7

System Stability
During the description of phase portrait construction in the preceding chapter, we
made note of a trajectory’s tendency to either approach the origin or diverge from
it. As one might expect, this behavior is related to the stability properties of the
system. However, phase portraits do not account for the effects of applied inputs;
only for initial conditions. Furthermore, we have not clarified the definition of
stability of systems such as the one in Equation (6.21) (Figure 6.12), which neither
approaches nor departs from the origin. All of these distinctions in system
behavior must be categorized with a study of stability theory. In this chapter, we
will present definitions for different kinds of stability that are useful in different
situations and for different purposes. We will then develop stability testing
methods so that one can determine the characteristics of systems without having
to actually generate explicit solutions for them.

7.1 Lyapunov Stability

The first “type” of stability we consider is the one most directly illustrated by the
phase portraits we have been discussing. It can be thought of as a stability
classification that depends solely on a system’s initial conditions and not on its
input. It is named for Russian scientist A. M. Lyapunov, who contributed much
of the pioneering work in stability theory at the end of the nineteenth century. So-
called “Lyapunov stability” actually describes the properties of a particular point
in the state space, known as the equilibrium point, rather than referring to the
properties of the system as a whole. It is necessary first to establish the
terminology of such points.

7.1.1 Equilibrium Points

Many of the concepts introduced in this chapter will apply equally well to
nonlinear systems as to linear systems. Even though much of the control system
design we will perform will be applied to linear models of perhaps nonlinear
270 Part II. Analysis and Control of State Space Systems

systems, it is a useful skill to analyze the stability properties of nonlinear systems

also. So consider the general form of an arbitrary system, perhaps nonlinear, and
perhaps time-varying:

x ( t )  f ( x ( t ), u( t ), t ) (7.1)

Whereas we were interested in the behavior of the phase trajectories relative to

only the origin in the previous chapter, for general nonlinear systems the point of
interest is not necessarily the zero vector, but rather a general equilibrium point:

Equilibrium point: An equilibrium point x e is a constant

vector such that if x ( t 0 )  x e and u( t )  0 in (7.1), then
x ( t )  x e for all t  t 0 . (7.2)

Because an equilibrium point is a constant vector, we have

x ( t )  0  f ( x e ,0, t ) (7.3)

Equilibrium points are sometimes also referred to as critical points or stationary

points. They represent a steady-state constant solution to the dynamic equation,
(7.1). In discrete-time systems, an equilibrium point is a vector x e such that
x ( k  1)  x ( k ) (  x e ) , rather than x ( t )  0 .
Considering now the special case of a linear system, an equilibrium point(s)
can be computed from

x ( t )  0  Ax e (7.4)

From this expression, it is apparent that the origin x e  0 will always be an

equilibrium point of a linear system. Other equilibrium points will be any point
null(A) that lies in the null spaceM of the system matrix A. Sets of equilibrium points will
therefore constitute subspaces of the state space, and we can have no isolated
equilibrium points outside the origin.

Example 7.1: Equilibrium Points for Linear and Nonlinear Systems

Find the equilibrium points for the following systems:

a) x 
LM1 3OP x  LM0OPu b) x( k  1) 
LM2 0OP x( k ) c) x   sin x
N3 9Q N1Q N0 .5Q
 Chapter 7. System Stability 271

Solution:
a) The b-matrix is irrelevant because we will assume a zero input. The A-
matrix has a zero eigenvalue so it has a nontrivial null space, which can be
found to lie along the line x e  [3 1] . This is the equilibrium set.

b) The discrete-time equilibrium implies that x ( k  1)  x ( k )  Ax ( k ) , so we

seek the solutions to the system ( A  I ) x e  0. For the A-matrix given,
A  I has no zero eigenvalues, so it has no nontrivial null space. The only
equilibrium solution is therefore x e  [0 0] .

c) The third system is nonlinear. Setting x  0   sin x e , we see that

x e   n  , where n  0,1, . This constitutes an infinite sequence of
isolated equilibria, which we cannot encounter in a linear system.

We are now prepared for the stability definitions pertinent to equilibrium

points. For the linear systems we treat here, we will speak of only isolated
equilibria at the origin, but sets of equilibria such as lines through the origin may
be interpreted similarly.

Stability in the Sense of Lyapunov: An equilibrium point x e

of the system x  A( t ) x is stable in the sense of Lyapunov, or
simply stable, if for any   0 , there exists a value ( t 0 ,  )  0
such that if x ( t 0 )  x e   , then x ( t )  x e   , regardless of
t ( t  t 0 ). The equilibrium point is uniformly stable in the sense
of Lyapunov if   (  ) , i.e., if the constant  does not depend
on initial time t 0 . (7.5)

As one might expect, time-invariant systems that are stable are uniformly
stable because the initial time t 0 should not affect the qualitative behavior of the
system. In words, the above definition means that for a stable in the sense of
Lyapunov system, anytime we desire that the state vector remain within distance
 of the origin, it is possible to accomplish this by simply giving it an initial
condition away from x e by an amount no larger than  . This scenario is depicted
in Figure 7.1. Clearly, we must have    or else the system will have an initial
condition that has already violated the desired bounds of the state vector.
Lyapunov stability is sometimes referred to as internal stability because its
definition describes the behavior of the state variables rather than the output
272 Part II. Analysis and Control of State Space Systems

variables, which are by contrast considered external. We will see stability

definitions related to system inputs and outputs in Section 7.3.
Note that the notion of Lyapunov stability does not require that the system in
question approach the equilibrium (usually, the origin), only that it remain within
the bounding radius  . For a system that is Lyapunov stable, we further qualify
the equilibrium point as asymptotically stable if the state vector tends toward zero,
i.e., if x ( t )  x e  0 as t   . In addition, if we can find constants   0 and
  0 such that for all t  t 0 ,

x ( t )  x e  e   ( t  t0 ) x ( t 0 )  x e

then the equilibrium is said to be exponentially stable.

As one might suspect, an LTI system that is asymptotically stable is also
exponentially stable because all solutions of LTI systems that approach the origin
do so by exponential decay. This is not necessarily the case for time-varying
systems and is certainly not always the case for nonlinear systems.

Asymptotically
stable


xe

Lyapunov Lyapunov
stable stable

Figure 7.1 Representative trajectories illustrating Lyapunov stability. The picture

shows a two-dimensional space such that the dotted circles indicate
bounds on the 2-norm of the state vector.

We can also speak of systems that are stable locally or stable globally (also,
stable in the large). An equilibrium that is stable globally is one such that the
constant  can be chosen arbitrarily large. This implies that if an equilibrium is
 Chapter 7. System Stability 273

asymptotically stable, then the trajectory from any initial condition will
asymptotically approach it. This in turn implies that there can be only one
equilibrium that is asymptotically stable in the large.
For linear systems, wherein equilibria are either isolated points or subspaces,
all stability is global. There is no other equilibrium that a large initial condition
might approach (or diverge from). However, for nonlinear systems, wherein
multiple isolated equilibria might exist, one sometimes has to define regions of
attraction around attractors, i.e., stable equilibria. These are sometimes called
basins of attraction or insets, denoting the region in the state space from which
an initial condition will approach the equilibrium point. Outside that region of
attraction, the trajectory might go elsewhere. Again, that will not happen in linear
systems.
As an illustration, the phase portrait of the nonlinear system


  0.4   sin   0 (7.6)

is shown in Figure 7.2.

 0

-5
-10 -5 0 5 10

Figure 7.2 Phase portrait of the damped pendulum of Equation (7.6). The symbols
(+) indicate the equilibria. (What would the phase portrait of an
undamped pendulum look like?)

