This action might not be possible to undo. Are you sure you want to continue?

**www.elsevier.com/locate/laa
**

H

∞

-control of linear state-delay descriptor

systems: an LMI approach

ୋ

E. Fridman

∗

, U. Shaked

Department of Electrical Engineering-Systems, Tel Aviv University, Ramat Aviv 69978, Tel Aviv, Israel

Received 20 January 2001; accepted 30 October 2001

Submitted by P. Van Dooren

Abstract

For continuous-time, linear descriptor system with state-delay a H

∞

-control problem is

solved. Sufﬁcient conditions for delay-dependent/delay-independent stability and L

2

-gain

analysis are obtained in terms of linear matrix inequalities (LMIs). A bounded real lemma

and state-feedback solutions are derived for systems which may contain polytopic parameter

uncertainties. The ﬁltering problem is also solved and an output-feedback controller is then

found by solving two LMIs. The ﬁrst LMI is associated with a proportional-derivative state-

feedback control. The second LMI is derived in two different forms, the ﬁrst one corresponds

to the adjoint of the system that describes the estimation error and the other stems from the

original system. These two forms lead to different results. Numerical examples are given

which illustrate the effectiveness of the new theory. © 2002 Elsevier Science Inc. All rights

reserved.

Keywords: Time-delay systems; Descriptor systems; H

∞

-control; LMI; Delay-dependent criteria

1. Introduction

Delay differential–algebraic equations, which have both delay and algebraic con-

straints, often appear in various engineering systems, including aircraft stabilization,

chemical engineering systems, lossless transition lines, etc. (see e.g. [4,13,14,22,28],

ୋ

This work was supported by the Ministry of Absorption of Israel and by C&M Maus Chair at Tel Aviv

University.

∗

Corresponding author. Tel.: +972-3640-5313; fax: +972-3640-7095.

E-mail address: emilia@eng.tau.ac.il (E. Fridman).

0024-3795/02/$ - see front matter 2002 Elsevier Science Inc. All rights reserved.

PII: S0024- 3795( 01) 00563- 8

272 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

and the references therein). Depending on the area of application, these models are

called singular or implicit or descriptor systems with delay. As has been pointed out

in [6,7], descriptor systems with delay may in fact be systems of advanced type.

Descriptor systems may be destabilized by small delay in the feedback [20].

There are only few papers on descriptor systems with delay [6,7,10,15,22,23,28].

A particular case of these systems (the so-called lossless propagation models), de-

scribed by

˙ x

1

(t ) = Ax

1

(t ) + Bx

2

(t − h), x

2

(t ) = Cx

1

(t ) + Dx

2

(t − h), (1)

has been treated as a special class of neutral systems either by letting x

2

(t ) = ˙ y

2

(t )

[23] or by writing the second equation as [15,22]

d

dt

_

x

2

(t ) − Cx

1

(t ) − Dx

2

(t − h)

_

= 0. (2)

The stability of a general neutral type descriptor system with a single delay de-

scribed by

E˙ x(t ) + Ax(t ) + B ˙ x(t − h) + Cx(t − h) = 0 (3)

with a singular matrix E has been studied in [28] by analyzing its characteristic

equation

det

_

sE + A + (sB + C) exp(−hs)

_

= 0

and ﬁnding frequency domain conditions which guarantee that all roots of the latter

equation have negative real parts bounded away from 0. A Lyapunov-based approach

to stability analysis of descriptor system with delay has been introduced in [10],

where delay-independent and delay-dependent linear matrix inequalities (LMIs) con-

ditions have been derived. For information on LMI approach to control, see [3].

All the above-mentioned results only analyze the existence and the stability of so-

lutions of descriptor systems with delay. To the best of our knowledge no

control problem solution has been derived for this class of systems. For descriptor

systems without delay, H

∞

-control problems have been treated in the frequency

domain [19,27] and in the time-domain [21,26,29]. In [21,26] an LMI approach has

been proposed. For nondescriptor systems with state-delay, LMI delay-dependent

and delay-independent H

∞

-controllers were derived in [12,16,24] (see also the refer-

ences therein). These ﬁnite-dimensional LMIs provide sufﬁcient conditions only for

inﬁnite-dimensional systems with state-delay. Unlike inﬁnite-dimensional methods

(see e.g. [1,11]) they lead to effective numerical algorithms and may be applied for

systems with polytopic uncertainties.

In the present paper, we adopt the ﬁnite-dimensional LMI approach to H

∞

-con-

trol of descriptor system with delay. Our objective is to obtain delay-dependent so-

lutions which are less conservative than the delay-independent ones. We apply the

descriptor model transformation that has been introduced recently for delay-depen-

dent stability and H

∞

-control of nondescriptor systems [9,12]. We derive bounded

real lemmas (BRLs) and ﬁnd solutions to the H

∞

ﬁltering, the state-feedback and

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 273

the output-feedback H

∞

-control problems. The solutions are delay-dependent with

respect to the ‘slow’ variable and delay-independent with respect to the ‘fast’ one.

The latter guarantees the robustness of the system behavior with respect to the small

changes in the delay.

Notation. Throughout the paper the superscript ‘T’ stands for matrix transposi-

tion, R

n

denotes the n-dimensional Euclidean space with vector norm | · |, R

n×m

is

the set of all n × m real matrices, and the notation P > 0 for P ∈ R

n×n

means that

P is symmetric and positive definite. The space of functions in R

q

that are square in-

tegrable over [0 ∞) is denoted by L

q

2

[0, ∞) with norm ·

L

2

. Let C

n

[a, b] denote

the space of continuous functions φ : [a, b] →R

n

with the supremum norm | · |. We

also denote x

t

(θ) = x(t + θ) (θ ∈ [−h, 0]) and j =

√

−1.

2. Problem formulation

Given the following system:

E˙ x(t ) =

2

i=0

A

i

x(t − h

i

) + B

1

w(t ) + B

2

u(t ), x(t ) = 0, ∀t 0, (4)

¯ y(t ) =

¯

C

2

x(t ) + D

21

w(t ), (5a)

z(t ) = col

_

¯

C

1

x(t ); D

12

u(t )

_

, (5b)

where x(t ) = col{x

1

(t ), x

2

(t )}, x

1

(t ) ∈ R

n

1

, x

2

(t ) ∈ R

n

2

is the system state vector,

u(t ) ∈ R

**is the control input, w(t ) ∈ L
**

q

2

[0, ∞] is the exogenous disturbance signal,

¯ y(T ) ∈ R

r

is the measurement vector and z(t ) ∈ R

p

is the state combination (objec-

tive function signal) to be attenuated. The time delays h

0

= 0, h

i

> 0, i = 1, 2, are

assumed to be known. We took for simplicity two delays, but all the results are easily

generalized for the case of any ﬁnite number of delays h

1

, . . . , h

m

. The singular

matrix E and the matrices A

i

, B

i

are constant matrices of appropriate dimensions.

Denote n n

1

+ n

2

.

Following [21,26], we assume for simplicity that

E =

_

I

n

1

0

0 0

_

. (6)

Every descriptor system can be represented in a form satisfying this assumption.

Note that in [21] there is =

T

> 0 instead of I

n

1

, but from such a system, the

system with E of (6) follows immediately. The matrices in (4), (5a) and (5b) have the

following structure:

A

i

=

_

A

i1

A

i2

A

i3

A

i4

_

, i = 0, 1, 2,

B

i

=

_

B

i1

B

i2

_

,

¯

C

i

=

_

¯

C

i1

¯

C

i2

_

, i = 1, 2.

(7)

274 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

A descriptor system without delay

E˙ x(t ) = A

0

x(t ) + B

1

w(t ) (8)

is regular if the characteristic polynomial det(sE − A

0

) does not vanish identically

in s ∈ C. It is well known that descriptor system may have impulsive solutions. The

existence of the latter solutions is usually studied in terms of the Weierstrass canon-

ical form and the index of the system which are deﬁned as follows [5,8,20]: there

exist nonsingular matrices P, Q ∈ R

n×n

such that

QEP =

_

I

n

1

0

0 N

_

, QA

0

P =

_

J 0

0 I

n

2

_

, (9)

where N ∈ R

n

2

×n

2

and J ∈ R

n

1

×n

1

are in Jordan form. The matrix N is nilpotent of

index ν, i.e. N

ν

= 0, N

ν−1

/ = 0. The index of (8) is the index of nilpotence ν of N.

The index of the system with delay

E˙ x(t ) =

2

i=0

A

i

x(t − h

i

) + B

1

w(t ) (10)

is deﬁned in [10] to be equal to the index of (8). The descriptor system (10) admits

impulsive solutions iff ν > 1 [10].

We do not require A

04

in (4) to be nonsingular. If A

04

is singular, then (8) has

index greater than 1 (see e.g. [5,8]). Hence, the index of the open loop system (4)

with delay is higher than one. Such a system may have an impulsive solution. The

nonsingularity of A

04

guarantees the existence and the uniqueness of solution to (4)

with u = 0 (see Proposition 3.1).

The following class of neutral descriptor systems

_

˙ x

1

(t ) −

2

i=1

F

i

˙ x

1

(t − h

i

)

0

_

=

2

i=0

A

i

x(t − h

i

) + B

1

w(t ) + B

2

u(t ), (11)

can be reduced to the form of (4) and (6). This follows from the fact that the aug-

mented system

˙ x

1

(t ) = y,

_

y(t ) −

2

i=1

F

i

y(t − h

i

)

0

_

=

2

i=0

A

i

x(t − h

i

) + B

1

w(t ) + B

2

u(t ),

(12)

is a particular case of (4) and (6).

For a prescribed scalar γ > 0, we deﬁne the performance index

J(w) =

_

∞

0

_

z

T

z − γ

2

w

T

w

_

dτ. (13)

The problem is to ﬁnd a controller such that the resulting closed loop system has in-

dex at most one is internally stable (i.e. asymptotically stable for w = 0) and J(w) <

0 for all disturbances w(t ) ∈ L

q

2

[0, ∞].

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 275

3. Stability and L

2

-gain analysis of a descriptor system with delay

BRLs will be obtained for systems with discrete and distributed delays. Given the

following system:

E˙ x(t ) =

2

i=0

A

i

x(t − h

i

) +

_

0

−d

A

d

(s)x(t + s) ds + B

1

w(t ), (14a)

x(t ) = 0 ∀t 0, (14b)

z(t ) = col

_

C

0

x(t ), C

1

x(t − h

1

), C

2

x(t − h

2

)

_

, (15)

where E is deﬁned in (6), x(t ) = col{x

1

(t ), x

2

(t )}, x

1

(t ) ∈ R

n

1

, x

2

(t ) ∈ R

n

2

, is the

systemstate vector, w(t ) ∈ L

q

2

[0, ∞] is the exogenous disturbance signal and z(t ) ∈

R

p

is objective function signal, A

d

(s) is a piecewise-continuous and uniformly

bounded (n

1

+ n

2

) × (n

1

+ n

2

)-matrix-function. We assume that the matrices in

(14a) and (14b) have the structure of (7) and

C

i

=

_

C

i1

C

i2

_

, i = 0, 1, 2, A

d

=

_

A

d1

A

d2

A

d3

A

d4

_

.

Denote h = max{h

1

, h

2

, d}. By solution of (14a) and (14b) on the segment [0, t

1

]

(t

1

> 0) we understand a pair of functions {x

1

(t ), x

2

(t )}, such that x

1

is absolutely

continuous and x

2

is integrable on [0, t

1

], these functions satisfy system (14a) almost

for all t ∈ [0, t

1

] and the initial conditions (14b).

Proposition 3.1. Assume that A

04

is nonsingular. For w(t ) ∈ L

q

2

[0, ∞) the solu-

tion to (14a) and (14b) exists and is unique on [0, t

1

] for all t

1

> 0.

Proof. By denoting ˙ y

2

= x

2

we obtain from (14a) a neutral type system with the ze-

ro initial conditions x

1

(t ) = 0, ˙ y

2

(t ) = 0 ∀t 0. Hence, y

2

(t ) = c ∀t 0, c ∈ R

n

2

.

This initial value problem for neutral system has a piecewise absolutely continuous

solution x

1

(t ), y

2

(t ) on [0, t

1

] such that x

1

(t ) is absolutely continuous on [0, t

1

] (see

[17, p. 143]). Therefore, x

2

(t ) is integrable and solution of (14a) and (14b) exists.

To prove the uniqueness we assume that there are two solutions of (14a) and

(14b). Then their difference satisﬁes the homogeneous equations (14a) and (14b)

with w = 0, which has a unique solution x ≡ 0 [10]. Hence, solution of (14a) and

(14b) is unique.

3.1. Stability of the difference operator and of the descriptor system

We assume:

A1. The matrix A

04

is nonsingular and the difference operator D : C

n

2

[−h, 0] →

R

n

given by

276 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

D(x

2t

) = x

2

(t ) +

2

i=1

A

−1

04

A

i4

x

2

(t − h

i

) +

_

0

−d

A

−1

04

A

d4

(s)x

2

(t + s) ds

is stable for all delays h

1

and h

2

(i.e. equation Dx

2t

= 0 is asymptotically stable for

all h

1

and h

2

).

A sufﬁcient condition for A1 is the following inequality:

2

i=1

¸

¸

A

−1

04

A

i4

¸

¸

+

_

0

−d

¸

¸

A

−1

04

A

d4

(s)

¸

¸

ds < 1,

where | · | is any matrix norm.

In the case of single delay (e.g. h

1

) in the fast variable x

2

we assume instead of

A1 the following:

A1

**. All the eigenvalues of A
**

−1

04

A

14

are inside of the unit circle.

In the case of multiple discrete delays in D, where A

d4

= 0, A1 is equivalent to

the following one (see [14, Theorem 6.1, p. 286]):

A1

**. If σ(B) is the spectral radius of matrix B, then σ
**

0

< 1, where

σ

0

sup

_

σ

_

2

k=1

A

−1

04

A

i4

e

iθ

k

_

: θ

k

∈ [0, 2π], k = 1, 2

_

. (16)

Evidently A1

is equivalent to A1

**in the case of single delay h
**

1

. A sufﬁcient

LMI condition for A1

**is given by the following:
**

Lemma 3.2 [10]. If there exist n

2

× n

2

-matrices P

f

, U

1f

, U

2f

that satisfy the fol-

lowing LMI:

_

_

P

T

f

A

04

+ A

T

04

P

f

+

2

i=1

U

if

P

T

f

A

14

P

T

f

A

24

∗ −U

1f

0

∗ ∗ −U

2f

_

_

< 0, (17)

then A

04

is nonsingular and

(i) A1

holds;

(ii) the difference operator

D(x

2t

) = x

2

(t ) +

2

i=1

A

−1

04

A

i4

x

2

(t − h

i

)

is stable for all h

1

, h

2

;

(iii) under additional assumption that P

f

> 0 the “fast system”

˙ x

2

(t ) = A

04

x

2

(t ) +

2

i=1

A

i4

x

2

(t − h

i

) (18)

is asymptotically stable for all h

1

, h

2

.

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 277

The following result on stability of (14a) and (14b) with w = 0 has been obtained

recently [10].

Lemma 3.3. Under A1 if there exist positive numbers α, β, γ and a continuous

functional V : C

n

[−h, 0] →R such that

β|φ

1

(0)|

2

V(φ) γ |φ|

2

, (19a)

˙

V(φ) −α|φ(0)|

2

, (19b)

and the function

¯

V(t ) = V(x

t

) is absolutely continuous for x

t

satisfying (14a) and

(14b) with w = 0, then (14a), (14b) with w = 0 is asymptotically stable.

3.2. Delay-independent BRL (with respect to discrete delays)

Descriptor type Lyapunov–Krasovskii functional for system (14a), (14b) has the

following form:

V(x

t

) = x

T

(t )EPx(t ) + V

1

+ V

2

, (20)

where

P =

_

P

1

0

P

2

P

3

_

, P

1

= P

T

1

> 0, (21)

V

1

=

2

i=1

_

t

t −h

i

x

T

(s)U

i

x(s) ds, U

i

> 0, (22)

and

V

2

=

_

0

−d

_

t

t +θ

x

T

(s)A

T

d

(θ)RA

d

(θ)x(s) ds dθ, R > 0. (23)

The ﬁrst term of (20) corresponds to the descriptor system, V

1

corresponds to the

delay-independent stability with respect to the discrete delays and V

2

—to delay-

dependent stability with respect to the distributed delays [18]. The functional (20) is

degenerated (i.e. nonpositive-definite) as it is usual for descriptor systems.

We obtain analogously to [10] the following:

Theorem 3.4. Under A1 (14a) and (14b), (15) is internally asymptotically stable

and for a prechosen γ > 0 J(w) < 0 for all nonzero w(t ) ∈ L

q

2

[0, ∞] and for all

h

1

0, h

2

0 if there exist n × n-matrix P of (21) with n

1

× n

1

-matrix P

1

and

n

2

× n

2

-matrix P

3

and n × n matrices U

i

= U

T

i

, i = 1, 2, R = R

T

that satisfy the

following LMI:

278 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

_

_

_

_

P

T

A

0

+ A

T

0

P +

2

i=1

U

i

+

_

0

−d

A

T

d

(s)RA

d

(s) ds P

T

B

1

P

T

A

1

P

T

A

2

dP

T

∗ −γ

2

I

q

0 0 0

∗ ∗ −U

1

0 0

∗ ∗ ∗ −U

2

0

∗ ∗ ∗ ∗ −dR

_

¸

¸

_

< 0.

(24)

Remark 1. From Theorem 3.4 it follows that the system

E˙ x(t ) = A

0

x(t ) + A

1

x(t − h) (25)

is asymptotically stable for all h 0 if the following LMI is feasible:

_

P

T

A

0

+ A

T

0

P P

T

A

1

A

T

1

P 0

_

+

_

U

1

0

0 −U

1

_

< 0. (26)

Multiplying (26) by from the left and by

∗

from the right, where

=

_

A

T

1

(−jωE − A

T

0

)

−1

I

_

,

∗

=

_

(jωE − A

0

)

−1

A

1

I

_

, ω ∈ R,

the following frequency domain inequality is readily obtained

diag

_

U

1

, −U

1

_

∗

< 0

or

A

T

1

_

− jωE − A

T

0

_

−1

U

1

(jωE − A

0

)

−1

A

1

< U

1

. (27)

Therefore, if LMI (24) is feasible (and thus, (26) is feasible), then for all ω ∈ R

the frequency domain inequality (27) holds. Hence the H

∞

-norm of U

1/2

1

(jωE −

A

0

)

−1

A

1

U

−1/2

1

is less than 1. This is a counterpart of the Kalman–Yakubovich–

Popov lemma for descriptor systems.

3.3. Delay-dependent BRL

We are looking for delay-dependent conditions with respect to slow variable x

1

.

With respect to discrete delays in the fast variables the results will be delay-inde-

pendent. The latter guarantees robust stability with respect to small changes of delay

[10]. Following [9,10] we represent (14a) and (14b) in the equivalent form:

˙ x

1

(t ) = y(t ),

_

y(t )

0

_

=

_

2

i=0

A

i1

A

02

2

i=0

A

i3

A

04

_

x(t ) +

2

i=1

_

A

i2

A

i4

_

x

2

(t − h

i

)

−

2

i=1

_

A

i1

A

i3

_ _

0

−h

i

y(t + s) ds +

_

0

−d

A

d

(s)x(t + s) ds + B

1

w(t ).

(28)

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 279

The latter system can be represented in the form:

¯

E

˙

¯ x(t ) =

2

i=0

¯

A

i

¯ x(t − h

i

) +

2

i=1

h

i

H

i

_

0

−h

i

y(t + s) ds

+

_

0

−d

¯

A

d

(s) ¯ x(t + s) ds +

¯

B

1

w(t ), (29)

where

¯ x =

_

_

x

1

y

x

2

_

_

,

¯

E =

_

_

I

n

1

0 0

0 0

n

1

×n

1

0

0 0 0

n

2

×n

2

_

_

,

¯

A

0

=

_

_

_

0 I 0

2

i=0

A

i1

−I

n

1

A

02

2

i=0

A

i3

0 A

04

_

¸

_

,

¯

A

i

=

_

_

0 0 0

0 0 A

i2

0 0 A

i4

_

_

, i = 1, 2,

H

i

=

_

_

0

A

i1

A

i3

_

_

,

¯

A

d

=

_

_

0 0 0

A

d1

0 A

d2

A

d3

0 A

d4

_

_

,

¯

B

1

=

_

0

B

1

_

.

