Lecture 2

Remark 4.1
Unlike Lyapunov theorems, LaSalle’s theorem does not require the function V ( x ) to be
positive definite.

Most often, our interest will be to show that x ( t ) → 0 as t → ∞ . For that we will need to
establish that the largest invariant set in E is the origin, that is: M = {0} . This is done by
showing that no solution can stay identically in E other than the trivial solution x ( t ) ≡ 0 .

Theorem 4.1 (due to Barbashin and Krasovskii)

Let x = 0 be an equilibrium point for (3.1). Let V : D → R be a continuously
differentiable positive definite function on a domain D ⊂ R n containing the origin, such
that V ( x ( t ) ) ≤ 0 in D . Let S = { x ∈ D : V ( x ) = 0} and suppose that no other solution
can stay in S , other than the trivial solution x ( t ) ≡ 0 . Then the origin is locally
asymptotically stable. If, in addition, V ( x ) is radially unbounded then the origin is
globally asymptotically stable.

Note that if V ( x ) is negative definite then S = {0} and the above theorem coincides with
the Lyapunov 2nd theorem. Also note that the LaSalle’s invariant set theorems are
applicable to autonomous system only.

Example 4.2
Consider the 1st order system
x = a x+u
together with its adaptive control law
u = − kˆ ( t ) x
The dynamics of the adaptive gain k̂ ( t ) is
k̂ = γ x 2
where γ > 0 is called the rate of adaptation. Then the closed-loop system becomes:
(
⎧ x = − kˆ ( t ) − a x
⎪ )
⎨
⎪⎩ kˆ = γ x 2
The line x = 0 represents the system equilibrium set. We want to show that the
trajectories approach this equilibrium set, as t → ∞ , which means that the adaptive
controller regulates x ( t ) to zero in the presence of constant uncertainty in a . Consider
the Lyapunov function candidate:

13
 ( )1
V x, kˆ = x 2 +
2
1 ˆ
2γ
k −b ( )
where b > a . Time derivative of V along the trajectories of the system is given by

( ) 1
( ) ( ) ( )
V x, kˆ = x x + kˆ − b kˆ = − x 2 kˆ − a + kˆ − b x 2 = − x 2 ( b − a ) ≤ 0
γ
Since V ( x, kˆ ) is positive definite and radially unbounded function, whose derivative

V ( x, kˆ ) ≤ 0 is semi-negative, the set Ω = {( x, kˆ ) ∈ R : V ( x, kˆ ) ≤ c} is compact,

c
2

positively invariant set. Thus taking Ω = Ωc , all the conditions of LaSalle’s Theorem are
satisfied. The set E is given by E = {( x, kˆ ) ∈ Ω : x = 0} . Because any point on the line
c

x = 0 is an equilibrium point, E is an invariant set. Therefore, in this example M = E .

From LaSalle’s Theorem we conclude that every trajectory starting in Ωc approaches E ,
( )
as t → ∞ , that is x ( t ) → 0 as t → ∞ . Moreover, since V x, kˆ is radially unbounded,
the conclusion is global, that is it holds for all initial conditions ( x ( 0 ) , kˆ ( 0 ) ) because the
constant c in the definition of Ω can be chosen large enough that ( x ( 0 ) , kˆ ( 0 ) ) ∈ Ω .
c c

5. Boundedness and Ultimate Boundedness

Consider the nonautonomous system
x = f (t, x ) (5.1)
where f :[ 0, ∞ ) × D → R n is piecewise continuous in t , locally Lipschitz in x on
[0, ∞ ) × D , and D ⊂ R n is a domain that contains the origin x = 0 .

Note that if the origin is an equilibrium point for (5.1) then by definition:
f ( t , 0 ) = 0, ∀ t ≥ 0 . On the other hand, even if there is no equilibrium at the origin,
Lyapunov analysis can still be used to show boundedness of the system trajectories. We
begin with a motivating example.

Example 5.1
Consider the IVP with nonautonomous scalar dynamics
x = − x + δ sin t
(5.2)
x ( t0 ) = a > δ > 0
The system has no equilibrium points. The IVP explicit solution can be easily found and
shown to be bounded for all t ≥ t0 , uniformly in t0 , that is with a bound b independent of
t0 . In this case, the solution is said to be uniformly ultimately bounded (UUB), and b is
called the ultimate bound, (prove it).

