IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 34, NO. 3.

MARCH 1989

293

LQG Control with an H , Performance Bound: A Riccati Equation Approach

Abstruct-An

LQG control-design problem involving a constraint on

H, disturbance attenuation is considered. The H, performance con-

straint is embedded within the optimization process by replacing the covariance Lyapunov equation by a Riccati equation whose solution leads to an upper bound on L2 performance. In contrast to the pair of separated Riccati equations of standard LQG theory, the H,-constrained gains are given by a coupled system of three modified Riccati equations. The coupling illustrates the breakdown of the separation principle for the H,constrained problem. Both full- and reduced-order design problems are considered with an H , attenuation constraint involving both state and control variables. An algorithm is developed for the full-order design problem and illustrative numerical results are given.

I. INTRODUCTION HE fundamental differences between Wiener-Hopf-Kalman (WHK) control design (for example, LQG theory [11) and H , control theory [2]-[4] can be traced to the modeling and treatment of uncertain exogenous disturbances. As explained by Zames in the classic paper [2], LQG design is based upon a stochastic noise disturbance model possessing a fixed covariance (power spectral density), while H , theory is predicated on a deterministic disturbance model consisting of bounded power (square-integrable) signals. Since LQG design utilizes a quadratic cost criterion, it follows from Plancherels theorem that WHK theory strives to minimize the L2 norm of the closed-loop frequency response, while H , theory seeks to minimize the worst-case attenuation. For systems with poorly modeled disturbances which may possess significant power within arbitrarily small bandwidths, H , is clearly appropriate, while for systems with well-known disturbance power spectral densities, WHK design may be less conservative. In addition to the fact that H , design embodies many classical design objectives [5], it also presents a natural tool for modeling plant uncertainty in terms of normed H , plant neighborhoods. In contrast, the H2 topology has been shown in [6] to be too weak for a practical robustness theory, while the H , norm is not only suitable for robust stabilization but is also conveniently submultiplicative. Within the WHK state-space theory, however, the appropriate robustness model appears not to be a nonparametric normed plant neighborhood as in H , theory, but rather a parametric uncertainty model. The principal technique for bounding the effects of real parameters within state-space models is Lyapunov function theory (see, e.g., [7]-[ 161 and the references therein). Such structured uncertainties are difficult to capture nonconservatively within H, theory except with specialized refinements [171.

Manuscript received November 24, 1987; revised August 1, 1988 and April

20, 1988. Paper recommended by Associate Editor, A. C. Antoulas. This work was supported in part by the Air Force Office of Scientific Research
under Contracts F49620-86-C-OOO2 and F49620-87-C-0108. D. S. Bernstein is with the Government Aerospace Systems Division, Harris Corporation, Melbourne, FL 32902. W. M. Haddad is with the Department of Mechanical and Aerospace Engineering, Florida Institute of Technology, Melbourne, FL 32901. IEEE Log Number 8824525.

In spite of the fundamental differences between WHK design and H , theory, a significant connection was discovered in [18]. Specifically, Petersen observed that a modified algebraic Riccati equation developed for parameter-robust full-state-feedback control can be reinterpreted to yield controllers satisfying H , disturbance attenuation bounds. This relationship was further explored in [I91 where it was shown that the H,-optimal static full-state-feedback controller is also optimal over the class of dynamic full-state-feedback controllers. The results of [181-[20] thus solve the standard problem considered in [3] and [4] for the full-state-feedback case. The extension of these results to the standard problem for dynamic output-feedback compensation, however, was not given in [18]-[20]. Within the realm of quadratic robust stabilization, the dynamic output-feedback problem was addressed in [7]. The results of [7] involve a pair of decoupled modified Riccati equations along with an auxiliary inequality. Using different techniques, a more complete solution was obtained in [13] and [14] involving a coupled system of three modified Riccati equations for full-order dynamic compensation and a coupled system of four modified Riccati and Lyapunov equations in the fixed-order (i.e., reduced-order) case as in [21]. The results of [I31 and [14] thus raise the following question: What is the relevance of this system of coupled design equations to the problem of H, disturbance attenuation via fixed-order compensation? To begin to address this question, the goal of the present paper is to develop a design methodology for fixed-order, i.e., full- and reduced-order, L2 optimal control which includes as a design , constraint a bound on H disturbance attenuation. There are three principal motivations for developing such a theory. First, the results of [ 181-[20] present full-state-feedback controllers whose form is directly analogous to the standard LQR solution. However, no L2 interpretation was provided in [18]-[20] to explain this similarity. The present paper thus provides an L2 interpretation within the context of an H , design constraint. A novel feature of this mathematical formulation is the dual interpretation of the disturbances. That is, within the context of L2 optimality the disturbances are interpreted as white noise signals while, simultaneously, for the purpose of H , attenuation the very same disturbance signals have the alternative interpretation of deterministic Lz functions. This dual interpretation is unique to the present paper since stochastic modeling plays no role in [18][20]. We also note recent results obtajned in [22] which essentially show that the Hz plant topology can be induced by imposing L2 and L, topologies on the disturbance and output spaces, respectively. For further investigation into the relationships between L2 and H , control, see [22a]. The second motivation for our approach is the simultaneous treatment of both Lz and H , performance criteria which quantitatively demonstrates design tradeoffs. Specifically, in order to enforce the H , constraint we derive an upper bound for the Lz criterion. Minimization of this upper bound shows that the enforcement of an H- disturbance attenuation constraint leads directly to an increase in the L2 performance criterion. The third motivation for our approach is to provide a characterization of fixed-order dynamic output-feedback control-

0018-9286/89/03oO-0293$01.oO @ 1989 IEEE

294

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 3 ,NO. 3, MARCH 1989 4

lers yielding specified disturbance attenuation. Existing optimal Notation H, design methods tend to yield high-order controllers. IntuiNote: Al matrices have real entries. l tively, solving the fixed-order design equations for progressively Real numbers, r x s real matrismaller H , disturbance attenuation constraints should, in the expected value ces, W x l , limit, yield an H,-optimal controller over the class of fmed-order r x r identity matrix, transpose, stabilizing controllers. Although our main result gives sufficient r x s zero matrix, O r x r conditions, we also state hypotheses under which these conditions Trace, spectral radius are also necessary (Proposition 4.1). It should also be noted that r x r symmetric, nonnegativethe inherent coupling among the modified Riccati equations shows definite, positive-definite matrices that the classical separation principle of LQG theory is not valid 22-21 E M', 2 2 - 2 E P, 1 ' for the H,-constrained full- and reduced-order design problems. 2 , z E 8' 1 2 In the full-order case involving equalized L2 and H , perfomPositive integers; n + n,(n, I n) ance weights, we also show that the H,-constrained gains are n, m,I, n,, A-dimensional vectors given by two rather than three equations (Section V). These two equations are precisely those given in [26] for the pure H, problem without an L2 interpretation. Since the results of [26] are 2 necessary as well as sufficient, these connections show that our sufficient conditions (at least in this special case) are also n x n, n x m, I x n matrices necessary. The authors are indebted to Prof. J. C. Doyle for n, x n,, n, x I , m x n, matrices pointing out these relationships and to D. Mustafa for providing a preprint of [45] which further clarifies these C O M ~ X ~ ~ O ~ S . Besides establishing connections with robust stabilizability in A [ic state-space systems, an immediate benefit of the modified Riccati equation characterization of H,-constrained controllers is the p-dimensional standard white opportunity to develop novel computational algorithms for connoise troller synthesis. To this end a continuation algorithm has been n x p, 1 x pmatrices; DIDT = 0 developed for solving the coupled system of three modified DID:, 40:; V2 E P' Riccati equations. In a numerical study (see Section VIII) we have demonstrated convergence of the algorithm and reasonable d computational efficiency. Homotopy methods were suggested for the coupled Riccati equations because of their demonstrated effectiveness in related problems which also involve coupled P control problems modified Riccati equations [23]-[25]. Since Hm are solvable by established numerical methods [4], it should be stressed that the objective of these numerical studies is not to make direct comparisons with existing H synthesis algorithms, but , rather to demonstrate solvability of the coupled modified Riccati equations. The contents of the paper are as follows. After presenting notation at the end of this section, the statement of the H,R constrained LQG control problem is given in Section II. The principal result of Section 1 (Lemma 2.1) shows that if the 1 algebraic Lyapunov equation for the closed-loop covariance is replaced by a modified Riccati equation possessing a nonnegativedefinite solution, then the closed-loop system is asymptotically stable, the H, disturbance attentuation constraint is satisfied, and the L2 performance is bounded above by an auxiliary cost function. The problem of determining compensator gains which minimize this upper bound subject to the Riccati equation 1 constraint is considered in Section 1 1 as the auxiliary minimizaBR;lBT, C T V ; l C x, c tion problem. Necessary conditions for the auxiliary minimization Nonnegative constant, positive problem (Theorem 3.1) are given in the form of a coupled system 0, Y constant of three modified Riccati equations. In Section IV the necessary conditions of Theorem 3.1 are combined with Lemma 2.1 to yield sufficient conditions for closed-loop stability, H , disturbance II. STATEMENT THE PROBLEM OF attenuation, and bounded L2 performance. In Section V we derive In this section we introduce the LQG dynamic output-feedback alternative forms of the design equations and specialize the results to the simpler case in which the LQG weights are equal to the H , control problem with constrained H , disturbance attenuation weights. To achieve further design flexibility, the reduced-order between the plant and sensor disturbances and the state and control-design problem is considered in Section VI. A simplified control variables. Without the L2 performance criterion, the qualitative analysis of the full-order design equations is given in problem considered here essentially corresponds to the standard Section VI1 to highlight important features with regard to problem of [3] and [4]. For simplicity we restrict our attention to existence and multiplicity of solutions. Finally, a numerical controllers of order n, = n only, i.e., controllers whose order is algorithm is presented in Section VIII along with illustrative equal to the dimension of the plant. This constraint is removed in numerical results. A series of designs is obtained to illustrate Section VI where controllers of reduced order are considered. tradeoffs between the L2 and H, aspects and the conservatism of Hence, throughout Sections 11-V the controller dimension n, and the L2 performance bound. Although in the present paper the closed-loop plant dimension A 4 n n, should be interpreted as numerical results are limited to the case of full-order dynamic n and 2n, respectively. Controllers of order greater than n are compensation, reduced-order designs have been obtained in [27] generally of less interest in practice and thus are not considered in using Theorem 6.1. this paper.

