SIAM J. CONTROL OPTIM.

c 2000 Society for Industrial and Applied Mathematics


Vol. 38, No. 3, pp. 971–994

ANY DOMAIN OF ATTRACTION FOR A LINEAR CONSTRAINED

SYSTEM IS A TRACKING DOMAIN OF ATTRACTION∗
FRANCO BLANCHINI† AND STEFANO MIANI‡
Abstract. In the stabilization problem for systems with control and state constraints a domain
of attraction is a set of initial states that can be driven to the origin by a feedback control without
incurring constraints violations. If the problem is that of tracking a reference signal, that converges to
a constant constraint-admissible value, a tracking domain of attraction is a set of initial states from
which the reference signal can be asymptotically approached without constraints violation during
the transient. Clearly, since the zero signal is an admissible reference signal, any tracking domain of
attraction is a domain of attraction. We show that the opposite is also true. For constant reference
signals we establish a connection between the convergence speed of the stabilization problem and
tracking convergence which turns out to be independent of the reference signal. We also show that
the tracking controller can be inferred from the stabilizing (possibly nonlinear) controller associated
with the domain of attraction.

Key words. constrained control, domain of attraction, Lyapunov functions, tracking, convex
analysis

AMS subject classifications. 93B52, 93D30, 93B51, 93B50

PII. S036301299834661X

1. Introduction. In this paper we consider linear systems with state and control
constraints. For this class of systems we say that a certain convex and compact set
including the origin in its interior is a domain of attraction to the origin if there
exists a feedback control such that for any initial state in this set the state is driven
asymptotically to the origin without constraints violation. Several previous references
have dealt with the construction of such domains (see for instance [3], [6], [11], [12],
[17], [18], [19], [20], [21], [26], [28]).
Here we consider the problem of tracking a reference signal including the zero
one. We assume that the system is square (i.e., it has as many inputs as outputs)
and that it is free of invariant zeros at one in the discrete-time case (or it is free
from zeros at the origin in the continuous-time case). We define a set of reference
vectors which are constraints-compatible in the sense that they are associated with
state vectors which are in the interior of the domain of attraction to the origin and
to feasible input vectors. The only condition required for the signal to be tracked is
that it asymptotically converges to a value which is in this set. A signal satisfying the
above requirement will be said to be admissible. We stress that the signal, during the
transient, may assume values outside this set and it is a goal of the control to avoid
constraints violations. For the zero reference signal, such a goal can be achieved only
if the initial state is inside a domain of attraction to the origin. Any initial state from
which we can track asymptotically an admissible reference signal is said to belong to
a tracking domain of attraction. The question which hence arises is whether there
are initial states inside the domain of attraction to the origin from which we cannot

∗ Received by the editors October 29, 1998; accepted for publication (in revised form) August 3,

1999; published electronically March 15, 2000.

http://www.siam.org/journals/sicon/38-3/34661.html
† Dipartimento di Matematica ed Informatica, Università di Udine, via delle Scienze 208, 33100

Udine, Italy ([email protected]).

‡ Dipartimento di Ingegneria Elettrica, Gestionale e Meccanica, Università di Udine, via delle

Scienze 208, 33100 Udine, Italy ([email protected]).

971
972 FRANCO BLANCHINI AND STEFANO MIANI

solve the problem of tracking an admissible reference signal. Put in other words this
means to verify whether or not the largest tracking domain of attraction is a proper
subset of the assigned domain of attraction. We show that the answer is negative.
The basic ideas we develop here are related to previously published results [2], [15],
[14], [16]. The main difference relies on the fact that in those references it is assumed
that a stabilizing linear nominal compensator (often referred to as precompensator) is
applied to the system. To this nominal compensator the so called reference governor
is added which is a device that possibly attenuates the effect of the reference signal in
order to avoid violations. This is basically equivalent to considering a stable system
(i.e., a system to which a precompensator is applied) and to managing the reference
for this system. Although this assumption is quite reasonable, it turns out that the
resulting constructed invariant sets depend on the precompensator. Therefore, an
unsuitable choice of the compensator can produce a very small domain of attraction.
Conversely, here we do not assume the existence of any precompensator and thus
we may take as a domain of attraction the largest one in absolute in the sense that, if
the initial state is outside that region, there is no guarantee that constraints violation
can be avoided, even for zero reference signal. We show that this is a tracking domain
of attraction, in the sense that there is a (possibly nonlinear) compensator, which is
not a priori fixed, that solves the tracking problem for any initial state in this region.
Furthermore, we show that the control strategy can be inferred from the original
stabilizing controller associated with the given domain of attraction. We use as a
technical support a Lyapunov function suitably constructed by “reshaping” the one
associated with such a domain.
Finally we show that, for constant reference signals and symmetric domains, the
speed of convergence can be estimated a priori from the speed of convergence achieved
by the stabilizing controller, and it does not depend on the particular reference.
The paper is organized as follows. In section 2 we provide the basic definitions.
In section 3 we state the main results which will be proven later in section 4 and
section 5. We particularize the results to domains of attraction of special shapes in
section 6. We show how to extend the results to nonsquare systems in section 7. We
finally present two examples in section 8 to illustrate the method, and we derive our
conclusions in section 9.
2. Definitions and problem statement. In what follows we denote by intP
the interior of a set P and by ∂P its boundary. With the term C-set we denote a
convex and compact set containing the origin as an interior point. It is known that
any C-set P induces a positively homogeneous convex function which is known as
Minkowski functional (see [22]):
(2.1) ψP (x) = inf{ξ ≥ 0 : x ∈ ξP}.
The function ψP (x) ≥ 0 is such that ψP (x) = 0 iff x = 0, and ψP (λx) = λψP (x) for
any λ ≥ 0. If P is 0-symmetric (i.e., x ∈ P implies −x ∈ P), then ψP is a norm. The
index P will be omitted for brevity when the inducing set P of ψP is clear from the
context.
In the following we consider both discrete and continuous-time square systems
δx(t) = Ax(t) + Bu(t)
(2.2)
y(t) = Cx(t) + Du(t),
where δ represents the shift operator δx(t) = x(t + 1) in the discrete-time case and the
derivative operator in the continuous-time case. The vector y(t) ∈ Rq is the system
 TRACKING DOMAIN OF ATTRACTION 973

output, the vector x(t) ∈ Rn is the system state, and u(t) ∈ Rq is the control input.
We assume that (A, B) is stabilizable and that x and u are subject to the constraints

(2.3) x(t) ∈ X ,
(2.4) u(t) ∈ U,

where X ⊂ Rn and U ⊂ Rq are assigned C-sets.

Due to the presence of the constraints, the solution of the stabilization problem
implies restrictions on the admissible initial conditions.
Definition 2.1. The C-set S ⊂ X is a domain of attraction for the system (2.2)
if there exists a continuous feedback control function u(t) = φ(x(t)) such that any
trajectory x(t) with initial condition x(0) ∈ S is such that x(t) ∈ S, u(t) ∈ U, and

lim x(t) = 0.
t→∞

It is well known that the above definition is fundamental for the stabilization
problem under constraints. An initial condition can be driven to the origin without
constraints violations iff it belongs to a domain of attraction. However, in practice,
simple convergence is not enough, but it is important to provide an index of the speed
of convergence to the origin. Thus we introduce the following two definitions.
Definition 2.2. The C-set S ⊂ X is a domain of attraction with speed of
convergence λ for the discrete-time system (2.2) if there exists 0 ≤ λ < 1 and a
continuous feedback control function u(t) = φ(x(t)) such that any trajectory x(t) with
initial condition x(0) ∈ S is such that

ψS (x(t)) ≤ λt ψS (x(0)) .

