European Journal of Control (2003)9:322±360

# 2003 EUCA

The Complementarity Class of Hybrid Dynamical Systems

W.P.M.H. Heemels1 and B. Brogliato2,y
1
Department of Electrical Engineering, Eindhoven University of Technology, P.O. Box 513, 5600 MB Eindhoven, The Netherlands;
2
INRIA, ZIRST Montbonnot, 655 avenue de l'Europe, 38334 Saint-Ismier, France

This paper gives an introduction to the field of friction, and one-sided springs; dynamic versions of
dynamical complementarity systems. A summary of linear and nonlinear programming problems; and
their main applications and properties together with dynamic Walrasian economies. In many of these
connections to other hybrid model classes is provided. applications a prominent role is played by a special
Moreover, the main mathematical tools which allow combination of inequalities, which is similar to the
one to study complementarity systems are presented linear complementarity problem (LCP) [31] of math-
briefly. Many examples illustrate the developments. ematical programming. Coupling such ``com-
The available results on modeling, simulation, controll- plementarity conditions'' to differential equations
ability, observability and stabilization are presented leads to dynamical extensions of the LCP that are
and further suggestions for reading can be found in this called complementarity systems (CS) [45,92,93]. These
overview. systems have already a long history within various
application fields like unilaterally constrained
Keywords: Hybrid systems; Complementarity; Piece- mechanical systems [22,39,65,74,75,82,88]. The main
wise linear systems; Dynamics and inequalities; Well- aim of the current paper is to introduce CS to readers
posedness; Stability; Controllability; Constrained who are unaware of what they are and what physical
mechanics; Switched circuits; Convex analysis or abstract systems they are able to model, it presents
their main features and results, which have already
1. Introduction been obtained in the domain of systems and control
theory. It has to be noted that there is a considerable
In many technical and economic applications one inherent complexity in systems of differential equa-
encounters systems of differential equations and tions and inequalities, since nonsmooth trajectories
inequalities. For a quick roundup of examples, one and possibly jumps have to be taken into account.
may think of the following: motion of rigid bodies Actually, the combinations of differential or differ-
subject to unilateral constraints; switched electrical ence equations and inequalities gives rise to systems
networks; optimal control problems with inequality that switch between modes on the basis of certain
constraints in the states and/or controls; dynamical inequality constraints and that behave within each
systems with piecewise linear characteristics, such as mode as ordinary differential systems. This ``multi-
saturation functions, dead zones, relays, Coulomb modal'' way of thinking is natural in a number of
applications: in the study of Coulomb friction, one has
the transition between stick mode and slip mode; in
This work has been supported by the European project SICONOS the study of electrical networks with ideal diodes,
IST2001-37172 (http://maply.univ-lyon1.fr/siconos)
y
E-mail: [email protected]
Correspondence and offprint requests to: W.P.M.H. Heemels,
Department of Electrical Engineering, Eindhoven University of Received 21 March 2003; Accepted 17 April 2003.
Technology, P.O. Box 513, 5600 MB Eindhoven, The Netherlands. Recommended by I. Postlethwaite, J.M. Maciejowski, K. Glover and
E-mail: [email protected] P.J. Fleming.
Hybrid Dynamical Systems 323

there is the transition between the conducting and the Definition 2.1. LCP (q, M): Given an m-vector q and
blocking mode of each diode; and in the context of an m  m matrix M find an m-vector  such that
dynamic optimization, one has mode transitions when
an inactive constraint becomes active, or vice versa. A  > 0; w :ˆ q ‡ M > 0; T w ˆ 0: …1†
similar point of view may be found in the literature on
hybrid systems encompassing both continuous and The LCP(q, M) can equivalently be written as
discrete dynamics, which have recently been a popular 0 6  ? q ‡ M > 0, …2†
subject of study both for computer scientists and
for control theorists. Actually, the class of com- where the notation w ?  expresses the orthogonality
plementarity systems has connections to various other between w and  and the inequalities should be
class of hybrid dynamical models; as shown in [50] interpreted componentwise. An interesting general-
in discrete-time strong links exist to piecewise linear ization of the LCP is the so-called ``cone com-
and affine systems [96], min-max-plus-scaling systems plementarity problem'', which has been mentioned
[34] and mixed logic dynamic systems [8] and in for instance in [31, p. 31]).
continuous-time to evolutional variational inequalities
[41], differential inclusions, piecewise smooth systems Definition 2.2. LCP (q, M): Let C be a cone and
[38], projected dynamical systems [35,84] and so on. C? ˆ fz j zT v 6 0, 8v 2 Cg its polar cone1 [51,91]. Given
For the previously mentioned classes several system an m-vector q and an m  m matrix M find an
and control theoretic issues such as existence and uni- m-vector  such that
queness of solutions (well-posedness), controllability,
 2 C; w :ˆ q ‡ M 2 C? ; T …q ‡ M† ˆ 0:
observability, stability and feedback control are dif-
ficult to settle due to the hybrid behavior. CS have the …3†
great advantage of being large enough in terms of
We say that the LCPC …q, M† is solvable if such a
applications, and being small enough in order to allow
 exists. In this case, we also say that  solves …is a
one to lead deep and complete theoretical investiga-
solution of † LCPC …q, M†. The set of all solutions of
tions (which is much more difficult for larger classes of
LCPC …q, M† is denoted by SOLC …q, M†. If C ˆ Rm‡ then
hybrid systems without any additional assumptions).
LCPC … q, M† becomes the ordinary LCP(q, M)
Given the wealth of possible applications it is of
de®ned in De®nition 2.1.
interest to overcome these difficulties. Roughly speak-
ing, what makes dynamical complementarity systems The LCP and its extensions play a key role in many
so specific is that tools from complementarity pro- economic and engineering areas [37] and an extensive
blems and convex analysis are at the core of their literature [31,83] is available on this problem. After
study. This is clearly not the case for some other introducing its dynamical generalizations in the next
classes of hybrid systems. Moreover, it is important section, we will go into more details of the application
when dealing with CS to keep the right mathematical areas.
formalisms in order not to lose these particular fea-
tures. This is the message that this paper (see also [22]) 2.2. Complementarity Systems
tries to transmit.
In this paper we focus on a class of nonsmooth
2. Complementarity Modeling dynamical systems, which can on one hand be inter-
preted as a specific class of hybrid dynamical systems
Since the word complementarity is so important in the and on the other as a dynamical extension of the LCP
context of this paper and the systems under study are or cone complementarity problem mentioned above.
the dynamical extension of it, let us start by introdu- These systems are called complementarity systems (CS)
cing some basic tools of complementarity analysis. and can be formulated in a general form by [22,92]
Later on we shall see how convex analysis, monotone
multi-valued mappings, and other tools interfere with x_ ˆ f…x, t, u, †, …4a†
complementarity. y ˆ …x, u, †, …4b†

2.1. Complementarity Problems

1
In the cone complementarity problem [31, p. 31] one uses the dual
In mathematical programming a key role is played by cone, which is defined as fz j zT v > 0, 8v 2 Cg. Note that this is just a
matter of convention as LCPC … q, M† as defined here would then
a special combination of inequalities and equations coincide with the cone complementarity problem as defined in [31,
that is called the LCP, which is defined as follows. p. 31].
324 W.P.M.H. Heemels and B. Brogliato