[:I

7 1

[lh2]

BERNSTEIN AND HADDAD: LQG CONTROL

295
For convenience we thus define R I 9 ErEl and R2 P ErE2 which apperir in subsequent expressions. Although an L2 crossweighting term of the form 2xT(t)R12u(t) also be included, can we shall not do so here to facilitate the presentation. For the H , performance constraint, the transfer function (2.5) involves weighting matrices E l , and EZm the state and control for variables. The matrices R I , P ET,EIm and R2, 42 E,',E2, are thus the H , counterparts of the L2 weights R I and R2. Although we do not require that R I , and R2, be equal to R I and R,. we shall require in the next section that R2, = P2R2, where the nonnegative scalar P is a design variable. Finally, the condition ET,,Ez, = 0 precludes an H , cross-weighting term which again facditates the presentation. Before continuing, it is useful to note that if A is asymptotically stable for a given compensator (A,, B, ,C,), then the performance (2.7) is given by J(A,, B,, Cc)=tr QR

H,-Constrained LQG Control Problem: Given the nth-order stabilizable and detectable plant
X(

t ) = A x ( t )+ Bu( t ) +Dl w ( t ) ,
y ( t ) = Cx(t)+D2 w ( t )

(2.1) (2.2)

determine an nth-order dynamic compensator

X J t ) =A,xc(t)+B,y(t),

(2.3) (2.4)

u ( t )= Ccxdt)

which satisfies the following design criteria: i) $e closed-loop system (2.1)-(2.4) is asymptotically stable, i.e., A is asymptotically stable; ii) the qm x p closed-loop transfer function

H ( s ) 8 Em(sZa-A)-ID
from w(t) to E1,x(t)

(2.5)

(2.10)

+ EZmu(t) satisfies the constraint

IIn(s)llm S Y
(2.6)

where the steady-state closed-loop state covariance defined by

QA

where y > 0 is a given constant; and iii) the performance functional

1-m

t-m

lim E [ f ( t ) f T ( t ) ]

(2.1 1)

satisfies the fi x fi algebraic Lyapunov equation

J ( A , , B,, C,)A lim ~ [ x T ( t ) R I x ( t ) + u T ( t ) R 2 ~ ( t ) l ( 2 . 7 )

o=AQ+QAT+ v.

(2.12)

is minimized. Note that the closed-loop system (2.1)-(2.4) can be written as

k(t) = Af(t ) + d w ( t )

(2.8)

Remark 2.5: Using (2.10) and (2.12) it can be shown that the L2 cost criterion (2.7) can be written in terms of the L2 norm of the impulse response of the closed-loop system. Specifically, by writing Q satisfying (2.12) as

and that (2.7) becomes J(A,, B,, C,)= lim E[(l%(t))T(&(t))]

f-m

(2.10) becomes

= lim
1-m

E[fT(t)Rf(t)l.

(2.9)

Remark 2. I : Since (A, B, C) is assumed to be stabilizable and detectable the set of nth-order stabilizing compensators is nonempty. Remark 2.2: It is easy to show that the performance functional (2.7) is equivalent to the more familiar expression involving an averaged integral, i.e.,
J ( A , , B,, C,)= - E lim 1-m t

[s,

[xT(s)R1x(s)

+ uT(s)R2u(s)lds

Remark 2.3: For convenience we assume D I D ; = 0, which and effectively implies that the plant disturbance and sensor noise are uncorrelated. Remark 2.4: One may also consider a general L2 optimization problem of the form min 11 T - VQVl12, where Q is a Then parameterization of stabilizing controllers. In this case, without a constraint on the MacMillan degree of Q, it may be possible to satisfy (2.6) with smaller values of y. Note that the problem statement involves both L2 and H , if and only if performance weights. In particular, the matrices R I and R2 are the L2 weights for the state and control variables. By introducing L2weighted variables

where 11 . !IF denotes the Frobenius matrix norm The key step in enforcing the disturbance attenuation constraint (2.6) is to replace the algebraic Lyapunov equation (2.12) by an algebraic Riccati equation which overbounds the closed-loop steady-state covariance. Justification for this technique is provided by the following result. Lemma 2. I : Let (A,, B, , C,) be given and assume there exists Q E W f i x f i satisfying
Q E RIfi

(2.13)

o =AQ + Sp

+ Y - 2 ~ RQ+ V . m

(2.14)

(A", d)is stabilizable A is asymptotically stable.

(2.15)

(2.16)

z ( t ) = El X ( t ) ,
the cost (2.7) can be written as

U (t ) =E2 U( t )

(2.17)

J(A,, B,, C,)=lim ~ [ ~ ~ ( t ) z ( t ) + u ~ ( t ) u ( t ) l .

t+m

QSQ.

(2.18)

296

IEEE TRANSACTIONS ON AUTOMATIC CONTROL,

VOL. 34, NO. 3, MARCH

1989

Consequently,

(2.14) and such that (2.15) or, equivalently, (2.16) holds. Then

2(PQ)I . y
where

(2.24)

3W,, B,, C,,Q) A tr QR.

(2.20)

Proof: Since Q is assumed to be invertible, (2.14) is equivalent to

Proof: It follows fromJ28, Theorem 3.61 that (2.15) implies is also stabilizable. Using the that (A, [y-2QR,Q + assumed existence of a nonnegative-definite solution to (2.14) and [28, Lemma 12.21, it now follows that A is asymptotically stable. The converse is immediate. The proof of (2.17) follows from a standard manipulation of (2.14); for details see [29, Lemma 11. To prove (2.18), subtract (2.12) from (2.14) to obtain

O = y 2 A T ~ + -p&-IA+ y I -

2 ~ I

PQ- 1 +R,.
-P
2

(2.25)

Subtracting (2.23) from (2.25) d ~ ~ that YQ- w s equivalently, y2Zfi 1 Q/2PQ/2. Thus,

0 Or, 9

Y2~~~~(Q2PQ2)=~max(~2Q~2)~X~1~2Q~~2)
= X-(PQ)

~ = A ( Q - + (Q- Q)AT+ Q)

(2.21) ~ ~

which yields (2.24).