(30)

A Lyapunov–Krasovskii functional for system (28) has the form:

V(t ) = ¯ x

T

(t )

¯

EP ¯ x(t ) +

2

i=1

_

t

t −h

i

x

T

1

(τ)S

i

x

1

(τ) dτ

+

2

i=1

_

t

t −h

i

x

T

2

(τ)U

i

x

2

(τ) dτ

+

2

i=1

_

0

−h

i

_

t

t +θ

y

T

(s)

_

A

T

i1

A

T

i3

_

R

i3

_

A

i1

A

i3

_

y(s) ds dθ

+

_

0

−d

_

t

t +θ

¯ x

T

(s)

¯

A

T

d

R

d

¯

A

d

¯ x(s) ds dθ, (31)

where P has the structure of (21) with P

1

∈ R

n

1

×n

1

, P

3

∈ R

n×n

and

0 < S

i

∈ R

n

1

×n

1

, 0 < U

i

∈ R

n

2

×n

2

,

0 < R

i3

∈ R

n×n

, R

d

∈ R

(n

1

+n)×(n

1

+n)

.

The ﬁrst term of (31) corresponds to the descriptor system, the second and the fourth

terms—to the delay-dependent conditions with respect to x

1

and the third—to the

delay-independent conditions with respect to x

2

, the ﬁfth term corresponds to delay-

dependent with respect to distributed delay.

280 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

We obtain the following:

Theorem 3.5. Under A1 the system of (14a), (14b) and (15) is internally asymptot-

ically stable and for a prechosen γ > 0, J(w) < 0 for all nonzero w(t ) ∈ L

q

2

[0, ∞]

if there exist matrices P ∈ R

(n

1

+n)×(n

1

+n)

of (21), 0 < P

1

∈ R

n

1

×n

1

, P

2

, P

3

∈

R

n×n

, S

i

= S

T

i

∈ R

n

1

×n

1

, U

i

= U

T

i

∈ R

n

2

×n

2

, W

i

∈ R

(n

1

+n)×(n

1

+n)

and R

i

= R

T

i

∈

R

(n

1

+n)×(n

1

+n)

, i = 1, 2 that satisfy the following LMI:

_

_

_

_

_

_

_

_

_

_

_

¯

P

T

_

0

B

1

_

h

1

X

1

h

2

X

2

−W

T

1

_

_

0

A

11

A

13

_

_

−W

T

2

_

_

0

A

21

A

23

_

_

P

T

_

_

0

A

12

A

14

_

_

P

T

_

_

0

A

22

A

24

_

_

dP

T

∗ −γ

2

I

q

0 0 0 0 0 0 0

∗ ∗ −h

1

R

1

0 0 0 0 0 0

∗ ∗ ∗ −h

2

R

2

0 0 0 0 0

∗ ∗ ∗ ∗ −S

1

0 0 0 0

∗ ∗ ∗ ∗ ∗ −S

2

0 0 0

∗ ∗ ∗ ∗ ∗ ∗ −U

1

0 0

∗ ∗ ∗ ∗ ∗ ∗ ∗ −U

2

0

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ −dR

d

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

< 0,

(32)

where

X

i

= W

T

i

+ P

T

, i = 1, 2,

¯

= +

2

i=0

_

_

C

T

i1

C

i1

0

n

1

0

0 0 0

0 0 C

T

i2

C

i2

_

_

+

2

i=1

W

T

i

_

_

0 0 0

A

i1

0 0

A

i3

0 0

_

_

+

2

i=1

_

_

0 A

T

i1

A

T

i3

0 0 0

0 0 0

_

_

W

i

and

P

T

_

_

_

0 I

n

1

0

2

i=0

A

i1

−I

n

1

A

02

2

i=0

A

i3

0 A

04

_

¸

_

+

_

_

_

0 I

n

1

0

2

i=0

A

i1

−I

n

1

A

02

2

i=0

A

i3

0 A

04

_

¸

_

T

P

+

2

i=1

_

_

_

_

S

i

0 0

0 h

i

_

A

T

i1

A

T

i3

_

R

i3

_

A

i1

A

i3

_

0

0 0 U

i

_

¸

¸

_

+

_

0

−d

¯

A

T

d

(s)R

d

¯

A

d

(s) ds (33)

and where R

i3

∈ R

n×n

is the (2, 2) block of R

i

.

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 281

Proof. Since

¯ x

T

(t )

¯

EP ¯ x(t ) = x

T

1

(t )P

1

x

1

(t )

differentiating the ﬁrst term of (31) with respect to t we have

d

dt

¯ x

T

(t )

¯

EP ¯ x(t ) = 2x

T

1

(t )P

1

˙ x

1

(t ) = 2¯ x

T

(t )P

T

_

_

˙ x

1

(t )

0

0

_

_

. (34)

Substituting (28) into (34) we obtain

dV(x

t

)

dt

+ z

T

(t )z(t ) − γ

2

w

T

(t )w(t )

= ξ

T

_

_

_

_

_

_

_

_

P

T

_

0

B

1

_

P

T

_

_

0

A

12

A

14

_

_

P

T

_

_

0

A

22

A

24

_

_

∗ −γ

2

I

q

0 0

∗ ∗ −U

1

0

∗ ∗ ∗ −U

2

_

¸

¸

¸

¸

¸

¸

_

ξ + z

T

z +

2

i=0

η

i

−

2

i=1

_

x

T

1

(t − h

i

)S

i

x

1

(t − h

i

) + x

T

2

(t − h

i

)U

i

x

2

(t − h

i

)

+

_

t

t −h

i

y

T

(s)

_

A

T

i1

A

T

i3

_

R

i3

_

A

i1

A

i3

_

y(s) ds

_

−

_

0

−d

¯ x

T

(t + θ)

¯

A

T

d

(θ)R

d

A

d

(θ) ¯ x(t + θ) dθ, (35)

where

ξ col

_

¯ x(t ), w(t ), x

2

(t − h

1

), x

2

(t − h

2

)

_

,

is deﬁned by (33) and

η

i

(t ) − 2

_

t

t −h

i

¯ x

T

(t )P

T

_

_

0

A

i1

A

i3

_

_

y(s) ds, i = 1, 2,

η

0

(t ) − 2

_

t

t −d

¯ x

T

(t )P

T

¯

A

d

(s) ¯ x(t + s) ds.

For any (n

1

+ n) × (n

1

+ n)-matrices R

i

> 0 and M

i

the following inequality holds

[25]:

−2

_

t

t −h

i

b

T

(s)a(s) ds

_

t

t −h

i

_

a(s)

b(s)

_

T

_

R

i

R

i

M

i

M

T

i

R

i

(2, 2)

_ _

a(s)

b(s)

_

ds (36)

282 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

for a(s) ∈ R

n

1

+n

, b(s) ∈ R

n

1

+n

given for s ∈ [t − h

i

, t ]. Here

(2, 2) = (M

T

i

R

i

+ I)R

−1

i

(R

i

M

i

+ I).

Denoting W

i

= R

i

M

i

P and using this inequality for a(s) = col{0 A

i1

A

i3

}y(s)

and b = P ¯ x(t ) we obtain for i = 1, 2

η

i

(t ) h

i

¯ x

T

(t )(W

T

i

+ P

T

)R

−1

i

(W

i

+ P) ¯ x(t )

+2

_

x

T

1

(t ) − x

T

1

(t − h

i

)

_ _

0 A

T

i1

A

T

i3

_

W

i

¯ x(t )

+

_

t

t −h

i

y

T

(s)

_

A

T

i1

A

T

i3

_

R

i3

_

A

i1

A

i3

_

y(s) ds. (37)

Similarly

η

0

(t ) d¯ x

T

(t )P

T

R

−1

d

P ¯ x(t )

+

_

t

t −d

¯ x

T

(t + s)

¯

A

T

d

(s)R

d

¯

A

d

(s) ¯ x(t + s) ds. (38)

We substitute (37), (38) into (35) and integrate the resulting inequality in t from 0

to ∞. Because V(x

0

) = 0, V(x

∞

) 0 and

_

∞

0

z

T

z dt =

2

i=0

_

∞

0

x

T

(t − h

i

)C

T

i

C

i

x(t − h

i

) dt

=

2

i=0

_

∞

0

x

T

(t )C

T

i

C

i

x(t ) dt,

we obtain (by Schur complements) that

z

2

L

2

− γ

2

w

2

L

2

¯

ξ

T

¯

ξ < −αx

2

L

2

, α > 0,

where is the matrix in the left-hand side of (32) and

¯

ξ col{ ¯ x(t ), w(t ), x

2

(t − h

1

), x

2

(t − h

2

), ¯ η},

where ¯ η is vector of ﬁctitious states. Hence, for w(t ) ∈ L

2

[0, ∞] we have x(t ) ∈

L

2

[0, ∞] and J(w) < 0 if (32) holds. Moreover, V of (31) satisﬁes (19a), (19b) and

hence (14a) and (14b) are internally stable.

3.4. Another delay-independent BRL (with respect to discrete delays)

For

W

i

= −P, R

i

=

εI

2n

h

i

, i = 1, . . . , m, (39)

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 283

LMI (32) implies for ε → 0

+

the following delay-independent LMI:

_

_

_

_

_

_

_

_

_

_

_

_

_

_

P

T

_

0

B

1

_

P

T

_

_

0

A

11

A

13

_

_

P

T

_

_

0

A

21

A

23

_

_

P

T

_

_

0

A

12

A

14

_

_

P

T

_

_

0

A

22

A

24

_

_

dP

T

∗ −γ

2

I

q

0 0 0 0 0

∗ ∗ −S

1

0 0 0 0

∗ ∗ ∗ −S

2

0 0 0

∗ ∗ ∗ ∗ −U

1

0 0

∗ ∗ ∗ ∗ ∗ −U

2

0

∗ ∗ ∗ ∗ ∗ ∗ −dR

d

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

< 0,

(40)

where

= P

T

_

_

0 I 0

A

01

−I

n

1

A

02

A

03

0 A

04

_

_

+

_

_

0 I 0

A

01

−I

n

1

A

02

A

03

0 A

04

_

_

T

P

+

2

i=1

_

_

S

i

0 0

0 0 0

0 0 U

i

_

_

+

_

0

−d

¯

A

T

d

(s)R

d

¯

A

d

(s) ds.

If LMI (40) is feasible, then (32) is feasible for a small enough ε > 0 and for R

i

and W

i

that are given by (39). Thus, from Theorem 2.1 the following corollary

holds:

Corollary 3.6. Under A1 the system of (14a), (14b), and (15) is stable for all

h

i

> 0, i = 1, 2 and J < 0 if there exist 0 < P

1

= P

T

1

, P

2

, P

3

, U

i

= U

T

i

and S

i

=

S

T

i

, i = 1, 2 that satisfy (40).

Remark 2. As we have seen above, the delay-dependent BRL of Theorem 2.1 is

most powerful in the sense that it provides sufﬁcient conditions for both the delay-

dependent and the delay-independent cases (where (40) holds). In the latter case,

(32) is feasible for h

i

→ ∞, i = 1, 2.

3.5. Delay-dependent BRL for systems with polytopic uncertainties

The BRL of Theorem 2.1 was derived for system (14a), (14b) where the system

matrices A

i

, C

i

, i = 1, 2, B

1

, A

d

are all known. However, since the LMI of (32) is

afﬁne in the system matrices, the theorem can be used to derive a criterion that will

guarantee stability and the required attenuation level in the case where the system

matrices are not exactly known and they reside within a given polytope.

284 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

Denoting

=

_

A

i

A

d

B

1

C

i

i = 0, 1, 2

_

we assume that ∈ Co{

j

, j = 1, . . . , N}, namely,

=

N

j=1

f

j

j

for some 0 f

j

1,

N

j=1

f

j

= 1,

where the N vertices of the polytope are described by

j

=

_

A

(j)

i

A

(j)

d

B

(j)

1

C

(j)

i

i = 0, 1, 2

_

.

We readily obtain the following:

Corollary 3.7. Assume that for all j = 1, . . . , N, A1 holds. Consider the system

of (14a), (14b), where the system matrices reside within the polytope . For a pre-

scribed γ > 0, the cost function (13) achieves J(w) < 0 over for all nonzero

w ∈ L

q

2

[0, ∞) if there exist n × n-matrices 0 < P

(j)

1

, W

(j)

i1

, W

(j)

i2

, W

i3

, W

i4

, j =

1, . . . , N, P

2

, P

3

, and 0 < R

(j)

i

, 0 < R

(j)

d

, 0 < U

(j)

i

, 0 < S

(j)

i

, i = 1, 2, j = 1,

. . . , N that satisfy (32) for j = 1, . . . , N, where the matrices

A

i

, A

d

, B

1

, C

i

, P

1

, W

1

, W

2

, R

1

, R

2

, R

d

, S

1

, S

2

, i = 0, 1, 2,

are taken with the upper index j.

3.6. On LMI conditions in the case of discrete delays

We consider A

d

= 0. Even in this simpler case condition A1 is not easily veriﬁ-

able. That is why instead of A1 one can assume that the fast LMI (17) is feasible for

some P

f

, U

kf

, k = 1, 2. Another possibility is to look for P

3

in Theorem 3.4 in the

diagonal form:

P

3

= diag

_

P

31

, P

32

_

, P

32

∈ R

n

2

×n

2

. (41)

In the latter case if the full-order LMI (24) holds for P

3

of (41), then (17) holds for

P

f

= P

32

, where U

kf

are (2, 2) blocks of U

k

.

Consider now a difference continuous system

0 = A

04

x

2

(t ) +

2

i=1

A

i4

x

2

(t − h

i

) + B

12

w(t ),

z(t ) = col

_

C

02

x

2

, C

12

x

2

(t − h

1

), C

22

x

2

(t − h

2

)

_

.

(42)

From Theorem 3.4, the following (delay-independent) BRL follows:

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 285

Corollary 3.8. Given γ > 0, if there exist n

2

× n

2

-matrices P

f

, U

1f

, U

2f

that sat-

isfy the following LMI:

_

_

_

_

P

T

f

A

04

+ A

T

04

P

f

+

2

i=1

U

if

+

2

i=0

C

T

i2

C

i2

P

T

f

A

14

P

T

f

A

24

P

T

f

B

12

∗ −U

1f

0 0

∗ ∗ −U

2f

0

∗ ∗ ∗ −γ

2

I

_

¸

¸

_

< 0,

(43)

then for all h

1

> 0, h

2

> 0 the difference system (42) is internally stable and

J(w) < 0.

3.7. H

∞

-norm of the ‘adjoint’ system

We begin by noting that the H

∞

-norm of the system

1

of (14a) and (14b), where

z = C

0

x(t ), A

d

= 0, (44)

is given by (see e.g. [2, vol. 2, p. 32]):

1

∞

= sup

ω∈R

¯ σ

_

_

_

C

0

_

jωE − A

0

−

2

i=1

A

i

e

−jwh

i

_

−1

B

1

_

_

_

, (45)

where ¯ σ{D} denotes the largest singular value of D. Since

¯ σ

_

H(jω)

_

= ¯ σ

_

H

T

(−jω)

_

for all the transfer function matrices H(s) with real coefﬁcients, it follows that the

H

∞

-norm of

1

is equal to the H

∞

-norm of the following system:

−E

˙

ξ(t ) =

2

i=0

A

T

i

ξ(t + h

i

) + C

T

0

˜ z(t ), ˜ w(t ) = B

T

1

ξ(t ),

ξ = 0 ∀t ∈ [0 h]

(46)

where ξ(t ) ∈ R

n

, ˜ z(t ) ∈ R

p

and ˜ w(t ) ∈ R

q

. Note that the latter systemrepresents the

backward adjoint of

1

(as deﬁned for nondescriptor case in [2, vol. 1]). Its forward

representation,

2

, is described by

E

˙

ξ(τ) =

2

i=0

A

T

i

ξ(τ − h

i

) + C

T

0

˜ z(τ), ˜ w(τ) = B

T

1

ξ(τ),

ξ = 0 ∀τ ∈ [−h 0].

(47)

Since the characteristic equations of

2

and

1

are identical, the former system is

asymptotically stable iff

1

is.

286 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

Sufﬁcient conditions of Theorem 3.5 for (14a), (14b), (44) and for its ‘adjoint’

may lead to different results. Therefore, one can apply Theorem 3.5 for the original

system and for its ‘adjoint’ and then choose the less conservative result.

Example 1. We consider the following system

_

˙ x

1

(t )

0

_

˙ x(t ) =

_

0.5 0

0 −1

_

x(t ) +

_

−1 1

0 0.5

_

x(t − h) +

_

1

0.5

_

w(t ),

z(t ) =

_

0.5 1

_

x(t ),

(48)

where x(t ) = col{x

1

(t ), x

2

(t )} ∈ R

2

. Applying the LMI condition of Theorem 3.5

to (48) and its ‘adjoint’ we obtain in both cases that the system is internally stable

for h 1.15. The minimum achievable value of γ is however different in the two

cases. For h = .1 we obtain for both systems γ = 2.3, while for h = 1 we obtain

for (48) γ

0

= 9 and for its ‘adjoint’ γ

t

= 6. For h = 1.12 the corresponding results

are γ

0

= 40 and γ

t

= 28, respectively. We see that in this example the conditions of

Theorem 3.5 for the ‘adjoint’ system are less conservative than those obtained for the

original system. Note that the same results are obtained by choosing block-diagonal

P

3

with P

32

> 0.

4. Delay-dependent state-feedback control

We apply the results of the previous section to the inﬁnite-horizon H

∞

-control

problem. Given system (4), (6) with the objective vector (5b). For a prescribed scalar

γ > 0, we consider the performance index of (13). We look for the state-feedback

gain matrix K which, via the control law

u(t ) = Kx(t ), K = [K

1

, K

2

] (49)

achieves J(w) < 0 for all nonzero w ∈ L

q

2

[0, ∞). Substituting (49) into (4), we

obtain the structure of (14a) and (14b) with A

0

+ B

2

K instead of A

0

and

C

T

0

C

0

=

¯

C

T

1

¯

C

1

+ K

T

D

T

12

D

12

K. (50)

Applying the BRL of Section 3 to the above matrices, results in a nonlinear matrix

inequality because of the terms P

T

2

B

2

K and P

T

3

B

2

K. We therefore consider another

version of the BRL which is derived from (32).

