5

Lyapunov Stability Theory

5.1 Introduction
The study of the stability of dynamical systems has a very rich history. Many
famous mathemat icians, physicists, and astronomers worked on axiomatizing the
concepts of stability. A problem, which attracted a great deal of early interest was
the problem of stability ofthe solar system, generalized under the title "the N-body
stability problem." One of the first to state formally what he called the principle of
"least total energy" was Torricelli (1608-1647), who said that a system of bodies
was at a stable equilibrium point if it was a point of (locaIly) minimal total energy.
In the middle of the eighteenth century, Laplace and Lagrange took the Torricelli
principle one step further: They showed that if the system is conservative (that is, it
conserves total energy-kinetic plus potential), then astate corresponding to zero
kinetic energy and minimum potential energy is a stable equilibrium point. In turn,
several others showed that Torricelli 's principle also holds when the systems are
dissipative, i.e., total energy decreases along trajectories of the system . However,
the abstract definition of stability for a dynamical system not necessarily derived
for a conservative or dissipative system and a characterization of stability were
not made till 1892 by a Russian mathematician/engineer, Lyapunov, in response
to certain open problems in determining stable configurations of rotating bodies
of fluids posed by Poincare. The original paper of Lyapunov of 1892, was trans-
lated into French very shortly there after, but its English translation appeared only
recently in [I 93l. The interested reader may consult this reference for many inter-
esting details, as weIl as the historical and biographical introduction in this issue
ofthe International Journal ofControl by A. T. Fuller. There is another interesting

S. Sastry, Nonlinear Systems

© Springer Science+Business Media New York 1999
 5.2 Definitions 183

survey paper about the impact of Lyapunov 's stability theorem on feedback control
by Michel [208].
At heart, the theorems of Lyapunov are in the spirit of Torricelli 's principle.
They give a precise characterization of those functions that qualify as "valid energy
functions" in the vicinity of equilibrium points and the notion that these "energy
functions" decrease along the trajectories of the dynamical systems in question.
These precise concepts were combined with careful definitions of different notions
of stability to give some very powerful theorems . The exposition of these theorems
is the main goal of this chapter.

5.2 Definitions
This chapter is concemed with general differential equations of the form

x = f t», t) , x(to) = xo, (5.1)

where x E ]Rn and t ~ O. The system defined by (5.1) is said to be autonomous

or time-invariant if f does not depend explicitly on t , It is said to be linear if
f t», t) = A(t)x for some A(·) : ]R+ t-+ ]Rnx n and nonlinear otherwise. In this
chapter, we will assurne that f tx , t) is piecewise continuous with respect to t ,
that is, there are only finitely many discontinuity points in any compact set. The
notation Bi; will be short-hand for B(O, h) , the ball of radius h centered at O.
Properties will be said to be true

• locally if they are true for all Xo insome ball Bi,

• globally if they are true for all Xo E R".
• semi-globally if they are true for all Xo E Bh with h arbitrary.
• uniformly if they are true for all to ~ O.
In the development that folIows, unless explicitly specified, properties are true
locally. The first few definitions and estimates are repeated from Chapter 3.

5.2.1 The Lipschit: Condition and Consequences

Definition 5.1 Lipschitz Continuity. The function f is said to be locally
Lipschitz continuous in x iffor some h > 0 there exists I ~ 0 such that

(5.2)

for all Xl , X2 E Bi, t ~ O. The constant I is called the Lipschitz constant. A

definitionfor globally Lipschitz continuousfunctionsfollows by requiring equation
(5.2) to holdfor XI, X2 E R". The definition ofsemi-globally Lipschitz continuous
functions holds as well by requiring that equation (5.2) hold in Bhfor arbitrary h
but with I possibly a function ofh. The Lipschitz property is by default assumed to
be uniform in t .
184 5. Lyapunov Stability Theory

If I is Lipschitz continuous in x , it is continuous in x. On the other hand, if I

has bounded partial derivatives in x, then it is Lipschitz. Formally, if

o.r«,t) := [aj;] aXj

denotes the partial derivative matrix of I with respect to x (the subscript 1 stands
for the first argument of I (x, t)), then ID) I (x , t) I :s I implies that f is Lipschitz
continuous with Lipschitz constant I (again locally, globally or semi-globally de-
pending on the region in x that the bound on ID 2 f (x , t)1 is valid). We have seen
in Chapter 3 that if I is locally bounded and Lipschitz continuous in x, then the
differential equation (5.1) has a unique solution on some time interval (so long as
x E Rh).

Definition 5.2 Equilibrium Point. x* is said to be an equilibrium point 01(5.1)

if I(x* , t) == Olor alt t ::: O.
If I (x, t) is Lipschitz continuous in x, then the solution x (t) == x* for all t is
called an equilibrium solution. By translating the origin to the equilibrium point x*
we can make 0 an equilibrium point. Since this is of great notational help, we will
henceforth assurne that 0 is an equilibrium point of (5.1). One of the most important
consequences of the Lipschitz continuity hypothesis is that it gives bounds on the
rate of convergence or divergence of solutions from the origin:
Proposition 5.3 Rate of Growth/Decay. 11 x = 0 is an equilibrium point 01
(5.1) I is Lipschitz in x with Lipschitz constant land piecewise constant with
respect to t, then the solution x (t) satisfies
Ixolel(t-to) ::: Ix(t)1 ::: Ixole-l(t-to) (5.3)

as lang as x(t) remains in Rh.

Proof: Since Ixl 2 = x T x , it follows that

I:t IXl21 = 2 IX I !:tlXII (5.4)

T
= 2/X :tXI s 2lxll:tXI ,
so that

Since I (x, t) is Lipschitz continuous, and I (x, 0) = 0 it follows that

d
-/lxl :s dt [x] :s Ilxl (5.5)

Using the Bellman Gronwalliemma twice (refer back to Chapter 3 and work out
the details for yourself) yields equation (5.3), provided that the trajectory stays in
the ball Rh where the Lipschitz condition holds. 0
 5.2 Definitions 185

The preceding proposition implies that solutions starting inside B h will stay in
Bh for at least a finite time . Also, if f(x, t) is globally Lipschitz, it guarantees that
the solution has no finite escape time, that is, it is finite at every finite instant. The
proposition also establishes that solutions x(t) cannot converge to zero faster than
exponentially.
We are now ready to make the stability definitions. lnformally x = 0 is stable
equilibrium point if trajectories x (t) of (5.1) remain elose to the origin if the initial
condition Xo is elose to the origin. More precisely, we have the following definition.
Definition 5.4 Stability in the sense ofLyapunov. The equilibrium point x = 0
is called a stable equilibrium point 0/(5.1) iffor all to ~ 0 and E > 0, there exists
8(to, E) such that

IXol < 8(to, E) =} Ix(t)! < E Vt ~ to. (5.6)

where x(t) is the solution 0/(5.1) starting from Xo at to.
The definition is illustrated in Figure 5.1, showing the trajectories starting in a
ball Bö and not leaving the ball B€. Sometimes this definition is also called stability
in the sense of Lyapunov (i.s.L.) at time to.
Definition 5.5 Uniform Stability. The equilibrium point x = 0 is called a
uniformly stable equilibrium point 0/ (5.1) if in the preceding definition 8 can be
chosen independent 0/ to.
Intuitively, the definition of uniform stability captures the notion that the equi-
librium point is not getting progressively less stable wilh time. Thus, in particular,
it prevents a situation in which given an E > 0, the ball of initial conditions of
radius 8 (to, E) in the definition of stability required to hold trajectories in the E ball
tends to zero as to ~ 00 . The notion of stability is weak in that it does not require
that trajectories starting elose to the origin to tend to the origin asymptotically.
That property is ineluded in adefinition of asymptotic stability:
Definition 5.6 Asymptotic Stability. The equilibrium point x = 0 is an
asymptotically stable equilibrium point 0/(5.1) if

Be - - - - - - - - - - - -

FIGURE 5.1. Illustrating the definition of stability

 186 5. Lyapunov Stability Theory

• x = 0 is a stable equilibrium point of(5.1),

• x = 0 is attractive, that is for all to ::: 0 there exists a 8 (to) such that
IXol < 8::} lim Ix(t)1 = O.
/-+00

An interesting feature of this definition is that it requires two separate condi-

tions: one, that the equilibrium point be stable and two that trajectories tend to the
equilibrium point as t ~ 00. Though, it may superficially appear to be the case,
the requirement that trajectories converge to the origin does not imply the stability
of the equilibrium point. To illustrate this, we consider the following example:
• 2 2
XI = XI -x2'
(5.7)

The phase portrait of this system (also presented as an example of an equilibrium

point with index 2 in Chapter 2) is as shown in Figure 5.2. All trajectories tend to
the origin as t ~ 00, except for the trajectory that follows the positive XI axis to
+00. By assuming that this "point at infinity" is the same as the point at infinity
at XI = -00 we may assert that all trajectories tend to the origin.' However, the
equilibrium point at the origin is not stable in the sense of Lyapunov: Given any
E > 0, no matter how small a 8 we choose for the ball of initial condition, there
are always some initial conditions close to the XI axis which will exit the E ball
before converging to the origin. The trajectory starting right on the XI axis gives a
hint to this behavior.
Definition 5.7 Uniform Asymptotic Stability. The equilibrium point X = 0 is
a uniformly asymptotically stable equilibrium point of(5.1) if

FIGURE 5.2 . An equilibrium point, which is not stable, but for wh ich all trajectories tend
to the origin if one includes the point at infinity

'This procedure is referred to in mathematical terms as the AlexandrotT or one -point

compactification of]R2. Another way of conceptualizing this procedure is to imagine that
the plane is mapped onto a sphere with the origin corresponding to the north pole and the
"point at infinity" to the south pole; that is, the state space of the system is really the sphere
S2.
 5.2 Definitions 187

• x = 0 is a uniformly stable equilibrium point of(5.1).

• The trajectory x(t) converges uniformly to 0, that is, there exists 8 > 0 and
a function y (r , xo) : R+ x IR" ~ IR+ such that lim r -+ oo Y (r, xo) = 0 for all
Xo E Bb and

IXol < 8 => Ix (t) 1 S y(t - te. xo) V t 2: to.

The previous definitions are local, since they concern neighborhoods ofthe equi-
librium point. Global asymptotic stability and global uniform asymptotic stability
are defined as folIows:

Definition 5.8 Global Asymptotic Stability. The equilibrium point x = 0

is a globally asymptotically stable equilibrium point of (5.1) if it is stable and
limHOOx(t) = Oforall Xo E IR".

Definition 5.9 Global Uniform Asymptotic Stability. The equilibrium point

x = 0 is a globally, uniformly, asymptotically stable equilibrium point of (5.1)
if it is globally asymptotically stable and if in addition, the convergence to the
origin of trajectories is uniform in time, that is to say that there is a function
y : IR" x R+ ~ IR such that

Ix(t)1 s y(xo , t - to) V t 2: O.

It is instructive to note that the definitions of asymptotic stability do not quantify

the speed of convergence of trajectories to the origin. For time-invariant linear
systems, the speed of convergence of trajectories either tü or from thc origin is
exponential, but for time varying and nonlinear systems, the rate of convergence
can be of many different types, for example as
I
..ji'
It is a good exercise to write down examples of this behavior. There is a strong
form of stability which demands an exponential rate of convergence:

Definition 5.10 Exponential Stability, Rate ofConvergence. The equilibrium

point x = 0 is an exponentially stable equilibrium point of (5.1) if there exist
m, a > 0 such that

(5.8)

for all Xo E Bh , t 2: to 2: O. The constant a is calied (an estimate of) the rate of
convergence.

Global exponential stability is defined by requiring equation (5.8) to hold for aII
Xo E R". Semi -global exponential stability is also defined analogously except that
m, tx are aIIowed to be functions of h. For linear (possibly time-varying) systems
it will be shown that uniform asymptotic stability is equivalent to exponential
stability, but in general, exponential stability is stronger than asymptotic stability.
188 5. LyapunovStability Theory

5.3 Basic Stability Theorems of Lyapunov

The so-called second method of Lyapunov enables one to determine the stability
properties of a system (5.1) without explicitly integrating the differential equation.
The method is a generalization of the basic notion that some measure of "energy
dissipation" in a system enables us to conclude stability. To make this precise we
need to define exactly what one means by a " measure of energy,' that is, energy
functions . This needs the following preliminary definitions (due, in the current
form to Hahn [I 24]):

5.3.1 Energy-Like Functions

Definition 5.11 Class K, KR Functions. Afunction a(·) : JR+ ~ JR+ belongs
to class K (denoted by a (.) E K) if it is continuous, strictly increasing and a (0) =
O. Thefunction a(·) is said to belong to dass KR if a is 0/ dass K and in addition,
a(p) ~ 00 as p ~ 00.

