M. VIDYASAGAR
VlDYASAGAR, 1947-M
Nonlinear systems analysis.
10 9 8 7 6 5 4
Preface, ix
Introduction, 1
1 .1 General considerations, 1
Second-order systems, 10
2.1 Preliminaries, 10
2.2 Linear systems, 14
2.3 Nonlinear systems, 21
2.4 Periodic solutions and limit cycles, 20
2.5 Two analytical approximation methods, 41
4
3.3 Contraction mapping theorem, 73
3.4 Nonlinear differential equations, 78
3.5 Solution estimates for linear equations, 88
Index, 297
Preface
ix
X Preface
Montreal M. VlDYASAGAR
Notes to the Reader
6 = tan
-1 —
*i
1.1
GENERAL CONSIDERATIONS
Nonlinear physical systems, that is, systems that are not necessarily
linear, differ from two important respects:
linear systems in
1. Generally speaking, one can usually obtain closed-form
expressions for solutions of linear systems, whereas this is not
always possible in the case of nonlinear systems. More often,
one is forced to be content with obtaining sequences of approxi-
mating functions that converge to the true solution or with
generating estimates for the true solution. As a result, one may
not have a good “feel” for what makes a nonlinear system
“tick,” compared with a linear system.
2. The analysis of nonlinear systems generally involves mathe-
matics that is more advanced in concept and more messy in
detail than is the case with linear systems.
A mathematical model that describes a wide variety of physical
nonlinear systems is an rcth-order ordinary differential equation of the
type
-l
d ny(t) d n y(t)_
dt
n
=h t, y(t), y(t\ '
-
dr 1
u(t) t> 0
where t is the time parameter, u(-) is the input function (the terms
control function and forcing function are also used), and y(*) is the
. . .
2 Chap. 1 Introduction
2 x^t) = y(t)
3 X 2 (t) = j)(0
4 X„(t ) = "- 1
dt
5 xft) =x 2 {t)
6 x 2 (t) =x 3 (/)
7 *„-i(0 = *„(0
8 xff) = Xj(0, * 2 (0> . .
. ,
w(0]
10 f(t, x, w) [x 2 ,
X3 ,
• . . ,
/z(?, Xj, . . . ,
xni w)]
12 = h ^U ^i
mi_1>
J2OX
~dt^ • • • > (0> • • •
• • •
uft), • • • ? Wp(0]> 19 . . . 9 k
where uf *), . . . ,
are the input functions and yf *), . .
,y k (*)
are the output functions. As before, define
Sec. 1.1 General Considerations 3
16 n =m +1
. . . + mk
Then x(f) is a state vector for the system described by (12), and the
system of equations (12) can once again be equivalently represented by a
single first-order vector differential equation of the form
if x >0
20 sign x(t) =
if x <0
It is easy to verify that no continuously differentiable function *(•)
exists such that (19) is satisfied. Thus statement 1 does not hold for this
system.
iThe exception is Chap. 6, where we shall study distributed systems, e.g., sys-
23 xft) = t
112
24 x 2 (t) = -t U2
Thus statement 1 is true, but 2 is false.
26 x(t ) = +x 1
2
(t), t> 0; x(0) =0
Then over the interval [0, 1), this equation has the unique solution
27 x(t) = tan t
suitable sense, of the solutions to (17). This is usually called the question
of stability. Ideally, we would like to know whether or not the solutions
to (17) are well behaved without actually solving the system equations
(17). The stability question is studied in depth in Chaps. 5 and 6.
Finally, as a prelude to these more advanced subjects, we shall
study second-order systems in Chap. 2. As we shall see there, a “geo-
metric” approach to second-order systems yields much intuition and
insight.
1.2
AUTONOMY, EQUILIBRIUM POINTS
In this section, we shall introduce two definitions that are frequently
used in the sequel. Before proceeding to these definitions, we shall clear
up one small point. Many of the definitions, theorems, etc., that follow
are stated for differential equations of the type
Comparing with (17) of Sec. 1.1, 2 we see that in [1.1(17)] the depen-
(1)
dence of the right-hand side on an input u(*) is explicitly identified,
whereas this dependence, if any, is suppressed in (1). This might mislead
one into thinking that [1.1(17)] describes a “forced” system, whereas
(1) describes an “unforced” system. However, this is not necessarily
the case. In problems of system analysis as opposed to optimal control ,
6 Chap. 1 Introduction
x(t) = x0 , V > t t1
The terms stationary point and singular point are also used in place of
equilibrium point.
x (ty
i x 2 (t)
_— i/l) sin X, (t)_
Notice first of all that the systemis autonomous. Next, we have that
•^20 === 0
sin x 10 =0
i.e., the set of equilibrium points of (12) is the set of points in R2 of the
form
3
Notice that we need not specify the time because the system is autonomous.
)
is described by
_ f[r(t)l
m
m lf=r(r)
x(t)
Xi (0 A
|
X0"
x 2 (0- AO.
Then the state equations are
Xi (0 r x 2(t )
~Vf[xmim_
From (19), we see that the set of equilibrium points of this (autonomous)
system consists of all points of the form ( r 0 0), where fir]) 0. There-
,
=
fore this system is in an equilibrium state if the particle has zero velocity
and is at a position where the force is zero, i.e., if the potential energy
is stationary.
than x 0 .
8 Chap. 1 Introduction
X)
25 A(? 0 ) =
dx x=Xo
29 lim
llx-x 0 2-»0 ||
M#
II
X Xq 1 2
=0
r(/ 0 , •) satisfies the condition
31
ll'x^il; ^ T whenever ||x - x 0 1| 2 < d
Such a choice for d is always possible in view of the limit condition (29).
Let N
be the neighborhood of x 0 defined by
32 N = {x G Rn :
II x - x0 || 2 < d]
>c||x — x0 1[ 2 - y||x x || 2
= yllx - x0 || 2
x^x
> 0 whenever 0
x2 = a 2 x 2 + b 2 X\X 2
(a) Show that (0, 0) is an equilibrium point.
(b) Show that (0, 0) is an isolated equilibrium point if and only if both
a x and a 2 are nonzero.
Problem 1.3. Consider the tunnel-diode circuit of Figure 1.1, where
FIG. 1.1
§ = —Gv d -f(vd)
(b) Find all the equilibrium points of this system, when (i) G= 0,
Second-order systems
2.1
PRELIMINARIES
In this chapter, we shall study various specialized techniques that are
available for the analysis of second-order systems. In subsequent chap-
ters, we shall remove this restriction on the order of the system and
study some techniques of analysis that can be applied to systems of any
order. Obviously, the latter techniques are also applicable to second-
order systems. However, second-order systems occupy a special place
in the study of nonlinear systems, for many reasons. The most important
reason is that the solution trajectories of a second-order system can be
represented by curves in the plane. As a result, many of the nonlinear
systems concepts such as oscillations, vector fields, etc., have a simple
geometric interpretation, in the case of second-order systems. (All the
technical terms used above will be defined shortly.) For these and other
reasons, second-order systems, by themselves, have been the subject of
much research, and in this chapter we shall present some of the simpler
results that are available.
In general, a second-order system under study is represented by
two scalar differential equations
state-plane plot. The state plane is the usual two-dimensional plane with
the horizontal axis labeled x and the vertical axis labeled x2 Suppose
1
.
3 jci (0 = *2(0
it is customary to refer to the state plane as the phase plane. Corre-
spondingly, in this case one also refers to phase-plane plots or phase-
plane trajectories. This special case arises quite commonly in practice.
In particular, if the system under study is governed by a scalar differ-
5 *i (0 = y{t)
6 x 2 (t) = y(t)
In this case, the system equation (4) is equivalent to the following two
first-order equations
7 Xj(0 =x 2 (t)
time t 0 (see Fig. 2.1). Typically, t 0 might be the initial time and (x 10 x 20 ) ,
h— *0 + f *Ei
Je *2
where the integral in (9) is taken along 0. If x 2 does change sign along
0, then the integral in (9) has to be evaluated as the sum of several
integrals, one corresponding to each segment of 0 along which x does
2
not change sign (see Fig. 2.2). Note that, as x 2f > 0, the integral (9) —
becomes an improper integral. The proof of the relationship (9) is
easily obtained starting from (7) and is left as an exercise for the reader
(see Problem 2.1).
and f2 in (1)— (2) are continuously differentiable. This ensures (as shown
in Chap. 3) that (l)-(2) have a unique solution at least locally in t.
That is, given the equations (l)-(2) together with some arbitrary initial
conditions
14 Xj(0) =x 10
15 x2 (0) = x 20
Problem 2.1. Prove the relationship (9). Hint: Use (7) to write
x x
(t + At) = Xi(t) + At x 2 (t) + o(At)
Eliminate At as A* — > 0.
constitute a vector field? Justify your answer. (Hint: Consider the behavior of
f near the origin.)
2.2
—
LINEAR SYSTEMS
We shall begin by studying linear systems, which are simpler to analyze
than nonlinear systems and yet provide much insight into the behavior
(0 =a lx x x (f) + a 12 x 2 (t)
*l(0) = ^10
x2 (0) = x 20
x(0 = Ax(t)
:
7 z (0 -
where M is a constant nonsingular 2x2 matrix with real coefficients.
In terms of the transformed variables z, (5)-(6) become
8 zif) = M AMz(Y) -1
z(0) = M x
_1
9 0
Xl 0
10 M-‘AM=
0 2-2
X 1
11 Mr ‘AM =
0 A
a B
H
12 M-‘AM =
~B «
where a + jfi, a — j ft
are the complex conjugate eigenvalues of
A (and we choose ft < 0 to be definite).
We shall study each of these cases in detail.
13 zft) = Aj zft)
14 z 2 (t) = X 2 z 2 (t )
15 z,00) = 2, 0
16 2 (0) = z2Q
2
17 2,(0 = 2,06*“
18 z2 (t) = 220 e*“
16 Chap. 2 Second-Order Systems
19 *2 Z 20
signs, the trajectories are as in Fig. 2.5. The arrowheads in Fig. 2.4
correspond to the case where X 2 < < 0; if k and A 2 are both posi-
x
are illustrated for, respectively, (1) and X 2 of the same sign and (2)
At and A 2 of the opposite sign. In the case where A t and A 2 are both of
the same sign, the equilibrium point (0, 0) is referred to as a stable node
if A and A 2 are both negative and as an unstable node if A and A are
x
2 x
both positive. The rationale is that if A and A 2 are both negative, the
1
21 z2 (t) = Xz 2 (t)
22 z i(0) — ^io
23 z (0) = z 20
2
18 Chap. 2 Second-Order Systems
25 z2 (t ) = z 20 eu
Once again, can be eliminated from (24)— (25) the resulting expression
t ;
28 zf0) = z 10
29 z (0) = z 20
2
Sec. 2.2 Linear Systems 19
30 r= (zj + z 2 1/2
2)
31 6 = tan" ^ 1
32 r(t) = ar(t)
33 0(0 = -fi
which has the solution
34 r(t) =r e 0
at
35 0(0 = -pt
4> 0
z2
FIG. 2.11
20 Chap. 2 Second-Order Systems
TABLE 2.1
Type of
Eigenvalues of A Equilibrium Point
x x (t)= — 2x (0 - x2 (t) +
1
2 v(t)
in i a
AAAAr- -AAAAr
0 1 F iz *\ 1 H
FIG. 2.13
Sec. 2.3 Nonlinear Systems 21
(0, 0), and find the matrix M that transforms the equations into the appro-
priate canonical form.
(i) (ii) A =
-r (iii) A=
’1
r
2_ _0 -i_
(iv) A= (v) A=
-r (vi) A =
"0 -r
o 2 -2
2.3
NONLINEAR SYSTEMS
In this section, we shall study four methods for obtaining the state-
plane trajectories of a second-order autonomous system of the type
1 (0 =/ [* 0 * (01
i i ( > 2
2 ± 2 (t) =f [x (t\x
1 1 1 (t)\
becomes the point (0,0) in the new coordinate system: More precisely, let jci =•
Xl — Xio X2,
= X2 — X 2 Q.
22 Chap. 2 Second-Order Systems
3 On ,
= §Il
dXj xi = 0,X2 = 0
a \\ a 12
4 A
- a l\ a 22_
= a Xi xl + a 12 x + fiOi, x
2 2)
6 f2 {x u x 2 ) = a 21 x l + a 22 x + r (x u x
2 2 2)
where rt and r 2 are the remainder terms. [Note that we have used the
fact that /,(0, 0) =/ 2 ( 0, 0) = 0.] Now, associated with the nonlinear
system (l)-(2), define the linear system
7 = a n 4(0 + a £
4(0 12 2 (t)
Clearly = 0, g = 0 an equilibrium
2 is point of the system (7)-(8),
which is commonly known as the linearization of the system (1)— (2)
around the equilibrium point (0, 0). The linearization method is based
on the fact (proved in Chap. 5) that in most cases the trajectories of the
nonlinear system (1)— (2) have, in some suitably small neighborhood of
the origin, the same qualitative features as the trajectories of the linear
system (7)-(8). Table 2.2 summarizes the situation.
TABLE 2.2
the last entry in the table can be explained as follows If the equilibrium :
Sec. 2.3 Nonlinear Systems 23
11 Xl (t) = y(t)
12 x 2 (t) = y(t)
(10) is transformed into the pair of first-order equations
13 Xl (t) =x 2 (t)
14 x 2 (t) = —x x
(t) + jx[ 1 — x\(t)\x 2 (t)
15 Ut) = Ut)
16 Lit) = j(0 + n$ 2 (t)
The eigenvalues of the associated A matrix satisfy the characteristic
equation
17 A2 — //yl +1=0
Hence, for all positive values of ju<2, the roots of (17) are complex
with positive real parts, so that the equilibrium point Si = 0, £ 2 = 0 of
(15) -(16) is an unstable focus. Referring to Table 2.2, we see that the
equilibrium point x x = 0, x 2 = 0 of (13)-(14) is also an unstable focus.
The Van der Pol equation is studied further in a subsequent example.
The simple method presented below is quite well suited for con-
structing a single trajectory of (1)— (2), starting from a given initial
point, and is especially suitable for implementation on a digital com-
puter. This method actually amounts to a graphical interpretation of
the forward Euler numerical integration procedure and is an example of
i
x x
(t + At) by x 1 + Ax and 1 x 2 and x 2 + Ax 2 Then,
similarly define .
(1 8)-(l 9) to get
20 x2 + Ax 2 = x2 + JiKXu X Ax
2)
=X + 2 s(x u x 2 ) Ax,
where
fi(x i, x 2 )
21 (Yv
s{Xl v ^
’ X2)
=A 7 &^ 2)
the end of the line segment as the new starting point, and repeating the
procedure. There is only one small point to be cleared up. The quantity
s(X}, x 2 ) specifies the slope of the trajectory but does not specify its
direction. But this is easily determined by examining the signs of the
quantities x 2 ) and f2 (x u x 2 ). We can assume without loss of
f1 (x u
generality that not both f (x u x 2 ) and f2 (x u x 2 ) are zero, because if
x
both fx (x l9 x 2 ) and f2 (x l9 x 2 ) are nonzero, one will reach the same con-
clusions regarding the direction of the trajectory at (x l9 x 2 ) whether he
examines the sign of f1 (x 1 x 2 ) or that of f2 (x u x 2 ).
,
l9 x 2 ) corresponds to increasing
is positive, the trajectory direction at (x
*2
j/h< /i <o
0
FIG. 2.15 I
22 x —= X 2
i
23 X2 =f (x „ x )
2 2
26 Chap. 2 Second-Order Systems
24 s(x 1; x2 ) = '
x2
’ —
Thus the first thing to be noticed is that whenever x 2 = 0 we have
^(^i, x 2 )
= oo unless /2 (x 1? x 2) = 0. Geometrically, this means that
whenever a trajectory of (22)-(23) meets the x r axis [at say (x 1? 0)] either
(x 1? 0) is an equilibrium point of the system (22)-(23) or else the tangent
to the trajectory of (22)-(23) at (x l9 0) is vertical. In addition, the direc-
tion of the trajectories of (22)-(23) can be determined by inspection:
Whenever the point (x l5 x2) lies in the first or second quadrants, we
have x 2 > 0, so that the direction of the trajectory of (22)-(23) should
be so chosen as to correspond to increasing x t . Similarly, whenever
(jc 15 x2) lies in the third or fourth quadrants, we have x 2 < 0, so the
direction of the trajectory at (x 1? x 2 ) corresponds to decreasing x If t
.
along which the solution trajectories of (1)— (2) have the slope c. Thus,
whenever a solution trajectory of (1)—(2) crosses the curve defined by
s(x u x 2) = c,
must do so with a slope of c. The procedure is to plot
it
the curve j(x 1; x 2 ) = c in the x 2 -x 2 plane and along this curve draw
“short” line segments having the slope c. Such a curve is known as an
isocline.The directions of these line segments are determined as in Sec.
2.3.2. The procedure is repeated for sufficiently many values of the
constant c, so that the x -x 2 plane is filled with isoclines, and one is
1
then able to rapidly sketch the trajectories of (1)— (2) starting from any
initial point.
Sec. 2.3 Nonlinear Systems 27
26 Example. Consider again the equations of the Van der Pol oscil-
lator, namely,
27 *i = *2
28 X2 = + /^(l - *l)*2
The “slope function” s(x u x 2 ) associated with this system of equations
29
is
*(*,. *,) - =a + ^
The curves x 2 ) = c corresponding to fi = 1 and for various
values of c are shown in Fig. 2.16. With the aid of this construction, one
can easily sketch the trajectory of the Van der Pol oscillator starting
from, say, x^O) = —2, x 2 (0) =
Also one can see that there
3. may
possibly be a periodic solution as indicated by the bold line.
32 x2 = —bx x 1 2
The system (31)— (32) is a very elementary model for the spreading of
disease within a population. Here x x
denotes the number of infected
people, while x 2 denotes the number of noninfected or “susceptible”
people. Equation (32) states that noninfected people become infected at
a rate proportional to x t x 2 which is a measure of the interaction be-
,
tween the two groups. The expression (31) for consists of two terms:
(1) —ax x which is the rate at which people die from disease or survive
,
and become forever immune, and (2) bx x 2 which is the rate at which
x ,
tion. By examining the vector field, we see that as t * oo, x^it) > 0 — —
FIG. 2.17
:
Problem 2.8. Find all the equilibrium points of the Volterra predator-
prey equations
*i = —x + t
x x2
x
x2 = x2 — x x x2
Linearize the system around each of the equilibrium points, and determine,
if possible, the nature of each equilibrium point. ( Answer one center, one
saddle.)
*i = x2
*2 = —X + t 6 (x 2 —y)
for (a) e = 1, (b) e = 0.1
Problem 2.10. Plot the velocity vector field for the pendulum equation
xx = x2
x2 = — sin*!
Problem 2.11. Consider the nonlinear circuit in Fig. 2.18. Suppose the
voltage-current relationship of the nonlinear resistor is given by
vr = ij - 3i'i + 3 ir AfOr)
FIG. 2.18
x x = v — x — x2 x
*2 = *1 — fix 2 )
(b) With v = 0, calculate the equilibrium points of the system.
(c) Linearize the system around each of the equilibrium points, and
determine the nature of the trajectories around each point.
)
2.4
PERIODIC SOLUTIONS AND LIMIT CYCLES
2.4.1 Introduction
1 Xi(0 =x 2 (t
2 x2 (t) = -x x (t)
3 *i(0) = x lQ
4 x2 (0) = x 20
is given by
5 x t (t) =r 0 cos(—t +0 0)
6 x 2 (t) =r o sin (— t +0 O)
where
7 r0 = (xf 0 + xl 0 )
1/2
8 0O = tan- 1 iaa
*10
Thus the solution of (l)-(2) is periodic regardless of what the initial
9 x t
= x + a x^f} — x\ — x\)
2
2
10 x2 = —x + a x (P — xf — x\)
x 2
2
field f(x) is radial if the vector f(x) is always aligned with the vector x,
for all x.] Now, if we introduce the polar coordinates
0 = tan —
-1
12
x i
13 r — —r
(X,r(fi
2 2
)
14 (j)
= —l
It can be easily verified that the solution of (13)— (14) is
15 =
+ c f-(1 0
2 ^ a ') 1/2
16 0(0 = 0 ~t O
where
Thus the system (9)— ( 1 0) has only one periodic solution, namely r 0 = /?,
i.e., x\ + x\ = /?
2
. Furthermore, whenever r0 ^ 0, all solutions of
(9)-(10)approach this periodic solution as t » oo. This example differs —
from the example of a simple harmonic oscillator in that the periodic
solution in the present case is isolated; i.e., there exists a neighborhood
of it that does not contain any other periodic solutions.
20 Vf(x) = X2 ) + *2)
4 A connected region can be thought of as a set that is in one piece, i.e., one in
which every two points in the set can be connected by a curve lying entirely within the
set. A set is simply connected if (1) it is connected and (2) its boundary is connected.
One can also think of a simply connected set as one that can be obtained by con-
tinuously deforming a circle. For example, an annular region is connected but not
simply connected.
.
is not identically zero over any subregion of D and does not change
sign in D. Then D contains no closed trajectories of [2.3(1 )]— [2 3(2)] .
22 f (x) • n(x) dl =0
J
But by the divergence theorem,
23 J^f(x)*n(x) dl = JJVf(x) ds (= 0)
s
where S is the area enclosed by /. Therefore, for (23) to hold, we must have
either Vf (x) = 0 V x e S or (ii) Vf (x) changes sign over the region S.
(i)
25 Xi =a xl x x + 012*2
26 X2 == 021*1 ~ 022*2
27 x = Ax
We know from Sec. 2.2 that a necessary and sufficient condition for the
system (25)-(26) to have periodic solutions is that the matrix A has two
nonzero imaginary eigenvalues. Because the eigenvalues of A satisfy the
characteristic equation
28 A2 — (a + 0
tl 2 2)A + (011022 01202 1 ) =0
it is clear that the system (25)-(26) has periodic solutions if and only if
29 0n + 02 2 = 0
31 Vf(x) = a + a 22
lx V x g R2
Hence we conclude that if a n + a 12 ^ 0, then no periodic solutions
exist, which is in accordance with the previous discussion.
33 x x
= x + x x\
2 x
34 x2 = —X + x\xx 2
35 x 1
=x 2
36 x2 = —x 1
44 S = {x(Y), > 0} t
then either
(i) S is a periodic solution of [2.3(l)]-[2.3(2)], or
(ii) L is a periodic solution of [2.3(1 )]— [2 3 (2)] .
book.
and such that all limit points of some trajectory S are contained in
M. Then M
contains at least one periodic solution of [2.3(1)]—
[2.3(2)]. In practice, it isM contains
very difficult to verify that all
M closed, can
the limit points of a trajectory. However, because is it
49 X l
= —X + x
l 2
50 x2 = —X — X
x 2
The velocity vector field for this system is sketched in Fig. 2.20. If we
chooseM to be the unit disk centered at the origin, then all along the
boundary of M the velocity vector field points into M. Hence all trajec-
*2
FIG. 2.20
36 Chap. 2 Second-Order Systems
54 0(0 = 0o — t
The concept of index is a very powerful one, and the results given
below only scratch the surface of the many results that are available.
Unfortunately, the arguments involved in index theory are well beyond
the scope of this book. Hence we shall content ourselves with the few
results given below, most of which are given without proof. For further
results, see [18] Nemytskii and Stepanov.
The definition below introduces the concept of the index of a
vector field.
57 /,(/) = 2- dd,(x u x2 )
where 0 (x u x 2 ) t
is the direction of the vector field f at the point
Ol. x2 ).
61 fact The index of a center, focus, and node is 1 and that of a saddle is
— 1. (This fact can be easily verified by a sketch of the velocity vector field
near a center, focus, node, and saddle.)
63 It (J) = ± I (Pi ) t
i=i
64 fact Let f and g be two vectorfields on R 2 and let 0 and 9 g denote the, f
65 x(/) = g[x(t )]
Then
66 It (J) = Tg (J)
67 fact Let J be a simple, closed, positively oriented trajectory of (56). Then
68 /,(/) = 1
equations
71 = —x + x t x
x2
72 X2 =X —Xx 2 t 2
38 Chap. 2 Second-Order Systems
We shall digress briefly to discuss the rationale behind the above model.
In (71)— (72), x x denotes the number of predators (foxes, say), and x
2
denotes the number of prey (rabbits). If x 2 = 0, (71) reduces to x x
= ~x u which implies that in the absence of prey the number of preda-
tors will dwindle exponentially to zero. If x ^ 0, (71) shows that x
2 x
73 *i=/i(*i,*2)
74 x2 — fi(x ij * 2)
78 {(x l5 x 2 ): V(x u x 2 ) = c)
80 V(x u x 2 ) = hfx + x ) h 2 (x 2)
82 h\(x i )x l (x 2 — 1) + h' (x )x 2 2 2 (l
— Xj) = 0
which can be rearranged as
84 h\(x 0 j-^1
1 ——
x
=c
1
85 h'2 (x 2 ) 7 -^— = c
1 — X2
where c is a real constant. It can be easily verified that the solution of
(84)-(85) is
86 h x (x x )
= c(ln x j
— Xj)
87 h 2 (x 2 ) = c( In x2 —x 2)
89 In x t
— x + In x — x = const.
1 2 2
91 Xj =x 2
92 x2 — — sin Xi
Let us once again choose V(x l9 x 2 ) to be of the form (80). Then, for
(75) to be satisfied, we must have
93 h\{x 1 )x 2 — h'2 {x 2 ) sin x x
=0
which implies, as in Example (79), that
94 ) = hMl = const. = c
sm Xi x2
Solving (94) gives
Sec. 2.5 Two Analytical Approximation Methods 41
95 ^i(-^i)
— — C COS X 1
96 h 2 (x 2 ) = c^~
97
^ — cos Xj = const.
constitute a continuum of closed trajectories for the simple pendulum.
98 Remarks Examples (79) and (90) illustrate how the method presented
:
here can sometimes yield good results. However, it should be clear that
(1) a function V(x u x 2 ) of the type (80) does not always work and (2)
even if it does, there is no guarantee that all closed trajectories of the
system are of the form (78). Despite these limitations, however, the
method is of some value, as indicated by the two examples.
*1 = x2
*2 = -*(*i) - Kx 2 )
Problem 2.13. Sketch a vector field with exactly one node and one
saddle point. Show that it is not possible to continuously deform this vector
field in such a way that there is a periodic solution enclosing both the node and
the saddle point.
Problem 2.14. Using the method of Sec. 2.4.5, derive an expression for
the closed trajectories of the system
*1 = x2
x2 = —g(x i)
25
TWO ANALYTICAL APPROXIMATION METHODS
In this section, we shall describe two techniques for obtaining analytical
expressions that approximate the periodic solution of second-order
nonlinear differential equations. In contrast with the method presented
in Sec. 2.4.5 (which gives exact expressions, if it works), the two methods
presented here are only approximate. However, they have the advantage
of having a wide range of applicability and of enabling one to study the
so-called “slowly varying” oscillations. It should be emphasized that,
^ t t
2 y(t) = a sin (t + i)
<t
where a( •) and 0(-) are “slowly varying”; i.e., a(t) and 0(0 are “small.”