In order to obtain an LMI we have to restrict ourselves to the case of the diag-

onal matrix P

3

of (41) and (as well as in the nondescriptor problem) to the case

of W

i

= ε

i

P, i = 1, 2, where ε

i

∈ R is a scalar parameter. Note that for ε

i

= −1

(32) yields the delay-independent condition of Corollary 3.6. It is obvious from the

requirement of 0 < P

1

, and the fact that in (32) (P

3

A

T

04

+ A

04

P

T

3

) must be negative

definite, that P is nonsingular. Deﬁning

P

−1

= Q =

_

Q

1

0

Q

2

Q

3

_

, (51a)

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 287

Q

2

=

_

Q

21

Q

22

_

, (51b)

Q

3

= diag

_

Q

31

, Q

32

_

, Q

32

∈ R

n

2

×n

2

, (51c)

and = diag{Q, I

q+4n+2n

1

} we multiply (32) by

T

and , on the left and on the

right, respectively. Applying the Schur formula to the quadratic term in Q, we obtain

the following inequality:

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

1

+

2

_

0

B

1

_

h

1

(ε

1

+ 1)I

n+n

1

h

2

(ε

2

+ 1)I

n+n

1

ε

1

_

_

0

A

11

A

13

_

_

ε

2

_

_

0

A

21

A

23

_

_

_

_

0

A

12

A

14

_

_

_

_

0

A

22

A

24

_

_

Q

T

_

_

_

_

C

T

01

0

C

T

02

_

¸

¸

_

∗ −γ

2

I

q

0 0 0 0 0 0 0

∗ ∗ −h

1

R

1

0 0 0 0 0 0

∗ ∗ ∗ −h

2

R

2

0 0 0 0 0

∗ ∗ ∗ ∗ −S

1

0 0 0 0

∗ ∗ ∗ ∗ ∗ −S

2

0 0 0

∗ ∗ ∗ ∗ ∗ ∗ −U

1

0 0

∗ ∗ ∗ ∗ ∗ ∗ ∗ −U

2

0

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ −I

p

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

Q

T

_

I

n

1

0

_

Q

T

_

I

n

1

0

_

Q

T

_

0

I

n

2

_

Q

T

_

0

I

n

2

_

h

1

Q

T

_

_

_

_

0

n

1

×n

1

0 0

0 A

T

11

A

T

13

0 0 0

0 0 0

_

¸

¸

_

h

2

Q

T

_

_

_

_

0

n

1

×n

1

0 0

0 A

T

21

A

T

23

0 0 0

0 0 0

_

¸

¸

_

0 0 0 0 0 0

0 0 0 0 0 0

0 0 0 0 0 0

0 0 0 0 0 0

0 0 0 0 0 0

0 0 0 0 0 0

0 0 0 0 0 0

−S

−1

1

0 0 0 0 0

∗ −S

−1

2

0 0 0 0

∗ ∗ −U

−1

1

0 0 0

∗ ∗ ∗ −U

−1

2

0 0

∗ ∗ ∗ ∗ −h

1

R

−1

1

0

∗ ∗ ∗ ∗ ∗ −h

2

R

−1

2

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

< 0,

(52)

where C

0

= [C

01

C

02

],

1

=

_

_

_

0 I

n

1

0

2

i=0

A

i1

−I

n

1

A

02

2

i=0

A

i3

0 A

04

_

¸

_

Q + Q

T

_

_

_

0 I

n

1

0

2

i=0

A

i1

−I

n

1

A

02

2

i=0

A

i3

0 A

04

_

¸

_

T

+

2

i=1

ε

i

_

_

0 0 0

A

i1

0 0

A

i3

0 0

_

_

Q +

2

i=1

ε

i

Q

T

_

_

0 A

T

i1

A

T

i3

0 0 0

0 0 0

_

_

, (53)

288 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

2

=

_

0

B

2

_

_

K

1

0

n

1

K

2

_

Q + Q

T

_

_

_

K

T

1

0

n

1

K

T

2

_

¸

_

_

0 B

T

2

_

.

We substitute (50) into (52), denote

K

1

Q

1

+ K

2

Q

22

= Y

1

, K

2

Q

32

= Y

2

, (54)

and obtain the following:

Theorem 4.1. Consider the system of (4), (6), (5b) and the cost function of (13).

For a prescribed 0 < γ, the state-feedback law of (49) achieves, J(w) < 0 for

all nonzero w ∈ L

q

2

[0, ∞) if for some prescribed scalars ε

1

, ε

2

∈ R, there exist

0 < Q

1

∈ R

n

1

×n

1

, 0 <

¯

S

k

= S

−1

k

∈ R

n

1

×n

1

, 0 <

¯

U

k

= U

−1

k

∈ R

n

2

×n

2

, k =1, 2, Q

2

∈ R

n×n

1

and Q

3

∈ R

n×n

of (51a)–(51c), 0 <

¯

R

1

= R

−1

1

, 0 <

¯

R

2

= R

−1

2

∈

R

(n+n

1

)×(n+n

1

)

, Y

1

∈ R

×n

1

and Y

2

∈ R

×n

2

that satisfy the following LMI:

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

1

+

¯

_

0

B

1

_

h

1

(ε

1

+ 1)

¯

R

1

h

2

(ε

2

+ 1)

¯

R

2

ε

1

_

_

0

A

11

A

13

_

_ ¯

S

1

ε

2

_

_

0

A

21

A

23

_

_ ¯

S

2

_

_

0

A

12

A

14

_

_ ¯

U

1

_

_

0

A

22

A

24

_

_ ¯

U

2

Q

T

_

_

_

_

¯

C

T

11

0

¯

C

T

12

_

¸

¸

_

∗ −γ

2

I

q

0 0 0 0 0 0 0

∗ ∗ −h

1

¯

R

1

0 0 0 0 0 0

∗ ∗ ∗ −h

2

¯

R

2

0 0 0 0 0

∗ ∗ ∗ ∗ −

¯

S

1

0 0 0 0

∗ ∗ ∗ ∗ ∗ −

¯

S

2

0 0 0

∗ ∗ ∗ ∗ ∗ ∗ −

¯

U

1

0 0

∗ ∗ ∗ ∗ ∗ ∗ ∗ −

¯

U

2

0

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ −I

p

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

_

_

_

_

Y

T

1

0

Y

T

2

_

¸

¸

_

D

T

12

_

Q

1

0

_ _

Q

1

0

_

_

_

_

_

Q

T

22

0

Q

T

32

_

¸

¸

_

_

_

_

_

Q

T

22

0

Q

T

32

_

¸

¸

_

h

1

_

_

_

Q

T

21

Q

T

31

0

_

¸

_

_

0 A

T

11

A

T

13

_

h

2

_

_

_

Q

T

21

Q

T

31

0

_

¸

_

_

0 A

T

21

A

T

23

_

0 0 0 0 0 0 0

0 0 0 0 0 0 0

0 0 0 0 0 0 0

0 0 0 0 0 0 0

0 0 0 0 0 0 0

0 0 0 0 0 0 0

0 0 0 0 0 0 0

I

p

0 0 0 0 0 0

∗ −

¯

S

1

0 0 0 0 0

∗ ∗ −

¯

S

2

0 0 0 0

∗ ∗ ∗ −

¯

U

1

0 0 0

∗ ∗ ∗ ∗ −

¯

U

2

0 0

∗ ∗ ∗ ∗ ∗ −h

1

¯

R

1

0

∗ ∗ ∗ ∗ ∗ ∗ −h

2

¯

R

2

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

< 0,

(55)

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 289

where

¯

=

_

0

B

2

_

_

Y

1

0

n

1

Y

2

_

+

_

_

_

Y

T

1

0

n

1

Y

T

2

_

¸

_

_

0 B

T

2

_

.

The state-feedback gain is then given by

K

2

= Y

2

Q

−1

32

, K

1

= (Y

1

− K

2

Q

22

)Q

−1

1

. (56)

Example 2. We consider the system

E˙ x(t ) = A

1

x(t − h) + B

1

w(t ) + B

2

u(t ), x

0

= 0,

z(t ) =

¯

C

1

x(t ) + D

12

u(t ),

(57)

where

E =

_

1 0

0 0

_

, A

1

=

_

−1 0

1 −1

_

,

B

1

=

_

1

1

_

, B

2

=

_

−0.5

1

_

,

C

1

=

_

1 0.2

_

, D

12

=

_

0.1

_

.

Note that in this example A

04

= 0. We ﬁrst ﬁnd the state-feedback solution. We

obtained a near minimum value of γ = 21 for h = 1.2 and ε

1

= −0.255. The state-

feedback control law that achieves the later bound on the H

∞

-norm of the closed

loop is u = Kx, where K = [175.62 − 430680].

The LMI in Theorem 4.1 is afﬁne in the system matrices. It can thus be applied

also to the case where these matrices are uncertain and are known to reside within a

given polytope. Considering the system of (4) and denoting

=

_

E A

0

A

1

A

2

B

1

B

2

¯

C

1

D

12

_

,

we assume that ∈ Co{

j

, j = 1, . . . , N}, where the N vertices of the polytope

are described by

(j)

=

_

E A

(j)

0

A

(j)

1

A

(j)

2

B

(j)

1

B

(j)

2

¯

C

(j)

1

D

(j)

12

_

.

We obtain the following:

Theorem 4.2. Consider the system of (4), (6), (5b), where the system matrices

reside within the polytope and the cost function of (13). For a prescribed 0 < γ,

the state-feedback law of (49) achieves, J(w) < 0 for all nonzero w ∈ L

q

2

[0, ∞)

290 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

and for all the matrices in if for some prescribed scalars ε

1

, ε

2

∈ R there

exist 0 < Q

1

∈ R

n

1

×n

1

, 0 <

¯

S

k

= S

−1

k

∈ R

n

1

×n

1

, 0 <

¯

U

k

= U

−1

k

∈ R

n

2

×n

2

, k = 1,

2, Q

2

∈ R

n×n

1

and Q

3

∈ R

n×n

of (51a)–(51c), 0 <

¯

R

1

= R

−1

1

, 0 <

¯

R

2

= R

−1

2

∈

R

(n+n

1

)×(n+n

1

)

, Y

1

∈ R

×n

1

and Y

2

∈ R

×n

2

that satisfy LMIs (55) for j = 1, . . . , N,

where the matrices

A

i

, i = 0, 1, 2, B

1

, B

2

,

¯

C

1

, D

12

,

¯

R

1

,

¯

R

2

are taken with the upper index j. The state-feedback gain is then given by (56).

5. Delay-dependent ﬁltering

We consider system (4) with the measurement law of (5a). We seek a ﬁlter of the

following observer form:

E

˙

ˆ x(t ) =

2

i=0

A

i

ˆ x(t − h

i

) + K

f

( ¯ y(t ) −

¯

C

2

x(t )) (58)

such that the H

∞

-norm of the resulting transference between the exogenous signal

w and the estimation error z is less than a prescribed value γ , where

z(t ) L(x(t ) − ˆ x(t )). (59)

From (4), (5a) and (58) it follows that the estimation error e(t ) = x(t ) − ˆ x(t ) is

described by the following model:

E˙ e(t ) = (A

0

− K

f

¯

C

2

)e(t ) +

2

i=1

A

i

¯ e(t − h

i

) + (B

1

− K

f

D

21

)w,

z(t ) = Le(t ).

(60)

The problem then becomes one of ﬁnding the ﬁlter gain K

f

such that J(w) < 0.

We consider the ‘adjoint’ to (60) system described by

E

˙

ξ(τ) =

_

A

T

0

−

¯

C

T

2

K

T

f

_

ξ(τ) +

2

i=1

A

T

i

ξ(τ − h

i

) + L

T

˜ z(τ),

˜ w(τ) =

_

B

T

1

− D

T

21

K

T

f

_

ξ(τ), ξ = 0 ∀τ ∈ [−h 0].

(61)

Analogously to Theorem 4.1 (by applying BRL of Theorem 3.5 to (61)) we obtain:

Theorem 5.1. Consider the system of (4), (5a) and the cost function J(w). For a

prescribed 0 < γ, the ﬁlter gain achieves, J(w) < 0 for all nonzero w ∈ L

q

2

[0, ∞)

if for some prescribed scalars ε

1

, ε

2

∈ R, there exist 0 < Q

1

∈ R

n

1

×n

1

, 0 <

¯

S

k

∈

R

n

1

×n

1

, 0 <

¯

U

k

∈ R

n

2

×n

2

, k = 1, 2, Q

2

∈ R

n×n

1

and Q

3

∈ R

n×n

of (51a)–(51c),

0 <

¯

R

1

, 0 <

¯

R

2

∈ R

(n+n

1

)×(n+n

1

)

, Y

1

∈ R

r×n

1

and Y

2

∈ R

r×n

2

that satisfy the fol-

lowing LMI:

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 291

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

1

+

¯

_

0

L

T

_

h

1

(ε

1

+ 1)

¯

R

1

h

2

(ε

2

+ 1)

¯

R

2

ε

1

_

_

_

_

0

A

T

11

A

T

12

_

¸

¸

_

¯

S

1

ε

2

_

_

_

_

0

A

T

21

A

T

22

_

¸

¸

_

¯

S

2

_

_

_

_

0

A

T

13

A

T

14

_

¸

¸

_

¯

U

1

∗ −γ

2

I

r

0 0 0 0 0

∗ ∗ −h

1

¯

R

1

0 0 0 0

∗ ∗ ∗ −h

2

¯

R

2

0 0 0

∗ ∗ ∗ ∗ −

¯

S

1

0 0

∗ ∗ ∗ ∗ ∗ −

¯

S

2

0

∗ ∗ ∗ ∗ ∗ ∗ −

¯

U

1

∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗

_

_

_

_

0

A

T

23

A

T

24

_

¸

¸

_

¯

U

2

Q

T

_

_

_

B

11

0

B

12

_

¸

_ −

_

_

_

_

Y

T

1

0

Y

T

2

_

¸

¸

_

D

21

_

Q

1

0

_ _

Q

1

0

_

_

_

_

_

Q

T

22

0

Q

T

32

_

¸

¸

_

_

_

_

Q

T

22

0

Q

T

32

_

¸

_ h

1

A

1

h

2

A

2

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

¯

U

2

0 0 0 0 0 0 0

∗ −I

q

0 0 0 0 0 0

∗ ∗ −

¯

S

1

0 0 0 0 0

∗ ∗ ∗ −

¯

S

2

0 0 0 0

∗ ∗ ∗ ∗ −

¯

U

1

0 0 0

∗ ∗ ∗ ∗ ∗ −

¯

U

2

0 0

∗ ∗ ∗ ∗ ∗ ∗ −h

1

¯

R

1

0

∗ ∗ ∗ ∗ ∗ ∗ ∗ −h

2

¯

R

2

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

< 0,

(62)

where

1

=

_

_

_

_

0 I

n

1

0

2

i=0

A

T

i1

−I

n

1

A

T

03

2

i=0

A

T

i2

0 A

T

04

_

¸

¸

_

Q + Q

T

_

_

_

_

0 I

n

1

0

2

i=0

A

T

i1

−I

n

1

A

T

03

2

i=0

A

T

i2

0 A

T

04

_

¸

¸

_

T

+

2

i=1

ε

i

_

_

_

0 0 0

A

T

i1

0 0

A

T

i2

0 0

_

¸

_

Q +

2

i=1

ε

i

Q

T

_

_

0 A

i1

A

i2

0 0 0

0 0 0

_

_

292 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

¯

= −

_

0

¯

C

T

2

_

_

Y

1

0

n

1

Y

2

_

−

_

_

_

Y

T

1

0

n

1

Y

T

2

_

¸

_

_

0

¯

C

2

_

,

A

i

=

_

_

_

Q

T

21

Q

T

31

0

_

¸

_

_

0 A

i1

A

i2

_

, i = 1, 2.

The ﬁlter gain is then given by

K

T

2

= Y

2

Q

−1

32

, K

T

1

= (Y

1

− K

T

2

Q

22

)Q

−1

1

, K

f

= col

_

K

1

, K

2

_

. (63)

The LMI in Theorem 5.1 is afﬁne in the system matrices. Similarly to Theorem

4.2, it can thus be reformulated also to the case of matrices with polytopic uncertain-

ties.

6. Delay-dependent output-feedback control

We adopt in this section the dissipation approach to the solution of the output-

feedback problem. It applies a controller of a state-feedback—observer structure and

requires a solution of two LMIs. We assume:

A2. The matrices B

1

and D

21

are orthogonal, i.e. B

1

D

T

21

= 0, and

˜

R = D

T

12

D

12

is

not singular.

6.1. The ﬁrst phase: a state-feedback controller design

Lemma 6.1. Assume A2. Consider system (4), (5b). For a prescribed γ > 0, the

feedback law

u(t ) = −

_

0

˜

R

−1

B

T

2

_

_

P

1

0

P

2

P

3

_

_

_

x

1

(t )

y(t )

x

2

(t )

_

_

, (64a)

_

P

1

0

P

2

P

3

_

=

_

Q

1

0

Q

2

Q

3

_

−1

(64b)

achieves J(w) < 0 for all nonzero w ∈ L

q

2

[0, ∞) if for some prescribed scalars

ε

1

, ε

2

∈ R, there exist 0 < Q

1

∈ R

n

1

×n

1

, 0 <

¯

S

k

= S

−1

k

∈ R

n

1

×n

1

, 0 <

¯

U

k

= U

−1

k

∈ R

n

2

×n

2

, k = 1, 2, Q

2

∈ R

n×n

1

and Q

3

∈ R

n×n

of (51a)–(51c), 0 <

¯

R

1

= R

−1

1

and 0 <

¯

R

2

= R

−1

2

∈ R

(n+n

1

)×(n+n

1

)

that satisfy the following LMI:

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 293

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

1

+

¯

_

0

B

1

_

h

1

(ε

1

+ 1)

¯

R

1

h

2

(ε

2

+ 1)

¯

R

2

ε

1

_

_

0

A

11

A

13

_

_ ¯

S

1

ε

2

_

_

0

A

21

A

23

_

_ ¯

S

2

_

_

0

A

12

A

14

_

_ ¯

U

1

_

_

0

A

22

A

24

_

_ ¯

U

2

Q

T

_

_

¯

C

T

11

0

¯

C

T

12

_

_

∗ −γ

2

I

q

0 0 0 0 0 0 0

∗ ∗ −h

1

¯

R

1

0 0 0 0 0 0

∗ ∗ ∗ −h

2

¯

R

2

0 0 0 0 0

∗ ∗ ∗ ∗ −

¯

S

1

0 0 0 0

∗ ∗ ∗ ∗ ∗ −

¯

S

2

0 0 0

∗ ∗ ∗ ∗ ∗ ∗ −

¯

U

1

0 0

∗ ∗ ∗ ∗ ∗ ∗ ∗ −

¯

U

2

0

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ −I

p

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

_

Q

1

0

_ _

Q

1

0

_

_

_

Q

T

22

0

Q

T

32

_

_

_

_

Q

T

22

0

Q

T

32

_

_

h

1

_

_

_

Q

T

21

Q

T

31

0

_

¸

_

_

0 A

T

11

A

T

13

_

h

2

_

_

_

Q

T

21

Q

T

31

0

_

¸

_

_

0 A

T

21

A

T

23

_

0 0 0 0 0 0

0 0 0 0 0 0

0 0 0 0 0 0

0 0 0 0 0 0

0 0 0 0 0 0

0 0 0 0 0 0

0 0 0 0 0 0

0 0 0 0 0 0

¯

S

1

0 0 0 0 0

∗ −

¯

S

2

0 0 0 0

∗ ∗ −

¯

U

1

0 0 0

∗ ∗ ∗ −

¯

U

2

0 0

∗ ∗ ∗ ∗ −h

1

¯

R

1

0

∗ ∗ ∗ ∗ ∗ −h

2

¯

R

2

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

< 0,

(65)

where

1

is given by (53) and where

ˆ

= −

_

0

B

2

_

˜

R

−1

_

0 B

T

2

_

.

Proof. The proof readily follows by choosing V as in (31) and applying (36) for

R

i

M

i

= ε

i

I, i = 1, 2. Denote by

u

the matrix in the left-hand side of (65). We

obtain by integrating dV(t )/dt that

J

_

∞

0

¯

ξ

T

p

¯

ξ dt +

_

∞

0

(u

T

− u

∗T

)

˜

R(u − u

∗

) dt

−γ

2

_

∞

0

(w

T

− w

∗T

)(w − w

∗

) dt, (66)

294 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

where

p

= diag

_

P

T

, I

8n+4n

1

+p+q

_

u

diag

_

P, I

8n+4n

1

+p+q

_

,

w

∗

= γ

−2

_

0 B

T

1

_

P

_

_

x

1

y

x

2

_

_

,

u

∗

= −

˜

R

−1

_

0 B

T

2

_

P

_

_

x

1

y

x

2

_

_

(67)

and where the relation between P and Q

i

, i = 1, . . . , 3, is given in (51a)–(51c) and

¯

ξ = col{x

1

, y, x

2

, w, η

1

}, with η

1

representing the ﬁctitious states that emerge when

applying Schur formula to construct

p

.

Unfortunately, the feedback law of (64a) and (64b) cannot be implemented even

when there exists a solution to (65), namely when the ﬁrst term in the right-hand side

of (66) is negative for all

¯

ξ ∈ R

8n+4n

1

+p+q

.