Definition 5.12 Locally Positive Definite Functions. A continuous function

v(x , t) : JRn X lR+ ~ lR+ is called a locally positive definite function (l.p.d.f) if,
for some h > 0 and some a(·) 0/ dass K,

v(O, t) = 0 and v(x, t) ::: a(lxl) Vx E Rh, t::: O. (5.9)

An I.p.d.f. is locally Iike an "energy function." Functions which are globally Iike
"energy functions" are called positive definite functions (p.d.f.s) and are defined
as folIows:

Definition 5.13 Positive Definite Functions. A continuous function v(x, t) :

lRn x lR+ ~ JR+ is called a positive definite function (p.d.f.) if for some a(·) 0/
dass KR ,

v(O, t) =0 and v(x, t) ::: a(lxl) V x E JRn , t::: 0 (5.10)

and, in addition, ai p) ~ 00 as p ~ 00.

In the preceding definitions of l.p.d.f.s and p.d.f.s, the energy was not bounded
above as t varied. This is the topic of the next definition

Definition 5.14 Decrescent Functions. A continuous function v(x, t) : JRn X

lR+ ~ JR+ is called a decrescent function if, there exists a function ß (.) 0/ dass
K , such that

v(x , t) ::; ß(!xl) Vx E Rh, t::: O. (5.11)

Example 5.15 Examples of Energy-like Functions. Here are some examples

0/ energy like functions and their membership in the various classes introduced
above. It is an interesting exercise to check the appropriate functions 0/ dass K
and KR that need to be used to verlfy these properties.
 5.3 Basic Stability Theorems of Lyapunov 189

1. v (x , t) = jx l2 : p .d.f. , decrescent .
2. v(x , t) = x T P x, with PE IR n x n > 0: p .d.f. , decrescent.
3. v (x , t) = (t + 1)lx I2 : [Link].
4. v(x, r) = e- ' lx I2 : decrescent.
5. v (x , t ) = sin 2 ( jx I2 ) : [Link]., decrescent.
6. v (x, t) = e' x T P X with P not po sitive definite: not in any of the abo ve classes.
7. v (x , t ) not explicitly depending on time t: decrescent.

5.3.2 Basic Theorems

Generally speaking, the basic theorem of Lyapunov states that when v( x , t) is
a p.d.f. or an l.p.d.f. and dv(x , I)/dt ~ 0 then we can conclude stability of the
equilibrium point. The time derivative is taken along the traje ctor ies of (5.1), i.e.,
dv(x , I) I av(x, t) av(x, t)
= + f t», t) . (5.12)
dt (5.1) at ax
The rate of change of v(x, t) along the trajectories ofthe vector field (5.1) is also
called the Lie derivative of v (x, t) along f (x , 1). In the statement of the following
theorem recall that we have translated the orig in to lie at the equilibrium point
under consideration .
Theorem 5.16 Basic Lyapunov Theorems.

Condltlons on Conditions on Conclusions

v(x , t ) -v(x, t )

I. I.p.d.f. ::: Olocally stable

2. l.p.d.f., decrescent ::: Olo cally uniformly stable

3. l.p.d.f. , decrescent l.p.d.f, uniformly asymptotically stable

4. p.d.f. , decrescent p.d.f. globally unif. asymp. stable

Proof:
1. Since v is an l.p.d.f., we have that for some a(·) E K,

v( x, t) ::: a(lxl) V x E Bs • (5.13)

Also, the hypothesis is that
ü(x , t ) ~ 0, V t ::: to, V x E B, (5.14)

Given E > 0, define EI = min (E, r, s ). Choose 8 > 0 such that

ß(to, 8) := sup v(x, to) < a (Ed.

lxi::.

Such a8 always exists, since ß(to, 8) is a continuous function of8 andrr (s j ) > O.
We now claim that Ix( to)1 ~ 8 implies that Ix(t)1 < EI vt ::: to. The proof is
190 5. Lyapunov Stability Theory

by contradiction. Clearly since

it follows that Ix(to) I < EI . Now, if it is not true that Ix(t) I < EI for all t, let
t) > to be the first instant such that Ix(t) I 2: EI. Then

(5.15)

Hut this is a contradiction, since v(x(t), t) ~ 0 for all [x] < E) . Thus,

Ix(t)1 < EI V t 2: to.

2. Since v is decrescent,

(5.16)

is nondecreasing and satisfies for some d

ß(~) < 00 for 0 ~ ~ ~ d.

Now choose ~ such that ß(~) < a(E)) .

3. If -v(x , t) is an l.p.d.f. , then v(x, t) satisfies the conditions of the previous
proof so that 0 is a uniformly stable equilibrium point. We need to show the
existence of a ~ I > 0 such that for E > 0 there exists T (E) < 00 such that

The hypotheses guarantee that there exist functions a (.), ß(.), y (.) E K such
that Vt 2: to, Vx E B, such that

a(lxl) ~ v(x, t) ~ ß(lxl),

v(x, t) s -y(lxl).

Given E > 0, define ~1, ~2 and T by

ß(~)) < a(r),

ß(~2) < min(a(E), ß(~))),

T = a(r)/Y(~2) .

This choice is explained in Figure 5.3. We now show that there exists at least
one instant tz E [tl, t) + Tl when IXol < ~2' The proof is by contradiction.
Recall the notation that rjJ(t, xo, to) stands for the trajectory of (5.1) starting
from Xo at time to. Indeed, if
 5.3 Basic Stability Theorems of Lyapunov 191

a(lxl)

FIGURE 5.3. The choice of constants in the proof of Lyapunov's Theorem.

then it follows that

oS + T. Xo. td. t1 + T)

r
a(~2) S V(S(tl

= V(tl ' xo) + ver, ,pCr, Xo. t) )dr

11

S ß(OI) - TY(02)
s ß(od - a(r)

< 0,

establishing the contradiction (compare the ends of the preceding chain of

inequalities to see this). Now, if t ~ tl + T, then

a(I,p([Link]) S v(t.,p([Link],td)

S V(t2. ,p(t2, Xo, «».

since v(x. t) S O. (Actually, the definition of 01 guarantees that the trajectory
stays in B, so that v(x, t) sO.) Thus,

a(I,p(t , xo , tl)1) S V(t2, ,p(t2. XO, t)) S ß(I,p(t2' Xo. tIl)

s ß(02)
< aCE)

so that 1,p(t2, xo, tdl < E for t ~ t) + T. o

Remarks:
1. The tabular version of Lyapunov's theorem is meant to highlight the following
correlations between the assumptions on v(x, t), v(x , t) and the conclusions:
a, Decrescence of v(x, t) is associated with uniform stability and the 10-
cal positive definite character of v(x, t) being associated with asymptotic
stability.
b, -v(x, t) is required to be an I.p.d.f. for asymptotic stability,
c, v(x. t) being a p.d.f. is associated with global stability.
192 5. Lyapunov Stability Theory

However, we emphasize that this correl ation is not perfect, since v(x, t) being
I.p.d.f. and - iJ (x, t) being l.p.d.f. does not guarantee local asymptotic stability.
(See Problem 5.1 for a counterexample from Massera [227]) .
2. The proof of the theorem, while seemingly straightforward, is subtle in that it
is an exercise in the use of contrapositives.

5.3.3 Examples ofthe Application ofLyapunov's Theorem

1. Consider the following model of an RLC circuit with a linear inductor, nonlinear
capacitor, and inductor as shown in the Figure 5.4. This is also a model for a
mechanical system with a mass coupled to a nonlinear spring and nonlinear
damper as shown in Figure 5.4. Using as state variables Xl, the charge on the
capacitor (respectively, the position of the block) and X2 , the current through
the inductor (respectively, the velocity of the block) the equations describing
the system are

(5.17)

Here f(-) is a continuous function modeling the resistor current-voltage char-

acteristic, and g (.) the capacitor charge-voltage characteristic (respectively the
friction and restoring force models in the mechanical analog) . We will assume
that f, g both modellocally passive elements, i.e., there exists a ao such that
af(a) 2: 0 Va E[-ao,ao],
ag(a) 2: 0 Va E [-ao, aol.

The Lyapunov function candidate is the total energy of the system, namely,
X2 ( XI
v(x) = ; + 10 g(a)da.

The first term is the energy stored in the inductor (kinetic energy of the body)
and the second term the energy stored in the capacitor (potential energy stored in

IH Xl
--. 1

~FrictioJ [(Xz)

FIGURE 5.4. An RLC circuit and its mechanical analogue

 5.3 Basic Stability Theoremsof Lyapunov 193

the spring). The function v(x) is an l.p.d .f., provided that g(x,) is not identically
zero on any interval (verify that this follows from the passivity of g). Also,

v(x) = X2[- !(X2) - g(xl)] + g(XI)X2

= -X2!(X2):::: 0

when IX21 is less than Go. This establishes the stability but not asymptotic sta-
bility of the origin. In point of fact, the origin is actually asymptotically stable,
but this needs the LaSalle principle, which is deferred to a later section.
2. Swing Equation
The dynamics of a single synchronous generator coupled to an infinite bus is
given by

() = w,
(5.18)
w= -M- 1Dto - M-1(P - B sin(O».

Here 0 is the angle of the rotor of the generator measured relative to a syn-
chronously spinning reference frame and its time derivative is to. Also M is
the moment of inertia of the generator and D its damping both in normalized
units; P is the exogenous power input to the generator from the turbine and B
the susceptance of the line connecting the generator to the rest of the network,
modeled as an infinite bus (see Figure 5.5) A choice of Lyapunov function is

1
v(O, w) = "2MW2 + PO + B cos(O).
The equilibrium point is 0 = sin -I (~), w = O. By translating the origin to this
equilibrium point it may be verified that v(O, w) - v(Oo , 0) is an l.p.d.f. around
it, Further it follows that

yielding the stability of the equilibrium point. As in the previous example, one
cannot conclude asymptotic stability of the equilibrium point from this analysis.
3. Damped Mathieu Equation
This equation models the dynamics of a pendulum with sinusoidally varying

y
La
I jB ----11 LO
Admittance
p Input ofbus
power

FIGURE 5.5. A generatorcoup1ed to an infinitebus

194 5. Lyapunov Stability Theory

length; this is a good model for instance of a child pumping a swing.

(5.19)
X2 = -X2 - (2 + sin rjx].
The total energy is again a candidate for a Lyapunov funct ion, namely,
2
v(x, t) = XI
2
+ 2+smt
X2
.'
It is an l.p.d.f, since
x2 x2
x2+ x2 > x2+ > x 2 + -.1..
2 + sin t - I
2
I 2 - I 3
Further, a small computation yields
. xi ( 4 + 2 sin t + cos 1) 0
v(x , t) = - (2 + sin r)? :5.

Thus, the equilibrium point at the origin is uniformly stable. However, we

cannot conc1ude that the origin is uniformly asymptotically stable.
4. System with a Limit Cycle
Consider the system

XI = -xz + XI (x~ + xi - I),

X2 = XI + xz(x~ + xi - 1).

A choice of v(x) = x~ + xi yields for ü(x) = 2(x~ + xi)(x~ + x i - 1), which

is the negative of an l.p.d.f. for {x : xf + xi < I}. Thus, 0 is a locally asymptot-
ically stable equilibrium (though it is not globally, asymptotically stable, since
there is a limit cyc1e of radius I , as may be verified by changing into polar
coordinates, see also Chapter 2).
Comments.
I. The theorems of Lyapunov give sufficient conditions for the stability of the
equilibrium point at the origin of (5.1). They do not, however, give a prescription
for determining the Lyapunov function v(x , t). Since the theorems are only
sufficient conditions for stability the search for a Lyapunov function establishing
stability of an equilibrium point could be long. However, it is a remarkable fact
that the converse of theorem (5.16) also exists: for example, if an equilibrium
point is stable then there exists an l.p.d.f. v(x, 1) with ü(x , t) :5 O. However,
the utility of this and other converse theorems 0/ Lyapunov is limited by the
fact that there is no general and computationally non-intensive technique for
generating these Lyapunov functions (an example of such a theorem, called
Zubov's theorem , is given in the Exercises (Problem 5.14) . We will not give the
details of the construction of converse Lyapunov functions in general, but we
focus on exponentially stable equilibria in the next subsection. However, the
 5.3 Basic StabilityTheoremsof Lyapunov 195

metbod of construction of tbe Lyapunov function in Section 5.3.4 is prototypical

of other converse theorems.
2. If tbe system has multiple equilibrium points, then by translating the equilibria
in turn to the origin one may individually ascertain their stability.

5.3.4 Exponential Stability Theorems

The basic theorem of Lyapunov, Theorem 5.16, stops short of giving explicit rates
of convergence of solutions to the equilibria. It may however be modified to do
so in the instance of exponentially stable equilibria. We will pay special attention
to exponentially stable equilibria, since they are robust to perturbation and are
consequently desirable from tbe viewpoint of applications. We will now state
necessary and sufficient conditions for the existence of an exponentially stable
equilibrium point, that is, we state a converse theorem for exponentially stable
systems:
Theorem 5.17 Exponential Stability Theorem and Its Converse. Assume
that f(x, t) : R+ X Rn ~ Rn has continuous first partial derivatives in x and is
piecewise continuous in t , Then the two statements below are equivalent:

1. x = 0 is a locally exponentially stable equilibrium point of

x= f(x , t) ,

i.e., if x E Bh for h small enough, there exist m , a > 0 such that

14>('r,x,t)1 ::::me-a(r-tl.