If we take λ = 1, the set S is simply said to be positively invariant (often referred to

as controlled–invariant). We say that Sλ ⊂ X is the largest domain of attraction if
for any domain of attraction S in X with speed of convergence λ we have S ⊂ Sλ .
Definition 2.3. The C-set S ⊂ X is a domain of attraction with speed of con-
vergence β for the continuous-time system (2.2) if there exists β > 0 and a continuous
feedback control function u(t) = φ(x(t)) such that any trajectory x(t) with initial con-
dition x(0) ∈ S is such that

ψS (x(t)) ≤ e−βt ψS (x(0))

If we take β = 0, the set S is simply said to be positively invariant (often referred to

as controlled-invariant). We say that Sβ ⊂ X is the largest domain of attraction if
for any domain of attraction S in X with speed of convergence β we have S ⊂ Sβ .
A stabilizable system always admits a domain of attraction P ⊂ X . The knowl-
edge of such a domain is fundamental in the stabilization problem under constraints,
since if x(0) ∈ P, then the conditions x(t) ∈ P ⊂ X for t ≥ 0, and x(t) → 0 as t → ∞
can be assured. Thus, once we have computed a domain of attraction P to solve the
problem (possibly the largest one [4], [6], [7]), we can replace the constraint x(t) ∈ X
by the new constraint

x(t) ∈ P.

In this paper we deal with the problem of tracking a certain class of reference signals.
To this aim we make the assumption that the system is free from zeros at one (zeros
at the origin).
974 FRANCO BLANCHINI AND STEFANO MIANI

Assumption 2.1. The square matrix



A−I B A B
Md = resp., Mc =
C D C D

is invertible.
The general case in which Md (Mc ) is not invertible or even nonsquare will be
considered in section 7. Under Assumption 2.1 the system has the property that for
any constant reference r ∈ Rq there is an unique state-input equilibrium pair (x̄, ū)
such that the corresponding equilibrium output is r. Such a pair is the unique solution
of the equation





x̄(r) 0 x̄(r) 0
Md = resp., Mc = .
ū(r) r ū(r) r

Thus we can define the set of admissible constant reference vectors

R = {r ∈ Rq : ū(r) ∈ U, x̄(r) ∈ P}.

Being that P and U are both bounded, R is also bounded. The set R is of funda-
mental importance. It is the set of all reference vectors for which the corresponding
input and state equilibrium pairs do not violate the constraints ū ∈ U and x̄ ∈ P.
While the reason for imposing the former is obvious, the second deserves some expla-
nation. Indeed we have imposed x ∈ X . However, note that if x̄(r) is not included in
a domain of attraction, then the condition x(t) → x̄(r) can cause a violation of the
constraints if, for instance, after a sufficiently long period, the reference r(t) switches
to zero.
We are now going to introduce the set of all the admissible signals to be tracked,
formed by the signals r(t) having a finite limit r∞ , with the condition that r∞ has
some admissibility condition with respect to the constraints.
Definition 2.4. Assume that a small 0 <  < 1 is given. A reference signal r(t)
is admissible if it is continuous1 and if it is such that
.
lim r(t) = r∞ ∈ (1 − )R = R .
t→∞

The parameter , as we will see later, is introduced to avoid singularities in the

control. Such  may be arbitrarily small, and thus it does not practically affect the
problem. We stress that an admissible reference signal does not need to assume its
values in R , but only its limit r∞ needs to do this. Now we can state the following
basic definition.
Definition 2.5. The set P ⊂ X is a tracking domain of attraction if there exists
a (possibly nonlinear) feedback control

u(t) = Φ(x(t), r(t))

such that for any x(0) ∈ P and for every admissible reference signal r(t)
(i) x(t) ∈ P and u(t) ∈ U,
(ii) y(t) → r∞ as t → ∞.

1 The continuity requirement for the function r(t) is limited to the continuous-time case only. It

is not essential but avoids unnecessary complications.

 TRACKING DOMAIN OF ATTRACTION 975
1

1
1 0 1

Fig. 1. The largest domain of attraction when λ = .9.

Since r(t) = 0 is an admissible reference signal, any tracking domain of attraction

is a domain of attraction. The main result of this paper is that of showing that every
domain of attraction P is also a tracking domain of attraction. The importance of
this assertion lies in the fact that the tracking problem can be solved once one has
found a domain of attraction. Since the latter operation is a well-established topic,
this allows for the solution of a more general problem.
Remark 2.1. Note that under Assumption 2.1 the condition y(t) → r∞ as t → ∞
. .
is equivalent to the two conditions x(t) → x̄∞ = x̄(r∞ ) and u(t) → ū∞ = ū(r∞ ).
2.1. A simple example. Consider the discrete-time system



1 −1 1  
A= B= C= 0 1 ,
1 1 0

together with the constraints sets X = {x : x∞ ≤ 1} and U = {u : |u| ≤ 1.2}. The
largest domain of attraction with speed of convergence λ for λ ≥ 0.9 is given by

P = {x : |x1 | ≤ 1, |x2 | ≤ 1, |x1 + x2 | ≤ λ}

and is reported in Figure 1 for λ = .9. Such a domain is associated with the polyhedral
function ψ(x) = max{|x1 |, |x2 |, |x1 + x2 |/λ}.
If we take λ < 1 we assure that P is a domain of attraction. Note that for λ
approaching 1 the set P approaches the largest invariant set. Therefore there is a
tradeoff between the choice of the largest invariant set and the convergence speed.
Once this tradeoff is fixed by a choice of λ, we have that there exists a stabilizing
control which avoids constraints violation for all initial states inside P. Now if we
want to track a reference signal r(t) the situation is different. For a given fixed r̄, the
corresponding state input pair is


−1

x̄ A−I B 0
= .
ū C 0 r̄
If our goal were asymptotic tracking only, in view of the linearity of the system we
could just use the translated Lyapunov function

ψ(x − x̄) = max{|x1 − x̄1 |, |x2 − x̄2 |, |x1 − x̄1 + x2 − x̄2 |/λ}.

But in this way the constraints could not be satisfied. The basic idea we pursue is to
“deform” the function ψ in such a way that the surface of level one is unchanged and
the minimum point (the zero) is assumed on x̄.
976 FRANCO BLANCHINI AND STEFANO MIANI

The set R is the interval

R = {r : |r| ≤ λ − }.

For every reference signal r(t) → r̄ ∈ R and x(0) ∈ P, our goal is that of letting
y(t) → r̄ and x(t) ∈ P. According to Remark 2.1, for the first condition it is necessary
and sufficient that x(t) → x̄ = [0 r]T and u(t) → r (in this case ū(r) = r).
3. Main results. We will now introduce the functions which will subsequently
be used for tracking every admissible reference signal. Suppose that a C-set P, which
is a domain of attraction, is given and consider its Minkowski functional which, from
(2.1), can be rewritten as

1
(3.1) ψP (x) = inf{α > 0 : x ∈ P}.
α
For every x̄ ∈ intP and x ∈ P, we introduce the following function:

1
(3.2) ΨP (x, x̄) = inf{α > 0 : x̄ + (x − x̄) ∈ P}.
α
It is immediate that the just introduced function Ψ recovers the values of ψ when
x̄ = 0, say ΨP (x, 0) = ψP (x). For fixed x̄, ΨP (x, x̄) is convex. Furthermore, the
function ΨP (x, x̄) for (x, x̄) ∈ P × intP is such that

(3.3) ΨP (x, x̄) = 0 iff x = x̄,

(3.4) ΨP (x, x̄) < 1 iff x ∈ P,
(3.5) ΨP (x, x̄) = 1 iff x ∈ ∂P.

A sketch of the function ΨP (x, x̄) for fixed x̄ is in Fig. 2. One further relevant property
of this function is that ΨP is Lipschitz in x and positively homogeneous of order one
with respect to the variable z = x − x̄ ∈ Rn , x̄ ∈ intP, i.e.,

(3.6) ΨP (ξz + x̄, x̄) = ξΨP (z + x̄, x̄).