6.2. The second phase: ﬁltering via the adjoint system

Denoting ¯ r = w − w

∗

we represent (4) and (5a) in the form:

_

_

˙ x

1

(t )

0

0

_

_

=

2

i=0

ˆ

A

i

_

_

x

1

(t − h

i

)

y(t − h

i

)

x

2

(t − h

i

)

_

_

+

ˆ

B

1

¯ r(t ) +

_

_

B

21

B

21

B

22

_

_

u(t ),

¯ y(t ) =

ˆ

C

2

_

_

x

1

(t )

y(t )

x

2

(t )

_

_

+ D

21

¯ r(t ),

(68)

where

ˆ

A

0

=

_

_

A

01

0 A

02

A

01

−I

n

1

A

02

A

03

0 A

04

_

_

+ γ

−2

ˆ

B

1

_

0

q×n

1

B

T

1

_

P,

ˆ

A

i

=

_

_

A

i1

0 A

i2

A

i1

0 A

i2

A

i3

0 A

i4

_

_

, i = 1, 2,

ˆ

B

1

=

_

_

B

11

B

11

B

12

_

_

,

ˆ

C

2

=

_

¯

C

21

0

r×n

1

¯

C

22

_

.

(69)

and where P solves (65). The objective function of (13) and (66) will then be negative

if there exist ˆ x(t ) and ˆ y(t ) in R

n

that satisfy

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 295

J

a

=

_

∞

0

(¯ z

T

¯ z − γ

2

¯ r

T

¯ r) dt < 0 ∀¯ r ∈ L

q

2

[0, ∞),

¯ z(t ) =

˜

R

−1/2

_

0 B

T

2

_

P

_

_

x

1

(t ) − ˆ x

1

(t )

y(t ) − ˆ y(t )

x

2

(t ) − ˆ x

2

(t )

_

_

.

(70)

The problem of ﬁnding ˆ x(t ) and ˆ y(t ) is, in fact, a H

∞

ﬁltering problem for the

descriptor system (68).

Consider the following ‘innovation’ ﬁlter

_

_

˙

ˆ x

1

(t )

0

0

_

_

=

2

i=0

ˆ

A

i

_

_

ˆ x

1

(t − h

i

)

ˆ y(t − h

i

)

ˆ x

2

(t − h

i

)

_

_

+ K

f

¯ y(t )

−K

f

ˆ

C

2

_

_

ˆ x

1

(t )

ˆ y(t )

ˆ x

2

(t )

_

_

+

_

_

B

21

B

21

B

22

_

_

u(t ). (71)

Denoting

e =

_

_

e

1

e

0

e

2

_

_

=

_

_

x

1

y

x

2

_

_

−

_

_

ˆ x

1

ˆ y

ˆ x

2

_

_

(72)

and using the assumption on D

21

and the definition of w

∗

in (67) we ﬁnd

_

_

˙ e

1

(t )

0

0

_

_

= (

ˆ

A

0

− K

f

ˆ

C

2

)e(t ) +

2

i=1

ˆ

A

i

e(t − h

i

) + (

ˆ

B

1

− K

f

D

21

)¯ r(t ),

¯ z(t ) =

˜

R

−1/2

_

0 B

T

2

_

Pe(t ).

(73)

The problem now becomes one of ﬁnding the gain matrix K

f

that will ensure the

stability of system (73) and that the H

∞

-norm of the transference from ¯ r to ¯ z is less

than γ . This problem was solved in Section 5. By applying Theorem 5.1 we obtain

the following result:

Theorem 6.2. Assume A2. Consider the system of (4), (5a), (5b) and the cost

function of (13). For a prescribed 0 < γ, there exists an output-feedback controller

that achieves, J(w) < 0 for all nonzero w ∈ L

q

2

[0, ∞) if for some prescribed sca-

lars ε

1

, ε

2

∈ R, there exist 0 < Q

1

∈ R

n

1

×n

1

, 0 <

¯

S

k

∈ R

n

1

×n

1

, 0 <

¯

U

k

∈ R

n

2

×n

2

,

k = 1, 2, Q

2

∈ R

n×n

1

and Q

3

∈ R

n×n

of (51a)–(51c), 0 <

¯

R

1

and 0 <

¯

R

2

∈

R

(n+n

1

)×(n+n

1

)

that satisfy (65) and for some prescribed scalars ˆ ε

1

, ˆ ε

2

∈ R, there ex-

ist 0 <

ˆ

Q

1

∈ R

n

1

×n

1

, 0 <

ˆ

S

k

∈ R

n

1

×n

1

, 0 <

ˆ

U

k

∈ R

n×n

, k = 1, 2,

ˆ

Q

2

∈ R

(n+n

1

)×n

1

and

ˆ

Q

3

∈ R

(n+n

1

)×(n+n

1

)

of the form

296 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

ˆ

Q

2

=

_

ˆ

Q

21

ˆ

Q

22

_

,

ˆ

Q

22

∈ R

n×n

1

,

ˆ

Q

3

= diag

_

ˆ

Q

31

,

ˆ

Q

32

_

,

ˆ

Q

31

∈ R

n

1

×n

1

,

ˆ

Q

32

∈ R

n×n

,

0 <

ˆ

R

1

, 0 <

ˆ

R

2

∈ R

(2n

1

+n)×(2n

1

+n)

, Y

1

∈ R

r×n

1

, and Y

2

∈ R

r×n

that satisfy the following LMI:

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

_

1

+

¯

_

_

0

n

1

×r

P

_

0

B

2

_

˜

R

−1/2

_

_

h

1

(ˆ ε

1

+ 1)

ˆ

R

1

h

2

(ˆ ε

2

+ 1)

ˆ

R

2

ˆ ε

1

_

_

_

_

0

ˆ

A

T

11

ˆ

A

T

12

_

¸

¸

_

ˆ

S

1

ˆ ε

2

_

_

_

_

0

ˆ

A

T

21

ˆ

A

T

22

_

¸

¸

_

ˆ

S

2

_

_

_

_

0

ˆ

A

T

13

ˆ

A

T

14

_

¸

¸

_

ˆ

U

1

_

_

_

_

0

ˆ

A

T

23

ˆ

A

T

24

_

¸

¸

_

ˆ

U

2

∗ −γ

2

I

r

0 0 0 0 0 0

∗ ∗ −h

1

ˆ

R

1

0 0 0 0 0

∗ ∗ ∗ −h

2

ˆ

R

2

0 0 0 0

∗ ∗ ∗ ∗ −

ˆ

S

1

0 0 0

∗ ∗ ∗ ∗ ∗ −

ˆ

S

2

0 0

∗ ∗ ∗ ∗ ∗ ∗ −

ˆ

U

1

0

∗ ∗ ∗ ∗ ∗ ∗ ∗ −

ˆ

U

2

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

ˆ

Q

T

_

_

0

B

11

B

1

_

_

_

_

_

_

Y

T

1

0

Y

T

2

_

¸

¸

_

D

21

_

ˆ

Q

1

0

_ _

ˆ

Q

1

0

_

_

_

_

_

ˆ

Q

T

22

0

ˆ

Q

T

32

_

¸

¸

_

_

_

_

_

ˆ

Q

T

22

0

ˆ

Q

T

32

_

¸

¸

_

h

1

ˆ

A

1

h

2

ˆ

A

2

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0

I

q

0 0 0 0 0 0 0

∗ −I

q

0 0 0 0 0 0

∗ ∗ −

ˆ

S

1

0 0 0 0 0

∗ ∗ ∗ −

ˆ

S

2

0 0 0 0

∗ ∗ ∗ ∗ −

ˆ

U

1

0 0 0

∗ ∗ ∗ ∗ ∗ −

ˆ

U

2

0 0

∗ ∗ ∗ ∗ ∗ ∗ −h

1

ˆ

R

1

0

∗ ∗ ∗ ∗ ∗ ∗ ∗ −h

2

ˆ

R

2

_

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

_

< 0,

(74)

where

P

1

= Q

−1

1

, P

2

= −Q

−1

3

Q

2

Q

−1

1

,

P

3

= Q

−1

3

= diag

_

Q

−1

31

, Q

−1

32

_

= diag

_

P

31

, P

32

_

,

P =

_

P

1

0

P

2

P

3

_

,

(75)

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 297

1

=

_

_

_

0 I 0

2

i=0

ˆ

A

T

i1

−I

n

1

ˆ

A

T

03

2

i=0

ˆ

A

T

i2

0

ˆ

A

T

04

_

¸

_

ˆ

Q +

ˆ

Q

T

_

_

_

0 I 0

2

i=0

ˆ

A

T

i1

−I

n

1

ˆ

A

T

03

2

i=0

ˆ

A

T

i2

0

ˆ

A

T

04

_

¸

_

T

+

2

i=1

ˆ ε

i

_

_

_

0 0 0

ˆ

A

T

i1

0 0

ˆ

A

T

i2

0 0

_

¸

_

ˆ

Q +

2

i=1

ˆ ε

i

ˆ

Q

T

_

_

_

0

ˆ

A

i1

ˆ

A

i2

0 0 0

0 0 0

_

¸

_

,

¯

= −

_

0

ˆ

C

T

2

_

_

Y

1

0

n

1

Y

2

_

−

_

_

_

Y

T

1

0

n

1

Y

T

2

_

¸

_

_

0

ˆ

C

2

_

,

ˆ

A

i1

= A

i1

,

ˆ

A

i2

=

_

0 A

i2

_

,

ˆ

A

i3

=

_

A

i1

A

i3

_

,

ˆ

A

i4

=

_

0 A

i2

0 A

i4

_

, i =1, 2,

ˆ

A

01

= A

01

+ γ

−2

B

11

B

T

1

P

2

,

ˆ

A

02

=

_

0 A

02

_

+ γ

−2

_

B

11

B

T

11

P

31

B

11

B

T

12

P

32

_

,

ˆ

A

03

=

_

A

01

A

03

_

+ γ

−2

_

B

11

B

T

1

P

2

B

12

B

T

1

P

2

_

,

ˆ

A

04

=

_

−I

n

1

A

02

0 A

04

_

+ γ

−2

_

B

11

B

T

11

P

31

B

11

B

T

12

P

32

B

12

B

T

11

P

31

B

12

B

T

12

P

32

_

,

ˆ

A

i

=

_

_

_

ˆ

Q

T

21

ˆ

Q

T

31

0

_

¸

_

_

0 A

i1

0 A

i2

_

, i = 1, 2.

The ﬁlter gain is then given by

K

T

2

= Y

2

ˆ

Q

−1

32

, K

T

1

= (Y

1

− K

T

2

ˆ

Q

22

)

ˆ

Q

−1

1

, K

f

= col

_

K

1

, K

2

_

. (76)

If a solution to (65) and (74) exists, then the output-feedback controller is ob-

tained by:

u(t ) = −

˜

R

−1

B

T

2

_

P

2

ˆ x

1

(t ) + P

3

col{ ˆ y(t ) ˆ x

2

(t )}

_

, (77)

where ˆ x and ˆ y are obtained by (71).

Example 3. We consider the following system:

E˙ x(t ) =

1

i=0

A

1

x(t − h

i

) + B

1

w(t ) + B

2

u(t ), x

0

= 0,

z(t ) =

¯

C

1

x(t ) + D

12

u(t ),

y(t ) =

¯

C

2

x(t ) + D

21

w(t ),

(78)

298 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

where

E =

_

1 0

0 0

_

, A

0

=

_

0 0

0 2

_

, A

1

=

_

−1 0

1 −1

_

,

B

1

=

_

1 0

1 0

_

, B

2

=

_

0

1

_

, C

1

=

_

1 0.2

_

, C

2

=

_

1 0

_

,

D

12

= 0.1, D

21

=

_

0 0.1

_

.

We ﬁrst ﬁnd the state-feedback solution. We obtained a near minimum value of

γ = 11 for h

1

= 1.2 and ε

1

= −0.3. The state-feedback control law that achieves

the later bound on the H

∞

-norm of the closed loop is u = Kx, where K = [42.6854

−0.3426]. Using (65) we ﬁnd that the feedback law u

∗

does not apply y because

B

21

= 0 and P

3

is diagonal. We thus obtained the same near minimum value of

γ = 11 for h = 1.2 and ε

1

= −0.3. The feedback control law that achieves the later

bound on the H

∞

-norm of the closed loop is

u

∗

= 9.72x

1

+ 3.22x

2

.

The output-feedback control is derived for the same values of h, ε

1

and ˆ ε

1

= −1. A

minimum value of γ = 2.4 is obtained. The resulting output-feedback has the form

9.72ˆ x

1

+ 3.22ˆ x

2

, where ˆ x is obtained by (71) with

K

f

=

_

6.1814 1.1882 −3.1710

_

T

.

6.3. The second phase: direct ﬁltering

The ﬁltering of the previous section suffers from an additional overdesign that

stems from the use of the adjoint system which must be stable independently of the

delays in the variable e

0

. The advantage of the approach of Section 6.2 in comparison

with [12], where nondescriptor systems were considered, lies in the fact that it applies

the efﬁcient bounds introduced by Park [25]. For smaller values of h the method of

[12] may lead to less conservative results (see Example 4 below). In the present sec-

tion, we generalize the method of [12] to the case of descriptor systems with delay.

Denoting ¯ r = w − w

∗

we represent (4) in the form:

_

_

˙ x

1

(t )

0

0

_

_

=

ˆ

A

_

_

x

1

(t )

y(t )

x

2

(t )

_

_

−

2

i=1

_

_

0

A

i1

A

i3

_

_

_

t

t −h

i

y(s) ds

+

_

0

B

1

_

¯ r(t ) +

_

0

B

2

_

u(t ), (79)

where

ˆ

A =

_

_

_

0 I 0

2

i=0

A

i1

−I A

02

2

i=0

A

i3

0 A

04

_

¸

_

+ γ

−2

ˆ

B

1

ˆ

B

T

1

P,

ˆ

B

1

=

_

0

B

1

_

,

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 299

ˆ

C

2

=

_

¯

C

21

0

¯

C

22

_

and where P solves (65). The objective function of (13) and (66) will then be negative

if there exist ˆ x(t ) and ˆ y(t ) in R

n

that satisfy (70).

Consider the following ‘innovation’ ﬁlter

_

_

˙

ˆ x

1

(t )

0

0

_

_

=

ˆ

A

_

_

ˆ x

1

(t )

ˆ y(t )

ˆ x

2

(t )

_

_

−

2

i=1

_

_

0

A

i1

A

i3

_

_

_

t

t −h

i

ˆ y(s) ds

+K

f

[ ¯ y(t ) −

ˆ

C

2

ˆ x(t )] +

_

0

B

2

_

u(t ). (80)

Using the notation of (72), assumption A2 and the definition of w

∗

in (67) we ﬁnd

_

_

˙ e

1

(t )

0

0

_

_

= [

ˆ

A − K

f

ˆ

C

2

]e(t ) −

2

i=1

_

_

0

A

i1

A

i3

_

_

_

t

t −h

i

ˆ e

0

(s) ds

+(

ˆ

B

1

− K

f

D

21

)¯ r(t ),

¯ z(t ) =

˜

R

−1/2

_

0 B

T

2

_

Pe(t ).

(81)

The problem now becomes one of ﬁnding the gain matrix K

f

that will ensure the

stability of system (81) and that the H

∞

-norm of the transference from ¯ r to ¯ z is less

than γ . Similarly to [12] we obtain the following result:

Theorem 6.3. Consider the system of (4), (5a), (5b) and the cost function of (13).

For a prescribed 0 < γ, there exists an output-feedback controller that achieves,

J(w) < 0 for all nonzero w ∈ L

q

2

[0, ∞) if for some prescribed scalars ε

1

, ε

2

∈ R,

there exist 0 < Q

1

∈ R

n

1

×n

1

, 0 <

¯

S

k

∈ R

n

1

×n

1

, 0 <

¯

U

k

∈ R

n

2

×n

2

, k = 1, 2, Q

2

∈

R

n×n

1

and Q

3

∈ R

n×n

of (51a)–(51c), 0 <

¯

R

1

and 0 <

¯

R

2

∈ R

(n+n

1

)×(n+n

1

)

that

satisfy (65) and there exist 0 <

ˆ

P

1

∈ R

n

1

×n

1

, 0 <

ˆ

U

k

∈ R

n

2

×n

2

, k = 1, 2,

ˆ

P

2

∈

R

n×n

1

and

ˆ

P

3

∈ R

n×n

of the form

ˆ

P

2

=

_

ˆ

P

21

ˆ

P

22

_

,

ˆ

P

22

∈ R

n

2

×n

1

,

ˆ

P

3

= diag

_

ˆ

P

31

,

ˆ

P

32

_

,

ˆ

P

31

∈ R

n

1

×n

1

,

ˆ

P

32

∈ R

n

2

×n

2

,

0 <

ˆ

R

1

, 0 <

ˆ

R

2

∈ R

n

1

×n

1

, Y ∈ R

(n+n

1

)×r

that satisfy the following LMI:

300 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

_

_

_

_

_

_

_

_

_

¯

1

ˆ

P

T

_

0

B

1

_

− YD

21

h

1

ˆ

P

T

_

_

0

A

11

A

13

_

_

h

2

ˆ

P

T

_

_

0

A

11

A

13

_

_

_

P

T

2

P

T

3

_

B

2

ˆ

P

T

_

_

0

A

12

A

14

_

_ ˆ

P

T

_

_

0

A

22

A

24

_

_

∗ −γ

2

I

q

0 0 0 0 0

∗ ∗ −h

1

ˆ

R

1

0 0 0 0

∗ ∗ 0 −h

2

ˆ

R

2

0 0 0

∗ ∗ ∗ ∗ −

˜

R 0 0

∗ ∗ ∗ ∗ ∗ −

ˆ

U

1

0

∗ ∗ ∗ ∗ ∗ ∗ −

ˆ

U

2

_

¸

¸

¸

¸

¸

¸

¸

_

< 0,

(82)

where P, P

1

, P

2

, P

3

are deﬁned by (75) and

¯

1

=

ˆ

P

T

ˆ

A +

ˆ

A

T

ˆ

P − Y

ˆ

C

2

−

ˆ

C

T

2

Y

T

+

_

_

0

n

1

×n

1

0 0

0

2

i=1

h

i

ˆ

R

i

0

0 0

2

i=1

ˆ

U

i

_

_

.

If a solution to (65) and (82) exists, then the output-feedback controller is obtained

by (77), where ˆ x and ˆ y are obtained by (80) with K

f

=

ˆ

P

−T

Y.

Example 4. We consider the nondescriptor system of [12]

˙ x(t ) = A

0

x(t ) + A

1

x(t − h) + B

1

w(t ) + B

2

u(t ),

z(t ) = col

_

C

1

x(t ), D

12

u(t )

_

,

¯ y(t ) = C

2

x(t ) + D

21

w(t ),

(83)

where

A

0

=

_

0 0

0 1

_

, A

1

=

_

−1 −1

0 −.9

_

, B

1

=

_

1 0

1 0

_

, B

2

=

_

0

1

_

,

C

1

=

_

0 1

_

, D

12

= 0.1, C

2

=

_

0 1

_

, D

21

=

_

0 0.1

_

.

(84)

We compare the output-feedback controller designs achieved by the two methods

presented in Sections 6.2 and 6.3. By the counterpart of Theorem 6.3 for the non-

descriptor case, the output-feedback control was derived in [12] for h = 0.999 and

ε

1

= −0.29. A minimum value of γ = 0.86 was obtained. By Theorem 6.2 a greater

value of γ = 11 is found for the same value of h and ε

1

= ε

2

= −0.3.

For h 1 the LMI of Theorem 6.3 is not feasible for any γ > 0. By Theorem 6.2,

the output-feedback control is obtained for h = 1.28, ε

1

= ˆ ε

1

= −.3. A minimum

value of γ = 20 is then achieved.

This example shows that for greater values of h Theorem 6.2 is less conservative

due to Park’s inequality [25] that is used for bounding the cross terms. For smaller

values of h Theorem 6.2 leads to more conservative results owing to the fact that the

adjoint systemfor e contains e

0

with delays and that the results are delay-independent

with respect to delays in e

0

.

