Automatica 40 (2004) 2003 – 2009 www.elsevier.com/locate/automatica Technical Communique New dilated LMI characterizations for continuous-time multiobjective controller synthesis夡 Yoshio Ebiharaa,∗ , Tomomichi Hagiwaraa a Department of Electrical Engineering, Kyoto University, Kyotodaigaku-Katsura, Nishikyo-ku, Kyoto 615-8510, Japan Received 10 May 2003; received in revised form 22 March 2004; accepted 14 June 2004 Available online 13 August 2004 Abstract This paper provides new dilated linear matrix inequality (LMI) characterizations for continuous-time controller synthesis. The dilated LMIs enable us to parametrize controllers without involving the Lyapunov variables in the parametrizations. Taking advantage of this feature, we can readily design multiobjective controllers with non-common Lyapunov variables, whereas we are forced to employ a common one in the well-known Lyapunov shaping paradigm. In particular, it is shown that the proposed dilated-LMI-based approach to H2 /D-stability synthesis encompasses the corresponding Lyapunov shaping paradigm as a special case. In this sense, the results of this paper can be viewed as partial counterparts of those obtained in the discrete-time setting. 䉷 2004 Elsevier Ltd. All rights reserved. Keywords: Multiobjective control; Robust control; Non-common Lyapunov variables; Linear matrix inequalities; Dilation 1. Introduction In recent studies, Oliveira, Bernussou, and Geromel (1999) opened a new horizon for linear matrix inequality (LMI)-based controller synthesis under the discrete-time setting. They showed that a matrix A is Schur stable iff there exist a Lyapunov variable X and a multiplier G satisfying AG −X < 0. (1) GT AT X − G − GT Moreover, nice properties of this “dilated” LMI in the light of the standard one −X + AXAT < 0 can be claimed as follows: (i) The Lyapunov variable X has no multiplication relations with A, whereas G does. 夡 This paper was not presented at any IFAC meeting. This paper was recommended for publication in revised form by Associate Editor Carsten W. Scherer under the direction of Editor Paul Van den Hof. ∗ Corresponding author. Department of Electrical Engineering, Kyoto University, Yoshida, Sakyo-ku, Kyoto 606-8510, Japan. Tel.: +81-75-3832255; fax: +81-75-383-2251. E-mail address: [email protected] (Y. Ebihara). 0005-1098/$ - see front matter 䉷 2004 Elsevier Ltd. All rights reserved. doi:10.1016/j.automatica.2004.06.009 (ii) The dilated LMI reduces to the standard one by letting (X, G) = (X, X). The benefit from the property (i) is the new controller parametrization that is independent of the Lyapunov variables. By virtue of this new parametrization, multiobjective controller synthesis with non-common Lyapunov variables and robust controller synthesis with parameter-dependent Lyapunov variables have been established, both of which are promising for less conservative results (Oliveira et al., 1999; Oliveira, Geromel, & Bernussou, 2002). The property (ii) is still of great importance, since this ensures that dilated-LMI-based synthesis approaches always encompass the corresponding standard-LMI-based ones. These achievements, however, heavily rely on the structural properties of LMIs for discrete-time system synthesis. Despite intensive research efforts (Apkarian, Tuan, & Bernussou, 2001; Dettori & Scherer, 2000; Oliveira & Skelton, 2001; Peaucelle, Arzelier, Bachelier, & Bernussou, 2000), corresponding results in the continuous-time system case are still open to this date. In this paper, we present new results on the dilated-LMIbased controller synthesis for continuous-time systems. In 2004 Y. Ebihara, T. Hagiwara / Automatica 40 (2004) 2003 – 2009 stark contrast with the discrete-time case, we derive dilated LMIs with the property (i) via a particular application of the Schur complement technique (Boyd, Ghaoui, Feron, & Balakrishnan, 1994) to the standard ones. In addition, as for the property (ii), an exact relation between the solutions of the dilated LMIs and the standard ones can be