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

**Systems with Polytopic Uncertainties
**

Guisheng Zhai

1∗

Hai Lin

2

and Panos J. Antsaklis

2

1

Faculty of Systems Engineering, Wakayama University

930 Sakaedani, Wakayama 640-8510, Japan

Email: zhai@sys.wakayama-u.ac.jp

Tel:+81(73)457-8187, Fax:+81(73)457-8201

2

Department of Electrical Engineering, University of Notre Dame

Notre Dame, IN 46556, USA

Email: {hlin1, antsaklis.1}@nd.edu

Tel:+1(574)631-5792, Fax:+1(574)631-4393

Abstract: In this paper, we consider quadratic stabilizability via state feedback for

both continuous-time and discrete-time switched linear systems that are composed

of polytopic uncertain subsystems. By state feedback, we mean that the switchings

among subsystems are dependent on system states. For continuous-time switched

linear systems, we show that if there exists a common positive deﬁnite matrix for

stability of all convex combinations of the extreme points which belong to diﬀer-

ent subsystem matrices, then the switched system is quadratically stabilizable via

state feedback. For discrete-time switched linear systems, we derive a quadratic

stabilizability condition expressed as matrix inequalities with respect to a family of

nonnegative scalars and a common positive deﬁnite matrix. For both continuous-

time and discrete-time switched systems, we propose the switching rules by using

the obtained common positive deﬁnite matrix.

Keywords: Continuous-time switched system, discrete-time switched system,

quadratic stabilizability, switching rule, state feedback, matrix inequality.

∗

Corresponding author.

1

1 Introduction

By a switched system, we mean a hybrid dynamical system that is composed of a family of

continuous-time or discrete-time subsystems and a rule orchestrating the switching between

the subsystems. In the last two decades, there has been increasing interest in stability analysis

and control design for switched systems (for example, Branicky, 1994 & 1998; Liberzon

and Morse, 1999; DeCarlo et al., 2000; Zhai, 2001 & 2003; Zhai et al., 2001a & 2002a;

Pettersson and Lennartson, 2002). The motivation for studying switched systems is from

the fact that many practical systems are inherently multimodal in the sense that several

dynamical subsystems are required to describe their behavior which may depend on various

environmental factors (Dayawansa and Martin, 1999; Pettersson and Lennartson, 2002), and

that the methods of intelligent control design are based on the idea of switching between

diﬀerent controllers (Morse, 1996; Liberzon and Morse, 1999; Hu et al., 2000 & 2002; Zhai

et al., 2002b). For recent progress and perspectives in the ﬁeld of switched systems, see the

survey papers Liberzon and Morse (1999), DeCarlo et al. (2000) and the references cited

therein.

As also pointed out in Feron (1996), switched systems have been studied from various

viewpoints. One viewpoint is that the switching signal is an exogenous variable, and then

the problem is to investigate whether there exists a switching signal such that the switched

system has desired performance (stability, certain disturbance attenuation level, etc.). An-

other viewpoint, which is of interest here, is that the switching signal is available to control

engineers, and thus it may be used for control purposes. In the last decade, many practical

methods have been proposed to use available switchings between various modes for control

purposes. In particular, a theoretical framework based on Lyapunov stability theory has been

established in the existing works (for example, Wicks et al., 1994; Branicky, 1994 & 1998).

This paper also assumes that the switching signal can be designed by control engineers,

and the performance index under consideration is quadratic stability of the switched systems.

More precisely, we aim to investigate whether there exists a switching signal such that the

switched system is quadratically stable. It is well known that quadratic stability requires for

uncertain systems a quadratic Lyapunov function which guarantees asymptotical stability for

all uncertainties under consideration, and is thus a kind of robust stability with very good

property, yet usually needs more restrictive condition. However, for complex real systems

with multiple control speciﬁcations, control engineers usually consider quadratic stabilization

so that more design freedom can be gained. For detailed study on quadratic stability and

stabilization, see for example Barmish (1985) and Khargonekar et al. (1990).

There are a few existing results concerning quadratic stabilization of switched linear sys-

2

tems that are composed of several unstable linear time-invariant subsystems. In Wicks et

al. (1994), it has been shown that the existence of a stable convex combination of the sub-

system matrices implies the existence of a state-dependent switching rule that stabilizes the

switched system along with a quadratic Lyapunov function that proves it. It has been proved

in Feron (1996) that when the number of subsystems is two, the existence of a stable convex

combination of the subsystem matrices is necessary and suﬃcient for quadratic stabilizabil-

ity of the switched system by state-dependent switching. An extension to output-dependent

switching for quadratic stability has also been made with a robust detectability condition

in Feron (1996). In Zhai (2001), the results in Wicks et al. (1994) and Feron (1996) have

been extended to the case of discrete-time switched linear systems, by giving a quadratic

stabilizability condition as a nonnegative combination of subsystems’ Lyapunov inequalities.

In that context, a signiﬁcant diﬀerence between continuous-time systems and discrete-time

ones has been pointed out. More precisely, we can easily ﬁnd a stable convex combination

condition of subsystem matrices as in Wicks et al. (1994) for quadratic stabilizability of

continuous-time switched systems, but for discrete-time switched systems we can not derive

such a combination condition without involving a Lyapunov matrix.

Motivated by the results in Wicks et al. (1994), Feron (1996) and Zhai (2001), we consider

in this paper quadratic stabilizability of switched linear systems that are composed of several

uncertain subsystems of polytopic type. It is well known that polytopic uncertainties exist in

many real systems, and most uncertain control systems can be approximated by systems with

polytopic uncertainties. Although there have been many existing results on switched systems

and quadratic stability/stabilization, to the best of our knowledge, there is no existing result

concerning quadratic stabilizability of switched systems with polytopic uncertainties.

Both continuous-time systems and discrete-time ones will be dealt with here. The reasons

of considering discrete-time switched systems have been enumerated in the recent paper

(Zhai et al., 2002a). For example, a multimodal dynamical system may be composed of

several discrete-time dynamical subsystems due to its physical structure, and even when all

subsystems are of continuous-time, the case of considering sampled-data control for the entire

system can be dealt with in the framework of discrete-time switched systems; see Hu and

Michel (2000), Rubensson and Lennartson (2000) for detailed discussions. Furthermore, we

ﬁnd that the extension from continuous-time switched systems to discrete-time ones is not

obvious in most cases, and the results may be quite diﬀerent, as also pointed out in Zhai

(2001) and will be seen later in this paper.

The contribution of the present paper is that, under the assumption that no subsystem

is quadratically stable (otherwise, the switching problem will be trivial by always choosing

the stable subsystem), we derive conditions under which the switched system is quadrati-

cally stabilizable by appropriate state-dependent switching (state feedback). In the case of

3

continuous-time switched system, the condition comes up with a requirement of a common

positive deﬁnite matrix for stability of all convex combinations of the extreme points which

belong to diﬀerent subsystem matrices. An example is used to demonstrate the result. In

the case of discrete-time switched systems, we express the quadratic stabilizability condition

as matrix inequalities with respect to a family of nonnegative scalar variables and a common

positive deﬁnite matrix. However, in discrete-time switched system case we can not regard

the condition as a requirement of a common positive deﬁnite matrix for stability of all convex

combinations of the extreme points. Furthermore, the stabilizability condition for discrete-

time switched systems has to deal with interference among the extreme points that belong to

the same subsystem matrices. These two signiﬁcant diﬀerences make quadratic stabilization

of discrete-time switched systems much more diﬃcult compared with continuous-time ones.