14
The UUB property of (5.2) can be established via Lyapunov analysis and without using
x2
the explicit solution of the state equation. In fact, starting with V ( x ) = , we can
2
calculate the time derivative of V , along the system trajectories.
V ( x ) = x x = x ( − x + δ sin t ) = − x 2 + δ x sin t ≤ − x 2 + δ x = − x ( x − δ )
It immediately follows that
V ( x ) < 0, ∀ x > δ
In other words, the time derivative of V is negative outside of the compact set
Bδ = { x ≤ δ } , or equivalently, all the solutions that start outside of Bδ will enter this set
within a finite time, and will remain inside the set for all future times. Formally, it can be
δ2
stated as follows. Choose c > . Then all the solutions that start in the set
2
Bc = {V ( x ) ≤ c} ⊃ Bδ
x ≤ 2c

will remain therein, for all future time. The latter takes place since V is negative on the
boundary. Hence, the solutions are uniformly bounded. Moreover, an ultimate bound of
these solutions can also be found. Choose ε such that
δ2
<ε <c
2
Then V is negative in the annulus set {ε ≤ V ( x ) ≤ c} , which implies that in this set
V ( x ( t ) ) will decrease monotonically in time, until the solution enters the set
{V ( x ) ≤ ε } . From that time on, the trajectory will remain in the set, because V is
x2
negative on the boundary: V ( x ) = ε . Since V ( x ) = , we can conclude that this
2
solution is UUB with the ultimate bound x ≤ 2 ε .

Definition 5.1

The solutions of (5.1) are

• uniformly bounded if there exists a positive constant c , independent of t0 ≥ 0 , and
for every a ∈ ( 0, c ) , there is β = β ( a ) > 0 , independent of t0 , such that

x ( t0 ) ≤ a ⇒ x ( t ) ≤ β , ∀ t ≥ t0 (5.3)

• globally uniformly bounded if (5.3) holds for arbitrarily large a

15
 • uniformly ultimately bounded with ultimate bound b if there exist positive
constants b and c , independent of t0 ≥ 0 , and for every a ∈ ( 0, c ) , there is
T = T ( a, b ) , independent of t0 , such that

x ( t0 ) ≤ a ⇒ x ( t ) ≤ b, ∀ t ≥ t0 + T (5.4)

• globally uniformly ultimately bounded if (5.4) holds for arbitrarily large a .

Figure 5.1: UUB Concept

In the definition above, the term uniform indicates that the bound b does not depend on
t0 . The term ultimate indicates that boundedness holds after the lapse of a certain time
T . The constant c defines a neighborhood of the origin, independent of t0 , such that all
trajectories starting in the neighborhood will remain bounded in time. If c can be chosen
arbitrarily large then the UUB notion becomes global.

Basically, the notion of UUB can be considered as a “milder” form of stability in the
sense of Lyapunov (SISL). A comparison between SISL and UUB concepts is given
below.

• SISL is defined with respect to an equilibrium, while UUB is not.

• Asymptotic SISL is a strong property that is very difficult to achieve in practical
dynamical systems.
• SISL requires the ability to keep the state arbitrarily close to the system
equilibrium by starting sufficiently close to it. This is still too strong a
requirement for practical systems operating in the presence of unknown
disturbances.

16
 • The main difference between UUB and SISL is that the UUB bound b cannot be
made arbitrarily small by starting closer to the equilibrium or the origin. In
practical systems, the bound b depends on disturbances and system uncertainties.

