z i (t), z i+1 (t), u(t)) f i (z i (t), z i+1 (t), u(t)) i cni (y h(z 1 (t), u(t)))
= gi+1 (t)i+1 + f i (z i (t), z i+1 (t), u(t)) f i (z i (t), z i+1 (t), u(t)) i cni g1 (t)1 .
7.3 High-Gain Nonlinear Observers 213
in which
Fi = f i (z i (t), z i+1 (t), u(t)) f i (z i (t), z i+1 (t), u(t))
for i = 1, 2, . . . , n 1 and
ei = ni i = ni (z i z i ), i = 1, 2, . . . , n. (7.14)
The right-hand side of this equation consists of a term which is linear in the vector
e = col(e1 , e2 , . . . , en )
and of a nonlinear term. The nonlinear term, though, can be bounded by a quantity which
is linear in
e
. In fact, observe that, because of Assumption (i), there is a number L such
that
|Fi | L
z i z i
for i = 1, 2, . . . , n.
Clearly,
e12 e22 ei2
z i z i
= 12 + 22 + + i2 = + + +
( n1 )2 ( n2 )2 ( ni )2
As far as the properties of the linear part are concerned, the following useful lemma can
be invoked.
and suppose there exists two real numbers , , with 0 < < , such that
Then, there is a set of real numbers c0 , c1 , . . . , cn1 , a real number > 0 and a symmetric
positive-definite n n matrix S, all depending only on and , such that
Remark 7.4 Note that, if the gi (t)s were constant, the matrix in question could be seen
as a matrix of the form
0 g2 0 0 0 cn1
0 0 g3 0 0 cn2
A GC =
g1 0 0 0 .
0 0 0 0 gn c1
0 0 0 0 0 c0
If all gi s are nonzero, the pair A, C is observable and hence there exists a vector G,
namely a set of coefficients c0 , c1 , . . . , cn1 , such that the matrix A GC is Hurwitz
and an inequality of the form (7.18) holds. Lemma 7.1 considers the more general case in
which the gi (t)s are functions of time, but bounded as in (7.17). It is stressed that, even
if in this case the matrix A(t) is time dependent, the matrix S is not, and only depends on
the numbers and . The proof of the lemma, which also describes how the coefficients
ci s are to be chosen, is given in Sect. 7.4.
Note that, in view of (7.13), the hypothesis that the gi (t)s satisfy (7.17) is in the
present case fulfilled as a straightforward consequence of Assumption (ii). With this
result in mind consider, for system (7.15), the candidate Lyapunov function
V (e) = eT Se.
Then, we have
V (e(t)) = eT (t)[A(t)S + S AT (t)]e(t) + 2eT (t)S F(t)
7.3 High-Gain Nonlinear Observers 215
the vector
having denoted by F(t)
F(t) = col( n1 F1 , n2 F2 , . . . , Fn ).
it is seen that
V (e(t)) ( a0 )
e
2 .
V (e(t)) aV (e(t))
lim e(t) = 0.
t
Theorem 7.1 Consider a nonlinear in uniform observability canonical form (7.11) and
an observer defined as in (7.12). Suppose assumptions (i) and (ii) hold. Then, there is a
choice of the coefficients c0 , c1 , . . . , cn1 and a value such that, if > ,
In summary, the observer (7.12) asymptotically tracks the state of system (7.11) if
the coefficients c0 , c1 , . . . , cn1 are such that the property indicated in Lemma 7.1 holds
(which is always possible by virtue of Assumption (ii)) and if the number is sufficiently
large. This is why the observer in question is called a high-gain observer.
In this section, we give a proof Lemma 7.1.6 The result is obviously true in case n = 1.
In this case, in fact, the matrix A(t) reduces to the scalar quantity
A(t) = c0 g1 (t).
2c0 g1 (t)
for some > 0, which is indeed possible if the sign7 of c0 is the same as that of g1 (t).
The proof for n 2 is by induction on n, based on the following result.
Lemma 7.2 Let and be fixed real numbers satisfying 0 < < . Suppose there
exists a choice of numbers a0 , a1 , . . . , ai1 , a positive-definite matrix Si and a number
i > 0, all dependent only on and , such that, for any set of continuous functions
{g1 (t), g2 (t), . . . , gi (t)} bounded as
the matrix
ai1 g1 (t) g2 (t) 0 0 0
ai2 g1 (t) 0 g3 (t) 0 0
Ai (t) =
a1 g1 (t) 0 0 0 gi (t)
a0 g1 (t) 0 0 0 0
satisfies
Si Ai (t) + AiT (t)Si i I.
the matrix
bi g1 (t) g2 (t) 0 0 0
bi1 g1 (t) 0 g3 (t) 0 0
Ai+1 (t) =
b1 g1 (t) 0 0 0 gi+1 (t)
b0 g1 (t) 0 0 0 0
satisfies
Si+1 Ai+1 (t) + Ai+1
T
(t)Si+1 i+1 I. (7.20)
Proof Let {g1 (t), g2 (t), . . . , gi (t), gi+1 (t)} be any set of continuous functions bounded
as in (7.19). Set
7 The g (t)s are continuous functions that never vanish. Thus, each of them has a well-defined sign.
i
7.4 The Gains of the Nonlinear Observer 217
0 g3 (t) 0 0 0 bi1
0 0 g4 (t) 0 0 bi2
A(t) =
, B=
0 0 0 0 gi+1 (t) b1
0 0 0 0 0 b0
ai1
ai2
C(t) = g2 (t) 0 0 0 0 K =
a1
a0
obtain
bi g1 (t) C(t)K C(t)
A i+1 (t) = T Ai+1 (t)T 1 =
(B + K bi )g1 (t) L(t)K L(t)
in which
L(t) = A(t) + K C(t).
The matrix L(t) has a structure identical to that of Ai (t), with g j (t) replaced by the
g j+1 (t), for j = 1, . . . , i. Thus, by the hypothesis of the lemma, the matrix L(t) satisfies
Si L(t) + L T (t)Si i I.
and set
Q(z) = z T ( Si+1 A i+1 (t) + A i+1
T
(t) Si+1 )z.
i
Q(z)
z
2 . (7.21)
2
218 7 Nonlinear Observers and Separation Principle
Choose
B = K bi
Since g1 (t) by assumption has a well-defined sign, let bi have the same sign as that of
g1 (t). Then, the previous inequalities hold if
1 1
|bi | |g1 (t)| > + |C(t)K | + (
C(t)
+
Si
L(t)
K
)2 . (7.22)
4 i
The proof above provides a recursive procedure for the calculations of the coefficients
c0 , c1 , . . . , cn1 for which the result of Lemma 7.1 holds, as illustrated in the following
simple example.
7.4 The Gains of the Nonlinear Observer 219
z 1 = f 1 (z 1 , z 2 , u)
z 2 = f 2 (z 1 , z 2 , z 3 , u)
z 2 = f 3 (z 1 , z 2 , z 3 , u)
y = h(z 1 , u)
h f1 f2
<| | < , < | | < , < | | < ,
z 1 z 2 z 3
for all z R3 and u R. Thus, the (unavailable) functions g1 (t), g2 (t), g3 (t) satisfy
A1 (t) = a0 g1 (t).
Suppose, without loss of generality, that g1 (t) > 0. Take S1 = 1, that yields
Following the procedure indicated in the proof, observe that (recall that a0 = 1)
1 0
B = b0 , K = 1, T = , L(t) = g2 (t), S2 = I2 .
1 1
b0 = b1 .
1 1
|b1 g1 (t)| > + |g2 (t)| + (2|g2 (t)|)2 ,
4 2
220 7 Nonlinear Observers and Separation Principle
which is fulfilled if
1 2
b1 > b1 := + +2 2.
4
As shown in the lemma, an inequality of the form (7.20) holds for i = 2, with
2 1
S2 = T T S2 T = T T T =
1 1
and 2 = min (T T T ). Letting the pair (b0 , b1 ) thus found by denoted by (a0 , a1 ), we
see that (a0 , a1 ), together with 2 and S2 are such that the assumption of the lemma holds
for i = 2. Observe that a0 = a1 = a, with a > b1 .
The next (and final) step is to find numbers b0 , b1 , b2 such that the 3 3 matrix
b2 g1 (t) g2 (t) 0
A3 (t) = b1 g1 (t) 0 g3 (t)
b0 g1 (t) 0 0
has the desired properties. Following the procedure indicated in the proof, observe that
1 00
b a
B= 1 , K = , T = a 1 0 ,
b0 a
a 0 1
1 0 0
0 g3 (t) a
L(t) = + g2 (t) 0 , S3 = 0 2 1 .
0 0 a
0 1 1
The magnitude of b2 should satisfy the inequality (7.22), which in the present case is
fulfilled if (observe that
K
= 2a,
S2
< 3 and 2 > 0.25)
1 1
b2 g1 (t) > + |g2 (t)| 2a + (|g2 (t)| + 3(|g3 (t)| + 2a |g2 (t)|))2 .
4 0.25
In this section, we show how the theory of nonlinear observers presented earlier can
be used to the purpose of achieving asymptotic stability via dynamic output feedback.8
Consider a single-input single-output nonlinear system in observability canonical form
(7.2), which we rewrite in compact form as
z = f (z, u)
(7.23)
y = h(z, u),
with f (0, 0) = 0 and h(0, 0) = 0 and suppose there exists a feedback law u = u (z),
with u (0) = 0, such that the equilibrium z = 0 of
D = diag(, 2 , . . . , n ) (7.26)
8 In this section, we continue to essentially follow the approach of [1]. A slightly alternative approach
0 0 1
We have seen that, to secure asymptotic convergence of the observation error e(t) to zero
is necessary to increase . This, even if the initial conditions z(0) and z (0) of the plant and
of the observer are taken in a compact set, may entail large values of e(t) for (positive)
times close to t = 0. In fact,
and
e(0)
grows unbounded with increasing . Since e(t) is a continuous function of t,
we should expect that, if is large, there is an initial interval of time on which
e(t)
is
large.9
Now, note that feeding the system (7.23) with a control u = u (z ) would result in a
system
z = f (z, u (z n D e)).
This is viewed as a system with state z subject to an input n D e(t). Now, if is large,
the matrix n D remains bounded, because all elements of this (diagonal) matrix are
non positive powers of . In fact
n D
= 1 if 1, as an easy calculation shows.
However, has remarked above,
e(t)
may become large for small values of t, if is large.
Since the system is nonlinear, this may result in a finite escape time.10 To avoid such
inconvenience, as a precautionary measure, it is appropriate to saturate the control, by
choosing instead a law of the form
u = g (u (z ))) (7.27)
to range, so as to prevent finite escape times, has been originally suggested by Khalil in [4, 5], see
also [6].
7.5 A Nonlinear Separation Principle 223
z = f (z, g (u (z )))
(7.28)
z = f (z , g (u (z ))) + D G 0 (h(z, g (u (z ))) h(z , g (u (z )))).
Replacing z by its expression in terms of z and e, we obtain for the first equation a
system that can be written as
z = f (z, g (u (z n D e)))
= f (z, u (z)) f (z, u (z)) + f (z, g (u (z n D e)))
= F(z) + (z, e)
in which
F(z) = f (z, u (z))
(z, e) = f (z, g (u (z n D e))) f (z, u (z)).
Note that, by the inverse Lyapunov Theorem, since the equilibrium z = 0 of (7.24) is
globally asymptotically stable, there exists a smooth function V (z), satisfying
(
z
) V (z) (
z
) for all z,
and
V
F(z) (
z
) for all z,
z
c = {z Rn : V (z) c} B R ,
= max u (z) + 1.
zc+1
Since e enters in (z, e) through the bounded function g (), it is easy to realize that
there is a number 1 such that
(z, e)
1 , for all z c+1 and all e Rn
g (u (z n D e)) = u (z n D e) (7.29)
224 7 Nonlinear Observers and Separation Principle
n D
= 1 and hence
e
small implies
n D e
small. Therefore, there are num-
bers 2 , such that
(z, e)
2
e
, for all z c+1 and all
e
.
These numbers 1 , 2 , are independent of (so long as 1) and only depend on the
number R that characterizes the radius of the ball B R in which z(0) is taken.
Let z(0) B R c . Regardless of what e(t) is, so long as z(t) c+1 , we have
V V
V (z(t)) = [F(z) + (z, e)] (
z
) +
1 .
z z
Setting
V
M = max
zc+1 z
V (z(t)) M1
V (z(t)) V (z(0)) + M1 t c + M1 t.
From this it is deduced that z(t) remains in c+1 at least until time T0 = 1/M1 . This
time may be very small but, because of the presence of the saturation function g (), it is
independent of . It rather only depends on the number R that characterizes the radius
of the ball B R in which z(0) is taken.
Recall now that the variable e satisfies the estimate established in Sect. 7.3. Letting
V (e) denote the quadratic form V (e) = eT Se, we know that
V (e(t)) 2 V (e(t))
in which
1 1
= 2
S
L n := ( a0 )
2 2
is a number that can be made arbitrarily large by increasing (recall that and
S
only
depend of the bounds and in Assumption (ii) and L on the Lipschitz constants in
Assumption (i)). From this inequality, bearing in mind the fact that
a1
e
2 V (e) a2
e
2
7.5 A Nonlinear Separation Principle 225
e(0)
2 n1 R.
Consequently
e(t)
2 A R e t n1 ,
e(T0 )
.
e(t)
2 A R e (tT0 ) e T0 n1 ,
V V
V (z(t)) = [F(z) + (z, e)] (
z
) +
(z, e)
.
z z
Pick , so that
e(t)
for all t T0 and hence, so long as z(t) c+1 ,
V (z(t)) (
z(t)
) + M2 .
Let r be
r = min
z
.
zSdc+1
226 7 Nonlinear Observers and Separation Principle
By construction
(
z
) (r ) for all z Sdc+1 .
If is small enough
1
M2 (r ),
2
and hence
1
V (z(t)) (r ),
2
so long as z(t) Sdc+1 . By standard arguments,12 this proves that any trajectory z(t)
which starts in B R , in a finite time (which only depends on the choice of R and d), enters
the set d and remains in this set thereafter. Observing that for any (small) there is
a number d such that d B , it can be concluded that, for any choice of R
there exist a number and a time T such that, if > , all trajectories with initial
conditions (z(0), z (0)) B R B R are bounded and satisfy
Moreover, limt e(t) = 0. If is small enough, on the set B B , the first equation
of (7.28) becomes (see (7.29))
Thus, using Lemma B.4 of Appendix B, it is concluded that also limt z(t) = 0. This
proves the following result.
It remains to discuss the role of the Assumptions (i) and (ii). Having proven that the
trajectories of the system starting in B R B R remain in a bounded region, it suffices
to look for numbers and and a Lipschitz constant L making Assumptions (i) and
(ii) valid only on this bounded region. To this end, recall that the parameters c and
considered in the proof of the above proposition only depend on the number R and on
the Lyapunov function V (z) of (7.24), and not on the observer. Let these parameters be
fixed as in the previous proof, define
and let f c (z, u) and h c (z, u) be smooth functions, expressed in observability canonical
form, satisfying
f c (z, u) = f (z, u), (z, u) Sc,
h c (z, u) = h(z, u), (z, u) Sc, .
agrees with system (7.23) on the set Sc, . The assumption that system (7.23) is expressed
in uniform observability canonical form implies that properties (7.3) hold for such system.
Thus, since Sc, is a compact set, it is always possible to define f c (z, u) and h c (z, u), at
points (z, u)
/ Sc, , in such a way that Assumptions (i) and (ii) are fulfilled.13 Note also
that the system
z = f c (z, u (z))
coincides with system (7.24) for all z c+1 . Thus, in particular, the Lyapunov function
V (z) introduced in the previous discussion satisfies
V
f c (z, u (z)) (
z
) z c+1 .
z
From this, repeating the proof of Proposition 7.3, one reaches the conclusion that if system
(7.30) is controlled by
and initial conditions are in B R B R , there is a number such that if > all
trajectories are bounded and converge to the equilibrium (z, z ) = (0, 0). This controller
generates an input always satisfying
u(t)
2 and induces in (7.30) a trajectory which
always satisfies z(t) c+1 . Since (7.30) and the original plant agree on the set Sc, , the
controller constructed in this way achieves the same result if used for the original plant.
7.6 Examples
x1 = x3 + (x1 + x13 + x2 )3 u
x2 = x1 3x12 x3 + (x2 + x13 )u 3x12 (x1 + x13 + x2 )3 u
x3 = (x2 + x13 )2
y = x2 + x13 .
1 (x, u) = x2 + x13
2 (x, u) = x1 + (x2 + x13 )u
3 (x, u) = x3 + (x1 + x13 + x2 )3 u + (x1 + (x2 + x13 )u)u.
Moreover
x2 + x13
z = n (x, 0) = x1
x3
z 1 = z2 + z1 u
z 2 = z 3 + (z 2 + z 1 )3 u
z 3 = z 12
y = z1 ,
x1 = x2 + x1 x2 u
x2 = x3
x3 = (x2 + x13 )2 + u
y = x1
This system is not transformable into a system in uniform observability canonical form.
In fact, we have
1 (x, u) = x1
2 (x, u) = x2 + x1 x2 u
and
1 (x, u)
1 0 0
2(x,
x = .
u) x2 u 1 + x1 u 0
x
Hence
0 0
K1 (x, u) = Im 1 0
0 1
but
0
K2 (x, u) = Im 0 if 1 + x 1 u = 0
1
0 0
K2 (x, u) = Im 1 0 if 1 + x1 u = 0
0 1
and the uniformity conditions are not fulfilled. The transformation is not possible.
Example 7.4 Consider again the system (6.42) of Example 6.4. This system can be
transformed into uniform observability canonical form. Note that
1 (x, u) = x2
2 (x, u) = x2 + x3
3 (x, u) = x1 + x2 + x22 + x3 + u
z 1 = z2
z 2 = z3 + u
(7.32)
z 3 = f 3 (z 1 , z 2 , z 3 ) + u
y = z1
in which
f 3 (z 1 , z 2 , z 3 ) = z 1 z 2 + 2z 3 z 12 + 2z 1 z 2 (z 3 z 2 z 12 )3 .
z 1 = z 2 + c2 (y z 1 )
z 2 = z 3 + 2 c1 (y z 1 ) + u (7.33)
z 3 = f 3 (z 1 , z 2 , z 3 ) + 3 c0 (y z 1 ) + u.
Now, recall that the equilibrium x = 0 of system (6.42) can be globally asymptotically
stabilized by means of a state feedback law u = (x), whose expression was determined
in Example 6.4. Passing to the z coordinates, the equilibrium z = 0 of (7.32) is globally
asymptotically stabilized by the feedback law u (z) = ( 1 (z)). In view of the results
presented in Sect. 7.5, it is concluded that a dynamic output feedback consisting of (7.33)
and of
u = g (( 1 (z)))
is able to asymptotically stabilize the equilibrium (z, z ) = (0, 0) of the associated closed-
loop system, with a region of attraction that contains an arbitrarily fixed compact set.
References
1. J.P. Gauthier, I. Kupka, Deterministic Observation Theory and Applications (Cambridge Uni-
versity Press, Cambridge, 2001)
References 231
2. A.R. Teel, L. Praly, Global stabilizability and observability imply semi-global stabilizability by
output feedback. Syst. Control Lett. 22, 313325 (1994)
3. A.R. Teel, L. Praly, Tools for semi-global stabilization by partial state and output feedback.
SIAM J. Control Optim. 33, 14431488 (1995)
4. F. Esfandiari, H.K. Khalil, Output feedback stabilization of fully linearizable systems. Int. J.
Control 56, 10071037 (1992)
5. H.K. Khalil, F. Esfandiari, Semiglobal stabilization of a class of nonlinear systems using output
feedback. IEEE Trans. Autom. Control 38, 14121415 (1993)
6. A.N. Atassi, H.K. Khalil, A separation principle for the stabilization of a class of nonlinear
systems. IEEE Trans. Autom. Control 44, 16721687 (1999)
Chapter 8
The Small-Gain Theorem for Nonlinear
Systems and Its Applications to Robust
Stability
In Sect. 3.5, we have determined conditions under which the pure feedback intercon-
nection of two stable linear systems is stable. As a matter of fact, we have shown
(see Theorem 3.2 and also Corollary 3.1) that the interconnection of two stable linear
systems is stable if the product of the respective L2 gains is small.1 In this section,
we discuss a similar property for the pure feedback interconnection of two nonlinear
systems. Instrumental, in this analysis, is the notion of input-to-state stability and
the associated notion of gain function.2
Consider a nonlinear system modeled by equations of the form
x 1 = f 1 (x1 , x2 )
(8.1)
x 2 = f 2 (x1 , x2 ),
in which x1 Rn1 , x2 Rn2 , and f 1 (0, 0) = 0, f 2 (0, 0) = 0 (see Fig. 8.1). This is
seen as interconnection of a system 1 with internal state x1 and input x2 and of a
system 2 with internal state x2 and input x1 .
We assume that both 1 and 2 are input-to-state stable. According to Defini-
tion B.1, this means that there exists two class K L functions 1 (, ), 2 (, ) and
two class K functions 1 (), 2 () such that, for any bounded input x2 () and any
x1 (0) Rn1 , the response x1 (t) of
x 1 = f 1 (x1 , x2 )
for all t 0, and for any bounded input x1 () and any x2 (0) Rn2 , the response x2 (t)
of
x 2 = f 2 (x1 , x2 )
for all t 0.
In what follows we shall prove that if the composite function 1 2 () satisfies3
3 A function : [0, ) [0, ) satisfying (0) = 0 and (r) < r for all r > 0 is called a simple
and let T be such that (8.5) holds. Consider, for system 1 , an input x 2 () defined
as x 2 (t) = x2 (t) for t [0, T ] and x 2 (t) = 0 for t > T . This input is bounded. Using
(8.2), it is seen that
Likewise, consider for system 2 an input x 1 () defined as x 1 (t) = x1 (t) for t [0, T ]
and x 1 (t) = 0 for t > T . This input is bounded. Using (8.3) it is seen that
Let us now replace the estimate (8.8) into (8.7) and observe that, if a max
{b, c, (a)} and (a) < a, then necessarily a max{b, c}. This, using the hypothesis
that 1 2 (r) < r, yields
The hypothesis (8.4) also implies that 2 1 (r) < r for r > 0, and therefore an
argument identical to the one used before shows that
4 Related
results can be found in [6, 7, 10]. See also [24] and [11] for earlier versions of a Small-
Gain Theorem for nonlinear systems.
236 8 The Small-Gain Theorem for Nonlinear Systems and Its Applications
combining which, and using the property that 1 2 () is a simple contraction, one
obtains
x1 (t) max{1 (x1 , 0), 1 2 (x2 , 0)}
x2 (t) max{2 (x2 , 0), 2 1 (x1 , 0)},
for all t 0. This, since i (r, 0) and i (r) are continuous functions vanishing at
r = 0, shows that the equilibrium (x1 , x2 ) = (0, 0) of (8.1) is stable in the sense of
Lyapunov.
To prove asymptotic stability, knowing that x2 (t) is bounded on [0, ), let
1 (r2 + h) = 1 (r2 ) + .
By definition of r2 , there is a time T > 0 such that x2 (t) r2 + h for all t T .
Using (8.2), observe that
Combining (8.12) with (8.13) and using the fact that 1 2 () is a simple contraction,
we obtain
lim sup x1 (t) 1 2 (lim sup x1 (t)) < lim sup x1 (t)
t t t
The condition (8.4), i.e., the condition that the composed function 1 2 () is a
simple contraction, is the nonlinear analogue of the condition determined in Theo-
rem 3.2 and Corollary 3.1 for the analysis of the stability of the pure feedback inter-
connection of two stable linear systems. As it was the case in the context of linear
systems, the condition in question is only a sufficient condition for global asymptotic
stability. From a practical viewpoint, the result of Theorem 8.1 can be implemented
as follows. First of all, one has to check that both 1 and 2 are input-to-state stable
and find estimates of the two gain functions 1 () and 2 (). This can be obtained by
seeking the existence (see Sect. B.2 in Appendix B) of an ISS-Lyapunov function
for each system, i.e., a pair of positive definite and proper functions V1 (x1 ), V2 (x2 )
such that, for some 1 (), 2 () and 1 (), 2 ()
V1
x1 1 (x2 ) f 1 (x1 , x2 ) 1 (x1 )
x1
and
V2
x2 2 (x1 ) f 2 (x1 , x2 ) 2 (x2 ).
x2
If such functions can be found, in fact, 1 and 2 are input-to-state stable and, as
shown in Theorem B.3, estimates of the gain functions can be obtained.
Example 8.1 Consider the pure feedback interconnection of two systems 1 and
2 , modeled by equations of the form
x 1 = x1 + x22
1 : x 2 = x1 x2 x23 + x2 (u1 ) (8.14)
y1 = x2 ,
1 = 3 + 2 u2
2 : (8.15)
y2 = 1 .
2 (r) = 2 r,
The function (x12 + x24 ) is positive definite and proper. Hence, there exists a class K
function (r) such that
x12 + x24 (x)
(r) = r 4 for 0 r 1
2 (8.16)
(r) = r2 1
4 for r 1
2
1 1
V (x) + (|u1 |),
2 2
from which it can be deduced that also system 1 is input-to-state stable. In fact,
with (r) defined as
(r) = 1 (1 (r))
1 1
x (|u1 |) V (x).
21
So long as an estimate of the gain is concerned, bearing in mind the fact that V (x) =
1
2
x2 and using Theorem B.3, we obtain
1 (r) = (r).
in which, we recall, 1 and 2 are larger than 1 (but otherwise arbitrary). Thus, it
is concluded that a sufficient condition for the interconnected system to be asymp-
totically stable is the existence of a class K function () and a pair of numbers
1 > 1, 2 > 1 such that the function (r) in (8.14) satisfies
1 1
max{ (r), (r)} <
2 2
r for all r > 0,
1 2
The Small-Gain Theorem for input-to-state stable systems can be used in various
ways when dealing with the problem of globally asymptotically stabilizing a nonlin-
ear system. The simplest application is found in the context of the stabilization prob-
lem discussed in Sect. 6.5, namely the problem of globally asymptotically stabilizing,
240 8 The Small-Gain Theorem for Nonlinear Systems and Its Applications
using output feedback, a system having relative degree 1. Consider again system
(6.27), rewritten here for convenience
z = f 0 (z, )
= q(z, ) + b(z, )u (8.17)
y = ,
f 0 (0, 0) = 0
q(0, 0) = 0.
Assume also that the system is strongly minimum phase, which means that system
z = f 0 (z, ) (8.18)
viewed as a system with state z and input is input-to-state stable. Suppose a control
u = k( ) can be found such that
viewed as a system with state and input z is input-to-state stable. If this were the
case and if the two gain functions of (8.18) and (8.19), which we denote as z () and
(), respectively, were such that
it would follow from Theorem 8.1 that the resulting closed-loop system, namely
system (8.17) controlled by u = k( ), is globally asymptotically stable.
In this context, it is observed that the gain function z () associated with (8.18) is
fixed, i.e., there is not control on it. Thus, the proposed design strategy is applicable
if it is possible to find a control u = k( ) such that system (8.19) is rendered input-
to-state stable, with a gain function () that respects the constraint (8.20). As a
matter of fact, this is actually possible, as shown below.