2 Quadratic Stabilizability for Continuous-Time Case

In this section, we consider the continuous-time switched linear system

˙ x(t) = A

σ(x,t)

x(t) , (2.1)

where x(t) ∈

n

is the state, σ(x, t) is a switching rule deﬁned by σ(x, t) :

n

×

+

→{1, 2},

and

+

denotes nonnegative real numbers. Therefore, the switched system is composed of

two continuous-time subsystems

CS

1

: ˙ x(t) = A

1

x(t) , (2.2)

and

CS

2

: ˙ x(t) = A

2

x(t) . (2.3)

Here, we assume that both CS

1

and CS

2

are uncertain systems of polytopic type described

as

A

i

=

N

i

j=1

µ

ij

A

ij

, i = 1, 2 (2.4)

where µ

i

= (µ

i1

, µ

i2

, · · · , µ

iN

i

) belongs to

_

µ

i

:

N

i

j=1

µ

ij

= 1 , µ

ij

≥ 0

_

, (2.5)

and A

ij

, j = 1, 2, · · · , N

i

are constant matrices denoting the extreme points of the polytope

A

i

, and N

i

is the number of the extreme points.

If CS

1

or CS

2

is quadratically stable, we can always activate the stable subsystem so

that the entire switched system is quadratically stable. Therefore, to make the switching

problem nontrivial, we make the following assumption.

4

Assumption 1: Both CS

1

and CS

2

are quadratically unstable, i.e., there does not exist

P

1

> 0 such that

A

T

1i

P

1

+ P

1

A

1i

< 0 , i = 1, 2, · · · , N

1

, (2.6)

and there does not exists P

2

> 0 such that

A

T

2j

P

2

+ P

2

A

2j

< 0 , j = 1, 2, · · · , N

2

. (2.7)

This assumption is obviously true when one of A

1i

’s and one of A

2j

’s are unstable. Fur-

thermore, even if all the matrices A

1i

’s and A

2j

’s are stable, there usually does not exist a

common Lyapunov matrix for them. Narendra and Balakrishnan (1994) showed that when

all A

1i

’s (or A

2j

’s) are stable and commutative pairwise, there exists a common Lyapunov

matrix. Another simple case is that all A

1i

’s (or A

2j

’s) are stable and symmetric (Zhai, 2003).

For most real control systems, this assumption is a reasonable one.

Now, we need the deﬁnition of quadratic stabilizability via state feedback for the switched

system (2.1).

Deﬁnition 1: The system (2.1) is said to be quadratically stabilizable via state feedback if

there exist a positive deﬁnite function V (x) = x

T

Px, a positive number and a switching

rule σ(x, t) depending on x such that

d

dt

V (x) < −x

T

x (2.8)

holds for all trajectories of the system (2.1).

Then, our problem in this section is to ﬁnd a state feedback (state-dependent switching

rule) σ(x, t) such that the switched system (2.1) is quadratically stable. We state and prove

the following main result.

Theorem 1: The switched system (2.1) is quadratically stabilizable via state feedback if there

exist constant scalars λ

ij

’s (i = 1, 2; j = 1, 2, · · · , N

i

) satisfying 0 ≤ λ

ij

≤ 1 and P > 0 such

that

[λ

ij

A

1i

+ (1 −λ

ij

)A

2j

]

T

P + P [λ

ij

A

1i

+ (1 −λ

ij

)A

2j

] < 0 (2.9)

holds for all i = 1, 2; j = 1, 2, · · · , N

i

.

Proof: For the beneﬁt of notation simplicity, we only give the proof in the case of

N

1

= N

2

= 2. The extension from N

1

= N

2

= 2 to general case is very obvious.

From (2.9), we know that there always exists a positive scalar such that

[λ

ij

A

1i

+ (1 −λ

ij

)A

2j

]

T

P + P [λ

ij

A

1i

+ (1 −λ

ij

)A

2j

] < −I . (2.10)

5

Then, for any x = 0, we obtain

x

T

[λ

11

A

11

+ (1 −λ

11

)A

21

]

T

Px + x

T

P [λ

11

A

11

+ (1 −λ

11

)A

21

] x < −x

T

x

x

T

[λ

12

A

11

+ (1 −λ

12

)A

22

]

T

Px + x

T

P [λ

12

A

11

+ (1 −λ

12

)A

22

] x < −x

T

x

x

T

[λ

21

A

12

+ (1 −λ

21

)A

21

]

T

Px + x

T

P [λ

21

A

12

+ (1 −λ

21

)A

21

] x < −x

T

x

x

T

[λ

22

A

12

+ (1 −λ

22

)A

22

]

T

Px + x

T

P [λ

22

A

12

+ (1 −λ

22

)A

22

] x < −x

T

x ,

(2.11)

which can be rewritten as

λ

11

x

T

_

A

T

11

P + PA

11

_

x + (1 −λ

11

)x

T

_

A

T

21

P + PA

21

_

x < −x

T

x (2.12)

λ

12

x

T

_

A

T

11

P + PA

11

_

x + (1 −λ

12

)x

T

_

A

T

22

P + PA

22

_

x < −x

T

x (2.13)

λ

21

x

T

_

A

T

12

P + PA

12

_

x + (1 −λ

21

)x

T

_

A

T

21

P + PA

21

_

x < −x

T

x (2.14)

λ

22

x

T

_

A

T

12

P + PA

12

_

x + (1 −λ

22

)x

T

_

A

T

22

P + PA

22

_

x < −x

T

x . (2.15)

It is very easy to verify from the above inequalities that either

x

T

_

A

T

11

P + PA

11

_

x < −x

T

x , x

T

_

A

T

12

P + PA

12

_

x < −x

T

x (2.16)

or

x

T

_

A

T

21

P + PA

21

_

x < −x

T

x , x

T

_

A

T

22

P + PA

22

_

x < −x

T

x (2.17)

is true. For example, if (2.16) is not true with x

T

_

A

T

11

P + PA

11

_

x ≥ −x

T

x, then we get

x

T

_

A

T

21

P + PA

21

_

x < −x

T

x from (2.12) and x

T

_

A

T

22

P + PA

22

_

x < −x

T

x from (2.13).

The same is true for other cases.

Now, we deﬁne the switching rule as

σ(x, t) ∈

_

i

¸

¸

¸x

T

_

A

T

ij

P + PA

ij

_

x < −x

T

x , j = 1, 2

_

. (2.18)

Then, based on the above discussion, we get

x

T

_

A

T

σ1

P + PA

σ1

_

x < −x

T

x , x

T

_

A

T

σ2

P + PA

σ2

_

x < −x

T

x , (2.19)

and thus

x

T

_

A

T

σ

P + PA

σ

_

x < −x

T

x (2.20)

since A

σ

is a linear convex combination of A

σ1

and A

σ2

. Clearly, (2.20) implies that the

inequality

d

dt

V (x) = x

T

_

A

T

σ

P + PA

σ

_

x < −x

T

x (2.21)

is true for all trajectories of (2.1) and thus the switched system is quadratically stable.

6

Remark 1: In fact, the condition (2.9) requires a common positive deﬁnite matrix for

stability of all convex combinations of the extreme points which belong to diﬀerent subsystem

matrices. When there is no uncertainty in (2.1), i.e., N

1

= N

2

= 1, (2.9) shrinks to a single

matrix inequality

[λ

11

A

11

+ (1 −λ

11

)A

21

]

T

P + P [λ

11

A

11

+ (1 −λ

11

)A

21

] < 0 , (2.22)

which means that λ

11

A

11

+ (1 − λ

11

)A

21

is Hurwitz stable. In Feron (1996), it has been

pointed out that in this case the switched system is quadratically stabilizable if and only if

there exists a positive scalar λ

11

< 1 such that λ

11

A

11

+ (1 − λ

11

)A

21

is Hurwitz stable. In

this sense, Theorem 1 is the extension of Feron (1996).

Remark 2: Although we made Assumption 1 so as to make our switching problem nontrivial,

Theorem 1 has covered the case where Assumption 1 is not true. That is, when CS

1

is

quadratically stable, we can choose λ

ij

= 1 for all i, j in (2.9). On the contrary, when CS

2

is quadratically stable, we can choose λ

ij

= 0 for all i, j in (2.9).

