h c w d l n g e of the 3ovI Conference on Decision and Control Brighton, England December 1991

TI-7 10:40

Stabilization of L37 Systems

Way M. Lu; Kemin Zhout and John C. Doyle*

Abstract
The problem of parametrizing all stabilizing controllers for general LFT systems is studied. The LFT systems can be variously interpreted as multidimensional systems or uncertain systems, and the controller is allowed t o have the same dependence on the frequency/uncertainty structure as plant. In the multidimensional systems case, this means the controller is allowed dynamic feedback, while the uncertain system case can be given a "gain scheduling" interpretation. Both p and Q stability are considered, altough the latter is emphasized. In both cases, the output feedback problem is reduced by a separation argument t o two simpler problems, involving the dual problems of full information (FI) and full control (FC). In the case of Q stabi1ity;the F I and F C stabilization problems can be characterized completely in terms of Linear Matrix Inequalities (LMIs). In the standard 1D system case with no uncertainty, the results in this paper reduce to the standard Youla parametrization, although the development here appears to be much simpler, and does not require coprime factorizations.

Introduction

The basic block diagram considered in this paper is

where G is the generalized plant with two sets of inputs: the exogenous inputs w and the control inputs U , and two sets of outputs: the measured outputs y and the regulated outputs z. The control problem in this setup is t o design feedback controller K such that the closed loop structure is stabilized in some sense and the signal z is specified (cf. Doyle et al, 91; Glover and Doyle, 1989 and Zhou et al, 1990). In this paper we are only concerned with the basic stabilization problem, so we will generally ignore z and w and focus on G as an LFT on A with input u and output y:

terminology does not, strictly speaking, apply, because the meaning is always clear from context. The aim of this paper is t o apply the machinery from Doyle et al( 1989 and 1991) t o parametrize all stabilizing controllers for general LFT systems. In this setting, both G and li are LFTs on A, which may be interpreted in a number of ways. One is t o view the A as transform variables in a multidimensional system. We then produce all stabilizing controllers for such systems. A more useful interpretation, as in Doyle et al(1991), is t o view one part of A, (eg. 61 = l / z ) as the transform variable in an uncertain system with the remaining part of A viewed as norm-bounded perturbations. The results in this paper could then be given a "gain-scheduled" interpretation, as the controllers depend on the same perturbations as does the plant. The terms frequency and uncertainty structures will be used interchangeably. We will focus on Q-stabilization under this general setting because the conditions for &-stability, stabilizability and detectability can be elegantly characterized using LMIs. Except for these LMI results, all the arguments and procedures can be applied t o p-stabilization case as well, although the set of controllers in the Q and p case and their interpretations would be different. The construction of stabilizing controllers for the output feedback (OF) problem is done via a sequence of special problems: full information (FI) problem, disturbance feedforward (OF) problem, full control (FC) problem and output estimation (OE) problem, together with a separation argument. These special problems are also of importance in their own right. The parameterization of all stabilizing controllers was first introduced for the one-dimensional case by Youla et al(197G) and extended to the multi-dimensional case by Guiver and Bose(1985). Both used the coprime factorization technique. In this paper we consider the general LFT setting without the need t o introduce coprime factorizations. In section 2 some notions related to &-@)stability are defined and characterized. In section 3 we will examine the &-stabilizations of different special problems, the OF problem is solved via separation arguments. In section 4 the O,-stabilizing controllers for O F problem will be characterized. A longer version of this paper is available (Lu et al, 1991).

2
The frequency/uncertainty structure A is in a set A E Cnxn. For concreteness, we will assume that

LFT Systems, tect ability

Stabilizability and De-

2.1

Linear Systems and Stabilities

A = {diag [611q,, . ..,6JqP] : 6; E C }

Consider G(A) and A E A from above as well as the commutative matrix set V of A defined as
2, =

B A = {A E A

: @ (A)

5 1)

as in Doyle et al(1991). This notation is a direct generalization of the now standard notation for state-space realizations of transfer functions. One of the advantages of the use of LFTs with this notation is that it facilitates manipulation using state-space-like machinery. Thus we often refer t o "state" and "state tra.nsformations" even when this
'Electrical Engineering 116-81, California Institute of Technology, Pasadena, CA
91125.

{D E C"'"

: D A = AD,det[D]

# 0).

The "state" variable transformation I I := TI is said t o be adH ' missible if the transformation matrix T E V. Then the corresponding "realization" is

'Department of Electrical and Computer Engineering, Louisiana State University, Baton Rouge, LA 70803.