C? 3 w ? …† 2 C, …4c†
g…w, , u, x, t† ˆ 0, …4d†
State x re-initialization rule, …4e†

where x 2 Rn is the state of the system, t is the time,

y 2 Rp is some measurable output signal available for
feedback and specified through the mapping , u 2 Rl
is a control signal to be chosen in some admissible
set U, the slack variable  2 Rm and the signal w 2 Rm
constitute a pair of complementary variables as indi-
cated in (4c), where (
) is some function. Just as in the
definition of the complementarity problem, the
symbol ? means that w and () have to be ortho- Fig. 1. A multivalued and nonsmooth feedback loop.
gonal. The cones C and C? ˆ fz j zT v 6 0, 8v 2 Cg are a
pair of polar convex cones. For a given x, t and u,
systems (LCS) [45,92,93] given by
Eqs (4c) and (4d) generalize the notion of LCPC …q, M†
in Definition 2.2 into a nonlinear variant defined by C x_ ˆ Ax ‡ B ‡ Eu, …6a†
and the functions g and . In fact, even the state re-
initialization rule may often be written in a com- w ˆ Cx ‡ D ‡ Fu, …6b†
plementarity framework, as we shall see. The necessity 0 6 w ?  > 0: …6c†
of a state re-initialization rule is obvious if one con-
siders some of the application domains of CS being 2.3. A Subclass of Hybrid Dynamical Systems
constrained mechanical systems (impacts), switched
circuits (``short circuits'' and ``sparks''), and optimal A way to consider CS is to use a hybrid point of view.
control problems in which the adjoint variable may Indeed, take for instance the system
jump. Similar as for smooth systems, to specify a P (5). The condi-
tions (5c) state that w…t†T …t† ˆ m iˆ1 wi …t†i …t† ˆ 0
particular trajectory one has to give an initial time  0 and consequently, for all times t and for all i ˆ 1, . . . , m
and an initial state x0, i.e. x( 0) ˆ x0. it holds that wi(t) ˆ 0 or i(t) ˆ 0 due to the non-
In many applications (and actually most of the negativity of w and . The set of indices I  {1, . . . , m}
examples in the paper) the set C can be taken as the for which wi(t) ˆ 0 at time t is called the mode or active
?
positive cone Rm ‡ . Note that in this case C ˆ R‡ and index set. Note that the active index set may change
hence, we obtain the description (where we drop the during the time evolution of the system. The system
time dependence, and the output equations and added may therefore switch from one ``operation mode'' to
a minus to obtain w 2 R ‡ ) another. To define the dynamics of (5) completely, one
has to specify what the effect of mode switches will be
x_ ˆ f…x, u, †, …5a† on the state variables (which is given partly by the re-
w ˆ g…x, u, †, …5b† initialization rule). The system has 2m modes. Each
mode is characterized by the active index set I 
0 6 w ?  > 0, …5c† {1, . . . , m}, which indicates that wi ˆ 0, i 2 I, and
i ˆ 0, i 2 = I. For each such mode the laws of motion are
State x re-initialization rule: …5d† given by systems of differential and algebraic equa-
tions (DAEs). Specifically, in mode I they are given by
Note that the inequalities need to be interpreted
_ ˆ f…x…t†, u…t†, …t††,
x…t† …7a†
componentwise. The system (5) can be considered as a
smooth dynamical system given by (5a)±(5b) with a w…t† ˆ g…x…t†, u…t†, …t††, …7b†
feedback loop between  and w via a nonsmooth wi …t† ˆ 0, i 2 I, …7c†
nonlinearity (a multivalued mapping) as in Fig. 1.
i …t† ˆ 0, i2
= I: …7d†
More details will be given in Sections 4.2.1 and 7.4 on
such interconnections (see in particular Fig. 16). The In this way we get into the realm of multi-modal or
input u and output y ˆ (x, y, ) are still free and can hybrid system for which the hybrid automaton model
be used for control purposes. Moreover, if f and g are is a widely accepted unifying modeling framework.
linear, we obtain so-called linear complementarity In [1] hybrid automata are described as follows. The
Hybrid Dynamical Systems 325

discrete part of the dynamics is modeled by means of a combinations of inequalities are found in all kinds of
graph whose vertices are called locations, discrete optimization problems [61], contact problems in
states or modes, and whose edges are transitions. The mechanics [59,74], resistive switched electrical circuits
continuous state takes values in a vector space Rn. For [12,64], piecewise linear (PWL) maps [36] and so on.
each mode there is a set of trajectories, which are Actually, in [37] one states that ``the concept of com-
called the activities in [1], and which represent the plementarity is synonymous with the notion of system
continuous dynamics of the system. The ``sets of equilibrium,'' which suggest that the system itself (so
activities'' may be compared to the well-known not only the equilibrium) might be described by a
``behaviors'' advocated by J. Willems and they are dynamical model in which complementarity plays a
typically described by differential or difference alge- role as well. These are exactly the complementarity
braic equations like the ones in Eq. (7). Interaction systems that we have introduced in the previous sec-
between the discrete dynamics (mode transitions) and tions, which might even be more fascinating to study.
the continuous dynamics takes place through invar- Historically, complementarity has been introduced in
iants and transition relations in which the latter is optimization by Karush, John, Kuhn and Tucker as
sometimes split in a set of guards and reset maps for early as 19392 [56,61], in mechanics by Moreau in the
each transition. Each mode has an invariant asso- sixties [74] and in electrical circuit theory by Van
ciated to it, which describes the conditions that the Bokhoven [12]. The development of methods based on
continuous state has to satisfy at this mode. In (5) complementarity, is closely linked to the develop-
these are given by the inequalities wi(t) > 0, i 2 = I and ments of convex analysis which have been led in the
(t) > 0, i 2 I. Each transition has an associated tran- second part of the twentieth century by Rockafellar
sition relation, which describes the conditions on the [91] (motivated by optimization) and Moreau [76]
continuous state under which that particular transi- (motivated by mechanics). Before going into the
tion may take place (called the guard ) and the effect details about these theoretical developments, let us
that the transition will have on the continuous state first consider several applications of CS.
(called the reset map). Invariants and transition rela-
tions play supplementary roles: whereas invariants
3.1. Electrical Networks with Diodes
describe when a transition must take place (namely
when otherwise the motion of the continuous state as
Consider a linear electrical network consisting of
described in the set of activities would lead to viola-
resistors, capacitors, inductors, gyrators, transformers
tion of the conditions given by the invariant), the
and of k ideal diodes [48]. The RLCGT components
transition relations (in particular the guards) serve as
form a multi-port network, which can under certain
``enabling conditions'' that describe when a particular
mild conditions be described by a state space repre-
transition may take place. Note that in (5) the tran-
sentation x_ ˆ Ax ‡ B, w ˆ Cx ‡ D [2] with state
sition relations consist of the re-initialization rule and
variable x representing for instance, fluxes through
a mapping that will describe which mode has to be
the inductors and charges at the capacitors and input/
selected when the invariant will be violated.
output variables  and w representing the port vari-
As we see, CS fit nicely in the description of these
ables connected to the diodes, i.e. i ˆ Vi, wi ˆ Ii or
hybrid automata (see also [22] for a discussion). On
i ˆ Ii, wi ˆ Vi, where Vi and Ii are the voltage across
one hand CS can profit from the theory in this field,
and current through the i-th diode, respectively.
but on the other can contribute to the development of
Finally, the ideal diode characteristics of the i-th diode
a general hybrid systems theory. Note that the above
are given by Vi 6 0, Ii > 0, (Vi ˆ 0 or Ii ˆ 0). By suitable
``hybrid'' point of view can be used to define trajec-
substitutions the following system description is
tories of the system, though this is not at all necessary
obtained:
(see [6,7,62,87]). However, before we go into these
details of the behavior of CS we will first indicate the _ ˆ Ax…t† ‡ B…t†,
x…t† …8a†
application domains and the relations to other w…t† ˆ Cx…t† ‡ D…t†, …8b†
(hybrid) modeling formalisms.
0 6 w…t† ? …t† > 0, …8c†

which is an LCS (without control inputs). Since Eqs (8a)

