Asian Journal of Control, Vol. 5, No. 4, pp. 467-475, December 2003 467 SLIDING MODES, ∆-MODULATORS, AND GENERALIZED PROPORTIONAL INTEGRAL CONTROL OF LINEAR SYSTEMS Hebertt Sira-Ramírez ABSTRACT In this article, we attempt a reapproachment between sliding mode control of linear systems and classical control through the possibilities of evading state measurements and circumventing the use of asymptotic state observers in the sliding surface synthesis. This is shown to be possible thanks to the use of integral state reconstructors combined with iterated integral output error compensation. The proposed scheme is also robust with respect to unmatched perturbation inputs. A connection between sliding modes and classical analog ∆-modulators, and their natural generalization, is also brought to attention as a tool for the realization of integral reconstructor based sliding mode control schemes. KeyWords: Generalized PI control, sliding mode control, delta-modulators. I. INTRODUCTION The many advantages of sliding mode control are well reported, founded, and illustrated, in the existing literature which, incidentally, has reached a significantly advanced state of maturity. Among the shortcomings of sliding mode control, we simply mention that sliding mode control technique is, fundamentally, a state-based discontinuous feedback control technique. The lack of complete knowledge of the state vector components forces the designer to use asymptotic state observers, of the Luenberger, or of the sliding mode type, or to resort to direct output feedback control schemes. Unfortunately, the first approach is not robust with respect to unforseen exogenous perturbation inputs, even if they happen to be of the “classical type” (by this we mean: steps, ramps, parabolas, etc.) while the second approach, while being quite limited in nature, is not applicable in non-minimum phase systems and it fails in unmatched perturbation input cases. For general background on sliding mode control, we refer the reader to the seminal books by Utkin, [8,9] and the recent books by Utkin, Guldner and Shi [10] and that by Edwards and Spurgeon [1]. Recent Manuscript received January 9 2002; accepted February 3 2003. The authors are with Cinvestav-Ipn, Sección de Mecatrónica, Department Ingeniería Eléctrica, Avenida IPN, # 2508, Col. San Pedro Zacatenco, A.P. 14740 México, D.F., México. This research was supported by the Centro de Investigación y Estudios Avanzados del IPN (CINVESTAV-IPN), México, D.F., México. developments, advances and applications, of the sliding mode control area are found in the book by Perruquetti and Barbot [6]. In this article, we propose a new approach to the synthesis of sliding mode feedback control schemes for linear, time invariant, controllable and observable Single Input Single Output (SISO) systems. The approach consists in using integral reconstructors of the state vector for the sliding surface synthesis. Integral reconstructors were introduced, in the realm of continuous feedback strategies, addressed as Generalized Proportional Integral (GPI) control schemes, in the work of Fliess [3]. A complete theoretical account of integral reconstructions and GPI control has been presented in recent articles by Fliess et al (see [4,5]). Integral reconstructors use only inputs, outputs and iterated integrals of such signals while neglecting the influence of the constant, but unknown, initial conditions and the effect of classical perturbation inputs. Hence, our approach naturally leads to dynamic input-output feedback schemes for the synthesis of the sliding surface and, due to its inherent input-output nature, it requires no matching conditions whatsoever. Due to the unstable nature of the state reconstruction errors associated with the integral reconstructos, a given suitable state-dependent sliding surface can be synthesized via integral reconstructors provided its expression is suitably modified with a sufficiently large linear combination of iterated output error compensation terms. The aim is to synthesize the sliding surface so as to exhibit an exponentially asymptotically stable dynamics under, closed loop, ideal