6 Stability Analysis
6 Stability Analysis
6 Stability Analysis
Stability Analysis
Objectives: To understand the various notions of stability and Lyapunov based analysis
in establishing them.
Lesson objectives
This module helps the reader in
understanding the various notions of nonlinear stability such as stability, asymptotic
stability, global asymptotic stability and exponential statbility
the use of Lyapunov stability theory in proving stability of equilibrium points
estimating the region of attraction of an equilibrium point
the use of La Salles invariance principle to infer asymptotic stability
establishing the stability of linear systems by computing the eigen-spectrum of the
linearized system matrix
establishing the stability of linear systems by solving the Lyapunov equation
instability result.
96
Suggested reading
Nonlinear System Analysis: M. Vidyasagar
Nonlinear Systems: H. K. Khalil
Linear Systems Theory: Joao P. Hespanha
97
Lecture-28
Stability definitions
Consider a system described by
x = f (x)
(28.1)
that contains the origin x = 0. Let xe D be an equilibrium point, that is, f (xe ) = 0.
Without loss of generality, we assume xe = 0, the origin. We next state the various
notions of stability for the system (28.1).
Definition 28.0.22 The equilibrium point x = 0 of (28.1) is stable if, for each > 0,
there is a = () > 0 such that
||x(0)|| < = ||x(t)|| < , t 0.
(28.2)
The notion of stability is depicted in Figure 28.1, where a trajectory originating from
-ball with initial condition x(t0 ) is confined to -ball for all t 0. Note that, there is no
requirement that trajectories approach the origin as time evolves.
Example 28.0.23 Let us revisit the linear oscillator system
x 1 = x2
x 2 = x1
(28.3)
and we know that the origin is the unique equilibrium point and further the trajectories
of the system (28.3) describe concentric circles centered at the origin. Thus, give any
> 0, there exists a > 0 such the condition (29.1) holds. In fact, any satisfying
meets the requirement. Hence, the origin is a stable equilibrium point.
98
x2
x(t0)
x1
x(t0)
x2
x(t)
xe
x1
1
0
1
2
3
4
5
5
0
x
x 1 =
x 2
99
(28.4)
and globally asymptotically stable if lim x(t) = 0 is satisfied for any x(0).
t
The notion of asymptotic stability is depicted in Figure 28.3, where the trajectories eventually approach the origin.
x2
x(t0)
x1
x(t0)
xe
x2
x(t)
x1
100
(28.5)
Lecture-29
We next present Lyapunovs stability theorem concerning the stability of the equilibrium
point x = 0 of (28.1).
Theorem 29.0.28 Let V : D IR be a continuously differentiable function such that
V is positive-definite on D and V 0 in D. Then x = 0 is stable. Further if V < 0 in
We have to establish a -Ball such that for all initial conditions in it, the corresponding
trajectories of (28.1) are bounded by for all t 0. Choose r (0, ] such that Br =
{x IRn : ||x|| r} D. Let = min V (x). Since V (x) > 0 for x 6= 0, > 0. Take
||x||=r
Then V (p) > , which is not true. Hence, the claim. The set along with various
other sets are depicted in Figure 29.1.
Since V 0 on D, the following holds V (x(t)) V (x(0)) , which further implies
that if x(0) then x(t) for all t 0. This establishes the positive invariance of
follows from the fact that a locally Lipschitz function on a domain D is Lipschitz on a
compact subset of D.
The continuity of V implies that there exists a > 0 such that ||x|| < = ||V (x)|| =
x(t) Br for all t 0. Therefore, ||x(0)|| < = ||x(t)|| < r for all t 0, thus
establishing the stability of x = 0.
101
D
Br
0.1
0.08
0.06
0.04
0.02
0
3
2
0.2
0.1
0
0
1
0.1
2
x2
0.2
x1
equivalently, for every a > 0, there is a T > 0 such that ||x(t)|| < a for all t > T . By the
same argument, for every a > 0, we can choose a b > 0 such that b Ba . Therefor it
102
for all t 0. Let = max V (x). The existence of is ensured because of the fact that
d||x||r
a continuous function (V ) on a compact set ({x : d ||x|| r}) achieves its maximum and
Rt
minimum. Since V < 0, < 0. Since V (x(t)) = V (x(0))+ V (x( ))d V (x(0))t,
0
V (x(t)) will eventually become negative and hence contradicts V (x(t)) c 0. Thus
c = 0.