Actually, if >’(•) is given by (3), we have
6 a sin (t + 0) + a0 cos (t + 0) = 0
where we have suppressed the dependence of a and 0 on t in the interests
of brevity. Substituting for y and y from (3) and (4) into (1) gives
Equations (6) and (7) represent two linear equations in the unknowns a
and 0. Solving for a and 0 gives
Ifwe want to find solutions of (1) of the form (3) where o(-) is
g(T) - g( 0) _ Q
T
or, equivalently,
11 dt = 0
T
where T is the period of «(•)• Unfortunately, (11) cannot be used
directly because the period Tis in general dependent on p and is hence
unknown. To get around this difficulty, we observe that «(•) goes
through one complete period as the phase 0 = t + 0(0 goes from 0 to
2n. Thus we change the variable of integration in (11) from t to 9; then
the limits of integration become 0 and 2n and ,
the integrand d(t)
becomes [using (8)]
d9__df
13
2
n~ T
Equation (13) expresses the fact that as t varies by T, 9 varies by 2n.
Thus (11) becomes
2n
C
14
1
where a and <5 are now constants. In this case, (14) and (15) simplify to
r 2 tc
17 cos 9 •/(a sin 0, a cos 0) d9 =0
Jo
19 d(t)
~ a(T - ) a(0)
T
20 0(0
2i f(y y),
= y( - y 1
2
)
Hence
28 f(a sin 9, a cos 9) = a cos 9(1 — a 2 sin 2 9)
= (a — cos 9 -j-
-^jr cos 3 9
so clearly
30 ~ 6 -f(a 6 a cos 9) dd =0
^J sin sin ,
31
32 0 = 0
If we are interested in finding the steady-state periodic solutions of Van
der Pol’s equation, we set a = 0, 0 = 0, which gives a = 2. Hence, to
first order in //, the limit cycle of the Van der Pol oscillator is described
by
33 y(t) = 2 sin ( +0 O)
To get the slowly varying solution, we solve (31) and (32), which
results in
1/2
34 a(t) =2 1 ~]
_1 + cexp(— fit)]
35 0(0 = 00
where c is a constant determined by the initial conditions. Thus we see
that even though the parameter ji does not affect the steady-state solu-
tion, it does affect the rate at which the transient solution approaches
the steady-state solution.
37 j(0) =a 0
38 j(0) = 0
This equation represents the oscillations of a mass constrained by a
nonlinear spring. If jn > 0, the spring is said to be hard, whereas if
ju < 0, it is said to be soft.
39 y 0 (t) =a 0 cos co 0 t
46 Chap. 2 Second-Order Systems
43 Jo(0 = cos co 0 t
sponding to pi =
0. Now the equation for yj(.) becomes
44 y x
(t) + co 2
Q y x
{t) = ~yl{t) = -a\ cos 3 co 0 t
= —\al cos co 0 t — cos 3 g> 0 *
3 /7 3 n3 n3
45 y {t) = * sin “ 32cof cos + 32^ cos 3
'
-st
The t sin a> 0 t term on the right-hand side of (45) is the secular term,
which arises because the forcing function of the equation for
y j
( )
contains a component of angular frequency co 0 , while the system itself
has a resonant frequency at o) 0 . Combining the above expression for
To(*) and /!(•), the approximate solution of (36)-(38), which is good to
However, ^ 0,
the angular frequency of the periodic solution of
if pi
functions j0( #
)> T i( #
)> • •
•] has an angular frequency co 0 , it is clear that
all functions j; f (*) consist of summations of terms of the type sin nco Q t
“small,” and let us express cos [(co 0 + S)t] as a power series in <5. This
leads to
S 2t 2
47 cos [(co 0 + S)t ] = cos co — Qt St sin co 0 t j- cos co 0 t . .
interval, but if the series is terminated after a finite number of terms, the
resulting finite summation contains secular terms. Moreover, the peri-
odicity and boundedness properties of the function cos [(co 0 + 5)t\ are
not at allapparent from the power series expansion in 5.
To alleviate this difficulty, we shall suppose that the solution X*)
of (36)— (38) is periodic with an angular frequency co, which itself is
48 CO
2
= CO + 2
0 //f iOo) + /^Oo) + • • •
50 To + ATi + • • • + C0 y - 2
0 a£iTo + MC0 yi
2
+
. . .+ juyl + . . . =0
Equating the coefficients of each power of ii to zero gives
51 y0 + co y = 0;
2
0
=
y 0 (0) a 0 y 0 ( 0)
,
=0
+ co y = —To + £i To; Ti(°) = Ti(°) = 0
2
52 y x x
Hence
53 To(0 = a cos o
54 y x
(t) + (o y = — al cos cot + ^ a cos cot
2
x
{t )
3
0
For y x {^) not to contain any secular terms, the forcing function on the
right-hand side of (54) must not contain any terms of angular frequency
co. Hence we must have
55 f1 = %d 2
0
48 Chap. 2 Second-Order Systems
where
58 co
2 = col + \vtal
59 Example. Consider the simple pendulum equation
60 y + sin y =0
Equation (60) can be approximated by
61 y +y -lL = 0
which is of the form (36), with // =— Using the foregoing analysis,
we see that the frequency of oscillation of the simple pendulum is
2 = -
62 O)
f
1
Solve the same equation using the perturbation method, and show that both
methods give the same solution to the first order in ji.
Problem 2.16. Apply the perturbation method to the Van der Pol
equation
y +y= vy(i - y 2)
Problem 2.17. Apply the Krylov-Boguliubov method to the pendulum
equation
Show that the expression derived for the frequency of oscillation is the same
as (62).
Sec. 2.5 Two Analytical Approximation Methods 49
y + y = nf(y, y)
XO) = 0; j>(0) = a 0
arguments. Show that, to first order in ji , both the perturbation method and
the Krylov-Boguliubov method give the same results.
Nonlinear differential equations
3 .1
MATHEMATICAL PRELIMINARIES
3 (V2) x + (y + z) = (x + y) + z 9 V x, y e V
(associativity of addition)
(V 5) For each rt ,
r 2 in R (resp. c x ,
c 2 in C) and each x e F,
8 (''i + r 2 )'X =r x *x + r 2 -x
9 1 • x =x
This axiomatic definition of a linear vector space is best illustrated
by several examples.
11 x + y = (*i +j„...,x + y J, n)
12 rx = (rx u rx„)
linear vector space if we make the following definitions: Let x(-) and
^(•) be elements of F\a b ]. Then the sum of x(-) and j(-), denoted by
,
16 (rx)(f) = rx(t)
In other words, the sum of two functions in F[a b\ ,
is obtained by point-
wise addition, and the product of a scalar and a function is obtained by
pointwise multiplication.
-
i
x = Ox
27 (N2) ||a*|| = a V* |
|
• ||
jc||, e X, for all scalars a
28 (N3) ||* + y|| < ||*|| + ||y ||, Vx,yeX
(triangle inequality)
29 [ definition] A
sequence (x„)7 in a normed linear space (X, || • ||)
is said to converge to an element * 0 e if ||*„ *0 0 as X — 1|
— >
30 1
1*„ —* 0 1|
<e whenever n > N(e).
This basic definition of convergence can be interpreted in many
ways. Clearly the sequence of vectors (*„)7 converges to * 0 if and only
if the sequence of nonnegative numbers (||*„ * 0 1|)7 converges to zero. —
Alternatively, let B(x 0 ,
e) denote the ball in X defined by
31 B(x o, e) = {* e X: ||* — * 0 II
< e}
Then the sequence (x„)T converges to *0 if, for any e, the ball B(x 0 ,
e)
1 1
f(x o) ~ f(y)\\r<£ whenever 1 1
x0 —y \ \ x < d(s, x 0)
1 1
f(x) - f(y) Hr < £ whenever 1 1
x -y \ \ x < 5(e)
We shall postpone examples of these concepts until we have given
a few examples of normed linear spaces.
Remarks:
1. The concept of a continuous function from one normed linear
space to another is a natural extension of the concept of a
continuous real-valued function of a real variable; in a general
normed linear space setting the norm plays the same role as
the absolute value does in the set of real numbers. Similar
statements apply to uniform continuity.
2. The important difference between continuity and uniform
continuity is that in that latter case 8 depends only on s and
not on x or y.
whenever we we
try to iteratively solve a nonlinear equation.) Thus
need a characterization of sequences that does not involve the (possibly
unknown) limit of the sequence. This is provided by the concept of a
Cauchy sequence.
a Banach space, then (as stated in the above definition) every Cauchy
sequence is convergent. (2) Even if ( X ||)
is not a Banach space, it , ||
•
43 1 1
x |
— max \x t \
1 <i<n
(The reason for the subscript oo will become clear later). The function
||
•
|
axioms (Nl) through (N3), as can be easily verified. In
|oo satisfies
< max |
x , | + max | y t \
= ||x|U + ||y||„
,
46 ll*lli = i=i
SI*/l
Clearly ||
•
||j
also satisfies axioms (Nl) and (N2). To verify (N3),
suppose x, y e Rn . Then
47 ll
x + ylli = i=2 1* + *1 < i=l
2 (1*1 + 1*1)
1
= i=l
21*1 + 21*1
i=l
= 11*11. + llylli
Hence the ordered pair ( R n Hj) , ||
• is also a normed linear space and can
in fact be shown to be a Banach space.
even though the underlying linear vector space is the same in both cases,
(namely R n ), because even though the linear vector space on which the
two norm functions are defined is the same in both cases, the norm
functions themselves are different.
50 = ,jc
J’
ii*ii'
{§ 'iT
where p ranges between .
p becomes the norm
1 and oo. If /? = 1, || ||
(N2) and can be shown to satisfy (N3) whenever 1 oo. Thus <p <
the ordered pair (R n •
|[p)
is a normed linear space
, whenever
|| 1 p <
< oo. (Of course, for different values of
p, the corresponding normed
linear spaces are different.)
In particular, if p = 2, we have
[
rt
) 1/2
2
51 IMI 2 = {Sl*fl
}
we shall now discuss. For ease of statement, we shall give the facts
only for Rn ,
but they are equally true for C”.
fact Let ||
•
|| a and be any two norms on
||
•
\\ fi
Rn . Then there exist finite
positive constants k t and k 2 such that
&i||x|| a < 1
1
X 1
1^
<k 2 ||x IU, V X e Rn
Now
us consider some consequences of fact (52). Let
let • a || ||
suppose x„ || —
x 0 a * 0 as n—> oo. Thus (xjr converges to x 0 in ||
—
(. Rn , ||
-
|| a );
i.e., (x B )“ converges to x0 in the sense of the norm ||
•
|| a.
so that x„ —x 0 |U
— > 0 as n — oo. Hence (x„)7 also converges to x0
in the sense of the
||
norm ||
-
|U-
Conversely, suppose 1 1
x„ —x 0 1 U
— > 0
as n—+ oo. Then, because ||
x„ —x 0 JU < (l/^i)||x — x M 0 |U ||x„ —x 0 || a
also approaches zero as n — > oo. Thus we see that (x„)~ converges to
x0 in the sense of ||
•
|| a if and only if it converges to x 0 in the sense of
IU This means that convergence in R is independent of the norm
n
•
||
chosen. Now, let us for the moment use the specific norm • IU to test ||
||
.
|| a if, whenever (Or is a sequence in a b] such that [< ,
tn —> t as n — oo,
1
the continuity of x(*) in one norm is precisely the same as the con-
tinuity of x(*) in another norm. In particular, using fact (52), we can
state the following.
Note that facts (52), (56), and (57) hold also for C”. But they
do not hold for an arbitrary normed linear space.
denote the set of all continuous real- valued functions over [a, b ]. Define
a function ||
•
|| c over C[a, b] as follows: If 40 £ C[a, b], then
59 II 40 lie
= max |
x(0
t(=[a,b ]
< max t
|
x(0 1
+ max y(t) = ||x(-)llc +
t
| \ II lie
where all maxima are taken over [a, b ]. Thus the ordered pair
( C[a b\, c) is a normed linear space. Notice that a sequence of
, ||
•
||
^he set Cn [a ,
b] is not to be confused with C (n) [a, 6], which is the set of all
continuous real-valued functions over [a, b] that are /z-times continuously differen-
tiable. We shall not have much occasion to use the latter symbol.
:
62 ||x(-)llc = max||x(OII
re [a, 6]
Once again, ||
•
|| c constitutes a norm on C"[a, b\, as can be readily veri-
fied.Axioms (Nl) and (N2) are easy. To test (N3), let x(.), y(.) e
C-[a, b]. Then
63 ||
x(.) + y(-)|| c = remax||x(0 + y(/)|| [a, 6]
inequality on R)
n
= l|x(.)llc + lly(-)llc
The ordered pair ( C n
[a , b], ||
•
|| c) is in fact a Banach space.
In this example, it is essential to note the difference between
||
•
||
and ||
•
|| c
— ||
•
||
is a norm on Rn c is a norm on
,
while ||
\a , b\.
•
||
Cn
The former has an n - vector as its argument, while the latter has a vector-
valued function as its argument. When we study nonlinear differential
equations in Sec. 3.3 this distinction becomes crucial.
The next result shows that an inner product space can be made
into a normed linear space in a natural way.
proof Clearly ||
•
||
satisfies (Nl) and (N2). To prove (N3), we need the
following extremely useful inequality, known as Schwarz’s inequality.
72 |<*,y>|<|MHbll
73 |<x, y}\ = ||x||-||y|| if and only if ax + fly = 0* for some scalars a,
/?, not both zero.
proof of lemma (71) We shall only prove the case where is a real inner X
product space. The complex case is handled in much the same fashion.
Consider the function
2
75 <x,y> <||x|| 2 .||y|| 2
Furthermore, suppose ax ^
+ Py 0* whenever either a or /? is nonzero.
Then/(a, p) > 0 whenever either a or /? is nonzero. This is the case if and
only if the discriminant of (74) is strictly negative, i.e.,
76 <x,y> 2
< ||x||
2
-||y||
2
= (MI +||y||) 2
Hence (N3) is satisfied, and ||
•
||
is a norm on X. —
] , , ,1
Theorem (69) shows that every inner product space can be made
into a normed linear space in a natural way. It is therefore possible to
speak of the completeness of an inner product space.
x„) and y = (y u . .
. y n) belong to R\ Then
80 <X, y>„ =E
=
Xy
1 1
t t
It is routine to verify that the function <•, •>„ satisfies axioms (II)
through (14), so that the ordered pair(i£",<(-, -)„)is a real inner product
space. It is in fact a Hilbert space. The norm on Rn corresponding to
this inner product is
81 ||
x || = {£*?}
which is the Euclidean norm discussed earlier.
83 <x(*)» y(*)>c = [
<x(0, y(t)}n dt
is uniformly continuous on X.
proof Use definition (32) of uniform continuity, and let <5(e) = 8, for
each 6. Because, for each x, y e X, we have
85 <11* -y\\
,
it follows that
88 fact Suppose (X, <• , •» is an inner product space. Then, for each y e X,
the function x t-> is uniformly continuous.
proof Use definition (32) of uniform continuity, and let S(8) = 8/\\ y\\,
8. Now, suppose ||x — z\\ < 5(8). By Schwarz’s inequality, we
for each
have
89 |<x, y> - <z, j»| = |<* - z,y>\ < \\x - z\\-\\y\\ < <5(fi)||y|| =8
Hence the function jc i-» y) is uniformly continuous. M
Problem 3.2. Show that, in a linear vector space, the additive inverse of
an element is unique.
Problem 3.3. Let S be the sequence space of Example (20), and let S r
be the subset of S consisting of sequences converging to r. For what values of
r is S r a subspace of *S?
{x G R 1
1
x | |p
= 1}
2 x, £ SI
=
i l
4 4
Ja
f x(t) dt II
II
< Ja f 1 1
x(r) || dt
y0 ,
respectively, then the sequence of real numbers «a:„, converges to
<* 0 , J'o) (Hint: Write <x„, y„}
- <x, y) as
3.2
INDUCED NORMS AND MATRIX MEASURES
In this section,we shall introduce the concepts of the induced norm
of a matrix and the measure of a matrix. These concepts are vitally
needed in Sec. 3.5, where we shall study methods for estimating the
solutions to nonlinear differential equations without actually solving
them.
LetC nxn (Rnxn ) denote all n X n matrices with complex the set of
nXn
(real) elements. Then C
can be made into a linear vector space if
addition and scalar multiplication are done componentwise. Moreover,
each element A e C nxn defines a corresponding linear mapping a from
Cn into Cny according to the rule
1 a(x) = Ax, V x e Cn
Conversely, every linear mapping P from C into itself can be associated
n
follows Let : e,- be the vector in C" which has all zero components, except
for the yth component, which is equal to 1. Now, p(ey ) is an element of
C because p maps C
n
,
n
into itself. Accordingly, let (fe l7 , i = 1
3 B= [b 1 |b 2 |...|bJ
then one can easily verify that
4 P(x) = Bx, V x g Cn
Therefore, there is a one-to-one correspondence between elements of
C nxn and linear mappings from C* into C”. 2 We do not in general dis-
possible for the same linear mapping to have different matrix representations with
respect to different bases. However, in this book we shall not explore such subtleties
of linear algebra.
Sec. 3.2 Induced Norms and Matrix Measures 65
5 [definition] Let ||
•
||
be a given norm on C”. Then for each
matrix A e C nXn ,
the quantity ||
A || f
defined by
6 l|A||, = sup
X
sup || Ax|| = sup || Ax||
x^O II |
| x|= 1 H x B<1
xGC”
and the other is the induced norm function mapping C nxn into R. ||
•
||,
The induced norm of a matrix A (or, what is the same, the induced
norm of a linear mapping A) can be given a simple geometric inter-
pretation. Equation (6) shows that A is the least upper bound of the || || £
7 B = {xe C n
: ||x|| < 1}
function ||
maps C nXn into [0, oo), satisfies axioms (Nl) through (N3),
•
|| f
proof Clearly, A 0 || || t
- > V A e C nXn and axioms (Nl) and (N2) can ,
9 || A+B ||/
= sup 1 1
(A + B)x|| = sup || Ax + Bx||
11x11 =1 II
xH=1
< sup=
Ix!
[|
l
|
Ax + Bx || || ||] (by the triangle inequality on C”)
< lxsup 1
1 Ax 1 1
+ sup 1
1 Bx ||
= II All,- + ||B||,
H=1 llx||=l
i,j
x
Then one can easily verify that defines a norm on C" ". However, ||
- 1|,
||
(Note that some authors use the term bound norm instead of
•
II;.
induced norm.)
In fact [3.1(52)], it is stated that any two norms on C n are topo-
logically equivalent. By the same token, any two norms on C nXn are
also equivalent. In particular, an induced norm is equivalent to another
“noninduced” norm. But a special property of induced norms is bought
out below.
11 lemma Let ||
•
|| f
- be an induced norm on C nxn corresponding to the
norm ||
*
||
on Cn . Then
12 1
1 AB 1 1/ < A 1
1
| h 1
1
B | | f, V A, Be C«*»
proof We have
13 1
1 AB |
|j* = sup || ABx ||
II
X =1 11
However, by definition,
14 ||
Ay || < A 1
1
|
h-
1 1 y 1 1, Vy g Cn
So, in particular,
15 ||
ABx || < 1
1
A | f 1
1 Bx 1
Similarly,
16 1 1
Bx 1 1 < 1
1
B 1
1,-
1 1
x 1
As a result, we get
17 ||
ABx 1 1 < A || ||;
||
B ||; |
|x ||
Thus induced norms have the special feature that they are
submultiplicative i.e., the induced norm of the product of two matrices
A and B is less than or equal to the product of the induced norms of
A and B. It can be readily veiified by example that the norm 1|, ||
-
defined in (10) does not have this property (and hence cannot be an
induced norm).
Sec. 3.2 Induced Norms and Matrix Measures 67
TABLE 3.1
* j
Let ||
• ||f
be an induced matrix norm on CnXn Then the correspond-
.
e->0 + S
19 Theorem Whenever ||
• || f
is an induced norm on C nXn ,
the cor-
responding measure //,(•) has the following properties:
(i) For each A e C" x ", the limit indicated in (18) exists and is
well defined.
(ii) -|| A||, < -/*,(- A) < ^.(A) < || A|L, V A € C"*"
(iii) /«,(aA)= a//, (A), V a > 0, V A e C n/n
(iv) max [/i,(A) — #,(— B), —/!,(— A) + /i,(B)] < /i,(A + B)
< Vi(A) + a,(B)
3 Some authors also use the term logarithmic derivative.
;
= ll
I + fi A|li - 1
21 f(e A)
22 /(Si ;
A) < f(e 2 ;
A) whenever 0 < <e 2
By definition, we have
23 Ci/0>i l A) = 1
1 1 H~ Ci A ||/ 1
A
=ig< +*, >+( -g) i-i
i i i
<fi||I + e 2 A||,- + l — —1
02 02
= s f(s ;A) 1 2
it follows that
25 — A 1
1
1 1/ </(£; A) < A 1
1
| | f, V e> 0
27 — M- A) < M A)
Now, clearly,
Therefore
Therefore,
This proves the validity of (iii) in the case a > 0. If a = 0, (iii) fol-
First, we have
34
||I + e(A + = ^(||! + 2eA + 1 + 2eB||, - 2)
^ ||I + 2eA||,- — 1 ||I + 2eB||, — 1
^ 28
+
,
2fi
36 M A) - M- B) < MA + B)
By symmetry, we must also have
= A) + (1 - a)//;(B) v a e [0,1]
by (iii)
Therefore,
— 1
1 4- cA |
e — e
As £_ —
0 + the left-hand side of (41) converges to ///(A), while the
,
that the norm used in (18) is an induced norm. It is possible, given any
norm ||
on C nxn (induced or not), to define a corresponding measure
•
||
nxn
function //(•) mapping C into R. In this case, all the properties of
Theorem (19) still hold except for (vi). However, for the purpose of
estimating solutions of vector differential equations, only measure
functions corresponding to induced norms prove useful.
Given a particular vector norm on C", it is in general a very ||
•
||
||
•
|| 1? and ||
•
|| 2 can be calculated explicitly and are displayed in Table
3.2.
TABLE 3.2
iixii. = i)i*/i
i=l
jUnW = max {ajj + "Z |«/y|}
j i*j
172
11*11* = (t k.-l
2
)
A2(A) = Amax [(A* + A)]/2
42 Example. Let
L-i -3J
Using the formulas given in Table 3.2, one can verify that
^ioo(A) = -1 ; /Loo(-A) = 4
This clearly shows that the measure of a given matrix depends on the
particular norm used to define the measure function. The eigenvalues of
A are
/1.2
1
- _a
-
2
±7,vr
+ —
If we use property (vi) of Theorem (19) to obtain bounds on the real
parts of the eigenvalues of A, we get
43 Example. Let
0 1
72 Chap. 3 Nonlinear Differential Equations
T3 13,11
L2
ju ioo gives an exact lower bound, ju n gives an exact upper bound, while
li i2 gives the interval of the smallest “width.”
In practice, to estimate the range of the real parts of the eigen-
values of a given matrix, one would utilize various measure functions
to obtain various intervals on the real line. Then, because property
(vi) of Theorem (19) is valid for all measure functions corresponding to
induced norms, one would take the intersection of all intervals in order
to obtain the best estimate. In Example (43), this intersection is [1, 2],
which happens to be an exact estimate.
Before we close this section, some comments are in order regard-
ing the use of the matrix measure in connection with norms on R n
and
nxn
the corresponding induced norms on R . First, it is easy to see that
every element in R nxn corresponds to a linear mapping from R n
into
itself and vice versa. Next, suppose ||
•
||
is a norm on R n
. Then, given
any A e RnXn we can define its induced norm A in a manner entirely
, || || f
the bounds on the real parts of the eigenvalues of A) essential use was
made of the fact that A maps C n into itself, because 2 might be a complex
eigenvalue of A, with a corresponding eigenvector having some complex
components. In spite of this fact, however, it can be shown that, given
any norm on R n and any A e jR bx ", it is possible to “extend”
||
•
||
itself. The details are omitted here, but the end result can be sum-
marized as follows.
Theorem Let ||
be a norm on R and let
•
||
R nXn —» R and
n
, ||
• -:
|| t
A=
-2 1 o r
0) (ii) A
2 — 3_ -l o_
4 1
2~
(iii) A= 0 -2 1
_-l 1 — 3__
Problem 3.8. Calculate /* n (A), p i2 {A), // IOo(A), and // n (- A), // /2 (- A),
// f0 o(— A), for the three matrices above. Find an interval in the real line that
contains the real parts of all eigenvalues of A.
33
CONTRACTION MAPPING THEOREM
In this section, we shall state and prove a very important theorem,
which we shall use in Sec. 3.4 to derive the existence and uniqueness
of solutions to a class of nonlinear vector differential equations.
The theorem that we shall prove is generally known as the con-
traction mapping theorem (and sometimes as the Banach fixed point
theorem) and is usually given in two forms: the global version and the
local version. The local theorem assumes a weaker hypothesis than the
global theorem and obtains correspondingly weaker conclusions. We
shall first give the global version.
Note that, hereafter, we use mapping function , and operator
,
interchangeably.
1 Theorem Let (X s ||
•
||)
be a Banach space, and let T : X X be
a mapping for which there exists a fixed constant p < 1 such that
4
R 1 1
6 ll^m Aw ||
= H-Vh+j- AVj|| ^ S= / 0
l|A"n+i + l A n +,-||
< 1=0
S P n+i
‘11*1 -A 0 ||< S
1=0
p
n+i
\\Xt Xq 1
= x~-p\\
Xi ~
8 ||
A* - All =|| 7** - Tx\\<p\\x* - All
which can hold only if ||a* — a|| = 0, i.e., a* = a (recall that p < 1).
Finally, to prove the estimate (4), consider the inequality (6). Because the
norm function ||: X is continuous, we have
||
•
9 A'
* - *„ll = II
(lim aJ - xn 1
m~* °o
where we have used the fact that the right-hand side of (6) is independent
of m. —
Note that in general it is not possible to replace (2) by the weaker
condition
10 ||
Tx — Ty || < ||x — y\\ V x, y e X such that x y
It is easy to show that any mapping satisfying (10) has at most one
fixed point, but quite possibly it may not have any at all. As a simple
example, let X= R, and let /: R—+ R be defined by
11 /(x) = x + -y
— arc tan x
Then
12 f\x)
J v y
= — 1 1
1
p
+X
— *2 = -r—^ <
+X
1
^
1
1 f° r all x g R
Thus by the mean value theorem,
13 fix) - f(y ) =f'(<£)(x - y) for some { e [x, y]
as the starting point and then setting x /+1 =/(x ) V / > 0. The f
FIG. 3.1
76 Chap. 3 Nonlinear Differential Equations
The application of Theorem (1) is limited by the fact that the opera-
tor T is required to satisfy (2) for all x, y in X. In other words, Thas to
be a global contraction. In Theorem (16), we examine the case where
T satisfies (2) only over some region M in X ,
i.e., the case where Tis
a local contraction, and we derive correspondingly weaker results.
—
and let T: X >X. Suppose there exists a constant p 1 such that <
17 II
Tx — Ty\\ < p\\x — y\\, \/ x, y <e M
and suppose one can find an element x 0 e X such that the ball
B= -x Tx °y*°
18
{* G X: II X 0 II
< II
l
II
21 Remarks:
1. The significance of Theorem (16) lies in the fact that 7^is only
required to be a contraction over the set M, and not over all
of X. The price we pay is that the conclusions of Theorem (16)
are also weaker than those of Theorem (1).