For convenience, let the lower subsystem of (8.17) be rewritten in more general
terms as
x = q(z, x) + b(z, x)u, (8.21)
Let () be a class K function. Then there exists a strictly increasing function k(x),
with k(0) = 0 and k(x) = k(x) which is continuously differentiable everywhere
except at x = 0, where it is only continuous, such that the system
viewed as a system with state x and input z is input-to-state stable, with gain function
(). If, in addition, 0 () and 1 1 () are locally Lipschitz at the origin, the
result holds with a function k(x) which is continuously differentiable everywhere.
Proof Choose
1
k(x) = (x + (x))
b0
If the two functions 0 () and 1 1 () are locally Lipschitz at the origin, a con-
tinuously differentiable function () exists that satisfies (8.23).
Now, consider, for the controlled system (8.22), the candidate ISS-Lyapunov
function V (x) = x 2 and observe that7
V b(z, x)
[q(z, x) + b(z, x)k(x)] = 2x q(z, x) (x + (x))
x b0
2|x| + 2|x|[max{0 (|x|), 1 (z)} (|x|)].
2
the right-hand side of the previous inequality is bounded by 2|x|2 . To this end,
observe that (8.23) implies
(r) 1 (r),
V
|x| (z) [q(z, x) + b(z, x)k(x)] 2|x|2 . (8.25)
x
Bearing in mind Theorem B.3 and the fact that, in this case, (r) = (r) = r 2 , it is
concluded that system (8.22) is input-to-state stable, with gain function ().
This result provides the desired answer to the question posed at the beginning,
i.e., to what extent it is possible to find a control u = k( ) that renders system (8.19)
input-to-state stable, with a gain function () that respects the constraint (8.20). To
this end, in fact, it suffices to pick any function () satisfying
Proposition 8.1 Consider a system of the form (8.17). Suppose the system is strongly
minimum phase, with gain function z (). Suppose there exists a number b0 > 0 and
two class K functions 0 () and 1 () such that
b(z, ) b0
for all (z, ) Rn1 R.
|q(z, )| max{0 (| |), 1 (z)}
The design method outlined in the proof of Lemma 8.1 is applied in the following
simple example.
Example 8.2 In the light of the result obtained above, we revisit Example 6.3, namely
the system
z = z3 +
= z + u, (8.26)
V
[z3 + ] = z4 + z.
z
1
z4 + |z| | | < z4 .
2
Such constraint holds if
1 4
z > |z| | |,
2
which is the case if
1
|z| (2| |) 3 .
Thus, the gain function of the upper subsystem of (8.26) can be estimated by
1
z (r) = (2r) 3 .
1 3
(r) = r .
4
We proceed now with the design of the function k( ) according to the construction
shown in the proof of Lemma 8.1. In the present case, 0 (r) = 0 and 1 (r) = r.
Thus, we can pick the function () as
Note that the function k( ) is only continuous, and not even locally Lipschitz, at
= 0.
This result should be compared with Example 6.3, in which it was shown that,
using a linear control law u = k , system (8.26) could only be semiglobally and
practically stabilized. The current example shows that the system in question can
actually be globally asymptotically stabilized, but at the price of using a nonlinear
control law u = k( ), which moreover happens to be only continuous at = 0.
The Small-Gain Theorem for input-to-state stable systems can be used to the purpose
of robustly stabilizing a system in the presence of model uncertainties in the same way
as the Small-Gain Theorem for linear systems was used to this purpose in Sect. 3.5.
This is the case when the system to be controlled can be seen as interconnection of
an accurately modeled system
244 8 The Small-Gain Theorem for Nonlinear Systems and Its Applications
x = f (x, v, u)
z = k(x, v, u)
y = h(x, v)
x p = f p (xp , z)
v = hp (xp , z).
x c = f c (xc , y)
u = hc (xc , y),
that forces the fulfillment of the small-gain condition (see Fig. 8.3).
The difference with the case studied in Chap. 4, though, is that general methods
such as the method described in Sect. 3.6 are not available, and one must proceed
on a case-by-case basis. In what follows, we describe a simple application, which
essentially extends to the case of a multi-input system, the method for gain assignment
described in the previous section. Consider a system modeled as in
x p = f p (xp , x)
(8.27)
x = q(xp , x) + b()u
with state x and input xp , is input-to-state stable with a gain function () such that
the composition p () is a simple contraction.
Note that, in this example of application, in the controlled part of the system it is
assumed that state and input have the same number m of components, which indeed
renders the design problem substantially simpler. It is also assumed that the full state
x is available for feedback, in which case a memoryless control u = kc (x) is sought.
In this setting, the matter then is to find a function kc (x) such that the system
for some b0 . Then there exist a nonsingular diagonal m m matrix E, whose entries
are equal either to 1 or to 1, a symmetric positive definite m m matrix S() of
continuous functions of and positive numbers a1 , a2 satisfying
Proof It is well known that a nonsingular m m matrix b() admits the decompo-
sition
b() = (I + L1 ())D()(I + L2T ()) (8.32)
in which
2 m
D = diag 1 , ,..., ,
1 m1
8 Note that the resulting system is a special case of the system in Fig. 8.2, namely the interconnection
of
x = q(v1 , x) + b(v2 )u z = x y = x
and
x p = f p (xp , z) v1 = xp
= 0 v2 =
with control u = hc (y).
9 Forsimilar results, see also [5, 8].
246 8 The Small-Gain Theorem for Nonlinear Systems and Its Applications
and both L1 () and L2 () are strictly lower triangular matrices, whose entries are
rational functions of the entries of b(). Since ranges over a compact set and all
|i ()|s are bounded from below, the diagonal elements dii () of D() satisfy
d0 |dii ()| d1 ,
for some pair 0 < d0 < d1 . Having defined D+ () = diag(|d11 ()|, . . . , |dmm ()|),
the matrix D() can be written as D() = D+ ()E in which E is a diagonal matrix
whose elements can only be 1 or 1. Then, consider the matrix
Proof Since
k m u12 ()k m1 . . . u1m ()k
0 k m1 . . . u2m ()k
[I + U()]Gk =
...
0 0 ... k
it is seen that
8.3 An Application to Robust Stability 247
2k m k u12 ()k m1 . . . u1m ()k
m1 2k m1 k . . . u2m ()k
[I + U()]Gk + GT T u12 ()k .
k [I + U()] kI = ...
u1m ()k u2m ()k ... k
If this matrix is positive definite, the lemma is proven. To see that this is the case, we
compute the leading principal minors. The first leading principal minor is
M1 = 2k m k
Clearly, there exist k1 > 0 such that, for any k > k1 , this number is positive. The
second leading principal minor is
The coefficient of the highest power of k is positive and ranges over a compact
set. Hence, there exists k2 > 0 such that if k > k2 , then M2 > 0 for all . Proceeding
in this way, we can find a set of positive numbers k1 , k2 , . . . , km , and a number
k0 = max{k1 , k2 , . . . , km } such that, if k > k0 ,
Lemma 8.4 Consider system (8.28). Suppose b() is such that the assumptions of
Lemma 8.2 hold. Suppose also that there exists two class K functions 0 () and
1 () such that
q(xp , x) max{0 (x), 1 (xp )} for all (xp , x).
V V
x = [q(xp , x) ES()(I + U())K(x)Ex]
x x
= 2x T E T S 1 ()Eq(xp , x) 2x T E T (I + U())G(x) Ex
2x T E T S 1 ()Eq(xp , x) (x)Ex2
2x S 1 () q(xp , x) (x)x2 .
(r) 2S 1 () max{0 (r), 1 1 (r)} for all r 0 and all M.
Since 0 () and 1 1 () are locally Lipschitz at the origin, one can assume that
such () is continuously differentiable at the origin and therefore the limit of (r)/r
as r 0 + exists, is finite and positive. Define (r) as
(r)
(r) = k + ,
r
V
x k0 x2 + x 2S 1 () max{0 (x), 1 (xp )} (x)
x
from which arguments identical to those used in the proof of Lemma 8.1 prove that
V
x (xp ) x k0 x2 .
x
8.3 An Application to Robust Stability 249
Proposition 8.2 Consider system (8.27). Suppose the upper subsystem, viewed as a
system with state xp and input x, is input-to-state stable, with a gain function p ().
Suppose the assumptions of Lemma 8.2 hold. Suppose also that there exists two class
K functions 0 () and 1 () such that
q(xp , x) max{0 (x), 1 (xp )} for all (xp , x),
and 0 () and 1 p () are locally Lipschitz at the origin. Then there exists a contin-
uous function (r) such that system (8.27) with control u = G(x) Ex is globally
asymptotically stable.
References
1. J.M. Coron, L. Praly, A.R. Teel, Feedback stabilization of nonlinear systems: sufficient con-
ditions and Lyapunov and inputoutput techniques, in Trends in Control, ed. by A. Isidori
(Springer, London, 1995), pp. 293348
2. D. Hill, P.J. Moylan, The stability of nonlinear dissipative systems. IEEE Trans. Autom. Control
AC21, 708711 (1976)
3. D. Hill, P.J. Moylan, Stability results for nonlinear feedback systems. Automatica 13, 377382
(1977)
4. D. Hill, P.J. Moylan, Connections between finite gain and asymptotic stability. IEEE Trans.
Autom. Control AC25, 931935 (1980)
5. A.K. Imai, R.R. Costa, L. Hsu, G. Tao, P.V. Kokotovic, Multivariable adaptive control using
high-frequency gain matrix factorization. IEEE Trans. Autom. Control 49(7), 11521157
(2004)
6. Z.P. Jiang, A.R. Teel, L. Praly, Small-gain theorem for ISS systems and applications. Math.
Control Signal Syst. 7, 95120 (1994)
7. Z.P. Jiang, I.M.Y. Mareels, Y. Wang, A Lyapunov formulation of the nonlinear-small gain
theorem for interconnected ISS systems. Automatica 32, 12111215 (1996)
8. A.S. Morse, A gain matrix decomposition and some of its applications. Syst. Control Lett. 21,
110 (1993)
9. E.D. Sontag, On the input-to-state stability property. Eur. J. Control 1, 2436 (1995)
10. A. Teel, A nonlinear small gain theorem for the analysis of control systems with saturations.
IEEE Trans. Autom. Control AC41, 12561270 (1996)
11. M. Vidyasagar, Decomposition techniques for large-scale systems with nonadditive interac-
tions: stability and stabilizability. IEEE Trans. Autom. Control AC25, 773779 (1980)
12. L. Wang, A. Isidori, H. Su, Global stabilization of a class of invertible MIMO nonlinear systems.
IEEE Trans. Autom. Control 60, 616631 (2015)
Chapter 9
The Structure of Multivariable Nonlinear
Systems
9.1 Preliminaries
In this chapter, we shall see how the theory developed for single-input single-output
systems can be extended to nonlinear systems having many inputs and many outputs,
usually referred to as multivariable systems. To simplify matters, we shall restrict
our analysis to the consideration of systems having the same number m of input and
output channels. The multivariable nonlinear systems we consider are described in
state-space form by equations of the following kind
m
x = f (x) + gi (x)ui
i=1
y1 = h1 (x)
ym = hm (x)
in which f (x), g1 (x), . . . , gm (x) are smooth vector fields, and h1 (x), . . . , hm (x)
smooth functions, defined on Rn . Most of the times, these equations will be rewritten
in the more compact form
x = f (x) + g(x)u
(9.1)
y = h(x)
having set
u = col(u1 , . . . , um )
y = col(y1 , . . . , ym )
and where
g(x) = (g1 (x) . . . gm (x))
h(x) = col(h1 (x), . . . , hm (x))
x = f(x) + g (x)v
(9.4)
y = h(x),
which has the same structure as (9.1). As a nontriviality condition, it will be always
assumed that (x) is nonsingular for all x.
1 The algorithm in question, originally introduced by Silverman in [4] to the purpose of analyzing
the structure of the zeros at infinity of the transfer function matrix of a multivariable linear system,
was extended in [5, 6] to the purpose of analyzing the property of invertibility of a nonlinear system.
Additional features of such algorithm are discussed in [7] and [10]. Related results on nonlinear
invertibiliy can be found in [8, 9].
9.2 The Basic Inversion Algorithm 253
normal forms determined earlier in Chap. 6 and in the study of the problem of left
invertibility. The term left invertibility, generally speaking, denotes the possibility
of uniquely recovering the input of a system from the knowledge of its output and of
its internal state. For a single-input single-output system, a sufficient condition for
left invertibility is that the system has uniform relative degree. In this case in fact, it
is possible to write
1
u(t) = [y(r) (t) L rf h(x(t))].
Lg L r1
f h(x(t))
In the case of a multivariable system, though, the issue is much more involved and a
property such as the one described above can only be decided at the end of a sequence
of recursive calculations, described in what follows, known as nonlinear Structure
Algorithm or also as nonlinear Inversion Algorithm A byproduct of such calculations
is also the possibility of determining functions of x that can be used for a change of
coordinates in the state space, by means of which the system is brought to a form that
can be considered as the multivariable version of the normal form of a single-input
single-output system.
Step 1. Set 0 = 0, consider the function
S0 (y, x) := y + h(x),
and define
S0 (1) S0
S 0 (y(1) , x, u) := y + [ f (x) + g(x)u].
y x
The function S 0 (y(1) , x, u) thus defined is an affine function of u and can be written
as
S 0 (y(1) , x, u) = L1 (y(1) , x) + M1 (x)u
in which
S0 S0
L1 (y(1) , x) = y(1) + f (x) M1 (x) = g(x).
x x
Assume that M1 (x) has constant rank 1 and that there exists a fixed set of 1 rows
(empty if 1 = 0) that are linearly independent for all x. After a permutation of
rows in S0 (y, x) if necessary, let such rows be the first 1 rows of M1 (x). Partition
L1 (y(1) , x) and M1 (x) as
254 9 The Structure of Multivariable Nonlinear Systems
L1 (y(1) , x) M1 (x)
L1 (y(1) , x) = M1 (x) = (9.5)
L1 (y(1) , x) M1 (x)
where in both cases the upper blocks have 1 rows (and hence the lower blocks have
m 1 rows). Consistently, let S 0 (y(1) , x, u) be partitioned as
(1)
S (y , x, u)
S 0 (y(1) , x, u) = 0 (1) .
S0 (y , x, u)
Define
S1 (y(1) , x) := L1 (y(1) , x) + F1 (x)L1 (y(1) , x),
S1 S1
S 1 (y(1) , y(2) , x, u) := (1) y(2) + [ f (x) + g(x)u].
y x
The function S 1 (y(1) , y(2) , x, u) thus defined is an affine function of u and can be
written as
S 1 (y(1) , y(2) , x, u) = L2 (y(1) , y(2) , x) + M2 (y(1) , x)u
in which
S1 (2) S1 S1
L2 (y(1) , y(2) , x) = y + f (x) M2 (y(1) , x) = g(x).
y(1) x x
Define2
M1 (x)
J2 (y(1) , x) := (9.8)
M2 (y(1) , x)
and assume that this (square) matrix has constant rank 2 and that there exists a fixed
set of 2 1 rows (empty if 2 = 1 ) of M2 (y(1) , x) that, together with the first 1
rows (that are linearly independent by assumption) form a linearly independent set
for all (y(1) , x). After a permutation of rows in S1 (y(1) , x) if necessary, let such rows
be the first 2 1 rows of M2 (y(1) , x). Partition L2 (y(1) , y(2) , x) and M2 (y(1) , x) as
(1) (2) L2 (y(1) , y(2) , x) (1) M2 (y(1) , x)
L2 (y , y , x) = M2 (y , x) = (9.9)
L2 (y(1) , y(2) , x) M2 (y(1) , x)
where in both cases the upper blocks have 2 1 rows (and hence the lower blocks
have m 2 rows). Consistently, let S 1 (y(1) , y(2) , x, u) be partitioned as
S 1 (y(1) , y(2) , x, u)
S 1 (y(1) , y(2) , x, u) = .
S 1 (y(1) , y(2) , x, u)
As a consequence of the assumption on the matrix (9.8), if 2 > 1 > 0 there exist
a (m 2 ) 1 matrix F21 (y(1) , x) and a (m 2 ) (2 1 ) matrix F22 (y(1) , x)
such that
M2 (y(1) , x) = F21 (y(1) , x)M1 (x) F22 (y(1) , x)M2 (y(1) , x). (9.10)
Define
If 2 > 1 but 1 = 0, the term F21 (y(1) , x)M1 (x) is missing in (9.10) and so is
the term F21 (y(1) , x)L1 (y(1) , x) in the definition of S2 (y(1) , y(2) , x). Moreover, for-
mula (9.11) must be adapted in a obvious manner, removing the terms S 0 (y(1) , x, u),
L1 (y(1) , x) and M1 (x) that do not exist. If 2 1 = 0 and 1 > 0, the upper blocks
in the partition of L2 (y(1) , y(2) , x) and M2 (y(1) , x) do not exist. Consequently, (9.10)
is rewritten as
M2 (y(1) , x) = F21 (y(1) , x)M1 (x)
and
256 9 The Structure of Multivariable Nonlinear Systems
Step k+1. Consider the function Sk (y(1) , . . . , y(k) , x) defined at step k and define
Sk Sk Sk
S k (y(1) , . . . , y(k+1) , x, u) := (1) y(2) + + (k) y(k+1) + [ f (x) + g(x)u].
y y x
The function S k (y(1) , . . . , y(k+1) , x, u) thus defined is an affine function of u and can
be written as
in which
Sk (2) Sk Sk
Lk+1 (y(1) , . . . , y(k+1) , x) = (1)
y + + (k) y(k+1) + f (x)
y y x
Sk
Mk+1 (y(1) , . . . , y(k) , x) = g(x).
x
Define
M1 (x)
Jk+1 (y(1) , . . . , y(k1) , y(k) , x) :=
Mk (y(1) , . . . , y(k1) , x)
(9.12)
Mk+1 (y(1) , . . . , y(k1) , y(k) , x)
and assume that this (square) matrix has constant rank k+1 and that there exists a
fixed set of k+1 k rows (empty if k+1 = k ) of Mk+1 (y(1) , . . . , y(k) , x) that,
together with the first k rows (that are linearly independent by assumption)
form a linearly independent set for all (y(1) , . . . , y(k) , x). After a permutation of
rows in Sk (y(1) , . . . , y(k) , x) if necessary, let such rows be the first k+1 k
rows of Mk+1 (y(1) , . . . , y(k) , x). Partition Lk+1 (y(1) , . . . , y(k+1) , x) and Mk+1 (y(1) ,
. . . , y(k) , x) as
Lk+1 (y(1) , . . . , y(k+1) , x)
Lk+1 (y(1) , . . . , y(k+1) , x) =
Lk+1 (y(1) , . . . , y(k+1) , x)
Mk+1 (y(1) , . . . , y(k) , x)
Mk+1 (y(1) , . . . , y(k) , x) =
Mk+1 (y(1) , . . . , y(k) , x)
9.2 The Basic Inversion Algorithm 257
where in both cases the upper blocks have k+1 k rows (and hence the lower blocks
have m k+1 rows). Consistently, let S k (y(1) , . . . , y(k+1) , x, u) be partitioned as
S k (y(1) , . . . , y(k+1) , x, u)
S k (y(1) , . . . , y(k+1) , x, u) = .
S (y(1) , . . . , y(k+1) , x, u)
k
As a consequence of the assumption on the matrix (9.12), if k+1 k > 0 there exist
(m k+1 ) (i i1 ) matrices Fk+1,i (y(1) , . . . , y(k) , x), for i = 1, . . . , k + 1 (recall
that 0 = 0), such that
k+1
Mk+1 (y(1) , . . . , y(k) , x) = Fk+1,i (y(1) , . . . , y(k) , x)Mi (y(1) , . . . , y(i1) , x).
i=1
(9.13)
Define
Moreover, formula (9.14) must be adapted in a obvious manner, removing the terms
S i1
(y(1) , . . . , y(i) , x, u), Li (y(1) , . . . , y(i) , x) and Mi (y(1) , . . . , y(i1) , x) that do not
exist. If k+1 k = 0, then the upper blocks in the partition of Lk+1 (y(1) , . . . ,
258 9 The Structure of Multivariable Nonlinear Systems
k
(1) (k)
Mk+1 (y , . . . , y , x) = Fk+1,i (y(1) , . . . , y(k) , x)Mi (y(1) , . . . , y(i1) , x),
i=1
and
where the term Fk+1,i (y(1) , . . . , y(k) , x)Li (y(1) , . . . , y(i) , x) may be missing if i
i1 = 0. In particular, if 1 = = k+1 = 0, then
Definition 9.1 The structure algorithm has no singularities if the matrix J1 (x) :=
M1 (x) has constant rank 1 and there exists a fixed set of 1 rows (empty if 1 = 0) that
are linearly independent for all x and, for all k = 1, 2, . . . , the matrix (9.12) has con-
stant rank k+1 and there exists a fixed set of k+1 k rows of Mk+1 (y(1) , . . . , y(k) , x)
(empty if k+1 = k ) that, together with the first k rows of (9.12), form a linearly
independent set for all (y(1) , . . . , y(k) , x).
Definition 9.2 The system (9.1) is uniformly invertible (in the sense of Singh) if the
structure algorithm has no singularities and there exists a integer k n such that
k = m.
Remark 9.1 Note that, in the above definition, the integer k is taken to be less than or
equal to n, the dimension of the state space of (9.1). This is not a restriction, though.
In fact, it will be shown later (see Proposition 9.2) that, if the structure algorithm has
no singularities and k = m for some k , then this integer k is necessarily less than
or equal to n. Knowing such bound on k it is useful because, if the algorithm is run
up to step n and it is found that n < m, then it is concluded that the system is not
uniformly invertible.
Note that, if the system is invertible in the sense of this definition, it is possible
to write3
3 Clearly,if at step k one finds that k = m, then the partitions of Lk () and Mk () are trivial
partitions, in which the upper blocks coincide with the matrices themselves, and this explains the
notation used in the following formula.
9.2 The Basic Inversion Algorithm 259
S 0 (y(1) , x, u) L1 (y(1) , x)
S 1 (y(1) , y(2) , x, u) L (y(1) , y(2) , x)
= 2
S k 1 (y , . . . , y , x, u)
(1) (k ) (1) (k )
Lk (y , . . . , y , x)
M1 (x)
M2 (y(1) , x)
+
u,
(9.15)
Mk (y(1) , . . . , y(k 1) , x)
and the matrix that multiplies u on the right-hand side is a m m matrix that is
nonsingular for all (y(1) , . . . , y(k 1) , x).
With this in mind, consider again system (9.1), in which we assume that
f (x), g(x), h(x) are smooth, let u : R Rm be a smooth4 input function, let x(0) be
the value of the state at time t = 0, let x(t) denote the value at time t of the resulting
state trajectory and let y(t) = h(x(t)) denote the value at time t of resulting output.
Clearly, for all t for which x(t) is defined,
S0 (y(t), x(t)) = 0.
Since the derivative of S0 (y(t), x(t)) with respect to time coincides, by construction,
with the function S 0 (y(1) (t), x(t), u(t)), it is seen that also
for all t for which x(t) is defined. This, in turn, in view of (9.7) implies that
for all t for which x(t) is defined. Continuing in this way, it is seen that, for any k 2,
the function Sk (y(1) (t), . . . , y(k) (t), x(t)) vanishes for all t for which x(t) is defined
and so does its derivative with respect to time.
Thus, it is concluded from (9.15) that, for all t for which x(t) is defined,
L1 (y(1) (t), x(t)) M1 (x(t))
L (y(1) (t), y(2) (t), x(t)) M2 (y(1) (t), x(t))
0=
2 +
u(t).
(1) (k ) (1) (k 1)
Lk (y (t), . . . , y (t), x(t)) Mk (y (t), . . . , y (t), x(t))
Since the matrix that multiplies u(t) is nonsingular, this relation can be used to
recover explicitly the value of u(t) as a function of the value of x(t) and of the values
of y(t) and its derivatives with respect to time, up to order k . It is for this reason
4 It
suffices to choose a C k function with k large enough so that all derivatives of y(t) appearing in
the developments which follow are defined and continuous.
260 9 The Structure of Multivariable Nonlinear Systems
that a system having the property indicated in Definition 9.2 is considered to be left
invertible.
Remark 9.2 Observe that, in the case of a single-input single output-system, the
structure algorithm has no singularities if and only if the system has uniform relative
degree. In this case,
for all k r 1,
x 1 = x6 + u1
x 2 = x5 + x2 x6 + x2 u1
x 3 = x6 + x1 x6 + x1 u1
x 4 = u2
x 5 = x4
x 6 = x8
x 7 = u3
x 8 = x7 + x5 u2 ,
y1 = x1
y2 = x2
y3 = x3 .
and obtain
y1(1) + x6 1 00
S 0 (y , x, u) = L1 (y , x) + M1 (x)u =
(1) (1) y(1)
2 + x5 + x2 x6 + x2 0 0 u.
y3(1) + x6 + x1 x6 x1 0 0
The matrix
1 00
J1 (x) = M1 (x) = x2 0 0
x1 0 0
9.2 The Basic Inversion Algorithm 261
and
x2
F1 (x) = ,
x1
so that
(1)
(1) (1)
(1) y2 + x5 + x2 x6 x2 (1) y2 + x5 + x2 y1
S1 (y , x) = + (y1 + x6 ) = .
(1)
y3 + x6 + x1 x6 x1 (1)
y + x6 + x1 y
(1)
3 1
The matrix
1 00
M1 (x)
J2 (y(1) , x) = = x2 y1(1) 0 0
M2 (y(1) , x)
y1(1) 0 0
has still rank 2 = 1 = 1. Hence, M2 (y(1) , x) and L2 (y(1) , y(2) , x) do not exist,
(1)
x2 y1 0 0
M2 (y(1) , x) =
y1(1) 0 0
(2)
y2 + x4 + (x5 + x2 x6 )y1(1) + x2 y1(2)
L2 (y(1) , y(2) , x) =
y3(2) + x8 + x6 y1(1) + x1 y1(2)
and
x2 y1(1)
F21 (y(1) , x) = .
y1(1)
262 9 The Structure of Multivariable Nonlinear Systems
This yields
(2) (1) (2)
(1)
y2 + x4 + (x5 + x2 x6 )y1 + x2 y1 x2 y1
S2 (y (1)
,y (2)
, x) = (2) (1) (2)
+ (1)
(y1(1) + x6 ),
y3 + x8 + x6 y1 + x1 y1 y1
i.e.
(2)
(1) (2) y2 + x4 + x5 y1(1) + x2 y1(2) + x2 [y1(1) ]2
S2 (y , y , x) = .
y3(2) + x8 + x1 y1(2) + [y1(1) ]2
The matrix
1 0 0
M (x)
J3 (y(1) , y(2) , x) = 1 = x2 y1(2) + x2 [y1(1) ]2 1 0
M3 (y(1) , y(2) , x)
y1(2) x5 0
and
F31 (y(1) , y(2) , x) = x5 (x2 y1(2) + x2 [y1(1) ]2 ) y1(2)
F33 (y(1) , y(2) , x) = x5 .
