Communications in Applied Analysis xx (200x), no.

N, xxxxxx
GENERALIZED GAP METRICS AND ROBUST STABILITY OF
NONLINEAR SYSTEMS
WENMING BIAN
School of Electronics and Computer Science, University of Southampton
Southampton SO17 1BJ, United Kingdom
E-mail: wb@ecs.soton.ac.uk
To Professor J. R. L. Webb on the occasion of his retirement
ABSTRACT. A gap metric of Georgiou and Smith (IEEE Trans. Auto. Control, 42(9):1200
1229, 1997), which does not need causal and surjective mapping between graphs to dene, is studied
and generalized based on the notion of biased norm, the corresponding robust stability theorem
is presented in the notion of stability with bias terms. The obtained results are then applied to
studied the stability of linear system realizations, semilinear systems with bounded nonlinearities
and a nonlinear system with time delay in the inputs.
AMS (MOS) Subject Classication. 93D09, 93D25, 93C10.
1. Introduction
A basic property of stable feedback control systems is that they tolerate uncer-
tainties which are suciently small in an appropriate sense, and reduce the eects
of uncertainties. A signicant approach to deal with uncertainties is the gap metric
theory, it provides a measure of distance between dynamical systems which are not
required to be stable in themselves, small distance between open-loop systems would
correspond to small errors in norm in the closed loop.
In the context of nonlinear systems, a fundamental framework developed by Geor-
giou and Smith in [9] provides a generalization of linear gap metric([8, 16]) and as-
sociated robust stability results on the basis of a robust stability margin taken to be
the inverse of the induced norm of a closed loop operator. Two types of gap metrics
were introduced in [9], the

metric and the

0
metric as denoted in the paper. the
rst one is dened using causal and surjective mappings between graphs of open loop
operators, but for the

0
metric, which is further studied in [11], no such mapping is
involved in its denition. Under appropriate well posedness assumptions, the main
robust stability theorems state that, given two plants P, P
1
and a controller K, if the
closed loop [P, K] is gain stable, i.e. the corresponding closed loop operator has a
Submitted November 3, 2010 1083-2564 $15.00 c _Dynamic Publishers, Inc.
2 W. BIAN
nite induced norm, and if the gap (either

or

0
) between P and P
1
is smaller than
the robust stability margin, then the closed loop [P
1
, K] is also gain stable and its
gain can be estimated in terms of the gap and the gain of the closed loop operator of
[P, K]. This motivates many further studies on gap metric and its applications, see
[1, 2, 5, 6, 10, 11, 12] and references therein.
The framework requires that both the plant and controller map zero inputs to
zero outputs (i.e., P0 = 0, K0 = 0) and that the closed loop operator has an induced
norm. However, there are important instances in which these sucient conditions for
robust stability generically fail; and yet for which robustness results should apply and
for which, to date, either relatively ad-hoc methods have been utilized to establish
robust stability, or no such such robust stability certicates have been established.
Many such systems can be handled by developing a robust stability theory based
on an underlying notion of stability which includes bias terms; for such notions of
stability see [3, 15]. The second class of systems are those for which P(0) = 0,
K(0) = 0 but whose closed loop operator is discontinuous at 0, thus precluding the
existence of a (local) nite gain. Most adaptive controllers fall within this category
[4]. A third class of examples includes systems which include inherent osets, arising
e.g. from quantization errors, sensors biases etc. Another such class of feedback
systems include nonlinear high gain controller designs which attenuate the eects of
unknown nonlinearities by high gain feedback, and which do not cancel the eect of
the nonlinearities.
The notion of gain function stability given in [9] might be an alternative approach
to those problems, but it is too general and fails to produce a clear description to the
stability. So several generalized gap metrics and theorems have been introduced. In
[10], based on shift operation, a theorem is given dealing with systems whose response
depends on a non-zero initial condition, and which do not start at an equilibrium. In
[5], notions of stability and gap matric with bias terms are presented under uniqueness
assumption instead of the stronger well-posedness assumption. In [12], stability is
dened via a biased norm but the gap metric remains the same as in [9]. All of
those generalizations are based on the

gap metric which needs causal and surjective
mappings between graphs. We note that, comparing with the

gap metric, the

0
gap metric has two advantages: rstly, it is smaller and therefore theoretically allows
a wider range of perturbations; secondly, its denition does not needs causal and
surjective mappings between graphs of which the causality is not easy to verify in
applications (so an alternative causality is used in [2, 5, 6, 7]).
So, in this paper, we will generalize Georgiou and Smiths

gap metric and the
corresponding robust stability theorem. In our setting, both the gap and the stability
will be based on norm with bias, the systems do not need to have zero outputs
for zero inputs, nor the system operator needs to be continuous at any point. This
GAP METRICS AND ROBUST STABILITY 3
generalization does not require causal mappings between graphs, but can be described
using surjective mappings. Using the obtained results, we study the stability of a
type of semilinear systems with bounded nonlinearities and linear system realizations.
The robust stability of integrator system with saturation is also given to show the
advantage of our results.
The paper is organized as follows. Section 2 introduces the basic signal spaces
setting and closed loop systems. Generalized gap metrics and robust stability results
are presented in Section 3. Stability of semilinear systems and linear system realiza-
tions are addressed in Sections 4 and 5, respectively, and in Section 6, we consider
the integrator system with saturation.
2. The closed loop system
Let T denote either the discrete half-axis time set N or the continuous time
counterpart, R
+
. In both cases T is totally ordered in the natural manner.
For T , let o