2. Everything is contingent on finding a suitable element x0 in
M such that the B defined contained in M. In
ball in (18) is
If we choose M as the
interval [—1, 1], then Tis a contraction
on M. However, not possible to find an x 0 e
it is such that M
the ball B defined by (18) is contained within M. Accordingly,
T has no fixed point in M.
3. Suppose we do succeed in finding an x 0 e M such that the
hypothesis of Theorem (16) holds. Then the particular sequence
]
(x 0 ,
x 1? . .
.) converges to x* (the unique fixed point of Tin M).
However, if y is some other element of M, the sequence ( y Ty , ,
T 2 y, . .
.) may or may not converge to x*. In contrast, if T
is a global contraction, then the sequence (y, Ty, T 2 y, . .
.)
proof of Theorem (16) First, it is clear that Thas at most one fixed point
in M, because of (17). If x0 e M
chosen in such a way that the ball
is B
defined in (18) is contained within M, then clearly the sequence (xf )~ is
contained within M [apply the inequality (6) with n = 0]. Because the con-
traction condition (17) holds in B, one can show, as in the proof of theorem
(1), that (xt ) is a Cauchy sequence and therefore converges, say, to x*.
Because of the continuity of the norm function, x* e B. The rest of the
proof exactly follows that of Theorem (1). —
23 Example. Consider once again the case where X= R, and let
24 sup
*e[-l,i]
I
f'(x) |
A p< 1
= \f(Xp) l/Oo) — Xp
B \ + I
1 -p = [a, b
is a subset of [—1, 1]. Then Theorem (16) states that there is a unique
for some z E (.
, IHI) and some r Let P: X — X be
> an operator
satisfying the following conditions:
(i) P maps B into itself, i.e., Px e B whenever x e B.
(ii) There is a constant p < 1 such that
29 ||JS„-X*||< -*oll
3.4
NONLINEAR DIFFERENTIAL EQUATIONS
3.4.1 Introduction
1 x(0 = i[t ,
x(/)], t > 0; x(0) =x 0
Sec. 3.4 Nonlinear Differential Equations 79
conditions under which (1) has exactly one solution over [0, oo).
One small point is to be cleared up before we proceed to the
theorems. First, if x(*) is a solution of (1) over [0, T], then x(*) also
satisfies
4 ||f(7, x) - i(t, y) ||
<k\\x- y||, V x, y e B, \/te[0,T]
5 ||
f<>, x 0 )|| < h, V f e [0, T]
where B is a ball in R n
of the form
6 B= [x e R n
: || x — x0 1|
< r}
Then (1) has exactly one solution over [0, 8] whenever
8 S ^ mia ( T’f’FTT?)
for some constant p < 1.
[ [
9 Remarks :
in C n
[ 0, 8].
Clearly x(-) is a solution of (2) over [0, 8] if and only if (Px)(-) = x(-);
arbitrary elements of S ;
then x(0 and y (t) lie in the ball B, V t e [0, 8 ].
Thus
13 1 1
(Px)(t) - (Py)(t) 1 1
< J' 1 1
f t, x(r)] - f [r, y(T)] 1 1
dx
<Pl|x(-) -y(-)llc
so that P is a contraction on S.
Next, we show that P maps S into itself. Let x(-) e S. Then
J0
= *
{ f[r, x(t)] - f(T, x 0) + f(r, x 0) } dx
Jo
- Xoll < {||f[T, X(T)] - f(T, X 0 )|| + l|f(T, Xo)|D dx
J‘
<kbd +hd<r
by (7). Hence
point being that S is a proper subset of C n [ 0, 5]). Thus the proof is com-
plete if we show that any fixed point of P in C n [0, d] must in fact be in S .
18 ||x(0 — x0 < 1 1
A| | x(t) — x(0)|| dx +ht
<hd + k\ |
x(t) — x(0) 1 1
dx
with respect to its second argument (x) and continuous one-sided partial
derivatives with respect to its first argument ( t ). Then (1) has exactly one
satisfies (4) and (5) for some set of finite constants r , t, k and,
h. M
Example. Consider the Van der Pol oscillator, which can be
described by the set of two first-order equations
*i= *2
X = — Xi —
2 1
— 1)*2
Theorem Suppose that for each Te [0, oo) there exist finite
||f(t, x) - f(t, y) ||
< kT ||
\
x -y ||, V x, y e Rn V , t e [0, T]
||f(t, x0 1|
< hT ,
V ? e [0, T]
Then (1) has exactly one solution over [0, T], VTe [0, oo).
30 fl (t, x) = f(/ + 5, x)
Thenf! also satisfies (26) and (27); therefore, once again by Theorem ( 3 ),
(29) has a unique solution over [0, S], where 8 same as before. Denote
is the
this solution by y 2 (*)- It is ea sy to verify that the function x 2 (-) defined by
yi(?>, o<t<S
31
.y i(t — 5), d < t<25
is the unique solution of (1) over the interval [0, 28]. Proceeding by induc-
tion, let x m (-) denote the unique solution of (1) over the interval [0, mS\
and consider the differential equation
where
33 f mit, x) = fit + mS ,
x)
Let ym+ i(*) denote the unique solution of (32) over the interval [0, (5] (the
same 8 as before). Then the function x m+1 (>) defined by
x m(0j 0 < < m8
t
34 m+1
^ .
mS < < mS + t 8
is the unique solution of (1) over the interval [0, m8 + <5]. In this manner,
the unique solution can be extended to all of [0, T].
39 X m (/)|| < 2 1 1
Xm + * + 1 ) X„, +l-(^)||
*m+i + l
< s hT kr\m + i+ i)i
m+p fi
= =S
m+
i
Wtt
40 |x m+ p(-) - x m (-)llc = sup ||x m+J,0) - x m (OII
t€[0,T]
m+p T7 00
< =2m + i 1
h T kif 4t- <c 1
^ • i
2 M#-
—m+
1
/!
41
<£jT||z,(-) - z 2 (.)llc
Next, to show that x*(-) is the only solution of (2), suppose y(-) also
satisfies (2). Then
48 ||y(/)-x*(OI|=0 V te [0, T]
i.e., y(*) = x*(*). This shows that x*(-) is the unique solution of (2). —
49 Remarks :
m
1. The sequence [P x 0 (*)] that converges to the solution x*(-)
satisfies
52 x(T) = z
Then f, also satisfies the hypotheses of Theorem (25), so that (53) has a
unique solution over[0, T]. Denote this solution by y(-), and define z 0 =
y(T). Then one can easily verify that the functions y^-) defined by
satisfies (51) and also satisfies (52). To prove the uniqueness of the element
z 0 corresponding to a particular z, assume by way of contradiction that
there exist two functions >',(•) and y 2 (-) in C”[0, T] that satisfy (51) and
(52). Let y x (0) = zu y 2 (0)
= z2 . Then the functions y„(-) and y6 (-) defined
by
56 y a (t) = yi(r - t)
57 y b (t) = y (T - 2 t)
must both satisfy (53). However, because the solution to (53) is unique, it
59 = x(t)];
x(t) x(0) = x
f[t, 0
Then for each e > 0, there exists a T) > 0 such that c5(e,
ous on Rn . In the same vein, let y/: R n — > R n be the mapping that
associates, with each initial condition x0 e R n the value at time T
,
tinuous.
Also, for every x 0 there exists a finite h T such that (27) holds. Hence, by
,
Theorem (25), (69) has a unique solution over [0, oo) corresponding to
each x 0 moreover, this solution depends continuously on x 0
; .
72 II
A(0 II,
< kT , V / G [0, T]
Hence we have
73 ||A(r)x — A(0y||</cr ||x - y II, V x, y e R"; V t e [0, T]
74 II
A0)x 0 1|
< kT \\
x0 1|, V(e [0, T]
x{t) = —x 2
; x(0) = 1
:
This equation has a unique solution over [0, oo) [namely, x{t) =
1 j(t + 1)], even though the function f(x ) = x 2
is not globally Lip-
schitz-continuous.
if and only if, for each finite T there exists a finite kbT such that
dx
< kT , V x e R, V t e [0, T]
i.e., df(t, x)/dx is bounded independently of x over each finite interval [0, T].
dfi(t, x)
n
^ kT i
j V i,j, V X G Rn , V t e [0, T]
dxj
3.5
SOLUTION ESTIMATES FOR LINEAR EQUATIONS
In this section, we shall give a method for obtaining both upper and
lower bounds on the norm of the solution of a given differential equa-
tion. The Bellman-Gronwall inequality (Appendix I) does give an easily
applicable upper bound on the norm of the solution of a linear differ-
ential equation, and a similar inequality known as Langenhop’s inequal-
ityprovides a lower bound. However, both of these bounds suffer from
the deficiency of being sign-insensitive ; i.e., they give exactly the same
Sec. 3.5 Solution Estimates for Linear Equations 89
estimates for
x(Y) = A(/)x(/)
and
x(0 == — A(/)x(f)
They do so because both the Bellman-Gronwall inequality and Langen-
hop’s inequality utilize ||A(f)||, which is, of course, sign-insensitive.
In contrast, the method given here utilizes the concept of the matrix
measure, which is sign-sensitive. As a result, the bounds derived in this
section are always “tighter” than (or the same as) those given by the
Bellman-Gronwall and Langenhop’s inequalities.
11 x(/o) II
exp - Hi- A(r)] c/r| <
{£
II
x(/) || < ||x(/ 0) ||
exp
{ £
^,[A(t)] JtJ
where o(<5) is used to denote an error term with the property that
italMM _o
<5 -+ 0
+ O
x(t + d) = [I +
SA(t)]x(t) + o (5)
4 By this we mean that, over each finite interval, A(-) is continuous at all
except a finite number of points; at each point of discontinuity, A(«) has well-defined
left and right limits and the norm of the difference of these limits is finite.
;
£
9 £||x(r)|| = lim
1|x(f
< ^[A(0]||x(0il
where d + /dt[ ] denotes the right-hand derivative. If we multiply both sides
10 exp j—
The proof of the left-hand inequal-
we get the right-hand inequality in (3).
13 A(rt = r- 2 * ll
L-i -rj
14 x(0) = [l 0]'
15 y«„[A(0] = /UA(01 = ~t + 1
Thus, applying inequalities (3) with each of the above measures gives
19 exp (-t - t )
2
< |x t (0l + 1
x 2 (t) |
< exp (--t — -2 -
>
for ||
• iii»
20 exp (-* - t
2
) < |
xi(t) |, |
x 2 (t) |
< exp (-t 2 j
)
for ||
||oo, /tj-oo
Thus the same two inequalities (3), when applied with different mea-
23
24 x(0) = [l 2]'
25 x(0 = li~e
-‘ 2
- ^e 5,2 ' !
26 /*n[A(0] = -t\ = 51
30
2
a/3” exp (— 2.52r ) < {| x (0
;
2
+ |
x 2 (t) |
2 1/2
} < ^/T exp (-1.49/*)
for ||x(/)|| 2
2 exp (— 3t 2 ) < |
*,(/) |, |
x 2 (0 1
< 2 exp (-t 2
)
for ||
x(/> |U
These bounds are depicted for the case t = 0.5 in Figs. 3.6, 3.7, and
3.8.
1
34 1 1
x(t ) ||
< exp £ a,[A(t)] rfrj
|
1 1
x(f 0 ) 1
38
39
If
||x(/ - 15)
JM£)J1
So
< ||i - 5A(<)IW|x(t)|| + £||v(OII + o(S)
||
- \Ml -
^[||x(0||]> -M-A(0]||x(0ll
we
^)ll
> -1 III - ^AC
those in (3). In (34) and (35), the first term on the right-hand side
while the
can be thought of as the effect of the initial condition x(f 0 ),
of the forcing function
second term can be thought of as the effect
v(0-
We close this subsection with an observation on systems of the
type (2), i.e., unforced linear systems. Suppose x(/ 0 ) ^0, so that
t, the extreme left-hand
side of (3) is
||
x(/ 0 ) ||
> 0. Then for all finite
while the extreme right-hand side is always
always strictly positive,
means that a system of type (2) cannot
finite. Physically speaking, this
have a finite settling time (i.e., it cannot go from a nonzero initial state
42 x* +1 = Akxk
settling time (though not finite escape time). For
can exhibit finite
—1 "^(0" ~2~
pCi(0"j r-2 + 2 sin/ 1
~*i(0)
^ _*2(0- - 2 —3 — 2 cos t_ _x 2 (0- ’
-* 2 (0)- _ 1 _
Problem 3.12. Calculate upper and lower bounds for ||x(/)|| l5 ||x(0IU,
||x(0l|oo, given
“
~ii(0~
~ — + 2 sin
t t 0 ~*i(0~
" 1 ~
#
only some of the many varieties of techniques that are available for the
approximate analysis of nonlinear systems. Moreover, even with regard
Sec. 4. 1 Describing Functions 97
to the three subject areas mentioned above, the presentation here only
scratches the surface, and references are given, at appropriate places, to
works that treat the subjects more thoroughly.
4.1
DESCRIBING FUNCTIONS
form (1) memoryless is that the output at time t [namely (7Vbc)(r)] depends
solely on the input value at the same time t [namely x(t)] and not on —
the past or future values of x(«).
The problem studied is the following: Suppose
in this subsection
that a particular function which might be called the
x 0 («) in C[0, oo),
2 (JVxoXO = f
Wo (t — t)xq(t) dx
Note that (Wx Q )(>) can be interpreted as the output of a linear time-
invariant system with input x 0 (-) and impulse response w 0 (>y. Thus the
objective is to approximate the output of the nonlinear operator
[namely (iVx 0 )(«)] by means of the output of a linear system [namely
(JFxoX-)]. Furthermore, we would like the approximation to be the
“best possible,” in a sense to be made precise next. We would like to
choose the impulse response w 0 («) in such a way as to minimize the
error criterion
assuming, of course, that the indicated limit exists and is finite. The error
criterion e is recognized as a measure of the average mean squared devia-
tion between (Nx 0 )(-) and (Wx <,)(•)•
In the sequel, we make the following assumption regarding the
nonlinearity N: For each finite constant b there exists a corresponding ,
4 |
(Nx)(t) |
<m Q (b) whenever x(-) e C[0, oo) and |
x(t) |
<b
Assumption (4) is not overly restrictive and places the
mathematical
manipulations that follow on a firm mathematical footing. For the same
2
reason, we restrict the impulse response w( ) by requiring that •
5 f |
w(7) dt |
< oo
Jo
These two assumptions, together with the restriction that the reference
input x 0 (-) is bounded, are enough to assure that all the integrals
encountered in the following arguments exist and are finite.
To determine the optimal choice for the impulse response w(-),
we use a standard variational argument. Assume that, for the problem
at hand, an optimum choice of w(-) exists, and denote it by w 0 (-)-
Suppose w(*) is any other impulse response; then, because we are as-
suming that w 0 («) is the best possible choice, we must have
Now, by definition,
— (Nx — o WoXo)
2
(t)\ dt
(W0 x 0 - Nx 0 )(t)\ dt
where we define
2 Assumption
(5) implies that the linear system represented by W is bounded-
input/bounded-output stable. See Sec. 6.3.
Sec. 4. 1 Describing Functions 99
11 [w V—0—W 0 {t
— r)]x 0 (r) ch
(W0 x 0 — Nx 0 )(t) dt = 0
Next, it is easy to verify that
12 f [w(t — t) — w 0 (t
— t)]x 0 (t) dr = f [w(r) —w 0 (t)]x 0 (7
— t) dr
Jo Jo
Substituting from (12) into (11) and interchanging the order of integra-
tion, we obtain
16 </>n(z ) = lim ~
r- ~ i Jr
f x a (t — T)(Nx 0 )(t) dt
tions x 0 (») and (^Fo^oXO? and similarly for N (>). With these definitions, (j)
17 V T >0
Equation (17) represents an important principle, namely: A
necessary and sufficient condition for an impulse response w 0 (») to be
optimal [in the sense of minimizing the error criterion in (3)] is that the
cross-correlation function between x 0 (*) and (fiE0 x 0 )(») is the same as
the cross-correlation function between x 0 (-) and (iVx 0 )(*)-
A few comments are in order regarding the above result: Given a
particular reference input x 0 (*), if some w 0 («) satisfies (14), this can be
taken to mean that a linear system with impulse response w 0 (*) is the
best possible linear approximation to the nonlinear operator N ,
with
the reference input x 0 (*)- However, (1) for a given x 0 («), there is in gen-
eral more than one w 0 (-) that satisfies (14), and (2) a function w 0 («) that
satisfies (14) for one choice of reference input x 0 (*) need not do so for
100 Chap. 4 Approximate Analysis Methods
18 x 0 (t) = V t > 0
for some real number k. Suppose also that the nonlinearity N is memory-
less and time-invariant; i.e., suppose that
22 *>„(*) = ^<5(0
A system whose impulse response is given by (22) is nothing but a con-
stant gain of value [n(k)/k]. Thus we have shown the following.
Conclusion. If N is a memoryless time-invariant nonlinearity of
the form (19), then an optimal quasilinearization of N with respect to
a constant reference input k ^0 is a constant gain of value [n(k)/k].
The latter constant gain is sometimes referred to as the equivalent lin-
earization of the nonlinearity N
with respect to the constant reference
input k.
Sec. 4. 1 Describing Functions 101
23 x 0 (0 = a sin cot
where we take a >
0 without loss of generality. In (23), we do not
include a phase angle [i.e., we do not choose x 0 (f) = a sin (cot + 0)]
because, as will be evident, 0 can be taken to be zero.
If N
is of the form (19), then (Nx 0 )(*) is a periodic function with
period Injco. Hence (Nx 0 )(>) can be expanded in a Fourier series, of the
form
24 (Nx 0 )(t) = cQ 2
+ i=i (c t cos icot +d
t
sin icot)
In general, there is no reason to assume that the d.c. bias term c Q in (24)
is zero. However, if the nonlinear element N, in addition to satisfying
(19), is also odd, i.e.,
25 n(— a) = —n(cj), V a e R
then not only is c 0 zero, but all the c-s in (24) are zero. In this case, (24)
simplifies to
26 (Nx 0 )(t) =2
i=i
d t sin icot
= if i =j
29 <5,7
li if i ^j
In other words, 0^(0 is the same as the cross-correlation between the
functions a sin cot and d t
sin cot. Thus (17) is satisfied if w 0 («) is chosen
such that
30 (W0 x 0 )(t) = d t
sin cot
31 w a (t) = A <5(0
i.e., if w 0 (-) corresponds to a constant gain of value dja.
102 Chap. 4 Approximate Analysis Methods
32 ya sin cot
Comparing the function (32) with the output (26) of the nonlinear ele-
ment N we ,
see that if we choose y = dja ,
then the first harmonic of the
where
34 + jy = w 2 0 (jco)
35 yx a = d\
36 y2 a — cx
39 r,(a;o,) = [+j^
42 2)
*o (0 = a sin co 2 t
both of which have the same amplitude but different frequencies. If both
co and co 2 are nonzero, we can write
43 *Sf
,
(0 = 4 (^)
1,
memoryless, we have
44 (Nxrm =
Hence, if
45 (Nx 0v )(t)
(
= c0 + S cos iCO\t + dt sin iCO x t)
i=i
then
47 ll(a;C0 1 ) =t](a;C0 2 )
Because co x and co 2 are arbitrary, (47) shows that rj(a; co) is independent
of co. M
48 fact If «(•) is odd in addition to being memoryless and time-invariant,
then t](a) is a real number.
104 Chap. 4 Approximate Analysis Methods
proof If N satisfies (25) and x 0 (-) is of the form (23), then (Nx 0 )(-) is
FIG. 4.1
is (as long as a ^
0). The first harmonic of a square wave of amplitude
50 tjiia) = —
na
52 =m ri(a) 1
if 0 <a<8
However, if a > <5, the output of the nonlinearity is of the form shown
in Fig. 4.3. In this case, it can be verified through laborious but straight-
forward calculations that
Figure 4.2 depicts the nonlinearity n{>) in the case where m >m
1 2.
Sec. 4. 1 Describing Functions 105
However, expression (53) for tj{a) is valid for any m u m 2 (See also .
Problem 4.5.) Thus (52) and (53) completely characterize the describing
function ;/(•)•
get the dead-zone nonlinearity w 2 (-) shown in Fig. 4.4. The corre-
sponding describing function tj 2 (>) is obtained from (52) and (53) as
0 <a<8
54 ^O) = Y'l
a
+ a \ a 2) J
„>
Similarly, if we let m 2 = 0, we get the limiter nonlinearity n 3 (>) shown
in Fig. 4.5. The describing function of n 3 {>) is
0 <a<3
55 i2
[
—m
n
1
1
_
•
sm
— i
1
<5
a
.
b
5
a \
i
<5
a 2)
2
\
1/2
1
a > 8
for > x 1
56 /(*) A |-^-[sin 1
x + x(l - x 2 1/2
for 0 < x < 1
) ]
106 Chap. 4 Approximate Analysis Methods
A sketch of f(x) is found in Fig. 4.6. With the above definition, tj 2 (a)
57 rj 2 (a) =m 2
58 rh(a) = m,/^j
While (54) and (57) give the precise expressions for q 2 (a), one can
readily derive the approximate shape of t] 2 (a) just by using common
sense. If the input amplitude a is less than <5, the output is zero, so that
rj 2 (a) is also zero. On the other hand, as a becomes larger, the effect of
the dead zone becomes smaller, so that fi 2 {a) is a monotonically in-
creasing function of a. As a becomes extremely large compared with 8 ,
the effect of the dead zone all but disappears, and n 2 (’) begins to look
like a linear element of gain m2 . Hence t] 2 (a) —m >
2 as a — * oo. The
characteristic shape of t] 3 (a) can be rationalized in the same way.
We shall close this subsection by deriving some bounds on the
describing function. Suppose the function «(•) in (19) satisfies a condi-
tion of the form
59 k x cr
2
< on(<j) < k a 2
2
, V o e R
for some real numbers k x and k 2 (We say that . n(>) lies in the sector
Sec. 4. 1 Describing Functions 107
[k 1? k 2 ].) If, in addition, «(•) also satisfies (25) [i.e., «(•) is odd], we show
below that
k i
< 77 (a) <k 2, \/ a e R
In other words, if «(•) lies in the sector [k u k 2 ], its describing function
also lies between the limits k and k 2
{
.
the above result states that if the graph of a nonlinear element lies
between two straight lines of slopes k and k 2 its describing function x ,
The bound (60) is rather easy to prove. Because /?(•) is odd, 77 (a)
is real. Moreover,
1n!<o
=—
f*
=
na\
n^a S^ n ^ S^ n^ ^ (letting 9 = cot )
>
~— f k (a x sin 0)
2
dO [by (59)]
Jo
=k x
FIG. 4.8
sible to have a nonzero periodic solution for x(-). If the two blocks
“N” and “G” are described by the input-output relationships
then, with zero input, the closed-loop system is described by the in-
tegral equation
where
66 tl(a;co) =y l +jy 2
is the describing function of «(-)• Next, let
61 i(jco) = g (co)
r +Mi(co)
Sec. 4. 1 Describing Functions 109
71 1 + u + jv = 0
This procedure is also referred to by some authors as the principle of
harmonic balance. Note that it is a very approximate procedure because
even if (63) has a periodic solution, it is highly unlikely that this periodic
solution a pure sinusoid of the form (64). However, the rationale is
is
73 x 0 (t) =a 0 sin co 0 t
74 x (t) = -(PGNx)(t)
where P is the operator that associates, with each periodic function, its
first harmonic. Of course, (74) is not the same equation as (63), because
there is no operator P in (63). However, if the operator G is a low-pass
filter , i.e,. if G rapidly attenuates the higher harmonics of (Nx)(-), then
(so the reasoning goes) there is not much error in replacing (63) by
(74), and of course x 0 (*) solves (74).
110 Chap. 4 Approximate Analysis Methods
Special Case
75 on(c) >0 , V o e R
i.e., in the case of the so-called first and third quadrant nonlinearities , the
harmonic balance equation (72) is particularly easy to solve. In this
case, the describing function rj is real, nonnegative and is independent of
co. Thus (72) reduces to
76 1 + g (m)tl(a) + jg,(co) = 0
r
77 + g (co)ri(a) = 0
1 r
78 giico) = 0
79 rj(a) = .
~ x
' g(jCOi)
Let a i \ (
. . . ,
a i] be the solutions of
{
(79), if any. Then we predict that
the system (63) has periodic solutions at frequencies co u ,
co m , with
amplitudes a[°, a fi corresponding to the frequency c»
. .
. ,
(
t
..
{
Two things are to be noted about the above procedure. (1) The
possible frequencies of oscillation co 1 , ... ,co m do not depend on the
nonlinearity «(•) in any way; rather, they are determined solely by the
transfer function g(-). (2) Once the possible frequencies of oscillation
co u .co m are determined, the possible amplitudes corresponding to
. . ,
each frequency are easily found from (79). Thus, in the case of odd non-
linearities, the harmonic balance equation (72) is effectively decoupled.
§(s) = s(s
+ lX-y + 2)
«(f) =£ 3
Sec. 4. 1 Describing Functions 111
and tj{a) is independent of co. Thus the harmonic balance equation (11)
becomes, in this case,
i 1 3a
©I i
^ jco(l+jco)(2+jco) 4
+ gr(co)fj(a) =
1 0
82 g {co) = 0
t
co = */2
Note that we ignore the solution co = —*f~2 because, in physical
terms, it is no different from the solution co = */2 Now, . substituting
co = into (77) gives [because gr (*/2) = —1/6]
Problem 4.3. Verify that the nonlinearity n(-) in Fig. 4.2 is the sum of
the nonlinearities « 2 (-) and az 3 (-) (in Fig. 4.4 and 4.5, respectively). Verify that
the describing function tj(>) is the sum of // 2 (-) and tJsO).
112 Chap. 4 Approximate Analysis Methods
Problem 4.4. Using the results of Problems 4.2 and 4.3, derive an
expression for the describing function of the dead-zone limiter shown in
Fig. 4.9. ( Answer : tj 4 (a) = m[f(S 2 /a) — /(St/a)].)
Problem 4.5. Using the results of Problems 4.2 and 4.3, derive an
expression for the describing function of the piecewise-linear element shown
in Fig. 4.10. [Answer:rj 5 (a) (m 1 2 )f{d 1 la) ( =
2
—m
m^)f(S 2 la) 3 .] + m — +m
Generalize the answers to a piecewise-linear element with / different slopes.
Problem 4.6. Let «(•), «/(•), and «„(•) be odd functions such that
88 g(s) = 5(3 + 1)
s 2 (s + 2) 2
Analyze the possible existence of periodic solutions in the cases where
(a) «(•) is the sign nonlinearity of Example (49),
(b) «(•) is the dead-zone nonlinearity of Example (51), with 5 = 1 and
m2 = 2, and
(c) n(-) is the five-segment nonlinearity of Fig. 4.10, with m = 2,
x
m2 = 1, m = 3 0, 5 1 = 1, and S 2 = 2.
Sec. 4.2 Numerical Solution Techniques 113
4.2
NUMERICAL SOLUTION TECHNIQUES
In this section, we shall discuss some methods for approximately solving
a given set of nonlinear vector differential equations. The techniques
presented here are particularly well suited for implementation on a
digital computer. Due to limitations of space, we shall discuss here only
the rationale of the various methods and omit the more detailed numer-
ical analytic aspects, e.g., an analysis of the errors created by the appro-
ximations. For an in-depth discussion of these points, see [8] Chua and
Lin.