Note that G(A) for A E A is unchanged by the transformation, as are the stability properties. Because we will make frequent use of this fact,

CH3076-7/91/0000-1239$01 0 1991 IEEE .OO

1239

Authorized licensed use limited to: Universita degli Studi di Roma La Sapienza. Downloaded on April 25,2010 at 00:10:21 UTC from IEEE Xplore. Restrictions apply.

++

I-*[

[-%q-Y]

we will state it as a theorem. The proof follows immediately from the definitions in Doyle et al( 1991) T h e o r e m 1 The p-stability and Q-stability of LFT systems are invariant under admissible state variable tmnsformations. Another important structural property of LFT systems is expressed in the following theorem. T h e o r e m 2 Let Al and A2 be two system matrices with corresponding uncertainty structures A: and A2 respectively. Then the system

2.3

Characterizations

of

Q-Stabilizability

and
p is the

Q-Detectability
Assume that rank(B) = p 5 n and rank(C) = q dimension of U and q is the dimension of y . P r o p o s i t i o n 1 Consider the system G(A) =

5 n where

[*]

with fw-

[ $ "A,' ]

with any compatible dimensioned matrix A12 is Q (or

p)-stable with respect to structure

z:

i f and only i f A1 and

quency structure A and mnk(B) = p < n. Assume B l E RnX("-p) i such that B ' B l = 0 and B B A is invertible. Then there exists s a static feedback F such that the closed loop system matrix A + B F is &-stable with respect to frequency structure A if and only if there exists a matrix P E V with P = P' > 0 such that

A2 are Q (or p)-stable. Note that in this theorem, the frequency structures A: and A2 may or may not depend on each other. We can easily sketch the proof
of this theorem. The state transformation matrix T =

BiAPA'BI - B i P B l

< 0.

[y

!]is

Moreover, if there exists P satisfied above requirements, then state feedback matrix F = -( B* P-'B)-' B'P-IA &-stabilizes the given system.
The proof of this proposition needs the following lemma: L e m m a 2 Consider the pair ( A ,B ) E RnX" x RnXP and runk(B) = p < n. Let B l E RnX("-P) and BO E RPxn are such that B f B = 0 and BO B l is unitary. Then

admissible. By conducting this transformation, the transformed system matrix TAT-' =

dA'z tends t o as tends 0 Az t o 0. By continuity, the stability of the resulting system is guaranteed by the stability of A I and Az. So A is stable since the admissible state transformation does not change stability. This argument holds for both p and Q. This theorem also implies that a cascade system is Q(or p)-stable if and only if each subsystem is Q(or p)-stable. 2.2

[2 i z ]

inf
F'ERPX"

a(A+ B F ) = a(B1A)

and the infimum is attained by F = -(B,'B)-'B,'A.

Proof. Since U := inf
FERPXn

Stabilizability and Detectability

The general stabilization problem is t o design a (possibly dynamical) output feedback controller Zi(A) = with frequency

I&#-[

[ BO

Bl

] is unitary, then
inf
FERPXn

8 ( A+ B F ) =

a(U*(A+ B F ) )

structure A0 such that the feedback system with G(A) is p-stable or &-stable with respect t o the induced new frequency structure

AN =

[ lo].

Here the structure A0 depends on A, that is,

it is made up of possibly multiple copies of the elements ot A. Now consider the following two special structures,

Moreover the infimum is attained by B,'A -( B,' B ) - 1 B,' A.

+ B,+BF = 0 or F

=
U

where the frequency structures in both cases are the same as the one for G(A). We give following definitions: Definition 1 The system G(A) with frequency structure A is pstabilizable(or Q-stabilizable) if there exists a dynamical controller for the corresponding system GsF(A):

Proof. of P r o p o s i t i o n 1 There exists a static feedback F such that the closed loop system matrix A B F is Q-stable with respect t o frequency structure A if and only if

1 > inf 8 ( D ( A B F ) D - ' ) = inf a(DAD-'

