IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO.

4, APRIL 2012 1035

because f (P; S \ M )  U . Since Sp is subspace-representable under Strict Lyapunov Functions for

P , we have the Super-Twisting Algorithm

f (P; Sp \ Cstb ) = f (P; S \ M ) \ f (P; Cstb ) = U \ f (P; Cstb ): (8) Jaime A. Moreno, Member, IEEE, and Marisol Osorio

Recall that

U= X 2 (s)
n 2n j Aj 1 X = 0 8j = 1; . . . ; nz nw 0 n Abstract—A method to construct a family of strict Lyapunov functions,
i.e., with negative definite derivative, for the super-twisting algorithm,
without or with perturbations, is provided. This second order sliding
modes algorithm is widely used to design controllers, observers and exact
where Aj are as defined in Definition 11. Then U is a variety be- differentiators. The proposed Lyapunov functions ascertain finite time
cause it is defined by a system of affine polynomial equations in vari- convergence, provide an estimate of the convergence time, and ensure
V
ables x11 ; . . . ; xn n . Note that m+1 ( (J 0 ))  U by definition of the robustness of the finite-time or ultimate boundedness for a class of
V 0 V
(J ). Therefore, by Lemma 10, we have m+1 ( (J ))  U . Since perturbations wider than the classical ones for this algorithm. Since
V V
(Jm+1 ) is the smallest variety containing m+1 ( (J )), we have
the Lyapunov functions and their derivatives are quadratic forms, the
V (Jm+1 )  U . Therefore
operation with them is as simple as for linear time invariant systems.
Index Terms—Discontinuous systems, Lyapunov functions, robust ob-
m+1 (V(J 0 ))  V(Jm+1 )  U:
servers, second order sliding modes (SOSM).
(9)

V
Now, note that m+1 ( (J 0 )) \ f (P; Cstb ) = U \ f (P; Cstb ) by I. INTRODUCTION
Lemma 9 and (8). Therefore, by (8) and (9)
The Super-Twisting Algorithm (STA) is a well-known second order
f (P; Sp \ Cstb ) = V(Jm+1) \ f (P; Cstb): sliding modes (SOSM) algorithm introduced in [9], and it has been
widely used for control [1], [4], [7], [9], [11], [12], observation [5],
We also have that V
(Jm+1 ) \ f (P; Cstb ) is convex since and robust exact differentiation [10], [15]. The STA can be written as
U \ f (P; Cstb ) is convex by convexity of U .
x_ 1 = 0 k1 jx1 j1=2 sign (x1 ) + x2 + %1 (x; t)
x_ 2 = 0 k2 sign (x1 ) + %2 (x; t) (1)
REFERENCES
[1] Y.-C. Ho and K.-C. Chu, “Team decision theory and information struc- where xi are the scalar state variables, ki are gains to be designed, and
tures in optimal control problems – Part I,” IEEE Trans. Autom. Con- %i are the perturbation terms. Under some conditions on k1 , k2 (1) is
trol, vol. 17, no. AC-1, pp. 15–22, Feb. 1972. robust against a bounded perturbation %1 (x; t) = 0, j%2 (x; t)j  L.
[2] M. Rotkowitz and S. Lall, “A characterization of convex problems in
decentralized control,” IEEE Trans. Autom. Control, vol. 51, no. 2, pp. Since the righthand side of (1) is discontinuous, its solutions will be
1984–1996, Feb. 2006. understood in the sense of Filippov [6].
[3] R. Radner, “Team decision problems,” Annals Math. Stat., vol. 33, no. Finite time convergence and robustness for the STA has been proved
3, pp. 857–881, 1962. by geometrical methods [12], or by means of the Homogeneity prop-
[4] H. S. Witsenhausen, “A counterexample in stochastic optimum con- erties of the algorithm [11], [14]. Vw (x) = k2 jx1 j + (1=2)x22 is a
trol,” SIAM J. Control, vol. 6, no. 1, pp. 131–147, 1968.
weak Lyapunov function for the system (1) without perturbations, since
V_ (x) = 0k1 k2 jx1 j1=2 is only negative semidefinite. (Finite time)
[5] V. D. Blondel and J. N. Tsitsiklis, “A survey of computational com-
plexity results in systems and control,” Automatica, vol. 36, no. 9, pp.
1249–1274, 2000. convergence can only be asserted by using a generalization of LaSalle’s
[6] M. Rotkowitz and S. Lall, “Decentralized control information struc- invariance principle for discontinuous systems [14], but it is not pos-
tures preserved under feedback,” in Proc. IEEE Conf. Decision Con- sible to provide robustness results, or to estimate the convergence time
trol, 2002, vol. 1, pp. 569–575. from it. By a detailed analysis of the weak Lyapunov function Vw (x)
[7] D. Cox, J. Little, and D. O’Shea, Ideals, Varieties, and Algorithms: An
in [18] finite time and robust convergence for the STA is proved. A
Introduction to Computational Algebraic Geometry and Commutative
Algebra, 3/e. New York: Springer, 2007. strict (or strong) Lyapunov function, i.e., a positive definite (p.d.) func-
[8] D. Youla, H. Jabr, and J. J. Bongiorno Jr., “Modern Wiener-Hopf de- tion with negative definite (n.d.) derivative, has been proposed in [16],
sign of optimal controllers – Part II: The multivariable case,” IEEE from which an estimation of the convergence time for the perturbed
Trans. Autom. Control, vol. 21, no. AC-3, pp. 319–338, Jun. 1976. STA is drawn. However, the form of the function makes it difficult to
[9] K. Zhou, J. C. Doyle, and K. Glover, Robust and Optimal Control. operate with it for applications or further developments.
Englewood Cliffs, NJ: Prentice-Hall, 1995.
[10] S. Boyd and C. Barratt, Linear Controller Design: Limits of Perfor-
mance. Englewood Clifs, NJ: Prentice-Hall, 1991. Manuscript received February 19, 2010; revised October 26, 2010; accepted
[11] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, August 10, 2011. Date of publication February 03, 2012; date of current version
U.K.: Cambridge Univ. Press, 2004. March 28, 2012. This work was supported in part by DGAPA-UNAM, project
[12] H. S. Shin and S. Lall, “Optimal decentralized control of linear systems PAPIIT IN117610, and CONACyT, Project 51244. Recommended by Associate
via Gröbner bases and variable elimination,” in Proc. Amer. Control Editor A. Ferrara.
Conf., 2010, pp. 5608–5613. J. A. Moreno is with Coordinación de Eléctrica y Computación, Instituto de
Ingeniería, Universidad Nacional Autónoma de México, 04510 México D.F.,
Mexico (e-mail: [email protected]).
M. Osorio is with Escuela de Ingeniería, Grupo de Investigación en Au-
tomática y Diseño, Universidad Pontificia Bolivariana, Cir. 1 Num. 70-01,
Medellín, Colombia (e-mail: [email protected]).
Color versions of one or more of the figures in this technical note are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TAC.2012.2186179

