Convergent LMI relaxations for robust analysis of uncertain linear systems using lifted polynomial parameter-dependent Lyapunov functions☆
Introduction
In the last decade, research in systems and control have witnessed an effort to compute better and better approximate solutions to robust optimization problems [1]. Take for instance the problem of robust stability analysis of uncertain linear systems [2]. Using Lyapunov stability theory, this problem can be cast as the feasibility of parameter-dependent Linear Matrix Inequalities (LMIs), which are optimization problems of infinite dimension. Much effort went into trying to produce finite dimension tests, through the use of relaxations or other techniques. The following works, among many others, effectively reduced the gap between quadratic stability, which is the simplest parameter-independent form of robust stability test, and robust stability [3], [4], [5], [6], [7], [8], [9], [10]. Related problems, such as robust and gain-scheduled synthesis, filtering, including performance indexes as the and the norms, can also be expressed in terms of parametrized LMIs.
Recent approaches to parameter-dependent LMIs work by transforming the original infinite-dimensional problem into a sequence of finite dimension LMI relaxations, in some cases with guarantees of convergence. Most relaxations are based on finding polynomial parameter-dependent Lyapunov matrix solutions of large enough degree. Indeed, as shown in [11], polynomial solutions completely characterize parameter-dependent LMIs with parameters in compact sets. Within the context of robust stability and and performance analysis, the following are some of the works that follow such polynomial approach [6], [8], [12], [13], [14], [15], [16], [10]. A related strategy is the use of sum-of-squares decompositions [9], [17].
With several convergent LMI relaxations for robust analysis available in the literature, it seems that the challenge now is to improve the quality of the relaxations. For instance, in terms of efficiency, i.e. the ability to obtain the same results with less computational burden; or the easiness to extend a given stability condition to cope with similar problems, such as control and filtering design; and even practical aspects such as the difficulties associated with programming the proposed conditions. Typically, polynomial approaches require a significant amount of preliminary manipulations before the conditions can be numerically solved.
Recently, interesting LMI relaxations have been introduced in [7], [18] that cope with the robust stability analysis of continuous-time time-invariant linear systems in polytopic domains. In these works, robust stability conditions are reformulated as parameter-dependent LMIs by imposing a special parameter-dependent polynomial Lyapunov function. This Lyapunov function depends polynomially on the uncertain system matrices and on a generic polynomial parameter-dependent matrix (Lyapunov matrix) to be determined. Constraining the particular structure of the parameter-dependent Lyapunov matrix, i.e. constant, affine, etc. the associated Lyapunov function can be viewed as a special polynomial whose degree is given as a function of a certain integer . For each , finite dimension LMI relaxations can be obtained by means of Finsler’s Lemma [19]. Numerical experiments show that conservativeness of such conditions is reduced as increases. Extensions to deal with performance analysis have also been presented in [20], [21]. In [20], the authors raise the possibility of changing the dynamic matrix in the Lyapunov function by an arbitrary matrix, possibly reducing the conservativeness of the evaluations.
This paper investigates the structure of the Lyapunov function originally proposed in [7], [18] in more detail. One of the main contributions is to show that the Lyapunov matrix can be constrained to be constant, i.e. parameter independent, without loss of generality if is allowed to be large enough. This can explain the good practical behavior of the relaxations proposed in [7], [18], [20], [21], where only sufficient conditions for robust stability have been provided. It is also shown how simple yet numerically effective LMI conditions can be obtained in a lifted parameter-space. The use of the proposed Lyapunov function to compute guaranteed costs is also discussed and some conditions that ensure convergence to necessity in this case are also given. Numerical aspects are discussed and illustrated by examples.
Section snippets
Problem statement
Consider the linear time-invariant continuous-time uncertain system with , , , , , and . The triplet is assumed to belong to a convex bounded (polytope-type) uncertain domain where is the unit simplex given by In other words, any uncertain matrix triplet can be written as a convex combination of vertices ,
-Robust stability
Robust stability and performance analysis of the uncertain system (1) is verified using the particular polynomial parameter-dependent quadratic Lyapunov function where is an integer, is such that the matrix is defined by and the matrix , a polynomial in , is partitioned as where each block , for all
-robust performance
In this section the previous discussion is extended to the question of robust performance. Parameter-dependent LMIs which are equivalent to the inequalities of Lemma 1 can also be obtained by using the concept of -robust stability. The following Lemma provides the condition of -robust performance analysis.
Lemma 9 Matrix is -robustly stable and the inequality holds for all if and only if there exists a polynomial parameter-dependent matrix in the form
Numerical complexity
The numerical complexity associated with an optimization problem based on LMIs can be estimated by the number of scalar variables and the number of LMI rows involved. Table 1 shows the values of and for Theorem 8 (T8) on -robust stability and Theorem 11 (T11) on -robust performance.
Note that Theorem 11 considers affine on . If is made constant, can be subtracted from . Moreover, if it is known a priori that matrix is Hurwitz in , the constraints
Numerical experiments
All numerical tests have been performed in a Pentium IV 2.6 GHz, 512 MB RAM, using the SeDuMi [30] and Yalmip [31] within Matlab. The source code can be downloaded from http://www.dt.fee.unicamp.br/~ricfow/robust.htm.