The graph plots trajectories on the  versus  axis. The system (7.6) represents
the equation of a damped pendulum, i.e., with viscous friction. If we compute the
equilibria of this system by solving sin  e  0 , which is similar to Example 7.1c),
274 Part II. Analysis and Control of State Space Systems

we find equilibrium points at x e   n  , where n  0,1, Of these, it is

apparent from the figure that zero and the even multiples of  are stable, whereas
the odd multiples of  are unstable. Note, in particular, that each stable
equilibrium has a well-defined region of attraction, which defines the set from
which an initial condition will produce a trajectory that approaches that
equilibrium. For the equilibria that are stable, they are not globally stable. This
being the equation for a pendulum, we can identify the stable equilibria as points
at which the pendulum hangs directly down. The unstable equilibria are points at
which the pendulum is balancing straight up, which are intuitively clearly
unstable.

7.1.2 Classification of Equilibria

Given our definition of Lyapunov stability, we can reconsider the phase portraits
of Chapter 6 in terms of these definitions. Because it is the eigenvalues of the A-
matrix that determine the directions of the phase trajectories, it will be the
eigenvalues that determine the type of stability exhibited.
In the captions to the phase portraits in Chapter 6, the origins of each plot
were given labels, e.g., stable node, or focus. Each such label applies to any
equilibrium of a phase portrait, and is a characteristic of the stability type of that
equilibrium. Recalling the system examples from Chapter 6 that had two
eigenvalues with negative real parts, i.e., the systems given by Equations (6.9),
(6.12), (6.18), (6.19), and (6.20), we can immediately see from the corresponding
figures that all trajectories asymptotically approach the origin (in fact, they do so
exponentially). Such equilibria are clearly Lyapunov stable. Those that have at
least one real invariant subspace (i.e., eigenvector), namely (6.9), (6.12), (6.18),
and (6.19), have equilibria known as stable nodes. The equilibrium in the
asymptotically stable system with complex conjugate eigenvalues (6.20) is known
as a stable focus. The origin of the system described by (6.21) is stable but not
asymptotically. It is known as a center. For systems with one negative and one
positive eigenvalue, such as in (6.15) and (6.17), the origin is known as a saddle
point.
Most of the phase portraits drawn (with the exception of the saddle points)
show stable systems. Unstable systems, for example, those with unstable nodes,
can easily be imagined by reversing the arrowheads on the portraits with stable
nodes. Likewise, an unstable focus spirals out of the origin, the opposite of Figure
6.10. In the case of the center, there is no real part to the eigenvalues, so no such
distinction needs to be made. However, with an appropriate choice of the A-
matrix, the trajectories can be made to progress in the opposite direction from
those shown (i.e., counterclockwise rather than clockwise).

7.1.3 Testing For Lyapunov Stability

While the definitions above give us the meaning of stability, they do little to help
us determine the stability of a given system (except that it is now easy to anticipate
 Chapter 7. System Stability 275

the stability conditions for a time-invariant system based on the locations of its
eigenvalues). In this section, we will give some results that will constitute stability
tests, useful for examining a system at hand.
Considering a general time-varying zero-input system,

x ( t )  A( t ) x ( t ) (7.7)

we present the first result as a theorem:

THEOREM: The zero-input system in (7.7) is stable in the

sense of Lyapunov if and only if there exists a constant
( t 0 )   such that for all t  t 0 , the state-transition matrix
( t , t 0 ) satisfies the relation

( t , t 0 )   ( t 0 ) (7.8)

If the constant  does not depend on t 0 , then the origin of the

system described by (7.7) will be uniformly stable in the sense
of Lyapunov.

PROOF: Recall that x ( t )  ( t , t 0 ) x ( t 0 ) . Then by the

Cauchy-Schwarz inequality (see Section 2.2.7), we have

x ( t )  ( t , t 0 ) x ( t 0 ) (7.9)

(Sufficiency) Suppose ( t , t 0 )   ( t 0 ) . Given a value  , we

are free to choose the value ( t 0 ,  )    ( t 0 ) . Then from (7.9),


x( t )  ( t , t 0 ) x( t 0 )  ( t 0 ) x( t 0 )  ( t 0 )
( t 0 )

or

x ( t )   for all t  t 0

(Necessity) Assume that the system is stable and show that

( t , t 0 ) is bounded. Suppose that ( t , t 0 ) were not
bounded. Then by definition, given x( t 0 ) such that
276 Part II. Analysis and Control of State Space Systems

x ( t 0 )  ( t 0 ,  ) and an arbitrary value   M   , there

exists a time t such that

( t , t 0 ) x ( t 0 )  M

But since x ( t )  ( t , t 0 ) x ( t 0 ) , this would imply

x(t )  M  

which violates the stability condition, a contradiction. Thus,

necessity is proven.

For asymptotic stability, we give the following result:

THEOREM: The origin of the system (7.7) is uniformly

asymptotically stable if and only if there exist two constants
 1  0 and  2  0 such that

( t , t 0 )   1e   2 ( t  t0 ) (7.10)

PROOF: (Sufficiency) If (7.10) holds and if x( t 0 )   , then

x ( t )  ( t , t0 ) x ( t0 )  1e   2 ( t  t0 )

so clearly x( t )  0 as t   .

(Necessity) We will divide the time axis into equal steps of size
 . We will then look at the state of the system at a particular
time t, considered to satisfy

t 0  k  t  t 0  ( k  1) (7.11)

Knowing that the equilibrium is uniformly asymptotically

stable (globally, because it is linear), then we know that for any
starting point x ( t 0 )  1 , given an arbitrary value  1 , there
must exist a time  such that x ( t 0   )  1 . This being the
 Chapter 7. System Stability 277

case, we are free to choose x ( t 0 )  1 , and select 1  e 1 .

Then there will exist  such that

x ( t 0   )  ( t 0  , t 0 ) x ( t 0 )  e 1

By the definition of an operator norm,

( t 0  , t 0 )  sup ( t 0  , t 0 ) x( t 0 )  e 1
x ( t0 ) 1

So now if we consider time t 0  k ,

( t0  k, t0 )  ( t0  k, t0  ( k  1)t0 )

 ( t0  ( k  1), t0  ( k  2)t0 )
 ( t0  2, t0   )( t0  , t0 )
 ( t0  k, t0  ( k  1)t0 ) (7.12)
 ( t0  ( k  1), t0  ( k  2)t0 ) 
  ( t 0  2  , t 0   )  ( t 0  , t 0 )
 e k

Now consider the quantity ( t , t 0  k ) . Because () is

bounded, we know that ( t , t 0  k ) is a finite number. Call
it

( t , t 0  k )   1e 1 (7.13)

for some constant  1   . Now we have that

( t , t 0 )  ( t , t 0  k )( t 0  k, t 0 )

so

( t , t0 )  ( t , t0  k )  ( t0  k, t0 )
(7.14)
 1e 1e  k  1e b g   e(1  ) t0 b k 1gt0
 k 1
1

by (7.13) and (7.12). However, from (7.11) we have asserted

that t  t 0  ( k  1) , so
278 Part II. Analysis and Control of State Space Systems

e  ( t  t0 )  e
b g
 t0  k 1   t0

and therefore,

1e
b g
 (1  ) t0  k 1   t0
 1e b g
 (1  ) t  t0


Now defining  2  (1  ) , we have proven necessity by showing
that

( t , t 0 )   1e b g
  2 t  t0

As usual, we can do a little better than this for time-invariant systems. Given
the proof above, the proofs of the following two theorems are relatively easy and
are left as an exercise.

THEOREM: The equilibrium of the LTI system x  Ax is

Lyapunov stable if and only if all the eigenvalues of A have
nonpositive real parts, and those that have zero real parts are
nonrepeated. (7.15)

THEOREM: The equilibrium of the LTI system x  Ax is

asymptotically Lyapunov stable if and only if all the
eigenvalues of A have negative real parts. (7.16)

It is these last two results that are the most commonly used to test for
Lyapunov stability. However, as we will see in the next section, one must be
careful not to apply the above results for time-invariant systems to time-varying
systems.