sliding condi- Asian Journal of Control, Vol. 5, No. 4, December 2003 468 u W ξ −W x 1 _ s Fig. 1. Classical analog ∆-modulator. tions. In this article, we show, and advocate, that the sliding surface design problem is greatly facilitated by resorting to the system flatness property (For the definitions and theoretical basis of differential flatness, the interested reader is referred to the work of Fliess and his coworkers [2]. See also the Appendix). We bring to the attention of the reader, the close connection between sliding modes and classical analog ∆-modulators. In particular, we show the relevance of ∆-modulators, and their generalizations, in the dynamic synthesis of state-dependent sliding surfaces using integral state reconstructors. A complete account of ∆modulators, which never benefited from the theoretical basis of sliding mode control, is found in the classical book by Steele [7]. Section 2 presents a review of the simplest analog ∆-modulator and its connection with sliding mode control. In this section we explore the links between ∆-modulators and sliding mode control of simple integration plants naturally involving GPI control schemes. Section 3 deals with a generalization of ∆-modulators for the synthesis of sliding surfaces in general linear SISO controllable and observable systems. Section 4 is devoted to present a design example involving the normalized model of two wheeled, frictionless, cars joined by a spring. We illustrates the robust and convenient features of our proposed GPI sliding mode controllers via digital computer simulations. The appendix presents some generalities regarding the concept of flatness and integral state reconstructors in observable linear, n-dimensional, SISO systems. II. ∆-MODULATORS AND GPI CONTROLLERS Consider the basic block diagram of Fig. 1 depicting a classical analog ∆-modulator, traditionally used in voice encoding systems. The following theorem summarizes the relation of ∆-modulators with sliding mode control and depicts the basic features of performance of the basic modulator. Theorem 2.1. Consider the ∆-modulator of Fig. 1. Given a bounded C1 signal ξ(t), with bounded first order time derivative, ξ , there exists a strictly positive gain, W, such that x(t) → ξ(t) in a finite amount of time th, pro- vided the following encoding condition is satisfied, W > sup ξ(t ) (1) Moreover, from any arbitrary initial value of the tracking, or local encoding, error e(t0) = x(t0) − ξ(t0), a sliding motion exists on the perfect encoding condition e = 0 for all t > th, where the quantity th is bounded by th ≤ T, with T e(t0 ) satisfying, T ≤ t0 + . W − sup ξ(t ) Proof. From the figure, the variables in the ∆-modulator satisfy the following relations: x = u u = W sign(ξ − x) (2) e = x −ξ Clearly, e = −W sign (e) − ξ(t ) and since ξ(t ) is assumed to be bounded, choosing W > sup ξ(t ) we have, for e > 0: ee = −W e − eξ(t ) = −W e − e ξ sign(e) ≤ −W e + e sup ξ = −(W − sup ξ ) e < 0 (3) A sliding regime exists on e = 0 for all time t after the hitting time th (see [8]). Under ideal sliding, or encoding, conditions, e = 0; e = 0, we have that x = ξ(t) and the equivalent (average) value of the coded output signal u is given by ueq = ξ(t ) for all t ≥ th. The ∆-modulator output u ideally differentiates the modulator input signal ξ(t) in an equivalent control sense ([8]). The role of the ∆-modulator in sliding mode control schemes avoiding full state measurements of simple linear plants will be clear from the following examples. 2.1 A second order example Consider the unperturbed second order plant ÿ = u, where only the output y and the input u are available for use on a stabilizing sliding mode feedback scheme. A traditional sliding mode controller with sliding surface: σ = y + λy is thus forbidden due to the unavailability of the signal y . An integral reconstructor for y is obt tained directly from the expression: yˆ (t ) = ∫0 u ( ρ )d ρ which, we know, differs from the actual value of y by the unknown constant initial condition, y 0 , i.e. y = ŷ + y 0 . A modified sliding surface, inspired on the GPI control technique (see [5]), can now be proposed as: t t σˆ = yˆ + λ1 y + λ0 ∫0 y ( ρ ) d ρ = ∫0 u ( ρ )d ρ + λ1 y t +λ0 ∫0 y ( ρ )d ρ (4) S.R. Hebertt: Sliding Modes, ∆-Modulators, and Generalized Proportional Integral Control ∆-Modulator ∆-Modulator GPI Compensator PI Compensator 0 λ1+ 0 Plant − σ^ u W λ0 __ s −W 469 1 _ s2 − σ^ (a+λ ) ___ λ _____ (b+λ2) + 1 + 0 2 s s y ξ u W −W Plant 1 __ s2 y 1 _ s y 1 _ s Fig. 3. GPI ∆-modulator sliding mode controller scheme for a perturbed plant. Fig. 2. GPI sliding mode controller exhibiting a ∆-modulator. 