Remark 29.0.29 Stability of the equilibrium is also referred to as stability in the sense
of Lyapunov or simply Lyapunov stable.
103
Lecture-30
The use of Lyapunovs theorem 29.0.28 will be shown through the following examples.
The theorem provides only sufficient conditions under which stability holds. It means that
if the conditions are not met, then one cannot infer anything about the stability result and
further investigation is required to arrive at a Lyapunov function. Even in the plane, the
search for Lyapunov functions is nontrivial and several methods like the variable gradient,
have been proposed to provide a way of arriving at a Lyapunov function. If the energy
of the system is known, then it serves as the candidate Lyapunov function as can be seen
through the simple pendulum example.
Example 30.0.30 Consider the simple pendulum system
x 1 = x2
x 2 = a sin x1
(30.1)
Consider the candidate Lyapunov function V = a(1 cos x1 ) + x22 /2. Then V = 0.
Therefore the lower equilibrium point (0, 0) is Lyapunov stable. If friction is considered
in the model, then the equations of motion are
x 1 = x2
x 2 = a sin x1 bx2
(30.2)
where, b = kl and k being the friction co-efficient. If we use the same V , then V = bx22
0, that is V is negative semi-definite. Therefore, we cannot infer about the asymptotic
stability of the origin, which the system possesses from the physics of the system or from
the phase-portrait. Hence, we search for candidate Lyapunov function such that V is
negative-definite. Consider V (x) = 12 x P x + a(1 cos x1 ), where P = P > 0 should be
selected in a manner that V < 0. The derivative of V along the trajectories of (30.2) is
given by
V
= a(1 p22 )x2 sin x1 ap12 x1 sin x1 + (p11 p12 b)x1 x2 + (p12 p22 b)x22 .
104
To eliminate the sign indefinite terms, we set p22 = 1 and p11 = p12 b. To retain the
(30.3)
Clearly (0, 0) is the unique equilibrium point. Consider the candidate Lyapunov function
x2
x4
V = 1 + 2 . Then V = x4 0. Therefore V is a Lyapunov function and the origin is
4
Lyapunov stable.
Example 30.0.32
x 1 = x2
x 2 = h1 (x1 ) x2 h2 (x3 )
(30.4)
x 3 = x2 x3 .
where, h1 and h2 are locally Lipschitz functions that satisfy hi (0) = 0 and zhi (z) > 0,
i=1,2, for all z 6= 0. Clearly (0, 0, 0) is the unique equilibrium point. Consider the
Rx
Rx
x2
candidate Lyapunov function V = 0 1 h1 (z)dz + 0 3 h2 (z)dz + 22 . Then V = h1 x 1 +
105
Lecture-31
Variable gradient method
and fix its components to
In this method, we assume a structure for the gradient V
x
render V < 0. Let V be the candidate Lyapunov function whose gradient we assumed to
have the form
V
= g(x)
x
where, g(x) = [g1 (x) g2 (x) . . . gn (x)] . Poincare Lemma gives us the condition under
which a given vector is a gradient of some scalar function.
Lemma 31.0.33 Given a smooth vector field v(x), v : IRn IRn , there exists a smooth
The construct g(x) is such that V is negative definite on a domain D. Finally, the
Lyapunov function V is extracted from the line integral
Z x
Z xX
n
V (x) =
g (s)ds =
gi (s)dsi .
0
(31.1)
i=1
Using the fact that the line integral of a gradient vector g : IRn IRn is path inde-
pendent, the line integration in (31.1) can be taken along any path joining the origin to
x. Choosing the path made up of line segments parallel to the coordinates axes, (31.1)
becomes
V (x) =
x1
x2
g1 (s1 , 0, 0, . . . , 0)ds1 +
g2 (x1 , s2 , 0, . . . , 0)ds2
0
Z xn
+...+
gn (x1 , x2 , x3 , . . . , sn )dsn .
