Professional Documents
Culture Documents
www.elsevier.com/locate/automatica
Technical communique
Structurally constrained H2 and H∞ control: A rank-constrained
LMI approach夡
Seog-Joo Kim a,∗ , Young-Hyun Moon b
a I&C Group, Korea Electrotechnology Research Institute (KERI), Sung Joo Dong, Changwon, Kyungnam 641-120, Korea
b Department of Electrical and Electronic Engineering, Yonsei University, Shinchon Dong, Seoul 120-749, Korea
Received 26 September 2005; received in revised form 8 March 2006; accepted 19 March 2006
Available online 6 June 2006
Abstract
This paper describes an iterative linear matrix inequality (LMI) approach to the design of a structurally constrained output feedback controller.
The structured synthesis is formulated as a novel rank-constrained LMI optimization problem, where the controller parameters are explicitly
described so as to impose structural constraints on the parameter matrices. An iterative penalty-function method based on heuristics is used to
solve the rank-constrained LMI problem. A numerical experiment illustrates the practicality of the proposed method.
䉷 2006 Elsevier Ltd. All rights reserved.
Keywords: H∞ control; Linear matrix inequality (LMI); Penalty-function method; Structured synthesis
1. Introduction (e.g., Skelton, Iwasaki, & Grigoriadis, 1997) converts SOF syn-
thesis to the well-known rank-constrained LMI problem that
Structurally constrained control involves designing a lin- can be solved relatively efficiently using various iterative meth-
ear time-invariant (LTI) controller subject to structural con- ods (Iwasaki & Skelton, 1995; Ghaoui, Oustry, & AitRami,
straints on its parameter matrices. This is sometimes called 1997; Grigoriadis & Skelton, 1996). However, the elimination
fixed-structure control or structured synthesis. Structured con- lemma cannot be used for general structured control problems,
trol problems have long been recognized as practically impor- since these require an LMI formulation in which the controller
tant because structural constraints are inevitable in many fields parameters can be directly manipulated. Several suitable LMI
of control and systems engineering, such as decentralized con- formulations along with the associated numerical methods have
trol. Static output feedback (SOF) stabilization also belongs to been proposed in the literature (de Oliveira, Camino, & Skel-
this class of synthesis problems since it can be regarded as an ton, 2000; Ebihara & Hagiwara, 2003; Han & Skelton, 2003;
unstructured fixed-order control problem. Tan & Grigoriadis, 2000), and the present paper presents an-
Despite its usefulness in practical control applications, struc- other heuristic algorithm.
tured controller synthesis is a challenging task due to its in- In this paper, motivated by Ebihara and Hagiwara (2003)
herent nonconvexity, and no complete solution has been found and Han and Skelton (2003), we first propose a unified LMI
yet. Nonetheless, SOF controller design is a well-studied field representation for the design of structurally constrained con-
in the linear matrix inequality (LMI) framework. The coupled trollers. The synthesis is characterized as a rank-constrained
LMI formulation employing the celebrated elimination lemma LMI optimization problem that differs from the coupled LMI
formulation. The problem is subsequently solved iteratively
using a penalty-function method (PFM) (Kim, Moon, Kwon,
夡 This paper was not presented at any IFAC meeting. This paper was
& Kim, 2005b). Our method can find a suboptimal solution
recommended for publication in revised form by Associate Editor Keqin Gu without using a low-efficiency bisection method or a specific
under the direction of Editor André Tits.
∗ Corresponding author. Tel.: +82 55 280 1472; fax: +82 55 280 1476. initialization procedure. The solution of our algorithm moves
E-mail addresses: sjkim@keri.re.kr (S.-J. Kim), moon@yonsei.ac.kr towards the region satisfying the rank condition as the penalty
(Y.-H. Moon). parameter increases.
