Journal of Process Control 10 (2000) 363±385 www.elsevier.com/locate/jprocont Review A tutorial on linear and bilinear matrix inequalities
Jeremy G. VanAntwerp, Richard D. Braatz* Large Scale Systems Research Laboratory, Department of Chemical Engineering, University of Illinois at Urbana-Champaign, 600 South Mathews Avenue, Box C-3, Urbana, Illinois 61801-3792, USA
Abstract This is a tutorial on the mathematical theory and process control applications of linear matrix inequalities (LMIs) and bilinear matrix inequalities (BMIs). Many convex inequalities common in process control applications are shown to be LMIs. Proofs are included to familiarize the reader with the mathematics of LMIs and BMIs. LMIs and BMIs are applied to several important process control applications including control structure selection, robust controller analysis and design, and optimal design of experiments. Software for solving LMI and BMI problems is reviewed. # 2000 Published by Elsevier Science Ltd. All rights reserved.
1. Introduction One of the main reasons for this is that process control engineers are generally unfamiliar with the mathematics A linear matrix inequality (LMI) is a convex con- of LMI/BMIs, and there is no introductory text avail- straint. Consequently, optimization problems with con- able to aid the control engineer in learning these vex objective functions and LMI constraints are mathematics. As of the writing of this paper, the only solvable relatively eciently with o-the-shelf software. text that covers LMIs in any depth is the research The form of an LMI is very general. Linear inequalities, monograph of Boyd and co-workers [22]. Although this convex quadratic inequalities, matrix norm inequalities, monograph is a useful roadmap for locating LMI and various constraints from control theory such as results scattered throughout the electrical engineering Lyapunov and Riccati inequalities can all be written as literature, it is not a textbook for teaching the concepts LMIs. Further, multiple LMIs can always be written as of LMIs to process control engineers. Furthermore, no a single LMI of larger dimension. Thus, LMIs are a existing text covers BMIs in any detail. useful tool for solving a wide variety of optimization This tutorial is an extension of a document used to and control problems. Most control problems of inter- train process control engineers at the University of Illi- est that cannot be written in terms of an LMI can be nois on the mathematical theory and applications of written in terms of a more general form known as a LMIs and BMIs. Besides training graduate students, the bilinear matrix inequality (BMI). Computations over tutorial is also intended for industrial process control BMI constraints are fundamentally more dicult than engineers who wish to understand the literature or use those over LMI constraints, and there does not exist o- LMI software, and experts from other ®elds (for exam- the-shelf algorithms for solving BMI problems. How- ple, process optimization) who wish to initiate investi- ever, algorithms are being developed for BMI problems, gations into LMI/BMIs. The only assumed background the best of which can be applied to process control is basic calculus, a course in state space control theory problems of modest complexity. [74,37], and a solid foundation in matrix theory [16,66]. The many ``nice'' theoretical properties of LMIs and The tutorial includes the proofs of several main BMIs have made them the emerging paradigm for for- results on LMIs. These are included for several reasons. mulating optimization and control problems. While First, many of the proofs are dicult to locate in the LMI/BMIs are gaining wide acceptance in academia, literature in the form that is most useful for applications they have had little impact in process control practice. to modern control problems. Second, the simplicity of the proofs provides some insights into the underlying * Corresponding author. Tel; +1-217-333-5073; fax: +1-217-33- geometry that manifests itself in terms of properties of 5052. the LMIs. Third, working through these proofs is the only E-mail address: [email protected] (R.D. Braatz). way to become suciently experienced in the algebraic
0959-1524/00/$ - see front matter # 2000 Published by Elsevier Science Ltd. All rights reserved. PII: S0959-1524(99)00056-6 364 J.G. VanAntwerp, R.D. Braatz / Journal of Process Control 10 (2000) 363±385