2.2 A perturbed second order plant The proposed sliding mode controller is clearly synthesized via an output PI compensating signal feeding a ∆-modulator as depicted in Fig. 2. According to the results of Theorem 2.1, under ideal sliding conditions, the average value of the output signal, ueq, of the ∆-modulator is given by ueq = −λ1 y − λ0 y , i.e. the second order plant is, on the average, being regulated by a PD controller. Note that the integral output error compensation term in equation (4) compensates for the constant error, incurred in the estimation of the state y through the integral reconstructor, when the sliding condition σ̂ = 0 is enforced. The modified sliding surface, σ̂ , is seen to be equivalent to t σˆ = y + λ1 y + λ0 ∫0 y ( ρ )d ρ − y 0 which, under ideal sliding conditions: σˆ = 0 ; σˆ = 0 , yields the following equivalent control and the following asymptotically stable closed loop dynamics, provided appropriate choice of the design parameters λ0 and λ1 is made. ueq = −λ1 y − λ0 y, y + λ1 y + λ0 y = 0 t Indeed, defining ζ = ∫0 y ( ρ )d ρ − y 0 / λ0 , the ideal sliding condition σ̂ = 0 yields the following linear homogeneous system with one unknown initial condition, y = −λ1 y − λ0ζ , y (0) = y0 y ζ = y, ζ (0) = 0 λ0 clearly, y + λ1 y + λ0 y = 0. A sliding mode controller for the stabilization of the system is then of the form: u = −W sign(σˆ ) t t σˆ = ∫0 u ( ρ )d ρ + λ1 y + λ0 ∫0 y ( ρ )d ρ = λ1 y + ∫ [u ( ρ ) + λ0 y ( ρ )]d ρ The above developments allow one, for instance, to readily propose a sliding mode feedback controller for the following perturbed second order plant ÿ = ay + by + u + ξ1(t − τ), with ξ being a constant, unknown, non-zero perturbation input appearing at the unknown instant τ > 0, and a and b being known parameters. We know that a PID controller of the form: u = −(b + λ2) y − t (a + λ1)y − λ0 ∫0 y ( ρ )d ρ , exponentially asymptotically stabilizes y to zero, irrespectively of ξ and τ, for an appropriate choice of the set of parameters, λ0, λ1 and λ2. The closed loop dynamics is given by (s3 + λ2s2 + λ1s + λ0)y = 0. If we regard the proposed PID controller as an average (equivalent) controller synthesis, then the results of Theorem 2.1 allow us to propose the following GPI compensation plus ∆-modulation scheme which recovers, on the average, the previous PID controller features. u = −W sign σˆ t t σˆ = ∫0 u ( ρ )d ρ + (b + λ2 ) y + (a + λ1 ) ∫0 y ( ρ )d ρ t ρ +λ0 ∫0 ∫0 y (ζ ) dζ d ρ Indeed, according to the results of the theorem, the equivalent control ueq, obtained from the condition σ̂ = 0 is given by t ueq = −(b + λ2 ) y − (a + λ1 ) y − λ0 ∫0 y ( ρ )d ρ (5) t Note that ∫0 [u ( ρ ) + ay ( ρ )]d ρ + by qualifies as an integral reconstructor, ŷ , of the state, y . Such reconstructor is related to the actual value y of the unmeasured state by the expression y = yˆ + y 0 − by0 + ξ(t – τ)1(t – τ). The sliding surface σ̂ is then equivalent to t t ρ σˆ = y + λ2 y + λ1 ∫0 y ( ρ )d ρ + λ0 ∫0 ∫0 y (ζ )dζ d ρ − y 0 + by0 − ξ (t − τ )1(t − τ ) (6) Under the sliding condition: σ̂ = 0, the closed loop dynamics is given by the following linear system with Asian Journal of Control, Vol. 5, No. 4, December 2003 470 some unknown initial states and subject to an exogenous sudden step input perturbation, υ, at time τ. y = −λ2 y − λ1ζ 1 , y (0) = y0 − σ^ 0 −1 −1 SP(s )+r(s ) λ y − by0 ξ , υ = 1(t − τ ) ζ1 = y + 0 ζ 2 + υ , ζ 1 (0) = − 0 λ1 λ1 λ1 ζ = y, ζ (0) = 0 2 Generalized ∆-Modulator W −W "Classical perturbation" input ξ Plant y u c(sI − A)−1b SQ(s−1) y 2 The unperturbed ideal sliding motions exhibits a characteristic polynomial q(s) given by: q(s) = s3 + λ2s2 + λ1s + λ0, which can be made Hurwitz, and