0005-1098/$ - see front matter 䉷 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.automatica.2006.03.017
1584 S.-J. Kim, Y.-H. Moon / Automatica 42 (2006) 1583 – 1588
We use the following notation. Sn denotes the set of all From > 0 and rank(W̃ ) = m, the following hold:
symmetric matrices in Rn×n , and In denotes the n × n identity
matrix (the dimension n is omitted where it can be inferred I ≺ W̃3 ⇔ −1 I − W̃3−1 0,
from the context). For a matrix A, its transpose, trace, and rank
W̃1 = W̃2 W̃3−1 W̃2T .
are denoted by AT , tr(A), and rank(A), respectively. If A is
a symmetric matrix, A 0 (A0) means that A is positive Also, it follows that
definite (semidefinite). For a rectangular matrix A, A⊥ stands
for an orthogonal complement of A; i.e., A⊥ A=0, A⊥ A⊥T 0. (W̃3 − I )−1 = W̃3−1 + W̃3−1 (−1 I − W̃3−1 )−1 W̃3−1 .
Finally, Tzw means the transfer function from w to z.
Applying the Schur complement to (3) yields
2. Preliminaries 0 M − W̃1 + W̃2 (W̃3 − I )−1 W̃2T
This section describes a basic tool for fixed-structure control = M + W̃2 W̃3−1 (−1 I − W̃3−1 )−1 W̃3−1 W̃2T
problems. In this framework, all LMI variables are explicitly M.
expressed in the LMIs at the expense of an increase in the
This ends the proof.
decision variables and a nonconvex rank condition. Lemma 1
establishes the basis of our method.
In the next section we show that the structurally constrained
controller synthesis can be described in the form of LMI (1)
Lemma 1. For matrices Q ∈ Sn and B ∈ Rn×m with
subject to the rank condition (2).
rank(B) < n, the following statements are equivalent:
3. Rank-constrained LMI formulation of a structured
(i) B ⊥ QB ⊥T ≺ 0.
synthesis
(ii) There exists a scalar ∈ R such that Q − BB T ≺ 0.
(iii) There exist a positive semidefinite matrix W ∈ Sn+m and
Consider the LTI system with the following state-space rep-
a scalar > 0 such that
resentation:
Q B
≺ W, (1) ẋ = Ax + B1 w + B2 u,
B T Im
rank(W ) = m. (2) z = C1 x + D11 w + D12 u,
y = C2 x + D21 w, (5)
Proof. Statements (i) and (ii) are equivalent to each other due
to Finsler’s lemma (e.g., Skelton et al., 1997). By applying a where x ∈ Rn is the state, w ∈ Rnw is the exogenous input, u ∈
congruence transformation to (1), condition (iii) can be equiv- Rnu is the control input, z ∈ Rnz is the output to be regulated,
alently described as the existence of W̃ 0 and > 0 such that and y ∈ Rny is the measured output. Our aim is to design an
SOF controller, u = Ky that places all the eigenvalues of the
Q B M 0
T T =
T
≺ W̃ , (3) closed-loop system in a stability region described by
B T I 0 I
if and only if the following rank-constrained LMI problem is Proof. See Iwasaki, Skelton, and Geromel (1994) and Chilali
feasible in P 0 (P ∈ Sn ), W 0 (W ∈ S3n+2nz +nw ), and a and Gahinet (1996). The proof is analogous to that for
scalar > 0: Lemma 2.
⎡ pP 0 qP 0 ATcl CclT ⎤ In the next section, we describe a computation method for
⎢ 0 Q 0 S BclT T ⎥
Dcl rank-constrained LMI problems.
⎢ ⎥
⎢ ∗ ⎥
⎢q P 0 rP 0 −In 0 ⎥
⎢ ⎥ ≺ W, (8)
⎢ 0 ST R −Inz ⎥
4. PFM for rank-constrained LMI problems
⎢ 0 0 ⎥
⎢ ⎥
⎣ Acl Bcl −In 0 In 0 ⎦ With a slight abuse of notation, we consider the following
Ccl Dcl 0 −Inz 0 Inz problem:
Lemma 3 states the structured SOF synthesis for H2 optimal Lemma 4. The rank of a positive semidefinite matrix W ∈ Sn
control. is less than or equal to r if and only if there exists a matrix
V ∈ Rn×(n−r) with V T V = In−r such that
Lemma 3. For system (5) with D11 = D21 = 0, we can find
an SOF control law such that the closed-loop system satisfies tr(V T W V ) = 0. (15)
the H2 performance, Tzw 2 = Ccl (sI − Acl )−1 Bcl 2 < , and
also the pole constraint (6) if and only if there exist P 0 Proof. Suppose that the eigenvalues of W are 1 · · · n .