Remark 3: Although we have assumed tacitly in the switched system (2.1) that the number

of subsystems is two, Theorem 1 can be extended to the case of more than two subsystems

in an obvious way. In that case, if the number of subsystems is M and the number of the

extreme points of each A

i

is N

i

, then the condition for quadratic stabilizability is described

as Π

M

j=1

N

j

matrix inequalities with respect to a family of nonnegative scalars and a common

positive deﬁnite matrix. With the increase of subsystem number or extreme point number, the

computation complexity increases and the possibility of quadratic stabilizability decreases.

Remark 4: Obviously, the condition (2.9) consists of bilinear matrix inequalities (BMIs)

with respect to λ

ij

’s and P > 0. As is well known, it is not an easy task to solve BMI.

Though in some special case we can convert a BMI into LMIs (Boyd et al., 1994) which can

be easily solved by many existing softwares (for example, LMI Toolbox (Gahinet et al., 1995)

in Matlab), there is not a globally eﬀective method for general BMI presently. For numerical

computation method of (2.9) in real control problems, we suggest trying the branch and

bound methods proposed in Goh et al. (1994) or the homotopy-based algorithm in Zhai et

al. (2001b). For example, the algorithm proposed in Zhai et al. (2001b) can be outlined as

follows. We ﬁrst ﬁnd an initial solution of

[λ

ij

A

1i

+ (1 −λ

ij

)A

2j

]

T

P + P [λ

ij

A

1i

+ (1 −λ

ij

)A

2j

] < βI (2.23)

for some positive scalar β = β

0

, which is always possible if we ﬁx certain P > 0, λ

ij

’s and

use β

0

large enough. Then, we deform the initial solution and β gradually until β reaches a

nonpositive scalar. At each step of the deformation, we solve the above matrix inequality by

ﬁxing P or λ

ij

’s .

7

In the end of this section, we give an example to demonstrate our result.

Example: Consider the switched linear system (2.1) composed of two subsystems where

A

11

=

_

−1 2

2 −1

_

, A

12

=

_

−1 1

1 −1

_

(2.24)

and

A

21

=

_

−1 −2

−2 −1

_

, A

22

=

_

−1 −1

−1 −1

_

. (2.25)

Since both A

11

and A

12

are unstable, there is not P

1

> 0 satisfying (2.6). Similarly, since

both A

21

and A

22

are unstable, there is not P

2

> 0 satisfying (2.7). Therefore, both CS

1

and

CS

2

are quadratically unstable.

It is easy to observe that

1

2

A

11

+

1

2

A

21

= −I

2

,

1

2

A

12

+

1

2

A

22

= −I

2

1

3

A

11

+

2

3

A

22

= −I

2

,

2

3

A

12

+

1

3

A

21

= −I

2

,

(2.26)

which implies that λ

11

= λ

22

=

1

2

, λ

12

=

1

3

, λ

21

=

2

3

together with a positive deﬁnite matrix

P satisfy the condition (2.9) in Theorem 1. Therefore, the switched linear system with (2.24)

and (2.25) is quadratically stabilizable via state feedback.

−0.5 0 0.5 1 1.5 2 2.5 3

0

1

2

3

4

x

1

x

2

0 0.5 1 1.5 2 2.5 3 3.5 4

0

1

2

3

4

5

Time

N

o

r

m

o

f

X

Figure 1. The state and the state’s norm of the switched system in the example

8

Let us investigate the system state trajectory using the two speciﬁc subsystems

A

1

= 0.2A

11

+ 0.8A

12

=

_

−1.0 1.2

1.2 −1.0

_

,

A

2

= 0.4A

21

+ 0.6A

22

=

_

−1.0 −1.4

−1.4 −1.0

_

,

(2.27)

which belong to CS

1

and CS

2

, respectively, and are both unstable.

We suppose that the initial state is x

0

= [3 4]

T

, and set P = I

2

. Then, since

x

T

0

(A

T

11

P + PA

11

)x

0

= 46 , x

T

0

(A

T

12

P + PA

12

)x

0

= −2 ,

x

T

0

(A

T

21

P + PA

21

)x

0

= −146 , x

T

0

(A

T

22

P + PA

22

)x

0

= −98 ,

(2.28)

we choose σ(x

0

, 0) = 2 according to the switching rule (2.18).

In the same way, we choose the subsystem mode for every time instant using the switching

rule (2.18) and evolve the system forth on. Although both A

1

and A

2

are unstable here, Figure

1 shows that both the system state and the norm of the state converge to zero very quickly

under the switching rule we proposed.

3 Quadratic Stabilizability for Discrete-Time Case

In this section, we consider the discrete-time switched linear system

x[k + 1] = A

σ(x,k)

x[k] , (3.1)

where x[k] ∈

n

is the state, σ(x, k) is a switching rule deﬁned by σ(x, k) :

n

×N

+

→{1, 2},

and N

+

denotes nonnegative integers. Therefore, the switched system is composed of two

discrete-time subsystems

DS

1

: x[k + 1] = A

1

x[k] , (3.2)

and

DS

2

: x[k + 1] = A

2

x[k] . (3.3)

As in Section 2, we assume here that both DS

1

and DS

2

are uncertain systems of polytopic

type described as (2.4) and (2.5), and make the following assumption.

Assumption 2: Both DS

1

and DS

2

are quadratically unstable, i.e., there does not exist

P

1

> 0 such that

A

T

1i

P

1

A

1i

−P

1

< 0 , i = 1, 2, · · · , N

1

, (3.4)

and there does not exist P

2

> 0 such that

A

T

2j

P

2

A

2j

−P

2

< 0 , j = 1, 2, · · · , N

2

. (3.5)

9

Deﬁnition 2: The system (3.1) is said to be quadratically stabilizable via state feedback

if there exist a positive deﬁnite function V (x[k]) = x

T

[k]Px[k], a positive number and a

switching rule σ(x, k) depending on x such that

V (x[k + 1]) −V (x[k]) < −x

T

[k]x[k] (3.6)

for all trajectories of the system (3.1).

Then, the problem in this section is to ﬁnd a state feedback (state-dependent switching

rule) σ(x, k) such that the switched system (3.1) is quadratically stable.

Theorem 2: The switched system (3.1) is quadratically stabilizable via state feedback if there

exist constant scalars λ

ij

’s (i = 1, 2; j = 1, 2, · · · , N

i

) satisfying 0 ≤ λ

ij

≤ 1 and P > 0 such

that

λ

ij

(A

T

1i

PA

1i

−P) + (1 −λ

ij

)(A

T

2j

PA

2j

−P) < 0 , i = 1, 2; j = 1, 2, · · · , N

i

(3.7)

A

T

1k

PA

1l

+ A

T

1l

PA

1k

−2P ≤ 0 , k, l = 1, 2, · · · , N

1

, k = l (3.8)

A

T

2u

PA

2v

+ A

T

2v

PA

2u

−2P ≤ 0 , u, v = 1, 2, · · · , N

2

, u = v . (3.9)

Proof: As in the proof of Theorem 1, we only give the proof in the case of N

1

= N

2

= 2.

From (3.7), we know that there always exists a positive scalar such that

λ

ij

(A

T

1i

PA

1i

−P) + (1 −λ

ij

)(A

T

2j

PA

2j

−P) < −I . (3.10)

Then, for any x[k] = 0, we obtain from (3.10) that

λ

ij

x

T

[k]

_

A

T

1i

PA

1i

−P

_

x[k] + (1 −λ

ij

)x

T

[k]

_

A

T

2j

PA

2j

−P

_

x[k] < −x

T

[k]x[k] . (3.11)

It is very easy to verify that either

_

_

_

x

T

[k]

_

A

T

11

PA

11

−P

_

x[k] < −x

T

[k]x[k]

x

T

[k]

_

A

T

12

PA

12

−P

_

x[k] < −x

T

[k]x[k]

(3.12)

or _

_

_

x

T

[k]

_

A

T

21

PA

21

−P

_

x[k] < −x

T

[k]x[k]

x

T

[k]

_

A

T

22

PA

22

−P

_

x[k] < −x

T

[k]x[k]

(3.13)

is true. From (3.8) and (3.9), we obtain that

x

T

[k](A

T

11

PA

12

−P)x[k] ≤ 0 , x

T

[k](A

T

21

PA

22

−P)x[k] ≤ 0 . (3.14)

10

Now, we deﬁne the switching rule as

σ(x, k) ∈

_

i

¸

¸

¸x

T

[k]

_

A

T

ij

PA

ij

−P

_

x[k] < −x

T

[k]x[k] , j = 1, 2

_

. (3.15)

Thus, we obtain from (3.12)-(3.15) that

x

T

[k]

_

A

T

σ

PA

σ

−P

_

x[k] < −x

T

[k]x[k] . (3.16)

This implies that the inequality

V (x[k + 1]) −V (x[k]) = x

T

[k]

_

A

T

σ

PA

σ

−P

_

x[k] < −x

T

[k]x[k] (3.17)

is true for all trajectories of (3.1) and thus the switched system is quadratically stable.

Remark 5: Although we can regard (3.7) in Theorem 2 as a parallel condition of (2.9)

in Theorem 1, there is essential diﬀerence between them. (2.9) represents a stable convex

combination of A

1i

and A

2j

, but (3.7) does not. From this observation we know that the

extension from continuous-time switched systems to discrete-time ones is far from trivial.

Remark 6: From Theorem 2, we see that discrete-time switched systems are much more

diﬃcult to deal with than continuous-time ones. The reason is that the Lyapunov matrix in-

equality A

T

PA−P < 0 for the discrete-time system x[k+1] = Ax[k] is not linear with respect

to A. When there is no switching, we can convert A

T

PA − P < 0 into

_

P A

T

P

PA P

_

> 0 ,

which is linear with respect to both A and P. By this change, we can deal with polytopic

perturbations in A. However, when there is switching, we can not make such manipulation

for x

T

(A

T

PA−P)x < 0 (x = 0). For this reason, we have to take care of x

T

(A

T

1k

PA

1l

−P)x

(k = l) and x

T

(A

T

2u

PA

2v

− P)x (u = v) for any x = 0. It is known from (3.14) that (3.8)

and (3.9) are proposed for this purpose, i.e., to deal with the interference among the extreme

points that belong to the same subsystem matrices.

4 Concluding Remarks

In this paper, we have considered quadratic stabilizability via state feedback for both continuous-

time and discrete-time switched linear systems that are composed of polytopic uncertain sub-

systems. For continuous-time switched linear systems, we have shown that if there exists a

common positive deﬁnite matrix for stability of all convex combinations of the extreme points

which belong to diﬀerent subsystem matrices, then the switched system is quadratically sta-

bilizable via state feedback. For discrete-time switched linear systems, we have derived a

quadratic stabilizability condition expressed as matrix inequalities with respect to a family

of nonnegative scalars and a common positive deﬁnite matrix. For both continuous-time

11

and discrete-time switched systems, we have given the switching rules by using the common

positive deﬁnite matrix.

There are several important issues which should be studied in future work. First, to solve

the proposed matrix inequalities (2.9) for continuous-time switched linear systems, one may

ﬁrst consider how to ﬁnd the stable convex combinations of the extreme points, as we did in

the example. This problem is still open for general case, but some geometric method dealing

with interval matrices of constant matrices may be eﬀective. For discrete-time switched linear

systems, the stabilizability conditions (3.8) and (3.9) are quite conservative, and thus need

to be relaxed. An extension to the present paper is to consider quadratic stabilizability via

output feedback (output-dependent switching) for uncertain switched systems. To the best

of our knowledge, there are very few results concerning output-dependent switching, though

it is a practical problem in real control systems. For this issue, some of the ideas in Feron

(1996) and Zhai (2001) should be useful.

Acknowledgements

The authors would like to thank the editor and the anonymous reviewers for their detailed

comments which greatly contributed to this paper. This work has been supported in part by

Japanese Society for the Promotion of Science under the Grant-in-Aid for Encouragement of

Young Scientists 11750396, and by the National Science Foundation (NSF ECS99-12458 &

CCR01-13131).

References

Barmish, B. R., 1985, Necessary and suﬃcient conditions for quadratic stabilizability of an

uncertain linear system. Journal on Optimization Theory and Applications, 46, 399–408.

Boyd, S., El Ghaoui, L., Feron, E., and Balakrishnan, V., 1994, Linear Matrix Inequalities

in System and Control Theory (Philadelphia, PA: SIAM).

Branicky, M. S., 1994, Stability of switched and hybrid systems. In Proceedings of the 33rd

IEEE Conference on Decision and Control, pp. 3498–3503.

Branicky, M. S., 1998, Multiple Lyapunov functions and other analysis tools for switched and

hybrid systems. IEEE Transactions on Automatic Control, 43, 475–482.

Dayawansa, W. P., and Martin, C. F., 1999, A converse Lyapunov theorem for a class of

dynamical systems which undergo switching. IEEE Transactions on Automatic Control,

44, 751–760.

12

DeCarlo, R. A., Branicky, M. S., Pettersson, S., and Lennartson, B., 2000, Perspectives and

results on the stability and stabilizability of hybrid systems. Proceedings of the IEEE, 88,

1069–1082.

Feron, E., 1996, Quadratic stability of switched systems via state and output feedback.

Technical report CICSP-468 (MIT).

Gahinet, G., Nemirovski, A., Laub, A. J., and Chilali, M., 1994, LMI Control Toolbox for

Use with Matlab (the MathWorks Inc.).

Goh, K. C., Safonov, M. G., and Papavassilopoulos, G. P., 1994, A global optimization

approach for the BMI problem. In Proceedings of the 33rd IEEE Conference on Decision

and Control, pp. 2009–2014.

Hu, B., and Michel, A. N., 2000, Stability analysis of digital feedback control systems with

time-varying sampling periods. Automatica, 36, 897–905.

Hu, B., Zhai, G., and Michel, A. N., 2000, Hybrid output feedback stabilization of two-

dimensional linear control systems. In Proceedings of the 2000 American Control Confer-

ence, pp. 2184–2188.

Hu, B., Zhai, G., and Michel, A. N., 2002, Stabilizing a class of two-dimensional bilinear sys-

tems with ﬁnite-state hybrid constant feedback. International Journal of Hybrid Systems,

2, 189–205.

Khargonekar, P. P., Petersen, I. R., and Zhou, K., 1990, Robust stabilization of uncertain

linear systems: Quadratic stabilizability and H

∞

control theory. IEEE Transactions on

Automatic Control, AC-35, 356–361.

Liberzon, D., and Morse, A. S., 1999, Basic problems in stability and design of switched

systems. IEEE Control Systems Magazine, 19, 59–70.

Morse, A. S., 1996, Supervisory control of families of linear set-point controllers-Part 1: Exact

matching. IEEE Transactions on Automatic Control, 41, 1413–1431.

Narendra, K. S., and Balakrishnan, V., 1994, A common Lyapunov function for stable LTI

systems with commuting A-matrices. IEEE Transactions on Automatic Control, 39, 2469–

2471.

Pettersson, S., and Lennartson, B., 2002, Hybrid system stability and robust veriﬁcation

using linear matrix inequalities. International Journal of Control, 75, 1335–1355.

13

Rubensson, M., and Lennartson, B., 2000, Stability and robustness of hybrid systems using

discrete-time Lyapunov techniques. In Proceedings of the 2000 American Control Confer-

ence, pp. 210–214.

Wicks, M. A., Peleties, P., and DeCarlo, R. A., 1994, Construction of piecewise Lyapunov

function for stabilizing switched systems. In Proceedings of the 35th IEEE Conference on

Decision and Control, pp. 3492–3497.

Zhai, G., 2001, Quadratic stabilizability of discrete-time switched systems via state and

output feedback. In Proceedings of the 40th IEEE Conference on Decision and Control,

pp. 2165–2166.

Zhai, G., 2003, Stability and L

2

gain analysis for switched symmetric systems, in D. Liu and P.

J. Antsaklis (eds) Contemporary Issues in Systems Stability and Control with Applications

(Boston, MA: Birkh¨ auser), Chapter 7, pp. 143–166.

Zhai, G., Hu, B., Yasuda, K., and Michel, A. N., 2001a, Disturbance attenuation properties

of time-controlled switched systems. Journal of The Franklin Institute, 338, 765–779.

Zhai, G., Hu, B., Yasuda, K., and Michel, A. N., 2002a, Stability and L

2

gain analysis

of discrete-time switched systems. Transactions of the Institute of Systems, Control and

Information Engineers, 15, 117–125.

Zhai, G., Ikeda, M., and Fujisaki, Y., 2001b, Decentralized H

∞

controller design: A matrix

inequality approach using a homotopy method. Automatica, 37, 565–572.

Zhai, G., Takai, S., Michel, A. N., and Xu, X., 2002b, Improving closed-loop stability of

second-order LTI systems by hybrid static output feedback. In Proceedings of the 41st

IEEE Conference on Decision and Control, pp. 2792–2797.

14

. Pettersson and Lennartson. This paper also assumes that the switching signal can be designed by control engineers. In the last decade. One viewpoint is that the switching signal is an exogenous variable.). control engineers usually consider quadratic stabilization so that more design freedom can be gained. and thus it may be used for control purposes. Branicky. Liberzon and Morse. we aim to investigate whether there exists a switching signal such that the switched system is quadratically stable. Another viewpoint. for complex real systems with multiple control speciﬁcations. More precisely. 2001 & 2003. 2002). Hu et al. (2000) and the references cited therein. For detailed study on quadratic stability and stabilization. see the survey papers Liberzon and Morse (1999). Branicky. a theoretical framework based on Lyapunov stability theory has been established in the existing works (for example. 1999. certain disturbance attenuation level. Liberzon and Morse. 2001a & 2002a. see for example Barmish (1985) and Khargonekar et al.. Pettersson and Lennartson. There are a few existing results concerning quadratic stabilization of switched linear sys- 2 . 1999. yet usually needs more restrictive condition. is that the switching signal is available to control engineers. The motivation for studying switched systems is from the fact that many practical systems are inherently multimodal in the sense that several dynamical subsystems are required to describe their behavior which may depend on various environmental factors (Dayawansa and Martin. (1990). Zhai et al.. In the last two decades. which is of interest here. 1996. 2000. Wicks et al. and is thus a kind of robust stability with very good property. For recent progress and perspectives in the ﬁeld of switched systems. 1999. DeCarlo et al. we mean a hybrid dynamical system that is composed of a family of continuous-time or discrete-time subsystems and a rule orchestrating the switching between the subsystems. However. etc. switched systems have been studied from various viewpoints. 2000 & 2002. many practical methods have been proposed to use available switchings between various modes for control purposes. there has been increasing interest in stability analysis and control design for switched systems (for example.. Zhai. 1994 & 1998. Zhai et al.1 Introduction By a switched system. 2002). and then the problem is to investigate whether there exists a switching signal such that the switched system has desired performance (stability. DeCarlo et al. 2002b). It is well known that quadratic stability requires for uncertain systems a quadratic Lyapunov function which guarantees asymptotical stability for all uncertainties under consideration. 1994. As also pointed out in Feron (1996). and the performance index under consideration is quadratic stability of the switched systems. 1994 & 1998). and that the methods of intelligent control design are based on the idea of switching between diﬀerent controllers (Morse. In particular..

. the switching problem will be trivial by always choosing the stable subsystem). Rubensson and Lennartson (2000) for detailed discussions. The reasons of considering discrete-time switched systems have been enumerated in the recent paper (Zhai et al. Furthermore. More precisely. (1994). we derive conditions under which the switched system is quadratically stabilizable by appropriate state-dependent switching (state feedback). there is no existing result concerning quadratic stabilizability of switched systems with polytopic uncertainties. we consider in this paper quadratic stabilizability of switched linear systems that are composed of several uncertain subsystems of polytopic type. as also pointed out in Zhai (2001) and will be seen later in this paper. An extension to output-dependent switching for quadratic stability has also been made with a robust detectability condition in Feron (1996). In that context. the results in Wicks et al. it has been shown that the existence of a stable convex combination of the subsystem matrices implies the existence of a state-dependent switching rule that stabilizes the switched system along with a quadratic Lyapunov function that proves it.tems that are composed of several unstable linear time-invariant subsystems. a signiﬁcant diﬀerence between continuous-time systems and discrete-time ones has been pointed out. (1994) and Feron (1996) have been extended to the case of discrete-time switched linear systems. In the case of 3 . Feron (1996) and Zhai (2001). see Hu and Michel (2000). Although there have been many existing results on switched systems and quadratic stability/stabilization. Motivated by the results in Wicks et al. It has been proved in Feron (1996) that when the number of subsystems is two. It is well known that polytopic uncertainties exist in many real systems. and even when all subsystems are of continuous-time. In Zhai (2001). and most uncertain control systems can be approximated by systems with polytopic uncertainties. 2002a). In Wicks et al. The contribution of the present paper is that. we ﬁnd that the extension from continuous-time switched systems to discrete-time ones is not obvious in most cases. but for discrete-time switched systems we can not derive such a combination condition without involving a Lyapunov matrix. under the assumption that no subsystem is quadratically stable (otherwise. the existence of a stable convex combination of the subsystem matrices is necessary and suﬃcient for quadratic stabilizability of the switched system by state-dependent switching. Both continuous-time systems and discrete-time ones will be dealt with here. (1994) for quadratic stabilizability of continuous-time switched systems. (1994). and the results may be quite diﬀerent. the case of considering sampled-data control for the entire system can be dealt with in the framework of discrete-time switched systems. by giving a quadratic stabilizability condition as a nonnegative combination of subsystems’ Lyapunov inequalities. For example. a multimodal dynamical system may be composed of several discrete-time dynamical subsystems due to its physical structure. to the best of our knowledge. we can easily ﬁnd a stable convex combination condition of subsystem matrices as in Wicks et al.

2. the condition comes up with a requirement of a common positive deﬁnite matrix for stability of all convex combinations of the extreme points which belong to diﬀerent subsystem matrices. we express the quadratic stabilizability condition as matrix inequalities with respect to a family of nonnegative scalar variables and a common positive deﬁnite matrix. These two signiﬁcant diﬀerences make quadratic stabilization of discrete-time switched systems much more diﬃcult compared with continuous-time ones. t) : n × + → {1. the stabilizability condition for discretetime switched systems has to deal with interference among the extreme points that belong to the same subsystem matrices. in discrete-time switched system case we can not regard the condition as a requirement of a common positive deﬁnite matrix for stability of all convex combinations of the extreme points. In the case of discrete-time switched systems. we assume that both CS1 and CS2 are uncertain systems of polytopic type described as Ni x(t) = A1 x(t) . i = 1.1) where x(t) ∈ n is the state. σ(x. However. ˙ (2. · · · . ˙ (2. j = 1. we make the following assumption. 2 Quadratic Stabilizability for Continuous-Time Case In this section. If CS1 or CS2 is quadratically stable.4) where µi = (µi1 .3) Here. the switched system is composed of two continuous-time subsystems CS1 : and CS2 : x(t) = A2 x(t) . An example is used to demonstrate the result. Therefore. and Ni is the number of the extreme points. we consider the continuous-time switched linear system x(t) = Aσ(x. Therefore. 4 .t) x(t) . Ni are constant matrices denoting the extreme points of the polytope Ai . t) is a switching rule deﬁned by σ(x. and + denotes nonnegative real numbers.5) and Aij . to make the switching problem nontrivial. µij ≥ 0 . Furthermore. (2. µi2 .continuous-time switched system. µiNi ) belongs to Ni µi : j=1 µij = 1 .2) Ai = j=1 µij Aij . ˙ (2. we can always activate the stable subsystem so that the entire switched system is quadratically stable. 2}. 2 (2. · · · .

Assumption 1: Both CS1 and CS2 are quadratically unstable. there usually does not exist a common Lyapunov matrix for them.6) 1i and there does not exists P2 > 0 such that AT P2 + P2 A2j < 0 . i.. t) depending on x such that d V (x) < − xT x dt holds for all trajectories of the system (2. i = 1. 2. N2 . Narendra and Balakrishnan (1994) showed that when all A1i ’s (or A2j ’s) are stable and commutative pairwise. Proof: For the beneﬁt of notation simplicity. there does not exist P1 > 0 such that AT P1 + P1 A1i < 0 .1). Then. even if all the matrices A1i ’s and A2j ’s are stable. (2. Theorem 1: The switched system (2.1). j = 1.e.1) is quadratically stable. (2. 5 . Deﬁnition 1: The system (2.1) is quadratically stabilizable via state feedback if there exist constant scalars λij ’s (i = 1. a positive number and a switching rule σ(x. there exists a common Lyapunov matrix.10) (2. 2.7) This assumption is obviously true when one of A1i ’s and one of A2j ’s are unstable. we need the deﬁnition of quadratic stabilizability via state feedback for the switched system (2. 2. this assumption is a reasonable one. · · · . Furthermore. t) such that the switched system (2. For most real control systems. we only give the proof in the case of N1 = N2 = 2. From (2. 2. we know that there always exists a positive scalar such that (2. 2003). Ni .9) holds for all i = 1. our problem in this section is to ﬁnd a state feedback (state-dependent switching rule) σ(x. Ni ) satisfying 0 ≤ λij ≤ 1 and P > 0 such that [λij A1i + (1 − λij )A2j ]T P + P [λij A1i + (1 − λij )A2j ] < 0 (2.1) is said to be quadratically stabilizable via state feedback if there exist a positive deﬁnite function V (x) = xT P x. N1 . We state and prove the following main result. The extension from N1 = N2 = 2 to general case is very obvious. 2j j = 1.8) [λij A1i + (1 − λij )A2j ]T P + P [λij A1i + (1 − λij )A2j ] < − I . · · · . Another simple case is that all A1i ’s (or A2j ’s) are stable and symmetric (Zhai.9). 2. 2. · · · . j = 1. · · · . Now.

t) ∈ i xT AT P + P Aij x < − xT x .21) σ dt is true for all trajectories of (2.16) is not true with xT AT P + P A11 x ≥ − xT x.16) (2. for any x = 0. which can be rewritten as λ11 xT AT P + P A11 x + (1 − λ11 )xT AT P + P A21 x < − xT x 11 21 λ12 xT AT P + P A11 x + (1 − λ12 )xT AT P + P A22 x < − xT x 11 22 λ21 xT AT P + P A12 x + (1 − λ21 )xT AT P + P A21 x < − xT x 12 21 λ22 xT AT P + P A12 x + (1 − λ22 )xT AT P + P A22 x < − xT x . ij Then. we obtain xT [λ11 A11 + (1 − λ11 )A21 ]T P x + xT P [λ11 A11 + (1 − λ11 )A21 ] x < − xT x xT [λ12 A11 + (1 − λ12 )A22 ]T P x + xT P [λ12 A11 + (1 − λ12 )A22 ] x < − xT x xT [λ21 A12 + (1 − λ21 )A21 ]T P x + xT P [λ21 A12 + (1 − λ21 )A21 ] x < − xT x xT [λ22 A12 + (1 − λ22 )A22 ]T P x + xT P [λ22 A12 + (1 − λ22 )A22 ] x < − xT x .13) (2. Clearly.20) implies that the inequality d V (x) = xT AT P + P Aσ x < − xT x (2.19) j = 1.Then. 21 xT AT P + P A22 x < − xT x 22 (2.14) (2. For example. we deﬁne the switching rule as σ(x.18) xT AT P + P A12 x < − xT x 12 (2.11) 6 . based on the above discussion. if (2.17) is true. xT AT P + P Aσ2 x < − xT x . we get xT AT P + P Aσ1 x < − xT x .15) (2.12) and x AT P + P A22 x < − xT x from (2.20) since Aσ is a linear convex combination of Aσ1 and Aσ2 . σ2 (2. (2.1) and thus the switched system is quadratically stable. 12 22 It is very easy to verify from the above inequalities that either xT AT P + P A11 x < − xT x . then we get 11 T T T T x A21 P + P A21 x < − x x from (2. 22 The same is true for other cases. 11 or xT AT P + P A21 x < − xT x . σ1 and thus xT AT P + P Aσ x < − xT x σ (2. Now. 2 .12) (2. (2.13).

which is always possible if we ﬁx certain P > 0. That is. the condition (2. the algorithm proposed in Zhai et al. Remark 3: Although we have assumed tacitly in the switched system (2. λij ’s and use β0 large enough. Remark 2: Although we made Assumption 1 so as to make our switching problem nontrivial. For example..e.9). N1 = N2 = 1. Then. 1994) which can be easily solved by many existing softwares (for example. we solve the above matrix inequality by ﬁxing P or λij ’s . we suggest trying the branch and bound methods proposed in Goh et al. On the contrary.9) shrinks to a single matrix inequality [λ11 A11 + (1 − λ11 )A21 ]T P + P [λ11 A11 + (1 − λ11 )A21 ] < 0 . In Feron (1996). (2001b).. As is well known. Theorem 1 is the extension of Feron (1996). it has been pointed out that in this case the switched system is quadratically stabilizable if and only if there exists a positive scalar λ11 < 1 such that λ11 A11 + (1 − λ11 )A21 is Hurwitz stable. we can choose λij = 1 for all i. Theorem 1 can be extended to the case of more than two subsystems in an obvious way.22) which means that λ11 A11 + (1 − λ11 )A21 is Hurwitz stable. the condition (2. (1994) or the homotopy-based algorithm in Zhai et al.. 1995) in Matlab). then the condition for quadratic stabilizability is described as ΠM Nj matrix inequalities with respect to a family of nonnegative scalars and a common j=1 positive deﬁnite matrix. when CS1 is quadratically stable. the computation complexity increases and the possibility of quadratic stabilizability decreases. j in (2. (2001b) can be outlined as follows. Theorem 1 has covered the case where Assumption 1 is not true.9) consists of bilinear matrix inequalities (BMIs) with respect to λij ’s and P > 0.23) for some positive scalar β = β0 . when CS2 is quadratically stable. (2.1) that the number of subsystems is two. For numerical computation method of (2.9) requires a common positive deﬁnite matrix for stability of all convex combinations of the extreme points which belong to diﬀerent subsystem matrices. we deform the initial solution and β gradually until β reaches a nonpositive scalar. we can choose λij = 0 for all i. At each step of the deformation. With the increase of subsystem number or extreme point number. Remark 4: Obviously. it is not an easy task to solve BMI. there is not a globally eﬀective method for general BMI presently. 7 . When there is no uncertainty in (2. i.1). j in (2.9) in real control problems.Remark 1: In fact. LMI Toolbox (Gahinet et al. Though in some special case we can convert a BMI into LMIs (Boyd et al. if the number of subsystems is M and the number of the extreme points of each Ai is Ni . In that case.9). (2. In this sense. We ﬁrst ﬁnd an initial solution of [λij A1i + (1 − λij )A2j ]T P + P [λij A1i + (1 − λij )A2j ] < βI (2.

5 3 3. we give an example to demonstrate our result.25) Since both A11 and A12 are unstable.5 1 1. Therefore. A12 = −1 1 1 −1 −1 −1 −1 −1 (2. (2. It is easy to observe that 1A + 2 11 1A + 3 11 1 A = −I . both CS1 and CS2 are quadratically unstable. 4 3 x2 2 1 0 −0.9) in Theorem 1.7). λ12 = 3 . 2 2 21 2 A = −I . the switched linear system with (2. 2 3 22 1A + 2 12 2A + 3 12 1 A = −I 2 2 22 1 A = −I .6).26) 1 1 2 which implies that λ11 = λ22 = 2 . The state and the state’s norm of the switched system in the example 8 . there is not P2 > 0 satisfying (2.In the end of this section. A22 = . λ21 = 3 together with a positive deﬁnite matrix P satisfy the condition (2.25) is quadratically stabilizable via state feedback. Example: Consider the switched linear system (2.5 2 Time 2.5 3 5 4 Norm of X 3 2 1 0 0 0. Similarly. Therefore.5 0 0. there is not P1 > 0 satisfying (2.5 2 2. since both A21 and A22 are unstable.5 1 x1 1. 2 3 21 (2.24) and (2.1) composed of two subsystems where A11 = and A21 = −1 2 2 −1 −1 −2 −2 −1 .5 4 Figure 1.24) .

Then. · · · . respectively. the switched system is composed of two discrete-time subsystems DS1 : x[k + 1] = A1 x[k] .2 −1. (2.4A21 + 0. 0 11 xT (AT P + P A21 )x0 = −146 .3) As in Section 2. · · · .28) we choose σ(x0 .4) 1i and there does not exist P2 > 0 such that AT P2 A2j − P2 < 0 . (3.6A22 = −1. 0 21 xT (AT P + P A12 )x0 = −2 . 2j j = 1. 0 12 xT (AT P + P A22 )x0 = −98 . we consider the discrete-time switched linear system x[k + 1] = Aσ(x.0 −1. i = 1. (3. Therefore. since xT (AT P + P A11 )x0 = 46 .0 1.4) and (2. and set P = I2 .k) x[k] . 3 Quadratic Stabilizability for Discrete-Time Case In this section.1) where x[k] ∈ n is the state. In the same way.e. there does not exist P1 > 0 such that AT P1 A1i − P1 < 0 .4 −1.5). Assumption 2: Both DS1 and DS2 are quadratically unstable. (3.27) .Let us investigate the system state trajectory using the two speciﬁc subsystems A1 = 0. we choose the subsystem mode for every time instant using the switching rule (2. and N + denotes nonnegative integers. 2.0 .2A11 + 0. 2. N2 . 0 22 (2. which belong to CS1 and CS2 . and make the following assumption. We suppose that the initial state is x0 = [3 4]T .2 1. we assume here that both DS1 and DS2 are uncertain systems of polytopic type described as (2.. N1 .8A12 = A2 = 0. (3. 2}.2) and DS2 : x[k + 1] = A2 x[k] . Figure 1 shows that both the system state and the norm of the state converge to zero very quickly under the switching rule we proposed. (3. k) is a switching rule deﬁned by σ(x. σ(x.18). Although both A1 and A2 are unstable here. k) : n ×N + → {1.5) 9 . and are both unstable. i.4 −1.0 −1. 0) = 2 according to the switching rule (2.18) and evolve the system forth on.

(3. 2.10) that λij xT [k] AT P A1i − P x[k] + (1 − λij )xT [k] AT P A2j − P x[k] < − xT [k]x[k] . 2u 2v i = 1. for any x[k] = 0. we know that there always exists a positive scalar such that (3.7).13) is true. l = 1.11) 1i 2j It is very easy to verify that either xT [k] AT P A11 − P x[k] < − xT [k]x[k] 11 or xT [k] AT P A12 − P x[k] < − xT [k]x[k] 12 (3. k = l u. u = v . we obtain from (3. Ni ) satisfying 0 ≤ λij ≤ 1 and P > 0 such that λij (AT P A1i − P ) + (1 − λij )(AT P A2j − P ) < 0 . the problem in this section is to ﬁnd a state feedback (state-dependent switching rule) σ(x. k) such that the switched system (3.6) k.8) and (3.1).1) is quadratically stable. N2 . k) depending on x such that V (x[k + 1]) − V (x[k]) < − xT [k]x[k] for all trajectories of the system (3. Then. From (3. Ni (3. 2. 11 xT [k](AT P A22 − P )x[k] ≤ 0 . · · · . From (3. v = 1. j = 1. 2.1) is quadratically stabilizable via state feedback if there exist constant scalars λij ’s (i = 1. 21 (3. 2. 1i 2j Then.8) (3.14) 10 .Deﬁnition 2: The system (3. 1i 2j AT P A1l + AT P A1k − 2P ≤ 0 .1) is said to be quadratically stabilizable via state feedback if there exist a positive deﬁnite function V (x[k]) = xT [k]P x[k]. Theorem 2: The switched system (3. we obtain that xT [k](AT P A12 − P )x[k] ≤ 0 . j = 1. · · · . we only give the proof in the case of N1 = N2 = 2. a positive number and a switching rule σ(x. 2. 2. N1 .10) λij (AT P A1i − P ) + (1 − λij )(AT P A2j − P ) < − I .12) T x [k] AT P A21 − P x[k] < − xT [k]x[k] 21 x [k] T AT P A22 22 − P x[k] < − x [k]x[k] T (3.9). · · · . · · · . 1k 1l AT P A2v + AT P A2u − 2P ≤ 0 .9) (3.7) (3. Proof: As in the proof of Theorem 1.

9) are proposed for this purpose. there is essential diﬀerence between them. However. we have considered quadratic stabilizability via state feedback for both continuoustime and discrete-time switched linear systems that are composed of polytopic uncertain subsystems. 4 Concluding Remarks In this paper. we have shown that if there exists a common positive deﬁnite matrix for stability of all convex combinations of the extreme points which belong to diﬀerent subsystem matrices. (2. k) ∈ i xT [k] AT P Aij − P x[k] < − xT [k]x[k] . 2 . then the switched system is quadratically stabilizable via state feedback. Remark 6: From Theorem 2.15) that xT [k] AT P Aσ − P x[k] < − xT [k]x[k] .17) (3. For discrete-time switched linear systems. but (3. Remark 5: Although we can regard (3. For this reason.7) in Theorem 2 as a parallel condition of (2. For both continuous-time 11 .8) and (3. we can not make such manipulation for xT (AT P A − P )x < 0 (x = 0). For continuous-time switched linear systems. we can convert AT P A − P < 0 into > 0. we can deal with polytopic perturbations in A. to deal with the interference among the extreme points that belong to the same subsystem matrices. σ This implies that the inequality V (x[k + 1]) − V (x[k]) = xT [k] AT P Aσ − P x[k] < − xT [k]x[k] σ (3.9) represents a stable convex combination of A1i and A2j . we have to take care of xT (AT P A1l − P )x 1k T T (k = l) and x (A2u P A2v − P )x (u = v) for any x = 0.9) in Theorem 1.14) that (3. we see that discrete-time switched systems are much more diﬃcult to deal with than continuous-time ones. From this observation we know that the extension from continuous-time switched systems to discrete-time ones is far from trivial. It is known from (3. we have derived a quadratic stabilizability condition expressed as matrix inequalities with respect to a family of nonnegative scalars and a common positive deﬁnite matrix.1) and thus the switched system is quadratically stable.7) does not.15) is true for all trajectories of (3. we deﬁne the switching rule as σ(x. j = 1. i. we obtain from (3.16) (3. By this change.e..Now.12)-(3. PA P which is linear with respect to both A and P . When there is no switching. The reason is that the Lyapunov matrix inequality AT P A−P < 0 for the discrete-time system x[k+1] = Ax[k] is not linear with respect P AT P to A. ij Thus. when there is switching.