inputs. In many cases, it is impossible to even predict etch performance for the same system on two dierent runs. For this reason, it is impossible to maintain con- sistent etch quality without the use of feedback control. The feedback controller must be designed to be robust to the variability in process behavior as well as the Fig. 1. Reactive ion etcher in classical feedback form. nonlinear nature of the reactive ion etching process. Here, we consider the laboratory reactive ion etcher studied by Vincent et al. [146]. The manipulated variables manipulations necessary to be able to formulate the LMI/ were the power of the applied rf voltage and the throttle BMIs to solve new process control problems. Finally, valve position which speci®es the input gas ¯owrate, and students learn far more by working through problems the controlled variables were the ¯uorine concentration or proofs than from reading theorem after theorem. and the bias voltage. Like many other chemical processes This paper is organized as follows. First an example is described in the literature, the plasma dynamics of a used to motivate studies in LMI/BMIs. The second sec- reactive ion etching process were reasonably well descri- tion de®nes the LMI and discusses some of its basic bed as a static input nonlinearity N followed by a linear properties. The third section shows how inequalities of time-invariant (LTI) plant PL (see Fig. 1), which is the many dierent types can be written as LMIs or BMIs. well known Hammerstein model structure [51,106,134]. The fourth section discusses optimization problems over This nonlinear model was identi®ed using an iterative LMI or BMI constraints, and why such optimization least squares algorithm with data obtained from an problems can be eciently solved numerically. The ®fth experimental system by exciting it with a pseudo-random section reviews algorithms and software packages used binary signal with varying amplitude [146]. The identi®ed to solve LMI/BMI optimization problems, and the sixth LTI plant for their experimental process was section lists LMI/BMI problems that are important in 2 3 process control applications. This is followed by con- 1:89e :5s s 38:2 35:9 s 37:8 6 7 cluding remarks. 6 s 5:37 s 0:160 s2 6:5s 20:2 7 6 7 PL6 7 1 4 0:0239e :5s s 9:6 0:143 s 38:9 5 2. Motivating example: a reactive ion etcher s 1:05 s 0:214 s2 3:28s 4:14
A large number of control problems can be written in The natural controller structure to use has the form ^ 1 terms of LMIs or BMIs that cannot be solved using K N KL where KL is designed to stabilize the linear ^ 1 Lyapunov equations, Riccati equations, spectral factor- portion of the plant PL and N is an approximate ization, or other classical techniques. The following is inverse of the static nonlinearity N. If the input non- an industrial process control problem in which the only linearity N were identi®ed perfectly then N^ 1 would be tractable solution is via an optimization over LMI and an exact inverse of N and there would be an identity BMI constaints. mapping from KL to PL. However, in practice the iden- Etching is known to be a highly nonlinear multi- ti®cation is not perfect, and there is a nonlinear map- variable process that is strongly dependent on reactor ping from KL to PL. Furthermore, it is highly unlikely geometry. Attempts to control etch characteristics that the system is nonlinear only at the process input. usually manipulate the reactor pressure, gas ¯ow rate, Output nonlinearity is also a probability. and the power applied to the electrodes. However, due Nonlinearities in both the input and the output can be to many disturbances, complicated reaction dynamics, rigorously accounted for by the uncertainty description and the general lack of detailed fundamental under- shown in Fig. 2. The operators ÁI and ÁO can vary standing of the plasma behavior, it is impossible to within set bounds as functions of time, and can achieve predict etch performance for a system given a set of an identical input-output mapping for any possible
Fig. 2. Reactive ion etcher with input and output nonlinearities modeled as uncertainty. J.G. VanAntwerp, R.D. Braatz / Journal of Process Control 10 (2000) 363±385 365 nonlinearity within the magnitude of the bounds set by equivalent to n polynomial inequalities. To see this, the uncertainty weights WI and WO. consider the well-known result in matrix theory (e.g. The only known method for designing a globally page 951 of [154]) that an n by n real symmetric matrix optimal robust nonlinear controller for this process is A is positive de®nite if and only if all of its principal by formulating the controller design as an optimization minors are positive. Let Aij be the ijth element of A. over LMI and BMI contraints (see Section 7.5 for Recall that the principal minors of A are details). The formulation allows a direct optimization of 02 31 the worst-case closed loop performance over the set of A11 A12 A13 plants described by the nonlinear uncertainty descrip- A11 A12 B6 7C A11; det ; det@4 A21 A22 A23 5A; tion. The BMI-based controller responded more than A21 A22 twice as fast to set point changes than a carefully tuned A31 A32 A33 classical controller (linear quadratic control, whose 02 31 computation was via a Ricatti equation), while at the A11 ... A1n same time providing guaranteed robustness [143]. B6 . . 7C ...; det@4 . . 5A 4