0
Example 31.0.34
x 1 = x2
x 2 = h(x1 ) ax2 , a > 0
(31.2)
where, h( ) is locally Lipschitz, h(0) = 0 and yh(y) > 0 for all y (b, c) for some
b, c > 0. Assume
g(x) =
"
x1 (x) + x2 (x)
x1 (x) + x2 (x)
(x)
(x)
(x)
(x)
+ x1
= (x) + x1
+ x2
x2
x2
x1
x1
(31.3)
(31.5)
=
h(y1 )dy1 + y12 + x1 x2 + y22
2
2
0
"
#
Z x1
a
1
=
h(y1 )dy1 + x
x.
2
0
| {z }
>0
107
108
Lecture-32
Global asymptotic stability
The existence of a Lyapunov function V (x) which proves the origin to be asymptotically
stable does not guarantee that all trajectories originating from D will converge to the
origin. For the origin to be globally asymptotically stable the Lyapunov function should
be radially unbounded.The Lyapunov function V (x) is said to be radially unbounded if
V (x) as x , uniformly in x.
for any c > 0 (assured by radial unboundedness) and if D = Rn then the origin x = 0 is
globally asymptotically stable. Note that the notion of GAS is applicable only to systems
with isolated unique equilibrium point.
Region/Domain of attraction
Definition 32.0.35 Let (t, x0 ) be the flow associated with a system.Then the region/domain
of attraction RA is defined as
RA = {x : (t, x) is defined for all t > 0 and (t, x) = 0 as t }
Lemma 32.0.36 The set RA is open, connected and positive invariant. Moreover, the
boundary of RA is formed by trajectories.
Example 32.0.37 Consider the system
x 1 = x2
x 2 = 4(x1 + x2 ) h(x1 + x2 )
109
(32.1)
where h(0) = 0 and h(p)p 0 |p| 1. The origin (0, 0) is the unique equilibrium point
asymptotically stable. From the phase portrait it can be seen that any trajectory within
the strip |(x1 + x2 )| 6 1 converges to the origin.
Finding the exact region of attraction is an onerous task even for second-order systems.
However, an estimate of the region of attraction still provides guaranteed stability. One
way of obtaining the estimate is to use the closed and bounded level sets of the Lyapunov
function itself. A closed form solution for the estimates can be found when the Lyapunov
function is quadratic in x.
110
Lecture-33
Estimate of RA for quadratic Lyapunov functions
Let the quadratic Lyapunov function for a system be given as V = x P x; P = P > 0.
Further, it is assumed that
V
f (x)
x
min x P x = min (P )r 2
kxk=r
min x P x
|b x|6r
which can be solved using the necessary conditions for optimality. Setting the cost function
as L = x P x + {(b x)2 r 2 } where IR is the Lagrange multiplier. The necessary
conditions for optimality are
L
x
= 0 and
= 0. Now,
L
= 2P x + 2(b x)b = 0
x
P x = (b x)b
(33.1)
L
= (b x)2 r 2
b x = r.
111
(33.2)
Hence J = x P x =
1
rP 1b
; = 1 .
1
b P b
b P b
|b
i x|6ri
ri2
1 b .
b
i
i P
(33.3)
We need to first show that the origin is asymptotically stable. Taking V (x) = 1 2 2 results
in V (x) = (x21 + x22 )(1 (x21 + x22 )). Thus V < 0 in D = {(x1 , x2 ) : x21 + x22 < 1}. Hence
(0, 0) is asymptotically stable and an estimate of the domain of attraction is given by
(33.4)
It is required to show that the origin is asymptotically stable and estimate its region of
attraction.
"
#The asymptotic stability of the origin is established by considering V (x) =
5 1
x
x and noting that V < 0 in D = {x IR2 : |x2 | 1}. The estimate of the
1 2
domain of attraction is given by c = {x IR2 : V (x) c}, where
1
c<
[0 1]
"
5 1
1 2
#1
9
5
[0 1]
.
Example 33.0.40 Consider a second-order system x = f (x) with asymptotically stable
origin. Let V (x) = x21 + x22 , and D = {x IR2 : |x2 | < 1, |x1 x2 | < 1}. Suppose that
V
f (x)
x
112
c1 =
[1 1]
"
#1
1 0
0 1
=
[1 1]
12
c2 =
[10 1]
"
1 0
0 1
1
2
#1
1
2
= 1.