established. Unfortunately, however, this relation is not so simple compared with the discrete-time results and poses some limitations on the scope of the controller synthesis by means of the proposed dilated LMIs. Nevertheless, when designing multiobjective H2 /D-stability controllers Chilali & Gahinet, 1996; Masubuchi, Ohara, & Suda, 1998; (Scherer, Gahinet, & Chilali, 1997), this relation enables us to ensure that the proposed dilated-LMI-based approach always encompasses the corresponding Lyapunov shaping paradigm (Scherer et al., 1997). In this sense, the results in this paper can be viewed as partial counterparts of those obtained in the discrete-time setting (Oliveira et al., 1999, 2002). We use the following notations. For a matrix A ∈ Rn×n , (A) means the set of the eigenvalues of A. The notation He{A} := A + AT will also be used. 2. Dilation lemma Various control performances of continuous-time systems can be characterized in terms of LMIs (Boyd et al., 1994; A I +He 0 G [ I 0 0 −bI bI I bT ] < 0. (3) Moreover, for every solution X > 0 of (2), (X, G) = (X, −a(A − aI )−1 X) is a solution of (3). Conversely, for every X > 0 such that (3) holds for some G, X = X satisfies (2). Proof. We first show that the condition (i) implies (ii). Applying the Schur complement technique (Boyd et al., 1994) to (2) with the given scalar a > 0, we have −2a X −2a X 0 0 0 −AX −XT −X −2a X He{(A − aI )X} 0 −1 −X −1 X 0 0 −1 T 0 −XA 0 − 2 X 0 0 −X 0 0 −I < 0. (4) Here, (A − aI ) is non-singular because A is stable by the condition (i). Hence, via congruence transformation with diag I, (A − aI )−1 , I, I, I on (4), we obtain −2a X −2a X(A − aI )−T 0 0 0 T −1 −1 −1 −1 −1 −2a(A − aI ) X He{(A − aI ) X} −(A − aI ) X −X − a(A − aI ) X −(A − aI ) X −T < 0. 0 − X ( A − aI ) −−1 0 0 1 X −1 −T 0 −X − a X(A − aI ) 0 − 2 X 0 −T 0 0 −I 0 −X(A − aI ) Skelton, Iwasaki, & Grigoriadis, 1997). For those standard(5) type LMIs, our first goal is to derive corresponding dilated LMIs. To this end, we introduce the following lemma, which The above inequality can be rearranged into the following is useful for deriving dilated LMIs from the standard ones := −(A − aI )−1 X. form by defining G and plays an important role in our study. Lemma 1. Let a matrix A ∈ Rn×n , scalars where the scalars 1 > 0, 2 > 0 and a matrix of column dimension n be given. Then, the following two conditions are equivalent, where b = a −1 > 0 is an arbitrarily prescribed number. (i) There exists X > 0 such that AX + XA + 1 X + 2 AXA + X X < 0. T T T (2) (ii) There exist X > 0 and G such that 0 −X X 0 X −X 0 0 −X 0 X 0 −−1 1 X 0 0 0 −X 0 −−1 2 X 0 XT 0 0 0 −I 2a X −X X aX XT 0 0 −X 0 −X X 0 −−1 X 0 0 1 −1 a X −X 0 −2 X 0 0 0 0 −I X A − aI I +He 0 G[ 2aI −I I aI 0 0 T ] < 0. (6) Performing again congruence transformation with diag I aI , I, I, I on (6), we arrive at (3) in the condition 0 I (ii) with X = X and G := a G. Y. Ebihara, T. Hagiwara / Automatica 40 (2004) 2003 – 2009 It remains to show that (ii) implies (i). This is straightforward since (3) implies I −A 0 0 0 I 0 0 0 0 0 I 0 0 0 0 AX + X AT = X XAT X I 0 0 0 L(X, G) 0 0 I 0 X −−1 1 X 0 0 −A 0 0 0 I 0 0 0 I 0 0 0 AX X T 0 0 < 0, −−1 X 0 2 0 −I 0 T 0 0 I (7) where L(X, G) denotes the left-hand side of inequality (3). Applying the Schur complement technique to the last inequality in (7), we have (2) with X = X. This completes the proof. In Lemma 1, the inequality (2) can be regarded as a standard LMI for the analysis and synthesis of continuous-time systems frequently used in the previous studies (Boyd et al., 1994; Skelton et al., 1997), while (3) is a dilated LMI corresponding to (2). The variables X and X in these LMIs denote the Lyapunov variables, and G in (3) is a multiplier introduced for dilation. By specializing parameters 1 , 2 and appropriately, we can obtain dilated LMIs for various control performances as we will see in the next section. It is important to note that the dilated LMI (3) has nice properties in comparison with the standard one (2). First, the