An1 ... Ann 3. The linear matrix inequality We apply this result to give that the LMI (2) is equiva- Here we de®ne the LMI and some of its basic prop- lent to: erties. We will use upper case Roman to refer to matrices, lower case Roman to refer to vectors or scalars, lower Xm F x F > 0 a linear inequality case Greek to refer to scalars, and upper case calligraphic 0;11 i i;11 i1 to refer to sets. The symbol 8 should be read ``for all'' and the symbol 2 should be read ``is an element of''. The ! ! notation Rm denotes the set of real vectors of length m, Xm Xm and RnÂn denotes the set of real n  n matrices. F0;11 xiFi;11 F0;22 xiFi;22 i1 i1 ! ! 3.1. De®nition Xm Xm F0;12 xiFi;12 F0;21 xiFi;21 > 0 A linear matrix inequality (LMI) has the form: i1 i1
Xm a quadratic inequality Fx F0 xiFi > 0 2 i1 . . where x 2Rm; F 2RnÂn. The inequality means that i 02 31 Fx is a positive de®nite matrix, that is, Fx 11 ... Fx 1k B6 7C det . . > 0 zTFx z > 0; 8z 6 0; z 2Rn: 3 @4 . . 5A Fx k1 ... Fx kk
The symmetric matrices Fi; i 0; 1; ...; m are ®xed kth order polynomial inequality and x is the variable. Thus, Fx is an ane function of the elements of x. . Eq. (2) is a strict LMI. Requiring only that Fx be . positive semide®nite is referred to as a nonstrict LMI. The strict LMI is feasible if the set fgxjFx > 0 is det Fx > 0 nth order polynomial inequality nonempty (a similar de®nition applies to nonstrict LMIs). Any feasible nonstrict LMI can be reduced to an equiva- lent strict LMI that is feasible by eliminating implicit The n polynomial inequalities in x range from order 1 equality constraints and then reducing the resulting LMI to order n. by removing any constant nullspace ([22], page 19). We will therefore focus our attention on strict LMIs. 3.3. Convexity
3.2. LMI equivalence to polynomial inequalities A set C is said to be convex if lx 1 l y 2 C for all x; y 2 C and l 2 0; 1 [107]. An important property It is informative to represent the LMI in terms of of LMIs is that the set fgxjFx > 0 is convex, that is, scalar inequalities. More speci®cally, the LMI (2) is the LMI (2) forms a convex constraint on x. To see this, 366 J.G. VanAntwerp, R.D. Braatz / Journal of Process Control 10 (2000) 363±385 let x and y be two vectors such that Fx > 0 and where ÈÉ Fy > 0, and let l 2 0; 1 . Then 1 2 q Fi diag Fi ; Fi ; ...; Fi ; 8i 0; ...; m 12 Xm F lx 1 l y F lx 1 l y F ÈÉ 0 i i i and diag X ; X ; ...; X is a block diagonal matrix i1 1 2 q Xm with blocks X1; X2; ...; Xq. This result can be proved lF0 1 l F0 l xiFi from the fact that the eigenvalues of a block diagonal i1 matrix are equal to the union of the eigenvalues of the Xm blocks, or from the de®nition of positive de®niteness. 1 l yiFi i1 lFx 1 l Fy 4. The generality of LMIs and BMIs
> 0: 5 This section shows how many common inequalities can be written as LMIs. In addition, it shows how many control properties of interest can be written exactly in 3.4. LMIs are not unique terms of the feasibility of an LMI. Such a problem is referred to as an LMI feasibility problem. The same set of variables x can be represented as the feasible set of dierent LMIs. For instance, if Ax is 4.1. Linear constraints can be expressed as an LMI positive de®nite then Ax subject to a congruence transformation (see section 14.7 of [154]) is also positive Linear constraints are ubiquitous in process control de®nite: applications. Model Predictive Control has become the most popular multivariable controller design method in A > 0 () xTAx > 0; 8x 6 0 6 many industries precisely because of its ability to address linear constraints on process variables [32,48,61,95,110, () zTMTAMz > 0; 8z 6 0; M nonsingular 7 114]. The standard linear programming and quadratic programming model predictive control formulations () MTAM > 0 8 can be written in terms of LMIs. Here we show the ®rst step, which is to write the linear constraints on process variables as LMI constraints. This implies, for example, that some rearrangements Consider the general linear constraint Ax < b written of matrix elements do not change the feasible set of the as n scalar inequalities: LMI. Xm AB 0 I AB 0 I bi Aijxj > 0; i 1; ...; n 13 > 0 () > 0 j1 CD I 0 CD I 0 DC where b 2Rn, A 2RnÂm, and x 2Rm. Each of the n () > 0 9 scalar inequalities is an LMI. Since multiple LMIs can BA be written as a single LMI, the linear inequalities (13) can be expressed as a single LMI. 3.5. Multiple LMIs can be expressed as a single LMI 4.2. Stability of linear systems One of the advantages of representing process control problems with LMIs is the ability to consider multiple Stability is one of the most basic needs for any closed control requirements by appending additional LMIs. loop system. Some methods for analyzing the stability Consider a set de®ned by q LMIs: of linear systems are covered in undergraduate process control textbooks [102,133]. Moreover, some nonlinear F1 x > 0; F2 x > 0;...;Fq x > 0 10 processes can be analyzed (at least to some degree) with linear techniques by performing a change of variables, Then an equivalent single LMI is given by such as in binary distillation [91] and pH neutralization [68,101]. Xm ÈÉThe Lyapunov method for analyzing stability is 1 2 q Fx F0 xiFi diag F x ; F x ; ...; F x > 0; described in most texts on process dynamics [70,108]. i1 The basic idea is to search for a positive de®nite func- 11 tion of the state (called the Lyapunov function) whose J.G. VanAntwerp, R.D. Braatz / Journal of Process Control 10 (2000) 363±385 367 time derivative is negative de®nite. A necessary and LTI techniques, including reactive ion etching [140], sucient condition for the linear system packed bed reactors [46], and most batch processes [9]. : In Section 4.2, we showed how testing the stability of x Ax 14 a linear system could be posed as an LMI feasibility problem. Now let us consider a generalization of that to be stable is the existence of a Lyapunov function problem to testing the stability of a set of linear time Vx xTPx where P is a symmetric positive de®nite varying systems that are described by a convex hull of matrix such that the time derivative of V is negative for matrices (a matrix polytope): all x 6 0 [108]: : x At x; At 2CofgA1; ...; AL 19 dVx : : xTPx xTPx dt An alternative way of writing this is [105]: Á xT ATP PA x < 0; 8x 6 0 15 : XL XL x At x; At liAi; 8li50; li 1: 20 () ATP PA < 0 16 i1 i1
A necessary and sucient condition for the existence This is an LMI, where P is the variable. To see this, of a quadratic Lyapunov function Vx xTPx that select a basis for symmetric n  n matrices. proves the stability of (20) is the existence of P PT > As an example basis, for i5j de®ne Eij as the matrix 0 that satis®es: with its i; j and j; i elements equal to one, and all of dVx : : its other elements equal to zero. There are m xTPx xTPx < 0; 8x 6 0; ij dt nn 1 =2 linearly independent matrices E and any 21 symmetric matrix P can be written uniquely as 8At 2CofgA1; ...; AL Xn Xn ij P PijE ; 17 Âà T T j1 i5j () x At P PA t x < 0; 8x 6 0; 22 8At 2CofgA1; ...; AL ij where Pij is the i; j element of P. Thus the matrices E form a basis for symmetric n  n matrices (in fact, if the ij T columns of each E are stacked up as vectors, then the () At P PA t < 0; 8At 2CofgA1; ...; AL resulting vectors form an orthogonal basis, which could be made orthonormal by scaling). 23 Substituting for P in terms of its basis matrices gives ! ! the alternative form for the Lyapunov inequality XL T XL XL ! ! () liAi P P liAi < 0; 8li50; li 1 Xn Xn Xn Xn T T ij ij i1 i1 i1 A P PA A PijE PijE A j1 i5j j1 i5j 24 Xn Xn Á P ATEij EijA < 0 XL Á XL ij T j1 i5j () li Ai P PAi < 0; 8li50; li 1 25 18 i1 i1
T which is in the form of an LMI (2), with F0 0 and () Ai P PAi < 0; 8i 1; ...; L 26 T ij ij Fk A E E A; for k 1; ...; m. The elements of the vector x in (2) are the Pij; i5j, stacked up on top of The search for P that satis®es these inequalities is an each other. LMI feasibility problem. This condition is also a su- cient condition for the stability of nonlinear time vary- 4.3. Stability of nonlinear and time varying systems ing systems where the Jacobian of the nonlinear system is contained within the convex hull in (20) [84]. There Many of the processes commonly encountered in are several diculties in applying the LMI condition for process control applications can be adequately modeled analyzing stability of nonlinear systems. First, it is very as being linear time invariant (LTI). However, many dicult to construct a convex hull for which the Jaco- chemical processes cannot be adequately analyzed using bian of a nonlinear system is provably contained within. 368 J.G. VanAntwerp, R.D. Braatz / Journal of Process Control 10 (2000) 363±385
Second, such a description will usually be highly con- Here Ax corresponds to Sx in the LMI (28), and servative, since the convex hull overbounds the Jacobian Qx and Rx correspond to I. of the real nonlinear system. Third, each new vertex adds another matrix inequality to the LMI feasibility 4.6. Ellipsoidal inequality problem (26). For a system with a large number of states (which is equal to the dimension of A) and ver- Ellipsoid constraints are important in process identi- tices (L), solving the LMI feasibility problem (26) can ®cation, parameter estimation, and statistics [15,27, become computationally prohibitive. The strength of 41,85]; as well as certain fast model predictive control the approach is that LMIs for controller synthesis for algorithms [138,139]. Applications recently described in systems of the form (20) are relatively easy to construct the literature include crystallization processes [88,93], [22,77,131]. polymer ®lm extruders [54], and paper machines [138, 139]. 4.4. The Schur complement lemma An ellipsoid described by