Discrete-Time Systems
These results are directly extensible to discrete-time systems. In particular, all the
boundedness conditions for the state-transition matrix ( t , t 0 ) extend as well to
( k , j ) . In the case of time-invariant systems, we can still derive eigenvalue
conditions, but we replace all reference to the left half of the complex plane with
the inside of the unit circle. Thus, the origin of a discrete-time LTI system is
asymptotically stable if the eigenvalues of Ad are strictly inside the unit circle,
and it is Lyapunov stable if the eigenvalues are inside or on the unit circle. Those
eigenvalues that are on the unit circle, however, must not be repeated.
 Chapter 7. System Stability 279

7.1.4 Eigenvalues of Time-Varying Systems

In general, one should avoid making conclusions about the stability of linear,
time-varying systems based on the eigenvalues at any point in time. The following
statements are true of time-varying systems:
• Time-varying systems whose “frozen-time” eigenvalues, i.e.,
eigenvalues computed at any single instant, all have negative real parts
are not necessarily stable.

• If the eigenvalues of the matrix A( t )  A  ( t ) have real parts that are

always negative, the system is asymptotically stable (this is a sufficient
but not necessary condition).

• If all the eigenvalues of A( t )  A  ( t ) are always positive, the system is

unstable [2].
• A time-varying system with one or more eigenvalues whose real part is
positive is not necessarily unstable.
• If a time-varying system’s eigenvalues have real parts such that
Re(  )    0 for all  and all time t, and the system is sufficiently
“slowly varying,” then the system will be asymptotically stable. By
slowly varying, we mean that there exists a value    such that
A ( t )   that ensures this stability. Likewise, a slowly varying system
with an eigenvalue in the right half plane that never crosses the
imaginary axis can be shown to be unstable.
It is not easy to prove all these statements without the theory presented in the
next section. We will demonstrate one of these results using an example.

Example 7.2: An Unstable System with Negative Eigenvalues

Determine the eigenvalues and closed-form solution to the system given by

LM x OP  LM
1 e2tOP L x O
M P
1

Nx Q MN 0  PQ N x Q
(7.17)
2 2

where   0 .

Solution:
From its triangular form, we know that the eigenvalues are  1   2   . In
order to solve the system, we can recognize that the variable x 2 ( t ) is decoupled
280 Part II. Analysis and Control of State Space Systems

bg
from x1 ( t ) and has the solution x 2 ( t )  e t x 2 0 . Then the first equation can
be written as

x1 ( t )  x1 ( t )  e t x 2 ( 0)

This equation can be shown (how?) to have the solution

LM
x1 ( t )  x1 ( 0) 
1 OP
x 2 ( 0) e  t 
1
x 2 ( 0)e t
N 1  Q 1 

which is clearly unbounded because of the second term. This system is therefore
an example of an unstable time-varying system with negative eigenvalues.

7.2 Lyapunov’s Direct Method

A different method for testing the stability of linear, zero-input systems is called
Lyapunov’s direct method. It is based on the concept of energy and dissipative
systems. The physically intuitive explanation is that an isolated system will have
a certain amount of a nonnegative abstract quantity that is similar to energy. If the
system has no input and its energy is always decreasing, we think of it as being
stable. If its energy increases, we think of it as being unstable. However, because
not all state space systems are descriptions of physical phenomena normally
endowed with real energy, we generalize the concept. This is the basis for
Lyapunov’s direct method.
First, we introduce the above physical model with a mechanical example.
Consider the mass-spring-damper system pictured in Figure 7.3. The dynamic
equation that describes this system, in the absence of an input force, is

mx( t )  bx ( t )  kx ( t )  0 (7.18)

The characteristic equation of such a system is therefore ms 2  bs  k  0 .

Knowledge of basic systems would indicate that if the three parameters m, b, and
k are all of the same algebraic sign, then the roots of the characteristic polynomial
will have negative real parts and the system will be stable in the sense of
Lyapunov. Because the three parameters have physical interpretations of mass,
damping, and restoring force constants, we know that all three will be positive. A
state space version of (7.18) can be written by defining x ( t ) and x ( t ) as the two
state variables, giving

LMx(t )OP  LM 0 1 OP Lx(t )O

M P
Nx(t )Q MN k m mPQ N x ( t )Q
b (7.19)
 Chapter 7. System Stability 281

x
k

Figure 7.3 An unforced sliding system with a mass, spring, and damper. The
position x  0 is defined as the position of the mass when the string is
unstretched and uncompressed.

Now consider the energy in the system, which we will call V ( x , x ) :

V ( x , x )  Vkinetic ( x , x )  Vpotential ( x )
(7.20)

1
2
e
mx 2  kx 2 j
To determine how this energy changes over time, we compute

V ( x , x )  mxx
  kxx

Evaluating this energy change “along the trajectory,” i.e., rearranging the original
differential equation and substituting in for 
x,

b g
V ( x, x )  x bx  kx  kxx
(7.21)
 bx 2

If indeed b  0 , then the value of V ( x , x ) will be nonpositive, and will only be

zero when x  0 , i.e., when the system is stationary. If we should encounter a
situation such as an initial condition where x 0  0 and x 0  0 , then we can
check this condition against the invariant subspaces (eigenvectors) of the state
 
space description (7.19). Because x ( 0) x ( 0)  x0 0 is not an invariant
subspace, the system will soon generate a nonzero velocity, and the energy will
282 Part II. Analysis and Control of State Space Systems

continue to decrease until x ( t )  0 and x ( t )  0 , i.e., move toward the

equilibrium position.
Hence, the energy argument provides a valid means of testing this simple
physical system for asymptotic stability. We can also argue that if, in this system,
there were no damping, i.e., b  0 , then the energy is not decreasing but is instead
constant. This situation would describe an oscillator whose energy is continually
being exchanged between kinetic and potential forms (in the mass and spring,
respectively). This would be a system that is Lyapunov stable. The electrical
counterpart would be, of course, an RLC circuit in which the resistance is
removed, leaving energy to bounce back and forth between the capacitor and the
inductor.

7.2.1 Lyapunov Functions and Testing

In order to generalize this stability concept to more abstract, perhaps time-varying
forms, we provide the following theorems. They are stated in terms of generic,
possibly nonlinear systems, and they must apply within an open region
surrounding the origin. We will examine the linear and LTI cases later:

THEOREM: The origin of the linear time-varying system

x ( t )  f ( x ( t ), t ) is Lyapunov stable if there exists a time-
varying function V ( x , t ) such that the following conditions are
satisfied:
1. V ( x , t )   1 ( x )  0 for all x  0 and for all t  t 0 .
V ( x , t )  0 only when x  0 . The function  1 ( x ) is any
continuous, nondecreasing function of the state x, but not
time t explicitly, such that  1 ( 0 )  0.

2. V ( x , t )    2 ( x )  0 for all x  0 and for all t  t 0 , where

V ( x, t ) V ( x, t )
V ( x, t )  f ( x, t )  (7.22)
x t

The function  2 ( x ) is a continuous, nondecreasing

function of the state x, but not time t explicitly, such that
 2 ( 0 )  0. (7.23)

THEOREM: The origin of the system is uniformly Lyapunov

stable if, in addition to the conditions above,  3 ( x )  V ( x , t )
for all x and all t  t 0 . The function  3 ( x ) is a continuous,
 Chapter 7. System Stability 283

nondecreasing function of the state x, but not time t explicitly,

such that  3 ( 0 )  0. (7.24)

THEOREM: The origin of the system is globally Lyapunov

stable if, in addition to conditions 1 and 2 above, V ( x , t ) is
unbounded as x   . (7.25)

If all three theorems hold, the system is globally, uniformly,

Lyapunov stable.

THEOREM: If condition 2 is altered as follows, then stability

in the above theorems is strengthened to asymptotic stability:

2a. V ( x , t )    2 ( x )  0 for all x  0 and for all t  t 0 . (7.26)

In these theorems, condition 1 is equivalent to the function V ( x , t ) being

positive definite, condition 2 is equivalent to V ( x , t ) being negative semidefinite,
and condition 2a is equivalent to V ( x , t ) being negative definite. For linear
systems, these terms are used in the same sense as in Chapter 5, where we
considered the positive- or negative-definiteness of quadratic forms only.
Although we present no proofs of the above theorems, we assert that they
simplify for time-invariant systems. For time-invariant systems (where all
stability is uniform), we have the considerably simpler statements given in the
following theorem. For the purposes of this theorem, we say that any function
V (  ) is positive (semi)definite if V (  )  0 [ V (  )  0 ] for all   0 and
V ( 0 )  0 . A function V (  ) is said to be negative (semi)definite if V (  ) is
positive (semi-)definite.