3. Application Areas and (8b) is a model for the RLCGT-multi-port

It is well known that the LCP and its various

variations and ramifications have many engineering 2
The result of W. Karush was never published except as a master
and economical applications [31,37]. These specific thesis of the university of Chichago in 1939.
326 W.P.M.H. Heemels and B. Brogliato

network, (A, B, C, D) satisfies a passivity condition as One sees that the control input u and the multiplier 
we will use later (see Section 7.3.2). In this framework enter in various places of the dynamics in each case.
one could also add pure switches and control inputs to This may not be without consequences on stabiliza-
regulate such switching systems [49]. Examples can be tion, controllability, and observability properties.
found in the area of power converters.
Let us illustrate these developments with the
following simple electrical circuits. 3.2. Constrained Mechanical Systems
Example 3.1. The circuit in Fig. 2(a) has the dynamics:
In mechanics, many fields of application concern
x_ 1 ˆ x2 , multi-body systems with unilateral contacts and fric-
R u  1 tion. To realize this we only have to look at our
x_ 2 ˆ x2 ‡ x1 , …9†
L L L LC everyday environment. For instance, home circuit
0 6 ? x2 > 0, breakers are constituted of complex kinematic chains
(10±20 bodies) with several (20±30) unilateral con-
where u denotes the voltage applied to the system, x1 is
tacts and friction. In such mechanisms, the uni-
the charge of the capacitor and x2 is current through
laterality and the friction are essential phenomena,
inductor. A second example is displayed in Fig. 2(b)
which cannot be disregarded for virtual prototyping.
and has the dynamics:
  Nuclear plant reactors have cooling bars which are
1 1 also subject to such phenomena and may induce very
x_ 1 ˆ x2 ‡  u ‡ x1 ,
R C particular (and possibly dangerous) dynamics with
 
1 1 complicated bifurcation phenomena. Some compact
x_ 2 ˆ  u ‡ x1 , discs players possess a stabilization system based on
L C
  steel balls constrained to move in a circular rig (a so-
1 1
0 6 x2  u ‡ x1 ?  > 0: called automatic balancing unit). Once again shocks
R C
between the balls and friction are the fundamental
…10† physical phenomena that make this mechanical device
A final circuit is given in Fig. 2(c) and leads to the compensate for the irregularities of the disc. Watches
following dynamics: form also nice examples of everyday devices, which
1 1 contain complex mechanics inside. Also aerospace
x_ 1 ˆ x2 ‡ x1 u, (models for planes landing and taking-off, liquid slosh
RC R
1 2  …11† phenomena in satellites) and automotive applications
x_ 2 ˆ x1 ‡ u ‡ , (clutches, engine dynamics or for virtual reality
LC L L
0 6  ? x2 > 0: applications like virtual mounting and assembly),
form a source of numerous impact and friction pro-
blems. Evidently robotics is a vast field of applica-
tions, as many robotic tasks involve contact and
intermittent motion (drilling, hammering, polishing,
and other machining tasks). Among them, bipedal
locomotion is certainly a well-known one, at least
from the academic side.
More will be said about mechanical systems in
Section 7.4 and on Moreau's work. To give one way
of modeling this type of systems let us take a con-
servative mechanical system in which q denotes
the generalized coordinates and p the generalized
momenta. The free motion dynamics can be expressed
in terms of the Hamiltonian H(q, p), which reflects
the total energy in the system. The equations are

@H
q_ ˆ …q, p†, …12a†
@p

@H
p_ ˆ …q, p† …12b†
Fig. 2. Electrical circuits. @p
Hybrid Dynamical Systems 327

with @H/@p and @H/@q denoting partial derivatives. is equivalent to LCP(q, M) with w the slack variable.
The system is subject to the geometric inequality [31±83, Section 9.3.1].
constraints given by h(q) > 0. Dynamical extensions of this result are also well-
Friction effects are not modeled here, for simplicity, known via Pontryagin's maximum principle, which
even though Coulomb's characteristic can be recast was developed in the early 1960 [89]. Before showing
into a complementarity framework (as many other this connection to CS, we will first go to dynamical
piecewise linear characteristics, in fact). To obtain a version in discrete-time that is directly connected to
complementarity formulation, we introduce (as in the classical KKT conditions.
[45,92,93]) the Lagrange multiplier  generating the
constraint forces needed to satisfy the unilateral con- Remark 3.2. Some other links between com-
straints h(q) > 0. According to the rules of classical plementarity and optimization exist as well. For
mechanics, the system can then be written as follows: instance, Moreau introduced complementarity in
@H Lagrangian systems and proved with convex analysis
q_ ˆ …q, p†, …13a†
@p tools that Gauss' principle still holds in this case
@H [74,75] (Gauss' principle of mechanics is a minimiza-
p_ ˆ …q, p† ‡ ‰rh…q†ŠT , …13b† tion problem involving the acceleration as the
@p
unknown). This was rediscovered by LoÈdstedt much
w ˆ h…q†, …13c† later [65]. Some of the material presented in
Section 5.5 wil also use the link between optimization
together with the complementarity conditions 0 6 and complementarity.
w? > 0. Note that rh(q) ˆ (@h/@q)(q). These condi-
tions express the fact that the Lagrange multiplier i is
only nonzero, if the corresponding constraint is active 3.3.1. Model Predictive Control and
(wi ˆ 0). Vice versa, if the constraint is inactive (wi > 0), Optimal Control
the corresponding multiplier i is zero. One also has to
add the impact rules to make the dynamics complete. The relationship between quadratic programming and
Another way to write the dynamics is via the linear complementarity problems has the immediate
Lagrangian formalism, which is actually the more consequence that the well-known control methodo-
useful and used formalism in mechanics and control: logy model predictive control in a linear quadratic
M…q† _ ˆ ‰rh…q†ŠT ,
q ‡ F…q, q† setting leads to closed-loop systems that can be con-
0 6 h…q† ?  > 0, …14† sidered as discrete-time complementarity systems.
Collision mapping, Indeed, in [11] one studies the control of the discrete-
time linear time-invariant system
_ contains Coriolis, centrifugal
where the vector F…q, q†
and conservative generalized forces. The collision
mapping can be chosen in various ways, as proposed x…k ‡ 1† ˆ Ax…k† ‡ Bu…k†,
…16†
by Moreau [78,81], Pfeiffer and Glocker [88], y…k† ˆ Cx…k†,
FreÂmond [39], Stronge [98], etc. Some mappings will
be discussed later.
where x(k) 2 Rn, u(k) 2 Rm, and y(k) 2 Rp are the state,
3.3. Optimization and Complementarity input, and output vector, respectively. In particular,
one is interested in the problem of tracking the
It is well-known in mathematical programming that a output reference signal r(k) 2 Rp while ful®lling the
necessary and sufficient condition for the optimality constraints
of a quadratic program (QP) with a positive definite
matrix in the cost criterion is equivalent to a set of D1 x…k† ‡ D2 u…k† ‡ D3
u…k† 6 d4 …17†
complementarity conditions by introducing so-called
slack variables and applying the Karush±John±
Kuhn±Tucker's (usually known as the KKT) theorem. at all time instants k > 0, where
u(k): ˆ
In particular, the quadratic optimization problem u(k) u(k 1) are the increments of the input.
Assume for the moment that a full measurement of
min imize
n
qT  ‡ 12 T M the state x(k) and the previously implemented control
2R
value xu(k) :ˆ u(k 1) (which might be considered as
subject to  > 0 …15† an additional state) are available at the current time k.
328 W.P.M.H. Heemels and B. Brogliato