the equilibrium point of the perturbed ideal sliding dynamics is seen to ξ be given by: y = 0, ζ1 = 0, ζ2 = − . λ0 Fig. 4. GPI Generalized ∆-modulator sliding mode controller scheme for a classically perturbed linear plant. K F M III. GPI SLIDING MODE CONTROL AND GENERALIZED ∆-MODULATORS The previous developments point to the fact that state dependent sliding surfaces can be effectively designed on the basis of state integral reconstructors, which are linear combinations of iterated integrals of inputs and outputs. Since the integral reconstructors neglect the initial state conditions and the external perturbation inputs, the integrally reconstructed sliding surface must be sufficiently compensated in order to counteract their de-stabilizing effects, when the ideal sliding conditions are valid. It is not difficult to show (See the Appendix) that in an observable linear system of the form: x = Ax + bu; y = cx, the state vector x can always be reconstructed, modulo the effect of neglected initial conditions and classical external perturbation inputs, in terms of vectors of integral polynomials: P(s−1) and Q(s−1) as follows: xˆ = P ( s −1 ) y + Q ( s −1 )u (7) A sliding surface, σ, which may have been synthesized from a viewpoint such as atness, of the form: σ = Sx with S being a constant row vector, can then be modi_ed by considering, σˆ = Sxˆ + r ( s −1 ) y = [ SP( s −1 ) + r ( s −1 )] y + SQ( s −1 )u (8) where r(s−1) is an integral polynomial scalar function which compensates for the neglected state initial conditions and the influence of classical external perturbation inputs. The complete control scheme is shown in Fig. 4 where the ∆-modulator of the previous examples is now addressed as a “Generalized ∆-modulator”. The sliding surface synthesis problem may then be stated as follows: Given a controllable and observable linear plant y = c(sI − A)−1bu = [n(s)/d(s)]u whose state vector x defining the realization: x = Ax + bu; y = cx, M x1 x2 Fig. 5. A two mass spring system. can always be integrally parameterized, modulo initial conditions, by the integral reconstructor x̂ = P(s−1)y + Q(s−1)u, find a row vector S and an integral polynomial r(s−1) such that the closed loop dynamics, corresponding to the ideal sliding conditions σ̂ = 0, σ̂ = 0, is exponentially asymptotically stable. The involved relations are: σˆ = [ SP( s −1 ) + r ( s −1 )] y + SQ( s −1 )u u = −W sign σˆ (9) d ( s) y = n( s)u + ξ The choice of the row vector S and the iterated integral compensation term r(s−1) is greatly facilitated by resorting to the system flatness property, (see [2]) which due to the controllability assumption is, therefore, guaranteed (see the Appendix). IV. A MASS-SPRING SYSTEM Consider the following unperturbed normalized model of two wheeled carts joined by a spring shown in Fig. 5 x1 = u + ( x2 − x1 ) x2 = −( x2 − x1 ) (10) y = x2 A state dependent exponentially stabilizing sliding surface can be shown to be given by σ ( x1 − x2 ) + γ 2 ( x1 − x2 ) + γ 1 x2 + γ 0 x2 (11) S.R. Hebertt: Sliding Modes, ∆-Modulators, and Generalized Proportional Integral Control with the coefficients γ2, γ1 and γ0 chosen so that the characteristic polynomial s3 + γ2s2 + γ1s + γ0 is Hurwitz. Note that the state expression for σ is motivated by the desirable closed loop dynamics y (3) + γ 2 y + γ 1 y + γ 0 y = 0 The system is found to be observable since y = x2 qualifies as the flat output. The complete differential parameterization of the states and the input in terms of the at output is given by x2 = y, x2 = y , x1 = ÿ + y, x1 = y(3) + y , u = y(4) + ÿ. The last flatness relations, the system equations and the observability relations: y = x2, y = x2 , ÿ = −(x2 − x1) and y(3) = − ( x2 − x1 ) , help us in obtaining the integral reconstructors of the state. These are found to be1: (2) (3) 1 = ( ∫ u ) + 2( ∫ u ) − ( ∫ u ) xl1 = ( ∫ u ) − y, xl (12) (3) 2 = ( ∫ u ) − 2( ∫ y ) xl2 = y, xl or, in complex variable, integral polynomial vector notation: xl1 s −2 −1 −1 −3 − 1 xl 1 2s s − s u y = + 0 l 1 x2 l −2 s −1 s −3 x 2 (13) (2) σ=m y (3) + k5 yˆ + k4 yˆ + k3 y + k2 ( ∫ u ) + k1 ( ∫ y ) + k0 ( ∫ (3) (14) y) which is equivalent to the following