THEOREM: The origin of a time-invariant system

x ( t )  f ( x ( t )) is Lyapunov stable if there exists a positive-
definite function V ( x ) such that its derivative

dV ( x ) dx dV ( x )
V ( x )   f (x)
dx dt dx

is negative semidefinite. If V ( x ) is negative definite, the origin

is asymptotically stable. (7.27)

In all these definitions and theorems, discrete-time systems can be easily

accommodated. The only difference from continuous-time systems is that instead
284 Part II. Analysis and Control of State Space Systems

of speaking of the derivative V ( x ) or V ( x , t ) , we speak instead of the first

difference: V ( x , k )  V ( x ( k  1), k  1)  V ( x ( k ), k ) .
If the functions V ( x , t ) , or V ( x ) , exist as defined above, they are called
Lyapunov functions. As we shall see, although it can be shown that stable or
asymptotically stable systems do indeed have Lyapunov functions, the theorems
do nothing to help us find them. For nonlinear systems, in particular, finding such
a function in order to prove stability can be a chore for which experience is the
best assistance. The usual procedure is to guess at a Lyapunov function (called a
candidate Lyapunov function) and test it to see if it satisfies the above conditions.
However, for linear systems, Lyapunov functions generally take quadratic forms.

Example 7.3: Lyapunov Function for a Nonlinear System

Use the candidate Lyapunov function

1 2
V ( x1 , x 2 )  x1  x 2 2
2

to test the stability of the system

x1   x1  2 x2 2
(7.28)
x 2  x1x2  x23

Solution:
First we note that the system is time-invariant, so we need not go to the trouble of
finding the functions  1 , etc. Instead, we proceed to the simplified tests, the first
of which is positive definiteness of V ( x ) . Clearly, V ( x )  0 and V ( 0 )  0 , so V
is positive definite. Now we compute

V ( x )  x1 x1  2 x 2 x 2

e j e
 x1  x1  2 x 2 2  2 x 2 x1 x 2  x 2 3 j
  x12  2 x1 x 2 2  2 x1 x 2 2  2 x 2 4
  x12  2 x 2 4

The derivative V ( x ) is therefore negative definite, and we can claim that the
system in (7.28) is asymptotically stable. Furthermore, because V ( x )   as
x   , and the system is time-invariant, we can also say that the system is
globally, uniformly asymptotically stable.
 Chapter 7. System Stability 285

As the first step toward generating Lyapunov functions for linear systems,
reconsider the linear system given in the mass-spring-damper example offered
above. Recall that we chose as an “energy” function the true energy of the
mechanical system

V ( x , x ) 
1
2
emx 2  kx 2 j
This is a positive-definite function of the state of the system. Furthermore, we
computed V ( x , x )   bx 2 , which, if b  0 , is strictly negative for all x  0 , but
may be zero for x  0 . We therefore do not make the claim that V is negative
definite. Rather, we implicitly resorted in that section to LaSalle’s theorem (see
[12] and [13]), which is useful in many similar situations and which justifies our
claim of asymptotic stability in this example.

THEOREM (LaSalle’s theorem): Within the region of the state

space for which the derivative of the candidate Lyapunov
function is such that V ( x )  0 , let Z denote the subset of the
state space in which V ( x )  0 . Within Z, let M denote the
largest invariant subset. Then every initial state approaches M,
even if V is not positive-definite. (7.29)

As we apply this theorem to the mass-spring-damper problem, we note that

Z  {( x , x ) | x  0} . The only invariant subspaces are the eigenvectors e1 and
e2 . These invariant sets intersect with Z only at the origin, so M is just the origin.
Therefore, as we concluded, the origin is asymptotically stable.

7.2.2 Lyapunov Functions for LTI Systems

In the case of linear systems, it is sufficient to consider quadratic forms as
Lyapunov functions. This is largely because for linear systems, stability implies
global stability, and the parabolic shape of a quadratic function satisfies all the
criteria of the above theorems. We will begin considering the special case of linear
systems with the time-invariant case. LTI systems, when examined with
Lyapunov’s direct method as described here, result in some special tests and
equations.
Consider the LTI system

x  Ax (7.30)

and the candidate Lyapunov function V ( x )  x  Px , where matrix P is positive

definite. Then testing the system (7.30), we compute
286 Part II. Analysis and Control of State Space Systems

V ( x )  x  Px  x  Px
 ( Ax )  Px  x  P( Ax )
(7.31)
 x  A  Px  x  PAx

e
 x  A  P  PA x j
To satisfy the theorem, then, it is necessary that the matrix A  P  PA be negative
definite for asymptotic stability, and negative semidefinite for Lyapunov stability.
That is, for some positive (semi)definite matrix Q, it is sufficient to show that

A  P  PA   Q (7.32)

to demonstrate stability of the homogeneous system in (7.30). Equation (7.32) is

lyap(A,C) known as the Lyapunov equation.M
However, one must be cautious in the application of this test because if the
quantity A  P  PA is not negative (semi)definite, nothing should be inferred
about the stability. Recall that Lyapunov’s direct method asserts that if a
Lyapunov function is found, then the system is stable. It does not say that if a
candidate Lyapunov function such as x  Px fails, the system is not stable. So if
we choose a positive-definite matrix P and we compute a Q that is indefinite, we
have shown nothing. (If Q is negative definite, however, instability can be
shown.)
Instead, we resort to the reverse process: We select a positive (semi) definite
Q and compute the solution P to the Lyapunov equation (7.32). This result is
stated as a theorem:

THEOREM: The origin of the system in (7.30) described by

matrix A is asymptotically stable if and only if, given a positive
definite matrix Q, the matrix Lyapunov equation (7.32) has a
solution P that is positive definite. (7.33)

PROOF: (Sufficiency) If P and Q are both positive definite,

then it has already been shown by Lyapunov’s direct method
that the system will be asymptotically stable, because x  Px
will have been demonstrated to be a Lyapunov function.

(Necessity) If the origin is known to be a stable equilibrium and

Q is positive definite, we will prove necessity by first proving
that P is the unique, positive definite result of
 Chapter 7. System Stability 287

z


P  e A t Qe At dt (7.34)
0

To demonstrate this, we substitute (7.34) into the left-hand side

of (7.32):

z z
 
 
A  P  PA  A  e A t Qe At dt  e A t Qe At A dt
0 0

z LMN OP

 A t 
 Qe At  e A t Qe At A dt
0
A e
Q
LM
d e A t Qe At
 OP
 z

N Q dt (7.35)
0
dt
 
 e A t Qe At
0
 0Q
 Q

where convergence of the integral is guaranteed by the fact that

the eigenvalues of A are in the left half of the complex plane.

To show that this solution is unique, suppose that there were a

second solution P such that A  P  PA
   Q . Then we would
have

( A  P  PA)  ( A  P  PA
 )0

implying

e j e
A  P  P  P  P A  0 j
t
Multiplying this equation by e A from the left and by e At
from the right,
288 Part II. Analysis and Control of State Space Systems


e j 
e
e A t A P  P e At  e A t P  P Ae At  j
d A t
P  P e At LM e j OPQ
dt
e
N
0

Now integrating this equation over [0,  ) ,

z

0  d eALM e P  P je OP dt  e e P  P je
t At A t At


0
N Q 0

 0  e P  P j

 P  P

so we must have P  P .

Finally, to demonstrate that P is positive definite, consider the

quadratic form in terms of an arbitrary vector  :

LM e OP
MNz

A t
  P    Qe At dt 
0 PQ (7.36)

z

 A t
  e Qe  dt
At

Because matrix Q is positive definite, the integrand of (7.36) is

positive definite, so

z


  e A t Qe At  dt  0
0

for all   0 . Therefore, P will be positive definite and hence,

will define a valid Lyapunov function. This therefore
constitutes a constructive proof of the theorem.

