You are on page 1of 29

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/281401563

Barbalat's Lemma and Stability – Misuse of a correct mathematical result?

Technical Report · September 2015

CITATIONS READS
6 3,397

1 author:

Itzhak Barkana
BARKANA Consulting
166 PUBLICATIONS 3,852 CITATIONS

SEE PROFILE

All content following this page was uploaded by Itzhak Barkana on 01 September 2015.

The user has requested enhancement of the downloaded file.


1

Barbalat’s Lemma and Stability – Misuse of a correct mathematical result?

Itzhak Barkana

BARKANA Consulting - Internal Report

September 1, 2015 DRAFT


2

Barbalat’s Lemma and Stability – Misuse of a


correct mathematical result?
Itzhak Barkana

Abstract

Even though Lyapunov methodology and the use of Lyapunov functions are the basis for almost
any stability analysis of nonlinear systems, the use of Lyapunov Stability theorem is limited by the
fact that the derivative of the Lyapunov function is required to be negative definite, while in most
cases it is hardly negative semidefinite. Because first extensions to the semidefinite case only covered
autonomous or periodic systems, Barbalat’s Lemma has been adopted from the theory of functions and
applied to stability analysis. Although new works of LaSalle and followers have extended the Invariance
Principle to nonautonomous systems, they have remained largely unknown and so, Barbalat’s Lemma
has remained the main stability analysis tool, in spite of the fact that it either requires all signals to
be uniformly continuous or requires finding solutions that could prevent the assumed negative effect of
discontinuities. This paper shows that the apparent need for continuity might have been the result of a
pure misuse of Barbalat’s Lemma.

Index Terms

Stability, Nonlinear systems, Autonomous and Nonautonomous systems

I. I NTRODUCTION

II. I NTRODUCTION

We recall that, even though now it is the basis for any modern stability analysis, the Lyapunov
stability approach itself [1] needed some 60 years before people started really paying attention
to it [2]–[6]. Moreover, after its first happy and successful applications, it soon became clear

I. Barkana is with BARKANA Consulting, 11/3 Hashomer St., Ramat Hasharon 47209, Israel. Phone: +972-3-5405973, Email:
ibarkana@gmail.com, i.barkana@ieee.org.

September 1, 2015 DRAFT


3

that finding an adequate Lyapunov function which has a negative definite derivative is not that
common in other than class-room examples. However, after many small steps that followed
Lyapunov, and in particular, based on largely unknown extensions of LaSalles Invariance Principle
to nonautonomous nonlinear systems [7]–[11], recent developments in stability analysis have
mitigated or even eliminated most apparently necessary prior conditions, thus adding confidence
in the robustness of adaptive and nonlinear control schemes in realistic situations [12]–[15].
Nevertheless, before these results of LaSalle and followers were available, Barbalat’s Lemma
has been adopted from functions theory and used for stability analysis. Moreover, because these
results of LaSalle and followers have remained largely unknown, Barbalat’s Lemma has remained
the main tool for stability analysis of nonlinear nonautonomous systems. This paper intends to
show that the tough uniform continuity condition that were deemed necessary for stability are
the result of a misinterpretation of a correct mathematical result from the theory of functions
when applied to stability of systems.
Section II thus briefly presents the various stability analysis tools with special emphasis on
Barbalat’s Lemma.

III. O N N ONLINEAR S YSTEMS S TABILITY A NALYSIS

Unless mentioned otherwise, this work deals with stability of non-autonomous nonlinear
systems (where the time explicitly appears) of the form

ẋ(t) = f(x, t). (1)

Although most works start by assuming a Lipschitz continuity condition or a somewhat relaxed
Caratheodory condition for f(x, t), we will only assume that the equation can be integrated and
has solutions, without questioning their continuity, uniqueness, etc. In this context, if impulse
response is the basic concept in linear systems, we see no reason to avoid impulse input
commands in nonlinear systems, i.e., to prevent the use of impulses in such equations as
ẋ = −x3 + δ(t), just because they are ”dangerously” nonlinear.
Because of the extreme importance of proper stability analysis of the Nonlinear Control
systems, here we briefly present a few old and new results [13]–[15].
a) Lyapunov direct method: The so-called Lyapunov direct method is the methodology that
allowed analysis of nonlinear systems stability without requiring to actually solve the nonlinear

September 1, 2015 DRAFT


4

differential equations for all initial conditions. Lyapunov proposed to associate with the system
an appropriate function, say positive definite and radially unbounded. Such functions have later
been called “Lyapunov functions” V (x), where x is the entire state vector. The Lyapunov theorem
states that if the derivative of V (x) along the trajectories of the system is negative definite, then
the system is globally asymptotically stable.
Although Lyapunov works have been published at the start of 20th century [1], it took more
than another half a century before they became the basic tool upon which modern stability anal-
ysis is based (for good presentations of Lyapunov direct method, see, for example the excellent
books [4] and [6]). Nevertheless, in spite of the initial enthusiasm, pretty soon developers were
forced to realize that in real world applications it is not easy to find an appropriate Lyapunov
function with a negative definite derivative. The main difficulty that limited the applicability
of the direct Lyapunov approach was the fact that, in most applications, the derivative of the
Lyapunov function usually was at most negative semidefinite. Here, things started looking pretty
complex.
b) Barbashin-Krasovskii-LaSalle Invariance Principle: First extensions to Lyapunov-style
approach for the case when the derivative of the Lyapunov function is only negative semidefinite
were first attributed in the West to LaSalle [2], yet now are also attributed to Krasovskii and
Barbashin [3], [5]. Their result has become known as the Barbashin-Krasovskii-LaSalle Invari-
ance Principle, that was only covering strictly autonomous functions of the form ẋ(t) = f (x)
(where the time does not explicitly show). In this case, they show that bounded trajectories still
end within the domain that is defined by V̇ (t) = 0. We notice that, unlike the strictly negative
definite case, this result does not necessarily imply asymptotic stability as its implications depend
on the meaning of V̇ (t) = 0.
Even though the Barbashin-Krasovskii-LaSalle Invariance Principle was an important exten-
sion of the Lyapunov approach to the case of semidefinite Lyapunov derivative and has also
been extended to nonautonomous yet periodic systems, most non-trivial control systems are
simply nonautonomous of the general form (1) and, as one may usually read in the professional
literature “unfortunately, this system is nonautonomous and so, the Invariance Principle cannot
be applied.”
Therefore, because control systems are nonautonomous, new extensions to the basic Lyapunov
stability theory have been sought.