In all robust stability examples, Theorem 8 provided a positive answer for . The worst case norm has been reached by the LMI conditions of Theorem 11 for (with affine on ) and (with ).
Example 1 Theorem 8 is used to check robust stability of 1000 stable polytopes
Conclusion
A new approach to produce convergent LMI relaxations to cope with the robust stability analysis of continuous-time uncertain linear systems in polytopic domains has been presented. A special polynomial parameter-dependent Lyapunov matrix is used to derive equivalent parameter-dependent LMIs which can be lifted into a larger dimensional space. In this space, efficient LMI relaxations can be obtained. An extension to cope with guaranteed cost computation is also provided. The results are
Acknowledgments
Supported by CNPq and FAPESP. The authors thank Pierre-Alexandre Bliman for many fruitful discussions and also the anonymous reviewers.
References (32)
- et al.
LMI characterization of structural and robust stability
Linear Algebra and Its Applications
(1998) An existence result for polynomial solutions of parameter-dependent LMIs
Systems & Control Letters
(2004)- et al.
LMI conditions for robust stability analysis based on polynomially parameter-dependent Lyapunov functions
Systems & Control Letters
(2006) - et al.
Stability of polytopes of matrices via affine parameter-dependent Lyapunov functions: Asymptotically exact LMI conditions
Linear Algebra and Its Applications
(2005) - et al.
Convex Optimization
(2004) - et al.
- et al.
Affine parameter-dependent Lyapunov functions and real parametric uncertainty
IEEE Transactions on Automatic Control
(1996) - et al.
An LMI condition for the robust stability of uncertain continuous-time linear systems
IEEE Transactions on Automatic Control
(2002) - et al.
Polynomially parameter-dependent Lyapunov functions for robust stability of polytopic systems: An LMI approach
IEEE Transactions on Automatic Control
(2005) - R.C.L.F. Oliveira, V.J.S. Leite, M.C. de Oliveira, P.L.D. Peres, An LMI characterization of polynomial...
Matrix sum-of-squares relaxations for robust semi-definite programs
Mathematical Programming Series B
Parameter-dependent LMIs in robust analysis: Characterization of homogeneous polynomially parameter-dependent solutions via LMI relaxations
IEEE Transactions on Automatic Control
A convex approach to robust stability for linear systems with uncertain scalar parameters
SIAM Journal on Control and Optimization
Cited by (63)
Graph Lyapunov function for switching stabilization and distributed computation
2020, AutomaticaCitation Excerpt :It should be pointed out that the PLF has been used earlier for the robust stability/stabilization of linear time-invariant (LTI) systems in Ebihara, Peaucelle, and Arzelier (2011) and Trégouët, Peaucelle, Arzelier, and Ebihara (2013), nonlinear systems (Kruszewski, Wang, & Guerra, 2008), and periodic systems (Bittanti, Bolzern, & Colaneri, 1985). For continuous-time nonlinear systems, the notion of non-monotonic Lyapunov functions traces back to the use of the higher order derivatives of the Lyapunov functions developed in early work (Butz, 1969; Meigoli & Nikravesh, 2012), which was further explored in the recent papers (Ahmadi & Parrilo, 2011; Ebihara, Hagiwara, Peaucelle, & Arzelier, 2005; Oliveira, de Oliveira, & Peres, 2008). In spite of the extensive literature in this area, to the authors’ knowledge, the GLF method has not been applied to the stabilization problem so far.
Robust stability of Markov jump linear systems through randomized evaluations
2019, Applied Mathematics and ComputationCitation Excerpt :The task of checking whether (1) is robustly mean square stable is cumbersome because it requires the solution of an infinite dimensional problem, based on parameter-dependent linear matrix inequalities (LMIs). Such a problem is difficult to solve even for small-dimensional systems [22]. A strategy to attack the problem relies on robust LMIs—they can be approximated by polynomial solutions of arbitrary degree on the uncertain parameters, but only in few cases upper bounds for the degree are known [3].
Robust performance for uncertain systems via Lyapunov functions with higher order terms
2019, Journal of the Franklin InstituteStatic output-feedback robust gain-scheduling control with guaranteed H<inf>2</inf> performance
2018, Journal of the Franklin InstituteCitation Excerpt :At this end, it is necessary to relax the infinite dimensional optimization problem into finite set of LMIs to be solved. Fortunately, solvability of PLMIs is manageable in the presence of recent theoretical and computational relaxation tools such as matrix Sum-Of-Squares [30], Slack variable approach [31], and matrix coefficient check methods [32,33]. In this paper, a specialized parser ROLMIP [34] (developed as a tool to perform such PLMIs manipulation and LMI relaxations) has been used to implement the conditions of Theorems 3 and4 to obtain the optimal solution of the convex optimization problem.
- ☆
This research has been supported in part by grants from “Fundação de Amparo à Pesquisa do Estado de São Paulo” — FAPESP and from “Conselho Nacional de Desenvolvimento Científico e Tecnológico” — CNPq, Brazil.