2. There exists a function v(x, t) and some constants h, a" a 2, a3, a4 > 0 such
thatfor all x E Rh, t ~ 0

a,lxl 2 :::: v(x , t) :::: a21x1 2,

dv(x , t) I :::: -a3lxl ,
2
dt (5 .l)
(5.20)

Iav~:, t) I s a4lxl ·

Proof:
(I) ::} (2)
Wc prove the three inequalities of (5.20) in turn, starting from the definition of
v(t,x) :

• Denote by 4>(r, x , t) the solution of (5.1) at time r starting from x at time t,

r
and define

v(x , t):= I 14>(r, x, t)12dr, (5.21)

196 5. Lyapunov Stability Theory

where T will be defined later. From the exponential stability of the system at
rate a and the lower bound on the rate of growth given by Proposition 5.3, we
have
mlxle-a(r-I) ::: It/J(r, x, t)1 ::: Ixle-J(r-I) (5.22)

for x E Rh for some h. Also 1the Lipschitz constant of f (x, t) exists because of
the assumption that f (x , t) has continuous first partial derivatives with respect
to x. This, when used in (5.21), yields the first inequality of (5.20) for x E Rh'
(where h' is chosen to be hlm) with
(I - e- 2/T) 2
(I - e- 2aT)
al := a2 := m (5.23)
21 2a
• Differentiating (5.21) with respect to t yields

dv(x, t) = It/J(t + T, x, t)12 - It/J(t, x, t)12

dt
t+ T d (5.24)
2)dr.
+[ I dt (It/J(r, x(t), t)1

Note that d jdt is the derivative with respect to the initial time t along the
trajectories of (5.1). However, since for all bot the solution satisfies

t/J(T, x(t + bot), t + bot) = t/J(r, x(t), t),

we have that that ;1r-(It/J(T, x(t), t)J2) == O. Using the fact that t/J(t, x, t) = x
and the exponential bound on the solution, we have that

dV~, t) < -(I _ m2e-2aT)lxI2.

t -
The second inequality of (5.20) now folIows, provided that T > (lla) Inm
and

• Differentiating (5.21) with respect to Xj , we have

8v(x,t) _2[I+T f--.A. .( )8t/Jj(r,x,t)d

- L...J'I'J T ,x,t r. (5.25)
8x, I j=1 8Xi

By way of notation define

. 8(jJj(T,x,t)
Qij (T, x, t ) .= --''-----
8xj
and
..( ) ._ 8fi(t, x)
A IJ x, t .- .
8xj
 5.3 Basic Stability Theorems of Lyapunov 197

Interchanging the order of differentiation by r , with differentiation by Xj yields

that
d
- Q (. , x, t) = A(,p(., x, t) , t) .Q(., x, t) . (5.26)
d.
Thus Q(., x, 1) is the state transition matrix associated with the matrix
A(,p(., x, r) , t). By the assumption on boundedness of the partials of f with
respect to x, it follows that IA(·, ')1 s k for some k, so that
IQ(.,x,t)1 s ek(r-t).

using this and the bound for exponential convergence in (5.25) yields

Iav~~ x) I s 2 ft+T mlxle(k-a)(r-tld.,

which is the last equation of (5.20) if we definc
2m(e(k-a)T - I)
a4 :=
(k - a)

This completes the proof, but note that v(x, t) is only defined for XE Bh' with
h' = h/ m, to guarantee that ,p(r, x, t) E B h for all r ~ t (convince yourself of
this point).

(2) => (I)

This direction is straightforward, as may be vcrified by noting that equation (5.20)
implies that
. a3 (
v(x , r) S - - v x, t). (5.27)
a2
This in turn implies that

v(t,x(t» S v(to,x(to»e -*(t-to). (5.28)

Using the lower bound for v(t, x(t» and the upper bound for v(to, x (to» we get

al!x(t)1 2 S a2Ix(to)12e-*(t-to). (5.29)

Using the estimate of (5.29) it follows that

Ix(t) 1 S m Ix(to) le-a(t -t o) .

with

D
198 5. LyapunovStability Theory

5.4 LaSalle's Invariance Principle

LaSalle 's invariance principle has two main applications:
1. It enables one to conclude asymptotic stability even when -v(x, t) is not an
I.p.d.f.
2. It enables one to prove that trajectories of the differential equation starting in a
given region converge to one of many equilibrium points in that region.
However, the principle applies primarily to autonomous or periodic systems, which
are discussed in this section . Same generalizations to the non-autonomous case
are also discussed in the next section. We will deal with the autonomous case first.
We recall some definitions from Chapter 2:
Definition 5.18 ca limit set. A set S C IRn is the co limit set of a trajectory
r/>(', xo, to) iffor every y E S, there exists a sequence of times tn -+ 00 such that
r/>(tn, Xo, to) -+ y.
Definition 5.19 Invariant set , A set M C IRn is said to be an invariant set if
whenever y E M and to 2: 0, we have
r/>(t, y, to) E M Vt 2: to.
The following propositions establish some properties of co limit sets and invariant
sets.
Proposition 5.20. Ifr/>(·, xo, to) is a bounded trajectory, its w limit set is compact.
Further, r/> (r, xo, to) approaches its w limit set as t -+ 00.
Proof: As in Chapter 2, see pages 46-50 of [329]. The proofis no different in IRn .
Proposition 5.21. Assume that the system (5.1) is autonomous and let S be the w-
limit set ofany trajectory. Then S is invariant.
Proof: Let y E Sand t) 2: 0 be arbitrary. We need to show that r/>(t, y, t1) E S
for all t 2: tl . Now y ES=} 3tn -+ 00 such that r/> (tn, Xo, to) -+ Y as n -+ 00 .
Since trajectories are continuous in initial conditions, it follows that

= lim r/>(t
n-->oo
+ tn - t1 , XO, to),

since the system is autonomous. Now, tn -+ 00 as n -+ 00 so that by Proposition

5.20 the right hand side converges to an element of S. 0

Proposition 5.22 LaSalle's Principle. Let v : IRn -+ IR be continuously

differentiable and suppose that
Qe = (x E IR n : v(x) S c}

is bounded and that v S 0 for all x E Q e' Define S C Qe by

S = (x E Qe : v(x) = O}
 5.4 LaSalle's Invariance Principle 199

and let M be the largest invariant set in S. Then, whenever Xo E Oe> l/J(t, XO , 0)
approaches M as t -+ 00.
Proof: Let Xo E Oe' Now, since v(l/J(t. Xo. 0» is a nonincreasing function of
time we see that l/J(t. Xo. 0) E Oe Vt. Further, since Oe is bounded v(l/J(t. Xo. 0»
is also bounded below. Let
Co = lim v(l/J(t, Xo. 0»
t.....00

and let L be the to limit set of the trajectory. Then, v(y) = Co for y E L. Since L
is invariant we have that iI(y) = OVy E L so that L c S. Since, M is the largest
invariant set inside S, we have that L c M . Since set.
Xo. 0) approaches L as
t -+ 00, we have that s(t. Xo. 0) approaches M as t -+ 00. o.
Theorem 5.23 LaSalle's Principle to Establish Asymptotic Stabllity, Let v :
]Rn ~ ]R be such that on Oe = {x E ]Rn : v(x) ::s c}, a compact set we have
iI(x) ::s O. As in the previous proposition define
s= {x E Oe: iI(x) = O}.
Then, if S contains no trajectories other than x = 0 then 0 is asymptotically stable.

Proof: follows direct1y from the preceding lemma .

An application of LaSalle 's principle is to prove global asymptotic stability is
as folIows:
Theorem 5.24 Application ofLaSalle's Principle to prove Global Asymptotic
Siabiiiiy. Let v(x) : jRn ~ jR be a [Link]. und v(x) ::: 0 [or all x E !Rn. Also , let
the set
S = {x E ]Rn : iI(x) = O}
contain no nontrivial trajectories. Then 0 is globally, asymptvtically stahle.
Examples:
1. Spring-mass system with damper
This system is described by

(5.30)
X2 = - f(X2) - g(xt>.
If f. g are locally passive. i.e.,

af(a)::: 0 Va E [-[Link].
then it may be verified that a suitable Lyapunov function (I.p.d.f.) is

V(XI. X2) = xi
2
+ r
Jo
l
g(a)da.

Further, it is easy to see that

iI(x" X2) = -x2!(X2) ::s 0 for X2 E [-ao, aol.
200 5. Lyapunov Stability Theory

Now choose

c = min(v( -ao, 0), v(ao, 0)).

Then V S 0 for x E Q c = ((XI , X2) : V(XI, X2) Sc}. As a consequence

of LaSalle's principle, the trajectory enters the largest invariant set in Q c n
((XI, X2) : V = O} = Q c n {XI, O}. To obtain the largest invariant set in this
region note that X2(1) == 0 => XI (I) == XIO => XI (I) = 0 = - /(0) - g(xlO)'
Consequently, we have that g(xlO) = 0 => XIO = O. Thus, the largest invariant
set inside Q c n {(XI, X2) : V = O} is the origin . Thus, the origin is locally
asymptotically stable.
The application ofLaSalle's principle shows that one can give interesting con-
ditions for the convergence of trajectories of the system of (5.30) even when
g(.) is not passive. It is easy to see that the arguments given above can be eas-
ily modified to obtain convergence results for the systern (5.30), provided that
J;' g(a)da is merely bounded below. The next examples are in this spirit.
2. Single generator coupled to an infinite bus
If () is the angle of the rotor of a generator with respect to a synchronously
rotating frame of reference and w = 0, then the equations of the dynamics of
the synchronous generator are given by

Mii + DO + B sin () = Pm - Pe' (5.31)

Here, M, D stand for the moment of inertia and damping of the generator rotor ;
Pm stands for the exogenous power input , and P, stands for the local power
demand. This equation can be written in state space form with XI = (), X2 = w.
The equilibrium points of the system are at () = sin "" (Pm; Pe), W = O. Note
that the equilibria repeat every 2rr radians. Consider a choice of

Mw 2
v(() , w) = -2- + (Pe - Pm)() - B cos(()) (5.32)

with

Since v((), w) is not bounded below, Q c = {CO , w) : v((), w) S c} is not

compact. However, if we know apriori that a trajectory is bounded, then we
know from LaSalle 's principle that it converges to the largest invariant set in
a compact region with w = O. Reasoning exactly as in the previous example
guarantees that the trajectory converges to one of the equilibrium points of
the system. (These are the only invariant sets in the region where v = 0.)
This behavior is referred to by power systems engineers as "trajectories that
skip only finitely many cycles (2rr multiples in the () variable) converge to an
equilibrium" owing to the periodic dependence of the dynamics on (). However,
ifwe view the dynamics as resident on the state space SI x IR I , we can no longer
use the function v((), w) to conclude convergence of the trajectories. (Why? See
Problem 5.11 for details on this point.)
 5.4 LaSaIle's Invariance Principle 201

FIGURE 5.6. An interconnected power system

3. Power system swing equations

The dynamics of a power system modeled as shown in Figure 5.6 as a collection
of generators interconnected by transmission lines of susceptance Bij between
generators i and j is a generalization ofthe swing equation for a single generator
coupled to an infinite bus :
n
M;Ö; + D/}; + L Bij sin(O; - Oj) = Pm; - Pa (5.33)
j=I.j#.;

for i = I, . .. , n. Here M;, D; stand for the moment of inertia and damping
of the ith generator. and Ps«, Pe; stand for the mechanical power input and
electrical power output from the i -th generator. The total energy function for
this system is given by

I' T ~
v(O, w) = 20 MO - . L:- Bij cos(O; - Oj) + 0 (P, - Pm) ,
• T
(5.34)
l < j .j=1

with its derivative given by

. 'T '
V = -0 DO.
As before the function v( 0, w) is not bounded below so that we cannot guarantee
that Qc is compact for any c. However, we may establish as before that all a
priori bounded trajectories converge. The same comments about convergence
of finite-cycle-skipping trajectories 2rr variations in one or more of the 0; hold
here as weIl. Also v(O, w) is no longer a valid function to consider when the
dynamics of the power system are considered on the state space T" x lRn (see
Problem 5.11 for details).
4. Hopfield Neural Networks [1431
Figure 5.7 shows a circuit diagram of a so-called Hopfield network consisting of
N operational amplifiers interconnected by an Re network. The operational am-
plifier input voltages are labeled u, and the outputs x., The relationship between
202 5. Lyapunov Stability Theory

Neuron I

Neuron i
Tilx J
T i2 x 2 -:-I~2.}---.-----r---'-~
Ti NxN

FIGURE 5.7. A Hopfield network

g ( U)

-------::>i'~-----_U

FIGURE 5.8. A sigmoidal nonlinearity

u, and x, for each i is given by a single , so-called sigmoidal nonlinearity: i.e., a

monotone increasing odd-symmetric nonlinearity of the form shown in Figure
5.8 asymptotic to ± I as its argument tends to ±oo, respectively. Examples of
sigmoidal functions g(u) are
2 I A1fU
g(u) = - tan" - ,
1f 2
(5.35)
eAU _ e- AU
g(u) = eAU + e- AU •

The equations of the Hopfield network are given by

du ; N
Ci d = LT;jxj - c,», + t.,
t j=1 (5.36)
x, =g(Ui).
Here Ci > 0 is the capacitance of the i th capacitor, and G i » 0 the conductance
of the i -th amplifier, Ti) stands for the strength of the interconnection between the
Xj and the i-th capacitor. I, stands for the current injection at the i-th node. Using
 5.4 LaSalle's Invariance Principle 203

the relationship between u, and x., we get the equations of the Hopfield network:

Xi = hi(x;) [tT;jXj - G;g-I(X;) + l iJ, (5.37)

}=I

where

I -
h;(x) = - dg I > 0 Vx.
C; du u=g -J(x)

Consider the funetion

1
v(x) = --xTTx+ LG;
N lXig-I(y)dy- LI;x;. N
(5.38)
2 ;= 1 0 ;= 1

It is now easy to verify that equation (5.37) may be rewritten as

. = -h;(x;)-
X;
av . (5.39)
Bx,
Thus, it follows that the Hopfield network is a gradient flow of v (along the "metrie"
given by the h, (.)) and

VI(5 .37) = - Lh;(X;)

N (av)2
-a. ~ O.
;=1 X,

Further, it follows that

v= 0 => dv = 0 Vi
ax ;
=> X= o.
Thus, to app1y LaSalle's principle to eonclude that trajeetories eonverge to one of
the equilibrium points, it is enough to find a region which is either invariant or on
whieh v is bounded. Consider the set
Q . = {x ERn: Ix;! ~ 1-E}.
We will prove that the region Q. is invariant for E small enough. Note that

d;/ = 2x;h;(x;) (t } =I
T;jxj - g-I (x;) + I;) .

Sinee g -I (x;) ~ ±oo as X; ~ ± 1, it follows that for E > 0 small enough,

dx~
_I < 0 for 1- E ~ Ix; I < 1,
dt
thus establishing that Q . is invariant. In turn, this implies that v(x) is bounded
below on Q., so that initial eonditions beginning inside Q. tend to the equilibrium
points inside Q e - Thus, for symmetrie intereonneetions T;j the Hopfield network
does not oseillate, and all trajeetories eonverge to one of the equilibria. It is of great
interest to estimate the number of equilibria of a Hopfield network (see [143] and
a very large literature on neural networks for more details).
204 5. Lyapunov Stability Theory

There is a globalization of LaSalle's principle, which is as folIows:

Theorem 5.25 Global LaSalle's Principle. Consider the system of (5.1). Let
v(x) be a p.d.j. with il ::5 OVx ERn . !fthe set
s = {x E ]Rn , ilex) = O}
contains no invariant sets other than the origin , the origin is globally asymptoti-
cally stable.
There is also aversion of LaSalle's theorem that holds for periodic systems as
weIl.
Theorem 5.26 LaSalle's Principle for Periodic Systems. Assume that the
system of (5.1) is periodic, i.e.,
fex, t) = fex, t + T), Vt Vx E ]Rn .

Further, let v(x, t) be a p.d.j. which is periodic in t also with period T. Define
S = {x E ]Rn : ilex, t) = 0, Vt :::: O}
Then if il(x, t) ::5 0 Vt 2: 0, Vx E ]Rn and the largest invariant set in S is the
origin, then the origin is globally (uniformly) asymptotically stable.

5.5 Generalizations of LaSalle's Principle

LaSalle's invariance principle is restricted in applications because it holds only
for time-invariant and periodic systems. For extending the result to arbitrary time-
varying systems, two difficulties arise:
1. {x : il(x, t) = O} may be a time-varying set.
2. Tbe w limit set of a trajectory is itself not invariant.
However, if we have the hypothesis that
il(x,1) ::5 -w(x) ::5 0,
then the set S may be defined to be
{x : w(x) = O},
and we may state the following gencralization of LaSalle 's theorem:
Theorem 5.27 Generalization ofLaSalle's Theorem. Assume that the vector
field f t», t) of(5.1} is locally Lipschitz continuous in x, uniformly in t, in a ball
ofradius r. Let v(x , t) satisfyfor functions eYl, eY2 ofclass K
eYl (lxI) ::5 v(x, t) ::5 eY2(lxl) . (5 .40)
Further,for some non-negative function w (x) , assume that
.
v(x, t) = at + av
av
ax ft» , t) ::5 -w(x) ::5 O. (5.41 )
 5.5 Generalizationsof LaSalle's Principle 205

Thenfor alllx(to)1 ::::: a il (al (r)), the trajectories x(·) are bounded and

lim w(x(t)) = O. (5.42)

t .....00

Proof: The proof of this theorem needs a fact from analysis called Barbalat's
lemma (which we have left as an exercise: Problem 5.9; see also [259]), which
states that if ifJ (-) : IR f-+ IR is a uniformly continuous integrable function with

1 00

ifJ(t) < 00,

then

lim ifJ(t) = O.
t ..... 00

The requirement ofuniform continuity of ifJ(·) is necessary for this lemma, as easy
counterexamples will show. We will use this lemma in what folIows.
First, note that a simple contradiction argument shows that for any p < r ,

Ix(to)1 ::::: ai' (a, (p)) =} Ix(t)1 ::::: p V t ~ to·

Thus /x(t)/ < r for all t ~ to, so that v(x(t), t) is monotone decreasing. This
yields that

1,~)
w(x(r))dr ::::: -
['0
t
v(x(r), r)dr

= v(x(to) , to) - v(x(t) , t) .

(5.43)

Since v(x, t) is bounded below by 0, it follows that

1'0
00 w(x(r))dr < 00.

By the continuity of f (x, t) (Lipschitz in x, uniformly in t) and the boundedness

of x(t) , it follows that x(t) is uniformly continuous, and so is w(x(t) . Using
Barbalat's lemma it follows that

lim w(x(t)) = O. o
t ..... 00

Remarks: The preceding theorem implies that x(t) approaches a set E defined
by

E := (x E B, : w(x) = O}.

However, the set E is not guaranteed to be an invariant set; so that one cannot, for
example, assert that x(t) tends to the largest invariant set inside E. It is difficult ,
in general, to show that the set E is invariant. However, this can be shown to be
the case when f (x , t) is autonomous, T-periodic, or asymptotically autonomous.
Another generalization of LaSalle 's theorem is given in the exercises (Problem
5.10).
206 5. Lyapunov Stability Theory

5.6 Instability Theorems

Lyapunov's theorem presented in the previous section gives a sufficient condition
for establishing the stability of an equilibrium point. In this section we will give
some sufficient conditions for establishing the instability of an equilibrium point.
Definition 5.28 Unstable Equilibrium. The equilibrium po int 0 is unstable at
to. if it is not stable.
This definition is more subtle than it seems: we may parse the definition, by
systematically negating the definition of stability: There exists an E > 0, such that
for all 8 balls of initial conditions (no matter how small the ball), there exists at
least one initial condition, such that the trajectory is not confined to the E ball; that
is to say,
V8 3xo E Bo
such that 3t o with

IX IS I ~ E.
Instability is of necessity a local concept. One seldom has adefinition for uniform
instability. Note that the definition of instability does not require every initial
condition starting arbitrarily near the origin to be expelied from a neighborhood
of the origin , it just requires one from each arbitrarily small neighborhood of the
origin to be expelled away. As we will see in the context of linear time-invariant
systems in the next section, the linear time-invariant system
x= Ax
is unstable, if just one eigenvalue of A lies in C~! The instability theorems have
the same flavor: They insist on iJ being an l.p.d.f. so as to have a mechanism for
the increase of v. However, since we do not need to guarantee that every initial
condition elose to the origin is repelled from the origin, we do not need to assurne
that v is an l.p.d.f.
We state and prove two examples of instability theorems:
Theorem 5.29 Instability Theorem. The equilibrium point 0 is unstable at time
to if there exists a decrescent function v : lRn x lR+ t-+ lR such that
1. iJ(x, t) is an [Link].
2. v(O, t) = 0 and there exist po ints x arbitrarily close to 0 such that v(x, to) > O.
Proof: We are given that there exists a function v(x, t) such that
v(x, t) s ß(lxl) xE e.,
iJ{x, t) ~ a(lxl) x E Bs •
We need to show that for some E > 0, there is no 8 such that

IXol < 8 => Ix(t)1 < E Vt ~ to.

 5.7 Stabilityof Linear Time-VaryingSystems 207

Now choose E = min(r, s) . Given 8 > 0 choose xs with IXol < 8 and v(xo , to) > O.
Such a choice is possible by the hypothesis on v(x , to). So long as 4J(t, xo, to) lies
in B, we have v (x (t ) , t) 2: 0, which shows that

v(x(t), t) 2: v(xo, to) > 0

This implies that Ix(t)1 is bounded away from O. Thus v(x(t), r) is bounded away
fromzero. Thus, v(x(t), t) will exceed äts) in finite time. In turn this will guarantee
that Ix(t) I will exceed E in finite time. 0

Theorem 5.30 Chetaev's theorem. The equilibrium point 0 is unstable at time

to if there is a decrescent function v : IR+ X IRn 1-+ IR such that

1. v(x , t) = AV(X, t) + VI (x, r), where A > 0 and VI (x, t) 2: 0 Vt 2: 0 Vx E Br •

2. v(O, t) = 0 and there exist points x arbitrarily close to 0 such that v(x , to) > O.

Proof: Choose E = r and given 8 > 0 pick Xo such that Ixo I < 8 and V(xo, to) >
O. When Ix(t)1 ~ r , we have
v(x, t) = AV(X, t) + VI (x, t) 2: AV(X, r).

If we multiply the inequality above by integrating factor e:", it follows that

dv(x , t)e-At
---->0.
dt -
Integrating this inequality from to to t yields
v(x(t), t) 2: eA(t-tO)v(xo, to).

Thus , v(x(t), t) grows without bound . Since v(x, t) is decrescent,

v(x, t) 2: ß(lxl)

for some function of dass K, so that for some ts. v(x(t), t) > ß(E) , establishing
that IX(t8) I > E . 0

5.7 Stability of Linear Time-Varying Systems

Stability theory for linear time varying systems is very easy and does not really
need the application of the theorems of Lyapunov. However, Lyapunov theory
provides valuable help in understanding the construction of Lyapunov functions
for non linear systems. Thus, in this section we will consider linear time varying
systems of the form

x = A(t)x, x(to) = xo. (5.44)

where A (t) E IRn xn is a piecewise continuous bounded function . As a consequence
the system of (5.44) satisfies the conditions for the existence and uniqueness of
solutions.
208 5. Lyapunov Stability Theory

Definition 5.31 State Transition Matrix. The state transition matrix cJ>(t, to) E
IR n x n associated with A(t) is by definition the unique solution of the matrix
differential equation
$(t, to) = A(t)cJ>(t, to), cJ>(to, to) = I. (5.45)
The flow of (5.44) ean be expressed in terms of the solutions to (5.45) as
x(t) = cJ>(t, to)x(to) . (5.46)
In partieular, this expression shows that the trajeetories are proportional to the
size of the initial eonditions, so that loeal and global properties are identieal. Two
important properties of the state transition matrix are as folIows:
1. Group Property

cJ>(t, to) = cJ>(t, r)cJ>(r , to) V t , r. to (5.47)

2. Inverse
cJ>(t, tO) -1 = cJ>(to, t) Vt , to. (5.48)
Reealling the formula for differentiating the inverse of a matrix :

we have
d [ (cJ>(to,1))- I]
d cJ>(t, to) = -d
-d
to to
= -cJ>(to , t)-I A(to)cJ>(to, t)cJ>(to , t)-I
= -cJ>(t, to)A(to) .

The following theorems are elementary eharaeterizations of stability, uniform

stability, and asymptotic stability of linear systems. Uniform asymptotic stability
is eharaeterized in the next theorem. Reeall that for linear systems loeal and global
stability notions are identieal.

Theorem 5.32 Stability ofLinear Systems. The right-hand-side ofthe follow-

ing table gives the stability conclusions of the equilibrium point 0 of the linear
time-varying system (5.44).

