Math. Control Signs Syst.

(2006) 18:369394
DOI 10.1007/s00498-006-0005-1
ORI GI NAL ARTI CLE
A performance comparison of robust adaptive
controllers: linear systems
Ahmad Sanei Mark French
Received: 10 March 2003 / Revised: 5 April 2005 /
Published online: 25 July 2006
Springer-Verlag London Limited 2006
Abstract We consider robust adaptive control designs for relative degree one,
minimum phase linear systems of known high frequency gain. The designs are
based on the dead-zone and projection modications, and we compare their
performance w.r.t. a worst case transient cost functional with a penalty on the
L

norm of the output, control and control derivative. We establish two quali-
tative results. If a bound on the L

norm of the disturbance is known and the

known a priori bound on the uncertainty level is sufciently conservative, then
it is shown that a dead-zone controller outperforms a projection controller. The
complementary result shows that the projection controller is superior to the
dead-zone controller when the a priori information on the disturbance level is
sufciently conservative.
Keywords Robust adaptive control Non-singular performance
1 Introduction
It is well known that adaptive controllers are suitable for systems whose nomi-
nal mathematical model contains an uncertain parameter . A common feature
of adaptive designs is the construction of a time varying parameter

() whose
value is controlled by an adaptive law. In contrast with most adaptive con-
trol mechanisms which would attempt to identify or estimate the uncertain
A. Sanei M. French (B)
School of Electronics and Computer Science, University of Southampton,
Southampton SO17 1BJ, UK
e-mail: [email protected]
A. Sanei
e-mail: [email protected]
370 A. Sanei, M. French
parameter of the plant by a parameter estimator

(), the objective of a
non-identier-based adaptive controller is to use certain information about the
plant to nd suitable methods of systemregulation. In other words, the adaptive
law makes no attempt to identify or estimate the unknown plant parameter ,
but merely attempts to seek out a stabilising value for the so-called tuning
function

(). See e.g. [12, 20] and [5] for an overview.
However, this method, like other adaptive controllers, is susceptible to phe-
nomena such as parameter drift even when small disturbances are present. To
overcome such problems, a number of standard techniques are widely utilised,
such as dead-zones, modication, projection modication, etc. [13].
Each of these techniques has advantages and drawbacks. For example, dead-
zone modications typically require a priori knowledge of the disturbance level,
and only achieve convergence of the state/output/error to some pre-specied
neighbourhood of the origin (whilst keeping all signals bounded). In partic-
ular if the disturbance vanishes, then the dead-zone controller does not typ-
ically achieve convergence of the output to zero, the convergence remains
to the pre-specied neighbourhood of the origin. On the other hand, projec-
tion modications generally achieve boundedness of all signals, and further-
more have the desirable property that if no disturbances are present, then the
state/output/error converges to zero, however, an arbitrarily small L

distur-
bance can completely destroy any convergence to zero.
This illustrates that in the case of asymptotic performance, there are some
known characterisations of good and bad behaviour. However, there are
many situations in which we cannot denitively state whether a projection or
dead-zone controller is superior. Although the asymptotic performance of such
modications has often been investigated, very little of this research is con-
cerned with transient performance. Furthermore, the known results, see e.g.
[9], as with most results in adaptive control, are conned to singular perfor-
mance, i.e. without any consideration of the control signal. The primary results
of this paper take control effort into account.
The present paper extends the line of our work on developing a comparison
theory of robust adaptive controllers [2, 17, 21]. We compare dead-zone and
projection-based adaptive controllers for nite dimensional, minimum phase,
linear systems of relative degree one with positive high frequency gain. The
comparison has been made with respect to a worst case non-singular transient
cost functional P penalising the state (x), the input (u) and the derivative of the
input ( u) of the plant. We will identify circumstances in which the dead-zone
controller is superior to the projection controller with respect to P and vice
versa.
The paper is structured as follows. In Sect. 2 we introduce the system class
and the basic adaptive controller. Section 3 denes two standard classes of
robust modications to the basic adaptive controller, namely the dead-zone
modication and the projection modication, and states their main properties.
The proofs of these results are deferred to Appendices 1 and 2. Our main results
are presented in Sect. 4, where situations are described in which the dead-zone
controller outperforms the projection controller and vice-versa. Sections 5 and
A performance comparison of robust adaptive controllers 371
6 contain the proofs of the results of Sect. 4, and Sect. 7 contains a short discus-
sion on alternative choices of dead-zones. Section 8 contains a brief summary
and conclusions.
2 System and basic control design
Suppose is a SISO linear time invariant plant described by
y =
b
n1
s
n1
+b
n2
s
n2
+ +b
0
s
n
+a
n1
s
n1
+ +a
0
(u +d), (1)
where a
i
, b
i
R, 0 i n 1, are unknown constants and d() belongs to
a class of bounded disturbances D L

= L

[R
0
] = L

[0, ). Note that

for notational convenience, we adopt the standard shorthand of letting u, y, d
denote both the time domain signal and corresponding element in frequency
domain. We assume that only the output y() is available for measurement.
Consider the following assumptions:
C1. The plant is minimum phase, i.e. b
n1
s
n1
+ +b
0
is Hurwitz.
C2. The plant order n is known, and the high frequency gain is positive (i.e.
b
n1
> 0).
It is convenient for the analysis in this paper to consider two different state space
representation of Eq. 1. The observer canonical form is obtained as follows:
(x
0
, , d()) : x(t) = Ax(t) +B(u(t) +d(t)) x(0) = x
0
R
n
, d D,
y(t) = Cx(t),
(2)
in which x(t), B, C
T
R
n
, A R
nn
, and
A =

a
n1
1 0 . . . 0
a
n2
0 1 0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
a
1
0 0 . . . 1
a
0
0 0 . . . 0

, B =

b
n1
.
.
.
b
1
b
0

, C =
_
1 0 0
_
, 3 (3)
where
= (a
0
, . . . , a
n1
, b
0
, . . . , b
n1
) S, S : ={ R
2n
| (, , ) satises C1,C2 },
(4)
and S represents the uncertain system parameters. We emphasise that by
non-identier-based control, we will not be estimating unknown parameter .
372 A. Sanei, M. French
The second useful state space form is obtained by utilizing the coordinate
transformation matrices:
S :=
_
B(CB)
1
, T
_
, S
1
=
_
C
T
, N
T
_
T
, (5)
where
T =
_
0 . . . 0
I
n1
_
R
n(n1)
, N =

b
n2
/b
n1
.
.
. I
n2
b
0
/b
n1

R
(n1)n
.
Applying Eq. 5 to Eq. 2, we have
x(t):=(y(t), z(t)
T
)
T
=S
1
x(t), S
1
BCS=
_
b
n1
0
0 0
_
, S
1
AS=
_
a
1