September 1, 2015 DRAFT


5

c) Barbalat’s lemma: One of these extensions was provided by Barbalat’s lemma that we
state here as it appears in [16]:
Lemma 1 (Barbalat’s Lemma): If the differentiable function V (t) has a finite limit as t → ∞
and if V̇ (t) is uniformly continuous (or V̈ (t) is bounded), then V̇ (t) → 0 as t → ∞.
Barbalat’s Lemma is very simple and thus, very attractive. Furthermore, under some conditions,
it allowed to finally show that the function V̇ (t) ultimately vanishes and in many cases even it
allowed reaching the desirable asymptotic stability or asymptotically perfect tracking conclusion.
Nevertheless, it also leaves the burdensome impression that any input command, distortion or
disturbance that may affect the uniform continuity of Lyapunov derivative may affect the proof
and thus, the very guarantee of stability of nonlinear systems. However, as we show below, it is
only because Barbalat’s lemma deals with the functions and not with the systems that it imposes
those strict conditions on continuity of functions and even of their derivative. These conditions
may happen to hold in some systems, yet if they are not satisfied under less than ideal conditions,
it is not necessarily a result of some lack of stability.
The need for Barbalat’s Lemma is motivated by smart counterexamples:
I. Having f (t) = dF (t)/dt → 0 does not imply that F (t has a finite limit.
Example 1: F (t) = sin(ln(t)) and f (t) = dF (t)/dt = cos(ln(t))/t. In spite of the fact that
f (t) = dF (t)/dt → 0, the function F (t) keeps moving up and down.
II. Even if the function F (t has a finite limit as t → ∞, this does not necessarily imply that
f (t) = dF (t)/dt → 0.
Example 2: F (t) = sin(t2 )/t (Fig. 1), and f (t) = dF (t)/dt = 2cos(t2 ) − sin(t2 )/t (Fig.
2). While second term in f (t) vanishes in time, first term keeps oscillating up and down and
maintains same amplitude of oscillations, while the frequency of successive oscillations keeps
increasing without bound.
Example 3: Given
F (t) = e−t sin(e2t )

results in
f (t) = dF (t)/dt = −e−t sin(e2t ) + e−t cos(e2t )e2t ∗ 2

or
f (t) = dF (t)/dt = −e−t sin(e2t ) + et cos(e2t ).

September 1, 2015 DRAFT


6

F(t) vs time
1

0.5

-0.5
0 1 2 3 4 5 6 7 8 9 10

Fig. 1. Example 2 - F (t) = sin(t2 )/t

f(t) vs time
2.5

1.5

0.5

-0.5

-1

-1.5

-2

-2.5
0 1 2 3 4 5 6 7 8 9 10

Fig. 2. Example 2 - f (t) = Ḟ (t) = 2cos(t2 ) − sin(t2 )/t

Here, the claim is that F (t) → 0 as t → ∞, while f (t) → ∞ as t → ∞. This example


seems to be a “decisive” illustration of what threats may affect stability of nonlinear systems if
uniform continuity is not invoked.
III. Most important in the context of stability, the claim is that even if F (t) is bounded from
below and decreasing (i.e., f (t) = dF (t)/dt ≤ 0), then F (t) converges to a limit, yet this does
not say whether or not f (t) = dF (t)/dt → 0. This is illustrated by two examples, as shown
below.
Example 4. f (t) = g(t) + h(t), g(t) = −e−t , h(t) = 0 almost everywhere, except for the
discrete-time sequence tk = 0, 1, 2, ... where h(tk ) = −1. Here, the argument is that because

September 1, 2015 DRAFT


7

f(t) vs time
0

-0.2

-0.4

-0.6

-0.8

-1

-1.2

-1.4
0 0.5 1 1.5 2 2.5 3 3.5 4

Fig. 3. Example 4 - f (t) = g(t) + h(tk )

F(t) vs time
1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5 2 2.5 3 3.5 4

Fig. 4. Example 4 - F (t)

h(t) does not contribute to the integral, the integral F (t) is decreasing and has a finite limit
(Fig. 4), while its non-smooth derivative f (t) = g(t) + h(t) keeps jumping for ever and does
not imply f (t) → 0 (Fig. 4).

Example 5. f (t) = ∞ k=0 e
−4k (t−k)2
(Fig. 5) and its integral F (t) (Fig. 6). Here, the argument
is that f (t) is continuous yet not uniformly continuous as its peaks become spikes as t →
∞. Because the spikes do not contribute to the integral, the claim is that the integral F (t) is
decreasing and has a finite limit, while its non-smooth derivative f (t) keeps jumping for ever.
Before we move to next results that cover general nonautonomous systems, we revisit the
counterexamples that have motivated for all of us the need of using Barbalat’s Lemma for

September 1, 2015 DRAFT


8

f(t) vs time
1.5

0.5

0 1 2 3 4 5 6 7 8 9 10

∑∞ k
(t−k)2
Fig. 5. Example 5 - f (t) = k=0 e−4

F(t) vs time
3

2.5

1.5

0.5

0
0 1 2 3 4 5 6 7 8 9 10

Fig. 6. Example 5 - F(t)

stability analysis of nonautonomous nonlinear systems and, in particular, its apparent need for
uniform continuity of the Lyapunov derivative.
Example 1 is less important, because in the context of stability we usually do guarantee that
F (t) reaches a finite limit in general. Nevertheless, even in this context, writing
( )
ln(t)
F (t) = sin(ln(t)) = sin t
t
shows a sinusoidal wave with ever increasing period. Therefore, as the period of this “sinusoidal”
function ultimately reaches infinity, it is not very clear if one should claim that there is no finite
limit, or if it is not more exact to claim that the wave reaches a constant though unknown limit
rather than no finite limit.