which, since A is asymptotically stable,is equivalent to

As discussed in the previous section, the replacement of (2.12) by (2.14) enforces the Hm disturbance attenuation constraint and Finally, (2.19) follows immediately from (2.18). 0 yields an upper bound for the L2 performance criterion. That is, Remark 2.6: Note that (2.15) is actually a closed-loop given a compensator (A,, B,, C,) for which there exists a disturbability condition which is not concerned with control as nonnegative-definite solution to (2.14), the actual L2performance such. This condition guarantees that the syjtem does not possess J(A,, B, , C,) of the compensator is guaranteed to be no worse undisprbed unstable modes. Of course, if Vis positive definite or than the bound given by 3 (A,, B,, C,, Q ) . Hence, 3 (A,, B,, C,, Q ) can be interpreted as an auxiliary cost which leads to the (A, D) is controllable, then (2.15) is satisfied. Lemma 2.1 shows that the H , disturbance attenuation con- following mathematical programming problem. Auxiliary Minimization Problem: Determine (A,, B,, C,, Q ) straint is automatically enforced whe! a nonnegative-definite solution to (2.14) is known to exist and A is asymptotically stable. which minimizes 3(A,, B,, C,,Q ) subject to (2.13) and (2.14). It follows from Lemma 2.1 that the satisfaction of (2.13) and Furthermore, all such solutions provide upper bounds for the actual closed-loop state covariance Q along with a bound on the L2 (2.14) along with the generic condition (2.15) leads to: 1) closedperformance criterion. Next, we present a partial converse of loop stability; 2) prespecified H , performance attenuation; and 3) Lemma 2.1 which guarantees the existence of a unique minimal an upper bound for the L2 performance criterion. Hence, it nonnegative-definite solution to (2.14) when (2.17) is satisfied. remains to determine (A,, B,, C,) which minimizes J(A,, B,, The minimal solution is desirable since it yields the least C,, Q ) , and thus provides an optimized bound for the actual L2 performance J(A,, B,, C,). Rigorous derivation of the necessary performance bound in (2.19). This was first pointed out in [$5]. Lemma 2.2: Let (A,, B,, C,) be given, suppose A is conditions for the auxiliary minimization problem requires addiasymptotically stable, and assume the disturbance attenuation tional technical assumptions. Specifically, we restrict (A,, B,, constraint (2.17) is satisfied. Then there exists a unique nonnega- C,, Q ) to the open set tive-definite solution Q satisfying (2.14) and such that A + X P {(A,., B,, C,, Q ) : Q E Pfi,A+y-QI?, y-QRm is asymptotically stable. Furthermore, this solution is also minimal. is asymptotically stable, Proof: The result is an immediate consequence of [30, pp. and (A,, B,, C,) is controllable and observable}. (3.1) 150 and 1671, using Theorems 3 and 2. along with the dual version of [28, Lemma 12.21. The proof of miniGality is given in Remark 3.1: The set X constitutes sufficient conditions under r29i. 0 which the Lagrange multiplier technique is applicable to the - Remark 2.7: To further clarify the relationships between the L2 auxiliary minimization problem. Specifically, the requirement and Hm aspects of the problem, we note that the closed-loop that Q be positive definite replaces 12.13) by an open set system can be represented by two possibly different transfer constraint, the stability of A + y-2QRmserves as a normality functions. Specifically, with respect to the L2 cost criterion, the condition, and (A,, B,, C,) minimal is a nondegeneracy condiclosed-loop transfer function between disturbances and controlled tion. Note that the stabilizability condition (2.15) and stability variables is given by the triple (A, d;, 8)while for the Hm condition (2.16) play no role in determining solutions of the constraint $e c_los$-loop transfer function (2.5) corresponds to auxiliary minimization problem. the triple (A, D, E,). The following result presents the necessary conditions for Finally, it can be shown that the closed-loop Riccati equation optimality in the auxiliary minimization problem. The proof of (2.14) also enforces a constraint on the norm of @e H a e e l this result is given in the Appendix as a special case of the operator corresponding to the closed-loop system ( A , D, E,) corresponding result for reduced-order dynamic compensation when Q is positive definite. Thus, let P E Nfl denote the solution considered in Section VI. As mentioned previously, we assume to that Rz, = P2RZ, where /3 1 0. Furthermore, for arbitrary Q, P E M define 0 =ATP+ PA + 8, (2.23) s A (Ifl+/32y-2QP)-. (3.2) and note that P and Q [satisfying (2.12)] are the observ_abil$y_and controllability Gramians, respectively, of the system (A, D, Em). Since the eigenvalues of QP coincide with the eigenvalue: of the As show_n in_[31], the norm of the Hankel operator corresponding nonnegative-definite matrix P 1 / 2 Q P 1 / it , follows that QP has 2 nonnegative eigenvalues. Thus, the eigenvalues of I,, + pzy-2QP to (A, D, E,) is given by Xgx(Po). Proposition 2.1: Suppose there exists Q E Pi satisfying are all greater than one so that the above inverse exists.

Q-e=1 :

III. THEAUXILIARY MINIMIZATION PROBLEM NECESSARY AND FOR CONDITIONS OY~IMALITY

eAt[y-2Ql?mQ]eATfr 0 . dt

(2.22)

BERNSTEIN AND HADDAD: LQG CONTROL

297
system of

Theorem 3.1: If (A,, B,, C,, Q) E 3c solves the auxiliary minimization problem then there exist Q, P, Q E RIn such that
A,= A - QC - CPS+ - ~ Q R ~ , ,
B,=QCTV;',
C,= -R;'BTPS,

Q, P, and Q. The Q and P equations are similar to the estimato: and regulator Riccati equations of LQG theory, while the Q

three modified algebraic Riccati equations in variables

and such that Q, P, Q satisfy

O = A Q + Q A T +vl + ~ Q R , , Q - Q . Z Q .

O=(A + y - 2 [ Q + Q ] R I , ) ~ P P ( A + T - ~ [ Q +QIRl,) +

equation has no counterp-&t in the standard theory. Note that the Q equation is decoupled from the P and Q equations and thus can (3.4) be solved independently. The P equation, however, depends on Q. Thus, regulator/estimator separation holds in only one (3.5) direction which clearly shows that the certainty equivalence principle is no longer valid for the L2/H, design problem. (3.6) Furthermore, since the P and Q equations are coupled, they must be solved simultaneously. Finally, note that if the H , disturbance then the attenuation constraint is sufficiently relaxed-, i.e., y -+ a, P,equation becomes decoupled from the Q equation and thus the Q equation becomes superfluous. Furthermore, the remaining Q (3*7) and P equations separate and coincide with the standard LQG result.

(3.3)

+RI - S TPCPS, (3.8)

IV. SUFFICIENT CONDITIONS H , DISTURBANCE FOR ATTENUATION

In this section we combine Lemma 2.1 with the converse of Theorem 3.1 to obtain our main result guaranteeing closed-loop stability, H , disturbance attenuation, and an optimized L2 per+ y - 2 Q ( ~ l m + p ~ T ~ ~ ~ ~ ) (3.9) ~ E ~ . Q + formance bound. Theorem 4.1: Suppose there exist Q, P , Q E RIn satisfying Furthermore, the auxiliary cost is given by (3,7)53.9) and let (A,, B,, C,, Q) bg given by ( . ) ( . ) Then 33-36. 3 ( A , , B,, C,, Q)=tr [ ( Q + Q ) R 1 + Q S T P Z P S ] . (3.10) (A, D ) is stabilizable if and only if A is asymptotically stable. In this case, the closed-loop transfer function H(s) satisfies the H , Conversely, if there exist Q, P, Q E RIn satisfying ( . ) ( . ) attenuation constraint 37-39, then (A,, B,, C,, Q) given by ( . ) ( . ) satisfies (2.13)and 33-36 (2.14)with auxiliary costJ2.20) given by (3.10). I1H ( S ) l l m s y (4.1) Remark 3.2: If Q and Q are nonnegative definite, then the fact that the definiteness condition (2.13)is satisfied can easily be seen and the L2 performance criterion (2.7)satisfies the bound by writing Q as J ( A , , B,, C,)str [ ( Q + Q ) R 1 + @ j T P C P S ] . (4.2)
-

o = ( A - C P S+

QR ) Q + &(A - CPS+ y - QR )

and 8 which yield the minimal solution to (2.14). Remark 3.3: Setting P = 0, or equivalently, Ezm = 0, specializes Theorem 3.1 to the cheap H , control case in which H , attenuation between disturbances and controls is not con0 strained. In this case S = I,,, Q is given by (3.6),and (3.3)-(3.5) Remark 4.1: In applying Theorem 4.1 it is not actually become necessary to check (2.15)which holds generically. Rather, it suffices to check the stability of A directly which is guaranteed to A,=A -Q ' S ; - X P + ~ - ~ Q R ~ , , (3.11) be equivalent to (2.15). In applying Theorem 4.1 the principal issue concerns condiB,=QCTV;', (3.12) tions on the problem data under which the coupled Riccati c,= - R ; ' B ~ P (3.13) equations (3.7)-(3.9) possess nonnegative-definite solutions. Clearly, for y sufficiently large, (3.7)-(3.9)approximate the standard LQG result so that existence is assured. The important where Q satisfies ( . ) and (3.8) (3.9) 37, and become case of interest, however, involves small y so that significant H , disturbance attenuation is enforced. Thus, if (4.1)can be satisfied O = ( A + y -2[ Q+ &]Rim) P +P ( A + y -2[ Q + & ] R I , ) T for a given y > 0, it is of interest to know whether one such + R , - P C P , (3.14) controller can be obtained by solving (3.7)-(3.9).Lemma 2.2 guarantees that (2.14)possesses a solution for any controller o = ( A - C P+ y - QR',) Q + &(A - X P+ - ~ Q R ,),T satisfying (2.17).Thus, our sufficient condition will also be + - 2 Q ~ l m Q + QZQ. (3.15) necessary as long as the auxiliary minimization problem possesses at least one extremal over 3c. When this is the case we have the Finally, the auxiliary cost reduces to following immediate result. over Proposition 4. I: Let y * denote the infimum of 11H(s)11 , d(& B,, C,, Q)=tr [ ( Q + Q ) R I + Q P C P I . (3.16) all stabilizing nth-order dynamic compensators and suppose that the auxiliary minimization problem has a-solution for all y > y * . Numerical solution of (3.7),(3.14),and (3.15)is discussed in Then for all y > y * a e r e exist Q, P, Q E M" satisfying (3.7)Section VIII. (3.9). Remark 3.4: Note that if both = 0 and R I , = 0, then Unlike the standard LQG result involving a pair of separated Theorem 3.1 specializes to the standard LQG result. Riccati equations, the new result enforcing H , disturbance Theorem 3.1 presents necessary conditions for the auxiliary attenuation involves a nonstandard coupled system of three minimization problem which explicitly synthesize extremal con- modified Riccati equations. The asymmetry of these equations trollers (A,, B,, C,). These necessary conditions comprise a suggests the possibility of a dual result in which the modifications