A
2

A
3

A
4
_
,
(6)
where a
1
R,

A
T
2
,

A
3
R
n1
and

A
4
R
(n1)(n1)
. This gives
y(t) = a
1
y(t) +

A
2
z(t) +b
n1
(d(t) +u(t)) , y(0) = Cx
0
,
z(t) =

A
3
y(t) +

A
4
z(t), z(0) = Nx
0
.
(7)
It has been shown that

A
4
is stable [5], i.e. there exists a positive denite
matrix R = R
T
> 0 satisfying the Lyapunov equation
R

A
4
+

A
T
4
R = I
n1
. (8)
Note that there exists k

> 0 such that for all k k

,
(S
1
(AkBC)S)
T
P +P(S
1
(AkBC)S) Q,
where the symmetric positive denite matrices P, Q are dened as
P =
_
1
2
0
0 R
_
, Q =
_
1 0
0
1
2
I
n1
_
. (9)
This can be shown by observing that for all k > k

and for all x R

n
,
x
T
_
(S
1
(AkBC)S)
T
P +P(S
1
(AkBC)S)
_
x
= (b
n1
k a
1
)y
2
+y
_

A
2
+2

A
T
3
R
_
z z
2
(b
n1
k M)y
2

1
2
z
2
,
x
T
Q x
A performance comparison of robust adaptive controllers 373
where
M := | a
1
| +
_

A
2
+2R

A
3

_
2
/2, k

:= (M+1/2)/b
n1
. (10)
Therefore, by considering the Lyapunov function V( x) = x
T
P x, it follows
that if S then the system 2 is high-gain stabilizable, i.e. it is stabilized by the
feedback u = ky for sufcient large k.
It was shown by Willems and Byrnes [20] that disturbance free (D = {0})
systems of the form 2, i.e. (x
0
, , 0) which satisfy C1,C2, are stabilised by the
following simple adaptive high-gain controller:
: u(t) =

(t) y(t),

(t) = y(t)
2

(0) = 0.
(11)
Special features of this strategy are its simplicity and the absence of any plant
identication mechanism. This idea formed the foundation of the theory of
universal adaptive control which has continued to develop since 1985 in the
areas of linear systems, nonlinear systems, and innite dimensional systems, see
for example [7, 11, 15, 18] for some representative papers, and [4] for an early
survey. These controllers have also been considered for application studies
[3, 10].
3 Robust modications to the control design
For systems of the form 2, it is well known that even a small L

disturbance
can cause a drift of the tunable function

(), see e.g. [1, 13]. To overcome this
problem, two distinct approaches have been proposed [13]: (1) using an appro-
priately rich reference input, and (2) modifying the adaptation law. In this
section, we briey explain two common methods in modifying the adaptive law,
in particular dead-zone and parameter projection modications (see e.g. [13]
for details).
Whilst a subset of results on asymptotic behaviour of dead-zone and pro-
jection modications presented in this section is known, we remark that the
boundedness properties of the closed loop signals (properties D2 and P2 of
Theorems 3.1 and 3.2) do not appear in the literature, to the best knowledge of
the authors. Moreover, these properties play a key role in the main results.
3.1 Dead-zone modication
Consider the unmodied adaptive law of the form

(t) = y(t)
2
,

(0) = 0. The
idea of the dead-zone is to modify the adaptive law so that the adaptive mech-
anism is switched off when system output y() lies inside a region where the
disturbance has a destabilising effect on the dynamics [14]. A priori knowledge
374 A. Sanei, M. French
of the size of the disturbance is typically used to dene the size of the dead-
zone. Let d
max
be the a priori known upper bound of the disturbance level,
i.e. d
max
d()
L
for all d() D. For SISO output feedback systems 2, the
standard denition of the modied adaptive law is

(t) = y(t)
2

D

(y(t)),

(0) = 0, := d
max
> 0 (12)
where

D

(y) := 0 if y [ , ] and

D

(y) := 1 if y [ , ], > 0. It should

be observed that there are other valid choices for , see Sect. 7 where this is
discussed further.
Note that the discontinuous switching activity of the above denition intro-
duces a r.h.s. discontinuity in the related differential equations resulting in the
need for a more delicate analysis. In this paper, we take the advantages of so-
called smooth dead-zone for which, the local existence and uniqueness of the
solution of the corresponding closed loop system follows directly from the clas-
sical theory of differential equations. However, modulo these technical issues,
the authors expect that similar results to those in this paper can also be obtained
for the standard dead-zone 12. The smooth dead-zone is dened by
D

(y) =
_
0, y [ , ]
|y| , y [ , ]
(13)
where > 0 and leads to the modied adaptive law of the form,

D
(d
max
) : u(t) =

(t)y(t)

(t) = |y(t)| D

(y(t)),

(0) = 0, := d
max
> 0.
(14)
The following theorem establishes the properties of such controllers:
Theorem 3.1 Consider the closed loop system ((x
0
, , d()),
D
(d
max
)) dened
by Eqs. 24 and 14, where x
0
R
n
, S, d() L

and d
max
0. Then the
following properties hold:
D1. There exists a unique solution (x(),

()) : R
0
R
(n+1)
.
D2. There exists a continuous function M : R
n
S R
0
R
0
R
0
such
that the closed loop signals x(),

(), u() satisfy
x()
L
+

()
L
+u()
L
M(x
0
, , d()
L
, d
max
)
D3. lim
t
inf
x[d
max
,d
max
]
|y(t) x| = 0.
Proof See Appendix 1.
A performance comparison of robust adaptive controllers 375
3.2 Projection modication
The projection modication [8] is an alternative method to eliminate parameter
drift by keeping the adaptive parameter within some a priori dened bounds.
In identier-based adaptive control, this can be accomplished by projecting
the parameter estimator into a given compact convex set containing the true
parameter vector. The general denition for such method can be found in, e.g.
[9, p. 511]. In our non-identier-based case, the denition is as follows. Let

max
> k

where k

is dened by Eq. 10. The modied adaptive law is dened

by

(t) =
_
y(t)
2

(t) <
max
0

(t)
max
(15)
Given an output y: [0, ) R, where [0, ] = [0, ) {}, we dene the
projection controller as follows:

P
(
max
) : u(t) =

(t)y(t)

(t) = y(t)
2
,

(0) = 0, t [0, T
m
),

(t) = 0, t [T
m
, ),
(16)
where T
m
= inf{t [0, ] |

(t) =
max
}. Note that T
m
is the rst time instance
that

hits the boundary
max
, and after this time

remains at constant level

max
and hence the adaptive law in Eq. 16 coincides with Eq. 15. We denote the
respective closed loop system by ((x
0
, , d()),
P
(
max
)). The stability of the
closed loop is examined in the following theorem.
Theorem 3.2 Consider the closed loop system ((x
0
, , d()),
P
(
max
)) dened
by Eq. 2 and 16, where x
0
R
n
, S, d() L

and let
max
> k

where k

0
is given by Eq. (10). Then the following properties hold:
P1. There exists a solution (x(),