denote the set of all locally integrable maps [0, ) A where

A is a nonempty set. For ease of notation dene o := o

. For T , T ,
0 < < dene a truncation operator T

and a restriction operator R

as follows:
T

: o

o , v T

v : (T

v)(t) =
_
v(t), t [0, )
0, otherwise
R

: o

, v R

v := (t v(t), t [0, )) .
Both operators are for considering signals over nite time intervals. The results of
this paper will be based on the use of truncation, but remain true if it is replaced by
restriction.
We dene 1 o to be a signal space if, and only if, it is a vector space. Suppose
additionally that 1 is a normed vector space and that the norm | | = | |
V
is (also)
dened for signals of the form T

v, v 1

, > 0. We can dene a norm | |

on o

by |v|

= |T

v|, for v o

. We associate spaces as follows:

1
e
= v o [ > 0 : T

v 1, the extended space;

1

= v o

[ (0, ) : T

v 1, for 0 < ; and

1
a
=

(0,]
1

, the ambient space.

For example, in the case when 1 = L
p
(R
+
, R
n
) with 1 p , we have 1
e
=
L
p
loc
(R
+
, R
n
) and 1
a
=

(0,]
L
p
loc
((0, ), R
n
). So the ambient space 1
a
includes
signals with nite blow up.
For signal spaces A, 1, dene the following:
(i) An operator Q: A
a
1
a
is called causal if, and only if,
x, y |
a
, dom(x) dom(y) : [T

x = T

y T

(Qx) = T

(Qy)] . (2.1)
4 W. BIAN
(ii) An operator Q: A
a
1
a
is called stabilizable if for all x A
a
, v 1
a
satisfying
Qx = v over dom(x) dom(v) and for all dom(x) dom(v), there exists
x A, v 1 such that Q x = v and T

(x, v)

= T

( x, v)

.
(iii) A causal operator Q: A
a
1
a
is called gain stable if Q(A) 1, Q(0) = 0 and
|Q| := sup
_
|T

Qx|

|T

x|

: x A , > 0 , T

x ,= 0
_
< .
(iv) A causal operator Q: A
a
1
a
is called (, )-gain stable, with , 0, if
Q(A) 1 and
|T

Qx|

|T

x|

+ , x A , > 0.
(v) A causal operator Q: A
a
1
a
is called -gain stable with bias if, and only if,
there exists 0 such that Q is (, )-gain stable with bias.
We now consider the closed loop system
[P, K] : y
1
= Pu
1
, u
2
= Ky
2
, u
0
= u
1
+ u
2
, y
0
= y
1
+ y
2
, (2.2)
as depicted in Figure 1, where P : |
a
}
a
and K : }
a
|
a
are causal mappings
u
0
u
1
y
1
P
K
y
0
u
2
y
2

+
+

Figure 1. The closed-loop [P, K].

representing the plant and the controller, respectively, and |, } are given two normed
signal spaces. We write J := | } and dene the norm in the product space J as
|(u, y)
T
| = max|u|, |y|, (u, y)
T
J.
For w
0
= (u
0
, y
0
)
T
J, a pair (w
1
, w
2
) = ((u
1
, y
1
)
T
, (u
2
, y
2
)
T
) J
a
J
a
, J
a
:=
|
a
}
a
, is a solution to (2.2) if, and only if, (2.2) holds on dom(w
1
, w
2
) := dom(w
1
)
dom(w
2
). Let
A
w
0
:= (w
1
, w
2
) J
a
J
a
[ (w
1
, w
2
)is a solution to (2.2)
be the set of all solutions, which may be empty. The closed loop system [P, K] is
said to have the existence property, if A
w
0
,= for all w
0
J, and the uniqueness
property, if
w
0
J : ( w
1
, w
2
), ( w
1
, w
2
) A
w
0
= ( w
1
, w
2
) = ( w
1
, w
2
) on dom( w
1
, w
2
) dom( w
1
, w
2
) .
GAP METRICS AND ROBUST STABILITY 5
For each w
0
J, dene
w
0
T by the property
[0,
w
0
) :=
_
( w
1
, w
2
)Xw
0
dom( w
1
, w
2
)
and dene (w
1
, w
2
) J
a
J
a
, with dom(w
1
, w
2
) = [0,
w
0
), by the property
R
t
(w
1
, w
2
) A
w
0
for all t [0,
w
0
). This induces the operator
H
P,K
: J J
a
J
a
, w
0
(w
1
, w
2
)
and the projection operators