0018-9286/$31.00 © 2012 IEEE

1036 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO. 4, APRIL 2012

The objective of this technical note (from which [13] is a preliminary [14] cannot be used. They require a continuously differentiable, or at
conference version) is to propose a method to construct a class of strict least a locally Lispchitz continuous Lyapunov function, but V (x) (3)
Lyapunov functions for the STA, that has the following distinguishing is continuous but not locally Lipschitz. However, as it is explained in
features compared to previous results: i) It provides a method to gen- the proof of Theorem 1 in Appendix A, it is possible to show the con-
erate a whole family of strict Lyapunov functions. ii) The family of Lya- vergence properties by means of Zubov’s theorem [17, Theorem 20.2,
punov functions is very simple to work with, since they are quadratic p. 568.], that requires only continuous Lyapunov functions.
forms, and with time derivatives that are also quadratic forms, whose Theorem 1: Consider the unperturbed system (1), with %1 = %2 =
matrices, in the unperturbed case, are related by an Algebraic Lyapunov 0, and constant gains k1 , k2 . Then the following statements are
Equation (ALE), in exactly the same manner as in the Linear Time In- equivalent:
variant case. iii) For each member of the family finite time convergence i) The origin x = 0 of (1) is finite-time stable.
of the STA can be derived, and an estimation of the Convergence Time ii) Matrix A (2) is Hurwitz, i.e., all its eigenvalues have negative
is a simple function of the Lyapunov function. iv) The class of per- real parts.
turbations, for which finite time stability can be established, is much iii) The constant gains are positive, i.e., k1 > 0 , k2 > 0.
larger than the classical ones [16], since here %1 (x; t) 6= 0 in (1) is iv) For every symmetric and positive definite matrix Q = QT > 0,
allowed. Moreover, for a larger class of perturbations ultimate bound- the ALE (5) has a unique symmetric and positive definite solution
edness (”practical” stability) can be ascertain. v) For the perturbed case P = P T > 0.
a method to design the gains and to construct robust (quadratic) Lya- In this case the function V (x), defined in (3), is a global, strict
punov functions is also provided. vi) The method is also able to provide Lyapunov function for system (1), for every P = P T > 0 that is
Chetaev’s functions to prove instability. solution of the ALE (5) with an arbitrary matrix Q = QT > 0.
The rest of the technical note is organized as follows. In Section II a Its time derivative V_ along the trajectories of the system satisfy
method to construct quadratic Lyapunov functions for the unperturbed the differential equation (4) almost everywhere. Moreover, a tra-
STA is presented, and finite time stability and instability of the STA jectory starting at x0 at time t = 0 reaches the origin in a time
can be proved by means of the Lyapunov functions. In Section III Lya- smaller than T (x0 ), given by
punov functions for the perturbed STA are constructed, from which a =2
2 1min fP g min fQg
gain design method is derived. For different classes of perturbations T (x0 ) = V 1=2 (x0 ) ;  = (6)
finite-time stability and ultimate boundedness results are obtained. An  max fP g
illustrative Example is presented in Section IV. The Appendix contains
where  is a constant depending on the gains k1 , k2 and Q.
all the proofs.