Then, the optimization problem these type of controlled systems comes down to
Ny
! studying CS.
X X
N u 1

min 0k‡tjk Qk‡tjk ‡
u0k‡t R
uk‡t

U
tˆ1 tˆ0 3.3.2. Optimal Control Problems in Continuous-time:
subject to Pontryagin's Maximum Principle
D1 xk‡tjk ‡ D2 uk‡t ‡ D3
uk‡t 6 d4 ,
t ˆ 0, 1, . . . , Nc , Also in a continuous-time setting complementarity
xk‡t‡1jk ˆ Axk‡tjk ‡ Buk‡t , t > 0, plays a central role. Consider the class of optimal
control problems consisting of maximizing the
yk‡tjk ˆ Cxk‡tjk , t > 0,
criterion
uk‡t ˆ uk‡t 1 ‡
uk‡t , t > 1, Z T

uk‡t ˆ 0, Nu 6 t < Ny , J…x0 , u† :ˆ ‰F…x, u, t†Šdt ‡ S…x…T †, T †
0
xkjk ˆ x…k†, uk ˆ u…k 1† ‡
uk …18†
by choosing an appropriate control function u satis-
is solved with respect to the column vector fying the control constraints u…t† 2 U, state constraint
U :ˆ ‰
u0k , . . . ,
u0k‡Nu 1 Š0 2 Rs , s :ˆ mNu, at each h(x, t) > 0, the dynamics x_ ˆ f…x, u, t† for all t 2 [0, T ]
time k, where xk ‡ tjk denotes the predicted state vector and the initial condition x(0) ˆ x0.
at time k ‡ t, obtained by applying the input sequence In the survey paper [42] the application of
uk, . . . , uk‡t 1 to model (16) starting from the state Pontryagin's maximum principle [89] to such optimal

control problems results in necessary conditions for
x(k), and k‡tjk ˆ yk‡tjk r…k† is the predicted tracking
error3. In (18), we assume that Q ˆ Q0  0, R ˆ R0  0 a control input to be optimal. Introduce the
(``'' denotes matrix positive definiteness), Ny, Nu, Nc Hamiltonian H(x, u, , t) :ˆ F(x, u, t) ‡ Tf(x, u, t).
are the output, input, and constraint horizons, The optimal control uopt satisfies:
respectively, with Nu 6 Ny and Nc 6 Ny 1. uopt ˆ arg maxu H…xopt , u, t†, …22a†
The MPC control law is based on the following
idea: At time k compute the optimal solution @H
0 x_ opt ˆ …xopt , uopt , t†, xopt …0† ˆ x0 , …22b†
U …k† ˆ ‰
u0 0
k , . . . ,
uk‡Nu 1 Š to problem (18), apply
@
@H @hT
u…k† ˆ xu …k† ‡
uk …19† _ ˆ …xopt , uopt , t† …xopt , t†, …22c†
@x @x
as input to system (16), and repeat the optimization w ˆ h…xopt , t† …22d†
(18) at the next time step k ‡ 1, based on the new
with complementarity conditions holding between the
measured (or estimated) state x(k ‡ 1). Note that
multiplier  and constraint variables w. The variable 

uk ˆ I1 U  …k†, …20† is called the adjoint or costate variable. Note that
these equations resemble the Hamiltonian formula-
where I1P:ˆ [Im 0


0]. By substituting xk‡tjk ˆ tion for unilaterally constrained mechanical systems
t 1 j
At x…k† ‡ jˆ0 A Buk‡t 1 j in (18), this can be written as as in Section 3.2. The equations (22) are completed by
minimize 1 0
U HU ‡ 0 …k†FU ‡ 12 0 …k†Y…k† boundary conditions resulting in a two-point
2
U boundary problem. For instance, in case of a
subject to GU 6 W ‡ S…k†, linearR plant x_ ˆ Ax ‡ Bu, quadratic criterion
T
…21† 1=2 0 ‰x…t†T Qx…t† ‡ u…t†T u…t†Š dt with Q positive


semi-de®nite, constraints U ˆ Rm , Cx…t† > 0 and
where …k†ˆ‰x 0
…k† x0u …k† r0 …k†Š0 ,
H ˆ H  0, and H, F, 0
initial condition x(0) ˆ x0 one obtains uopt(t) ˆ BT(t)
Y, G, W, S are easily obtained from (18). and
Note that (21) is a QP depending on the current state
      
x(k), past input xu(k) ˆ u(k 1), and reference r(k). x_ A BBT x 0
As the QP can be transformed into an LCP via the ˆ ‡ 
_ Q AT  CT
KKT conditions, we can rewrite the closed-loop    
x…0† x0
system as a linear system subject to complementarity with ˆ , …23a†
relations as was outlined in [11]. Hence, the analysis of …T† 0
 
x
w ˆ ‰C 0Š , …23b†
3
If the reference is known in advance, one can replace r(k) with 
r(k ‡ t), with a consequent anticipative action of the resulting MPC
controller. Otherwise, we set r(k ‡ t) ˆ r(k) for t > 0. 06w ?  > 0, …23c†
Hybrid Dynamical Systems 329

which is a LCS. Note that we have boundary condi- Eqs. (25b) and (25c) form a LCP. It can easily be seen
tions for t ˆ 0 and t ˆ T. Interestingly, it is possible that this problem has two characteristics (just as in the
that jumps occur in the adjoint variable  (which physical system) given by:
corresponds to ``Dirac pulses'' in ). Also for these
Inactive spring: If q > 0, then we have  ˆ 0 and
jumps additional relations are available. We do not
w ˆ kq. This means that we have the ``free motion
specify all the available conditions, but only would
dynamics'' mq ˆ u.
like to illustrate that this kind of optimal control
Active spring: If q 6 0, then it holds that  ˆ kq
problems ®t in the class of CS.
and w ˆ 0. Hence, the system evolves according to
The general formulation in [42] is called an informal
mq ˆ u kq.
theorem, because the result is not rigorously estab-
lished. It might be interesting to see how the study on Hence, certain dynamical (PWL or PWA) systems can
CS can contribute to settle the difficult issues that play be recast as CS. This is in agreement with the com-
a role in this challenging field. bination of the facts that closed-loop MPC systems
For a further overview of connections between turned out to be CS (cf. Section 3.3.1) and that in [10]
optimization and complementarity, see [95]. it was shown that the closed-loop systems can also be
described by discrete-time piecewise affine systems. As
a consequence, the MPC context indicates that there
3.4. Piecewise Affine Systems
must be a strong relationship between PWA and
(linear) CS in discrete-time. In the static case the
To link the optimal control problem in the previous
connection between piecewise affine mappings and
section to the next application domain of CS being
LCP is already known for a long time [36]. Also in
piecewise affine (PWA) systems, one can see that the
circuit theory many static piecewise-linear (PL) circuit
first condition in (22) for a constrained set U ˆ Rm‡ elements play a role (next to the ideal diode char-
and a quadratic criterion gives uopt ˆ max(0, BT).
acteristics of Section 3.1), which were suitably for-
This is a PWL relationship, that is equivalent to the
mulated as complementarity models. Indeed,
complementarity description
Kevenaars en Leenaerts (see e.g. [64]) showed that all
0 6 uopt ? uopt BT  > 0: …24† the explicit piecewise linear canonical representations
proposed by Chua and Kang, GuÈzelis and GoÈknar,
In this case additional complementarity conditions and Kahlert and Chua used for PWL circuit elements
would be added to (23). Observe that we have actually are all covered by one implicit model based on the
rewritten a simple piecewise linear function as a LCP. This implicit model was developed by Van
complementarity model. Before we go to more general Bokhoven [12].
piecewise linearities, let us ®rst consider an application To give some example consider the rather arbitrary
of the max-relation in the piecewise linear mechanical piecewise affine function as given in Fig. 4, which is
system below. described by
Example 3.3. A simple mechanical example is v ˆ g ‡ r1 z ‡ …r2 r1 †max…d a1 , 0†
depicted in Fig. 3 with mass m, position q and spring ‡ …r3 r2 † max…d a2 , 0†: …26†
stiffness k. This system is described by
By using the same trick (24) as for rewriting the max-
m
q ˆ u ‡ , …25a† relation, we obtain
w ˆ kq ‡ , …25b† u ˆ g ‡ r1 z ‡ …r2 r1 †1 ‡ …r3 r2 †2 …27†

0 6  ? w > 0: …25c†

Fig. 3. Simple mechanical system. Fig. 4. Piecewise affine function.