September 1, 2015 DRAFT


9

Actual dF(t)/dt=g(t) vs time


0

-0.1

-0.2

-0.3

-0.4

-0.5

-0.6

-0.7

-0.8

-0.9

-1
0 0.5 1 1.5 2 2.5 3 3.5 4

Fig. 7. Example 4 - ActualdF (t)/dt = g(t)

Nevertheless, the other examples are more interesting, because their common claim is that the
integral does reach a finite limit, while the derivative does not end at zero.
Example 4 revisited: We here start with Example 4 because in the integrant f (t) of this
example it is easy to separate the term g(t), which contributes to the integral, from the term
h(t), which does not contribute to the integral. It is easy to see that, even if one formally denotes

Ḟ (t) = f (t) = g(t) + h(t),

when we integrate we get


∫ t ∫ t
F (t) = f (τ )dτ = (g(τ ) + h(τ )) dτ
0 0

and, because h(t) does not contribute to the integral,


∫ t
F (t) = g(τ )dτ = exp(−t)
0

Thus, the real derivative of F (t), which we get by actually differentiating F (t) is

dF (t)/dt = −exp(−t)

and, therefore, the actual derivative does reach zero as t → ∞ (Fig. 7).
In general, for any function f (t) = g(t) + h(t), where h(t) does not contribute to the integral
∫t
F (t) = 0 f (τ )dτ , we actually get
∫ t ∫ t
F (t) = f (τ )dτ = g(τ )dτ
0 0

September 1, 2015 DRAFT


10

and therefore, the actual derivative of F (t) is

dF (t)/dt = g(t).

As g(t) ≤ 0 keeps contributing to the integral as long as it is not zero, the integral function
cannot stop decreasing unless g(t), its actual derivative, stops being negative, or, in other words,
unless dF (t)/dt = g(t) → 0 as t → ∞. Example 4 was the last attempted counterexample that
was meant to show why stability could not be based on Lyapunov derivative unless one requires
uniform continuity of f(x, t). However, in spite of the initial confusion and failed attempts, we
managed to show that it does not motivate the need for uniform continuity and ultimately clearly
showed that, in spite of the customary formal notations, those terms of the integrant which do
not contribute to the integral do not appear in the actual derivative of the integral function.
therefore this attempted counterexample ultimately served as a positive trigger which led us to
review the other customary counterexamples that we all got used to take for granted.
Example 5 revisited: The analysis of Example 5 is more delicate, because here it is more
difficult to separate between the terms which contribute to the integral and the terms which do
not contribute. Nevertheless, because in this example f (t) ≥ 0, the integral function F (t) reaches
a finite limit because the ultimate contribution of those terms that ultimately become just local
spikes is zero. Therefore, in spite of the formal notation, the actual derivative of F (t) is not its
integrant f (t), but rather another function, call it g(t) = dF (t)/dt, which is obtained by actually
differentiating F (t) and which ultimately stops contributing to F (t) because it ultimately does
reach zero, or g(t) → 0 as t → ∞.
The right implications from these examples is that, if the integrant function f (t) contains
∫t
terms that do not contribute to the integral F (t) = 0 f (τ )dτ , then the integral can reach a
finite limit even though the integrant f (t) does not tend to zero, as it keeps jumping for ever.
However, the actual derivative of the function which reaches a finite limit is not identical with
the integrant and it does finally tend to reach zero.
Therefore, examples 4 and 5 seem to show that a better formulation of Barbalat’s Lemma
should be:
Lemma 1b (The alternative Barbalat’s Lemma): Given the function f (t) and given that its
∫t
integral function F (t) = 0 f (τ )dτ has a finite limit as t → ∞, then if the integrant function
f (t) is uniformly continuous (or f˙(t) is bounded), then f (t) → 0 as t → ∞.

September 1, 2015 DRAFT


11

However, in spite of the correct interpretation of the counterexamples, the first formulation
of Barbalat’s Lemma which appears in most Nonlinear Systems books and publications is also
a perfectly correct mathematical result, only its implications in the context of systems stability
might have gone wrong. In particular, its apparent need for uniform continuity in the context of
stability seems to be the result of a formal use of mathematical notations, which confuses and
mixes between the integrant function f (t) and the real derivative of F (t) that must be obtained
by actually differentiating F (t).
Examples 2 and 3 are mode difficult to contradict, because they seem to use well-established
mathematical formulas. However, even though the formulas seem to “work,” they have been
established for decent functions, while here they are formally used with less than decent functions.
Indeed, it is very clear that the derivative of sin(ωt) is ωcos(ωt). However, even if this is
true for ω = 1 and ω = 2, can one really differentiate the function sin(ωt) when ω actually is
infinity? Can one really refer to the derivative of the functions sin(t2 ) = sin(t ∗ t) or sin(e2t ) =
sin((e2t /t)t) when t actually reaches infinite? In our opinion, the answer is negative, unless the
Dirichlet function has suddenly become both integrable and differentiable. Can one invoke the
absolute need for uniform continuity and, as its motivation, use the derivative of a “sinusoidal”
function of infinite frequency and then multiply and divide the result by time when time is
infinity, (when anything other than the peculiar differentiation tells you that, when time actually
reaches infinity, F (t) = sin(t2 )/t is nothing else but just another mathematical representation of
a constant), and assume that its derivative also results in sinusoidal functions of infinite frequency
and finite or even infinite amplitude?
Nevertheless, while the function sin(t2 ) becomes ”indecent” as time tends to infinity, F (t) =
sin(t2 )/t remains decent. Therefore, for this specific situation we must go back and apply the
basic differentiation rule to the decent function F (t). We do not expect different results for finite
t, and indeed
sin((t+h)2 ) sin(t2 )
dF F (t + h) − F (t) t+h
− t
= lim = lim
dt h→0 h h→0 h
sin(t2 +2ht+h2 ) sin(t2 )
dF t+h
− t sin (t2 + 2ht + h2 ) t − (t + h) sin (t2 )
= lim = lim
dt h→0 h h→0 th (t + h)

dF t sin (t2 ) cos(2ht + h2 ) + t cos (t2 ) sin(2ht + h2 ) − t sin (t2 ) − h sin (t2 )
= lim
dt h→0 th (t + h)

September 1, 2015 DRAFT


12

Towards the limit, we ignore h versus t and h2 versus th to get


dF t sin (t2 ) cos(2ht) + t cos (t2 ) sin(2ht) − t sin (t2 ) − h sin (t2 )
= lim
dt h→0 t2 h
( )
dF t sin (t2 ) (cos(2ht) − 1) + t cos (t2 ) sin(2ht) sin (t2 )
= lim −
dt h→0 t2 h t2
For t finite, we have sin(2ht) ≈ 2ht and cos((2ht) ≈ 1
dF 2ht2 cos (t2 ) − h sin (t2 ) 2t2 cos (t2 ) − sin (t2 )
= lim = lim
dt h→0 t2 h h→0 t2
and finally get
dF ( ) sin (t2 )
= 2 cos t2 −
dt t2
However, as t → ∞,
 
( ) sin((t+h)2 ) sin(t2 )
dF F (t + h) − F (t) −
lim = lim lim = lim  lim t+h t 
t→∞ dt h→0 t→∞ h h→0 t→∞ h