P//K
: J J
a
, w
0
w
1
, and
K//P
: J J
a
, w
0
w
2
.
Clearly,
H
P,K
=
_

P//K
,
K//P
_
and
P//K
+
K//P
= I. (2.3)
The graphs of the plant P and the controller K, denoted by (
P
and (
K
(or
graph(P) and graph(K)) respectively, are dened as follows:
(
P
=
__
u
Pu
_

u |, Pu }
_
, (
K
=
__
Ky
y
_

Ky |, y }
_
.
The closed loop system [P, K] given by (2.2), is said to be:
(i) locally well posed if, and only if, it has the existence and uniqueness properties
and the operator H
P,K
: J J
a
J
a
, w
0
(w
1
, w
2
), is causal;
(ii) globally well posed if, and only if, it is locally well posed and H
P,K
(J) J
e

J
e
;
(iii) regularly well posed if it is locally well posed and for all w
0
J with
w
0
< ,
we have
|H
P,K
w
0
|

as
w
0
.
(iv) BIBO stable if, and only if, it is globally well posed and H
P,K
(J) J J.
(v) gain stable, (, )-gain stable or -gain stable with bias if, and only if,
P//K
is
gain stable, (, )-gain stable or -gain stable with bias respectively.
We remark that in the above denitions, the operator H
P//K
can be replaced by

P//K
or
K//P
due to the relations between the three operators given in (2.3).
3. Robustness
In this section, we deal with robustness of globally gain stability with bias of
feedback systems in the sense that if, for given nominal plant P and controller K,

P//K
is (, )-gain or -gain stable with bias, then
P
1
//K
is also (, )-gain or -
gain stable with bias for a suitable perturbation P
1
. The allowed perturbations are
measured by distances which are generalizations of a gap metric given in [9]. It is
6 W. BIAN
proved that if the distance is smaller than the inverse of the gain , then the feedback
stability is preserved.
Let |, } be two signal spaces and let P, P
1
: |
a
}
a
be the input-to-output
operators of two control systems. Given a number 0, we dene the following four
gap metrics from P to P
1
:

(P, P
1
) = limsup

inf
_
> 0 :
for any y (
P
1
, there exists x (
P
such that |y x|

|x|

+
_
;

(P, P
1
) = limsup

inf
_

_
> 0 :
there exists 0 such that, for any
y (
P
1
, there exists x (
P
satisfying
|y x|

|x|

+
_

_
;

(P, P
1
) = sup
>0
inf
_
> 0 :
for any y (
P
1
, there exists x (
P
such that |y x|

|x|

+
_
;

d(P, P
1
) = sup
>0
inf
_

_
> 0 :
there exists 0 such that, for any
y (
P
1
, there exists x (
P
satisfying
|y x|

|x|

+
_

_
.
Clearly,

(P, P
1
)

(P, P
1
),

d(P, P
1
)

d

(P, P
1
),

(P, P
1
)

d(P, P
1
) and

(P, P
1
)

(P, P
1
) for any > 0. In the case when (
P
= (
P
1
, we have

(P, P
1
) =

(P, P
1
) =

d(P, P
1
) =

d

(P, P
1
) = 0.
It is noted that in the case when = 0, we have

0
(P, P
1
) = limsup

sup
yG
P
1
y =0
inf
xG
P
x =0
|y x|

|x|

,
which is exactly the alternative gap metric studied in [9] (where the setting is slightly
dierent but wont aect the results) and its generalizations in [11]. So the two -
gap metrics above and the corresponding robust stability theorems below are direct
generalizations of those in [9, 11].
To describe the gap metrics alternatively, we denote by
O(P, P
1
; T) =
_
:
: T (
P
(
P
1
is a surjective
set-valued mapping with bounded values
_
,
and for any mapping Q : A J
a
J
a
and number 0, we let:
|Q|
()
= limsup

inf > 0 : |T

Qx|

|T

x|

+ , x A,
|Q|
()
= sup
>0
inf > 0 : |T

Qx|

|T

x|

+ , x A,
where, in the case when Q is set-valued, |T

Qx|

= sup|T

y|

: y Qx. Note,
both |Q|
()
and |Q|
()
are generalizations of the induced operator norm |Q| and
|Q|
()
is called the -gain in [2] or biased gain in [12].
GAP METRICS AND ROBUST STABILITY 7
Proposition 3.1.

(P, P
1
) = inf
_
|( I)[
D
|
()
: O(P, P
1
; T), T (
P
_
, (3.1)

(P, P
1
) = inf
_
|( I)[
D
|
()
: O(P, P
1
; T), T (
P
_
. (3.2)
Proof. We only prove (3.1) since the proof for (3.2) is similar.
For convenience, we denote the right hand sides of (3.1) and (3.2) by

(P, P
1
)
and

d

(P, P
1
), respectively.
Let : T (
P
(
P
1
is a surjective mapping. Then for any y (
P
1
, there
exists x
y
(
P
such that y x
y
and, therefore