(P ∈ Sn ), W 0 (W ∈ S3n+2nz ), and a scalar > 0 such that Then, the following inequality holds (Horn & Johnson, 1985,
p. 191):
tr(B1T P B 1 )2 , (11)
⎡ ⎤ 1 + · · · + n−r tr(V T W V ) (16)
pP qP 0 ATcl CclT
⎢ q ∗P rP −In 0 ⎥ for any matrix V ∈ Rn×(n−r) such that V T V = In−r . The
⎢ 0 ⎥
⎢ ⎥ equality in (16) holds when the columns of V consist of the
⎢ 0 −Inz ⎥
⎢ 0 Inz 0 ⎥ ≺ W, (12) eigenvectors corresponding to the n − r smallest eigenvalues
⎢ ⎥
⎣ Acl −In 0 In 0 ⎦ of W . Therefore, the sufficiency is trivial from (16). To prove
Ccl 0 −Inz 0 Inz the necessity, suppose that a matrix W 0 with rank(W ) r is
given. Then, for V consisting of the eigenvectors corresponding
rank(W ) = n + nz , (13) to the n − r smallest eigenvalues of W , we have tr(V T W V ) =
1 + · · · + n−r = 0. This ends the proof.
where
Also, notice that if rank(W ) > r, then tr(V T W V ) > 0.
Acl = A + B2 KC 2 , Ccl = C1 + D12 KC 2 . Therefore, tr(V T W V ) can be used as a penalty function for
1586 S.-J. Kim, Y.-H. Moon / Automatica 42 (2006) 1583 – 1588
the rank condition of problem (14). We now define the penalty Proof. We know from (16) that p(xk ; Vk ) p(xk ; Vk−1 ) holds.
function as From the optimality condition, the following inequalities hold:
p(x; V ) = tr(V T W V ), (17) (xk+1 ; , , Vk+1 ) (xk+1 ; , , Vk )
(xk ; , , Vk ).
where V ∈ Rn×(n−r) , V T V = In−r . From (16), we see that the
proposed penalty function (17) is an upper bound on the sum This ends the proof.
of the n − r smallest eigenvalues of W . Furthermore, note that
the penalty function is linear in x for a given V . As will be Lemma 5 implies that for fixed and , the sequence
seen below, we iteratively solve the following penalized convex {(xk ; , , Vk )} is nonincreasing. When the value of the
optimization to obtain a solution to (14) instead of minimizing penalty function p(xk ; Vk ) is not sufficiently small with fixed
(17) and cT x and , we can make it smaller by increasing based on the
following lemma.
xk = arg min{(x; k ,k ,Vk−1 ) : x ∈ C}, k=1, 2, . . . , (18)
x
Lemma 6. In (18), suppose that and V are fixed. If
where k+1 > k , the following inequalities hold:
(x; , , V ) = cT x + tr(W ) + p(x; V ) (19) (xk+1 ; , k+1 , V ) (xk ; , k , V ), (23)
and p(xk+1 ; V ) p(xk ; V ), (24)
C = {x : W (x)0, L(x) 0}. c xk+1 + tr(W (xk+1 )) c xk + tr(W (xk )).
T T
(25)
In (18), k > 0 is the penalty parameter and k > 0 is the opti- Proof. Let us denote
mization weight. Also, Vk−1 is constructed from the eigenvec-
tors corresponding to the n−r smallest eigenvalues of W (xk−1 ) c̃T x = cT x + tr(W ).
since the eigenvectors of W can be taken to be orthonormal to
The inequality (23) can be proved by the following inequalities
each other.
using the optimality condition and k+1 > k :
To understand the necessity of the additional term tr(W ) in
(19), consider the following simple feasibility problem where (xk ; , k , V ) = c̃T xk + k p(xk ; V )
c = 0: find x1 , x2 satisfying the convex condition
c̃T xk+1 + k p(xk+1 ; V )
x 0 c̃T xk+1 + k+1 p(xk+1 ; V )
W 1 0, x2 1 (20)
0 x2 = (xk+1 ; , k+1 , V ).