To demonstrate how Lyapunov analysis can be used to study UUB, consider a

continuously differentiable positive definite function V ( x ) . Choose 0 < ε < c . Suppose
that the sets Ωε = {V ( x ) ≤ ε } and Ωc = {V ( x ) ≤ c} are compact. Let
Λ = {ε ≤ V ( x ) ≤ c} = Ω c − Ωε
and suppose that it is known that the time derivative of V ( x ( t ) ) along the trajectories of
the nonautonomous dynamical system (5.1) is negative definite inside Λ , that is
V ( x ( t ) ) ≤ −W ( x ( t ) ) < 0, ∀ x ∈ Λ, ∀ t ≥ t0
where W ( x ( t ) ) is a continuous positive definite function. Since V is negative in Λ , a
trajectory starting in Λ must move in the direction of decreasing V ( x ( t ) ) . In fact, it can
be shown that in the set Λ the trajectory behaves as if the origin was uniformly
asymptotically stable, (which it does not have to be in this case). Consequently, the
function V ( x ( t ) ) will continue decreasing until the trajectory enters the set Ωε in finite
time and stays there for all future time. Hence, the solutions of (5.1) are UUB with the
ultimate bound b = max x . A sketch of the sets Λ , Ωc , Ωε is shown in Figure 5.2.
x∈Ωε

Figure 5.2: UUB by Lyapunov Analysis

In many problems, the relation V ( t , x ) ≤ −W ( x ) is derived and shown to be valid on a

domain which is specified in terms of x . In such cases, the UUB analysis involves
finding the corresponding domains of attraction and an ultimate bound.

6. Barbalat’s Lemma and Invariance – like Theorems

17
For autonomous systems, LaSalle’s invariance set theorems allow asymptotic stability
conclusions to be drawn even when V is only negative semi-definite in a domain Ω . In
that case, the system trajectory approaches the largest invariant set E , which is a subset
of all points x ∈ Ω where V ( x ) = 0 . However the invariant set theorems are not
applicable to nonautonomous systems. In the case of the latter, it may not even be clear
how to define a set E , since V may explicitly depend on both t and x . Even when
V = V ( x ) does not explicitly depend on t the nonautonomous nature of the system
dynamics precludes the use of the LaSalle’s invariant set theorems.

Example 6.1

The closed-loop error dynamics of an adaptive control system for 1st order plant with one
unknown parameter is
e = −e + θ w ( t )
θ = −e w ( t )
where e represents the tracking error and w ( t ) is a bounded function of time t . Due to
the presence of w ( t ) , the system dynamics is nonautonomous. Consider the Lyapunov
function candidate
V ( e,θ ) = e 2 + θ 2
Its time derivative along the system trajectories is
V ( e,θ ) = 2 e e + 2θ θ = 2 e ( −e + θ w ( t ) ) + 2θ ( −e w ( t ) ) = −2 e 2 ≤ 0
This inequality implies that V is a decreasing function of time, and therefore, both e ( t )
and θ ( t ) are bounded signals of time. But due to the nonautonomous nature of the
system dynamics, the LaSalle’s invariance set theorems cannot be used to conclude the
convergence of e ( t ) to the origin.

In general, if V ( t , x ) ≤ −W ( x ) ≤ 0 then we may expect that the trajectory of the system

approaches the set {W ( x ) = 0} , as t → ∞ . Before we formulate main results, we state a
lemma that is interesting in its own sake. The lemma is an important result about
asymptotic properties of functions and their derivatives and it is known as the Barbalat’s
lemma.

Definition 6.1 (uniform continuity)

A function f ( t ) : R → R is said to be uniformly continuous if
∀ ε > 0 ∃δ = δ ( ε ) > 0 ∀ t2 − t1 ≤ δ ⇒ f ( t2 ) − f ( t1 ) ≤ ε (6.1)
Note that t1 and t2 play a symmetric role in the definition of the uniform continuity.

Lemma 6.1 (due to Barbalat)

18
Let f ( t ) : R → R be differentiable and has a finite limit as t → ∞ . If f ( t ) is uniformly
continuous then f ( t ) → 0 as t → ∞ .

Lemma 6.2
If f ( t ) is bounded then f ( t ) is uniformly continuous.

An immediate and practical corollary of Barbalat’s lemma can now be stated.

Corollary 6.1
If f ( t ) : R → R is twice differentiable, has a finite limit, and its 2nd derivative is bounded
then f ( t ) → 0 as t → ∞ .

In general, the fact that derivative tends to zero does not imply that the function has a
limit. Also, the converse is not true.
f (t ) → C ⇔ f (t ) → 0

Example 6.2
cos ( ln t )
• As t → ∞ , f ( t ) = sin ( ln t ) does not have a limit, while f ( t ) = →0.
t
• As t → ∞ , f ( t ) = e −t sin ( e 2 t ) → 0 , while f ( t ) = −e − t sin ( e 2 t ) + et sin ( e 2 t ) → ∞ .