T 1 T The Schur complement lemma converts a class of x xc P x xc < 1; P P > 0 32 convex nonlinear inequalities that appears regularly in control problems to an LMI. The convex nonlinear can be expressed as an LMI using the Schur comple- inequalities are ment lemma with Qx 1, Rx P, and Sx T x xc : 1 T Rx > 0; Qx Sx Rx Sx > 0; 27 T 1 x xc T T > 0: 33 where Qx Qx ; Rx Rx ,andSx depend x xc P anely on x. The Schur complement lemma converts this set of convex nonlinear inequalities into the equivalent LMI 4.7. Algebraic Riccati inequality Qx Sx > 0: 28 Algebraic Riccati equations are used extensively in Sx T Rx optimal control, as described in textbooks on advanced process control [111,130], which describe applications to A proof of the Schur complement lemma using only chemical reactors, distillation columns, and other pro- elementary calculus is given in the Appendix. In what cesses. A result involving a Riccati equation can be follows, the Schur complement lemma is applied to replaced with an equivalent result where the equality is several inequalities that appear in process control. replaced by an inequality [151]. More speci®cally, these optimal controllers can be constructed by computing a 4.5. Maximum singular value positive de®nite symmetric matrix P that satis®es the algebraic Riccati inequality: The maximum singular value measures the maximum gain of a multivariable system, where the magnitude of ATP PA PBR 1BTP Q < 0 34 the input and output vector is quanti®ed by the Eucli- dean norm [130]. It is also very useful for quantifying where A and B are ®xed, Q is a ®xed symmetric matrix, frequency-domain performance and robustness for and R is a ®xed symmetric positive de®nite matrix. multivariable systems [96,130]. Process applications are The Riccati inequality is quadratic in P but can be provided in many popular undergraduate process con- expressed as a linear matrix inequality by applying the trol textbooks [102,126]. Schur complement lemma: The maximum singular value of a matrix A which ATP PA QPB anely depends on x is denoted by Ax , which is > 0: 35 the square root of the largest eigenvalue of Ax Ax T. BTPR The inequality Ax < 1 is a nonlinear convex con- straint on x that may be written as an LMI using the The next two sections provide examples of algebraic Schur complement lemma: Riccati inequalities for analyzing the properties of linear Ax < 1 () Ax Ax T< I 29 or nonlinear systems. () I Ax I 1Ax T> 0 30 4.8. Bounded real lemma IAx The Bounded real lemma forms the basis for LMI () T > 0 31 Ax I approaches to robust process control which have been J.G. VanAntwerp, R.D. Braatz / Journal of Process Control 10 (2000) 363±385 369 applied to reactive ion etching [140,143], polymer columns [129], packed bed reactors [46] and a reactive extruders [141], and paper machines [141], and gain ion etching [143]. A property that is regularly exploited scheduling which has been applied to chemical reactors in the development of robustness analysis tools [14,72] [12,13]. Although the Bounded real lemma has applica- for linear systems subject to linear or nonlinear pertur- tion to the control of both linear and nonlinear pro- bations is passivity. The linear system (36) is said to be cesses, the actual result is based on the state space passive if system representation of a linear system : ut Tyt dt50 42 x Ax Bu; y Cx Du; x 0 0; 36 0 where A 2RnÂn; B 2RnÂp; C 2RpÂn,andD 2RpÂp for all u and 50. This property is equivalent to the are given data. Assume that A is stable and that existence of P PT > 0 such that [22] A; B; C is minimal [74]. The transfer function matrix is "# ATP PA PB CT Gs CsI A 1B D: 37 40: 43 BTP C DT D The worst-case performance of a system measured in terms of the integral squared errors of the input and It is instructive to show the connection between the output is quanti®ed by the H1 norm [157]: bounded real lemma and the positive real lemma [5], especially since it is often referred to in the robust con- k Gs k1 sup Gs sup Gj ! : 38 Re s >0 !2R trol literature. A standard result from network theory [10,45,125] is that passivity is equivalent to Gs in (37) The H1 norm can be written in terms of an LMI. To being positive real, that is, see this, we will use a result from the literature [158] that à 2 Gs Gs 50 8Refgs > 0 44 the H1 norm of Gs is less than if and only if I DTD > 0 and there exists P PT > 0 such that Á Á where Gs à is the complex conjugate transpose of Gs . T T T A P PA C C PB C D The relationship between bounded real and positive 1 Á 1 Á real is that I Gs I Gs is strictly positive real  2I DTD BTP DTC < 0 39 if and only if Gs is strictly bounded real. This follows from [87] The Schur complement lemma implies that this Ric- à cati inequality is equivalent to the existence of P PT > A < 1 () A A < I 45 0 such that the following LMI holds: à 1 à 1 "#ÂÃÂà () I A 2I 2A A I A > 0 46 ATP