330 W.P.M.H. Heemels and B. Brogliato

with 0 6 i d ‡ ai ? i > 0. Having these type of non- 4.1.1. Linear Complementarity (LC) Systems
smooth functions in a dynamical setting, e.g. control
systems with saturations or dead zones, gives rise to In view of Eq. (33) discrete-time LCS are given by the
CS. Another dynamical example in this context are equations
linear relay systems given by
x…k ‡ 1† ˆ Ax…k† ‡ B1 u…k† ‡ B2 …k†, …33a†

x_ ˆ Ax ‡ Bu, …28†
y…k† ˆ Cx…k† ‡ D1 u…k† ‡ D2 …k†, …33b†
y ˆ Cx ‡ Du, …29†
uˆ sign y, …30†
w…k† ˆ E1 x…k† ‡ E2 u…k† ‡ E3 …k† ‡ g4 , …33c†

where sign(a) is a set-valued mapping given by

0 6 …k† ? w…k† > 0 …33d†
(
1 if a > 0,
sign…a† ˆ …31† with w(k), (k) 2 Rm the complementarity variables.
‰ 1, 1Š if a ˆ 0,
Moreover, the variables u(k) 2 Rl, x(k) 2 Rn and
1 if a < 0.
y(k) 2 Rp denote the input, state and output, respec-
tively, at time k (this notation also holds for the other
See the left picture in Fig. 7 for its graph. hybrid system models that will be introduced).
It can be verified that this system is equivalent to In [50] one has also formulated so-called extended
the following affine CS (e denotes the column vector LCS in which not only products of two terms can be
with all entries equal to 1 and I the identity matrix): negative as in the orthogonality conditions, but any
product of an arbitrary number of terms.
x_ ˆ Ax ‡ Be 2Ba , …32a†
 a     a  4.1.2. PWA Systems
w Cx De 2D I 
ˆ ‡ ,
wb e I 0 b PWA systems [96] are described by
…32b†
x…k ‡ 1† ˆ Ai x…k† ‡ Bi u…k† ‡ fi
a a
with the complementarity conditions 0 6 w ?  > 0 y…k† ˆ Ci x…k† ‡ Di u…k† ‡ gi
and 0 6 wb ? b > 0. " #
x…k†
In Section 4 we will return to PWA and piecewise for 2
i , …34†
smooth systems together with several other hybrid u…k†
model classes, which have strong connections to CS.
where
i are convex polyhedra (i.e. given by a ®nite
number of linear inequalities) in the input/state space.
4. Relations to Other Model Classes
4.1.3. Mixed Logical Dynamical (MLD) Systems
There are many different mathematical formalisms
which are used for the analysis and control design of
In [8] a class of hybrid systems has been introduced in
hybrid systems. It is of great interest to investigate the
which logic, dynamics and constraints are integrated.
relationships between these models. Let us provide
This resulted in the description
some preliminary views on this point. Note that in the
previous section we have already indicated strong x…k ‡ 1† ˆ Ax…k† ‡ B1 u…k† ‡ B2 …k† ‡ B3 z…k†,
connections between PWA systems and linear/affine
CS. We will continue along these lines, starting with …35a†
the discrete-time case. y…k† ˆ Cx…k† ‡ D1 u…k† ‡ D2 …k† ‡ D3 z…k†,
…35b†
4.1. Discrete-time LCS
E1 x…k† ‡ E2 u…k† ‡ E3 …k† ‡ E4 z…k† 6 g5 ,
In [50] one studied the following hybrid model classes. …35c†
Hybrid Dynamical Systems 331

where x(k) ˆ [xrT(k) xbT (k)]T with xr(k) 2 Rnr and and/or boundedness of certain variables. For all the
xb(k) 2 {0, 1}nb (y(k) and u(k) have a similar structure), details on the transformations, consult [50]. To illus-
and where z(k) 2 Rrr and (k) 2 {0, 1}rb are auxiliary trate the results here, we consider an example taken
variables. Inequalities (35c) have to be interpreted from [8,50].
componentwise. Systems of the form (35) are called
MLD systems. Example 4.1.

0:8x…k† ‡ u…k† if x…k† > 0,
4.1.4. Max-Min-Plus-Scaling Systems x…k ‡ 1† ˆ
0:8x…k† ‡ u…k† if x…k† 6 0
…37†
The last discrete-time class that was described in [50]
consists of discrete-event systems, which can be with m 6 x(k) 6 M. In [8] it is shown that Eq. (37) can
modeled using the operations maximization, mini- be written as
mization, addition and scalar multiplication. In this
x…k ‡ 1† ˆ 0:8x…k† ‡ u…k† ‡ 1:6z…k†,
way they extended the well-known max-plus systems.
Expressions that are built using these operations are m…k† 6 x…k† m, x…k† 6 M…k†,
called max-min-plus-scaling (MMPS) expressions, e.g. z…k† 6 M…k†, z…k† > m…k†,
5x1 3x2 ‡ 7 ‡ max(min(2x1, 8x2), x2 3x3). z…k† 6 x…k† m…1 …k††,
Consider now the MMPS systems, which are
z…k† > x…k† M…1 …k†† …38†
described by
x…k ‡ 1† ˆ Mx …x…k†, u…k†, d…k††, …36a† and the condition (k) 2 {0, 1}.
One can verify that Eq. (37) can be rewritten as the
y…k† ˆ My …x…k†, u…k†, d…k††, …36b† MMPS model
Mc …x…k†, u…k†, d…k†† 6 c, …36c† x…k ‡ 1† ˆ 0:8x…k† ‡ 1:6 max…0, x…k†† ‡ u…k†
…39†
where Mx , My and Mc are MMPS expressions in and as the LC formulation
terms of the components of x(k), the input u(k) and
the auxiliary variables d(k), which are all real valued. x…k ‡ 1† ˆ 0:8x…k† ‡ u…k† ‡ 1:6…k†, …40a†

4.1.5. Equivalences 0 6 w…k† :ˆ x…k† ‡ …k† ? …k† > 0: …40b†

In Fig. 5 the relationships between the classes as Note that the MLD representation (38) requires
derived in [50] are indicated. An arrow going from bounds on x(k), u(k) (although such bounds can be
class A to class B means that A is a subset of B. arbitrarily large). The PWA, MMPS, and LC
Moreover, arrows with a star (?) require conditions to expressions do not require such a restriction.
establish the indicated inclusion. The conditions are
related to well-posedness (i.e. existence and unique-
ness of solution trajectories given an initial condition) 4.2. Continuous-time CS

Several interesting links exist also between continuous-

time CS and other model classes. Of course,
the dynamical extensions of various problems with
relationships to LCP like variational inequalities,
monotone multi-valued mappings and so on, give
related dynamical model classes. We will first start a
discussion on these types of (static) problems before
we will present a preliminary view on its dynamical
counterparts.