F,D F.D

+ DBFD-')

where infimum is obtained over possible F E RPx" and D E V. Now take V i = ( B f ( D * D ) - ' B l ) - ? B f D - ' then it is easy t o check that V i V l = I and V i ( D B ) = 0. Now by lemma 2, we have 1 > inf U ( D ( A B F ) D - ' ) = inf a(VTDAD-')
F,D

such that the closed loop system is p-stable (or Q-stable) with respect to the induced frequency structure. We can characterize this property by following fact.
L e m m a 1 The given system is p-stabilizable(or Q-stabilizable), i.e. its corresponding system GSF can be p-stabilized(or Q-stabilized) by

or there exists a D E

V such that

some K(A) =

I*]

(ViDAD-')(ViDAD-')* < Z
Take P = (D*D)-' then P E V and P = P'

with frequency structure A0 related to A

> 0, hence we have

if and only if the augmented system of GsF(A)

r A O I B

(B;PB~)-?B;APA*B~(B;PB~)-? o -z <
or B i A P A ' B l - B i P B l

01
=

<0

Using the above result we can easily get

can be p-stabilited(or &-stabilized) by static feedback F

[ 2: 2; ]

T h e o r e m 3 G(A) =

with respect to frequency structure AN =

[%]

is Q-stabilizable if and only if there

We can also define detectability in the obvious (dual) way.

exists a static feedback matrix F such that the closed loop system matrix A + B F is Q-stable with respect to the same frequency structure.
1240

Authorized licensed use limited to: Universita degli Studi di Roma La Sapienza. Downloaded on April 25,2010 at 00:10:21 UTC from IEEE Xplore. Restrictions apply.

Proof. If B is square and of full rank, then the result is trivial. NOW we only consider the case where r a n k ( B ) = p < n . Assume that the system can be &-stabilized by a dynamical controller I<(A) = F,(Fo,Ao) with FO =

[ 2; 2: ]

and A0 is related Where all of these special systems have the same frequency structures as G(A). We assume all physical variables have the compatible dimensions. The motivations for these different problems in one-dimensional case were given in Doyle et al (1989). To examine their relationships, we will use a simple notion of algebraic duality. Consider the standard block diagram. Now consider another system shown below

t o the system frequency structure A. By lemma 1, this is equivalent t o the augmented system

being Q-stabilized by static feedback frequency structure AN =

I lo 1.

Fzz Fiz

Fzl with respect t o the

41

Now assume the commutative

matrix set of AN is V N ,then by the above proposition, there exists a PN =:

PT Po (positive), such that

V N ,which is positive definite and P E V

whose plant and controller are obtained by transposing G(A) and Ii(A). And we can check via LFT formula that Tw = [Ft(G,IC)IT = : 3p(GT,IiT)= Tztwt. It is not difficult tosee that Ii &-stabilizes G with respect to the induced frequency structure AN =

i.e. BiAPA'Bl - B i P B l < 0 which means that the system can be &-stabilized by a static feedback matrix via the previous proposition.
U

We can also get the dual results for &-detectability.

3
3.1

Stabilization and Special Problems

Problem Statement, Assumptions, Problems and Special

The stabilization problem is t o find a feedback mapping Ii such that the closed-loop system is well-posed and stable in some sense. For the rest of this paper will focus on the &-stabilization problem, and controllers that &-stabilize systems will be said to be admissible. We call this general synthesis problem output feedback(0F). All of the remaining results, including the proofs, hold for p-stability as well, and the relationship will be discussed in the conclusions. Assume G(A) has a realization of the form

only if KT &-stabilizes GT with respect to the frequency structure A;. And we say that these two control structures are algebraically dual, in particular, GT and K T are dual objects of G and Ii, respectively. So as far as stabilization or other synthesis problems are concerned, we can obtain the results for GT from its dual object G if available. So it is not difficult t o see that the FI and FC structures ( D F and O E structures) are algebraically dual. We will also see that F I and D F structures ( F C and O E structures) are equivalent in a sense that will be made precise in the next subsection.

] [ lo

if and

3.2

FI and DF Problems

For the FI problem, we only need to assume ( A ,E*) is &-stabilizable t o guarantee the solvability. And the conditions for the solvability of D F problem are both ( A ,Bz) is &-stabilizabie and (Cz,A ) is Q-detectable. Suppose that we have controllers K F I and ICDF and let TFIand TDFdenote the closed-loop Tzws in
YFI