S3 (y(1) , y(2) , y(2) , x) = L3 (y(1) , y(2) , y(3) , x) + F31 (y(1) , y(2) , x)L1 (y(1) , x)
+ F33 (y(1) , y(2) , x)L3 (y(1) , y(2) , y(3) , x).
in which the (1 3) row vector M4 (y(1) , y(2) , y(3) , x) has the form
M4 (y(1) , y(2) , y(3) , x) = 1 .
9.2 The Basic Inversion Algorithm 263
has rank 4 = 3 and the algorithm terminates. The system is uniformly invertible.
In the previous section, we have described the structure algorithm in rather general
terms, to the purpose of identifying the required regularity assumptions and to intro-
duce and motivate the property of invertibility. In this section, we revisit the various
steps of the algorithm in more detail, showing that the various functions generated at
each step can be given more explicit expressions, highlightingin particularthe
role of the individual components of the output and of their higher order derivatives
with respect to time.
In what follows, given a vector y Rm , for any pair of indices 1 i j m, we
use y[i,j] to denote the vector
S0
Set H1 (x) = f (x) and split it as
x
H1 (x)
H1 (x) =
H1 (x)
in which5
(1) (1)
Q1 (y[1,1]
, x) = H1 (x) + F1 (x)[y[1,1]
+ H1 (x)].
in which
(1) (2) Q1 (2) Q1
H2 (y[1,1]
, y[1,1]
, x) = (1)
y[1,1]
+ f (x)
y[1, x
1]
(1)
while M2 (y(1) , x) is seen to be independent of y[ 1 +1,m]
, and accordingly rewritten as
(1)
M2 (y[1,1 ] , x). Note that
2 (y[1,
(1) Q1
M 1]
, x) = g(x).
x
All of the above yields
S 1 (y(1) , y(2) , x, u) = y[
(2)
1 +1,m]
(1)
+ H2 (y[1,1]
(2)
, y[1,1]
2 (y[1,
, x) + M (1)
1]
, x)u.
(1)
Let now H2 (y[1, (2)
, y[1, 2 (y[1,
, x) and M (1)
, x) be partitioned as in (9.9), which yields
1] 1] 1]
and
2 (y[1,
M (1)
, x)
(1) 1]
M2 (y , x) = .
2 (y[1,
M (1)
] , x)1
(1)
Since M2 (y(1) , x) is independent of y[ 1 +1,m]
, so are the matrices F21 (y(1) , x)
and F22 (y(1) , x). Letting the latter be denoted, respectively, as F21 (y[1,
(1)
1]
, x) and
(1)
F22 (y[1,1 ] , x), it is concluded that
2 (y[1,
M (1)
1]
, x) = F21 (y[1,
(1)
1]
, x)M1 (x) F22 (y[1,
(1)
1]
2 (y[1,
, x)M (1)
1]
, x)
in which6
(1)
Q2 (y[1,1]
(2)
, y[1,2]
(1)
, x) = H2 (y[1,1]
(2)
, y[1,1]
, x) + F21 (y[1,
(1)
1]
(1)
, x)[y[1,1]
+ H1 (x)]
+ F22 (y[1,
(1)
1]
(2)
, x)[y[1 +1,2 ]
(1)
+ H2 (y[1,1]
(2)
, y[1,1]
, x)].
Continuing in this way, it is seen that L3 (y(1) , y(2) , y(3) , x) can be expressed in the
form
(3) (1) (2) (3)
L3 (y(1) , y(2) , y(3) , x) = y[2 +1,m]
+ H3 (y[1,1]
, y[1,2]
, y[1,2]
, x),
(1) (2)
while M3 (y(1) , y(2) , x) is seen to be independent of y[ 1 +1,m]
and of y[2 +1,m]
, and
(1) (2) (1) (2) (3)
accordingly rewritten as M3 (y[1,1 ] , y[1,2 ] , x). As a consequence S3 (y , y , y , x)
can be expressed in the form
(3) (1) (2) (3)
S3 (y(1) , y(2) , y(3) , x) = y[3 +1,m]
+ Q3 (y[1,1]
, y[1,2]
, y[1,3]
, x).
and
(k) (1) (2) (k)
Sk (y(1) , . . . , y(k) , x) = y[k +1,m]
+ Qk (y[1,1]
, y[1,2]
, . . . , y[1,k]
, x).
Suppose now the system is invertible in the sense of the Definition 9.2. As
explained in the conclusions of the last section, for each k the functions S k1
(y(1) , . . . ,
y(k) , x, u) and Sk (y(1) , . . . , y(k) , x) vanish along the trajectories of system (9.1). There-
fore, for each k
(k) (1) (2) (k1) (k)
y[k1 +1,k ]
= Hk (y[1,1]
, y[1,2]
, . . . , y[1,k1 ]
, y[1,k1 ]
, x)
k (y[1,
+M (1) (2)
, y[1, (k1)
, . . . , y[1, , x)u (9.16)
1] 2] k1 ]
while if 2 = 1 = 0 we have
(1) (2) (1) (2)
Q2 (y[1, 1]
, y[1, 2]
, x) = H2 (y[1, 1]
, y[1, 1]
, x).
266 9 The Structure of Multivariable Nonlinear Systems
and
(k) (1) (2) (k)
y[k +1,m]
= Qk (y[1,1]
, y[1,2]
, . . . , y[1,k]
, x). (9.17)
In particular, writing all (9.16) together, it can be concluded that the identity
(1)
y[1,1]
H1 (x)
y(2) (1)
H2 (y[1, (2)
, y[1, , x)
[1 +1,2 ] 1] 1]
(3) (1) (2) (3)
y[2 +1,3 ] = H3 (y[1,1 ] , y[1,2 ] , y[1,2 ] , x)
..
..
. .
(k ) (1) (k 1) (k )
y[ k1 +1,k ]
H k (y[1,1 ] , . . . , y , y
[1,k 1 ] [1,k 1 ] , x)
M1 (x)
M 2 (y[1,
(1)
, x)
1]
3 (y[1, ] , y[1, ] , x)
(1) (2)
+ M u
1
..
2
.
(y[1,
M (1) (k 1)
, . . . , y[1, , x)
k 1] k 1 ]
holds, in which the m m matrix that multiplies u on the right-hand side is invertible
for all values of the arguments.
Remark 9.3 It is worth stressing the underlying structure involved in the expressions
of the higher order derivatives of y. In fact, the first derivative is expressed as
(1)
(1) y[1,1]
H1 (x) + M1 (x)u
y = (1) = (1) .
y[1 +1,m]
Q1 (y[1, 1]
, x)
and so on.
In this section, it is shown that if a system of the form (9.1) is invertible, in the sense
of Definition 9.2, it is possible to define a special change of coordinates, by means of
which the equations that describe the system can be transformed in equations having
a special structure, that can be regarded as a multivariable version of the normal
forms introduced in Chap. 6. The changes of coordinates in question are based on
9.4 Partial and Full Normal Forms 267
and define
T1 (x)
T2 (x)
(x) =
.
(9.18)
Tk (x)
d = m + (m 1 ) + (m 2 ) + + (m k 1 )
: Rn Rd
x = (x)
Proof The first part of the proposition is a trivial consequence of Proposition 9.1. To
prove the second part, observe that m i 1 for all i = 0, . . . , k 1. Thus,
k 1
k (m i ) = d n.
i=0
Assumption 9.1 System (9.1) is uniformly invertible, in the sense of Definition 9.2,
and there exists a map
Z : Rn Rnd
x z = Z(x)
T1
T 1 (x, u) := [ f (x) + g(x)u]
x
consistently with the partition used in (9.5), recall that the upper block has 1 rows
while the lower block has m 1 rows, and observe that
in which the 1 m matrix M1 (x), by hypothesis, has rank 1 for all x. Observe also
that, by definition,
This formula can be fruitfully used to determine the expression of T 1 (x, u). In fact,
using this formula and property (9.6), it is seen that
This expression relates the derivative with respect to time of the lower block of
the set of new coordinates T1 (x) to the set of new coordinates T2 (x) and to the term
[L1 (0, x) + M1 (x)u] that, as shown above, expresses the derivative with respect to
time of the upper block of the set of new coordinates T1 (x). In other words
The pair of expressions (9.19) and (9.20) hold if 1 > 0. If 1 = 0, we have instead
T 1 (x, u) = L1 (0, x)
and
S1 (0, x) = L1 (0, x)
T2
T 2 (x, u) := [ f (x) + g(x)u]
x
consistently with the partition used in (9.9), recall that the upper block has 2 1
rows while the lower block has m 2 rows, and observe that
T3 (x) = S2 (0, 0, x) = L2 (0, 0, x) + F21 (0, x)L1 (0, x) + F22 (0, x)L2 (0, 0, x).
270 9 The Structure of Multivariable Nonlinear Systems
This formula can be fruitfully used to determine the expression of T 2 (x, u). In fact,
using this formula and property (9.10), it is seen that
Again, this establishes a relation between the derivative with respect to time of the
lower block of the set of new coordinates T2 (x), the set of new coordinates T3 (x)
and the terms [L1 (0, x) + M1 (x)u] and [L2 (0, 0, x) + M2 (0, x)u] that express the
derivatives with respect to time of the upper block of the set T1 (x) and the upper
block of the set T2 (x). In other words
T 2 (x, u) = T3 (x) F21 (0, x)T 1 (x, u) F22 (0, x)T 2 (x, u). (9.23)
The pair of expressions (9.19) and (9.20) hold if 2 > 1 > 0. If 2 1 > 0 but
1 = 0, the term F21 (0, x)T 1 (x, u) in the formula above is missing. If 2 1 = 0
and 1 > 0, we have
and hence
T 2 (x, u) = L2 (0, 0, x) + M2 (0, x)u
= T3 (x) F21 (0, x)L1 (0, x) + M2 (0, x)u
= T3 (x) F21 (0, x)[L1 (0, x) + M1 (x)u],
that is
T 2 (x, u) = T3 (x) F21 (0, x)T 1 (x, u). (9.24)
If 2 = 1 = 0, we have
T 2 (x, u) = T3 (x). (9.25)
Tk (x)
T k (x, u) = [ f (x) + g(x)u]
x
it is found that (see (9.15))
The equations found in this way can be organized in various formats. Let
Xk = Tk (x), for k = 1, . . . , k ,
define the new (sets of) coordinates. The vector Xk has dimension m k1 , in which
k1 the rank of the matrix Jk1 ()is an integer increasing (but not necessarily
strictly increasing) with k (with 1 not necessarily positive). With the sequence of
integers 1 , 2 , . . . we associate a set of integers r1 , r2 , . . . , r , with r = k , defined
as follows. Let r1 be the smallest integer such that r1 > 0. Let r2 > r1 be the smallest
integer such that
r1 = r1 +1 = = r2 1
r2 > r2 1 ,
r 1 > r 1 1
r 1 = r 1 +1 = = r 1
m = r > r 1 .
Counting the number of components of the vectors Xk , it is seen that all r1 vectors
have m 1 components, and so on, ending with the conclusion that the r r 1
vectors
Xr 1 +1 (x), Xr 1 +2 (x), . . . , Xr (x) (9.29)
have m r 1 components.
Then, using arguments identical to those used above to deduce the expressions of
T 1 (x, u) and T 2 (x, u), it is found that the sets of new coordinates X1 , . . . , Xr1 satisfy
equations of the form
X 1 = X2
X 2 = X3
(9.30)
X r1 1 = Xr1
X r1 = a1 (x) + b1 (x)u
X r
1
= Xr1 +1 + Rr1 +1,1 (x)[a1 (x) + b1 (x)u],
272 9 The Structure of Multivariable Nonlinear Systems
the sets of new coordinates Xr1 +1 , . . . , Xr2 satisfy equations of the form
and so on, until it is found that the sets of new coordinates Xr 1 +1 , . . . , Xr satisfy
equations of the form
1
X r 1 +1 = Xr 1 +2 + Rr 1 +2,j (x)[aj (x) + bj (x)u]
j=1
1
X r 1 +2 = Xr 1 +3 + Rr 1 +3,j (x)[aj (x) + bj (x)u]
j=1 (9.32)
1
X r 1 = Xr + Rr ,j (x)[aj (x) + bj (x)u]
j=1
X r = a (x) + b (x)u.
y = X1 .
In these equations, the coefficient matrices Ri,j (x) coincide with the values
at (y(1) , . . . , y(i2) ) = (0, . . . , 0) of the matrices Fi1,rj (y(1) , . . . , y(i2) , x) deter-
mined at the various stages of the structure algorithm. Moreover, the matrices bi (x)
are partitions of Jk (0, . . . , 0, x), that is
b1 (x)
= Jk (0, . . . , 0, x). (9.33)
b (x)
mi := ri ri1
Xj = col(i,j , . . . , ,j ).
in which k,j , for i k , has mk components. The splitting thus defined can be
organized as shown in the following table, where all vectors on ith row have the same
number mi of components,
1,1 1,r1
2,1 2,r1 2,r1 +1 2,r2
3,1 3,r1 3,r1 +1 3,r2 3,r2 +1 3,r3
,1 ,r1 ,r1 +1 ,r2 ,r2 +1 ,r3 ,r 1 +1 ,r
y = col(y1 , y 2 , . . . , y ),
i,1 = i,2
i,r1 1 = i,r1
i,r1 = i,r1 +1 + i,r
1
1 +1
(x)[a1 (x) + b1 (x)u]
i,r2 1 = i,r2 + i,r
1
2
(x)[a1 (x) + b1 (x)u]
i,r2 = i,r2 +1 + i,r
1
2 +1
(x)[a1 (x) + b1 (x)u] + i,r2
2 +1
(x)[a2 (x) + b2 (x)u]
(9.34)
i1
i,ri1
j
= i,ri1 +1 + i,ri1 +1 (x)[aj (x) + bj (x)u]
j=1
i1
i,ri 1 = i,ri +
j
i,ri (x)[aj (x) + bj (x)u]
j=1
i,ri = ai (x) + bi (x)u,
while
y i = i,1 .
It is worth observing that, for any choice of vectors vi Rmi , the set of equations
v = A(x) + B(x)u,
where the matrix B(x) is invertible. In fact (see (9.33)) the matrix B(x) coincides
with Jk (0, . . . , 0, x) and the latter is nonsingular a consequence of the assumption
that the system is invertible. Clearly, the solution u of (9.35) is given by a control of
the form (9.2), in which
To conclude the analysis, it remains to discuss conditions under which the map
Z() considered in Assumption 9.1 exists. We do this in the special case in which all
mi s in the partitions above are equal to 1, i.e., in the special case in which all j,k s
are one-dimensional, and hence = m.8 Using the vector (x) and the matrix (x)
defined in (9.37), define f(x) and g (x) as in (9.3). Then, define recursively a set of
d vector fields as follows
rm rm1
1
Ym1 (x) = g m1 (x) m,r
m1
m i+1
(x)Ymi+1
i=1
i
Ym1 = (1)i1 adi1
f
1
Ym1 (x) for i = 2, . . . , rm1 ,
i=1 i=1
i
Ym2 = (1)i1 adi1
f
1
Ym2 (x) for i = 2, . . . , rm2 ,
and so on.
Then, the following result holds, which is the multivariable version of Proposi-
tion 6.3.9
Yji (x), 1 j m, 1 i rj
Z = {x Rn : (x) = 0} (9.38)
Z : Rn Rnd
x z = Z(x)
Remark 9.4 The partial normal form, among other things, plays an important role
in the problem of finding the input that forces the output to remain identically zero.
In fact, from the set (9.30), in which X1 (t) = y(t), it is seen that if y(t) is identically
zero, so are X1 (t) and X2 (t), . . . , Xr1 (t). Then, from the last two equations of this set,
in which Xr1 (t) and Xr1 (t) are partitions of Xr1 (t), it is seen that
a1 (x) + b1 (x)u = 0,
and also that Xr1 +1 (t) is identically zero. Using these two facts in the set (9.31),
it is seen that Xr1 +2 (t), . . . , Xr2 (t) are all identically zero. Then, from the last two
equations of this set it is seen that
a2 (x) + b2 (x)u = 0,
9 For a proof, see [2] and also [1, pp. 114115]. Related results can also be found in [11].
276 9 The Structure of Multivariable Nonlinear Systems
and that Xr2 +1 (t) is also identically zero. Continuing in this way, it is concluded that,
if y(t) is identically zero, then (x(t)) = 0, i.e., x(t) evolves on the set Z defined
by (9.38) and the input u necessarily satisfies
0 = A(x) + B(x)u.
Thus, u(t) = u (x(t)) in which u (x) is the function (compare with (9.37))
u (x) = B1 (x)A(x).
The union of (9.40) and all (9.34) characterizes the normal form of the system.
Note that if a complementary set z = Z(x) of coordinates can be found such that
Lg Z(x) = 0, then Eq. (9.40) reduces to
z = f 0 (z, ), (9.41)
and the associated normal form will be referred to as a strict normal form.
10 The notation used in this formula is defined as follows. Recall that Z(x) is a (n d) 1 vector,
whose ith component is the function zi (x), and that g(x) is a n m matrix, whose jth column is the
vector field gj (x). The vector L f Z(x) is the (n d) 1 vector whose ith entry is L f zi (x) and the
matrix Lg Z(x) is the (n d) m matrix whose entry on the ith row and jth column is Lgj zi (x).
9.4 Partial and Full Normal Forms 277
Remark 9.5 Note that, while Assumption 9.1, which includes the assumption that
the system is uniformly invertible, implies the existence of a globally defined normal
form (9.40)(9.34), the converse is not true in general. In fact, a system possessing a
normal form having the structure indicated above may fail to be uniformly invertible.
This is shown, for instance, in the following simple example.
x 1 = a1 (x) + b1 (x)u
x 2 = x3 + (x)[a1 (x) + b1 (x)u]
x 3 = a2 (x) + b2 (x)u
y1 = x1
y2 = x2
in which
b (x)
B(x) = 1
b2 (x)
we see that
S0 (1) S0 y1(1) + a1 (x)
L1 = y + f (x) =
y x y2(1) + x3 + (x)a1 (x)
S0 b1 (x)
M1 = g(x) = .
x (x)b1 (x)
The matrix M1 has rank 1 = 1, and its the first row is nowhere zero. Thus,
M1 = b1 (x)
M2 = (x)b1 (x)
L1 = y1(1) + a1 (x)
L2 = y2(1) + x3 + (x)a1 (x)
Accordingly
At step 2, we have
S1 (2) S1
L2 = y + f (x)
y(1) x
and
S1
M2 = g(x) = b2 (x) + b1 (x) + (x)b1 (x) + b2 (x) y1(1)
x x1 x2 x3
(1)
= + (x) y1(1) b1 (x) + 1 + y b2 (x).
x1 x2 x3 1
The rank of this matrix is not constant for all x, y1(1) . In fact, for y1(1) = 0 the rank is
2 (recall that B(x) has rank 2), while if
(1)
y = 1
x3 1
the rank is 1. Thus, in general, the structure algorithm may have singularities and it
cannot be claimed that the system is uniformly invertible in the sense of definition
(9.36).
From this analysis it can also be concluded that the system is uniformly invertible
if and only if
=0
x3
i.e., if the coefficient (x) is independent of the component x3 of the state vector x.
We will return in more detail on this property in Sect. 11.4.
Remark 9.6 Note that, while in a single-input single-output system the hypotheses
of Proposition 6.3 guarantee the existence of a strict normal form, in a multi-input
multi-output system the hypotheses of Proposition 9.3 do not guarantee the existence
of a strict normal form. As a matter of fact, much stronger assumptions are needed
to this end.11
Example 9.3 Consider the case of a two-input two-output system, with n = 4. Sup-
pose Lg h(x) is nonsingular, so that the inversion algorithm terminates at step 1, with
1 = 2. The partial normal form is simply a set of two equations
11 = a1 (x) + b1 (x)u
21 = a2 (x) + b2 (x)u
in which
11 a1 (x) b1 (x)
= y, = L f h(x) = Lg h(x).
21 a2 (x) b2 (x)
Lg Z(x) = 0.
Letting z1 (x), z2 (x) denote the two components of Z(x), the condition in question
reads
zi (x)
Lgj zi (x) = gj (x) = 0 for i = 1, 2 and j = 1, 2. (9.42)
x
By virtue of a known property of the Lie bracket of vector fields, these conditions
imply12
zi (x)
[g1 , g2 ](x) = 0 for i = 1, 2.
x
which, together with both (9.42), yields
Z(x)
g1 (x) g2 (x) [g1 , g2 ](x) = 0. (9.43)
x
This expression shows that the three columns of
G(x) = g1 (x) g2 (x) [g1 , g2 ](x)
Z(x)
K(x) = .
x
12 If 1 (x) and 2 (x) are vector fields defined on Rn and (x) is a real-valued function defined on
Rn , the property in question consists in the identity
If (x) is a globally defined diffeomorphism, the two rows of the matrix K(x)
are linearly independent and hence its kernel has dimension 2. Thus, (9.43) can be
fulfilled only if the three columns of the matrix G(x) are linearly dependent, for each
x R4 . If these columns are independent, a map Z(x) with the properties indicated
above cannot exist. Now, it is easy to find examples of systems for which this is the
case. Consider for instance the case of a system in which
0 x3 1
0 0 0 x
g1 (x) =
1 , g2 (x) = 0 , [g1 , g2 ](x) = 0 , h(x) = 3
x4
0 1 0
and f (x) is unspecified. In this system Lg h(x) is the identity matrix, but the system
does not admit globally defined strict normal forms.
L rf h(x) 1
(x) = , (x) = ,
Lg L r1
f h(x) Lg L r1
f h(x)
one obtains
y(r) (t) = v(t).
is related to its input v(t) by a linear differential equation. The effect of the feedback
law is to force a linear inputoutput behavior on the resulting closed-loop system. For
this reason, the law in question is usually referred to as an inputoutput linearizing
feedback law.
9.5 The Case of InputOutput Linearizable Systems 281
Wk (x) = Lg L k1
f h(x).
Proposition 9.4 The inputoutput behavior of system (9.1) is that of a linear system
if and only if, for k = 1, . . . , 2n, the matrices Wk (x) are independent of x.
k (x) = Lg L k1 h(x)
W f
Proposition 9.5 Suppose system (9.1) satisfies Assumption 9.1. There exists a feed-
back law of the form (9.2), in which (x) is a m m matrix invertible for all x, such
that the resulting system (9.4) has a linear inputoutput behavior if and only if all
matrices Fij (), with i = 1, . . . , k 1 and j = 1, . . . i, determined by the structure
algorithm are constant matrices.15
13 See [3].
14 The matrix W m matrix whose entry on the ith row and jth column is Lgj L k1
k (x) is the m f hi (x).
With reference to the functions f(x) and (x)
defined in (9.3), note that, for any m 1 vector (x),
we have L f (x) = L f (x) + [Lg (x)](x) and Lg (x) = [Lg (x)](x).
15 For consistency, we rewrite the matrix F1 (x) determined at the first step of the algorithm as F11 (x).
282 9 The Structure of Multivariable Nonlinear Systems
i,1 = i,1
i,r1 1 = i,r1
i,r1 = i,r1 +1 + i,r
1
v
1 +1 1
i,r2 1 = i,r2 + i,r
1
v
2 1
i,r2 = i,r2 +1 + i,r
1
v + i,r
2 +1 1
2
v
2 +1 2
(9.44)
i1
i,ri1
j
= i,ri1 +1 + i,ri1 +1 vj
j=1
i1
i,ri 1 = i,ri +
j
i,ri vj + i,r
2
v
2 +1 j
j=1
i,ri = vi ,
with
y = i,1 .
(x) = 1 (x)
(x)
= 1 (x)(x)
By assumption
Kk := Lg L k1
f
h(x)
is a constant matrix for all k = 1, . . . 2n. A simple induction argument16 shows that
From this, taking the derivatives of both sides along g(x) it follows that, for any
k > 1,
16 Simply take the derivative of both sides along f (x), use the fact that L f L k h(x) = L k+1
h(x) +
f f
Lg L k h(x)(x)
and the fact that the Ki s are constant.
f
9.5 The Case of InputOutput Linearizable Systems 283
Lg L k1 k1
f h(x) = Lg L f h(x)(x) + Kk1 Lg (x)
+ Kk2 Lg L f (x)
+ + K1 Lg L k2
f (x)
i.e.,
Lg L k1
f h(x) = Kk (x) + Kk1 Lg (x)
+ Kk2 Lg L f (x) + + K1 Lg L k2
f (x).
(9.45)
The consequences of this expression on the structure algorithm are now exam-
ined.17 Observe that, by definition
and that
M1 (x) = Lg h(x) = Lg h(x)(x)
= K1 (x).
The matrix K1 is constant, while the matrix (x) is nonsingular for all x. Therefore,
the matrix M1 (x) has constant rank 1 = rank(K1 ) . After a permutation of rows if
necessary, let the first 1 rows of K1 be linearly independent, partition K1 as
K1
K1 =
K1
K1 = F1 K1 .
Then, clearly,
M1 (x)
K1 (x)
=
M1 (x) K1 (x)
and
S1 (y(1) , x) = F1 Im1 y(1) + F1 Im1 L f h(x).
i.e., (2) 2
L2 (y(1) , y(2) , x) = F1 Im
1 y + F1 I m1 L f h(x)
(1)
M2 (y , x) = F1 Im1 Lg L f h(x) .
17 We consider in what follows the case in which > 0 and 2 > 0. The other cases require simple
1
adaptations of the arguments.
284 9 The Structure of Multivariable Nonlinear Systems
from which it is seen that, since the matrix (x) is nonsingular for all x, the matrix
(9.8) has constant rank
K
2 = rank 1 . (9.46)
M2
By construction
F21 0(m2 )(m1 ) K1 + F22 Im2 F1 Im1 K2
= F21 K1 + F22 Im2 M 2 =0
F1 Im1 K1 = 0
and therefore
3 (x)
M3 () = M
for some constant matrix M 3 . As a consequence, the matrix (9.12), written for k = 2,
is a matrix of the form
M1 (x) K1
M2 (y(1) , x) = M 2 (x),
(1) (2)
M3 (y , y , x)
M3
from which it is deduced that F31 (), F32 (), F33 () are constant matrices. Continuing
in the same way, it is not difficult to arrive at the conclusion that all matrices Fij (),
with i = 1, . . . , k 1 and j = 1, . . . i, determined by the structure algorithm are
constant matrices.
Example 9.4 Consider the system with three inputs and three outputs
x 1 = x52 + u1
x 2 = x4 + x52 + u1
x 3 = x6 + 3x52 + 3u1
x 4 = u3
x 5 = x13 + u1 + (1 + x22 )u2
x 6 = x5 + 2u3 ,
y1 = x1
y2 = x2
y3 = x3 .
and obtain
y1(1) + x52 100
S 0 (y(1) , x, u) = L1 (y(1) , x) + M1 (x)u = y2(1) + x4 + x52 + 1 0 0 u.
y3(1) + x6 + 3x52 300
286 9 The Structure of Multivariable Nonlinear Systems
The matrix
10 0
J1 (x) = M1 (x) = 1 0 0
30 0
and
1
F1 = ,
3
so that
(1)
y2(1) + x4 + x52 1 (1) y2 + y1(1) + x4
S1 (y(1) , x) = + (y + x 2
) = .
y3(1) + x6 + 3x52 3 1 5
y3(1) + 3y1(1) + x6
The matrix
100
M1
J2 = = 0 0 1
M2
002
and
F21 = 0, F22 = 2.