(P, P
1
) = limsup

inf
_
> 0 : |y x
y
|

|x
y
|

+ for all y (
P
1
_
limsup

inf
_
> 0 : |( I)x
y
|

|x
y
|

+ for all y (
P
1
_
limsup

|( I)[
D
|
()
,
which

(P, P
1
)

(P, P
1
).
To show the reverse inequality, we let > 0. Then for any y (
P
1
, there exists
at least one x (
P
such that
|y x|

(P, P
1
) + )|x|

+ , for large > 0. (3.3)

Let T = x (
P
: there exists y (
P
1
such that x, y satisfy (3.3) and dene a
mapping : T (
P
(
P
1
as:
(x) = y : x, y satisfy (3.3) for x T.
Clearly, is surjective mapping from (
P
to (
P
1
with bounded values and
|( I)x|

(P, P
1
) + )|x|

+
for all x T and large > 0. So

(P, P
1
) |(I)[
D
|
()

(P, P
1
) +. Since
is arbitrary, we see

(P, P
1
)

(P, P
1
). This completes the proof.
We remark that

(P, P
1
) is a generalization of the main gap metric of [9] where,
and for all of other generalizations, the mappings in O(P, P
1
; T) are required to
be causal. In the case when the mappings are required causal, we can only have
the inequality

(P, P
1
)

(P, P
1
). So, the

-gap metric dened in this paper is the
smaller than those dened using causal and surjective mappings between graphs and,
therefore, better in theory and applications.
The following theorems generalize the standard results from both linear and non-
linear robust control, see [9, 11] and the references therein.
8 W. BIAN
Theorem 3.2. Consider the feedback system described in Figure 1. Let P, P
1
: |
a

}
a
, K: }
a
|
a
be stabilizable. Let 0, > 0 and [P, K] be globally well-posed
and [P
1
, K] is either globally or regularly well-posed. If [P, K] is (, r)-gain stable and

(P, P
1
) <
1
,
then [P
1
, K] is globally well-posed and (
1
, r
1
)-gain stable with

1
=
(1 +

(P, P
1
))
1

(P, P
1
)
, r
1
=
(1 +

(P, P
1
))(r + )
1 (

(P, P
1
))
+ .
The same conclusion holds if the gap metric

(P, P
1
) is replaced by

(P, P
1
),

(P, P
1
) or

d(P, P
1
).
Proof. Let =

(P, P
1
). By assumption, there exists a
0
> 0 such that (+
0
)r < 1.
Let w J, (0,
0
) and 0 < <
w
. By the well-posedness assumption, we
may suppose w = w
1
+w
2
with unique w
1
= (u
1
, y
1
)

, w
2
= (u
2
, y
2
)

J
a
such that
y
1
= P
1
u
1
, u
2
= Ky
2
. This tells
P
1
//K
w = w
1
. By the stabilization assumptions,
there exist w

1
(
P
1
, w

2
(
K
such that T

w
1
= T

1
, T

w
2
= T

2
. By the denition
of

(P, P
1
), there exists x (
P
such that
|w

1
x|

( + )|x|

+ . (3.4)
Write w = x + w

2
. Since [P, K] is well-posed, we see

P//K
w = x.
Therefore, by (3.4) and the (, r)-gain stability of
P//K
, we see
|
P
1
//K
w|

= |w
1
|

= |w

1
|

(1 + + )|x|

+
= (1 + + )|
P//K
w|

+
(1 + + )| w|

+ (1 + + )r + . (3.5)
Again by (3.4), there exists z
w
J such that |z
w
|

1 and
T

w = T

1
+ T

2
= T

x + T

2
+ (( + )|x|

+ )T

z
w
= T

w + (( + )|x|

+ )T

z
w
,
from which it follows:
| w|

|w|

+ ( + )|x|

+ = |w|

+ ( + )|
P//K
w|

+
|w|

+ ( + )| w|

+ ( + )r +
and therefore
| w|

|w|

1 ( + )
+
( + )r +
1 ( + )
.
GAP METRICS AND ROBUST STABILITY 9
Substituting into (3.5) to obtain
|
P
1
//K
w|

(1 + + )
1 ( + )
|w|

+
(1 + + )(r + )
1 ( + )
+ .
Since w J, the right hand side is uniformly bound for all 0 < <
w
, it follows
that if [P
1
, K] is regularly well-posed, it is also globally well-posed. By letting 0,
we obtain |
P
1
//K
w|

1
|w|

+ r
1
, i.e., [P
1
, K] is (
1
, r
1
)-gain stable.
If the gap metric

(P, P
1
) is replaced by

(P, P
1
), then the assumption

(P, P
1
) <

1
implies that there exists > 0 such that

(P, P
1
) <
1
. Therefore, the same
conclusion holds.
The proofs for the cases when