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 301

7. Conclusions

An LMI solution is proposed for the problem of stability and H

∞

-control of linear

time-invariant descriptor systems. This solution is based on the Lyapunov function

approach to descriptor systems with delay introduced in [10] and on the LMI ap-

proach to H

∞

control of nondescriptor systems of [12]. The LMI sufﬁcient condi-

tions that are obtained allow solutions to the H

∞

-control problem in the uncertain

case where the system parameters lie within an uncertainty polytope. As a byprod-

uct, new LMI conditions for stability and H

∞

-control of difference continuous time

equations are obtained.

The design of the output-feedback controller is achieved by two methods: one is

based on the BRL for the adjoint of the system that describes the estimation error;

the other applies the BRL directly to the system of the estimation error (and thus

generalizes the result of [12] to the descriptor case). Both methods suffer from an

additional overdesign that stems from the need to estimate the state and its derivative.

These methods lead to complementary results: for greater values of the delay the ﬁrst

method is less conservative, while for smaller values of the delay—the second one

provides less conservative results. In the special case where a result is sought which

is delay-independent with respect to the process and delay-dependent with respect

to observer, the latter overdesign can be removed since the estimate of the state (and

not of its derivative) is needed.

One question that often arises when solving control and estimation problems for

systems with time-delay is whether the solution obtained for certain delays h

i

will

satisfy the design requirements for all delays

¯

h

i

h

i

. In the problems of state-feed-

back and ﬁltering the answer is the afﬁrmative since the LMIs in Theorems 4.1 and

5.1 are convex in the time delays. The situation in the output-feedback control case

is however different, in spite of the seemingly convexity of the LMI of Theorems

6.2 and 6.3 in the delay parameters. The fact that the P

2

and P

3

depend nonlinearly

on the delay implies that the output-feedback controller that is derived for a certain

delay will not necessarily satisfy the design speciﬁcations for smaller delays.

References

[1] T. Azuma, K. Ikeda, K. Uchida, Inﬁnite-dimensional LMI approach to H

∞

-control synthesis for

linear systems with time delay, in: Proceedings of the 5th European Control Conference, Karlsruhe,

Germany, September 1999.

[2] A. Bensoussan, G. DoPrato, M.C. Delfour, S.K. Mitter, Representation and control of inﬁnite di-

mensional systems, vol. 1–2, Birkhäser, Basel, 1992.

[3] S. Boyd, L. El Ghaoui, E. Feron, V. Balakrishnan, Linear matrix inequality in systems and control

theory, SIAM Frontier Series (April 1994).

[4] R. Brayton, Small signal stability criterion for electrical networks containing lossless transmission

lines, IBM J. Res. Develop. 12 (1968) 431–440.

[5] A. Bunse-Gerstner, R. Byers, V. Mehrmann, N. Nickols, Feedback design for regularizing descriptor

systems, Linear Algebra Appl. 299 (1999) 119–151.

302 E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

[6] S. Campbell, Singular linear systems of differential equations with delays, Appl. Anal. 2 (1980)

129–136.

[7] S. Campbell, 2-D (differential-delay) implicit systems, in: Proceedings of IMACS World Congress

on Scientific Computation, Dublin, 1991, pp. 1828–1829.

[8] L. Dai, Singular Control Systems, Springer, Berlin, 1989.

[9] E. Fridman, New Lyapunov–Krasovskii functionals for stability of linear retarded and neutral type

systems, Systems Control Lett. 43 (4) (2001) 309–319.

[10] E. Fridman, Stability of linear descriptor systems with delay: a Lyapunov-based approach, J. Math.

Anal. Appl. (under review).

[11] E. Fridman, U. Shaked, H

∞

-state-feedback control of linear systems with small state-delay, Systems

Control Lett. 33 (1998) 141–150.

[12] E. Fridman, U. Shaked, A descriptor system approach to H

∞

control of linear time-delay systems,

IEEE Trans. Automat. Control 47 (2002).

[13] A. Halanay, V. Rasvan, Stability radii for some propagation models, IMA J. Math. Control Inform.

14 (1997) 95–107.

[14] J. Hale, S. Lunel, Introduction to Functional Differential Equations, Springer, New York, 1993.

[15] J. Hale, P. Martinez Amores, Stability in neutral equations, J. Nonlinear Anal. Theory, Meth. Appl.

1 (1977) 161–172.

[16] D. Ivanescu, J-M. Dion, L. Dugard, S.I. Niculescu, Dynamical compensation for time-delay systems:

an LMI approach, Internat. J. Robust Nonlinear Control 10 (2000) 611–628.

[17] V. Kolmanovskii, A. Myshkis, Applied Theory of Functional Differential Equations, Kluwer Aca-

demic Publishers, Dordrecht, 1999.

[18] V. Kolmanovskii, S.I. Niculescu, J.P. Richard, On the Liapunov–Krasovskii functionals for stability

analysis of linear delay systems, Internat. J. Control 72 (1999) 374–384.

[19] H. Kwakernaak, Frequency domain solution of the H

∞

problem for descriptor systems, in: S. Hara,

Y. Yamamoto (Eds.), Learning Control and Hybrid Systems, Lecture Notes in Control and Informa-

tion Sciences, vol. 241, Springer, New York, 1999, pp. 317–336.

[20] H. Logemann, Destabilizing effects of small time delays on feedback-controlled descriptor systems,

Linear Algebra Appl. 272 (1998) 131–153.

[21] I. Masubuchu, Y. Kamitane, A. Ohara, N. Suda, H

∞

control for descriptor systems: a matrix in-

equalities approach, Automatica 33 (1997) 669–673.

[22] S.-I. Niculescu, V. Rasvan, Delay-independent stability in lossless propagation models with appli-

cations, in: Proceedings of MTNS 2000, Perpignan, France, 2000.

[23] V. Rasvan, Absolute Stability of Time Lag Control Systems, Ed. Academiei, Bucharest, 1975 (in

Romanian).

[24] C.E. de Souza, X. Li, Delay-dependent robust H

∞

control of uncertain linear state-delayed systems,

Automatica 35 (1999) 1313–1321.

[25] P. Park, ADelay-dependent stability criterion for systems with uncertain time-invariant delays, IEEE

Trans. Automat. Control 44 (1999) 876–877.

[26] A. Rehm, F. Allgower, Descriptor and nondescriptor controllers in H

∞

control of descriptor sys-

tems, in: Proceedings of ROCOND, Prague, 2000.

[27] K. Takaba, N. Morihira, T. Katayama, H

∞

control for descriptor systems—a J-spectral factorization

approach, Proceedings of the 33rd IEEE Conf. Decision and Control, 1994, pp. 2251–2256.

[28] W. Zhu, L. Pitzold, Asymptotic stability of linear delay differential–algebraic equations and numer-

ical methods, Appl. Numer. Math. 24 (1997) 247–264.

[29] H. Xu, K. Mizukami, Upper and lower bounds of H

∞

-optimal performance for a class of continuous

type descriptor systems, Proceedings of the 38th Conf. Decision and Control, Cobe, Japan, vol. 1,

1996, pp. 1021–1026.

272

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

and the references therein). Depending on the area of application, these models are called singular or implicit or descriptor systems with delay. As has been pointed out in [6,7], descriptor systems with delay may in fact be systems of advanced type. Descriptor systems may be destabilized by small delay in the feedback [20]. There are only few papers on descriptor systems with delay [6,7,10,15,22,23,28]. A particular case of these systems (the so-called lossless propagation models), described by x1 (t) = Ax1 (t) + Bx2 (t − h), ˙ x2 (t) = Cx1 (t) + Dx2 (t − h), (1) has been treated as a special class of neutral systems either by letting x2 (t) = y2 (t) ˙ [23] or by writing the second equation as [15,22] d x2 (t) − Cx1 (t) − Dx2 (t − h) = 0. (2) dt The stability of a general neutral type descriptor system with a single delay described by E x(t) + Ax(t) + B x(t − h) + Cx(t − h) = 0 ˙ ˙ (3)

with a singular matrix E has been studied in [28] by analyzing its characteristic equation det sE + A + (sB + C) exp(−hs) = 0 and ﬁnding frequency domain conditions which guarantee that all roots of the latter equation have negative real parts bounded away from 0. A Lyapunov-based approach to stability analysis of descriptor system with delay has been introduced in [10], where delay-independent and delay-dependent linear matrix inequalities (LMIs) conditions have been derived. For information on LMI approach to control, see [3]. All the above-mentioned results only analyze the existence and the stability of solutions of descriptor systems with delay. To the best of our knowledge no control problem solution has been derived for this class of systems. For descriptor systems without delay, H∞ -control problems have been treated in the frequency domain [19,27] and in the time-domain [21,26,29]. In [21,26] an LMI approach has been proposed. For nondescriptor systems with state-delay, LMI delay-dependent and delay-independent H∞ -controllers were derived in [12,16,24] (see also the references therein). These ﬁnite-dimensional LMIs provide sufﬁcient conditions only for inﬁnite-dimensional systems with state-delay. Unlike inﬁnite-dimensional methods (see e.g. [1,11]) they lead to effective numerical algorithms and may be applied for systems with polytopic uncertainties. In the present paper, we adopt the ﬁnite-dimensional LMI approach to H∞ -control of descriptor system with delay. Our objective is to obtain delay-dependent solutions which are less conservative than the delay-independent ones. We apply the descriptor model transformation that has been introduced recently for delay-dependent stability and H∞ -control of nondescriptor systems [9,12]. We derive bounded real lemmas (BRLs) and ﬁnd solutions to the H∞ ﬁltering, the state-feedback and

E. Fridman, U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302

273

the output-feedback H∞ -control problems. The solutions are delay-dependent with respect to the ‘slow’ variable and delay-independent with respect to the ‘fast’ one. The latter guarantees the robustness of the system behavior with respect to the small changes in the delay. Notation. Throughout the paper the superscript ‘T’ stands for matrix transposition, Rn denotes the n-dimensional Euclidean space with vector norm | · |, Rn×m is the set of all n × m real matrices, and the notation P > 0 for P ∈ Rn×n means that P is symmetric and positive definite. The space of functions in Rq that are square inq tegrable over [0 ∞) is denoted by L2 [0, ∞) with norm · L2 . Let Cn [a, b] denote the space of continuous functions φ : [a, b] → Rn with the supremum norm | · |. We √ also denote xt (θ) = x(t + θ) (θ ∈ [−h, 0]) and j = −1.

**2. Problem formulation Given the following system:
**

2

E x(t) = ˙

i=0

Ai x(t − hi ) + B1 w(t) + B2 u(t),

x(t) = 0, ∀t

0,

(4) (5a) (5b)

**¯ y(t) = C2 x(t) + D21 w(t), ¯ ¯ z(t) = col C1 x(t); D12 u(t) ,
**

n1 n2

where x(t) = col{x1 (t), x2 (t)}, x1 (t) ∈ R , x2 (t) ∈ R is the system state vector, q u(t) ∈ R is the control input, w(t) ∈ L2 [0, ∞] is the exogenous disturbance signal, y(T ) ∈ Rr is the measurement vector and z(t) ∈ Rp is the state combination (objec¯ tive function signal) to be attenuated. The time delays h0 = 0, hi > 0, i = 1, 2, are assumed to be known. We took for simplicity two delays, but all the results are easily generalized for the case of any ﬁnite number of delays h1 , . . . , hm . The singular matrix E and the matrices Ai , Bi are constant matrices of appropriate dimensions. Denote n n1 + n2 . Following [21,26], we assume for simplicity that E= In1 0 0 . 0 (6)

Every descriptor system can be represented in a form satisfying this assumption. Note that in [21] there is = T > 0 instead of In1 , but from such a system, the system with E of (6) follows immediately. The matrices in (4), (5a) and (5b) have the following structure: Ai = Bi = Ai1 Ai3 Ai2 , Ai4 i = 0, 1, 2, (7) ¯ Ci2 , i = 1, 2.

Bi1 ¯ ¯ , Ci = Ci1 Bi2

i. Such a system may have an impulsive solution. The index of (8) is the index of nilpotence ν of N. N ν−1 = 0. then (8) has index greater than 1 (see e. We do not require A04 in (4) to be nonsingular. . ∞].8. ˙ y(t) − 2 i=1 Fi y(t 2 − hi ) 0 = i=0 Ai x(t − hi ) + B1 w(t) + B2 u(t). The index of the system with delay 2 E x(t) = ˙ i=0 Ai x(t − hi ) + B1 w(t) (10) is deﬁned in [10] to be equal to the index of (8).274 E. asymptotically stable for w = 0) and J (w) < q 0 for all disturbances w(t) ∈ L2 [0. In2 (9) where N ∈ Rn2 ×n2 and J ∈ Rn1 ×n1 are in Jordan form. For a prescribed scalar γ > 0. Q ∈ Rn×n such that QEP = In1 0 0 . The existence of the latter solutions is usually studied in terms of the Weierstrass canonical form and the index of the system which are deﬁned as follows [5.e. Fridman. (11) can be reduced to the form of (4) and (6). The descriptor system (10) admits impulsive solutions iff ν > 1 [10].g. The matrix N is nilpotent of / index ν. It is well known that descriptor system may have impulsive solutions. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 A descriptor system without delay E x(t) = A0 x(t) + B1 w(t) ˙ (8) is regular if the characteristic polynomial det(sE − A0 ) does not vanish identically in s ∈ C. the index of the open loop system (4) with delay is higher than one. N ν = 0. Hence. The nonsingularity of A04 guarantees the existence and the uniqueness of solution to (4) with u = 0 (see Proposition 3. This follows from the fact that the augmented system x1 (t) = y. If A04 is singular. we deﬁne the performance index J (w) = 0 ∞ zT z − γ 2 w T w dτ. The following class of neutral descriptor systems x1 (t) − ˙ 2 ˙ i=1 Fi x1 (t − hi ) 2 0 = i=0 Ai x(t − hi ) + B1 w(t) + B2 u(t). (13) The problem is to ﬁnd a controller such that the resulting closed loop system has index at most one is internally stable (i. [5. (12) is a particular case of (4) and (6).20]: there exist nonsingular matrices P .1).8]). U.e. N QA0 P = J 0 0 .

y2 (t) = 0 ∀t 0. such that x1 is absolutely continuous and x2 is integrable on [0. Hence. t1 ]. Assume that A04 is nonsingular. 2. Ad4 Denote h = max{h1 . d}. By denoting y2 = x2 we obtain from (14a) a neutral type system with the ze˙ ˙ ro initial conditions x1 (t) = 0. (14a) (14b) (15) x(t) = 0 z(t) = col C0 x(t). ∞) the solution to (14a) and (14b) exists and is unique on [0. w(t) ∈ L2 [0. C2 x(t − h2 ) . i = 0. t1 ] (see [17. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 275 3. Hence. 143]). y2 (t) on [0. Proof. ∞] is the exogenous disturbance signal and z(t) ∈ Rp is objective function signal. Fridman. is the q system state vector. Ad (s) is a piecewise-continuous and uniformly bounded (n1 + n2 ) × (n1 + n2 )-matrix-function. By solution of (14a) and (14b) on the segment [0. The matrix A04 is nonsingular and the difference operator D : Cn2 [−h. t1 ] (t1 > 0) we understand a pair of functions {x1 (t). This initial value problem for neutral system has a piecewise absolutely continuous solution x1 (t). We assume that the matrices in (14a) and (14b) have the structure of (7) and Ci = Ci1 Ci2 . x1 (t) ∈ R . Proposition 3. t1 ] and the initial conditions (14b). Stability of the difference operator and of the descriptor system We assume: A1. c ∈ Rn2 . y2 (t) = c ∀t 0. t1 ] such that x1 (t) is absolutely continuous on [0. Therefore. q 3. n1 n2 where E is deﬁned in (6). To prove the uniqueness we assume that there are two solutions of (14a) and (14b). x2 (t)}. For w(t) ∈ L2 [0. x2 (t) is integrable and solution of (14a) and (14b) exists. Then their difference satisﬁes the homogeneous equations (14a) and (14b) with w = 0. U. x2 (t)}. 1. x(t) = col{x1 (t). Stability and L2 -gain analysis of a descriptor system with delay BRLs will be obtained for systems with discrete and distributed delays. which has a unique solution x ≡ 0 [10]. t1 ] for all t1 > 0. h2 . solution of (14a) and (14b) is unique.E. 0 −d Ad (s)x(t + s) ds + B1 w(t). p.1.1. x2 (t) ∈ R . C1 x(t − h1 ). Ad = Ad1 Ad3 Ad2 . these functions satisfy system (14a) almost for all t ∈ [0. Given the following system: 2 E x(t) = ˙ i=0 Ai x(t − hi ) + ∀t 0. 0] → Rn given by .

(ii) the difference operator 2 D(x2t ) = x2 (t) + i=1 A−1 Ai4 x2 (t − hi ) 04 is stable for all h1 . (iii) under additional assumption that Pf > 0 the “fast system” 2 x2 (t) = A04 x2 (t) + ˙ i=1 Ai4 x2 (t − hi ) (18) is asymptotically stable for all h1 . Theorem 6. h1 ) in the fast variable x2 we assume instead of A1 the following: A1 . where Ad4 = 0. k = 1. 04 In the case of multiple discrete delays in D. A1 is equivalent to the following one (see [14. Fridman.276 E. U. In the case of single delay (e. 2π]. p. A sufﬁcient LMI condition for A1 is given by the following: Lemma 3. U2f that satisfy the following LMI: T T T Pf A04 + AT Pf + 2 Uif Pf A14 Pf A24 i=1 04 (17) ∗ −U1f 0 < 0. 04 (16) Evidently A1 is equivalent to A1 in the case of single delay h1 . If σ (B) is the spectral radius of matrix B. h2 . Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 2 D(x2t ) = x2 (t) + i=1 A−1 Ai4 x2 (t − hi ) + 04 0 −d A−1 Ad4 (s)x2 (t + s) ds 04 is stable for all delays h1 and h2 (i.2 [10]. . 2 .g. equation Dx2t = 0 is asymptotically stable for all h1 and h2 ). A sufﬁcient condition for A1 is the following inequality: 2 i=1 0 −d A−1 Ai4 + 04 A−1 Ad4 (s) ds < 1. 04 where | · | is any matrix norm. then σ0 < 1. 286]): A1 . where 2 σ0 sup σ k=1 A−1 Ai4 eiθk : θk ∈ [0.1. All the eigenvalues of A−1 A14 are inside of the unit circle. If there exist n2 × n2 -matrices Pf . U1f .e. ∗ ∗ −U2f then A04 is nonsingular and (i) A1 holds. h2 .

2.e. P3 t t−hi T P1 = P1 > 0.2. (19a) (19b) −α|φ(0)|2 . Under A1 (14a) and (14b). We obtain analogously to [10] the following: Theorem 3. (14b) with w = 0 is asymptotically stable. U. Delay-independent BRL (with respect to discrete delays) Descriptor type Lyapunov–Krasovskii functional for system (14a). β. Lemma 3. (14b) has the following form: V (xt ) = x T (t)EPx(t) + V1 + V2 . then (14a). ¯ and the function V (t) = V (xt ) is absolutely continuous for xt satisfying (14a) and (14b) with w = 0. The functional (20) is degenerated (i. Ui > 0. R = R T that satisfy the following LMI: . where P = P1 P2 2 (20) 0 . Under A1 if there exist positive numbers α. h2 0 if there exist n × n-matrix P of (21) with n1 × n1 -matrix P1 and n2 × n2 -matrix P3 and n × n matrices Ui = UiT . ∞] and for all h1 0. (21) V1 = i=1 x T (s)Ui x(s) ds. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 277 The following result on stability of (14a) and (14b) with w = 0 has been obtained recently [10]. (15) is internally asymptotically stable q and for a prechosen γ > 0 J (w) < 0 for all nonzero w(t) ∈ L2 [0. 0] → R such that β|φ1 (0)|2 ˙ V (φ) V (φ) γ |φ|2 .3.4. γ and a continuous functional V : Cn [−h. (23) The ﬁrst term of (20) corresponds to the descriptor system. 3. Fridman. nonpositive-definite) as it is usual for descriptor systems. i = 1. (22) and V2 = 0 −d t t+θ x T (s)AT (θ)RAd (θ)x(s) ds dθ. V1 corresponds to the delay-independent stability with respect to the discrete delays and V2 —to delaydependent stability with respect to the distributed delays [18].E. d R > 0.