9.5 The Case of InputOutput Linearizable Systems 287
This yields
S2 (y(1) , y(2) , x) = y3(2) + 3y1(2) + x5 2 y2(2) + y1(2)
= y3(2) + y1(2) + 2y2(2) + x5 .
In step 3, we obtain
The matrix
M1 1 0 0
J3 (x) = M2 = 0 0 1
M3 1 (1 + x2 )2 0
has rank 3 = 3 for all x R6 and the algorithm terminates. The system is uniformly
invertible. Moreover, since F1 and F22 are constant matrices, the system can be
transformed, by means of a feedback of the form (9.2), into a system whose input
output behavior is that of a linear system.
To see how this can be achieved, we determine the normal form of the system.
According to the procedure described in Sect. 9.4, we define
11
21 = X1 = S0 (0, x), 22
= X2 = S1 (0, x), 33 = X3 = S2 (0, 0, x),
32
31
11 = x52 + u1
21 = 22 + [x52 + u1 ]
22 = u3
31 = 32 + 3[x52 + u1 ]
32 = 33 + 2u3
33 = x13 + u1 + (1 + x22 )u2 ,
A multivariable nonlinear system of the form (9.1) is said to have vector relative
degree {r1 , r2 , . . . , rm } at a point x if:
(i) Lgj L kf hi (x) = 0 for all 1 i m, for all k < ri 1, for all 1 j m, and for
all x in a neighborhood of x ,
(ii) the m m matrix
Lg1 L rf1 1 h1 (x) Lgm L rf1 1 h1 (x)
Lg L r2 1 h2 (x) Lg L r2 1 h2 (x)
B(x) =
1 f m f
(9.47)
rm 1 rm 1
Lg1 L f hm (x) Lgm L f hm (x)
is nonsingular at x = x .18
Remark 9.7 It is clearly seen that this definition includes the one given in Chap. 6
for a single-input single-output nonlinear system. As far as the sequence of integers
{r1 , r2 , . . . , rm } is concerned, observe that each ri is associated with the ith output
channel of the system. By definition, for all k < ri 1, all m entries of the row vector
Lg1 L kf hi (x) Lg2 L kf hi (x) . . . Lgm L kf hi (x)
are zero for all x in a neighborhood of x and, for k = ri 1, in this row vector at
least one entry is nonzero at x . Thus, in view of condition (i), it can be said that,
for each i, there is at least one choice of j such that the (single-input single-output)
system having output yi and input uj has exactly relative degree ri at x . However,
such property is only implied by (i) and (ii). As a matter of fact, property (ii) is much
stronger, as it includes the assumption that the matrix B(x ) is nonsingular.
Remark 9.8 The nonsingularity of B(x ) may be interpreted as the appropriate mul-
f h(x ) is nonzero in a
tivariable version of the assumption that the coefficient Lg L r1
single-input single-output system having vector relative degree at x = x .
System (9.1) is said to have uniform vector relative degree if there exists a sequence
r1 , r2 , . . . , rm such that system (9.1) has vector relative degree {r1 , r2 , . . . , rm } at each
x Rn .
Proposition 9.6 System (9.1) has uniform vector relative degree if and only if it
is uniformly invertible, in the sense of Definition 9.2, and all matrices Fij (), with
i = 1, . . . , k 1 and j = 1, . . . i, determined by the structure algorithm are zero.
18 There is no abuse of notation in using here symbols (specifically r1 , . . . , rm in (i) and (ii), and
B(x) in (9.47)) identical to symbols used earlier in Sect. 9.4. As a matter of fact, we will show in a
moment that the notations are consistent.
9.6 The Special Case of Systems Having Vector Relative Degree 289
Proof Suppose the system has uniform vector relative degree and let the outputs be
reordered in such a way that r1 r2 rm . Then, it is straightforward to see
that, if r1 = 1, then J1 (x) = Lg h(x) while, if r1 > 1,
J1 () = = Jr1 1 () = 0
and
Jr1 () = Lg L rf1 1 h(x).
Let m1 be the number of nonzero rows in Lg L rf1 1 h(x). Then r1 = r2 = = rm1 and
rm1 +1 > rm1 . By assumption (ii), the upper m1 rows of Jr1 () are linearly independent
at each x, while the lower m m1 rows are identically zero. Therefore, all matrices
Fi,j () with i = 1, . . . , r1 and j = 1, . . . , i are missing, with the exception of Fr1 ,r1 ()
which is zero. As a consequence
Continuing in a similar way, it is easily concluded that the structure algorithm has no
singularities, that all matrices Fij (), with i = 1, . . . , k 1 and j = 1, . . . i, deter-
mined by the structure algorithm are either missing or zero, and that Jrm () coincides
with the matrix B(x) defined in (9.47), which is nonsingular for all x. Hence the
system is uniformly invertible.
Conversely, suppose all matrices Fij (), with i = 1, . . . , k 1 and j = 1, . . . i,
determined by the structure algorithm are zero. Then, in the set of equations (9.34),
j
all multipliers i,k (x), for k = r1 + 1, . . . , ri and j = 1, . . . , i 1, are zero. As a
consequence,
y i(ri ) = i,1
(ri )
= ai (x) + bi (x)u.
From this, using the fact that the matrix B(x) in (9.36) is nonsingular for all x, it is
immediate to conclude that the system has uniform vector relative degree, with
It is seen from this proposition that the class of those multivariable systems that
have uniform vector relative degree is a subclass of those multivariable systems that
can be inputoutput linearized, which in turn is a subclass of those multivariable
systems that are invertible.
If Assumption 9.1 holds, the equations describing a system having uniform vector
relative degree {r1 , r2 , . . . , rm } can be transformed, by means of a globally defined
diffeomorphism, into equations of the form
290 9 The Structure of Multivariable Nonlinear Systems
z = f 0 (z, ) + g0 (z, )u
1,1 = 1,2
1,r1 1 = 1,r1
1,r1 = a1 (x) + b1 (x)u
m,1 = m,2 (9.48)
m,rm 1 = m,rm
m,rm = am (x) + bm (x)u
y1 = 1,1
ym = m,1 ,
= col(1 , 2 , . . . , m )
i = col(i,1 , i,2 , . . . , i,ri ) i = 1, 2, . . . , m.
B,
Such equations can be put in compact form, by letting A, C denote the matrices
A = diag(A 1 , . . . , A m )
B = diag(B 1 , . . . , B m ) (9.49)
C = diag(C 1 , . . . , C m )
With such notations, in fact, the equations in question can be rewritten in the form
z = f 0 (z, ) + g0 (z, )u
+ B[A(x)
= A + B(x)u] (9.51)
y = C,
where A(x) and B(x) are the m 1 vector and, respectively, the m m matrix defined
in (9.36). Finally, setting
q(z, ) = A( 1 (z, ))
b(z, ) = B( 1 (z, ))
9.6 The Special Case of Systems Having Vector Relative Degree 291
with x = 1 (z, ) the inverse of the map (x) considered in Assumption 9.1, the
equations in question can be further rewritten as
z = f 0 (z, ) + g0 (z, )u
+ B[q(z,
= A ) + b(z, )u] (9.52)
y = C.
References
In this and in the following chapter we consider nonlinear systems having m inputs
and m outputs, modeled by equations of the form (9.1) and we suppose that Assump-
tion 9.1 holds. This assumption guarantees the existence of a globally defined change
of coordinates x = (x) in which x can be split in two subsets z Rnd and Rd .
The set in turnis split in subsets, each one of which consists of a string of ri
of components
i,1 , i,2 , . . . , i,ri
which are seen to satisfy equations of the form (9.34).1 The collection of all (9.34), for
i = 1, . . . , , is a set of m 1r1 + m 2 r2 + + m r = d equations that characterizes
what has been called a partial normal form. A full normal form is obtained by adding
the Eq. (9.40) that models the flow of the complementary set z of new coordinates. It
should be observed, in this respect, that a strict normal form, namely a normal form
in which (9.40) takes the simpler structure (9.41), exists only under additional (and
strong) assumptions (see Remark 9.6).
The problem addressed in the present and in the following chapter is the design
of feedback laws yielding global stability or stability with guaranteed domain of
attraction. To this end it is appropriate to consider a property that can be viewed
as an extension, to multivariable systems, of the property of strong minimum phase
considered in Definition 6.1 and that in fact reduces to such property in the case of
a single-input single-output system.
Definition 10.1 Consider a system of the form (9.1), with f (0) = 0 and h(0) = 0.
Suppose the system satisfies Assumption 9.1 so that a normal form can be globally
holds.
It is seen from the previous definition that, if a strict normal form exists, the role
of the input u() in the previous definition is irrelevant and the indicated property
reduces to the property that system (9.41), viewed as a system with state z and input
, is input-to-state stable. As in Chap. 6, it will be useful to consider also the stronger
version of the property in question in which () and () are bounded as in (6.18).
Definition 10.2 A system is stronglyand also locally exponentiallyminimum-
phase if, for any x(0) Rn and any admissible input function u() : [0, ) Rm ,
so long as x(t) is defined an estimate of the form (10.1) holds, where (, ) and ()
are a class K L function and, respectively, a class K function bounded as in (6.18).
Remark 10.1 If a strict normal form exists, in which (see (9.41))
z = f 0 (z, ),
and
V0
f 0 (z, ) (z) for all (z, ) Rnd Rd such that z ( ).
z
However, if the system does not possess a strict normal form (for instance because
conditions analogous to those outlined in Example 9.3 do not hold), a similar criterion
for determining the property of strong minimum phase is not available.
Example 10.1 Consider the system
z 1 = z 1 + z 2 1 u 2
z 2 = z 2 + 2 z 1 1 u 2
1 = u1
2 = u2.
This is a system having vector relative degree {1, 1} expressed in normal form.
However, this system cannot be expressed in strict normal form. In fact, since
10.1 The Hypothesis of Strong Minimum-Phase 295
0 z 2 1 z2
0 z 1 1 z 1
g1 (x) =
1 , g2 (x) = 0 , [g1 , g2 ](x) = 0
0 1 0
has rank 3 for all nonzero z. Hence, as shown in Example 9.3, a strict normal form
cannot exist.
This system is strongly minimum phase. In fact, taking the positive definite func-
tion V (z) = z 12 + z 22 , it is seen that along the trajectories of (9.40)
V = 2(z 12 + z 22 ) + 2z 2 2 .
Form this, it is readily seen that a bound of the form (10.1) holds, regardless of what
the input u() is.2
z = f 0 (z, ) + g0 (z, )u
+ B[q(z,
= A ) + b(z, )u] (10.2)
y = C,
in which
K = diag( K 1 , . . . , K m ),
A + B K = diag( A 1 + B 1 K 1 , . . . , A m + B m K m ).
Note also that, since each pair ( A i , B i ) is a reachable pair, the eigenvalues of A + B K
can be freely assigned by properly choosing the matrix K .
Then, the following resulta multivariable version of Proposition 6.4holds.
Proposition 10.1 Suppose system (10.2) is strongly minimum phase. If the matrix K
is such that ( A + B K ) C , the state feedback law (10.3) globally asymptotically
stabilizes the equilibrium (z, ) = (0, 0).
Proof We begin by showing that, for any initial condition (z(0), (0)) the trajectory
of the closed loop systems is defined (and, actually, bounded) for all t 0. As a
matter of fact, suppose that this is not the case. Then, for any R, there exists a time
TR such that the trajectory is defined on the interval [0, TR ] and
for some suitable M > 0 and > 0, if we pick R > M(0) the first inequality
is contradicted. As far as the second one is concerned, using the property that the
system is strongly minimum phase, it is seen thatsince the trajectory is defined on
the interval [0, TR ]the estimate
holds. Thus, if
R > (z(0), 0) + (M(0)),
Having proven that the trajectory is defined for all t 0, we use the estimate (10.1)
to claim that the upper subsystem of (10.4), viewed as a system with state z and input
, is input-to-state stable. Since the lower subsystem is globally asymptotically stable,
the conclusion follows from standard results.3
Remark 10.2 From the arguments used in the previous proof, it is also readily seen
that, if system (10.2) is stronglyand also locally exponentiallyminimum phase,
the equilibrium (z, ) = (0, 0) of the closed loop system (10.4) is globally asymp-
totically and also locally exponentially stable.
As observed right after Proposition 6.4, the feedback law (10.3), although intuitive
and simple, is not useful in a practical context because it relies upon exact cancellation
of nonlinear functions and requires the availability of the full state (z, ) of the system.
We will see in the subsequent sections of this chapter how such limitations canto
some extentbe overcome.
We describe now a design method, based a work of Khalil and Freidovich,4 according
to which the feedback law (10.3) can be effectively replacedto the purpose of
achieving stability with a guaranteed domain of attractionby a (dynamic) feedback
law in which the availability of (z, ) and the knowledge of the functions b(z, ) and
q(z, ) are not required.
To this end, we need to put a restriction on the matrix b(z, ) appearing in (10.2).
Specifically, we consider the case in which the matrix in question has the property
indicated in the following assumption.
Assumption 10.1 There exist a constant nonsingular matrix b Rmm and a num-
ber 0 < 0 < 1 such that
for some pair bmin bmax . Actually, in case m = 1, Assumption 10.1 is equivalent
to (10.6). In fact, if m = 1, Assumption 10.1 reduces to the assumption of existence
of a number b such that
b(z, ) b
0 < 1
b
which is always possible if b(z, ) is bounded as in (10.6). Thus, Assumption 10.1
can be regarded as a multivariable version of the (very reasonable) assumption that
the so-called high-frequency gain coefficient b(z, ), of a single-input single-output
nonlinear system, is bounded from below and from above.5
(, ) = b1 [ K ],
in which
5 Inthe control of systems having a vector relative degree, it is inevitable to put assumptions on the
matrix b(z, ), the so-called high-frequency gain matrix. Such assumptions are essentially meant to
make sure that the various components of b(x, )u have the right sign. Different hypotheses have
been proposed in the literature to this purpose, see e.g., [1013]. We consider Assumption 10.1 a
reasonable alternative.
10.3 A Robust Observer 299
= col(1 , 2 , . . . , m )
= col(1 , 1 , . . . , m )
where, for i = 1, 2, . . . , m,
In these equations, the coefficient and ci,0 , ci,1 , . . . , ci,ri are design parameters,
while bi is the ith row of the matrix b.
The dynamical system thus defined has a structure similarto some extentto
that of the observer considered in Chap. 7. In view of performing a similar conver-
gence analysis, the states i,1 , i,2 , . . . , i,ri , i will be replaced by variables defined
as (compare with (7.14))
Setting
e = col(e1 , . . . , em )
ei = col(ei,1 , ei,2 , . . . , ei,ri +1 ), i = 1, . . . , m
and
ext = col( , ),
T : Rd+m Rd+m
(10.10)
ext e = T (z, , ext )
Lemma 10.1 Suppose Assumption 10.1 holds. Then the map (10.10) is globally
invertible.
Proof Observe that the first ri of the identities (10.9) can be trivially solved for i, j ,
expressing it as a function of ei, j and i, j . To complete the inversion, it remains to find
from the last identities. This can be achieved if b has suitable properties. Setting
the last equations of (10.9), written for i = 1, . . . , m, can be organized in the single
equation
= q(z, ) + [b(z, ) b]G((, )) .
Adding and subtracting K on the right-hand side and bearing in mind the definition
of (, ), this becomes
Set now
(z, , ) = b1 [ K q(z, ) + ]
= F().
Since G(s) is bounded and so is b(z, ), the map F(s) is proper (that is,
F(s) as s ). Thus, according to Hadamards Theorem, the map
F(s) has a globally defined inverse if the Jacobian of F(s) is everywhere nonsingu-
lar. The Jacobian of F(s), in turn, has the following expression
F
= b1 [b(z, ) b]G (s) + I (10.14)
s
F
= b1 [b(z, ) b]G (s) + b = b1 (M(z, , s) + I )b
s
in which
M(z, , s) = [b(z, ) b]G (s)b1 . (10.15)
M(z, , s) + I. (10.16)
Clearly, a sufficient condition for (10.16) to be invertible is that M(z, , s) < 1.6
The property that M(z, , s) < 1 is implied by Assumption 10.1. In fact, since
G (s) is diagonal and satisfies |G (s)| 1, we have (under such Assumption)
M(z, , s) max [b(z, ) b]b1 0 < 1 for all (z, , s).
diagonal, |
|1
In other words, Assumption 10.1 guarantees that the Jacobian matrix (10.14) of the
map F(s) is nowhere singular and hence the map in question is globally invertible.
Having shown that F(s) is globally invertible, we can express (, ) as
(, ) = F 1 ( (z, , )),
= K bF 1 ( (z, , ))
showing how can be expressed as a function of e (and, of course, of (z, )). This
completes the proof that the map e = T (z, , ext ) is globally invertible in ext .
x = col(z, )
6 By contradiction, suppose M(z, , s) + I were singular. Then, there would exist a vector x sat-
isfying M(z, , s)x = x which would imply M(z, , s) 1 which contradicts the condition
M(z, , s) < 1.
302 10 Stabilization of Multivariable Nonlinear Systems: Part I
in which case the vector e is seen as a function of (x, ext ) and, conversely, the vector
ext is seen as a function of (x, e).
Remark 10.4 Using the various expressions introduced above, the fact that q(0, 0) =
0 and the fact that F(0) = 0, it is easy to check that, if x = 0, then e = 0 if and only
if ext = 0.
in which qi (z, ) and bi (z, ) denote the derivatives with respect to time of qi (z, )
and bi (z, ). In view of the developments which follow, it is useful to separate, in
this expression, the terms that depend on the design parameters. This is achieved by
setting
which yields
in which 2,i (x, e) denotes the ith row of the vector 2 (x, e). Looking at the expres-
sion of , we can write
10.3 A Robust Observer 303
c1,0 e1,1
c2,0 e2,1
[bi (z, ) bi ]G ((, ))b1 = [bi (z, ) bi ]G ((, ))b1 ..
.
cm,0 em,1
c1,0 e1,1
c2,0 e2,1
:= 0,i (x, e) ..
.
cm,0 em,1
in which
A = diag(A1 , A2 , . . . , Am )
B1 = diag(B1,1 , B1,2 , . . . , B1,m )
B2 = diag(B2,1 , B2,2 , . . . , B2,m )
C = diag(C1 , C2 , . . . , Cm )
and 0 (x, e), 1 (x, e), 2 (x, e) are the functions indicated above.
The special structure of this equation will be exploited in the next section, where
a complete analysis of the asymptotic properties of the resulting closed-loop system
is conducted. In this analysis, it is of paramount importance to take advantage of
appropriate bounds on the functions 0 (x, e), 1 (x, e), 2 (x, e), described in what
follows.
Lemma 10.2 Suppose Assumption 10.1 holds and 1. There exist numbers 0 <
1 and 1 such that
Lb K e.
1 1
From this, it is seen that the second bound in (10.22) holds for some 1 > 0.
So long as the bound in (10.23) is concerned, observe that, if system (10.2) is
controlled by (10.7), we have
2,i (x, e) = q
z [ f 0 (z, ) + g0 (z, )G(( , )]
i
qi
+ [ A + B[q(z, ) + b(z, )G(( , ))]]
+ b
z [ f 0 (z, ) + g0 (z, )G(( , )]G((, ))
i
bi
+ [ A + B[q(z,
) + b(z, )G(( , ))]]G((, ))
+[bi (z, ) bi ]G ((, ))b1 K [ A
+ B[q(z,
) + b(z, )G(( , ))]].
Bearing in mind the expressions of G() and G () and the fact that g() and g ()
are bounded functions, it is seen that, so long as x R, the function 2,i (x, e)
remains boundedregardless of what e actually is8 by a number M R that only
depends on R and is independent of . This completes the proof of the lemma.
We proceed now with the proof that, if system (10.2) is controlled by (10.7), in
which and are states of (10.8), the saturation level L in g() and the design
parameters ci,0 , ci,1 , . . . , ci,ri and in (10.8) can be chosen in such a way that the
resulting system has appropriate asymptotic properties. Specifically, it will be shown
that the equilibrium point (x, ext ) = (0, 0) of the resulting closed-loop system can
be semiglobally and practically stabilized and, under appropriate additional assump-
tions, also semiglobally asymptotically stabilized. The discussion below essentially
follows the paradigm adopted in the proof of Proposition 7.3. However, adaptation
to the present context requires number of non-negligible modifications.
We begin by rearranging the equations that describe the closed-loop system. First
of all, as done in the previous section, we use e to replace the states and of (10.8),
in which case the Eq. (10.21) holds. Moreover, we add and subtract the function
(10.3) to the control u defined in (10.7), which is therefore written as
in which
3 (x, e) = G(( , )) [b(z, )]1 (q(z, ) + K ),
b
x (x)
7 If b(x) is a m-dimensional row vector, the notation used in the next formula stands for
b b1 b2 bm
(x) := (x) (x) (x) .
x x x x
where, consistently with the notation used in the previous section, as arguments of
3 () we have used x, e instead of z, , , .
As a consequence, the equations of system (10.2) controlled by (10.7) can be
regarded as equations of the form
in which
f 0 (z, ) + g0 (z, )[b(z, )]1 (q(z, ) + K ) g0 (z, )
F(x) = G(x) = .
( A + B K ) Bb(z, )
It is known from Sect. 10.2 that, if system (10.2) is strongly minimum phase and K
is chosen as indicated in Proposition 10.1, the equilibrium x = 0 of
x = F(x) (10.25)
Remark 10.5 In this respect, it is important to stress that the set C is assigned in
terms of the original state variables, namely the state x of the controlled plant (10.2)
and the state ext of (10.8). In the Eq. (10.26), instead, the vector e replaces ext .
Thus, bearing in mind the change of variables defined in (10.9), it should be taken
into account how the (compact) set to which the initial values of x and e belong is
influenced by the choice of . We will return on this issue later on.
With C being the fixed compact set of initial conditions of the system, pick a
number R > 0 such that
9 Tothis end, use the properties that f 0 (0, 0) = 0 and q(0, 0) = 0 and the observation in Remark
10.4 to show that 3 (0, 0) = 0 and 2 (0, 0) = 0. From Lemma 10.2, observe that 1 (x, 0) = 0.
10.4 Convergence Analysis 307
V
F(x) (x)
x
for some class K function (). Pick a number c > 0 such that
c = {x : V (x) c} B R ,
We will show in what follows that, if the remaining design parameters are appropri-
ately chosen, for all times t 0 the vector x(t) remains in c+1 .
We begin with the analysis of the evolution of x(t) for small values of t 0.
To this end observe that, since G() is a bounded function, so long as x c+1 the
perturbation term 3 (x, e) in the upper system of (10.26) remains bounded. In other
words, there exists a number 3 , such that
Note that the number 3 is independent of the choice of . Actually, it depends only
on c+1 and hence, indirectly (via the choice of R), on the choice of the set C in
which the initial conditions are taken.
Consider now a trajectory with initial condition x(0) B R c . In view of the
above bound found for 3 (x, e), it can be claimed that, so long as x(t) c+1 ,
V V
V (x(t)) = [F(x) + G(x)3 (x, e)] (x) + G(x)3 .
x x
Let
V
M = max G(x)
xc+1 x
V (x(t)) M3
V (x(t)) V (x(0)) + M3 t c + M3 t.
This inequality shows that x(t) remains in c+1 at least until time T0 = 1/M3 . This
time might be very small but it is independent of , because so are M and 3 . It
follows from the previous remarks that such T0 only depends on c+1 and hence
only on the choice of the set C .
During the time interval [0, T0 ] the state e(t) remains bounded. This is seen from
the bottom equation of (10.26), using the bounds determined for 0 (x, e), 1 (x, e),
2 (x, e) in Lemma 10.2 and the fact that x(t) c+1 for all t [0, T0 ]. The bound
on e(t), though, is affected by the value of . In fact, looking at the definitions of the
various components of e, it is seen that e(0) grows with (despite of the fact that,
by assumption, x(0) R and ext (0) R).
Having established that trajectories of the system exists on the time interval [0, T0 ],
we analyze next the behavior of e(t) for times larger than T0 . To this end, the two
results that follow are instrumental.
Lemma 10.3 There exist a choice of the coefficients ci,0 , . . . , ci,ri for 1 i m, a
positive definite and symmetric (d + m) (d + m) matrix P and a number > 0
such that
e = Ae + B2 u
(10.29)
y = Ce
Note that Ti (0) = 1. Therefore, if the ci, j s are chosen in such a way that the poles
of Ti (s) are all real and have negative real part, |Ti ( j)| 1 for all R. As a
consequence
T () H = 1.
10.4 Convergence Analysis 309
Consider now the number 0 defined in Assumption 10.1. Since this number is strictly
less than 1, we can claim that
T () H < 01 .
This being the case, it follows from the Bounded Real Lemma (see Theorem 3.1)
that there exist a positive definite and symmetric matrix P and a number > 0 such
that, along the trajectories of (10.29), the estimate
holds.
System (10.28) is obtained from (10.29) by picking u = 0 (x, e)y for which,
using the bound (10.22), we have
Lemma 10.4 Let the ci, j s be chosen so as to make (10.27) satisfied. Suppose x(t)
c+1 for all t [0, Tmax ) and suppose that ext (0) R. Then, for every 0 < T
Tmax and every > 0, there is a such that, for all ,
in which is any arbitrary (positive) number. Let now R be a number such that
c+1 B R .
310 10 Stabilization of Multivariable Nonlinear Systems: Part I
2 (x, e) M R .
Choose as
M R2 P
= .
2
Bearing in mind the estimates in (10.30), set
(21 + )P
= (10.31)
2a2
2 a1 > 1,
U (e(t)) 2 U (e(t)) + 2 ,
holds, for any t [0, Tmax ). From this, by means of standard arguments, it can be
concluded that10
t a2
e(t) Ae e(0) + , in which A = .
a1
At this point, it is necessary to obtain a bound for e(0). Bearing in mind the
Definitions (10.9), the fact that x(0) R and ext (0) R, and assuming 1,
it is seen that
rmax
e(0) R
t rmax +
e(t) A Re
for all t [0, Tmax ). Let now T be any time satisfying 0 < T < Tmax , andusing
(10.31)observe that
2
U (t) e2 t U (0) +
2
from which, using the estimates (10.30) and the fact that 2 a1 > 1, the claimed inequality follows.
10.4 Convergence Analysis 311
(21 +)PT T
T rmax = A Re
2a rmax
A Re 2a2
e 2 .
T
2a rmax
Clearly, the function f : f () = e 2 decays to 0 as . Thus, there
is a number such that
(21 +)PT
A Re
T rmax = A Re 2a2
f () .
e(T ) 2.
Finally, bearing in mind the fact that > 0, we see that for t [T, Tmax )
This lemma will now be used to show that the parameter can be chosen in
such a way that x(t) c+1 for all t [0, ). We know from the previous analysis
that x(t) c+1 for t T0 . Suppose that, for some Tmax > T0 , x(t) c+1 for all
t [T0 , Tmax ). From the lemma we know that for any choice of , there is a value of
such that, if , e(t) 2 for all t [T0 , Tmax ).