(P, P
1
) is replaced by the

d-gap metrics are
similar.
To close this section, we next present an estimate to gap metric for operators with
certain graph representations. This estimate will be useful when the above theorem is
applied to study the robust stability of some special systems in the rest of this paper.
Proposition 3.3. Let P, P
1
: |
a
}
a
be given. Suppose that there exist operators
M, M : T | } and N, N : T | | such that
(
P
=
__
M
N
_
v : v T
_
, (
P
1
=
__
M +M
N +N
_
v : v T
_
. (3.6)
If there exist k 0, 0 such that
_
_
_
_
_
_
M
N
_
u
_
_
_
_
_

k
_
_
_
_
_
_
M
N
_
u
_
_
_
_
_

+ for all u T and large > 0, (3.7)

then

(P, P
1
) k. If, in addition, k < 1, then

(P, P
1
)
k
1k
.
If (3.7) are satised for all > 0, then the same conclusions hold for the gap
metric

d

(P
1
, P) as well.
Proof. For any y =
_
M +M
N +N
_
u (
P
with u T, let x =
_
M
N
_
u (
P
. By
the assumptions, we see that
|y x|

=
_
_
_
_
_
_
M
N
_
u
_
_
_
_
_

k
_
_
_
_
_
_
M
N
_
u
_
_
_
_
_

+
for large > 0, which implies that

(P, P
1
) k.
If k < 1, then for any u T, (3.7) implies
_
_
_
_
_
_
M
N
_
u
_
_
_
_
_

k
_
_
_
_
_
_
M +M
N +N
_
u
_
_
_
_
_

+
_
_
_
_
_
_
M
N
_
u
_
_
_
_
_

+
10 W. BIAN
and therefore
|x y|

k
1 k
|y|

+
which implies

(P
1
, P)
k
1k
.
If (3.7) is satised for all > 0, using the same method, we can show the same
conclusions hold for the gap metric

d

(P
1
, P).
Note that the operators M, N, M and N are not required to be stable nor
invertible. Hence, (M, N) (resp. (M +M, N +N)) is not necessarily the coprime
factorisation for P (resp. P
1
) (see [1, 13, 14]). However, if the operators admit
(right) coprime factorisations, then the graph representations (3.6) can be achieved
with (M, N), (M + M, N + N) the coprime factorisations. Interestingly, when
(M, N) is uniformly bounded, k may be taken as 0 and we will have

(P, P
1
) = 0.
4. Semilinear systems with bounded nonlinearities
In the rest of this paper, as applications of results in the last section, we will
address some stability problems in nonlinear feedback control regarding. We will
estimate the gap metrics between certain specic nominal plants and their perturba-
tions, and study the robust stability of the underlying control systems.
First of all, in this section, we consider the eect of initial state value to the sta-
bility of a control system. There has always been a creative tension in control theory
between state space and input-output methods, but the original formulation of gap
metric approach precludes the case of non-zero initial conditions as the assumptions
require both plant and controller map zero to zero. Although this condition has been
relaxed later and, particularly, recently in the biased notions [2, 5, 10], a problem
that is not addressed within the gap metric framework is: if a controller stabilizes
a system with a given initial state condition, does it stabilize the system when the
initial state condition changes? We will prove that under our framework, for semiliner
systems with bounded nonlinearities, changing initial state condition does not change
the robust stability and the stability margin.
The systems concerned are of the following form, denoted by (f, x
0
)
(f, x
0
) : u y :
x

(t) = Ax(t) + f(t, x(t)) + Bu(t), x(0) = x

0
Y
y(t) = Cx(t) + Du(t)
(4.1)
where U, Y are normed vector spaces, A : dom(A) Y Y is a linear operator,
C : Y Y , B, D : U Y are bounded linear operators and f(t, x) : R Y Y
is measurable in t and Lipschitz in x. Let | = L

(R
+
, U), } = L

(R
+
, Y ) be the
signal spaces. We assume the following two assumptions.
Assumption 1. T | and for each u T, (4.1) has a solution in }.
GAP METRICS AND ROBUST STABILITY 11
Assumption 2. There exists a linear operator F : Y U such that A + BF
generates a c
0
-semigroup of bounded linear operators S(t) and, for each v |, the
equation
x

(t) = (A+ BF)x(t) + f(t, x(t)) + Bv(t), x(0) = x

0
. (4.2)
has unique solution x }.
We rst give a graph representation for (f, x
0
), which is a generalization of the
corresponding results in [13, 14].
Lemma 4.1.
graph((f, x
0
)) =
__
M
N
_
v : v |
_
with
Mv(t) = FS(t)x
0
+ F
_
t
0
S(t s)[Bv(s) + f(s, x(s))]ds + v(t),
Nv(t) = (C + DF)S(t)x
0
+ (C + DF)
_
t
0
S(t s)[Bv(s) + f(s, x(s))]ds + Dv(t),
where x is the solution to (4.2) corresponding to input v.
Proof. It is known that M : v u and N : v y are the input-output mappings to
the closed loops
x

(t) = (A + BF)x(t) + f(t, x(t)) + Bv(t), x(0) = x

0
(4.3)
u(t) = Fx(t) + v(t) (4.4)
y(t) = (C + DF)x(t) + Dv(t). (4.5)
By our assumptions, both M and N are well-dened operators. Furthermore, M is
invertible with the inverse M
1
: u v given by:
x