III. LYAPUNOV FUNCTIONS FOR THE

II. LYAPUNOV FUNCTIONS FOR THE STA
STA WITH PERTURBATIONS
WITHOUT PERTURBATIONS
Since the derivative of the vector  T
It is well known [7], [9], [11], [12], that the STA (1) is robustly stable
= [1 ; 2 ] =
1=2 to perturbations globally bounded by
jx1 j sign (x1 ) ; x2 is given by
1 1 %1 (x; t) = 0 ; j%2 (x; t)j  L
2 k1
1 0 (7)
_ = A ; A = 2 (2)
1 j
j 0k2 0
for any constant L > 0, when the gains k1 , k2 are appropriately se-
where j1 j = jx1 j1=2 , it is appealing to consider the quadratic function lected. In Theorem 2 this fact is established by means of a (robust)
strict Lyapunov function, where robust means that it decreases mono-
V (x) =  T P  (3) tonically along the trajectories of the system with perturbations.
Theorem 2: Suppose that the perturbation terms of system (1) are
with P a constant, symmetric and positive definite matrix, as a strict globally bounded by (7). Then for every positive value of L > 0 there
Lyapunov function candidate for the STA (1) without perturbations exist gains k1 , k2 such that the origin x = 0 is a robustly, globally
(%1 = %2 = 0). Compared to the Lyapunov function proposed in [16], finite-time stable equilibrium point. Moreover, there exists a symmetric
V (x) in (3) seems more simple and appropriate for applications, since and positive definite matrix P = [pij ] such that V (x) =  T P  is a
it is a quadratic form in  , and its derivative along the trajectories of quadratic, strict and robust Lyapunov function for the perturbed system,
the unperturbed system (1) is satisfying

V_ = x1 j01=2  T Q
0j (4) V_  0j x1 j01=2  T QR  (8)

(almost everywhere), where P and Q are related by the Algebraic Lya- almost everywhere, for some symmetric and positive definite matrix
punov Equation (ALE) QR = [qRij ]. Furthermore, a trajectory starting at x0 will converge to
the origin in a finite time smaller than T~ (x0 )
AT P + PA = 0 Q: (5)
=2
2 1min fP g min fQR g
T~ (x0 ) = V 1=2 (x0 ) ; 
~ = :
Since A is Hurwitz iff k1 > 0, k2 > 0, for every Q = QT > 0 there 
~ max fP g
exists a (unique) solution P = P T > 0 of the ALE (5), so that V (x) is
a strict Lyapunov function. So our results resembles the classical results From the proof of Theorem 2 in Appendix B one can derive the fol-
for Linear Time Invariant systems. Next Theorem 1 shows in fact, that lowing selection rules for the gains k1 , k2 and for the matrices P and
the stability of the equilibrium x = 0 of (1) is completely determined QR of the Lyapunov function.
by the stability of the matrix A: it is stable or unstable if and only if A Algorithm 1: i) Choose positive constants ( ; ), so that 0 < <
is stable or unstable. This is a remarkable fact. 1, > 1. ii) Find positive constants ( ; ) satisfying the inequality
Although the results of Theorem 1 are very simple, a careful analysis 2 2 1 2
is required, since the classical versions of Lyapunov’s theorem [2], [6], 0 > 0 (1 + ) + (1 + ) : (9)
4
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO. 4, APRIL 2012 1037

bounded for 2 small enough, i.e., 2 < min f Rg

Q = , and arbitrary
1 > 0, with final bound

max f g
P 1 
b =
min f g (1 0
P ) (min f Rg 0
Q 2  )
(14)

where  2 , 0 <  < 1, and P , Q are given in (12).