-@ -

ys

I~FI iD which has compatible dimensions with the related physical variables. Assume ( A ,Bz) is &-stabilizable and (Cz,A ) is &-detectable. We have following theorem. In addition, let the state-space realization of I<(A) be T h e o r e m 4 ( i ) Feedback matrix IiDF Q-stablizes D F structurr, I<(A) = I then I<DF[C~] can be used as a &-stabilizing feedback matriz for the corresponding FI structure, which makes ~ ~ ( G DK D, F )= F with the frequency structure A0 a function of A, which will be assumed Ff(GFr, I i D d C z 11). throughout t o be Q-stabilizable and &-detectable. The well-posedness (ii) Suppose that A - BlCz is &-stable. For any &-stabilizing feedof this interconnection implies I - DzzD is invertible. back matrix IiFI for the FI structure, a Q-stabilizing feedback We will see that ( A ,Bz) is &-stabilizable and ( C z , A ) is &s IiFI) controller for GDF i ~ ~ ( P D F , with detectable with respect t o frequency structure A, is both sufficient and necessary for the solvability of O F problem. Now, we consider some special problems which are related t o the general 0 F problems, they all pertain t o the standard block diagram, but with different structures from G. The problems are labeled as FI, DF, FC and OE which stand for full information, disturbance feedI~ ) Ff(PDF7 ZiFI)). Mowover, ~ ~ ( G F I ,F I= ~~FP(GDF, f o r w a d , full control and output estimation respectively. Their corresponding plants are Proof. (i) is easy. Now we prove (ii). As~~ F sume ~ ~ ( G D F ,( P DKF ,I ) ) =: .Fe(GkI,IiFI). We will first prove that GkI = GFI. We examine the system GkI, let z and f denote the state of GDF and PnF. respectively, take e := z - 2 and f as the states of the

[%]

1241

Authorized licensed use limited to: Universita degli Studi di Roma La Sapienza. Downloaded on April 25,2010 at 00:10:21 UTC from IEEE Xplore. Restrictions apply.

resulting interconnected system Gb,, then its realization is

with respect t o the frequency structure

transfer matrix is exactly GFI as claimed. Since A - BIG2 is &-stable, so the theorem follows.
0

[ 11.

The resulting

where

is &-stable, and

Remark 1 Note that, given a feedback matrix K D Ffor the D F structure, then the corresponding FI controller can be obtained from (i) z and now we return it to D F problem by (ii), as K D F C I

1;

then ~ ~ ( P DKDF F,

fi(PDF,KFI) c I = I ~ F I . 2 Remark 2 This theorem shows that if A - B1C2 is &-stable, then problems FI and DF ape equivalent in the above sense. 3.3 FC and OE Problems FC(0E) problem is dual t o the FI(DF) case and has the dual solvability condition t o FI(FD) problem. The solutions t o this kind of control problems can be obtained by first transposing GFC(GOEand solving the corresponding FI(DF) problem then transposing it back. Consider the following FC and O E diagrams

[c 2 ]

1) = K D F . w e also have the fact that

From theorem 2, we know that K &-stabilizes G if and only if K &-stabilizes Gtmp. Now Gt,, has an OE structure, so let L be such that A+ LC2 is Q-stable then

corresponding FC problem. Since A + B2 F is &-stable by construction, by Theorem 5 we have a controller given by K(A) = 3 4 J , where

[:I

is a &-stabilizing controller for the

rA+BzFIO

-B21

we have the dual results t o the ones in the last subsection:

This proves the following theorem, which is stated without the assumption that D Z Z= 0. Theorem 0 Consider the geneml O F problem. Let F and L be such that A + LC2 and A + B2F are &-stable, then the controller

Theorem 5 (i) Given an O E feedback matrix KOE which &stabilizes 0 Estructure, then the corres nding &-stabilizing FC controller can be chosen as

~ ~ ( G o IE ,o E = ~ ~ ( G F c , ~ )

[ T]

F~oE,

K(A) =

+ B2F + LC2 + LDzzF I - L

F

which makes

ICOE).

(ii) Suppose that A - BzCl is &-stable. For any &-stabilizing FC controller K F C , a &-stabilizing O E controller can be taken as ~ ~ ( P o K F C )with E,
PoE(A) =

with the frequency structure A &-stabilizes the given system. We can also get the same result by the dual procedure t o above construction. Now we examine the separation property. We will denote the state variable of G(A) as x and state variable of resulting controller Ii(A) as E. We will take the state variable of the closed loop system as
r 7

z=

, and the corresponding realization is

Momover, ~ ( ( G F CI,~ F c = ~ ~ ( G o E , ~ ~ ( IP o E) .) . ) ~Fc

Now we conduct the admissible state transformation Z

Remark 3 Given controllers I ~ F C and K O E for FC and O E structures, respectively, we have the following facts:

ization is

1,

then T =

JI

1.