(t) = Ax(t) + f(t, x(t)) + Bu(t), x(0) = x

0
(4.6)
v(t) = u(t) Fx(t). (4.7)
By Assumption 1, we see that dom(M
1
) T. If u dom(M
1
), then v = M
1
u
} and the x satisfying (4.6)-(4.7) is the solution to (4.2). From Assumption 2, it
follows x } and therefore dom(M
1
) = T. Hence, the composition NM
1
is the
input-out mapping associated with
x

(t) = Ax(t) + f(t, x(t)) + Bu(t), x(0) = x

0
(4.8)
v(t) = u(t) Fx(t), (4.9)
z

(t) = (A+ BF)z(t) + f(t, z(t)) + Bv, z(0) = x

0
(4.10)
y(t) = (C + DF)z(t) + Dv(t). (4.11)
12 W. BIAN
Substituting (4.9) into (4.10) and subtracted by (4.8), we have
z

= (A+ BF)(z x) + f(t, z(t)) f(t, x(t)).

Since S(t) is a c
0
-semigroup, there exist , r 0 such that |S(t)| re
t
for all t 0.
Since f is Lipschitz and x, z have the same initial value, we see
[z(t) x(t)[ =

_
t
0
S(t s)[f(s, z(s)) f(s, x(s))]ds

cr
_
t
0
e
(ts)
[z(s) x(s)[ds with some c > 0
and
e
t
[z(t) x(t)[ cr
_
t
0
e
s
[z(s) x(s)[ds, for all t 0.
By Gronwalls Inequality, x z. Therefore, NM
1
= (f, x
0
) and
graph((f, x
0
)) =
__
u
(f, x
0
)u
_
: u T
_
=
__
Mv
Nv
_
: v |
_
.
Remark: since no more information is given about the left invertibility of (M, N)

,
(M, N) is not necessarily the coprime factorisation of (f, x
0
).
We now in the position to estimate the gap between (f, x
0
) and (f, x
0
) with
x
0
, x
0
Y and x
0
,= x
0
.
Corollary 4.2. Consider the system given by equation (4.1). Let x
0
, x
0
Y and
x
0
,= x
0
. Let P = (f, x
0
) be the nominal plant and P
1
= (f, x
0
) the perturbation.
Suppose F is a bounded operator and the semigroup of linear operators S(t) generated
by A+ BF is such that
|S(t)| re
t
with r 1, > 0 for all t > 0. (4.12)
If either
|f(t, x)| c with c 0 for all t 0, x Y, (4.13)
or there exists d [0, /r) such that
|f(t, x
1
) f(t, x
2
)| d|x
1
x
2
| for all t 0, x
1
, x
2
Y, (4.14)
then

(P, P
1
) = 0.
Proof. As shown in Lemma 4.1, there exist operators N, N
1
: } } and M, M
1
:
| } such that
graph(P) =
__
M
N
_
u : u |
_
, graph(P
1
) =
__
M
1
N
1
_
u : u |
_
GAP METRICS AND ROBUST STABILITY 13
and
Mu(t) =: (M
1
M)u(t) = FS(t)( x
0
x
0
)
+ F
_
t
0
S(t s)
_
f(s, x(s)) f(s, x(s))

ds, (4.15)
Nu(t) =: (N
1
N)u(t) = (C + DF)S(t)( x
0
x
0
)
+ (C + DF)
_
t
0
S(t s)[f(s, x(s)) f(s, x(s))]ds. (4.16)
Here x is the solution to equations (4.2) and x is the solution to
x

(t) = (A + BF)x(t) + f(t, x(t)) + Bu(t), x(0) = x

0
.
Therefore
x

(t) x

(t) = (A+ BF)( x(t) x(t)) + f(t, x(t)) f(t, x(t)). (4.17)
If condition (4.13) hold, then, for any > 0, we have
|Mu|

|FS(t)( x
0
x
0
)|

+ 2c|F|
_
t
0
e
(ts)
ds
r|F|| x
0
x
0
| + 2c|F|
1
,
|Nu|

|(C + DF)S(t)( x
0
x
0
)|

+ 2c|C + DF|
_
t
0
e
(ts)
ds
r|C + DF|| x
0
x
0
| + 2c|C + DF|
1
for all u |, and by Proposition 3.3 with k = 0,

(P, P
1
) = 0.
If condition (4.14) hold, then, for any > 0, from (4.17), it follows
| x(t) x(t)| |S(t)( x
0
x
0
)| +
_
t
0
|S(t s)f(s, x(s)) f(s, x(s))|ds
re
t
| x
0
x
0
| + rd
_
t
0
e
(ts)
| x(s) x(s)|ds.
This gives | x(t) x(t)| r| x
0
x
0
|e
(rd)t
r| x
0
x
0
| and, therefore
|f(t, x(t)) f(t, x(t))| rd| x
0
x
0
|.
Using the same argument as used above, we see