1 + p12 R
Moreover, if 1 = 0 then the origin x = 0 of (1) is an equilibrium point
that is robustly globally finite-time stable, and all trajectories converge
in a time smaller than T~ (x0 )

~ 2 1=2 min f Rg 0
Q 2
2
1 + p12
Fig. 1. Ellipse describing the boundary of the set (9) for = 1=4, = 6. T (x0 ) =
~

V (x0 ) ~=
; 
1=2
max f g
P
(15)

~ is a constant depending on the per-

where x0 is the initial state and 
turbation coefficient 2 .
This inequality represents the interior of an ellipse in the plane ( ; ),
parameterized by ( ; ), with center at the point (see Fig. 1)
IV. EXAMPLE
0 1 0 21 +1 To illustrate the previous results consider a forced Van der Pol equa-
10 0
c = 2 ; c = 2 : (10) tion [8]
1

If > 1 it is always possible to find positive values of ( ; ) satis- _1

x = "h (x1 ) + x2 _2
; x = 0 x1 + F (x; t)
fying (9), since in this case the center ( c ; c ) of the ellipse is on the
first quadrant, i.e., c > 0, c > 0. iii) Given such values of ( ; ) with h (x1 ) = x1 0x13 =3+ (t), " > 0, where  (t) is an uncertain and
and ( ; ) the gains bounded perturbation with j (t)j  1 for all t  0, and the forcing
term F (x; t) is also uncertain and bounded, i.e., jF (x; t)j  L for all
2 p ( + 1)
(x; t). If y = x1 is measured the following STA-based observer can be
k1 =
(1 0 )
L ; k2 =
(1 0 )
L (11) used to estimate x2 :
_1
^
x 0 1 j ^1 0 j sign (^1 0 )
= "h (y ) + x
^2 k x y x y ;
assure the robust, finite-time stability of the origin of the STA (1). iv)
Matrices P and QR can be selected as
_2 =
^
x 0 0 2 sign (^1 0 )
y k x y

The observation error, where i = ^i 0 i for = 1, 2, is given by

e x x i

p11 = 1; p22 =
(1 0 )
; p12 = 0 (1 0 )
= 0 p22
;
[see (1)]
2L 2
0 1 j 1 jsign( 0 0 0
L
_1 =
e k e e1 )+ e2 (t); e
_2 = k2 sign(e1 ) F (x; t):
qR11 = k1 + 2p12 (k2 + L) + 2L (1 0 k1 p12 )
p22

p12
;

In Fig. 2 simulation results are shown for the state, estimation and
qR12 = 0 21 (1 0 k1 p12 ) + (k2 + L) p22 ; qR22 = 0
p12 : (12) error trajectories. In Algorithm 1 selecting = 1=4, = 6 and
L = 3, the ellipse describing the boundary of the set (9) is represented

Note that the set of gains for a given value of L obtained from Algo- in Fig. 1. Any point in its interior provides a suitable pair ( ; ). Se-
rithm 1 is much bigger than the known recipes in [7], [9], [11], [12]. lecting the center point of the ellipse (10) ( c ; c ) = (1:044; 0:089)
The next Theorem 3 shows that if the STA is robust against per- leads to the gains k1 = 24:27, k2 = 5. For the simulation " = 1,
turbations satisfying (7), then it will also endure perturbations with F (x; t) = 2:5 sin (10t) and  (t) = 0 for t 2 [0; 5] and switches to

%1 (x; t) 6= 0 that are globally bounded by  (t) = 2:5 sin (20t) for t 2 [5; 10]. The states of the plant follow

the typical limit cycle behavior, slightly perturbed by the presence of

j 1j 
% 1 + 2 j 1j +
x
2
x2 ; j 2j 
% L (13)
the perturbations F (x; t) and  (t). The Super-Twisting Observer esti-
mation converges in finite-time to both states of the plant, despite of the
unknown perturbation F (x; t). However, when the perturbation term
for some non negative constants 1 ; 2 ; L  0. If 1 = 0, so that  (t) appears at t = 5 s. the estimation errors are no longer zero but

%1 (x; t) vanishes at the origin, and 2 is sufficiently small, then the ultimately bounded.
trajectories will still converge to zero in finite time. If 1 6= 0, so that
%1 (x; t) does not vanish at the origin, then the trajectories will (in gen-
V. CONCLUSION
eral) not converge to the origin, but they will be globally ultimately
bounded [8, Chapter 9], that is, there exists a positive constant b, and In this technical note a method to construct a family of quadratic
for every a > 0, there is T = T (a; b)  0 such that strict Lyapunov functions for the unperturbed and the perturbed STA
has been provided, and for each element of the family an estimation

k (t0 ) k a = )k  (t) k b ; 8  t t0 +T :
of the reaching time is given. The freedom in the election of the Lya-
punov function can be used for different purposes, as for example to op-
timize the reaching time estimation in the class of quadratic functions.
Theorem 3: Suppose that the perturbation terms of the system (1) Furthermore, the reaching time estimation relations are also useful to
are globally bounded by (13), and that the gains k1 , k2 satisfy the con- design the gains of the algorithm in such a way, that after a certain
ditions of the Theorem 2 for %1 (x; t) = 0, j%2 (x; t)j  L. Under prefixed time the algorithm has converged, for a compact set of initial
these conditions the trajectories of system (1) are globally ultimately conditions.
1038 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO. 4, APRIL 2012