4.2.1. Monotone Multi-valued Mappings and

Convex Analysis

Fig. 5. Graphical representation of the links between the Let us notice an important fact about the
classes of hybrid systems considered in this paper. complementarity condition C? 3 w ? …† 2 C. This
332 W.P.M.H. Heemels and B. Brogliato

set of conditions means that w and () have to be Further links exist to convex analysis as the com-
perpendicular, while belonging to C? and C, respec- plementarity conditions are closely related to the
tively. This can be thought of as defining a mapping subdifferentiation of convex functions. Indeed, from
Mcomp that associates to each w a set of possible values basic convex analysis one obtains the equivalence
of (). Let us consider two couples (w1, 1) and
(w2, 2) that satisfy the complementarity conditions. 0 6  ? w > 0 () 2@ …R‡ †m …w†, …42†
Then the following holds:
where  and w are m-dimensional vectors. The func-
hw1 w2 , 1 2 i > 0, …41† tion K(x) is the indicator function of the closed
convex set K, de®ned as
where h., .i is the scalar product in Rn. Indeed a 
straightforward calculation yields hw1 w2 , 1 2 i ˆ 0 if x 2 K,
K …x† ˆ …43†
wT1 1 ‡ wT2 T2 wT1 2 wT2 1 ˆ wT1 2 wT2 1 > 0 ‡1 if x 2 = K,
since the cones C? and C are polar cones. Inequality
(41) states that Mcomp is a monotone multi-valued and given a convex function F : R ! Rn its sub-
mapping [51,91]. The multi-valued feature of Mcomp is differential @F(x) at x is de®ned as
easily deduced by noticing that for a given w, there
@F…x† ˆ f 2 Rn j hy x, F…y† F…x†i
may in general be an in®nity of  which satisfy the
complementarity conditions. For instance in the > h , y xi for all y 2 Rn g: …44†
scalar case and with C ˆ R :ˆ …1; 0Š and C? ˆ R‡ ˆ
‰0, 1†, one gets 0 6 w ?  > 0. The mapping w 7!  Subdifferentiation is a generalization of the notion of
is monotone and its graph is the so-called corner law as a derivative of a function, which applies to convex
depicted in Fig. 6. Another important notion is the functions which may be nondifferentiable (in ordinary
maximality of a monotone operator, which has to be sense), and may even take in®nite values. It can also be
understood in terms of graph inclusion. Roughly extended to set-valued mappings. An example in the
speaking, a monotone operator is maximal if its graph former case is the absolute value function abs : x 7! |x|
is ``completely ®lled in.'' For instance, the graph in that is subdifferentiable at x ˆ 0 with
Fig. 7(a) is maximal (note that this is the sign-char- (
1 if x < 0,
acteristic of Eq. (31)), but the graph in Fig. 7(b) is not @abs…x† ˆ ‰ 1, 1Š if x ˆ 0, ˆ sign x,
maximal. 1 if x > 0
and the mapping whose graph is the corner law, is also
subdifferentiable.
Computing the subdifferential of the indicator,
leads to the set-valued function (being a convex cone
for any x) given by

@ K …x† ˆ NK …x†, …45†

where

NK …x† ˆ fv j hv, w xi 6 0 for all v 2 Kg …46†

Fig. 6. The corner law (a multi-valued monotone mapping). is the normal cone [51,91] to K at x. Outside K one
de®nes Nk(x) ˆ @ k(x) :ˆ ;.
Convex functions taking infinite values like the
indicator function have been introduced by Moreau in
relation with unilateral constrained mechanics (the
reader may think of K(x) as a potential function
associated with nonpenetration of the ``position'' x
outside the domain K). See [76,91] for details. Exam-
ples in the plane are depicted in Fig. 8. In unilaterally
constrained mechanics, this means that the contact
force should be zero in the interior of K, and should
belong to the normal cone when the system is on the
Fig. 7. A maximal and a nonmaximal graphs. boundary of K.
Hybrid Dynamical Systems 333

 In [46] it has been shown that under some mild

conditions the CS (49) is equivalent to the projected
dynamical system (PDS) [35,84] given by
x_ ˆ K …x, f…t, x††, …50†
where

PK …x ‡ v† x
K …x, v† ˆ lim , …51†
!0 

and PK the projection operator that assigns to each

vector x in Rn the vector in K that is closest to x in
Fig. 8. Normal and tangent cones (cf. Eq. (94) below for a
the Euclidean norm k
k (i.e. PKx ˆ arg mink2Kkx
definition of a tangent cone). kk). These systems are used for studying the beha-
vior of oligopolistic markets, urban transportation
networks, traf®c networks, international trade, and
From the above relationships between the various agricultural and energy markets (spatial price
static problems, we can now discuss the related equilibria) [35,84].
dynamical systems for which we start from a closed  Since we saw the relationship (42) we arrive from
convex set K. Eq. (49) at the DI given by
 Evolution variational inequalities (EVI) are dyna- x_ ‡ f…t, x† 2 @ K …x†, …52†
mical systems of the form:
by using the differential rule @ K …x† ˆ
hx_ ‡ f…t, x†, v xi > 0 for all v 2 K: …47† rh…x†@ Rm‡ …h…x†† of convex analysis [91, Theorem
23.9]. Since K(x) ˆ NK(x), we see that we arrived at
EVI may represent the dynamics of various systems (48) again by following the route as outlined.
like oligopolistic markets [84], and some electrical
circuits [41].
 By de®nition of a normal cone Eq. (46), we directly Remark 4.2. The DI x 2 @jxj, _ which models a mass
obtain from (47) the differential inclusion (DI) subject to Coulomb friction, cannot be written as an
EVI as in (47). Indeed this would imply that the set
x_ ‡ f…t, x† 2 NK …x†: …48† _ is equal to the cone NK(x) for all x 2 K and for a
@jxj
certain closed convex set K. However, clearly @jxj_ is
In case f(t, x) ˆ 0 and K ˆ K(t, x), the equations turn
not a cone for @j0j ˆ [ 1, 1].
into the Sweeping Process as de®ned by Moreau
[62,77]. Let us note that having a varying set K
complicates the analysis a lot, but may be necessary Remark 4.3. Measure Differential Equations (MDE)
in some control applications (see [22, Remark 6] are dynamical systems of the form x…t† _ ˆ f…x…t†, u…t††
and [41, Remark 2]). if t 6ˆ tk , x…t‡
k † ˆ g…x…t k †† if t ˆ t k . The instants tk,
 In case K is given by {x 2 Rn j h(x) > 0} with k > 0, may be constant or may depend on x(t). There is
h: Rn7!Rk convex and real-analytic, it can be shown some resemblance between MDEs and CS whose
[46] that trajectories are composed of ``free-motion'' paths
 X   separated by isolated state jumps. For mechanical
T 
 systems this corresponds to having h(q(t)) > 0 for all
NK …x† ˆ ‰rhi …x†Š i  i 6 0
i2J…x†
t 6ˆ tk, and tk‡1 tk >  for some  > 0. Such systems
are sometimes called vibro-impact systems (especially
and J(x)={ j j hj (x) ˆ 0} the active constraint set at in the Russian literature). Although the analytical
x. Hence, this means that we can obtain the CS tools (e.g. for stability studies) that are developed for
speci®ed by MDEs may sometimes be used for ``vibro-impact'' CS,
in general this is by far not the case.
x_ ‡ f…t, x† ˆ rh…x†, …49a†
Just as for the discrete-time case, where we consider
w ˆ h…x†, …49b† the connection between LC systems and PWA sys-
tems, we would like to state a similar result for
0 6  ? w > 0: …49c† continuous-time piecewise smooth system (PSS). Note
334 W.P.M.H. Heemels and B. Brogliato