Because the sinusoidal functions are bounded, yet t tends to infinite, the numerator vanishes
and so ( )
dF 0
lim = lim lim =0
t→∞ dt h→0 t→∞ h

Example 3 revisited: Given the decent function

F (t) = e−t sin(e2t )

we only compute its derivative as t → ∞ to get


( ) ( ( ) )
dF F (t + h) − F (t) e−(t+h) sin e2(t+h) − e−t sin (e2t )
lim = lim lim = lim lim
t→∞ dt h→0 t→∞ h h→0 t→∞ h
Again, because the sinusoidal functions are bounded, while the multiplying exponential func-
tions vanish as the time t tends to infinite, the numerator again vanishes and so
( )
dF 0
lim = lim lim =0
t→∞ dt h→0 t→∞ h

Therefore, in both cases, the actual derivative of the integral function F (t) does not contain
those spikes and actually does reach zero.
Nevertheless, in spite of these comments above, the problem is not with those examples,
even if all were perfectly right, which they are not. The problem is with their demanding and

September 1, 2015 DRAFT


13

V
1

0.9

0.8

0.7

0.6

V
0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time [sec]

Fig. 8. Smooth Lyapunov function

V
1

0.9

0.8

0.7

0.6
V

0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time [sec]

Fig. 9. Smooth Lyapunov derivative

restricting implications for stability analysis of nonlinear system. Why should Barbalat’s Lemma
and these counterexamples prevent someone from using an assumed dangerously discontinuous
square-wave input command with a robot, only because the robot is nonlinear?
The problem with Barbalat’s Lemma probably is the one word that is missing: “nicely.” Indeed,
one could decide to only look for such uniformly continuous and differentiable functions as
F (t) = e−t (Fig. 8) and its uniformly continuous derivative f (t) = dF (t)/dt = −e−t (Fig. 9),
where the derivative f (t) is negative definite, the function F (t) is decreasing and reaches zero
and its derivative reaches zero “nicely.”
However, if we now sample-and-hold F (t), thus making it into a staircase (i.e., a sequence of

September 1, 2015 DRAFT


14

Vzoh
1

0.9

0.8

0.7

0.6

Vzoh
0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time [sec]

Fig. 10. Staircase Lyapunov function

steps), doesn’t the “dangerously” discontinuous F (t) 10 keep decreasing in ever decreasing steps
and doesn’t it ultimately reach 0 in spite of its discontinuities? Doesn’t the derivative remain
negative semidefinite and ultimately reach 0, in spite of it being discontinuous, unbounded and
even containing impulses?
These counterexamples seem all to be claiming that common sense could be wrong. This
claim could be right yet, as we can see, it does not imply that one can forget one’s common
sense and just follow the formula. So, if common sense made you think that the derivative of a
constant is simply zero, this could be right, even if counterexamples seem to prove you wrong!
In particular, playing with infinity may generate peculiar results: one can ”prove” that 1=0, 1=2
or that a+b=c in any triangle.
So, the use of Barbalat’s Lemma in stability analysis seems to have largely divided the world
into two large groups. The Optimists work very hard to guarantee that all signals are uniformly
continuous. The Pessimists, who want to be realistic and do not think they can afford to assume
uniform continuity, work even harder, using great Math, to find “solutions” to the threats of
discontinuity.
Even though another group and other works have tried to ask “Who needs continuity?” they
seem to have been just ignored.
One of the most interesting approaches that have been developed instead is Impulsive Control,
where the position x(t) itself can jump from A to B and back, all that without affecting the
velocity and so, even though the position jumps, the velocity function f(x, t) in (1) remains

September 1, 2015 DRAFT


15

Lipschitz continuous. As an abstract mathematical exercise, this is very impressive. Nevertheless,


with all due respect to Math, one cannot really think that one can instantaneously move one’s
robot from one position to another, without even having to think of the infinite speed, not to
mention the acceleration and the energy that must be invested.
Moreover, Barbalat’s Lemma and its demanding conditions are deemed needed because, as
most colleagues and some of the best publications put it, “Unfortunately, LaSalle’s Invariance
Principle (and you’d better call it ‘Barbashin-Krasovskii-LaSalle Invariance Principle’) only
covers ‘time-invariant’ (more exactly, autonomous) systems.”
Is it indeed so? Only if one still sticks to the 60s and ignores the relevant works of LaSalle
and followers.
The examples should show the difference between the Barbalat’s Lemma, an important the-
oretical result for the theory of functions, and Lyapunov-LaSalle stability analysis approach.
While all conditions invoked by Barbalat’s Lemma are imposed in order to make sure that
the derivative tends to zero, the Lyapunov-LaSalle approach is only interested in the system
trajectories and only tries to locate the limit points of those trajectories, even though, to this
end and only to this end, it may make use of derivatives, yet only as a useful means for locating
the limit points. Here, it is worth mentioning that, in spite of the customary use of Lyapunov
functions and Lyapunov derivatives, there are also direct uses of the Lyapunov function, without
using the derivative at all [17].
d) (The real) LaSalle’s Invariance Principle: As a matter of fact, extensions of LaSalle’s
Invariance Principle to nonautonomous systems have been available at least since 1976 [7]–[11],
and have been immediately adopted and used since 1980 for such special problems as adaptive
control of large space structures and other applications (see for example [18] for a proof and
a brief presentation of the theory along with some early examples). Nonetheless, as classical
books in nonlinear systems [16], [19]–[21] and even recent publications [22], [23] seem to show,
either the Principle is largely unknown or, at least, has remained misunderstood.
Therefore, it is important to first emphasize LaSalle’s simple and ingenious idea. Instead of
dealing with the properties of some general function, LaSalle’s Invariance Principle goes back to

establishing some milder conditions on the system. Using the notation |f | = f12 + f22 + ...fn2 ,
satisfaction of one of the following two assumptions along trajectories of a system is checked:
1) |f (x(t), t)| is uniformly bounded for any bounded x.