We note that this theorem does not directly extend to Lyapunov stability; that
is, choice of an arbitrary positive semidefinite Q will not necessarily generate a
positive definite P for a system that is Lyapunov stable. However, by an extension
of LaSalle’s theorem above, we can provide the following generalization:
 Chapter 7. System Stability 289

THEOREM: The origin of the system (7.30) described by

matrix A is asymptotically stable if and only if, given a positive
semidefinite matrix Q, such that x  Qx is not identically zero
when evaluated on a nonzero trajectory of (7.30), the matrix
Lyapunov equation (7.32) has a solution P that is positive
definite. (7.37)

The proof of this theorem, while not given explicitly here, results from the
application of LaSalle’s theorem.

Example 7.4: Stability Test for a Parameterized System

An LTI system is given by

x 
LM3k 3k OP x (7.38)
N 2k 5k Q

Use the Lyapunov stability theorem to find a bound on k such that the origin of
this system is asymptotically stable.

Solution:
While the Lyapunov stability theorem is not commonly used for stability testing
(because it is much easier to simply test the eigenvalues), it can be used to
determine stability bounds on matrix parameters such as this one. Because the
theorem says that any positive definite matrix Q will suffice, it is sufficient to
choose Q  I . Then because we know that the required solution P will have to
be positive definite and, hence, symmetric, we construct the equation

LM3k 2k OP LM p
11 p12 OP  LM p
11 p12 OP LM3k 3kOP  LM1 0OP
N 3k 5k Q N p 12 p22 Q Np 12 p22 Q N 2k 5k Q N0 1 Q
Multiplying out the left-hand side, we arrive at three independent scalar equations
(the two off-diagonal scalar equations will be identical):

3kp11  2kp12  3kp11  2kp12  6kp11  4kp12  1

3kp12  2kp22  3kp11  5kp12  3kp11  8kp12  2kp22  0
3kp12  5kp22  3kp12  5kp22  6kp12  10kp22  1

As these constitute three linear equations in three unknowns, they can be readily
solved as
290 Part II. Analysis and Control of State Space Systems

19 7 15
p11  p12  p22 
72k 48k 80k

To verify the positive definiteness of this solution, we form the determinant

19 7 LM19 7 OP
72k 48k 1 M18 12
PP
b4k g MM 7
P 
15 P
2
7 15
48k 80k MN12 20 PQ

To ensure that both leading principal minors are positive, we can choose k  0 .
Therefore, we may conclude that the origin of the system is asymptotically stable
for all k  0 .

Example 7.5: Discrete-Time LTI Lyapunov Equation

Derive the Lyapunov equation for a discrete-time homogeneous LTI system:

x ( k  1)  Ax ( k ) (7.39)

Solution:
Assuming that we can again use a quadratic form as the candidate Lyapunov
function:

V x ( k )  x  ( k ) Px ( k ) (7.40)

we can find the first difference as

V ( x , k )  V x ( k  1)  V x ( k )
 x  ( k  1) Px ( k  1)  x  ( k ) Px ( k )
 x  ( k ) A PAx ( k )  x  ( k ) Px ( k )

e j
 x  ( k ) A PA  P x ( k )

So the term in parentheses must be a negative definite matrix if the quadratic form
in (7.40) is to be a valid Lyapunov function. The discrete-time Lyapunov
dlyap(A,Q) equationM is therefore

A  PA  P   Q (7.41)
 Chapter 7. System Stability 291

7.2.3 Unstable Systems

It is a common misconception that the theory of internal stability, and in particular
Lyapunov stability, is limited in its ability to conclude that a system is unstable,
i.e., not stable in the sense of Lyapunov. However, there are some results that one
may use to label a system (actually, its equilibium) as unstable:

THEOREM: The origin of the linear system in (7.7) is unstable

if there exists a function V ( x , t ) and a continuous
nondecreasing function of the state only,  ( x ) , such that
 ( x )  V ( x , t ) for all x and t  t 0 ,  ( 0 )  0 , and the following
three conditions are satisfied:

1. V ( 0, t )  0 for all t  t 0 .

2. V ( x , t 0 )  0 at any point arbitrarily close to the origin.

3. V ( x , t ) is positive definite in an arbitrary region around the

origin. (7.42)

Note that the function  ( x ) is the same one that appears in the original
theorems on Lyapunov stability, e.g., conditions 1 or 2. Its existence is
unnecessary for time-invariant systems. We can illustrate the process on a trivial
example.

Example 7.6: An Unstable System

Determine a function V ( x , t ) that, according to the theorem above, demonstrates
the instability of the following system:

x 
LM2 0 OP x
N0 1Q
Solution:
It is clear from the self-evident eigenvalues that the system will not be Lyapunov
stable. To illustrate the existence of V ( x , t ) consider

V ( x, t )  x12  x 22  x 
LM1 0 OP x  x
 
Px (7.43)
N0 1Q
One should be careful not to refer to such a function as a Lyapunov function,
292 Part II. Analysis and Control of State Space Systems

because it is not; it is clearly not positive definite. However, it is zero at the origin,
and is positive arbitrarily near the origin [consider approaching the origin from
the ( x1 ,0 ) direction]. Now compute

V ( x , t )  2 x1 x1  2 x 2 x 2
b g
 2 x1 2 x1  2 x 2  x 2 b g
 4 x12  2 x 22

which is clearly positive definite. The system has therefore been shown to be
unstable.

7.3 External Stability

In each of the stability definitions, theorems, and examples above, linear systems
were considered in the absence of any input. However, any student of systems
and controls knows that the input generally has an effect on the qualitative
behavior of the output. From this perspective on the system, i.e., the input/output
or external behavior, we will need new definitions and tests for system stability.
Such tests may disregard the behavior of the internal dynamics (i.e., the state
vector) altogether.

7.3.1 Bounded Input, Bounded Output Stability

Our working definition for external stability will use the concept of bounded
input, bounded output (BIBO) stability. We will consider the boundedness of the
input function u( t ) and the output function y ( t ) . For a given function of time
f ( t ) , we consider the function to be bounded if for all time t, there exists a finite
constant M such that f ( t )  M . We only consider the stability of systems of
the general form

x ( t )  A( t ) x ( t )  B( t )u( t )
(7.44)
y ( t )  C( t ) x ( t )  D( t )u( t )

We further restrict our attention to systems wherein A( t ) and B( t ) are

themselves bounded operators (otherwise instability might be obvious).

BIBO Stability: The system in (7.44) is BIBO stable if for any

bounded input u( t ) , u( t )  M , and for any initial condition
x 0  x ( t 0 ) , there exists a finite constant N ( M , x 0 , t 0 ) such
that y( t )  N for all t  t 0 . (7.45)
 Chapter 7. System Stability 293

One must be careful when applying this definition. If any bounded input results
in an unbounded output, then the system is considered not BIBO stable. Some
texts will provide the definition of so-called “bounded input, bounded state”
(BIBS) stability, which we give here for completeness, but which we will not
discuss in much depth:

BIBS Stability: The system in (7.44) is BIBS stable if for any

bounded input u( t ) , u( t )  M , and for any initial condition
x 0  x ( t 0 ) , there exists a finite constant N s ( M , x 0 , t 0 ) such
that x ( t )  N s for all t  t 0 . (7.46)

Testing for BIBO Stability

To investigate the conditions under which a system is BIBO stable, we first
reconsider the expression for system output:

z
t
y( t )  C( t )( t , t 0 ) x( t 0 )  C( t )( t ,  ) B(  )u(  )d  D( t )u( t ) (7.47)
t0

where, of course, ( t ,  ) is the state-transition matrix discussed in Chapter 6. In

order to move the feedthrough term into the integrand, we can exploit the Dirac
delta function to achieve

z
t
y( t )  C( t )( t , t 0 ) x( t 0 )  C( t )( t ,  ) B(  )  ( t   ) D(  ) u(  ) d
t0
(7.48)

This allows us to refer to the bracketed term


H ( t ,  )  C( t )( t ,  ) B(  )  ( t   ) D(  ) (7.49)

as the impulse-response matrix. Now considering the initial condition x ( t 0 ) to

be the result of an infinite history of past inputs, we use the expression

z
t
y( t )  H ( t ,  )u(  ) d

294 Part II. Analysis and Control of State Space Systems

in the following theorem:

THEOREM: The linear system in (7.44) is BIBO stable if and

only if there exists a finite constant N H such that

z
t
H ( t ,  ) d  N H (7.50)


PROOF: We will prove this theorem by initially considering

the scalar case wherein we have a scalar impulse-response
matrix denoted h( t ,  ) and h( t ,  )  h( t ,  ) .