Conditions on <I> (r, (0) Conclusions

SUPt",'o 1<1>(1, (0)1 := m(10) < 00 Stable at 10

SUP 'o"'OSUP' ''''o 1<1>(1 , (0)1 < 00 Unifonnly Stable

limH x 1<1>(1, (0)1 = 0 Asyrnptotically stable

 5.7 Stability of Linear Time-Varying Systems 209

Theorem 5.33 Exponential and Uniform Asymptotic Stability. The equilib-

rium point 0 0/ (5.44) is unifonnly asymptotically stable if it is uniformly stable
and 1<I>(t , to)1 ~ 0 as t ~ 00 , uniformly in to. The point x = 0 is a uniform
asymptotically stable equilibrium point 0/ (5.44) iff x = 0 is an exponentially
stable equilibrium point 0/ (5.44).
Proof: The first part of the theorem is a statement of the definition of uniform
asymptotic stability for linear time varying systems. For the second half, the fact
that exponential stability implies uniform asymptotic stability follows from the
definition. For the converse implication let us assume that the equilibrium is uni-
fonnly asymptotically stable, From the first part of the theorem, we have that
uniform stability implies that Vt), there exist mo, T such that

1<I>(t,t)I:5mo, i e «.
Uniform convergence of <I> (t, to to 0 implies that
1
1<I>(t , t)1 :5 2' Vt 2: t) + 1'.

Now given any t, to pick k such that

to + k'I' :5 t :5 to + (k + 1)1'.
Since

<I>(t, to) = <I>(t, to + kT) n k

j= l
<I>(to + jT, to + jT - T),

it follows that

1<I>(t, to)1 :5 1<I>(t, to + k'I"; n k

j=1
1<I>(to + j'I', to + jT - 1')1

:5 moT k
:5 2m 02-(l-to)/ T
:5 me- ).(I-to)

where m = 2mo and J... = log2/1'. o

5.7.1 Autonomous Linear Systems

The results stated above can be specialized to linear time- invariant systems of the
form

x= Ax (5.49)

with x E !Rn .
210 5. Lyapunov Stability Theory

Theorem 5.34 Stability ofLinear Time-Invariant Systems.

1. The equilibrium at the origin of(5.49) is stable iffall the eigenvalues of Aare
in C_ and those on the jto axis are simple zeros ofthe minimal polynomial of
A.
2. The equilibrium at the origin of (5.49) is asymptotieally stable iff all the
eigenvalues lie in C~ .
An alternative way of studying the stability of (5.49) is by using the Lyapunov
funetion with symmetrie, positive definite P E JRnxn ,

v(x) = x T Px,
An easy ealculation yields that
iI(x) = xT(A T P + PA)x. (5.50)
If there exists a symmetrie, positive definite Q E JRn xn sueh that
ATp +PA = -Q , (5.51)
then we see that -il(x) is a p.d.f. Equation (5.51), whieh is a linear equation of
the form

ar: = Q
nxn nxn
where E is a map from IR f--+ IR referred to as aLyapunov equation. Given a
symmetrie Q E IR n x n , it may be shown that (5.51) has a unique symmetrie solution
P E JRn xn when E is invertible , that is, iff all the n 2 eigenvalues of the linear map
E are different from 0:
'A i + 'Aj =1= 0 V 'A i, 'A j E a(A) .

Here a(A) refers to the speetrum, or the set of eigenvalues of A.

Claim 5.35. lf A has all its eigenvalues in C~ , then the solution of(5.51) is given
by

Proof: Define

Set) = 11 eATrQeArdT,

Using a ehange of variables, we may rewrite this as

Set) = 11 eAT(t-r)QeA(I-r)dT.

Differentiating this expression using the Leibnitz rule yields

S(t)=ATS+SA+Q .
 5.7 Stability of Linear Time-Varying Systems 211

Also, since A has eigenvalues in C e:. the eigenvalues of the linear operator

S~ ATS+SA

Iie in C~ as weil, so that S ~ 0 as t ~ 00, whence P = S( 00) exists, and satisfies

AT S(oo) + S(oo)A + Q = O.
Hence the P ofthe claim satisfies equation (5.51). o
Theorem 5.36 Lyapunov Lemma. For A E IRn xn the following three
statements are equivalent:
1. The eigenvalues ofA are in Ce:. .
2. There exists asymmetrie Q > 0 sueh that the unique symmetrie solution to
(5.51 )satisfies P > O.
3. For all symmetrie Q > 0 (5.51) the unique symmetrie solution satisfies P > O.
4. There exists C E ]Rmxn (with m arbitrary) with the pair A , C observable.?
there exists a unique symmetrie solution P > 0 ofthe Lyapunov equation with
Q = CT[Link] 0,
(5.52)
5. For all C E ]Rmxn (with m arb itrary), sueh that the pair A, C is observable,
there exists a unique solution P > 00f(5.52).
Proof: (iii) => (ii) and (v) => (iv) are obvious.
(ii) => (i) Choose v(x) = x T Px a p.d.f. and note that -il(x) = x T Qx is also a
p.d.f,
(iv) => (i) Choose v(x) = x T Px a p.d .f. and note that -il(x) = xTCTCx . By
LaSalle's principle, the trajectories starting from arbitrary initial conditions con -
verge to the largest invariant set in the null space of C, where iI(x) O. By the =
assumption of observability of the pair A, C it follows that the largest invariant set
inside the null space of C is the origin.
(i) => (iii) By the preceding claim it follows that the unique solution to (5.51) is

P = 1 00

e
ATI
QeAldt.

That P ::: 0 is obvious. It only needs to be shown that P > O. Indeed, if x T P x =0

for some x E ]Rn then it follows with Q = MT M that

0= x T Px = 1 00

x Te
ATI
MT MeAlxdt = 100

IMe Alxl 2dt

Thus we have that M e AIX == O. In particular, Mx = O. But M is nonsingular since

Q is. Hence x = 0, establishing the positive definiteness of P.

2Recall from linear systems theory that a pair A. C is said to be observable itTthe largest
A-invariant space in the null space of Cis the origin.
212 5. Lyapunov Stabi1ity Theory

(i) :::} (v) By the preceding claim it follows that the unique solution to (5.51) is

P = 100 eATICTCeAldt.

That P ::: 0 is obvious. It only needs to be shown that P > [Link], if x T P X =0

for some x E lRn then it follows that C e A1x = O. In particular, by differentiating
Ce A1x n-times at the origin, we get Cx = C Ax = ... = C A n - 1x = O. Since the
pair A, C are observable, it follows that x = 0 establishing the positive definiteness
of P. 0
The following generalization ofthe Lyapunov lemma, called the Taussky lemma,
is useful when a(A) </. Ce:. We drop the proof since it is not central to the
developments of this chapter (see [296]).
Lemma 5.37 Taussky Lemma. For A E lRnxn and given Q E !Rnxn positive
definite, ifa(A) n {jw : ta E ]-00, oo[} = '" the unique symmetrie solution P to
the Lyapunov equation (5.51)

has as many positive eigenvalues as the number 0/ eigenvalues 0/ A in Ce: and as

many negative eigenvalues as the number 0/ eigenvalues 0/ A in C~.

5.7.2 Quadratic Lyapunov Functions for Linear Time Varying

Systems
The time varying counterpart to the claim of (5.35) may be derived after a few
preliminary results.
Claim 5.38. Consider the system 0/ (5.44) with uniformly asymptotieally stable
equilibrium 0:
x= A(t)x, x(to) = xo.
Further, assume that Q(.) is a eontinuous, boundedfunetion. Then for t ::: 0 the
matrix

P(t) = /00 cl>T (r, t)Q(r)<I>(r, tyd : (5.53)

is well-defined. Further. P(t) is bounded.

Proof: Uniform asymptotic stability of (5.44) guarantees exponential stability
so that
1<I>(r, t)1 s me-J..(r-I) Vr ::: t,

This estimate may be used to show that P(t) as defined in (5.53) is weIl defined
and bounded. 0

Claim 5.39. I/Q(t) > 0 V t::: Oandfurther,

axTx~xTQ(t)x, VXE!Rn, t:::O,
 5.7 Stability of Linear Time-Varying Systems 213

and A(t) is bounded, say by m , then PU) of(5.53) is uniformly positive definite
for t ~ 0, i.e., there exists ß > 0 such that
ßx Tx :::: x T P(t)x.
Proof:

x T P(t)x = f oo x T<l>T(T, t)Q(T)<I>(T , t)x d t

~ a f oo I<I> (T, t)xl 2dT

~ a foo x Txe- 2k(t-r )dT.
Here the last inequality follows from the Bellman Gronwalllemma which gives a
lower bound on I<I>(T, t)xl from the norm bound 1A(t)1 :::: k, as
I<I>(T, t)xl ~ e-k(r-t)lxI.

From the last inequality, it also follows that

T a T
x P(t)x ~ 2k x x,

establishing the claim. D

Theorem 5.40 Time-Varying Lyapunov Lemma. Assume that A (.) is bounded.

lffor some Q(t)~ a I , pet) as defined by (5.53) is bounded, then the origin is the
uniformly asymptotically stable equilibrium point of (5.44).
Proof: By the preceding claim, pet) satisfying (5.53) is bounded below, so that
2 T 2
Y lxl ~ x P(t)x ~ ß lx l
whence x T P (t)x is a decrescent p.d.f. Further, it is easy to see that
v(x, t) = x T (ht) + AT (t)P(t) + PU)A(t»x
= _x T Q(t)x
2
:::: -alxI •

This establishes exponential stability of the origin and as a consequence uniform

asymptotic stability. D

Uniform Complete Observability and Stability

Aversion of the time -varying Lyapunov lemma for system s that are uniformly
completely observable mayaIso be stated and proved (see Exercise 5.12 for the
definition of uniform complete observability and the statement of the lemma). The
key philosophical content of this result and the corresponding time invariant one
stated extremely loosely is that iffor some Lyapunov function candidate v(x, t),
the nonlinear system
x= f(x ,t)
214 5. Lyapunov Stability Theory

with output map

y(t) = h(x , t) := - + - f(x, t)

av av
at ax
is "locally uniformly observable" near the origin, then the origin is uniformly
asymptotically stable. The reason this statement is loose is that the eoneept of
"local uniform observability" needs to be made precise. This is straightforward
for the linear time varying ease that we are presently diseussing, as diseussed in
Problem 5.12.

5.8 The Indirect Method of Lyapunov

Consider an autonomous version ofthe system of (5.1), namely,
x= f(x)

*
with
Ix=o
f (0) = O. so that 0 is an equilibrium point of the system. Define A =
E jRn xn to be the Jaeobian matrix of f(x). The system

i; = Az
is referred to as the linearization of the system (5.1) around the equilibrium point
O. For non-autonomous systems, the development is similar: Consider
x= f(x,t)
with f(O, t) =0 for all t ~ o. With
A(t ) = af(x, t) I
ax x=O
it follows that the remainder
f1 (x , t) = f(x , t) - A(t)x
is o(lxl) for eaeh fixed t ~ o. It may not, however, be true that
·
I1m sup
If, (x, t)1 =
0
. (5.54)
[x] ..... O t ~O [x]
that is f(x, t) is a funetion for which the remainder is not uniformly of order o(lx 1).
For a sealar example of this eonsider the funetion
f(x , t) = -x + t x ".
In the instanee that the higher order terms are uniforrnly, o(lx 1), i.e., if (5.54) holds,
then the system of (5.55)
i; = A(t)z (5.55)
is referred to as the linearization of (5.1) about the origin . It is important to note
that analysis of the linearization of (5.55) yields conclusions about the nonlinear
system (5.1) when the uniform higher-order condition of(5.54) holds.
 5.8 The Indirect Method of Lyapunov 215

Theorem 5.41 Indirect Theorem of Lyapunov: Stability from Linearization.

Consider the system of(5.1) and let
·
Ittn sup
1ft (x, t)1
=
0
.
Ixl..... O I~O [x]
Further. assume that A(·) is bounded. 1f 0 is uniformly asymptotically stable
equilibrium point of (5.55), then it is a locally uniformly asymptotically stable
equilibrium point of(5.1).

Proof: Since A(·) is bounded and 0 is a uniformly asymptotically stable equi-

Iibrium point of (5.55), it follows from Claims 5.38, 5.39 that pet) defined
by

pet) = fOO <pT (r, t)<P(r, t)dr

is bounded above and below, that is, it satisfies
ßx Tx::: x T P(t)x ::: axTx

for some a , ß > O. Thus v(x, t) = x T P(t)x is a decrescent p.d.f. Also,

iI(x, t) = xT[J>(t) + AT (t)P(t) + P(t)A(t)]x + 2x T P(t)f, (x , t),

= _x Tx + 2x T P(t)ft(X, t) .