Fig. 2. Simulation of state, estimation and error trajectories for the Van der Pol system, for Plant’s parameter "= 1 and observer gains k = 24:27, k = 5.
( ) = 2 5sin (10 )
Unknown perturbations were F x; t : ()=0
t and  t [0 5]
for t 2 ; and switches to  t ( ) = 2 5sin (20t) for t 2 [5; 10].
:

APPENDIX the problematic term to assure that V (' (t; x0 )) is AC is the term
Proof of Theorem 1: Note first that x = 0 is an equi- jx1 j1=2 sign (x1 )'1 (t; x0 ) = j'1 (t; x0 )j1=2 sign ('1 (t; x0 )), since
librium point, since (1) is a differential inclusion x_ 2 f (x), jx j1=2 sign (x ) is not Lipschitz at x = 0. However, as we will show
1 1 1
and 0 2 f (0)[2]. Equivalence of items (ii), (iii) and (iv) is a in the next paragraph, '1 (t; x0 ) is monotone when it goes through
well-known fact from Lyapunov stability of Linear Time Invariant '1 (t; x0 ) = 0. This proves that V (' (t; x0 )) is indeed an AC func-
Systems [8]. Now it will be shown that (ii) =) (i). For any tion of t, and therefore its derivative is defined almost everywhere and
Q = QT > 0, with A Hurwitz, it follows that the ALE (5) has a it is given by (4), where it is defined.
unique solution P = P T > 0. The proposed Lyapunov function In order to show that '1 (t; x0 ) is monotone when crossing
V (x) =  T P  = p11 jx1 j + 2p12 x2 jx1 j1=2 sign (x1 ) + p22 x22 , zero, suppose that '2 (; x0 ) 6= 0 at the instant t =  , when
where pij are the components of the matrix P , is an absolutely con- '1 (t; x0 ) crosses zero. Then from the differential inclusion
tinuous (AC) function of x, and it is also positive definite and radially x_ 1 2 0k1 jx1 j1=2 sign (x1 ) + x2 it follows that '1 (t; x0 ) will
unbounded, i.e. be monotonically increasing or decreasing during an interval con-
taining  . If '2 (t; x0 ) = 0 when '1 (t; x0 ) crosses zero, then
min fP g k k22  V (x)  max fP g k k22 (16) '1 (t; x0 ) will stay in zero. In both cases V (' (t; x0 )) is AC.
Since V (' (t; x0 )) is AC and V_ (4) is n.d. almost every-
where k k22 = jx1 j + x22 is the Euclidean norm of  . To ensure conver- where, it follows that V (' (t; x0 )) is a monotonically de-
gence to zero it is necessary to show that, along the trajectories ' (t; x0 ) creasing function [2, p. 207]. Moreover, from (16) and the fact
of the system (1), V (' (t; x0 )) decreases monotonically. When the 01=2
that jx1 j1=2  k k2  min fP g V 1=2 (x) it follows that
Lyapunov function is continuously differentiable, or at least locally _V (' (t; x0 ))  0V 1=2 (' (t; x0 )), where  is given by (6). Since
Lipschitz continuous, this is usually done by means of Lyapunov’s the- an AC function is the integral of its derivative [3], then
orem [2], [6], [8], [14]. However, V (x) does not satisfy these con- t
ditions on the set S = (x1 ; x2 ) 2 2 j x1 = 0 , due to the term V (' (t; x0 )) 0 V (' (0; x0 )) = V_ (' (; x0 )) d
jx1 j1=2 sign (x1 ). V (x) can still be used as a Lyapunov function, in 0
t
the same spirit as in the theorem of Zubov [17, Theorem 20.2, p. 568.],  0 V 1=2 (' (; x0 )) d :
if we are able to show that it decreases monotonically along the trajec- 0