= AT (−jωE − AT )−1 1 0 ∗ = the following frequency domain inequality is readily obtained diag U1 . Following [9. The latter guarantees robust stability with respect to small changes of delay [10]. From Theorem 3. .3. 3. if LMI (24) is feasible (and thus. (27) Therefore. ∗ from the right. then for all ω ∈ R 1/2 the frequency domain inequality (27) holds. This is a counterpart of the Kalman–Yakubovich– Popov lemma for descriptor systems. U.278 E. (26) is feasible). −U1 or AT − jωE − AT 1 0 −1 <0 U1 (jωE − A0 )−1 A1 < U1 . With respect to discrete delays in the fast variables the results will be delay-independent. Fridman. (24) Remark 1. Delay-dependent BRL We are looking for delay-dependent conditions with respect to slow variable x1 .4 it follows that the system E x(t) = A0 x(t) + A1 x(t − h) ˙ is asymptotically stable for all h P T A0 + AT P 0 AT P 1 Multiplying (26) by 0 if the following LMI is feasible: 0 < 0. −U1 ∗ (25) U P T A1 + 1 0 0 (26) from the left and by I . where (jωE − A0 )−1 A1 . ˙ y(t) = 0 − i=1 2 i=0 Ai1 2 i=0 Ai3 2 A02 A04 0 −hi x(t) + 2 i=1 Ai2 x (t − hi ) Ai4 2 0 −d (28) Ai1 Ai3 y(t + s) ds + Ad (s)x(t + s) ds + B1 w(t). Hence the H∞ -norm of U1 (jωE − −1/2 A0 )−1 A1 U1 is less than 1.10] we represent (14a) and (14b) in the equivalent form: x1 (t) = y(t). I ω ∈ R. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 2 U + 0 AT (s)RA (s) ds d i=1 i −d d P T A0 + AT P + 0 ∗ ∗ ∗ ∗ P T B1 −γ 2 Iq ∗ ∗ ∗ P T A1 0 −U1 ∗ ∗ P T A2 0 0 −U2 ∗ dP T 0 0 0 −dR < 0.

the second and the fourth terms—to the delay-dependent conditions with respect to x1 and the third—to the delay-independent conditions with respect to x2 . Ai = 0 i=0 2 0 0 A04 i=0 Ai3 0 0 0 0 ¯ ¯ Hi = Ai1 . 0 < Ri3 ∈ Rn×n . B1 Ai3 Ad3 0 Ad4 In1 ¯ E= 0 0 0 0 0 0 0 0 = 0 Ai2 . the ﬁfth term corresponds to delaydependent with respect to distributed delay. Fridman. ¯ x2 ¯ ¯ ¯ Ad (s)x(t + s) ds + B1 w(t). Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 279 The latter system can be represented in the form: ¯˙ E x(t) = ¯ 2 i=0 ¯ ¯ Ai x(t − hi ) + 0 −d 2 0 hi Hi i=1 −hi y(t + s) ds (29) + where x1 x = y . ¯ where P has the structure of (21) with P1 ∈ Rn1 ×n1 .E. i = 1. Ai4 0 . 0n1 ×n1 0 0n2 ×n2 0 I 0 0 ¯ ¯ A0 = 2 Ai1 −In1 A02 . B1 (30) A Lyapunov–Krasovskii functional for system (28) has the form: ¯ ¯ V (t) = x T (t)EP x(t) + ¯ 2 2 i=1 t t−hi T x1 (τ )Si x1 (τ ) dτ + i=1 2 t t−hi 0 T x2 (τ )Ui x2 (τ ) dτ t t+θ + + i=1 −hi 0 t −d t+θ y T (s) AT i1 AT Ri3 i3 Ai1 y(s) ds dθ Ai3 (31) ¯d ¯ ¯ x T (s)AT Rd Ad x(s) ds dθ. 2. . 0 < Ui ∈ Rn2 ×n2 . Rd ∈ R(n1 +n)×(n1 +n) . P3 ∈ Rn×n and 0 < Si ∈ Rn1 ×n1 . Ad = Ad1 0 Ad2 . The ﬁrst term of (31) corresponds to the descriptor system. U. .

P3 ∈ Rn×n . i = 1. T 2 Ci1 Ci1 0n1 0 ¯ = + 0 0 0 + TC 0 0 Ci2 i2 i=0 T T 2 0 Ai1 Ai3 0 + 0 0 Wi 0 0 i=1 0 and PT 0 2 i=0 Ai1 2 i=0 Ai3 0 WiT Ai1 Ai3 i=1 2 0 0 0 0 0 0 In1 −In1 0 0 hi AT i1 A02 + A04 0 0 2 i=0 Ai1 2 i=0 Ai3 In1 −In1 0 A02 P A04 0 T Si 2 0 + i=1 0 + 0 −d 0 Ai1 Ai3 AT Ri3 i3 0 0 Ui (33) ¯d ¯ AT (s)Rd Ad (s) ds and where Ri3 ∈ Rn×n is the (2. (14b) and (15) is internally asymptotq ically stable and for a prechosen γ > 0. Si = SiT ∈ Rn1 ×n1 . Under A1 the system of (14a). . 2. (32) where Xi = WiT + P T . 2 that satisfy the following LMI: ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ¯ PT 0 B1 −γ 2 Iq ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0 0 0 0 T T −W1 A11 −W2 A21 P T A12 P T A22 dP T A13 A23 A14 A24 0 0 0 0 0 0 0 −h1 R1 0 0 0 0 0 0 ∗ −h2 R2 0 0 0 0 0 ∗ ∗ −S1 0 0 0 0 ∗ ∗ ∗ −S2 0 0 0 ∗ ∗ ∗ ∗ −U1 0 0 ∗ ∗ ∗ ∗ ∗ −U2 0 ∗ ∗ ∗ ∗ ∗ ∗ −dRd h1 X1 h2 X2 < 0. 0 < P1 ∈ Rn1 ×n1 . J (w) < 0 for all nonzero w(t) ∈ L2 [0. P2 . Ui = UiT ∈ Rn2 ×n2 . Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 We obtain the following: Theorem 3. Wi ∈ R(n1 +n)×(n1 +n) and Ri = RiT ∈ R(n1 +n)×(n1 +n) . i = 1. ∞] if there exist matrices P ∈ R(n1 +n)×(n1 +n) of (21). Fridman. 2) block of Ri .5. U.280 E.

¯ Ai3 ¯ x T (t)P T Ad (s)x(t + s) ds. U. For any (n1 + n) × (n1 + n)-matrices Ri > 0 and Mi the following inequality holds [25]: −2 t t−hi bT (s)a(s) ds t t−hi a(s) b(s) T Ri MiT Ri Ri Mi (2. w(t). Since T ¯ ¯ x T (t)EP x(t) = x1 (t)P1 x1 (t) ¯ differentiating the ﬁrst term of (31) with respect to t we have x1 (t) ˙ d T ¯ T ¯ ˙ ¯ x (t)EP x(t) = 2x1 (t)P1 x1 (t) = 2x T (t)P T 0 . ¯ ¯ col x(t). x2 (t − h1 ). x2 (t − h2 ) . Fridman. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 281 Proof. ¯ 0 x T (t)P T Ai1 y(s) ds.E. 2) a(s) ds b(s) (36) . ¯ dt 0 Substituting (28) into (34) we obtain dV (xt ) + zT (t)z(t) − γ 2 w T (t)w(t) dt 0 0 PT P T A12 B1 A14 = ξT ∗ −γ 2 Iq 0 ∗ ∗ −U1 ∗ ∗ ∗ 2 (34) 0 P T A22 A24 ξ + zT z + 0 0 −U2 2 ηi i=0 − i=1 T T x1 (t − hi )Si x1 (t − hi ) + x2 (t − hi )Ui x2 (t − hi ) + − where ξ 0 −d t t−hi y T (s) AT i1 AT Ri3 i3 Ai1 y(s) ds Ai3 (35) ¯d x T (t + θ)AT (θ)Rd Ad (θ)x(t + θ) dθ. 2. ¯ ¯ is deﬁned by (33) and ηi (t) η0 (t) −2 −2 t t−hi t t−d i = 1.

V of (31) satisﬁes (19a). ¯ ¯ (38) We substitute (37). ∞] we have x(t) ∈ ¯ L2 [0. x2 (t − h2 ). Denoting Wi = Ri Mi P and using this inequality for a(s) = col{0 Ai1 Ai3 }y(s) and b = P x(t) we obtain for i = 1. Hence. for w(t) ∈ L2 [0. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 for a(s) ∈ Rn1 +n . η}. t]. . 3. (19b) and hence (14a) and (14b) are internally stable. . hi i = 1. V (x∞ ) 0 and ∞ 0 2 zT z dt = i=0 2 0 ∞ x T (t − hi )CiT Ci x(t − hi ) dt x T (t)CiT Ci x(t) dt. U. m. 2 ¯ ηi (t) hi x T (t)(WiT + P T )Ri−1 (Wi + P )x(t) ¯ ¯ T T + 2 x1 (t) − x1 (t − hi ) 0 AT i1 ¯ AT Wi x(t) i3 (37) + Similarly η0 (t) t t−hi y T (s) AT i1 AT Ri3 i3 Ai1 y(s) ds. Ai3 −1 dx T (t)P T Rd P x(t) ¯ ¯ + t t−d ¯d ¯ x T (t + s)AT (s)Rd Ad (s)x(t + s) ds. (39) . b(s) ∈ Rn1 +n given for s ∈ [t − hi . ∞] and J (w) < 0 if (32) holds. Fridman. Because V (x0 ) = 0. 2) = (MiT Ri + I )Ri−1 (Ri Mi + I ).282 E. Here (2. = i=0 0 ∞ we obtain (by Schur complements) that z where ¯ ξ 2 L2 − γ2 w 2 L2 ¯ ¯ ξ T ξ < −α x 2 L2 . w(t).4. Ri = εI2n . . . ¯ where η is vector of ﬁctitious states. x2 (t − h1 ). (38) into (35) and integrate the resulting inequality in t from 0 to ∞. Moreover. is the matrix in the left-hand side of (32) and ¯ col{x(t). Another delay-independent BRL (with respect to discrete delays) For Wi = −P . α > 0.

Ad are all known.1 the following corollary holds: Corollary 3. then (32) is feasible for a small enough ε > 0 and for Ri and Wi that are given by (39).5. However. from Theorem 2. (14b).1 is most powerful in the sense that it provides sufﬁcient conditions for both the delaydependent and the delay-independent cases (where (40) holds). P2 . . Under A1 the system of (14a). the delay-dependent BRL of Theorem 2. P3 . 3. In the latter case. i = 1. Thus. U. the theorem can be used to derive a criterion that will guarantee stability and the required attenuation level in the case where the system matrices are not exactly known and they reside within a given polytope.E. Fridman. (32) is feasible for hi → ∞. Si Remark 2. (14b) where the system matrices Ai .1 was derived for system (14a). Ci . 2. since the LMI of (32) is afﬁne in the system matrices. −d 0 0 Ui i=1 T If LMI (40) is feasible. and (15) is stable for all T hi > 0. Ui = UiT and Si = T . B1 . 2 and J < 0 if there exist 0 < P1 = P1 . i = 1. i = 1. 2 that satisfy (40). Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 283 LMI (32) implies for ε → 0+ the following delay-independent LMI: 0 0 0 0 0 PT P T A11 P T A21 P T A12 P T A22 dP T B1 A13 A23 A14 A24 ∗ −γ 2 Iq 0 0 0 0 0 ∗ ∗ −S1 0 0 0 0 < 0. As we have seen above. 2. i = 1. ∗ 0 0 0 ∗ ∗ −S2 ∗ 0 0 ∗ ∗ ∗ −U1 ∗ ∗ ∗ ∗ ∗ −U2 0 ∗ ∗ ∗ ∗ ∗ ∗ −dRd (40) where T 0 I 0 0 I 0 A01 −In1 A02 + A01 −In1 A02 P =P A03 0 A04 A03 0 A04 2 Si 0 0 0 ¯d ¯ 0 0 0 + + AT (s)Rd Ad (s) ds.6. Delay-dependent BRL for systems with polytopic uncertainties The BRL of Theorem 2.

. 2.j = 1. 0 < Ui . On LMI conditions in the case of discrete delays We consider Ad = 0. . Another possibility is to look for P3 in Theorem 3. . P3 . W1 . (14b). 2. A1 holds. Ci .4. S1 . j = (j ) (j ) (j ) (j ) 1. . 1. P32 . . Wi4 . i = 1. . j =1 fj = 1. . N. . Consider the system of (14a). 2.7. 2) blocks of Uk . Rd . Wi2 . . Even in this simpler case condition A1 is not easily veriﬁable. j = 1. Ad . W2 . Wi1 . i = 0. . . where Ukf are (2. R2 . 3.4 in the diagonal form: P3 = diag P31 . That is why instead of A1 one can assume that the fast LMI (17) is feasible for some Pf . U. and 0 < Ri . namely. k = 1. . 0 < Rd . . 1. P1 . S2 . Ukf . Wi3 . N. the following (delay-independent) BRL follows: . . Assume that for all j = 1. We readily obtain the following: Corollary 3. ∞) if there exist n × n-matrices 0 < P1 . P32 ∈ Rn2 ×n2 . 2 ∈ Co{ j j. C12 x2 (t − h1 ).6. where the matrices Ai . where the N vertices of the polytope are described by j = A(j ) i Ad (j ) B1 (j ) Ci (j ) i = 0.284 E. . (41) In the latter case if the full-order LMI (24) holds for P3 of (41). B1 . P2 . Consider now a difference continuous system 2 0 = A04 x2 (t) + z(t) = col Ai4 x2 (t − hi ) + B12 w(t). . From Theorem 3. N. where the system matrices reside within the polytope . then (17) holds for Pf = P32 . R1 . N}. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 Denoting = Ai we assume that N Ad B1 Ci i = 0. C22 x2 (t − h2 ) (42) . 0 < Si . Fridman. For a prescribed γ > 0. i=1 C02 x2 . 2 . . are taken with the upper index j. . . the cost function (13) achieves J (w) < 0 over for all nonzero q (j ) (j ) (j ) w ∈ L2 [0. N = j =1 fj for some 0 fj 1. 1. N that satisfy (32) for j = 1. .

U2f that satisfy the following LMI: T T T T T Pf A04 + AT Pf + 2 Uif + 2 Ci2 Ci2 Pf A14 Pf A24 Pf B12 i=1 i=0 04 ∗ −U1f 0 0 < 0.8.7. 2. ∗ ∗ −U2f 0 ∗ ∗ ∗ −γ 2 I (43) then for all h1 > 0. U. and 1 are identical. ˜ i=0 (46) ∀t ∈ [0 h] ˜ ˜ where ξ(t) ∈ Rn . where (44) B1 . Given γ > 0. h2 > 0 the difference system (42) is internally stable and J (w) < 0. Its forward representation. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 285 Corollary 3. is described by ˙ E ξ (τ ) = ξ =0 2 T AT ξ(τ − hi ) + C0 z(τ ). Fridman. vol. 32]): ¯ 1 ∞ = sup σ C0 jωE − A0 − ω∈R 2 i=1 Ai e−jwhi where σ {D} denotes the largest singular value of D.g. 3. ˜ (47) ∀τ ∈ [−h 0]. it follows that the H∞ -norm of 1 is equal to the H∞ -norm of the following system: ˙ −E ξ (t) = ξ =0 2 T AT ξ(t + hi ) + C0 z(t). U1f . (45) is given by (see e. Since ¯ σ H (jω) = σ H T (−jω) ¯ ¯ for all the transfer function matrices H (s) with real coefﬁcients. p. vol. −1 1 of (14a) and (14b). Ad = 0. z(t) ∈ Rp and w(t) ∈ Rq . 2 .E. 2 Since the characteristic equations of asymptotically stable iff 1 is. [2. H∞ -norm of the ‘adjoint’ system We begin by noting that the H∞ -norm of the system z = C0 x(t). ˜ i T w(t) = B1 ξ(t). the former system is . 1]). Note that the latter system represents the backward adjoint of 1 (as deﬁned for nondescriptor case in [2. ˜ i i=0 T w(τ ) = B1 ξ(τ ). if there exist n2 × n2 -matrices Pf .

one can apply Theorem 3.1 we obtain for both systems γ = 2.5 for the original system and for its ‘adjoint’ and then choose the less conservative result.5 (48) Applying the BRL of Section 3 to the above matrices.5 x1 (t) ˙ x(t) = ˙ 0 0 z(t) = 0.5 to (48) and its ‘adjoint’ we obtain in both cases that the system is internally stable for h 1. 2. where x(t) = col{x1 (t). Applying the LMI condition of Theorem 3. 0. Fridman. via the control law u(t) = Kx(t). It is obvious from the T requirement of 0 < P1 . For h = . Therefore. results in a nonlinear matrix T T inequality because of the terms P2 B2 K and P3 B2 K. ∞). We see that in this example the conditions of Theorem 3. that P is nonsingular. K2 ] (49) (4). Note that the same results are obtained by choosing block-diagonal P3 with P32 > 0. For h = 1. x2 (t)} ∈ R2 . Delay-dependent state-feedback control We apply the results of the previous section to the inﬁnite-horizon H∞ -control problem. (44) and for its ‘adjoint’ may lead to different results.5 1 x(t).15.12 the corresponding results are γ0 = 40 and γt = 28. while for h = 1 we obtain for (48) γ0 = 9 and for its ‘adjoint’ γt = 6. Deﬁning P −1 = Q = Q1 Q2 0 . Note that for εi = −1 (32) yields the delay-independent condition of Corollary 3.5 0. K = [K1 . i = 1. (14b).5 for the ‘adjoint’ system are less conservative than those obtained for the original system. where εi ∈ R is a scalar parameter.3.5 for (14a). Given system (4). 0 −1 x(t) + −1 0 1 1 x(t − h) + w(t). Q3 (51a) . Example 1. Substituting (49) into obtain the structure of (14a) and (14b) with A0 + B2 K instead of A0 and T T ¯T ¯ C0 C0 = C1 C1 + K T D12 D12 K.6. we (50) q achieves J (w) < 0 for all nonzero w ∈ L2 [0. we consider the performance index of (13). For a prescribed scalar γ > 0. 4. We consider the following system 0. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 Sufﬁcient conditions of Theorem 3. The minimum achievable value of γ is however different in the two cases.286 E. In order to obtain an LMI we have to restrict ourselves to the case of the diagonal matrix P3 of (41) and (as well as in the nondescriptor problem) to the case of Wi = εi P . respectively. U. We look for the state-feedback gain matrix K which. and the fact that in (32) (P3 AT + A04 P3 ) must be negative 04 definite. We therefore consider another version of the BRL which is derived from (32). (6) with the objective vector (5b).