expression, for some unknown constants ρ1, ρ2 and ρ3 σˆ = y (3) + k5 y + k4 y + k3 y + k2 ( ∫ y ) + k1 ( ∫ + k0 ( ∫ (3) y ) + ρ1 + ρ 2 t + ρ3t (2) y) (i) We use the following notation: ( ∫ φ ) = (1) t with ( ∫ φ ) = ( ∫ φ ) = ∫0 φ (σ )d σ . ζ 1 = (∫ y) + k k1 ρ ρ (∫ ζ 2 ) + 1 , ζ 2 = (∫ y) + 0 (∫ ζ 3 ) + 2 , k2 k2 k1 k1 ζ 3 = (∫ y) + ρ3 k0 (16) we have y (3) = −k5 y − k4 y − k3 y − k2ζ 1 , y (0) = y0 , y (0) = y 0 , y (0) = y0 ρ k ζ1 = y + 1 ζ 2 , ζ 1 (0) = 1 k2 k2 (15) (17) k ρ ζ2 = y + 0 ζ 3 , ζ 2 (0) = 2 k1 k1 ρ ζ3 = y, ζ 3 (0) = 3 k0 The characteristic polynomial of this linear homogeneous system coincides with the desired one. Suitable choice of the design parameters renders the characteristic polynomial Hurwitz. The proposed choice of σ̂ results then in: (3) (2) σˆ = (k4 − 2)( ∫ u ) + k5 ( ∫ u ) + ( ∫ u ) + (k3 − 2k5 ) y (2) y ) + k0 ( ∫ (3) (18) y) or, in complex Laplace transform notation 4 − 2k4 + k2 k1 k0 σˆ = (k3 − 2k5 ) + + 2 + 3y s s s k − 2 k 1 + 4 3 + 52 + u s s s (19) 4.1 Simulation results Figure 6 depicts the controlled responses of the normalized system to a given set of adverse initial conditions. The chosen closed loop characteristic polynomial, corresponding to σ̂ = 0 was set to be, p(s) = (s2 + 2ζωns + ω n2 )3, with ξ = 0.8, ωn = 1.2. The gain W was set to 12. 4.2 Unmatched perturbations 2 Under ideal sliding motions on σ̂ = 0, the closed loop system is a linear homogeneous system in y with charac 1 teristic polynomial, s6 + k5s5 + k4s4 + k3s3 + k2s2 + k1s + k0. Indeed, defining, +(4 − 2k4 + k2 )( ∫ y ) + k1 ( ∫ The integral reconstructors of the state differ from the actual values of the states by at most a second order time polynomial of the form: α + β t + γ t2. Thus, the modified sliding surface requires three iterated integrals of the output stabilization error. We propose then a sliding surface achieving, in closed loop, an exponentially asymptotically stable dynamics for the at output y. Such a surface is given by 471 t σ1 0 0 ∫∫ σ "∫0 i −1 φ (σ i )d σ i … d σ 1 To see that the proposed synthesis technique requires no matching condition of the exogenous perturbation inputs, consider the case in which a constant, unknown, perturbation input suddenly affects the underactuated part of the system. Assume such a perturbation appears at the unknown time, τ, as follows, Asian Journal of Control, Vol. 5, No. 4, December 2003 472 0.5 0.5 x1(t) x2(t) 0 −0.5 0 2 0 −2 −4 0 ξ 0 2 4 6 8 10 12 −0.5 0 2 5 x1(t) 10 15 20 25 30 10 15 20 25 30 10 15 20 25 30 0 σ(t) σ(t) −2 2 20 4 6 8 10 12 −4 0 5 20 u(t) u(t) 0 0 −20 0 x2(t) 2 4 6 8 10 12 Fig. 6. GPI sliding mode controlled responses of two car system. −20 0 5 Fig. 7. GPI sliding mode controlled responses of perturbed two car system. x1 = u + ( x2 − x1 ) x2 = −( x2 − x1 ) + ξ 1(t − τ ) (20) y = x2 Then it can be seen that the integral reconstructors of the state differ from their actual values, at most, by a third order time polynomial of the form a + β t + γ t2 + κt3. This implies that the sliding surface expression must be of the form: (2) σˆ = m y (3) + k6 yˆ + k5 yˆ + k4 y + k3 ( ∫ y ) + k2 ( ∫ y ) + k1 ( ∫ (3) y ) + k0 ( ∫ (4) (21) y) The integral input-output parameterization of such a sliding surface is now given by: 4 − 2k5 + k3 k2 k1 k0 + 2 + 3 + 4y σˆ = (k4 − 2k6 ) + s s s s k − 2 k 1 + 5 3 + 62 + u s s s 4.4 The case of unavailable at output Consider now the following perturbed system with a different output signal, x1 = u + ( x2 − x1 ) + ξ 1(t − τ ) x2 = −( x2 − x1 ) y = x1 The system is observable from the given output and hence an integral reconstructor of the state vector is possible to obtain. The integral reconstructors of the unperturbed state components are given by: (2) (3) 2 = −( ∫ u ) + 2 y xl2 = ( ∫ u ) − y, xl (3) 1 = ( ∫ u ) + ( ∫ u ) − 2( ∫ y ) xl1 = y, xl (22) Note that under such a constant perturbation, the perturbed differential parameterization of the state is: x1 = ÿ + y − ξ1(t − τ). This means that in steady state x1 will converge towards the unknown constant value −ξ. The output y = x2 will, nevertheless, converge to zero as desired. 