Since b(z, ) is nonsingular for all (z, ), there exists a number b0 such that
[b(z, )]1 b0 for all x c+1 . Suppose satisfies
1
[b(z, )]1 [b(z, )]1 e b0 2 < ,
2
for all t [T0 , Tmax ), which in turn implies, because of the choice of L,
1
[b(z, )]1 [ K q(z, ) + ] L . (10.33)
2
Using (10.12) and the definition of the function F(s) in (10.13), observe that
11 Note that depends on and, actually, increases as decreases. This fact, however, does not
affect the prior conclusion.
312 10 Stabilization of Multivariable Nonlinear Systems: Part I
F(s) = [b(z, )]1 bF(s) = [I [b(z, )]1 b]G(s) + [b(z, )]1 bs,
1
F((, )) L . (10.34)
2
The function F(s) is globally invertible (because so is F(s), as shown earlier in the
proof of Lemma 10.1). Moreover, it can be easily seen that the function F(s) is an
identity on the set
In fact,
s CL G(s) = s,
and this, using the expression shown above for F(s), proves that F(s) = s for all
s C L . As a consequence, the pre-image of any point p in the set C L is the point
p itself. Since the inequality (10.34) implies F((, )) C L , it is concluded that
F((, )) = (, ) and hence
1
(, ) L , (10.35)
2
for all t [T0 , Tmax ).
We use this property to show that, for all t [T0 , Tmax ),
G(( , )) = ( , ), (10.36)
i.e. that on the time interval t [T0 , Tmax ), none of the components of the control
(10.7) is saturated. To this end, observe that
= D()e (10.37)
and note that D() 1 if, without loss of generality, it is assumed 1. Since,
by definition,
( , ) = (, ) b1 K D()e,
10.4 Convergence Analysis 313
if > 1 we have
Suppose satisfies
4b1 K < 1. (10.38)
Then, using (10.35), we see that ( , ) < L , and this proves that (10.36) holds
on the time interval [T0 , Tmax ).
We return now to Eq. (10.24) and observe that, for all t [T0 , Tmax ),
This estimate is now used in the evaluation of the derivative of V (x) along the
trajectories of (10.24). Setting
4 = M(b1 K + b0 ),
V
[F(x) + G(x)3 (x, e)] (x) + 4 e (x) + 24 .
x
Suppose satisfies
24 < min (x). (10.39)
xc+1
If this is the case, then V (x(t)) is negative on the boundary of c+1 . As a consequence,
x(t) c+1 for all t 0 and Tmax = . In summary, from Lemma 10.4 we deduce
that, given a number that satisfies (10.32), (10.38), (10.39), there exists a number
314 10 Stabilization of Multivariable Nonlinear Systems: Part I
such that, if , then for any initial condition in C the trajectories of the
system are such that x(t) c+1 for all t 0 and e(t) 2 for all t T0 .
This being the case, arguments similar to those used in the proof of Proposition
6.6 can be invoked to show that also x(t) enters, in finite time, an arbitrarily small
target set. Let be a (small) given number, pick d such that
d B c+1 ,
and let
a = min (x).
xSdc+1
If satisfies12
44 < a, (10.40)
we see that
1
V (x(t)) a, for all x Sdc+1 .
2
This, as shown in a similar context in the proof of Proposition 6.6, implies the
existence of a time T1 > T0 such that
ultimate bound for x(t) and ext (t) has been chosen, can be fixed, which in
turn determines the minimal value of . .
If asymptotic stability is sought, one can proceed as in the case considered in
Sect. 6.5, by strengthening the assumption of strong minimum phase on the controlled
system (10.2).
Proposition 10.3 Consider system (10.2). Suppose the system is stronglyand also
locally exponentiallyminimum phase and Assumption 10.1 holds. Let the system be
controlled by (10.7)(10.8), with b chosen so as to satisfy the condition in Assumption
10.1. Let K such that A + B K is Hurwitz. For every choice of a compact set C there
is a choice of the design parameters L and ci,0 , ci,1 , . . . , ci,ri , with 1 i m, and a
number such that, if , the equilibrium (x, ext ) = (0, 0) is asymptotically
stable, with a domain of attraction A that contains the set C .
Proof The proof uses arguments very similar to those used in the proof of Proposition
6.7. We limit ourselves to give only a sketch of such arguments. It is known from
Remark 10.2 that the equilibrium x = 0 of (10.25) is locally exponentially stable.
Hence, it can be assumed that the function V (x), for some positive numbers b, k0 , d >
0 satisfies13
V V
F(x) bx2 and k0 x for all x d.
x x
Consider the set
Sd = {(x, e) : x d, e d}.
Looking at the definitions of 3 (x, e) and 2 (x, e) it is seen that, if d is small enough,
there is a positive number k1 (independent of if 1) such that
V
V (x(t)) = [F(x) + G(x)3 (x, e)] bx2 + k0 k1 x e
x
and
so long as (x, e) Sd . Using standard estimates, it is seen from this that there is
a number such that, if , the function W (x, e) is negative definite for all
(x, e) Sd . From this, the conclusion follows as in the proof of Proposition 6.7.
As anticipated, the feedback law used to obtain the stabilization results described
in the propositions above uses only the output y of the controlled system as measured
variable, and does not rely upon exact knowledge of the functions of (z, ) that
characterize the model (10.2). Rather, it only relies upon the knowledge of the values
of the integers r1 , . . . , rm , on the assumption that the system is strongly minimum
phase and on the assumption that the high-frequency gain matrix b(z, ) satisfies
Assumption 10.1. Note, in particular, that the construction of the observer (10.8)
only requires knowing the values of r1 , . . . , rm : everything else only consists of
design parameters. In this respect, it can be claimed that the feedback law in question
is robust, as opposite to the feedback law discussed in Sect. 7.5.
References
13. H. Rodriguez, A. Astolfi, R. Ortega, On the construction of static stabilizers and static output
trackers for dynamically linearizable systems, related results and applications. Int. J. Control
79 (2006)
14. F. Esfandiari, H.K. Khalil, Output feedback stabilization of fully linearizable systems. Int. J.
Control 56, 10071037 (1992)
Chapter 11
Stabilization of Multivariable Nonlinear
Systems: Part II
In this chapter, we consider the problem of achieving global stability, or stability with
a guaranteed region of attraction, for systems that do not possess a vector relive degree
but are uniformly invertible, in the sense of Definition 9.2 and satisfy Assumption 9.1.
It has been shown in Chap. 9 that, under such assumptions, the equations describing
the system can be transformedby means of a globally defined diffeormorphism
into equations consisting of a set of the form (9.40) and of sets of the form (9.34).
In the first part of the chapter we address the case of a system that belongs to the
special class considered in Sect. 9.5. The peculiar feature of such class of systems is
j
that, in the Eq. (9.34), the coefficients i,k (x) are independent of x. This property is
of a great help in the design of stabilizing laws.
We begin by observing that, for this class of systems, the design of full state glob-
ally stabilizing feedback laws is relatively straightforward, if the system is assumed
to be strongly minimum phase, in the sense of Definition 10.1. To this end, in fact,
one could proceed as follows.1 Suppose, to simplify matters, that = m, in which
case all mi s are equal to 1 (in fact, the general case in which one or more of the
mi s is larger than 1 can be handled exactly in the same way). Consider the set (9.34)
corresponding to i = 1, which is a system of the form
1,1 = 1,2
1,r1 1 = 1,r1
1,r1 = a1 (x) + b1 (x)u.
1 The design method described hereafter, that leads to the state feedback control (11.4), essentially
follows the procedure suggestedin a more general contextin [1], where the property of minimum
phase introduced in [5] is exploited.
1 = (A 1 + B 1 K 1 )1
in which A 1 and B 1 are matrices whose structure is the same as that of the matrices
(9.50). Clearly, it is possible to pick K 1 so as to make (A 1 + B 1 K 1 ) a Hurwitz matrix.
Next, consider the set (9.34) corresponding to i = 2, whichin view of (11.1)is
a system of the form
2,1 = 2,2
2,r1 1 = 2,r1
2,r1 = 2,r1 +1 + 2,r
1
K
1 +1 1 1
1
i,r2 1 = i,r2 + i,r2
K1 1
i,r2 = a2 (x) + b2 (x)u.
2 = (A 2 + B 2 K 2 )2 + A21 1
in which A 2 and B 2 are matrices whose structure is the same as that of the matrices
(9.50) and A21 is a fixed matrix. Clearly, it is possible to pick K 2 so as to make
(A 2 + B 2 K 2 ) a Hurwitz matrix.
This procedure can be iterated up to a last stage, corresponding to i = , in which
it is assumed that u is such that
A(x) + B(x)u = K ,
for some M > 0 and > 0. If the system is strongly minimum phase, this implies
as shown in the proof of Proposition 10.1 in a similar contextthat the equilibrium
(z, ) = (0, 0) is globally asymptotically stabilized.
The feedback law thus found, though, suffers of limitations that have already been
outlined in similar contexts before: it relies upon exact cancelation of nonlinear terms
and presupposes the availability of the full state (z, ). To overcome this limitation
we proceed differently, by means of a design method consisting of a blend of the
procedure used in Sect. 6.5 to derive an equation of the form (6.35), and of the robust
stabilization method described in Chap. 10.
To this end, we show first how the equations that describe the system can be
brought to a form that can be seen as a multivariable version of the form (6.35).2 To
this end, we begin by rewriting the equations (9.34) in more compact form.3 Setting
i,1 0 Imi 0 0 0
i,2 0 0 Imi 0 0
i =
... , A =
i
Bi = . . . .
i,ri 2 0 0 0 Imi 0
i,ri 1 0 0 0 0 Imi
we obtain4
1 = A1 1 + B1 1,r1
1,r1 = a1 (x) + b1 (x)u
and, for i = 2, . . . , ,
2 The transformations described hereafter, that lead to equations having the form (11.19), have been
proposed in [2].
3 Inwhat follows we return, for completeness, to the general case in which one or more of the mi s
can be larger than 1.
4 We tacitly assume that r > 1. If r = 1 the vector is missing and so is the associated dynamics.
1 1 1
322 11 Stabilization of Multivariable Nonlinear Systems: Part II
i1
i = Ai i + Bi i,ri + j=1 ij [aj (x) + bj (x)u]
i,ri = ai (x) + bi (x)u
where the ij s are (ri 1)mi mj matrices, partitioned into (mi mj )-dimensional
blocks, defined as follows
i1 = col(0, . . . , 0, i,r
1
1 +1
, . . . , i,r
1
i
)
i2 = col(0, . . . , 0, i,r2 +1 , . . . , i,r
2 2
i
)
i,i1 = col(0, . . . , 0, i,r
i1
i1 +1
, . . . , i,r
i1
i
).
Define a (d m) d matrix T0 as
I (r1 1)m1 0
0
0 0
0 21
I (r2 1)m2 0
0
0
T0 =
0 31 0 32 I(r3 1)m3 0 0
(11.6)
0 1 0 2 0 3 I(r 1)m 0
and a m d matrix T1 as
K1 Im1
0
0 0
0 K2 21 0
K2 Im2
0
T1 = 0 K3 31 0 K3 32 0
K3 Im3 . (11.7)
0 K 1 0 K 2 0 K 3 K Im
is nonsingular. Set
= T0
(11.9)
= T1 .
Then,
11.1 Handling Invertible Systems that Are InputOutput Linearizable 323
= F + G
(11.10)
= H + K + [A(x) + B(x)u],
in which
(i) the (d m) (d m) matrix F has a lower triangular structure
F11 0 0
F21 F22 0
F =
F1 F2 F
where, for i = 1, . . . , , the matrix Fii is a (ri 1)mi (ri 1)mi matrix
whose characteristic polynomial coincides with [di ()]mi , and hence is a Hurwitz
matrix,
(ii) the m 1 vector A(x) and, respectively, the m m matrix B(x) are those defined
in (9.36).
Proof Define
1 = 1
(11.11)
i = i i1
j=1 i,j j,rj , for i = 2, . . . , ,
This yields
1 = A1 1 + B1 1,r1
(11.12)
i = Ai i + i1
j=1 Ai i,j j,rj + Bi i,ri , for i = 2, . . . , .
i = Ki i + i,ri , (11.13)
1 = (A1 B1 K1 )1 + B1 1
i1
i = (Ai Bi Ki )i i1
j=1 Ai i,j Kj j + j=1 Ai i,j j + Bi i , for i = 2, . . . , .
(11.14)
Setting, for i = 1, . . . ,
Fii = Ai Bi Ki (11.15)
(from which it is easily seen that the characteristic polynomial of the matrix Fii
coincides with [di ()]mi ),
Gii = Bi
and, for j = 1, . . . , i 1,
324 11 Stabilization of Multivariable Nonlinear Systems: Part II
Fij = Ai i,j Kj ,
Gij = Ai i,j ,
1 = F11 1 + G11 1
i (11.16)
i = Fii i + i1
j=1 Fij j + j=1 Gij j , for i = 2, . . . , .
Moreover
= col(1 , 2 , . . . , ), = col(1 , 2 , . . . , ),
F11 0 0 G11 0 0
F21 F22 0 G21 G22 0
F =
, G = ,
F1 F2 F G1 G2 G
we have in fact
= F + G, (11.18)
and
= H + K + [A(x) + B(x)u],
in which H and K are suitable matrices and A(x) and B(x) are those defined in (9.36).
By construction = T0 , in which T0 is the matrix (11.6) and = T1 in which T1
is the matrix (11.7). Inspection of T0 and T1 shows that the matrix (11.8) is invertible
and this concludes the proof of the lemma.
j
It is seen from this lemma that the system under consideration, if all ik ()s are
independent of x, can be modeled by equations that have the form
z = f 0 (z, ) + g0 (z, )u
= F + G (11.19)
= H + K + [A(x) + B(x)u].
vector as if it were the output of (11.19). The first and more straightforward
consequence of this viewpoint is that, since the matrix B(x) is nonsingular for all x,
if is regarded as output of a system whose dynamics have the structure (11.19),
a system having vector relative degree {1, 1, . . . , 1} is obtained.
As a matter of fact, the equations thus found can be seen as characterizing a normal
form for such system. In fact, if we set5
z f 0 (z, ) g0 (z, )
z = , f0 (z, ) = , g 0 (z, ) =
F + G 0
and
q (z, ) = H + K + A(x), z, ) = B(x)
b(
z = f0 (z, ) + g 0 (z, )u
z, )u, (11.20)
= q (z, ) + b(
Lemma 11.2 Let T1 be defined as in Lemma 11.1. Consider system (11.19) and
suppose there exist a class K L function (, ) and a class K function () such
that, for any x(0) Rn and any admissible input function u() : [0, ) Rm , so
long as x(t) is defined the estimate
holds.
Proof As shown in Lemma 11.1, the matrix F in (11.18) is Hurwitz. If this is the
case, the linear system (11.18), viewed as a system with input and state , is input-
to-state stable. In particular, there exist positive numbers M, , L such that, for any
piecewise-continuous bounded function () : [0, ) Rm , the response (t) of
(11.18) from the initial state (0) satisfies
Bearing in mind the fact that is a linear function of (, ), it follows that (t)
satisfies an estimate of the form
so long as it is defined.
Combining the estimate (11.21) with the estimate (11.23), it is concluded that
there exist a class K L function (, ) and a class K function () such that, for
any x(0) Rn and any admissible input function u() : [0, ) Rm , so long as
x(t) is defined the an estimate of the form (11.22) holds.
System (11.20), which is nothing else than the original system (11.19) with the sets
of coordinates z and grouped together, regarded as a system with output has
vector relative degree {1, 1, . . . , 1}. If (11.19) is strongly minimum phase and the
matrix T (that defines the variable = T1 ) is chosen as indicated in Lemma 11.1,
then also system (11.20) is strongly minimum phase. Thus, the stabilization problem
can be simply handled by means of the procedures described in Chap. 10.
Specifically, the system can be globally stabilized by means of the full state
feedback law
u = [b( z, )]1 (q(z, ) )
u = [B(x)]1 (H K A(x) ).
and this, in view of the fact that system (11.20) is strongly minimum phase, suffices
to obtain global asymptotic stability of the equilibrium (z, ) = (0, 0) (see Proposi-
11.2 Stabilization by Partial-State Feedback 327
tion 10.1 in this respect). The control law in question, though, is not useful as it relies
upon exact cancelations and availability of the full state of the system. The results of
Chap. 10 have shown, however, that a more robust version of such control law can
be constructed, that yields asymptotic stability with a guaranteed region of attraction.
As a matter of fact, by appealing to the construction presented in Chap. 10, it
possible to claim that the desired goal can be achieved by means of a dynamic
control law of the form (10.7) and (10.8). In view of the fact that the vector relative
degree is now {1, 1, . . . , 1}, the argument (, ) of (10.7), in which is replaced
by , becomes b1 [ ] and the dynamic control law becomes6
u = G(b1 [ ])
= + bG(b1 [ ]) + c1 ( ) (11.24)
= 2 c0 ( ).
Then, results identical to those indicated in Propositions 10.2 and 10.3 hold. In
particular, so long as the issue of asymptotic stability is concerned, the following
result holds.
Assumption 11.1 There exist a constant nonsingular matrix b Rmm and a num-
z, ) in (11.20) satisfies
ber 0 < 0 < 1 such that the matrix b(
max z, ) b]b1 0
[b( for all (z, ). (11.25)
diagonal, 1
6 For
convenience, we write here instead of .
328 11 Stabilization of Multivariable Nonlinear Systems: Part II
Therefore, so long as all components of are available, the control law in question
can be regarded as robust. We will discuss in the next section how such partial state
can be estimated.
The stabilization method described in the previous section reposes on the availability
of the redesigned output . This, in turn, is a linear function of all d components of
the vector = (x) defined in (9.18). In this section, we show how this vector can
be approximately estimated, by means of a high-gain observer driven only by the
j
actual measured output y. We retain the assumption that in the Eq. (9.34) the ik (x)s
are constant coefficients.
The matter is to estimate the vectors 1 , 2 , . . . , . As far as the vector 1 is
concerned, observe that its components 1,1 , 1,2 , . . . , 1,r1 satisfy
1,1 = 1,2
1,2 = 1,3
1,r1 1 = 1,r1
1,r1 = a1 (x) + b1 (x)u.
in which the coefficients 1 and c1,0 , c1,2 , . . . , c1,r1 1 are design parameters. Defining
a vector e1 of estimation errors as (recall that y 1 = 1,1 )
r1 1
1 Im1 0 0 1,1 1,1
0 r1 2
0
e1 =
1 Im1 1,2 1,2 (11.27)
0 0 Im1 1,r1 1,r1
in which
c1,r1 1 Im1 Im1 0 0 0
c1,r1 2 Im1 0 Im1 0 0
M11 =
,
N1 =
.
c1,1 Im1 0 0 Im1 0
c1,0 Im1 0 0 0 Im1
Let the parameters c1,0 , c1,2 , . . . , c1,r1 1 be chosen in such a way that M11 is a Hurwitz
matrix.
As far as the set 2 is concerned, let this vector be split as col( 2 , 2,r2 ), as already
done before Lemma 11.1, and recall that
2 = 2 2,1 1,r1 .
This yields
2 = A2 2 + A2 2,1 1,r1 + B2 2,r2
2,r2 = a2 (x) + b2 (x)u.
To the purpose of expressing these equations in detail, observe that 21 and A2 2,1
can be split, consistently with the partition of 2 in r2 1 blocks, each one of
dimension m2 , as
1 2
21 =
r2 2 , A2 21 =
r2 1 .
r2 1 0
In this way, it is seen that the previous set of equations has the following structure:
2,1 0 Im2 0 0 2,1 2 1,r1
2,2 0 0
0 Im2 2,2 3 1,r1
... = ... + .
2,r2 1 0 0 0 Im2 2,r2 1 0
2,r2 0 0 0 0 2,r2 a2 (x) + b2 (x)u
This structure would suggests the use, for the estimation of 2,1 , , . . . , 2,r2 1 , 2,r2 ,
of a system of the form
330 11 Stabilization of Multivariable Nonlinear Systems: Part II
in which the coefficients 2 and c2,0 , c2,1 , . . . , c2,r2 1 are design parameters. In this
expression, however, 1,r1 and 2,1 are not directly available. Hence, we replace them
by the estimates provided by (11.26), i.e., we replace 1,r1 by 1,r1 and 2,1 , which by
definition is equal to y 2 1 1,r1 , by y 2 1 1,r1 . This yields
this yields
e 2 = 2 M22 e2 + M21 (2 )e1 + N2 [a2 (x) + b2 (x)u] (11.31)
in which
c2,r2 1 Im2 Im2 0 0 0
c2,r2 2 Im2 0 Im2 0 0
M22 =
,
N2 =
,
c2,1 Im2 0 0 Im2 0
c2,0 Im2 0 0 0 Im2
and
11.3 Stabilization via Dynamic Output Feedback 331
0 0 (2r2 1 2 2r2 c2,r2 1 1 )
0 0 (2r2 2 3 2r2 c2,r2 2 1 )
M21 (2 ) =
.
0 0 r2
(2 c2,1 1 )
0 0 (2r2 c2,0 1 )
Suppose, for the time being, that = 2. In this case, the design of the estimator
is concluded and we see that the estimation errors e1 , e2 satisfy
e 1 1 M11 0 e1 N1 0
= + [A(x) + B(x)u], (11.32)
e 2 M21 (2 ) 2 M22 e2 0 N2
in which M11 and M22 are Hurwitz matrices. System (11.32) can be seen as a linear
system driven by the input [A(x) + B(x)u]. With the results of Sect. 10.4 in mind, it
is important to prove a property similar to the property proven in Lemma 10.4, i.e.,
that, for any choice of a time T > 0 and a number , there is a choice of the gain
parameters 1 , 2 such that e(t) 2 for all t T . To this end, the following
result is useful.
Lemma 11.3 Consider a matrix of the form
A1 (1 ) 0
A(1 , 2 ) =
B(2 ) 2 A2
P1 A1 (1 ) + AT1 (1 )P1 1 I
Suppose 1 > 1 and compute the derivative of V (x) = x T Px along the solutions of
x = A(1 , 2 )x, to obtain
V 2 (x1 2 + x2 2 ),
a1 > 2
(11.33)
(a1 2 )2 > [P2 B(2 )]2 .
From this, it is seen that if 1 = 22r1 and a is large enough, there exist a number
2 such that (11.33) hold for all 2 > 2 , which proves the lemma.
p
Using this result, it is straightforward to conclude that, having chosen 1 as 1 = 2
(with p = 2r2 1, as shown in the proof of the lemma), there is a positive-definite
matrix P and a number 2 such that, if 2 > 2
T
1 M11 0 1 M11 0
P + P 2 I.
M21 (2 ) 2 M22 M21 (2 ) 2 M22
From this, the reader should have no difficulties to prove a result similar to that of
Lemma 10.4.
With this in mind, we return to the control law (11.24) which, as we have observed,
is driven by a quantity
K1 1 + 1,r1
= 1 =
2 K2 2 + 2,r2
that is not directly available for measurement. Following the design paradigm illus-
trated in Sect. 7.5, in the expression of i the (unavailable) states i and i,ri are
to be replaced by the estimates i and i,ri provided by the estimators (11.26)
and (11.29).
11.3 Stabilization via Dynamic Output Feedback 333
In the arguments used in Sect. 7.5, the actual control insisting on the system was
obtained through appropriate saturation functions, whose role was essentially that
of guaranteeing boundedness of the trajectories over the initial time interval [0, T0 ],
an interval of time during which, as a consequence of the choice of large value of ,
the estimation error e(t) could become excessively large (and hence, possibly, be the
cause of finite escape times). The same strategy should be followed in the present
context, and this suggests to replace the control (11.24) by a control of the form
u = G(b1 [ ])
= + bG(b1 [ ]) + c1 (S( ) ) (11.34)
= 2 c0 (S( ) ),
with 1 , 1,r1 , 2 , 2,r2 generated by (11.26) and (11.29). This results in a dynamical
system driven only by the measured output y.
From this, the reader should have no difficulties in arriving at conclusions essen-
tially identical to those of Propositions 10.2 and 10.3, i.e., to prove that the free
design parameters can be chosen in such a was as to secure semiglobal practical
stability, or even semiglobal asymptotic stability. Details are omitted and left to the
reader.
In the discussion above we have assumed, for convenience, = 2. The procedure
can be recursively extended to cases in which > 2. If = 3, one considers the
change of variables defined in (11.11) for i = 3, namely
3 = 3 31 1,r1 32 2,r2 .
7 This lemma has been, in fact, stated in such a way that a recursive usage is possible.
334 11 Stabilization of Multivariable Nonlinear Systems: Part II
which case the collection of all (9.34) characterizes a full normal form. The second
j
of these hypotheses is that the multipliers i,k (x) in (9.34) depend on the various
components of x in a special triangular fashion. To simplify matters, we consider
the case of a system having only two inputs and outputs8 and we use xi,k instead
of i,k to denote the various components of the state vector x. For convenience, we
let [xi ]k denote the vector
The model of the system consists in the aggregate of a set of the form (9.34)
written for i = 1, namely
x 1,1 = x1,2
x 1,2 = x1,3
(11.35)
x 1,r1 1 = x1,r1
x 1,r1 = a1 (x) + b1 (x)u
y1 = x1,1
and of a set of the form (9.34) written for i = 2 (note that r2 > r1 ). For the latter
set, it is assumed that the multiplier 2,k
1
(x) only depends on x1 and on [x2 ]k1 . More
explicitly, it is assumed that the set in question has the following structure:
x 2,1 = x2,2
x 2,2 = x2,3
x 2,r1 1 = x2,r1
x 2,r1 = x2,r1 +1 + 2,r1 +1 (x1 , [x2 ]r1 )(a1 (x) + b1 (x)u) (11.36)
x 2,r2 1 = x2,r2 + 2,r2 (x1 , [x2 ]r2 1 )(a1 (x) + b1 (x)u)
x 2,r2 = a2 (x) + b2 (x)u
y2 = x2,1 .
The relevance of this particular structure, in the context of the problem of designing
feedback stabilizing laws, resides in the facts that, for such special class of systems:
(i) there exists a state feedback law that globally stabilizes the equilibrium x = 0,
(ii) the state x can be expressed as a function of the derivatives of the output y with
respect to time, up to order r2 1. These two facts, together, make it possible to
develop an output feedback stabilization scheme, as shown in the second part of this
section. For the time being, we proceed with the proof of the two properties outlined
above.9
Lemma 11.4 There exists a feedback law u = (x) that globally asymptotically
stabilizes the equilibrium x = 0 of (11.35) and (11.36).
Proof Consider the set (11.35). Pick a function v(x1 ) so that the equilibrium x1 = 0
of
x 1,1 = x1,2
x 1,r1 1 = x1,r1
x 1,r1 = v1 (x1 )
x 2,1 = x2,2
x 2,2 = x2,3
x 2,r1 1 = x2,r1
x 2,r1 = x2,r1 +1 + 2,r1 +1 (x1 , [x2 ]r1 )v1 (x1 )
x 2,r2 1 = x2,r2 + 2,r2 (x1 , [x2 ]r2 1 )v1 (x1 )
x 2,r2 = a2 (x) + b2 (x)u.
This system has a structure similar to that of the system considered in Remark 6.8. In
particular, in the jth equation, the variable x2,j+1 can be regarded as a virtual control.