In view of property (3.3) we have that the function is a suitable Lyapunov candidate
for tracking, and from (3.4)–(3.5) we have that this function is suitable to prevent
constraints violations, as we will show later.
Let us now consider the function ΨP (x, x̄) and for every x ∈ P and x̄ ∈ intP,
with x = x̄, set

. 1
x̃ = x̄ + (x − x̄) ∈ ∂P.
ΨP (x, x̄)

The vector x̃ is the intersection of ∂P with the half line starting from x̄ and passing
through x (see Fig. 2). Assume that P is a given domain of attraction and that
φ(x) is the control law associated with this set. As shown in [4], [7], there exists
always a positively homogeneous control of order 1 which can be associated with such
a domain. Therefore, without restriction, we introduce the following assumption.
Assumption 3.1. The stabilizing control law φ(x) associated with the domain of
attraction P is Lipschitz and positively homogeneous of order 1, i.e., φ(αx) = αφ(x)
for α ≥ 0 and x ∈ P.
 TRACKING DOMAIN OF ATTRACTION 977

P
x

~
x

Fig. 2. The function ΨP (x, x̄) for fixed x̄.

The next step in the derivation of a feedback control law Φ(x, r) is the definition
of a saturation map Γ : Rq → R as follows:

rψR
(r)−1 when ψR
(r) > 1
Γ(r) =
r otherwise
say
 
1
Γ(r) = r ∗ sat ,
ψR
(r)
where sat(·) is the saturation function. Note that Γ(r) is the identity if we restrict
r ∈ R . Conversely, for r ∈ R , Γ(r) is the intersection of ∂R and the segment
having extrema 0 and r.
The control we propose has the form
(3.7) Φ(x, r) = φ(x̃)ΨP (x, x̄) + (1 − ΨP (x, x̄))ū,
where





x̄(r̄) . 0 x̄(r̄) . 0
(3.8) = Md−1 resp., −1
= Mc
ū(r̄) r̄ ū(r̄) r̄
and
(3.9) r̄ = Γ(r).
Note that, for r ∈ R , (3.9) does not play any role. Note also that, since r̄ = Γ(r) ∈
R , then x̄ ∈ intP, thus the term ΨP (x, x̄) in (3.7) is defined. However, the expression
(3.7) is not defined for x = x̄ because of the critical term φ(x̃)ΨP (x, x̄). Nevertheless,
in view of the homogeneity of φ and from the expression of x̃, we have that
(3.10) φ(x̃)ΨP (x, x̄) = φ(ΨP (x, x̄)x̃) = φ (x + (ΨP (x, x̄) − 1)x̄) .
Then φ(x̃)ΨP (x, x̄) → 0 as x → x̄ so that we can extend the function by continuity
by assuming
φ(x̃)ΨP (x̄, x̄) = 0.
978 FRANCO BLANCHINI AND STEFANO MIANI

The introduced control law inherits most of the properties from φ(x) according to
the next proposition which assures existence and uniqueness of the solution of (2.2)
when the control Φ(x, r) is used, provided that the admissible reference signal r(t) is
measurable.
Proposition 3.1. Suppose φ(x) is Lipschitz and homogeneous of order 1. Then
Φ(x, r) : P × Rq → U, defined as in (3.7)–(3.9), is continuous and it is Lipschitz with
respect to x.
To have an idea of how this control works, note that as long as the condition

(3.11) ΨP (x(t), x̄(t)) ≤ 1

is satisfied, we have that x(t) ∈ P. Moreover, for x̄ ∈ intP, the control is just a convex
combination of the control ū(r̄) and φ(x̃). By construction, ū(r̄) ∈ U and φ(x̃) ∈ U;
thus Φ(x, r) ∈ U.
Our effort will be devoted to proving that the proposed control law guarantees
(3.11) as well as the limit condition

(3.12) ΨP (x(t), x̄(r∞ )) → 0,

where x̄(r∞ ) is the steady state associated with r∞ ∈ R (note that Γ(r∞ ) = r∞ ).
Indeed such a limit condition implies x(t) → x̄(r∞ ) and, from (3.7) and (3.10),
Φ(x(t), r(t)) → Φ(x̄(r∞ ), r∞ ) = ū(r∞ ). Therefore if (3.12) holds, we have that
y(t) → r∞ .
For evident practical reasons, we introduce an extended concept of speed of con-
vergence, appropriate for the condition (3.12). In the continuous-time case, given
a fixed x̄ ∈ intP we say that the speed of convergence is β > 0 if the Lyapunov
derivative Ψ̇P (x(t), x̄) of ΨP (x(t), x̄) is such that

. ΨP (x + h(Ax + Bu), x̄) − ΨP (x, x̄)

Ψ̇P (x, x̄, u) = lim+ ≤ −βΨP (x, x̄),
h→0 h
where the existence of the limit is assured by the convexity of ΨP with respect to x.
In the discrete-time case we say that the speed of convergence is λ < 1 if

ΨP (Ax + Bu, x̄) ≤ λΨP (x, x̄).

We start by considering a special case, namely the one in which the reference
signal is constant. In this case we can show that, if we have a domain of attraction
to the origin with a certain speed of convergence, we can achieve the tracking goal
without constraints violation for all the initial states in such a domain. Furthermore,
for symmetric domains, we can guarantee a speed of convergence which is independent
of the reference input.
Theorem 3.2. Let P be a domain of attraction with speed of convergence λ
for the discrete-time dynamic system (2.2) associated with the control φ(x) satisfying
Assumption 3.1. Then, for every admissible constant reference signal r(t) = r̄, the
control law (3.7)–(3.8) is such that for every initial condition x(0) ∈ P we have that
x(t) ∈ P and u(t) ∈ U for every t ≥ 0 and limt→∞ y(t) = r̄. Moreover, if P is
0-symmetric, the speed of convergence λT R = λ+1 2 is guaranteed.
The next theorem is the continuous-time version of the above.
Theorem 3.3. Let P be a domain of attraction with speed of convergence β for
the continuous-time dynamic system (2.2) associated with the control φ(x) satisfying
Assumption 3.1. Then, for every admissible constant reference signal r(t) = r̄, the
 TRACKING DOMAIN OF ATTRACTION 979

control law (3.7)–(3.8) is such that for every initial condition x(0) ∈ P, we have that
x(t) ∈ P and u(t) ∈ U for every t ≥ 0 and limt→∞ y(t) = r̄. Moreover, if P is
0-symmetric, the speed of convergence is at least βT R = β2 .
The proposed control law can be successfully used even when the reference r(t)
is allowed to vary provided that it is admissible according to Definition 2.4.
Theorem 3.4. Let r(t) be admissible as in Definition 2.4. Any domain of attrac-
tion P, with speed of convergence β > 0 (0 ≤ λ < 1), for system (2.2) is a tracking
domain of attraction. Moreover, the control law in (3.7)–(3.9) assures the conditions
(i) and (ii) of Definition 2.5.
Remark 3.1. If we consider a constant reference r ∈ R and the corresponding
steady state vectors derived from x̄ and ū by means of (3.8), then we can apply a
state and control translation by considering the system δ x̂ = Ax̂ + B û with the new
constraints û = u − ū ∈ U − ū = Û and x̂ = x − x̄ ∈ X − x̄ = X̂ . From this algebraic
point of view, our result amounts to proving that the largest domain of attraction of
the translated problem is just achieved by translating the original largest domain of
attraction as P̂ = P − x̄.
4. Properties of the function Ψ(x, x̄). To prove the results of the previous
section we first need to present some properties of the function Ψ(x, x̄). First we prove
its continuity. For the simple notation we will drop the index P.
Lemma 4.1. For (x, x̄) ∈ P × intP, the function Ψ(x, x̄) is continuous.
Proof. Consider the case x = x̄. Take any point (x, x̄) and a close pair (x , x̄ ).
Assume that x ∈ Vx andx̄ ∈ Vx̄ ⊂ intP where Vx and Vx̄ are close neighborhoods
of x and x̄ such that Vx Vx̄ = ∅, the empty set. This implies that x − x̄  ≥ M
for some positive M . Assume now x ∈ P and take α > 0 (which exists because
x̄ ∈ intP) such that x̄ + (x − x̄ )/α ∈ P . Then (x − x̄ )/α ≤ m, for some positive
m. This means that α ≥ M/m; thus it is lower bounded. From (3.2) we have