PA CTC PB CTD Âà > 0 40 1 BTP DTC 2I DTD () I Aà I Aà I A I Aà I A which is equivalent to  I A 1 > 0 47 "# ATP PA CTCPB CTD < 0: 41 () I Aà 1 I Aà I A I A 1> 0 48 BTP DTCDTD 2I Âà () I A I A 1 à I A I A 1> 0 49 It is common to incorporate weights on the input u and output y so that the condition of interest is whether the H1 norm of W1 s Gs W2 s is less than 1. A sys- 4.10. The S procedure tem with an H1 norm less than one is said to be strictly bounded real. This condition is checked by testing the The S procedure greatly extends the usefulness of feasibility of the LMI using the state-space matrices for LMIs by allowing non-LMI conditions that commonly the product W1 s Gs W2 s . arise in nonlinear systems analysis to be represented as LMIs (although with some conservatism). This techni- 4.9. Positive real lemma que has been applied to the analysis of pH neutraliza- tion processes [119] and crystallization processes [116]. Robustness analysis has been widely applied in the First we will describe the S procedure as it applies to process control literature. Examples include distillation quadratic functions, and then discuss its application to 370 J.G. VanAntwerp, R.D. Braatz / Journal of Process Control 10 (2000) 363±385 quadratic forms. Let 0; ...; p be quadratic scalar 4.11. Stability of linear systems with nonlinear functions of x 2Rn: perturbations
T T T i x x Tix 2ui x i; i 0; ...; p; Ti Ti The derivation of LMI feasibility problems to analyze the stability or performance of linear systems subject to 50 linear/nonlinear time invariant/varying perturbations is The existence of 150; ...;p50 such that rather straightforward conceptually [22], although the algebra can be messy for more complex systems Xp [115,118]. For continuous time systems, the basic x x 50; 8x; 51 0 i i approach is to postulate a positive de®nite Lyapunov i1 function of the state and some undetermined matrices, implies that and then apply the S procedure (if necessary) to derive LMI conditions on the undetermined matrices which 0 x 50; 8x such that i x 50; i 1; ...; p: 52 imply that the time derivative of the Lyapunov function is negative de®nite. For discrete time systems, the divi- To see why this is true, assume there exists ded dierence of the Lyapunov function is used instead 150; ...;p50 such that (51) holds for all of the time derivative. Here we show how this approach i x 50; i 1; ...; p. Then is applied to a system of especial relevance to process control applications. Xp Consider a discrete time system subject to slope- x 5 x 50; 8x: 53 0 i i restricted static nonlinearities: i1 xk 1 Ax k B qk Note that (51) is equivalent to qk Cx k 61
Xp T0 u0 Ti ui with the nonlinearities described by T i T 50 54 u 0 u i 0 i1 i i qi k i qi k qi k 40; for i 1; ...; m 62 since with the local slope restrictions xTTx 2uTx 50; 8x 55 i qi k 1 i qi k 0 < < Tii; for i 1; ...; m 63 x T Tu x qi k 1 qi k () 50; 8x 56 1 uT 1 th where Tii is the maximum slope of the i nonlinearity. This can be used to represent a linear process with x T Tu x () 50; 8x; 57 actuator limitation nonlinearities which is controlled by uT an antiwindup compensator [36,78,30], or a closed loop system with each component being either a linear sys- Tu () 50: 58 tem or a dynamic arti®cial neural network [117,120]. uT Both of these types of closed loop systems have been extensively studied in the process control literature (see Hence the above S procedure can be equivalently the above references and citations therein). written in terms of quadratic forms. Instead of writing The Lur'e-Lyapunov function is de®ned by the above version which is completely in terms of non- m strict inequalities, we will provide here a version that X qi k Vxk xT k Px k 2 Q d 64 applies to the case where the main inequality is strict i ii i1 0 (the proof is similar). Let T0; ...; Tp be symmetric matrices. If there exists 150; ...;p50 such that where P is positive de®nite and the Qii are nonnegative Xp so that the Lyapunov function is positive de®nite. The T T > 0; 59 0 i i ®rst term is the standard quadratic Lyapunov function i1 which is discussed in many state space systems text- then books [100] and in textbooks on process analysis [70,108], which describe applications to polymerization T T x T0x > 08x 6 0 such that x Tix50; i 1; ...; p: 60 and other chemical reactors. The second term was J.G. VanAntwerp, R.D. Braatz / Journal of Process Control 10 (2000) 363±385 371 introduced by Lur'e [86] to include the nonlinearities of ®nding some matrix  of compatible dimensions (62) explicitly in the Lyapunov function. such that The method of Lyapunov for discrete time systems is to write the divided dierence for the Lyapunov function: A PTÂTQ QTÂP < 0 71