[0 1]
x2
x1
113
Lecture-34
Example 34.0.41 Consider the system
x 1 = x1 x31 + x2
x 2 = 3x1 x2
(34.1)
Next, we perform a coordinate change so that the equilibrium point (2, 6) is the origin
of the new coordinate system and the objective is establish the stability using a quadratic
114
(34.2)
z2 = 3z1 z2
(34.3)
To eliminate the sign-indefinite terms, we set p12 = 0. Then (34.3) can be expressed as
V
= z Q(z1 )z
(34.4)
where,
Q(z1 ) =
"
(p11 +3p22 )
2
(p11 +3p22 )
2
p22
It is easy to see that for p11 , p22 > 0, the matrix Q(z1 ) is locally positive-definite around
z1 = 0. Fix, p11 = 3, p22 = 1 (see subsection 34.0.41 ). Then Q(z1 ) > 0 in the set
D1 = {(z1 , z2 ) : z1 (2, )}. Thus V < 0 in D1 and it follows that (z1 , z2 ) = (0, 0) is
asymptotically stable.
To compute the region of attraction RA and compare it with the estimate of the domain
of attraction c , we plot the vector field plot of (34.2) as shown in Figure 34.1. A local
analysis about the saddle point shows that the stable eigenvectors satisfy z2 z1 = 0
and the unstable eigenvectors satisfy 3z1 + z2 = 0. The region of attraction is given by
RA = {(z1 , z2 ) IR2 : 3z1 + z2 > 0}. The estimate of the domain of attraction of the
origin (z1 = 0, z2 = 0) is given by c = {(z1 , z2 ) (3z1 +z2 > 0)D1 : 3z12 +z22 c, c > 0}
is compact.
115
15
10
x2
10
15
10
0
x1
10
+ 6z1 + 11)
116
p11 + 3p22
2
2
(34.5)
The function h(z1 ) has minima at z1 = 3 for all values of p11 and p22 . So by substituting
z1 = 3 in (34.5)
p11 + 3p22
2
2
In order to render Q(z1 ) > 0, the det(Q(z1 )) > 0, this implies h(3) > 0, hence
p211 9p222 + 2p11 p22 > 0.
Let us set the minimum value of h(z1 ) = 0, thus
p211 9p222 + 2p11 p22 = 0
2
p11
p11
2
+ 9 = 0.
p22
p22
Substituting
p11
p22
(34.6)
of p11 and p22 the function h(z1 ) can achieve a positive minima. Further, by putting
z1 = 2 in (34.5) and solving we get
h(2) = p11 p22 (4 12 + 11)
= p211 9p222 + 6p11 p22
p11 + 3p22
2
2
6
+ 9 = 0.
p22
p22
Again, substituting
p11
p22
(34.7)
positive and equal. Thus for z1 6 2, h(z1 ) cannot have positive minima for all positive
real values of p11 and p22 . Hence we have
117
p11
p22
= y = 3 p22 = 1, p11 = 3.
Lecture-35
La Salles invariance principle
In many problems, it is easy to establish the negative semi-definiteness of V rather than the
negative definite property. In such a case, Lyapunovs theorem only establishes stability
and nothing can be said about asymptotic stability. La Salles invariance principle aids
in establishing asymptotic stability property of the origin even when V is negative semidefinite.
Theorem 35.0.42 Consider the system x = f (x) and let be a compact set that is
invariant with respect to this system. Let V : D R be continuously differentiable such
that V 0 in . Let E be the set of points in such that V = 0. Further let M be the
largest invariant set in E. Then every solution starting in converges to M as t .