PA: SIAM). Multiple Lyapunov functions and other analysis tools for switched and hybrid systems. Acknowledgements The authors would like to thank the editor and the anonymous reviewers for their detailed comments which greatly contributed to this paper. E. 12 . Necessary and suﬃcient conditions for quadratic stabilizability of an uncertain linear system. B. and Martin. 1985. S. For this issue.8) and (3. Stability of switched and hybrid systems. First. 1994. to solve the proposed matrix inequalities (2.and discrete-time switched systems. some of the ideas in Feron (1996) and Zhai (2001) should be useful. 46. M. 44. 43.. W. Branicky. and by the National Science Foundation (NSF ECS99-12458 & CCR01-13131). Dayawansa. though it is a practical problem in real control systems. pp. There are several important issues which should be studied in future work.. IEEE Transactions on Automatic Control. 475–482. For discrete-time switched linear systems. C. References Barmish. This work has been supported in part by Japanese Society for the Promotion of Science under the Grant-in-Aid for Encouragement of Young Scientists 11750396. the stabilizability conditions (3.9) are quite conservative. A converse Lyapunov theorem for a class of dynamical systems which undergo switching. Feron. Boyd. there are very few results concerning output-dependent switching. V. S. as we did in the example. In Proceedings of the 33rd IEEE Conference on Decision and Control.. S. 3498–3503. we have given the switching rules by using the common positive deﬁnite matrix. 1994. 1998. 1999.. and thus need to be relaxed.. M. 399–408. but some geometric method dealing with interval matrices of constant matrices may be eﬀective.. R.. Branicky. This problem is still open for general case. L.. 751–760. Journal on Optimization Theory and Applications. and Balakrishnan.. El Ghaoui. To the best of our knowledge. P. one may ﬁrst consider how to ﬁnd the stable convex combinations of the extreme points. Linear Matrix Inequalities in System and Control Theory (Philadelphia.9) for continuous-time switched linear systems. F. IEEE Transactions on Automatic Control. An extension to the present paper is to consider quadratic stabilizability via output feedback (output-dependent switching) for uncertain switched systems.

. 39. Goh. A... and Lennartson.). and Balakrishnan. 1069–1082. Supervisory control of families of linear set-point controllers-Part 1: Exact matching. IEEE Control Systems Magazine.. A. G. 75. G. and Michel. 88. Morse. AC-35. G. and Michel. In Proceedings of the 2000 American Control Conference. S. 189–205. 59–70. B. 2469– 2471. M. 1996.. Robust stabilization of uncertain linear systems: Quadratic stabilizability and H∞ control theory. S. C. Hu.. K. Pettersson. S. S. Petersen. International Journal of Control. 2. B. A. K. P. R... In Proceedings of the 33rd IEEE Conference on Decision and Control. Zhai. 19. V.. and Michel. 356–361. Quadratic stability of switched systems via state and output feedback. Hybrid system stability and robust veriﬁcation using linear matrix inequalities. A. 1990. 1994.. pp.. Branicky. N.. B.. M. and Chilali. 2009–2014. Automatica. IEEE Transactions on Automatic Control. I. J. A.. Hybrid output feedback stabilization of twodimensional linear control systems. pp. Stability analysis of digital feedback control systems with time-varying sampling periods. 41. Perspectives and results on the stability and stabilizability of hybrid systems.. P... 897–905.. Stabilizing a class of two-dimensional bilinear systems with ﬁnite-state hybrid constant feedback.. and Lennartson. G.. 2184–2188. E. Pettersson. A. 1335–1355.. A. IEEE Transactions on Automatic Control. Feron.. A global optimization approach for the BMI problem. P.. IEEE Transactions on Automatic Control. 1994. 2000. Khargonekar. 2002. and Morse. Gahinet. and Papavassilopoulos. N.. R. Hu. 36. Safonov. 2002. G.DeCarlo. Nemirovski. B. Proceedings of the IEEE. N. A common Lyapunov function for stable LTI systems with commuting A-matrices. 1999. 1413–1431.. Basic problems in stability and design of switched systems. Laub. LMI Control Toolbox for Use with Matlab (the MathWorks Inc. International Journal of Hybrid Systems. A.. Liberzon. D. S. and Zhou.. 13 . Zhai. B. K.. 2000. 1994. M.. 2000. 1996. S. Hu. Technical report CICSP-468 (MIT). Narendra..

and Michel.. Y. 2000. R. and Xu. pp. G. 565–572. A. P. Takai. and Fujisaki.. X.. Wicks. pp. K. B. 210–214. B. Improving closed-loop stability of second-order LTI systems by hybrid static output feedback. Hu. S.. pp... Liu and P. J. In Proceedings of the 2000 American Control Conference.. G.. 2001a. N. Yasuda. K.. Zhai. Disturbance attenuation properties of time-controlled switched systems. pp. Stability and L2 gain analysis for switched symmetric systems. and Michel.. Construction of piecewise Lyapunov function for stabilizing switched systems. M. in D.. Journal of The Franklin Institute. Zhai.. and Lennartson. 37.. B. 143–166. 14 . 3492–3497. Antsaklis (eds) Contemporary Issues in Systems Stability and Control with Applications a (Boston.. 2002b. Zhai. 2165–2166. Peleties.. Control and Information Engineers. 2002a. A. G. Zhai. 2003. Stability and robustness of hybrid systems using discrete-time Lyapunov techniques. Chapter 7.. 117–125. Hu. M.. Yasuda. A. and DeCarlo. 338.. Quadratic stabilizability of discrete-time switched systems via state and output feedback. M. A. 765–779. 1994. In Proceedings of the 40th IEEE Conference on Decision and Control.Rubensson. A. In Proceedings of the 35th IEEE Conference on Decision and Control. 2792–2797. G. In Proceedings of the 41st IEEE Conference on Decision and Control. Automatica. MA: Birkh¨user). G. Ikeda.. 2001. pp.. Stability and L2 gain analysis of discrete-time switched systems. G. Michel. N. N. 15... Decentralized H∞ controller design: A matrix inequality approach using a homotopy method. Zhai. Transactions of the Institute of Systems. Zhai. 2001b.

Sign up to vote on this title

UsefulNot useful- A Lyapunov Analysis of Stability Robustness for Discrete Linear Descriptor Systemsby mamos99blue
- sliding manifolds Matrixby alamimoham
- Robust Stability and Stabilization of Fractional-Order Interval Systems with the Fractional Order alpha.pdfby karpagasenthilpandy
- Statistical Mechanical Analysis of a Typical Reconstruction Limit of Compressed Sensingby Sameh Ahmed

- Survey of Decentralized Control Methods for Large Scale Systems
- Stability and Stabilizability of Discrete-time Switched Linear Systems With State Delay - ACC 2005
- 10.1007_s11045-011-0166-z
- A Survey of Recent Results
- phd_thesis.pdf
- A design methodology for switched discrete time linear systems with applications to automotive roll dynamics control.pdf
- laksh
- 366-TAC09
- A Lyapunov Analysis of Stability Robustness for Discrete Linear Descriptor Systems
- sliding manifolds Matrix
- Robust Stability and Stabilization of Fractional-Order Interval Systems with the Fractional Order alpha.pdf
- Statistical Mechanical Analysis of a Typical Reconstruction Limit of Compressed Sensing
- A+SURVEY+ON+SEMI-TENSOR+PRODUCT+OF+MATRICES
- Construction of Lyapunov-Krasovskii Functional for Time-Varying Delay
- tmp3AD9.tmp
- Manuel de La Sen
- Patter Generation in Diffusive Networks
- 3. State Vaiable Analysis of Discrete-data Systems
- Stochastic Stability Properties
- 1-s2.0-S0005109804000925-main
- System Analysis Through State Space Model - GATE Study Material in PDF.compressed
- Fault-Partition-Dependent Fault-Tolerant Control for T-S Fuzzy Time-Delay Systems
- LinearLyapunov
- fuzzy lyapunov stability
- he rr1
- rez527
- automatic control theory
- 4364-104-CSII
- Asymptotic Properties and Stability Criteria of Zeros of Sampled-Data Models for Decouplable MIMO
- 10.1.1.52
- 291-IJC03