(Sufficiency) Suppose input u( t ) is bounded by finite scalar

M. Then we have

z z
t t
y( t )  h( t ,  ) u(  ) d  M h( t ,  ) d
 

Then if scalar N H exists such that

z
t
h( t ,  ) d  N H (7.51)


then clearly

y ( t )  MN H  

and BIBO stability is demonstrated.

(Necessity) If the system is BIBO stable, then for any bounded

input u( t ) , there will exist a finite constant N such that

y(t )  N (7.52)

To construct an argument by contradiction, suppose that (7.50)

were not true, i.e., for any finite scalar N H
 Chapter 7. System Stability 295

z
t
h( t ,  ) d  N H (7.53)


Because we have asserted BIBO stability, we can choose our

input function, so long as it is bounded. We select

u( t )  sgn h( t ,  )

Then we have

z z
t t
y( t )  h( t ,  )u(  ) d  h( t ,  ) d  N (7.54)
 

from (7.52). Yet this result contradicts (7.53), so necessity is

demonstrated.

The extension of this proof for sufficiency to the vector and

multivariable case is straightforward. The necessity part is
based on the argument that if a matrix operator is unbounded,
then at least one of its individual entries must also be
unbounded. This is used to justify a term-by-term application
of the proof above to the impulse-response matrix [15].

When using a result such as this, we should remember that the matrix norm
may be computed in a number of ways, as discussed in Section 3.2.1. One of the
most common techniques is to use the singular value decomposition, recalling
that M   1 , i.e., the largest eigenvalue of M  M .

BIBO Stability for Discrete-Time Systems

As one might expect from the development of the BIBO stability definition, test,
and proof above, the test for discrete-time systems is analogous to that for
continuous-time systems. In order to state this result, we must determine an
expression for the impulse-response matrix that is similar to (7.49) for
continuous-time. Recalling Equation (6.49), we can construct the output of the
generic (i.e., possibly time-varying) discrete-time system as
296 Part II. Analysis and Control of State Space Systems

F A (i)I x( j )
k 1
y ( k )  Cd ( k ) GH  JKi j
d

L O k 1
  C ( k )M A ( q )PB( i  1)u( i  1)  D ( k )u( k )
k

i  j 1 MN d
PQ qi
d d

or, considering the initial condition x ( j ) to have arisen from an infinite past
history of inputs,

L k 1 O
 Cd ( k )M Ad (q )PB(i  1)u(i  1)  Dd ( k )u( k )
k
y( k ) 
i  MN qi PQ
Performing the same trick of moving the feedthrough term into the summation by
introducing the discrete-time Dirac delta function,

y( k ) 
k
|R L k 1 O
 SCd ( k )M Ad (q )P B(i  1)  (i  k  1)Dd (i  1)Vu(i  1)
|U
i  T| MN qi PQ W|
(7.55)

Now the term in braces in (7.55) can be identified as the discrete-time impulse-
response matrix, which we will denote H d ( k , i ) .

LM A (q)OP B(i  1)  (i  k  1)D (i  1)

k 1

MN PQ
H d ( k , i )  Cd ( k ) d d (7.56)
qi

With this notation, we can state the result as follows:

THEOREM: The discrete-time linear state space system

x( k  1)  Ad ( k ) x( k )  Bd ( k )u( k )
y( k )  Cd ( k ) x( k )  Dd ( k )u( k )

with impulse-response matrix H d ( k , i ) as given in (7.56) is

BIBO stable if and only if there exists a finite constant N d such
that
 Chapter 7. System Stability 297

k
 H d ( k, i)  N d (7.57)
i 

for all k.

7.3.2 BIBO Stability for Time-Invariant Systems

Consider a state space system described by (7.44) that is time-invariant, i.e., one
for which the system matrices are not explicit functions of time. Then, recalling
the definition of the impulse-response matrix (7.49), we have


H ( t ,  )  Ce A( t   ) B  ( t   ) D (7.58)

Then, by the theorem giving the condition for BIBO stability, we see that the LTI
system will be BIBO stable if and only if there exists a finite constant N H such
that

z z
t t
H ( t ,  ) d  Ce A( t   ) B  ( t   ) D d  N H
 

The first remark we must make about such a system is that for testing purposes,
the D-matrix is often not included in the integrand. Its contribution to the integral
is clearly additive, and its integral

z
t
( t   ) D d  D


will clearly be unbounded if and only if D itself is infinite. Such infinite

feedthrough–gain systems rarely occur, and if they do, the reason for the overall
system instability would be obvious (recall that we are considering only bounded
matrices anyway).
It remains therefore to test

F e I
z GH z
t t


Ce A( t   ) B d  C

A( t   )
JK
d B (7.59)

for boundedness. Realizing that e A( t   )  ( t ,  ) for an LTI system, the middle

factor in (7.59) is reminiscent of the test for Lyapunov stability given in (7.8).
However, Lyapunov stability and BIBO stability of LTI systems are not
298 Part II. Analysis and Control of State Space Systems

equivalent. The reason for this is that one cannot neglect the effects of
premultiplication by C and postmultiplication by B in (7.59). Suppose, for
example, that

F e I
GH z
t
A( t   )
JK
d B  N C bg (7.60)


where N( C ) is the null space of C. Then no matter what the size of the integral,
Equation (7.60) would imply the boundedness of (7.59). Therefore, it would be
possible for a system that is not asymptotically (or even Lyapunov) stable to
indeed be BIBO stable. This applies to time-varying systems as well. However,
the reverse is true, as it is relatively easy to show that asymptotic stability implies
BIBO stability, i.e., boundedness of the state does imply boundedness of the
output (assuming bounded system matrices).
This leads to a relatively simple conclusion about the concept of BIBS
stability. One can readily demonstrate that a linear system is BIBS stable if, for
all bounded inputs, there exists a finite constant N s such that

z
t
( t ,  ) B(  ) d  N s (7.61)
t0

for all t.
A better way to approach BIBO stability is in the frequency-domain. One can
show that for an LTI system,

e At  L1 sI  A {b g }
1

(see Exercise 6.6). Therefore, demonstration of the boundedness of (7.59) can be

achieved by considering the frequency-domain counterpart,

{b
CL1 sI  A g }B
1
(7.62)

Boundedness of (7.62) is therefore equivalent to the condition of having the poles

of the expression in (7.62) in the closed left-half plane.* However, providing for
the existence of a bounded input, which might have characteristic frequencies on
the imaginary axis, we must restrict the poles of (7.62) to the open left-half plane.