T% =

After the transformation, the real-

Remark 4 W e can see that if A - B2Cl is Q-stable, then FC and OE problems are equivalent in the above sense.
3.4

OF Problem and Separation Property

Consider the general OF problem. Without loss of generality, we further assume the realization of G(A) has D22 = 0. Let z denote the "state" of system G(A). Since ( A , & ) is &stabilizable, there is a constant matrix F such that A+ B2 F is &-stable. Note that F 0 is actually a special FI stabilizing controller. Now let v = U - Fx, then the system can be broken into two subsystems GI and Gtmnas shown below

i.e. the system is decoupled into two separated subsystems which are &-stable with respect t o the same frequency structure A by assumption. Hence the closed-loop system after admissible state variable transformation is also &-stable with respect to the new frequency structure AN =

[ 1]

by theorem 2, so is the original closed-loop

system as desired. The separation property is clear now, since the synthesis of O F problem can be reduced t o FI and FC problems, i.e. the latter two problems can be designed independentlv.

1242

Authorized licensed use limited to: Universita degli Studi di Roma La Sapienza. Downloaded on April 25,2010 at 00:10:21 UTC from IEEE Xplore. Restrictions apply.

4
4.1

The Characterization of All Q-Stabilizing Controllers For Output Feedback Problems

Problem Statement and Motivations

U;

it is Q-stable by the Q-stability of the closed loop system. Then 0 = F X U = FZ Qw, so I<F~(A) F Q(A) Z

1.

Consider again the standard system block diagram

Remark 5 In the above proof, we have used a standard technique of

changing variables. A similar change of variables appeared in Doyle et al(1989).
Now we consider the dual F C problem whose diagram is given C in the last section. We say controllers KFC and ~ F are equivalent in the sense that the same injection inputs YFCS p d u c e the same outputs 2s. This also guarantees the identity of their resulting closed loop transfer matrices from w to z. We also have P r o p o s i t i o n 3 Let L be a constant matrix such that A LC2 is &stable. Then the set of equivalent classes of all admissible controllers for FC in the above sense can be parameterized as

with

Suppose ( A , Bz) is Q-stabilizable and (Cz, A ) is &-detectable, in this section we discuss the following problem:

Given a plant G(A), parametrize all controllers Zi(A) that Q-stabilize G.

More specially, the characterization problem is t o find a suitable matrix J(A) such that any admissible controller I i ( A ) for G(A) can be expressed as an LFT with respect to a Q-stable parameter matrix Q(A) with coefficient matrix J(A), i.e. K = &(J,Q). Note that for a Q-stable parameter matrix Q(A), it will be assumed that its realizations are &-stabilizable and Q-detectable. In this section, we only consider the characterization of all admissible controllers for O F problems. For this purpose it is unnecessary t o parametrize all admissible controllers for the special problems. Instead, we only parameterize a set of controller equivalence classes which generate the same control action. We say two controllers, Ii and K, are of equivalent control actions i f their corresponding closed loop tmnsfer matrices are identical, i.e. Fl(G, Ii) = Fr(G, Ii), written as A E A. The controller equivalence is an equivalence relation. We will see that, for different special problems, we have different refined versions of this relation. We will also see that the characterizations of equivalent classes of admissible controllers for special problems are good enough t o construct the characterization of all admissible controllers for O F problem.

with any Q-stable Q(A) 4.3

Admissible Controllers for Problems DF and OE

The D F block diagram is given as in the last section. We will further assume that A - BlCz is Q-stable in this subsection. This will simplify the solution, and can be relaxed. We say that the controllers KDF(A) and IibF(A) are equivalent for the D F problem i f the two transfer matrices from w to U in the above diagram are the same . Of course the resulting two closed loop transfer matrices from w t o z are identical.