Thus, the reader should have no difficulties in proving the existence of a function
v2 (x1 , x2 ) such that, if
a2 (x) + b2 (x)u = v2 (x1 , x2 ), (11.38)
is also easy to check that, if the 2,k (x)s have the indicated structure, the system is uniformly
9 It
x1i = y1(i1)
for i = 1, . . . , r1 . (11.40)
x2i = y2(i1)
in which the various components of the arguments x1 and [x2 ]r1 of 2,r1 +1 () coincide
with y1 , . . . , y1(r1 1) and, respectively, with y2 , . . . , y2(r1 1) , as shown in (11.40). Thus,
it is concluded that there exists a function 2,r1 +1 () such that
x2,r1 +1 = 2,r1 +1 y1 , . . . , y1(r1 ) , y2 , . . . , y2(r1 ) . (11.41)
x2,r1 +2 = x 2,r1 +1
2,r1 +2 (x1 , [x2 ]r1 +1 )[a1 (x)+ b1 (x)u]
= 2,r1 +1 y1 , . . . , y1(r1 ) , y2 , . . . , y2(r1 ) 2,r1 +2 (x1 , [x2 ]r1 +1 )y1(r1 ) .
The first term 2,r1 +1 () is a function y1 , . . . , y1(r1 +1) , y2 , . . . , y2(r1 +1) , while the argu-
ments x1 and [x2 ]r1 +1 of 2,r1 +2 () are functions of y1 , . . . , y1(r1 ) , y2 , . . . , y2(r1 ) , as shown
in (11.40) and (11.41). Thus it is concluded that there exists a function 2,r1 +2 ()
such that
x2,r1 +2 = 2,r1 +2 y1 , . . . , y1(r1 +1) , y2 , . . . , y2(r1 +1) . (11.42)
The procedure can be iterated, until the existence of a map 2,r2 () is shown, such
that
x2,r2 = 2,r2 y1 , . . . , y1(r2 1) , y2 , . . . , y2(r2 1) . (11.43)
Putting all expressions (11.40), (11.41), (11.42), . . . , (11.43) together proves the
lemma.
The two properties just proven show that the system can be stabilized by means of a
feedback law of the form u = ( (y1r2 , y2r2 )), in which y1r2 and y2r2 could be estimated
by means of a high-gain observer. In this respect, though, it must be stressed that the
arguments of () consist of y1 , y2 and all their higher order derivatives up order
r2 1. In particular, this requires the estimation of the derivatives of y1 from order
11.4 Handling More General Classes of Invertible Systems 337
r1 to order r2 1 and such derivatives, in turn, depend on the input u and a few of
its higher order derivatives, up to order r2 r1 1. To circumvent this problem, it is
convenient to dynamically extend the system, by adding a chain of r2 r1 integrators
on both input channels.10 In other words, the system is extended by setting
u = 1
1 = 2
(11.44)
r2 r1 1 = r2 r1
r2 r1 = v,
i1 [bi (x)1 ]
i2 (x, 1 ) = [ f (x) + g(x)1 ] + [ f (x) + g(x)1 ]
x x
i,j i,j
j1
[bi (x)j ]
i,j+1 (x, 1 , . . . , j ) = [ f (x) + g(x)1 ] + k+1 + [ f (x) + g(x)1 ].
x k x
k=1
It is easy to check that the change of variables thus defined is invertible. In fact, using
the property that the matrix B(x) is invertible, it is seen that11
1 = B1 (x)[1 1 (x)]
2 = B1 (x)[2 2 (x, 1 )]
r2 r1 = B1 (x)[r2 r1 r2 r1 (x, 1 , . . . , r2 r1 1 )].
1 = A(x) + B(x)1
for j = 1, . . . , r2 r1 1. Moreover
r2 r1 = r2 r1 +1 (x, 1 , . . . , r2 r1 ) + B(x)v,
which we rewrite as
r2 r1 = r2 r1 +1 (x, ) + B(x)v
and
x 2,1 = x2,2
x 2,r1 1 = x2,r1
x 2,r1 = x2,r1 +1 + 2,r1 +1 (x1 , [x2 ]r1 )1,1
x 2,r2 1 = x2,r2 + 2,r2 (x1 , [x2 ]r2 1 )1,1 (11.46)
x 2,r2 = 2,1
2,1 = 2,2
2,r2 r1 = 2,r2 r1 +1 (x, ) + b2 (x)v
y2 = x2,1
This system has a structure similar to that of (11.35) and (11.36). Therefore, a
result similar to that of Lemma 11.4 holds.
Lemma 11.6 There exists a feedback law v = (x, ) that globally asymptotically
stabilizes the equilibrium (x, ) = 0 of (11.45) and (11.46).
12 As usual, let
= col(1 , . . . , r2 r1 )
= col(1 , . . . , r2 r1 ).
.
11.4 Handling More General Classes of Invertible Systems 339
Note that since can be expressed as a function of and x, as shown above, the
feedback law determined in this lemma can also be expressed as a function of x and
. In other words, the lemma says that there exists a function v (x, ) such that, if
system (11.35) and (11.36), extended by (11.44), is controlled by
v = v (x, ).
A little computational effort is needed to find an equivalent relation for y2(r2 ) . To this
end, observe first of all that
Thus
y2(r1 +1) = x2,r1 +2 + 2,r1 +2 (x1 , [x2 ]r1 +1 )1,1 +
+ 2,r1 +1 (x1 , [x2 ]r1 )1,1 + 2,r1 +1 (x1 , [x2 ]r1 )1,2 .
The quantity 2,r1 +1 (x1 , [x2 ]r1 ) is a function of x1 , [x2 ]r1 +1 , 1,1 . Thus
y2(r1 +1) = x2,r1 +2 + r1 +2 (x1 , [x2 ]r1 +1 , 1,1 ) + 2,r1 +1 (x1 , [x2 ]r1 )1,2
for some function r1 +2 (x1 , [x2 ]r1 +1 , 1,1 ). Next, we see that
y2(r1 +2) = x2,r1 +3 + r1 +3 (x1 , [x2 ]r1 +2 , 1,1 , 1,2 ) + 2,r1 +1 (x1 , [x2 ]r1 )1,3
for some function r1 +3 (x1 , [x2 ]r1 +2 , 1,1 , 1,2 ). Continuing in the same way, one
obtains for y2(r2 ) an expression of the form
In summary, we find that y1(r2 ) and y2(r2 ) can be given expressions of the form (we have
replaced by its expression as function of (x, ))
340 11 Stabilization of Multivariable Nonlinear Systems: Part II
Note that such expressions involve the input v but not its derivatives.
The quantities yi , yi(1) , . . . , yi(r2 1) , needed for the implementation of the feed-
back law, are estimated by the states i,1 , i,2 , . . . , i,r2 of two identical high-gain
observers modeled by equations of the form
References
1. D. Liberzon, Output-input stability implies feedback stabilization. Syst. Control Lett. 53, 237
248 (2004)
2. L. Wang, A. Isidori, H. Su, Global Stabilization of a Class of Invertible MIMO Nonlinear
Systems. IEEE Trans. Automat. Control 60, 616631 (2015)
3. L. Wang, A. Isidori, L. Marconi, H. Su, Stabilization by output feedback of multivariable invert-
ible nonlinear systems (submitted)
4. A.R. Teel, L. Praly, Tools for semi-global stabilization by partial state and output feedback.
SIAM J. Control Optimiz. 33, 14431488 (1995)
5. D. Liberzon, A.S. Morse, E.D. Sontag, Output-input stability and minimum-phase nonlinear
systems. IEEE Trans. Autom. Control 47, 422436 (2002)
Chapter 12
Regulation and Tracking in Nonlinear
Systems
12.1 Preliminaries
The nonlinear equivalent of the design problem addressed in Chap. 4 can be cast
in the following terms. The controlled plant is a finite-dimensional, time-invariant,
nonlinear system modeled by equations of the form
x = f (w, x, u)
e = he (w, x) (12.1)
y = h(w, x),
lim e(t) = 0,
t
w = s(w) (12.2)
obtained when its initial condition w(0) is allowed to vary on a prescribed set W . For
convenience, it is assumed that the set W on which the state of (12.2) is allowed to
range is a compact set, invariant for the dynamics of (12.2). If this is the case, then
the -limit set of W under the flow of (12.2) is the set W itself.1 If this is the case,
it can be said that the exosystem is in steady-state and this can be regarded as a
nonlinear counterpart of the assumption that all the eigenvalues of a linear exosystem
have zero real part.
The control law for (12.1) is to be provided by a system modeled by equations of
the form
x c = f c (xc , y)
(12.3)
u = hc (xc , y)
with state xc Rnc . The initial conditions x(0) of the plant (12.1), and xc (0) of
the controller (12.3) are allowed to range over a fixed compact sets X Rn and,
respectively Xc Rnc .
In this setting, the problem of output regulation can be cast as follows. Consider
the closed-loop system
w = s(w)
x = f (w, x, hc (xc , h(w, x))) (12.4)
x c = f c (xc , h(w, x)),
e = he (w, x).
B.
12.1 Preliminaries 343
This is equivalent to assume that, in the closed-loop system, for each given exogenous
input function w(t), there exists a unique steady-state response, which therefore can
be expressed as
xss (t) = (w(t))
xc,ss (t) = c (w(t)).
Moreover, for convenience, we also assume that the maps (w) and c (w) are con-
tinuously differentiable. This enables us to characterize in simple terms the property
that the steady-state locus is invariant under the flow of the closed-loop system (12.4).
If this is the case, in fact, to say that the locus (12.5) is invariant under the flow of
(12.4) is the same as to say that (w) and c (w) satisfy4
s(w) = f (w, (w), hc (c (w), h(w, (w))))
w w W . (12.6)
c
s(w) = f c (c (w), h(w, (w)))
w
These equations are the non-linear counterpart of the Sylvester equations (4.17).
From item (ii) in the characterization of the problem of output regulation, it is
also seen that limt e(t) = 0, uniformly in the initial condition. This entails an
important consequence on the set (W X Xc ), expressed as follows.5
lim e(t) = 0,
t
In other words, if the controller solves the problem of output regulation, the steady-
state locus, which is asymptotically approached by the trajectories of the closed-loop
system, must be a subset of the set of all pairs (w, x) for which he (w, x) = 0 and
hence the map (w) necessarily satisfies
the first equation of (12.6) and (12.8) can be rewritten in controller-independent form
as
s(w) = f (w, (w), (w))
w w W . (12.10)
0 = he (w, (w))
These equations, introduced in [2] and known as the nonlinear regulator equations,
are the nonlinear counterpart of the Francis equations (4.6).
In order to better understand the meaning of the second equation in (12.6), it is
convenientas done in a similar context for linear systemsto split the vector y of
measured outputs in two parts as
e he (w, x)
y= =
yr hr (w, x)
x c = f c (xc , e, yr )
u = hc (xc , e, yr ).
Using these notations, the second equation in (12.6) and (12.9), in the light of
(12.8), become
c
s(w) = f c (c (w), 0, hr (w, (w)))
w w W . (12.11)
(w) = hc (c (w), 0, hr (w, (w)))
12.2 Steady-State Analysis 345
These equations are the nonlinear counterpart of the equations (4.20) and have a
similar interpretation. Observe also that, in the special case in which y = e, the
equations (12.11) reduce to equations of the form
c
s(w) = f c (c (w), 0)
w (12.12)
(w) = hc (c (w), 0).
At is was the case for linear systems, the conditions thus found can be exploited in
the design of a controller that solves the problem of output regulation. The substantial
difference with the case of linear systems, though, is that one no longer can appeal to
results such as those used in Sect. 4.4, which were dependent on some fundamental
properties of linear systems. In fact, the arguments used in the proof of Proposition 4.3
consisted of Lemma 4.2, expressing the properties of stabilizability and observability
of the (linear) augmented system (4.30), and of Lemma 4.3, which was proved using
(4.27). Both such arguments that have no general counterpart for nonlinear systems.
In addition, it is stressed that, since the controlled system is affected by the exogenous
input w and the latter is not available for feedback, stabilization methods that do not
rely upon measurement of w, i.e., robust stabilization methods, are mandatory.
This being the case, we consider in what follows only the case in which y = e
(and, consistently, we identify he (w, x) with h(w, x)) and we pursue the method
indicated in the second part of Sect. 4.6. In fact, the proof that the method in question
is successful does not appeal to the result of Lemma 4.3. Mimicking the design
philosophy used in that context, where the controller consisted of a preprocessing
internal model of the form (4.53) with input u provided by a stabilizer of the form
(4.54), we consider now a controller consisting of a preprocessing internal model
described by equations of the form
= () + Gu
(12.13)
u = () + u
and this property was instrumental in determiningin the proof of Proposition 4.6
the expression of the center eigenspace in the associated closed-loop system (4.55).
In order to obtain a similar result in the present context, we postulate (for the time
being) the fulfillment of equations that can be seen as a nonlinear counterpart of the
two linear equations above, i.e., we assume the existence of a map : W Rd
346 12 Regulation and Tracking in Nonlinear Systems
satisfying
s(w) = ( (w))
w w W . (12.15)
(w) = ( (w))
= As + Bs e
(12.16)
u = Cs + Ds e.
w = s(w)
x = f (w, x, () + Cs + Ds e)
= () + G(Cs + Ds e) (12.17)
= As + Bs e
e= h(w, x).
Using (12.10) and (12.15), it is an easy matter to check that the graph of the
(nonlinear) map
x (w)
w W
= (w) ,
0
6 Strictlyspeaking, it is not necessary to consider here a linear stabilizer. In fact, as it will be seen
in the sequel, it suffices that the stabilizer has an equilibrium state yielding u = 0. However, since
essentially all methods illustrated earlier in the book for nonlinear (robust) stabilization use linear
stabilizers, in what follow we will consider a stabilizer of this form.
7 Compare with a similar conclusion obtained in the proof of Proposition 4.6.
12.2 Steady-State Analysis 347
Assumption 12.2 There exist a pair ((), ()) such that (12.15) hold for some
(w).
w = s(w)
z = f 0 (w, z, 1 , . . . , r )
1 = 2
(12.19)
r1 = r
r = q(w, z, 1 , . . . , r ) + b(w, z, 1 , . . . , r )u
e = 1
in which r is the relative degree of the system, z Rnr and b(w, z, 1 , . . . , r ), the
so-called high-frequency gain, is nowhere zero.
If the model of the plant is available in normal form, the nonlinear regulator
equations (12.10) can be analyzed by means a technique that extends the technique
used in Example 4.1. Let (w) be partitioned, consistently with the partition of the
state (z, 1 , . . . , r ) of (12.19), into
0
s(w) = f 0 (w, 0 (w), 1 (w), . . . , r (w))
w
i
s(w) = i+1 (w) i = 1, . . . , r 1
w
r
s(w) = q(w, 0 (w), 1 (w), . . . , r (w)) + b(w, 0 (w), 1 (w), . . . , r (w))(w)
w
0 = 1 (w).
1 (w) = = r (w) = 0
0
s(w) = f 0 (w, 0 (w), 0, . . . , 0) (12.20)
w
Moreover
q(w, 0 (w), 0, . . . , 0)
(w) = . (12.21)
b(w, 0 (w), 0, . . . , 0)
In summary, the nonlinear regulator equations have a solution if and only if there
exists a map 0 (w) that satisfies (12.20). If this is the case, the map (w) is given
by (12.21).8
We are now in a position to appropriately formulate the assumption of minimum
phase. The difference with the case dealt with in Sect. 6.3 is that now we should no
longer consider a property of input-to-state stability to an equilibrium, but rather to
a compact invariant set. In fact, as shown by (12.20) the set
A0 = {(w, z) : w W , z = 0 (w)}
w = s(w)
(12.22)
z = f 0 (w, z, 1 , . . . , r )
still has relative degree r, but is not in strict normal form. In order to determine
whether or not the property of being minimum phase holds it is appropriate to put
first the system in strict normal form, which is not difficult.
To this end, set
F() = () G () (12.25)
= F() + G[ () + u ].
= GM(w, z, ) (12.26)
in which r
1
M(w, z, ) = ds. (12.27)
0 b(w, z, 1 , . . . , r1 , s)
Note that M(w, z, ) vanishes at r = 0 and that its derivative with respect to time
has an expression of the form
r
d 1 d 1
M(w, z, ) = r + ds
dt b(w, z, 1 , . . . , r1 , r ) 0 dt b(w, z, 1 , . . . , r1 , s)
q(w, z, )
= + [ () + u ] + N(w, x, )
b(w, z, )
350 12 Regulation and Tracking in Nonlinear Systems
in which
r
r1
1 b b b
N(w, x, ) = s(w) + f 0 (w, z, ) + i+1 ds,
0 [b(w, z, 1 , . . . , r1 , s)]2 w z i
i=1
q(w, z, )
= F( + GM(w, x, )) G GN(w, x, ),
b(w, z, )
w = s(w)
z = f 0 (w, z, )
q(w, z, )
= F( + GM(w, x, )) G GN(w, x, )
b(w, z, )
1 = 2 (12.28)
r1 = r
r = q(w, z, ) + b(w, z, )[ ( + GM(w, x, )) + u ]
e = 1 ,
w = s(w)
z = f 0 (w, z, )
q(w, z, ) (12.29)
= F( + GM(w, x, )) G GN(w, x, )
b(w, z, )
viewed as a system with state (w, z, ) and input . We observe that, if Assumption
12.2 holds, the set
0
f 0 (w, 0 (w), 0) = s(w)
w
and, respectively, the right-hand side of the third equation becomes (use here the fact
that M(w, z, 0) = 0 and N(w, z, 0) = 0 and (12.21) and (12.15))
12.3 The Case of SISO Systems 351
q(w, 0 (w), 0)
F( (w)) G = ( (w)) G ( (w)) + G(w) = s(w).
b(w, 0 (w), 0) w
Thus, it makes sense to consider the case in which the subsystem in question
is input-to-state stable to the set A0a . As far as the z component is concerned, the
required property is precisely the property considered in Definition 12.1. Thus, it
remains to postulate a similar property for the component, which we do in the form
of an assumption.
Assumption 12.3 Set (t) = (t) (w(t)). There exist a class K L function
1 (, ) and a class K function 1 () such that, for any (w0 , 0 ) W Rd , any
piecewise-continuous bounded function (w0 (), z0 ()) : [0, ) W Rnr and
any continuous bounded function 0 () : [0, ) Rr , the response (w(t), (t)) of
(12.29) from the initial state (w(0), (0)) satisfies
Clearly, if the system (12.19) is strongly minimum phasein the sense of Defin-
ition 12.1and Assumption 12.3 holds, then also the augmented system (12.28) is
minimum phase as well. As a matter of fact, composing the estimates in (12.23) and
(12.30), it is easily seen that the quantity
z z 0 (w)
w, z, )A0a := =
(w)
(w, z, )A0a (t) ((w, z, )A0a (0), t) + (0 ()[0,t] )} for all t 0
As shown in Sect. 5.6, in the case of a linear system Assumptions 12.2 and 12.3 can
always be fulfilled. In fact, choosing and G as in (4.23) and (4.24) and picking
in such a way that the matrix F is Hurwitz, the (linear version of) Assumption 12.2
is fulfilled and the preprocessing internal model can be expressed as
= F + [ + u ]
u = + u .
The functions M(w, z, ) and N(w, z, ) introduced in the change of variables (12.26)
become
1
M(w, z, ) = r N =0
b
while (see Example 4.1)
q(w, z, ) 1 1
= [P1 w + A10 z + A11 ] = w + [A10 (z 0 w) + A11 ]
b(w, z, ) b b
1
= w + [A10 z + A11 ].
b
As a consequence, the dynamics of become
1 1
= F( + r ) + G w G [A10 z + A11 ],
b b
and, since (use here (12.14))
F + G = (F + G ) = = S,
1
= F + [Fr GA10 z GA11 ].
b
The matrix F is Hurwitz by construction and hence Assumption 12.3 trivially holds.
This construction has a nice nonlinear counterpart, which is based on the following
important result, proven in [12].
Lemma 12.2 Let d 2nw + 2. There exist an > 0 and a subset S C of zero
Lebesgue measure such that if the eigenvalues of F are in { C : Re[] }\S,
then there exist a differentiable function : W Rd and a continuous bounded
function : Rd R such that9
9 The function () is only guaranteed to be continuous and may fail to be continuously differentiable.
Closed-forms expressions for () and other relevant constructive aspects are discussed in [13].
12.4 The Design of an Internal Model 353
s(w) = F (w) + G(w)
w for all w W . (12.31)
(w) = (( w))
With this result in mind it is easy to see how the properties indicated in Assump-
tions 12.2 and 12.3 can be fulfilled. In fact, choosing
() = F + G ()
the property in Assumption 12.2 trivially holds. With such choice, the function F()
in (12.25) becomes a linear function, i.e., F() = F. Then, the dynamics of =
(w) become
q(w, z, )
= F + F (w) + FM(w, z, ) G GN(w, z, ) s(w)
b(w, z, ) w
q(w, z, )
= F + F (w) + FM(w, z, ) G GN(w, z, ) F (w) G(w)
b(w, z, )
q(w, z, )
= F + FM(w, z, ) GN(w, z, ) G[ + (w)].
b(w, z, )
q(w, z, )
v(w, z, ) = FM(w, z, ) GN(w, z, ) G[ + (w)] (12.32)
b(w, z, )
v(w, z, ) max{( z)},
), ( (12.33)
for some class K function (), also the property indicated in Assumption 12.2
holds.
The above setup is pretty general and, essentially, does not require specific assump-
tions (other than (12.33)). However, the construction of the function () is not
immediate. To overcome this difficulty, another approach is available, which how-
ever requires a specific assumption, that essentially postulates the existence of a
regression-like relation between the higher derivatives (with respect to time) of
the function (w).
Assumption 12.4 There exists an integer d and a globally Lipschitz smooth function
: Rd R such that
Under this assumption,10 it is very easy to construct a pair of functions ((), ())
such that the properties indicated in Assumption 12.2 hold for some (w). In fact,
the reader should have no difficulties in checking that this is the case if
2
3
() = = A
+ B()
d
(1 , 2 , . . . , d )
() = 1 = C
B,
with A, C matrices of the form (3.8) and
(w)
Ls (w)
(w) =
.
Lsd1 (w)
This being the case, it remains to see whether the vector G can be chosen in such
a way that also Assumption 12.3 holds. To this end observe that, if () and ()
are chosen in this way, then
F() = () G () = (A GC)
+ B().
= (A GC)(
+ (w) + M(w, z, )) + B( + (w) + M(w, z, ))
q(w, z, )
G GN(w, z, ) s(w)
b(w, z, ) w
= (A GC)(
+ M(w, z, )) + (A GC) (w) + B( + (w) + M(w, z, ))
q(w, z, )
G GN(w, z, ) A (w) B( (w))
b(w, z, )
= (A GC)
+ B[(
+ (w) + M(w, z, )) ( (w)]
q(w, z, )
+ (A GC)M(w,
z, ) G[ + (w)] GN(w, z, ).
b(w, z, )
= (A GC)
+ B[(
+ (w) + M(w, z, )) ( (w)] + v(w, z, ),
10 SinceW is a compact set, in the condition above only the values of () on a compact set matter.
Thus, the assumption that the function is globally Lipschitz can be taken without loss of generality.
12.4 The Design of an Internal Model 355
G = D G0
in which D is the matrix (compare with Sect. 2.5 and with Sect. 7.5)
D = diag(, 2 , . . . , d ) (12.35)
= D1 .
B,
Bearing in mind the definition of A, C and D , observe that
D1 AD D1 B = 1 B,
= A, CD
= C, (12.36)
d
from which we obtain
= (A G0 C)
+ 1 B[(D
1
+ (w) + M(w, z, )) ( (w)] + D v(w, z, ).
d
+ (A G0 C)
P(A G0 C) T P = I.
Moreover, bearing in mind the fact that the function () is globally Lypschitz and
assuming, without loss of generality, that 1, it is seen that
1
B[(D + (w) + M(w, z, )) ( (w)]
d
1 1
d LD + M(w, z, ) L + d LM(w, z, ),
for some L > 0. Thus, along trajectories of the system, the function V ( ) = T P
satisfies
V ( ) 2 + 2 P L
+ LM(w,
z, ) + v(w, z, ) .
This being the case, it is clear that, if is large enough, conclusions similar to those
obtained in the previous case hold. More specifically, define
v (w, z, ) = LM(w, z, ) + v(w, z, ),
observe that
356 12 Regulation and Tracking in Nonlinear Systems
2 P LM(w, z, ) + v(w, z, ) 2 P2 + v(w, z, )2
choose so that
a = 2LP P2
is positive. Then,
V ( ) a 2 + v(w, z, )2 .
From this, it is concluded that if v (w, z, ) can be estimated as in (12.33), also the
property indicated in Assumption 12.3 holds.11
Example 12.1 Consider a controlled Van der Pol oscillator
x 1 = x2
(12.37)
x 2 = x1 + (1 x12 )x2 + u
w = Sw
1 = 2
(12.38)
2 = q(w, ) + u
e= 1 .
in which
Since in this case dim(z) = 0, the nonlinear regulator equations do have a trivial
solution, in which (w) = 0 and
11 Note that G is present in the function v(w, z, ) and hence in v (w, z, ). Thus, this function depends
on and, actually, its magnitude grows with . However, this does not affect the conclusion. Once
is fixed, the bound on v (w, z, ) is fixed as well. The gain function 1 () in (12.30) is influenced
by the value of , but an estimate of this form holds anyway.
12.4 The Design of an Internal Model 357
in which
= (1 a2 ) a a 0 0 0 .
in which
0 a 0 0 0 0
a 0 0 0 0 0
0 0 0 2a a 0
S=
0
.
0 2a 0 0 a
0 0 3a 0 0 0
0 0 0 3a 0 0
(1 , 2 , 3 , 4 ) = 9a4 1 10a2 3 .
w = Sw
= + B()
A + Gu
1 = 2 (12.39)
2 = + u ]
q(w, ) + [C
e= 1 .
This is a system having relative degree 2, but not in strict normal form. To get a strict
normal form, we change into
= G2
and obtain
w = Sw
= (A GC)(
+ G2 ) + B( + G2 ) Gq(w, )
1 = 2 (12.40)
2 = + G2 ) + u ]
q(w, ) + [C(
e= 1 .
satisfies
(w) + B(
s(w) = A (w))
w
(w) = C (w).
which we rewrite as
= (A GC)
+ B(
) + v(w, )
= (A GC)
+ B(
)
On this manifold the regulation error e is zero and hence the problem of output
regulation is solved, by the controller
+ B()
= A + D G0 [Cs + Ds e]
= As + Bs e
+ Cs + Ds e.
u = C
We conclude the chapter with a concise discussion of how the results presented in
Chap. 5 can be extended to the case in which consensus is sought in a network of
nonlinear agents. If the network is heterogeneous, mimicking the approach described
in Sect. 5.6, the problem can be addressed in two steps: inducing consensus among
360 12 Regulation and Tracking in Nonlinear Systems
a set of identical local reference generators and then using (decentralized) output
regulation theory to synchronize the output of each agent with that of the local
reference generator. The second step of this procedure only entails simple adaptations
of the design method described earlier in this chapter. Therefore, in this section, we
limit ourselves to show how the first step of the procedure can be completed. We
refer the reader to Chap. 5 for all details and notations about networks of agents and
interconnections.