Bihari’s inequality [17, p. 509] implies that V (' (t; x0 )) 

tories of the system, and converges to zero.
If we are able to show that V (' (t; x0 )) is AC as a function of 2
time, then V (' (t; x0 )) is a monotonically decreasing function if and V 1=2 (x0 ) 0 (=2)t , so that V (' (t; x0 )) and ' (t; x0 ) con-
only if V_ is negative definite almost everywhere [2, p. 207]. Note that verge to zero in a finite time smaller than T (x0 ) given in (6).
V (' (t; x0 )) = V  ' (t; x0 ) is the composition of V (x), an AC func- We show that (i) =) (ii), or equivalently, that  (ii) =) (i). As-
tion of x, and ' (t; x0 ), that by definition of a solution of the differen- sume that A is not Hurwitz, so we will show that x = 0 is not asymp-
tial inclusion (1) is an AC function of time. However, the composition totically stable for (1). Several cases are possible here, depending on
of two AC functions is in general not an AC function [3, p. 391]. Re- the position of the eigenvalues 1;2 = 0k1 =4 6 k12 =16 0 k2 =2 of
call that the addition and product of two AC functions is always an AC matrix A.
function, but the composition of two (scalar) AC functions h  g can be 1) 1;2 are imaginary, i.e., k1 = 0, k2 > 0. P = diag fk2 ; 1g > 0,
assured to be AC if either h is Lipschitz or g is monotone [3, p. 391]. In and Q = 0 are solutions of (5), so that V (x) is locally Lipschitz,
p.d. and V_ = 0. This means that V (' (t; x0 )) remains constant
V (' (t; x0 )) = p11 j'1 (t; x0 )j along the solutions of the system, and so it does not converge to
+2p12 '2 (t; x0 ) j'1 (t; x0 )j1=2 sign ('1 (t; x0 )) + p22 '22 (t; x0 ) zero.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO. 4, APRIL 2012 1039

2) 1;2 = 0, i.e., k1 = 0, k2 = 0. A has two eigenvalues at the are satisfied, where

origin. In this case the system is linear, and the origin is clearly
0k1 p12 ; p22
p22 (k2 + L) ;
(k2 0 L) :
2 ;
unstable.
3) 1;2 are real with different signs, i.e., k12  8k2 and k2 < 0. Then p12 (k2 + L) (20)

Pc =
0k1 1 ; Qc =
2k2 0 k12 k1
<0 As discussed in the Algorithm 1 the last inequality in (19) represents
1 0 k1 01 the interior of an ellipse in the plane ( ; ), parameterized by ( ; ).
If > 1 it is always possible to select a point ( ; ) such that all in-
are solutions of (5). An extension of Chetaev’s Instability The- equalities (19) are satisfied. In particular, this will happen for the center
orem [8, Theorem 4.3] will be used to prove instability. For this point ( c ; c )(10) of the ellipse. Given a quadruple ( ; ; ; )
consider Chetaev’s function Vc (x) =  T Pc  , that is positive in that satisfies (19), the set of gains (11) and the values of the compo-
the set U =  2 2 j 1 (22 0 k1 1 ) > 0 . U contains points nents of P (12) are directly derived from (20). The rest of the proof is
arbitrarily near to the origin, and has a boundary at f1 = 0g and analogous to the proof of Theorem 1.
f22 = k1 1 g, where Vc (x) = 0. For any initial point x0 2 Proof of Theorem 3: From Theorem 2 it is known that V (x) =
U \ Br , with Br =  2 2 j k k2  r , Vc (x0 ) > 0, and,  T P  is a Lyapunov function for the (structured) perturbed system
since V_ c (x) =  T Qc  > 0 in U , the trajectory ' (t; x0 ) has to (with %1 (x; t) = 0, j%2 (x; t)j  L), and (8) is satisfied for P and
leave U \ Br through the boundary of Br . This implies that the QR in (12). V_ along the solutions of the perturbed system (1), when it
origin is unstable. exists, is given by
4) One of the eigenvalues 1;2 is zero, i.e., k1 6= 0 and k2 = 0. It
follows that 2 (t) = 2 (0) for all t  0, so that no asymptotic V_  0 11 2  Q  0 [ %1 0 ] P 
T

stability is possible. jx1 j =

R

5) 1;2 have both positive real part, so that matrix A is Anti-Hurwitz.

In this case the ALE (5) has a positive definite solution P for every  0 11 2 min fQ gk k22 0 1 + 2 k k2  k k2
jx1 j =
R

negative definite matrix Q, and therefore V (x) > 0, V_ (x) >

0. Instability follows form the previous extension of Chetaev’s  0 k k122 (min fQ g 0 2 ) k k2 0 1 
theorem. jx1 j =
R

Proof of Theorem 2: Consider a quadratic Lyapunov function  0 (min fQ g 0 2 ) k k2 + 1 

R

candidate V (x) =  T P  (3), with P given in (12). If the perturba-

tion satisfy (7) then '1 (t; x0 ) is monotone when it crosses zero, and where  1 + p122 and the fact that jx j1=2
1  k k2 has been used.
so V (' (t; x0 )) is AC and its time derivative along the solutions of When 2 is small, i.e.
system (1), when it exists, is given by
min fQR g
2 < (21)