that some kind of PSS was obtained from CS by using 4.2.3. From LCS to PWL systems
a multi-modal or hybrid point of view as given in
Section 2.3. However, one has to notice that often one To show the connection between continuous-time
does not include the possibility of discontinuities in LCS and PWL systems we consider the bimodal LCS
the state trajectory in a PWA or PSS framework.
Moreover, one normally assumes that the smooth x_ ˆ Ax ‡ b ‡ eu, …55a†
submodels of a PWA system or a PSS live on a part of
T
the state space which has a non-trivial interior, while w ˆ c x ‡ d, …55b†
smooth phases of CS (5) given by the DAE (7) might 0 6 w ?  > 0, …55c†
be defined on lower-dimensional subspaces of the
total space due to the presence of the algebraic equa-
tions. For instance, the constrained motion of a robot where A 2 Rnn, b 2 Rn1, c 2 Rn1, d 2 R, and
arm that is in contact with its environment looses two 0 6ˆ e 2 Rn1.
degrees of freedom (the relative distance between We assume that either d > 0 or (d ˆ 0 and cTb > 0).
environment and arm and the corresponding velocity This assumption guarantees that for each input u 2 L2
are zero). This is also nicely illustrated by the two- and initial state x0 (with cTx0 > 0 in case d ˆ 0), there
carts example below, where the constrained motion exists a unique (absolutely) continuous state trajectory
evolves on a two-dimensional space, while the system x with x(0) ˆ x0 and a unique pair …, w† 2 L1‡12 such
has a four-dimensional state space. that Eqs. (55) hold almost everywhere. For a treat-
ment of the well-posedness of LCS with external
inputs, we refer to [25,26,49].
The complementarity conditions (55c) imply that
4.2.2. From PSS to CS either  or w is zero at (almost) each time instant (see
Section 2.3). As a consequence, this gives a system
Consider the following example with four smooth with two modes, i.e. a bimodal system. Indeed, to see
regimes valid in their own part of the state space the bimodal structure more explicitly, consider first
specified by certain inequalities. the case d > 0. Then, one can rewrite (55) as

8 
> f11 …x†, when h1 …x† > 0, h2 …x† > 0, Ax ‡ eu if cT x > 0,
>
< x_ ˆ …56†
f10 …x†, when h1 …x† > 0, h2 …x† < 0, …A bd 1 cT †x ‡ eu if cT x 6 0.
x_ ˆ
>
> f …x†, when h1 …x† < 0, h2 …x† > 0,
: 01
f00 …x†, when h1 …x† < 0, h2 …x† < 0: For the case d ˆ 0 and cTb > 0, one can rewrite (55) as
…53† 8
< Ax ‡ eu if …cT x, cT Ax ‡ cT eu†  0,
x_ ˆ P…Ax ‡ eu† if cT x ˆ 0 and
This can be rewritten as :
cT Ax ‡ cT eu 6 0,
…57†
x_ ˆ 1 2 f11 …x† ‡ 1 …1 2 †f10 ‡ …1 1 †2 f01 …x†
‡ …1 1 †…1 2 †f00 …x†, …54† where P ˆ I b(cTb) 1cT [92]. Indeed, the mode w ˆ
cTx ˆ 0 yields by differentiating that cTAx ‡ cTb ‡ cT
with i 2 12 ‡ 12sign (hi(x)). Hence, we obtain (again) eu ˆ 0. Hence, one can solve  as a function of x and u
a differential inclusion. As these type of sign- and by proper substitutions one gets the result above.
characteristics can easily be transformed into com- Note that this illustrates the remark made earlier that
plementarity conditions (see Eq. (32)), this shows how within a mode the state variable x might ``live'' on a
to transform PSS to CS. Note that the right-hand side lower-dimensional subspace (in Eq. (57) on the
of Eq. (54) is a set-valued function at the separating (n 1)-dimensional subspace {x j cTx ˆ 0}) and actu-
boundaries as one is allowed to take any convex ally if the linear system has higher relative degree
combination of the vector ®elds in the neighboring (d ˆ 0, cTb ˆ 0), then also resets of the state variable
regions. In this manner the solution concept can be are necessary as for instance in the two-carts systems
taken as Filippov's convex de®nition as presented in studied below. This is a distinctive feature of CS,
[38]. In a similar way one can also use Utkin's which is hardly studied in the framework of PWL
equivalent control de®nition to de®ne the ``sliding systems. The material in this section provides a pre-
behavior'' at the switching surface. liminary answer to the problem OP 4 in [22].
Hybrid Dynamical Systems 335

the left cart is attached to a wall by a spring. More-

over, the motion of the left cart is constrained by a
completely inelastic stop.
For simplicity, the masses of the carts and the spring
constants are set equal to 1. The stop is placed at the
(unconstrained) equilibrium position of the left cart.
By x1, x2 we denote the deviations of the left and right
cart, respectively, from their equilibrium positions and
x3, x4 are the velocities of the left and right cart,
respectively. By , we denote the reaction force exer-
ted by the stop. Furthermore, the variable w is set
equal to x1. Simple mechanical laws lead to the
dynamical relations
Fig. 9. Some links between various formalisms. x_ 1 …t† ˆ x3 …t†,
x_ 2 …t† ˆ x4 …t†,
4.3.1. In Summary x_ 3 …t† ˆ 2x1 …t† ‡ x2 …t† ‡ …t†, …58†
x_ 4 …t† ˆ x1 …t† x2 …t†,
This section clearly indicates the various relationships
that exist between CS and other description formats, w…t† ˆ x1 …t†:
which are summarized by Fig. 9.
To model the stop in this setting, the following rea-
soning applies. The variable w(t) ˆ x1(t) should be
nonnegative because it is the position of the left cart
5. Dynamics and Well-posedness
with respect to the stop. The force exerted by the stop
can act only in the positive direction implying that (t)
In Section 2.3 we have already introduced a hybrid
should be nonnegative. If the left cart is not at the stop
point of view for considering CS. This observation can
at time t (w(t) > 0), the reaction force vanishes at time
be used to define solution concepts for CS. Similarly,
t, i.e. (t) ˆ 0. Similarly, if (t) > 0, the cart must neces-
one can use the connections to differential inclusions
sarily be at the stop, i.e. w(t) ˆ 0. This is expressed by
to come up with more classical solution concepts. We
the complementarity conditions
will give some possibilities to define the notion of
solution trajectories after giving an example in the 0 6 w…t† ? …t† > 0: …59†
next section.
The system consists of two modes: the unconstrained
mode ((t) ˆ 0) and the constrained mode (w(t) ˆ 0).
5.1. Dynamics The dynamics of these modes are given by the fol-
lowing differential and algebraic equations:
To illustrate the dynamical behavior of CS, we start by
a simple example. Unconstrained

x_ 1 …t† ˆ x3 …t†,
5.1.1. Two-carts System x_ 2 …t† ˆ x4 …t†,
Let us illustrate the dynamics of CS by an example x_ 3 …t† ˆ 2x1 …t† ‡ x2 …t†,
consisting of two carts as depicted in Fig. 10 as was used x_ 4 …t† ˆ x1 …t† x2 …t†,
also in [45,92]. The carts are connected by a spring and z…t† ˆ 0:

Constrained

x_ 1 …t† ˆ x3 …t†,
x_ 2 …t† ˆ x4 …t†,
x_ 3 …t† ˆ 2x1 …t† ‡ x2 …t† ‡ …t†,
x_ 4 …t† ˆ x1 …t† x2 …t†,
Fig. 10. Two-carts system. w…t† ˆ x1 …t† ˆ 0:
336 W.P.M.H. Heemels and B. Brogliato