Consider the problem of achieving consensus in a network of N identical nonlinear
agents modeled by equations of the form12
x k = f (xk ) + uk
(12.41)
yk = h(xk ),
in which xk Rn , uk R, yk R and
+ B(x)
f (x) = Ax
h(x) = Cx,
B,
with A, C matrices of the form (2.7), and (x) a nonlinear function.
The control of such systems is chosen as
uk = Kvk
N
vk = akj (yj yk ). (12.42)
j=1
Assumption 12.5 The function (x) is globally Lipschitz and there exists a compact
set X Rn , invariant for x = f (x), such that the system
x = f (x) + v
12 The model (12.41) should be compared with the case of a model (5.24) with B = I, that is precisely
K = D K0 ,
N
k + B(x
x k = Ax k ) D K0
j.
kj Cx (12.43)
j=1
x = (IN A) (L D K0 C)
x + (IN B)(x)
where
(x) = col((x1 ), . . . , (xN )).
x = (T 1 In ) (IN A)
(L D K0 C)
(T In )x + (T 1 In )(IN B)((T
In )x )
(L D K0 C)
= (IN A) 1
x + (T B)((T In )x ).
x = col(x1 , x2 x1 , . . . , xN x1 )
and define
x = col(x2 x1 , x3 x1 , . . . , xN x1 ).
1 + B(x
x 1 = Ax
1 ) (L12 D K0 C)x
(L22 D K0 C)
x = (IN1 A) x + (x1 , x )
in which
(x2 ) (x1 )
(x3 ) (x1 ) .
(x1 , x ) = (IN1 B)
(xN ) (x1 )
Lemma 12.3 Suppose Assumption 12.5 hold. Suppose the interconnection graph is
connected. Let P be the unique positive definite symmetric solution of the algebraic
Riccati equation
+ PA T 2PC T CP
AP + aI = 0
K0 = PC T . (12.44)
Then, there is a number such that, if , for any initial condition x1 (t) is a
bounded function and the equilibrium x = 0 of
(L22 D K0 C)
x = (IN1 A) x + (x1 , x )
= (IN1 D1 )x ,
= L
(x1 , (IN1 D ) ) LD n .
V ( ) 2 + 2LQ 2 .
12.5 Consensus in a Network of Nonlinear Systems 363
Form this it is deduced that, if > 2LQ, as long as trajectories are defined, V ( (t))
is steadily decreasing. Let be fixed in this way. Since x1 obeys
1 + B(x
x 1 = Ax
1 ) (L12 D K0 CD ),
using Assumption 12.5, it is seen that x1 (t) is bounded for all t 0 (by a quantity
that only depends on initial conditions and ). Thus. trajectories of the entire system
are defined for all t 0 and the equilibrium = 0 is globally exponentially stable,
with a (fixed) quadratic Lyapunov function.
Proposition 12.1 Let the hypotheses of the previous lemma hold and let K0 be chosen
as in (12.44). There is a number > 0 such that, if , the states of the N
systems (12.43) reach consensus, i.e., for every xk (0) Rn , k = 1, . . . , N, there is a
function x : R Rn such that
References
14. A. Isidori, C.I. Byrnes, Steady-state behaviors in nonlinear systems, with an application to
robust disturbance rejection. Annu. Rev Control 32, 116 (2008)
15. A. Isidori, L. Marconi, G. Casadei, Robust output synchronization of a network of hetero-
geneous nonlinear agents via nonlinear regulation theory. IEEE Trans. Autom. Control 59,
26802691 (2014)
Appendix A
Background Material in Linear
Systems Theory
In this section, a few fundamental facts about symmetric matrices and quadratic
forms are reviewed.
Symmetric matrices. Let P be a n n symmetric matrix of real numbers (that is, a
matrix of real numbers satisfying P = PT ). Then there exist an orthogonal matrix Q
of real numbers1 and a diagonal matrix of real numbers
1 0 0
0 2 0
=
0 0 n
such that
Q1 PQ = QT PQ = .
PQ = Q
P1 Q = Q1 ,
Q1 = QT .
Springer International Publishing Switzerland 2017 365
A. Isidori, Lectures in Feedback Design for Multivariable Systems,
Advanced Textbooks in Control and Signal Processing,
DOI 10.1007/978-3-319-42031-8
366 Appendix A: Background Material in Linear Systems Theory
from which it is seen that 1 is a Jordan form of P1 and the ith column qi of Q is
also an eigenvector of P1 , associated with the ith eigenvalue 1
i .
n
n
V (x) = x Px =
T
pij xi xj
i=1 j=1
1 T T 1
V (x) = x Px = [x Px + x Px] = x
T T
(P + P ) x
T
2 2
Note that the inequalities are tight (hint: pick, as x, the last and, respectively, the first
column of Q).
Sign-definite symmetric matrices. Let P be symmetric. The matrix P is said to be
positive semidefinite if
x T Px 0 for all x.
We see from the above that P is positive semidefinite if and only if min 0 and
is positive definite if and only if min > 0 (which in turn implies the nonsingularity
of P).
There is another criterion for a matrix to be positive definite, that does not require
the computation of the eigenvalues of P, known as Sylvesters criterion. For a sym-
metric matrix P, the n minors
Appendix A: Background Material in Linear Systems Theory 367
p11 p12 p13
p11 p12
D1 = det(p11 ), D2 = det , D3 = p21 p22 p23 , . . .
p21 p22
p31 p32 p33
Proof Observe that a necessary condition for (A.1) to be positive definite is R > 0.
Hence R is nonsingular and (A.1) can be transformed, by congruence, as
T
I 0 Q S I 0 Q SR1 S T 0
= .
R1 S T I ST R R1 S T I 0 R
stand for the matrix P R is positive semidefinite and, respectively, for the matrix
P R is positive definite.
in which the off-diagonal elements are negative, is positive definite (use Sylvesters criterion above).
368 Appendix A: Background Material in Linear Systems Theory
P = MTM
In this section, we discuss the existence of solutions of two relevant linear matrix
equations that arise in the analysis of linear systems. One of such equation is the
so-called Sylvesters equation
AX XS = R (A.3)
Another instance in which an equation of this kind arises is the analysis of the stability
of a linear system, where this equation assumes the special form AX + XAT = Q,
known as Lyapunovs equation.
Another relevant linear matrix equation is the so-called regulator or Franciss
equation
S = A + B + P
(A.4)
0 = C + Q
A1 = A, q1 () = 1, A2 = I, q2 () = ,
Equations of the form (A.5) are also known as Hautus equations.4 Noting that
the left-hand side of (A.5) can be seen as a linear map
d
H : Rm Rn d
: X H (X) := A1 Xq1 (S) + + Ak Xqk (S),
A() = A1 q1 () + + Ak qk ()
are linearly independent for each which is an eigenvalue of S. If this is the case
and n = m,
the solution X of (A.5) is unique.
4 See [3].
5 Note that, if this is the case and n = m,
the map is also injective, i.e., it is an invertible linear map.
In this case the solution X of (A.5) is unique.
370 Appendix A: Background Material in Linear Systems Theory
are linearly independent for each which is an eigenvalue of S. If this is the case
and m = p, the solution pair (, ) is unique.
is negative definite. Then, all eigenvalues of the matrix A have negative real part.
Proof Let be an eigenvalue of A and x an associated eigenvector. Let xR and xI
denote the real and, respectively, imaginary part of x, i.e., set x = xR + jxI and let
x = xRT jxIT . Then
x Px = (xR )T PxR + (xI )T PxI .
Since P is positive definite and xR and xI cannot be both zero (because x = 0), we
deduce that
x Px > 0. (A.6)
x (AT P + PA)x = x Px + x Px = ( + )x Px
+ = 2Re[] < 0.
Remark A.1 The criterion described in the previous theorem provides a sufficient
condition under which all the eigenvalues of a matrix A have negative real part. In the
analysis of linear systems, this criterion is used as a sufficient condition to determine
whether the equilibrium x = 0 of the autonomous system
x = Ax (A.8)
d V dx
V (x(t)) = = 2x T (t)PAx(t) = x T (t)(PA + AT P)x(t).
dt x x=x(t) dt
If P is positive definite and PA+AT P is negative definite, there exist positive numbers
a1 , a2 , a3 such that
d
V (x(t)) V (x(t))
dt
with = a3 /a2 > 0 and therefore
Thus, for any initial condition x(0), limt x(t) = 0. This proves that all eigenvalues
of A have negative real part.
Theorem A.3 (Converse Theorem) Let A Rnn be a matrix of real numbers.
Suppose all eigenvalues of A have negative real part. Then, for any choice of a
symmetric positive-definite matrix Q, there exists a unique symmetric positive-definite
matrix P such that
372 Appendix A: Background Material in Linear Systems Theory
PA + AT P = Q. (A.9)
Proof Consider (A.9). This is a Sylvester equation, andsince the spectra of A and
AT are disjointa unique solution P exists. We compute it explicitly. Define
T
M(t) = eA t QeAt
lim M(T ) = 0
T
and T
P := lim M(t)dt < .
T 0
We have shown in this way that P satisfies (A.9). It is the unique solution of this
equation.
To complete the proof it remains to show that P is positive definite, if so is Q. By
contradiction, suppose is not. Then there exists x0 = 0 such that
x0T Px0 0 ,
Setting
x(t) = eAt x0
this is equivalent to
x T (t)Qx(t)dt 0 ,
0
Bearing in mind the expression of x(t), this implies x0 = 0 and completes the proof.
In this section, a few fundamental facts about the stabilization of linear systems are
reviewed.7 Consider a linear system modeled by equations of the form
x = Ax + Bu
(A.10)
y = Cx
x c = Ac xc + Bc y
(A.11)
u = Cc xc + Dc y,
Definition A.1 The pair (A, B) is stabilizable if there exists a matrix F such that
(A + BF) has all the eigenvalues in C .
Definition A.2 The pair (A, C) is detectable if there exists a matrix G such that
(A GC) has all the eigenvalues in C .
Noting that the closed-loop system (A.12) can be written as x c = Ac xc , with
xc = col(x, xc ) and
(A + BDc C) BCc
Ac = , (A.13)
Bc C Ac
Theorem A.4 There exist matrices Ac , Bc , Cc , Dc such that (A.13) has all the eigen-
values in C if and only if the pair (A, B) is stabilizable and pair (A, C) is detectable.
Proof (Necessity) Suppose all eigenvalues of (A.13) have negative real part. Then,
by the converse Lyapunovs Theorem, there exists a unique, symmetric, and positive-
definite solution Pc of the matrix equation
consistently with the partition of Ac (note, in this respect, that the two diagonal blocks
may have different dimensions n and nc ). Note also that P and Pc are necessarily
positive definite (and hence also nonsingular) because so is Pc . Consider the matrix
I 0
T= .
Pc1 S T I
P0 = P SPc1 S T , F = Dc C Cc Pc1 S T
The matrix P0 is positive definite, because it is the upper-left block of the positive-
definite matrix (A.15). The matrix (A.17) is negative definite, because it is the upper-
left block of the negative-definite matrix (A.16). Thus, by the direct criterion of
Lyapunov, it follows that the eigenvalues of A + BF have negative real part. This
completes the proof that, if Ac has all eigenvalues in C , there exists a matrix F
such that A + BF has all eigenvalues in C , i.e., the pair (A, B) is stabilizable. In a
similar way it is proven that the pair (A, C) is detectable.
(Sufficiency) Assuming that (A, B) is stabilizable and that (A, C) is detectable, pick
F and G so that (A + BF) has all eigenvalues in C and (A GC) has all eigenvalues
in C . Consider the controller
x c = (A + BF GC)xc + Gy
(A.18)
u = Fxc ,
i.e., set
Ac = A + BF GC, Bc = G, Cc = F, Dc = 0.
x = Ax + BFxc
x c = GCx + (A + BF GC)xc .
This system is in block-triangular form and both diagonal blocks have all eigenvalues
in C . Thus the controller (A.18) guarantees that the matrix Ac has all eigenvalues
in C .8
To check whether the two fundamental properties in question hold, the following
tests are useful.
8 Observe that the choices of F and G are independent of each other, i.e., F is only required to place
the eigenvalues of (A + BF) in C and G is only required to place the eigenvalues of (A GC)
in C . For this reason, the controller (A.18) is said to be a controller inspired by a separation
principle.
376 Appendix A: Background Material in Linear Systems Theory
or, what is the same, if and only if the condition (A.19) holds for all (A) (and
not just for all such s having nonnegative real part). The linear system (A.25) is
observable if and only if
C
CA
rank
= n (A.22)
CAn1
or, what is the same, if and only if the condition (A.20) holds for all (A) (and
not just for all such s having nonnegative real part).
It is seen from this that, in general, reachability is a property stronger than stabiliz-
ability and observability is a property stronger than detectability. The two (pairs of)
properties coincide when all eigenvalues of A have nonnegative real part. If the rank
of the matrix on the left-hand side of (A.21) is n1 < n, the system is not reachable
and there exists a nonsingular matrix T such that9
1 A11 A12 B1
TAT = , TB =
0 A22 0
in which A11 Rn1 n1 and the pair (A11 , B1 ) is reachable. This being the case, it
is easy to check that the pair (A, B) is stabilizable if and only if all eigenvalues
of A22 have negative real part. A similar criterion determines the relation between
detectability and observability. If the rank of the matrix on the left-hand side of (A.22)
is n1 < n, the system is not observable and there exists a nonsingular matrix T such
that
A11 A12
TAT 1 = , CT 1 = 0 C2
0 A22
in which A22 Rn1 n1 and the pair (A22 , C2 ) is observable. This being the case, it is
easy to check that the pair (A, C) is detectable if and only if all eigenvalues of A11
have negative real part.
Let d < n denote the dimension of V and let {v1 , v2 , . . . , vd } be a basis of V , that is
a set of d linearly independent vectors vi Rn such that
V = Im(V )
Then, it is an easy matter to check that V is invariant under A if and only if there
exists a d d matrix AV such that
AV = V AV .
The map z AV z characterizes the restriction to V of the map x Ax. This being
the case, observe that if (0 , z0 ) is a pair eigenvalue-eigenvector for AV (i.e., a pair
satisfying AV z0 = 0 z0 ), then (0 , V z0 ) is a pair eigenvalue-eigenvector for A.
Let the matrix A have ns eigenvalues in C , na eigenvalues in C + and nc eigen-
values in C 0 = { C : Re[] = 0}, with (obviously) ns + na + nc = n. Then
(passing for instance through the Jordan form of A) it is easy to check that there exist
three invariant subspaces of A, denoted Vs , Va , Vc , of dimension ns , na , nc that are
complementary in Rn , i.e., satisfy
Vs Va Vc = Rn , (A.23)
x = Ax (A.24)
378 Appendix A: Background Material in Linear Systems Theory
with x Rn . It is easy to check that if subspace V is invariant under A, then for any
x V , the integral curve x(t) of (A.24) passing through x at time t = 0 is such
that x(t) V for all t R.10 Because of (A.23), any trajectory x(t) of (A.24) can
be uniquely decomposed as
lim xs (t) = 0
t
w = Sw
(A.26)
u = Qw
in which w R2 and
0 0
S= , Q = u0 1 0 ,
0 0
10 Note that also the converse of such implication holds. If V is a subspace with the property that,
for any x V , the integral curve x(t) of (A.24) passing through x at time t = 0 satisfies x(t) V
for all t R, then V is invariant under A. The property in question is sometimes referred to as the
integral version of the notion of invariance, while the property indicated in the text above is referred
to as the infinitesimal version on the notion of invariance.
Appendix A: Background Material in Linear Systems Theory 379
In this way, the forced response of the given linear system, from any initial state
x(0), to the input (A.25) can be identified with the free response of the composite
system
w S 0 w
= (A.28)
x BQ A x
from the initial state (x(0), w(0)) with w(0) given by (A.27).
Since A has all eigenvalues with negative real part and S has eigenvalues on the
imaginary axis, the Sylvester equation
S = A + BQ (A.29)
has a unique solution . The composite system (A.28) possesses two complementary
invariant eigenspaces: a stable eigenspace and a center eigenspace, which can be
respectively expressed as
0 I
V s = span , V c = span .
I
The latter, in particular, shows that the center eigenspace is the set of all pairs (w, x)
such that x = w.
Consider now the change of variables x = x w which, after a simple calculation
which uses (A.29), yields
w = Sw
x = Ax .
lim x (t) = 0 ,
t
which shows that the (unique) projection of the trajectory along the stable eigenspace
asymptotically tends to zero. In the original coordinates, this reads as
11 To check that this is the case, simply bear in mind that the solution w(t) of (A.26) is given by
cos(0 t) sin(0 t) w1 (0)
w(t) = eSt w(0) =
sin(0 t) cos(0 t) w2 (0)
.
380 Appendix A: Background Material in Linear Systems Theory
from which we see that the steady-state response of the system to any input generated
by (A.26) can be expressed as
It is worth observing that the steady-state response xss (t) thus defined can also be
identified with an actual forced response of the system to the input (A.25), provided
that the initial state x(0) is appropriately chosen. In fact, since the center eigenspace is
invariant for the composite system (A.28), if the initial condition of the latter is taken
on V c , i.e., if x(0) = w(0), the motion of such system remains confined to V c for
all t, i.e., x(t) = w(t) for all t. Thus, if x(0) = w(0), the actual forced response
x(t) of the system to the input (A.25) coincides with the steady-state response xss (t).
Note that, in view of the definition of w(0) given by (A.27), this initial state x(0) is
nothing else than the first column of the matrix .
The calculation of the solution of the Sylvester equation (A.29) is straightfor-
ward. Set
= 1 2
i.e.,
= Re[(j0 I A)1 B]u0 Im[(j0 I A)1 B]u0 .
As shown above, the steady-state response has the form (A.30). Hence, in particular,
the periodic input
u(t) = u0 cos(0 t)
in which
T (j) = C(jI A)1 B + D.
In this section, a few fundamental facts about algebraic Riccati equations are
reviewed.12 An algebraic Riccati equation is an equation of the form
AT X + XA + Q + XRX = 0 (A.33)
in which all matrices involved are n n matrices and R, Q are symmetric matrices.
Such equation can also be rewritten the equivalent form as
A R I
X I = 0.
Q AT X
From either one of these expressions, it is easy to deduce the following identity
A R I I
= (A + RX) (A.34)
Q AT X X
and to conclude that X is a solution of the Riccati equation (A.33) if and only if the
subspace
I
V = Im (A.35)
X
Lemma A.5 The spectrum of the Hamiltonian matrix (A.36) is symmetric with
respect to the imaginary axis.
Proof Set
0 I
J=
I 0
Suppose now that the matrix (A.36) has no eigenvalues on the imaginary axis.
Then, the matrix in question has exactly n eigenvalues in C and n eigenvalues
in C + . As a consequence, there exist two complementary n-dimensional invariant
subspaces of H: a subspace V s characterized by property that restriction of H to V s
has all eigenvalues in C , the stable eigenspace, and a subspace V a characterized
by property that restriction of H to V a has all eigenvalues in C + , the antistable
eigenspace. A situation of special interest in the subsequent analysis is the one in
which the stable eigenspace (respectively, the antistable eigenspace) of the matrix
(A.36) can be expressed in the form (A.35); in this case in fact, as observed before, it
is possible to associate with this subspace a particular solution of the Riccati equation
(A.33).
If there exists a matrix X such that
I
V s = Im ,
X
and the matrix A+RX has all eigenvalues in C . This matrix is the unique13 solution
of the Riccati equation (A.33) having the property that A + RX has all eigenvalues
in C and for this reason is called the stabilizing solution of the Riccati equation
(A.33).
Similarly, if there exists a matrix X + such that
I
V a
= Im ,
X+
13 If Vis an n-dimensional subspace of R2n , and V can be expressed in the form (A.35), the matrix
X is necessarily unique.
Appendix A: Background Material in Linear Systems Theory 383
and the matrix A + RX + has all eigenvalues in C + . This matrix is the unique solution
of the Riccati equation (A.33) having the property that A + RX has all eigenvalues in
C + and for this reason is called the antistabilizing solution of the Riccati equation
(A.33).
The existence of such matrices X and X + is discussed in the following statement.
Proposition A.1 Suppose the Hamiltonian matrix (A.36) has no eigenvalues on the
imaginary axis and R is a (either positive or negative) semidefinite matrix.
If the pair (A, R) is stabilizable, the stable eigenspace V s of (A.36) can be
expressed in the form
I
V = Im
s
X
The following proposition describes the relation between solutions of the algebraic
Riccati equation (A.33) and of the algebraic Riccati inequality
AT X + XA + Q + XRX > 0 ,
is not empty and any X in this set satisfies X < X (respectively, X > X + ).
References
1. P.S. Antsaklis, A.N. Michel, Linear Systems (McGraw Hill, New York, 1997)
2. P. Lancaster, L. Rodman, Algebraic Riccati Equations. (Oxford University Press, Oxford, 1995)
3. M. Hautus. Linear matrix equations with applications to the regulator problem, in Outils et
modles mathematiques pour lautomatique, lanalyse de syst mes et le traitement du signal ed.
by I. Landau, C.N.R.S. Paris, 3 (1983), pp. 399412
Appendix B
Stability and Asymptotic Behavior
of Nonlinear Systems
We assume in what follows that the reader is familiar with basic concepts concerning
the stability of equilibrium in a nonlinear system. In this section we provide a sketchy
summary of some fundamental results, mainly to the purpose of introducing notations
and results that are currently used throughout the book.14
Comparison functions. A continuous function : [0, a) [0, ) is said to belong
to class K if it is strictly increasing and (0) = 0. If a = and limr (r) = ,
the function is said to belong to class K . A continuous function : [0, a)
[0, ) [0, ) is said to belong to class K L if, for each fixed t, the function
: [0, a) [0, )
r (r, t)
: [0, ) [0, )
t (r, t)
is a class K L function.
The Theorems of Lyapunov. Consider an autonomous nonlinear system
x = f (x) (B.1)
If
V
f (x) 0 for all x Bd , (B.3)
x
the equilibrium x = 0 of (B.1) is stable.
If, for some class K function (), defined on [0, d),
V
f (x) (x) for all x Bd , (B.4)
x
the equilibrium x = 0 of (B.1) is locally asymptotically stable.
If d = and, in the above inequalities, (), () are class K functions, the
equilibrium x = 0 of (B.1) is globally asymptotically stable.
Remark B.1 The usefulness of the comparison functions, in the statement of the
theorem, is motivated by the following simple arguments. Suppose (B.3) holds. Then,
so long as x(t) Bd , V (x(t)) is non-increasing, i.e., V (x(t)) V (x(0)). Pick < d
and define = 1 (). Then, using (B.2), it is seen that, if x(0) < ,
which implies x(t) . This shows that x(t) exists for all t and the equilibrium
x = 0 is stable.
Suppose now that (B.4) holds. Define (r) = ( 1 (r)), which is a class K
function. Using the estimate on the right of (B.2), it is seen that (x) (V (x))
and hence
V
f (x) (V (x)).
x
Since V (x(t)) is a continuous function of t, non-increasing and nonnegative for each
t, there exists a number V 0 such that limt V (x(t)) = V . Suppose V is
strictly positive. Then,
d
V (x(t)) (V (x(t)) (V ) < 0.
dt
Integration with respect to time yields
V (x(t)) V (x(0)) (V )t
for all t. This cannot be the case, because for large t the right-hand side is negative,
while the left-hand side is nonnegative. From this it follows that V = 0 and therefore,
using the fact that V (x) vanishes only at x = 0, it is concluded that limt x(t) = 0.
Note also that identical arguments hold for the analysis of the asymptotic properties
of a time-dependent system
x = f (x, t)
c = {x Rn : V (x) c}.
The function V (x), which is positive definite (i.e., is positive for all nonzero x
and zero at x = 0) is said to be proper if, for each c R, the sublevel set c is a
compact set. Now, it is easy to check that the function V (x) is proper if and only the
inequality on the left-hand side of (B.2) holds for all x Rn , with a function ()
which is of class K . Note also that, if V (x) is proper, for any c > 0 it is possible
to find a numbers c1 > 0 and c2 > 0 such that
Bc1 c Bc2 .
388 Appendix B: Stability and Asymptotic Behavior of Nonlinear Systems
A typical example of how sublevel sets can be used to analyze boundedness and
decay of trajectories is the following one. Let r1 and r2 be two positive numbers, with
r2 > r1 . Suppose V (x) is a function satisfying (B.2), with () a class K function.
Pick any pair of positive numbers c1 , c2 , such that
c1 Br1 Br2 c2 .
Scc12 = {x Rn : c1 V (x) c2 }.
V
f (x) a for all x Scc12 .
x
Then, for each initial condition x(0) Br2 , the trajectory x(t) of (B.1) is defined for
all t and there exists a finite time T such that x(t) Br1 for all t T . In fact, take
any x(0) Br2 \ c1 . Such x(0) is in Scc12 . So long as x(t) Scc12 , the function V (x(t))
satisfies
d
V (x(t) a
dt
and hence
V (x(t)) V (x(0)) at c2 at.
Thus, at a time T (c2 c1 )/a, x(T ) is on the boundary of the set c1 . On the
boundary of c1 the derivative of V (x(t)) with respect to time is negative and hence
the trajectory enters the set c1 and remains there for all t T .
It is well known that the criterion for asymptotic stability provided by the previous
theorem has a converse, namely, the existence of a function V (x) having the properties
indicated in Theorem B.1 is implied by the property of asymptotic stability of the
equilibrium x = 0 of (B.1). In particular, the following result holds.
It is well known that, for a nonlinear system, the property of asymptotic stability
of the equilibrium x = 0 does not necessarily imply exponential decay to zero of
x(t). If the equilibrium x = 0 of system (B.1) is globally asymptotically stable
and, moreover, there exist numbers d > 0, M > 0 and > 0 such that
for all s B .
x = f (x, u) (B.5)
Definition B.1 System (B.5) is said to be input-to-state stable if there exist a class
K L function (, ) and a class K function (), called a gain function, such that,
for any bounded input u() and any x(0) Rn , the response x(t) of (B.5) in the initial
state x(0) satisfies
x(t) (x(0), t) + (u() ) (B.6)
for all t 0.
15 The concept of input-to-state stability, its properties and applications have been introduced in the
sequence of papers [10, 11, 14]. A summary of the most relevant aspect of the theory can also be
found in [5, p. 1731].
390 Appendix B: Stability and Asymptotic Behavior of Nonlinear Systems
for all t 0. Note also that, letting u()[0, t] denote the supremum norm of the
restriction of u() to the interval [0, t], namely
and
V
f (x, u) (x) for all (x, u) Rn Rm satisfying x (u). (B.10)
x
An equivalent form in which the notion of an ISS-Lyapunov function can be
described is the following one.
V
f (x, u) (x) + (u) for all (x, u) Rn Rm . (B.11)
x
The existence of an ISS-Lyapunov function turns out to be a necessary and suffi-
cient condition for input-to-state stability.