dilated LMI achieves the “separation” of the Lyapunov variable X from A in the sense that X is free from multiplications with A. Second, the dilated LMI “recovers” the standard one by letting (X, G) = (X, −a(A − aI )−1 X) irrespective of 1 , 2 and . Although we have proved Lemma 1 based on the Schur complement technique, the equivalence of (i) and (ii) can also be shown by the Elimination Lemma (Gahinet & Apkarian, 1994; Iwasaki & Skelton, 1994; Skelton et al., 1997). If we directly work on the Elimination Lemma, however, it is hard, or at least quite impossible, to clarify the recovery property of the dilated LMI (3) represented by (X, G) = (X, −a(A − aI )−1 X). This is because the Elimination Lemma originally moves from (3) to (2) to eliminate the variable G and hence this lemma itself provides no constructive way to derive the dilated LMI (3) from (2). It is this particular application of the Schur complement technique that leads us to this nice property. In the next section, this property plays a key role in ensuring an explicit advantage of the dilated LMIs over the standard ones when designing multiobjective controllers. 3. Multiobjective controller synthesis using dilated LMIs In this section, we first derive dilated LMIs for the H2 performance and the D-stability constraints (Chilali & Gahinet, 2005 1996). Then, we show new controller parametrizations for the dilated-LMI-based controller synthesis. It turns out that these new parametrizations enable us to design multiobjective H2 /D-stability controllers with non-common Lyapunov variables. As the main result of the paper, we further prove that this dilated-LMI-based synthesis approach always encompasses the corresponding Lyapunov shaping paradigm (Scherer et al., 1997) as a special case. 3.1. Dilated LMIs for the H2 performance and the D-stability The first result concerns the dilated LMI for the H2 performance. Theorem 2 (H2 Performance). Let us consider the continuous-time system T (s) := C(sI − A)−1 B. Then, the following three conditions are equivalent, where b=a −1 > 0 is an arbitrarily prescribed number. (i) The matrix A is stable and the H2 cost ||T ||2 is bounded by 2 > 0. Namely, ||T ||2 < 2 . (ii) There exist X2 > 0 and Z2 > 0 such that T T AX + X2 C CX2 < 0, 2 + X2 A T Z2 B > 0, trace(Z2 ) < 22 . B X2 (iii) There exist X2 > 0, Z2 > 0 and G2 such that 0 −X2 0 −X2 0 0 0 0 −I A + He I G2 [ I −bI 0 ] < 0, C Z2 BT > 0, trace(Z2 ) < 22 . B X2 (8) (9) Moreover, for every solution (X2 , Z2 ) of (8), (X2 , Z2 , G2 ) = (X2 , Z2 , −a(A − aI )−1 X2 ) is a solution of (9). Conversely, for every pair of the matrices X2 > 0 and Z2 > 0 such that (9) holds for some G2 ; X2 = X2 and Z2 = Z2 satisfy (8). Proof. The equivalence of (i) and (ii) is well-known (Boyd et al., 1994; Skelton et al., 1997). The equivalence of (ii) and (iii) and the latter assertion of Theorem 2 immediately follow from Lemma 1 by letting 1 = 2 → 0, = C and X = X2 , X = X2 , G = G2 . We next consider the dilated LMIs for the D-stability (Chilali & Gahinet, 1996; Peaucelle et al., 2000). Since the matrix A should be stable in Lemma 1, the regions D are restricted to those contained in the open left half-plane. In Theorems 3–5 given below, the LMIs in the condition (ii) are from Chilali and Gahinet (1996). Applying Lemma 1 2006 Y. Ebihara, T. Hagiwara / Automatica 40 (2004) 2003 – 2009 to those standard LMIs, corresponding dilated LMIs in (iii) can be obtained. Although the dilated LMI in Theorem 5 is not a direct consequence of Lemma 1, we can derive it by following similar arguments to the proof of Lemma 1. Theorem 3 (-Stability region). Let a matrix A ∈ Rn×n be given. Then, the following three conditions are equivalent, where b = a −1 > 0 is an arbitrarily prescribed number. (i) The matrix A satisfies (A) ⊂ H(), where H() := { ∈ C : Re() < − } ( > 0). (ii) There exists XH > 0 such that AXH + XH AT + 2XH < 0. (10) (iii) There exist XH > 0 and GH such that 0 −XH XH −XH 0 0 1 −1 XH 0 − XH 2 A +He I GH [ I −bI bI ] < 0. 