()) : R
0
R
(n+1)
.
P2. Let V = {(,
max
) S R
>0
| k

<
max
}. Then there exists a continu-
ous function M : R
n
V R
0
R
0
such that the closed loop signals
x(),

(), u() satisfy
x()
L
+

()
L
+u()
L
M(x
0
, ,
max
, d).
Proof See Appendix 2.
4 Statement of the main results
A goal in control theory is to design control laws which achieve good perfor-
mance for any member of a specied class of systems. Consider a system
376 A. Sanei, M. French
which belongs to the set of all admissible systems S

. The performance of a
controller is measured by a cost functional J dependent on some measurable
signals (state/output/input). In this paper, we are interested in a worst case
scenario, i.e. a performance P which is dened over the power set of S

and is
formulated as a supremum of all possible costs. Furthermore, the non-singular
performance measure will penalise the state (x) and the input and its derivative
(u, u) of the plant; specically we will consider cost functionals of the form:
P( (X
0
( ), , D()), )= sup
x
0
X
0
( )
sup

sup
dD()
(x()
L
+u()
L
+ u()
L
) ,
(17)
where
X
0
( ) := {x
0
R
n
| x
0
}, > 0,
D() := {d() L

| d()
L
}, 0,
(18)
and is any compact subset of (), where
() := { S | k

}, 0, (19)
where A, B, C are dened in Eq. 3 and is given by Eq. 4. Note that () is not
bounded, and that there are elements on the boundary of () which do not
satisfy C1, C2 and for which the closed loop is not stable, hence generating an
innite cost. Therefore, the second supremum cannot be taken over (). This
is reected in the bounds obtained in Theorems 3.1 and 3.2 (see relations 45,
52, and 58).
4.1 Main results
Both designs considered require particular types of a priori information on the
system or disturbance environment. The projection-based design requires an
explicit a priori upper bound
max
on the stabilizing value of , whereas the
dead-zone design requires an explicit a priori bound d
max
on the magnitude
of the disturbance signals d D(). We therefore consider the situations where
either
max
or d
max
are large with respect to the true values or , respectively.
This corresponds to the cases where the control designer has made conservative
choices for the projection/dead-zone levels.
The following theorems are the main results of the paper:
Theorem I Consider the system (x
0
, , d()) and the controllers
D
() and

P
() dened by Eqs. 2, 14 and 16 respectively, where x
0
R
n
, S, d L

.
Let , > 0 and = () be compact. Consider the transient performance
cost functional 17. Then for all d
max
, there exists

max
such that for all
A performance comparison of robust adaptive controllers 377
Fig. 1 Statement of the main results: Theorem I (left), Theorem II (right)

max

max
,
P( (X
0
( ), , D()),
P
(
max
) ) > P( (X
0
( ), , D()),
D
(d
max
) ).
This theorem can be interpreted as stating that if the a priori knowledge of
the parametric uncertainty level
max
is sufciently conservative (
max

max
),
then the dead-zone based design will outperform the projection-based design,
see Fig. 1(left).
Theorem II Consider the system (x
0
, , d()) and the controllers
D
() and

P
() dened by Eqs. 2, 14 and 16, respectively, where x
0
R
n
, S, d L

.
Then for all > 0 there exists > 0 such that for all
max
, there exists
d

max
> 0 such that for all d
max
d

max
,
P( (X
0
( ), , D()),
D
(d
max
)) > P( (X
0
( ), , D()),
P
(
max
)),
where P is the transient performance cost functional 17, () is compact
and (0).
This theorem can be interpreted as stating that above a certain uncertainty
level , if the a priori knowledge d
max
of the disturbance level is sufciently
conservative (d
max
d

max
), then the projection design will outperform the
dead-zone design, see Fig. 1 (right).
As an aside, we observe that it is natural to ask whether there is any benet
of adaptive control when a bound
max
is known, as with this information, the
natural static output feedback u(t) =
max
y(t) achieves stabilization. One sce-
nario in which adaptive strategies can be shown to outperformsuch memoryless
feedbacks is when
max
is a conservative estimate for , see, e.g. [2] for such an
argument in a related context. In the context of this paper, similar arguments
also show that one should utilize dead-zone adaptive controllers in preference
to memoryless feedbacks in the scenario whereby
max
is a conservative esti-
mate for , and d
max
is small. In the scenario where d
max
is a conservative
bound on , the motivation for adaptation, e.g. using the projection design over
the memoryless design is less clear-cut. However, if both
max
and d
max
are
conservative estimates of and , respectively, and classes of disturbances are
378 A. Sanei, M. French
considered in which adaptive law does not drift, as for example in the distur-
bance free case, the projection design will in general not cause

to reach the
projection value
max
, and hence can be expected to generate a lower cost than
the high gain memoryless design.
5 Proof of Theorem I
Firstly, we show that P = for the unmodied design 11 (Proposition 5.3).
From this we can show that the projection modication design
P
(
max
) has
the property that P as
max
(Proposition 5.4). Finally we show that
P < , and that the bound is independent of
max
, for the dead-zone design,

D
(d
max
) (Proposition 5.5). This sufces to establish Theorem I.
Proposition 5.1 Consider the closed loop system ((x
0
, , d()), ) dened by
Eqs. 2 and 11, where x
0
R
n
, S. Let d() for some = 0. Then
x(t) 0 as t

(t) as t .
Proof Let (x
T
,

)
T
denote the unique global solution [20] of the closed loop
((x
0
, , d()), ) given by Eqs. 2 and 11.
) Suppose for a contradiction that

(t) . Then

(t)

< , since

() is monotonically increasing by Eq. 11. Therefore, since by assump-

tion x 0 and since the r.h.s. of Eqs. 2and 11 is continuous, it fol-
lows that (x(),

()) (0,

) is an equilibrium point of the closed loop

((x
0
, , d()), ). Hence (x(),

()) (0,

) must be a solution of the

following equations:
x
2
(t) a
n1
x
1
(t) +b
n1
(

(t)x
1
(t)) = 0,
.
.
.
a
0
x
1
(t) +b
0
(

(t)x
1
(t)) = 0, (20)
x
1
(t)
2
= 0.
Since the system is minimum phase, it follows that b
0
= 0. We also have
= 0. Therefore (x(),

()) (0,

) cannot be a solution of Eq. 20, hence

contradiction.
) Let > 0 and recalling that y = Cx, z = Nx, dene
V(t) =
1
2
y(t)
2
+z(t)
T
Rz(t), t 0,
A performance comparison of robust adaptive controllers 379
where R is dened by Eq. 8. The time derivative of V along the solution
of Eqs. 7 and 11 is

V(t) = y(t)
_
( a
1
b
n1

(t))y(t) +

A
2
z(t) +b
n1

_
+2y(t)

A
T
3
Rz(t) +z(t)
T
(

A
T
4
R +R

A
4
)z(t)
= (b
n1

(t) a
1
)y(t)
2
+
b
n1

y(t)
_
+y(t)
_

A
2
+2

A
T
3
R
_
z(t) z(t)
2
(b
n1

(t) M)y(t)
2

1
2
z(t)
2
+
1
2

2
,
where M := | a
1
| +b
2
n1
/2 +
_

A
2
+2R

A
3

_
2
/2. We choose t

> 0
such that b
n1

(t) M > 1/2 for all t t

. Therefore

V(t) V(t) +
1
2

2
t t

,
where = min{1, 1/2(R)}. It follows that
lim
t
sup V(t)

2
2
.
Since is arbitrary, this proves lim
t
V(t) = 0, hencelim
t
(y(t), z(t)
T
) =
0. This completes the proof.