Example 6.3
Consider an LTI system
x = Ax + Bu
with a Hurwitz matrix A and a uniformly bounded in time input u ( t ) . These two facts
imply that the state x ( t ) is bounded. Thus, the state derivative x ( t ) is bounded. Let
y = C x represent the system output. Then y = C x is bounded and, consequently the
system output y ( t ) is a uniformly continuous function of time. Moreover, if the input
u ( t ) = u0 is constant then the output y ( t ) tends to a limit, as t → ∞ . The latter
combined with the fact that y is uniformly continuous implies, (by Barbalat’s Lemma),
that the output time derivative y asymptotically approaches zero.

To apply Barbalat’s lemma to the analysis of nonautonomous dynamic systems we state

the following immediate corollary.

Corollary 6.2 (Lyapunov-like Lemma)

If a scalar function V = V ( t , x ) is such that
• V ( t , x ) is lower bounded
• V ( t , x ) is negative semi-definite along the trajectories of x = f ( t , x )

19
 • V ( t , x ) is uniformly continuous in time
then V ( t , x ) → 0 , as t → ∞ .

Notice that the first two assumptions imply that V ( t , x ) tends to a limit. The latter
coupled with the 3rd assumption proves (using Barbalat’s lemma) the corollary.

Example 6.4
Consider again the closed-loop error dynamics of an adaptive control system from
Example 6.1. Choosing V ( e,θ ) = e 2 + θ 2 , it was shown that along the system trajectories:
V ( e,θ ) = −2 e 2 ≤ 0 . The 2nd time derivative of V is
V ( e,θ ) = −4 e e = −4 e ( −e + θ w ( t ) )
Since w ( t ) is bounded by hypothesis, and e ( t ) and θ ( t ) were shown to be bounded, it is
clear that V is bounded.. Hence, V is uniformly continuous and by the Barbalat’s lemma
(or the Lyapunov-like lemma), V → 0 which in turn indicates that the tracking error e ( t )
tends to zero, as t → ∞ .

Often, the Barbalat Lemma is formulated with respect to an integral of a function, rather
than to the function itself.

Lemma 6.3 (due to Barbalat)

If f : R + → R is uniformly continuous for all t ≥ 0 , and if the limit of the integral

t
lim ∫ f (τ ) dτ (6.2)
t →∞
0

exists and finite, then

lim f ( t ) = 0 (6.3)
t →∞

Definition 6.2
The L p norm of a function f : R + → R is defined as:
1
⎛∞ ⎞p
f (t ) = ⎜ ∫ f (τ ) dτ ⎟
p
p
(6.4)
⎝0 ⎠
for all p = [1, ∞ ) , while
f (t ) ∞
= max f ( t ) (6.5)
t ≥0

denotes the L∞ norm of f .

20
 ∞
We say that f ∈ Lp if f ( t ) ∫ f (τ )
2
p
exists and finite. In particular, f ∈ L2 if dτ < ∞ .
0

The next statement directly follows from Lemma 6.3, if one chooses f ( t ) = g ( t ) .
p

Corollary 6.2
If g , g ∈ L∞ and g ( t ) ∈ Lp , for some p = [1, ∞ ) , then lim g ( t ) = 0 .
t →∞

21
Part II: System ID Methods

1. Introduction
Read [Ioannou], Chapters 1 – 2.

The goal of any parameter estimation method is to infer the values of uncertain system
parameters from the measurements of input and output signals of the system. Parameter
estimation can be done either on-line or off-line. The latter may be preferable if the
parameters are constant and there is a sufficient time for estimation to be performed.
However, for parameters that vary during operation, on-line parameter estimation is
necessary to keep track of the parameter values. Since problems in robust and adaptive
control context usually involve time-varying parameters, on-line estimation methods are
thus more relevant.

In what follows, several basic methods of on-line estimation will be presented. Note that,
although the main purpose of the on-line estimators may be to provide parameter
estimates for adaptive / self-tuning control, they can also be used for other purposes, such
as system health monitoring and failure detection.