4.3 Simulation results Figure 7 depicts the performance of the proposed sliding mode feedback controller with the desired characteristic polynomial, corresponding to σ̂ = 0, set to be, p(s) = (s2 + 2ζωns + ω n2 )3(s + r), with ζ = 0.85, ωn = 1, r = 1.5 and W = 10. The perturbation amplitude was prescribed as ξ = 0.15. The perturbation appears at time τ = 12 time units. (23) (24) It is easy to see that the discrepancy between the integrally reconstructed states and their actual values, due to the effect of the initial conditions and of the constant unmatched perturbation input, is, at most, of the form: α + β t + γ t2. This means that the expression of a state dependent sliding surface must be compensated by a linear combination involving at most three iterated integrals of the output stabilization error. Contrary to our previous design example we assume the objective is now to stabilize the variable y = x1 to zero. Note that the perturbed differential parameterization of the output y = x1, in terms of the at output x2, is given by: y = x2 + x2. Consider then, as a sliding surface coordinate function, the expression: (2) (3) 1 + λ3 y + λ2 ( ∫ y ) + λ1 ( ∫ y ) + λ0 ( ∫ y ) σˆ = xl (25) In terms of the actual value of the at output, the sliding S.R. Hebertt: Sliding Modes, ∆-Modulators, and Generalized Proportional Integral Control surface coordinate function σ̂ is equivalent to σˆ = x1 + λ3 y + λ2 ( ∫ y ) + λ1 ( ∫ (2) y ) + λ0 ( ∫ (3) y) (26) + r1 + r2 t + r3t 2 On σ = 0, the resulting closed loop dynamics is given by y (4) + λ3 y (3) + λ2 y + λ1 y + λ0 y = 0 (27) which can be guaranteed to be an exponentially asymptotically stable dynamics upon an appropriate choice of the design coefficients. V. CONCLUSIONS A new approach to sliding mode control of linear time-invariant SISO controllable and observable systems has been presented. The proposed technique, which evades the need for asymptotic state observers, exploits an input-output viewpoint, much as in classical linear control theory, granted by the help of integral state reconstructors involving only iterated integrals of input and output signals aided by output error iterated integral compensation. A suitable generalization of analog ∆-modulators, widely used in the early days of voice encoding systems, allows one to obtain a dynamic input-output based synthesis of the sliding surface and demonstrates a natural connection with the recently introduced GPI control design technique. The proposed design approach suffers from none of the limitations of direct output based sliding mode control and it is robust even with respect to unmatched classical exogenous perturbation inputs due to its underlying input-output character. We presented several simple design examples illustrating the advantages and potential of the approach. The proposed GPI-sliding mode control schemes, through generalized ∆-modulators, are most adequately implemented when the flatness of the system is suitably exploited in the prescription of a desired homogeneous (closed loop) linear dynamics for the at output. The results here explained are easily extendable to multivariable continuous-time linear systems and also to linear discrete time systems. These will be the subject of forthcoming publications. APPENDIX A.1 Flatness in linear systems Flatness has been introduced in [2] from a general nonlinear multivariable systems viewpoint. A smooth single input system, in state space representation, x = f(x, u), x ∈ Rn, u ∈ R, is said to be flat if there exists a scalar output y, which is a function of the state x, such 473 that there exist an n-vector function φ and a scalar function ψ such that: x = φ(y, y , …, y(n−1)) and u = ψ(y, y , …, y(n)). We say that y completely differentially parameterizes all system variables (states, input and, eventually, the system output) and usually refer to φ and ψ as a differential parametrization. Evidently, in the adopted context, a flat single input system is (locally) exactly linearizable by means of static state feedback and state and input coordinates transformations, i.e., it is (locally) equivalent to the linear Brunovsky system, y(n) = υ, with υ being the transformed input. Flat outputs are thus devoid of any zero dynamics. In