1
Ψ(x , x̄ ) = inf α > 0 : x̄ + (x − x̄ ) ∈ P
(4.1)
α
 

   

 1 1 1 

= inf α > 0 : x̄ + (x − x̄) ∈ P + (x̄ − x̄ ) 1 − + (x − x ) ,

 α α α 
   

=z
.
where P + z = {y = x + z, x ∈ P}. Since α as in (4.1) is lower bounded by M/m,
then z → 0 as (x , x̄ ) → (x, x̄).
Moreover, P is a C-set; thus for every 0 < ˆ < 1 there exists δ such that if
x − x  ≤ δ and x̄ − x̄  ≤ δ, then
(4.2) (1 − ˆ)P ⊂ P + z ⊂ (1 + ˆ)P.
Consider the following function:

1
α() = inf α > 0 : x̄ + (x − x̄) ∈ (1 + )P .
α
The inclusion (4.2) assures that for ˆ and δ as above
) ≤ Ψ(x , x̄ ) ≤ α(−ˆ
α(ˆ ),
and from (3.2) we have
α(ˆ
) ≤ Ψ(x, x̄) ≤ α(−ˆ
).
980 FRANCO BLANCHINI AND STEFANO MIANI

Thus, if we show that α() is continuous at 0, we have |α(ˆ ) − α(−ˆ)| → 0, as ˆ → 0,

and we have completed the proof.
Since x̄ ∈ intP, there exists 0 < ¯ < 1 such that x̄ ∈ int{(1 + )P} for all || < ¯.
The set of all points of the form x̄ + (x − x̄)/α, α > 0 is a ray originating
from x̄ which, for || ≤ ¯, intersects the boundary of (1 + )P at the unique point
x̄ + (x − x̄)/α(). We can write this condition as
(4.3) ψ (x̄ + (x − x̄)/α()) = 1 + .
Note also that for α > α() the point x̄ + (x − x̄)/α ∈ int{(1 + )P}; thus the function
α() is strictly decreasing on [−¯
, ¯], and so it is invertible. From (4.3) its inverse is
the function
.
(α) = ψ (x̄ + (x − x̄)/α) − 1
which is continuous (actually convex). Then α() is continuous.
In the case x = x̄ ∈ intP, say Ψ(x, x̄) = 0 it is very easy to show from (3.2) that
Ψ(x , x̄ ) → 0 as (x , x̄ ) → (x̄, x̄), and we skip the proof for brevity.
The function Ψ(x, x̄) is not necessarily differentiable and then, to use it as a
candidate Lyapunov function, we invoke the theory of subdifferentials. Given a convex
function F (x), the subdifferential of F in x is defined as the set [23]
∇F (x) = {w : w, z − x ≤ F (z) − F (x) for all z ∈ Rn } .
Each element w of ∇F is called subgradient. In what follows we denote by ∇Ψ(x, x̄)
the subdifferential of the function Ψ(x, x̄) with respect to x (i.e., x̄ is assumed con-
stant):
∇Ψ(x, x̄) = {w : w, z − x ≤ Ψ(z, x̄) − Ψ(x, x̄)} .
If Ψ(x, x̄) is differentiable with respect to x, then ∇Ψ(x, x̄) is a singleton including
the gradient vector

T
. ∂Ψ(x, x̄) ∂Ψ(x, x̄)
∇Ψ(x, x̄) = ,..., .
∂x1 ∂xn
For constant x̄, the Lyapunov derivative of the function Ψ(x, x̄), along the system
trajectory of the system (2.2), is given by
Ψ(x + τ (Ax + Bu), x̄) − Ψ(x, x̄)
(4.4) Ψ̇(x, x̄, u) = lim .
τ →0+ τ
From [23, Th. 23.4], we have that
(4.5) Ψ̇(x, x̄, u) = sup w, Ax + Bu.
w∈∇Ψ(x,x̄)

This fact implies that the condition

Ψ̇(x, x̄, u) ≤ −βΨ(x, x̄)
is equivalent to
w, Ax + Bu ≤ − βΨ(x, x̄) for all w ∈ ∇Ψ(x, x̄).
The next lemma will show some fundamental properties concerning the subdifferen-
tiability of Ψ. Its importance will be evident later when we deal with continuous-time
systems.
Lemma 4.2. Assume that x̄ ∈ intP is fixed. Then
 TRACKING DOMAIN OF ATTRACTION 981

(i) Its subdifferential is constant over any ray emanating from x̄

(4.6) ∇Ψ(x̄ + µy, x̄) = ∇Ψ(x̄ + y, x̄) for all µ ≥ 0;

(ii) The equality

Ψ(x, x̄) = ŵ, x − x̄

holds for all ŵ ∈ ∇Ψ(x, x̄).

(iii) For every x ∈ ∂P the positive cones generated by the subdifferentials ∇ψ(x)
and ∇Ψ(x, x̄) coincide. This means that each w ∈ ∇ψ(x) is aligned with
some ŵ ∈ ∇Ψ(x, x̄) and vice-versa. Furthermore, if P is 0-symmetric (i.e.,
P = −P) for every x̄ ∈ (1 − )P and x ∈ ∂P, the aligning factor γ(ŵ, w)
such that

ŵ = γ(ŵ, w)w, ŵ ∈ ∇Ψ(x, x̄), w ∈ ∇ψ(x)

satisfies the inequality

1 1
(4.7) γ(ŵ, w) ≥ ≥ .
2− 2
Proof. Property (i) is a straightforward derivation from the definition of the
subdifferential once we have established that ψ̃(y) = Ψ(x̄ + y, x̄) is a positively homo-
geneous function of order 1 according to (3.6). Property (ii) follows from the following
properties of the positively homogeneous functions of order 1 (see the appendix)

(4.8) ψ̃(y) = w, y for all w ∈ ∇ψ̃(y).

To prove (iii), let x̄ ∈ intP be fixed. From (3.4)–(3.5) the unit balls of the functions
Ψ(x, x̄) and ψ(x) are both equal to P. For x ∈ ∂P, the normal cone [13] to this set
is given by

N = {z ∈ Rn : z, x − x ≤ 0 for all x ∈ P}.

Since the normal cone for x ∈ ∂P is the convex positive cone generated by any of the
two subdifferentials ∇Ψ(x, x̄) and ∇ψ(x), we have the alignment of any element of
the former with an element of the latter and the existence of a positive real γ(ŵ, w)
such that ŵ = γw, with ŵ ∈ ∇Ψ(x, x̄) and w ∈ ∇ψ(x).
To prove (4.7) consider x ∈ ∂P so that Ψ(x, x̄) = 1 with x̄ ∈ (1 − )P. Let
ŵ ∈ ∇Ψ(x, x̄) and w ∈ ∇ψ(x) be as above. From property (ii)

Ψ(x, x̄) = 1 = ŵ, x − x̄ = γw, x − x̄.

Then, since w ∈ ∇ψ(x), ψ(x) = w, x = 1 and by the symmetry of P, ψ(−x̄) =

ψ(x̄) ≤ 1 −  we get

1
(4.9) = w, x − x̄ = w, −x̄ − x + 2w, x
γ
(4.10) ≤ ψ(−x̄) − ψ(x) + 2 = ψ(−x̄) + 1 ≤ 2 − 

which is equivalent to (4.7).