As mentioned in Section 1 , it is desirable to determine solutions Q 1

Proofi The converse portion of Theorem 3.1 implies that Q given by (3.6) satisfies (2.13) (2.14) and with auxiliary cost given 31) . by ( . 0 .It now follows from Lemma 2 1 that the stabiliza_bility is condition (2.15) equivalent to the asymptotic stability of A, the H , disturbance attenuation constraint (2.17) holds, and the performance bound (2.19),which is equivalent to ( . ) holds. 42,

298

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 34, NO. 3, MARCH 1989

to the standard P and Q Riccati equations are reversed. Such a dual result will generally be different from Theorem 4.1 and thus will yield an improved bound for particular problems. This poidt was demonstrated in [16] for the problem of robust performance analysis. Due to space limitations, however, we give only a brief outline of the dual H , results. Note that J(A,, B,, C,) given by (2.10) is also given by J(A,, B,, C,) = tr

C,= - R i l B T Z ,

(5.5)

3 ( A c ,B,, C,,Q ) = ~ ~ [ ( Q + Q ) R I + Q Z W . (5.6)

pv

(4.3)

Proof: Using the identities P = (Z, - P2y-2ZQ)- l Z = Z(Z"-PZy-2QZ)-L which follow from (5.1), equation (5.2) can be obtained by forming the new equation
( Z , - P2y-2ZQ)(3.8)(Z, P 2 y -'QZ) -

wheEe E RIfi is the unique solution to (2.23) with I, replaced ? by R . Next, utilizing (4.3) in place of (2.10), the H , disturbance attenuation constraint (2.6) can now be enforced by replacing (2.23) by the Riccati equation

+ P 2 y -2Z(3.9)Z.

(5.7)

0=AT6 + @A+ - 2 6 y

v"6 + R

(4.4)

where p has the same form as p but may involve weights VI, , and V2,. Note that (4.4) is merely the dual of (2.14). We also require the condition dual to (2.15) given by

Finally, (5.3)-(5.5) are restatements of (3.9), (3.3), and (3.5) with Z = PS. 0 Having obtained a single equation (5.2) for Z = PS by combining (3.8) and (3.9) for P and $, it is of interest to know whether (3.8) for P can be recovered from (5.2) and (5.3). Propo$tion 5.2: Let Q E M", fi > 0, suppose there exist Z E P" and Q E M" satisfying (5.2) and (5.3), and assume that
P ( Z Q )< P

-v.

(5.8)

(4.5) Then P 2 (Z-I - fi2y-'Q>-' is positive definite and satisfies (3.8). Furthermore, P satisfies Z = PS. Proof: If (5.8) holds, then it can be shown that P as defined and that A + y-2pm6 asymptotically stable. Once again, the be sufficient conditions for Ha disturbance attenuation involve a above is positive definite. Reversing the proof of Proposition 5.1, coupled system of three modified Riccati equations dual to (3.7)- (3.8) can be recovered by forming (3.9). Similar remarks apply to $e redlced-o_der case given by 0 Theorem 6.1 below, Finally, jf R , = R and V , = V, then it can ( z n - ~ ~ y - ~ z Q ) - ~ [ ( 5 . 2 ) - ~ ~ y - ~ z ( 5 . 3 ) z ] ( Z n - ~ ~ y - ~ Q z ) - ~ . be shown that tr QR = tr 6 V and thus the solutions to the primal Although Proposition 5.2 allows us to reconstruct (3.8) for P , it and dual problems coincide. can only be utilized when (5.8) holds. This fact raises a question as to the sufficiency of (3.7), (5.2), and (5.3) in the absence of v. ALTERNATIVE OF THE RICCATI EQUATIONS FORMS (3.8). It turns out that the matrices P and Z need not actually In this section we develop alternative forms of the Riccati satisfy (3.8) and (5.2) to enforce the H , performance constraint equations (3.7)-(3.9). These alternative forms provide further (2.17) since only the Q and Q equations are required. Rather, P insight into the structure of (3.7)-(3.9) and, in certain cases, are can be viewed as a parameterization of Z which, in turn, is a simpler and thus are easier to solve computationally. This section parameterization of the gains A, and C, given by (5.4) and (5.5) which yield a controller satisfying the desired H , performance. also provides connections between our approach and [26]. First we note that-the gains (3.3), ( 3 . 9 , and (3.10) do not These observations are summarized by the following result which depend upon P and Q individually, but rather only upon the term does not require that Z be obtained by solving (5.2). Proposition 5.3: Let Z E M" and suppose there exist Q, Q E Z A PS. Thus, it is of interest to know whether (3.8) and (3.9) can be transformed to yield an equation which characterizes Z M" satisfying (3.7) and (5.3). Then (A,, B,, C,, Q) given by directly. The following tesult summarizes useful properties of 2. (5.4), (3.4), ( 5 . 3 , and (3.6) satisfy (2.13) and (2.14). Thus, Lemma 5.1: Let P , Q E M" and define Z A PS. Then Z = (2.15) and (2.16) are equivalent, and, in this case, (2.17) and Z T = STP, where S T = (I, + fi2y-2PQ)-', and Z is (2.19) hold. Proof: The result follows by direct verification of (2.14).0 nonnegative definite. If, in addition, P is positive definite, then Z Proposition 5.3 shows that the H , constraint (2.17) is enforced is positive definite and for arbitrary Z E M as long as (3.7) and (5.3) can be solved for Z = (P - 1 + P 2 y - 2 Q) - 1. (5.1) Q and 0. The price we pay for using arbitrary Z is that we no longer are assured that Z is obtained from (5.2) or from Z = PS Proof: The result (5.1) is immediate. The remaining results where P satisfies (3.8). Since P arises from the Lagrange can be obtained by replacing P by P + el,,, where E > 0, and multiplier for the constraint (2.14) [see (A.3)], it follows that an 0. 0 arbitrary choice of P (or 2)may fail to minimize the L2 auxiliary taking the limit as E Prqposition 5.1: Let Q E M" and suppose there exist P E P" cost (2.20). Thus, regarding P and Z as free parameters and Q E M" satisfying (3.8) and (3.9). Then Z A PS satisfies effectively ignores the L2 aspect of Theorem 4.1.

( E , A ) is detectable

-+

O=(A + y - 2 Q R l m + ~ - 2 Q [ R 1 , - P 2 R l ] ) T Z

It is also of interest to introduce yet another transformation of (3.7)-(3.9) by defining

Y 4 (Z-

+ Z ( A + y-2QR1m + y - 2 Q [ R ~ m - P 2 R ~ 1 )
+ R , -z(C+P2y-4&[Rla-P2R1]Q)z

+ P2y-2 Q )

-I

= ( P- I

+ P2y- 2 [ Q + Q ] )- I

(5.9)

+ ~2~
o = ( A - cz + y -

-~ZQSQZ
Q R ~ , Q + &(A - cz + - ~ Q R , , ) ) T

and (3.9) is equivalent to

when P is positive definite. As in Lemma 5.1, Y is also positive (5.2) definite. Proposition 5.4: Let Q E M" and suppose there exist P E P n and Q E M" satisfying (3.8) and (3.9). Then Y defined by (5.9) satisfies
O=(A + y - 2 [ Q + & 1 t R ~ m - P 2 R ~ 1 ) T Y

+ y - 2 Q ( ~ l m + ~ 2 ~ ~ ~(5.3) + ~ E ~ . ) Q

+ Y ( A +V2t + Q l t R ~ m-P'RII) Q

Furthermore, (3.3), ( 3 3 , and (3.10) become

A,=A
-

+ R , + P2y - 2 YV, Y - YE Y
(5.4)
-P2y-4

QC - C Z + -~~ Q R ~ , ,

Y ( Q +Q)(RI,-P'RI)(Q+ Q) Y

(5.10)

BERNSTEIN AND HADDAD: LQG CONTROL

299
B,= QCTV;' ,

and (3.9) is equivalent to

O = ( A - C[ Y - ' - p Z y - ' Q ]

(5.22) (5.23)

-' + y - Z Q R ~ m ) $ -'
(5.11)

c,=-R,-,'BTZ,.

+Q(A-X[ Y - ' - ~ z y - 2 Q ] - ' + ~ - 2 Q R ~ m ) T

+y-'Q(RI,
*

+ p 2 [ Y - ' -BZy-'Q]

X[ Y - ' - / 3 2 y - 2 Q ] - ' ) Q + QSQ.

Then (A, D)is stabilizable if and only if A is asymptotically stable. In this case, the closed-loop transfer function H(s) satisfies the H , disturbance attenuation constraint

Furthermore, (3.3), ( 3 . 9 , and (3.10) become A , = A -QC-X( Y - ' - P 2 y - 2 Q ) - ' + y - 2 Q R I m , (5.12)

IIH ( s ) l l m

57

(5.24)

and the LZperformance criterion (2.7) satisfies the bound

J(Ac, Be, Cc)stdQRlm+QCQZm].

c,=- R - ' B = ( Y - ~ - B ~ ~ - ~ Q ) - ~ (5.13) ,

dU,,
B,, Cc,Q)=tr [(Q+&)Ri + $( Y - - PZy - Q ) - C ( Y - I - /3 'y

(5.25)

Proof: First note that it follows from (5.18) that

'

Q )- '1.