Vxk 1 Vxk 65 This equation is solvable for some  if and only if the following two conditions hold: with the state vector substituted in using (61). The T overall system is globally asymptotically stable if the WPAWP < 0 72 undetermined matrices P and Qii can be computed so that Vxk 1 Vxk is less than zero. The non- T WQAWQ < 0 73 linearities are bounded in the divided dierence using (63) and the mean value theorem, and the S procedure is where WP and WQ are matrices whose columns are used to convert the divided dierence subject to the bases for the null spaces of P and Q, respectively. A inequalities (62) to an LMI. With some algebra to col- proof of this result is given in [58]. lect the terms [115,118], it is found that a sucient con- dition for the global asymptotic stability of (61)±(63) is 4.13. Bilinear matrix inequality the existence of a positive-de®nite matrix P and diag- onal positive semide®nite matrices Q and R 2RhÂh such Bilinear inequalities arise in pooling and blending that problems [147], systems analysis [140], and nonlinear programming. A bilinear matrix inequality (BMI) is of M M 1;1 1;2 > 0 66 the form: M2;1 M2;2 Xm Xn Xm Xn Fx ; y F0 xiFi yjGj xiyjHij > 0 where i1 j1 i1 j1 74 T T T M1;1 A PA P A I C TQC A I 67 where Gj and Hij are symmetric matrices of the same T T T dimension as F , and y 2Rn. Bilinear matrix inequal- M1;2 A PB A I C TQCB i ities were popularized by Safonov and co-workers in a A I TCTQ CTR 68 series of proceedings papers [63±65,125], and ®rst applied to a nontrivial process description (i.e., a che- mical reactive ion etcher) by VanAntwerp and Braatz T T T M2;1 B PA B C TQC A I QC A I RC [140], and was later applied to paper machines [141]. A BMI is an LMI in x for ®xed y and an LMI in y for 69 ®xed x, and so is convex in x and convex in y. The bilinear terms make the set not jointly convex in x and y. T T T T T To see this, consider the simplest BMI which is the M2;2 B PB B C TQCB QCB B C Q 2R bilinear inequality 70 1 xy > 0; 75 and T diagfgTii . The new matrix R is introduced by the S procedure. This is an LMI feasibility problem that where x and y are scalar variables. One way to see that has been applied to the analysis of pH neutralization this set is nonconvex is to graph the set in the xy-plane processes [119] and crystallization processes [116] under and apply the de®nition of convexity. Another way to nonlinear feedback control. see this is to consider two elements of the set that con- tradict the de®nition of convexity. For example, con- 4.12. Variable reduction lemma sider x; y equal to the values 0:1; 7:9 and 7:9; 0:1 . Both values satisfy the bilinear inequality since The variable reduction lemma allows the solution of 1 0:1 7:9 1 7:9 0:1 0:21 > 0. But the point algebraic Riccati inequalities that involve a matrix of on the line half way between the two values unknown dimension. This often arises when ®nding the 1=2 0:1; 7:91=27 :9; 0:1 4; 4 does not satisfy controller that minimizes the H1 norm (see Section 7.5 the bilinear inequality: 1 4Á4 15 < 0. for an example). Besides bilinear and general quadratic inequalities Given a symmetric matrix A 2RnÂn and two matrices P and Q of column dimension n, consider the problem xTQx cTx p > 0; 76 372 J.G. VanAntwerp, R.D. Braatz / Journal of Process Control 10 (2000) 363±385 general polynomial inequalities can also be written as important to understand that being NP-hard is a property BMIs. Consider, for example, the nonlinear inequality of the problem itself, not of any particular algorithm. It is also important to understand that having a problem x3 yz < 1 77 be NP-hard does not imply that practical algorithms are not possible. Practical algorithms for NP-hard problems By de®ning x2 w, and x v, this inequality is exist and typically involve approximation, heuristics, equivalent to: branch-and-bound, or local search [35,62,104]. Deter- mining whether a problem is polynomial time or NP- 1 xw yz > 0 78 hard informs the systems engineer what kind of accu- racy and speed can be expected by the best algorithms, x v50 and what kinds of algorithms to investigate for provid- v x50 ing practical solutions to the problem. Suppose that a real valued function fx is de®ned on w vx50 a convex set C 2Rn. The function fx is convex on C if vx w50 [107] f lx 1 l y 4lfx 1 l fy 79 Since a BMI describes sets that are not necessarily convex, they can describe much wider classes of con- for all x; y 2 C and l 2 0; 1 .Aconvex optimization straint sets than LMIs, and can be used to represent problem has the form more types of optimization and control problems. The main drawback of BMIs is that they are much more inf fx ; 80 dicult to handle computationally than LMIs. x2C