982 FRANCO BLANCHINI AND STEFANO MIANI

Remark 4.1. If symmetry is removed, it is possible to show that there exists

m > 0 such that γ(ŵ, w) ≥ m > 0, although no lower bounds can be provided a
priori.
We recall that if P is a convex set in a real normed space X, then it is possible
to define on the dual space X ∗ the support functional h(x∗ ) of P
(4.11) h(x∗ ) = sup x, x∗ .
x∈P

If the set contains the origin as an interior point, then h(x∗ ) is positive definite, and
if the set is compact, there exists a finite constant K such that h(x∗ ) ≤ Kx∗  for
every x∗ . Once h(·) is known, it is possible to describe a closed convex set as the
intersection of all the half-spaces that contain it:

(4.12) P = {x : x, x∗  ≤ h(x∗ )}.
x∗

For every x belonging to a C-set P we have that −h(−x∗ ) ≤ x, x∗  ≤ h(x∗ ) for
every x∗ ∈ X ∗ . If P is 0-symmetric, then h(−x∗ ) = h(x∗ ) for all x∗ . The next lemma
characterizes Ψ(x, x̄) in terms of h.
Lemma 4.3. Let h(·) be the support functional of P. Then
x∗ , x − x̄
(4.13) Ψ(x, x̄) = sup ∗ ∗
.
x∗ =1 h(x ) − x , x̄

Proof. By a proper scaling of the vectors x∗ in (4.12), we know that x ∈ P iff

x, x∗  ≤ h(x∗ ) for every x∗ such that x∗  = 1. Then from (3.2)

1
Ψ(x, x̄) = inf α : x̄ + (x − x̄), x∗  ≤ h(x∗ ), for all x∗  = 1
α

1
= inf α : (x − x̄), x∗  ≤ h(x∗ ) − x̄, x∗ , for all x∗  = 1 .
α
from which follows (4.13).
The next lemma we present gives an upper bound to the values of the function
Ψ(x, x̄) when (x, x̄) ∈ αP × intP. It will be useful to guarantee a convergence speed
in the discrete-time case.
Lemma 4.4. Suppose P is 0-symmetric and let x ∈ αP and x̄ ∈ intP. Then
α+1
(4.14) Ψ(x, x̄) ≤ .
2

Proof. From Lemma 4.3 we need to prove that

x∗ , x − x̄ 1+α
(4.15) ∗ ∗
≤
h(x ) − x , x̄ 2
for every x∗ such that x∗  = 1. For every such x∗ we have that x, x∗  ≤ αh(x∗ )
whenever x ∈ αP. Thus
x∗ , x − x̄ αh(x∗ ) − x∗ , x̄
≤
h(x∗ ) − x∗ , x̄ h(x∗ ) − x∗ , x̄
x∗ ,x̄
α− h(x∗ )
(4.16) = x∗ ,x̄
1− h(x∗ )
 TRACKING DOMAIN OF ATTRACTION 983

for every x∗  = 1 (note that h(x∗ ) = 0). Now we notice that, since P is 0-symmetric,
x∗ , x̄
−1 ≤ ≤ 1.
h(x∗ )
Thus expression (4.16) is bounded by the extrema of the function
α−ξ
f (ξ) = with |ξ| ≤ 1.
1−ξ
Simple algebra shows that f (ξ) is continuous and decreasing thus attaining its maxi-
mum at ξ = −1. This proves (4.15).
Remark 4.2. The previous lemma gives a specific upper bound for Ψ(x, x̄) for
x̄ ∈ intP and x ∈ αP. If symmetry is removed, we have that there exists mα < 1
such that Ψ(x, x̄) ≤ mα < 1, although no upper bounds can be guaranteed a priori.
5. Proofs of the main results. For brevity, the next proofs will consider the
case of a symmetric P only. The general proofs follow immediately by considering
Remarks 4.1 and 4.2.
Proof of Proposition 3.1. In view of the homogeneity of φ(x) we have that

Φ(x, r) = φ(x̃)Ψ(x, x̄) + (1 − Ψ(x, x̄))ū

(5.1) = φ(Ψ(x, x̄)x̄ + x − x̄) + (1 − Ψ(x, x̄))ū.

By exploiting the Lipschitz condition on φ(x) and the triangular inequality we get

Φ(x1 , r) − Φ(x2 , r) ≤ LΨ(x1 , x̄)x̄ + x1 − x̄ − Ψ(x2 , x̄)x̄ − x2 + x̄

+(1 − Ψ(x1 , x̄))ū − (1 − Ψ(x2 , x̄))ū;

therefore, after proper rearrangement,

Φ(x1 , r) − Φ(x2 , r) ≤ Lx1 − x2  + (Lx̄ + ū)Ψ(x1 , x̄) − Ψ(x2 , x̄).

Let LP be the Lipschitz constant of the function Ψ(x, x̄) and let K be a constant such
that Lx̄ + ū ≤ K so that we can finally obtain

Φ(x1 , r) − Φ(x2 , r) ≤ (L + LP K)x1 − x2 .

The Lipschitz constant of Φ(x, r) is then LΦ = (L + LP K).

The continuity of Φ(x, r) can be inferred easily by considering expression (5.1)
which, in view of Lemma 4.1 and the Lipschitz condition of φ, is continuous with
respect to x, x̄, and ū. Now x̄ and ū are linear functions of r̄ = Γ(r), and Γ is
continuous. This completes the proof.
Proof of Theorem 3.2. We have already noticed that Φ(x, r) ∈ U, which is
an immediate consequence of the convexity of U and the fact that the control Φ(x, r)
as in (3.7)–(3.9) is the convex combination of φ(x̃) and ū, both belonging to U. The
condition x ∈ P for every t ≥ 0 will be proved by showing that Ψ(x, x̄) is a Lyapunov
function for system (2.2). Consider x ∈ P, x = x̄ and let

y = Ax + BΦ(x, r)
= A(x̃Ψ(x, x̄) + (1 − Ψ(x, x̄))x̄) + B(φ(x̃)Ψ(x, x̄) + (1 − Ψ(x, x̄)ū)
= Ψ(x, x̄)(Ax̃ + Bφ(x̃)) + (1 − Ψ(x, x̄)(Ax̄ + B ū)
= Ψ(x, x̄)(Ax̃ + Bφ(x̃)) + (1 − Ψ(x, x̄))x̄,
984 FRANCO BLANCHINI AND STEFANO MIANI

where we have replaced Ax̄ + B ū = x̄ due to the fact that (x̄, ū) is an equilibrium
pair. Define now ỹ = Ax̃ + Bφ(x̃), since P is domain of attraction as in definition 2.2.
Then ψ(Ax̃ + Bφ(x̃)) ≤ λψ(x̃) = λ, and thus ỹ ∈ λP. By virtue of (3.6),

Ψ(y, x̄) = Ψ(ỹΨ(x, x̄) + (1 − Ψ(x, x̄))x̄, x̄)

= Ψ((ỹ − x̄)Ψ(x, x̄) + x̄, x̄)
1+λ
= Ψ(x, x̄)Ψ(ỹ, x̄) ≤ Ψ(x, x̄),
2
where the inequality comes from Lemma 4.4. Thus

Ψ(Ax + BΦ(x, r), x̄) ≤ λT R Ψ(x, x̄).

Proof of Theorem 3.3 Let x ∈ P and x̄ ∈ (1 − )P. From (4.5) we need to

prove that Ψ̇(x, x̄) ≤ −β/2Ψ(x, x̄), which is true if

β
(5.2) ŵ, Ax + BΦ(x, r) ≤ − Ψ(x, x̄) < 0
2
is satisfied for all ŵ ∈ ∇Ψ(x, x̄) and x = x̄. Let, then, x = x̄. Note that

ŵ, Ax + BΦ(x, r)

=  ŵ , A(Ψ(x, x̄)x̃ + (1 − Ψ(x, x̄))x̄) + B(Ψ(x, x̄)φ(x̃) + (1 − Ψ(x, x̄))ū 
=  ŵ , [Ax̃ + Bφ(x̃)] Ψ(x, x̄) + (1 − Ψ(x, x̄)) [Ax̄ + B ū] .