0 Theorem 4 (Circular region). Let a matrix A ∈ Rn×n be given. Then, the following three conditions are equivalent, where b = a −1 > 0 is an arbitrarily prescribed number. (i) The matrix A satisfies (A) ⊂ C(c, r), where C(c, r) := { ∈ C : | − c| < r} (c < − r < 0). (ii) There exists XC > 0 such that AXC + XC AT − − r2 1 − AXC AT < 0. c c 0 −XC XC −XC 0 0 (12) XC 0 c XC c2 − r 2 0 −X C A I +He GC [ I 0 0 0 −bI (iii) There exist XS > 0 and GS XS 0 −kXS 0 0 −kXS XS 0 0 0 −XS −kX S A 0 0 GS I +He 0 0 I 0 A kI −I bI 0 GS bI −bkI I < 0. kI (15) 3.2. Change-of-variables for dilated-LMI-based controller synthesis Let us consider the continuous-time LTI plant described by x˙ = Ax + B1 w + Bu, y = Cx + D21 w. (16) Our goal is to design a full-order output-feedback controller K of the form x˙K = AK xK + BK y, 0 −XC 0 cX C −bkI −bI such that 0 −XS −kX S 0 Moreover, for every solution XS > 0 of (14), (XS , GS ) = (XS , −a(A − aI )−1 XS ) is a solution of (15). Conversely, for every XS > 0 such that (15) holds for some GS , XS = XS satisfies (14). z = C1 x + D12 u, XC (iii) There exist XC > 0 and GC such that (i) The matrix A satisfies (A) ⊂ S(k), where S(k) := { ∈ C : | Im()| < k| Re()|} (k > 0). (ii) There exists XS > 0 such that AXS − XS AT k(AXS + XS AT ) < 0. XS AT − AXS k(AXS + XS AT ) (14) (11) Moreover, for every solution XH > 0 of (10), (XH , GH ) = (XH , −a(A − aI )−1 XH ) is a solution of (11). Conversely, for every XH > 0 such that (11) holds for some GH , XH = XH satisfies (10). c2 Theorem 5 (Conic sector region). Let a matrix A ∈ Rn×n be given. Then, the following three conditions are equivalent, where b = a −1 > 0 is an arbitrarily prescribed number. u = CK xK . (17) In the state-feedback case (C = I , D21 = 0), we also design a static state-feedback controller u = Kx. With the plant and the controllers, the closed-loop system can be described by x˙cl = Axcl + Bw, I ] < 0. (13) Moreover, for every solution XC > 0 of (12), (XC , GC ) =(XC , −a(A−aI )−1 XC ) is a solution of (13). Conversely, for every XC > 0 such that (13) holds for some GC , XC =XC satisfies (12). z = Cxcl , (18) and we denote its transfer function from w to z by Tzw (s). For the full-order output-feedback controller K, the matrices in (18) are given by A BC K , A= BK C AK B1 , C = [ C1 D12 CK ] . B= (19) BK D21 Y. Ebihara, T. Hagiwara / Automatica 40 (2004) 2003 – 2009 On the other hand, for the static state-feedback controller K, they are A = A + BK, B = B1 , C = C1 + D12 K. (20) In the rest of this subsection, we show how to linearize (dilated) bilinear matrix inequalities (9), (11), (13) and (15) obtained by replacing (A, B, C) by the matrices in (19) or (20). To facilitate our statements, for the time being, we drop the subscript j of the multipliers Gj (j = 2, H, C, S) in (9), (11), (13) and (15) and represent them by G. 3.2.1. State-feedback case It suffices to take the change-of-variable W := KG to turn all inequalities in (9), (11), (13) and (15) into LMIs. The state-feedback gain K can be reconstructed by K = W G−1 . The nonsingularity of G is ensured since each of (9), (11), (13) and (15) requires G + GT > 0. 3.2.2. Output-feedback case (Ebihara, 2002) The change-of-variables technique stated below is based on the result by Scherer et al. (1997), and similar to its variants in Apkarian et al. (2001), Oliveira et al. (2002). Let us first partition G and its inverse H as G11 G12 H11 H12 −1 G= , H =G = , (21) G21 G22 H21 H22 where G11 ∈ Rn×n and H11 ∈ Rn×n and other submatrices have compatible dimensions. We assume that G21 and H21 are nonsingular without loss of generality (Apkarian et al., 2001). With (21) and the controller matrices given in (17), we further define the following matrices: I G11 H11 I , H := G := , 0 G21 H21 0 T T := H11 G11 + H21 G21 , C¯ K := CK G21 , A B G11 T ¯ + H21 BK ] AK G21 . C 0 C¯ K (22) T BK , B¯ K := H21 T A¯ K := [H11 (23) Then, appropriate congruence transformations with the matrix H can be applied to (9), (11), (13) and (15) so that the resulting constraints only involve