* We are assuming here that the reader is familiar with the relationship between the
characteristic frequencies of a signal and the qualitative behavior of its time-domain
representation.
 Chapter 7. System Stability 299

Otherwise, the characteristic frequencies of the output y ( t ) might be repeated on

the imaginary axis and therefore produce an unbounded result. By taking the
LaPlace transform of the LTI system equations, it is relatively easy to demonstrate
that the transfer function of a system (i.e., the LaPlace transform of its impulse-
response) is given by

b
H ( s )  C sI  A g 1
BD (7.63)

(see Section 1.1.5). Therefore, if the matrix D is bounded, our condition for BIBO
stability can be stated as follows: An LTI continuous-time system is BIBO stable
if and only if the poles of its (proper) transfer function are located in the open
left-half complex plane. In the case of an LTI discrete-time system, this statement
should be adjusted so that the poles must lie inside the open unit circle. Of course,
by speaking of the transfer function (singular), we are implying a SISO system.
However, the above statements apply if we add the further condition that a MIMO
system is BIBO stable if and only if each individual input-to-output subsystem
transfer function is BIBO stable.
Readers familiar with the Routh-Hurwitz method for testing stability of a
system may now consider this test in the context of the stability properties
discussed in this chapter [4]. Conventional Routh-Hurwitz testing determines
BIBO stability. Extended Routh-Hurwitz testing, i.e., with additional analysis of
the “auxiliary equation” in the case that a row of the Routh-Hurwitz array is
entirely zero, has the further ability to determine so-called “marginal stability.”
Marginal stability corresponds to stability that results when the output is
guaranteed to be bounded whenever the input asymptotically approaches zero,
i.e., when the input has no characteristic frequencies with zero real parts.

7.4 Relationship Between Stability Types

Given the discussions of the various kinds of stability in this chapter, the
relationship between different types of stability can be summarized by the Venn
diagram in Figure 7.4. Although it is easy to envision systems that are Lyapunov
stable but not BIBO stable, e.g., systems with poles on the imaginary axis, it is
sometimes not obvious how a system can be BIBO stable, but not asymptotically
stable, or even Lyapunov stable. Such a system is given in the following example.

Example 7.7: A BIBO, but Not Asymptotically Stable System

Consider the system given by the differential equation

y  y  2 y  u  u (7.64)

Show that this system is BIBO but not asymptotically stable.

300 Part II. Analysis and Control of State Space Systems

Solution:
First we will put the system into state variable form. From the simulation diagram
given in Figure 1.7, we can derive a state variable model as follows:

x 
LM0 1OP x  LM0OPu
N2 1Q N1Q (7.65)
y  1 1 x

Finding the eigenvalues of this system, we determine that ( A)  1,2 . From l q

these eigenvalues, we see that this system is neither asymptotically nor Lyapunov
stable. We will now check to see whether it is BIBO stable. It is an easy exercise
to check that

L e t  13 e 2t 1 t
 13 e 2t OP
M
2
3 3
e
MM PP
At
e
N 2
3
e t  23 e 2t 1 t
3
e  23 e 2t Q
and that

Ce At B  e 2t

Therefore,

z z
t t
1
Ce A( t   ) B d  e 2( t   ) d 
 
2

which is obviously bounded. The system is therefore BIBO stable.

As a further exercise, consider finding the transfer function from the original
differential equation in (7.64). If the initial conditions are zero, we have

( s 2  s  2)Y ( s )  ( s  1)U ( s )

giving

Y ( s) ( s  1) ( s  1) 1
 2  
U ( s ) ( s  s  2) ( s  2)( s  1) ( s  2)

The transfer function shows poles in the open left-half of the complex plane, so
again we have demonstrated that the system is BIBO stable.
 Chapter 7. System Stability 301

BIBO
BIBS

Lyapunov asymptotic
stable
exponential

Figure 7.4 Venn diagram showing the relationship between stability types.

7.5 Summary
Before reading this chapter, one may already have formed an intuitive concept of
system stability: A system is generally regarded as stable if its response remains
finite. Although this interpretation often suffices in the sense that unstable
systems are easy to recognize when they are encountered (we know one when we
see one), the mathematically rigorous definitions for stability are more complex.
Particularly, in the case of nonlinear or time-varying systems, we can encounter
systems that are stable in the absence of inputs but unstable with them (Lyapunov
but not BIBO stable); systems that are unstable when the internal dynamics (i.e.,
the states) are examined, but stable if only the output is of concern (BIBO but not
Lyapunov stable); or systems that are stable for some initial conditions and/or
initial times, but not from other initial states.
These distinctions are important for the study of newly modeled systems, for
it is now clear that observation of one characteristic of a system or of one example
of its behavior is not sufficient to characterize the stability of the system in
general. A linear system that is designed and placed into a practical application
should be assured of stability under all of its expected operating conditions.
External stability, while usually desirable, is often insufficient for practical
purposes. Usually, if internal states exceed their normal levels, the linear model
upon which the analysis is performed will begin to fail, and all predictions of
system performance will become invalid. Rarely, when internal states are
conceptual abstractions, the fact that a system is internally not stable may not be
important so long as the desired external behavior is stable.
The job of the engineer in system design is often predicated upon an analysis
of the system’s stability requirements. While undergraduate texts are fond of
302 Part II. Analysis and Control of State Space Systems

providing the many classical methods for achieving various performance

specifications, real-life applications can be much more difficult. In complex
systems, the goal of the designer is simply to achieve stability.
Tools for this purpose introduced in this chapter include the following:
• Stability in the sense of Lyapunov characterizes the system’s state
response in the absence of inputs, i.e., to initial conditions alone. For LTI
systems, this type of stability is tested by examining the eigenvalues of
a system to see that they are in the left-half plane (continuous-time
systems) or inside the unit circle (discrete-time systems). For time-
varying systems, as usual, the test is neither as simple to state nor to
perform, and one should be careful not to reach erroneous conclusions
based on the “frozen-time” eigenvalues.
• BIBO stability describes the behavior of the system in response to
bounded inputs. Thus, BIBO stability is fundamentally different from
Lyapunov stability, which neglects input. We have seen an example in
which a system that is BIBO stable results from a state variable model
whose zero-input counterpart is not Lyapunov stable. We will further
explore this phenomenon after the following chapter on controllability
and observability. As with Lyapunov stability, LTI systems prove to be
much easier to test for BIBO stability than time-varying systems.
• Perhaps the most powerful tool introduced in the chapter is Lyapunov’s
direct, or second method.* This tool is particularly helpful in the study
of nonlinear systems, for which there are relatively few other general
stability analysis methods available. Lyapunov’s method, wherein we
construct so-called “Lyapunov functions,” can be adapted to assist us in
designing controllers, placing bounds on system parameters, and
checking the sensitivity of a system to its initial conditions. We will
provide more examples of its uses in the exercise problems.

7.6 Problems
7.1 Prove that the equilibrium of the LTI system x  Ax is asymptotically
stable if and only if all the real parts of the eigenvalues of A are negative.
Hint: Use Jordan forms.

7.2 Draw (or program a computer to draw) a phase portrait for the system in
Example 7.3, which has the following nonlinear dynamic equations

* Lyapunov’s first method is known as the method of exponents. It and Lyapunov’s

linearization method [18] are used exclusively for nonlinear systems and thus are not
explored in detail here.
 Chapter 7. System Stability 303

x1   x1  2 x2 2
x 2  x1x2  x23

7.3 A certain real-valued nonlinear system is given by the equations

x1   x1  x2  x1( x12  x22 )

x 2   x1  x2  x2 ( x12  x22 )

a) Determine the equilibrium points of the system.

b) Choose a candidate Lyapunov function and use it to determine the
bounds of stability for the system. Sketch the region in the x1  x 2
plane in which stability is guaranteed by the Lyapunov function.
c) Draw a phase portrait for the system.

7.4 Determine whether the origin of the following system is Lyapunov stable:
x  tx .

7.5 An LTI system is described by the equations

x 
LMa 0 OP x
N1 1Q
Use Lyapunov’s direct method to determine the range of variable a for
which the origin is asymptotically stable.

7.6 An LTI system is described by the equations

x( k  1) 
LM1  a 1OP x( k )
N 0 1Q
Use Lyapunov’s direct method to determine the range of variable a for
which the origin is asymptotically stable.

7.7 Show that a sufficient condition for a linear time-varying system to be

asymptotically stable is for the matrix A  ( t )  A( t ) to have all its
eigenvalues in the open left-half of the complex plane.
304 Part II. Analysis and Control of State Space Systems

7.8 Find a Lyapunov function that shows asymptotic stability for the system
x  (1  t ) x .

7.9 Show that matrix A has eigenvalues whose real parts are all less than 
if and only if given an arbitrary positive definite matrix Q, there exists a
unique, positive definite solution P to the matrix equation:

A  P  PA  2P   Q

7.10 Determine whether the following system is BIBO stable: x   x t  u .

7.11 Prove that an asymptotically stable system is by necessity also BIBO

stable.