2. Static Linear in Parameters Model

An estimation model that relates the available data to the unknown parameters forms the
basis of any parameter estimation method. A quite general model for parameter
estimation applications is in the linear parameterization form:
y ( t ) = ΘT Φ ( x ( t ) ) + ε ( t ) (2.1)
where y ( t ) ∈ R m , x ( t ) ∈ R n are the system output and input signals, the ( N × m ) – matrix
Θ ∈ R N ×m contains the unknown parameters to be estimated, Φ ( x ( t ) ) ∈ R N (called the
“regressor”) represents the N – dimensional vector of chosen basis functions, and
ε ( t ) ∈ R m denotes the non-parametric uncertainties in the system (such as noise,
modeling errors, etc). Note that in (2.1) both y ( t ) and Φ ( x ( t ) ) are known quantities.
This implies that (2.1) is simply is a system of linear equations in terms of the unknown
Θ . For every time instant t there is such a system. So if continuous measurements of
y ( t ) and Φ ( x ( t ) ) are available on-line then, in general, there will be more equations
than the unknowns. However, this will facilitate parameter estimation in the presence of
inevitable noise and modeling error.

In off-line estimation, one collects the data points of y ( t ) and Φ ( x ( t ) ) for a period of
time, and solves the equations once and for all. In on-line estimation, equations are

22
solved recursively, implying that the estimated value of Θ ˆ ( t ) is updated once a new
ˆ =Θ

set of data y ( t ) and Φ ( x ( t ) ) becomes available.

Static linear in parameters model (2.1), although simple, is actually quite general, as seen
from the following example.

Example 2.1: Linear parameterization of 1st order dynamics

Consider the 1st order system

⎧⎪ x = − x + a x + b u + ξ ( t )
⎨ (2.2)
⎪⎩ x ( 0 ) = 0
where x , u are the scalar state and input respectively, a and b are the unknown
constants to be identified on-line using the available measurements of x and u . Also in
(2.2), ξ ( t ) represent bounded unknown disturbance, (such as process noise). The very
first step in the design of an on-line parameter identifier is to lump the unknown
parameters a and b in a vector, that is one needs to rewrite (2.2) in the form of (2.1).
Toward that end, one can easily integrate (2.2) and get:
t t t
x ( t ) = x ( 0 ) + a ∫ e −(t −τ ) x (τ ) d τ + b ∫ e−( t −τ )u (τ ) d τ + ∫ e−( t −τ )ξ (τ ) d τ (2.3)
0 0 0 0
x f (t ) u f (t ) ε (t )

or, equivalently:
x (t ) = a x f (t ) + b u f (t ) + ε (t ) (2.4)
Finally, the system can be written in the desired form of (2.1) :

⎛ x f (t ) ⎞
x = ( a b ) ⎜⎜ ⎟ + ε (t ) (2.5)
⎝ u f ( t ) ⎟⎠
ΘT
Φ

Example 2.2: Linear parameterization of aircraft pitch dynamics

Consider the following pitch dynamics of a rigid-body, fixed-wing aircraft.

IY q = M 0 (α ) + M q q + M δ δ + p r ( I z − I x ) + r 2 − p 2 I x z ( ) (2.6)
M aero M IC ( p , r )

where IY is the pitch inertia, ( q, q ) are the pitch acceleration and angular rate. Also in
(2.6), M aero denotes the aircraft total aerodynamic pitching moment which is comprised
of the baseline component M 0 (α ) , pitch damping M q q , and pitch increment due to
elevator deflection M δ δ . Finally, M IC ( p, r ) represents the sum of cross-coupling terms

23
that are due to the angular roll and yaw rates, ( p, r ) , scaled by the aircraft inertia
( )
components I x , I z , I x z .