the following terms. TH Xj H = X¯ j (j = 2, H, C, S), T H11 T T , H GH = H G = I G11 TH B = TB +B ¯ K D21 H11 1 , B1 CGH = CG = [ C1 C1 G11 + D12 C¯ K ], (24) 2007 TH AGH = TH AG T H11 A + B¯ K C A¯ K . = A AG11 + B C¯ K (25) We see that the above terms are affine with respect to X¯ j (j = 2, H, C, S), G11 , H11 , , A¯ K , B¯ K and C¯ K . Thus, the matrix inequalities (9), (11), (13) and (15) turn out to LMIs with respect to the variables X¯ j (j =2, H, C, S), G11 , H11 , , A¯ K , B¯ K , C¯ K , Z and 22 . Once the variables G11 , H11 , , A¯ K , B¯ K and C¯ K have been found, the output-feedback controller (17) can be reconstructed by −T ¯ BK = H21 BK , CK = C¯ K G−1 21 , A −T T ¯ AK = H21 A¯ K − H11 BK C B 0 G11 C¯ K G−1 21 , (26) where G21 and H21 are non-singular matrices satisfying T G = − HT G . H21 21 11 11 3.3. Multiobjective H2 /D-stability controller synthesis via dilated LMIs We now apply the dilated LMIs to the multiobjective H2 /D-stability controller synthesis. For the plant (16), the problem is to find a controller K, full-order output-feedback or static state-feedback, that minimizesTzw 2 subject to the D-stability constraint (A) ⊂ H() C(c, r) S(k). In tackling this problem, the standard and dilated LMIs lead to two different approaches. (i) Standard-LMI-based approach (Scherer et al., 1997): Minimize 22 subject to (8), (10), (12) and (14) by a common Lyapunov variable X := X2 = XH = XC = XS . (ii) Dilated-LMI-based approach: Minimize 22 subject to (9), (11), (13), and (15) by a common multiplier G := G2 = GH = GC = GS . In the approach (i), a common Lyapunov variable is enforced to convexify the synthesis problem via the Lyapunov shaping paradigm (Scherer et al., 1997). This restriction inherently brings conservatism into the design and only upper bounds of the cost functional will be minimized. On the other hand, in the approach (ii), we impose a common multiplier G. Since the controller parametrizations for dilated LMIs are dependent on the multipliers, this restriction is inevitable to convexify the problem. Due to this restriction, the approach (ii) is still conservative. Nevertheless, the approach (ii) is promising for the achievement of less conservative results since we can employ non-common Lyapunov variables for each design specification. Furthermore, the advantage of the dilated-LMI-based approach (ii) over the standard-based one (i) can be stated more rigorously by using the recovery properties of dilated LMIs stated in Theorems 2–5. 2008 Y. Ebihara, T. Hagiwara / Automatica 40 (2004) 2003 – 2009 Theorem 6. For the multiobjective H2 /D-stability synthesis problem, suppose that the standard-LMI-based approach with a common Lyapunov variable achieves an upper bound 2c > 0 of the cost functional. Then, the dilated-LMI-based approach with a common multiplier and a common prescribed scalar b=a −1 but with non-common Lyapunov variables always achieves an upper bound 2G that is lower than or equal to 2c , irrespective of a > 0. Namely, we have 2G 2c (∀a > 0). Proof. The assertion follows immediately from Theorems 2–5. Indeed, suppose that the standard-LMI-based approach achieves an upper bound 2c with the variables X2 = XH = XC = XS = X > 0 and Z2 > 0 in (8), (10), (12) and (14). Then, from the aforementioned theorems, the dilated-LMIbased approach ensures the achievement of the same upper bound 2c with the variables X2 = XH = XC = XS = X > 0, Z2 = Z2 > 0 and G2 = GH = GC = GS = G = −a(A − aI )−1 X in (9), (11), (13) and (15). This completes the proof. In the previous study (Apkarian et al., 2001), a dilated LMI for the H2 performance of continuous-time systems is derived and its usefulness is illustrated by studying multichannel H2 synthesis problems. Similar to the approach (ii), it was shown that we can design controllers that satisfy multiple H2 specifications by employing different Lyapunov variables for each channel. However, the approach in (Apkarian et al., 2001) is not proved to encompass the corresponding Lyapunov shaping paradigm, which is because the existence of “common G” as in the proof of Theorem 6 is not ensured explicitly. On the other hand, when we deal with multichannel H2 synthesis problems by means of the proposed dilated LMI (9), the existence of “common G” is obvious since the closed-loop matrix A is common to all channels. Hence, even when designing multichannel H2 controllers, we can ensure the advantage of (9) over (8) in the same way as Theorem 6. 