the context of single input linear systems of the form: x = Ax + bu, the system is at if and only if it is controllable in the sense of Kalman, i.e. the controllability matrix C = [b, Ab, … An−1b] is invertible. Furthermore, a flat output may be determined as the linear combination of the states obtained from the last row of the inverse of the controllability matrix, i.e. y = [0 " 0 1]C −1 x Evidently, any constant multiple of y also qualifies as a flat output and, also, the equivalence result is global. A.2 Integral reconstructors Here we use the developments in [3] and [5] to demonstrate that for any observable linear Single Input Single Output (SISO) system the state vector can always be expressed in terms of iterated integrals of the inputs and the outputs, modulo the influence of initial conditions. Such estimators are known as “integral reconstructors” of the state or, also, as an “integral input-output parameterization” of the state vector. Due to the initial conditions, integral reconstructors differ from the actual value of the state vector by constant errors and iterated integrals of such constant errors. This classical, unstable, reconstruction errors (constants, ramps, parabolas, etc.) can always be stably compensated in feedback control schemes using such reconstructors by the suitable addition of a sufficiently large number of iterated integrals of the output stabilization or tracking error. Consider the observable linear, time-invariant, SISO system x = Ax + bu, x(0) = x0 , y = cx (A.1) Integrating the system, in the sense of Mikusiński (see [5]) which neglects the effect of initial conditions, we obtain x u x = A +b s s (A.2) Iterating on this functional relation once more we obtain: Asian Journal of Control, Vol. 5, No. 4, December 2003 474 x(t ) = A2 x u u + Ab 2 + b 2 s s s Iterating a total of n − 1 times, we obtain the following implicit representation of the state vector: u ( s) x( s) n −1 x( s) = An −1 n −1 + ∑ Ai −1b i s s i =1 (A.3) On the other hand, consider the output signal y and its successive time derivatives " 0 1 " 0 s u ( s ) (A.4) ⋅ % 0 # " cb s ( n − 2) Integrating n − 1 times, the obtained expression (A.4), we have: 1 s n −1 1 x( s) sn−2 = + ( ) y s O M s n −1 # 1 s u(s) (A.7) 1 s n −1 1 y(s) − M s n − 2 # 1 s u (s) (A.8) where P and Q are vectors of integral polynomials. We address the expression (A.8), which does not take into account the influence of the initial states, the integral reconstructor of the state vector and denote it by x̂ . Such an “open loop” estimate of the state is based only on iterated integrals of inputs and outputs. The integral state reconstructor may be used, in principle, on an input-output synthesis of a sliding surface σ = Sx, as long as the expression for σ involving the state estimates is complemented with additional compensation terms which counteract the effect of the neglected initial states values and, possibly, of the unaccounted classical exogenous perturbation inputs. Such compensation terms only require further iterated integrations of the output stabilization error (or of the output tracking error in output tracking problems). REFERENCES (A.5) Thanks to the observability of the system we can uniquely solve for the quantity x(s)/sn−1 from (A.5) to obtain: 1 s n −1 1 x( s) −1 n − 2 = O s s n −1 # 1 s i =1 u ( s) u ( s) si x = P ( s −1 ) y + Q ( s −1 )u 1 s u(s) = Ox ( s ) + M # ( n − 2) I s 1 s n −1 1 sn−2 # 1 n −1 + ∑ Ai −1b 1 s n −1 1 sn−2 ( ) y s M − # 1 s In other words, we have that the state vector x can always be expressed as: 1 c s y ( s ) = cA x( s ) # # ( n −1) n −1 s cA 0 0 cb 0 + # # n−2 cA b " 1 s n −1 −1 1 n −1 x=A O n−2 s # 1 (A.6) We can now combine the expression (A.6) with (A.3) to obtain: 1. Edwards, C. and S. K. Spurgeon, Sliding Mode Control, Taylor and Francis, London (1998). 2. Fliess, M., J. Lévine, P. Martin, and P. Rouchon, “Flatness and Defect of Nonlinear Systems: Introductory Theory and Examples,” Int. J. Contr., Vol. 61, pp. 1327-1361 (1995). 3. Fliess, M., R. Marquez, and E. Delaleau, “State Feedbacks without Asymptotic Observers and Generalized PID regulators,” Nonlinear Control in the Year 2000, A. Isidori, F. Lamnabhi-Lagarrigue, W. Respondek (Eds.), Lecture Notes in Control and Information Sciences, Springer, London (2000). 