Note that in theorem 35.0.42, the only requirement on V is that it be continuously differentiable and V 0 on a compact set . Since we deal with positive-definite V , the choice
= c ensures the compactness of , at least for sufficiently small c > 0. The extension
118
x1 = x2
(35.1)
x2 = a sin x1 bx2
x2
(35.2)
x 2 = x31 x32
Taking V (x) =
x41
x2
+ 22 ,
4
x 1 = x2
(35.3)
x 2 = x1 x2 x2 |x2 |
Taking V (x) =
x21 +x22
2
Example 35.0.47
x 1 = x2
(35.4)
x 2 = h1 (x1 ) x2 h2 (x3 )
x 3 = x2 x3
Rx1
h1 (z)dz +
Rx2
0
x(t) S x2 0 x2 0 x3 0 x3 0
Hence, the origin (0, 0, 0) is asymptotically stable.
119
Lecture-36
Stability of Linear systems
Consider the linear system given by x = Ax and a corresponding V (x) = x P x where
P = P > 0. Then,
V
= x P x + x P x
= x A P x + x P Ax
= x (A P + P A)x
x =
"
0 1
1 1
120
(36.1)
Choosing P =
"
P11 P12
P12 P22
and Q =
"
1 0
0 1
0
2 0
P11
1
1 1 1 P12 = 0
P22
1
0
2 2
P =
A=
"
"
3
2
12
0 1
1 1
12
1
#
P11
3
2
= 1
P
12
2
1
P22
is symmetric, positive definite and unique. From Theorem 36.0.49
x IRn ,
(36.2)
Let x = 0 be the equilibrium point i.e f (0) = 0. Further assume f : D Rn is continuously differentiable on a domain D in the neighbourhood of the origin. Let A =
f
| .
x x=0
If the linearized system (36.2) is exponentially stable, then there exists a ball B D and
constants c, > 0 such that for every solution x(t) to the nonlinear system (36.2) that
starts at x(0) B, we have
k x(t) k ce(t) k x(0) k, t 0.
(36.3)
V
f (x)
x
e(t) V (x(0)), t 0.
Joint Initiative of IITs and IISc
Funded by MHRD
x B.
(36.4)
= f (x) P x + x P f (x)
= Ax + r(x) P x + x P Ax + r(x)
= x (A P + P A)x + 2xP r(x)
(36.5)
= k x k2 +2x P r(x)
k x k2 +2 k P k k x k k r(x) k .
Let > 0 be sufficiently small so that the ellipsoid centered at x = 0 satisfies the following:
Thus, for x E,
E , {x IRn : x P x } B.
1 2c k P k k x k
V
1
1
= k x k
.
2
4ckP k
k x k2 +2c k P k k x k3
= 1 2c k P k k x k k x k2
= 12 k x k2
(36.6)
< 0, x 6= 0.
For this choice of , the set E is positively invariant and the origin is asymptotically stable.
Further, from (36.6) and the fact that x P x k P k k x k2 , it follows that if x(0) starts
Joint Initiative of IITs and IISc
Funded by MHRD
122
E (inside B)
B (inside E)
xe
1
4 c kP k
inside E,
V
2 kP k
123
Lecture-37
Instability
An equilibrium is unstable if it is not stable. There are several results establishing instability, but we shall use the result by Chetayev, which is stated as follows.
Theorem 37.0.53 Let x = 0 be an equilibrium point of x = f (x). Let V : D IR be a
continuously differentiable function such that V (0) = 0 and V (x0 ) > 0 for some x0 with
arbitrarily small ||x0 ||. Define a set U = {x Br : V (x) > 0}, where Br = {x D :
||x|| r} and suppose that V (x) > 0 in U. Then x = 0 is unstable.
Example 37.0.54 Consider the following nonlinear system
x 1 = x31 + x2
(37.1)
x 2 = x61 x32
Show the region = {0 x1 1} {x2 x31 } {x2 x21 } is a positively invariant set
Note the system has two equilibrium points at (0, 0) and (1, 1). The region is bounded
by the curves (as shown in Figure 37.1) 1 = x2 x21 and 2 = x2 x31 . The derivative of
1 along the trajectorioes of (37.1) is,
124
Hence, the direction of the vector field f on the curve 1 points into . Similarly, the
derivative of 2 along trajectories of (37.1) is,
2 = x61 x32 3x21 (x31 + x2 )
Hence, the direction of the vector field f on the curve 2 also points into . Thus all the
trajectories eminating from a neighbourhood of are trapped in it, that is, is positively
invariant set.