and the rank condition rank(W ) = 1. The solution to this In addition, from the inequalities
problem is x1 = 0, x2 1. Assume that (18) is started with
x1 > 1, x2 =1 and that only penalty function (17) is minimized. c̃T xk + k p(xk ; V ) c̃T xk+1 + k p(xk+1 ; V ) (26)
Since the smaller eigenvalue is 1, and the associated eigenvec-
and
tor is [0 1]T , the resulting penalty function becomes x2 . Then,
minimizing only (17) yields the problem: c̃T xk+1 + k+1 p(xk+1 ; V ) c̃T xk + k+1 p(xk ; V ),
min x2 s.t. (20), (21) the following holds:
and its solution can be any value x1 0 and x2 = 1. Therefore, (k+1 − k )p(xk+1 ; V ) (k+1 − k )p(xk ; V ),
we observe that x1 is not guaranteed to be zero by (21). This
which implies (24). Finally, from (24) and (26), we can see that
implies that if the solution to (21) by existing optimization
(25) holds.
software is x1 > 1, x2 =1, we cannot obtain the desired solution
despite successive iterations. However, by minimizing (19) as In a similar manner, the following lemma holds for fixed
min x1 + x2 + x2 s.t. (20), and V .
we can obtain x1 = 0, x2 = 1 in a single iteration. Note that Lemma 7. In (18), suppose that and V are fixed. If
tr(W ) = ni=1 i . This example demonstrates why we should k+1 > k , the following inequalities hold:
minimize the sum of the n − r eigenvalues relative to the total
sum of the eigenvalues. (xk+1 ; k+1 , , V ) (xk ; k , , V ),
Let us now examine certain properties of the algorithm. cT xk+1 cT xk ,
Lemma 5. In (18), suppose that and are fixed, and that tr(W (xk+1 )) + p(xk+1 ; V ) tr(W (xk )) + p(xk ; V ).
only Vk is updated. Then, the following inequality holds:
Proof. The proof is omitted because it is similar to that for
(xk+1 ; , , Vk+1 )(xk ; , , Vk ). (22) Lemma 6.
S.-J. Kim, Y.-H. Moon / Automatica 42 (2006) 1583 – 1588 1587
Remark 1. Lemmas 6 and 7 shed light on the roles of the com- Table 1
putation parameters and . By increasing , we can reduce Comparison of the H2 norm for Example 1
the value of the penalty function, and by increasing , we can Method de Oliveira et al. Ebihara and Hagi- PFM
reduce the value of the original objective function. (2002) wara (2003)
p (x; V)
(1) Initialization: Find an initial x0 by solving x0 ={x : x ∈ C}. 10-4
Compute V0 from W (x0 ) using eigenvalue decomposition,
10-8
and choose 1 > 1, 1 > 1, ∈ (0, 1), >1, > 1,
> 1, 1 >1, 2 >1, and 3 >1. Set k = 1. 10-12
(2) Convex optimization: Compute xk by solving the convex
LMI optimization problem (18). H2 norm
(3) Computation of V : Compute Vk from W (xk ) using eigen- 0.3
value decomposition.
0.25
(4) Feasibility test: If p(xk ; Vk ) < 1 or p(xk ; Vk )/tr(W (xk ))
< 3 , an almost feasible solution is obtained and the com- 2
0.2
putation proceeds to step 6.
(5) Penalty-parameter update: If p(xk ; Vk ) cannot be
0.15
reduced to the prescribed extent with k , that is,
p(xk ; Vk ) > p(xk−1 ; Vk−1 ), then increase the penalty pa- Penalty parameter and optimization weight
rameter by k+1 =k ; otherwise, k+1 =k . Set k ← k +1 106
and go to step 2.
(6) Optimality test: If |cT xk −cT xk−1 |/|cT xk−1 | 2 , then cT xk
104
,
and the obtained controller was Ebihara, Y., & Hagiwara, T. (2003). Structured controller synthesis using
LMI and alternating projection method. Proceedings of the 42nd IEEE
−0.4104 −0.3536 0 0
K= . conference on decision and control, Hawaii, USA (pp. 5632–5637).