Since (5.54) holds, there exists r > 0 such that

I
If, (x, t)1 :5 3ß [x] Vx E e, i : 0,

and
21xl 2
12xT P(t)f) (x, t)1 :5 -3- Vx E s. ,
so that
xTx
iitx , t) :5 --3- Vx E s..
Thus -il(x, t) is an l.p.d.f, so that 0 is a locally (since -il is an l.p.d.f. only in Br )
uniformly asymptotically stable equilibrium point of (5.1). 0

Remarks:
1. The preceding theorem requires uniform asymptotic stability of the Iinearized
system to prove uniform asymptotic stability of the nonlinear system. Coun -
terexamples to the theorem exist if the Iinearized system is not uniformly
asymptotically stable.
2. The converse of this theorem is also true (see Problem 5.18).
3. If the Iinearization is time-invariant, then A(t) == A and a(A) C C~, then the
non linear system is uniformly asymptotically stable.
216 5. Lyapunov Stability Theory

4. This theorem provcs thatglobal uniform asymptotic stability ofthe linearization

implies loeal uniform asymptotic stability of the original nonlinear system. The
estimates of the theorem can be used to give a (conservative) bound on the
domain of attraction of the origin. For example, the largest level set of v (x, t)
that is unifonnly contained in B, is such abound and can be further estimated
using the bounds on P(t) as {x E lRn : lxi :::; ~} Also, the estim ates on r in
the preceding proof could be refined. Systematic techniques for estimating the
bounds on the regions of attraction of equilibrium points of nonlinear systems
is an important area ofresearch and involves searching for the 'best' Lyapunov
function s.
To prove the instability counterpart to the above theorem one needs to assume
that the linearized system is autonomous even though the original system is non-
autonomous.
Theorem 5.42 Instability from Linearization. Consider the nonlinear system
(5.1) and let the linearization satisfy the eonditions ofequation (5.54). Further, let
the linearization be time invariant, i.e. ,
öf(x, t)
AU) = öx Ix=o = A o.
lf A o ha s at least one eigenvalue in C~ , then the equilibrium 0 of the nonlinear
system is unstable.

Proof: Consider the Lyapunov equation

Al' P + PA o = 1

Assume, temporarily, that A o has no eigenvalues on the jta axis . Then by the
Taussky lemma the Lyapunov equation has a unique solution, and further more if
Ao has at least one eigenvalue in C~, then P has at least one positive eigenvalue.
Then v(x, t) = X T P x takes on positive values arbitrarily close to the origin and
is decrescent. Further,
T
iJ(x, t) = x (Al' P + PAo)x + 2x T Pfl(X, t)
= x T X + 2x T p I, (x, t).
Using the assumption of (5.54)it is easy to see that there exists r such that
xTx
iJ(x , t) ~ -3- vx E e.,
so that it is an l.p.d.f. Thus, the basic instability theorem yields that 0 is unstable.
In the instance that A o has some eigenvalues on the jt» axis in addition to at
least one in the open right half plane, the proof follows by continuity. 0

Remarks:
1. We have shown using Lyapunov 's basic theorems that if the the linearization
of a nonlinear system is time invariant then :
 5.9 Domainsof Attraction 217

• Having all eigenvalues in the open left half plane guarantees local uniform
asymptotic stability of the origin for the nonlinear system.
• If at least one of the eigenvalues of the linearization lies in the open right
half plane, then the origin is unstable .
The only situation not accounted for is the question of stability or instability
when the eigenvalues of the linearization lie in the closed left half plane and
include at least one on the jw axis. The study of stability in these cases is
delicate and relies on higher order terms than those of the linearization. This is
discussed in Chapter 7 as stability theorems on the center manifold.
2. The technique provides only loeal stability theorems . To show global stability
for equilibria of nonlinear systems there are no short cuts to the basic theorems
of Lyapunov.

5.9 Domains of Attraction

Once the loeal asymptotic stability of an equilibrium point has been established,
it is of interest to deterrnine the set of initial eonditions that eonverges to the
equilibrium point.
Definition 5.43 Domain of Attraction at 10. Consider the differential equation
(5.1) with equilibrium point Xo at 10. The domain of attraetion 0/ Xo at 10 is the set
0/ all initial conditions x at time 10, denoIe QJ (I, 10, x) satisfies
lim QJ(/, 10, x) = Xo.
1-+00

For autonomous systems,

x= /(x) (5.56)

the domain of attraetion of the equilibrium point Xo (assumed to be 0, without loss

of genera!ity) is a set that is independent of the initial time , since the ftow only
depends on the time differenee I - 10. For this case, we use the notation of Chapter
2, namely QJI(X), to mean the state at time I starting from x at time O. We may
charaeterize this set as follows:

Q := {x: !im QJI(X) =

1-+00
O} .
Proposition 5.44 Topological Properties of the Domain of Attraction. LeI
o be an asymptotically stable equilibrium point 0/ (5.56) . Then , the domain 0/
attraction, Q % is an open, invariant set. Moreover, the boundary o/Q is invarianl
as weil.
Proof: Let x E Q . For the invarianee of Q we need to show that QJ" (x) E Q for
all s . To this end , note that

QJI(QJ,,(X)) = QJI+., (x),

218 5. LyapunovStability Theory

so that for all s, we have

lim ~t(~.f(x» = 0,
t-+oo

establishing that ~.f (x) E n for all s.

To show that n is open, we first note that a neighborhood of the origin is
contained in n: Indeed, this follows from the definition of asymptotic stability.
Let us denote this by Ba, a ball of radius a. Now let x E n since ~t(x) ---+ 0 as
t ---+ 00, it follows that there exists T such that l ~t (x ) 1 < a/2 for t ~ T. From
the fact that solutions of the differential equation (5.56) depend continuously on
initial conditions on compact time intervals, it follows that there exists I) > 0 such
that for Iy - x] < 1), we have that I~T(Y) - ~T(x)1 < a/2. In turn, this implies
that ~r<y) E Ba, so that lim H oo ~t(Y) = O. Hence a ball of radius I) around x is
contained in n, and this proves that n is open.
To show that the boundary of n, labeled an is invariant, let x E an. Then
X n ---+ x with X n E n. Given any time t, then there exists a sequence X n such that
~t(xn) ---+ ~t(x) . Now, ~t(xn) E n, since n is invariant. Further, ~t(x) tI- n since
x E an. Hence, ~t(x) E an for all t. Hence , an is invariant. 0

The proof ofthe preceding proposition hints at a way of constructing domains of

attraction of an asymptotically stable equilibrium point. One starts with a small ball
of initial conditions elose to the equilibrium point and integrates them backwards in
time. More specifically, let Ba be the initial ball of attraction around the equilibrium
point at the [Link] ö.j-tS jj c« {~-T(X): xE Ba}correspondstointegrating
equation (5.56) backwards for T seconds and is contained in n. By choosing T
large, we get a good estimate for the domain of attraction.
Example 5.45 Power Systems: Critical Clearing Times, Potential Energy Sur-
faces, Alert States. The dynamics of an interconnected power system with n
buses may be modeled by the so-called swing equations, which we have seen
somewhat earlier in this chapter:

+ D/l; + L
n
M ;ij; B;j sin(e; - ej ) = Pm; - Pe;.
;=1

Here e; stands for the angle of the i -th generator bus, M; the moment of inertia
of the generator, D; the generator damping, Pm;, Pe; the mechanical input-power
and the electrical output-power at the i-th bus, respectively. During the normal
operation ofthe power system, the system dynamics are located at an equilibrium
point of the system, ee E JRn. In the event of a system disturbance, usually in the
form ofa lightning strike on a bus, say, [Link] equipment on the bus opens,
interrupting power transfer from generator i to generator j. In this instance, the
dynamics of the system move from ee, since ee is no longer an equilibrium point
of the modified system dynamies. Of interest to power utility companies is the
so-called reelosure time for the bus hit by lightning. This is the maximum time
before the trajectory of the system under the disturbance drifts out ofthe domain
 5.9 Domains of Attraction 219

ofattraetion ofthe equilibrium point (Je . It is ofinterest tofind the maximumsueh

time under a variety ofdifferentfault eonditions so as to give the fault as long as
possible to clear.
In thepast, utilities relied on lengthyoff-linesimulations to determinetheseeriti-
eal clearing [Link] reeently,analytieteehniquesforestimatingtheboundaries
of the domains ofattraetion of equilibria have been developed. In partieular, the
potential energyboundary surfaee (PEBS)methodrelieson the "potentialenergy"
funetion [239J.
n n
W«(J) = L - L Bij COS«(Ji - (Jj) + L(Pei - Pmi)()i (5.57)
i=1 j~ i i=1
whieh is a modifieation of the funetion of equation (5.34) in Seetion 5.4 to ex-
clude the "kinetic energy" terms. It may now be verified that loeal minima of
W «(J) eorrespond to stable equilibria of the swing dynamies and that level sets
of the funetion W «(J) about loeal minima, whieh do not interseet any saddles of
the funetion give eonservative estimates ofthe domain ofattraetion (see Problem
5.30). In particular; a great deal ofeffort is spent on eomputing the "nearest un-
stable" equilibrium ofthe swing dynamies(i.e.•the saddle extremum or maximum
ofthefunetion W «(J)) closest to a given operatingpoint (Jo. An operatingpoint or
equilibrium of the form «(Jo , 0) is ealled an alert state when the nearest unstable
equilibrium point is close (in Euclidean norm) to (Jo.
We will now illustrate some qualitative considerations about the computation of
domains of attraction of equili bria of power systems . This is an extremel y important
part ofthe operational considerations of most electrical utilities which spend a great
deal of resources on stability assessment of operating points [314]. We begin with
some general considerations and remarks. Dur treatment follows Chiang , Wu, and
Varaiya [62]. Let x, be a locally asymptotically stable equilibrium point of the
time -invariant differential equation x = fex) and define the domain of attraetion
of X s to be

A(xs ) := l-: lim cP,(x) = x,,} .

'-+00

As we have shown earlier in this section A(xs ) is an open set containing X s ' We
will be interested in characterizing the boundary of the set A(xs ) ' It is intuitive
that the boundary of A (x,) contains other unstable equilibrium points, labeled xi.
We define their stable and unstable manifolds (in analogy to Chapter 2) as

Ws(Xi) = {x : lim cP,(x) = Xi}'

'-+00

-l
(5.58)
Wu(Xi) = {x : lim cP,(x) =
' ....... -00

In Chapter 7 we will state the generalizations of the Hartman-Grobman and stable

unstable manifold theorems which guarantee that the sets Ws, Wu defined above
are manifolds. We now make three assumptions about the system = fex): x
220 5. Lyapunov Stability Theory

Al. All equilibrium points of the system are hyperbolic. Actually, we only need
the equilibrium points on the boundary of A(xs ) to be hyperbolic.
A2. If the stable manifold Ws(Xi) of one equilibrium point intersects the unstable
manifold of another equilibrium point Wu(Xj), they intersect transversally.
The reader may wish to review the definition of transversal intersection from
Chapter 3, namely, if X E Ws (r,) n Wu(Xj), then

TAWs(Xi)) + TAWu(Xj)) = R".

A3. There exists an energy like function v(x) satisfying:

a. v(x) < 0 x rf- {x : lex) = O} .

b. {x : v(x) = O} is the set of equilibrium points.
c, v(x) is a proper function, that is, the inverse image of compact sets is
compact.
The first two hypotheses are "generically valid hypotheses." The last hypothesis
is a strong "global" version of a converse Lyapunov hypothesis, which insists that
all trajectories converge to one or the other of the equilibrium points and that there
are no closed orbits or other complex w limit sets. The following proposition can
now be proved.
Proposition 5.46 Unstable Equilibrium Points on the Boundary of A(xs ) '
Consider the time-invariant nonlinear system satisfying assumptions AI-A3 above,
with Xs an asymptotically stable equilibrium point. Then , an unstable equilibrium
point x, belongs to the boundary 01 A(xs) ifand only if
(5.59)

Proof: The proof relies on the follow ing lemma, which we leave as exercise to
the reader (taking the help of [62] if necessary):
Lemma 5.47.

1. An equilibrium point x , belongs to aA (Xi) ifand only if

Wu(Xi) - {x;} n Ä(x s) =f. 0.
2. Let xi, Xj be two equilibriumpoints. Then ifthe unstable manifoldofx, intersects
the stable manifold ofx], that is
Wu(Xi) - {x;} n Ws(Xj) - {Xj} =f. 0,
then the dimension oIWu(xi) is greater than the dimension oIWu(xj).
3. Assumption A3 implies that every trajectory starting on the boundary 01 A (x s )
converges to one ofthe equilibrium po ints on the boundary 01 A(xs) .
w.,
4. Iffor equilibria XI , X2, X3, (XI) intersects W u (X2) transversely and W u(X2)
intersects W.,(X3) transversely, then Wuexj) intersects Ws(X3) transversely.
The necessity of condition (5.59) follows in straightforward fashion from the
first statement of Lemma 5.47.
 5.9 Domains of Attraction 221

For the sufficiency, we first start with the case that x, has an unstable manifold
of dimension I on the boundary of A (x,,) . We would like to make sure that W u (Xi)
is not entirely contained in 8A(x,,). Indeed, if this were true, since all points on
the boundary of A (x s ) converge to one of the equilibrium points on the boundary
A(x,,), say X, then we have that

Wu(Xi) - {x;} n W,(x) - {x} =1= 0,

and from the preceding lemma, we would have that

1 = dirn Wu(Xi) > dirn Wu(x),
implying that the dimension of Wu(x) = 0, or that xis a stable equilibrium point
on the boundary of A (x,,) . This establishes the contradiction.
If x; has an unstable manifold of dimension 2 on the boundary of A(xs ) , as
before assuming for the sake of contradiction that Wu(Xi) lies entirely in 8A(x,,),
then there exists x such that
W u (Xi) - {x;} n W, (x) - {x} =1= 0 .