Next, consider a ball centered at origin defined by B2/3 = {(x1 , x2 ) R2 : ||x||2 32 }.
differentiable function. V (x) = 0 on the boundary of U and since the origin also belongs
to the boundary of U, V (0) = 0. There exists x0 U, arbitrarily close to the origin,
such that V (x0 ) > 0 (for instance, consider x0 as (0.2,0.02), then V (x0 ) = 0.032). The
x2 = x31
x2 = x21
x2
(1,1)
B2/3
U
2/3
(0,0)
125
x1
for all x U.
All the hypothesis of Chetayevs theorem hold, thereby proving the origin to be unstable.
126
Exercise problems
1. Use a quadratic Lyapunov function candidate to show that the origin is asymptotically stable:
a.
x 1 = x1 + x1 x2
x 2 = x2
b.
x 1 = x2 (1 x21 )
(37.2)
J3 3 = (J1 J2 )1 2 + u3
where 1 , 2, 3 are the components of the angular velocity vector along the
principal axes, u1 , u2, u3 are the torque inputs applied about the principal axes, and
J1 > J2 > J3 > 0 are the principal moments of inertia. The rotational kinetic
energy of the spacecraft is given by T () = 21 (J1 12 + J2 22 + J3 32 ) while the total
p
angular momentum of the body is given by P () = J12 12 + J22 22 + J32 32 .
a. Find all the equilibria of the rotational dynamics.
127
b. Compute T and P along the trajectories of (37.2). What can you conclude
about stability?
c. Show that with ui = 0, i = 1, 2, 3, the origin = 0 is stable. Is it asymptotically
stable?
d. Use the Lyapunov function V () = (T () T (e))2 + [(P ())2 (P (e ))2 ]2 to
show that every equilibrium of the form e = [ 0 0] is Lyapunov stable.
x 1 =
x 2 =
x 3 =
1
(J2
J1
1
(J3
J2
1
(J1
J3
J3 )(x2 x3 + x3 )
J1 )(x1 x3 )
(37.3)
J2 )(x1 x2 + x1 )
V (x) = 0 and the sphere ||x|| = r. Since V (0) = 0, the origin lies on the
boundary of D1 . Further, in D1 , both V and V > 0. By Chetayevs theorem,
the origin is unstable.
g. Suppose that the torque inputs apply the feedback control ui = k1 i , where
gradient method, find a Lyapunov function that shows that the origin is globally
asymptotically stable.
Joint Initiative of IITs and IISc
Funded by MHRD
128
P A + A P + Q P BR1 B P = 0
where, R = R > 0, and Q = Q 0. Using V (x) = x P x as a Lyapunov function
candidate, show that the origin is globally asymptotically stable when
(1) Q > 0.
(2) Q = C C and the pair (A, C) is observable.
6. Consider the tunnel diode circuit as shown in figure 1. The tunnel diode has a
nonlinear voltage-controlled constitutive relationship that contains the region of
negative resistance. This characteristic made it possible to construct bistable circuits
which were used in switching or memory elements in early computers. The circuit
equations are given by
LI = E RI V
C V
).
= I I(V
C
L
))2 + [ 1 RI 2 EI
(E RI V )2 + (I I(V
2
2
2
Z V
+IV
Ig (Vg )dVg ]
0
where IR is free, find conditions on for asymptotic stability. Also find the
condition on the steepness of slope of the negative resistance region of the tunnel
diode.
7. Consider the following system
x 1 = (x2 1)x31
x41
x2
x 2 =
2 2
(1 + x1 )
(1 + x22 )
(a) Show that the origin is asymptotically stable. (b) Is V radially unbounded ?
Joint Initiative of IITs and IISc
Funded by MHRD
129
"
is asymptoti1 4
cally stable using the Lyapunov equation A P + P A = Q, where Q > 0.
10. The dynamics of a Spring-mass system with viscous and Coulomb friction and zero
external force is given by
M y = Mg ky c1 y c2 y|
y|
where, M is the mass, k is the spring constant, c1 , c2 > 0 are friction coefficients and
y is the displacement of the mass from a reference position. Show that the origin is
globally asymptotically stable.
130