V_ =0jx1 j01 2  Q +
= T 2 [ 1 %1 ; jx1 j1 2 %2 ] P  : (17)
=

jx1 j1 2 2
=
it follows that for every 1 > 0 and 0 <  < 1:
Using (7) and setting  (t; x) = %2 (t; x) sign (x1 ), (17) becomes V_  0 (min fQ g 0 2 )  k k2 + (1 0 )k k2 + 1 
R

V_ = jx1 j01 2  (A + N (t; x)) P + P (A + N (t; x)) ;

= T T  0  (min1f2Q g 0 2 ) V 1 2 (x) ; 8 k k2   ;
R =

max fP g =

where A + N (t; x) = 0 (k2002k1(t; x)) 02 : where

1 1
1 

(1 0 ) (min fQ g 0 2 ) :
Since  (t; x) is bounded V (x) is p.d. and V_ is n.d. if
R

Therefore any trajectory

enters the set =
p22 > p12 2 1
; p12 < 0 ; 0 > k1 p12 + (1 0 k1 p12 )2 + 1 1 1 x2 2
j V (x)  max fP g 2


in finite time T , that

4 using Bihari’s inequality (proof of Theorem 1) can be calculated
1 1 1 + 2p12
2
(k2 0  (t; x))0(10 k1p12 )p22 (k20  (t; x)) + 1 1 1 for an initial state x0 as
1 1 1 + (k2 0  (t; x))2 p22
2
: (18)
T
21max2 fP g
=  (min
=

V 1=2 (x0 ) 0 1max

=2
fP g 
Using the fact that (k2 0 L)  (k2 0  (t; x))  (k2 + L), and fQ g 0 2 )
R

noticing that k2 > L has to be satisfied (otherwise the origin would

be unstable for the perturbation  (t; x) = L), the last inequality in and stays there for all future times, so that k (t)k2  b for all t 
(18) will be satisfied if t0 + T , with b as in (14). If 1 = 0 and 2 satisfies (21), as in the

0 > k1 p12 + 41 (1 0 k1 p12 )2 + 2p12

proof of Theorem 1, finite-time convergence, with convergence time
2
(k2 + L)0 estimated by (15), follows.

0 (1 0 k1 p12 ) p22 (k2 0 L) + (k2 + L)2 p22

2
REFERENCES
or, equivalently, that the matrix QR given in (12) is p.d., i.e., QR > 0. [1] G. Bartolini, A. Ferrara, E. Usai, and V. Utkin, “On multi-input chat-
tering-free second-order sliding mode control,” IEEE Trans. Autom.
It follows that (18) will be satisfied if the inequalities
Control, vol. 45, no. 9, pp. 1711–1719, Sep. 2000.

>1; >0 ; 0< <1; >0;

[2] A. Baccioti and L. Rosier, Liapunov Functions and Stability in Control
Theory, 2nd ed. New York: Springer-Verlag, 2005.
02 > 2
0 (1 + ) + 14 (1 + )2 (19)
[3] V. I. Bogachev, Measure Theory. Berlin, Germany: Springer-Verlag,
2007, vol. I, p. 491.
1040 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO. 4, APRIL 2012