3.4. Comparisons with the results in the discrete-time setting We have shown the usefulness of the proposed dilated LMIs when designing multiobjective H2 /D-stability controllers. Obviously, the results in Theorem 6 can be viewed as partial counterparts of those obtained in the discrete-time setting (Oliveira et al., 1999, 2002). Nevertheless, the results in this paper are still deficient in comparison with the discrete-time case. We summarize what deficiencies there are in the present work. 1. The H∞ synthesis cannot be handled. This is due to the restricted form of the LMIs in Lemma 1. Recall that the standard LMI for the H∞ performance includes a constant matrix and does not conform to (2). If a constant matrix is added to (2), such a matrix will be multiplied by (A − aI )−1 in (5) and make it hard to obtain the corresponding dilated LMI. 2. The way as to how the dilated LMIs recover the standard ones, i.e., the choice (X, G) = (X, −a(A − aI )−1 X) depends on A. This prevents us from applying the dilated LMIs to robust controller synthesis problems for polytopic-type uncertain systems. More precisely, even though the proposed dilated LMIs enable us to design robust controllers with parameter-dependent Lyapunov variables, we cannot guarantee the improvement of the control performance over the standard-LMI-based approaches with parameter-independent Lyapunov variables, unless the scalar parameter a is also optimized (Ebihara & Hagiwara, 2002a, 2002b). The necessity of line search is also observed in (Dettori & Scherer, 2000; Peaucelle et al., 2000; Shaked, 2001). In the discretetime case, the dilated-LMI-based approach is proved to encompass the standard one without any further ado (Oliveira et al., 1999, 2002). Despite these deficiencies, the results in this paper are still important. In particular, we have shown that the dilated LMIs work better than the standard ones without any line search when designing multiobjective H2 /D-stability controllers. This achievement is novel and distinguishes the present work from other related studies (Apkarian et al., 2001; Dettori & Scherer, 2000; Peaucelle et al., 2000; Shaked, 2001). 4. Illustrative examples Consider the LTI plant described by −0.32 0.04 0.01 1 0.42 x˙ = 0.45 0.99 0.64 x + 0 w + 0.74 u, −0.76 −0.37 −0.40 1 0.31 0 2 0 0 z = 0 0 1 x + 0 u, 0 0 0 5 y = [ 0.72 0.85 0.88 ]x + 2w. Our goal here is to design a full-order output-feedback controller K that minimizes ||Tzw ||2 subject to the D-stability constraint (A) ⊂ H(0.3). Solving this problem by the two approaches discussed in Section 3.3, we obtain the results in Table 1. The standardLMI-based approach yields the controller Kc (s) = −15.93(s + 0.53)(s + 0.41) . (s + 6.21)(s − 1.56)(s + 0.55) (27) This controller places closed-loop poles at −0.36 ± 0.21i, −0.99 ± 0.67i, −1.09 and −1.15. On the other hand, the dilated-LMI-based approach with a = 1.0 leads to K(s) = −10.60(s 2 + 0.84s + 0.18) . (s + 4.99)(s − 1.19)(s + 0.44) (28) Y. Ebihara, T. Hagiwara / Automatica 40 (2004) 2003 – 2009 Table 1 Computation results Approach Standard-LMI Dilated-LMI dilated LMIs is ensured without any line search. This important achievement is the main contribution of this work. Upper bounds 115.62 73.60 Actual costs 79.09 68.49 110 100 90 upper bounds actual costs 80 70 10−4 10−2 2009 1 102 104 106 parameter a (logarithmic scale) Fig. 1. The H2 costs achieved by the dilated LMIs. The closed-loop poles are −0.34 ± 0.26i, −0.63, −0.76 ± 0.79i and −1.15. In this problem, the H2 cost achieved by the pure H2 optimal controller is 46.08. This