(5.14)

- ( A +y-ZQR1, + Z , ' R 1 , ) = Z , [ A

+ y -2QRlm
(5.26)

Proof: To obtain (5.10), form

Y [ z - '(5.2)z- + BZy -2(3.7)] Y.

+ ( y - ' Q S Q - C)Z,]Z,'

'

The following result allows us to recover (3.8) for P from (5.10) and (5.11). Propo$tion 5.5: Let Q E M", fi > 0, suppose there exist Y E P n and Q E M" satisfying (5.10), (5.11), and assume that
P(

and thus (5.19) implies that -(A y-'QR,, + Z;'RI,) is asymptotically stable. It now follows from (5.20) that there exists N E P" satisfying
O = - ( A +y-'QRl,

+ Z,'Rl,)'N-N(A

+y-'QRlm

+ Z , ' R ~ , ) + ~ - 2 ( R ~ , + Z , C Z , ) . (5.27)

Y [ Q + $1)<8-2r2.

(5.15)

Then p 4 (y-1 - BZy-z[Q + & ) - I is positive definite and satisfies (3.8). Proof: The result follows by reversing the proof of Proposition 5.4. 0 By specializing further, it is possible to achieve even greater simplification. Specifically, we consider the case in which the Lz and H , weights are equalized, i.e.,
RIm=RI, B = 1 .

It can now be shown that 0 = y Z Z i 1- N - satisfies (5.3)with /3 = 1 and Z = Z , . Furthermore, (5.8) is satisfied so that the hypotheses of Theorem 4.1 are verified. The expression (5.25) 0 now follows by direct substitution. Finally, we consider a simplified version of Proposition 5.4. Proposition 5.7: Assume (5.16) is satisfied and suppose there exist Q E M" and Y , E Pn satisfying
O=AQ+QAT+ Vi+y-'QRl,Q-QSQ, O=A'Y,+ Y,A+Rl,+y-2Y,VlYmY,CY,,

(5.28) (5.29) (5.30) (5.31)

(5.16)

P(Q Y In this case it is always possible io eliminate (5.3) and (5.11) by = y2Y-' - and such that noting that they are satisfied by Q = y2Z-' and Q, respectively. However, although this solution enforces the H , A + (Y-' VI- X) Y, is asymptotically stable constraint, it can be seen from the resulting form of 3 that this solution does not correspond to the m n m l solution Q of (2.14). and iia Hence, we impose additional assumptions which allow us to ( A + Y i ' R I , , y - ' [ R I , + ( Y ; ' - y - ' Q ) - ' directly characterize the solution which yields the minimal performance bound. We are indebted to D. Mustafa for clarifying X( Y , I - y - 2 Q ) ] " 2 ) is observable. this point in [45] where it is also shown that the auxiliary cost Furthermore, let (A,, B,, C,) be given by (2.20) is equivalent to an entropy integral. Proposition 5.6: Assume (5.16) is satisfied, suppose there A , = A - QC - X( ~ i 1 - - Z Q )- l + y - ~ Q R ' , , y exist Q E N" and 2 E P" satisfying ,

(5.32)

(5.33) (5.34)

o = A Q + QA T + V ,+ y - ~ Q R ' Q - QSQ, , O = ( A + ~ - ' Q R I , ) ~ Z ,Z,(A + y L Z Q R l m ) + + RI, - Z m ZZm + Y -'Zm QEQZ,

and such that A +y and

(5.17)

B, = QCTV ; I ,

c = - R z2 , BT(
(5.18)

U;' -

- 2 ~ -1. )

(5.35)

-' + (y -' Q - X)Z, QRlm QC

Then (A, D)is stabilizable if and only if A' is asymptotically stable. In this case, the closed-loop transfer function H(s) satisfies the Hmdisturbance attenuation constraint

is asymptotically stable

I1 H(s)llm 5 7
and the L2 performance criterion (2.7) satisfies the bound

(5.36)

(5.19)

J(A,, B,, C c ) s t r [ Q R l m + Q ~ Y ,(' - y - 2 Q ) - ' ] . Q (5.20)

Furthermore, let (A,, B,, C,) be given by A,= A - QC- X Z , + y - Z Q R ~ , , , , with

(5.37)

Pro08 The proof is similar to the proof of Proposition 5.6 0 = y z Y , - Q - N-', N satisfies where

o= - ( A + Y,'R,,)TN-N(A + Y,'R',)
(5.21)

-2

[R I ,

+ ( Y 2' - y -'Q) - 'I;( Y , ' - y -'Q)- 'I.

300

IEE:E TRANSACTIONS ON AUTOMATIC CONTROL,

VOL. 34, NO. 3, MARCH

1989

Remark 5.1: The solutions Q and Y, of (5.28) (5.29) and are analogous to the matrices Y, and X , of [26], while (5.30) corresponds to condition 5.2(iii) of [26]. Note that by letting y + 03, (5.25)and (5.37)coincide with 5-77a of [l]and the LQG result is recovered. Remark5.2: It is interesting to note that (5.17) (5.18) and with controller gains (5.21)-(5.23)are already known since they are identical to the optimality conditions for the linear-exponential-ofquadratic-Gaussian problem treated in [33] (see also [34] and [ 5 ) Specifically, see (3.1)and (4.1)on pp. 603 and 609, 3]. respectively. As shown in [33],minimizing the criterion
J=

(6.13)

Then, (A, 6) is stabilizable if and only if A is asymptotically stable. In this case, the closed-loop transfer function H(s) satisfies the disturbance attenuation constraint

e,

IIH(s)llm 5 Y
and the L2 performance criterion (2.7)satisfies the bound
J(A,, B,, C,) s tr[(Q+ Q)Rl + QS TPCPS].

(6.14)

(6.15)

lim
t+m

@&2(xTRi~+~TR24

Remark 6.1: It is easy to see that Theorem 6.1 is a direct To generalization of Theorem 4.1. recover Theorem 4.1,set n, = leads to the pair of modified Riccati equations (5.17)and (5.18) n so that r = G = r = I, and 7L = 0. In this case the last term with y-2 replaced by p . This implies that the exponential-of- in each of ( . ) ( . ) can be deleted and (6.8) becomes 65-68 quadratic design problem effectively enforces a bound of p - on superfluous. Furthermore, (6.5)-(6.7)now reduce to (3.7)-(3.9), the H , norm of the closed-loop transfer function. There also exist as expected. If, furthermore, 0 = 0 then S = I,, so that (6.5)fundamental connections with the problem of entropy maximiza(6.7)now reduce to the cheap H , control case given by (3.71, tion [43]-[45]. (3.14), (3.15). Alternatively, setting y = 03 and retaining the and reduced-order constraint n, < n yields the result of [21]. VI. EXTENSIONS REDUCED-ORDER DYNAMIC TO COMPENSATION Remark_6 2 By introducing a new variable Z = PS = ( P .: as in becomes In this section we extend Theorem 4.1 by expanding the + ,62y-2Q)-1Section V, (6.6) formulation of Section III to allow the compensator to be of fixed O = ( A + Y - ~ Q R+ Y - ~ Q [ R ~ ,2 R ~ ] ) T Z ~, -P dimension n, which may be less than the plant order n. Hence, in + Z ( A +y-'QRIm + y - 2 Q [ R ~ m - P 2 R ~ l ) this section define n' = n + n,, where n, 5 n. As in [21]this constraint leads to an oblique projection which introduces +RI - Z ( C + P 2 y - 4 Q [ R I , - P 2 R i ] Q ) Z additional coupling in the design equations along with an additional equation. _The_ following lemma is requifep. (6.16) + 7 : ZCZTL + P2y-2Z(QCQ- 71 Q 8 Q r ; ) Z Lemma 6.I: Let Q, E M" and suppose rank QP = n,. Then P there exist n, x n G , r, and n, X n, invertible M , unique except which specializes to (5.2)when n, = n, i.e., r L = 0. When (5.16)holds, (6.16)becomes for a change of basis in Kin,, such that QP=GTMT,
(6.1) (6.2) (6.3) (6.4)

O = ( A +y-2QR1m)TZm+Zm(A +y-'QRIm)

rG

= Znc.

Furthermore, the n x n matrices r P GTr,

71

+R I , - Z , CZ , + 7: Z , X Z , 7 1 + y -2Z,( QCQ - 7L QZQ7 T ) Z , .