[4] I. Boiko and L. Fridman, “Analysis of chattering in continuous sliding- Multiple Model Adaptive Mixing Control:
mode controllers,” IEEE Trans. Autom. Control, vol. 50, no. 9, pp. The Discrete-Time Case
1442–1446, Sep. 2005.
[5] J. Davila, L. Fridman, and A. Levant, “Second-order sliding-modes ob- Simone Baldi, Petros Ioannou, and Edoardo Mosca
server for mechanical systems,” IEEE Trans. Autom. Control, vol. 50,
no. 11, pp. 1785–1789, Nov. 2005.
[6] A. F. Filippov, Differential Equations with Discontinuous Right-Hand
Side. Dordrecht, The Netherlands: Kluwer, 1988. Abstract—Most of the work in multiple model adaptive control with var-
[7] L. Fridman and A. Levant, “Higher order sliding modes,” in Sliding ious forms of switching focused on continuous-time systems. The purpose
Mode Control in Engineering, J. P. Barbot and W. Perruquetti, Eds. of this technical note is to extend the results of one approach, the adaptive
New York: Marcel Dekker, 2002, pp. 53–101. mixing control (AMC), to discrete-time systems. Further, the technical note
[8] H. K. Khalil, Nonlinear Systems, Third ed. Upsaddle River, NJ: Pren- solves the tracking problem which has not been addressed in most schemes
tice–Hall, 2002, p. 750. of this class. Stability and robustness properties of the AMC scheme for
[9] A. Levant, “Sliding order and sliding accuracy in sliding mode control,” discrete-time systems are analyzed. It is shown that in the ideal case, when
Int. J. Control, vol. 58, no. 6, pp. 1247–1263, 1993. no disturbances or unmodeled dynamics are present, the tracking error
[10] A. Levant, “Robust exact differentiation via sliding modes techniques,” converges to zero. In the non ideal case, the mean-square tracking error is
Automatica, vol. 34, no. 3, pp. 379–384, 1998. of the order of magnitude of the modeling error provided the unmodeled
[11] A. Levant, “Homogeneity approach to high-order sliding mode de- dynamics satisfy a norm-bound condition. While these robustness results
are conceptually similar to those of traditional robust adaptive control, the
sign,” Automatica, no. 41, pp. 823–830, 2005.
proposed scheme does not suffer from the drawback of stabilizability of the
[12] A. Levant, “Principles of 2-sliding mode design,” Automatica, no. 43,
estimated plant and in addition performs much better in simulation studies.
pp. 576–586, 2007. Furthermore, it allows well developed results from robust control to be in-
[13] J. A. Moreno and M. Osorio, “A Lyapunov approach to second-order corporated in the design.
sliding mode controllers and observers,” in Proc. 47th IEEE Conf. De-
cision Control. CDC’08, 2008, pp. 2856–2861. Index Terms—Multiple model adaptive control, robust adaptive control.
[14] Y. Orlov, “Finite time stability and robust control synthesis of uncer-
tain switched systems,” SIAM J. Control Optim., vol. 43, no. 4, pp.
1253–1271, 2005. I. INTRODUCTION
[15] A. Pisano and E. Usai, “Globally convergent real-time differentiation
via second order sliding modes,” Int. J. Syst. Sci., vol. 38, no. 10, pp. All real systems are affected by uncertainties due to unmodeled dy-
833–844, 2007. namics, unknown parameters and process changes. A practical control
[16] A. Polyakov and A. Poznyak, “Reaching time estimation for “Super- design must be able to maintain performance and stability irrespec-
Twisting’ second order sliding mode controller via Lyapunov function tive of these uncertainties. Adaptive control was motivated to meet
designing,” IEEE Trans. Autom. Control, vol. 54, no. 8, pp. 1951–1955, the challenge of handling parametric uncertainties much larger than
Aug. 2009.
[17] A. S. Poznyak, Advanced Mathematical Tools for Automatic Control those robust control can handle, when the uncertainties are so large to
Engineers. Amsterdam, The Netherlands: Elsevier, 2008, vol. 1, De- the point that no single fixed LTI controller can guarantee satisfactory
terministic Techniques, p. 774. closed-loop behavior. This is typically achieved by estimating the un-
[18] V. Utkin, J. Guldner, and J. Shi, Sliding Mode Control in Electro-Me- known parameters on-line and using the estimates as the true ones in
chanical Systems, 2nd ed. Orlando, FL: CRC Press, Taylor & Francis, calculating the controller parameters. For the on-line calculation of the
2009.
controller parameters to be possible at all times, the estimated plant
has to satisfy the assumptions of stabilizability and detectability at
each instant of time. Such a property is not guaranteed by the tradi-
tional adaptive control schemes. In order to enforce it, a number of dif-
ferent modifications and new classes of adaptive controllers have been
proposed [1]–[3]. A class of adaptive controllers which bypasses the
stabilizability/detectability problem and can take advantage of the ro-
bust control tools for LTI systems involves a supervisory logic based
adaptive control architecture which consists of two main components:
(1) a multicontroller formed by a family of candidate controllers de-
signed off-line to meet the control objectives based on different pos-
sible sets of values of the plant parameters; and (2) a supervisor re-
sponsible of deciding which controller or combination of controllers

Manuscript received May 04, 2010; revised December 22, 2010; accepted
September 08, 2011. Date of publication September 26, 2011; date of current
version March 28, 2012. This work was supported in part by National Science
Foundation (NSF) Award CMS-0510921 and in part by the Cyprus University
of Technology. Recommended by Associate Editor A. Ferrara.
S. Baldi was with the Dipartimento Sistemi e Informatica, Università
di Firenze, Firenze 50139, Italy, and is now with the Computer Science
Department, University of Cyprus (UCY), Nicosia 1678, Cyprus (e-mail:
[email protected]).
P. Ioannou is with the Department of Electrical Engineering, University of
Southern California, Los Angeles, CA 90089 USA (e-mail: [email protected]).
E. Mosca is with Dipartimento Sistemi e Informatica, Università di Firenze,
Firenze 50139, Italy (e-mail: [email protected]).
Color versions of one or more of the figures in this technical note are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TAC.2011.2169620

0018-9286/$26.00 © 2011 IEEE