controller places closed-loop poles at −0.04, −0.10, −0.65, −0.68 ± 0.18i and −1.15 and hence does not meet the D-stability constraint. Recall that Theorem 6 ensures the advantage of the dilated-LMI-based approach irrespective of a > 0. To illustrate this, we applied the approach under different values of a and obtained the results shown in Fig. 1. This figure shows that, irrespective of a > 0, the dilated-LMI-based approach yields upper bounds that are lower than 115.62 achieved by the standard one. 5. Conclusion In this paper, we presented new results on the dilatedLMI-based controller synthesis for continuous-time systems. We have shown a constructive way to derive dilated LMIs from the standard ones, and through this particular derivation, some nice properties of the dilated LMIs have been clarified. By virtue of these properties, we proved an explicit advantage of the dilated LMIs over the standard ones when designing multiobjective H2 /D-stability controllers. In stark contrast with the existing results on the dilated-LMI-based continuous-time controller synthesis, the advantage of the References Apkarian, P., Tuan, H. D., & Bernussou, J. (2001). Continuoustime analysis and H2 multi-channel synthesis with enhanced LMI characterizations. IEEE Transactions on Automatic Control, 46(12), 1941–1946. Boyd, S. P., Ghaoui, L. E., Feron, E., & Balakrishnan, V. (1994). Linear matrix inequalities in system and control theory. Philadelphia, PA, SIAM. Chilali, M., & Gahinet, P. (1996). H∞ Design with pole placement constraints: An LMI approach. IEEE Transactions on Automatic Control, 41(3), 358–363. Dettori, M., & Scherer, C. W. (2000). New robust stability and performance conditions based on parameter dependent multipliers. Proceedings of the 39th CDC (pp. 4187–4192). Ebihara, Y. (2002). LMI-based multiobjective controller design with noncommon Lyapunov variables. Ph.D. dissertation, Kyoto University. Ebihara, Y., & Hagiwara, T. (2002a). New dilated LMI characterizations for continuous-time control design and robust multiobjective control. Proceedings of 2002 ACC (pp. 47–52). Ebihara, Y., & Hagiwara, T. (2002b). Robust controller synthesis with parameter-dependent Lyapunov Variables: A dilated LMI approach. Proceedings of the 41st CDC (pp. 4179–4184). Gahinet, P., & Apkarian, P. (1994). A linear matrix inequality approach to H∞ control. International Journal of Robust and Nonlinear Control, 4, 421–448. Iwasaki, T., & Skelton, R. E. (1994). All controllers for the general H∞ control problem: LMI existence conditions and state space formulas. Automatica, 30(8), 1307–1317. Masubuchi, I., Ohara, A., & Suda, N. (1998). LMI-based controller synthesis: A unified formulation and solution. International Journal of Robust and Nonlinear Control, 8, 669–686. Oliveira, M. C., Bernussou, J., & Geromel, J. C. (1999). A new discrete-time robust stability condition. Systems & Control Letters, 37, 261–265. Oliveira, M. C., Geromel, J. C., & Bernussou, J. (2002). Extended H2 and H∞ norm characterizations and controller parametrizations for discrete-time systems. International Journal of Control, 75, 666–679. Oliveira, M. C., & Skelton, R. E. (2001). Stability tests for constrained linear systems. Lecture notes in control and information sciences, Vol. 268. Perspectives in robust control. Berlin: Springer. Peaucelle, D., Arzelier, D., Bachelier, O., & Bernussou, J. (2000). A new robust D-stability condition for real convex polytopic uncertainty. Systems & Control Letters, 40, 21–30. Scherer, C. W., Gahinet, P., & Chilali, M. (1997). Multiobjective output-feedback control via LMI optimization. IEEE Transactions on Automatic Control, 42(7), 896–911. Shaked, U. (2001). Improved LMI representations for the analysis and the design of continuous-time systems with polytopic type uncertainty. IEEE Transactions on Automatic Control, 46(4), 652–656. Skelton, R. E., Iwasaki, T., & Grigoriadis, K. (1997). A unified algebraic approach to linear control design. London, Taylor & Francis.
© Copyright 2024