(6.17)

Analogous equations for Y defined by (5.9) can also be developed. VII. ANALYSIS THE DESIGN OF EQUATIONS Before developing numerical algorithms, it is instructive to analyze the design equations to determine existence and multiplicity of nonnegative-definite solutions. Although a detailed theoretical analysis remains an area for future research, in this section we present a simplified treatment which highlights important asymptotic properties of the equations. it turns out that several key properties are discernible by considering the scalar case n = 1. Although many of these properties can be developed for general n, the simplified scalar case suffices for obtaining a useful qualitative analysis. Here we consider only (3.7),(3.14),and

4 In-7

are idempotent and have rank n, and n - n,, respectively. Proof: Conditions (6.1)-(6.4)are a direct consequence of [36,Theorem 6.2.51. U Theorem 6.1: Let n, 5 n, suppose there exist Q,P,Q, E M" satisfying
0 = AQ + QA

+ V I+ y - 2 QR , - QZ:Q + Q
1

71

QC Q7 : ,

(6.5)

O = ( A + y - 2 [ Q + Q ] R ~ m ) T P + P (Y - ~ [ Q + Q ] R I , ) + R I + A
- S TPCPS+7 ; S TPEPS~L (6.6) ,

o = ( A - C P S+ y - Q R ~IQ + & ( A - G P S+ - ~ Q R ] ,T) , + y - Qwl, + p2sT~~ P S )Q + QC Q - Q C Q ~; ,

O = ( A- Q C + ~ - ~ Q R ~ , ) ~ P + P ( A y - 2 ~ ~ l m ) - QC +

(3.15).
(6.7)

+STPXPS-~TSTPCPS7,,(6.8)
rank Q=rankP=rank QP=n, and let (A,, B,, C,, Q) be given by
(6.9)

Since the Q equation (3.7) decoupled from (3.14) (3.15), is and it can be analyzed separately. It is easy to-see that there exists a as unique nonnegative solution for y > (Rl/Z)II2 in the case of a standard Riccati equation with stabilizability and detectability hypotheses. Furthermore, it can be seen that for
(RI /[ C + ( A2/ V i ) ] ) II2<y< (RI

A , = r ( A - QC - CPS + y -2QRlm) ', G B, = r QC V ; I ,

(6.10)
(6.11)

c,=- R ; I B ~ P S G T ,

(6.12)

there exist two nonnegative solutions when A is stable and zero nonnegative solutions when A is unstable. Below this lower bound for y, nonnegative solutions Q do not exist. This result thus indicates (as in LQG theory [42])a lower bound to the achievable H , disturbance attenuation as determined by the sensor noise intensity V2appearing in E. Since the P and Q equations (3.14) (3.15)are coupled, they and

BERNSTEIN AND HADDAD: LQG CONTROL

30 1

must be analyzed jointly. Since (3.15) is a standard Riccati two neutrally stable poles. The problem data are as follows: equation it follows under generic hypotheses that it possesses n = n , = 8 , m = l = 1, q = p = 2 , ezactly one nonnegative-definite solution for all values of Q and Q. The analysis of the Q equation is, however, more involved. It can be shown that the existence of real solutions is a complicated 1 0 0 0 0 0 0 - 0.161 function of y, Q , and P. When real solutions do exist, it follows 0 1 0 0 0 0 0 - 6.004 that there exist either zero or two nonnegative-definite soluGons. - 0.5822 0 0 1 0 0 0 0 To obtain further qualitative insight into the solutions P and Q, we fix y and allow R2 4 0, that is, the cheap L2 control case. It thus A = - 9.9835 0 0 0 1 0 0 0 - 0.4073 0 0 0 0 1 0 0 follows that P (RIC)2and that either Q 2y2(C/R1)12 or 0 0 0 0 0 1 0 - 3.982 Q 1/2CQ2(CR1)-12,which correspond to the previously 0 0 0 0 0 0 1 0 mentioned pair of solutions satisfying (3.15). This result thus 0 0 0 0 0 0 0 0 indicates that an arbitrarily small H , disturbance attenuation constraint y can be achieved [subject to the solvability of (3.7)] by sufficiently increasing the Lz controller authority. That is, since solutions exist in the cheap L2 control case, the H , disturbance attenuation constraint is achievable. The ability to achieve small y is also attributable to the fact that since 0 = 0, H , disturbance attenuation to the control variables is not limited in (3.7), (3.14), B = and (3.15) as in Theorems 3.1 and 6.1. Of course, as is w! e known, it is not possible to make y 0 by letting C 00 and C 0.1045 00 when the system possesses nonminimum phase zeros. Also, 0.9955 note that both of the asymptotic solutions to (3.15) are guaranteed to yield the bound (4.1). The solution of interest, however, is Q = O(C-_/2) since it clearly yields a lower value of & 4 , , B,, C,, Q) than Q = O(C12). 0 0 0 0 0.55 1 1 1.32 18 El = E l , = ~ 0 0 0 0 0 0 0 0 1

-+

AND RESULTS VIII. NUMERICAL ALGORITHM ILLUSTRATIVE

In this section we describe a numerical algorithm which has been developed and implemented for solving the coupled Riccati equations (3.7), (3.14), and (3.15). We also present numerical results for an illustrative example. Coupled modified Riccati equations arise in a variety of problems and homotopic continuation methods have been shown to be particularly successful [23]-[25]. To solve (3.7), (3.14), and (3.15) we have implemented a simplified continuation method involving the constraint constant y. The idea is to exploit the fact that for large y the problem is approximated by LQG which provides a reliable starting solution. The continuation parameter y is then successively decreased until either a desired value of y is achieved or no further decrease is possible. This algorithm is now summarized. Let E > 0 denote a convergence criterion. Algorithm 8.1: To solve (3.7), (3.14), and (3.15), perform the following steps: Step 1: Initialize y > 0. Step 2: Solve (3.7) for Q. Step 3: Let k = 0, Qo = 0 . Step 4: Solve (3.14) for p k + = P with Q = Qk. Step 5: Solve (3.15) for $ k + l = Q with P = P k + l . step 6: If k 2 1 check for I I P k + l - P k l I < E and 11 & + I -

Step 7 If convergence is not achieved in Step 6 (or k = 0) let k : k + 1 and go to Step 4; otherwise decrease y and go to Step 2. Steps 2, 4, and 5 were carried out using a standard Riccati solver [37] which proved to be reliable even when the quadratic term was indefinite or nonnegative definite. For instance, for the example considered below, the term Y - ~ R- c was indefinite for , all finite y. The crucial step in the algorithm is the decreasing of y in Step 7. Significant effort was devoted to providing a smooth transition to smaller values of y without sacrificing computational efficiency. The development of more sophisticated continuation algorithms remains an area for future research. The example considered was formulated in [38] and was considered extensively in [24], [25], and [39] to compare reducedorder design methods. The example is interesting since it possesses a complex pair of nonminimum phase zeros due to the fact that the physical system (coupled rotating disks) has noncolmated sensors and actuators. The plant is of eighth order and has
+

&k)(

<

E.

With the problem data as given, the LQG controller was found to yield a closed-loop H , performance of 1.39 (i.e., 2.87 dB above unity gain). Using Algorithm 8.1 we obtained a solution for y = 0.52 for a net H , performance improvement of 8.7 dB (see Fig. 1). Note that this result is consistent with [3, Proposition 8.11 which implies that the maximum ratio of the H , performance of the optimal L2 controller to the H , performance of the optimal H , controller can be no more than twice the number of right-halfplane zeros, which for the present problem with two nonminimum phase zeros corresponds to a factor of 4 (i.e., 12 dB). Our numerical experience revealed two interesting features. First, the loop between Steps 4 and 6 converged reliably. of However, a critical value ymin y was invariably found below appears to which solutions could not be computed. This value ymin represent the best achievable H , performance for the given Lz weights. Second, for each value of y 2 ymin which a solution for was computed, the actual H , performance was close to this value revealing that the H, bound is tight. For example, the actual worst-case attenuation of the y = 0.52 design shown in Fig. 1 is 0.511. Controller characteristics are given in Table I and are plotted in Fig. 2 for several values of y. Note that in each case the L2 performance bound is within 30 percent of the actual L2 performance. IX. FURTHER EXTENSIONS The results obtained herein can readily be extended in several directions. These include the treatment of parameter uncertainties [131-[ 151, [46], extensions to controllers with static feedthrough [32], and the inclusion of cross-weighting terms (xr(t)R12u(t)) and noise correlation (DIDT # 0). Finally, as mentioned in Remark 5.2, connections with the exponential-of-quadratic cost criterion [33]-[35] and entropy maximization [43]-[45] are of interest.

302
FREO

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 34, NO. 3, MARCH 1989

lHZl
1
10

10'-3

LQO

10'-2

1O'-I

-= P Tz QIZ + Pz Qz = 0,
-= PzB, Vz ( P Qi

ad: aAC

(A.5) (A.6)

ad: aBc

+ Tz + Pz Q Tz)CT = 0,

ad: -=RzCcQz+ aC C

P2y-zRzCc(PiQiz + P ~ z Q z ) ~ Q i z

z -40.0 U)

+ BT(PiQ i z +
Expanding (2.14) and (A.4) yields

PizQz)=O. (A.7)

-50.0-60.0-

Fig. 1.

O=AQI+ Q i A T +BCcQrz+QizCrBT+y-2Q1RimQi

+ P Z y -'Qiz CrRz CcQTz + Vi

+ 8%-

(A. 8)

O=AQn + QizA + BGQz + Qi CTBT+Y -'QIRim QIZ :

Z ~ I z ~~ 4 . 9 z ~ c ~ z ~ p )
(A.lO)

TABLE I
H , Attenuation
Constraint 7 Actual I , Attenuation

O=AcQz + QzA f + B J Q n + QT,CTBr+Y-2QTzRimQiz

Actual Lt Performance J(A.,B.,C.)
.143 ,146 ,151 .168 ,176

LI Performance
Bound

+ P2y-2QzCTR2CcQ2+ B, VZBT,

llH(4ll(WG)
2
1.5 1.0 9
8

J(Ac,B.,Cc,Q)
.159 ,171
,204

o = A T ~+ p i A + C I

~ B ~ P~~B,C +P

Tz

1.39

1.18
1.06 ,855

+ Y-zRi-(PiQi +PizQTJT + Y -'(Pi Qi + Pi2 Q ) Ri m + R i

(A. 11)

0 = A T P l+ PizAc + CTB,TPz+PI BC, ~

.797
.732

,217 ,236 ,262 ,299

.I87

+ Y -'RI= ( P T Qi + Pz Q yz> z + 13% - 2 ( p IQ~~ p I 2 ~C,TR~C,, + Q )

0 =A , T P ~+ P ~ A , +PT,BC,

(A. 12) (A. 13)

7 .52

,361 ,511

,203
262

+ c , T B ~ P ~+~c,TR~c,.