Suppose that the aerodynamic pitching moment components M 0 (α , q ) , M q , M δ are

unknown and need to be estimated on-line using available input-output measurements
(α , p, q, r , δ ) and q , respectively. It is further assumed that the aircraft inertia terms
are constant and known. Consequently, the inertial coupling term M IC ( p, r ) in the right
hand side of (2.6) is known, as well. Rewriting (2.6) using aerodynamic coefficients
yields:
IY ⎛ M IC ⎞ qc
⎜q − ⎟ = C M 0 (α ) + C M q + CM δ δ (2.7)
QSc ⎝ IY ⎠ 2V
In (2.7), Q denotes the dynamics pressure, S is the aircraft wing reference area, and c is
the mean geometric chord. For simplicity, the unknown pitch control effectiveness CM δ
and the pitch rate damping coefficient CM q are assumed to be constant. On the other
hand, the uncertain aerodynamic pitching moment component CM 0 (α ) depends on the
aircraft angle of attack α . Suppose that the pitching moment can be written as:
CM 0 (α ) = θ MT Φ (α ) + ε M 0 (α ) (2.8)
where θ M ∈ R N is the N – dimensional vector of unknown constant parameters, and
Φ (α ) ∈ R N is the N – dimensional regressor vector that consists of appropriately chosen
basis functions, such as splines, polynomials, RBF-s, etc. Also in (2.8), ε M 0 (α ) is the
non-parametric approximation error that can be made small by choosing a sufficiently
large number of basis functions N. Substituting (2.8) into (2.7) gives the desired linear in
parameters model.
IY ⎛ M IC ⎞ qc
⎜q − ⎟ = θ M Φ (α ) + C M q + CMδ δ + ε M 0 (α )
T
(2.9)
QSc ⎝ IY ⎠ 2V
or equivalently:
⎛ Φ (α ) ⎞
⎜ ⎟
IY ⎛ M IC ⎞
QSc ⎝
⎜q −
IY ⎠
(
⎟ = θ M CM q CM δ ⎜
T
) ⎜ qc ⎟
2V ⎟
+ ε M0 (2.10)
ΘT ⎜ ⎟
y(t )
⎝ δ ⎠
Ψ (α , q ,V , δ )

Note that the error term ε M 0 may contain other non-parametric uncertainties in the
system, such as measurement noise, numerical inaccuracies etc. As a result, the aircraft
pitch dynamics relation (2.6) is represented using the linear-in-parameters model:
y ( t ) = ΘT Ψ ( t ) + ε M 0 ( t ) (2.11)

24
Note that the static model (2.11) requires knowledge of the pitch acceleration signal q .
Most often in real-life applications, vehicle rotational accelerations are not available as
on board measurements. In this case, an alternative modeling approach can be utilized.
We rewrite (2.6) in the form,
⎡ M (α , q, δ ) + M IC ( p, r ) ⎤
q = −a f q + ⎢ a f q + aero ⎥ (2.12)
⎣ I Y ⎦
where a f > 0 , and integrate both sides, (assuming zero initial conditions).
t t
M aero (α (τ ) , q (τ ) , δ (τ ) ) t
M IC ( p (τ ) , r (τ ) )
q ( t ) = a f ∫ e −( t −τ ) q (τ ) d τ + ∫ e − (t −τ ) d τ + ∫ e− ( t −τ ) dτ (2.13)
0 0
IY 0
IY
q f (t ) M aero f (t ) M IC f (t )

Then,
q ( t ) − M IC f ( t ) − a f q f ( t ) = M aero f ( t ) (2.14)
y (t )

where y ( t ) is the known output, while M aero f ( t ) represents the vehicle’ filtered pitch
aerodynamics. This term can be modeled as a linear in unknown parameters function.
⎛ Φ (α ) ⎞
⎜ ⎟
M aero f = θ M CM q CMδ
T ⎜
(qc ⎟ QS c
⎜ 2 V ⎟ IY
+ε f )
(2.15)
ΘT ⎜ ⎟
⎝ δ ⎠
Ψ (α , q ,V , δ )

Thus, again we arrive at a static model in the form of (2.11).

Remark 1: In some practical applications, using a filtered version of the derivative, such
⎛ s ⎞
as q f = ⎜
⎜ τ q s + 1 ⎟⎟
q , in place of the derivative q works well because the signal q that is
⎝ ⎠
being filtered is continuous. Therefore, the signal energy is at low frequencies, where the
1
specified filter is a good approximation to the derivative, ( defines the bandwidth of
τq
the derivative filter). It is evident that the derivative approximation error can be folded
into the bounded non-parametric uncertainty vector ε M 0 .

25