0 0 −0.3492 −0.1648 Fares, B., Apkarian, P., & Noll, D. (2000). An augmented Lagrangian
method for a class of LMI-constrained problems in robust control
Table 1 lists the design result of the PFM and those of previous theory. Proceedings of the American control conference, Chicago, USA
studies; comparison of the PFM data with those obtained pre- (pp. 3702–3706).
viously demonstrates the good numerical performance of the Ghaoui, L. E., Oustry, F., & AitRami, M. (1997). A cone complementarity
linearization algorithm for static output feedback and related problems.
PFM. The behavior of the PFM for Example 1 is shown in
IEEE Transactions on Automatic Control, 42(8), 1171–1176.
Fig. 1, which indicates that the value of the penalty function Grigoriadis, K. M., & Skelton, R. E. (1996). Low-order control design for
decreases monotonically until the PFM converges to an almost LMI problems using alternating projection methods. Automatica, 32(8),
feasible solution in about 90 iterations. After the penalty func- 1117–1125.
tion tends to almost zero, adjusting the optimization weight Han, J., & Skelton, R. E., (2003). An LMI optimization approach for structured
linear controllers. Proceedings of the 42nd IEEE conference on decision
results in a suboptimal solution.
and control, Hawaii, USA (pp. 5143–5148).
Horn, R. A., & Johnson, C. R. (1985). Matrix analysis. Cambridge: Cambridge
6. Conclusion University Press.
Iwasaki, T., & Skelton, R. E. (1995). The XY-centering algorithm for the dual
We have described a practical procedure for finding struc- LMI problem: A new approach to fixed order control design. International
Journal of Control, 62(6), 1257–1272.
turally constrained H2 and H∞ suboptimal output feedback
Iwasaki, T., Skelton, R. E., & Geromel, J. C. (1994). Linear quadratic
controllers for linear systems. The synthesis is formulated as a suboptimal control with static output feedback. Systems & Control Letters,
new type of rank-constrained LMI optimization problem, which 23(6), 421–430.
was solved by an iterative PFM. Our method is applicable to Kim, S. J., Moon, Y. H., Kwon, S., & Kim, K. H. (2005a). Design of
both continuous- and discrete-time systems. Although the pro- a structurally constrained suboptimal controller using an LMI method.
Proceedings of the 44th IEEE conference on decision and control and the
posed algorithm does not exhibit global convergence, it per-
European control conference 2005, Seville, Spain (pp. 4318–4323).
formed well in the numerical experiment presented here. Kim, S. J., Moon, Y. H., Kwon, S., & Kim, K. H. (2005b). Rank-constrained
LMI approach to mixed H2 /H∞ static output feedback controllers.
Acknowledgments Proceedings of the 16th IFAC world congress, Prague, Czech.
Löfberg, J. (2004). YALMIP: A toolbox for modeling and optimization in
Matlab. Proceedings of the IEEE computer-aided control system design
The authors would like to thank the reviewers for their helpful
conference, Taipei, Taiwan.
and constructive comments. Scherer, C., Gahinet, P., & Chilali, M. (1997). Multiobjective output-feedback
control via LMI optimization. IEEE Transactions on Automatic Control,
References 42(7), 896–911.
Skelton, R. E., Iwasaki, T., & Grigoriadis, K. M. (1997). A unified approach
Chilali, M., & Gahinet, P. (1996). H∞ design with pole placement constraints: to linear control design. London: Taylor & Francis.
An LMI approach. IEEE Transactions on Automatic Control, 41(3), Sturm, J. F. (1999). Using SeDuMi 1.02, a Matlab toolbox for optimization
358–367. over symmetric cones. Optimization Methods and Software, 11–12(1–4),
de Oliveira, M. C., Camino, J. F., & Skelton, R. E. (2000). A convexifying 625–653.
algorithm for the design of structured linear controllers. Proceedings Tan, K., & Grigoriadis, K. M. (2000). Robust decentralized control using
of the 39th IEEE conference on decision and control, Sydney, Australia an alternating projection approach. Proceedings of the American control
(pp. 2781–2786). conference, Chicago, USA (pp. 801–805).
de Oliveira, M. C., Geromel, J. C., & Bernussou, J. (2002). Extended H2 and
H∞ norm characterization and controller parameterizations for discrete-
time systems. International Journal of Control, 75(9), 666–679.