Iq+4n+2n1 } we multiply (32) by right. Q22 Q32 ∈ Rn2 ×n2 . on the left and on the and = diag{Q.E. Applying the Schur formula to the quadratic term in Q. U. 0 . 0 0 0 0 0 0 0 0 0 0 0 0 −1 −h1 R1 0 −1 ∗ −h2 R2 (52) where C0 = [C01 C02 ]. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 287 Q2 = Q21 . 0 2 1 = i=0 Ai1 2 i=0 Ai3 In1 −In1 0 0 0 0 A02 Q + QT 0 0 Q + 0 A04 2 0 0 In1 −In1 0 AT i3 + 0 εi Ai1 Ai3 i=1 2 εi QT 0 0 i=1 2 i=0 Ai1 2 i=0 Ai3 0 AT i1 A02 A04 (53) 0 T 0 0 0 . T (51b) (51c) Q3 = diag Q31 . and . we obtain the following inequality: T 1+ 2 0 B1 −γ 2 Iq ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ QT In1 0 h1 (ε1 + 1)In+n1 0 −h1 R1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ QT 0 In2 QT h2 (ε2 + 1)In+n1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ In1 0 QT 0 In2 0 0 0 0 0 0 0 −1 −S1 ∗ ∗ ∗ ∗ ∗ 0 0 0 0 0 0 0 0 −1 −S2 ∗ ∗ ∗ ∗ 0 0 0 0 0 0 0 0 0 −1 −U1 ∗ ∗ ∗ 0 0 0 0 0 0 0 0 0 0 −1 −U2 ∗ ∗ C 0 0 0 0 01 ε1 A11 ε2 A21 A12 A22 QT 0 A13 A23 A14 A24 T C02 0 0 0 0 0 0 0 0 0 0 0 0 −h2 R2 0 0 0 0 0 ∗ −S1 0 0 0 0 ∗ ∗ −S2 0 0 0 ∗ ∗ ∗ −U1 0 0 ∗ ∗ ∗ ∗ −U2 0 ∗ ∗ ∗ ∗ ∗ −Ip ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0n1 ×n1 0 0n1 ×n1 0 0 0 0 0 AT A T A T A T 11 13 h2 QT 21 23 h1 QT 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 < 0. Fridman. respectively. Q32 .

Consider the system of (4). 0 < R1 = R1 . and obtain the following: Theorem 4. (5b) and the cost function of (13). there exist −1 −1 n1 ×n1 ¯ ¯ . ε2 ∈ R. the state-feedback law of (49) achieves. T B2 . Y1 ∈ R and Y2 ∈ R that satisfy the following LMI: ¯ CT ¯ 1+ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0 B1 −γ 2 Iq ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ Q1 0 0 0 0 0 0 0 0 0 ¯ −S1 ∗ ∗ ∗ ∗ ∗ 0 0 0 0 11 ¯ ¯ ¯ ¯ ¯ ¯ h1 (ε1 + 1)R1 h2 (ε2 + 1)R2 ε1 A11 S1 ε2 A21 S2 A12 U1 A22 U2 QT 0 A13 A23 A14 A24 ¯ CT 12 (54) 0 ¯ −h1 R1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ Q1 0 0 0 0 0 0 0 0 0 0 ¯ −S2 ∗ ∗ ∗ ∗ YT 1 T 0 D 12 T Y2 0 0 0 0 0 0 0 Ip ∗ ∗ ∗ ∗ ∗ ∗ 0 0 ¯ −h2 R2 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ QT QT 22 22 0 0 QT QT 32 32 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ¯ −U1 0 ¯ ∗ −U2 ∗ ∗ ∗ ∗ 0 0 0 ¯ −S1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0 0 0 0 ¯ −S2 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0 0 0 0 0 ¯ −U1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0 0 0 0 0 0 ¯ −U2 ∗ ∗ ∗ ∗ ∗ ∗ ∗ T Q21 h1 QT 0 AT AT 11 13 31 0 0 0 0 0 0 0 0 0 0 0 0 0 ¯ −h1 R1 ∗ T Q21 h2 QT 0 AT AT 21 23 31 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ¯ −h2 R2 0 0 0 0 0 0 0 −Ip ∗ ∗ ∗ ∗ ∗ ∗ < 0. For a prescribed 0 < γ . 0 < R2 = R2 ∈ (n+n1 )×(n+n1 ) ×n1 ×n2 R . Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 2 = 0 B2 K1 0n1 T K1 K2 Q + QT 0n1 0 T K2 K2 Q32 = Y2 . (6). U. 0 < Uk = Uk ∈ Rn2 ×n2 .288 E. 0 < Sk = Sk ∈ Rn1 ×n1 . Fridman. 2.1. J (w) < 0 for q all nonzero w ∈ L2 [0. denote K1 Q1 + K2 Q22 = Y1 . We substitute (50) into (52). ∞) if for some prescribed scalars ε1 . (55) . k = 1. Q2 0 < Q1 ∈ R −1 −1 ¯ ¯ ∈ Rn×n1 and Q3 ∈ Rn×n of (51a)–(51c).

where the N vertices of the polytope (j ) = B1 (j ) A1 ¯ (j ) C 1 A2 (j ) (j ) . J (w) < 0 for all nonzero w ∈ L2 [0. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 289 where ¯ = 0 B2 Y1 0n1 Y2 T Y1 + 0n1 0 T Y2 T B2 . (57) 1 .j we assume that are described by (j ) ∈ Co{ E A0 B2 (j ) (j ) = 1. C1 = 1 Note that in this example A04 = 0.1 . D12 We obtain the following: Theorem 4.5 . The state-feedback gain is then given by K2 = Y2 Q−1 . 1 (56) Example 2. . 1 B2 = 0.1 is afﬁne in the system matrices. We ﬁrst ﬁnd the state-feedback solution. The LMI in Theorem 4. Fridman. For a prescribed 0 < γ . q the state-feedback law of (49) achieves.62 − 430680].255. Considering the system of (4) and denoting = E B1 A0 B2 A1 ¯ C1 A2 . ∞) .2 and ε1 = −0. 0 A1 = −1 1 0 . Consider the system of (4). We consider the system E x(t) = A1 x(t − h) + B1 w(t) + B2 u(t). D12 j.E. 32 K1 = (Y1 − K2 Q22 )Q−1 . We obtained a near minimum value of γ = 21 for h = 1. −0. where E= B1 = 1 0 0 . where the system matrices reside within the polytope and the cost function of (13). It can thus be applied also to the case where these matrices are uncertain and are known to reside within a given polytope. U. . The statefeedback control law that achieves the later bound on the H∞ -norm of the closed loop is u = Kx. 1 D12 = 0. N}. . ˙ ¯ z(t) = C1 x(t) + D12 u(t).2 . (5b). where K = [175.2. −1 x0 = 0. . (6).

Analogously to Theorem 4. ξ = 0 ∀τ ∈ [−h 0]. . B2 . k = 1. ˜ i i=1 (61) ξ(τ ). We seek a ﬁlter of the following observer form: ˙ E x(t) = ˆ 2 i=0 ¯ Ai x(t − hi ) + Kf (y(t) − C2 x(t)) ˆ ¯ (58) such that the H∞ -norm of the resulting transference between the exogenous signal w and the estimation error z is less than a prescribed value γ . Q2 ∈ Rn×n1 and Q3 ∈ Rn×n of (51a)–(51c).1 (by applying BRL of Theorem 3. D12 . ∞) ¯ if for some prescribed scalars ε1 . Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 and for all the matrices in if for some prescribed scalars ε1 . 2. .1. 0 < Uk = Uk ∈ Rn2 ×n2 .290 E. Y1 ∈ R and Y2 ∈ R that satisfy LMIs (55) for j = 1. ¯ i=1 (60) The problem then becomes one of ﬁnding the ﬁlter gain Kf such that J (w) < 0. 2 Ai e(t − hi ) + (B1 − Kf D21 )w. Delay-dependent ﬁltering We consider system (4) with the measurement law of (5a). U. For a q prescribed 0 < γ . ˆ (59) From (4). . where z(t) L(x(t) − x(t)). We consider the ‘adjoint’ to (60) system described by ¯T T ˙ E ξ (τ ) = AT − C2 Kf ξ(τ ) + 0 w(τ ) = ˜ T B1 T T − D21 Kf 2 AT ξ(τ − hi ) + LT z(τ ). 5. Fridman. Q2 ∈ Rn×n1 and Q3 ∈ Rn×n of (51a)–(51c). . (5a) and (58) it follows that the estimation error e(t) = x(t) − x(t) is ˆ described by the following model: ¯ E e(t) = (A0 − Kf C2 )e(t) + ˙ z(t) = Le(t). 0 < Sk = Sk ∈ Rn1 ×n1 . 0 < R2 ∈ R(n+n1 )×(n+n1 ) . ε2 ∈ R there −1 −1 ¯ ¯ exist 0 < Q1 ∈ Rn1 ×n1 . there exist 0 < Q1 ∈ Rn1 ×n1 . where the matrices ¯ ¯ ¯ Ai . Y1 ∈ Rr×n1 and Y2 ∈ Rr×n2 that satisfy the following LMI: . i = 0. N.5 to (61)) we obtain: Theorem 5. ε2 ∈ R. R2 are taken with the upper index j. (5a) and the cost function J (w). 0 < R2 = R2 ∈ (n+n1 )×(n+n1 ) ×n1 ×n2 R . R1 . C1 . the ﬁlter gain achieves. 0 < Sk ∈ n1 ×n1 ¯ k ∈ Rn2 ×n2 . 2. . The state-feedback gain is then given by (56). Consider the system of (4). J (w) < 0 for all nonzero w ∈ L2 [0.0 < U R ¯ ¯ 0 < R1 . −1 −1 ¯ ¯ 2. B1 . 1. 0 < R1 = R1 . k = 1.

Fridman. U.E. (62) where 1 = 0 2 T i=0 Ai1 2 T i=0 Ai2 In1 −In1 0 0 0 0 0 AT Q + QT 03 AT 04 2 0 0 2 T i=0 Ai1 2 T i=0 Ai2 In1 −In1 0 Ai2 0 0 AT 03 AT 04 0 T 2 + εi AT i1 i=1 AT i2 0 0 Q + 0 0 εi QT 0 0 i=1 Ai1 0 0 . Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 291 ¯ 1+ 0 LT −γ 2 Ir ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ¯ h1 (ε1 + 1)R1 ¯ h2 (ε2 + 1)R2 0 T ¯ ε1 A11 S1 AT 12 0 0 0 ¯ −S1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ QT 22 0 QT 32 0 0 0 0 0 0 0 0 0 0 ¯ −U1 ∗ ∗ ∗ 0 T ¯ ε2 A21 S2 AT 22 0 0 0 0 ¯ −S2 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ T Q 22 0 QT 32 0 0 0 0 0 0 0 0 0 0 0 ¯ −U2 ∗ ∗ 0 T A U1 ¯ 13 AT 14 0 0 0 0 0 ¯ −U1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0 ¯ −h1 R1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0 0 ¯ −h2 R2 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ T 0 Y B11 T 1 A U2 QT 0 − 0 D21 ¯ 23 T B12 Y2 AT 24 0 0 0 0 0 0 ¯ U2 ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0 0 0 0 0 0 0 −Iq ∗ ∗ ∗ ∗ ∗ ∗ h1 A1 h2 A2 0 0 0 0 0 0 0 0 0 0 0 0 0 ¯ −h2 R2 Q1 0 0 0 0 0 0 0 0 0 ¯ −S1 ∗ ∗ ∗ ∗ ∗ Q1 0 0 0 0 0 0 0 0 0 0 ¯ −S2 ∗ ∗ ∗ ∗ 0 0 0 0 0 0 0 0 0 0 0 0 ¯ −h1 R1 ∗ < 0.

6. The ﬁlter gain is then given by T K2 = Y2 Q−1 . ∞) if for some prescribed scalars −1 −1 ¯ ¯ ε1 . 6. 1 Kf = col K1 . 32 T T K1 = (Y1 − K2 Q22 )Q−1 . 2. It applies a controller of a state-feedback—observer structure and requires a solution of two LMIs. (5b).1. 0 < R1 = R1 −1 (n+n1 )×(n+n1 ) ¯ and 0 < R2 = R2 ∈ R that satisfy the following LMI: .292 E. Ai1 Ai2 . 0 < Sk = Sk ∈ Rn1 ×n1 . B1 D21 = 0. and R = D12 D12 is not singular.1. the feedback law x1 (t) P1 0 T ˜ y(t) .1 is afﬁne in the system matrices. Delay-dependent output-feedback control We adopt in this section the dissipation approach to the solution of the outputfeedback problem.2. K2 . there exist 0 < Q1 ∈ Rn1 ×n1 . For a prescribed γ > 0.e. 0 < Uk = Uk −1 ¯ ∈ Rn2 ×n2 . ε2 ∈ R. 2. k = 1. (63) The LMI in Theorem 5. The matrices B1 and D21 are orthogonal. Q2 ∈ Rn×n1 and Q3 ∈ Rn×n of (51a)–(51c). U. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 ¯ = − 0T ¯ C2 Y1 0n1 Y2 T Q21 T Ai = Q31 0 0 T Y1 − 0n1 0 T Y2 ¯ C2 . i = 1. Assume A2. i. Similarly to Theorem 4. The ﬁrst phase: a state-feedback controller design Lemma 6. Consider system (4). u(t) = − 0 R −1 B2 (64a) P2 P3 x2 (t) P1 P2 Q1 0 = Q2 P3 0 Q3 −1 (64b) q achieves J (w) < 0 for all nonzero w ∈ L2 [0. Fridman. We assume: T T ˜ A2. it can thus be reformulated also to the case of matrices with polytopic uncertainties.

. Fridman. i = 1. Denote by u the matrix in the left-hand side of (65).E. (65) where 1 is given by (53) and where 0 ˜ −1 R 0 B2 T B2 . Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 293 ¯ 1+ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ T ¯ C11 0 0 0 0 ¯ ¯ ¯ ¯ ¯ ¯ h1 (ε1 + 1)R1 h2 (ε2 + 1)R2 ε1 A11 S1 ε2 A21 S2 A12 U1 A22 U2 QT 0 ¯T A13 A23 A14 A24 C12 0 0 0 0 0 0 0 −γ 2 Iq ¯ 0 0 0 0 0 0 ∗ −h1 R1 ¯ ∗ ∗ −h2 R2 0 0 0 0 0 ¯ 0 0 0 0 ∗ ∗ ∗ −S1 ¯ 0 0 0 ∗ ∗ ∗ ∗ −S2 ¯ ∗ ∗ ∗ ∗ ∗ −U1 0 0 ¯ 0 ∗ ∗ ∗ ∗ ∗ ∗ −U2 ∗ ∗ ∗ ∗ ∗ ∗ ∗ −Ip ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0 B1 T Q22 0 QT 32 0 0 0 0 0 0 0 0 0 0 ¯ −U1 ∗ ∗ ∗ T Q22 0 QT 32 0 0 0 0 0 0 0 0 0 0 0 ¯ −U2 ∗ ∗ T Q21 h1 QT 0 31 Q1 0 0 0 0 0 0 0 0 0 ¯ S1 ∗ ∗ ∗ ∗ ∗ Q1 0 0 0 0 0 0 0 0 0 0 ¯ −S2 ∗ ∗ ∗ ∗ AT 11 AT 13 T Q21 h2 QT 0 31 AT 21 AT 23 0 0 0 0 0 0 0 0 0 0 0 0 0 ¯ −h1 R1 ∗ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ¯ −h2 R2 < 0. U. ˆ =− Proof. We obtain by integrating dV (t)/dt that ∞ J 0 ¯ ξT ¯ pξ ∞ dt + 0 ∞ ˜ (uT − u∗T )R(u − u∗ ) dt (66) −γ2 0 (w T − w ∗T )(w − w ∗ ) dt. The proof readily follows by choosing V as in (31) and applying (36) for Ri Mi = εi I. 2.

A04 Ai2 B11 ˆ Ai2 . The second phase: ﬁltering via the adjoint system Denoting r = w − w ∗ we represent (4) and (5a) in the form: ¯ 2 x1 (t) ˙ x1 (t − hi ) B21 ˆ ¯ ˆ 0 = Ai y(t − hi ) + B1 r (t) + B21 u(t). is given in (51a)–(51c) and ¯ ξ = col{x1 . . (67) and where the relation between P and Qi . (68) (69) 0r×n1 and where P solves (65). Ai4 B12 ¯ C22 .2. w. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 where = diag P T . . x2 . 2. the feedback law of (64a) and (64b) cannot be implemented even when there exists a solution to (65). B22 0 x2 (t − hi ) i=0 x1 (t) ˆ y(t) = C2 y(t) + D21 r (t).294 E. The objective function of (13) and (66) will then be negative if there exist x(t) and y(t) in Rn that satisfy ˆ ˆ . B1 = B11 . ¯ ¯ x2 (t) where A01 ˆ A0 = A01 A03 Ai1 ˆ Ai = Ai1 Ai3 ˆ ¯ C2 = C21 0 −In1 0 0 0 0 A02 T ˆ A02 + γ −2 B1 0q×n1 B1 P . Unfortunately. . namely when the ﬁrst term in the right-hand side ¯ of (66) is negative for all ξ ∈ R8n+4n1 +p+q . U. i = 1. . i = 1. I8n+4n1 +p+q . 3. η1 }. I8n+4n1 +p+q x1 T w ∗ = γ −2 0 B1 P y . with η1 representing the ﬁctitious states that emerge when applying Schur formula to construct p . 6. x2 x1 T ˜ u∗ = −R −1 0 B2 P y x2 p u diag P . Fridman. y.

2. (5b) and the cost function of (13). This problem was solved in Section 5. ε2 ∈ R. a H∞ ﬁltering problem for the ˆ ˆ descriptor system (68). ε2 ∈ R. 0 < Sk ∈ Rn1 ×n1 . ¯ 2 i=1 (72) ˆ ˆ r Ai e(t − hi ) + (B1 − Kf D21 )¯ (t).E. 2. 0 < R1 and 0 < R2 ∈ k = 1.2. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 295 Ja = (¯ T z − γ 2 r T r ) dt < 0 ∀¯ ∈ L2 [0. For a prescribed 0 < γ . J (w) < 0 for all nonzero w ∈ L2 [0. there exists an output-feedback controller q that achieves. 0 < Uk ∈ Rn×n . U. Q2 ∈ R(n+n1 )×n1 ˆ and Q3 ∈ R(n+n1 )×(n+n1 ) of the form . k = 1. B22 x2 (t) ˆ Denoting x1 x1 ˆ e1 ˆ e = e0 = y − y e2 x2 x2 ˆ and using the assumption on D21 and the definition of w ∗ in (67) we ﬁnd e1 (t) ˙ ˆ ˆ 0 = (A0 − Kf C2 )e(t) + 0 T ˜ z(t) = R −1/2 0 B2 P e(t). z ¯ ¯ ¯ r x1 (t) − x1 (t) ˆ T ˜ ˆ z(t) = R −1/2 0 B2 P y(t) − y(t) . (5a). ∞). Consider the following ‘innovation’ ﬁlter ˙ 2 x1 (t − hi ) ˆ x 1 (t) ˆ ˆ 0 = ˆ Ai y(t − hi ) + Kf y(t) ¯ x2 (t − hi ) ˆ i=0 0 x1 (t) ˆ B21 ˆ ˆ (71) −Kf C2 y(t) + B21 u(t).1 we obtain the following result: Theorem 6. Assume A2. By applying Theorem 5. Fridman. in fact. ∞) if for some prescribed scan1 ×n1 ¯ ¯ lars ε1 . Q2 ∈ R ˆ ˆ R(n+n1 )×(n+n1 ) that satisfy (65) and for some prescribed scalars ε1 . Consider the system of (4). n×n1 n×n ¯ ¯ and Q3 ∈ R of (51a)–(51c). 0 < Sk ∈ Rn1 ×n1 . there exˆ ˆ ˆ ˆ ist 0 < Q1 ∈ Rn1 ×n1 . (73) The problem now becomes one of ﬁnding the gain matrix Kf that will ensure the stability of system (73) and that the H∞ -norm of the transference from r to z is less ¯ ¯ than γ . ¯ ˆ x2 (t) − x2 (t) 0 ∞ q (70) The problem of ﬁnding x(t) and y(t) is. there exist 0 < Q1 ∈ R . 0 < Uk ∈ Rn2 ×n2 .