September 1, 2015 DRAFT


16

or
∫β
2) α
|f (x, τ )|dτ = µ(β − α).
Theorem 3 (LaSalle’s Invariance Principle): Consider the nonlinear non-autonomous system
(1). Assume that there exists a Lyapunov function V (x), which is bounded from below, and
that its derivative V̇ (x, t) along the trajectories of (1) is Negative Semi-Definite and satisfies a
relation of the form V̇ (x, t) ≤ W (x) ≤ 0. Now, define the domain Ω = {x|W (x) = 0}. Then, if
one of the assumptions 1 or 2 holds, the entire state vector x(t) ultimately reaches the domain
Ω [7], [18].
e) The new Invariance Principle: As recently observed [12], even the milder conditions
that LaSalle’s original formulation still seems to impose are both difficult to satisfy in realistic
applications and also are not necessarily needed. Therefore, a new Invariance Principle for
nonautonomous systems was we presented [12] in a form that is appropriate for the problems
we discuss and that further relaxes even the milder conditions of LaSalle. Here, satisfaction of
one of the following two assumptions is checked along trajectories of a system:
A) |f (x(t), t)| is uniformly bounded for any bounded x.
or
∫β
B) α
|f (x(τ ), τ )|dτ is bounded along any bounded trajectory x(t) and for any finite time
interval p = β − α.
Note that, while Assumption A is identical to Assumption 1 in LaSalle’s formulation, our
presentation of Assumption B above is different from LaSalle’s original presentations, where
∫β
Assumption 2 is written in the form α |f (x, τ )|dτ = µ(β − α). In LaSalle’s formulation, the
function µ(τ ) is a “modulus of continuity,” from which one can imply that the trajectory is
supposed to be a continuous function of time. Here, we replaced the modulus of continuity by
a simple bound, because although continuity is desirable, it cannot be guaranteed in practical
environments and it is not necessarily needed for stability. As the proof of stability shows, the
only condition that is needed is that the trajectory, continuous or not continuous, cannot pass
an infinite distance in finite time [12].
Note: Because a number of reviewers and other readers questioned the functions that could
possibly satisfy the assumptions of either LaSalle’s Invariance Principle or the new Invariance
Principle and even offered counterexamples, it is worth emphasizing that those are only assump-
tions that must be checked along the specific system trajectories and do not relate to any general

September 1, 2015 DRAFT


17

properties of the function f(x, t).


On another point, LaSalle’s works could only deal with those Lyapunov derivatives that satisfy
a relation of the form V̇ (x, t) ≤ W (x) ≤ 0. However, although in many cases this relation
could be sufficient, it may still unnecessarily restrict the applicability of stability theory. For
example, while for V̇1 (x, t) = −x2 (2 + sint) one can define W (x) = −x2 and then write
V̇1 (x, t) ≤ W (x) ≤ 0, this is not possible for V̇2 (x, t) = −x2 (1 + sint). Nevertheless, one
can still define W (x) = −x2 and g(t) = 1 + sint and get V̇2 (x, t) ≤ W (x)g(t) ≤ 0. It is
clear that V̇2 (x, t) ≤ 0 or, in other words, that V̇2 (x, t) is uniformly negative semi-definite.
In a more general case, such as V̇3 (x, t) = −x21 (1 + sint) − x22 (1 + cost) it is still clear that
V̇3 (x, t) is uniformly negative semi-definite, although it cannot be written in any one of the (more
convenient) previous forms. Therefore, whenever needed, we will show that one can directly deal
with uniformly positive and negative definite explicit functions of time.
Theorem 4 (The new Invariance Principle): Consider the nonlinear non-autonomous system
(1). Assume that there exists a Lyapunov function V (x) which is bounded from below and that its
derivative V̇ (x, t) along the trajectories of (1) is Negative Semi-Definite, i.e., satisfies V̇ (x, t) ≤ 0.
Now, define the domain Ω where the Lyapunov derivative equals zero, Ω = {x|V̇ (x, t) = 0)},
and the restricted domain Ωi where the Lyapunov derivative is identically zero (i.e., not just
equal zero), Ωi = {x|V̇ (x, t) ≡ 0)}. Then, if one of the assumptions A or B holds, the entire
state vector x(t) ultimately reaches the domain Ω. In particular, equilibrium points and limit
cycles belong to the restricted domain Ωi [12].
In other words, under fairly mild conditions, the Invariance Principle extension to nonau-
tonomous systems guarantees that all trajectories ultimately reach the domain Ω. However,
as examples illustrate [12], [15], its significance and efficiency and sometimes even its mere
existence seems to have remained unknown to a large section of potential users.
f) The new Theorem of Stability: Note that stability theory based on the new Invariance
Principle approach eliminates the previous requirement for uniform continuity of the Lyapunov
derivative and actually any other requirement, except for the guarantee that any bounded trajec-
tory x(t) cannot pass an infinite distance in finite time.
We notice that Assumptions A and B that were required by the new Invariance Principle
were deemed needed only in order to show that limit points of type rosette, i.e., those isolated
rosette-type limit points, that the trajectories might reach, leave, and then come back to them an

September 1, 2015 DRAFT


18

infinite number of times, must belong to Ωe = {x|V̇ (x, t) = 0}.


In retrospect, it is amazing that, while so much thought and effort had been invested to
investigate what must happen at those particular rosette-type point locations, not much room
or thought was left for all those segments that a trajectory must pass after leaving the limit
point and before coming back to it and where the assumable non-zero derivative is supposed to
be negative. Because only if the trajectory revisits the point an infinite number of times would
a rosette-type point become a limit point, ultimately, V̇ (x, t) must tends to zero all along the
trajectory or, in other words, even what might have started looking as a rosette-type limit point
ultimately must also belong to Ωi , as part of a limit cycle or even as an equilibrium point.
In order to formulate the new Theorem of Stability in its most general form, the domain Ωi
is defined as follows [15]:
{ ( ) }
Ωi = x| lim V̇ (x, t) ≡ 0 . (2)
t→∞

The new Theorem of Stability can thus be written in the following simple formulation [15]:
Theorem 5 (The new Theorem of Stability) Consider the nonlinear non-autonomous system
(1). Let V (x) be a differentiable function bounded from below. (Note that V(x) is not required to
be Positive Definite.) Assume that its derivative V̇ (x, t) along the trajectories of (1) is Negative
Semidefinite, i.e., satisfies V̇ (x, t) ≤ 0. Then, all limit points of any bounded trajectory x(t)
belong to the domain Ωi [14], [15]. (For convenience and in order to make the report self-
contained, the Appendix also presents the proof of the new Theorem of Stability.)
Note that, for convenience, because V (x) is a selected function, one can assume that both
V (x) and V̇ (x, t) are continuous functions of x. However, as shown in [12] and [15], their
differentiability with respect to t implies the Dini derivatives (see for example [24], [25]). In
other words, while it is nice to have continuous functions that are also differentiable in the
classical sense, stability is not affected if eventual discontinuity of x(t) leads to discontinuity of
V (x(t)). In this context, a piece-wise continuous function may still be differentiable everywhere,
even if its derivatives at the points of discontinuity are δ-functions.