Throughout this section, we are concerned with finding (approxi-
mately) the solution to the nonlinear vector differential equation
which assures (see Theorem [3.4(25)]) that (1) has a unique solution over
[0, oo). Let x*(?) denote the actual solution of (1) and let (?„)f be a
monotonically increasing sequence of real numbers with t n approaching
oo as n — > oo. Our objective is to find various expressions that approxi-
mate x*(t„) n 9
= 1,2,... . That is, instead of attempting to approximate
the function x*(f), which is defined for all values of t > 0, we attempt
to approximate the sequence of vectors [x*(f n )]“. This is in keeping with
themodern trend toward the increased use of digital computers. In
many instances, the numbers t n are equally spaced i.e., ,
2 tn = nh
- for some number h > 0, which is referred to as the step size.
The various derivative terms in (3) can now be calculated using the
differential equation (1). For instance,
4 x*(0 = f(f^.xj)
114 Chap. 4 Approximate Analysis Methods
5 X*(/„) =
=f x(t„, x*)-f (t„, x*) +f ,(t„, X?)
that are retained determine the order of the Taylor series method if —
the series in (6) is truncated after the term involving h k where k is some ,
7 x„ +1 = x„ + hi{t„, x„)
Hence
11 h{a t f(t„, x„) +a
+ ah, x„ + )Sf
2 f[tn x„)]}
Comparing the right-hand sides of (8) and (11), we see that they differ
only by o(h 2 ), provided
12 a + a2 —
l
1
13 a2 a = -j
14 a 2 fi =A
Because (12)— (14) represent three equations in four unknowns, we can
assign onenumber arbitrarily. Thus (12)—(14) can be rearranged as
15 = — 1 a2
16 a = J_
2a2
= h
17 fi
2a 7
provided a 2 ^
0. Therefore the general second-order Runge-Kutta
algorithm is described by
18 = x, + h {(1
n a 2 )f(t„, x„) + a 2 f \n + _A_
i
’
xn .
M (t n , x„y
2a 2
'
la 2
116 Chap. 4 Approximate Analysis Methods
X„ +1 = x„ + h
{
f(
y n)
+ + h,\„ + hf(t„, x„)]J
Similarly, by setting a2 = 1, we get the modified Euler-Cauchy algori-
thm, which is described by
h_ hf(t n9 x„)
M
'
= x„ L n\
,
2 ’
Y" ,
'
It should be once again emphasized that formulas (19) and (20) do not
give identical results.However, if we start from the same x„, the x„ +1
2
given by (19) and that given by (20) would differ only by o (h ).
The most commonly used of the Runge-Kutta algorithms is one
3
of order 4, which is described by
where
x„, h) = go + 2gi + + g3
g 4 (t„,
go = X„)
g, = r( + A, X. + If)
r,
gl = f(l. + X„ + If)
g3 = f + h,x + hg
(t n n 2)
29 M(0 = e Kt =2A 1
'
We shall now show that when applied to the differential equation (28)
the fourth-order Runge-Kutta algorithm exactly reproduces the first
30 M(A) 2A'i
= i=o Z
34 g = 3 a[i + (a +A 2 A+A 3
= A + A2h + A 3 A+a A 4
35 g4 = 2G t + 2G 2 +G 3
^ i
_ ^2 h h2 ,
/z
3
6 2 6 24
36 Xjv+i = I + AA + A 2 A + A A + A*A 3
Thus, clearly, contains the first five terms in the Taylor series (30).
Incidentally, this example illustrates that the Runge-Kutta algori-
thm is applicable not only to vector differential equations but also to
matrix differential equations.
i.e., that
39 x(t) = c0 + cf + c2t 2
Then
40 xn = Cq + c t nh +c 2n
2
h2
41 x{t) = c1 + 2c 2t
so that
42 f(t n x„),
= c + lcx 2 nh
Substituting from (40) and (42) into (37) and choosing n 0 for con- =
venience, we obtain what is known as the exactness constraint for the
algorithm (37), namely,
X
+ i=- 1
hbi[c t + 2c 2 (—ih)]
k k
45 Cl = XI —ia c i 1 + =X-l bfil
i =0 i
k k
46 c2 = i
1]
=0
i
2
GiC 2 + =2-1 —2ibiC 2
I
Once again, because we would like (43) to hold for all second-order
polynomials, (44)-(46) must hold for all c 0 , c l9 c 2 . This implies that
47 2= a t
= 1
1 0
48 XI — i^i + XI b t
= \
i=-l i =-
49 X i2a + X- i
=
—216, = 1
1 =0 f 1
Thus on the 2k
(47)-(49) give three constraints 3 constants a 0 + ,
a k b_ u
,
b k and any choice of these constants such that (47)-(49)
. . . , ,
50 a0 = l
51 +b = 0 1
52 2b _ = x
1
53 ao = I? 1
=— , b2 =—
o io , ,
55 x(t) =2 CjV
= j
Then
m
56 x(t) = 2 jc j
tJ l
= j
Therefore,
57 Xi = 2 c ji J h j
= j
m
58 f(t„ x,) = j=£ jcjp- 'h'~' 1
Substituting from (57) and (58) into (37) and taking n =0 for con-
venience, we obtain
59
m
£ Cjti = £=
km£ at Cj(—ihy
j= 0 1 0 7 =0
+ —£~l i
bt £= jcj(—ihy-
7 1
l
h
60 1 =2 tf/(— 0 J + S bJi—iy- 1
, , j = 1, . .
.
m
= i 0 = -i »
k
61 1 =2 ai (corresponding to j = 0)
i= 0
If equality holds in (62), then we can solve for the a-s and b- s uniquely.
Now, if ra is odd, one can always choose k such that
63 2k ~f~ 3 = ra -j- 2
,
Adams-Bashforth Algorithms
67 a t
=0 for i = 1, . .
.
k
68 =0
Thus there are exactly k +2 constants to be determined, namely a 0 ,
69 #0 = 1
70 x„ +1 = x„ + hi(t x„) n, (m = 1)
71 x„ +1 = x„ + h[jf(t„, x„) - if((„.!, X _ n j)] (m = 2)
72 x„ +I = x„ + x„) - f|f (?„_!, x„_,) + x„_ 2 )]
(m = 3)
73 X„ +I = X„ + X„) - XB -l) + 24 f(*«- 2 > x„- 2)
- x„_ 3 )] (/M = 4)
Adams-Mou/ton Algorithms
77 X„ + , = x„ + hf(t„ +1 x„ +1 , ) (m = 1)
78 x„ + 1
= x„ + h[{f(t n+i x„ +1 ) + , x„)] (m = 2)
79 x„+i = x„ + h[^f(t„ +1 x„ +1 ) + , (t„, x„) - Tzf(t„-i, X„_,)]
(m = 3)
80 x„ + i = x„ + h[^f(t„ +1 x„ +1 ) , + x„) -
+ X„- 2 )] (m = 4)
Predictor-Corrector Algorithms
81 x„ + 1 = y„ + x„ +I )
85 x<'+ 1
,)
= x„ + X' + + x„)] i) (corrector)
In the foregoing, (84) predicts the value of x n+1 using an explicit for-
mula, while (85) corrects this predicted value of x n+1 In practice, if the .
/
/
/
x
/
/
t
/
I i
iii i i
>
FIG. 4.11 f
0 t\ 1 ^3 *4 t t t
5 6
43
SINGULAR PERTURBATIONS
In this section, we what is known as the problem of
shall briefly study
Note that for any value of e other than zero the system (l)-(2) consists
g[x(0, y(01 =0
Suppose it is possible to solve the m equations comprising (3) to obtain
an explicit expression for y (t) in terms of x(t), of the form
y(0 = h[x(/)]
where h: R" — Rm
* . Then (1) and (3) together reduce to
Setting e —
0 in (2) is called a singular perturbation because it
completely changes the nature of (2), viz., from a differential equation
to an algebraic equation. The objective of the theory of singular pertur-
bations (stated in very simplified terms) is to examine the simplified
system (5) and from this to draw conclusions about the original system
(l)-(2) with 6^0.
Physically speaking, singular perturbations arise from an attempt
to approximate a high-order nonlinear system with another one of
5
lower order. The following example, due to [9] Desoer and Shensa,
illustrates this point.
*1
1 H
FIG. 4.12
dance with standard practice, the inductor current x and the capacitor l
voltage x 2 are chosen as the state variables of the system. Using the
current and voltage laws of Kirchhoff, the dynamic equations of the
circuit in Fig. 4.12 can be written as
0 = x, -*»
r
x2 = rx x
Xi = (~ 1 — r)x x
Now, it is clear from (11) that, as long as r > —l, the solution x of t
X2 _ _l/e — l/er _x 2 _
Letting A denote the 2x2
matrix in (12), one can easily verify that
if r <0 ,
then for sufficiently small values of the parameter e the matrix
A has at least one eigenvalue with positive real part, so that the original
system (7)-(8) is unstable in this case. 6 In other words, if r e (—1, 0),
then the simplified system is stable, but the unsimplified system is
6 See Sec. 5.3 for a full discussion of the stability of linear systems.
2 : 6
Ai Al2
15 A(e) = i
19 1 /0 — Xi€ \
< r whenever 0 < € < e0
7 Note that when we solve the equation det [XI — M(e)\ = 0, the resulting n
solutions for X can be numbered in any order we choose.
,
Let {a 1# . .
. ,
a m} denote the (not necessarily distinct) eigenvalues of A 22 ,
21 P <r,t\ i = 1, . .
. , n
whenever
22 i =n+ 1, +m 0 < € < 60
26 At
<Xi-n
€
< € ’
/ =«+ 1, ...,« + m, whenever 0 < € < €x
27 1 62/ 1
< r, i = n ,
whenever 0 < 6 < €1
Note that the only change between (20) and (28) is that the second “row”
of matrices has been multiplied by e, which is permissible because 6^0.
Now, A i (6; A) is continuous in € at € = 0, and
29 A ff\ \
Ai(0; X)
1
= A
det
TAl1 ~ ^
A2 i A 2 2_
Noting that the determinant of the matrix
I
0“
30
— 22
!
A2 i
Ij
is 1, we have
"
An - Al Ai 2
r
I OF
31 det
A2 | A 22 _ A 22 A 2 .ji i
i
Ai — Ai 2 A 22 A 2 1 i < i -
det
0 A 2 2_
det (An — Ai 2 A 22 A 21 — Al) det A 22
where {7 1 , . .
. /„} is a subset of {1, . .
. , n + m}.
Now, if we pick the numbers r, € 1 ,
and e 2 carelessly, it is possible for the
same X to t satisfy both (26) and (32). However, suppose we choose r as
_ —min; [Re Oj
33 r
2
They study the behavior of system (39) as e —> 0 and fi —> oo. In this
case, the second equation in (39) becomes an algebraic equation, while
the third equation reduces to
40 x3 =0
The results proved by Desoer and Shensa are as follows: Define
41 B = An A 12 A 22 A 21
and let C be the matrix resulting from simplifying the set of equations
0 = AjjXjl T A 12 x + A x-
2 13 3
42 0 = A 21 Xj “T A 22 x + A 23 x 2 3
= A Xj + A 32 x + A 33 x
jli± 3 31 2 3
Ah A 12 Aj 3
45 A(6 ;
ju) = A 2 i/ 6 A 22 /f A 23 /6
_A 3 i//j a 32 fii A 33 /^_
have negative real parts whenever 0 <e< e0 and ju > ft 0 .
tems, the results derived here form the basis for the results given in Sec.
JtXll 0 -1 — 3_ _*2 _
€X 2 = 0-2-1 1 0 *2
5.1
BASIC DEFINITIONS
In this section, we shall state the precise definitions of the various
concepts of Liapunov stability; in addition, we shall introduce the vari-
ous kinds of positive definite functions.
131
]
shall assume that the function f is of such a nature that (1) has a unique
solution over [0, oo) corresponding to each initial condition for x(0)
and that this solution depends continuously on x(0). This is the case,
for example, if f satisfies a global Lipschitz condition (see Sec. 3.4).
Recall that an x„ e R" is said to be an equilibrium point of the system
(1) at time t 0 if
2 f(t, x 0) = 0. V t > t0
where
4 f ft, z) =f (t, z -x 0)
5 f(f, 0) = 0, M t> to
As discussed in Sec. 1.2, this means that if the system (1) is started
ing: Suppose the initial state x(t 0 ) is not 0 but is “close” to it; what is
condition ||
x(r 0 ) ||
such that any solution trajectory of (1) starting at
time tQ from an initial state lying in the ball of radius 8(t 0 e) always ,
lies in the ball of radius e at all times t>t 0 . Another way of stating
this is as follows : Arbitrarily small perturbations (about the equilibrium
state 0) of the initial state x(£ 0 ) result in arbitrarily small perturbations
of the corresponding solution trajectories [of (1)].
To give yet another interpretation, let Cn [t Q , oo) denote the linear
space of continuous ^-vector- valued functions on [* 0 ,
oo), and let
BC”[t 0 oo) denote the subspace of
, C n
[t 0 , oo) consisting of bounded
continuous functions. If we define the norm
I|x(.)|| t = sup II *(011, V x(.) e BCn [t 0 ,
oo)
[3.1(52)]), one can see that the stability status of an equilibrium point
does not depend on the particular norm used to verify (6).
Once the notion of stability is understood, it is easy to understand
what uniform stability means. According to definition (6), the equilib-
rium point 0 is stable at time t 0 if, for each e > 0, a corresponding
8 can be found such that (6) holds. In general, this 8 depends on both
e and / 0 However, if a 8 can be found that depends only on e and not
.
— > oo as
t —
> oo. While this is one way in which instability can occur,
means that, given any e > 0, one can always find a corresponding <5
such that (7) holds. Therefore, the equilibrium point 0 is unstable if,
for some e > 0, no 8 > 0 can be found such that (7) holds. Physically
speaking, the equilibrium point 0 is unstable if there is some ball B e
of radius e centered at 0 such that for every 8 0, no matter how small, >
there is a nonzero initial state x(/ 0 ) in Bs such that the trajectory starting
at x(f 0 ) eventually leaves Be This and only this is the definition of insta-
.
12 X 1
= —X 2
13 x2 = —x + 1 (1 — x\)x 2
librium point of this system, and as such, the solution trajectory starting
from Xj(0) =x 2 (0)
=0 is given by x (t) = x 2 (t)
t
0 = V t > 0. How-
ever, solution trajectories starting from any nonzero initial state all
approach the limit cycle, as shown in Fig. 5.1. Now, let us study the
stability status of the equilibrium point (0, 0) using definitions (6) and
(9). If we choose 6=1, say, then B = e {(x u x 2 ): (pc\ + xl) l/1
< 1} lies
entirely within the region enclosed by the limit cycle. Therefore all
Note that all trajectories of the system, no matter what the initial state
of the system is, are bounded, and none “blows up,” so the system is
lr
The last statement can be made more precise, as follows: Given any e > 0,
there exists a T(f) < 00 such that
O
^ > Position r
FIG. 5.3 r = 0
where we initially place the ball, it does not move. If we define the norm
of the state vector x = [r r]' as
20 ||
x ||
= (r
2
+r 2
)
1/2
suppose the ball is displaced to the dotted position and is let go (with
zero initial velocity). In the absence of friction, the ball will roll down
through the equilibrium position and up the other side to the same height
138 Chap. 5 Stability in the Sense of Liapunov
displaced to the dotted position and released, it will roll back and forth,
but because of the energy dissipation due to friction, the amplitude of
the oscillations gradually decreases and approaches 0 as t — » oo. In
r =r 3 ,
r =0 is stable. The details are left as an exercise.
we know that c(t 0 ) is a finite number, for any fixed t0 . Thus, given any
x n
such that (8) holds. Therefore the equilibrium point x = 0 is not uni-
formly stable over [0, 00 ).
proof Let s(t, x0 , t) denote the solution of (1) together with the initial
condition
29 x(T) = x0
as evaluated at the time t. The fact that the equilibrium point 0 is stable at
time 11 means > 0, there exists a
that, for every 6 <5 > 0 such that
30 ||
0 1|
< 5=>||s(/,x 0 ,fi)|| < € V t>t > x
3
The remainder of Sec. 5.1.1 may be skipped without loss of continuity.
1 1 ] 1 1
36 ||s (t, x0 , t) II
< 6, \f t> T whenever ||x 0 1|
< r0
Here we have made use of the obvious fact that 3 < €. Because the above
argument can be repeated for any e > 0, it follows that 0 is a stable
proof Clearly (ii) implies (i), so it remains only to show that (i)
40 /*(x 0 , t) = sup ||
s(t + y, x0 ,
t) ||
y>0
Because 0 is a stable equilibrium point all times T e [0, T], the function
jh(xq, t) is continuous at x 0 = 0, for each T (see Problem 5.4). There-
fore, if
then //(•) is continuous at x 0 =0. Thus, given any € > 0, we can find a
43 1
x0 1
< <5, T E [0, T] => 1
s(t + y, x 0 , t) 1
<6 vy>0
4 Note that, because of the periodicity of the function f, we have that if 0 is
51 B r
= {x: ||x|| < r}, r> 0
class K if
(i) a(*) is nondecreasing,
(ii) oc(0) = 0, and
(iii) cc(p) > 0 whenever p > 0.
Note that some authors replace (i) above by the more stringent
requirement that <%(•) is strictly increasing. It turns out, however,
54 (i) W(0) = 0
55 (ii) W(x) >0, Vx^O belonging to some ball B r, r > 0 [where
the ball B r is defined in (51)].
56 (iii) W(0) =0
57 (iv) W(x) >0, V x 7* 0
58 (v) W(x) > oo — as ||x|| — » oo, uniformly in x.
sense of definition (47). Then clearly and (ii) above hold. To prove the
(i)
The rest of the proof is the same as in the case of l.p.d.f.’s and is left as an
exercise. wm
5
The proof of this statement involves the so-called compactness property of
the unit ball in R n and can be found in any standard text on real analysis; see, e.g.,
[20] Royden.
Sec. 5. 1 Basic Definitions 143
where B r
is a ball in R n
.
Wi(x u x 2) = x\ + x\
144 Chap. 5 Stability in the Sense of Liapunov
^(x) = ||x|P
if we take ||
• || to be the Euclidean norm on R2 .
unbounded as t increases.
as t —
> oo This example shows that it is not
.
possible to weaken (63)
to the statement
76 Remarks :
(1). If we keep the same V but change the system (1), the
-(2/i + 3) h
FIG. 5.6
V(t, x) = x'M(t)x
where M(f) is a real symmetric n x n matrix, and t h-> M(0 is continuous.
Find the necessary and sufficient conditions for V to be a p.d.f. [Answer:
inf,>o AminMW > 0, where 2 min denotes the smallest eigenvalue.]
(i) W(x u x 2 ) = x\ + x 2
(ii) W(x 1} x 2 ) = x\ + x\
(iii) W(x u x2 ) = (x + x\) 2 t
5.2
LIAPUNOV'S DIRECT METHOD
The idea behind the various Liapunov theorems on stability, asymptotic
stability, and instability is as follows: Consider a system which is
“isolated” in the sense that there are no external forces acting on the
system. Equation [5.1(1)] is a suitable model for such a system because
no input is explicitly identified on the right-hand side of [5. 1(1)]. Suppose
that one can identify the equilibrium states of the system and that 0 is
one of the equilibrium states (or possibly the only equilibrium state).
Suppose it is also possible to define, in some suitable manner, the total
energy of the system, which is a function having the property that it is
zero at the origin and positive everywhere else. (In other words, the
energy of the system has its global minimum at 0.) Now suppose the
system, which is originally at the equilibrium state 0, is perturbed into
a new nonzero initial state (where the energy level is positive, by assump-
system dynamics are such that the energy of the system is
tion). If the
nonincreasing with time, then the energy level of the system never
increases beyond its initial positive value. Depending on the nature of
the energy function, this may be enough to imply the stability of the
equilibrium point 0. On the other hand, if the system dynamics are such
that the energy of the system is monotonically decreasing with time and
the energy eventually reduces to zero, it may be possible to conclude
the asymptotic stability of the equilibrium state 0, under suitable
additional assumptions. The theorems of Liapunov, as well as their
generalizations proved by later researchers, cast these ideas into a
mathematically precise form. In addition, they offer the flexibility of
allowing one to choose the energy function corresponding to a particular
system.
In this section, we shall present the three basic kinds of theorems
stemming from the so-called method of Liapunov, namely:
direct
stability theorems, asymptotic stability theorems, and instability theo-
rems. These are illustrated by several examples, some of which are of a
dynamics nature. These examples serve to show that, in some cases,
the total energy stored within the system is a logical choice for the
Liapunov function 6 of the system.
V(t, x) >a (|
|
x 1 1), \/ t>t 0 V ,
x.e B s for some ball B s
—
where a R > R is a function of class K. To show that 0 is a stable equilib-
:
rium point at time / 0 ,we must show that, given any € > 0, we can find a
S(t 0 €) > 0 such that [5.1(7)] holds. Accordingly, given e > 0, let
,
=
min [e, r, s} and pick 8 > 0 such that
9
Now, because
a(||x(0ll) < V[t, x(0]
(5) and (6) together imply that
1 1
x(r) 1 1
<6j <€, V t>t 0
Hence [5.1(7)] holds and 0 is a stable equilibrium point at time f0 - mm
m < co whenever 0
Thus, letting €i = min {e, r, s, d], we can always pick 8 > 0 such that
<5<d
m < a(60. One now proceeds exactly as in the proof of Theorem (1) to
show that [5.1(8)] holds with this choice of <5. The details are left as an
exercise. —
13 Remarks :
15 0 -j- sin 9 —0
:
16 x i
—X 2
17 x2 — sin Xj
Now, the total energy of the pendulum is the sum of the potential and
kinetic energies, which is
18 F(x l5 x 2 ) = (1 — cos x ) + t
where the term represents the potential energy and the second
first
represents the kinetic energy. One can readily verify that V is an l.p.d.f.
and is continuously differentiable, so that V is a Liapunov function
candidate for applying Theorem (1).
Next, we have
19 F(x l9 x 2 ) = sin XjXi +xx 2 2
21 Xj =x 2
22 x2 = -f(x - g(x,) 2)
where f(x 2 ) represents the friction and g(x,) represents the restoring
force of the spring. It is assumed that / and g satisfy the following
conditions
(i) / andg are continuous.
(ii) Whenever a belongs to some interval [— <7 0 ,
we have
25 V(x u x 2 ) = da
^+ J*
g(<r)
Note that the energy dissipated through friction is not “stored” in the
system (because it is not recoverable) and is therefore not included in
V. Note also that V as defined in (25) is a natural generalization of that
defined in (18). Now, by virtue of the assumptions on g(*), it is easy
to show that V is continuously differentiable and is an l.p.d.f., and
is hence a suitable Liapunov function candidate for applying
Theorem (1).
Next, we have
26 V(x u x 2 ) = x x + g(x )x
2 2 1 1
= x [—f(x - g(x
2 2) t )] + g(x 1
)x 2
= -x f(x 2 2)
30 X x
—X 2
31 x2 = —x — 2 (2 + sin f)*i
sm -
f
t
-f-
34 W (xi 2 , x2) = x\ + x\
152 Chap. 5 Stability in the Sense of Liapunov
cos
35 V(t,x u x 2 ) = -Xl
t
+L 9v
2. V
x vx, +2+
4- jjfil •)
J
(2 + sin 02 si nt
cos t
~~ X2
2
(2 + sinf)
2 + 2x^2
, + 2(2 + sin t)
.2
1
cos t
+ sin tf (2
~~ _ Xl 4 + 2 sin + cos
2 f t
(2 + sin 0
2
< 0, V t>. 0, V X u x2
38 X\ = x2
39 x2 = —p{t)x — e~’x,
2
40 V(t, x u x2) = x\ + e x\ l
41 W(x u x 2 ) = xj + xl
we see that V is a suitable Liapunov function candidate for applying
Theorem (1). However, because V is not decrescent, it is not a suitable
Liapunov function candidate for applying Theorem (9). Hence, using
this particular F-function, we cannot hope to establish uniform
stabil-
Differentiating F, we get
Sec. 5.2 Liapunov's Direct Method 153
— &x\[ — 2p(t) + 1]
In the same way, Theorems (1) and (9) are also qualitative in the
sense that they provide some conditions under which the existence of
a suitable 8 can be concluded. However, in principle at least, the
conditions fi(t 0 , 8) < oc(e) [if Theorem (1) is being applied] or ft (8)
< oc(e) [if Theorem (9) is being applied] can be used to actually deter-
mine a suitable value of 8 corresponding to a given e. But in practice
this is rather messy.
46 1 1
s(t j + t, x 0 Oil
,
<€ whenever t > T(e\ [|x 0 ||
< S u and V tx > t0
*1
54 Chap. 5 Stability in the Sense of Liapunov
The hypotheses on V and V imply that there are functions a(-), /?(•),
and y(0 belonging to class K such that
47 «(||x||)< V(t,x)< Allxll) w
V t
_ .
48 V(t, x)
51 II
whenever ||x 0 1|
< 8 1 and t1 >t 0
To prove (52), assume by way of contradiction that
Then
54 0 < a(<5 2 ) < V\t\ + J 7
rti+T .
— V(ti, x ) + 0 |
V[r, s(t, x0 , ti)] dr
Jt i
<0 by (49)
57 a[| s(r,I
x0 , *i)||] < p(d < a(6) 2) by (50)
proof Same as that of Theorem (45), except that one can take r = oo
in (47). m
Remarks :
y(p)
— > oo as p — oo.
63 Xi = x^xj + x\ — 1) — x 2
64 X2 = X + X (X + A ~ 1)
1 l \
65 V(x u x 2 ) = x\ + x\
Then V is a p.d.f. Further,
70 V(t, x u x2) = x\ + (1 + e~ 2,
)xl
156 Chap. 5 Stability in the Sense of Liapunov
71 V(t x l5 x 2 )
,
= —2[x\ + x x2 t + x\{ 1 + 2^“ 2f )]
72 W(x u x 2 ) = 2{x\ +x t
x2 + x\)
Because W satisfies (61), it follows that 0 is a globally asymptotically
stable equilibrium point.
73 [definition] A set S in R n
is said to be the positive limit set of
a trajectory s(«, x 0 ,
/0) of [5.1(1)] if, for every y e S, there exists
a sequence (/„)7 such that tn —+ oo and s(t„, x0 ,
t 0) — > y.
74 [definition] A set M in R n
is said to be an invariant set of the
system [5.1(1)] if, whenever y e M and > t0 0, we have
proof We shall give the proof only in the case of autonomous systems.
Suppose S is the positive limit set of the trajectory s(-, x 0 t 0 ) of [5.1(1)] ,
(see Problem 5.5). Now, as n — oo, tn — » oo, so that the right-hand side
of (80) converges to an element of S, by lemma (76). Hence s(r, y, 1 1 ) e S
for all t > tx . tm
83 S = {xe Q c : V(x) = 0}
and let M
be the largest invariant set of [5.1(1)] contained in S. Then,
whenever x 0 e £2 C? the solution s(t, x 0 0) 8 of [5.1(1)] approaches as ,
M
t — * oo.
8 Note that we can take the initial time to be 0 without loss of generality,
because the system is autonomous.