Proposition 5.2 Consider the closed loop system ((x

0
, , d()), ) dened by
Eqs. 2 and 11, where x
0
R
n
, S. Let d() for some = 0. If x() is
uniformly continuous, then
x(t) 0,

(t) as t .
Proof Let (x
T
,

)
T
denote the unique global solution [20] of the closed loop
((x
0
, , d()), ) given by Eqs. 2 and11, and recall that y = Cx, z = Nx. Firstly,
we show that y(t) 0 as t . From this we will prove that

(t) and
nally by Proposition 5.1, we conclude that x(t) 0 as t .
Suppose for a contradiction that y(t) 0 as t . Then there exists M > 0
and a positive sequence {t
k
}
k1
, t
k
as t for which y(t
k
) M. Since,
by assumption, x() is uniformly continuous, it follows that y() is uniformly
continuous, and hence
> 0 s.t. [0, ], and t > 0, |y(t) y(t +)| <
M
2
. (21)
Therefore |y(t
k
)y(t
k
+)| < M/2 and since y(t
k
) M, we have that y(t
k
+) >
M/2 i.e. y(t) M/2, for all t [ t
k
, t
k
+ ]. Without loss of generality, we may
380 A. Sanei, M. French
assume t
k+1
t
k
. It follows that

(t
k
+) =
t
k
+
_
0

()d =
t
k
+
_
0
y
2
()d
M
2
4
k, (22)
so

(t
k
+ ) as k , hence

(t) as t . It follows by Prop-
osition 5.1 that x(t) 0 as t , therefore y(t) 0 as t by Eq. 2,
hence contradiction.
Now we have y(t) = x
1
(t) 0 as t and we claim

(t) as t .
Suppose for contradiction

(t) as t . Then

(t)

< as t ,
since

() is monotonic by Eq. 11. Substitute this into Eq. 2, we have
x
1
(t) = x
2
(t) (a
n1
+

b
n1
)x
1
(t) +b
n1
, (23)
.
.
.
x
n
(t) = (a
0
+

b
0
)x
1
(t) +b
0
, (24)
where by minimumphase assumption, b
i
= 0, i {0, . . . , n1}. Since x
1
(t) 0
as t , Eq. 24 implies that x
n
(t) as t , since x() is uniformly
continuous. It follows that x
n1
(t) as t , and cascading the argument
shows to x
1
(t) as t , hence contradiction. Therefore

(t) as
t . From this and Proposition 5.1, the claim of the Proposition follows.
Proposition 5.3 Consider the closed loop system ((x
0
, , d()), ) dened by
Eqs. 2 and 11, where x
0
R
n
, S, d L

. Let , , > 0 and suppose

= () is compact. Consider the transient performance cost functional
17. Then
P( (X
0
( ), , D()), ) = . (25)
Proof Let lim denote lim sup
t
. Let x
0
X
0
( ) and . Let d
1
() = 0,
and consider ((x
0
, , d
1
), ). Consider the two cases: either 1. lim x(t) =
or 2. lim x(t) < :
1. Suppose lim x(t) = , i.e. lim Ax(t) +Bu(t) +B = . Therefore
either
(a). lim x(t) = , which implies that x()
L
= , or
(b). lim x(t) < , therefore lim u(t) = , i.e. u()
L
= .
2. Suppose lim x(t) < i.e. x() is uniformly continuous. Therefore by
Proposition 5.2

> 0 T > 0 s.t. t > T

(t) >

. (26)
Now we choose d
2
(t) := , t T, d
2
(t) := , for all t > T. Note that
d
2
(t) = d
1
(t) for all t T and d
2
D(). Consider ((x
0
, , d
2
), ). With
A performance comparison of robust adaptive controllers 381
this choice, and by continuity and causality, we have that
lim
tT
+
x(t) = x(T), lim
tT
+

(t) =

(T) (27)
where lim
tT
+ denote lim
tT,t>T
. It follows that
_
lim
tT
+
u(t)
_
u(T) = 2

(T)CB 2

b
n1
. (28)
By choosing a suitable

, it follows that

(T) can be made arbitrarily large
and hence the difference 28 is arbitrarily large. Then either u(T) is large or
lim
tT
+ u(t) is large, therefore u()
L
can be made arbitrarily large.
Therefore at least one component of Eq. 17 is unbounded, hence
P( (X
0
( ), , D()), ) = .

Proposition 5.4 Consider the closed loop system ((x

0
, , d()),
P
(
max
)) de-
ned by Eqs. 2 amd 16, where x
0
R
n
, S, d L

. Let , , > 0 and

suppose = () is compact. Consider the transient performance cost
functional 17. Then
P( (X
0
( ), , D()),
P
(
max
)) as
max
.
Proof Let M > 0. By Proposition 5.3 there exists x
0
X
0
( ), d() D(),
so that the closed loop ((x
0
, , d()), ) satises
x()
L

[0,)
+u()
L

[0,)
+ u()
L

[0,)
2M.
It follows that T > 0 s.t. x()
L

[0,T]
+ u()
L

[0,T]
+ u()
L

[0,T]
M.
By choosing
max
= 2

(T), we have that

max
>

(T), hence the solutions
((x
0
, , d()), ) and((x
0
, , d()),
P
(
max
)) are identical on[0, T], andthere-
fore
P( (X
0
( ), , D()),
P
(
max
)) x()
L

[0,T]
+u()
L

[0,T]
+ u()
L

[0,T]
M.
Since this holds for all M > 0, this completes the proof.
Proposition 5.5 Consider the closed loop ((x
0
, , d()),
D
(d
max
)) dened by
Eqs. 2 and 14, where x
0
R
n
, S, d L

. Let , , > 0 and suppose

= () is compact. Consider the transient performance cost functional
17. Then
P( (X
0
( ), , D()),
D
(d
max
)) < , d
max
> 0. (29)
382 A. Sanei, M. French
Proof Let x
0
X
0
( ), and d D(). A direct application of Property
D2 of Theorem3.1 guarantees the boundedness of x(),

(), u() as a continuous
function of x
0
, , d
max
. Since
u(t) = y(t)
2
D
d
max
(y(t))

(t)C
__
A

(t)BC
_
x(t) +Bd(t)
_
, t 0
it follows that u()
L
is bounded in terms of a continuous function of
x
0
, , d
max
. Therefore
P((x
0
, , d()),
D
(d
max
)) M(x
0
, , d
max
),
for some continuous function M : R
n
S R
0
R
0
. Taking the supremum
over system parameters x
0
, , d implies that for all d
max
,
P( (X
0
( ), , D()),
D
(d
max
)) sup
x
0
X
0
( )
sup

sup
dD()
M(x
0
, , d
max
) < .