A. Example 1

In order to illustrate the advantage of using the new Theorem of Stability we revisit an
example of [15], where the Lyapunov derivative may take more complex forms and the assumably

September 1, 2015 DRAFT


19

necessary conditions of continuity and differentiability are obviously violated. For example,
define the square function wave sα,β,T with lower value α, upper value β and period T and
assume that the system is

ė(t) = −e(t)s0,2,3 + θw(t)


(3)
θ̇(t) = −e(t)w(t)
The Lyapunov function
V (e, θ) = e2 (t) + θ2 (t)

results in the Lyapunov derivative

V̇ (e, θ, t) = −2e2 (t)s0,2,3 ≤ 0.

Because the Lyapunov derivative is only negative semidefinite, the original Lyapunov theorem
only tells us that all trajectories are bounded. This is a nonautonomous nonlinear systems, so
the Barbashin-Krasovskii-LaSalle Invariance Principle does not apply. Moreover, although the
input command w(t) does not show in the Lyapunov derivative, the square wave s0,2,3 does
and therefore the derivative function is not continuous in t any longer and Barbalat’s lemma
cannot be used. Moreover, LaSalle’s Invariance Principle also ends its utility because there is
no W (e, θ) such that V̇ (e, θ, t) = −2e2 (t)s0,2,3 ≤ W (e, θ) ≤ 0. However, the new Theorem of
Stability and the direct use of the nonautonomous Lyapunov derivative helps, because we still
get V̇ (e, θ, t) = −2e2 (t)s0,2,3 ≤ 0, so e(t) ≡ 0 guarantees that V̇ (e, θ, t) ≡ 0. Because the new
theorem of stability requires the identity condition, it implies that the derivatives of V̇ (x) are
also zero and so, ė(t) = 0. Substituting in the second equation leads to θ̇(t) = 0 and then from
first equation one gets θ(t)w(t) = 0. One can see that if the input w(t) ends with some nonzero
value, it implies that θ(t) ends at θ(t) = 0, while if the input w(t) ends at zero, it implies that
θ(t) still ends at some other constant value that depends on the initial conditions. Therefore, it
is important to notice that, because in this simple example θ(t) represents the ”adaptive gain,”
the result of the new Theorem of Stability, namely, that the adaptive gain θ(t) ultimately ends
at some constant value, was obtained without requiring any of the (otherwise) ”customary”
persistent excitation conditions.
Nonzero ultimate w(t) value, that here is the result of using some square-wave input command,
results in the constant values e(t) = 0 and θ(t) = 0, as illustrated in Fig. 11.

September 1, 2015 DRAFT


20

et vs thetat
0.5

θ
−0.5

−1
−1 −0.5 0 0.5 1 1.5 2
e

Fig. 11. Example 1 - Typical Trajectory with square-wave input command

et vs thetat
0.3

0.2

0.1

−0.1
θ

−0.2

−0.3

−0.4

−0.5
0 0.5 1 1.5 2
e

Fig. 12. Example 1 - Typical Trajectory with vanishing ramp input command

Instead, a zero ultimate value of w(t), that here is the result of using a ramp function that
starts at w(0)=-1 and ends after one second at w(1)=0, results in the constant values e(t) = 0
and θ(t) ̸= 0, as illustrated in Fig. 12.
Before we end with this example, it is worth mentioning that stability analysis of this example
is not affected if the input command w(t) or the time-varying parameter function contain δ-
functions or other discontinuous or unbounded terms, as long as they do not affect the negative
semi-definiteness of the Lyapunov derivative.

September 1, 2015 DRAFT


21

B. Example 2

Another example was recently offered to illustrate the apparent need for uniform continuity.
Instead, it seems to be another illustration for the new Theorem of Stability. The system is

ẋ1 (t) = u1 (t)


(4)
ẋ2 (t) = u2 (t)
Here, the input commands u1 (t) and u2 (t) are given by

u1 (t) = −1.5sign(x1 (t)) + sign(x2 (t))


(5)
u2 (t) = −sign(x2 (t)) − sign(x1 (t) + x2 (t))
The example was offered because the simulation plots show wild behavior of the input
commands and the state derivative violates any condition of continuity (Figs. 13 and 14).
Nevertheless, the new Theorem of Stability still allows choosing the Lyapunov function V (e, θ) =
x21 (t) + x22 (t), which results in the Lyapunov derivative

V̇ (x1 , x2 ) = 2x1 ẋ1 + 2x2 ẋ2 (6)

V̇ (x1 , x2 ) = −3x1 sign(x1 ) + 2x1 sign(x2 ) − 2x2 sign(x2 ) − 2x2 sign(x1 + x2 ) (7)

V̇ (x1 , x2 ) = −3|x1 | + 2x1 sign(x2 ) − 2|x2 | − 2x2 sign(x1 + x2 ) (8)

The presence of the absolute value terms shows that, the Lyapunov derivative is negative
semidefinite, V̇ (x1 , x2 ) ≤ 0. The new Theorem of Stability still tells that all trajectories end at
V̇ (x1 , x2 ) ≡ 0, which, even in this wild case, is equivalent with the origin, x1 = x1 = 0, or with
asymptotic stability (Figs. 15, 16, 17).
However, at a second thought, the wild behavior of the input commands looked peculiar. If
x1 > 0 and x2 = 0, then sign(x2 ) = 0 and both input commands should be pretty constant.
After a second analysis, the explanation was found in the numerical problems of the integration.
While an accidental spike of width zero of the derivative would have zero effect on the integral,
this is not so in the digital integration. Plotting x1 and x2 at larger scale (Fig. 18) shows a
numerical noise of about 2 • 10−7 in amplitude.
Therefore, we added the instructions