Remark 6 Since the equivalence between F and D F problems as Z

discussed in the last section, it is easy to show that if I i D F 2 IibF in the D F structure, then I<DF C I 4 IibF CZ I z in the corresponding FI structure. We also have that if IiFI g K k z , then

3 (PDF ZiFI) 4 Fl(PDF, zih). 1

7

Now we construct the parametrization of equivalence classes of stabilizing controllers for D F structure. Let IiDF(A) be an admissible Q-stabilizes the control for DF then ziFI(A) = ziDF(A) CZ I

4.2

Admissible Controllers for FI and FC Problems

We first examine the F structure as given in the last section. We Z I I say two controllers I ~ F and ~ F are equivalent if they produce the same closed loop transfer function from w to U . Obviously, it is also K ~ guaranteed that ~ ~ ( G F IF, I )= F~(GFI, F I ) . Since we have full information for feedback, our controller will have the following general form

corresponding GFI(A). Assume ziFI(A) Z Iikr(A) = F Q(A) Z for some &-stable Q(A), then ZihI(A) &-stabilizes GFI(A) and Fe(JDF(A), Q(A)) = J=P(PDF(A), ZihAA)) where

IiFr(A) =
with Iil(A) &-stabilizing

[w]
[F

[ Zii(A)

Iiz(A)

]
+

and arbitrary Q-stable Ziz(A).

P r o p o s i t i o n 2 Let F be a constant matrix such that A B 2 F is Qstable. Then all admissible controllers, in the sense of generating all Q-stabilizing control, for F I can be parametrized as

BzF is &-stable. Hence by Theorem 4, with F such that A IibF(A) := F~(JDF(A), Q(A)) stabilizes GDF(A) for any Q-stable Q(A). Since IiFr(A) 2 Iih/(A), so by Remarks 1 and 6 we know ZiDF(A) 2 IibF(A) = F~(JDF(A),Q(A)), this characterizes the equivalence class of all controllers for D F problem by the equivalence Z of F and D F . In fact, we have following proposition.
P r o p o s i t i o n 4 All admissible controllers for D F problem can be characterized by I ~ D F ( A )= .Ft( JDF(A),Q,J(A) with Q-stable Qo(A), where JDF(A) is given as above. P r o o f . It is easy to show the controllers expressed in the given LFT formula do &-stabilize GDF by transforming it to the corresponding F problem. Now Let ZiDF be any admissible controller for GDF, then Z F ~ ( ~ D F i,D F ) is Q-stable where I

I<FZ(A) 2
with any Q-stable Q(A).

Q(A)

Proof. It is easy t o see that the controller given in the above formula Q-stabilizes the system GFI(A). Hence we only need t o show that the given set of controllers parameterize all Q-stabilizing control action, U , i.e., there is a choice of Q-stable Q(A) such that the transfer functions from w t o U for any stabilizing controller Iipr(A) = Iil(A) Ziz(A) and for IiFz(A) = F Q(A) are the same. To show that, make a change of control variable as v = U - Fs, where z denotes the state of the system GF/(A). Consider the system with ) the controller Z ~ F ~ Alet. Q(A) be the transfer matrix from w to

1243

Authorized licensed use limited to: Universita degli Studi di Roma La Sapienza. Downloaded on April 25,2010 at 00:10:21 UTC from IEEE Xplore. Restrictions apply.

Jtmp can be obtained by using composition formula of LFT as

This theorem shows that any admissible controller K ( A ) can be characterized as an L F T of a Q-stable parameter matrix & ( A ) ,i.e., K ( A ) = F t ( J ( A ) , Q ( A ) )There is an alternative direct proof that this . parametrization produces all stabilizing controllers. To see this, recall from the inversion formulas for LFTs in Doyle et al( 1991) that we can ) solve the equation K ( A ) = F t ( J ( A ) , Q ( A )for Q ( A ) t o give

Q = F " ( J - ~ f ) = 3 [ ( j ,K ) i,
where a little algebra shows that Hence F ~ ( J D Qo) = F (JtmprK D F ) = K D F . This shows that F, t , any admissible controller can be expressed in the form of ~ ~ ( J DGFo ) for some &-stable Qo. 0 Since for any admissible X D F , we can get two parametrizations by the above two formulae. We relate them as

J-' =

[y]
-F

Proposition 5 Let KDF = F (J D F ,Qo) for some &-stable Q o be obt

tained from the constructive proof of Proposition 4 and let I i b F = F ~ ( J D F , Q )KDF for some Q-stable Q be obtained from its corre2 = sponding F I structure. Then Qo = Q . Moreover I ~ D F K b F .
Now we turn t o the O E problem, the system structure is given in the last section. We will assume that A - B2C1 is &-stable. We have the following result

Proposition 6 All admissible controllers f o r O E problem can be chart acterized as F (JOE, Qo) with any Q-stable Qo, where JOE i s defined

Note that Q is stable if and only if K stabilizes &. But j22 = Gzz, so Q is stable if and only if Z stabilizes G, as desired. f This last manipulation illustrates the power and simplicity of the LFT machinery. In contrast, the standard proof that the formula K(A) = Fp(J(A), ( A ) )yields al stabilizing controllers requires inQ l volved arguments using coprime factorizations. It is also possible t o obtain similar results using the chain-scattering version of LFTs (Kimura, private communication).