»
With the aid of lemma (81), we can state and prove a sufficient
S = {xE£l c : V(x) = 0}
contains no trajectories of [5.1(1)] other than the trivial trajectory
x(t) = 0. Then the equilibrium point 0 of [5.1(1)] is asymptotically
stable.
S = {x e Rn : V(x) = 0}
contains no nontrivial trajectories of [5.1(1)]. Then 0 is a globally
asymptotically stable equilibrium point of [5.1(1)].
f(t, x) =f + (t T, x), V t, V X g R n
,
for some T> 0
Sec. 5.2 Liapunov's Direct Method 159
Define
91 5 = {x g Rn : V(t, x) = 0, V t > 0}
Suppose V(t, x) < 0, V/>0, V x g Rn and that contains no
,
namely,
93 x 1
= —X 2
94 x2 = —f(x - g(*i) 2)
95 V( Xl ,x 2 ) = £ + £g(Od{
Conditions (i)— (iii) above ensure that V is a continuously differentiable
p.d.f., so that V is a suitable Liapunov function candidate for applying
Theorem (87). Calculating V, we get
96 K(x l5 x 2 ) = -x 2 f(x 2 )
Thus F(xj, x 2 ) < 0, V (xi, x 2) g i^
2
. Moreover,
97 S = {(x u x2) G R 2
: V(x u x 2 ) = 0} = {(x u x 2) G i?
2
: x2 - 0}
100 Xi =
101 x2 =— 2 + (2 + sin 0*1
This system is clearly periodic with period 2 n. Let us choose V as in
(32), i.e.,
xi
102 V(t, x t,
x 2) = xj + 2 sin t
-f-
103
tV/.
V(t, x u x 2)
n
= -x .. 2
a
4 + 2 sin + cos t t
(2 + s in0
'
2
[see (35)]. Thus V< 0 for all t, (x„ x 2 ). Moreover, the set S defined
in (91) is given by
104 S = {(x u x 2 ): x2 = 0}
As in the previous example, if a trajectory of (100)-(101) is entirely
contained in S, it must have x 2 (t) 0, which in turn implies that x, =
is a constant [by (100)], and that
We shall show that, if we let € = min [r, s], then no matter how small
x
The inequalities (107) and (108) together imply that ||x(/)|| is bounded
away from zero, which in turn implies, by the positive definiteness of V,
that V[t, x(01 is also bounded away from zero. Hence, within a finite
amount of time, V[t x(/)] will exceed the value /?(6), which means that
,
110 x t
=x —x +x
t 2 t
x2
111 X2 = —x — x\ 2
Thus V is an l.p.d.f. over the ball B _ 5 for any 8 e (0, 1). Hence Fis a l
positive values arbitrarily close to the origin, and (ii) V(t, x) is of the
form
u V[tMm> 0
119 ^-t {e-
Hence
Hence V[t, x(f)] increases without bound, and therefore ||x(t)|| must even-
tually equal e. This shows that 0 is an unstable equilibrium point. H
121 Example. Consider the system of equations
122 *1 = *i + 2x + XiXi 2
123 x2 = 2x + x — x\xx 2 2
and let
124 V(x u x 2 ) = x\ — x\
Then Visa suitable Liapunov function candidate for applying Theorem
(115). Calculating V, we get
= 2V{x x ,
x2 ) + Ax\x\
Because 4x\x\ >0 V (x u x2 ) e R2 ,
it follows by Theorem (115)
Sec. 5.2 Liapunov's Direct Method 163
to hold in some region for which the origin is a boundary point (and
not an interior point). Theoretically, however, Theorems (115) and
(126) are equivalent, as shown later. Theorem (126) is generally known
as Cetaev’s theorem.
FIG. 5.7
164 Chap. 5 Stability in the Sense of Liapunov
131 x i
= -f- x2 + x\
132 ±2 = *i — x2 +xx t 2
above), and
2. The Liapunov function F, especially in the theorems on stabil-
V(x 1 ,
x2 ) = c(l — cos Xi) + x\\2
show that (0, 0) is a stable equilibrium point if a > b > 0, and that (0, 0) is
m+f[y(t)]Xt)+g[y(t)] =0
(a) Transform this equation into state variable form by choosing x t =
y, *2 =X
166 Chap. 5 Stability in the Sense of Liapunov
Xi = x t + 2x\
x2 = 2x x 2 x + x\
Using the Liapunov function candidate
V(x u x 2 ) = x\ — x\
show that (0, 0) is an unstable equilibrium point.
xt = x\ — x 1 x2
x2 = — x\ — 2x t x 2
using the Liapunov function candidate
V(x u x 2 ) = x 1 (x 1 - x2 )
and Theorem (115), show that (0, 0) is an unstable equilibrium point.
5.3
STABILITY OF LINEAR SYSTEMS
In this section, we shall study the Liapunov stability of systems described
by linear vector differential equations. The results presented here not
only enable us to obtain necessary and sufficient conditions for the
stability of linear systems but also pave the way to deriving Liapunov’s
indirect method (which consists of determining the stability of a non-
linear system by linearization).
2 x(0 = t 0 )x(t 0 )
where <!>(•, •) is the state transition matrix associated with A(«) and
is the unique solution of the equation
4 O(t 0 , to) =I
With the aid of this explicit characterization for the solutions of (1),
it is possible to derive some useful conditions for the stability of the
equilibrium point 0. Because these conditions involve the state transi-
tion matrix <D, they are not of much computational value, because it is
ifand only if 9
6 sup
t>to
1 1
t0) 1 1,
A m(t < oo 0)
proof
(i) “If” Suppose (6) is true, and let e 0 be specified. If we define >
S(e, t 0 ) as e/m(t 0 ), then, whenever ||x(f 0 )|| <5(6, t 0 ), we have <
7 1 1
x(/) 1|
= ||
<D (t, t 0 )x(t 0 ) 1 1
< 1 1
<D (/, f 0 )IHI x(f 0 ) 1
<m(t 0 )-d(€,t 0 ) =€
so that [5.1(7)] is satisfied. This shows that the equilibrium point 0
is stable at time t0 .
“
(ii) Only If” Suppose 0(- / 0 ) is an unbounded
(6) is false, so that 1
1
, 1 1/
8 l|3>(Mo)ll,>|^
where <5i < <5 is some positive number. Next, select v to be a vector
of norm 1 such that
9 ll$(Mo)v|| HI$(Mo)lli
9 Recall that ||
• ||,- denotes the induced norm of a matrix.
*
—
and approaches 0 as t > oo, it follows that 1®(- / 0 ) k is bounded over the 1 , I
so that x(t )
— » 0 as t — oo. This shows that the equilibrium point 0 is
(i) => (iii) Suppose (iii) is false. Then at least one component of
<D(f, t 0 ), say 0 l7 (/, t 0 ), does not approach zero as t — > oo. Let x(/ 0 ) = dej,
where e is the yth elementary unit vector (i.e., e y has all zero components,
except for the yth component, which is 1). Then, regardless of how small we
make 5, it is clear that the /th component of the corresponding solution
x(-) does not approach 0 as t — > oo. Thus, if (iii) is false, then there are
trajectories starting from initial conditions arbitrarily close to 0 that do
not converge to 0. Thus (i) is also false. mi
13 Theorem With regard to the system (1), the following two state-
Hence
sup
t>t i
|| ©ten) || f < oo
implies stability at all times after t0 . The key point in the proof of
Theorem _1
(13) is that || [<!>(£, £ 0 )] IL-
is finite. In a general nonlinear
system, the transition map x 0 i-> s(/
1? x0 ,
* 0 ) is bounded, but (1) it
may not have an inverse, and (2) even if it has an inverse, the inverse
may not be bounded. Thus Theorem (13) is a special result for linear
systems. 10
prodf
(i) “If” Suppose m 0 is finite; then, for any e > 0, [5.1(8)] is satisfied
with 5 =
€lm 0 .
10 Theorem
(13) also holds in the case of nonlinear systems of the form ±(t)
= X (0L where f is continuous in t and Lipschitz-continuous in x; this is because
in this case also the map x 0 s(*i, x0 ,
tQ ) has a bounded inverse; see Theorem
[3.4(50)].
>
least one component of <!>(•, •), say 0 O (-, •), has the property that
sup |
(f>ij(t, f 0 )l is unbounded as a function of t0
t>to
Let x 0 = ey-, the yth elementary vector. Then, because of (22), the
ratio ||®(f, f 0 )xo||/||x 0 ||
cannot be bounded independently of t0.
Thus the equilibrium point 0 is not uniformly stable over [0, oo). —
Theorem The equilibrium point 0 of (1) is uniformly asymptotically
stable over [0, oo) if and only if
sup sup ||®(7, t 0 )\\t< OO
fo>0 t>t o
proof
” then the equilibrium point 0
(i) “ If By Theorem (19), if (24) holds,
is uniformly stable over [0, oo). Similarly, if (25) holds, the ratio
1 1
cD(f, 1 0 )x 0 1 1/| |
xo 1 1
approaches zero as t — oo, uniformly in 1 0 and
(ii) “ Only If
”
This part of the proof is left as an exercise. M
Theorem (27) below gives an alternative equivalent condition for
uniform asymptotic stability.
proof
(i) “If” Suppose (28) holds. Then clearly (24) and (25) also hold, so
that the equilibrium point 0 is uniformly asymptotically stable.
(ii) “ Only If” Suppose (24) and (25) hold. Then there exist m and T
such that
r
29 \\®(t,ti)\\i<m 0 , V t>t u V ti
Now, given any t0 and any t > *o> pick k such that tQ + kT < < to
t
+ kT -f T. Then we have
Hence
34 ||0(r, r 0 )||i 0 0
37 x(/) = A x(t)
In this special case, Liapunov theory is very complete, as demonstrated
by the following series of theorems.
proof From Appendix II, the state transition matrix 0(r, t0) of the
system (37) is given by
41 V(x) = x'Px
where P is a real symmetric matrix. Then V is given by
44 Theorem Let A e
and let [X l9 RnXn
2„} be the (not necessarily
. . . ,
45 Xt + Xf^O, V Uj
(where * denotes complex conjugation).
The proof of this theorem is not difficult but requires some con-
cepts from linear algebra not heretofore covered, and the reader is
therefore referred to [7] Chen for the proof.
Using Theorem (44), we can state the following corollary.
satisfied. _
174 Chap. 5 Stability in the Sense of Liapunov
49 R= r eA '
1
dt
Jo
r e A r'Me Az
'
51 S it) ) dr
Jo
52 S(0 = e A ‘Me At
'
+ [
eA
'
T
Me ATA dr
Jo Jo
= f (A'e A
'
r
Me Ar + eA
'
T
MAe Ar ) dr
£
^(e A,T Me AT ) dr = e
Aft
Me At - M
/
= s(0 M
so that (50) is satisfied.
Now, because all eigenvalues of A have negative real parts, we see from
(52) that S(0 — » 0 as t —* co. Also, as t — > oo, the integral in (51) remains
well defined and in fact approaches R as given in (49). Thus, as t — > oo,
in the steady state, (50) reduces to
54 0 = AR + RA + M
which is equivalent to (48). Hence (49) provides a solution to (48). By
Theorem (44), it is in fact the unique solution. ^
We can now state one of the main results for the Liapunov matrix
equation.
Sec. 5.3 S tability of Linear S ystems 175
positive definite.
(iii) For every positive definite matrix Q e Rnxn , (43) has a
unique solution for P, and this solution is positive definite.
56 P = f°° e A>t Qe At dt
Jo
P = eA
'
59 f
r
M'Me Ax dr
Jo
60 x'Px = f x'e A
'
r
M'M^ AT x c/t = f \\Me Ar x\\l dr >0
Jo Jo
where ||
•
|| 2 denotes the Euclidean norm. Also, (60) shows that if x'Px = 0,
then
61 Me Atx = 0, V t>0
If we take t = 0 in (61), we get Mx = 0, which implies that x = 0 because
M is nonsingular. Thus P is positive definite and (i) => (iii). H
62 Remarks :
A g Rnxn ,
pick Q g R nxn to be any positive definite matrix
(a logical choice is Q=I
or some other diagonal matrix),
and solve (43) for P. Suppose that (43) has no solution or that
it has more than one solution (i.e., it does not have a unique
11
63 Example . Let
and let
= P i Pi
P
.Pi Pi
1
!To avoid getting lost in an avalanche of numbers, we shall illustrate the use
of the Liapunov matrix equation only for second-order systems, even though for
n — 2 this is the hard way to do it.
Sec. 5.3 Stability of Linear Systems 177
Pi — Pi~ Ps = 0
2Pi + 2/7 = —1 3
Pi = Pi = —2
Thus
- 1/2
P=
3/2
Even though (43) has a unique solution in this case, the resulting matrix
P is clearly not positive definite. Hence not all eigenvalues of A have
negative real parts.
64 Example. Let
- "
'-2 -3“ 2 0 Pi Pi
A= , Q= P=
1 2_ _0 2_ . Pi P3_
Then (43) reduces to the three equations
- 4Pi + 2p 2 = -2
— 2>Pi + Pi — 0
— 6p 2 + 4/7 = —2 3
Adding 4 times the second equation and —3 times the first equation,
we get
—6p 2 + 4/7 = 3 6
which is clearly inconsistent with the third equation. Thus (43) has no
solution in this case. Thus we conclude that not all eigenvalues of A
have negative real parts.
65 Example. Let
n -n n on
2
42 -i r
-11 5
67 V(x) = — x'Px
Then
68 V(x) = x'Qx
Moreover, V and V assumes positive values arbi-
is a p.d.f.,
1
The equilibrium point 0 of (37) can be unstable in another way, namely,
all eigenvalues of A have nonpositive real parts, but some eigenvalue of A having a
zero real part is a multiple zero of the minimal polynomial of A.
Sec. 5.3 Stability of Linear Systems 179
J‘
A[A(t)] dx <w 0, V t>t „ V ti >t 0
P jU,[A(t)] .
-co as t —> oo
Jta
0 is uniformly asymptotically stable over [t Q , oo) if, for every M> 0, there
exists a T< oo such that
rh+t
/*t[A(T)] dr < — Af, V t>T, V ti > \
Jti
>
f
Jto
— fii[— A(x)] dx — > oo as t — oo
Remarks:
1. As we saw and 3.5, the measure of a matrix is
in Secs. 3.2
strongly dependent on the vector norm on Rn that is used to
define the matrix measure. In lemmas (71), (74), and (77), the
indicated conclusions follow if the appropriate conditions hold
for some matrix measure //,(•)• Thus there is a great deal of
flexibility in applying lemmas (71), (74), and (77), because one
is free to choose the matrix measure ///•)> and a wise user will
“cos t —sin f
A(0 =
_sin t cos t
-2 + 3e-'~
-3
and choose the matrix measure ju n ( •). It is easy to see that, as t —
-> oo,
The bound (84), together with the boundedness of Q(«), proves the
lemma. _
85 lemma Suppose that, in addition to the hypotheses of lemma (82),
we also have the following
(i) Q(0 is symmetric and positive definite for each t 0; moreover, >
there exists a positive constant a such that
proof Let x e Rn ,
and consider the triple product x'P(/)x. From (83),
we get
s'(t, x, t)Q(t)s(r, x, t) dr
where s(t, x, t) denotes (as before) the solution of (1), as evaluated at time
r, corresponding to the initial condition x at time t. Now, (86) and (89)
together imply that
90 xV(t)x >aJ ||
s(t, x, t) 1 1| dr
where ||
•
|| 2 is the Euclidean norm on Rn Now, by Theorem
. [3.5(1)], we
have
xx
2m o
The inequality (88) now follows by taking /? = a/2m 0 .
94 V(t, x) = x'P(r)x
where P(-) is given by (83), and applying Theorem [5.2(45)].
The utility of Theorem (93) lies in the fact that it provides the basis
for Liapunov’s indirect method (linearization), as applied to non-
autonomous systems.
can be expressed as
where
proof Define
= + T, to + T) &(to + T ) t
,
10 • exp ( M.T)
•exp [-M {t - / 0 )1
106 + T,t 0 + T) = t Q)
107 (/ + T ,
t 0) = ®(f, /0 )-exp [— M(f - t 0 )] = V( t, / 0 )
obtained, the results of Sec. 5.3.1 can be used to obtain necessary and
sufficient conditions for stability. These are given next.
Theorem Consider the system (1), and suppose A(«) satisfies (99).
Then the equilibrium point 0 of (1) is uniformly stable over [0, oo)
ifand only if all eigenvalues of have nonpositive real parts, and M
any eigenvalue of M
having a zero real part is a simple zero of the
minimal polynomial of M. The equilibrium point 0 of (1) is uniformly
asymptotically stable over [0, oo) if and only if all eigenvalues of M
have negative real parts.
cos t sin t
0) =
sin t '
cos t
P + (2 - a)k + (2 - a) = 0
Thus, if 1 <a< 2, the eigenvalues of A(?) have negative real parts for
Sec. 5.3 Stability of Linear Systems 185
all t , and in fact the eigenvalues of A(>) are bounded away from the
imaginary axis; yet, the equilibrium point 0 is unstable.
110 Theorem Consider the system (1); suppose that A(-) is continu-
ously differentiable and that the eigenvalues of A(t) all have negative
real parts for all t > 0. Under these conditions, the equilibrium
point 0 of (1) is uniformly asymptotically stable over [0, oo), provided
the quantity
111 sup
t> o
||
A(f)||/ Av
is sufficiently small.
±c — ~C °
- -1
Rl2 Xc
-X L _ .
_0 L_ R22- Lx l _
(a) Show that the total energy stored in the capacitors and inductors is
given by
V = yXcCx c + yXiLx L
(b) Using this V as a Liapunov function candidate, show that (0, 0) is
a stable equilibrium point if R n and R 22 are both positive semidefinite, and
*i (0 —2 + sin t
2 1
(a)
-*2 ( 0 - cos / — 1_
“
"*i or "It - t
2
1 - t 3 +n ~*i( 0
(b) *2(0 = t
2 - 1
3
0 *2(0
-•*3(0- „ 2 5 4 — t_ -*s(0-
Problem A system of the form [5.1(1)] is said
5.16. to be bounded at
time t 0 if for every S > 0 there exists an e(d, t 0 ) such that
x(/ 0
|| < <5=>||x(0|| < €(t 0
) || V t> t 0 , <5)
It is said to be uniformly bounded over [r 0 , oo) if for every <5 > 0 there exists
(a) Show that for a linear system of the form (1), the equilibrium point
0 is bounded at line t0 if and only if it is stable at time t0. [Hint : Use Theorem
(5).]
(b) Show that for a linear system of the form (1), the equilibrium point
0 is uniformly bounded over [/ 0 , oo) if and only if it is uniformly stable over
[to, oo). [Hint: Use Theorem (19).]
5.4
LIAPUNOV'S INDIRECT METHOD
In this section, we shall combine the results of the earlier sections to
obtain one of the most useful results in stability theory, namely Lia-
punov’s indirect method (also known as Liapunov’s first method). The
great value of this method lies in the fact that, under certain conditions,
it enables one to draw conclusions about a nonlinear system by studying
the behavior of a linear system.
We shall begin by studying the concept of linearizing a nonlinear
system around an equilibrium point. Consider first the autonomous
system
Sec. 5.4 Liapunov's Indirect Method 187
1 i(0 = f WO]
Suppose f(0) = 0, so that 0 is an equilibrium point of the system (1),
and suppose also that f is continuously differentiable. Define
2
a ran
3 fj(x) = f(x) — Ax
then
4 lim JliWIUo
||x||-»0 II
X ||
5 f(x) = Ax + f i(x)
is the Taylor’s series expansion of f(.) around the point x = 0 [recall
that f(0) = 0]. With this notation, we refer to the system
6 z(t) = Az (t)
suppose that
8 f(t, 0) = 0, v/>o
and that f(7, •) is continuously differentiable. Define
f
12 lim sup l!
ifr*)H
||X||
=Q
l|x|H0 f>0
13 m= a(/)z(o
0 o
15 *i(0 = — *i(f) + tx l
16 x 2 (t) = x t) - xt ( 2 (t)
_i
17 A(/)=r °i v<>
_
L i — iJ
However, the remainder term f\{t, x) is given by
18 f 1 (t,x) = [txl 0]'
19 i,0) = — Zj(0
20 i 2 (0 = z ~ z (0i(t) 2
Suppose that
23 f(/, 0) - 0
24
27 A(0 is bounded
Under these conditions, if the equilibrium point 0 of the system
28 z(0 = A(t)z(t)
is uniformly asymptotically stable over [0, oo), then the equilibrium
Sec. 5.4 Liapunov's Indirect Method 189
32 V(t x) ,
= xT(/)x + f'(t, x)P(r)x + x'P(0f(*, x)
= xTP(/) + A'(0P(0 + P(0A(01x + 2x'P(0fi(^, x)
However, from (29), it can easily be shown that
36 1
2x'P(/)fi(*, x) |
< V t >0 whenever 1 1
x 1 1
< r
Therefore
Thus all the hypotheses for Theorem [5.2(45)] are satisfied, and we con-
clude that the equilibrium point 0 of (22) is uniformly asymptotically
stable over [0, oo). ma
38 corollary Consider the autonomous system
39 ±(f) = f [x(01
Suppose that f(0) = 0 and that f is continuously differentiable. Define
and that (26) holds. Under these conditions, the equilibrium point 0
of (42) is unstable if at least one eigenvalue of A 0 has a positive real
part.
proof We shall give the proof only for the case where the matrix A 0
satisfies the condition [5.3(45)]. The proof for the general case can be
obtained from the one given below by using continuity arguments.
Because A 0 is assumed to satisfy [5.3(45)], and has at least one eigen-
value with a positive real part, we have from Theorem [5.3(66)] that the
equation
44 A 0 P + PA 0 = I
has a unique solution for P and that this matrix P has at least one positive
eigenvalue.Now, by arguments entirely analogous to those used in the
45 F(x) = x'Px
then V(t, x) is an l.p.d.f., so that, by Theorem [5.2(106)], the equilibrium
point 0 of (42) is unstable. The details are left as an exercise. H
Remarks: Theorems (21) and (41) are very useful because they
enable one to draw conclusions about the stability status of the equi-
librium point 0 of a given nonlinear system by examining a linear
system. The advantages of these results are self-evident. Some of the
limitations of these results are the following: (1) The conclusions
based on linearization are purely local in nature; to study global
asymptotic stability, it is still necessary to resort to Liapunov’s direct
method. (2) In the case where the linearized system is autonomous, if
some eigenvalues of A have zero real parts and the remainder have
negative real parts, linearization techniques are inconclusive, because
this case falls outside the scope of both Theorems (21) and (41).
(This can be rationalized as follows: In the case described above, one
Sec. 5.4 Liapunov's Indirect Method 191
can say that the linearized system is on the verge of stability or, if —
one is a pessimist, that it is on the verge of instability. Thus it stands
to reason that the stability status of the equilibrium point 0 is actu-
allydetermined by the higher-order terms, which are being neglected
in the linearization.) (3) In the case where the linearized system is
nonautonomous, if the equilibrium point 0 is asymptotically stable
but not uniformly asymptotically stable over [0, oo), the linearization
technique is once again inconclusive. It can be shown by means of
examples that the assumption of uniform asymptotic stability in
Theorem (21) is indispensable. Note that corollary (38) is the basis
of the linearization technique discussed in Sec. 2.3.
47 x 1
=x 2
48 x2 = ju( ~ xj)x — x
1
2 t
If > 0, the eigenvalues of this matrix A both have positive real parts.
Hence, by Theorem (41), the equilibrium point (0, 0) of the system
(47)-(48) is unstable.
51 Xi = —X + X + x 2 X x2
1
52 X2 = —x + x\ t
54 A'P + PA = —I
3 3
56 ||fi(x)||<^||x|| whenever ||
x ||
< r
where p < 1 is some number. Note that we chose p = f in (35) for the
purposes of proving the theorem, but here it is desirable to choose p as
close to 1 as possible in order to get the least conservative value for r.
59 u(0 = g[x(0]
in such a way that the equilibrium point 0 of the closed-loop system
1
This application requires some familiarity with linear systems and the linear
regulator problem in optimal control theory.
Sec. 5.4 Liapunov’s Indirect Method 193
1. f(0, 0) - 0.
2. f is continuously differentiable.
3. If we define
A= <7f(x, u) ~[
_ dx Jx= 0} n = 0
~
= df (x, u) ~|
B
- <?U _Lc=0,u = 0
then we have
rank [B| AB| . . . |
A" -1 B] =n
Assumption (3) above has a very simple interpretation In much :
the same way that (6) is called the linearization of (1) around the point
x = 0, one can think of the system
Thus, assumption (3) states that the linearized system (65) is controlla-
ble. For such systems, the following result is well known (see, e.g.,
[7] Chen).
fact If (64) holds, then there exists an n x n matrix K such that all eigen-
values of A — BK have negative real parts.
Thus fact (66) tells us that if we choose the feedback control law
v(0 = — Kz (t)
-P - A M - MA + MBQ _1 B'M = 0
for the unknown matrix M e Rnxn . It can be shown that, under the
stated assumptions, (69) has a unique solution for M, which is sym-
194 Chap. 5 Stability in the Sense of Liapunov
then all eigenvalues of A — BK have negative real parts. The last state-
ment can be proved using Liapunov theory, and we shall now do so.
71 fact Suppose that A, B satisfy (64), that P e Rnxn and Q e R mXm are
positive definite, that M satisfies (69), and that K is defined by (70). Then
all eigenvalues of A — BK have negative real parts.
proof We make use, without proof, of the fact that the solution M
of (69) is positive definite. We have
72 A — BK = A — BQ _1 BM
Therefore
feedback gain has been found for the linearized system (65), we have
already found a way to stabilize the original nonlinear system (58).
df~
76 A= (?X_ x=0>u =o
df~
77 B =
.^UJx=0,u = 0
78 u(0 = — Kx(r)
in (75), then 0 is an asymptotically stable equilibrium point of the
resulting system
81 x(r) = g[x(/)]
x2 = sin x — x + u t 2
satf a =
\a, M< 1
\
(sign
.
tr, 1
cr |
> 1
The A matrix above has one eigenvalue with a positive real part. How-
ever, the linearized system can be made stable by applying the control
law
v = [4
196 Chap. 5 Stability in the Sense of Liapunov
87 f(0, 0)=0
88 g(0,0) = 0
Define
df
89
~dx ( 0, 0)
90 _ <5f
<?y ( 0, 0)
91 _
<?x ( 0, 0)
92 — o*g
dy ( 0, 0)
93 v = h(x)
is a solution of
94 g(x, y) =0
Moreover, we have that
(i) If all eigenvalues ofand of A n — A 12 A 22 A 21 have nega-
A 22 1
96 = x (x + y — 1) + x\ + y
1 1
97 x = x (x\ — 5) + x (x — 3) + y(x + 1)
2 t 2 1 2
98 ey = x (x - 2) - x (x - 1) - y(y + 1)
1 1 2 1
A2 — i [ 2 1]
^22 — [“l]
An ^ 12 -^ 22^21
0 < < € €0 .
5 5
THE LUR'E PROBLEM
In this section, we of an important class of
shall study the stability
control systems, namely, feedback systems whose forward path contains
a linear time-invariant subsystem and whose feedback path contains a
memoryless (possibly time-varying) nonlinearity. Such a system is
shown in Fig. 5.8 and is described by the equations
1 x(0 = Ax(t) + b £(t), t> 0
3 m =
g(s)= c(sl - A)- b + d
l
1
We also assume, as in the rest of this chapter, that $(• , •) is of such a nature
that the system (1)— (3) has a unique solution corresponding to each initial condition
x(0) and that this solution depends continuously on x(0). This is the case, for exam-
ple, if 0(* , •) satisfies a global Lipschitz condition.