if abs(x1 ) < 2.1 • 10−7 x1 = 0

September 1, 2015 DRAFT


22

x1dot=u1
3

x1dot
0

-1

-2

-3
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t

Fig. 13. Example 2 - ẋ1 = u1

x2dot=u2
3

1
x2dot

-1

-2

-3
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t

Fig. 14. Example 2 - ẋ2 = u2

V(t)
0.3

0.25

0.2
V(t)

0.15

0.1

0.05

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t

Fig. 15. Example 2 - The Lyapunov function

September 1, 2015 DRAFT


23

dV/dt
0

-0.2

-0.4

-0.6

-0.8

dV/dt
-1

-1.2

-1.4

-1.6

-1.8

-2

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4


t

Fig. 16. Example 2 - The Lyapunov derivative

x1 x2
0.6
x1
x2
0.5

0.4

0.3
x1 x2

0.2

0.1

-0.1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t

Fig. 17. Example 2 - x1 and x2

×10-6 x1 x2
1
x1
0.8
x2

0.6

0.4

0.2
x1 x2

-0.2

-0.4

-0.6

-0.8

-1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t

Fig. 18. Example 2 - x1 and x2 at large scale

September 1, 2015 DRAFT


24

x1dot=u1
3

x2dot
0

-1

-2

-3
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t

Fig. 19. Example 2 after correction - ẋ1 = u1

x2dot=u2
3

1
x2dot

-1

-2

-3
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t

Fig. 20. Example 2 after correction - ẋ2 = u2

and
if abs(x2 ) < 2.1 • 10−7 x2 = 0

With this addition, simulation results correspond to the expected theoretical results (Figs. 19 and
20) and x2 remains zero even when plotted at very large scale (Fig. 21).

IV. C ONCLUSION

This paper revisits Barbalat’s Lemma and its use for stability analysis and shows that in
particular, its apparent need for uniform continuity or, alternatively, the apparent need for devel-
oping solutions for apparent problems with discontinuity might have simply been an unfortunate

September 1, 2015 DRAFT


25

×10-6 x1 x2
1
x1
0.8
x2

0.6

0.4

0.2

x1 x2
0

-0.2

-0.4

-0.6

-0.8

-1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
t

Fig. 21. Example 2 after correction - x1 and x2 at large scale

misuse of a correct mathematical result from function theory.

R EFERENCES

[1] A. M. Lyapunov, The General Problem of the Stability of Motion, Annales de la Faculté des Sciences de Toulouse, Second
Series. Toulouse, FRANCE: Faculté des Sciences de Toulouse, 1907, vol. 9.
[2] J. P. LaSalle and S. Lefschetz, Stability by Lyapunov Direct method with Applications. New York, NY: Academic Press,
1961.
[3] N. N. Krasovskii, Stability of Motion. Stanford, CA: Stanford University Press, 1963.
[4] W. Hahn, Stability of Motion. New York, NY: Springer-Verlag, 1967.
[5] E. A. Barbashin, Introduction to the theory of stability. Wolters-Noordhoff, 1970.
[6] N. Rouche, P. Habets, and M. Laloy, Stability Theory by Lyapunov’s Direct Method. New York: Springer-Verlag, 1977.
[7] J. P. LaSalle, “Stability of non-autonomous systems,” Nonlinear Analysis Theory Methods and Applications, vol. 1, no. 1,
pp. 83–90, 1976.
[8] ——, The Stability of Dynamical Systems. Philadelphia: SIAM, 1976.
[9] Z. Artstein, “Limiting equations and stability of nonautonomous ordinary differential equations, appendix a,” in The Stability
of Dynamical Systems. New York: SIAM, 1976, vol. 35, pp. 187–235.
[10] ——, “The limiting equations of nonautonomous ordinary differential equations,” Journal of Differential Equations, vol. 25,
pp. 184–202, 1977.
[11] ——, “Uniform asymptotic stability via the limiting equations,” Journal of Differential Equations, vol. 27, pp. 172–189,
1978.
[12] I. Barkana, “Defending the beauty of the invariance principle,” International Journal of Control, vol. 87, no. 1, pp. 186–206,
2014, (Published On-Line 6 September 2013) DOI:10.1080/00207179.2013.826385.
[13] ——, “Simple adaptive control - a stable direct model reference adaptive control methodology- brief survey,” International
Journal of Adaptive Control and Signal Processing, vol. 28, no. 7, pp. 567–603, 2014, (Published On-Line 17 June 2013)
DOI:101002/acs.2411.

September 1, 2015 DRAFT


26

[14] ——, “The beauty of simple adaptive control and new results in nonlinear systems stability analysis,” in Proceedings of
2014 ICNPAA World Congress, 10th International Conference on Mathematical Problems in Engineering, Aerospace and
Sciences, Narvik, Norway, Jul. 2014, Keynote speech.
[15] ——, “The new theorem of stability - Direct extension of Lyapunov theorem,” Mathematics in Engineering, Science and
Aeronautics - MESA, 2015, (To appear).
[16] J.-J. Slotine and W. Li, Applied Nonlinear Control. Englewood Cliffs, New Jersey: Prentice Hall, 1991.
[17] A. N. Michel and L. Hou, “Relaxation of hypotheses in LaSalle–Krasovskii–type invariance results,” SIAM Journal of
Control and Optimization, volume=49, number=4, pages=1383-1403, note=DOI: 10.1137/090749712.
[18] H. Kaufman, I. Barkana, and K. Sobel, Direct Adaptive Control Algorithms, 2nd ed. New York: Springer, 1998.
[19] S. Sastri, Nonlinear Systems. New York: Springer, 1999.
[20] H. K. Khalil, Nonlinear Systems, 3rd ed. Englewood Cliffs, NJ: Prentice-Hall, 2002.
[21] M. Vidyasagar, Nonlinear Systems Analysis. Philadelphia, PA: SIAM, 2002.
[22] M. Hou, G. Duan, and M. Guo, “New versions of Barbalats lemma with applications,” J. Control Theory Appl., vol. 8,
no. 4, pp. 545–547, 2010.
[23] T.-C. Lee, D.-C. Liaw, and B.-S. Chen, “A general invariance principle for nonlinear time-varying systems and its
applications,” IEEE Transactions on Automatic Control, vol. 46, no. 12, pp. 1989–1993, 2001.
[24] A. M. Bruckner, Differentiation of Real Functions, 2nd ed. Providence, RI: American Mathematical Society, 1994.
[25] K. B. Oldham and J. Spanier, The Fractional Calculus. Mineola, NY: Dover, 2006.