4. Fliess, M., and R. Márquez, “Continuous Time Linear Predictive Control and Flatness: A Module-Theoretic Setting with Examples,” Int. J. Contr., Vol. 73, pp. 606-623 (2000). 5. Fliess, M., R. Marquez, E. Delaleau, H. Sira- Ramírez, “Correcteurs Proportionneles Intégraux Généralisés,” ESAIM, Contr. Optim. Calculu. Variat., Vol 7, No. 2, pp. S.R. Hebertt: Sliding Modes, ∆-Modulators, and Generalized Proportional Integral Control 23-41 (2002). 6. Perruquetti, W. and J. P. Barbot, Sliding mode control in Engineering, Marcel Dekker Inc., New York (2002). 7. Steele, R., Delta Modulation Systems, Pentech Press, London (1975). 8. Utkin, V.I., Sliding Modes and Their Applications in Variable Structure Systems, Mir Publishers, Moscow (1978). 9. Utkin, V.I., Sliding Modes in Control and Optimization, Springer-Verlag, Berlin (1992). 10. Utkin, V.I., J. Guldner, and J. Shi, Sliding Mode Control in Electromechanical Systems, Talylor and Francis, London (1999). H. Sira-Ramírez was born in San Cristóbal (Venezuela). He obtained the Electrical Engineer’s degree from the Universidad de Los Andes in Mérida (Venezuela) in 1970. He later obtained the M.Sc. in EE and the Electrical Engineer degree, in 1974, and the Ph.D. degree, also in EE, in 1977, all from the Massachusetts Institute of Technology (Cambridge, USA). He has been a Visiting Professor at the University of Illinois at Urbana-Champaign (USA), Purdue University (USA), The Universities of Sheffeld and Leicester (UK), the Laboratory Heudyasic of the Universitè de Compiegne (CNRS, Compiegne-France), the Laboratoire des Signaux et Systèmes (CNRS, Plateau de Moulon-France), The Institute Nationale des Sciences Appliquées (Toulousse-France), The Universidad Nacional del Sur (Argentina), The Universidad de La Plata (Argentina), The Ecole Central de Lille (Lille-France), The Centre Automatique et Systèmes of the Ecole Nationale Superieure des Mines de Paris (Fontainebleau, France), The University of Delaware (Newark-DEL, USA), The Laboratoire d’Auotmatique de Grenoble (CNRS, Grenoble-France) 475 and the Laboratoire STIX of the Ecole Polytechnique (Palaiseau-France). Dr. Sira-Ramírez worked for 28 years at the Universidad de Los Andes from which he was, Head of the Control Systems Department, Head of the Graduate Studies in Control Engineering and Vicepresident of the University. Since 1998, he became a Professor Emeritus of said University. Currently, he is a Titular Researcher in the Centro de Investigación y Estudios Avanzados del Instituto Politécnico Nacional (CINVESTAV-IPN) in México City (México). Dr. Sira-Ramírez is a Senior Member of the Institute of Electrical and Electronics Engineers (IEEE), a Distinguished Lecturer from the same Institute and a Member of the IEEE International Committee. He is also a member of the Society for Industrial and Applied Mathematics (SIAM), of the International Federation of Automatic Control (IFAC) and of the American Mathematical Society (AMS). He has published more than 250 technical articles; over a 100 of them in credited journals and the rest in international conferences. Dr. Sira-Ramírez has authored chapters in eighteen contributed books and he is a coauthor of the books, Passivity Based Control of Euler-Lagrange Systems published by Springer-Verlag, in 1998, Algebraic Methods in Flatness, Signal Processing and State Estimation, Lagares 2003 and Differentially Flat Systems, Marcel Dekker 2004. He is the recipient of several regional and national scientific awards from his native country of Venezuela. He has been a member of the Academia de Mérida (Venezuela), where he served as Second and First Vice-president for several years. He is a level III member of the Sistema Nacional de Investigadores (Conacyt, México) and a level IV member of the Programa de Promoción del Investigador (Fonacyt, Venezuela). Dr. Sira-Ramírez is interested in the theoretical and practical aspects of feedback regulation of nonlinear dynamic systems with special emphasis in Variable Structure feedback control techniques.