APPENDIX

Proof: By a minor extension of results from [a], can (A. 10) be rewritten as

Lemma A.1: Q2 and Pz are positive definite.

PROOFOF THEOREM 6.1

To optimize (2.20) over the open set 3c subject to the constraint (2.14), form the Lagrangian
&(A,, B,, C,, Q, 6 , A)

O = ( A ,+ BcCQiz Q: Qz + Q
where

z V C

+ BcCQiz Q: 1 + 9

tr{XQZ?+[AQ+U-T +y-'QRmQ+ P]S} (A.l)

9
and

A Y -' T,RI m Qiz + P ' Q Y

-' C,TRzC,Qz +B VzBr Qz E

where the Lagrange multipliers X 2 0 and 6 E WAX' are not both zero. We thus obtain
-=

ad:

aQ

(A+y-2QZ?m)T6 S(A+y-2Ql?,)+ XR. +

(A.2)

Setting ad:/aQ = 0 yields

or Dr&in generalized inverse of Qz. Next note that since (Ac,B,) is controllable it follows from [28, Lemma 2.1 and Theorem 3.61 that (A, + B,CQizQ:, is also controllable. Now, since Qz and t are nonnegative definite, [28, Lemma 12.21 implies that Qzis positive definite. Using (A.13), similar arguments show that Pz is positive 0 definite. Since Rz, Vz, Qz, Pz are invertible, (A.S)-(A.7) can be written as

Ql Moore-Penrose is the

O = (A+y-zQZ?m)TS S(A+y-2QZ?m) XI?. + +

(A.3)

Since + y-2Ql?m assumed to be stable, X = 0 implies 6 = is 0. Hence, it can be assumed without loss of generality that h = 1 . Furthermore, 6 is nonnegative definite. Now partition ii x ii QS into n x n , n x ncr and n, x n, subblocks as
Qi

Qiz

PI

Qz] ' '=[PT,

PI2 pZ]

Thus, with h: = 1 the stationarity conditions are given by

ad: -=(A + y - z Q B m ) + @ ( A y-'QZ?,) T6 +

aQ

+R = 0,

(A.4)

BERNSTEIN AND HADDAD: LQG CONTROL

I LSO I

303

I.DO

1.20

2.goo

IT)

1.10

1 .COO

zz
! s
T

0.90

t =
0 -00

1 .LOO

0 .EO0
0 -70

0a .

160'.00

'

'

'

'

200'.00

RCIIJRL L 2

cow

'

'

.
I

260'.00

I* 10'-9

and the n, x n, n,

n,, and n, x n matrices

&, P, G, and r as
Q'=Q+Q, p,=p+P, Q12=
T,

G 6 Q ; ' Q L , M 4 Q2P2,r 4 - P ; ' P g .

Note that 7 = G T r , Clearly, Q, P, Q, and P are symmetric and Q and P are nonnegative definite. To show that Q and P are also nonnegative definite, note that Q is -upper left-hand block of the nonnegative definite matrix QQQT,where

(A.21) (A.22) (A.23)

PI2 = - PG

Q2 = r&r , P2 = GPG? T

Next note that by using (A.21)-(A.23) it can be shown that the right-hand coefficient of C, in (A.16) is given by

s 4 Zn,+@2y-21'QPGT.
To prove that is invertible use (A. 19) and (6.3) and note that Similarly, P is nonnegative definite. Next note that with the above definitions (A. 14) is equivalent to Z,,+@2y-21'~PGT=Z,,+@2y-21'~TPGT (6.2) and that (6.1) holds. Hence, 7 = G T ris idempotent, i.e., r 2

7.

It is helpful to note the identities

=Inc+ @2y-2(r@ ) ( G P G T ) . T
= - I' ' P L = 'PzI',

Q = Ql2G = G T Q $= GTQ2G, P = - PlJ

7GT=GT, rr=I',
Q=7Q,

(A. 17) (A.18) (A.19)

(A'20)

Since rQrTand GPGTare nonnegative definite, their product has nonnegative eigenvalues (see Lemma 5.1). Thus, each eigenvalue of Z + /32y-21'QPGTis real and is greater than unity. Hence, invertible. Now note that by using (6.2) and (6.3) it can be shown that

8;s

P=A,

QP= - Q12Pg.
Using (A.14) and Sylvester's inequality, it follows that rank G = rank r = rank Q I 2 rank P I 2 n,. = = Now using (A. 17) and Sylvester's inequality yields
n,=rank Q12+rank G-n,srank

The expressions (6.11), (6.12), and (6.13) follow from (A.15), (A.16), and the definition of Q. Next, computing either r(A.9)(A.lO) or G(A.12) + (A.13) yields (6.10). Substituting (A.21)(A.23) into (A.8)-(A.13) and the expression for A, into (A.9), (A.10), (A.12), and (A.13) it follows that (A.lO) = r(A.9) and (A.13) = G(A.12). Thus, (A.lO) and (A.13) are superfluous and can be omitted. Thus, (A.8)-(A.13) reduce to
O=AQ+ QA T + Vi +y-'(Q+ Q ) R l m ( Q + Q)

Qsrank Qlz=n,

which implies that rank Q = n,. Similarly, rank P = n,, and rank &P = n, follows from (A.20). The components of Q and 6 can be written in terms of Q, P,

+ @2y =PIPSQ -2Qs + ( A - I P S ) & + &(A - C P S ) T ,

(A.24)

304

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 34, NO. 3, MARCH 1989

o = [ ( A- Z P S ) & +& ( A - C P S )T + QCQ

+ Y - 2 ( Q+ Q ) R1- ( Q+ Q) - Y QRim Q + P~~ - 2 Q s T ~ ~ ~ Ts, Q ] r

(A.25)

O=(A + y - 2 [ Q + ] R I , ) ~ PP ( A + y - 2 [ Q +Q]Rl,)+R1 Q + + ( A - Q C + ~ - ~ Q R ~ , ) ~ B + P (CA ~ - ~ Q R ~ - ) (AX) -Q + ,

O = [ ( A- QC + Y - ~ Q R ~ , ) ~ P + - QA + Y - ~ Q R ~ - ) B( c

+ STPEPS]G T .

(A.27)

Next, using (A.24) + GTr(A.25)G - (A.25)G - [(A.25)GIT and GTr(A.25)G - (A.25)G - [(A.25)GIT yields (6.5) and (6.7). Similarly, using (A.26) + rTG(A.27)I - (A.27)r -[(A.27)r] and rTG(A.27)r - (A.27)r - [(A.27)F] yields (6.6) and (6.8). Finally, to prove @e converse we use (6.5)-(6.13) to-ob-tain (2.14) and (A.4)-(A.7). Let A,, B,, C,, G , r, 7,Q , P , Q , P , Q be as in the statement of Theorem 6.1 and define QI,Q12, Q 2 , P I , P12,Pz by (A.21)-(A.23). Using (6.2), (6.11), and (6.12) it is easy to ve$Q (A.6) and (A.7). Finally, substitute the definitions of Q, P , Q, P, G , r, and 7 into (6.5)-(6.8) using (6.2), (6.3), and (A.19) to obtain (2.14) and (A.4). Finally, note that
r
-m

Q Onxn, &= l o n c x on,

which shows that Q 2 0.

l;]
r m

ACKNOWLEDGMENT
The authors wish to thank Prof. P. P. Khargonekar for several helpful discussions, J. Straehla for transforming the original manuscript into TEX,A. Daubendiek, S . Greeley, S. Richter, and A. Tellez for developing the numerical algorithm and performing the calculations of Section VIII, D. Hyland, E. Collins, and L. Davis for helpful discussions and suggestions, Dr. A. N. Madiwale for providing simplifications of (6.6)-(6.8), the reviewers for several helpful comments, Prof. J. C. Doyle for helpful discussions, and D. Mustafa for providing a preprint of [45].