A PPENDIX

All theorems of stability attempt to locate the limit points or the points of accumulation of
bounded trajectories. Therefore, the first concept to ruminate about before going to the proof
of the theorem is the very definition of limit points. A limit point, or point of accumulation,
of a trajectory is such a point that any neighborhood, arbitrarily small, around it contains an
infinite number of points of the trajectory. When one first hears this definition, one could
be confused, because it seems pretty clear that any point of a continuous curve satisfies this
condition. Therefore, when one deals with dynamical system and trajectories, one defines a
discrete time sequence {tk } and, correspondingly, discrete-time points on the trajectory {x(tk )}.
In this context, a limit point is that point of the discrete sequence {x(tk )} that any neighborhood,
arbitrarily small, around it contains an infinite number of discrete points of the trajectory.
Next, any bounded trajectory that leads to the creation of such an infinite number of discrete
points, must contain at least one such accumulation. Therefore, as we only want to know where
these discrete-time limit points are located, although the ever diminishing distances between
points around a limit point may sound similar to the definition of continuity, they have nothing
in common with, neither do they need to even mention any continuity.

September 1, 2015 DRAFT


27

As illustrated in detail in [12], there are various ways the limit point may appear when we
start moving along any bounded trajectory from x(t0 ) to x(t1 ),..., x(ti ),...
The simplest form is when the trajectory has an isolated limit point x∗ that it only reaches as
time approaches infinity. In such a case, for any ε positive and arbitrarily small there exists a
time tm such that the distance |x(ti ) − x∗ | < ε for all i > m.
Another form of isolated limit point is best illustrated by a rosette with all its infinitely many
petals having a common point. Although the trajectory may reach the neighborhood |x(ti )−x∗ | <
ε of the limit point x∗ at some (finite or infinite) time tm , it goes away, yet it comes back to
same neighborhood and then leaves it and comes back an infinite number of times.
Last form of limit point is best illustrated by the trajectory that reaches the neighborhood
|x(ti ) − x∗ | < ε of the limit point x∗ at some (finite or infinite) time tm , then goes away and
comes back to same neighborhood an infinite number of times, yet keeps doing it via similar
points, namely, all point it meets on its way are limit points. This is an example of limit cycle,
a collection of un-isolated limit points.
However, attacking the problem from the point of view of limit points has required introducing
tough conditions that, while for a very long time deemed absolutely needed for the proofs of
the various theorems, new investigations raised question about their necessity and gradually
managed to mitigate some of the initial conditions. Finally, this paper attempts to eliminate all
those additional conditions, from Barbalat through LaSalle and ending with the new Theorem
of Stability and making it a direct extension of Lyapunov Theorem to the case of negative
semidefinite Lyapunov derivative.
It is important to emphasize that for all following developments we only consider bounded
trajectories x(t) that are not necessarily continuous yet that are piece-wise continuous. We also
consider Lyapunov functions V (x, t) that are continuous functions of x(t) yet, like x(t), are
(x,t) (x,t)
only piece-wise continuous in t. Similarly, we assume that the derivative dV dt = ∂V ∂t +
∂V (x,t) dx
∂x dt
= V̇ (x, t) along the trajectories of (1) is a continuous function of x. The main argument
for all following developments is the fact that for any point p along any bounded trajectory, the
selected function V(p,t) is bounded from below. If the point p is reached at some time t, then for
∫t
any p there exist a constant M, 0 ≤ M < ∞, such that V (p, t) = V (x(0)) + 0 V̇ (x(τ ), τ )dτ ≥
−M > −∞ and this remains true even if p is only reached as time tends to infinity when,
although the integral becomes improper, its value remains finite. In this case, for any limit point

September 1, 2015 DRAFT


28

∫∞
p there exist a constant N, 0 ≤ N < ∞, such that V (p) = V (x(0)) + 0
V̇ (x(τ ), τ )dτ ≥ −N >
−∞
Now, consider any bounded trajectory of (1) and let V̇ (x, t) ≤ 0. For any bounded trajectory
∫t
and for any finite time t1 , one gets V (x(t)) = V (x(t1 )) + t1 V̇ (x(τ ), τ )dτ . Here, the eventual
negativity of V̇ (x, t) ≤ 0 proves to be as strong as the uniform continuity that Barbalat’s Lemma
required and ultimately obviates the need for uniform continuity. Thus, V̇ (x, t) cannot be negative
along any other than finite lengths of time. Otherwise, lim (V (x(t), t)) would tend to −∞ unless
t→∞
for any ϵ positive and arbitrarily small there exists some finite time t2 such that |V̇ (x, t)| ≤ ϵ for
any t ≥ t2 . Thus, ultimately, V̇ (x, t) tends to zero all along any bounded trajectory. Furthermore,
V̇ (x, t) tends to zero in order to stay zero, so all dynamics vanishes along with V̇ (x, t). In other
words, all limit points of any bounded trajectory ultimately belong to Ωi .
For more complex cases, when due to the presence of transient terms the Lyapunov derivative
cannot be called even negative semidefinite, a Modified Theorem of Stability has been formulated.
We assume that V̇ (x, t) = W1 (x, t)+W2 (x, t). As before, we assume that W1 (x, t) is the negative
semidefinite term, W1 (x, t) ≤ 0, and that it is uniformly continuous in x (not in t, though). The
transient term W2 (x, t) is bounded for any bounded x and vanishes as t → ∞. Because of
this transient term, the Lyapunov derivative is not even negative semidefinite, so all previous
conditions seem to be violated.
Nevertheless, along any bounded trajectory and for any finite time t1 , one still gets V (x(t)) =
∫t
V (x(t1 ))+ t1 V̇ (x(τ ), τ )dτ . We want to prove that, as W2 (x, t) ultimately vanishes, W2 (x, t) also
ultimately vanish. However, because of the effect of indefinite W2 (x, t), this may not happen.
Still, if W1 (x, t) does not tend to vanish, the only other alternative is that the trajectory may
asymptotically approach a point or a domain where W1 (x, t) ≤ −d1 < 0 for some finite d1 > 0.
In such a case, because the bounded W2 (x, t) does eventually vanish, there exists a finite time
t1 such that V̇ (x, t) = W1 (x, t) + W2 (x, t) < −d2 < 0 for all t ≥ t1 and for some d2 , where
∫t
0 < d2 ≤ d1 . In this case, V (x(t)) = V (x(t1 )) + t1 V̇ (x(τ ), τ )dτ and thus V (x(t)) → −∞ as
t → ∞, which contradicts the lower boundedness assumption.
Therefore, the negative semidefinite part of the Lyapunov derivative cannot be negative along
any trajectory where it is supposed to lie for an unbounded amount of time. In conclusion all
bounded trajectories asymptotically approach the domain Ωw = {x| lim (W1 (x, t)) ≡ 0}.
t→∞

September 1, 2015 DRAFT

View publication stats

You might also like