In going through the steps of the previous argument, we see that x is either stable
or has an unstable manifold of dimension I. Since x cannot be stable, it should
have a stable manifold of dimension 1. But then, by the last claim of Lemma 5.47,
it follows that
Wu(Xi) - {x;} n W,(x) - {x} =1= 0 and Wu(x) n A(x.,) =1= 0
implying that
Wu(Xi) n A(x,) =1= 0.

By induction on the dimension ofthe unstable manifold of x., we finish the proof.
o
Using this proposition, we may state the following theorem
Theorem 5.48 Characterization of the Stability Boundary. For the time in-
variant nonlinear system x = /(x) satisfying assumptions AJ -A3 above, let x. ,
i = 1, .. . , k be the unstable equilibrium points on the boundary 0/ the domain 0/
attraction 8 A (r,') 0/ a stable equilibrium point. Then,
k
8A(x s ) = UW.,(Xi)' (5.60)
i= 1

Proof: By the third statement of Lemma 5.47 it follows that

k
8A(x,,) C UW,(Xi)'
i=1

To show the reverse inclusion, note from the previous proposition that
222 5. Lyapunov Stability Theory

Choose p E Wu(Xi) n A(xs ) and let p E D C A(xs ) be a disk of dimension

n - dirn Wu(Xi) transversal to Wu(Xi) at p. It is intuitive that as t -+ -00 the flow
r/J, (D) tends to the stable manifold of x.. The precise proof of this fact is referred
to as the A lemma of Smale ([277], [I40l) and is beyond the scope of this book.
This lemma along with the fact that the invariance of A (x s ) guarantees that

and yields that

Ws(Xi) = lim r/J,(D)
t.... -00
c Ä(xs ) .

Since Ws (x,) n A(xs ) = 0, it follows that

W,(Xi) c aA(xs ) '
Consequently, we have that
k
U W ,(Xi) c aA(x s) ,
i=1

completing the proof. o

This theorem characterizing the boundary ofthe stability region can be combined
with the Lyapunov function to establish the following theorem, whose proof we
leave to the exercises (Problem 5.29).
Theorem 5.49 Lyapunov Functions and the Boundary of the Domain of At-
traction, Consider the time invariant nonlinear system x
= I(x) satisfying
assumptions AI-A3, with X s a locally asymptotically stable equilibrium point and
xi, i = 1, . .. , k the equilibrium points on the boundary 01 its domain 01 attraction
A(xs ) ' Then we have
1. On the stable manifold W, (Xi) 01 an equilibrium point, the Lyapunov function
v(x) attains its minimum at the equilibrium point xi.
2. On aA(xs ) the minimum olv(,) is achieved at a type 1 equilibrium point, that
is, one whose stable manifold has dimension n - 1.
3. 11 the doma in 01 attraction A(xs ) is bounded, the maximum 01 v( ·) on the
boundary is achieved at a source (that is, an equilibrium with a zero dimensional
stable manifold).
Remark: The preceding theorem is extremely geometrical. It gives a picture
of a dynamical system satisfying assumptions AI-A3 as "gradient like" [277].
The stable equilibria lie at the base of valleys. The region of attraction of these
equilibria are separated by "ridges" which are the stable manifolds of saddles or
peaks. This is very much the landscape of the earth as we fly over it in an aircraft.

Numerical Methods of Computing Domains 01 Attraction

There are many interesting methods for computing domains of attraction. A de-
tailed discussion of these is deferred to the literature: An interesting set of methods
 5.11 Exercises 223

was proposed by Brayton and Tong in [41]. While we have shown that quadratic
Lyapunov functions are the most natural for linear systems, there has been arecent
resurgence in methods for obtaining piecewise quadratic Laypunov functions for
nonlinear systems [156]. In this paper, the search for piecewise quadratic Lya-
punov functions is formulated as a convex optimization problem using linear
matrix inequalities. In turn, algorithms for optimization with linear matrix in-
equality constraints have made a great deal of progress in recent years (see, for
example, the papers ofPackard [238], [19] and the monograph by Boyd, Balakr-
ishnan, EI Ghaoui, and Feron [36]). We refer the reader to this literature for what
is a very interesting new set of research directions.

5.10 Summary
This chapter was an introduction to the methods of Lyapunov in determining the
stability and instability of equilibria of nonlinear systems. The reader will enjoy
reading the original paper of Lyapunov to get a sense of how much was spelled
out over a hundred years aga in this remarkable paper. There are several textbooks
which give a more detailed version of the basic theory with theorems stated in
greater generality. An old classic is the book of Hahn [124]. Other more modem
textbooks with a nice treatment include Michel and Miller [209], Vidyasagar [317],
and Khalil [162]. The recent book by Liu and Michel [210] contains some further
details on the use of Lyapunov theory to study Hopfield neural networks and
generalizations to classes of nonlinear systems with saturation nonlinearities .

5.11 Exercises
Problem 5.1 A p.d.f, function v with -iJ not a p.d.f., Considerafunctiong 2 (t )
with the form shown in Figure 5.9 and the scalar differential equation
. g(t)
x = --x.
g(t)

width

o 2 3 4 5
FIGURE 5.9. A bounded function g2(t) converging to zero but not uniformly.
224 5. Lyapunov Stability Theory

Consider the Lyapunov funetion eandidate

v(x, t) = ~2
g (t)
[3 _10r g2(T:)dT:] .

Verify that v(x, t) is a p.d.f. and so is v(t, x) . Also, verify that the origin is stable
but not asymptotically stable. What does this say about the missing statement in
the table of Lyapunov's theorem in Theorem 5.16.

Problem 5.2 Krasovskii's theorem generalized. (a) Consider the differential

x
equation = fex) in IR". If there exist P, Q E IR"X", two constant, symmetric,
positive definite matriees sueh that
af aj T P
P-(x) + -(x) = -Q,
ax ax
then prove thatx = 0 is globally asymptotieally stable. Hint: Try both the Lyapunov
function candidates f(X)T P fex) and x T Px. Be sure to justify why theformer is
a p.d.f. ifyou decide to use it!
x
(b) Consider the differential equation = A (x)x in IR". If there exist P, Q E
IR" X", two constant, symmetrie, positive definite matriees sueh that

PA(x) + A(x)T P = -Q,

then prove that x = 0 is globa11y asymptotieally stable .

Problem 5.3. Considcr the seeond-order nonlinear system:

Show that the linearization is ineonc1usivein determining the stability ofthe origin.
Use the direet method of Lyapunov and your own ereative instinets to piek a V to
study the stability of the origin for -I ::: E ::: I.

Problem 5.4. Consider the seeond order nonlinear system:

X2 = -XI + (1 - xf - Xi)X2 '

• Diseuss the stability of the origin .

• Find the (only) limit eyc1e of the system .
• Prove using a suitablc Lyapunov funetion and LaSa11e 's prineiple that a11
trajeetories not starting from the origin eonverge to the limit eyc1e.

Problem 5.5 Rayleigh 's equation. Consider the seeond-order Ray leigh system:

XI = -E(X? /3 - XI + X2),
 5.11 Exerciscs 225

This system is known to have one equilibrium and one limit eyele for E :f:. O. Show
that the limit eycle does not lie inside the strip -I ~ XI ~ I. Show, using a suitable
Lyapunov funetion, that it lies outside a eircle of radius -/3.

Problem 5.6 Discrete-time Lyapunov theorems. Consider the diserete time

nonlinear system

x(k + I) = !(x(k) , k). (5.61)

Cheek the definitions for stability, uniform stability, ete. and derive diserete time
Lyapunov theorems forthis system with suitable eonditions on a Lyapunov funetion
eandidate v(x, k) and on

L\v(x(k), k) := v(x(k + I), k + I) - v(x(k), k)

= v(f(x(k), k), k + I) - v(x(k), k).

Also, diseuss theorems for exponential stability and instability. Is there a

eounterpart to LaSalle's principle for diserete time systems?
Problem 5.7 Discrete-time Lyapunov equation. Provc that the following
statements are equivalent for the diserete-time linear time invariant system

x (k + I) = Ax(k) :

1. X = 0 is exponentially stable.
2. lAi I < 1 for all the eigenvalues of A.
3. Given any positive definite symmetrie matrix Q E Rnxn, there exists a unique,
positive definite symmetrie matrix P E IRnxn satisfying

4. Given a matrix C E IRnoxn sueh that A, C is observable, there exists a unique

positive definite matrix P sueh that
A TpA - P = -CTC.

Problem 5.8 Discrete-time stability from the linearization. Give a proof of

the uniform loeal stability of the equilibrium at the origin of the diserete-timc
non linear system (5.61) when the linearization is uniformly stable .
Problem 5.9 Barbalat's lemma [259]. Assume that rPO : IR Ho IR is
uniformly continuous and integrable, that is,

1 00

rP(t)dt < 00.

Then show that

lim rP(t)
1--> 00
= o.
Give a eounterexample to show the neeessity of uniform eontinuity.
226 5. Lyapunov Stability Theory

Problem 5.10 Anotber generalization of LaSalle's princlple, Consider a

nonlinear time varying system Lipschitz continuous in x unifonnly in t. Let
v; R x Rn ~ R be such that for x E Bn 8 > O. and functions ClI, ClZ of dass K,

Cl] (lxI) ~ v(x, r) .s Clz(lxl),

v(t, x) ~ 0, (5.62)
tH
[
t v(s(r , t , x), r)dr ~ -AV(X, t) 0 < A < l.

Now show that for Ix(to)1 < Cl;' (ClI (r» show that the trajectories converge
uniformly, asymptotically to O.

Problem 5.11 Power system swing dynamics on T" x Rn [9]. Consider the
swing equations of the power system (5.33). Since the right-hand side is periodic
in each of the B; with period 2iT. we can assume that the state space of the system
is Si x . . . X Si X Rn or T" X Rn. Note that the function v(B, cu) of (5.34) is not
a weil defined function from T " x Rn ~ R. unless Pm; - Pe; = O. What are the
implications of the assertion in the text that bounded trajectories (on Rn X Rn)
converge to equilibrium points of the swing dynamics on the manifold T" x Rn?
Prove that trajectories of the swing equations are bounded on the manifold
T" x Rn. Can you give counterexamples to the assertion that bounded trajectories
on T" x Rn converge to equilibria (you may wish to review the Josephson junction
circuit of Chapter 2 for hints in the case that n = I)? Prove that all trajectories
of the swing equation converge to equilibria for Pm; - Pe; sufficiently small for
i = I , .. . , n .

Problem 5.12 Generalization of tbe time-varying Lyapunov lemma. Con-

sider the linear time varying system

x= A(t)x .

Define it to be uniformly completely observable from the output y(t) = C(t)x(t),

if there exists a /j > 0 such that the observability Gramian defined in (5.63) is
unifonnly positive definite.

j
tH
Wo(t, t + 8):= t <!>T(r, t)C T (r)C(r)<!>(r, t)dr) 2: kl Vt. (5.63)

Now use Problem 5.10 to show that if there exists a uniformly positive definite
bounded matrix P (-) : R ~ Rn x n satisfying

k]f ~ pet) s kzf,

(5.64)
-pet) = AT(t)P(t) + P(t)A(t) + CT(t)C(t) .
with A(t) , C(t) a unifonnly completely observable pair, then the origin is uni-
fonnly asymptotically stable and hence exponentially stable . Prove the converse
as weil.
 5.11 Exercises 227

Problem 5.13 Time-varying Lyapunov lemma generalized with output injec-

tlon [259]. Prove that if a pair (A(t) , C(t)) is uniformly completely observable,
that is the observability Gramian of (5.63) is uniformly positive definite, then for
any K (t) E ]R nx n ) that (A(t) + K (t)C(t), C(t)) is also uniformly completely ob-
servable. You will need to do a perturbation analysis to compare trajectories ofthe
two systems

x= A(t)x(t),
tU = (A(t) + K(t)C(t))w(t),
with the same initial condition at time to and give conditions on the terms of the
second observability Gramian using a perturbation analysis. You may wish to refer
to the cited reference for additional help.

Problem 5.14 Zubov's converse theorem.

1. For time-invariant nonlinear systems assume that there exists a function v :

n ~ [0, I [ with v(x) = 0 if and only if x = O. Further, assurne that if n is
bounded, then as x ~ an, v(x) ~ 1. On the other hand, if Q is unbounded
assurne that [x] ~ 00 => v(x) ~ 1. Assume also that there exists a p.d.f,
hex), such that for x E n

iI(x) = -h(x)(1 - v(x)) . (5.65)

Show that n is the domain of attraction ofthe asymptotica11ystable equilibrium

point O.
2. Let v , h be as in the previous part, but replace equation (5.65) by
I

iI(x) = - h ex ) (I - v(x»(1 + If(x)1 2 )i . (5.66)

Prove the same conclusion as in the previous part .