P32 . Q32 ∈ Rn×n . 3 1 (75) P3 = Q−1 = diag Q−1 . Q−1 = diag P31 .296 E. and Y2 ∈ Rr×n 0 T ˆ A U1 ˆ 13 ˆ AT 14 ˆ ˆ ˆ Q3 = diag Q31 . 3 31 32 P = P1 P2 0 . Q22 ˆ 0 < R1 . 1 P2 = −Q−1 Q2 Q−1 . ˆ ˆ Q31 ∈ Rn1 ×n1 . ˆ Q22 ∈ Rn×n1 . U. P3 . Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 ˆ Q ˆ Q2 = ˆ 21 . Q32 . that satisfy the following LMI: 0 0 T T ˆ ˆ ¯ ˆ ˆ ˆ ˆ ˆ ˆ ε ε 0 ˜ −1/2 h1 (ˆ 1 + 1)R1 h2 (ˆ 2 + 1)R2 ε1 A11 S1 ε2 A21 S2 1+ R P B2 ˆ ˆ AT AT 12 22 ∗ −γ 2 Ir 0 0 0 0 ˆ1 0 0 0 ∗ ∗ −h1 R ˆ 0 0 ∗ ∗ ∗ −h2 R2 ˆ 0 ∗ ∗ ∗ ∗ −S1 ˆ ∗ ∗ ∗ ∗ ∗ −S2 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0n1 ×r 0 ˆ T B11 Q B1 0 0 0 0 0 0 0 Iq ∗ ∗ ∗ ∗ ∗ ∗ ∗ YT 1 0 D21 T Y2 0 0 0 0 0 0 0 0 −Iq ∗ ∗ ∗ ∗ ∗ ∗ ˆ Q1 0 0 0 0 0 0 0 0 0 0 ˆ −S1 ∗ ∗ ∗ ∗ ∗ ˆ Q1 0 0 0 0 0 0 0 0 0 0 0 ˆ −S2 ∗ ∗ ∗ ∗ ˆ QT 22 0 ˆ QT 32 0 0 0 0 0 0 0 0 0 0 0 ˆ −U1 ∗ ∗ ∗ ˆ QT 22 0 ˆ QT 32 0 0 0 0 0 0 0 0 0 0 0 0 ˆ −U2 ∗ ∗ 0 T ˆ A U2 ˆ 23 ˆ AT 24 0 0 0 0 0 ˆ −U1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0 0 0 0 0 0 ˆ −U2 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ˆ h1 A1 0 0 0 0 0 0 0 0 0 0 0 0 0 ˆ −h1 R1 ∗ ˆ h2 A2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ˆ −h2 R2 < 0. Y1 ∈ Rr×n1 . Fridman. ˆ 0 < R2 ∈ R(2n1 +n)×(2n1 +n) . (74) where P1 = Q−1 .

0 T 0 ˆ AT 03 ˆT A 04 ˆ Ai1 0 0 Ai1 0 ˆ ˆ ˆ ˆ Ai1 = Ai1 . Ai4 = Ai3 0 T ˆ A01 = A01 + γ −2 B11 B1 P2 . Ai3 = . ˆ ˆ Example 3. . 2. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 297 1 = 0 2 ˆT i=0 Ai1 2 ˆT i=0 Ai2 I −In1 0 0 0 0 ˆT + εi Ai1 ˆ i=1 ˆ AT i2 2 ¯ = − 0T ˆ C2 Y1 0 2 ˆ ˆ T 0 0 0 Q + εi Q ˆ i=1 0 0 0 T Y1 ˆ 0n1 Y2 − 0n1 0 C2 . We consider the following system: 1 E x(t) = ˙ i=0 A1 x(t − hi ) + B1 w(t) + B2 u(t). Ai2 = 0 Ai2 . 2. T B11 B12 P32 . Fridman. (76) If a solution to (65) and (74) exists. T Y2 0 ˆ ˆ ˆ AT Q + QT 03 ˆ AT 04 0 2 ˆT i=0 Ai1 2 ˆT i=0 Ai2 I −In1 0 ˆ Ai2 0 . Ai4 0 The ﬁlter gain is then given by T ˆ K2 = Y2 Q−1 . x0 = 0. (77) where x and y are obtained by (71). U. T ˆ A02 = 0 A02 + γ −2 B11 B11 P31 T B11 B1 P2 A01 ˆ A03 = .E. i = 1. (78) ¯ z(t) = C1 x(t) + D12 u(t). A04 = + γ −2 T T 0 A04 B12 B11 P31 B12 B12 P32 T ˆ Q21 ˆT ˆ Ai = Q31 0 Ai1 0 Ai2 . ¯ y(t) = C2 x(t) + D21 w(t). K2 . 1 Kf = col K1 . Ai2 . + γ −2 T A03 B12 B1 P2 T T B11 B11 P31 B11 B12 P32 −In1 A02 ˆ . then the output-feedback controller is obtained by: T ˜ ˆ ˆ ˆ u(t) = −R −1 B2 P2 x1 (t) + P3 col{y(t) x2 (t)} . 32 T T ˆ ˆ K1 = (Y1 − K2 Q22 )Q−1 . i = 1.

6854 −0.1710 T .4 is obtained. For smaller values of h the method of [12] may lead to less conservative results (see Example 4 below).3. Using (65) we ﬁnd that the feedback law u∗ does not apply y because B21 = 0 and P3 is diagonal. 0 A0 = B2 = 0 0 0 . We ﬁrst ﬁnd the state-feedback solution. 0 . The second phase: direct ﬁltering The ﬁltering of the previous section suffers from an additional overdesign that stems from the use of the adjoint system which must be stable independently of the delays in the variable e0 .3.22x2 . The resulting output-feedback has the form ˆ ˆ 9.2 . Denoting r = w − w ∗ we represent (4) in the form: ¯ 2 0 x1 (t) ˙ x1 (t) t ˆ Ai1 0 = A y(t) − y(s) ds t−hi 0 x2 (t) i=1 Ai3 + where ˆ A= 0 0 r (t) + ¯ u(t). ε1 and ε1 = −1. where K = [42. We thus obtained the same near minimum value of γ = 11 for h = 1. U. −1 C2 = 1 0 . 0 0 . D21 = 0 0. The advantage of the approach of Section 6.1. We obtained a near minimum value of γ = 11 for h1 = 1.1 . lies in the fact that it applies the efﬁcient bounds introduced by Park [25]. 0 . 1 C1 = 1 D12 = 0. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 where E= B1 = 1 0 1 1 0 . The feedback control law that achieves the later bound on the H∞ -norm of the closed loop is u∗ = 9.72x1 + 3.1814 1.22x2 .3426].3. A04 . where nondescriptor systems were considered.2 and ε1 = −0. Fridman. In the present section. The state-feedback control law that achieves the later bound on the H∞ -norm of the closed loop is u = Kx. B2 B1 I −I 0 0 ˆ B1 = 0 . 6.2 in comparison with [12]. we generalize the method of [12] to the case of descriptor systems with delay. 2 A1 = −1 1 0.298 E.1882 −3. B1 (79) 0 2 i=0 Ai1 2 i=0 Ai3 ˆ ˆT A02 + γ −2 B1 B1 P .72x1 + 3. ˆ The output-feedback control is derived for the same values of h.2 and ε1 = −0. where x is obtained by (71) with ˆ Kf = 6. A minimum value of γ = 2.

ε2 ∈ R. Similarly to [12] we obtain the following result: Theorem 6. there exists an output-feedback controller that achieves. For a prescribed 0 < γ . ¯ 0 Ai1 i=1 Ai3 2 t t−hi e0 (s) ds ˆ (81) ˆ r +(B1 − Kf D21 )¯ (t). (5b) and the cost function of (13). ˆ ˆ P31 ∈ Rn1 ×n1 . Y ∈ R(n+n1 )×r ˆ ˆ ˆ P3 = diag P31 . Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 299 ˆ ¯ C2 = C21 0 ¯ C22 and where P solves (65). 2. P22 ˆ 0 < R1 . ¯ ¯ there exist 0 < Q1 ∈ Rn1 ×n1 . k = 1. The objective function of (13) and (66) will then be negative if there exist x(t) and y(t) in Rn that satisfy (70). P2 ∈ satisfy (65) and there exist 0 < P n×n1 ˆ3 ∈ Rn×n of the form and P R ˆ P ˆ P2 = ˆ21 . Q2 ∈ n×n1 n×n ¯ 1 and 0 < R2 ∈ R(n+n1 )×(n+n1 ) that ¯ and Q3 ∈ R of (51a)–(51c). 0 < Sk ∈ Rn1 ×n1 . P32 . q J (w) < 0 for all nonzero w ∈ L2 [0. assumption A2 and the definition of w ∗ in (67) we ﬁnd e1 (t) ˙ ˆ ˆ 0 = [A − Kf C2 ]e(t) − 0 T ˜ z(t) = R −1/2 0 B2 P e(t). The problem now becomes one of ﬁnding the gain matrix Kf that will ensure the ¯ ¯ stability of system (81) and that the H∞ -norm of the transference from r to z is less than γ . Consider the system of (4). B2 (80) Using the notation of (72).E. Fridman. that satisfy the following LMI: .3. U. 0 < Uk ∈ Rn2 ×n2 . ˆ ˆ Consider the following ‘innovation’ ﬁlter ˙ x1 (t) ˆ x 1 (t) ˆ ˆ ˆ 0 = A y(t) − x2 (t) ˆ 0 0 Ai1 i=1 Ai3 2 t t−hi y(s) ds ˆ ˆ ˆ +Kf [y(t) − C2 x(t)] + ¯ 0 u(t). ˆ P22 ∈ Rn2 ×n1 . ∞) if for some prescribed scalars ε1 . (5a). 2. k = 1. 0 < Uk ∈ Rn2 ×n2 . P32 ∈ Rn2 ×n2 . ˆ 0 < R2 ∈ Rn1 ×n1 . 0 < R R ˆ ˆ ˆ1 ∈ Rn1 ×n1 .

¯ where A0 = 0 0 0 .29. A1 = −1 0 −1 .3.2 leads to more conservative results owing to the fact that the adjoint system for e contains e0 with delays and that the results are delay-independent with respect to delays in e0 .300 E. D12 u(t) . P3 are deﬁned by (75) and 0n1 ×n1 ¯ 1 = P T A + AT P − Y C2 − C T Y T + 0 ˆ ˆ ˆ2 ˆ ˆ ˆ 0 2 ˆ i=1 hi Ri 0 0 2 ˆ i=1 Ui 0 0 . ˙ z(t) = col C1 x(t). For h 1 the LMI of Theorem 6. We consider the nondescriptor system of [12] x(t) = A0 x(t) + A1 x(t − h) + B1 w(t) + B2 u(t). A minimum value of γ = 0.3 for the nondescriptor case. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 ˆ PT 0 − Y D21 B1 −γ 2 Iq ∗ ∗ ∗ ∗ ∗ 0 ˆ T A11 h1 P A13 0 ˆ −h1 R1 0 ∗ ∗ ∗ 0 ˆ T A11 h2 P A13 0 0 ˆ −h2 R2 ∗ ∗ ∗ T P2 T P3 ∗ ∗ ∗ ∗ ¯1 B2 ∗ ∗ 0 0 0 ˜ −R ∗ ∗ 0 ˆ T A12 P A14 0 0 0 0 ˆ −U1 ∗ 0 ˆ T A22 P A24 0 0 0 0 0 ˆ −U2 < 0. If a solution to (65) and (82) exists. By the counterpart of Theorem 6. This example shows that for greater values of h Theorem 6. −. Fridman. 0 B2 = 0. ˆ ˆ Example 4. 0 .1 . then the output-feedback controller is obtained ˆ by (77).2 is less conservative due to Park’s inequality [25] that is used for bounding the cross terms.86 was obtained. ˆ the output-feedback control is obtained for h = 1. D21 = 0 We compare the output-feedback controller designs achieved by the two methods presented in Sections 6. For smaller values of h Theorem 6. P1 .2 and 6.3. ε1 = ε1 = −. (82) where P . where x and y are obtained by (80) with Kf = P −T Y . 1 1 0 . 1 (84) (83) C1 = 0 D12 = 0. U. the output-feedback control was derived in [12] for h = 0. P2 . By Theorem 6.3. A minimum value of γ = 20 is then achieved.3 is not feasible for any γ > 0. .2 a greater value of γ = 11 is found for the same value of h and ε1 = ε2 = −0. y(t) = C2 x(t) + D21 w(t). 1 1 .2.999 and ε1 = −0.1.9 C2 = 0 B1 = 1 . By Theorem 6.28.

new LMI conditions for stability and H∞ -control of difference continuous time equations are obtained. while for smaller values of the delay—the second one provides less conservative results. . Conclusions An LMI solution is proposed for the problem of stability and H∞ -control of linear time-invariant descriptor systems. K. Mitter. El Ghaoui.1 are convex in the time delays.C. S. the other applies the BRL directly to the system of the estimation error (and thus generalizes the result of [12] to the descriptor case). V. These methods lead to complementary results: for greater values of the delay the ﬁrst method is less conservative. Balakrishnan. Byers. [2] A. K. As a byproduct.K. Azuma. References [1] T. Nickols. IBM J. E. Birkhäser. 1–2.2 and 6. Both methods suffer from an additional overdesign that stems from the need to estimate the state and its derivative. Bunse-Gerstner. U. [5] A. the latter overdesign can be removed since the estimate of the state (and not of its derivative) is needed. R. L. 12 (1968) 431–440.E. Delfour. G. Mehrmann. Bensoussan. September 1999. Brayton. Inﬁnite-dimensional LMI approach to H∞ -control synthesis for linear systems with time delay.1 and 5. in: Proceedings of the 5th European Control Conference. Fridman.3 in the delay parameters. Feron. Uchida. Karlsruhe. Res. The situation in the output-feedback control case is however different. DoPrato. One question that often arises when solving control and estimation problems for systems with time-delay is whether the solution obtained for certain delays hi will ¯ satisfy the design requirements for all delays hi hi . vol. The design of the output-feedback controller is achieved by two methods: one is based on the BRL for the adjoint of the system that describes the estimation error. Feedback design for regularizing descriptor systems. Ikeda. The LMI sufﬁcient conditions that are obtained allow solutions to the H∞ -control problem in the uncertain case where the system parameters lie within an uncertainty polytope. Boyd. V. Basel. M. [3] S. The fact that the P2 and P3 depend nonlinearly on the delay implies that the output-feedback controller that is derived for a certain delay will not necessarily satisfy the design speciﬁcations for smaller delays. Representation and control of inﬁnite dimensional systems. In the problems of state-feedback and ﬁltering the answer is the afﬁrmative since the LMIs in Theorems 4. Linear matrix inequality in systems and control theory. Develop. in spite of the seemingly convexity of the LMI of Theorems 6. 299 (1999) 119–151. Germany. In the special case where a result is sought which is delay-independent with respect to the process and delay-dependent with respect to observer. Linear Algebra Appl. [4] R. This solution is based on the Lyapunov function approach to descriptor systems with delay introduced in [10] and on the LMI approach to H∞ control of nondescriptor systems of [12]. 1992. SIAM Frontier Series (April 1994). Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 301 7. Small signal stability criterion for electrical networks containing lossless transmission lines. N.

New York. [28] W. vol. J. [12] E. Ohara. X. A Delay-dependent stability criterion for systems with uncertain time-invariant delays. [7] S. Decision and Control. Math. Suda. Singular Control Systems. Appl. Masubuchu. J. N. Dordrecht. [14] J. H∞ control for descriptor systems: a matrix inequalities approach. 1. Hale. Introduction to Functional Differential Equations. Hara. S. 1999. 1993. V. 2-D (differential-delay) implicit systems. 1994. (under review). Morihira. Appl.). Destabilizing effects of small time delays on feedback-controlled descriptor systems. 1991. New Lyapunov–Krasovskii functionals for stability of linear retarded and neutral type systems. pp. [18] V. Springer. in: S. 1999. 14 (1997) 95–107. Kamitane. Pitzold. Y. Kwakernaak. Nonlinear Anal. U. Xu. Logemann. 2251–2256.I. [11] E. [29] H. Kolmanovskii. Kolmanovskii. in: Proceedings of MTNS 2000. 1996. L. Halanay. Lunel. 24 (1997) 247–264. Perpignan. Internat. IEEE Trans. [17] V. in: Proceedings of IMACS World Congress on Scientific Computation. Anal. France. J. . [9] E. Cobe. Shaked. 1989. 1828–1829. Fridman. Rasvan. Dai. [23] V. Appl. Automat. Stability of linear descriptor systems with delay: a Lyapunov-based approach. Rehm. Linear Algebra Appl. Katayama. Asymptotic stability of linear delay differential–algebraic equations and numerical methods. Martinez Amores. Numer. Control 47 (2002). 2000. Control 72 (1999) 374–384. Myshkis. 317–336. Shaked. J. vol. pp. Prague. Anal. [27] K. Springer. IEEE Trans. [8] L. Shaked / Linear Algebra and its Applications 351–352 (2002) 271–302 [6] S. Allgower. Absolute Stability of Time Lag Control Systems. Applied Theory of Functional Differential Equations. Stability radii for some propagation models. [24] C. S. U. Singular linear systems of differential equations with delays. A. Niculescu. Springer. 272 (1998) 131–153. Lecture Notes in Control and Information Sciences. A descriptor system approach to H∞ control of linear time-delay systems.P. Fridman. Descriptor and nondescriptor controllers in H∞ control of descriptor systems. Bucharest. 2000. Delay-independent stability in lossless propagation models with applications. de Souza. Richard. Kluwer Academic Publishers. 1 (1977) 161–172. S. Japan. Automatica 35 (1999) 1313–1321. Automat. J-M. Stability in neutral equations. pp. Dynamical compensation for time-delay systems: an LMI approach. [13] A. Campbell. Niculescu. P. V. Y. Dublin. Control 44 (1999) 876–877.I. [22] S. Robust Nonlinear Control 10 (2000) 611–628. Zhu. Fridman. L. [15] J. Dugard. [16] D. Systems Control Lett. Meth. Yamamoto (Eds. Math. pp. Proceedings of the 38th Conf. J. 1975 (in Romanian). Hale. Dion. Academiei.-I. U. Math. IMA J. Internat. On the Liapunov–Krasovskii functionals for stability analysis of linear delay systems. Appl. Automatica 33 (1997) 669–673. Ivanescu. Berlin. New York. H∞ control for descriptor systems—a J-spectral factorization approach. A. Takaba.E. Fridman. H∞ -state-feedback control of linear systems with small state-delay. Ed. 43 (4) (2001) 309–319. Upper and lower bounds of H∞ -optimal performance for a class of continuous type descriptor systems. Campbell. [10] E. Li. Mizukami. Theory. Park. [25] P. T. 33 (1998) 141–150. F. Rasvan. [26] A. K. N. Systems Control Lett. [19] H. Niculescu. Frequency domain solution of the H∞ problem for descriptor systems.302 E. Learning Control and Hybrid Systems. Proceedings of the 33rd IEEE Conf. 1021–1026. 2 (1980) 129–136. Delay-dependent robust H∞ control of uncertain linear state-delayed systems. 241. Rasvan. Decision and Control. Fridman. Control Inform. [20] H. [21] I. in: Proceedings of ROCOND.

Sign up to vote on this title

UsefulNot useful- ACC_2007
- detectability subspaces
- 1-s2.0-S0021869306002584-main (2)
- An Intrinsic Behavioural Approach to the Gap Metric
- (55-4)379
- On Strongly g(x)-Clean Rings
- Bài báo Toán
- LMI
- dm-intro
- Scientia Magna, Vol. 10, No. 1, 2014
- SCIENTIA MAGNA, book series, Vol. 10, No. 1
- Smarandache isotopy of second Smarandache Bol loops
- Ehrenfeucht-Fraïssé Games - Applications and Complexity - Montanari & Vitacolonna
- International Journal of Engineering Research and Development (IJERD)
- 10.1.1.158
- TRANS_2D
- Algebra Homework Solutions
- hmw9(MA504)
- LU Factorization
- Q1
- Mathematics for high school
- Stability Analysis of Systems With Generalized Frequency Variables
- A Bound on the Scrambling Index of a Primitive Matrix Using Boolean Rank
- Solutions to the Exercises
- Lecture 9
- Simulation Theory
- Problems
- Face Detection
- Time Integration Methods – Still Questions [Bajer, 2002]
- Inyection and surjection.pdf
- Laa

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd