Nonlinear Analysis Notes

ECE 528: Analysis of Nonlinear Systems1

Lecture 1

1.1

Topological Concepts in Rn

Definition 1.1. Given a point x Rn , its  neighborhood for  > 0 is the set {y Rn : |xy| < }.
Definition 1.2. A set D Rn is open if D contains some  neighborhood for every x D.
Definition 1.3. A set D Rn is closed if for every sequence of points {xn } D which converges
to a point x Rn , x D already.
Definition 1.4. The boundary D of D Rn is the set of points x Rn such that every neighborhood of x contains points both of D and Dc (i.e. if B(r, x) D 6= and B(r, x) Dc 6=

for every r, then x D). The interior of D, D:= D \ D.

If D is open then D D = . If D is closed then D D.
Definition 1.5. A set D is convex if for every x, y D and for all [0, 1], the point x + (1
)y D. D is connected if for any two points x, y D, there is an arc lying entirely in D joining
x and y.
[
Definition 1.6. A cover O of a set D is a set {U } such that
U D.

Definition 1.7. D is compact if any (and therefore in Rn all) of the following are satisfied:
1. for every cover O there is a finite subcover Of O.
2. D is closed and bounded.
3. Every sequence has a convergent subsequence with limit in D. In other words, D is sequentially
compact.
Note: in Euclidean space Rn , the conditions in this definition are all equivalent. In general,
2 6 1 (for example, unit sphere in -dimensional space is not compact).
Definition 1.8. A function f : Rn Rn is continuous at a point x Rn if f (xk ) f (x) whenever
xk x. More precisely, for every  > 0 there is a > 0 such that |y x| < |f (x) f (y)| < .
Important: x is fixed and y varies. If we allow both to vary, we have uniform continuity, i.e.
|x y| < |f (x) f (y)| < , where x, y D.
We say that f is continuous on D dom(f ) if f is continuous at every point in D.
1

Typeset by James Schmidt. Disclaimer : these notes are not edited.

Nonlinear Analysis Notes

Also, f is continuous on D if for every open set U in the range of f , the preimage f 1 (U ) is
open in the domain of f .
Note that for general continuity, = (, x) whereas for uniform continuity, = (), independent of x. In general, continuity does not imply uniform continuity but on a compact domain,
continuity does imply uniform continuity.
Proof. Recall that if f is continuous and D is compact, then f (D) is compact too. Therefore, the
image of f is compact. Let  > 0 be given and cover Im(f ) with B(, y) as y ranges over the image
of f . Since f (dom(f )) is compact, there is a finite subcover. Since f is continuous, the pre-image
of each of these open  balls is open. Take to be such that a ball is contained in each one.
Theorem 1.1 (Weierstrass). A continuous function over a compact set attains its extreme points.
In other words, there is an x, y dom(f ) such that f (x) = sup({f (y) : y dom(f )}).
Definition 1.9. A function f is injective on D if f (x) = f (y) x = y for every x, y D; in
other words, f 1 is well defined on f (D).
Definition 1.10. A function f : [a, b] R R is piecewise continuous if it is continuous everywhere except for a countable number of points in its domain, and has left and right limits f (x )
and f (x+ ) everywhere.
Definition 1.11. A function f : Rn Rm is differentiable if its partial derivatives
is continuously differentiable if

Example: f (x) =

f1
xj

fi
xj

exists, and

are continuous.

x2 sin(1/x) ifx 6= 0
0
else

C n (D1 , D2 ) is defined as the space of N -times continuously differentiable functions from D1

D2 ; e.g. C 0 are continuous functions, etc.

f
x1

For f : Rn R, its gradient vector f (x) = ... (also sometimes denoted

f : Rn Rm , the Jacobian matrix is again denoted

f
x

f
xn

or J(f ) with J(f )ij =

f
x ).

If

fi
xj .

Theorem 1.2 (Mean Value Theorem (for scalar valued functions)). Suppose a function f : Rn R
is C 1 . Then for every x, y Rn , there is a point z x + (1 )y for some [0, 1], such that
f (y) f (x) = f (z) (y x).
In general for a vector valued function f : Rn Rm , the Mean Value Theorem must be
applied component wise, i.e. fi (y) fi (x) = fi (zi ) (y x), for i = 1, . . . , m.
Theorem 1.3 (Bellman-Gronwall Lemma). Let : [a, b] R be continuous, : [a, b] R continZ t
uous and nonnegative, and y : [a, b] R is continuous. Suppose that y(t) (t) +
(s)y(s)ds,
a
Z t
Rt
then y(t) = (t) +
y(s)(s)e s ( )d ds.
a

Nonlinear Analysis Notes

Z t
Special case: when c R, then y(t) exp( ( )d ). If also d R, then y(t)
a

e(ta) . Note that in the hypothesis, y appears on both sides, so even though the conclusion of the
theorem is more complicated, the result is that y is bounded by an explicit function independent
of y.
Applications: take y = y, y(t0 ) = y0 . This is equivalent to the integral equation
Z

y(s)ds

y(t) = y0 +

(1.1)

t0

and this equation gives the implicit relation of the sort in the statement of the theorem. Of course,
the solution is y(t) = y0 e(tt0 ) which is exactly the form of the inequality relation when and
are constant; the B-G lemma is more general in that it gives the inequality version of this.
Z

Proof for c. Define v(t) := y(t) exp(

( )d ). By hypothesis, v(t) q(t) := ( +
a
Z t
Z t
Z t
(s)y(s)ds) exp(
( )d ). We differentiate, to get q(t)

= (t)y(t) exp(
( )d )
a
a Z
a
t
( )d ) q(t)) = (t)(v(t) q(t)) 0 because v q and 0.
(t)q(t) = (t)(y(t) exp(
a

So q(t) q(a) = , which implies that v(t) as well. Now just multiply through by
Z t
exp( ( )d ) to get the result.
a

Definition 1.12. A sequence {xk } X- (X, || ||) a normed vector space- is called Cauchy if
||xn xm || 0 as n, m 0. More precisely, for every  > 0, there is an N N such that
n, m > N ||xn xm || < .
If {xk } is convergent (i.e. xk x), then it is Cauchy. Proof is obvious: |xn xm | = |xn x +
x xn | |xn x| + |x xm | < /2 + /2 for appropriately chosen N and n, m > N .
The converse is not always true, for example in noncomplete spaces.
Definition 1.13. A normed vector space which is complete is called Banach, i.e. if every Cauchy
sequence is convergent (has its limit in the space).
Examples of Banach spaces:
1. (Rn , || ||p ) for 1 p .
2. (C([a, b], Rn ), || || .

Nonlinear Analysis Notes

Lecture 2

2.1

Contraction Mapping

Let (X, || ||) be a Banach space; a map P : X X is called a contraction [mapping] if there is a
< 1 such that ||P (x) P (y)|| ||x y|| for all x, y X. Comments: 0 is automatic because
im(|| ||) 0. Also, P is automatically uniformly continuous on X (i.e. for every  > 0 theres a
> 0 such that ||x y|| < ||P (x) P (y)|| < , for all x, y X); for contraction, let := /.
Theorem 2.1 (Contraction Mapping Theorem). Let S X be closed and let P : S S be a
contraction. Then
1. There exists a unique fixed point x S of P , i.e. a point x such that P (x ) = x .
2. x is the limit of successive approximations of x0 , x1 := P (x0 ), . . . , xn := P n (x0 ), . . . for all
x0 S.
Proof. Step 1: want to show that the limit of P k (x0 ) exists for every x0 . Step 2: show that the
limit is a fixed point. Step 3: show that the fixed point found in step 2 is unique.
Step 1: It suffices to show that the sequence {P k (x0 )} is Cauchy for any x0 S, because
the existence of limit follows [by definition] from the completeness of X. Let  > 0 be given, then
||P k (x) P k+1 (x)|| = ||P k (x) P k (P (x))|| k ||x P (x)|| = k C, where C := ||x P (x)||. Then
k 0 as k 0; let N be such that k < /C for any k > N . To show Cauchy, in general,
need to show ||P k (x0 ) P k+r (x0 )|| = ||P k (x0 ) P k (P r (x0 ))|| k ||x0 P r (x0 )||, and the same
argument as when r = 1 works; alternatively, apply triangle inequality to ||P k (x) P k+1 (x)|| =
||P k (x) P k+1 (x) + P k+1 (x) . . . + P k+r1 (x) + P k+r (x)|| to get (k+r1 + . . . + k )||P (x0 ) x0 || =

X
k
k ||P (x0 ) x0 ||(
k ) =
||P (x0 ) x0 || 0 as k .
1
k=0

Note that x S since S is closed.

Step 2: ||P (x ) x || = 0 by the previous argument; since P (x ) = P ( lim P k (x0 ) =
k

lim P k+1 (x0 ) = x , where the second equality holds by continuity of P .

Step 3 is obvious: ||P (x) P (y)|| ||x y||; but P (x) = x and P (y) = y implies that
||P (x) P (y)|| = ||x y||.

Nonlinear Analysis Notes

2.2

Fundamental Properties of Dynamical Systems

A dynamical system looks like:

x = f (t, x), x(t0 ) = x0

(2.1)

where x Rn , t [t0 , ), t0 denotes initial time, x0 Rn denotes the initial state, and f :
[t0 , ) Rn Rn .
Special case: the system is time invariant (or autonomous), namely
x = f (x)

(2.2)

d
x stands for dt
x(t) where x(t) is a solution of the system. We need to set out conditions for when
such a solution exists, is well defined, is unique, etc.

2.3

Existence and Uniqueness of Solutions

Want to know: from every chosen initial condition x0 that there exists a unique x(t) satisfying
x(t0 ) = x0 and x(t)

= f (x) satisfies the dynamic equation defining the system. We need to clarify
exactly what this means (e.g. what regularity conditions we impose on the system).
Example 1. Let x = ax where a R, and x R. For initial condition x(t0 ) = x0 , the solution is given by
x(t) = x0 ea(tt0 )

(2.3)

Notice that this solution exists (is defined) for all t > t0 . It decays for a < 0 and grows for
a > 0.
Example 2. Let
x =

(2.4)

with x0 = 0. Or you could take x(t) = t2 /4. The problem: solutions to this equation are

uniquely determined by the initial condition. Whats wrong? f (x) = x has infinite slope
(therefore, nonexistent derivative, i.e. f 0 (0) is not well defined).
Example 3. Let
x = x2

(2.5)

with x(0) = x0 6= 0; can solve this by integrating:

dx
= dt
x2

x0

dx
=
x2

dt
0

1 x
| = t x(t) =
x x0

1
x0

1
t

(2.6)

What does this look like? If x0 > 0, then at time t = 1/x0 , x(t) = ; we call this phenomenon
finite escape time. Thus the solution exists only locally around the initial condition. The
growth here for the solution is faster than the growth of exponential growth (which reflects
possible difference between linear and nonlinear dynamics).

Nonlinear Analysis Notes

2.4

Some Notions

First, consider autonomous system

x = f (x)

(2.7)

Definition 2.1. A function f : Rn Rn is locally Lipschitz if for any compact set D Rn , there
is a constant LD |R satisfying:
||f (x) f (y)|| L||x y|| x, y D

(2.8)

Definition 2.2. A function f : Rn Rn if there exists such an L satisfying the same property
above, except for all x, y Rn (not just compact subsets).
Examples
1. A contraction mapping is trivially Lipschitz.

2. f (x) = x is not locally Lipschitz at 0. This is obvious since f 0 (0) = . On the other
hand, it is at any other point. Note, also, that f is uniformly continuous on R0 . Therefore,
uniform continuity does not imply that the Lipschitz condition is satisfied.
3. If f is C 1 and f 0 is bounded on its domain, then f is globally Lipschizt.
Proof. M V T |f (x) f (y)| |f 0 (z)||x y| for some z.
4. f (x) = x2 is locally Lipschitz. Follows from previous, since f 0 is bounded on every compact
subset. On the other hand, its not globally Lipschitz. The same idea applies; since f 0 (x) = 2x
is unbounded on R. This doesnt prove this, but provides intuitive rationale.
Proposition 2.1. C 1 Local Lipschitz Continuous.
The first implication was already shown and the second is obvious.

Nonlinear Analysis Notes

Lecture 3

Recall C 1 Locally Lipschitzcontinuity.

The reverse implications do not hold. For example,

x is continuous but not Lipschitz.

Theorem 3.1 (Radamachers Theorem). A locally Lipschitz function is differentiable almost everywhere.
By default, we take Lipschitz to mean locally Lipschitz.
Definition 3.1. For x = f (t, x) we say that f is locally Lipschitz in x uniformly over t (for t [a, b]
or [a, )), if for every compact subset D Rn , there is an L > 0 sucxh that |f (t, x) f (t, y)|
L|x y| for every x, y D, and every t [a, b].
Similar definition applies for global Lipschitz.
Theorem 3.2. x = f (t, x), with x(t0 ) = x0 , and x Rn , t [t0 , t1 ] (when t1 = the right
bracket is open). Suppose that f is piecewise continuous in t for each fixed x, and that f is locally
Lipschitz in x uniformly over t. Then, for arbitrary initial x0 , there exists a unique solution
x : [t0 , t0 + ] Rn for some > 0. Moreover, if f is globally Lipschitz, then the solution is
correspondingly global ( = ).
Remark: in general, will depend on the initial condition. Also, we may not always be
interested in uniqueness (existence alone suffices), in which case local Lipschitzness can be replaced

by continuity of f (a much weaker assumption); take for example x = x).

Before the proof, lets clarify what we mean by solution. Our options:
1. x is differentiable everywhere and satisfies the ordinary differential equation x = f (x) for all
t.
2. x is differentiable almost everywhere and satisfies the ordinary differential equation x = f (x)
for all t.
Consider, for example, x = f (t) where f is a square wave. The solution is a sawtooth, whose
derivative exists almost everywhere but not everywhere (it is continuous everywhere, though, and
satisfies the ODE everywhere that x is defined). Nevertheless, we expect our results to apply in
this situation as well.
Instead of the ordinary differential equation x = f (t, x), and x(t0 ) = x0 , we can consider the
Z t
integral equation x(t) = x0 +
f (s, x(s))ds. The upshot is that this equation is that it doesnt
t0

have any derivatives. For the proof well work with it instead. By solution we mean a function x()
which satisfies this integral equation. Since f is piecewise continuous in t, and x is its integral, it
follows that x is differentiable a.e. and satisfies x = f (t, x) a.e. Such functions are called absolutely
continuous.

Nonlinear Analysis Notes

Proof. Strategy: Looking for a map P : X X which is a contraction and such that its fixed point
is a solution to the integral equation, where X = {x : [t0 , t0 +] Rn } = C 0 ([t0 , t0 +], Rn ). We need
Z t
f (s, x(s))ds.
to figure out what P is; Given x(), defined P (x) to be the function (P x)(t) := x0 +
t0

A function x() is a fixed point of P iff x is a solution in the sense which we defined above.
Now the name of the game is simply to apply the contraction mapping theorem; in order to
do so, we need a closed subset S X such that P : S S is a contraction. Given x0 define
S = {x : [t0 , t0 + ] Rn C([t0 , t0 + ], Rn ) : ||x x0 || < r} where r > 0 is fixed and arbitrary,
is to be chosen and || || is the maximum norm on C 0 , namely ||x x0 || = max |x(t) x0 |.
t0 tt0 +

Note that here x0 (t) x0 (where x0 denotes a function on left and a value in Rn on the right). To
enforce that S X, we need to change the function domain in x to [t0 , t0 + ]. Essentially, the rest
of the proof is an exercise in verifying that the conditions of Contraction Mapping are satisfied.
The first thing is that X must be complete. That (C 0 , ||||) with the maximal norm is a Banach
space is a fact from basic analysis. But what we need is for S to be a Banach space. Indeed, this
will follow immediately from the fact that S is a closed subset of X, a Banach space. Need to show
that if xk x is a sequence of function in S which converge to some function x, that x S, i.e.
that ||x x0 || r. But this is obvious.
Z

Now check that P : S S, which is equivalent to that |

f (s, x(s))ds| r for all t
t
0
Z t
Z t
|f (s, x(s)) f (s, x0 ) +
f (s, x(s))ds|
[t0 , t0 + ]. Here is where we have to specify . Well, |
t
t
0
0
Z t
Z t
Z t
f (s, x0 )|ds
|f (s, x(s)) f (s, x0 )ds +
|f (s, x0 )|ds L|x(s) x0 | +
|f (s, x0 )|ds, where
t0

t0

t0

the last
Z tinequality follows immediately from our local uniform Lipschitz assumption. The second
term
|f (s, x) |ds (t t0 ) max |f (s, x0 )| =: h(t t0 ) h. Putting this all together, our
t0

t0 st0 +

first integral is bounded by (Lr + h) which we want to be less than r. So Take to be such that

r
Lr + h

(3.1)

Nonlinear Analysis Notes

Lecture 4
Z

t1

Proof Contd. We have x(t) = x0 +

f (s, x(s))ds =: (P x)(t). Define S := {x : [t0 , t0 + ] Rn :

t0

|x(t) x0 | rt}. We checked that P : S S if

ball h := max |f (s, x0 )|.

r
Lr+h

where LL is a Lipschitz constant on r

t0 st1

What is left to show is that P is a contraction. Consider ||P x P y|| = max ||(P x)(t)
Zt0ttt0 +
Z t
Z t
L|x(s)y(s)|ds
|f (s, x(s))f (s, y(s))|ds
f (s, x(s))f (s, y(s))ds|| =
(P y)(t)|| = ||
t0

L||x y||. We need L < 1, so if <

1
L

t0

t0

then P is a contraction.

r
Combining these results, if < min{t1 t0 , Lr+h
, L1 }, then Contraction Mapping provides the
desired conclusion.

In the global case, we need to use the fact that f is globally Lipschitz; we still have that
r
< min{t1 t0 , Lr+h
, L1 }, but since L is a universal Lipschitz constant, the only thing left to verify
r
is that h doesnt grow too fast. But notice that Lr+h
L1 as r . Pick some < 1/L, at each
r
r
step for a given h we can pick r sufficiently large so that Lr+h
is close enough to L1 so that Lr+h
.
Then we will have a constant which can be used in repeated application in the Local case to give
a global solution.
Remark: global Lipschitz is generally a very strong condition, and sometimes unnecessarily stronger than needed to prove global existence of solutions. In practice, well use Lyapunov
functions.

4.1

Continuous Dependence on Initial Conditions and Parameters

Let x = f (t, x, ) with t [t0 , t1 ], x Rn and Rp a constant parameter, and initial condition
x(t0 ) = x0 .
Theorem 4.1. Suppose that f is continuous in each argument and locally Lipschitz in x uniformly
over t [t0 , t1 ] and over Brp (0 ) some 0 , r > 0 and B p Rp . Suppose that x = f (t, x, 0 ), and
x(t0 ) = x0 has a solution x() defined (and therefore automatically unique) on the whole interval
[t0 , t1 ]. Then for every  > 0 there is a > 0 such that |x0 x0 | < and | 0 | < , then
x = f (t, x, ), x(0 ) = x0 has a unique solution on [t0 , t1 ] satisfying |x(t) x(t)| <  for all
t [t0 , t1 ].
Consider, for elucidation, a counterexample; namely a dynamical system with bifurcation,
where the dynamics of the solution depend on discontinuous function. As a remark, it should be
enough to have piecewise continuity in t.
Before proving the theorem, first a lemma.

Nonlinear Analysis Notes

10

Lemma 4.1. Consider two systems x = f (t, x) with i.c. x(t0 ) = x0 , y = f (t, y) + g(t, y) with
y(t0 ) = y0 ); f satisfies usual hypotheses (those which guarantee existence and uniqueness of
solution) and g satisfies |g(t, y)| , some > 0. Suppose that the corresponding solutions x()
and y() are defined on [t0 , t1 ] and belong to some bounded set W . Then they satisfy the following
bound:

|x(t) y(t)| |x0 y0 | exp(L(t t0 )) + (exp(L(t t0 )) 1)

(4.1)
L
for all t [t0 , t1 ] where L is the Lipschitz constant on W .
Proof. Rewrite in integral form:
Z

f (s, x(s))ds

(4.2)

f (s, y(s)) + g(s, y(s))ds

(4.3)

|f (s, x(s)) f (s, y(s))| + |g(s, y(s))|ds

(4.4)

x(t) = x0 +
t0

and
Z

y(t) = y0 +
t0

Subtracting, we get
Z

|x(t) y(t)| |x0 y0 | +

t0

where the inequality follows from about a hundred applications of triangle inequality and passage
under the integral. Then we use the bound on g and local Lipschitzness of f with constant L; we
thus obtain
Z t
Z t
L|x(s)y(s)|ds (4.5)
|f (s, x(s))f (s, y(s))|+|g(s, y(s))|ds |x0 y0 |+(tt0 )+
|x0 y0 |+
t0

t0

Now use Bellman-Gronwell Lemma:

Z

|x(t) y(t)| |x0 y0 | + (t t0 ) +

(|x0 y0 | + (s t0 ))|LeL(ts) ds

(4.6)

t0
d L(ts)
The last thing to notice is that LeL(ts) = ds
e
and integrate by parts to obtain the result.
Heavy computation, but nothing extra conceptually.

Proof of Theorem. Strategy: reduce the statement in the theorem to the statement in the previous
lemma; then apply lemma.
Define epsilon tube U [t0 , t1 ] Rn Rn+1 defined by U := {(t, x) : t [t0 , t1 ] |x x(t)|
}, a compact subset of Rn+1 . We want to show that a solution to the system above will be
contained in the epsilon tube, or more precisely, that (t, x(t)) U for all t [t0 , t1 ]. We can write
f (t, x, ) = f (t, x, 0 )+f (t, x, )f (t, x, 0 ); the difference between the last two terms we can treat
as g. So write f (t, x, ) = f (t, x, 0 ) + g(t, x). By continuity (since we assume that f is continuous
in ), for every > 0 there is a > 0 such that |lambda 0 | < and (t, x) U implies that
|f (t, x, ) f (t, x, 0 )| < .
Now apply lemma with here y = x. As long as (t, x(t)) U , the lemma implies that
|x(t) x(t)| |x0 x0 |eL(tt0 ) + L (eL(tt0 ) 1). We want to show that the term in the left hand
side is less than . The time horizon is fixed but we can take |lambda 0 | and |x0 x0 | to be as
small as we like (), and arbitrarily small as well; and the right combination of small and

Nonlinear Analysis Notes

5
5.1

11

Lecture 5
Continuous Dependence on Parameters and Initial Conditions

Proof Contd. x(t) solution of x = f (t, x, 0 ) with initial condition x0 and similarly for x except
that x = f (t, x, 0 )+g(t, x). We arrived at a bound |x(t)x(t)| |x0 x0 |eL(tt0 ) + L (eL(tt0 ) 1).
We pick |x0 x0 | and small enough. Since |t1 t0 | is bounded, we can bound the right hand side
by , no matter what  is.
Remark: this argument only works for finite time interval, unless we have some stability
property as time goes to infinity. We only need |g| in tube.

5.2

Differentiability of Solutions; Sensitivity Functions

Let x = f (t, x, ) where f is C 1 w.r.t. x and (in addition to what we normally have). Then
solution x(t, ) is differentiable w.r.t. to . TO see this, first convert to integral equation
Z

x(t, ) = x0 +

f (s, x(s, ), )ds

(5.1)

t0

with x0 fixed. Then we can differentiate this with respect to .

We write

x(t, ) = x (t, ) =

t0

f
f
(s, x(s, ), )x (s, ) +
(s, x(s, ), )ds
x

(5.2)

Therefore x (t, ) is a solution of the differential equation t

x (t, ) = x (t, ) = x (t0 , ) +
f
f
f
f
x (t, x(t, ), )x (t, )+ (t, x(t, ), ) = x (t, x(t, ), )x (t, )+ (t, x(t, ), ) because x(t0 , )
x0 (alternatively, could see this from the integral). Let S(t) := x (t, 0 ), A(t) := f
x (t, x(t, 0 ), 0 ),
f
B(t) := (t, x(t, 0 ), 0 ). We generally call this S(t) the sensitivity function and it satisfies the
sensitivity equation:

S(t)
= A(t)S(t) + B(t)
(5.3)

The significance is the following: up to first order we can write

x(t, ) = x(t, 0 ) + S(t)( 0 ) + o()

(5.4)

Nonlinear Analysis Notes

5.3

12

Comparison Principle

Let u and v be two scalar valued signals satisfying both

u = f (t, u),

u(t0 ) = u0

(5.5)

v f (t, v),

v(t0 ) = v0

(5.6)

Then v(t) u(t) as long as the two signals exist.

(t, u) + for smlal > 0. If v crosses above u, then
To show this rigorously, can consider uf
by continuous dependence, v also crosses above u for small , which is a contradiction.

5.3.1

Some Examples

Example 1 x = x x4 x. By comparison principle, solution of x = x with same initial condition,

et c0 gives an upper bound for x = x x4 .
Example 2 x = x x3 . Instead of considering x directly, lets look at v(x) := x2 . Then v = 2xx =
2x2 2x4 2x2 = 2v, and now we have reduced this problem to the last example with
v 2v; namely v(t) e2t v(t0 ). This implies that x2 (t) e2t x2 (t0 ) |x(t)| et |x(t0 )|.

5.4

Lyapunov Stability

5.4.1

Stability Definitions

First the autonomous case where

x = f (x),

x Rn

(5.7)

and f is locally Lipschitz. Assume that x = 0 is an equilibrium, namely f (0) = 0.

Definition 5.1. A local equilibrium xe = 0 is Lyapunov stable if for every  > 0 there is a > 0
such that |x(0)| < implies that |x(t)| <  for all t 0.
Definition 5.2. A local equilibrium point xe = 0 is attractive if there is a > 0 such that for all
initial conditions x0 satisfying |x0 | < , x(t) 0 as t . More formally, for every x0 Rn ,
there is a such that for all  > 0 there exists a T (a function of x0 and ) for which |x(t)| < 
whenever t T and |x0 | < .
Definition 5.3. A local equilibrium point xe = 0 is [locally] asymptotically stable if it is stable and
locally attractive.
An example of stable but not attractive system is
  
 
x 1
0 1
x1
=
x 2
1 0
x2

(5.8)

Nonlinear Analysis Notes

For the converse non-implication, consider Vinograds counterexample

!
 
x2 (yx)+y 5
x
(x2 +y 2 )(1+(x2 +y 2 )2 )
=
y 2 (y2x)
y
2
2
2
2 2
(x +y )(1+(x +y ) )

here f is indeed Lipschitz and the system is attractive, and its not Lyapunov stable.

13

(5.9)

Nonlinear Analysis Notes

14

Lecture 6

Definition 6.1. Let x = f (x) with origin the equilibrium. We say the equilibrium is Lyapunov
stable if for every  > 0 there is a > 0 such that |x(0)| < |x(t)| <  for all t, and attractive
if there is a 2 such that |x0 | < 2 lim x(t) = 0; finally the equilibrium is asymptotically stable
t

if it is L. stable and attractive.

Attractivity: there is 2 > 0 such that for every x0 satisfying |x0 | < , for all  > 0 there is a
T (x0 , ) such that |x(t)| <  for every t > T .
Definition 6.2. Region of attraction is the set of all x0 from which x(t) as t .
Remark: asymptotic stability has a ball in its definition but the region need not be a ball.
Definition 6.3. Global Asymptotic Stability: asymptotically stable and region of attraction is entire
space.
Definition 6.4. Exponentially stable if there are , c, > 0 such that |x(0)| < |x(t)| c|x0 |et
for all t 0.
Remark: in a time invariant system, we can WLOG take any arbitrary equilibrium to be zero
by change of coordinates and that initial time is zero.
Definition 6.5. Globally Exponential Stability: Exponential Stability for all > 0.
Now for time varying case
x = f (t, x),

x(t0 ) = x0

(6.1)

We still assume that origin is equilibrium, namely f (t, 0) = 0.

Definition 6.6. Equilibrium point is Lyapunov stable if for every  > 0 and every initial time t0
there is a 1 (, t0 ) such that |x0 | < 1 |x(t)| <  for all t t0 .
Definition 6.7. Attractive: if for every t0 there is a 2 such that |x0 | < 2 x(t) 0, namely
for every t0 , there is a 2 (t0 ) such that for every x0 such tht |x0 | < 2 and for every  > 0 there is
a T (t0 , x0 , ) such that |x(t)| <  for all t t0 + T .
As before, asymptotic stability is stability with attractivity. The rest generalize in the obvious
way; just include initial time where needed (e.g. |x(t)| c|x0 |e(tt0 ) for all t t0 ).
With these definitions, convergence can get slower as we increase t0 ; also there is included the
possibility of a lack of robustness w.r.t. disturbances. Therefore, we seek stability properties which
are uniform with respect to t0 .
Definition 6.8. 0 is uniformly stableif for every  > 0 there is a 1 > 0 (independent of t0 ) such
that for all t0 , for all x0 satisfying |x0 | < 1 we have that |x(t)| <  for all t t0 .

Nonlinear Analysis Notes

15

(These new definitions, as usual with uniformity concepts, simply permute the order of some
quantifiers from previous definitions.)
Definition 6.9. Uniformly attractive: if there is a 2 > 0 such that for all  > 0 there is a T such
that for all t0 , |x0 | < 2 |x(t)| <  for all t t0 + T .
Note that T = T (2 , ) does not depend on t0 or specific choice of x0 in the 2 ball.
Uniformly Stable + Uniformly Attractive = Uniformly Asymptotically Stable
Similarly for GUAS (global uniform asymptotic stability): for all 2 > 0 and  > 0, there is
a T (2 , ) such that |x0 | < |x(t)| <  for all t t0 + T . GUES: |x(t)| c|x0 |e(tt0 ) for all
t t0 , already uniform w.r.t. t0 , because convergence depends only on t t0 , not t0 itself. Already
uniform w.r.t. x0 (in 2 ball) because it involves only the norm |x0 |.

6.1

Comparison Functions

Definition 6.10. A function : [0, ) [0, ) is called K if its continuous, (0) = 0 and
monotonically increasing.
Note that we can also define on a finite interval to be of class K if it has the same property.

Examples: x, x.
Definition 6.11. If : [0, ) [0, ) is of class K and (r) as r then is of class
K .
Definition 6.12. A continuous : [0, ) [0, ) [0, ) is called KL if (, t) is of class K for
each fixed t and (t, ) decreases to zero for each fixed r.
Example: (r, t) = cret .
Given a system
x = f (t, x),

x(t0 ) = x0 ,

f (t, 0) = 0

(6.2)

then we have the following lemma

Lemma 6.1 (Khalil Lemma 4.5). Uniform Stability [at origin] is equivalent to the existence of
K such that |x(t)| (|x0 |) for |x0 | < c some c R. Secondly, uniform asymptotic stability
is equivalent to the following: there is a KL such that |x(t)| (|x0 |, t t0 ) for |x0 | c some
c R.
WLOG, can take K , by increasing if necessary.
Note that exponential stability is a special case, with (r, s) = cres . Another example of a
r2
.
class KL function is (r, t) = 1+t

Nonlinear Analysis Notes

16

Proof. We prove the second: suppose that |x(t)| (|x0 |, t t0 ), KL. Since is decreasing
in the second argument, we have |x(t)| (|x0 |, 0) = (|x0 |) K. Now need to show that this
implies uniform stability (for every  > 0 there is a > 0 such that |x0 | < |x(t)| < ). Choose
such that () = . In other words, () = 1 (), the local inverse of .
Finally, since |x(t)| < (|x0 |, t t0 ) and for each fixed |x0 | we have (|x0 |, t t0 ) 0 as
t implies that x(t) 0; |x0 | < 2 , to have (2 , t t0 ) <  need t t0 T for large enough
T dependent on . But this follows from the definition of L

The point of defining these classes of functions is precisely to encode the information contained
in the definition of stability and attractivity.

6.2

Lyapunov Functions

Given a system x = f (x), x Rn , then a continuously differentiable function V : Rn R, V (x) is

positive definite for all x Rn \ 0, and whose derivative is negative semi definite v(x)

0, radially
unbounded (V (x) whenever ||x|| ) is said to be Lyapunov function.

Nonlinear Analysis Notes

17

Lecture 7

Let V : Rn R be C 1 function for system

x = f (x)

(7.1)

x Rn . We want derivative of V along solution of x = f (x). V (x(t)): consider as a function of

time for a given solution x(t); then
V
V
d
V (x(t)) =
x(t) x(t)

=
f |x(t)
dt
x
x
We work with
to x = f (x).

V
x

(7.2)

f (x) a function from Rn R; this is defined independently of a solution

Example: Let V (x) = xT P x, and x = Ax; then

V
f |x = x T P x + xT P x = (Ax)T P x + xT P (Ax) = xT (P A + AT P )x
V (x) =
x

(7.3)

Theorem 7.1 (Lyapunovs 2nd Method). Let

x = f (x),

f (0) = 0

(7.4)

with V : Rn R positive definite and C 1 . Then

1. The zero equilibrium is stable if
V (x) 0 x Rn \ 0

(7.5)

V (x) < 0 x Rn

(7.6)

V (x) < 0 x 6= 0

(7.7)

2. Asymptotically stable if

3. GAS if
and V is radially unbounded.
Remarks: V (0) < 0 is not possible since V (0) = V
x f |x=0 = 0. Secondly, for the first two
n
statements, the same holds locally for V : D R R and we just ask for V 0 for all x D.
Proof. We start with the first statement. We need to show that for every  > 0 there is a > 0 such
that |x0 | < |x(t)| <  for every t. Let  > 0 be given. Take positive b such that b < min v(x)
|x|=

which is well defined by positive definiteness of V . Then let 1 > 0 be such that if |x| < 1 , then
V (x) b which exists by continuity of V .
Now we claim that if |x0 | < 1 , then |x(t)| <  for all t. Well, |x0 | 1 implies that V (x0 ) t
and V (x(t)) 0 implies that V (x(t)) b for all t 0. Given this, then we have |x(t)| <  because

Nonlinear Analysis Notes

18

otherwise, there would be a time t such that |x(t)| =  which by choice of b implies that V (x(t)) > b,
contradiction.
For aysmptotic stability, pick some  > 0 and find 1 > 0 for this  as in the first part.
Let |x0 | 1 ; we need to show that x(t) 0. Consider V (x(t)): V (x(t)) < 0 except at 0 and
V (x(t)) 0 (monotonic) implies that V (x(t)) =: c exists. Two possibilities: c = 0 and c 6= 0.
Case 1: c = 0. Then x(t) ; then V (x(t)) 0, V = 0 only at zero implies that x(t) 0.
Case 2: c > 0 Then V (x(t)) c > 0 for all t. As before, there is an r > 0 such that for
all |x| r we have V (x) c. However, x(t) a solution to the system cannot enter the r-ball, i.e.
0 < r |x(t)| . The set {x : r |x| } is a compact set. Take max V in this set and call it
d < 0. For all t, then, we have V (x(t)) V (x0 ) dt for all t 0. But then V < 0 at some time
T , contradcting positive definiteness of V . So c > 0 isnt possible.
For global asymptotic stability, we want to show that 1 as  . As in proof of first
statement, as  , since V is radially unbounded, we can take b and therefore 1 as
well.

Example: x = (x) with x(x) > 0 for all x 6= 0. Then set V = 1/2x2 so that V = x(x)

<0
for all x 6= 0 which implies GAS.

Nonlinear Analysis Notes

19

Lecture 8

8.1

Laypunov Example

Last time we proved Lyapunovs theorem, namely that x = f (x) with x Rn , and f (0) = 0, for
V : Rn R a C 1 positive definite function and V (x) = V
x f (x) zero at equilibrium point: then

the system is L. stable if V (x) 0 for all x and asymptotically stable if V (x) < 0 for all x 6= 0.
Global asymptotic stability holds if V is radially unbounded. Recall that radial unboundedness
means that V (x) as |x| . As a counter example, (x1 x2 )2 is not radially unbounded.
Example:
x
+ x + (x) = 0
in state space form this is equivalent to
  

x 1
x2
=
x 2
x2 (x1 )

(8.1)

(8.2)

satisfying x(x) > 0 for all x 6= 0. We want to

whether or not this is stable or asymptotically
Z find
x
stable. Potential energy is given by (x) =
(z)dz and kinetic energy is 21 x22 . For Lyapunov
0

function lets first try

1
V (x1 , x2 ) = (x21 + x22 )
2

(8.3)

V = x1 x2 x2 (x2 + (x1 )) = x1 x2 x22 x2 (x1 )

(8.4)

Then
This function obviously isnt helpful because we know nothing about (x1 )x2 .
Then take V to be the total energy,
1
V (x) = (x1 ) + x22
2

(8.5)

which is positive definite if x(x) > 0 and radially unbounded if x(x) > kx2 for some k > 0. We
calculate
V (x) = (x1 )x2 x2 (x2 + (x1 )) = x22
(8.6)
which is negative semindefinite. Therefore, by Lyapunovs 2nd method the origin is L. stable.
However it is not negative definite because at x2 = 0, but x1 6= 0 we still have V (x) = 0. Therefore
we cannot conclude that the system is asymptotically stable. Nevertheless, our inability to conclude
a. stability does not imply that the system is not asymptotically stable. In other words, Lyapunov
stability is only a sufficient condition for a. stability. Perhaps we can find another Lyapunov
function V which would work.
We need to make sure that V is positive definite


a1 a3 /2
>0
a3 /2 a2

(8.7)

Nonlinear Analysis Notes

A choice that works:

V =

a 2
1
x + ax1 x2 + x22 + (x1 )
2 1
2

20

(8.8)

for 0 < a < 1; then

a a2
det a/2 a/2 a/2 1/2 =
>0
(8.9)
4
as long as a (0, 1). V is positive definite and also radially unbounded; need to check that its
derivative is negative definite:
V = (ax1 +ax2 +(x1 ))x2 (ax1 +x2 )(x2 +(x1 )) = ax1 x2 +ax22 +(x1 )x2 ax1 x2 ax1 (x1 )x22 x2 (x1 )
(8.10)
Simplifying, we have
V (x) = (1 a)x22 ax1 (x1 ) < 0
(8.11)
and this is negative definite.
Lesson: finding Lyapunov functions is hard. There is no algorithmic method for finding one
in general, but a general approach is trial and error. Next is another way to find Stability:

8.2

LaSalles Principle

Goal: conclude global asymptotic stability from only V 0 with some additional analysis. Let
x = f (x) and x = x(t) be a fixed solution of the system. For this trajectory x(t), define its positive
limit set + as the set of all limit (or accumulation) points of x(t); in math lingo:
[
+ =
{z Rn : lim x(tk ) = z}
(8.12)
{tk }2R+

tk

where we stipulate (notationally) that lim x(tk ) = whenever the limit doesnt exist (and further
tk
[
+
we say
:= ). Sometimes is denoted as L+ and also sometimes called the -limit set

(also going backward in time we have -limit set). The concept is useful really only for bounded
trajectories which dont converge to an equilibrium.
Another definition
Definition 8.1. A set M Rn is called [positive] invariant] if x0 M implies that x(t) M for
all t > 0.
Properties of positive limit set of +
Proposition 8.1. Assume that x(t) is a bounded solution of x = f (x). Then its positive limit set
+ has the following properties:
1. + 6= .
2. + is bounded.

Nonlinear Analysis Notes

21

3. + is closed (and therefore compact).

4. + is an invariant set.
5. x(t) + as t . More precisely, for lim x(t, + ) := lim inf |x(t) z| = 0.
t

Proof.

t z+

1. Bolzano-Weierstrass (a bounded set always has a limit point).

2. + is a subset of the closure of the evolution set.

3. Need to show that if xk + and xk x then x + as well. Define a sequence of tkj such
that x(tkj ) x by taking the diagonal of sequences {tjk } where tjk is the sequence of points
tk for which x(tjk ) xj for j fixed. Just make sure that |x(tkj ) xk | < 1/k and tkj .
It converges to x because |xk x| < /2 for k large enough and |x(tkj ) xk | <  for j large
enough, using the triangle inequality this implies that |x(tkj ) x| < . Therefore x + .
4. Need to show that for any y0 + , then the corresponding solution {y(t) : t R+ } + .
t
x(tj ) y0 = y(0)
y(t), then x(tj + t) y(t) too, by continuity of solution w.r.t. initial
condition because if we took x0
/ + but close enough to y0 , then x0 (t) should be close to
y0 (t) as well.
5. If not, then there is an  > 0 and sequence {tk } such that d(x(tk ), + ) ; in other
words, that |x(tk )z|  for all z + . But x(tk ) bounded implies it is sequentially bounded
(subsequence has a limit), call it x; then by definition x + , contradiction (since x
/ +
since |x z|  for all z + ).

Nonlinear Analysis Notes

22

Lecture 9

Last time we talked about positive limit set, as the set of limit points of a trajectory, and we
enumerated a litany of properties which this limit set satisfies.

9.1

LaSalles Invariance Principle

Theorem 9.1. Let V : Rn R be C 1 positive definite, and satisfy

V (x) 0, x Rn

(9.1)

Let S := {x Rn : V (x) = 0} and M be the largest invariant subset of S. Then every bounded
solution x(t) converges to M as t .
Corollary 9.1 (Barbashin-Krasovski). Under the same hypotheses of the previous theorem, if
M = {0} and all the solutions are bounded, then the system is GAS.
Proof of Theorem. Fix x(t) a (arbitrary) bounded solution. From Lyapunovs theorem, the hypothesis that V (x) 0 implies that x(t) is already Lyapunov stable and since V (x(t)) is (not necessarily
strictly) decreasing, it has a limit as t , call it C 0. Let + be the positive limit set of
x(t). By definition, every z + is the limit of {x(tk )}kN for some sequence of times {tk } R.
V (x) = V ( lim x(tk )) = lim V (x(tk )) = C where the second equation holds by continuity. Here z
k

was arbitrary which means this limit is for every z + , i.e. V (+ ) = {c}.
From property 4 (of + from last lecture), + is invariant which implies that V (x) = 0 on
because V c on + . From property 5, lim x(t) + . To summarize, + M S := {x :

+
t
V (x) = 0}, and x(t) + , which is exactly what we wanted to show.

Remarks: LaSalle states that x(t) M , not x(t) + because + is hard to find and
depends on choice of x(t); M is typically easier to find and works for all solutions x(t).
Boundedness of x(t) is crucial; unbounded solutions, if they exist, may not converge to M
(and in general probably will not). IF V is radially unbounded (an additional hypothesis), then
V 0 implies that each solution is bounded.
Locally, for x0 close enough to zero, solutions are bounded by Lyapunov stability.
Positive definiteness of V in theorem is needed for Lyapunov stability but not needed for
x(t) M claim; we only need a lower bound on V (x(t)), which exists because x(t) is bounded.
We want to use the corollary to show global asymptotic stability. Need to be able to show
that M = {0} iff there are no nonzero solutions along V 0.
Example:
  

x 1
x2
=
x 2
x2 (x1 )

(9.2)

Nonlinear Analysis Notes

23

satisfying x1 (x1 ) > 0 for all x1 6= 0. Energies are given by

1
K = x22
2
Z x1
U = (x1 ) =
(z)dz

(9.3)
(9.4)

We tried a Lyapunov equation as total (internal) energy, v = 12 x22 +(x1 ) and V = (x1 )x2 x2 (x2 +
(x1 )) = x22 0. Recall that this L. function did not work before using Lyapunov analysis. Lets
see how LaSalle can help. The set S is just the x axis, S = {(x1 , x2 ) R2 : x2 = 0}. Suppose there
is a nonzero solution along which V = 0, i.e. one that remains in the x1 axis, so x2 0. But if
x2 = 0 then x 2 0, so that by equation of x,
both x 2 and x 1 are identically zero. Then (x1 ) must
be identically zero which implies that x1 0. This establishes that the equilibrium is the only
solution for which V = 0, and by LaSalles invariance principle (or the corollary of B-K), provided
that V is radially unbounded, the system is globally asymptotically stable. Note, however, that V
need not be radially unbounded; so it is necessary to stipulate certain conditions on to ensure
radial unboundedness of V .
Another example from adaptive control:
x = x + u

(9.5)

where is fixed but unknown. Idea: introduce a guess for value of and define the control law
then the closed loop x = x which is nicely stable. If they are not equal,
u = ( + 1)x. If = ,
tune according to

= x2
(9.6)
which will grow fast if x does not converge to zero (thereby eventually will dominate and x will
= 1 (x2 + ( )2 ), then V = xx + ( ) = x2 0.
go to zero). Lets try V (x, )
2

clear all
theta=100;
f=@(t,x)[(thetax(2)1)*x(1);(x(1))2]
figure
hold all
[tp,xp]=ode45(f,[0,2],[1,1]);
plot(tp,xp(:,1));
figure
plot(tp,xp(:,2));

: V = 0} which is the set of points where x = 0. M = {0} iff there are no nonzero
S = {(x, )

solutions in axis; S consists entirely of equilibria, M = S, implies that LaSalle tells us nothing
about (see figure; here = 100 but 180). This is nevertheless a good result: V radially
unbounded and V 0 we can still say that remains bounded.

Nonlinear Analysis Notes

24

100

80

60

40

20

0
0

0.1

0.2

0.3

0.4

0.5

Figure 1: Simulation of behavior of x

200
180
160
140
120
100
80
60
40
20
0
0

0.5

1.5

Figure 2: Behavior of Adaptation

Nonlinear Analysis Notes

10

25

Lecture 10

10.1

Lyapunovs First Indirect Method

Idea: use quadratic Lyapunov function to show local asymptotic stability.

Let V (x) = xT P x, P = P T > 0 C 1 positive definite and radially unbounded. Also, suppose
that min (P )|x|2 V (x) max (P )|x|2 ;
x = f (x),

f (0) = 0

(10.1)

construct linearization; for each component fi (x), i = 1, . . . , n, apply mean value theorem to write
i
fi (x) = fi (0) + f
x (zi ) x where zi (0, xi ). Since 0 is equilibrium point the first term is zero, so
we can rewrite this as
fi
fi
fi
f (0)x + [
(zi )
(0)] x
(10.2)
x
x
x
The temr in brackets approaches zero as x 0 by continuity. Therefore, we have
fi
(0) x + gi (x)
x

(10.3)

where gi (x) = o(|x|), which means exactly that

gi (x)
=0
x0 |x|
lim

(10.4)

We can repeat this for each i, and we have

f (x) =
where

f
x

f
(0) x + g(x)
x

(10.5)

denotes the Jacobian matrix

f1

x1

= ...
x
f

n
x1

..
.

f1
xn

..
.

(10.6)

fn
xn

To avoid cumbersome notation, we call A := the Jacobian of f , and write f (x) = Ax + g(x) which
approximates the behavior of the original system near 0.
Theorem 10.1 (Lyapunovs First Method). Given system
x = f (x),

f (0) = 0

(10.7)

and linearization f (x) = Ax + g(x) with g(x) = o(|x|). Suppose A is Hurwitz; thne the system is
locally asymptotically stable at 0.

Nonlinear Analysis Notes

26

Proof. Recall from ECE 515 that A Hurwitz implies that theres a positive definite matrix P =
P T > 0 satisfying the Lyapunov equation
P A + AT P = Q < 0

(10.8)

for some positive definite matrix Q. Let V (x) = xT P x. Along solutions of x = f (x) = Ax + g(x),
we get V (x) = xT (P A+AT P )x+2xT P g(x) = xT Qx+2xT P g(x) min (Q)|x|2 +2|x|||P |||g(x)|
but we know that |g(x)|
|x| 0 as |x| 0; in other words |x| < |g(x)| |x| for given  and
there is such a related () ( . . .). Then for |x| < , we have V min (Q)|x|2 + 2||P |||x|2 =
(min (Q)2||P ||)|x|2 . Pick  small enough so that V < 0 for |x| < . This implies by Lyapunovs
2nd method local asymptotic stability.
Comments: min (P )|x|2 xT P x min (P )|x|2 , and V |x|2 2V for some > 0.
As  0, decay rate for linearized dynamics. By comparison principle, V (x(t)) e2t V (x0 ),
and this implies thta min (P )|x(t)|2 e2t max (P )|x0 |2 ; divide through by min and take square
roots to get
|x(t)| cet |x0 |
(10.9)
q
(P )
where c = max
, and this equation means that the system is not only locally Lyapunov stable,
min (P )
but logcally exponentially stable, and the convergence rate near zero is that of x = Ax.
|x(t)| cet |x0 |

(10.10)

Around the origin the lyapunov function has level sets which can take various shapes, e.g. ellipses,
and the more elongated the ellipses are the larger the ratio is between max and min .
If A = f
x (0) has at least one eigenvalue with positive real part, then the origin is not stable. Idea: unstable eigenvalue of A dominates behavior of g(x). A rigorous proof via (Chetaev)
instability theorem can be found in Khalil.
If Re(i ) 0 for all eigenvalues but at least one i has real part exactly 0, then the test is
inconclusive and theres nothing we can say from Lyapunovs test regarding stability or instability;
sometimes this is called the critical case. We do know, however, in this case that local exponential
stability is not possible, but nevertheless local asymptotic stability is possible. This will follow
later from converse Lyapunov theorem for exponential stability.

10.2

Some Examples
x = x x3

(10.11)

x = x

(10.12)

and the linearization is simply

and this is asymptotically stable, so the original system is locally exponentially stable. In fact, this
system is globally asymptotically stable by Lyapunovs second method, since V (x) = x2 gives
V = x2 x4 < 0 x 6= 0

(10.13)

Nonlinear Analysis Notes

27

it happens also to be globally exponentially stable since V x2 . On the other hand,

x = x + x3

(10.14)

is still locally exponentially stable (by both Lyapunovs 1st and 2nd method (use same V )).
x = x3

(10.15)

Here the linearization is zero so Lyapunovs first method doesnt help here; on the other hand,
the direct method with V (x) = x2 /2 gives V = x4 < 0 x 6= 0 which implies global asymptotic
stability, but convergence rate is not exponential. Solve the ode to get x(t) 1t .
Nonlinear damped spring, revisited
x
+ x + (x) = 0
where x(x) > 0 for all x 6= 0. Rewriting in state space

  
x2
x 1
=
x2 (x1 )
x 2

(10.16)

(10.17)

Compute linearization
f
A=
|x=0 =
x


0
1
0 (0) 1

(10.18)

and the linearized system is

  
 
x 1
x1
0
1
=
(10.19)
0
x 2
(0) 1
x2



1
0
= 2 + + and apply Routh criteria
and A is Hurwitz iff if (0) > 0 (for det
1
on coefficients). So under this condition, the first method gives local exponential stability. Compare this conclusion with the same example from a previous lecture when we used Lyapunovs
second method, from which we got global asymptotic stability with a creatively found V , under
the assumpion that x(x) > 0. We also used LaSalles invariance principle, and that gave us local
asymptotic stability under the same conditions and global asymptotic stability with strengthened
condition that x(x) > kx2 for positive k (where here we used V = 12 x22 + (x1 ) radially unbounded). For the first method, we get only local asymptotic stability, under the added condition
that 0 (0) > 0, a local condition of x(x) > kx2 . The tradeoff is that the first method is relatively
easy to apply but its conclusion is not as strong as could be with the second method (gives no
global conclusions, and gives nothing if 0 (0) = 0).

10.3

Stability of Nonautonomous Systems

Consider
x = f (t, x)

(10.20)

Nonlinear Analysis Notes

28

Looking a these systems is relevant for analysis of even autonomous systems. Consider x = f (x)
and we want to track a reference trajectory call it xref (t); for this problem we track the error
dynamics for e := x xref with
e = x x ref = f (x) x ref = f (e + xref ) x ref
Thus in state space our system is given by
  

e
f (e + xref (t)) x ref (t)
=
x
f (x)

(10.21)

(10.22)

where these terms describing dynamics of e are time dependent.

The properties that we want from x = f (t, x) will be:
1. Uniform stability
2. Uniform asymptotic stability
3. Global uniform asymptotic stability
Well see that V (x) is not always enough; we will sometimes need to work with V (t, x), from which
V
V = V
t + x f (t, x), and we will want this to be negative uniformly over time.

Nonlinear Analysis Notes

11
11.1

29

Lecture 11
Stability of Non-Autonomous Systems

x = f (t, x) with x Rn and t [t0 , ), and f (t, 0) 0 an equilibrium at the origin. Given
V
function V = V (t, x) : [t0 , ) Rn R, and V (t, x) := V
t (t, x) + x (t, x) f (t, x), along solution
x(t) for x = f (t, x) we have
d
V (t, x(t)) = V (t, x(t))
(11.1)
dt
Suppose there is a continuous positive definite functions W1 (x), W2 (x) such that
W1 (x) V (t, x) W2 (x) t, x

(11.2)

This implies that V (t, x) is positive definite in the sense that V (t, 0) = 0 and V (t, x) > 0 for all
x 6= 0, for all t; property V (t, x) W2 (x) is called decrescent property of V .
Theorem 11.1 (Lyapunovs Direct Method for Time-Varying Systems). Given system
x = f (t, x)

(11.3)

with equilibrium f (t, 0) 0, and V (t, x) C 1 positive definite and decresecent. Then the following
hold:
1. V (t, x) 0 origin is uniformly stable.
2. If V (t, x) W3 (x) < 0 for all x 6= 0, then origin is uniformly asymptotically stable.
3. If condition 2. holds and W1 is radially unbounded, then we have GUAS.
Proof. 1 is similar to the proof of 1. for Lyapunovs Theorem in the autonomous case x = f (x).
We need to show that for every  > 0 there is a > 0 such that |x0 | < |x(t)| <  for all t0 and
all t t0 . Fix  > 0 and pick positive b satisfying
b < min W1 (x)
|x|=

(11.4)

and pick > 0 such that if |x| we have V (x) b. Then if |x0 | , from V 0 we know that
V (t, x(t)) V (t0 , x0 ) W2 (x0 ) b. This means that |x(t)| remains less than  because if at some
t we have |x(t)| = , then W1 (x(t)) > b by definition of b, and hence V (t, x(t)) W1 (x(t)) > b,
which contradicts that V (t, x(t)) b for all t.
For 2., let = () for some  > 0, and let |x0 | . There is class K functions 1 , 2 on
[0, ] such that 1 (|x|) W1 (x) V (t, x) W2 (x) 2 (|x|). Similarly, V W3 (x) 3 (|x|),
where 3 K on [0, ]; 2 K has an inverse 21 . This gives V 3 (21 (V )) =: (V ) (a
composition of two class K functions and is therefore also class K).
Claim: the scalar system
y = (y),

yR

(11.5)

Nonlinear Analysis Notes

30

, for continuous positive definite, is asymptotically stable: KL such that |y(t)| (|y0 |, tt0 )
(at least for |y0 | some ). Proof of claim: use V (y) = 21 y 2 ; then V (y) = y(y) < 0, y > 0.
Note that can be explicitly constructed from (Lemma 4.4 K). Use comparison principle:
y = (y),

y(t0 ) = V (t0 , x0 ) 0

(11.6)

V (V )

(11.7)

V (t, x(t)) (V (t0 , x0 ), t t0 )

(11.8)

1 (|x(t)|) V (t, x(t)) (V (t0 , x0 ), t t0 ) (2 (|x0 |), t t0 )

(11.9)

it tells us that
Then
Applying 11 to the outer terms, we have
|x(t)| 11 ((2 (|x0 |), t t0 )) =: (|x0 |, t t0 ) KL

(11.10)

This shows U AS.

3. has the same proof as in the autonomous case.
Remark: If 1 , 2 , 3 are scalar multiples of some power of |x|,
k1 |x|a V (t, x) k2 |x|a

(11.11)

V (t, x) k3 |x|a

(11.12)

and
for some k1,2,3 , a > 0, then 0 is uniformly exponenentially stable (and globally so if the inequalities
holds globally).
We already made a similar observation earlier for quadratic bounds a = 2. Then
k3
(tt0 )
V V V (t, x(t)) e k2
V (t0 , x0 )
k2
k3

(11.13)

where the implication holds by the comparison principle. This implies that
k

k3 (tt0 )

k1 |x(t)|a e
and then

k2 |x0 |a

k2 1/a
) |x0 | =: ce(tt0 ) |x0 |
k1
where the last term is the desired UES estimate.

(11.14)

|x(t)| e

ak3 (tt0 )
2

(11.15)

For example: given linear time varying dynamics (LTV),

x = A(t)x

(11.16)

V (t, x) = xT P (t)x where 0 < c1 I P (t) c2 I for all t and some c1 , c2 > 0. This makes V positive
definite and decrescent. The derivative of V then is
P (t) + P (t)A(t) + AT (t)P (t) Q(t) c3 I, c3 > 0

(11.17)

Nonlinear Analysis Notes

31

Then remark applies with a = 2 and we have GU ES. In fact, for linear (even time-varying)
U AS GU ES.
Briefly, Lyapunovs First (indirect) method for time varying systems
x = f (t, x) = A(t)x + g(t, x)

(11.18)

f
where A(t) = f
x (t, 0). If x is uniformly continuous in t for each x, then g(t, x) = o(|x|). If
x = A(t)x is asymptotically stable (or GUES, automatic for linear systems), then x = f (t, x) is
locally GUES. There is no simple eigenvalue test for the non-autonomous case to check for this
condition.

11.2

LaSalle-Like Theorems for Time Varying Systems

There is some information which can be gotten from

V (t, x) W3 (x) 0

(11.19)

where W3 is positive semi definite). However, the conclusions are not as strong as Lasalles results
from time invariant systems.

Nonlinear Analysis Notes

12

32

Lecture 12

12.1

LaSalle Like Stability Theorem for Time Varying Systems

Theorem 12.1. Let

x = f (t, x)

(12.1)

satisfy the usual conditions, and let f (t, 0) be uniformly bounded, namely f (t, 0) C for some
C > 0 all t 0. Let V (t, x) be positive definite decrescent, i.e. there exists w1 and w2 positive
definite satisfying
w1 (x) V (t, x) w2 (x)
(12.2)
and that there is a w3 continuous positive semi definite:
V (t, x) w3 (x) 0, x

(12.3)

Then all bounded solutions converge to the set S := {x : w3 (x) = 0}.

Example: Consider

  
e2t x2
x 1
=
x 2
e2t x1 x2

(12.4)

and let V (x1 , x2 ) = 21 (x21 + x22 ), then V = x22 0, so S = {(x1 , x2 ) R2 : x2 = 0}; applying
the usual procedure, x2 0 x 2 0 which implies by the second equation that x1 0 and this
implies that there are no nonzero solutions in S M = {0}. Nevertheless it is not true that x
converges to 0. To see this, pick x0 such that x1 (t0 ) > 0 and x2 (t0 ) = 0. The first observation, is
that a solution x(t) will remain in the B||x0 || ball about 0. In other words, |x(t)| |x0 | for all t.
We solve the equations:
Z
t

e2s x2 (s)ds

x1 (t) = x1 (0) +

(12.5)

t0

and use |a b| |a| |b| to write

Z

|x1 (t)| |x1 (t0 )| |x1 (t0 )|

t0

1
|e2s |ds |x1 (t0 )|
2

(12.6)

Where the last inequality follows from that the integral, evaluated, is 12 e2s |tt0 12 (e2t 1) 12 .
Therefore, x1 (t) 6 0.
Boundedness of solutions follows locally from V 0, and global boundedness follows from w1
being radially unbounded (forces V to be radially unbounded).
Before proving the theorem, we will need the following lemma:
Lemma

is bounded, w is continuous,
Z 12.1 (Barbalats Lemma). Suppose that x(t) is bounded, x(t)
and
w(x(t))dt < . Then w(x(t)) 0 as t .
t0

Nonlinear Analysis Notes

33

Proof. Since w(x) is uniformly continuous over D, we have that for every  > 0 there is a x > 0
such that |x y| x |w(x) w(y)| 2 . Secondly, since x is bounded, x(t) is also uniformly
continuous, i.e. given any x > 0 there is a t such that |t1 t2 | t |x(t1 ) x(t2 )| x .
Combining, we have that |t1 t2 | t |w(x(t1 ) w(x(t2 ))| /2.
Now we show that w(x(t)) 0 as t . Suppose not; then there is an  > 0 and
sequence {tk } such that w(x(tk ))  for all k. From uniform continuity, |t tk | t , then
w(x(t)) /2. For simplicity, take w 0 (in the theorem w3 has this property). Then we integrate
Z t

X
w
and get
t /2 = .
t0

Proof of Theorem. Consider V (t, x) w3 (x) and integrate from t0 to t; then V (t, x(t))V (t0 , x0 ) =
Z t
Z t

w3 (x(s))ds, iff
w3 (x(s))ds V (t0 , x0 ) V (t, x(t)) V (t0 , x0 ) for all t where the last int0
t0
Z
equality holds because V is positive definite. Therefore
w3 (x(s))ds exists and is bounded. We
t0

need to show that w3 (x(s)) 0 as s . If x(t) is bounded, the image is contained in a compact
set D; since w3 (x) is continuous, it is uniformly continuous over D, i.e. for every  > 0 there is a
x > 0 such that |x y| x |w(x) w(y)| 2 .
Now we want to invoke the previous lemma: we have already that x(t) is bounded, that w
is continuous and that its integral is finite. But we still have to check that x is bounded. In our
system, x(t)

= f (t, x(t)) but Lipschitness of f implies that |f (t, x) f (t, 0)| L|x| for all x D
and all t. We assumed in the statement of the theorem that f (t, 0) c for all t, so these two
things together imply that f (t, x) must also be bounded for all x D and all t as desired. Now
we can apply the lemma, to get that w3 (x(t)) 0 as t , which is exactly what we wanted to
show.
As a special case, this proof also shows that for x = f (x) (time invariant), if V 0 for all x
then every bounded solution converges to S = {x : V 0}. Note that this is weaker than LaSalle,
but the proof does not involve properties of limit set + .
For general time varying systems, the stronger claim that x(t) M is not true. It is, however,
true for some special classes of time varying systems (see [1]).

Nonlinear Analysis Notes

13

34

Lecture 13

Las time we proved LaSalle-Yoschizawa,

w1 (x) V (t, x) w2 (x)

(13.1)

V (t, x) w3 (x) 0

(13.2)

and
both imply that along bounded solutions
lim w3 (x(t)) = 0

(13.3)

We want to know when w3 (x) 0 implies that x 0. Define output y = w3 (x), so want
y 0 x 0. This is observability. Interpretation of LaSalle-Yoshizawa, V w3 (x) and
observable w.r.t. y := w3 (x) implies (G)AS (G if all solutions are bounded).
Example, consider the linear time varying system
x = A(t)x

(13.4)

V (t, x) = xT P (t)x

(13.5)

V (t, x) = xT [P (t)A(t) + AT (t)P (t) + P (t)]x

(13.6)

Q(t) = C T (t)C(t)

(13.7)

and

for some C(t) such that the following equation holds

P (t)A(t) + AT (t)P (t) + P (t) C T (t)C(t) 0

(13.8)

Define y := C(t)x. Then V y T y. Observability Gramian:

Z

t0 +

M (t0 , t0 + ) =

T (t, t0 )C T (t)C(t, t0 )dt

(13.9)

t0

and we say that the system is uniformly observable if there is a positive c R such that M (t0 , t0 +
) cI for all t0 .
Uniform observability and equation 13.8 imply GES. See [1, Khalil] for proof.

Nonlinear Analysis Notes

13.1

35

Converse Theorems

We had several versions of Lyapunovs stability theorems. Conclusions, e.g.

1. V < 0 x 6= 0 AS
2. V < 0 x 6= 0 and V radially unbounded GAS.
3. K1 |x|2 V (x) k2 |x|2 and V k3 |x|2 GES.
4. x = f (t, x), V (t, x) satisfying V w3 (x) < 0 x 6= 0 implies U AS.
Recall from 515, if A Mn (R) is Hurwitz, then there is, for every positive definite symmetric
Q, a unique P = P T > 0 such that
P A + AT P = Q
(13.10)
For V = xT P x, V = xT (P A + AT P )x = xT Qx < 0x 6= 0, and P is in fact given by
Z
T
P =
eA QeA d

(13.11)

There is also an extension of this result to LTV systems: x = A(t)x and assume that its
GUES, i.e.
||(, t)|| ce( t) , c, > 0
(13.12)
which means
|x(t)| = |(, t)x0 | = ||(, t)|||x0 | c|x0 |e( t)
Then there is a quadratic Lyapunov function xT P (t)x where
Z
P (t) =
T (, t)Q( )(, t)d

(13.13)

(13.14)

where Q() is chosen to satisfy for some a, b > 0,

0 < aI Q(t) bI

(13.15)

The Lyapunov function is

V (t, x) = xT P (t)x =

xT T (, t)Q( )(, t)xd

(13.16)

Notice that the (, t)x term is the solution at time of the system starting at time t. We can
rewrite this as
Z
T ( ; t, x)Q( )( ; t, x)d
(13.17)
t

where ( ; t, x) denotes solution at time which starts in state x at time t.

Claim: V (t, x) < 0, and in fact c1 |x|2 V c2 |x|2 and V c3 |x|2 which proves GES. This
will follow from a more general converse theorem for GES nonlinear systems.

Nonlinear Analysis Notes

36

Theorem 13.1 (Converse Lyapunov For Exponential Stability). Take a nonlinear LTV
x = f (t, x), f (t, 0) 0, f C 1

(13.18)

and there are L, r > 0 such that

||

f
|| L, t, |x| r
x

(13.19)

Assume UES (U automatic), namely

|x(t)| c|x0 |e(tt0 ) , |x0 | r/c

(13.20)

Then locally in r/c-ball about 0, there is a function V (t, x) satisfying the following three properties
1. c1 |x|2 V (t, x) c2 |x|2 with c1 , c2 > 0.
2. V =

V
t

V
x f (t, x)

c3 |x|2 with c3 > 0.

3. | V
x | c4 |x|, c4 > 0 (V is quadratic-like).
This V can be defined by
t+

T ( ; t, x)( ; t, x)d

V (t, x) =

(13.21)

for > 0 sufficiently large. Moreover, if hypotheses hold for r = (globally) and the system is
GES, then this V also works for all x.
Note that this formula for V is not constructive because it requires knowledge of solutions. If
the system is time-invariant
x = f (x)
(13.22)
then the solutions ( ; t, x) 7 ( t, x) and V becomes (s = t)
Z
V =

T (s, x)(s, x)ds

(13.23)

which is independent of t: V = V (x).

Z

t+
2

|( ; t, x)| d

Proof. 1. V (t, x) =
t

t+

c2 e2( t) d |x|2 using the exponential stability

2

c
estimate. The right hand side is computable; it is 2
(1 e2 )|x|2 =: c2 |x|2 . This gives the
V
upper bound in the first statement. Using || x || L by the Mean Value Theorem we have
|f (t, x)| L|x| since f (t, 0) = 0. This implies that the solutions cannot decay faster than eLt ,
using the comparison principle. Solutions |( ; t, x)|2 |x|2 e2L( t) which implies that V (t, x)
Z t+
1
e2L( t) d |x|2 can compute this to get a constant 2L
(1 e2L )|x|2 =: c1 , and this proves
t

the lower bound in the first statement.

Nonlinear Analysis Notes

37

2. Need to differentiate the expression for V :

Z
V (t, x) = T (t+; t, x)(t+; t, x)T (t; t, x)(t; t, x)+

t+

2[T ( ; t, x)t ( ; t, x)+x ( ; t, x)f (t, x)]d

(13.24)
so
V (t, x) = T (t+; t, x)(t+; t, x)|x|2 +

t+

2[T ( ; t, x)t ( ; t, x)+x ( ; t, x)f (t, x)]d (13.25)

Claim: t + x f (t, x) = 0; for ( ; t, x) = ( ; t + t, x + tf (t, x)) differentiating with respect to

t gives the result 0 = t + x f . Therefore we have
V (t, x) = T (t + ; t, x)(t + ; t, x) |x|2 c2 |x|2 e2 |x|2 = (1 c2 e2 )|x|2
Then defined c3 := (1 c2 e2 ) for large enough (namely >
Proof of 3. is similar; see Khalil.

log(c2 )
2 ).,

proving 2.

(13.26)

Nonlinear Analysis Notes

14
14.1

38

Lecture 14
Applications of Converse Lyapunov Theorems

Last time we proved converse Lyapunov theorem for (L)ES:

(L)ES that there is a function V satisfying c1 |x|2 V (t, x) c2 |x|2 , V c3 |x|2 and
| V
x | c4 |x|.
Application to Lyapunovs 1st (indirect) method. x = f (x), linearization, x = Ax then
f (x) = Ax + g(x) where g(x) = o(|x|). We have x = Ax exponentially stable implies that x = f (x)
is LES (realied on converse Lyapunov functoin for x = Ax). In the other direction, we use converse
Lyapunov for LES for nonlinear systems. We can in fact prove this implication in Lyapunovs 1st
method for more general time-varying case, i.e. well show that if x = f (x) is locally exponentially
stable, then its linearization x = A(t)x must be exponentially stable.
Have: x = A(t)x the linearization, and x = f (t, x) = A(t)x + g(t, x) with |g(t, x)| = o(|x|)
(uniformly in t).
Since x = f (t, x) is locally exponentially stable, [by converse Lyapunov for exponential staV
V
2

bility] there is a V such that V

t + x f c3 |x| , and | x | c4 |x|. Now lets compute V along
the linearization x = Ax = f (t, x) g(t, x), and g o(|x|) means that for every  > 0 there is a
V
V
V
2
> 0 such that if |x| < then |g(t,x)|
< . Then V |x=Ax
= V

t + x f x g c3 |x| x g
|x|
2
2
c3 |x|2 + | V
x ||g| c3 |x| + c4 |x| when |x| < (). By taking  small enough (i.e.  <
0 < k < c3 , we get that
V |x=A(t)x
k|x|2

c3 k
c4 ),

for

(14.1)

This together with c1 |x|2 V c2 |x|2 implies that the linearization is exponentially stable, which
is what we wanted to prove.

14.2

Vanishing Perturbations (9.1[1])

Given system
x = f (t, x) + g(t, x)

(14.2)

with f the nominal system and g perturbation. Suppose that the nominal system x = f (t, x) is
exponentially stable (either locally or globally), and the perturbation is vanishing the following
sense: |g(t, x)| |x| for all t for some  > 0. By converse Lyapunov for exponential stability, there
is a V satisfying the same conclusions as we had before:
1. c1 |x|2 V (t, x) c2 |x|2
2. V f =

V
t

3. | V
x c4 |x|

V
x f (t, x)

c3 |x|2

Nonlinear Analysis Notes

39

V
V
with all c1 , . . . , c4 > 0. Now we consider V f +g = V
t + x f (t, x) + x g(t, x) and this we can bound,
2
2
as before, by c3 |x| + c4 |x| . The only difference between here and before is that  is fixed, not
arbitrarily small. Nevertheless, if  < c3 /c4 , then there is a k > 0 such that  < c3ck
and this
4
2
implies that V f +g k|x| , and therefore x = f + g is locally exponentially stable. Conclusion:
exponential stability is persevered under small enough vanishing perturbations. (Note that if we
only had asymptotic stability, we would not be able to make this argument.) Also, this doesnt give
us a very constructive means of analyzing robustness; it merely guarantees a sense of robustness
qualitatively.

14.3

Non-vanishing Perturbations (9.2[? ])

Same basic system

x = f (t, x) + g(t, x)

(14.3)

with nominal dynamics x = f (t, x) exponentially stable but this time the perturbation satisfies
only
|g(t, x)| 
(14.4)
a small but does not vanish as x 0. Proceed as before: converse lyapunov theorem gives V with
2

the usual properties, V f c3 |x|2 and V

x | c4 |x| and look at Vf +g c3 |x| + c4 |x|. The only
difference with what we had before is that the perturbation is linear, not quadratic. The reason
this is a problem is that for small x the dominating term will be the perturbation. We will show,
then, that though we cant get a handle on what happens near zero, at a large scale the system
still behaves nicely. Thus
c4
V f +g c3 |x|2 + c4 |x| = c3 |x|(|x| ) < 0
c3

(14.5)

if |x| > cc34  =: r. We want to claim that a level set which touches the ball Br is an attractive
invariant set, but not the ball itself. Consider then max V (x) =: cM and the level set cM := {x :
|x|=r

V (x) = cM }, the smallest level set of V which contains the r ball Br , i.e. |x| r V c2 r2 .
Outside of this set V is decreasing, and once we enter this set we will remain there. If we want
q a
2
ball which is attractive, then take r = maxxcM ||x||. In other words, V c2 r |x| cc21 r.
q
Conclusion: eventually, |x(t)| cc21 cc34 . This is called ultimate boundedness: exponential or even
asymptotic stability is not preserved under non-vanishing perturbations.
More on such analysis later (input to state stability). Note: recall vanishing perturbations
looked like |g| |x| and non-vanishing looked like |g| . Near origin vanishing is a stronger
property, but far away its weaker. Therefore, we have the same conclusion for g such that g 
for small x and g |x| for larger x.

14.4

Converse Lyapunov Theorems for Asymptotic Stability

We will discuss but not prove the following theorems (the first is 4.16[1]).

Nonlinear Analysis Notes

40

Theorem 14.1 (Massera). x = f (t, x), x = 0 equiliberium, f is C 1 , for |x| r: f

x is bounded
uniformly over t and assume that the system is UAS: i.e. there is a KL such that |x(t)|
(|x0 |, t t0 ), for every |x0 | r0 such that (r0 , 0) r. Then there is a C 1 function V satisfying:
1. 1 (|x|) V (t, x) 2 (|x|)
2.

V
t

V
x f (t, x)

3 (|x|)

3. | V
x | 4 (|x|)
with 1,...,4 K on [0, r0 ].
If f = f (x) (time invariant) then V can be chosen independently of t.
Z
G(|( ; t, x)|)d where G is
Comments: Masseras construction of V looks like V (t, x) =
t

class K with some other properties.

= f + g, we get (if |g| < ),

Bound | V
x | 4 (|x|) useful for perturbation analysis: for x

Vf +g 3 (|x|) + 4 (|x|) < 0 if 3 (|x|) > 4 (|x|); if |x| r and |x| 31 (4 (r)) then V < 0.
The next one is Theorem 4.17[1], and is only for time invariant systems.
Theorem 14.2 (Kurzweil). x = f (x) asymptotically stable around 0 equilibrium and f is locally
lipschitz. Then there is a continuously differentiable function V (x) satisfying
1. V is positive definite
2.

V
x f (x)

3 (|x|) < 0 for all x 6= 0 (3 positive definite)

3. if x = f (x) is GAS, then V is radially unbounded

Comments; Kurzweils construction looks like V (x) = sup{g((t; x))k(t)} (recall that (t; x)
t0

is the solution at time t starting at state x). With this construction, there is no integration; it is
global and gives radially unbounded V for GAS while Masseras construction works only locally
around zero. On the other hand, Kurzweils construction does not give a bound on the gradient,
| V
x |.
Reference: L. Vu, D. Liberzon, in Sytems and Control Letters, 54 (2005), pp 405-416, available
at http://liberzon.csl.illinois.edu/research/comm_sys_jn.pdf

Nonlinear Analysis Notes

15
15.1

41

Lecture 15
Stability of Interconnected Systems: Small Gain and Passivity

We consider internal stability of x = f (x) as it relates to inputs, outputs, feedback, x = f (x, u),
y = h(x).
u

(+d1 )1
(+d2 )2 .
Consider system 1
Absolute stability problem: Special case 1 is LT I system
x = Ax + Bu
y = Cx

(15.1)

and 2 is static nonlinearity u = (y). Closed loop system: x = Ax + B(Cx) has much more
structure than x = f (x).

15.2

Small Gain Theorems

View the following system u y from some input space U to output space Y passing
through some dynamics . Fix initial condition x0 . Need to pick U and Y normed vector spaces.
Generally well take U = Lp and Y = Lq for some
1 p, q , p = q or p 6= q. For example,
Z

|u(t)|2 dt. Similarly, p = q = , then in L ,

if p = q = 2, then U = Y = L2 and ||u||2L2 =

||u||L = sup |u(t)|. In general, we have

t[0,)

sZ
||u||Lp =

||u||p dt

(15.2)

Definition 15.1. We will say that system has finite Lp to Lq gain > 0 if ||y||Lq ||u||Lp + c
where c 0 depends on choice of x0 and equals 0 when x0 = 0.
By convention, we take the infimum of all which work in the definition; we call it the induced
||y||Lq
gain, = sup
. More generally, could have ||y|| (||u||) with K .
u6=0 ||u||Lp
When p = q, instead of Lp to Lq gain, well say merely Lp gain. Most often used: L2 gain
(or just gain, by default its L2 ). Relation to stability: if u 0, then ||y|| is bounded for all x0
and in particular, if we set zero initial condition, then y 0. Under observability property, can
conclude that x 0.
Want: Lyapunov (sufficient) conditions for induced gain.
Lyapunov sufficient condition for finite L2 gain: suppose there is a function V such that

V |y|2 + 2 |u|2 . We claim that then the systems L2 induced gain exists and is at most .

Nonlinear Analysis Notes

42

Proof. Similar to proof of LaSalle-Yoshizawa. We integrate from 0 to t:

Z t
Z t
2
2
|u|2 dt
V (t) V (0)
|y| dt +

(15.3)

and rearranging:
Z

|y|2 dt 2

|u|2 dt + V (0) V (t) 2

|u|2 dt + V (0)

(15.4)

wherethe last inequality holds from positive definiteness of V . Now we take square roots, using
that a2 + b2 a + b (for a, b 0) to get
sZ
sZ
t
t
p
2
|y| dt
|u|2 dt + V (0), t
(15.5)
0

As this holds for all time, we can take the limit as t to get
s
Z
lim

where c =

|y|2 dt = ||y||L2 ||u||L2 + c

(15.6)

p
V (x0 ).

Alternative Lyapunov condition for finite L2 gain: L2 gain is less than or equal to if for
every > there is an  > 0 and V (x) positive definite and C 1 such that
V |y|2 + 2 |u|2 |x|2

(15.7)

.
Some properties/facts of induced gain:
1. A linear (possibly time varying) GES system has finite L2 gain and in fact finite Lp to Lq gains
for any p, q. To show this, use variation of constants formula, i.e. for x = A(t)x + B(t)u, y =
C(t)x we have
Z t
y(t) = C(t)[(t, t0 )x0 +
(t, s)B(s)u(s)ds]
(15.8)
t0

and GES iff ||(t, s)|| ce(ts) || for c, > 0.

2. For LTI systems, L2 gain equals sup ||G(i)||2 = sup max (G(i)) where G denotes the trans

fer matrix, and max denotes the maximal singular value.

3. For LTI systems with A Hurwitz, L2 gain iff Eq 15.7 holds.
y2

+u

+u

2
1
Consider u1 1 y
2 2
1 ; 1 , 2 act on L2 for simplicity (could have
considered in more generality 1 : Lp Lq , and similarly for 2 ). Assume the following

Nonlinear Analysis Notes

43

1. 1 has gain 1 : ||y1 || 1 ||e1 || + c1

2. 2 has gain 2 : ||y2 || 2 ||e2 || + c2
3. small gain condition: 1 2 1

Then the overall system has finite gain from
 
 
u1
y1
from
to
.
u2
y2

u1
u2

 
e1
to
or what amounts to exactly the same
e2

In particular, if there are no external inputs u1 and u2 , then y1 and y2 have finite norms.
 
 
u1
e1
Proof. Show finite gain from
to
; we have e1 = u1 y1 , and e2 = u2 + y1 , and we want
u2
e2
to show finite gain from u0 s to e0 s; then ||e1 || ||u1 || + ||y2 || ||u1 || + 2 ||e2 || + c2 where we use
the gain from 2 . Since e2 = u2 + y1 we can use triangle inequality again to get
||e1 || ||u1 || + 2 ||u2 || + 2 ||y1 || + c2

(15.9)

Now using the second gain condition for 1 , we have

||e1 || ||u1 || + 2 ||u2 || + 2 1 ||e1 || + 2 c1 + c2
Recall that 1 2 < 1 we can solve for norm of e1 by rearranging to get
1
(||u1 || + 2 ||u2 || + 2c1 + c2 )
||e1 ||
1 1 2

(15.10)

(15.11)

which is exactly what we wanted to show, for e1 . A similar calculation proves the claim for e2 and
combine to get
 
 
e
u
|| 1 || || 1 || + c
(15.12)
e2
u2

Special case of linear system + static nonlinearity, absolute stability scenario: u 1 y

x = Ax + Bu
u, with 1 :
and is confined to region determined by two lines y = ax.
y = Cx
Assume that LTI system 1 has finite L2 gain and that static nonlinear satisfies sector condition
|(y)| 1 |y| for some > . The the closed loop system
x = Ax + B(Cx)

(15.13)

is asymptotically stable.
Proof. Recall: because 1 has gain , such that Equation 15.7 holds. For closed loop system, plug
in u = (y) with |(y)| 1 |y| and we get
1
V |y|2 + 2 2 |y|2 |x|2 < 0, x 6= 0

(15.14)

which implies asymptotic stability (and GAS if V is radially unbounded). This connects Equation
15.7 with external signals to standard Lyapunov condition.

Nonlinear Analysis Notes

16
16.1

44

Lecture 16
Passivity

u y with
x = f (x, u)
y = h(x)

(16.1)

Definition 16.1. A system is passive if there is a function V = V (x) such that

V uT y

(16.2)

We call V a storage function and uT y the suppy rate.

For example, electric circuit, u is the voltage, y the current, and uy the power inflow. Passive
system dissipates (does not generate) energy/power.
Integrating, V uT y, we get
Z

V (T ) V (0)

uT (s)y(s)ds

(16.3)

Z
Note that we could define passivity in terms of

uT y 0 (for x0 0) and look for storage

function.
Variant of passivity: if we have equality V = uT y, we say that the system is lossless. We
could have something like V u(u) + uT y, with u(u) > 0 for all u 6= 0, then this is called
input-strictly passive. Also, could have V y(y) + uT y, called output strictly passive. Finally,
with V W (x) + uT y with W positive definite, we call this strictly passive (or state strictly
passive).
Examples:
x = ax + u, a > 0
y=x
x R1

(16.4)

Try for a candidate storage function V (x) = 12 x2 so that V = ax2 + ux = ax2 + uy where the
second term is the supply rate and the term ax2 is our positive definite W term, so this is [state]
strictly passive.

Nonlinear Analysis Notes

16.2
y

45

Passivity of System Interconnections

u 1
1
2
Suppose that
1. 1 is strictly passive, so V W (x) + uT y (with x the state at 1 and W is positive definite
2. 2 is a state feedback function. For example, 2 can be a linear map (u = ky, with k 0).
Then for the closed loop system, V W (x) y T ky W (X) < 0 for all x 6= 0, which
implies that the closed loop system is asymptotically stable.
Alternatively, a static nonlinearity, u = (y), with sector condition y T (y) > 0 for all
y 6= 0 (or simply y(y) 0, y). For scalar u, y, this is the usual picture where the graph of
passes through third and first quadrants. Recall, for small gain we had a graph confined to
the region enclosed by two symmetric lines of positive and negative slopes ((y)| 1 |y|). For
closed loop, V W (x) y T (y) W (x) and therefore the closed loop system is again
asymptotically stable.
Note that if instead of strict passivity of 1 we had output strict passivity, then wed need
LaSalle observability to show asymptotic stability of closed loop, because it tells us that if we get
V W (y), we need to know something about how y relates to x (i.e. what is given by those
observability conditions).
Z
Can view static map u = (y) as passive because
uT y 0. More generally, we can have
any passive system as 2 and stability will follow from the next result.
+u1 (y2 )

+u

2
1 (x1 ) y
e2 2 (x2 ) y2 e1 1 Same setup as we had last lecture for
Small Gain Theorem.

Theorem 16.1. Feedback connection of two (strictly) passive systems is (strictly) passive.
T
T

V 1 (x1 ) W (x1 ) +
1 and
 e1 y
 similarly, V2 (x2 ) W2 (x2 ) + e2 y2 . Claim: closed loop is
u1
y1
strictly passive from
to
(note that we could also have chosen e for the output).
u2
y2

Proof. Need to show that there is a storage function V (x1 , x2 ) such that the derivative V
W (x1 , x2 ) + uT1 y1 + uT2 y2 with W positive definite. Try V = V1 + V2 (heuristically, we can
think of this as total stored energy). Then V = V 1 + V 2 W1 (X1 ) W2 (x2 ) + eT1 y1 + eT2 y2 ,
and we will define W (x1 , x2 ) := W1 (x1 ) + W2 (x2 ) which is positive definite. However, we need to
manipulate the expression to get ey in terms of uy; to do this we go back to the diagram, and write
e1 = u1 y2 , and e2 = u2 + y1 , and plugging into the previous expression, we get for eT1 y1 + eT2 y2 :
(u1 y2 )T y1 + (u2 + y1 )T y2 = uT1 y1 + uT2 y2
which is exactly what we wanted.

(16.5)

Nonlinear Analysis Notes

46

Corollary 16.1. If u1 = u2 = 0 (no external inputs), then instead of strict passivity, we have that
the closed loop system is asymptotically stable.
For once we get rid of input we have that the storage function V becomes Lyapunov function.

16.3

Absolute Stability Problem

u 1 ; Specifically,
1 LTI and 2 state nonlinearity, so we have 1 y 2
1 : x = Ax + bu, y = cT x
2 : u = (y)

(16.6)

a single input single output system with sector nonliniearity, which satisfies k1 y 2 y(y) k2 y 2
where 0 k1 < k2 . The absolute stability problem is to find conditions on (A, b, c, k1 , k2 )
guaranteeing global or local asysmptotic stability of all closed loop systems obtained in this way.
Aizermans conjecture (1949): system is absolutely stable if the linear system obtained from
u = ky is asymptotically stable for all k [k1 , k2 ]. The statement of the conjecture is not true.
Nevertheless, this [not true] conjecture was one of Aizermans most influential contribution to the
field.
This connects with passivity; for we already know that if LTI system 1 is (strictly) passive
and sector condition holds with k1 = 0 and k2 = , then the closed loop system is asymptotically
stable. (In fact, in that earlier result we did not require 1 to be linear). Still, there is a disconnect,
because this result only gives us conditions given k1 and k2 fixed at 0 and (which indeed is the
best possible pair for k1 and k2 ) but it tells us nothing about conditions on (A, b, c) for LTI system
to be passive. We want: given A, b, c when can we determine that there is a V without actually
looking for it. Good news, there are computable/checkable conditions, but unfortunately they
require working with the transfer function (i.e. we must step away from state space methods)
g(s) = cT (Is A)1 b.
Definition 16.2. The transfer function g(s) is called positive real if it satisfies the following two
conditions:
1. g(s) R for all s R.
2. Re(g(s)) 0 when Re(s) 0.
Note that this second condition is not easy to check. However, when all poles of g (eigenvalues
of A) are stable (i.e. in open left hand plane), it is enough to check the second condition along
the imaginary axis, i.e. for all s = i with R. Notice that this is reminiscent of Nyquist: the
second condition means that the Nyquist locus lies in the closed right half plane.
Definition 16.3. g(S) is strictly positive real if g(s ) is PR for some  > 0.
From 2 above its clear that SP R is stronger than PR. Connection with passivity.

Nonlinear Analysis Notes

17

47

Lecture 17

Recall from last time that we have for a transfer function

g(s) = cT (Is A1 b

(17.1)

is positive real (PR) if both

1. g(s) R for all all s
2. Re(g(s)) 0 (n the right hand plane (Re(s) 0)
when all poles of g have negative real parts, its enough to check the second condition for
s = i.
We say, also, that g(s) is strictly positive real (SPR) if g(s ) is P R for some  > 0. For
g(s) to be PR its necessary that its relative degree be zero or q, and g must have stable poles and
stable zeros (minimum phase).
q(s)
1
, with rel deg = deg(p) deg(q). For example, g(s) = s+a
, a > 0 (which comes
g(S) = p(s)
1
from x = ax + u, y = x), the first condition is satisfied. and for i = i, g(i) = i+a
have real
part 0 for all R. In fact, it is also SPR, using 0 <  < a.

17.1

Connection Between PR and Passivity

Theorem 17.1. Consider

x = Ax + bu
y = cT x

(17.2)

both controllable and observable (and therefore a minimum realization of g(s)). If A is Hurwitz and
g(s) = cT (Is A)1 b is SPR, then the system is strictly passive.
This is a consequence of KYP Lemma:
Lemma 17.1 (Kalman-Yakubovich-Popov). For LTI system as in the preceding theorem, if g(S)
is SPR, then there is a positive definite symmetric matrix P = P T > 0 which satisfies
1. P A + AT P = Q < 0 for some Q > 0
2. P b = c
The lemma is not trivial to prove, but well use it to show the theorem.

Nonlinear Analysis Notes

48

Proof of Theorem. Write V (x) := 21 xT P x with P as in the lemma; differentiating, we have

1
1
1
V = xT (P A + AT P )x + xT P bu = xT Qx + xT cu = xT Qx + uT y
2
2
2

(17.3)

which implies strict passivity by definition.

Last time, we had 1 y 2 = ()

u 1 with 1 strictly passive and [, )
sector implied that the system is asymptotically stable, because we had
V W (x) + uT y = W (x) y T (y) < 0, x 6= 0

(17.4)

In the absolute stability problem, we were looking at 1 linear. Now we have a method by means
of which we can determine passivity of 1 .
Proposition 17.1 (Passivity Criterion). IF g(S) is SPR, A is Hurwitz, and is a [0, ) sector
nonlinearity, then x = Ax b(cT x) is GAS- this is absolute stability.
This means, given 1 : (A, b, c) and 2 : (k1 , k2 )-sector, passivity criterion gives absolute
stability when g(s) = cT (Is A)1 b is SPR and k1 = 0, k2 = .

17.2

Loop Transformations

[7.1[1]]
If [k, -sector, then we can write (y) = ky + (y) where (y) [0, )-sector. In other
words, starting with overall system x = Ax b(cT x), if we can rewrite function as above, then
we can incorporate the ky term in linear part of the system, and the nonlinearity that remains will
be [0, )-sector. So e.g. x = x + (x) = x + kx + (x) = (1 + k)x + (x).
Another example: if [0, k]-sector, then

1 k1

[0, )-sector. Using this idea, we can

reduce [k1 , k2 ]-sector nonlinearities in feedback with an LTI system to [0, ]-sector nonlinearities
in feedback with another LTI system. After such a transformation, passivity criterion can be
applied, and it gives:
Proposition 17.2 (Circle Criterion). Given
1 : x = Ax + bu, y = cT x
2 : u = (y)

(17.5)

2 g(s)
T
1
2
assume that 1+k
1+k1 g(s) is SPR (g(s) = c (Is A) b) and that [k1 , k2 ]-sector (i.e. that k1 y
y(y) k2 y 2 ). Then x = Ax b(cT x) is absolutely stabe.

Nonlinear Analysis Notes

49

Figure 3: From http://www.engr.du.edu/vahid/EnPo2.html

Notes: this still relies on KYP lemma and gives a quadratic Lyapunov function. Also, A need
not be Hurwitz, but A bk1 cT must be.
1+k2 g
Geometric interpretation: Draw a circle with outer points 1/k1 and 1/k2 . Then 1+k
SPR
1g
means that the Nyqyuist plot of g lies outside this disk and encircles it m times CCW, where m is
the number of unstble eigenvalues of A.

Special cases: k1 = 0, k2 = , need g itself to be SPR; Nyquist plot of g must lie in RHP;
this matches previous condition in passivity criterion.
Or k1 = k2 = k; becomes linear, u = ky, and disk point 1/k, and Nyquist plot
encircles 1/k m times as in classical control and Nyquist criterion for stability of linear feedback
system.
Finally, consider k1 < 0 and k2 > 0 (so nonlinearity confined to [not necessarily symmetric]
lines of positive and negative slope). If k2 = k1 , then this is |(y)| < k|y| (we saw this in small
gain theorem). Then the circle is centered at origin. Now the Nyquist plot must be inside the disk;
LTI part must have gain 1/k.
Consider 1 y 2 u 2 where 1 is usual LTI system as above and similarly 2 is
[k1 , k2 ]-sector. Circle criterion gives quadratic Lyapunov function.
Proposition 17.3 (Popov Criterion). Suppose
1. g(s) = cT (Is A)1 b has one pole at 0 and all other poles in open LHP.
2. (1 + s)g(s) is PR for some 0

Nonlinear Analysis Notes

50

3. (0, ) sector: 0 < y(y) for all y 6= 0

Then x = Ax b(cT x) is absolutely stable.
In the statement above, is called the Popov multiplier. Were not testing stability of g
itself, but a linear function multiple of g(s). Note also that theres a strict inequality in the (0, )
sector. Lyapunov function constructed in proof is no longer quadratic:
T

cT x

V (x) = x P x +

(z)dz

(17.6)

and it depends on choice of nonlinearity !

Compare with circle criterion: V (x) = xT P x did not depend on .
There are lots of examples where circle criterion failes: e.g.

17.3

1+k2 g
1+k1 g ,

k1 = 0 1 + k2 g SPR.

ISS STability and Related Notions

[4.9[1]]
x = f (x, d)

(17.7)

x the state and d external input (disturbance and/or control). First we need to make sure that
x = f (x, d(t)) is well-posed (i.e. existence and uniqueness of solutions). We discussed this for
x = f (t, x). If we let f (t, x) := f (x, d(t)), we need conditions on f , like, that f is piecewise
continuous in t and locally lipschitz in x uniformly over t. What assumptions on f (, ) on d()
guarantee the desired conditions on f ?
Once we have these conditions, we will try to understand the notion of gain from d to x, but
not in the linear sense, but a notion more suitable for nonlinear systems.

Nonlinear Analysis Notes

18
18.1

51

Lecture 18
Input to State Stability

x = f (x, d(t)) where x is the state and d(t) is disturbance.

Define f(t, x) = f (x, d(t)) which satisfies the usual conditions, piecewise continuous in t and

f is Lipschitz in x uniformly in t. What properties of f guarantee this? To get the first, we need
both that d is piecewise continuous in t and f is continuous in d.
Then f (x, d()) is piecewise continuous because a continous function acting on piecewise continuous function is still piecewise continuous. On the other hand, piecewise continuity of d is
not enough since composition of two piecewise continuous functions is not necessarily piecewise
continuous.
To get the second, we need that f is locally Lipschitz in x uniformly over d, so that d locally
bounded:
|f (x, d(t)) f (x2 , d(t))| L|x x2 |
(18.1)
or we can assume that f (x, d) is locally Lipschitz as a function of (x, d).
Remark: 0-GES means GES under zero input (this is clear for lineary systems with A Hurwitz.)

18.2

Nonlinear Systems

Assume a system is 0-GES: x = f (x, 0) is GAS. Does this imply bounded input bounded state?, or
convergent input convergent state?
Implications and nonimplications:
1. 0-GAS does not imply BIBS. For example, x = x + dx for d = 2 is not bounded.
2. 0-GAS does imply CICS: x = x + dx, as d 0, then x 0 for this system only.
3. 0-GAS does not imply CICS: x = x + dx2 has finite escape time.

18.3

ISS Definition

Let x = f (x, d(t)).

1. 0-GAS for d 0, |x(t)| (|x0 |, t) for KL.

Nonlinear Analysis Notes

52

2. System Gain: ||x|| (||d||), for K ; Pick norm: ||d|| = sup0st d(s).
Causality means that x(t) depends on d(s) for s t. Rewrite gain condition as
|x(t)| (||d||[0,t] )

(18.2)

Caveat: only true for x0 = 0 (as we would otherwise need to add a term for nonzero x0 ).
Definition 18.1. ISS: A system is ISS if its solutions satisfy
|x(t)| (|x0 |, t) + (||d||[0,t] ), t

(18.3)

where KL and K
ISS implies BIBS, for d bounded implies that (sup(|d|)) is constant. Also, ISS implies CICS;
see Homework.

Nonlinear Analysis Notes

19
19.1

53

Lecture 19
Lyapunov Characterization ISS
x = f (x, d)

(19.1)

ISS definition: |x(t)| (|x0 |, t) + (||d||[0,t] ); note that this is a global property. We could have
given a local version as well.
Recall Passivity: V W (x) + uT y and L2 gain: V |y|2 + 2 |u|2 .
An ISS-Lyapunov function is a function V : Rn R which is positive definite and radially
unbounded iff 1 (|x|) V (x) 2 (|x|) for 1 , 2 K .
Now we ask that V :=

V
x f (x, d)

(|x|) + (|d|) with both , K .

Theorem 19.1. System is ISS iff there is an ISS Lypaunov function.

Example: x = x + xd x3 with both x, d R. Try V (x) = 12 x2 . Then V = x2 + x2 d x4 ;
we need to manipulate this expression to look like the definition of V for ISS Lypunov function; in
particular, we need to decouple x and d terms. In this case, we can easily do that by noticing that
the expression can be square completed: V = x2 d2 /4 + x2 d x4 + d2 /4 = (d/2 x2 )2 + d2 /4
x2 + d2 /4, and so we have = x2 and = d2 /4, and hence the system is ISS.
There is an equivalent way of writing the property for V , and it will be more convenient for
proving the theorem:
Lemma 19.1. V is an ISS Lyapunov function iff there is an 3 , K such that
|x| (|d|) V 3 (|x|)

(19.2)

Often is called the gain margin.

Proof. (): Write V (|x|)+(|d|) = 1/2(|x|)1/2(|x|)+(|d|) 1/2(|x|) if (|x|)
(|d|) which happens iff |x| 1 (2(|d|)) =: (|d|). SO we define 3 (|x|) = 1/2(|x|).
(): There are two cases. Case 1:, |x| (|d|); then V 3 (|x|), and were done automatically.
The second is more involved, |x| (|d|). Take this function V 3 (|x|) and build K
as follow:
(r) :=
max
{V (x) + 3 (|x|)}
(19.3)
|d|r,|x|(|d|)

which is nondecreasing; it can be upper bounded by a class K function.

Then V (x) + 3 (|x|) (|d|), and V 3 (|x|) + (|d|), as desired.

Nonlinear Analysis Notes

54

Proof of Theorem. (): This direction relies on converse Lyapunov Theorem for robust stability,
and we wont prove this direction.
(): We will show that system is ISS if there is a V satisfying the gain margin property.
Consider a ball of radius (||d||) in the x state space. The norm of d could either be on [0, ) or
on [0, t]. So either we take d to be bounded and take its sup value on the whole time interval or
we consider it on each time interval and the ball will be changing as time evolves. In the end, it
doesnt matter (on account of causality).
Consider an evolving trajectory of x; there are three (repeating) stages:
1. x(t) is outside (||d||) ball centered at zero.
2. x(t) is inside the ball.
3. x(t) is outside again.
We will construct and so that the ISS estimate holds at stage 1, 2, and 3 (and then by
extension for all future times as well). (Recall that we will be using: |x| (|d|) V 3 (|x|),
and we have 1 (|x|) V (x) 2 (|x|).)
Stage 1: |x(t)| (||d||) (|d(t)|), so the gain margin characterization applies here and we
have that V 3 (|x|) 3 (21 (V )); still, 3 21 K . We know that for systems evolving
on R, y = 3 21 (y), y(0) = V (x(0)), there is a KL, |y(t)| (y(0), t) (GAS) and by the
comparison principle, we get the same for V :
V (x(t)) (V (x0 ), t)

(19.4)

but we want a bound not on V but on x. Nevertheless, using the 1 , 2 bounds, we can easily
switch between them:
|x(t)| 11 ((V (x0 ), t)) 11 ((2 (|x0 |), t) =: (|x0 |, t)

(19.5)

Stage 2: when x(t) is inside the ball, |x(t)| (||d||) and the ISS estimate holds with =
(no function), so theres nothing to do in stage 2.
Stage 3: if x(t) exits the ball again, the situation we want to consider is the trajectory confined
to the smallest (sub)level set of V containing the ball, and find the radius of the smallest ball
containing this level set. If |x| (||d||) then V (x) 2 (||d||). If V (x) 2 (||d||), then
|x| 11 2 (||d||) =: (||d||). During stage 3, we have that
|x(t)| (||d||)

(19.6)

Theres another way (using the same approach as in stage 1) to arrive at the same result: we
consider x0 as a point on the boundary of the ball, and get |x(t)| ((||d||), 0) = (||d||); for
all t, |x(t)| (|x0 |, t) + (||d||) where the comes from stage 1 and from stage 3.

Nonlinear Analysis Notes

55

Remark, it would be more difficult to prove the theorem if we were working directly with the
definition of ISS, namely V (|x|) + (|d|).
More info on ISS can be found here http://citeseerx.ist.psu.edu/viewdoc/download?
doi=10.1.1.78.5962&rep=rep1&type=pdf

Nonlinear Analysis Notes

20
20.1

56

Lecture 20
Some applications of ISS to interconnected systems

ISS: |x(t)| (|x0 |, t) + (||d||[0,t] ). Cascade systems:

x

x = f (x)
z = g(z, x)

(20.1)

Question we ask is whether if x and z system are GAS and 0-GAS, respectively, then is the
overall cascades system GAS? The answer is no, with counter example
z = z + z 2 x
x = x

(20.2)

but we need to start with large enough initial condition x0 . x 0 indeed, but its too late: z blows
up in finite time. Solution: strengthen assumption: keep x system GAS, but make z-system ISS,
and then this does imply that the overall system is GAS.
Sometimes this claim is called the Cascade Theorem.
Proof. Since x system is GAS, we can write a bound
|x(t)| (|x0 |, t)

(20.3)

|z(t)| 2 (|z0 |, t) + (||x||[0,t] )

(20.4)

and for the z-system ISS implies that

These are the hypotheses, and we need to show that they imply that the overall system is GAS,
which means that we have to show

 

x
x(t)
|
| (| 0 |, t)
(20.5)
z(t)
z0
All that we need is a decay for z, and it is allowed to depend on both the initial condition of z and
x but it must decay with time.
We start by manipulating the norm of x in the z ISS equation:
||x||[0,t] = max |x(s)|
0st

(20.6)

and we know that |x(s)| 1 (|x0 |, s) for each s 0. Then

||x||[0,t] max 1 (|x0 |, s) = 1 (|x0 |, 0)

(20.7)

|z(t)| 2 (|z0 |, t) + (1 (|x0 |, 0))

(20.8)

0st

Then plugging in, we have

Nonlinear Analysis Notes

57

As we let time go to infinity, the first term decays but the second does not. So this approach is not
going to be fruitful.
Instead, consider |z(t)| 2 (|z0 |, t t0 ) + (||x||[t0 ,t] ). Trick: make both t0 and t go to infinity
but in such a way that they are coupled, so let t0 = t/2; then as t , t/2 as well, and
(this is the important part), |t t0 | also.
Now, what we have is that
|z(t)| 2 (|z(t/2)|, t t/2) + (||x||[t/2,t] )

(20.9)

Now we need to bound z at intermediate time, t/2, in terms of z at initial time 0:

|z(t/2)| 2 (|z0 |, t/2) + ||x||[0,t/2]

(20.10)

We break the latter term into two parts:

||x||[t/2,t] 1 (|x0 |, t/2)
||x||[0,t/2] 1 (|x0 |, 0)

(20.11)

The first term decays as t but the second does not. However, the second is inside the z(t/2)
function which is itself the 2 function which is already decaying, so this non-decay is not a problem.
We use the following fact: if K, then (r1 + rr ) (2r1 ) + (2r2 ), and we could continue
inductively to get
n
X
X
ri )
(nrj )
(20.12)
(
To see this e.g. r1 + r2 max{r1 , r2 } so (r1 + r2 ) max{(2r1 ), (2r2 )}.
Rest of the proof will be left as an exercise.
Ultimately, were not interested in only cascading systems; we want feedback. Thus we will
discuss how to incorporate feedback using the current framework for analysis.

20.2

Nonlinear ISS [Generalization of ] Small-Gain Theorem

Consider

z
x = f (x, z, u) x
z = g(z, x, v) z

(20.13)

Before in small gain we assumed that both individual subsystems had small gain and we figured
out how to combine them to get small gain of the whole system.
Suppose that the x system is ISS w.r.t. inputs both z and u:
 
z
|x(t)| 1 (|x0 |, t) + (||
|| )
u [0,t]

(20.14)

Nonlinear Analysis Notes

58

and that we have the same thing for z system w.r.t. inputs x and v:
 
x
|z(t)| 2 (|z0 |, t) + 2 (||
||[0,t] )
v

(20.15)

and that we have a small gain condition

1 2 (r) r, r 0

(20.16)

Note that the small gain condition is sometimes written as

(1 + 1 ) (2 + 2 ) id, some1 , 2 K

(20.17)

This is needed if we use + in ISS definition, but not if we use max.

The conclusion is then: The overall system is ISS w.r.t. (u,v) and in particular it is GAS if
there are no u and v.
The proof is similar to the cascade proof but theres a little bit more work; can be found in
paper Small Gain Theorem for ISS Systems and Applications by [Jiang-Teel-Praly, 1994] (here
http://link.springer.com/article/10.1007%2FBF01211469
An alternative proof is with ISS-Lyapunov functions V1 (x), V2 (x) V (x, z) := max{V1 (x), (V2 (z))}
for appropriately defined C 1 satisfying 1 2 id 2 11 (paper in which this technique is used can be found here http://www.sciencedirect.com.proxy2.library.illinois.
edu/science/article/pii/0005109896000519).

20.3

ISS related notions not covered in Class

Z

1. Integral input to state stability (iISS). Definition: (|x(t)|) (|x0 |, t)+

, K , KL. This is actually weaker than ISS; an example is

(|u(s)|)ds where
0

x = x + dx

(20.18)

which is not ISS because bounded input does not imply bounded state, but is iISS. (We will
see this in more detail in a to-be-distributed past exam)
2. ISS is equivalent to the existence of a function V satisfying V (|x|) + (|d|) where is
positive definite (not necessarily class K), and K .
3. Dual notions: output to state stability (OSS), which is written analogous to ISS:
|x(t)| (|x0 |, t) + (||y||[0,t] )
an observability like notion (detectability)

(20.19)

Nonlinear Analysis Notes

21
21.1

59

Lecture 21
Nonlinear Feedback Control
x = f (x, u), x Rn , u Rm

(21.1)

Goal is to achieve asymptotic stabilization to zero by feedback.

We will focus on systems affine in control, i.e. those of the form
x = f (x) + g(x)u

(21.2)

where G(x) is an n m matrix. Assume that f (0) = 0. State feedback means that u = k(x), we
will generally stipulate that k(0) = 0.
Example 1:
x = x2 + xu

(21.3)

Then a good choice of stabilizing feedback would be e.g. U = x 1. A better control, which
would satisfies u(0) = 0, would be u = x x2 to give x = x3 .
Example 2:
x = x + x2 u

(21.4)

We can stabilize by taking u = 2/x but at zero u(0) = and we cant easily fix it just by defining

2/x if
x 6= 0
(21.5)
u(x) =
0
else
(Discontinuity at zero is not removable).
Example 3:
x = x + x2 (x 1)u

(21.6)

2
A stabilizing feedback u = x(x1)
blows up not only at 0 but also at 1, so the control actually
ceases to exist at a finite time. This is worse than the situation in Example 2 since we cant really
approach where we want to be.

We want to work with scalar valued V (x) instead of x itself since usually this evolves in Rn
(and therefore is harder to analyze).
We can revisit the previous examples with V (x) =

x2
2 .

1.
V = x3 + x2 u
and using u = x x2 gives V = x4 < 0 for all V 6= 0.

(21.7)

Nonlinear Analysis Notes

60

2.
V = x2 + x3 u

(21.8)

and with u = 2/x we get V = x2 . However, since x3 is small compares to x2 for x 0, u

is discountinuous at 0.
3. We have V = x2 + x3 (x 1)u and x2 = 1 at x = 1, so theres no stabilizing feedback.
Returning the general system
x = f (x, u)

(21.9)

Let V : Rn R be C 1 , positive definite, and optionally radially unbounded (if global results are
desired). Then its derivative
V
V (x, u) =
f (x, u)
(21.10)
dx
Definition 21.1 (Artstein, 83). We call V a control Lyapunov function (CLF) if the following
holds:
inf V (x, u) < 0, x 6= 0
(21.11)
uU

Rm

with U
a set of admissible controls. In other words, for each x 6= 0 theres a u U such that
V (x, u) < 0.
For the affine system
x = f (x) + g(x)u

(21.12)

V a CLF function is equivalent to:

1. V (x, u) =

V
x f (x)

V
x g(x)u

2. CLF condition: For all x 6= 0, then

V
x f (x)

V
x g(x)

6= 0.

Equivalently, for all x 6= 0, either the first term is negative or we need to apply some control.
If U Rm is bounded, then the above statement no longer holds. Therefore, this equivalent
formulation is true only in the special case that U = Rm .
We say that a CLF V has a small control property (scp) if for every  > 0 theres a > 0 such
that for all x B n (0, ) there is a u U B m (0, ) such that V (x, u) < 0 (this concept describes
what we need for k(0) = 0 and k to be continuous at 0).
Recall: Example 2 where x = x + x2 u and V = x2 /2, V = x2 + x3 u. V is CLF but does not
have scp (easy to see).
Theorem 21.1 (Artstein,Songtag). Consider nonlinear system affine in u
x = f (x) + g(x)u

(21.13)

and suppose there is a CLF function V . Then there is an asymptotically stabilizing feedback u =
k(x) C 1 away from x = 0. If, moreoever, V has the scp, then k can be chosen continuous also at
zero with k(0) = 0.

Nonlinear Analysis Notes

61

Proof idea: V a CLF means by definition that

inf V (x, u) < 0, x 6= 0

uU

(21.14)

so that for all x 6= 0, theres a u such that V (x, u) < 0. Then paste these controls together to get
u = k(x) a stabilizing feedback.
What we want to ensure is continuous selection of u w.r.t. x satisfying V (x, u) < 0, i.e. the
resulting functioned obtained by pasting these values together is C 1 except possibly at zero.
Example (in paper by Sontag and Sussmann, 1980):
x = x[(u 1)2 (x 1)][(u + 1)2 + (x 2)]

(21.15)

Note that this system is not affine in control, so if it doesnt agree with the conclusion of the
theorem that is no contradiction because the requisite conditions are not the same.
Claim: V (x) = x2 /2 is CLF. To see this take
V = [(u 1)2 (x 1)][(u + 1)2 + (x 2)]

(21.16)

which is negative exactly when one of the terms in square brackets is negative, i.e. when (u 1)2 <
(x 1) and (u + 1)2 + (x 2) > 0 or the other way around. Having the first negative gives a region
represented by parabola lying on its side with vertex at (x, u) = (1, 1) with arms going toward
x = + (the region is inside the arms). The other region (corresponding to (u + 1)2 + (x 2) < 0)
is another parabola lying in its side with vertex at (x, u) = (2, 1) and arms going toward x = .
We need for each point on the x axis to have a region of at least one of the parabolas directly above
it or below (or both). In other words, the projection of the union of both regions equals the entire
x axis. Indeed, in this case that holds, so V is a CLF. However, the parabolas are disjoint and
separated by a positive distance (i.e.
inf
||x y|| > 0), so there is no continuous stabilizing
xR1 ,yR2

feedback. Any stabilizing feedback will be discontinuosu at some nonzero x.

First a topological fact.
Artsteins Proof.
V
V
f (x) +
g(x)u
(21.17)
V =
x
x
has for each x 6= 0 a u which makes it negative. Take an arbitrary state x1 6= 0, and for it consider
the subset of U which work, i.e. U1 := {u U : V (x1 , u) < 0} =
6 . Take your favorite element

from it, call it u1 . Note that V is continuous in x (by assumption of V C 1 ). So this u1 for nearby
states will also work: i.e. V (x1 , u1 ) < 0 V (x, u1 ) < 0 for all x B(x1 , 1 ) a small enough ball
around x1 . (B is open.)
We play the same game, for x2 and get some other ball B(x2 , 2 ). Repeating this for each x
in the statespace Rn \ 0, we generate a cover B(x1 , 1 ), B(x2 , 2 ), . . .. Rn is locally compact so each
point has finitely many balls containing it (if there are more we can just throw them away).
n
Now we use the partition of unity. Namely, there are functions
X pj (x) from R \ 0 to [0, 1]
which are smooth and satisfy pj (x) = 0 for all x
/ B(xj , j ) and
pj (x) = 1 for all x Rn \ 0

Nonlinear Analysis Notes

62

(this is called a partition of unity subordinate to a given cover). See e.g. http://en.wikipedia.
org/wiki/Partition_of_unity for more information.
Using a partition of unity, we can define
k(x) :=

pi (x)ui

(21.18)

for x 6= 0 and k(0) = 0. This is C 1 away from zero, and stabilizing since at each point were taking
a convex combination of stabilizing feedback values.
Now we need to show that V < 0:
V (x, u) =

V
x f (x)

V
x

pj (X)g(x)ui

X
j

V
pj (x)
(f (x) + g(x)ui ) < 0
x

, x 6= 0, B(xj )

Note that we use convextity of f + gu with respect to u.

(21.19)

Nonlinear Analysis Notes

22

63

Lecture 22
x = f (x)g (x)u

(22.1)

A CLF function V satisfies inf V (x, u) < 0 for all x 6= 0, with

uU

V
V
V (x, u) =
f (x) +
g(x)u
x
x
so a CLF means that if

V
x g(x)

(22.2)

= 0 then the drift term is stable.

Theorem 22.1. If there is a control Lyapunov function V , then there is stabilizing feedback u =
k(x), and if V has scp property, then k can be continuous at 0.
Last time we proved it using partition of unity; the proof was non constructive and abstract
(though quite nice, nonetheless). The next one, due to E. Sontag, gives an explicit formula for
u = k(x).
v
x

f (x); this is called

if
Lg V 6= 0
else

(22.3)

Sontags Proof. To start we need to establish some notation: let Lf V :=

the Lie Derivative of f , and similarly, Lg V : V
x g(x).
Now define
(
k(x) :=

Lf V +

(Lf (V ))2 +|Lg V |4

(Lg V )T
|Lg V |2

We have to show two things: first that this feedback actually stabilizes the system, which can
be seen by plugging in for V and seeing that its negative. Secondly, must show that its continuous.
To see stability, we consider by hypothesis only when Lg V 6= 0 (since we already assumed
that the drift is stable):
p
q
T
(L
V
)
L
V
(L
V
+
(Lf V )2 + |Lg V |4
g
g
f
V cl = Lf V + Lg V k(x) = Lf V
=

(Lf V )2 + |Lg V |4 < 0,

||Lg V ||2
(22.4)

for all nonzero x, where the last inequality holds because Lf V and Lg V are not simultaneously
zero except at the origin. Note that other than the choice of k, theres nothing smart going on in
the verification of stability.
Next we show
write k(x) = (a, b)(Lg V )T where a := Lf V and b = |Lg V |2 .
( regularity;

a+ a2 +b2
if
b 6= 0
b
Then (a, b) =
We claim that is smooth and in fact real analytic on
0
else
the set {(a, b) : b > 0 or a < 0} (exclude fourth quadrant). Indeed, satisfies the following implicit
relation
b2 2a b = 0
(22.5)

Nonlinear Analysis Notes

64

which can be checked easily by computation. Taking F (a, b, ) := b2 2a b, considering its

derivative

F
b 6= 0
2 a2 + b2 if
(22.6)
= 2b 2a =
2a
else

By Implicit Function Theorem, (a, b) is C 1 (and actually smooth analytic as well) function.
To summarize, we have mappings x 7 a and 7 |Lg |2 continuous and mapping of these to
(a, b) which is C , which by composition implies that u = k(x) depends continuously on x.
Analysis of scp can be found in Sontags book.
Remark: theres an interpretation of this formula in terms of linear quadratic optimal control:
u = K(x) achieves Lf V + Lg V k < 0 or a + Bk < 0 where B := Lg V (b = |B|2 ). This all depends
on x, but pretend that x is fixed. In other words, consider an auxiliary scalar linear system
z = az + Bv

(22.7)
Z

where z is the state and v the control. Introduce cost J() :=

v 2 + bz 2 . Optimal control is

obtained from Riccati equation:

bp2 2ap b = 0

(22.8)

and the optimal control is k = B T p. This LQR cost forces control to be small when b is small,
which forces control to be continuous.
Example 1 from last time
x = x2 + xu

(22.9)

where last time we had that u = x x2 x = x3 or u = x 1 x = x but u is not zero

at zero. Using V = x2 /2, Lf V = x3 and Lg V = x2 . Sontags formula gives
(

3
6
8
x + xx4 +x x2 if
x 6= 0
u=
(22.10)
0
else
This formula simplifies to
u = x |x|

p
1 + x2

(22.11)

The closed loop system is then

p
x = x2 sgn(x) 1 + x2

(22.12)

This control gives a balance between the two options we previously had, namely between control
speed and control effort.

Nonlinear Analysis Notes

22.1

65

More on Nonlinear Control Design: Backstepping

Given x = f (x) + g(x)u or even a more general system x = f (x, u), how to find a CLF? Affine
system + CLF gives (by Sontags formula) stabilizing feedback. We already know that given
x = f (x) trying to find a Lyapunov function can be difficult. On the other hand, CLFs are more
flexible: they can be a Lyapunov function, so long as you choose a control in the right way. (Recall
that x 6= 0u such that Lf V + Lg V u < 0, which is equivalent to requiring that Lf V < 0 when
Lg V = 0). For we only need V to be Lyapunov when the control Lie term vanishes (so in other
words, we do not require that the CLF is Lyapunov everywhere.)
Backstepping is a tool for generating CLF for control systems with specific structure (Meilakhs,
1975; Morse, 1976; Krstic-Kannelakopoulos-Kokotovi book, 1990): take
x = f (x) + g(x)u

(22.13)

Suppose there exists a stabilizing feedback u = k(x) and corresponding Lyapunov function V (x).
Consider an augmented system,
  

x
f (x) + g(x)
(22.14)
=
u

Can consider this a courser model of a finer system, e.g. kinematic and dynamic system. For
example
Z

x
u

x = f (x) + g(x)u

(22.15)
as opposed to u x = f (x) + . . .
For the augmented system we want to find a CLF Va (x, ) and a stabilizing feedback u =
ka (x, ). The idea is to start with xsystem (simple) and add integrators (or more general dynamics) to build up to the more complicated system. For example
x = x2 + xu

(22.16)

   2

x 1
x1 + x1 x2
=
x 2
u

(22.17)

Whcih we could augment as

so what we would do is consider first x 1 = x21 + x1 u pretending that x2 is u and then go back to
the original system. Moreover, as well see, we dont need to have pure integrators, it is possible to
have e.g.
x 1 = x21 + x1 x2
x 2 = x3 + . . .
(22.18)
x 3 = u

Nonlinear Analysis Notes

23
23.1

66

Lecture 23
Backstepping

Suppose for system

x = f (x) + g(x)u

(23.1)

that we have stabilizing feedback u = k(x), k(0) = 0, and corresponding Lyapunov function V (x)
V
V
f (x) +
g(x)k(x) W (x) < 0 x 6= 0
x
x

(23.2)

We take the augmented system

x = f (x) + g(x)
= u

(23.3)

Then we claim that

( k(x))2
(23.4)
2
is a control Lyapunov function for the augmented system; a stabilizing feedback can be defined by
u = ka (x, ) = k 0 (x)f (x) + k 0 (x)g(x) V
x g(x) + k(x); then we compute the derivative
Va (x, ) = V (x) +

V
0
0
V a = V
x f (x) + x g(x) + ( k(x))(u k (x)f (x) k (x)g(x)) =
V
V
V
0
0
x f (x) + x g(x)k(x) + ( k(x))( x g(x) + u k (x)f (x) k (x)g(x))
0
0
W (x) + ( k(x))( V
x g(x) + u k (x)f (x) k (x)g(x))
2
W (x) ( k(x)) < 0, (x, ) 6= (0, 0)

(23.5)

To see the last inequality, suppose that W (x) ( k(x))2 = 0, then W (x) = 0 x = 0
k(x) = 0 = 0.
The control function may not be unique but this technique gives us a systematic way of
generating a CLF.
We can extend this to
x = f (x) + g(x)
= F (x, ) + G(x, )u, G 6= 0

(23.6)

Exercise: carry out procedure to make what we had above work for this (same process). Systems
that look like this are said to be in strict feedback form.

23.2

ISS Disturbance Attenuation

From the original system

x = f (x)

(23.7)

x = f (x) + g(x)u

(23.8)

we took investigated both control problem

Nonlinear Analysis Notes

67

and disturbance problem

x = f (x) + g(x)d

(23.9)

where we had CLFs and ISS-LF, respectively. Now we want to combine these and consider control
design:
x = f (x) + g1 (x)d + g2 (x)u
(23.10)
We assume here that dynamics are affine w.r.t. both control and disturbance. The problem we
want to consider is how to choose u such that the closed loop with exogenous disturbance d will
still have ISS property, well call it ISS-CLF.
Recall that for x = f (x, d), V is an ISS Lyapunov function if
V
f (x, d) (|x|) + (|d|), , K
x

(23.11)

This suggests that we define for a system x = f (x) + g1 (x)d + g2 (x) that V is an ISS CLF if
inf

uU

V
V
V
f (x) +
g1 (x)d +
g2 (x)u (|x|) + (|d|)
x
x
x

(23.12)

From this we hope that we can find the desired feedback using Sontags universal formula:
inf a(x) + B(x)u < 0, x 6= 0 u = k(x) = (a, B)

(23.13)

inf a(x, d) + B(x)u (|x|) + (|d|)

(23.14)

uU

Now we have
u

The functional dependence of a on d requires that u = k(x, d) which by hypthesis this is not allowed
(generally we dont have control access to disturbances). So before applying Sontags formula, note
that the above ISS-CLF condition can be equivalently rewritten according to gain margin condition:
|x| (|d|) inf
u

V
V
V
f (x) +
g1 (x)d +
g2 (x)u 3 (|x|) < 0, x 6= 0
x
x
x

(23.15)

where , 3 K . But we still need to find a way to deal with the disturbance term inside.
Observer, however, that this condition is equivalent to:

 

 V
  V  1
V



 (|x|) + V g(x)u 3 (|x|)
inf
f (x) + 
g(x) + 
(23.16)
u x
x
x 
x
Hence we are choosing our disturbance to be bounded within a ball of a specified size, and the
worst case is then


 V

V

(23.17)
g1 (x) d = 
g1 (x) 1 (|x|)
x
x
Now, finally, we can apply Sontags formula to find an input to state stabilizing feedback
u = k(x).

Nonlinear Analysis Notes

23.3

68

Advanced Topics: Perturbation Theory and Averaging

We are going to generalize earlier results of vanishing and nonvanishing perturbations, which as a
corollary will give a result on continuity of solutions on infinite intervals. Well then apply these
results to averaging (this material is 9.1-9.4 and 10.3-10.4,10.6[1]).
Consider again a system of the form
x = f (t, x) + g(t, x)

(23.18)

where f is the nominal dynamics and g the perturbation. Goal: assuming that x = f (t, x) has
nice behavior (in a sense we will later make precise), we want to be able to say something about
the perturbed system.
Hypotheses: we assume that the equilibrium x(0) = 0 is exponentially stable for the nominal
system. Additionally, we need to say something about the disturbance, namely that it is bounded
as
|g(t, x)| (t)|x| + (t), t, x
(23.19)
where and are continuous, nonnegative and is bounded. More assumptions on will come
later.
This combines and generalizes earlier system descriptions which weve seen:
1. vanishing perturbations: |g(t, x)| |x| with  small. (In the current situation, wed have
(t)  and 0
2. Nonvanishing perturbation: |g(t, x)| , and this corresponds to 0 and |(t)| .
Since x = f (t, x) is exponentially stable, by converse Lyapunov theorem for ES, there is a
Lyapunov function V such that
c1 |x|2 V (t, x) c2 |x|2
V
V
2
+
x f (t, x) c3 |x|
t

 V  c4 ||x||, c1 , . . . , c4 > 0
x

(23.20)

where these inequalities hold for all x B(0, r), a ball around the origin of radius r. If the system
is GES, then take r = . The main idea is to differentiate V along solutions of the perturbed
system, and use assumptions to get some sort of decay; it wont be exponential, nor necessarily
attractive to zero.

Nonlinear Analysis Notes

24

69

Lecture 24

Consider perturbed system

x = f (t, x) + g(t, x)

(24.1)

and assume that

1. x=0 is exponentially stable for x = f (t, x)
2. |g(t, x)| (t)|x| + (T ), for , 0, continuous and boudned
By converse Lyapunov for exponential stability, we have
c1 |x|2 V (t, x) c2 |x|2
V
V
2
+
 x f (t, x) c3 |x|
t
 V  c4 |x|
x

(24.2)

where these hold locally, i.e. |x| r some r > 0, and c1 , . . . , c4 0.

Then the derivative of V for X = f + g is
V
t

V
x f (t, x)

V
x g(t, x)

 

c3 |x|2 +  V
x |g(t, x)|
2
c3 |x| + c4 (t)|x|2 + c4 q
(t)|x|

Define W (t) :=

[ cc23

c4
c1 (t)]V

+ c4 (t)

(24.3)

V
c1

p
V (t, x(t)) along solutions, so then the derivative of W is given by

V
= 2
V
12 ( cc32

c4
c1 (t))W

c4

2 c1 (t)

(24.4)

By the comparison principle, using




(s)ds

(24.5)

(t, )( )d

(24.6)

c3
c4
(t, t0 ) = exp
(t t0 ) +
2c2
2c1

t0

as the state transition matrix from t0 to t, we have

W (t) (t, t0 )w(t0 ) +
Since

c1 |x| W

c4

2 c1

t0

c2 |x|, we have as a result

r
Z t
c2
c4
|x(t)|
(t, t0 )|x0 | +
(t, )( )d
c1
2c1 t0

Need a bound on ; assume that

Z

(24.7)

(s)ds (t t0 ) + , ,  0
t0

(24.8)

Nonlinear Analysis Notes

Then
(t, t0 ) exp([

c4
c4
c3

](t t0 )) exp(
)
2c2 2c1
2c1

70

(24.9)

We want  to be small enough so that the expression in brackets is positive, which would give
exponential decay in time: we define this right hand side to be
e(tt0 ) , > 0, 1

(24.10)

where  is small enough, namely  < c1 c3 /(c2 c4 ).

Plug this into bound for x, and get
r
Z
c4 t (t )
c2
(tt0 )
|x0 |e
+
e
( )d
|x(t)|
c1
2c1 t0

(24.11)

But must make sure that |x| r. Therefore, we need to make sure that the last expression
has a particular bound; we need:
r
c2
c4
(1 e(tt0 ) ) sup (t) r
(24.12)
|x0 |ed(t=t0 ) +
c1
2c1
tt0
And it can be shown that the left hand side is bounded by
r
c2
c4
max{
sup((t))}
|x0 |,
c1
2c1

(24.13)

because et + b(1 et ) looks like a convex curve from a to b.

So we need to assume that x0 and () are small enough, and that
r
r c1
|x0 |
c2
(t)

2rc1
, t
c4

(24.14)
(24.15)

Then the bound refeq:bnd is true for t t0 .

We have thus proved the Perturbation Lemma:
Proposition 24.1. Let x = 0 be an exponential equilibrium of nominal system x = f (t, x), let
V = V (t, x) be a Lyapunov function provided by the converse Lyapunov theorem for exponential
stability
c1 |x|2 V (t, x) c2 |x|2 c1 , c2 0
(24.16)
V c3 |x|2
c3 0
| V

c
|x|
c

0
4
4
x
locally for |x| r, r > 0. Suppose further that
g(t, x)| (t)|x| + (t)

(24.17)

Nonlinear Analysis Notes

71

with , 0 and satisfying

Z t
c1 c3
, 0
(s)ds (t t0 ) + , 0 <
c2 c4
t0

(24.18)

Suppose that
r
|x0 |

c1
2rc1
, (t)
, t
c2
c4

(24.19)

Then: the solutions of x = f (t, x) + g(t, x) satisfy:

r
Z
c4 t
c2
(tt0 )
|x0 |e
+
( )d, t t0 , > 0, 1
|x(t)|
c1
2c1 t0

(24.20)

where and are given in the above proof (as is the proof of this statement)
c4
c3
c4

 > 0, := e 2c1
2c2 2c1

:=

(24.21)

For the GES case x = f (t, x), r = for every x0 , ().

Proof. Everything that precedes the statement of the proposition.
Vanishing perturbations: no (t) implies that the integral term in our bound disappears, and
we get exponential stability of the perturbed system. Before, we had |(t)| , but this is a little
bit more general ( allows to be large for a finite time period).
Z

(s)ds (t t0 ) + :

Some classes of () satisfying

t0

1. |(t)| , t (as vanishing perturbation earlier)

Z t
2. Any L1 function
(s)ds
t0

Z
3. Any function converging to zero (becomes  after some time, handles
Z
happens, and (t t0 ) upper bounds after)

before that

Example:
x = [A(t) + B(t)]x

(24.22)
Z

where x = A(t)x is GES. Then GES is preserved if B(t) 0 or

case (t) B(t).

||B(t)||dt < because in this

0

Nonlinear Analysis Notes

72

For the perturbation Lemma to give useful bound in presence of (t), we need to know someZ t
thing about
e(t ) ( )d . This is response of stable scalar linear system
t0

z = z + (t)

(24.23)

forced by (t) with initial condition z0 = 0.

We know for example that (t) bounded implies that this is bounded. Also if (t) 0 then
solution also 0. Staring at the bound in the proposition long enough convinces one of these
statements.
In this prolem, we are given the structure of the system x = f +g, but in a general problem that
we may be trying to figure out, we need to decide/find for ourselves exactly how to artificially
set


0 1
things up to apply this result (A similar principle applied in the homework for e.g. A =
,
0 1


 
0
1
0
and u = (x1 ), where A is not Hurwitz, but it can be made so by taking A =
B=
1
1
for positive , and u = (x1 ) + x, basically just adding zero.)

Nonlinear Analysis Notes

25

73

Lecture 25

Lemma 25.1 (Perturbation Lemma). x = f (T, x) exponentially stable around x = 0, |g(t, x)|
Z t
(s)ds (t t0 ) + , wiht , |x0 |, (t) small enough. Then solutions of
(t)|x| + (t) with
t0

x = f (t, x) + g(t, x) satisfy

|x(t)| c|x0 |e(tt0 ) + d

e(tt0 ) ( )d

(25.1)

t0

25.1

Continuity of Solutions on Infinite Time Horizon

x = f (t, x), x(t0 ) = x0
y = f (t, y) + g(t, y), y(t0 ) = y0

(25.2)

Theorem 25.1 (Continuity on Infinite Time Horizon). Let f C 1 and f

x locally LIpschitz in x
uniformly in t, |g(t, y)| , and x = 0 exponentially stable equilibrium of x = f (t, x). Then for
each compact set inside the exponentially stability region for x = f (t, x), if x0 is in this compact
set and if |x0 y0 | and are sufficiently small, then solutions x(t) and y(t) of x = f (t, x) and
y = f (t, y) + g(t, y) satisfy the following bound
|x(t) y(t)| ce(tt0 ) |x0 y0 | + , t 0

(25.3)

where c, , > 0.
Proof. In order to prove this we will use the perturbation lemma. Consider the error e := y x
and we want to derive a differential equation of e in the form nominal +perturbation.
We have
e
=
y x = f (t, y) + g(t, y) f (t, x)
= f (t, e) + (t, e) + g(t, y)
= f (t, x(t) + e) f (t, x(t)) f (t, e) + f (t, 0)
f
= f
x (t, x(t) + 1 e)e x (t, 2 , e)e

(25.4)

where we want to express the second line as a function of e (plus perturbation), (t, e) := f (t, x(t)+
e) f (t, x(t)) f (t, e), f (t, 0) = 0 (since 0 is equilibrium of x = f ), and the fifth line follows from
the mean value theorem with 1 , 2 [0, 1]. Since f
x is Lipschitz, uniformly in time, we have
|(t, e)| L(|e| + |x|)|e|

(25.5)

|(t, e) + g(t, y)| L(|e| + |x|)|e| +

(25.6)

for L 0.
Then Perturbation term is

which is exactly in the form we need (t)|e| + (t) as in the Perturbation Lemma, and where
(t) = L(|e| + |x|), (t) = . Recall that we assumed that the nominal system is exponentially

Nonlinear Analysis Notes

74

stable. Recall that x(t) 0 exponentially fast and e(0) is small . If e0 is small enough, for arbitrary
Z t
small, > 0, we have
(s)ds (t t0 ) + , where
t0

|x(s)|ds

(25.7)

t0

where |x(t)| < for all g T (true so long as |e(t)| r where r comes from Perturbation lemma).
Now apply the perturbation lemma,
Z t
(tt0 )
e(t ) d
(25.8)
|e(t)| c|e0 |e
+d
t0

where the integral term is less than r and the bound is valid for all t 0.

25.2

Periodic Perturbation of Autonomous Systems[Section 10.3 Khalil]

Take
x = f (x) + g(t, x, )

(25.9)

 > 0, f, g C 1 w.r.t. x (see Khalil for more details). Assume that x = 0 is exponentially stable
equilibrium for the nominal system x = f (x), that g is bounded and T -periodic in t for T > 0, i.e.
g(t + T, x, ) = g(T, x, ) for every t, x, .
Previous analysis: if x0 and  are small, then x(t) some neighborhood of zero whose size is
proportional to . What happens inside this neighborhood? When can x(t) be periodic?
Observations: define the following map: P (x) := (T, 0, x, ) the solution at time T at x at
time 0, with value  on right hand side of the system; P : Rn Rn .
Lemma 25.2. Perturbed system has a T -periodic solution iff the equation x = P (x) has a solution,
say x = p (fixed point).
Proof. Obvious.
Lemma 25.3. There are positive numbers k and  such that X = P (x) has a unique solution in
{x : |x| k|epsilon|} for every || <  .
Proof Idea. Use implicit function theorem and use fact that A := f
x |x=0 is Hurwitz (from exponential stability assumption and Lyapunovs 1st method). (Remember A, were gonna need it later.)
For x = 0 = 0, we have a trivial periodic solution x 0.
Lemma 25.4. If x(t, ) is a T -periodic solution of the perturbed sytsem satisfying the bound
|x(t, )| < k||, (where k is from lemma 2), then the solution is exponentially stable.

Nonlinear Analysis Notes

75

Proof Sketch. Set error variable z := x x(t, ) iff x = z + x(t, ), then
z = x x = f (x) + g(t, x, ) f (x) g(t, x, )
= f (z + x) f (x) + [g(t, z + x, ) g(t, x, )]
= f(t, z)

(25.10)

Now linearize about z = 0 to get

f
z |z=0

g
= f
x |z=0 +  x |=0
g
= A + [ f
x (x) A] +  x (t, x, )
g
=  x (t, x, )

(25.11)

0

0
where the first term vanishes by continuity of

f
x

Then

f
g
(x) A] +  (t, x, )
x
x
and the second term as we previously saw is a vanishing perturbation.
z = Az + [

(25.12)

By result on vanishing perturbations, linearized z system is exponentially stable and by Lyapunovs first method, original z system is exponentially stable.
Theorem 25.2. There exists k,  such that for every || <  , perturbed system has a unique
T -periodic solution x(t, ) such that |x(t, )| < k||, and this solution is exponentially stable.
Note that if g(t, 0, ) = 0, then the perturbed system has equilibrium at zero. Then x 0 is
this unique periodic solution, and its still exponentially stable.
Proof. The previous three lemmas.

Nonlinear Analysis Notes

26

76

Lecture 26

26.1

Averaging Theory
x = f (t, x)

(26.1)

f is T -periodic in t, f (t + T, x) = f (t, x) for all t, x and  small, slow system response.

Suppose we have a small bandwith system (acts like low pass filter, i.e. high frequency input
doesnt make it through). Well define an average for the system and approximate the behavior
of original system by that of average system.
Define average
Z

1
fav (x) :=
T

f (, x)d

(26.2)

Average system x = fav (x) an autonomous system. More generally, given a system
x = f (t, x, )

(26.3)

we define the average

1
fav (x) =
T

f (, x, 0)d

(26.4)

The goal is to study how the autonomous average system approximates the behavior of the
original time varying system.

Let h(t, x) := f (t, x, 0) fav (x), which is T periodic in t and has mean zero. Also let u(t, x) :=
h(, x)d which is also periodic in T . (Think sin(t) and 1 cos(t)). Hence u(t, x) is bounded

for all t and for all x in a compact set, and

u
u
= h(t, x),
=
t
x

Z
0

h
(, x)d
x

(26.5)

These are T periodic, zero average, and also bounded for all t and all x in a compact set.
Change of variables: x = y + u(t, y), y a new state variable, then
x = y + 

u
u
(t, y) +  (t, y)y
t
y

(26.6)

where x = f (t, y + u, ) and so we get

(I + 

u
u
(t, y))y = f (t, y + u, )  (t, y)
y
t

(26.7)

Now plug the definition of h in to get

(I +  u
y (t, y))y = f (t, y + u, ) f (t, y, 0) + fav (y)
= fav (y) + p(t, y, )

(26.8)

Nonlinear Analysis Notes

77

where we define p(t, y, ) := f (t, y + u, ) f (t, y, 0) (note that this is T periodic), which we
can rewrite as
p(t, y, ) = f (t, y + u, ) f (t, y, ) + f (t, y, ) f (t, y, 0)
f
= f
y (t, y + 1 u, )u +  (t, y, 2 ), 0 1 , 2 1

(26.9)

using the mean value theorem in the second line.

Consider the left hand side above:
I +

u
(t, y)
y

(26.10)

is nonsingular for  small enough because u

y is bounded over any compact set of x. So we can
u
1
consider its inverse (I +  y (t, y)) = I + O(), where O() means order , i.e.
|O()| c||

(26.11)

(I + A)1 = I A + 2 A2 . . .

(26.12)

as  0.

We can bring the implicit differential equation for y to the form

y = fav (y) + 2 q(t, y, )

(26.13)

where q is T periodic in t and 2 q term collects all the terms of order > 1 in . This is a perturbation
of the average system which is simply
y = fav (y)
(26.14)
We arent done yet, because we dont want to keep  explicitly on the right hand side. After all,
previously we didnt have it. The trick is to rescale the system, divide time by  to get a slow time
scale, w.r.t. the  will disappear. Another change of variables: s = t, where s is a new time, and
d
t = 1 s, and we get a differential equation with ds
:
dy
1
s
= y = fav (y) + q( , y, )
ds



(26.15)

Now q, which was T -periodic in t is now T periodic in s. We bring in three previous results to
arrive at the following three claims:
1. If y is a solution of the average system
dy
= fav (y)
ds

(26.16)

then by continuity of solutions on finite intervals w.r.t. initial conditions and perturbations,
if y(0, ) and y(0) are within O(), then for  small enough, we have
|y(s, ) y(s)| = O()
on some interval, say s [0, b].

(26.17)

Nonlinear Analysis Notes

78

Remark: in original time t, this interval is for t [0, b/] which is, though finite, pretty large.
Also, y is not the original variable so it may seem that this result is not what we want. But
recall the relation x = y + u(t, y) where u is T periodic and bounded and multiplied by ,
the same claim/approximation result holds for x and x of original system and its average. To
be precise: d(x, y) and d(x, y) are of order  so we can easily pass between the statement for
y and the corresponding statement for x. For the following two claims the same holds true.
2. Lets now assume that the average system
y
= fav (y)
ds

(26.18)

has y = 0 as an exponentially stable equilibrium, then by continuity on infinite horizon

intervals, as long as y(0) is in a compact set contained in the region of exponential stability
the above O() approximation is valid for all t [0, ), again for  small enough.
3. Now we can use the fact that q is periodic. From a result in last lecture on the existence
of periodic solutions under periodic perturbations, still assuming that y = 0 is an exponentially stable equilibrium of the average system, we have that in some neighborhood of that
equilibrium, for  small enough, there exists a unique exponentially stable periodic solution
of the perturbed system. Period is T in the s scale and therefore T in the original t scale,
as expected.
Note: if moreover f (t, 0, ) = 0 (meaning that 0 is still an equilibrium of perturbed system), then the unique periodic solution in the above statement is this equilibrium, and it is
exponentially stable.

26.2

Examples

Consider the LTV system

x = A(t)x

(26.19)

x = Ax

(26.20)

and average system

where A =

1
T

A( )d , still has equilibrium 0. Suppose that A is Hurwitz, then that guarantees

0

that we can apply each of the above three results, so in particular, the behavior of the system above
is close to that of the average system and there is unique equilibrium which is exponentially stable
for the original system since A LTI has unique 0 equilibrium. Note that this requires  to be small
enough (as we dont have the above results for arbitrary .
Example 10.9 from [1]: Scalar nonlinear system
1
x = (x sin2 (t) x2 )
2
where right hand side is periodic with period ; sin2 (t) = 21 12 cos(2t); also
Z
1
1
1
fav (x) =
x sin2 t x2 dt = (x x2 )
0
2
2

(26.21)

(26.22)

Nonlinear Analysis Notes

79

So the averaged system is

1
1
x = ( x x2 )
2
2

(26.23)

with two equilibria (0 and 1).

Jacobian:

1
fav
= x
(26.24)
x
2
which is 1/2 at x = 0 and 1/2 at x = 1. The second is exponentially stable. There is a unique
-periodic soluiton in the vicinity of x = 1.

Nonlinear Analysis Notes

27

80

Lecture 27
x = f (t, x, )

(27.1)

f is T periodic: f (t + T, x, ) = f (t, x, ) and we have average system

Z
1 T
f (t, x, 0)dt
fav (x) =
T 0

(27.2)

If x(0) xav (0) = O() then for  small enough, x(t) xav (t) = O(), t [0, b/]. Suppose that
x = 0 is an exponentially stable equilibrium of the average system; then the approximation is valid
for all t [0, ), as long as x(0), xav (0) stability region, and for  small enough there is a unique
T -periodic solution in the vicinity of x = 0 and it is exponentially stable. If f (t, 0, ) = 0, then
x = 0 is this exponentially stable periodic solution.
Example (continuing from last time): Van der Pol oscillator (in [1] 10.5, also 2.4)
x
+ x = x(1
x2 )

(27.3)

without the 1 x2 this would just be usual linear oscillation with damping. We rewrite as first
order differential equation:
x 1 = x2
(27.4)
x 2 = x1 + x2 (1 x21 )
and apply change of coordinates, into polar:
x1
= r sin()
x2 = r cos()

(27.5)

p
r = x21 + x22
= tan1 ( xx12 )

(27.6)

Solving for r and , we have

and the resulting differential equation is

=

x1 x 1 +x2 x 2
r

x 1 x2 x 2 x1

x22 (1+

x2
1)
x2
2

= 41 (x1 x2 x2 x1 + x22 (1 x21 )) = x22 (1 x21 )

= 1r r2 cos2 ()(1 r2 sin2 ())
= r cos2 ()(1 r2 sin2 ())
= r12 (x22 + x21 x1 x2 (1 x1 )2 )

(27.7)

= 1 r12 r2 cos() sin()(1 r2 sin2 ())

= 1  cos() sin()(1 r2 sin2 ())

and finally we have

dr
d

cos ()(1r sin ())

=  1 rcos()
sin()(1r2 sin2 ())

(27.8)

is the independent variable and this has period 2 for .

Now fav =

1
2

r cos2 ()(1 r2 sin2 ())d andthe average system

dr
= fav =
d

(27.9)

Nonlinear Analysis Notes

cos2 ()

1+cos(2)
2

1
2

Z
0

81

1
r cos2 ()d = r and
2

cos2 () sin2 () = 41 (1 + cos(2)(1 cos(2)) = 41 (1 cos2 (2))

= 41 14 ( 12 + 12 cos(4)) = 18 81 cos(4)

(27.10)

and finally we have

1
2

Z
0

1
r3 cos2 () sin2 ()d = r3
8

(27.11)

This system has three equilibria: r = 0, r = 2, but only two make sense (cant have a
negative radius). To see which is stable, we need to look at the jacobian:
dfav
1 3
= r2
dr
2 8

(27.12)

and evaluating at our equilibria points, we have dfdrav |r=0 = 12 and dfrav |r=2 = 1; the first is unsetable
and the second is exponentially stable. In the original system, in a vicinity (i.e.  small enough) of
the circle of radius 2, there is a unique 2-periodic exponentially stable solution. It is worth noting
that in the ordinary linear oscillator case, there are many such periodic stable solutions (they fill
the space) but here there is only one.
Can check the following observations by simulation in Matlab (or check [1]). For  0.2 the
limit cycle is pretty much circular.
For  = 1, the cycle shape is distorted but still the same property holds that nearby solutions
are attracted to it.
For  = 5, the cycle shape is really distorted, with sharp corners, but still attracting for nearby
trajectories.
Remark: these results are for  small. This means: there is an  such that for every  < 
the results hold, but generally  wont be arbitrarily large.
We think of x = f (t, x) + g(t, x) as being composed of fast dynamics (f ) and slow dynamics
(g). This can be studied using the theory of singular perturbations (advertisement: this is covered
in 517).

27.1

General Averaging
x = f (t, x, )

but we no longer have that f is periodic in t. We can nevertheless still define

Z
Z
1 t
fav =
f (, x, 0)d
t t 0

(27.13)

(27.14)

Nonlinear Analysis Notes

82

Now lets see if this limit exists; if it does, then it gives us the average system, the same as
before
x = fav (x)
(27.15)
which is also still going to be autonomous.
Examples: f (t, x) = g(x)
t+1 As a function of time, f (t, x)- for x fixed- decays. The average is
zero (trivial calculus computation). This example shows that we can use a function which is not
necessarily periodic but whose nonperiodic part decays in the average, like this one.
If the average exists, it can be used to approximate the original system, in a way similar to
what we did in the particular case, but the details are a little more involved (see 10.6[1]).

27.2

Things not Covered [i.e. summer bedtime reading]

1. General Averaging
2. Singular Perturbations
3. Center Manifold Theory
4. More advanced control design (e.g. backstepping)

Nonlinear Analysis Notes

References
[1] H. Khalil. Nonlinear Systems. Prentice Hall, 2002.

83