Problem 5.15 Modified exponential stability theorem. Prove the fo11owing

modification of the exponential stability theorem (5.17) of the text: If there exists
a function v(x, t) and some constants h, al, a2, a3, a4 such that for a11 x E Bh,
t ::: 0,

allxl 2 s v(x, t) s a21x1 2,

dv(x , t) I s 0,
dt (5 .1) (5.67)

f I
l+O dv(x , t)
dt
I
(5 .1)
:s -a3lx (t )1 2 ,
then x(t) converges exponentia11y to zero. This problem is very interesting in that
it allows for a decrease in v(x, t) proportional to the norm squared of x(t) over a
window of length 8 rather than instantaneously.
228 5. Lyapunov Stability Theory

Problem 5.16 Perturbations of exponentially stable systems. Consider the

following system on ]Rn
x= f t», t) + g(x, r). (5.68)
Assume that f(O, t) = g(O, t) == O. Further, assume that:
1. 0 is an exponentially stable equilibrium point of
x= [t» , t).
2.
ig(x,t)l::: J1-lxl "Ix E ]Rn.

Show that 0 is an exponentially stable equilibrium point of 5.68 for J1- small enough.
The moral of this exercise is that exponential stability is robust!
Problem 5.17 Bounded-input bounded-output or finite gain stability ofnon-
linear systems. Another application of Problem 5.16 is to prove L oo bounded
input bounded output stability of nonlinear systems (you may wish to review the
definition from Chapter 4). Consider the nonlinear control system

x = ft», t) + g(x, t)u,

(5.69)
Y = hex) .

Prove that ifO is an exponentially stable equilibrium point of f (», t), Ig(x, t)1 :::
J1-lx land finally
h(O) = 0, Ih(x)l::: ßlxl "Ix,

then the control system (5.69) is L oo finite gain stable. Give the bound on the L oo
stabiIity.
Problem 5.18 Loeal exponential stability implies exponential stability ofthe
Iinearization. Prove the converse ofTheorem 5.41, that is, ifO is a locally expo -
nentially stable equilibrium point ofthe non linear system (5.1) then it is a (globaIly)
exponentially stable equilibrium point of the Iinearized system, provided that the
Iinearization is uniform in the sense ofTheorem 5.41 . Use the exponential converse
theorem of Lyapunov.
Problem 5.19 Intereonnected systems. The methods of Problem 5.16 may
be adapted to prove stability results for interconnected systems. Consider an
interconnection of N systems on ]Rn;, i = 1, . . . N, given by
N
Xi = f;(Xi, t) +L gij(XJ, .. . ,XN, t) (5.70)
j=1

with f; (r,, t) , gij(XI , ... , XN, t) vector fields on ]Rn; for i = 1, . .. , N. Now
assume that in each of the decoupled systems
Xi = f;(Xi, t) (5.71)
 5.11 Exercises 229

the origin is an exponentially stable equilibrium point, and that for i, j = I , ... , N ,
N
Igij (xl, . .. , XN, t)1s L Yij IXj l.
j=1

Using the converse Lyapunov theorem for exponentially stable systems to generate
Lyapunov functions Vi (Xi , r) satisfying
i 2 i 2
0'Jix;I ::::: Vi (X i, t) ::::: 0'21x;I ,

-8Vi
8t
+ -8Vi
f ; (x -
8Xi I "
t) < -0'
-
i
3
lx-I"
2
(5.72)
I
8Vi ::::: O'~lx;I,
-
-
I
8Xi

give conditions on the 0'\' O'~, O'~, O'~, Yij to prove the exponential stability of the
origin. Hint: Consider the Lyapunov function
N
L diVi(Xi),
i= 1

where d, > 0 and Vi (z,) is the converse Lyapunov fun ctionfor each ofthe systems
(5.71).
Sec [162] for a generalization of this problem to the case when the stability
conditions of equations (5.72) are modified to replace lXi I by f/Ji (z,) where f/Ji(.) is
a positive definite function, and
N
Igij (x l , ... , xN )1 ::::: L Yijf/Jj(Xj).
j=1

Problem 5.20 Hopfield networks with asymmetrie intereonneetions. Give

a counterexample to the lack of oscilIations of the Hopfield network of (5.37)
for asymmetrie interconnections Tij. Modify the method of Problem 5.19 to give
conditions under which the Hopfield network with asymmetrie interconnections
does not have limit cycles.

Problem 5.21 Satellite stabilization. Consider the satellite rigid body B with
an orthonormal body frame attached to its center of mass with axes aligned with
the principal axes of inertia as shown in Figure 5.10. Let bo E R 3 be a unit vector
representing the direction of an antenna on the satellite and do E R 3 a fixed unit
vector representing the direction of the antenna of the receiving station on the
ground. Euler's equations of motion say that if w E R 3 is the angular velocity
vector of the satellite about the body axes shown in Figure 5.10, then

lw+wx!w=r (5.73)

Here! = diag (11, h h). The aim ofthis problem is to find a controllaw to align
be with -do. In the body frame, do is not fixed. 1t rotates with angular velocity -w
230 5. Lyapunov Stability Theory

FIGURE 5.10. Satellite and earth station

so that
do = -w x do (5.74)
Now consider the controllaw
r = -cxw + do x bo
applied to (5.73), (5.74). What are the equilibrium points ofthe resulting system in
]R3+3? Linearize the system about each equilibrium point and discuss the stability
of each. Is it possible for a system in ]R6 to have this number and type of equilibria.
Do the system dynamics actually live on a submanifold of ]R6? If so, find the state
space of the system. Prove that the controllaw causes all trajectories to converge to
one of the equilibrium points. It may help to use the Lyapunov function candidate
1 T 1 2
v(w,do) = 2: w lw+ 2:ldo +bol .
Another approach to satellite stabilization is in arecent paper by D' Andrea Novel
and Coron [77].

equiJibrium position: W(x. t).=O

----------------.,,-- -W{x, t) --- --- ----- - - +
,, utt)
....t--- - x ---:, x=l
(free end)
x=O
(Iixed end)
FIGURE 5.11. Boundary control stabilization of a vibrating string
 5.11 Exercises 231

Problem 5.22 Boundary control of a vibrating string, The equation of motion

of a linear, frictionless, vibrating string of length 1 is given by the following partial
differential equation:
a 2W(x, t) a 2W(x, t)
m at 2 - T ax 2 = 0, (5.75)

where W (x, t) denotes the displacement of the string at loeation x at time t, with
x E]O, 1[, t ::: O. The quantities m and T are the mass per unit length of the string
and the tension in the string, respectively. Let the string be fixed at x = 0 as shown
in Figure 5.11. Thus, W(O, t) = 0 for all t ::: O. Let a vertical control force u( ·)
be applied to the free end of the string at x = I. The balance of the forces in the
vertical direction (at x = I) yields:

aW(x, t)
u(t) = T - - -
I (5.76)
ax .r ee l

for all t ::: O. The energy of the spring at time t is

E(t) = ~
210
t m (aw(x,
at
t))2 dx + ~ [I T (aw(x , t))2 dx
210 ax
(5.77)

which consists ofthe kinetic energy (first integral in the expression above) and the
strain energy (second integral above). Show that if a boundary control

u(t) = -k aW(x, t) I (5.78)

ot x= J

is applied, where k > 0, then B(t) .::: 0 for all t ::: O. See also [274].
Problem 5.23 Brockett's H dot equation [26], [132].
1. Consider the matrix differential equation (H(t) E JRnxn

iI = [H , G(t)] (5.79)
Here G(t) E JRnxn is a skew-symmetric matrix. The notation [A , B] := AB -
BA refers to the Lie bracket, or commutator, between A and B. Show that if
H (0) is symmetrie, then H (r) is symmetrie for all t. Further, show in this case
that the eigenvalues of H (r) are the same as the eigenvalues of H (0). Equation
(5.79) is said to generate an iso-spectral flow on the space of symmetrie matrices
SS(n) .
2. Consider the so-called double bracket equation
iI = [H, [H , Nll (5.80)

with N E JRnxn a constant matrix. The Lie bracket is defined as above . Show
that if H (0) and N are symmetrie, then (5.80) generates an iso-spectral flow as
weIl on SS(n) .
3. Now using the Lyapunov funetion candidate
v(H) = - traee(H T N)
232 5. Lyapunov Stability Theory

and a norm on R'"?' givenby IIAII = (trace A" A)!,provethat

iJ(H) = -IHH , N] 11 2

Now use the fact that

and LaSalle 's principle to find the limits as t ---+ 00 of H(r).

4. Consider the double bracket equation with

N = diag()"1, A2, . . . , An)

with AI > A2 > . . . > An > O. Use the iso-spectrality of (5.80) and the
previous calculations to characterize the equilibrium points of (5.80). How
many are there?
5. Linearize (5.80) about each equilibrium point and try to determine the stability
from the linearization. What can you say if some of the eigenvalues of N, H (0)
are repeated?

Discuss how you might use the H dot equation to sort a given set of n numbers
AI , ... , An in descending order. How about sorting them in ascending order?

Problem 5.24 Controlled Lagrangians for the satellite stabilization problem

[28], [30]. Consider a satellite system with an internal rotor as shown in Figure
5.12 with equations given by

A,WI = A2W2W3 - (A3W3 + 1]ä)W2 ,

A2W2 = AIWjW3 - (A3W3 + hä)WI,
(5.81)
A3W3 + 1]& = (AI - A2)W,W2,
w
1] 3 +1]& = u,

where the body angular velocity is given by (Wj, W2, ( 3)T E IR 3 , and 1, ,/2,13 are
the satellite moments of inertia, J 1 = Ji- h the rotor moments of intertia, and
a the relative angle of the rotor. Furthermore, Ai = I, + Ji, It is easy to verify
that regardless of the choice of u we have an equilibrium manifold w, = W3 = 0
and W2 = m arbitrary. This corresponds to the satellite spinning freely about the
intermediate axis. Prove that the control law

stabilizes the satellite about each of these equ ilibria. Note that you are not required
to prove anything about the variables a, ä.

Problem 5.25 Rates of convergence estimates for the Lyapunov equation.

Consider the Lyapunov equation
 5.11 Exercises 233

Satellite cut open

FIGURE 5.12 . Satellite with a spinning rotor aligned with its central principal axis as
controller

Note that an estimate of the rate of convergence is given by

Amin(Q)
JLA(Q) := 2A
max(P)'

Show the following three properties of JLA(Q):

1. Show that JLA(kQ) = JLA(Q) for all k E 1R+.

2. Show that if Amin(Q) = 1, then JLA(/) 2: JLA(Q).
3. Using the two previous facts, show that JLA (/) 2: JLA (Q) for arbitrary Q.
Thus, if one were to use the Lyapunov equation to estimate the rate of convergence
of a linearequation, it is best to use Q = I! Now, use this to study the stability of
the perturbed linear system

x= Ax + f t», t). (5.82)

Here , A E IRn x n has its eigenvalues in C~, and If(x, 01 ::: Ylxl . Now prove that
the perturbed linear system is stable if Y < JLA (/).
Problem 5.26 Stability conditions based on the measure of a matrix. Recall
the definition ofthe measure ofa matrix JL(A) from Problem 3.2. Use this definition
to show that iffor some matrix measure JLO, there exists an m > 0, such that

l'0
'O+T
JL(A(r))dr < -m Yt 2: T to 2: 0,

then the origin of the linear time -varying system of (5.44) is exponentially stable .
Apply this test to various kinds of matrix measures (induced by different norms) .

Problem 5.27 Smooth stabilizability [279]. Assume that the origin x = °of
the system
x= [t», u)
234 5. Lyapunov Stability Theory

is stabilizable by smooth feedback, that is, there exists u = k(x) such that 0 is
x
a locally asymptotically stable equilibrium point of = f(x, k(x». Now prove
that the extended system

x= f(x, z),
(5.83)
z= h(x, z) +u
is also stabilizable by smooth feedback. Hint: Use the converse Lyapunov function
associated with the x system and try to get z to converge to k (x). Repeat the problem
when the original problem is respectively globally or exponentially stabilizable.
By setting h(x, z) == 0, we see that smooth stabilizability is a property which is
not lost when integrators are added to the input!
Problem 5.28. Prove Lemma 5.47.
Problem 5.29. Prove Theorem 5.49. Now apply this theorem to the following
generalization of the power system swing dynamics

(5.84)
Xz = -Dxz - Vw(xd.
Here XI, Xz E IRn, w(xd is a proper function with isolated non-critical stationary
points and D E IRn xn is a symmetric positive definite matrix. Is this "second order"
system gradient like? Characterize the domains of attraction of its stable equilibria.
Show that all the equilibria of the system are either stable nodes, unstable nodes,
or saddles (i.e., that the eigenvalues of their linearizations are real).
Problem 5.30 Potential energy boundary surfaces. Prove rigorously that the
method sketched in the Example 5.45, a level set of W (ß) around a local minimum
ßo just touching a saddle or local maximum of W , is a (conservative) estimate of
the domain of attraction of the equilibrium ßo of the swing dynamics. You may
wish to use the "Lyapunov" function of (5.34) for helping you along with the proof.