Concluding Remarks

with L such that A 4.4

+ LC2 is &-stable.

All Admissible Controllers for Problem OF

Consider the O F problem given in section 4.1, we have

Theorem 7 Let F and L be such that A + LC2 and A + BZF are Q-stable, then all controllers &-stabilize G( A ) can be Characterized as F ( J , Q ) with ,
A

J ( A )=

+ BzF + LCz + LDzzF

-(cz

-L
0

Bz

+ LDzz

+ DzzF)

We have considered the problems of stabilization and the parametrization of all stabilizing controllers for LFT systems. Although the focus was on &-stability, most of the results are identical in the p-stability case via simple change of notation. The exception is that the FI and OE solutions in the p case are not simply characterized by LMIs, and stabilization by dynamic feedback is not equivalent t o stabilization by constant gain. The separation theory holds in the p case exactly as in the & case, although the notation is a bit more cumbersome. The separation proof in this paper holds in greater generality than for just the & and p stability problems. All that is required for the separation proof is that the notion of stability satisfy two requirements: 1) stability invariance under a sufficiently rich set of similarity transformations, as in Theorem 1, and 2) a certain structural property as given in Theorem 2. It would clearly be possible t o develop a more abstract axiomatic stabilization theory using these 2 properties.

and with any &-stable Q ( A ) such that the structum is well-posed.

References
1. Doyle,J.C. (1984), Lecture Notes in Advances in Multivariable

Proof. We will assume again Dzz = 0 for simplicity. Let x denote the state of system G. Since (A, Bz) is &-stabilizable, there is a constant matrix F such that A BzF is &-stable. Note that F 0 is actually a special F I &-stabilizing controller. Now let v = U - Fx as in the proof of Theorem 6, we have K ( A ) &-stabilizes G ( A )if and only if it &-stabilizes

Control, ONR/Honeywell Workshop, Minneapolis, MN. 2. Doyle,J.C., K.Glover, P.Khargonekar and B.Francis (1989), StateSpace Solutions t o Standard 7 1 2 and 71, Control Problems, IEEE Trans., Vol.AC-34, pp.831 847.

Gtmp(A)=

I*[
4

-F

c z

0 Dzi

3. Doyle,J.C., A.Packard and K. Zhou(1991), Review of LFTs, LMIs and p , 1991 IEEE CDC, England. 4. Glover,K. and J.C.Doyle (1989), A State Space Approach t o U, Optimal Control, in Three Decades of Mathematical System Theory (H.Nijmeijer, J.M.Schumacher Eds.), Springer-Verlag, Berlin. 5. Guiver,J.P. and N.K.Bose (1985), Causal and Weakly Causal 2D Filters with Applications in Stabilization, in Multidimensional Systems Theory(N.K.Bose ed.), D.Reide1 Publishing Co., Holland, 1985.
6. Lu,W.M., K.Zhou and J.C.Doyle(lggl), On stabilization and Stabilizing Controller Characterization of LFT Systems, Report.

Now G t m p ( Ais of the OE structure, let L be such that A LC2 ) is &-stable then by proposition 6 all controllers &-stabilizing Gtmp( ) A are given by K ( A )= Fe(J(A),Q(A))

1
1

A + Bz:+ Lc2
2

I B21
-L O I 1 1 0 1
0

7. Youla,D.C., H.A.Jabr and J.J.Bongiorno (1976), Modern WienerHopf Design of Optimal Controllers -- The Multivariable Case, IEEE Trans., Vol.AC-21, pp.319 338.
N

This concludes our proof.

8. Zhou,K., J.C.Doyle, K.Glover and B.Bodenheimer (1990), Mixed 'HZ and 1-1, Control, Proc. of 1990 ACC.

1244

Authorized licensed use limited to: Universita degli Studi di Roma La Sapienza. Downloaded on April 25,2010 at 00:10:21 UTC from IEEE Xplore. Restrictions apply.