(P, P
1
) = 0.
By Theorem 3.2, we conclude:
Corollary 4.3. Consider the system given by equation (4.1). Under the assumptions
of Corollary, 4.2, if a controller K is such that [(f, x
0
), K] is -gain stable with bias,
then [(f, x
0
), K] is also -gain stable with bias for any x
0
.
14 W. BIAN
The existence of operator F satisfying (4.12) indicates that the nominal plant
(f, x
0
) is stabilizable with the feedback controller u = Fx + v. If the system is
of nite dimension, it is equivalent to that A + BF is Hurwitz and, in that case,
S(t) = exp[(A+BF)t] and is the maximal eigenvalue of A+BF. However, if the
nominal plant is not stabilizable, it is known that the initial state value is essential.
The next result shows that if a linear system is perturbed by a uniformly bounded
(nonlinear) function, its stability remains the same.
Corollary 4.4. Consider the system given by equation (4.1). Under the assumptions
of Corollary 4.2 and condition (4.13), we have

((0, x
0
), (f, x
0
)) = 0.
Proof. As shown in Lemma 4.1, there exist operators N, N
1
: } } and M, M
1
:
| } such that
graph((0, x
0
)) =
__
M
N
_
u : u |
_
, graph((f, x
0
)) =
__
M
1
N
1
_
u : u |
_
and
_
M
N
_
v(t) =:
_
M
1
M
N
1
N
_
v(t) =
_
F
_
t
0
S(t s)f(s, x(s))ds
(C + DF)
_
t
0
S(t s)f(s, x(s))ds
_
,
where x is the solution to the equation (4.2). Since f is uniformly bounded, as shown
in Corollary 4.2, there exists a constant c
f
> 0 such that
_
_
_
_
_
_
M
N
_
v
_
_
_
_
_

c
f
for all v |, > .
By Proposition 3.3 with k = 0,

((0, x
0
), (f, x
0
)) = 0.
5. Stability of realizations
It is known that, for a given transfer function, there are possibly innite many
dierent state space realizations of which the stability may vary. We now use Corol-
lary 4.2 to show that the gap between any two realizations which are stabilizable
is zero and, therefore, by Theorem 3.2, a controller stabilizing one realization also
stabilizes the other one.
We rst present a lemma regarding the estimate of gap metric.
Lemma 5.1. Let P
1
, P
2
, P
3
: |
a
}
a
and

(P
1
, P
2
) k
1
,

(P
2
, P
3
) k
2
with
k
1
, k
2
< . Then

(P
1
, P
3
) k
1
+ k
2
+ k
1
k
2
.
The same results hold if the gap metric

is replaced by

d.
GAP METRICS AND ROBUST STABILITY 15
Proof. Suppose > 0. By the denition of the gap metric, there are
1
> 0,
2
> 0
such that for each w
3
graph(P
3
), there exist w
2
graph(P
2
) and, therefore, w
1

graph(P
1
) satisfying:
|w
3
w
2
|

(k
2
+ )|w
2
|

+
2
, |w
2
w
1
|

(k
1
+ )|w
1
|

+
1
for all > 0. So
|w
3
w
1
|

|w
3
w
2
|

+|w
2
w
1
|

(k
2
+ )
_
|w
2
w
1
|

+|w
1
|

+
2
+ (k
1
+ )|w
1
|

+
1
(k
2
+ )
_
(1 + k
1
+ )|w
1
|

+
1

+
2
+ (k
1
+ )|w
1
|

+
1
.
So

(P
1
, P
3
) (k
2
+ )(1 + k
1
+ ) + k
1
+ . By let 0 we obtain

(P
1
, P
3
) k
1
+ k
2
+ k
1
k
2
.
Corollary 5.2. Suppose that G(s) is a transfer matrix having the following two re-
alizations of the same dimension
P
0
(x
0
) : x

(t) = Ax(t) + Bu(t), x(0) = x

0
,
y(t) = Cx(t) + Du(t),
P
1
(x
1
) : z

(t) = A
1
z(t) + B
1
u(t), z(0) = x
1
,
y(t) = C
1
z(t) + D
1
u(t),
where A, B, C, D and A
1
, B
1
, C
1
, D
1
are matrices of nite dimensions. If both P
0
(x
0
)
and P
1
(x
1
) are stabilizable, then

(P
0
(x
0
), P
1
(x
1
)) = 0. Consequently, if K is a
suitable controller such that [P
0
(x
0
), K] is -gain stable, then [P
1
(x
1
), K] is -gain
stable.
Proof. By our assumptions and Corollary 4.2, we see

(P
0
(x
0
), P
0
(0)) = 0 and

(P
1
(x
1
), P
1
(0)) = 0.
By Lemma 5.1

(P
0
(x
0
), P
1
(x
1
))

(P
0
(x
0
), P
0
(0)) +

(P
0
(0), P
1
(x
1
))
+

(P
0
(x
0
), P
0
(0))

(P
0
(0), P
1
(x
1
))
=

(P
0
(0), P
1
(x
1
))

(P
0
(0), P
1
(0)) +

(P
1
(0), P
1
(x
1
))
+

(P
0
(0), P
1
(0))