We remind the reader that in the continuous-time case we have that Ax̄ + B ū = 0
and that (i) of Lemma 4.2 holds so that ŵ ∈ ∇Ψ(x, x̄), iff ŵ ∈ ∇Ψ(x̃, x̄). Since
x̃ ∈ ∂P, then, by virtue of (iii) of Lemma 4.2, ŵ = γ(ŵ, w)w, with w ∈ ∇ψ(x̃) and
γ(ŵ, w) ≥ 1/(2 − ). Therefore

ŵ, Ax + BΦ(x, r) = Ψ(x, x̄)ŵ, Ax̃ + Bφ(x̃)

= γ(ŵ, w) w, Ax̃ + Bφ(x̃) Ψ(x, x̄).

Due to the fact that P is a domain of attraction, we have that

w, Ax̃ + Bφ(x̃) ≤ −β for all w ∈ ∇ψ(x).

From (iii) of Lemma 4.2 we have that, if P is symmetric, γ(ŵ, w) ≥ 1/(2 − ) ≥ 1/2,
and thus we have (5.2).
Remark 5.1. According to Remarks 4.1 and 4.2, for nonsymmetric domains we
have that λT R = mα < 1 and βT R = βm > 0.
Proof of Theorem 3.4 We will consider the control law (3.7)–(3.9) and we will
prove the result for the continuous-time case only, as the discrete-time version can be
obtained by following a similar argument.
To show that the proposed control law maintains the state inside P for every
x(0) ∈ P, we need to show its positive invariance for the closed-loop system. But this
is immediate because we can just consider the original Lyapunov function ψ(x). For
x ∈ ∂P we have that Φ(x, r) = φ(x), and thus the Lyapunov derivative of the system
with the tracking control is

ψ̇(x(t)) = sup w, Ax + Bφ(x),

w∈∇ψ(x)
 TRACKING DOMAIN OF ATTRACTION 985

which is the same as the Lyapunov derivative of the system with original stabilizing
control associated with P. By Definition 2.3, the derivative for x ∈ ∂P is such that
ψ̇(x) ≤ −β,
which obviously prevents the state x(t) from leaving P [1].
The fact that the control Φ(x, r) ∈ U has been shown at the beginning of the
proof of Theorem 3.2 (we remind the reader that ū is computed with respect to the
“saturated” reference r̄ = Γ(r)).
To complete the proof we need now to show that y(t) approaches r∞ when t → ∞.
In view of Remark 2.1, this amounts to showing that x(t) → x̄∞ , say Ψ(x, x̄∞ ) → 0.
Let Φ(x(t), r(t)) and Φ(x(t), r∞ ) be the control actions associated with the refer-
ence values r(t) and r∞ , resp., so that we can write
ẋ = Ax + BΦ(x, r) = Ax + BΦ(x, r∞ ) + B[Φ(x, r) − Φ(x, r∞ )].
To show that x(t) → x̄∞ we consider the candidate Lyapunov function Ψ(x(t), x̄∞ ).
Take any vector w ∈ ∇Ψ(x(t), x̄∞ ). Then
w, Ax + BΦ(x, r) = w, Ax + BΦ(x, r∞ ) + w, B[Φ(x, r) − Φ(x, r∞ )]
  
z
β
≤ − Ψ(x(t), x̄∞ ) + |z(t)|,
2
where we have exploited the condition w, Ax + BΦ(x, r∞ ) ≤ − β2 Ψ(x(t), x̄∞ ) proven
in Theorem 3.3. As the above inequality holds for all w ∈ ∇Ψ(x(t), x̄∞ ), from (4.5)
we get
Ψ̇(x(t), x̄∞ , Φ(x(t), r(t))) ≤ −β/2Ψ(x(t), x̄∞ ) + |z(t)|.
The vector w ∈ ∇Ψ(x(t), x̄∞ ) is bounded due to the fact that Ψ(x, x̄∞ ) is Lipschitz
with respect to the first argument x. Since r(t) → r∞ ∈ R , there exists T1 such that
r(t) ∈ R 2
for t ≥ T1 . Φ(x, r) is continuous, thus it is uniformly continuous on P × R .
This means that z(t) → 0 as r̄(t) → r∞ . Standard Lyapunov arguments lead to the
conclusion that Ψ(x(t), x̄∞ ) → 0, say x(t) → x̄∞ .
6. Special domains of attraction. The proofs of the preceding sections have
allowed us to show that every domain of attraction is indeed a tracking domain. The
actual implementation of the proposed control law requires the computation of the
functions Ψ(x, x̄) and ψR
(r). In the next subsections we will see how the above
arguments apply when we consider different families of domains.
6.1. Ellipsoidal domains. Ellipsoids are very popular candidate domains of
attraction. They have been deeply investigated in literature, see for instance [18],
[21]. A detailed exposition on the properties of such sets can be found in [10]. Assume
that an ellipsoidal domain of attraction
P = {x : xT Qx ≤ 1}, Q > 0,
is available and this domain is associated with the linear control law φ(x) = Kx.
Since by Assumption 2.1 Mc (resp., Md ) is invertible we denote with Kxr and Kur ,
resp., the first n lines and the last m lines of Mc−1 (resp., Md ) so that x̄(r) = Kxr r
and ū = Kur r. The set R is the intersection of the sets
Rx = {r : rT Kxr
T
QKxr r ≤ 1}
986 FRANCO BLANCHINI AND STEFANO MIANI

and
Ru = {r : Kur r ∈ U}.
The Minkowski functional of R = (1 − )R is easily computable. For instance, if U
is the unit ball of some norm  · U , then Ru = {r : Kur rU ≤ 1} so that
ψR
= (1 − )−1 max{(rT Kxr
T
QKxr r)1/2 , Kur rU }.
With the choice of an ellipsoidal tracking domain, denoting by y = x − x̄, the
expression of Ψ(x, x̄) reduces to the following:
1
x̄T Qy + [(x̄T Qy)2 + y T Qy(1 − x̄T Qx̄)] 2
Ψ(x, x̄) = ;
1 − x̄T Qx̄
thus its computation essentially involves the computation of the three quantities
xT Qx̄, xT Qy and y T Qy.
The operations involved on-line in the implementation of the proposed control
law are hence the following:
1. Compute the “constrained” reference value rc = Γ(r);
2. Compute x̄ = Kxr rc and ū = Kur rc ;
3. Compute Ψ(x, x̄);
4. Set
Φ(x, r) = Kx + (1 − Ψ(x, x̄))(ū − K x̄).
Although we have explicitly considered the case in which the original controller is
linear, we can easily consider the case in which such a controller is nonlinear. The
price we pay is a less neat expression of the controller.
6.2. Polyhedral domains. Polyhedral invariant sets have been deeply investi-
gated for constrained stabilization problems [3], [11], [12], [17], [19], [20], [26], [28].
The basic reason is that if linear state and control constraints are considered, the
largest domain of attraction can be arbitrarily closely approximated by a polyhedral
set which is a domain of attraction. Such a result holds for both discrete-time [17], [19]
and continuous-time systems [7]. It is also known that once such a domain of attrac-
tion is available, a feedback stabilizing control law can be inferred. Therefore, in the
constrained stabilization problem, considering polyhedral sets as candidate domains
of attraction is not restrictive.
We consider now the implementation of a tracking control strategy for a polyhe-
dral domain of attraction. A polyhedral C-set P can be represented as
P = {x : Fi x ≤ 1, i = 1, 2, . . . , s} = {x : F x ≤ 1̄},
.
where F is the matrix whose rows are Fi , while 1̄ = [1 1 . . . 1]T . The Minkowski
functional associated with P is given by [5], [28]
ψ(x) = max Fi x.
i

From the above expression it is evident that in this particular case the set is the
intersection of a finite number of half-spaces, thus if we consider the function Ψ(x, x̄),
its expression is given by
. Fi (x − x̄)
Ψ(x, x̄) = max , x ∈ (1 − )P,
i 1 − Fi x̄
 TRACKING DOMAIN OF ATTRACTION 987

where  > 0 is a small number. Note that 1 − Fi x̄ ≥ , so singularities are avoided.