Theorem B.3 System (B.5) is input-to-state stable if and only if there exists an ISS-
Lyapunov function. In particular, if such function exists, then an estimate of the form
(B.6) holds with (r) = 1 (( (r))).
x = Ax + Bu
is input-to-state stable, with a linear gain function. In fact, let P denote the unique
positive-definite solution of the Lyapunov equation PA + AT P = I and observe
that V (x) = x T Px satisfies
V
(Ax + Bu) x2 + cx u
x
for some c > 0. Pick 0 < < 1 and set = c/(1 ). Then, it is easy to see that
V
x u (Ax + Bu) x2 .
x
The system is input-to-state, with (r) = r. Since min (P)x2 V (x)
max (P)x2 , we obtain following the estimate for the (linear) gain function
max (P)
(r) = r.
min (P)
x = ax k + bx p u,
in which k N is odd, a > 0 and p N is such that p < k. Pick V (x) = 21 x 2 and
note that, since k + 1 is even,
392 Appendix B: Stability and Asymptotic Behavior of Nonlinear Systems
V
(ax k + bx p u) a|x|k+1 + |b||x|p+1 |u|
x
(r) = .
a
V
x (u) (ax k + bx p u) |x|k+1 .
x
The system is input-to-state stable, with (r) = (r).
Note that the condition k > p is essential. In fact, the following system, in which
k = p = 1,
x = x + xu
is not input-to-state stable. Under the bounded (constant) input u(t) = 2 the state
x(t) evolves as a solution of x = x and hence diverges to infinity.
Theorem B.4 System (B.5) is input-to-state stable if and only if there exist class K
functions 0 () and () such that, for any bounded input and any x(0) Rn , the
response x(t) satisfies
z = f (z, )
(B.12)
= g( ),
the stability properties of the two components subsystems determine the stability of
the cascade. If both systems were linear systems, the cascade would be a system
modeled as
z = Fz + G
= A,
and it is trivially seen that if both F and G have all eigenvalues in C , the cascade is
an asymptotically stable system. The nonlinear counterpart of such property, though,
requires some extra care.
The simplest scenario, in this respect, is one in which one is interested in seeking
only local stability. In this case, the following result holds.17
z = f (z, 0) (B.13)
It must be stressed, though, that in this lemma only the property of local asymptotic
stability of the equilibrium (z, ) = (0, 0) is considered. In fact, by means of a simple
counterexample, it can be shown that the global asymptotic stability of z = 0 as an
equilibrium of (B.13) and the global asymptotic stability of = 0 as an equilibrium
of = g( ) do not imply, in general, global asymptotic stability of the equilibrium
(z, ) = (0, 0) of the cascade. As a matter of fact, the cascade connection of two
such systems may even have finite escape times. To infer global asymptotic stability
of the cascade, a (strong) extra condition is needed, as shown below.
f (z, ) = z + z2
g( ) = .
17 More details and proofs of the results stated in this section can be found in [5, p. 1117 and
3136].
394 Appendix B: Stability and Asymptotic Behavior of Nonlinear Systems
z = z + z 2 (B.14)
z0
z (t) = exp(t).
z0 1 z0 exp(t)
Suppose z0 > 1. Then, the z (t) escapes to infinity in finite time. In particular, the
maximal (positive) time interval on which z (t) is defined is the interval [0, tmax (z0 ))
with
z0
tmax (z0 ) = ln .
z0 1
Now, return to system (B.12), with initial condition (z0 , 0 ) and let 0 be such that
z = z + z2 z + z2 .
z(t) z (t).
Hence z(t) escapes to infinity, at a time t tmax (z0 ). The lesson learned from this
example is that, even if (t) exponentially decreases to 0, this may not suffice to
prevent finite escape time in the upper system. The state z(t) escapes to infinity at a
time in which the effect of (t) on the upper equation is still not negligible.
The following results provide the extra condition needed to ensure global asymp-
totic stability in the cascade.
in which (t) is a continuous function, defined for all t 0 and suppose that
limt (t) = 0. Pick z0 S, and suppose that the integral curve z(t) of (B.15)
satisfying z(0) = z0 is defined for all t 0, bounded, and such that z(t) S for all
t 0. Then limt z(t) = 0.
This last result implies, in conjunction with Lemma B.3, that if the equilibrium
z = 0 of (B.13) is globally asymptotically stable, if the equilibrium = 0 of the
lower subsystem of (B.12) is globally asymptotically stable, and all trajectories of
Appendix B: Stability and Asymptotic Behavior of Nonlinear Systems 395
the composite system (B.12) are bounded, the equilibrium (z, ) = (0, 0) of (B.12)
is globally asymptotically stable.
To be in a position to use this result in practice, one needs to determine conditions
under which the boundedness property holds. This is indeed the case if the upper
subsystem of the cascade, viewed as a system with state z and input , is input-to-state
stable. In view of this, it can be claimed that if the upper subsystem of the cascade is
input-to-state stable and the lower subsystem is globally asymptotically stable (at the
equilibrium = 0), the cascade is globally asymptotically stable (at the equilibrium
(z, ) = (0, 0)).
As a matter fact, a more general result holds, which is stated as follows.
Theorem B.5 Suppose that system
z = f (z, ), (B.16)
viewed as a system with input and state z is input-to-state stable and that system
viewed as a system with input u and state is input-to-state stable as well. Then,
system
z = f (z, )
= g(, u)
is input-to-state stable.
Example B.4 Consider the system
x 1 = x1 + x2 1 2
x 2 = x2 + 12 x1 1 2
1 = 13 + 1 u1
2 = 2 + u2 .
The subsystem consisting of the two top equations, seen as a system with state
x = (x1 , x2 ) and input = (1 , 2 ) is input-to-state stable. In fact, let this system be
written as
x = f (x, )
1 2
V (x) = (x + x22 )
2 1
for which we have
V 1 1 1 1
f (x, ) = (x12 + x22 ) + x2 12 x12 x22 + 14 x2 + 4 .
x 2 2 2 2
396 Appendix B: Stability and Asymptotic Behavior of Nonlinear Systems
Thus, the function V (x) satisfies the condition indicated in Lemma B.2, with
1 2 1 4
(r) = r , (r) = r .
2 2
The subsystem consisting of the two bottom equations is composed of two separate
subsystems, both of which are input-to-state stable, as seen in Examples B.1 and B.2.
Thus, the overall system is input-to-state stable.
x = f (x) (B.18)
Definition B.3 Let x0 Rn be fixed. Suppose that x(t, x0 ) is defined for all t 0. A
point x is said to be an -limit point of the motion x(t, x0 ) if there exists a sequence
of times {tk }, with limk tk = , such that
lim x(tk , x0 ) = x.
k
The -limit set of a point x0 , denoted (x0 ), is the union of all -limit points of the
motion x(t, x0 ).
It is obvious from this definition that an -limit point is not necessarily a limit
of x(t, x0 ) as t , because the solution in question may not admit any limit as
t (see for instance Fig. B.2).18
However, it is known that, if the motion x(t, x0 ) is bounded, then x(t, x0 ) asymp-
totically approaches the set (x0 ), as specified in the lemma that follows.19 In this
respect, recall that a set S Rn is said to be invariant under (B.18) if for all initial
conditions x0 S the solution x(t, x0 ) of (B.18) exists for all t (, +) and
18 Figures B.2, B.3, B.4 are reprinted from Annual Reviews in Control, Vol. 32, A. Isidori and
C.I. Byrnes, Steady-state behaviors in nonlinear systems with an application to robust disturbance
rejection, pp. 116, Copyright (2008), with permission from Elsevier.
19 See [1, p. 198].
Appendix B: Stability and Asymptotic Behavior of Nonlinear Systems 397
x(t, x0 ) S for all such t.20 Moreover, the distance of a point x R from a set
S Rn , denoted dist(x, S), is the nonnegative real number defined as
Lemma B.5 Suppose there is a number M such that x(t, x0 ) M for all t 0.
Then, (x0 ) is a nonempty connected compact set, invariant under (B.18). Moreover,
Example B.5 Consider the classical (stable) Van der Pol oscillator, written in state-
space form as
x 1 = x2
(B.19)
x 2 = x1 + (1 x12 )x2
in which, as it is well known, the damping term (1 x12 )y can be seen as a model
of a nonlinear resistor, negative for small x1 and positive for large x1 (see [6]). From
the phase portrait of this system (depicted in Fig. B.3 for = 1) it is seen that all
motions except the trivial motion occurring for x0 = 0 are bounded in positive time
and approach, as t , the limit cycle L . As consequence, (x0 ) = L for any
x0 = 0, while (0) = {0}.
An important useful application of the notion of -limit set of a point is found in the
proof of the following result, commonly known as LaSalles invariance principle.21
20 We recall, for the sake of completeness, that a set S is said to be positively invariant, or invariant
in positive time (respectively, negatively invariant or invariant in negative time) if for all initial
conditions x0 X, the solution x(t, x0 ) exists for all t 0 and x(t, x0 ) X for all t 0
(respectively exists for all t 0 and x(t, x0 ) X for all t 0). Thus, a set is invariant if it is both
positively invariant and negatively invariant.
21 See e.g., [6].
398 Appendix B: Stability and Asymptotic Behavior of Nonlinear Systems
V
f (x) 0 for all x Rn . (B.20)
x
Proof A direct consequence of (B.20) is that, for any x0 , the motion x(t, x0 ) is
bounded in positive time. In fact, this property yields V (x(t, x0 )) V (x0 ) for all
t 0 and this in turn implies (see Remark B.1)
Thus, the limit set (x0 ) is nonempty, compact and invariant. The nonnegative-valued
function V (x(t, x0 )) is non-increasing for t 0. Thus, there is a number V0 0,
possibly dependent on x0 , such that
lim V (x(t, x0 )) = V0 .
t
By definition of limit set, for each point x (x0 ), there exists a sequence of times
{tk }, with limk tk = , such that limk x(tk , x0 ) = x. Thus, since V (x) is
continuous,
V (x) = lim V (x(tk , x0 )) = V0 .
k
Appendix B: Stability and Asymptotic Behavior of Nonlinear Systems 399
In other words, the function V (x) takes the same value V0 at any point x (x0 ).
Now, pick any initial condition x 0 (x0 ). Since the latter is invariant, we have
x(t, x 0 ) (x0 ) for all t R. Thus, along this particular motion, V (x(t, x 0 )) = V0
and
d V
0 = V (x(t, x 0 )) = f (x) .
dt x x=x(t,x0 )
This, implies
x(t, x 0 ) E , for all t R
x 1 = x2
x 2 = x1 (1 + x1 x2 ) ,
V
f (x) = 2(x1 x2 )2 .
x
The function on the right-hand side is not negative definite, but it is negative semi-
definite, i.e., satisfies (B.20). Thus, trajectories converge to bounded sets that are
invariant and contained in the set
E = {x R2 : x1 x2 = 0}.
Now, it is easy to see that no invariant set may exist, other than the equilibrium,
entirely contained in the set E . In fact, if a trajectory of the system is contained in E
for all t R, this trajectory must be a solution of
x 1 = x2
x 2 = x1 .
This system, a harmonic oscillator, has only one trajectory entirely contained in E ,
the trivial trajectory x(t) = 0. Thus, the equilibrium point x = 0 is the only possible
invariant set contained in E and therefore this equilibrium is globally asymptotically
stable.
400 Appendix B: Stability and Asymptotic Behavior of Nonlinear Systems
The number T in this expression depends on but also on x0 .22 The distance of
x(t, x0 ) from S is said to tend to 0, as t , uniformly in x0 on B, if for every
there exists T , which depends on but not on x0 , such that (B.23) holds for all
x0 B.
Example B.7 Consider again the Example B.5, in which the set defined by (B.22)
consists of the union of the equilibrium point {0} and of the limit cycle L and let B
be a compact set satisfying B L . All x0 B are such that dist(x(t, x0 ), ) 0
as t . However, the convergence is not uniform in x0 . In fact, observe that, if
x0 = 0 is inside L , the motion x(t, x0 ) is bounded in negative time and remains
inside L for all t 0 (as a matter of fact, it converges to 0 as t ). Pick
any x1 = 0 inside L such that dist(x1 , L ) > and let T1 be the minimal time
needed to have dist(x(t, x1 ), L ) for all t T1 . Let T0 > 0 be fixed and define
x0 = x(T0 , x1 ). If T0 is large, x0 is close to 0, and the minimal time T needed to have
dist(x(t, x0 ), ) for all t T is T = T0 + T1 . Since the time T0 can be taken
arbitrarily large, it follows that the time T > 0 needed to have dist(x(t, x0 ), )
for all t T can be made arbitrarily large, even if x0 is taken within a compact set.
Uniform convergence to the target set is important for various reasons. On the
one hand, for practical purposes it is important to have a fixed bound on the time
needed to get within an -distance of that set. On the other hand, uniform convergence
plays a relevant role in the existence Lyapunov functions, an indispensable tool in
analysis and design of feedback systems. While convergence to the set (B.22) is not
guaranteed to be uniform, there is a larger setthoughfor which such property
holds.
fact it is likely that, the more is x0 distant from S, the longer one has to wait until x(t, x0 )
22 In
Definition B.4 Let B be a bounded subset of Rn and suppose x(t, x0 ) is defined for
all t 0 and all x0 B. The -limit set of B, denoted (B), is the set of all points x
for which there exists a sequence of pairs {xk , tk }, with xk B and limk tk = ,
such that
lim x(tk , xk ) = x.
k
It is clear from the definition that, if B consists of only one single point x0 , all
xk s in the definition above are necessarily equal to x0 and the definition in question
returns the definition of -limit set of a point. It is also clear that, if for some x0 B
the set (x0 ) is nonempty, all points of (x0 ) are points of (B). In fact, all such
points have the property indicated in the definition, with all the xk s being taken equal
to x0 . Thus, in particular, if all motions with x0 B are bounded in positive time,
(x0 ) (B).
x0 B
It is well known that, in a nonlinear system, an equilibrium point which attracts all
motions with initial conditions in some open neighborhood of this point is not neces-
sarily stable in the sense of Lyapunov. A classical example showing that convergence
to an equilibrium does not imply stability is provided by the following 2-dimensional
system.25
Example B.9 Consider the nonlinear system
x 1 f (x1 , x2 )
= (B.24)
x 2 g(x1 , x2 )
for (x1 , x2 ) = (0, 0). The phase portrait of this system is the one depicted in Fig. B.4.
This system has only one equilibrium at (x, y) = (0, 0) and any initial condition
(x1 (0), x2 (0)) in the plane produces a motion that asymptotically tends to this point.
However, it is not possible to find, for every > 0, a number > 0 such that every
initial condition in a disc of radius produces a motion which remains in a disc of
radius for all t 0.
It is also knownthoughthat if the convergence to the equilibrium is uniform,
then the equilibrium in question is stable, in the sense of Lyapunov. This property is
a consequence of the fact that x(t, x0 ) depends continuously on x0 (see for example
[2, p. 181]).
We have seen before that bounded motions of (B.18) with initial conditions in a
bounded set B asymptotically approach the compact invariant set (B). Thus, the
question naturally arises to determine whether or not this set is also stable in the
sense of Lyapunov. In this respect, we recall that the notion of asymptotic stability
It is not difficult to show (see [12] or [15]) that if the set A is also bounded and
hence compact, and the convergence in (ii) is uniform in x0 , then property (ii) implies
property (i). This yields the following important property of the set (B).
Lemma B.8 Let B be a nonempty bounded subset of Rn and suppose there is a
number M such that x(t, x0 ) M for all t 0 and all x0 B. Then (B) is a
nonempty compact set, invariant under (B.18). Suppose also that (B) is contained
in the interior of B. Then, (B) is asymptotically stable, with a domain of attraction
that contains B.
We use the concepts introduced in the previous section to define a notion of steady
state for a nonlinear system.
Definition B.5 Consider system (B.18) with initial conditions in a closed subset
X Rn . Suppose that X is positively invariant. The motions of this system are said
to be ultimately bounded if there is a bounded subset B X with the property that,
for every compact subset X0 of X, there is a time T > 0 such that x(t, x0 ) B for all
t T and all x0 X0 .
Motions with initial conditions in a set B having the property indicated in the
previous definition are indeed bounded and hence it makes sense to consider the limit
set (B), whichaccording to Lemma B.6is nonempty and has all the properties
indicated in that lemma. What it is more interesting, though, is thatwhile a set B
having the property indicated in the previous definition is clearly not uniquethe
set (B) is a unique well-defined set.
Lemma B.9 26 Let the motions of (B.18) be ultimately bounded and let B be any
other bounded subset of X with the property that, for every compact subset X0 of X,
there is a time T > 0 such that x(t, x0 ) B for all t T and all x0 X0 . Then,
(B) = (B ).
It is seen from this that, in any system whose motions are ultimately bounded, all
motions asymptotically converge to a well-defined compact invariant set, which is
filled with trajectories that are bounded in positive and negative time. This motivates
the following definition.
Definition B.6 Suppose the motions of system (B.18), with initial conditions in a
closed and positively invariant set X, are ultimately bounded. A steady-state motion
is any motion with initial condition in x(0) (B). The set (B) is the steady-state
locus of (B.18) and the restriction of (B.18) to (B) is the steady-state behavior of
(B.18).
z = f (z, u) (B.25)
so long as the input u can be seen as the output of an autonomous input generator
w = s(w)
(B.26)
u = q(w).
In this way, the concept of steady-state response (to specific classes of inputs) can
be extended to nonlinear systems.
The idea of seeing the steady-state response of a forced system as a particu-
lar response of an augmented autonomous system has been already exploited in
Sect. A.5, in the analysis of the steady-state response of a stable linear system to
harmonic inputs. In the present setting, the results of such analysis can be recast as
follows. Let (B.25) be a stable linear system, written as
z = Az + Bu , (B.27)
in which z Rn , and let (B.26) be the input generator defined in (A.26). The
composition of (B.25) and (B.26) is the autonomous linear system (compare with
(A.28))
w S 0 w
= . (B.28)
z BQ A z
input, the motions of system (B.28), with initial conditions taken in X, are ultimately
bounded. In fact, let be the solution of the Sylvester equation (A.29) and recall
that the difference z(t) w(t) tends to zero as t . Then, any bounded set B
of the form
B = {(w, z) Wc Rn : z w d}
in which d is any positive number, has the property requested in the definition of
ultimate boundedness. It is easy to check that
(B) = {(w, z) Wc Rn : z = w} ,
that is, (B) is the graph of the restriction of the linear map x = w to the set Wc .
The set (B) is invariant for (B.28), and the restriction of (B.28) to the set (B)
characterizes the steady-state response of (B.27) to harmonic inputs of fixed angular
frequency , and amplitude not exceeding c.
A totally similar result holds if the input generator is a nonlinear system of the
form (B.26), whose initial conditions are chosen in a compact invariant set W . The
fact that W is invariant for the dynamics of (B.26) implies, as a consequence of
Lemma B.8, that the steady-state locus of (B.26) is the set W itself, i.e., that the
input generator is in steady state.27 The composition of (B.26) and (B.27) yields
an augmented system of the form
w = s(w)
(B.29)
z = Az + Bq(w) ,
in which (w, z) X := W Rn . Note that, since W is invariant for (B.26), the set
X is invariant for (B.29).
Since the inputs generated by (B.26) are bounded and the lower subsystem of
(B.29) is input-to-state stable, the motions of system (B.29), with initial conditions
taken in X, are ultimately bounded. In fact, since W is compact and invariant, there
exists a number U such that q(w(t)) U for all t R and all w(0) W .
Therefore, standard arguments can be invoked to deduce the existence of positive
numbers K, and M such that
for all t 0. From this, it is immediate to check that any bounded set B of the form
27 Note that the set Wc considered in the previous example had exactly this property.
406 Appendix B: Stability and Asymptotic Behavior of Nonlinear Systems
in which d is any positive number, has the property requested in the definition of
ultimate boundedness. This being the case, it can be shown that the steady-state locus
of (B.29) is the graph of the (nonlinear) map28
: W Rn
0
(B.30)
w (w) = eA Bq(w(,
w))d ,
i.e.,
(B) = {(w, z) W Rn : z = (w)} ,
w)) is by hypoth-
To check that this is the case, observe first of all thatsince q(w(t,
esis a bounded function of t and all eigenvalues of A have negative real partthe
integral on the right-hand side of (B.30) is finite for every w W . Then, observe that
the graph of the map z = (w) is invariant for (B.29). In fact, pick an initial state
for (B.29) on the graph of this map, i.e., a pair (w0 , z0 ) satisfying z0 = (w0 ) and
compute the solution z(t) of the lower equation of (B.29), via the classical variation
of constants formula, to obtain
0 t
z(t) = eAt eA Bq(w(,
w0 ))d + eA(t ) Bq(w(,
w0 ))d
0 00
= eA Bq(w(
+ t, w0 ))d = eA Bq(w(,
w(t, w0 )))d.
This shows that z(t) = (w(t)) and proves the invariance of the graph of ()
for (B.29). Since the graph of () is a compact set invariant for (B.29), this set is
necessarily a subset of the steady-state locus of (B.29). Finally, observe that, since the
eigenvalues of A have negative real part, all motions of (B.29) whose initial conditions
are not on the graph of () are unbounded in negative time and therefore cannot be
contained in the steady-state locus, which by definition is a bounded invariant set.
Thus, the only points in the steady-state locus are precisely the points of the graph
of ().
This result shows that the steady-state response of a stable linear system to an
input generated by a nonlinear system of the form (B.26), with initial conditions
w(0) taken in a compact invariant set W , can be expressed in the form
instance, the system in question could be a stable Van der Pol oscillator, with W
defined as the set of all points inside and on the boundary of the limit cycle. In this
case, it is possible to think of the steady-state response of (B.27) not just as of the
(single) periodic input obtained when the initial condition of (B.26) is taken on the
limit cycle, but also as of all (non periodic) inputs obtained when the initial condition
is taken in the interior of W .
Consider now the case of a general nonlinear system of the form (B.25), in which
z Rn , with input u supplied by a nonlinear input generator of the form (B.26).
Suppose that system (B.25) is input-to-state stable and that the initial conditions of
the input generator are taken in compact invariant set W . It is easy to see that the
motions of the augmented system
w = s(w)
(B.31)
z = f (z, q(w)) ,
with initial conditions in the set X = W Rn , are ultimately bounded. In fact, since
W is a compact set, there exists a number U > 0 such that
u() = q(w()) U
for all w(0) W . Since (B.25) is input-to-state stable, there exist a class K L
function (, ) and a class K function () such that
for all t 0. Since (, ) is a class K L function, for any compact set Z and any
number d > 0, there exists a time T such that (z(0), t) d (U) for all z(0) Z
and all t T . Thus, it follows that the set
Lemma B.10 Consider a system of the form (B.31) with (w, z) W Rn . Suppose
its motions are ultimately bounded. If W is a compact set invariant for w = s(w),
the steady-state locus of (B.31) is the graph of a (possibly set-valued) map defined
on W .
It should be stressed that the map whose graph characterizes the steady-state
locus of (B.31) may fail to be single-valued and, also, may fail to be continuously
differentiable, as shown in the examples below.
z = z3 + zu , (B.32)
w = 0
u=w
z = z3 + zw
w = 0
: wW (w) R
1 w 0 (w) = {0}
0 < w 1 (w) = {z R : |z| w}.
Example B.11 Consider the system
z = z3 + u
which is input to state stable, with input u provided by the harmonic oscillator
w 1 = w2
w 2 = w1
u = w1
for which we take W = {w R2 : w 1}. It can be shown29 that, for each
w(0) W , there is one and only one value z(0) R from which the motion of
the resulting augmented system (B.31) is bounded both in positive and negative
time. The set of all such pairs identifies a single-valued map : W R, whose
graph characterizes the steady-state locus of the system. The map in question is
continuously differentiable at any nonzero w, but it is only continuous at w = 0.
If the map whose graph characterizes the steady-state locus of (B.31) is single-
valued, the steady-state response of an input-to-state stable system of the form (B.25)
to an input generated by a system of the form (B.26) can be expressed as
dz(t)
= f (z(t), q(w(t))) = f ((w(t)), q(w(t)).
dt
29 See [16].
410 Appendix B: Stability and Asymptotic Behavior of Nonlinear Systems
dz(t) dw(t)
= = s(w(t))
dt w w=w(t) dt w w=w(t)
and hence it is seen that (w) satisfies the partial differential equation
s(w) = f ((w), q(w)) for all w W. (B.33)
w
References
A controller, 37, 41
Addition of zeros, 35 perturbation, 64
Adjacency matrix, 137
Affine systems, 167, 208
Agents, 135 E
Annular set, 388 Edge (in a graph), 137
Augmented system, 97, 106, 108, 127 Eigenspaces, 377
Escape time, 195, 222, 393
Exosystem
B linear, 85
Backstepping, 181 nonlinear, 341
B , 186 Exponential stability, 389
Bounded real lemma, 52
F
C Flag, uniform canonical, 205
Cancelation, 181, 194 Follower, 135
Communication graph, 137 Francis equation, 88, 369
Comparison function, 385 Function
Consensus, 137 class K , 385
Contraction, 234 class K , 385
Coordination class K L , 385
leaderfollower, 135
leaderless, 137
G
Gain
D function, 389
d.c. motor, 68 function (estimate of), 391
Detectable, 373 high frequency, 22, 25
Diffeomorphism -suboptimal H design, 71
global, 168 GauthierKupcas form, 202
local, 168 Graph
Differential of a function, 170 connected, 138
Dissipation inequality, 45 directed, 137
Dynamic undirected, 137
H symmetrix, 365
Hamiltonian matrix, 52, 381 Minimum-phase
Harmonic function, 50 globally, 178
Hautus equation, 369 globally and also locally exponentially,
Heterogeneous network, 147 179
High-gain linear system, 28
observer, 201, 211 strongly, 178, 293
output feedback, 32, 188 strongly and also locally exponentially,
H design, 71 179, 294
H norm, 51
Homogeneous network, 147
N
Neighbor (in a graph), 137
I Node (in a graph), 137
Input-to-state stability, 178, 233, 389 Nominal model, 64
Interconnected systems, 57 Nonresonance condition, 88
Interconnection, 58 Normal form
pure feedback, 58, 233 strict (of a SISO linear system), 25
well-defined, 58 global (of a SISO nonlinear system), 173
Internal model, 96, 112 of a SISO linear system, 21, 24
property, 95 strict (of a SISO nonlinear system), 172
tunable, 119
Invariant
set, 396 O
subspace, 377 Observability, uniform, 202
Inverted pendulum, 35 Observable linear system, 27, 202, 376
ISS-Lyapunov function, 390
P
L Parameter uncertainty, 28
L2 gain, 45, 46 Partial-state feedback, 34, 185, 194
L2 norm, 43 Path (in a graph), 137
Lf (x), 169 Peaking, 222
Laplacian matrix, 138 Positive definite and proper, 178, 387
LaSalles principle, 397 Postprocessor, 105, 107, 111
Leader, 135 Practical stabilization, 188
Lie bracket, 174 Preprocessor, 111
LMI, 47 Prime form, 24
coupling condition, 79 Principal minor, 367
in H design, 71
in L2 gain estimation, 48
Lyapunovs Q
equation, 369 Quadratic form, 366
theorems (for linear systems), 370
theorems (for nonlinear systems), 386
R
Reachable linear system, 27, 376
M Realization (minimal), 27
Matrix Regulator equation
negative definite, 367 linear, 88, 369
negative semidefinite, 367 nonlinear, 344
orthogonal, 365 Relative degree
positive definite, 366 of a linear system, 21
positive semidefinite, 366 of a nonlinear system, 168
Index 413