(P
1
(0), P
1
(x
1
))
=

(P
0
(0), P
1
(0)).
16 W. BIAN
As P
0
(0) and P
1
(0) have the same graph given by (u, y)

: y = Gu, we have

(P
0
(0), P
1
(0)) = 0. The -gain stability of [P
1
(x
1
), K] follows from Theorem 3.2.
This completes the proof.
6. A nonlinear system with time delay
In this section, we consider the stability of integrator system with saturation and
delay in the input. We let the nominal plant be the system without delay and the
perturbation be the system with delay, then applying the results in Section 3 to show
the stability of systems with delay.
Let | = } = L

(R
+
, R). The nominal and perturbed plants P, P
1
are described,
respectively, by
P : y
1
(t) = x(t), x

1
(t) = sat(u
1
(t)), x(0) = x
0
,
P
1
: y
1
(t) = x(t), x

1
(t) = sat(u
1
(t h)), x(0) = x
0
,
where sat(u
1
) = u
1
when [u
1
[ 1 and is equal to sign(u
1
) when [u
1
[ > 1.
It is known that
graph(P) =
__
M
N
_
u : u dom(P)
_
, graph(P
1
) =
__
M
1
N
1
_
u : u dom(P)
_
,
where M = M
1
= I (the identity), N = P and N
1
= P
1
. Moreover
_
M
N
_
u(t) =:
_
M
1
M
N
1
N
_
u(t) =
_
0
x(t h) x(t)
_
.
Since
[x(t h) x(t)[ sup
s[th,t]
[x

(s)[h sup
s[th,t]
[sat(u(s))[h,
we have
_
_
_
_
_
_
M
N
_
u
_
_
_
_
_

h|sat(u)|

h
for all > 0. Hence, by Proposition 3.3,

h
(P, P
1
) =

(P, P
1
) = 0.
Choose the feedback controller to be K = 1. Then both [P, K] and [P
1
, K]
are well-posed and [P, K] is 4-gain stable with zero bias (see [9]). By Theorem 3.2,
[P
1
, K] is 4-gain stable with bias for all h > 0.
However, if we apply the

0
-gap metric, since |sat(u)| |u|, we would have

0
(P, P
1
) h and [P
1
, K] is 4-gain stable with zero bias for h < 1/4.
GAP METRICS AND ROBUST STABILITY 17
REFERENCES
[1] W. Bian and M. French, Graph Topologies, Gap Metrics and Robust Stability for Nonlinear
Systems. SIAM Journal of Control and Optimization, 44 (2005), 418 - 443.
[2] W. Bian, M. French and A. Pr ugel-Bennett, Stability for Network Flow Control Against
Disturbances and Time-Delays. Submitted to SIAM Journal of Control and Optimization,
2009.
[3] C. Desoer and M. Vidyasagar. Feedback Systems Input Output Properties. Academic Press, 1st
edition, 1975.
[4] M. French. Smooth adaptive controllers have discontinuous closed loop operators. proceeding
of 16th Int. Symp. on Mathematical Theory of Networks and Systems, Leuven, 2004.
[5] M. French and W. Bian, A biased approach to nonlinear robust stability with applications in
adaptive control, Proceeding of 48th IEEE CDC/28th CCC, Shanghai, 2009 1393 - 1398
[6] M. French, Adaptive control and robustness in the gap metric. IEEE Transactions on Auto-
matic Control, 53 (2008), 461478.
[7] M. French, A. Ilchmann and M. Mueller, Robust stabilization by linear output delay feedback.
SIAM Journal on Control and Optimization, 48(2009), 25332561.
[8] T. T. Georgiou and M. C. Smith, Optimal robustness in the gap metric, IEEE Transactions
on Automatic Control, 35(1990), 673686
[9] T. T. Georgiou and M. C. Smith,, Robustness analysis of nonlinear feedback systems: an
input-output approach, IEEE Transactions on Automatic Control, 42(1997), 12001221
[10] T. T. Georgiou and M. C. Smith, Biased norms and robustness analysis for nonlinear feedback
systems, Proceeding of 36th IEEE CDC, San Diego, 1997, 642643
[11] M. R. James, M. C. Smith and G. Vinnicombe, Gap metrics, representations and nonlinear
robust stability, SIAM J. Control Optim., 43(2005), 1535-1582.
[12] D.Li, N. Hovakimyan and T. Georgiou, Robustness of L
1
adaptive controllers in the gap metric,
2010 American Control Conference, Baltimore, 2010, 3247 - 3252
[13] M. S. Verma, Coprime fractional representations and stability of nonlinear feedback systems,
Int. J. Control, 48(1988), 897918
[14] M. S. Verma and L. R. Hunt, Right coprime fractorizations and stabilization for nonlinear
systems, IEEE Transactions on Automatic Control, 38(1993), 222231
[15] M. Vidyasagar. Nonlinear Systems Analysis. Prentice Hall, 1993.
[16] Zames G. and El-Sakkary A. K., Unstable systems and feedback: The gap metric, Proceeding
of the Allerton Conf., (Oct. 1980), 380385