The construction of a polyhedral domain of attraction is a well-established topic
and allows the designer to derive a Lipschitz piecewise linear control law of the form
u = K(x)x according to [4], [17], which is positively homogeneous of order 1. Once
such a domain is known, the expression of the control can be derived by means of
(5.1) with the provided expression of Ψ.
After the determination of the value of Ψ, as for the other families considered
here, the first step necessary for the implementation of the tracking control (3.7)–
(3.9) concerns the evaluation of the function
ψR
(r) = (1 − )−1 max {ψ(Kxr r), Kur rU }
(we remind the reader that x̄ = Kxr r and ū = Kur r), whose computation is needed
for the determination of the “constrained” reference value rc = Γ(r).
6.3. Smoothed domains. In this section we consider the family of domains of
attraction whose Minkowski functional is given by
 s  2p
1

ψ(x) = σ2p (Fi x)
i=1

where

xr if x ≥ 0,
σr (x) =
0 otherwise.
This family has been analyzed in [9] for the constrained stabilization problem.
Similar functions have been used for tracking constant signals for continuous-time
systems in the presence of state constraints only in [8]. It is helpful to recall that, al-
though the functions used in [8] are similar to those proposed here, they are essentially
different. Their definition is based on an analytic expression and such definition does
not apply to the Minkowski function of a generic C-set. In particular the unitary level
surface is not invariant when the reference is allowed to vary. They are unsuitable for
the cases in which there are control constraints and variable reference signals.
Here we reconsider this class for tracking purposes for continuous-time systems
and we assume that the input constraints are of the form U = {u : u∞ ≤ 1}. The
reason for considering this class of functions comes from the fact that the differentia-
bility of this class of functions together with the assumption on U to be a box allows
for the determination of an explicit control law (as opposed to the control law which
can be associated with a polyhedral domain which is general implicit). Indeed it can
be shown that to every such domain of attraction P can be associated the Lipschitz
control function
φ(x) = −sat(k(B T F T Γ(x)T ))ψ(x)
where
.  
Γ(x) = ψ(x)(1−2p) (F1 x)2p−1 . . . (Fs x)2p−1
and k is a computable positive constant (see [9]). The drawback of the capability of
determining an explicit control law is given anyway by the fact that it is not possible
to give an explicit expression for Ψ(x, x̄), but this has to be computed by bisection
on the parameter α in (3.2).
Again, the control action requires the computation on-line of Ψ(x(t), x̄(t)) and
the application of (5.1).
988 FRANCO BLANCHINI AND STEFANO MIANI

7. Nonsquare systems. So far we have considered the case in which the system
has as many inputs as outputs. It is not difficult to extend the results to nonsquare
systems. Consider the nonsquare matrix Mc and the equation




A B x̄(r) 0
= .
C D ū(r) r̄

It is well known that, if the above equation does not admit an unique solution, two
cases are possible: for a given r̄ there may be infinitely many solutions, or there may
be none. In the first case one has to choose one vector [x̄(r)T ū(r)T ] among the
solutions. One obvious choice is to consider the minimum (possibly weighted) norm
solution given by


†

x̄(r) A B 0
(7.1) = ,
ū(r) C D r̄

where Mc† is the Moore–Penrose inverse of Mc .

In the case in which there are no solutions, the obvious extension, the least square
solution, is not suitable since for tracking purpose we need the steady-state condition


  x̄(r)
(7.2) A B =0
ū(r)

to be satisfied exactly, a condition which is in general not satisfied by the solution

(7.1). In this case we may take, among all the equilibrium points satisfying (7.2), those
which minimize ȳ − r̄. This is obtained by first parameterizing all the elements of
the kernel of [A B] as x = N z, u = M z, where AN + BM = 0 and then choosing for
a given reference r the least square solution of

min [CN + DM ]z − r.

z

Since the minimum is obtained for z = [CN + DM ]† r we finally obtain



x̄(r) N [CN + DM ]† r
= .
ū(r) M [CN + DM ]† r

Once the linear relation above is established, all the theory developed in the previous
chapters remains valid. Clearly in the over-determined (i.e., no exact solution) case,
asymptotic tracking, even for a constant signal, is not possible in general.
8. Examples.
8.1. Continuous-time example: linear state feedback. As a first example
we consider the following continuous-time system:



1 .4 0
ẋ(t) = x(t) + u(t)
.8 .5 1
 
y(t) = .2 .1 x(t).

We derived for this system a linear state feedback compensator u = Kx which

is such to stabilize the closed loop system while guaranteeing that for every initial
state x(0) ≤ 1 we have Kx ≤ Umax along the system trajectory. By solving the
 TRACKING DOMAIN OF ATTRACTION 989

15

10

5
x’Qx<1
Kx=5

0 Kx=5

10

15
5 4 3 2 1 0 1 2 3 4 5

Fig. 3. Domain of attraction.

corresponding set of linear matrix inequalities (LMI) [10] for Umax = 5, we obtained
the gain
 
K = −3.6845 −1.5943
which corresponds to the ellipsoidal domain of attraction (with speed of convergence
β = .0001) P = {x : xT Qx ≤ 1}, where


.7859 .2950
Q= .
.2950 .1172
The set P is the ellipse depicted in Figure 3 and is by construction contained in the
region where |Kx| ≤ Umax (also reported in Figure 3); thus the constraints on the
reference value r̄ can be derived by the set P alone and result in |r̄| ≤ .2403. We
slightly reduced this value to .24 to guarantee that x̄(r̄) belongs to the interior of P.
This means that r̄ = Γ(r) = 0.24 sat(r/0.24).
We report in Figure 4 the zero initial state time evolution of the output corre-
sponding to the reference signal (dashed line)

0.40 for 0 ≤ t ≤ 100,
r(t) =
0.20 + 0.20e−(t−100)/10 for 100 < t
and we report in Figure 5 the state space evolution.
8.2. Discrete-time example: linear variable structure state feedback.
As a second example we consider the discrete-time system



1 .3 .5
x(k + 1) = x(k) + u(k)
−1 1 1
 
y(k) = −.1 .3 x(k)
with the state and control constraint sets, resp., given by
X = {x : x∞ ≤ 1}
990 FRANCO BLANCHINI AND STEFANO MIANI

0.4

0.35

0.3

0.25

0.2
y,ref

0.15

0.1

0.05

0.05

0.1
0 50 100 150 200 250
t

Fig. 4. Output and reference (dashed) time evolution.

15

10

5
x2

10

15
5 4 3 2 1 0 1 2 3 4 5
x1

Fig. 5. Domain of attraction and state space evolution.

and

U = {u : |u| ≤ 1}.

For this system we computed a symmetric polyhedral domain of attraction P =

{x : F x∞ ≤ 1} with speed of convergence λ = .9 (which is the polyhedral region
in Figure 6), where
 
0 1
F =  1.8160 −0.2421  .
1.3140 0.1932

In this case the constraints on the reference value derive from the constraint that
x̄ ∈ P and translate in |r̄| ≤ .27. The linear variable structure controller associated
with P is given by u(x) = ki x with i = arg maxj |fj x|, where fi and ki are the ith
 TRACKING DOMAIN OF ATTRACTION 991

x2

1
0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8
x1

Fig. 6. Polyhedral domain of attraction and state space evolution.

rows of F and K, resp., where

 
−0.4690 −0.7112
K =  −0.6355 −0.7806  .
−1.3140 −0.1931

We applied the control law proposed in section 6.2 with  = .01, starting from zero
initial state, for the reference signal

r(t) = 0.2 + 0.4 ∗ sin(0.01 ∗ t) ∗ e−0.005∗t .