REFERENCES
H. Kwakemaak and R. Sivan, Linear Optimal Control Systems. New York: Wiley, 1972. G. Zames, Feedback and optimal sensitivity: Model reference transformations, multiplicative seminorms, and approximate inverses, ZEEE Trans. Automat. Contr., vol. AC-26, pp. 301-320, 1981. B. A. Francis and J. C. Doyle, Linear control theory with an H, optimality criterion, SIAM J . Contr. Optimiz., vol. 25, pp. 815844,1987. B. A. Francis, A Course in H , Control Theory. New York: Springer-Verlag, 1987. 1. C. Doyle and G. Stein, Multivariable feedback design: Concepts for a classical modem synthesis, ZEEE Trans. Automat. Contr., vol. AC-26, pp. 4-6, 1981. B. A. Francis, On robustness of the stability of feedback systems, ZEEE Trans. Automat. Contr., vol. AC-25, pp. 817-818, 1980. I. R. Petersen, A Riccati equation approach to the design of stabilizing controllers and observers for a class of uncertain linear systems, ZEEE Trans. Automat. Contr., vol. AC-30, pp. 904907, 1985. A. R. Galimidi and B. R. Barmish, The constrained Lyapunov problem and its application to robust output feedback stabilization, ZEEE Trans. Automat Contr., vol. AC-31, pp. 410-419, 1986. I. R. Petersen and C. V. Hollot, A Riccati equation approach to the stabilization of uncertain systems, Automatica, vol. 22, pp. 397411, 1986. K. Zhou and P. P. Khargonekar, Stability robustness bounds for linear state-space models with structured uncertainty, ZEEE Trans. Automat. Contr., vol. AC-32, pp. 621-623, 1987. D. Hinrichsen and A. J. Pritchard, Stability radius for structured perturbationsand the algebraic Riccati equation, Syst. Contr. Lett., vol. 8, pp. 105-113, 1987.

D. S. Bemstein and W. M. Haddad, t h e optimal projection equations with Petersen-Hollot bounds: Robust controller synthesis with guaranteed structured stability radius, in Proc. ZEEE Conf. Decision Contr., Los Angeles, CA, Dec. 1987, pp. 1308-1318. D. S. Bemstein, Robust static and dynamic output-feedback stabilization: Deterministic and stochastic perspectives, ZEEE Trans. Automat. Contr., vol. AC-32, pp. 1076-1084, 1987. D. S. Bemstein and W. M. Haddad, The optimal projection equations with Petersen-Hollot bounds: Robust stability and performance via fixed-order dynamic compensation for systems with structured real-valued parameter uncertainty, ZEEE Trans. Automat. Contr., vol. 33, pp. 578-582, 1988. , Robust stability and performance for fixed-order dynamic compensation via the optimal projection equations with guaranteed cost bounds, in Proc. Amer. Contr. Conf., Atlanta, GA, June, 1988, pp. 2471-2476; also in Math. Contr. Sig. Syst., vol. 2, 1989. , Robust stability and performance analysis for linear dynamic systems, ZEEE Trans. Automat. Contr., vol. 34, 1989. J. C. Doyle, Analysis of feedback systems with structured uncertainties, ZEE Proc., Part D, vol. 129, pp. 242-250, 1982. I. R. Petersen, Disturbance attenuation and H optimization: A design method based on the algebraic Riccati equation, ZEEE Trans. Automat. Contr., vol. AC-32, pp. 427429, 1987. P. P. Khargonekar, 1. R. Petersen, and M. A. Rotea, H optimal control with state-feedback, ZEEE Trans. Automat. Contr., vol. 33, pp. 786-788, 1988. K. Zhou and P. P. Khargonekar, An algebraic Riccati equation approach to H optimization, Syst. Contr. Lett., vol. 11, pp. 8591, 1988. D.C. Hyland and D. S. Bemstein, The optimal projection equations for fmed-order dynamic compensation, ZEEE Trans. Automat. Contr., vol. AC-29, pp. 1034-1037, 1984. D. A. Wilson, Convolution and Hankel operator norms for linear systems, ZEEE Trans. Automat. Contr., vol. 34, pp. 94-97, 1988. M. J. Grimble, Optimal H , robustness and the relationship to LQG design problems, Znt. J. Contr., vol. 43, pp. 351-372, 1986. M. Mariton and R. Bertrand, A homotopy algorithm for solving coupled Riccati equations, Optimal. Contr. Appl. Meth., vol. 6, pp. 351-357, 1985. S. Richter, A homotopy algorithm for solving the optimal projection equations for fixed-order dynamic compensation: Existence, convergence and global optimality, in Proc. Amer. Contr. Conf., Minneapolis, MN, June 1987, pp. 1527-1531. S. W. Greeley and D. C. Hyland, Reduced-order compensation: LQG reduction versus optimal projection using a homotopic continuation method, in Proc. ZEEE Conf. Decision Contr., Los Angeles, CA, Dec. 1987, pp. 742-741. J. C. Doyle, K. Glover, P. P::Khargonekar, and B. A. Francis, State-space solutions to standard H2 and H , control problems, in Proc. Amer. Contr. Conf., Atlanta, GA, June 1988, pp. 1691-1696. A. N. Madiwale, Some practical aspects of H , compensators, M.I.T. Lincoln Laboratory, in preparation. W. M. Wonham, Linear Multivariable Control: A Geometric Approach. New York: Springer-Verlag, 1979. J. C. Willems, Least squares stationary optimal control and the algebraic Riccati equation, ZEEE Trans. Automat. Contr., vol. AC-16, pp. 621-634, 1971. R. W. Brockett, Finite Dimensional Linear Systems. New York: Wiley, 1970. K. Glover, All optimal Harkel-norm approximations of linear multivariable systems and their L-error bounds, Znt. J. Contr., vol. 39, pp. 1115-1193, 1984. D. S. Bemstein, The optimal projection equations for static and dynamic output feedback: The singular case, IEEE Trans. Automat. Contr., vol. AC-32, pp. 1139-1143, 1987. A. Bensoussan and J. H. van Schuppen, Optimal control of partially observable stochastic systems with an exponential-of-integral performance index, SIAM J. Contr. Optimiz., vol. 23, pp. 599-613, 1985. D. H. Jacobson, Extensions of Linear-Quadratic Control, Optimization and Matrix Theory. New York: Academic, 1977. P. Whittle, Risk-sensitivelinear/quadratic/Gaussian control, Adv. Appl, Prob., vol. 13, pp. 764-777, 1981. C. R. Rao and S. K. Mitra, Generalized Inverse of Matrices and Its Applications. New York: Wiley, 1971. A. J. Laub, A Schur method for solving Riccati equations, ZEEE Trans. Automat. Contr., vol. AC-32, pp. 913-921, 1979. R. H. Cannon, Jr. and D. E. Rosenthal, Experiments in control of flexible structures with noncolocated sensors and actuators, AZAA J. Guid. Contr. Dynam., vol. 7, pp. 546-553, 1984. Y. Liu and B. D. 0. Anderson, Controller reduction via stable factorization and balancing, Znt. J. Contr., vol. 44, pp. 507-531, 1986. A. Albert, Conditions for positive and nonnegative definiteness in terms of pseudo inverse, SIAM J. Contr. Optimiz., vol. 17, pp.
134-440-, -. . .- . . . 1969 -.

[41] D. S. Bemstein and W. M. Haddad, LQG control with an H ,

BERNSTEIN AND HADDAD: LQG CONTROL

305
control of spacecraft with flexible appendages with particular emphasis on robust and nonlinear control techniques. Dr. Bemstein is a member of SIAM.

performance bound: A Riccati equation approach, in Proc. Amer. Contr. Conf., Atlanta, CA. June 1988, pp. 796-802. H. Kwakernaak and R. Sivan, The maximally achievable accuracy of linear optimal regulators and linear optimal filters, IEEE Trans. Automat. Contr., vol. AC-17, pp. 79-86, 1972. K. Glover and J . C. Doyle. State-space formulae for all stabilizing controllers that satisfy an H,-norm bound and relations to risk sensitivity, Syst. Contr. Lett., vol. 11, pp. 167-172, 1988. D. Mustafa and K. Glover, Controllers which satisfy a closed-loop H, norm bound and maximize an entropy integral, in Proc. IEEE Conf. Decision Contr., Austin, TX, Dec. 1988. D. Mustafa, Relations between maximum entropyiH, control and combined HJLQG control, preprint. A. N . Madiwale, W. M. Haddad, and D. S. Bernstein, Robust H, control design for systems with structured parameter uncertainty, in Proc. IEEE Conf. Decision Contr., Austin. TX, Dec. 1988, pp. 965-972

Dennis S. Bernstein (M82) received the Sc.B. degree in applied mathematics from Brown University, Providence, RI, and the M.S.E. and Ph.D. degrees from the Computer, Information and Control Engineering Program at the University of Michigan, Ann Arbor. After spending two years at Lincoln Laboratory, Massachusetts Institute of Technology, Lexington, he joined the Controls Analysis and Synthesis Group of the Government Aerospace Systems Division, Harris Corporation, Melbourne, FL. At Harris Corporation his research interests are directed primarily toward the

Wassim M. Haddad (S87-M87) was born in . ..~ Athens, Greece, on July 14, 1961. He received the B.S., M.S., and Ph.D. degrees in mechanical engineering in 1983. 1984, and 1987, respectively, from the Florida Institute of Technology, Melbourne. Since 1987 he has been a consultant for the Controls Analysis and Synthesis Group of the Government Aerospace Systems Division, Harris Corporation, Melbourne, FL, and he is currently a faculty member in the Department of Mechanical and Aerospace Engineering, Florida Institute of Technology. His research interests are in the area of robust estimation and control for aerospace svstems. Dr. Haddad is a member of Tau Beta Pi
~ ~~