where fx is a convex function in x, C is a convex set, 5. Optimization problems and inf refers to the in®mum over C. If the in®mum is achieved by an element in C, then the minimization Many optimization and control problems can be problem will be written as min. written in terms of ®nding a feasible solution to a set of Well known problems that can be formulated as con- LMIs or BMIs. Most problems, however, are best writ- vex optimization problems include linear programming ten in terms of optimizing a simple objective function and convex quadratic programming. The advantage of over a set of LMIs or BMIs. There is a fundamental formulating control problems in terms of convex opti- dierence between the computational requirements for mization problems (when possible) is that wide classes optimization problems over LMIs, and those over of convex optimization problems are in the class P [97]. BMIs. This section begins with an introduction to con- Being in P means that these problems can be provably vex optimization and computational complexity, which solved eciently on a computer. This makes convex provides a fundamental framework for understanding optimization problems desirable for solving large scale the relative complexities of optimization problems. This systems problems. Convex optimization problems often is followed by the de®nition of some optimization pro- occur in engineering practice and many can be written blems that appear when formulating and solving control as LMIs. This is the strength of using LMI formula- problems using LMIs/BMIs. tions. Convex optimizations over LMIs are solvable in polynomial time. 5.1. Computational complexity and convexity Other systems engineering problems cannot be written in terms of LMIs, but can be written in terms of BMIs. Optimization problems are generally characterized as Nearly every control problem of interest can be written being in one of two classes: P and NP-hard [62,104]. The in terms of an optimization problems over BMIs. These class P refers to problems in which the time needed to optimization problems, however, are NP-hard [135], exactly solve the problem can always be bounded by a which implies that it is highly unlikely that there exists a single function which is polynomial in the amount of polynomial-time algorithm for solving these problems. data needed to de®ne the problem. Such problems are This means that algorithms for solving optimization said to be solvable in polynomial time. Although the problems over BMIs are currently limited to problems exact consequences of a problem being NP-hard is still a of modest size. Algorithms and their expected perfor- fundamental open question in the theory of computa- mance will be discussed in more detail in Section 6. For tional complexity, it is generally accepted that a pro- the rest of this section we review the most common LMI blem being NP-hard means that its solution cannot be and BMI optimization problems that appear in control computed in polynomial time in the worst case. It is applications. J.G. VanAntwerp, R.D. Braatz / Journal of Process Control 10 (2000) 363±385 373
5.2. Semide®nite programming lB A > 0. As l is reduced from some suciently high value, at some point the matrix lB A will lose rank, at The following optimization problem is commonly which point there exists a nonzero vector y that solves referred to as a semide®nite program (SDP) [4]: (85), implying that this value of l is the largest general- ized eigenvalue. Hence inf cTx 81 x Fx >0 lmax min l inf l 86 lB A50 lB A>0
One SDP which often arises in control applications is the LMI eigenvalue problem (EVP). It is the minimiza- Often it is desired to minimize the largest generalized tion of the maximum eigenvalue of a matrix that eigenvalue of two symmetric matrices which depend depends anely on the variable x, subject to an LMI anely on a variable x, subject to an LMI constraint on constraint on x. Many performance analysis tests, such x. as computing the H1 norm in (38), can be written in terms of an EVP [144]. Two common forms of the EVP inf lmax Ax ; Bx : 87 Bx >0 are presented so that readers will recognize them: Cx >0
inf l 82 Here lmax Ax ; Bx is the largest generalized eigen- x;l lI Ax >0 value of two matrices, A and B, each of which depend Bx >0 anely on x. From (86) this optimization problem is equivalent to inf l 83 x;l inf l: 88 Ax ;l >0 lBx Ax >0 Bx >0 Cx >0 where Ax ; l is ane in x and l. The equivalence of (81), (82), and (83) will now be The problem of minimizing the maximum generalized demonstrated. The LMI eigenvalue problem (82) can be eigenvalue is a quasiconvex objective function subject to written in the form (83) by de®ning Ax ; l a convex constraint, where quasiconvexity means that diagfglI Ax ; Bx (recall that multiple LMIs can be written as a single LMI of larger dimension). To show lmax A x 1 z ; B x 1 z that a problem in the form (83) can be written in the ÈÉ 4 max lmax Ax ; Bx ; lmax Az ; Bz 89 form (81), de®ne x^ xT l T, Fx ^ Ax ^ ,andcT 0T1 T where 0 is a vector of zeros. To see that (81) transforms to (82) consider for all 2 0; 1 and all feasible x and z. To see that this is true, ®rst de®ne l^ equal to the right hand side of (89). inf cTx inf l inf l inf l: 84 Then Fx >0 cTx