Figure 7 depicts the time evolution of the output and the reference value, and
Figure 6 shows the corresponding state space trajectory.
9. Conclusions. In this paper it was shown that every set of initial states which
can be brought to the origin while assuring that no control and state constraints are
violated is indeed a tracking domain. This amounts to saying that it is possible to
track a given reference value only if its corresponding state and input equilibrium pair
belongs to a domain of attraction, whereas if this condition is not met the state is not
“trackable.” We have provided this result based on the system structure, without as-
suming the existence of a stabilizing compensator, by furnishing a Lyapunov function
which is derived by the domain of attraction and which allows us to derive a tracking
strategy that avoids state and input constraints yet guarantees an a priori speed of
convergence. Future directions in this area concern the possibility of deriving stabi-
lizing tracking strategies capable of maintaining the tracking error within prespecified
bounds and of improving the speed of convergence.
The presented approach can be applied as well to systems with control constraints
only. In this case the domain of attraction can be computed by considering a fictitious
(sufficiently large) state constraints region and by applying the techniques suggested
in [9], [17], [20]. We finally remind the reader that for systems with no state con-
straints and free from eigenvalues in the open right half plane (or with no eigenvalues
outside the closed unit disk) a completely different approach can be used for global
992 FRANCO BLANCHINI AND STEFANO MIANI

0.4

0.3
r, y

0.2

0.1

0 100 200 300 400 500 600 700 800 900 1000
t

Fig. 7. Output and reference time-evolution.

stabilization [25] or semiglobal stabilization [24], [27]. In this case those approaches
can be successfully applied to the solution of the tracking problem with an arbitrarily
large initial condition set.

Appendix. Given a differentiable positively homogeneous function of order 1

F : Rn → R it is well known that ∇F (y), y = F (y). This property can be extended
easily to our case of nondifferentiable functions under the convexity assumption.
Proof of the equality (4.8). Let F : Rn → R be a convex and positively
homogeneous function of order 1. Then for any y and any w ∈ ∇F (y), we have
w, y = F (y).
For y = 0 the property is obvious. So let y = 0. By definition any w ∈ ∇F (y) is
such that

F (z) − F (y) ≥ w, z − y

for all z ∈ Rn , that is

F (z) − w, z ≥ F (y) − w, y.

Now let z = λy with λ > 0. Then

λ [F (y) − w, y] ≥ F (y) − w, y for all λ > 0,

which necessarily implies

F (y) − w, y = 0.
 TRACKING DOMAIN OF ATTRACTION 993

Acknowledgments. The authors are grateful to Professor Zanolin for helpful

discussions on the paper.

REFERENCES

[1] J. P. Aubin and A. Cellina, Differential Inclusions, Springer-Verlag, Berlin, 1984.

[2] A. Bemporad, A. Casavola, and E. Mosca, Nonlinear control of constrained linear systems
with predictive reference management, IEEE Trans. Automat. Control, 42 (1997), pp. 340–
349.
[3] A. Benzaouia and C. Burgat, The regulator problem for a class of linear systems with con-
strained control, Systems Control Lett., 10 (1988), pp. 357–363.
[4] F. Blanchini, Ultimate boundedness control for discrete-time uncertain system via set-induced
Lyapunov functions, IEEE Trans. Automat. Control, 39 (1994), pp. 428–433.
[5] F. Blanchini, Non-quadratic Lyapunov functions for robust control, Automatica, 31 (1995),
pp. 451–461.
[6] F. Blanchini, Set invariance in control—a survey, Automatica, 35 (1999), pp. 1747–1767.
[7] F. Blanchini and S. Miani, Constrained stabilization for continuous-time systems, Systems
Control Lett., 28 (1996), pp. 95–102.
[8] F. Blanchini and S. Miani, Set based constant reference tracking for continuous-time con-
strained systems, in Advances of Stability Theory at the End of the XXth Century, A.A.
Martynyuk, ed., Stability Control Theory Methods Appl. 13, Gordon and Breach, London,
UK, 2000, pp. 267–278.
[9] F. Blanchini and S. Miani, Constrained stabilization via smooth Lyapunov functions, Systems
Control Lett., 35 (1998), pp. 155–163.
[10] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear Matrix Inequalities in Sys-
tem and Control Theory, SIAM Studies in Applied Mathematics 15, SIAM, Philadelphia,
PA, 1994.
[11] C. Burgat and S. Tarbouriech, Positively invariant sets for constrained continuous-time
systems with cone properties, IEEE Trans. Automat. Control, 39 (1994), pp. 401–405.
[12] E. B. Castelan and J. C. Hennet, Eigenstructure assignment for state constrained linear
continuous time systems, Automatica, 28 (1992), pp. 605–611.
[13] F. Clarke, Optimization and Nonsmooth Analysis, John Wiley & Sons, New York, 1983.
[14] E. G. Gilbert and K. T. Tan, Linear systems with state and control constraints: The theory
and the applications of the maximal output admissible sets, IEEE Trans. Automat. Control,
36 (1991), pp. 1008-1020.
[15] E. G. Gilbert, I. Kolmanowsky, and K. T. Tan, Discrete–time reference governors and
the nonlinear control of systems with state and control constraints, Internat. J. Robust
Nonlinear Control, 5 (1995), pp. 487-504.
[16] T. J. Graettinger and B. H. Krogh, On the computation of reference signal constraints for
guaranteed tracking performance, Automatica, 28 (1992), pp. 1125–1141.
[17] P. O. Gutman and M. Cwikel, Admissible sets and feedback control for discrete-time lin-
ear systems with bounded control and states, IEEE Trans. Automat. Control, 31 (1986),
pp. 373–376.
[18] P. O. Gutman and P. Hagander, A new design of constrained controllers for linear systems,
IEEE Trans. Automat. Control, 30 (1985), pp. 22–33.
[19] S. S. Keerthy and E. G. Gilbert, Computation of minimum-time feedback control laws for
discrete-time systems with state and control constraints, IEEE Trans. Automat. Control,
32 (1987), pp. 432–435.
[20] J. B. Lasserre, Reachable, controllable sets and stabilizing control of constrained systems,
Automatica, 29 (1993), pp. 531–536.
[21] Z. Lin, A. A. Stoorvogel, and A. Saberi, Output regulation for linear systems subject to
input saturation, Automatica, 32 (1996), pp. 29–47.
[22] D. G. Luenberger, Optimization by Vector Space Methods, John Wiley & Sons, New York,
1969.
[23] R. T. Rockafellar, Convex Analysis, Princeton University Press, Princeton, NJ, 1972.
[24] A. Saberi, Z. Lin, and A. R. Teel, Control of linear systems with saturating actuators, IEEE
Trans. Automat. Control, 41 (1996), pp. 368–378.
[25] H. J. Sussmann, E. D. Sontag, and Y. D. Yang, A general result on the stabilization of linear
systems using bounded controls, IEEE Trans. Automat. Control, 39 (1994), pp. 2411–2425.
994 FRANCO BLANCHINI AND STEFANO MIANI

[26] M. Sznaier and M. J. Damborg, Heuristically enhanced feedback control of constrained

discrete-time systems, Automatica, 26 (1990), pp. 521–532.
[27] A. R. Teel, Global stabilization and restricted tracking for multiple integrators with bounded
controls, Systems Control Lett., 18 (1992), pp. 165–171.
[28] M. Vassilaki and G. Bitsoris, Constrained regulation of linear continuous-time dynamical
systems, Systems Control Lett., 47 (1989), pp. 247-252.