Proof of Theorem I This is a simple consequence of Proposition 5.4 and Prop-

osition 5.5.
6 Proof of Theorem II
In order to prove Theorem II, we rst give the following Propositions:
Proposition 6.1 Consider the closed loop system ((x
0
, , d()),
D
(d
max
)) de-
ned by Eqs. 2 and 14, where x
0
R
n
, S, d L

. Let , , > 0 and suppose

= () is compact, and (0). Consider the transient performance
cost functional 17. Then > 0 such that
P( (X
0
( ), , D()),
D
(d
max
)) , as d
max
.
Proof Since (0), there exists such that A has an eigenvalue
with strictly positive real part or a purely imaginary eigenvalue. In either
case it follows that there exists x
0
X
0
( ) [ , ] and d D() such that
lim
t
sup| x(t)| = where x() is the solution to:

x(t) = A x(t) +Bd(t) x(0) = x

0
t [0, ).
Consider the closed loop system ((x
0
, , d()),
D
(d
max
)) and dene as
follows:
=
_
if y(t) [ , ] t 0
inf{t 0 | |y(t)| }, otherwise.
A performance comparison of robust adaptive controllers 383
Note that by dead-zone denition 14,

(t) = 0 for all t [0, ), hence

(t) = 0
for all t [0, ) since

(0) = 0. Therefore
x(t) = Ax(t) +Bd(t) x(0) = x
0
t [0, ).
Since x(t) = x(t) for t [0, ) it follows by the observability of (A, C) that
y(t) = Cx(t) hits the boundary of [ , ] in nite time, i.e. < . It follows
that
x()
L
|y()| = = d
max
.
Hence P((x
0
, , d()),
D
(d
max
)) as d
max
, and therefore
P( (X
0
( ), , D()),
D
(d
max
)) as d
max
.

Proposition 6.2 Consider the closed loop ((x

0
, , d()),
P
(
max
)) dened by
Eqs. 2 and 16, where x
0
R
n
, S, d L

. Let , , > 0 and suppose

= () is compact. Then for all
max
,
P( (X
0
( ), , D()),
P
(
max
)) < . (30)
Proof Let x
0
X
0
( ), , d D(). A direct application P2 of Theo-
rem 3.2 guarantees the boundedness of signals x(),

(), u() of the closed loop
((x
0
, , d()),
P
(
max
)) as a continuous function of x
0
, , d,
max
. It follows
that
u(t) = y(t)
3

(t)C(Ax(t) +B(u(t) +d(t)))
is bounded in terms of a continuous function of x
0
, , d,
max
, hence
P((x
0
, , d()),
P
(
max
)) M(x
0
, ,
max
, d) < , (31)
where M : R
n
V R
0
R
0
is continuous. Taking the supremum over
system parameters x
0
, , d implies that for all
max

P( (X
0
( ), , D()),
P
(
max
)) sup
x
0
X
0
( )
sup

sup
dD()
M(x
0
, ,
max
, d) < .

Proof of Theorem II This is a simple consequence of Proposition 6.2 and Prop-

osition 6.1.
384 A. Sanei, M. French
The proof of Theorem II is heavily based on the natural assumption that the
size of dead-zone is chosen to be equal to the a priori bound on the distur-
bance level d
max
. In particular, := d
max
implies that P( (X
0
( ), (), D()),

D
(d
max
)) as d
max
. In the following section we show that the other
choices of also yield the similar results.
7 Alternative choices of dead-zones
In this section we consider alternative choices for the dead-zone, and variations
on the denition of the controller. We consider the following smooth dead-zone
controller (compare to Eq. 14):

D,
(d
max
) : u(t) =

(t)y(t))

(t) = |y(t)| D

(y(t)),

(0) = 0, := (d
max
)
(32)
where : [0, ) [0, ), noting that we have previously considered the case
where (r) = r. We consider two possibilities for the growth of the function
1
:
1. lim
r
(r) = .
2. lim
r
(r) c for some constant c 0.
In case (1) it is straightforward to observe that
P := P
_
(X
0
( ), , D()),
D,
(d
max
)
_
as d
max
, (33)
by the same argument as in Proposition 6.1.
Choice (2) is also a viable option, since
D,
(d
max
) =
D
((d
max
)), and
by Theorem 3.1, P
_
(X
0
( ), , D()),
D,
(d
max
)
_
< and lim
t
inf
x[(d
max
),(d
max
)]
|y(t) x| = 0. This choice has been extensively studied
in the context of tracking, where the case (r) = is known as -tracking
[6]. The next result shows that if c is too small, then the projection controller
outperforms the dead-zone design. For simplicity we consider a scalar system:
Proposition 7.1 Consider the system(x
0
, (a, 1), d()) andthe controllers
D,
()
and
P
() dened by Eqs. 2,32 and 16, respectively, where x
0
R, a R, d L

and where lim

r
(r) c for c > 0. Let , > 0 and let = {(a, 1) R
2
:
|a| }. Consider the transient performance cost functional 17. Then for all
d
max
, and for all
max
, there exists c

> 0 such that for all 0 < c < c

,
P
_
(X
0
( ), , D()),
D,
(d
max
)
_
> P
_
(X
0
( ), , D()),
P
(
max
)
_
.
1
Other cases such as oscillatory but unbounded () can be handled suitably by considering
monotonic subsequences.
A performance comparison of robust adaptive controllers 385
Proof Let x
0
X
0
( ) and suppose (a, 1) is such that a > 0. Let d() = > 0
be constant. Consider ((x
0
, (a, 1), d),
D,
(d
max
)). Since y(t) [(d
max
),
(d
max
)] as t , there exists T > 0 such that y(T) 2(d
max
) and
(1/2)(d/dt)y(T)
2
= y(T)y(T) 0. It follows that,
(

(T) a)

2(d
max
)
.
Dene d
2
() as in Proposition 5.3 and consider ((x
0
, , d
2
),
D,
(d
max
)). It
follows that
_
lim
tT
+
u(t)
_
u(T) = 2

(T) =

2
(d
max
)

2
c
.
Hence for all P

> 0, there exists c

> 0 such that for all 0 < c < c

,
P
_
(X
0
( ), , D()),
D,
(d
max
)
_
= sup
x
0
X
0
( )
sup

sup
dD()
u()
L
,
> P

. (34)
The proof is completed by choosing P

= P( (X
0
( ), , D()),
P
(
max
)) <
.
Similar results can be obtained in the tracking setting, where the error signal
e() = y() y
ref
() replaces the state x() = y() and y
ref
() y
ref
() plays an
analogous role to the disturbance d(). See [16], [17] for further discussions.
8 Conclusion
In this paper we have established two results to demonstrate situations in which
we can compare the transient performance of projection and dead-zone based
controllers of non-identier-based adaptive designs. We have shown that
The dead-zone-based controller outperforms the projection-based control-
ler when the a priori information on the uncertainty level is sufciently
conservative.
The projection-based controller outperforms the dead-zone-based control-
ler when the a priori information on the disturbance level is sufciently
conservative.
Our results are based on the a priori information d
max
and
max
. We have also
shown that other choices of controllers with differing dependence on a priori
information such as -tracking controllers can be driven to similar conclusions.
There are a number of directions in which the result can be generalised, for
example: (1) generalisation of the result for adaptive controllers for higher rel-
ative degree plants, (2) generalisation of the results for nonlinear systems in the
386 A. Sanei, M. French
form of integrator chain [16], or in the strict feedback form, and (3) establishing
whether the same results can be given for the alternative costs, for example,
P = x()
L
+u()
L
.
It would also be of interest to determine explicit tight bounds on d

max
,

max
,
note however that this is challenging as it would require tight upper and lower
bounds on performance of both controllers.
Acknowledgements We would like to express our gratitude to the reviewers for their perceptive
comments and detailed reading of the draft manuscript. We would also like to thank Achim Il-
chmann for his extensive and detailed comments which led to a substantial improvement of the
manuscript.
Appendices
The proofs of Theorems 3.1 and 3.2 are given in this part.
Appendix 1: Proof of Theorem 3.1
Before we give the proof, we rst present a preliminary Lemma:
Lemma A-1 Suppose M R
nn
, n 1 is a stable matrix. Let the positive
denite matrix G R
nn
be the solution of the Lyapunov equation
GM+M
T
G = I.
Then for all t 0,
e
Mt
k
1
e
kt
, k := 1/(G), k
1
:=
_
(G)/(G),
where (G), (G) denote the minimum and maximum eigenvalue of G, respec-
tively.
Proof The proof is elementary and omitted for brevity.
The proof of Theorem 3.1 is derived from the technique used in the proof of
-tracking based on [5](pp 112117) but with a signicant extension to obtain
property D2.
Proof of Theorem 3.1 The continuity of D

implies by classical result of differ-

ential equations that there exists a unique solution (x(),

()) over a maximal
interval of existence [0, ) for some (0, ]. Using the transformation
dened by Eq. 5, we obtain
y(t) =
_
a
1
b
n1

(t)
_
y(t) +

A
2
z(t) +b
n1
d(t), y(0) = Cx
0
, (35)
z(t) =

A
3
y(t) +

A
4
z(t), z(0) = Nx
0
, (36)

(t) = |y(t)| D
d
max
(y(t)),

(0) = 0 (37)
A performance comparison of robust adaptive controllers 387
Dene the C
1
function
V(y) :=

0, y [ , ],
1
2
(|y| )
2
, y [ , ],
(38)
where = d
max
0. Observe that

V(y(t)) = (t) y(t) t [0, ), (39)

where the continuous function : [0, ) R is
(t) :=

0, y(t) [ , ].
(|y(t)| )
y(t)
|y(t)|
, y(t) [ , ].
(40)
Note that by Eq. 14

(t) = (t)y(t) = |(t)| |y(t)|, t [0, ) (41)

and by the continuity of Eq. 40, |y(t)| |(t)| |(t)|, or
|(t)|
1
|(t)| |y(t)| =
1

(t), t [0, ) (42)

Substituting Eq. 35 in Eq. 39, we have for all t [0, ),

V(y(t)) = ( a
1
b
n1

(t))y(t)(t) +

A
2
z(t)(t) +b
n1
d(t)(t)
( a
1
b
n1

(t))

(t) +b
n1
d()
L
|(t)| +

A
2
z(t) |(t)|
(M
1
b
n1

(t))

(t) +

A
2
z(t) |(t)|, (43)
where M
1
:= M
1
(, d
max
, d()
L
) := | a
1
| +
1
b
n1
d()
L
. Note that the
continuous dependency of M
1
on S follows from denition 4 and transfor-
mation 6 which also depends continuously on S. M
1
also depends continu-
ously on d
max
by denition 14.
Nowwe derive a relation between the second part of Eq. 43 and

(). Rewrite
Eq. 36 as
z(t) =

A
4
z(t) +

A
3
(t) +h(t), (44)
where
h(t) :=

A
3
(y(t) (t)).
388 A. Sanei, M. French
Note that since
y =

y, |y| <

y
|y|
, |y|
we have that |h(t)|

A
3
. Since

A
4
is exponentially stable, by Lemma A-1,
for all t [0, )
M
2
e
t
e

A
4
t
, M
2
:= M
2
() =
_
(R)
(R)
, := () = 1/(R), (45)
where R is dened in Eq. 8. M
2
and depend continuously on S since

A
4
depends continuously on S, hence R and its eigenvalues are continuously
dependent on S. Therefore by Eq. 44, we have for all t [0, )
z(t) M
2
e
t
z
0
+M
2
t
_
0
e
(ts)
(

A
3
|(s)| +|h(s)|) ds
M
2
e
t
z
0
+M
2

A
3

t
_
0
e
(ts)
|(s)|ds +M
2

A
3

1
(1 e
t
)
M
3

1 +
t
_
0
e
(ts)
|(s)|ds

, (46)
whereM
3
:= M
3
(x
0
, ) := M
2
_
z
0
+

A
3
(1 +
1
)
_
. Thecontinuous depen-
dency of M
3
on x
0
, S, d
max
follows from the denition of and the continu-
ous dependency of S
1
, and M
2
on S. Let t [0, ). From inequality (46)
it follows that
t
_
0
z(s) |(s)| ds M
3
t
_
0
|(s)| +M
3
t
_
0
|(s)|
s
_
0
e
(s)
|()| d ds. (47)
Using the Cauchy-Schwarz inequality, we have that
t
_
0
|(s)|
s
_
0
e
(s)
|()| d ds ()
L
2
(0, t)
_
_
_
_
_
_

_
0
e
()
|()|d
_
_
_
_
_
_
L
2
(0, t)
.
A performance comparison of robust adaptive controllers 389
An application of the following inequality (Theorem 6.5.54)[19]
_
_
_
_
_
_

_
0
e
()
|()|d
_
_
_
_
_
_
L
2
(0, t)

_
_
e

_
_
L
1
(0, t)
()
L
2
(0, t)

1
()
L
2
(0, t)
,
(48)
together with relations 42, 47, 48, implies
t
_
0
z(s) |(s)|ds M
3
(1 +
1
)
t
_
0
|(s)| +(s)
2
ds
M
3
(1 +
1
)
t
_
0
(1 +
1
)|(s)| |y(s)| ds
M
4

(t), (49)
where M
4
:= M
4
(x
0
, ) := M
3
(1 +
1
)(1 +
1
). Now, using Eq. 49, we can
calculate a bound on V(y(t)) in terms of

(t) by integrating Eq. 43 over [0, t]
V(y(t)) V(y(0)) +
t
_
0
_
M
1
b
n1

(s)
_

(s)ds +

A
2
M
4

(t)
V
0
+M
5

(t)
b
n1
2

(t)
2
,
where V
0
:= V(y(0)), and M
5
:= M
5
(x
0
, , d
max
, d()
L
) := M
1
+

A
2
M
4
.
The positive deniteness of V implies that

b
n1
2

(t)
2
+M
5

(t) +V
0
0. (50)
Solving the quadratic inequality 50 for

(t), we have that
M
5
M
6
b
n1

(t)
M
5
+M
6
b
n1
, (51)
where M
6
:= M
6
(x
0
, , d
max
, d()
L
) :=
_
M
2
5
+2b
n1
V
0
> M
5
. We discard
the non-positive lower bound of Eq. 51 due to the fact that by Eq. 37,

(0) = 0
and

() is non-decreasing. Therefore
0

(t)
M
5
+M
6
b
n1
. (52)
390 A. Sanei, M. French
The continuous dependency on x
0
, , d
max
of M
4
M
6
is a consequence of the
continuous dependency of M
3
, and the coordinate transformation 6. We
now dene a continuous function V

: R
n
S R
0
R
>0
R
0
by
V

(x
0
, , d()
L
, d
max
) = V
0
+M
5
(M
5
+M
6
)/b
n1
. Then
V(y(t)) V

(x
0
, , d()
L
, d
max
) t [0, ), (53)
The boundedness of y() in terms of V

(x
0
, , d()
L
, d
max
) on [0, ) fol-
lows from Eq. 38 and 53. Therefore by Eq. 40, () is bounded by a contin-
uous function of V

(x
0
, , d()
L
, d
max
) on [0, ). Hence by Eq. 46, z() is
bounded by a continuous function of V

(x
0
, , d()
L
, d
max
) on [0, ). It fol-
lows by Eq. 6 that x() L

(0, ) is bounded by a continuous function of

V

(x
0
, , d()
L
, d
max
). The continuity of the right-hand side of Eqs. 3537
and the boundedness of the solution (x(),

()) implies that = . Finally
the boundedness of u() as a continuous function of V

(x
0
, , d()
L
, d
max
)
follows from Eq. 14, thus establishing D1D2.
Since y() is bounded, it follows from the denition of

() that

() is uni-
formly continuous. Furthermore,

() L
1
, hence Barbalats lemma shows

(t) 0 as t . So lim
t
inf
x[,]
|y(t) x| = 0 hence proving D3, and
thus completing the proof.
Appendix 2: Proof of Theorem 3.2
Proof of Theorem 3.2 Since the (time-varying) right-hand side of the differ-
ential equations 2 and 16 is piecewise locally Lipschitz, a unique piecewise
absolutely continuous local solution exists. Let (x(),

()) denote a solution of
closed loop ((x
0
, , d()),
P
(
max
)) on a maximum interval of existence [0, )
for some [0, ). Using the transformation dened by Eq. 5, we obtain
y(t) =
_
a
1
b
n1

(t)
_
y(t) +

A
2
z(t) +b
n1
d(t), y(0) = Cx
0
, (54)
z(t) =

A
3
y(t) +

A
4
z(t), z(0) = Nx
0
, (55)

(t) =
_
y(t)
2

(t) <
max
0

(t)
max
,

(0) = 0 (56)
By denition of projection modication 16,

(t)
max
for all t 0. Let
k

<
max
be given by Eq. 10. The monotonicity of

() implies that either
(i)

(t) < k

for all t [0, ), or (ii) there exists t

[0, ) such that

(t) k

for all t [t

, ).
A performance comparison of robust adaptive controllers 391
(i) If

(t) < k

for all t [0, ) then by Eq. 16 we have

(t) = y(t)
2
for all
t [0, ). It follows that

_
0
y(s)
2
ds < k

,
and hence y() L
2
[0, ). Since

A
4
is stable, Eq. 55 can be considered as an
L
2
input y() to a stable system z(t) =

A
4
z(t)+

A
3
y(t) for all 0 t < , from
which follows that z(), z() L
2
[0, ). An explicit bound on z
L
2
[0,)
can
be obtained as follows. Note that since y() L
2
[0, ) and

A
4
is expo-
nentially stable, by Lemma A-1, there exists M
0
, 0 such that for all
t [0, )
M
0
e
t
e

A
4
t
M
0
:= M
0
() =
_
(R)/(R), := () = 1/(R),
(57)
where R is dened in Eq. 8. Therefore
2
z
L
2
[0,)
e

A
4

z
0

L
2
[0,)
+(sI

A
4
)
1

A
3

y
L
2
[0,)

M
0

z
0
+(sI

A
4
)
1

A
3

_
k

.
(58)
Abound on z() is also required to showthe boundedness of x(). Applying
Eq. 57 to the solution of Eq. 36, we have that for all t [0, ),
z(t)
M
0

z
0
+M
0
t
_
0
e
(ts)

A
3
|y(s)| ds
M
1

1 +
t
_
0
e
(ts)
|y(s)| ds

,
(59)
where M
1
:= M
1
(x
0
, ) := M
0
(1/z
0
+

A
3
). Since y() L
2
[0, ),
and y
L
2
[0,)
k

, it follows that z() is bounded as a continuous func-

tion of x
0
and . It remains to show y() L

[0, ), and to construct

an explicit uniform bound for y() which is continuously dependent on
x
0
, , d,
max
. To this end, consider Eq. 35 and dene:
V
1
(t) :=
1
2
y(t)
2
, t [0, ). (60)
2
Recall that G()
H
= ess sup
R
|G(jw)|.
392 A. Sanei, M. French
The time derivative of Eq. 60 along the solution of Eq. 5456 is bounded
by