;
(A2) A has a simple eigenvalue of zero, and the rest of its eigen-
values have negative real parts.
(A3) The pair (A, b) is controllable i.e.,
rank [c |
A'c |
. . .
|
(A')"
_1
c] =n
(A5) The nonlinearity 0(., .) satisfies
0(f, 0) = 0, V t >0
<J0(/, a) > 0, Va e R, V t >0
[The condition (7) is sometimes stated as “0(f, .) is a first and third-
quadrant nonlinearity,” because (7) ensures that the graph of cn->
(f)(t, o) lies in the first and third quadrants, for all t > 0.]
Subject to the above assumptions, the problem is to find condi-
tions on A, b, c, d which ensure that the equilibrium point x 0 of the =
system (l)-(3) is globally asymptotically stable whenever any non-
0 is
linearity satisfying (A5).
In contrast with the systems studied in Sec. 5.2, we are concerned
at present not with a particular system but with an entire class of sys-
tems, because we have not made any assumptions on
0 other than that
it satisfies (A5). For this reason, the problem under study is sometimes
not result in any loss of generality, because given any proper rational
function h(s), one can always find A, b, c, d such that
h(s) = c\sl - A) -1 b + d
and such that (A3), (A4) hold. See [7] Chen for full details.
In some cases, (7) is replaced by the following more stringent
condition: There exist constants k 2 '>k such that l
>0
k a2x
<a <f)(t, o) <ko 2
2
, V o e R, V * >0
Because of its importance, (9) is given a name.
200 Chap. 5 Stability in the Sense of Liapunov
Aizerman's Conjecture
invariant system
11 ±(t) = [A — bc'/:]x(/)
Kalman's Conjecture
15 As pointed out in Sec. 5.1, one can unambiguously say that the system is
globally asymptotically stable, rather than that the equilibrium point 0 is globally
asymptotically stable.
6
14 A'P + PA = — qq' — eQ
15 Pb — v = y q 1/2
16 h(s) A y + 2v'(sl - A) _1
b
16
satisfies
17 Re h(jco) >0, V co e R
proof See [16] Lefschetz [pp. 115-18].
1
Although we have not introduced the concept here, we mention that, with
the assumptions on A, (17) is equivalent to the requirement that h(s) is a strictly
positive real function.
202 Chap. 5 Stability in the Sense of Liapunov
18 F(x) = x'Px
Note that the nonlinear characteristic does not appear explicitly in <f>
2. 1 + kd 0.
pose 1 + kd < 0. Then one can always find a number k 0 e [0, k] such
that 1 + k d = 0. Suppose we now choose (p) = k o. Then
0 <f> 0 0 <j> 0
20 (1 + k d)a
= c'x Q
21 gfs) = c(.sT — A) -1 b +d
Then g is the open-loop transfer function. If we now apply a feedback
gain of k 0 ,
the closed-loop transfer function becomes
22 A tot
gA ’
=
1 +
M
k 0 g(s)
we get
„ . . i?(°°) = d
23 gA°°) = i fc 0 g(wj r+M
However, note that 1 + k 0 d = 0 (and that d 0). Hence g(oo) = oo, ^
which means that g c is not a proper rational function and so represents
an unstable system (see Sec. 6.3).
The conclusion, therefore, is that if cj) belongs to the sector [0, k],
Sec. 5.5 The Lur'e Problem 203
28 0 < G
< k, V o 0
30 ((la7 +0<0
Hence the last term in V is nonpositive. We now make use of Theorem
(13) in order to complete the square, so as to make the rest of the terms
in V negative. Specifically, apply Theorem (13), with Q any positive
definite matrix, A, b as at present, y 1 kd, and v Ax/2. Then = + =
Theorem (13) assures us that if (17) holds (the implications of which are
discussed later on), then there exist a positive definite matrix P and a
vector q such that
31 A'P + PA = — qq' — eQ
32 Pb — ^A:c = (1 + Ax/) q 1/2
Thus V becomes
33 F(x) = — ex'Qx — x'qq'x + 2£x'(l + kd) q — (1 + kd)( l/2 2
+ ((ka + 0
= —ex'Qx — [x'q — (1 + kd) (] + ((ka + 0 1/2 2
< -ex'Qx
204 Chap. 5 Stability in the Sense of Liapunov
34 [x'q - (1 + kdy /2
£]
2
>0
Thus, clearly, V is a p.d.f., and —V is a p.d.f., whence by Theorem
[5.2(58)] it follows that the system (l)-(3) is globally asymptotically
stable.
Now, let us come back to condition (17), which, as has just been
demonstrated, is a sufficient condition for global asymptotic stability.
where
36 g(s) A c(sl — A) -1 b + d
is the transfer function of the open-loop system (l)-(2). Therefore, a
sufficient condition for the global asymptotic stability of the system
(l)-(3) is
37 Re [1 + kg(jco)] > 0, V co e R
This is stated formally as a fact.
38 fact If (f)
belongs to the sector [0, k] and 1 + kd > 0, then (37) is a suf-
ficient condition for the absolute stability of the system (l)-(3).
If </) belongs to the sector [0, oo), then the corresponding absolute
stability criterion can be obtained from (37) by dividing through by k
and then letting k — > oo.
39 fact If </) belongs to the sector [0, oo) and d > 0, then a sufficient condi-
tion for absolute stability is
40 Re g(jco) >0, V co e R
41 1 + kd = Re [1 + kg(oo)] > 0
shown in Fig. 5.9. In this case, the transformed forward path transfer
function is
42 &0) = +gO)
ag(s) 1
Sec. 5.5 The Lure Problem 205
S t (s)
I I l
I J
43 = <f>(a) — aer
Now, belongs to the sector [0, /}
— a]. However, before we can
apply fact (38), we must ensure that the transfer function g (s) can be t
44 g (s)
t
= c£s\ - A,)->b, + d,
where A t
is a Hurwitz matrix. Now, if the pair (A„ b,) is controllable
and the pair (A c,) is observable, then the eigenvalues of A, are the
f,
same as the poles of g (s). In turn, if a 0, then the poles of g (s) are
t ^ t
the same as the zeroes of 1 + ag{s). Therefore, to apply fact (38) to the
transformed system (42)-(43), we must first ensure that 1 + ag{s) has
no zeroes in the closed right half-plane.
The last condition can be readily verified using the well-known
Nyquist criterion. Suppose (l)-(2) is a realization of the untransformed
transfer function g(s), and suppose (1) A has no eigenvalues on the
imaginary axis, and (2) A has v eigenvalues with positive real part. Then
gCs ) has no poles on the imaginary axis and has exactly v poles in the
1
open right half-plane. In this context, the Nyquist criterion states that
1 + agC?) has no zeroes in the closed right half-plane if and only if the
Nyquist plot of g(jco) does not intersect the point — 1/a + j0 and encir-
cles it in the counterclockwise sense exactly v times.
We are now ready to extend fact (38) to the general case of a non-
linearity in the sector [a, fi].
Based on the foregoing discussion, a suf-
206 Chap. 5 Stability in the Sense of Liapunov
ficient condition for the absolute stability of the system (1)— (3) can be
given as follows.
45 fact Suppose A in (1) has no eigenvalues with zero real part and has v
eigenvalues with positive real part ; suppose 0 in (3) belongs to the sector
[a, /}], a ^ 0. Then a sufficient condition for absolute stability is
46 The Nyquist plot of g(jco) does not intersect the point — 1/a + j0
and encircles it v times in the counterclockwise sense.
47 Re[l+(/? — «)
r
^g>Lj >0’ VCoeR
48 Re \ \
Ll +
Mm
± ag(ja>)J >0 , VcoeR
Suppose
49 g(jco) = u+jv
Then (48) becomes
50
(l + pu)(l + au) + a^ > R
(1 + aw) + v
2 2
If (46) holds, the denominator of (50) is never zero, so that (50) can be
simplified to
FIG. 5.10
Sec. 5.5 The Lur'e Problem 207
53 fact Suppose A in (1) has no eigenvalues with zero real part and has v
eigenvalues with positive real part; suppose 0 in (3) belongs to the sector
[a, fi], with a > 0. Then a sufficient condition for absolute stability is
54 The Nyquist plot of g(jco) does not enter the disk Z>(a, /?) and
encircles it in the counterclockwise sense exactly v times.
55 The Nyquist plot of g(jco) neither encircles nor intersects the point
~ l/ a +/0.
56 Re
H+
MM
H + ag(jco)_\ >0, ] V oe R
ft
57 (u + I)(« + +v < 2
0, V a) e R
j)
Thus the geometrical interpretation is that the Nyquist plot of g(jco)
lies in the interior of the disk T>(a, p). Moreover, if this is the case, (55)
is automatically satisfied. Thus we have
58 fact Suppose A is Hurwitz, and suppose 0 belongs to the sector [a, /?],
with a < 0 < p. Then a sufficient condition for absolute stability is
59 The Nyquist plot of g(jco) lies in the interior of the disk D( a, P).
(ii) 7/0 =a<p the Nyquist plot of i(jco) lies in the half-plane
61
|j:
Re s >
v)
fi
1
64 d= 0
^+a 2 ]’
65 0(cr, t) = oq cos co 0t
66 z(s) - S2 + 2ns + u
2
+ 2
67 (f)
belongs to the sector [—q,q]
radius 1 jq. Plotting g(jco), we note (see Fig. 5.11) that g(jco) attains its
maximum modulus at co =
) an d that this maximum mod-
=fc(fl
2 — M
2
ulus is 1/(2 jua). Thus the condition that the Nyquist plot of g(jco) lies in
the interior of the disk D( a, fi) reduces to
68 J_<±
Ilia q
Sec. 5.5 The Lure Problem 209
or, equivalently,
69 q < 2 pa
Thus (69) is a sufficient condition for the global asymptotic stability of
(63). Furthermore, because the circle criterion is a sufficient condition
for absolute stability, (69) actually implies the global asymptotic sta-
bility of all systems of the form
70 y + 2 py + ( ju
2
+a 2
)y - qf(f 9 y)
=0
whenever /is any nonlinearity belonging to the sector [—1, 1].
73 «(ff)A[>0«
Jo
P isa positive definite matrix, and /? 0. The function (72) is called a >
Liapunov function of the Lur’e-Postnikov type. Because <j) belongs to
the sector [0, oo), it is clear that 0>(<7) is always nonnegative, so that V
is a p.d.f.
Defining
76 v = Pb — -^-A'c
78 A'P + PA = -Q
where Q is a positive definite matrix of our choice. Then
F(x) = — x'Qx + 2<j;x'v — J?c'b
2
79 <^
We claim that if
80 fic'b
— v'Q
-1
v > 0
81 v = 0 == x = 0 , £ = 0
82 x —Q - 1
v<j; =0
83 1 = 0
V
Clearly, (82) and (83) together establish (81). Next, because V is a nega-
tive definite quadratic form in x and there exist positive constants
and a 2 such that
84 V(x) < — ocjx'x — oc 2 £
2
Hence
85 V(x) < — OGiX'x
which shows that — is a p.d.f. and establishes the global asymptotic
stability of (l)-(3).
The preceding discussion is summarized in the following.
90 fa) = -0[<7(O]
Clearly, the difference is between (90) and (3). By recasting (88)-(90)
in the form
i
x(0“ o —
x(0“ ~b~
91
+
-0 o_ m. _ 1 _
212 Chap. 5 Stability in the Sense of Liapunov
X (0
92 o(t) = \ c'
d]
7
'
L£(0J
93 u(t) = -0HO1
We see that the present system is of the form (l)-(3), where the A
matrix satisfies assumption (A2) instead of (Al).
Throughout this subsection, we shall make the following Assump-
tion: 0 belongs to the sector (0, oo), and d 0. ^
The following result shows that d > 0 is actually a necessary
condition for absolute stability.
94 fact Assuming that d ^0 and that 0 belongs to the sector (0, oo), a
necessary condition for the absolute stability of the system is d > 0.
95 0(0*) = eo
for all € > 0. With this choice of 0, the system becomes
we neglect the terms involving 6 on the right-hand side of (97), the system
is qualitatively similar to
±(01 _ I” A 0 1 r x(0”
98
ju£(t)_ c' — d_ _£(t)_
99
OITA 0 1 rx(0
p£,(i)_ _-c' -d] L«0
which has the solution
100 x(0 =o
101 Z(t) = <?(0) exp
In deriving (100), we use the fact that A is Hurwitz and is therefore non-
Finally, by continuity, the system (97) is unstable for all sufficiently small
values of €. This shows that d > 0 is a necessary condition for absolute
stability.
105 V= + 2a (a — c'x)d£ +
2x'Px
= x'(A'P + PA)x - 2x'Pb0 - 2ad(a - c'x)0
116 A P + PA = — 6Q - qq'
117 Pb - v = y^ 2 q
We shall return to condition (115) later on. Now, if (115) holds, then V
becomes
118 V = — ex'Qx — (x'q) 2 — 2x'q}> 1/2 0 — y0 2 — ladofy
= — 6xQx — (x'q + y 1/2 2 — 20Ldo(j)
(j))
Finally, we see that (122) is the same as (113). This proves the theorem. M
Sec. 5.5 The Lur'e Problem 215
As it stands, Theorem
(109) shows that (1 10) is a sufficient condi-
tion for the existence of aLiapunov function of the form (103) for the
system (88)-(90). However, Popov shows that (110) is also a necessary
condition for the existence of a Liapunov function of the form (103).
This should not be misinterpreted to mean that (1 10) is a necessary and
sufficient condition for the absolute stability of the system (88)-(90).
Rather, (110) is a necessary and sufficient condition for the existence of
a Liapunov function of the type (103).
Geometric Interpretation
Like the circle criterion, the Popov criterion can also be given a
geometric interpretation. Suppose we plot Re g(jco)\s. colmg(jco), with
co as a parameter. Note that we only need to do the plotting for co 0, >
because both Re g(jco) and co Im g(jco) are even functions of co and no ,
Re g(jco) vs. Im g(jco). Now, the condition (110) simply states that there
exists a nonnegative number r such that the Popov plot of g lies to the
right of a straight line of slope 1 jr passing through the origin. If r 0, =
this straight line is taken as the vertical axis of the Popov plot. The
situation is depicted in Fig. 5.12.
g ,(.v)
L
FIG. 5.13 01
Remarks: In the case where 0 belongs to the sector (0, k), we use
the sufficient condition (125), which is weaker than (110). Clearly,
Sec. 5.5 The Lur'e Problem 217
130 - 5(JTT?
In this case
la
R'^-frrl 5 5
)
2
132 co Im g(ja>) = — 1
(1
The Popov plot of g is shown in Fig. 5.14. It is clear from this figure
that (125) can never be satisfied unless — l/k < — i.e., k < 2. On the
other hand, if k < 2, (125) can always be satisfied by a suitable choice
of r. Hence this system is absolutely stable for all nonlinearities in the
sector (0, 2).
As is the case with the circle criterion, one of the main advantages
of Popov’s criterion is that it is applied directly to the transfer function
g(^) and is thus independent of the state variable realization (88)-(90).
In fact, there is no need to construct a state variable realization for a
given g(s).
In Chap. 6, we shall encounter another Popov criterion, which is
i(s) =
(s + l)(s + 2)(s + 3)
Using the circle criterion, determine the largest range [OC, /?] such that the
system is absolutely stable for all (possibly time-varying) nonlinearities <f>
in
^= (s - l)(s + l) 2
Using the circle criterion, determine the largest range [OC, /?] such that the sys-
tem is absolutely stable for all nonlinearities in the sector [OC, /?].
Compare.
Problem 5.22. Consider a system of the form (88)— (90), with
—
s(s -f 1)
Using the Popov criterion, show that this system is absolutely stable for all
finite number. Verify also that (110) does not hold in this
particular case.
'
5.6
SLOWLY VARYING SYSTEMS
In this section, we shall prove a very general result on the stability of
so-called slowly varying systems. The approach given here is
due to
[3] Barman.
Consider a general nonautonomous system described by
where f : R x Rn — *•
R". If we now consider the autonomous system
as a particular case of the system (1), with its time dependence “frozen”
at time p. This concept can be clarified by considering linear systems.
Suppose we are given the nonautonomous system
3 x(t) = A(t)x(t)
Then, for any p > 0, the autonomous system
4 x(t) = A(p)x(t)
can be thought of as describing the system (3), frozen at time p.
In this section, we shall address ourselves to the following ques-
5 D i(t, x) =
t
8 ||f(7, x) — f(t, y) ||
< &||x — y ||, V t > 0, V x, y e R*
Suppose the family of frozen systems (2) satisfies
V > V > t t0 , *0 o, v x g Rn v P
,
>o
220 Chap. 5 Stability in the Sense of Liapunov
10 \\D t
f(t, x) ||
< <5||x||, \/ t >0, \/ x g Rn
In fact, 8 can be so chosen that the solution trajectories of (1) satisfy
13 s p {t, x, 0) =x+
JQ
f [p, s^Ct, x, 0)] dx, V t >0
Now, (8) and (9) together imply that
15 1 1
s p {t, x, 0) 1 1
> whenever t e [o,
the induced /2 - norm of the Jacobian matrix D^i{t, x) always less than or
is
17 1 1
DpSpit, X, 0) 1 1
<J ‘
[*1 1
D p s p {t, X, 0)|| + <5|| s p (r, x, 0) ||] dr
0
18 II
-DpSjO, x, 0)|| < ||x||^ exp kt
family of frozen systems (2) as well as for the original system (1). Choose
y > 0 such that
19
and choose
21
^
Vp (-)
V/>>0,
along the solutions of
VxeR-
(2), we observe
that
22 V(x) = lim
n
o+
Now
23 Vp [sp (h, x, 0)] = 1
s p (h, x, 0), 0] |
p' dx
Jo 1
J”
small. First, because (21) holds, we see that V is both decrescent and a
p.d.f. Next, calculating the derivative of V along the trajectories of (1), we
get
= -||x|| *? + D t V(t,x)
c 1
Hence
31 |
D, V(t, x) |
< J“ y\ s t (r, |
x, 0) 1
2 » 1 1
s,(t, x, 0) 1 1
•
1 1
As f (r, x, 0) 1 1
dr
32 |
D,V(t, x)|
<J
r°°
y||x|| >7n
2 2 >'- 1
exp [-(2 y - l)at] • —
8m
exp kr dr
Hence the integral on the right-hand side of (32) is finite; let us denote it
it is clear from (29) that — V is a p.d.f. Hence V satisfies all the hypotheses
of Theorem [5.2(58)]. Therefore we conclude that V is a Liapunov func-
tion for the system (1) and that 0 is a globally asymptotically stable equi-
librium point of (1).
35 Cl \x\\
|
2 r<V(t,x)<c 2 \\x\\ 2 y
36 V(t, x) <• — 3 || x ||
2y
37 -jj
V\t ,
s (t, x, /o)] < -c 3 ||s(r, x, r 0 )||
2y
From (37), we get, after using the integrating factor exp [— c 3 (t — t Q )/c 2 ],
~ C3 ~"~
38 V[t, s(t, x, f 0 )] < V(t, x) exp
[
'
c2 ]
and was pioneered during the 1960’s by Sandberg and Zames (see [21]-
[23] Sandberg and [27, 28] Zames). There exists a large
amount of litera-
ture on this subject, and a thorough treatment can be found in [10]
Desoer and Vidyasagar and [26] Willems.
Before proceeding to the study of input-output stability it is ,
224
1
At one may well ask, why not simply use one of the
this stage,
approaches —why
both? The answer is that while the two
use
approaches are related they do not give identical results. Because, in
,
6-1
INTRODUCTION TO ^-SPACES AND
THEIR EXTENSIONS
In this section, we shall give a brief introduction to L p -spaces and their
[ i= 1 i=
where the Efs are all open intervals. Note that ijl*(A) is always nonnegative
and may be infinite for some subsets of R. The function u* is called an j
The label LJfi, oo) denotes the set of all measurable functions
/(•)• [0> °°) — R that are essentially bounded on 1
[0, oo).
Thus, for 1 < p < oo, Lp [ 0, oo) denotes the set of measurable
functions whose pth powers are absolutely integrable over [0, oo),
whereas Z^O, oo) denotes the set of essentially bounded measurable
functions.
4 m= L*
1/2
(i +
i
log 0-
12 /p
where p e [1, oo), belongs to the set Lp [ 0, oo) but does not belong to
Lq [ 0,
oo) for any p.
5 fact Whenever p e [1, oo], the set Lp [ 0, oo) is a real linear vector space,
in the sense of definition [3.1(1)].
Thus, whenever /(•),#(•) belong to Z^fO, oo), we have that (/ + #)(•)
also belongs to L p [0, oo), as does (&/)(•), for all real numbers a. This fact
follows from what is known as Minkowski’s inequality (see [20] Royden).
7 11 /( 011 ,= T
-Jo
i/(o i'*
l/p
The function ||
• ||„: T„[0, oo) — > [0, oo) is defined by 2
9 fact (Minkowski's inequality) Let p E [1, oo], and let /(•), g(')
e Lp [ 0, oo). Then
10 \\(f+ g)(-)\\ P <\\f(-)\\ P + \\g(-)\U
Thus fact (9) shows that p satisfies the triangle inequality. ||
•
\\
Because axioms (Nl) and (N2) of definition [3.1(25)] are easily verified,
we see that the pair (Lp [ 0, oo), ||
. \\ p ) is a normed linear space, for each
P e [1, oo].
11 fact For each p e [1, oo], the normed linear space (L p [0, oo), ||
•
|p is
12 </(•),*(•)>» = \j{t)g{t)dt
Thus (L 2 [0, oo), <•, «> 2 ) is a complete inner product space, i.e., a Hilbert
space.
Fact (11) brings out the main reason for dealing with L^-spaces
in studying input-output stability. We could instead study the set
Cp 0, [
oo), which consists of all continuous functions /(•) satisfying
(2). Then Cp 0,
is a subset of L 0, oo), and the pair (C [ 0, oo),
p [
oo)
p [
||
•
IP normed linear space. However, it is not complete.
is also a
Hence, if we require a Banach space as the setting for the stability
theory, it is better to work with the spaces Lp [ 0, oo). Note that
(IJ0, oo), |p is the completion of the normed linear space
|| .
15 h(t) A f(t)g(t)
belongs to LfiO, oo). Moreover
16
Jo
"
i m i
•
i
sit) dt i
< [ J“ i m \> dtj*
[ J“
i
g(t) i« dt
1/9
]
The inequality (16) can be concisely expressed as
17 !l/(-k(-)lli^ll/(-)IUk(-)ll 9
/
18 [definition] Let x(*): [0, oo) — > R be measurable. Then for all
19
0 <t<T
T< t
20 [definition] The set of all measurable functions /(•): [0, °°) > —
R ,
such that /r («) e Lp [ 0 ,
oo) for all T, is denoted by L pe [0 oo) ,
22 m= t
Then for each T< oo, the truncated function r (-) is the triangular
pulse function defined by
23
T< t
It is clear that, for each finite value of T, the function /r (-) belongs to
all thespaces ZJ0, oo). Hence the original function /(•) belongs to
each of the extended spaces L pe [0, oo), for each p e [1, oo]. However,
/(•) does not belong to any of the unextended spaces Lp [0, oo).
The relationships between the unextended spaces Lp [0, oo) and
the extended spaces Lpe [ 0,
oo) are further brought out in the following.
24 lemma For each p e [1, oo], the set L pe [0, oo) is a real linear vector
space. If p e [1, oo] and /(•) e L pe [0, co), then
(0 ||/H*)lh> isa nondecreasing function of T.
(ii) /(•) E L p [0, oo) (the unextended space) if and only if there exists a
finite constant m such that
In this case
only a linear vector space and cannot be made into a normed linear
space.
27 [ definition The symbol L"[0, oo) denotes the set of all n-tuples
f = [/i . . . where e Lp [0, oo) for i — 1,
f, n. L pe [0,
n
. . .
oo)
28 l|f(-)ll,A[gll/XOII5]
6.1.3 Causality
31 lemma Let A:L npe > L pe Then A is causal — . in the sense of definition
n
(29) if and only if, whenever /, g e L pe and
we have
33 (Af)r = (Ag)r
proof
u n
(i) If Suppose A satisfies (32) and (33); we must show that A sat-
isfies (30). Accordingly, let / L pe and T < oo be arbitrary. Then
e
230 Chap. 6 Input-Output Stability
36 (Ag)r — (Agr)T
Moreover, because fT = gT , (35) and (36) together imply that
37 ( Af) T = (Ag) T
which is the same as (33). H
Definition (29) and lemma (31) provide two alternative (but
entirely equivalent) interpretations of causality. Definition (29) states
that amapping A is causal if any truncation of Af to an interval [0, T]
depends only on the corresponding truncation fT of /. To put it
another way, the values of Af over the interval [0, T] depend only on
the values of / over [0, T Lemma
]. (31) states that A is causal if and
only if, whenever / and g are equal over an interval [0, T], we have that
Af and Ag are also equal over [0, T].
(d) fit) =
fO, if t =0
(e) /(/) = l/t 1/2
,
if 0 <f< 1
|
[l It, if 1 <t
Problem 6.2. Prove lemma (24).
dr
(Af )(t) = f a(t, X)n[f(x)}
Sec. 6.2 Input-Output and State Representation of Systems 231
where n: R —>R is not identically zero. Show that A is causal if and only if
6.2
INPUT-OUTPUT AND STATE REPRESENTATION
OF SYSTEMS
In this section, we shall give a brief introduction to some results on
constructing a state representation of a given input-output relation-
ship. The purpose in presenting these results is to once again emphasize
that the input-output representation and the state representation are
two different ways of describing the same system.
Given is a system described in state variable form by the set of
equations
Suppose (1) has a unique solution for x(.), for each u(-) belonging
to a prespecified class of inputs. [This is the case, for example, if f is
3 y(t) = f
H(f — t)u(t) dr
Jo
4 y(s) = &(s)Q(s)
fact Given a proper rational matrix H(Y), one can always find A, B, C, D
such that
HO) - COI - A) _1 B + D, Vs
and such that
rank [B AB . . . A” -1 B] =n
(A')”“ C ]=«
/
1
rank [C' A'C' ...
where n is the order of the matrix A. In this case, the set of eigenvalues of
A is the same as the set of poles of HO); i-e., a complex number k is an
eigenvalue of A if and only if H(*) has a pole at k.
the form (5)-(6) which has the input-output relationship (3). Further,
we can ensure that the system (5)-(6) is controllable and observable
and that the set of eigenvalues of A coincides with the set of poles of
HO).
y(0 = f H(7,
Jo
t)u(t) dx
3
Note that if D ^ 0, then the impulse response of the system (5)— (6) contains
an impulse distribution at t = 0.
Sec. 6.3 Definitions of Input-Output Stability 233
14 y(t) = C(t)x(t)
such that the system of equations (13)— (14) has the input-output rela-
tionship (12).The conditions under which this can be done are given
next.
15 fact Given H(-, •), one can find A(«), B(*), C(«) such that the system of
equations (13)— (14) has the input-output relationship (12) if and only if
16 H(/, T) = *(*)(T)
proof See [7] Chen.