7.12 Prove the sufficiency of the condition given in Equation (7.45) for BIBS
stability, i.e., that there exists a finite constant N s such that

z
t
( t ,  ) B(  ) d  N s
t0

7.13 Determine whether the system shown in Figure P7.13 is BIBO stable.
Consider the input to be the current u( t ) , and the output to be the voltage
v(t ) .

u (t ) L C R v (t )

P7.13
 Chapter 7. System Stability 305

Repeat if R   (i.e., remove the resistor).

7.14 Consider a system given by the matrix

A
LM0 0OP
N 0 0Q
Determine all of the equilibrium points. Are they Lyapunov stable? If
b  [1 1] , c  [1 0] and d  0 , is the system BIBO stable? If instead
c  [0 0] , is the system BIBO stable?

7.15 Consider the system given by

x 
LM2 0OP x  LM1OPu
N 0 0Q N1Q
y  1 2 x

a) Find all equilibrium solutions x e .

b) Determine which equilibria are asymptotically stable.

c) Determine if the equilibrium solutions are Lyapunov stable.
d) Determine if the system is BIBO stable.
e) Let z1  x1 , z 2   x1  x 2 , and u( t )  0 . If we denote
  , find the equilibrium solutions z and
z  [ z1 z 2 ] and z  Az e
sketch them on the 1 z  z 2 plane.

f) Draw, on the same plane, a phase portrait.

7.16 For each of the following system matrices for systems of the form
x  Ax  Bu , determine a bounded input that might excite an unbounded
output.
306 Part II. Analysis and Control of State Space Systems

LM1 3 9 0 0 OP
MM2 3 0 0 0P LM1 2 3 OP
A 0 0P A  M4 6P
a)
MM0 2 0
0 0
0
0 3P
P b)
MN4
5
5 6PQ
MN0 0 0 3 0PQ

(Hint: Very little computation is necessary. Note that the B -matrix is

undetermined in each case.)

7.17 For the three systems given below, determine the stability (i.e., Lyapunov,
asymptotic, or BIBO).
LM 1 0 OP LM 1 OP
x ( k  1) 
a) MN0.5 0.5Q
P x( k )  MN1PQu( k )
y( k )  5 5 x( k )

LM7 2 6OP LM1 1 OP

x  M 2 3 2P x  M1 1Pu
b) MN2 2 1PQ MN1 0 PQ
yM
L1 1 2 OP x
N 1 1 1Q
c) AM
L 2 5OP b  LM 1 OP c  1 1
N4 0Q N1Q
7.18 Given the system

LM25 45 55 OP LM 2 OP

M
x  M 5 15
P M P
15P x  M 0.2 Pu
MM P MM PP
N 15 35 35PQ N0.8Q
y 1 0 0x
 Chapter 7. System Stability 307

determine whether the system is Lyapunov stable, asymptotically stable,

and/or BIBO stable. If it is not BIBO stable, give an example of a bounded
input that will result in an unbounded output.
308 Part II. Analysis and Control of State Space Systems

7.7 References and Further Reading

Although stability for time-invariant systems is relatively easy to test,
determining the stability of time-varying and nonlinear systems can be very
difficult. General texts used to distill the material presented here include [3], [7],
and [15]. An interesting analysis of stability in the sense of Lyapunov is given in
[14], where the authors establish the equivalence between Lyapunov stability and
the stability determined by the more familiar Routh-Hurwitz test. Further ties
between stability types are established in [1] and [16], where the relationships
between BIBO, BIBS, and asymptotic stability are established for time-varying
systems, although the controllability and observability concepts introduced in
Chapter 8 are required. BIBO stability tests based on the impulse-response matrix
were introduced in [16].
The issue of slowly varying systems, as briefly mentioned in Section 7.1.4,
is addressed in [5], [11], and [17].
Lyapunov’s direct method is the topic of much study, owing partially to the
difficulty with which Lyapunov functions are found for time-varying and
nonlinear systems. A good introductory discussion is included in [7], [9], [12],
[13], [18], and [19], with [10] and [20] concentrating on discrete-time systems.
The issue of decrescence, i.e., the importance of the  i -functions described in
the theorems of Section 7.2.1, is explicitly discussed in [7] and [18]. These
functions are critical for generating Lyapunov functions for time-varying systems,
but were not discussed in depth here.
Also useful for nonlinear and time-varying systems are other types of
stability tests, such as the small gain theorems used to determine input/output
stability, as discussed in [6], [12], and [19].

[1] Anderson, Brian D. O., and J. B. Moore, “New Results in Linear System
Stability,” SIAM Journal on Control and Optimization, vol. 7, no. 3, 1969, pp.
398-414.
[2] Brogan, William L., Modern Control Theory, 3rd edition, Prentice-Hall, 1991.
[3] Chen, Chi-Tsong, Linear System Theory and Design, Holt, Rinehart, and Winston,
1984.
[4] D’Azzo, John D., and Constantine H. Houpis, Linear Control System Analysis and
Design: Conventional and Modern, 4th edition, McGraw-Hill, 1995.
[5] Desoer, Charles A., “Slowly Varying System x  A( t ) x ,” IEEE Transactions on
Automatic Control, vol. AC-14, 1969, pp. 780-781.
[6] Desoer, Charles A., and M. Vidyasagar, Feedback Systems: Input-Output
Properties, Academic Press, 1975.
[7] Hahn, Wolfgang, Stability of Motion, Springer-Verlag, 1967.
[8] Kailath, Thomas, Linear Systems, Prentice-Hall, 1980.
 Chapter 7. System Stability 309

[9] Kalman, Rudolph E., and J. E. Bertram, “Control System Analysis and Design Via
the Second Method of Lyapunov: I. Continuous-Time Systems,” Transactions of
the ASME, Journal of Basic Engineering, vol. 82D, June 1960, pp. 371-393.
[10] Kalman, Rudolph E., and J. E. Bertram, “Control System Analysis and Design Via
the Second Method of Lyapunov: II. Discrete-Time Systems,” Transactions of
the ASME, Journal of Basic Engineering, vol. 82D, June 1960, pp. 394-400.
[11] Kamen, Edward W., P. P. Khargonekar, and A. Tannenbaum, “Control of Slowly-
Varying Linear Systems,” IEEE Transactions on Automatic Control, vol. 34, no.
12, 1989, pp. 1283-1285.
[12] Khalil, Hassan, Nonlinear Systems, Macmillan, 1992.
[13] LaSalle, J., and S. Lefschetz, Stability by Liapunov’s Direct Method, with
Applications, Academic Press, 1961.
[14] Puri, N. N., and C. N. Weygandt, “Second Method of Liapunov and Routh’s
Canonical Form,” Journal of The Franklin Institute, vol. 276, no. 5, 1963, pp.
365-384.
[15] Rugh, Wilson, Linear System Theory, 2nd edition, Prentice Hall, 1996.
[16] Silverman, Leonard M., and Brian D. O. Anderson, “Controllability,
Observability, and Stability of Linear Systems,” SIAM Journal on Control and
Optimization, vol. 6, no. 1, 1968, pp. 121-130.
[17] Skoog, Ronald A., and Clifford G. Y. Lau, “Instability of Slowly Time-Varying
Systems,” IEEE Transactions on Automatic Control, vol. AC-17, no. 1, 1972, pp.
86-92.
[18] Slotine, Jean-Jacques, and Weiping Li, Applied Nonlinear Control, Prentice Hall,
1991.
[19] Vidyasagar, Mathukumalli, Nonlinear Systems Analysis, 2nd edition, Prentice
Hall, 1993.
[20] Weiss, Leonard, “Controllability, Realization, and Stability of Discrete-Time
Systems,” SIAM Journal on Control and Optimization, vol. 10, no. 2, 1972, pp.
230-251.