V
1
(t) = ( a
1
b
n1

(t))y(t)
2
+

A
2
z(t)y(t) +b
n1
d(t)y(t)
=
1
2
y(t)
2
+( a
1
+1)y(t)
2
b
n1

(t)y(t)
2
+

A
2
z(t)y(t)
1
2
y(t)
2
+b
n1
d(t)y(t)
V
1
(t) +( a
1
+1 +

A
2

2
)y(t)
2
+|z(t)|
2
+
b
2
n1
2
d()
2
L

, t [0, ) (61)
where Eq. 61 follows from two applications of Youngs inequality and by
observing that b
n1

(t) > 0 for all t [0, ). Hence for all t [0, ),

1
2
y(t)
2
V
1
(t)
e
t
V
1
(0) +
t
_
0
e
(ts)
_
( a
1
+1 +

A
2

2
)y(s)
2
+|z(s)|
2
+
b
2
n1
2
d()
2
L

_
ds

y(0)
2
2
+( a
1
+1 +

A
2

2
)y()
2
L
2
[0,)
+z()
2
L
2
[0,)
+
b
2
n1
2
d()
2
L

. (62)
Since y()
2
L
2
[0,)
< k

which is continuously dependent on , and we have

shown that z()
2
L
2
[0,)
is bounded as a continuous function of |x
0
| and ,
it follows that y() is bounded on [0, ) in terms of a continuous function
of x
0
, , d and
max
. It follows by Eqs. 59 and 6 that x() is bounded on
[0, ) as a continuous function of x
0
, , d,
max
.
(ii) Suppose there exists t

[0, ) such that

(t) k

for all t [t

, ). Dene
the Lyapunov function
V
2
(t) = x(t)
T
P x(t), t [0, ) (63)
where x(t), P are dened by Eqs. 6 and 9, respectively.
Then for all t [0, ), the time derivative of V
2
(t) w.r.t. Eqs. 5456 is given
A performance comparison of robust adaptive controllers 393
by

V
2
(t) = x(t)
T
P

x(t) +

x(t)
T
P x(t),
= x(t)
T
P
_

D(

(t)) x(t) +

Bd(t)
_
+
_

D(

(t)) x +

Bd(t)
_
T
P x(t),
= x(t)
T
_
P

D(

(t)) +

D
T
(

(t))P
_
x(t) +2 x(t)
T
P

Bd(t),
(64)
where S is given by Eq. 5 and

B := S
1
B,

D() := S
1
(ABC) S. Let
(, d) :=
2
_
(P)
(Q)
|P

B| d()
L
,
where Q is dened in Eq. 9. Then is continuously dependent on S
by the continuity of the transformation (6). By Eq. 9, we observe that

V
2
(t) x(t)
T
Q x(t) +2 x(t)
T
P

Bd(t), (65)

_
(Q)
2
_
(P)

V
2
(t)
|P

B| d()
L

_
x(t)
2
, (66)
(Q)
_
1
(, d)

V
2
(t)
_
x(t)
2
, t [t

, ). (67)
Inequality 67 implies that

V
2
(t) < 0 for all V
2
(t) > (, d)
2
, i.e.
V(t) V

( x

, , d) := max
_
V
2
( x

) , (, d)
2
_
t [t

, ), (68)
where x

:= x(t

). Therefore, by Eq. 63, x() is bounded on [t

, ) in terms
of V

( x

, , d). A bound on x() on [0, t

) and a bound on the end point

x(t

) as a continuous function of x
0
, , d,
max
follows directly by the
argument of part (i) where

(t) < k

<
max
for all t [0, t

). Hence by the
continuity of the transformation 6, x() is bounded on [0, ) as a continuous
function of x
0
, , d,
max
.
The boundedness of x() over [0, ) implies that x() cannot have a nite
escape time. By denition 16,

() is also known never to leave the set [0,
max
].
Hence = , i.e. the solution ( x(),

() ) exists for all t [0, ). Furthermore,
the solution is bounded as a continuous function of x
0
, , d,
max
. Finally, the
boundedness of u() as a continuous functionof x
0
, , d,
max
follows by Eq. 16.

References
1. Egardt B (1979) Stability of adaptive controllers. Lecture notes in control and information
sciences. Springer, Berlin Heidelbergh New York
394 A. Sanei, M. French
2. French M (2002) An analytical comparison between the nonsingular quadratic performance of
robust and adaptive backstepping designs. IEEE Trans on Automat Control 47(4):670675
3. Georgieva P, Ilchmann A (2001) Adaptive -tracking control of activated sludge processes. Int
J Control 74:12471259
4. Ilchmann A (1991) Non-identier-based adaptive control of dynamical systems: a survey. IMA
J Math Control Inform 8:321366
5. Ilchmann A (1993) Non-identier-based high-gain adaptive control. Lecture notes in control
and information sciences, Springer, Berlin Heidelberg New York
6. Ilchmann A, Ryan EP (1994) Universal -tracking for nonlinearly perturbed systems in the
presence of noise. Automatica 30:337346
7. Ilchmann A, Townley S (1999) Adaptive sampling control of high-gain stabilizable systems.
IEEE Trans Automat Control 44(10):19611966
8. Kreisselmeier G, Narendra KS (1982) Stable model reference adaptive control in the presence
of bounded disturbances. IEEE Trans Automat Control 27(6):11691175
9. Kristic M, Kanellakopoulos I, Kokotovic P(1995) nonlinear and adaptive control design. Wiley,
New York
10. Lee KW, Khalil HK (1997) Adaptive output feedback control of robot manipulators using
high-gain observer. Int J Control 67(6):869886
11. Logemann H, Townley S (1997) Adaptive stabilization without identication for distributed
parameter systems: An overview. IMA J Math Control Inform 20:175206
12. Mrtensson B (1985) The order of any stabilizing regulator is sufcient a priori information for
adaptive stabilization. Syst Control Lett 6:8791
13. Narendra KS, Annaswamy AM(1989) Stable adaptive systems. Prentice-Hall, EnglewoodCliffs
14. Peterson BB, Narendra KS (1982) Bounded error adaptive control. IEEE Trans Automat
Control 27(6):11611168
15. Ryan EP (1991) A universal adaptive stabilizer for class of nonlinear systems. Syst Control Lett
16:209218
16. Sanei A (2002) Towards a performance theory of robust adaptive control. PhD thesis, Depart-
ment of Electronics and Computer Science, University of Southampton, UK
17. Sanei A, French M (2004) Towards a performance theory of robust adaptive control. Int J
Adapt Control signal Process 18:403421
18. Townley S (1996) Topological aspects of universal adaptive stabilization. SIAM J Control
Optim, 34:10441070
19. Vidyasagar M (1978) Nonlinear systems analysis. Prentice-Hall, Englewood Cliffs
20. Willems JC, Byrnes CI (1984) Global adaptive stabilization in the absence of information on
the sign of the high-frequency gain. Lecture notes in control and information science. Springer,
Berlin Heidelberg New York, 62:4957
21. Xie C, French M (2004) A performance comparison between two design techniques for non-
linear output feedback control. Int J Control 77:264276