6.3
DEFINITIONS OF INPUT-OUTPUT STABILITY
In this section, we shall introduce the basic definitions of input-output
stability.
input f g Lnpe ,
the output of the system is Af g L pe Now, we say
.
that the system is Lp -stable if, whenever the input f belongs to the
unextended space L p the resulting output Af belongs to L p and more-
,
over the norm of the output Af is no larger than k times the norm of the
input f plus the offset constant b.
/e
the input (is bounded), the output Af e L (is bounded) and
4 (Af)(t) = (
,-«<»-*>/(T
) dx
Jo
(see definition [6.1(20)]. Hence, for each finite L, there exists a finite
constant mT such that
6
5 \f{t)\<m T a.e., V t e [0, T]
6 git = (
dr
7 ['\e-^\.\f{r)\dr
|g(0l< Jo
< f mT |
e~ a{t
~ r)
1
dr = mT ^ ~~ e
a
—
Jo
— a
a.e.
The inequality (7) shows that g( •) is bounded over [0, T]. Because this
reasoning is valid for every finite L, it follows that g(-) g L^, i.e.,
that A L^e
.
>
Looe •
9 \g(t)\<^’
a
Vt> 0
6 “a.e.” stands for “almost everywhere,” which means “except on a set of mea-
sure zero.”
7 Hereafter we shall drop the phrase “almost everywhere,” it being implicitly
understood.
\ ^
Loo-stable.
12 \f(t)\<m, V/g[0J]
then
13 \(Af)(t)\<m\ V * e [0, T]
This mapping A also maps into itself. To see this, let / e L^.
Then, for each finite L, there exists a finite constant mT such that
16 \f(t)\<m T , V f e [0, T]
17 | ( Af)(t ) |
< m T Jo f |
e
{t ~ r)
|
dr = m T(er — 1)
Now, for each finite T the right-hand side of (17) is a finite number.
Hence Af g L^.
On the other hand, A is not Loo-stable, as can be seen by setting
f(t) = 1, V t. Then /(•) e L^, but
18 (Af)(t) = f e {t
~ T)
dr = e* — l
Jo
^
u 1? u 2 e 1? e 2 y l5 y 2 are
, ,
all members of some extended space L” e and the ,
operators G t
and G map L pe
2 into itself. The equations describing this
feedback system are
19 e, =u -y x 2
20 e = u + yi
2 2
21 y, = G^
22 y2 = G2 e 2
The idea is that G t
and G2 represent two given subsystems which are
interconnected in the manner shown, u t and u 2 represent the (known)
inputs, e 1and e 2 are the (unknown) “errors,” and y and y 2 are the 1
(unknown) outputs.
In analyzing the system described by (1 9)— (22), two distinct
types of questions arise. It is clear that we can eliminate y l5 y 2 from
(19)-(22) and rewrite the system equations as
23 ej = Uj — G e 2 2
24 e 2 = u + Gjej 2
Sec. 6.3 Definitions of Input-Output Stability 237
25 y, = G^U! - y 2)
26 y2 = G (u + y,) 2 2
The system descriptions (19)— (22), (23)-(24), and (25)-(26) are all
equivalent. With regard to this system, the first type of question that
one can ask is the following: Given inputs u and u 2 in L pe
n
does there 1 ,
equations are satisfied, and if so, is this set of solutions unique? This is
usually called the existence and uniqueness question. The second type
of question takes the following form: Given inputs u 1? u 2 in L”, assum-
ing that solutions in L npe exist for e 1? e 2 y 1? y 2 do these solutions actually
, ,
belong to L np ? This is called the stability question. The reason for making
the distinction between the two types of questions is that, in practice,
different types of conditions arise in answering the two types of ques-
tions. For instance, existence and uniqueness of solutions can be assured
under very mild conditions [usually some type of global Lipschitz
conditions; see Theorem (27).] On the other hand, stability requires
stronger conditions, which are the main subject of this chapter.
Accordingly, we shall state below a theorem regarding the exist-
ence and uniqueness of solutions to ( 1 9)— (22) and then proceed to a
study of stability conditions. Theorem (27) is not the most general of its
kind, but it is easy to prove and is adequate for most situations.
R+ x R — R
n
>
n
30 n t
(t, 0) = 0 V >0 t
31 n 2 (^, 0) = 0 V >0 t
(ii) Satisfy global Lipschitz conditions; i.e., there exist finite con-
stants k and k 2 such
x
that
32 II n^, x) -n 1
(r,y)||<fc 1 ||x- y||,
V/>0, V x, y e Rn
33 \\n 2 (t,x) - n2 (f,y)||<fc 2 ||x- y||,
V t > 0, V x, y g Rn
238 Chap. 6 Input-Output Stability
whenever u l9 u 2 y l5 y 2 ,
satisfy (25)-(26).
Remarks:
1. Notice that, according to definition
(34), the question of the
Lp -stability
of the system (25)-(26) is quite divorced from
the question of existence and uniqueness of solutions to
(25)-(26). If u 1? u 2 belong to Z”, then in order for the system
(25)-(26) to be Z^-stable, one of two alternatives must occur:
(i) either there do not exist any y l5 y 2 both belonging to
,
Sec. 6.3 Definitions of Input-Output Stability 239
Lpe ,
such that (25)-(26) are satisfied, or (ii) there do exist
y l9 y 2 in L pe
n
such that (25)-(26) are satisfied, and all such
solutions in fact belong to the unextended space L np and
satisfy (35)— (36). This, and only this, is what is meant by
Instability. There no a priori assumption that, corre-
is
39 yi =e -u 2 2
40 y2 = Uj - ej
Hence, if u 1? u 2 e l9 e 2 belong to
,
L pn and if e 1? e 2 satisfy
conditions such as (37)— (38), then y l9 y 2 also belong to L np
and satisfy conditions similar to (35)— (36).
3. Comparing definitions (1) and (34), one may feel that there
are two distinct concepts of Instability, one for open-loop
stability and one for feedback stability. However, this is
not the case. It can be shown that if we take the system
equations (25)-(26) as defining a relation between the input
pair (u 1? u 2 ) and the output pair (y l5 y 2 ), then definition (34)
requires the same properties of this relation as definition
~ r)
42 ( G x
x){t = f
Jo
e~ a{t x{x) dr
43 (G 2 x)(t) = kx(t)
~ T)
44 y i(0 = f e~ (a+k)(t u 1 (z) dx
Jo
45 y 2 (t) = kyft)
First, we see from Examples (3) and (14) that y u y 2 e L„ e whenever
u1 e L^. Next, based on the same examples, it follows that^ 1? y 2 e
whenever u e L provided a + k > 0. Thus the system is Z^-stable,
x
64
RELATIONSHIPS BETWEEN I/O STABILITY AND
LIAPUNOV STABILITY
In some cases, input-output stability methods can be used to establish
Liapunov stability. In this section, we shall present some results that
state such results precisely.
3 e(r) = u(/) — f e Mt
~ r]
y(t) dr
Jo
~ r)
5 x{t) = e At x 0 — f e A{t { [t, x(t)] dr
Jo
6 u (t) = e At x 0
in the system (3)-(4) (i.e., the system of Fig. 6.3), then the “error” e(-) is
7 ||e
Ar
x0 1|
< me~ at
, V f >0
where ||
•
||
denotes the Euclidean norm. Thus the particular input u(*)
defined by (6) belongs to L n2 Therefore, if the system (3)-(4) is I 2 -stable
.
(in the sense of definition [6.3(34)]), then both e(*) and y(-) belong to L\.
Let us define
~ r)
8 z (t) = e Mt y(r) dr
[ - >>
Then, by (3),
rt/2 rt
= e Mt
~ x)
y(r) dr + e Mt
~ x)
y(r) dr
JO Jt/2
after changing the dummy variable of integration in the first integral. Thus
by the triangle inequality, we get
where || •
|| f
denotes the induced /2 -norm on RnXn (see Sec. 3.2). Next, by
Schwarz’s inequality on L 2 we ,
get
' 2
12 ||z(0||
< [J (||e
At
||i)
2
dxY •
||y(f - r)||
2
dxj'
( 2 [JJ 2
+ 1/2
(||eA< '' i>
||i)2
1/2
dx
T •
[£ 2
||y(T)| 2
1/2
'
dx
T
< [J “ (lk A 1l0 ^]
;
2
•
[J~
||y(T)||
2
rfr]
U1 " 1/2
eAt d *]
+ [J0
°°(H lli)
2 •
[J(/ 2
||y(T)|prfr]
Sec. 6.4 Relationships Between I/O Stability and Liapunov Stability 243
16 rank [c' |
A'c' |
. . . |(A')”
_1
c'] =n
Associated with the system (15), define the single-input/single-
output feedback system (shown in Fig. 6.4) described by
proof The proof closely parallels that of Theorem (1). First, (15) can
be expressed as an equivalent nonlinear integral equation
19 x(/) = 6?
Af
x0 — 1
e A(
'~ T)
b0[T, c'x(t)] dr
Jo
Therefore
21
—
II o **
^ 'w'
which is the same as (17)-(18). Because w(*) e I 2 if the system (17)— (18)
,
v(0 — > 0 as t — oo, we can show by standard arguments from linear sys-
tem theory that x(f) — 0 as t — oo. —
Theorems (1) and (14) provide a glimpse of how one can prove
Liapunov stability results using the input-output approach. Specifically,
any criterion for the Instability of the type of system shown in Figs.
6.3 and 6.4 also yields a corresponding result for the global asymptotic
stability of systems of the type (2) and (15). The advantage of using
input-output techniques to tackle Liapunov stability problems is that
some of the stability criteria are better motivated in the input-output
framework. However, the disadvantage is that input-output techniques
yield global asymptotic stability or nothing at all. It is difficult to
tinuous; and the pair (A, c) is observable. Suppose 0(-, •) belongs" to the
sector [a, JJ], i.e.
a<7 2
< 0(7, a) < P<7 2
, V t > 0, V g e R
c'(5l — A) -1 b + d
FIG. 6.5
Sec. 6. 5 Open - L oop S tability of Linear S ystems 245
y e L2 ,
w(t) — dy(t) — > 0 as t —* co and hence w(t) — d(j)[t, —w(t)] — 0 as
t — > co. Thus conclude that w(t) — > 0 and hence x(t) — » 0 as t — » oo.}
65
OPEN-LOOP STABILITY OF LINEAR SYSTEMS
Before attempting to study the stability of interconnected systems such
as [6.3(19)-(22)], it is helpful to first obtain conditions under which the
operators Gi and G2 represent L p -stable subsystems. In this section, we
shall concentrate on linear systems for the most part and obtain necessary
and sufficient conditions for a linear system to be L p -stable. The term
open-loop stability refers to the fact that we are studying the subsystems
Gi and G 2 rather than the , overall closed-loop system, which is described
by [6.3(1 9)— (22)1 -
(This is much harder to achieve using Liapunov theory.) Then, once the
single-input/single-output case is thoroughly analyzed, the results for
the multi-input/multi-output case follow easily.
To do this, we introduce the sets Cfc and Basically, (as shown
later) a is the set of stable impulse responses, and & is the set of stable
transfer functions. The precise definitions are given next.
/
fO, t < 0
2 f(t) = J oo
t> 0
U= 0
where <?(•) denotes the unit delta distribution, tt are nonnegative
constants, fl (.) is a measurable function, and, further,
3
1
S
= 0
l/d < 00
4 [\m\dt<™
Jo
The norm ||
•
|| a of a distribution /(•) in 0, is defined by
5 ll/(-)lla =L
= 1 0
I/I + Jof \m\dt
The convolution of two distributions /(•) and g(*) in a, denoted
by f *g, is defined by
Thus, the set a consists of all distributions that vanish for t < 0,
and for t > 0 consist of a sum of delayed impulses and a measurable
function, with the additional property that the weights of the impulses
form an absolutely summable sequence and the measurable function
is absolutely integrable. One can think of Ct as the space L augmented x
by delayed impulses.
Note that, in computing the convolution of two distributions,
one should take
i =0
10 g(t) = S giS(t - t!
g)
) + g (t)a
Sec. 6.5 Open-Loop Stability of Linear Systems 247
11 (/ * g)(t) = tt AgAt -
i =0 j= 0
t\
f) - & y)
+ £figjLt-tn
i=0
+ t=0
£g,f.(t-tP)
+ j0
f/aO - PgaiPdT
14 f2 ( t) =2
= 1 0 (z +
1
^
1)
— zT), T >0 a given number
does not belong to d because the sequence [l/(z + 1)]q is not absolutely
summable. The distribution
16 /4 W = <5(0 +
belongs to d and has a norm of 2.
17 Remarks :
1. Note that L 1
is a subset of d; further, if /(•) e L ls then
18 ll/(*) lla
= 11/(0 Hi
(i) | H
la is a norm on d, and & is complete under this norm.
(ii) The convolution * is distributive; i.e.,
20 /* O + h) = f*g +/* h, V f,g,h g a
21 ( f+g)*h=f*h+g*h , V f, g, h G d
(iii) The convolution * is commutative; i.e.,
23 \\f*g\\a<\\f\\a-\\g\\a
24 5*f = f* 8 =f V/e d
(vi) Cfc has no divisors of the zero element; i.e.,
25 f g * = 0=>/ = 0 or ^ = 0
26 (/**xo =
i
±
=
£= figAt - p -
0 j 0
t 1</>)
S figait -
+ 1=0 t\
f)
) £ g,ut -
+ 1=0 1!*>)
+ f'/M - T)ga(r) dr
separately. First,
27
i
S
= 0 7=0
SI/,Mftl = (SI/«l)-(S
V =
'*’= 0 / 7 0
k,l)
/
Next
28
f“
Jo
I
I 1
S
=0
fiUt - tl
f)
) dt<±
= 1 0
\ft \
r
J0
| g/f
- /P)l dt
Similarly,
29
f Ii. *- (,§
i*i) •
*
I [j;
u
Finally,
f° (t ~ dx dt
— Jo I
“ T) H I
dr dt
Jo I Jo I Jo
I
C+ = {s: Re s > 0}
With this background, we define the set &.
proof If (i) and (ii) hold, it is clear that /(•) [the inverse Laplace
transform of/(*)l consists of (i) a measurable function that is bounded by
a decaying exponential, and (ii) possibly an impulse at t = 0. Hence
/(•) e i.e.,/(-) e &. On the other hand, if (i) does not hold, then /(•)
contains higher-order impulses and hence does not belong to a, while if
(ii) does not hold, then the measurable part of
/(•) is not absolutely
integrable.
250 Chap. 6 Input-Output Stability
The set a, has some very useful properties. Most physical systems,
even “unstable” systems, have impulse responses that belong to
=
a e [e.g., h(t) e1]. Moreover, it can be shown that if /;(•) is any dis-
tribution that vanishes for t < 0, then the operator H defined by
36 (Hf )(t) = f* h(t t)/(t) dx
d0
maps Lpe into Lpe for all p e [1, oo] if and only if /?(-) e GL e 9 In other .
proof We shall first prove that (iv) implies each of (i), (ii), and (iii).
38 ll/Olla = ||/(’)lli
9 The proof of this statement is very similar to that of Theorem (37) and is
therefore left as an exercise.
<
Next, because /(•) contains no impulses, neither does h*f, which means
that h *f e L u and
40 P*/IU = P*/lli
Now, (39) and (40) together imply that
41 P*/lli<£PMI/lli
which shows that is L j -stable. H
(iv) => (ii) Suppose that/(-) e L„ and that h( •
) is of the form
42 hit) = i—0
S h,5{t - t t) + ha (t)
Then
= ll/(-)IU-P(-)ll«
Because (44) holds for (almost) all ?, we see that h *f e L„ and that
45 P*/IU<Plla-||/ll~
This shows that H is L„ -stable.
(iv) This part of the proof is rather involved and can be skipped
=> (iii)
without loss of continuity. Because we have already established (i) and (ii),
we shall concentrate on Z^-stability for 1 p < oo. Suppose /(•) G L < p .
(42); then h* /is of the form (43). Now, if /(•) e L p then clearly the ,
function
belongs to L p and ,
48 ha (t - x)f(x) dx A g 2 (t)
jo
Let q = pj(p — 1) so that (l/q) + (1/p) = 1. Then
49 1
82(f) I
<£ I
K(t - t) I
*|/(t) j
dx
=£ I
K(t - x)\'/p\ha (t - xY'f\f(x)\dx
< |J 1
Kit - T)1 rfrj
/9
• ha (t - t) * f(x) p dx
17 '"
0 [Jj I | \
J
252 Chap. 6 Input-Output Stability
J0
‘
m- T)|.|/(T)P dx
Now, if /(•) e Lp ,
then the function r i—
> |/(t) p \
e L { . Thus, using the
estimate (41), we get
9
51
J~
|* 2 ( 0 I
P dt < [J“ \h (x)\ dx^' a
•
f~
\ha(x)\ dx •
Jo |
f(x)\> dx
now prove the converse, namely, that (i), (ii), or (iii) each implies (iv).
(i) => (iv) Suppose (i) is true. Because H is linear, (i) implies that H
is continuous. Now, because can be approximated every element in &
arbitrarily closely (in the sense of distributions) by an element of L u this
implies that iTmaps (2 into itself. Let f(t) = 8{t) then h * f e (2, because
,
\
(ii) => (iv) This is a special case of a result contained in Theorem (70).
(iii) => (iv) Suppose (iii) is true, i.e., that H is L p -stable for all
implies (iv).
<^J l/(/®)l
2
^=« 2
ll/ll!
(i) Testable, (ii) Z^-stable, (iii) Z^-stable for all p e [1, oo] if
and only if /?(.) e d.
3. We have established the useful bounds (53) and (55). These
show that, given a system with impulse response h(>) e d 9
where the impulse response matrix H(.) e a™*"; 10 i*e., the entries of
the m X n matrix of distributions H all belong to d e In analogy with .
1
°To avoid confusion, we write H for the operator and H(«) for the associated
impulse response.
. ,
where
63 «i = l|Mj ||( 2
||
•
|| i2 indicates the /2 -induced matrix norm on Rmxn and ,
M t
is the
mX n matrix whose ijth entry is ||/z l7 (.)|| a If p = 2, then
.
where
65 a2 = ||M 2 ||, 2
66 (M 2 ) i7 = sup |
hij(jco) |,
i=l,..., rn,j = 1, . . n
CO
(64) reduce to (53) and (55), respectively, in the case of scalar systems.
68 (Gf)(t)= X)
iei(t)
gi(t)f(t — t,)+ [
Jo
ga(t, x)f (t) dx
where
69 I(t) = {/: t t < /}
In other words, (68) is obtained from (67) by taking the summation only
over those indices i such that t t t. <
Theorem (70) gives necessary and sufficient conditions for an
operator G of the form (68) to be Z^-stable.
71 t
E
iei(t)
1
g,{t) |
€ Lme
72 x i->
ga(t, x) e Lu Vt>0
73 t t-y f | ga (t, r) dx |
G L me
Jo
74 sup
t
I 2
\iei(t)
| gi (t) |
+ Jof |
g„(t, x) dx)
|
A c„ < oo
)
(i) “T/ ” Suppose that (74) holds and that /(•) e Then
<^||/(-)IU
Because the right-hand side of (75) is independent of t, we see that
(00(0 G Too and that
76 11(00(011°° < Coo ||/(0IU
Hence G is Too-stable.
(ii) “Ow/y If' We shall show the contrapositive, namely, that if (74)
does not hold, then (GO ||oo/||/||oo can be made arbitrarily large by
||
77 r{t)A Z
iei(t)
k*(0l+ [
Jo
\ga (t, T)\dx
256 Chap. 6 Input-Output Stability
> k. The first step is to pick some € > 0 and to choose 9 k such that
r(9 k ) >
k + e. This is always possible because r(-) is unbounded.
Next, pick intervals S of the form t
78 = Wk — ti —G i, 6k — ti + Gil, i E I(6 k)
such that
79 2
i&1 ^) JrGSt
f \ga{9 k , T)|rfT<4-
1
80 m =
(
l
sign gi{9k),
sign ga (6 k ,t),
if t
if t
e
i
<5,
U
teHM
$t
+I ga{0k, T)A(T) dr
TG[O,0 t]
xt U Si
iei(d k )
- 2 f \ga(0 k ,r)\dr
iei(O k) J T edi
83 (Gfk)(6 k ) > ( 2
(iene k )
gi(9 k )+ C\ga(9k,r)\dr]
Jo )
k -\- 6 — € =k
because the first term in the braces is not less than k + € by the way
we chose 9 k ,
while the second term in the braces is not more than €
by (79). Finally,
that G
not Loo-stable. This shows that (74)
is is a necessary condi-
tion for G to be Loo-stable. —
The next theorem gives necessary and sufficient conditions for
G to be Lj-stable.
86 2
iei(t)
| g (t
t + t *) |
G L oe 0
89 sup
r
{£
U=0
\gfy + 01 + f
Jt
I gaif, t)| dt) Ac^oo
)
90 rmm
Jo
dt < f s
Jo fe/(r)
\ g,(t)\-\f{t-td\dt
+ f Jo
Jo
i)\-\m\dTdt
=s
= i 0
r \g (f)v\ Kt -
Jo
t t t)\dt
+ \~\ga(t,T)l\m\dtdT
+ J
\ 8a(J ’ dt dZ
Jo
< sup r £
=
i
gi(x + /,)i
T Ll 0
+ {Jj/(r)|rfr}
= Cl||/(0lll
This shows that G is Li -stable.
(ii) “ Only If” Suppose G is Testable. Because G is linear, this implies
that G is continuous on L and
x hence that G maps Ct into Ct. Let
258 Chap. 6 Input-Output Stability
91 ( Gf)(t )
= 2 Si(f)5{t —T— tt) + ga (t, t)
iei(t)
where we use the fact that ga (t, t) 0 for t T. Now, the In- = <
stability of G
not only implies that (92) holds for each fixed T but
also that ||G/(-)||a is bounded independently of r; i.e.,
95 \\Gf\\ p ll/ll,
proof We shall give the proof only in the case where g t = 0 for all i.
Suppose /(•) g L p We . can assume that 1 < < 00, because (95) is
98 |(G/)(0|< [ \gait,z)V\m\dz
Jo
= fkaaT)|^ka (/,T)|^|/(T)UT
Jo
1/9 1/P
< {£
I
gait, T) |
</rj
{ £ |
r) |
• |/(t) 1?
Sec. 6.5 Open-Loop Stability of Linear Systems 259
99 I
(G/)(0 1*
< git, X) rf*} | |
git, T) |
• [/(T) |* dx
{ Jj { J“ j
fV.a,T)M/(T)|'rfT
Jo
100
f"
Jo
mm* * < r Jo \git, z)\-\m\
Jo
p dx dt
= p
c~ q
\gln)V\m\ p dt dx
= <£• I
git, t) dt~\ | \mv dx
J“ [j;
< dH*c x ll/ll?
3. We
have proved the useful inequality (95), which gives an
upper bound for the Z^-norm of ((!/)(•)•
We shall conclude this subsection by showing that if G is a time-
invariant operator, then Theorems (70), (85), and (94) together reduce to
Theorem (37). Accordingly, suppose
101 g (t )
t
=h t
(a constant), V t >0
102 ga (t, T) = ha(t — t), V t, T > 0
104 sup
t
(
Uei(t)
2 \K\ + Jofl h (t — xa ) |
dx)
)
< oo
we have
106 cx = = P(-)IU
in this case, (95) reduces to (53).
\f(s)\<\\f(-)\\a
Problem 6.7. Show that iff(-) e then the function 5 e~ sTf(s) also
belongs to & for all positive T. [Hint: consider the inverse Laplace transform
of e~ M]
sT
(i) m=£ i =0
S(t — iT), T> 0
= e~ sT s 2 + 5^ + 5
2
(0 f(s) s + s + 10
(ii) f(s) = 1/cosh V ^ (Hint: use partial fractions.)
6.6
LINEAR TIME-INVARIANT FEEDBACK SYSTEMS
we study the conditions under which a feedback inter-
In this section,
connection of linear time-invariant subsystems results in a stable sys-
tem. Since the results for linear time-invariant systems are the ones
most often used in practice, and since many of the stability results for
nonlinear and/or time-varying systems use the results for linear time-
invariant systems as a point of departure, it is important to have a good
understanding of what makes a linear time-invariant feedback system
stable or unstable.
)
The following result is the key tool used to derive all the stability
criteria of this section.
5 l^y^lkCOIIa, V s S C+
But this shows that
6 lfwls >0 V s e C.
iT«M -
(ii) “T/ ” The proof of the “if” part is far beyond the scope of this book
and can be found in [14] Hille and Phillips [p. 150]. M
Lemma (1) states the following: Suppose p(-) is the Laplace
transform of an element of ft. Then its reciprocal 1 /p(*) is also the
Laplace transform of an element of ( if and only if /?(•) is bounded
away from zero over the right half-plane. In particular, this means that
| |
—> oo can be found such that p(s ) t
— > 0 as / —> oo.
proper and has no poles in C+ Then]. l/p(s) belongs to (I if and only if (1)
The proof is
left as an exercise.
Notice that in proving the “only if” part of lemma (1) we never
made use of the fact that p e &. In view of this, we can state the fol-
lowing.
lemma Suppose p(*) is a distribution, and let p(*) denote its Laplace
transform. Under these conditions, if !/]?(•) belongs to then (2) holds.
ficient).
Us) = ,
i+ g,(s)g
,
444r ,-M - +f^7vw
2 (s)
ii
^i(«)f
) 1
i 2 (-y)
yi(s)
iMUs) Mi (s) - Ms) u 2 (s)
i + ii(s)g 2 (s) 1 + £i(^)£zO)
FIG. 6.6
2
hij(-) E (3 for /, j = 1, 2.
,
1
14 K-)A 1+ #l(0#2(0
e <3
15 A*) EE 1
11 Or, equivalently, (1) L\ -stable or (2) L^-stable for all p e [1, 00].
0
21 lemma Suppose
22 g 2 (s) = g 20 =£ 0 V s
A A
Then e A for i,j = 1,2 if and only if red, where r is defined in (14).
proof
(i) “If” Suppose fed. Then,
- gl(s)g2
23 1 r(s) =
[1 + #l(^20]
also belongs to &. This shows that /* 12 , ^ 2 i e S. Also, since g 2
is a nonzero constant, we have that
^n0) = [1 “ r(s)]/g 20
belongs to &. Finally,
24 ^ 22 ^) = gioffs)
belongs to d.
zeroes of d x
are in C+. Under these conditions, the system (9)— (10)
12
is Loo-stable if and only if
inf |1
sec+
+g 20 liO)| > 0
Remarks:
(i) The condition (27) has exactly the same interpretation as
(20); namely, for the class of systems under study, a necessary
and sufficient condition for stability is that the magnitude of
the return difference is bounded away from zero as s varies
over the closed right half-plane.
(ii) Note that there is noassuming that all
loss of generality in
zeroes of d 1
are in C+
g has some poles in the open
,
since if t
28 p(s)
_
= «i0)
(s + 1)«
29 ? 0) = dM
(s + 1)«
= + gf “
(J)
30 Us) - #,. W + 1
f«
,,, q(s)
31 ns ~
{?('’) + gio[p(s) + ?0)£i&0)]}
We now claim that, if (27) is true, then
32 inf \q(s)
sec+
+g 2 o[p(s) + q(x)gib(s)] > \
0
33 inf
SEC+
|^»[1 + g 2 o#i(s)]| > 0
Hence we conclude that no such sequence fe)? exists, i.e., that (32) is
true. Next, by lemma (1) and the fact that q + g 2 o(p + qgn) e d, we
have that 1 /[q + g 20 (p + qgn)] e d, whence red, by (31). By lemma
(21), this shows that the system (9)-(10) is Loo-stable. M
Consider next the case where g x and g 2 are both of the form (26).
In this case, the necessary and sufficient conditions for stability are
slightly more complicated.
39 (i) inf 1 1 +g t
(s) g 2 (s) |
> 0,
s GC+ ^
40 (ii) d (s) g (s)^0 whenever d 2 (s)
t x
= 0,
41 (iii) d2 (s) g 2 (s) ^ 0 whenever d (s) x
= 0.
Remarks: Theorem (37) states that three conditions together are
necessary and sufficient for stability:
(i) the return difference condition (39), which is the same as
before;
(ii) the quantity d g does not vanish at the poles of g 2
1 1
in C+ ;
and
(iii) the symmetrical condition about d2 g 2 and g l9
Conditions (ii) and (iii) assure the complete shifting of the C+ poles
of ii and g 2 and can be rationalized as follows: suppose s 0 e C+ is
,
(40) and can be checked easily, and only (39) is involved. The rest
(41)
of this subsection is devoted to the presentation of a graphical stability
criterion for verifying a condition of the form (20) [or (27) or (39)], by
examining only 1 +
gfjto) § 2 (jco) as co varies over R. The criterion here
is a natural generalization of the well known Nyquist criterion, and has
the same advantages associated with the Nyquist criterion, namely (i) :
42 r(s ) = + giC^OO = + S
1
=
rfi~
sU
+ rfs) 1
/ 0
"
+ SE
^Pi(s- Piy
where tt are nonnegative constants, p e C+ is
t ;
the multiplicity of
the pole pp, and n p is the total number of poles of f in C+ ;
in addition,
43 Sk/I<°°
/=o
44 r a { •) E Li
Basically, in (42), we have splitup r into a part that belongs to &
(namely the first three terms on the right side), plus the contribution of
the right half-plane poles of r expressed in partial fraction form.
,
Now,
if Theorem (19) or Theorem (25) is being applied, then the return dif-
268 Chap. 6 Input-Output Stability
45 f(s)
— pf
(s
46 inf
sec+
|
r(s ) |
>0
In general, this problem is difficult to solve. However, if the “delays”
t, in (42) are rationally related, i.e., if there exist a number T and inte-
gers such that
47 t, = n[T , V i
49 inf
s G=5i
|
r(s) |
>0
If we repeat thisprocedure at each j co-axis pole of P, we get a finite num-
ber of half-disk regions where (49) holds. Let us now delete these half-
]
disks from C+ ,
and denote the resulting indented right half-plane as Ci+
(see Fig. 6.8). Now, because of (49), we have
50 inf
sECi+
|
r(s) |
> 0
FIG. 6.8
270 Chap. 6 Input-Output Stability
57 inf
sEC+
1 1 + & g(y) > |
0
FIG. 6.9
14 Or equivalently, L\ -stable,
(i) (ii) Z^-stable for all p e [1,°°].
Sec. 6.6 Linear Time-Invariant Feedback Systems
271
= e - ^ + 4—* + 2
2
60 $,(*) 2
s 1
62 Si(s)
A{s)
line. It is known that, for a general nonuniform line, A(s) is of the form
63 A(s) =n
i=
0 + Pi)
1
64 pi ~ a/
2
Hence applying Theorem (58) to the present case, we see that the sys-
tem under study is Testable if and only if the Nyquistplot of g,(yco)
neither intersects nor encircles the critical point — \\k + /().
_1
65 lemma Let F(*) e & nxn Then
. [F(-)] e & nXn if and only if
66 inf |
det F (s) > |
0
iGC +
proof
(i) “/jf” We can write
1
67 [F (s)r l = Adj F (s)
det FCy)
Sec. 6.6 Linear Time-Invariant Feedback Systems 273
where Adj F(V) denotes the adjoint matrix of F (s), i.e., the matrix
consisting of the cofactors of F(V). Because every cofactor of F(s )
consists of sums and products of the elements of F(s) and because
every element of F(-) belongs to <3, it follows that Adj F(-) e &
nXn .
Similarly, det F(-) e (3. Now, if (66) holds, then by lemma (1),
-1 nXn and
1/det F(*) e (3. Hence [F(-)] is the product of matrix in &
69 = u, - y 2
70 e2 =u +f 2 1
71 Yi —G e 1 1
72 y2 =Ge 2 2
The system (69)-(72) is the same as that in Fig. 6.5, except that gj and
g 2 are replaced by the matrix transfer functions Gj and G 2 As in the .
where
75 e = [«', ei]'
76 a = [a; KY
77 y = [y'i y'lY
"
78 G(s) = "O.W 0
0 G 2 (s)_
79 H.r°_ ‘i
-I 0
274 Chap. 6 Input-Output Stability
81 [I + h6(-)] _1 e
or (2)
82 G(-)[I + HG(.)]
_1
e & xn
83 e = (I + HG) _1 u
84 y = G(I + HG) _1 u
The conditions (81) and (82) now follow readily. M
Thus our objective is to find conditions on the subsystem transfer
functions G and G, which ensure that (81) or (82) holds. In the case
,
where both 6] and G 2 belong to fi" ” (i.e., the open-loop stable case),
x
86 inf
s EC+
|
det[I + h6(s)] > 0 |
r j g n
= det (I + G
^ ^
+ HG) =
2
87 det (I det a 2 6i)
G i
1 _
15 Or, equivalently,
(1) Li-stable, (2) Z^-stable for p e [1, oo].
16 Or, equivalently, (1) Li-stable, (2) L^-stable for all p e [1, oo].
S ec. 6. 6 Linear Time-In variant Feedback Systems 275
u2 — Si —
ku a nonzero constant. Then the closed-loop transfer
function from u x to is
88 h ll() M- 1
liw
+ kg^s)
Now, suppose s 0 e C+ is a pole of |i Because we can write
89 * ll(j)
k + 1/^(5)
it is clear that h lx is analytic at s 0 and that R u (s 0 ) = 1/k. Thus, if we
place a nonzero feedback around all poles of £, in C+ are shifted
from their original locations. Now, if we can assure that the new pole
locations are not in C+, the system is stable (loosely speaking). This is
the purpose of condition (39). Conditions (40) and (41) ensure that the
effective feedback at the pole locations of g, and
g2 is nonzero, so that
pole shifting does take place.
In contrast, in the multivariable case, it is quite possible for both
the open-loop and closed-loop transfer functions to have poles at the
same point in C. For example, let
"
1 /(s - 0
'
'2 2
"
Gifc) =
1)
; ^ 2 (s) =
0 Vis + 1)J 2 0
and suppose u 2 =
0. Then G 2 represents a constant nonsingular feed-
back matrix, and the transfer function matrix from u, to y is ,
s + 1 -2
90 Hn (5) = 6,(1 + 626,)- (s - Vis + 3) - Vis + 3)
-2 s + 1
Problem 6.12. Consider the feedback system shown in Fig. 6.9. Deter-
mine the range of values of the constant k for which the feedback system is
Coo-stable, with
(i) g(s) = e -I (s 2 + 1)
s2 + 2s
= 1 — e~ 2ss
(ii) g(s)
s + 1
As (s 2
= e~ -f 4s + 4)
(iii) g(s)
s2 — 2s — 3
(iv) #(s) = 1 - e~ + e~ + s 3s
s + 5
Chap. 6 Input-Output Stability
276
Suppose
" S
e s
1
= s +2
Gi (j)
Lj
—+TT 1
1 - e
'3*
- n 11
=
1
1 —-1
+
1
7“*
< i)G ^ )=
Lo 2J
(ii) Ga (j) s 3
_2 0 -
01
(iii) G 2 O) = P
L_0 e 'J
.7
~
TIME-VARYING AND/OR NONLINEAR SYSTEMS
In the previous section, we studied the feedback stability of linear time-
invariant systems. In this section, we study nonlinear and/or time-
varying systems. Our attention is centered exclusively on single-input/
single-output systems. The stability of multi-input/multi-output sys-
tems, which requires more elaborate techniques, is studied in detail in
[10] Desoer and Vidyasagar.
In this section, we prove two well known results, known as the
circle criterion and the (generalized) Popov criterion, respectively. The
circle criterion provides a sufficient condition for the L 2 - stability of a
feedback system containing a linear time-invariant element in the for-
ward path and a memoryless time-varying and/or nonlinear element in
the feedback path. We show that the circle criterion is also a necessary
condition for stability, in a sense to be made precise later on. The gen-
between the straight lines passing through the origin with slopes a
and p, respectively. This is illustrated in Fig. 6.12.
and, moreover,
5 II
* u) T 1 2 < sup \h(jCO)\-\\uT
03
\\ 2
6 (h * u) T = (h * uT)T
Thus
7 ||(^ * u)t\\i — II
(h* u T) T || 2 < \
\h * uT \\ 2
< sup 03
|
h(j q)\*\\ut \\ 2
FIG. 6.13 F
(i.e., the “delays” are all equally spaced). The element F is memory-
less and is described by
where 0(., •) is continuous and belongs to the sector [a, /?]. Under
these conditions, the system under study is L 2 -stable if one of the
following sets of conditions, as appropriate, holds : If a, ft ^ 0, let
D be the disk in the complex plane centered on the real axis and
passing through the points 1/a j 0 and 1 //? j 0. — + — +
(i) IfO <a< p: (a) When the 70-axis is indented as in Sec.
6.6.2, the graph of g(ja>) is bounded away from the disk D ;
i.e.,
13 inf g(y'o)
|
—z >0 |
(OdzR
ZED
n
_ Arg [g = 2 nvp
14
^ D
lim Arg [gSBl
^ f -
)
,
\/ z
(ii) I/O = a < fi: (a) v p = 0, i.e., g has no poles with positive
real parts, and (b)
Remarks:
1. Suppose
16 g(s) = c'(sl - A) _1 b + d
where the matrix A is Hurwitz (i.e., all eigenvalues of A have
negative real parts), the pair (A, b) is controllable, and the
pair (A, c) is observable. Then Theorem (8) gives a sufficient
condition for the L 2 -stability of the system of Fig. 6.5. By
Problem 6.5, the same conditions also guarantee the global
asymptotic stability of the equilibrium point 0 of the system
[6.4 (27)]— [6 4(29)] .
;
Thus Theorem (8)
see also Problem 6.14.
contains, as a special case, the circle criterion for Liapunov
stability (Theorem [5.5(60)]). The details are left as an exercise.
On the other hand, Theorem (8) is more general than Theorem
[5.5(60)], because the former is applicable even to distributed
proof Define
17 c =^+ «
2
II fc*
1 &
18 >
2
so that
19 =c—r
OC
20 P=c+r
280 Chap. 6 Input-Output Stability
Rearrange the system of Fig. 6.13 as in Fig. 6.14. In this way, the system
under study is transformed into another equivalent system whose forward
element G t has the transfer function
21 gt(s) =
1 + cg(s)
and whose feedback element F t is described by
22 (F y)(t)
t = 01 \t, y(t)] - cy(t) A y(t)]
FIG. 6.14
Clearly, the new function 0,0, •) belongs to the sector [— r, r], Now,
regardless of which set of conditions (i)-(iv) is applicable, one can verify,
using Theorem [6.6(58)], that g e & and
t that
23 sup [#,0'«)l
CO
= < v '
The details of the verification are left as a problem (see Problem 6.15).
Now, with regard to the rearranged system of Fig. 6.14, we have that
g g
t Qj and that 0,0, •) belongs to the sector [— r, r]. The latter condition
means that y lt e L le whenever e 2t e L le and that
26 eu = u - F e 2t x t
27 e 2t = u 2 + G e u t
— [
28 (.
— U \T ~~
e lt)r \Fti.e 2t)AT
29 ( e 2t)r
— u2 t ~ [Gt(e u)r]T
< 1 +
1 1 12 1 1
-^*(^ 2 f)r II 2
<ll«irll 2 + r\\(e ) T 2 2t \\
[by (24)]
31 1 1 2 r)r 1 1 2 ^ Il
w 2rll2 + [^(^lf)r]rll2
II
< Il
w 2 r ||2 + \\G (e U T t ) \\2
< II
«2 r|| 2 + yo||(eif)rll 2 [by (25)]
obtain
34
get
35 ll (e2()r || 2
<lMk^MM
ry 0
s(s )
s.
= e_ 5 s
2
~+2
+ 4s
s2
Suppose it is desired to find the largest sector [a, fi] such that the circle
criterion is satisfied. Note that g has one pole with a positive real part,
vp = 1 ;
therefore if a < 0, the conditions for cases (ii)-(iv) of Theorem
17
(8) can never be satisfied. Thus, for the circle criterion to hold, we
must have a > 0.
The plot of g(jco) isshown in Fig. 6.10. For the system to be L 2 -
stable, the plot of g(j co) must encircle the critical disk D exactly once, in
a counterclockwise sense. From Fig. 6.10, we see that this is the case
provided 0.5 <a< < 0.55.
Hence the system is Z, 2 -stable for all
ft
nonlinearities in the sector [a, /?] with a > 0.5 and < 0.55.
We now show that the circle criterion of Theorem (8) is also a
necessary condition for the stability of a certain class of systems. First
of all, by applying Theorem (8), case (iii), we get the following result:
CO
39 g(s) = g + g (s) +
0 a ga (-) G L x
FIG. 6.15
Thus g is of the form (10)— (1 1), except that there are no delay terms.
Suppose we ask: What conditions must g satisfy in order for the system
of Fig. 6.15 to be Z^-stable for all gains k e [— r, r] and for all delays
t>0? To put it another way, under what conditions does g(s)/
[1 + k e~ g(s)\ e (I for all k e [— r, r] and for all x
sx
0? The answer >
is readily obtained from an application of the graphical stability crite-
proof
(i) “ If ” If g e & and (38) is satisfied, then the polar plot of g(jco) is
plot of e~ jcor g(jco) is bounded away from, and does not encircle, the
point —
1 /(k +/0) whenever k e [— r, r]. Also e~ ST g(s) e <5,
——
S ec. 6 7
. Time- Varying AND/OR Nonlinear Systems 283
V T > 0.
Hence by Theorem [6.6(58)], we see that g(y)/[l
+ ke~ ST
fi, V k G [— r, r], V T
g(s)] E 0. >
(ii) “ Only If” Suppose g(s)/[l + ke~ STg(s )] e & V k e [— r, r],
>
V T 0. Then, by setting k = 0, t = 0, we see that g e &. Next,
by the Riemann-Lebesgue lemma, ga (jco) >0 as |o| * oo, —
which means that g(jol) >g 0 as |co —
» oo. The hypothesis |
—
implies that g 0 < r~\ because if |g 0 > r” 1 then the function
| 1 | ,
Hence the polar plot of e~ Jcorg(jco) intersects the point |z| + j0,
which shows that £(s)/[l + ke~ srg(s)] ^ ft with k = \z\~ 1 , because
the stability criterion [6.6(58)] is violated. This contradicts the
hypothesis that g(s)/[l ke~ sz g(s)\ g S + V k e [— r, r ] and
V t > 0. Hence we conclude that no such number z exists, i.e.,
that (38) holds. —
Thus Theorem (40) shows that the circle criterion is not at all
conservative when applied to systems of the form shown in Fig. 6.15.
Actually, we have restricted k to belong to the symmetrical interval
[— r, r] only to keep things simple, and we can prove similar results for
the case where k in Fig. 6.15 varies over an arbitrary interval [a, ft].
Combining Theorems (40) and (8), we get the following important
result.
t > 0. Then the system of Fig. 6.13 is L 2 -stable for all nonlineari-
ties 0(«, •) in the sector [—r,r].
43 Theorem Suppose
44 gC?) = c'(sl — A) -1 b + d
where the matrix A is Hurwitz, the pair (A, b) is controllable, and the
pair (A, c) is observable. Suppose the system of Fig. 6.15 is L
284 Chap. 6 Input-Output Stability
47 u(t) = j(?)]
48 lemma Suppose <j>: R —* R belongs to the sector [0, k]\ i.e., suppose
51 <3>(<7) = 0(<7) do
J*
Then <D(cr) >0, V o e R, because of the sector condition (49). From
Fig. 6.17, we have
52 y 2 (t) = 2 (t)]
</>[e
= ^(0 dt + q
T Jo\ (*
Jo
e 2 (t)<f>[e 2 {t)\ dt
T
60 P r(t)f(t) dt> 5 Jo f\t)
Jo
[ dt, V T> 0
where r = h*f
= \~(h*fT)(t)fT(t)d t
,
i
2n \fT (jco)\*dco
I
= <5|IMII
where fT denotes the Fourier transform of /r (*). Clearly (62) is the same
as (60). ^
Now we come to the following.
64 g(-),g(-)ea
Suppose the nonlinearity 0(«) satisfies (49). Under these conditions,
thissystem with inputs u and z is L 2 -stable if there exist constantsx
e1 and e2 belong to L2 ,
because these are the quantities of interest in
the original system (shown in Fig. 6.16). This is not very difficult. If
q > 0 and u 2
an arbitrary element
is of L 2 the “input” z may or may ,
66 fact Under the conditions of Theorem (63), the system of Fig. 6. 1 6 has the
property that e u e 2 E I 2 whenever u u u2 e L2 ,
with the added assump-
tion that u 2 E L2 if q > 0.
67 <\fufi)T = f /i OO/iW dt
Jo
Note that </i,/2 >r is the inner product (in the L 2 -sense) of the L 2 - func-
tions and f2T , so that by Schwarz’s inequality, we have
68 K/ij/iX 1
! < ll/irlta’IIArlk
For the system of Fig. 6.17, we have
69 et = u — y2t
70 v =z+w
Hence
71 <<?i, w> r + <>, y 2 >r = <e u v — z) T + <v, «! — ei> r
= — (e z) T + (e u v) T ~ (v, e y T +
l9 t (v 9 u t y T
Now, if we define
74 m = sup I
(1 + jcoq)g(jco) |
coER
we have
75 ||wr l| 2 <m\\e 1T \\ 2
76 (e l9 wy T >V 1 1 tfirlli — (v
2"^ ll^irlll
+ "j" Ikirlli 1
>0-4)11^+2^11^
= (y-l)\\e 1T \\l + £-2 \\(v-z)T \\l
Because y 2 = u — eu
1 (77) becomes
+ ^\\{v-z)T \\l
where c l9 c 2 c 3 are
, finite constants. Now, by (65),
and of course 3/2m 2 > 0. Thus the left-hand side of (80) is a positive
definite form in e 1T and vT This implies that \\e 1T 2 \\v T 2 are bounded
.
\\ , \\
81 Example. Let
^= s2 + 4s + 4
The Nyquist and Popov plot of g are shown in Figs. 6.19 and 6.20,
plot
respectively. From Fig. 6.20, we see that no matter how small 1 fk is
(i.e., no matter how large k is) we can always construct a suitable Popov
line, as indicated. Hence the system of Fig. 6.16 isZ 2 -stable for all time-
invariant nonlinearities 0(-) in the sector [0, k\ 9 for all finite k. On the
other hand, if we apply the circle criterion, case (ii), we see that
inf
oER
c
Re g(ja>) = OA
)
stable for all nonlinearities </>(•) in the sector [0, k]. In general, Aizer-
man’s conjecture is false. However, Popov’s criterion provides a means
of identifying a large class of transfer functions g(-) for which Aizer-
man’s conjecture holds.
Suppose g, g e 6t. If g contains any impulses, g would contain
higher-order distributions and hence would not belong to a. Thus
g, g E Ct implies that g does not contain any impulses, i.e., that g e L 1 .
Reg(/a;)
Then the stability of the nonlinear feedback system is assured for all
still does not follow that Aizerman’s conjecture is false for such a g.
Thus, in summary, Popov’s criterion provides a readily verifiable suf-
ficient condition for determining whether Aizerman’s conjecture is valid
for a particular transfer function g(«).
RegO'co)
Problem 6.14. Suppose g(-) is of the form (44), and that the criteria of
Theorem (8) are satisfied. Show that in this case 1 +k ^0 V k e [ -a, JJ]
Then, using the results of Problem 6.5, show that Theorem (8) contains
Theorem [5.5(60)] as a special case.
Problem 6.15. Show that if the criteria of Theorem (8) are satisfied, then
(23) holds.
Problem 6.16. Using the circle criterion, determine the largest range
[<%, /?] such that the system of Fig. 6.13 is Z, 2 -stable for all 0(«, •) in the
sector [a, /?] with
0) &(s)
s s )
g(s) =e 3s s
2
+ 3s + 2
s2 - 4
(ii)
mew - -^, 2
Problem 6.17. Using the Popov criterion, determine the largest con-
stant k such that the system of Fig. 6.16 is £ 2 -stable for all 0(-) in the
sector [0, k], with
(i) £ 0) =e j +4
i2 + +6
s 2
= e~ 2s i
(ii) g(s)
T+W
(
s 2 + 9s + 20
(iii) g(s) =e 3s
s3 + 6s 2 + 11s + 6
Do any of these transfer functions satisfy Aizerman’s conjecture?
Appendices
BELLMAN-GRONWALL INEQUALITY
1 Theorem Suppose c > 0, r(-) and k(-) are nonnegative valued
continuous functions, and suppose
2 r{t) <c+ f
•/o
k(r)r(r) dr, V t e [0, T]
Then
proof Let 5(0 denote the right-hand side of (1). Then (1) states that
d
8 5(0 exp -k(l) dx <0 V G T]
dt
X t [0,
292
)
The utility of this inequality lies in that, given the implicit bound
(2) for r(t) [notice that the right-hand side of (2) also involves /*(•)],
II
SUMMARY OF MATRIX EXPONENTIALS
AND STATE TRANSITION MATRICES
Given an n X n matrix A, the polynomial c(X) defined by
c(A) = det (21 — A)
c(A) =0
A polynomial m(X of the lowest possible degree such that m(A) 0 =
is minimal polynomial of A. A minimal polynomial of A is a
called a
divisor of the characteristic polynomial of A, and every eigenvalue of
A is also a zero of a minimal polynomial of A.
Given an n X n matrix A, the n x n matrix-valued function
e
Af = 2 AT/z*!
= i 0
fact e At
is bounded as a function of t if and only if Re A 0 V /, and t <
m =
t whenever Re At = 0 (i.e., if all eigenvalues of A have nonpositive
1
real parts, and all eigenvalues of A having zero real parts are simple zeros
of a minimal polynomial of A) e At » 0 as t ;
—
* oo if and only if Re At < 0 —
V / (i.e., all eigenvalues of A have negative real parts).
:
294 Appendix
/
S= A
0
i+1
t‘/i\
is given by
References
295
296 References
[14] Hille, E., and Phillips, R. S. Functional Analysis and Semigroups, Amer.
Math. Soc., Providence, R.I., 1957.
[15] Jury, E. I. Inners and Stability of Dynamic Systems, John Wiley & Sons,
New York, 1975.
[16] Lefschetz, S. Stability of Nonlinear Control Systems, Academic Press,
New York, 1962.
[17] Narendra, K. S. and Taylor, J. H. Frequency Domain Criteria for Absolute
Stability, Academic Press, New York, 1973.
[18] Nemytskii, V. V. and Stepanov, V. V. Qualitative Theory of Differential
Equations, Princeton Univ. Press, Princeton, 1960.
[19] Ralston, A. A First Course in Numerical Analysis, McGraw-Hill, New York,
1965.
[20] Roy den, New York, 1963.
H. L. Real Analysis, Macmillan,
[21] Sandberg, W. “On the L 2 -Boundedness of Solutions of Nonlinear Func-
I.
tional Equations,” Bell Sys. Tech J., vol. 43, pp. 1581-1599, 1964.
.
erned by Nonlinear Functional Equations,” Bell Sys. Tech. J ., vol. 44, pp.
871-898, 1965.
[24] Taussky, O. “A Generalization of A Theorem of Liapunov,” Siam J. Appl.
Math., vol, 9, pp. 640-643, 1961.
[25] Wilkinson, J. H. The Algebraic Eigenvalue Problem, Clarendon Press, Ox-
ford, 1965.
[26] Willems, J. C. The Analysis of Feedback Systems, MIT Press, Cambridge,
Mass., 1970.
[27] Zames, G. “On the Input-Output Stability of Time-Varying Nonlinear
Feedback Systems, Part I: Conditions Derived Using Concepts of Loop
Gain, Conicity and Positivity,” IEEE Trans. Auto. Control, vol. AC-11,
pp. 228-238, 1966.
[28] Zames, G. “On the Input-Output Stability of Time-Varying Nonlinear
Feedback Systems, Part II: Conditions Involving Circles in the Frequency
Plane and Sector Nonlinearities,” IEEE Trans. Auto. Control, vol. AC-11,
pp. 465-476, 1966.
Index
Index
A
a 246
, C*, 52
a ,249 Cn x », 64
tte, 250 C[a, b], 59
Absolute stability, 199 C"[a,
, b], 59
Adams-Bashforth algorithms, 121 Cauchy sequence, 55
Adams-Moulton algorithms, 121 Causality, 229
Aizerman’s conjecture Center, 19, 20
in input-output setting, 290 Cetaev’s theorem, 163
(see also Circle criterion, input- Circle criterion, input-output, 278
output, necessity of) necessity of, 282
Asymptotic stability, 135 Circle criterion, Liapunov, 207
theorems, 153, 158 Complete inner product space (see
Asymptotic stability in the large (see Hilbert space)
Global asymptotic stability) Complete normed space (see Banach
Autonomous system, 5 space)
Continuity (in a normed space), 55
uniform, 55
B Continuous dependence on initial con-
ditions
BCn [ta ,
oc
), 133 example of lack of, 4
Banach algebra, 247 theorem, 86
Banach space, 56 Contraction mapping theorem
Bellman-Gronwall inequality, 292 global, 73
Bendixson’s theorem, 31 local, 76, 78
299
300 Index
Index 301
L Lipschitz constant, 80
Loop transformations, 204, 216
Lur’e problem, 199
Lp , 226 Lur’e-Postnikov criterion, 211
extended, 228
1$ 229
Instability, 233 M
Liapunov function, 149
common, 202 Mathieu equation, 151, 159, 208
existence of quadratic, 175, 181 Measure of a matrix, 67
Lur’e-Postnikov type, 210 formulae, 71
Popov type, 213 properties, 67
Liapunov function candidate, 149 Memoryless nonlinearity, 97
Liapunov matrix equation, 172 odd, 101
closed-form solution, 174 Minkowski’s inequality, 227
stability conditions based on, 175
Multistep algorithms, 118
uniqueness of solutions, 173 exactness constraints, 119
Liapunov’s direct method, 147
Liapunov’s indirect method, 186
Limit cycle (see also Periodic solution), 31 N
analytical method for, 38
302 Index
State-plane trajectory, 11
State variable, 2
R State vector, 2
Stationary point (see Equilibrium point)
R n 52
,
Subspace, 53
R nxn , 64
RC transmission line, 271
Rayleigh’s equation, 29, 48
Region of attraction, 135 T
Return difference, 264
Return difference matrix, 274 Taylor series method, 113
Runge-Kutta method Topological equivalence, 58
fourth order, 116 Truncation, 228
second order, 115
S U