Remark 5.1. Note that the unconstrained mode is differently, from (0, 1, 0, 0)T smooth continuation in
directly stated as an ordinary differential equation the unconstrained mode is possible.
(ODE) in the state variable x. However, the con-
strained mode is a DAE (of high index) as the variable This last transition is a special one in the sense that,
 is not explicit. Since w ˆ x1  0 in this mode, also first, the constrained mode is active, causing the cor-
w_  0 holds, which yields x3  0. Similarly, we must responding state reset. After the reset, no smooth
have that w   0, which leads to 2x1 ‡ x2 ‡  ˆ 0. continuation is possible in the constrained mode
Hence, we have  ˆ x2. By substituting this in the resulting in a second mode change back to the
DAE of the constrained mode, we obtain x1 ˆ x3 ˆ 0 unconstrained mode.
and x2 ˆ x2 (note that the constrained mode has a A possible trajectory is given in Fig. 11 for initial
state variable (x2, x4)T of dimension 2). Hence, by state
differentiating the variable w, we can rewrite the DAE
x0 ˆ e A …0 1 10†T
as an ODE.
 …0:3202, 0:4335, 0:3716, 1:0915†T
When the system is represented by either of these
modes, the triple (, x, w) is given by the corre- on the interval [0, 3].
sponding dynamics as long as the following inequal- Note the complementarity between  (denoted by
ities in (59): ``u'' in Fig. 11) and w ˆ x1 and the discontinuity in the
derivative of x1 at time t ˆ 1.
Unconstrained
w…t† ˆ x1 …t† > 0, 5.2. Solution Concepts
Constrained For CS one may develop several solution concepts
…t† ˆ x2 …t† > 0: (see [47]), which may be similar to the notion of an
execution for hybrid automata [54], or to the solution
are satis®ed. A change of mode is triggered by viola- concept for differential inclusions as used for differ-
tion of one of these inequalities and during a transi- ential equations with discontinuous right-hand sides
tion from the unconstrained to the constrained mode [38]. A solution concept of the first type can for
we assume that the re-initialization rule is inelastic, i.e. instance be formulated as follows.
x3 :ˆ 0. The mode transitions that are possible for the
two-carts systems are described below. Definition 5.2. A set E  R‡ is called an admissible
 Unconstrained ! constrained. The inequality event times set, if it is closed and countable, and 0 2 E.
w(t) > 0 tends to be violated at a time instant t ˆ . To each admissible event times set E, we associate a
The left cart hits the stop and stays there. The collection of intervals between events E ˆ f…t1 , t2 † 
velocity of the left cart is reduced to zero instanta- R‡ j t1 , t2 2 E [ f1g and …t1 , t2 † \ E ˆ ;g.
neously at the time of impact: the kinetic energy of
the left cart is totally absorbed by the stop due to a
purely inelastic collision. A state for which this 1.5

happens is, for instance, x() ˆ (0, 1, 1, 0)T.

 Constrained ! unconstrained. The inequality 1
(t) ˆ x2(t) > 0 tends to be violated at t ˆ . The u

right cart is located at or moving to the right of its 0.5

x_1

equilibrium position, so the spring between the carts

is stretched and pulls the left cart away from the
stop. This happens, e.g. if x() ˆ (0, 0, 0, 1)T. 0

 Unconstrained ! unconstrained with re-

initialization according to constrained mode. The –0.5

inequality w(t) > 0 tends to be violated at t ˆ . As

an example, consider x() ˆ (0, 1, 1, 0)T. At the 1 x_2
time of impact, the velocity of the left cart is reduced
to zero just as in the ®rst case. Hence, a state reset
(re-initialization) to (0, 1, 0, 0)T occurs. The right
–1.5
0 0.5 1 1.5 2 2.5 3 3.5
time
cart is at the right of its equilibrium position and
pulls the left cart away from the stop. Stated Fig. 11. Solution trajectory of two-carts system.
Hybrid Dynamical Systems 337

Note that both left and right accumulations4 of 5

event times are allowed by the above definition. 4

Definition 5.3. [25]. A quadruple (E, , x, w) where E is 3

an admissible event times set, and (, x, w) : R ‡ 7!
Rm‡n‡m is said to be a hybrid solution of the auton- 2

omous CS

x_2
1
x_ ˆ f…x, †,
w ˆ h…x, †, 0

06w ? >0 –1

with initial state x0, if x(0) ˆ x0, x is continuous on R ‡ –2

and the following conditions hold for each  2 E :

–3
1. The triple (, x, w)j is real-analytic. 1 0 1 2 3
x_1
4 5 6 7

2. For all t 2 , it holds that

Fig. 12. Trajectory in the phase plane with initial
_ ˆ f…x…t†, …t††,
x…t† state (5, 5)T.
w…t† ˆ h…x…t†, …t††,
3

0 6 w…t† ? …t† > 0:

2.5
The occurrence of accumulation points in the event
x1
x
x

times set of a solution trajectory is a particular 2

instance of Zeno behavior (an in®nite number of

events in a ®nite length time-interval). Let us consider
1.5

x
some CS in which this phenomenon shows up.
2
1

Example 5.4. (Time-reversed Filippov's example). 0.5

Consider a time-reversed version of a system studied

by Filippov [38, p. 116] (mentioned also in [66]), i.e. 0

x_ 1 ˆ sign…x1 † ‡ 2sign…x2 †, …61a† –0.5

x_ 2 ˆ 2sign…x1 † sign…x2 †, …61b† 1

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
time

that can be written as a LCS as seen before (see

Eq. (32)). Solutions of this piecewise constant system Fig. 13. A sample trajectory of Filippov's example.
are spiraling towards the origin, which is an
equilibrium. Since d=dt…jx1 …t†j ‡ jx2 …t†j† ˆ 2, when allows the zero trajectory starting in the origin. For a
x(t) 6ˆ 0, solutions reach the origin in ®nite time. See trajectory in the phase plane, see Fig. 12. Note that the
Fig. 12 for a trajectory. However, solutions cannot trajectories now move away from the origin. From
arrive at the origin without going through an in®nite this it can be seen that there are (in®nitely many)
number of mode transitions (relay switches). Since solutions corresponding to initial state x0 ˆ 0 starting
these mode switches occur in a ®nite time interval, the up with a left accumulation point. See Fig. 13 for a
event times contain a right-accumulation point (i.e. time-trajectory.
the time that the solution reaches the origin) after In this respect we can use the following terminology.
which the solution stays at zero. Definition 5.6. An admissible event times set E is said
Example 5.5. (Filippov's original example). The time- to be left (right) Zeno free if it does not contain any
reverse of Eqs (61) (which is the original example in left (right) accumulation points. A hybrid solution5 is
[38]) given by said to be left (right) Zeno free if the corresponding
event times set is left (right) Zeno free. It is said to be
x_ 1 ˆ sign…x1 † 2sign…x2 †, …62a† left (right) Zeno if it is not left (right) Zeno free, and
x_ 2 ˆ 2sign…x1 † ‡ sign…x2 †, …62b† non Zeno if it is both left and right Zeno free.

5
We assume that (E, , x, w) is nonredundant, i.e. there does not
4
An element t of a set E is said to be a left (right) accumulation exist a t 2 E and t0 , t00 with t0 < t < t00 such that (, x, w) is analytic
point if for all t0 > t…t0 < t†…t, t0 † \ E……t0 , t† \ E† is not empty. on (t0 , t00 ), i.e. there are no void event times in E.
338 W.P.M.H. Heemels and B. Brogliato

Note that if a left-Zeno free solution concept is to x(‡) ˆ (0, 1, 0, 0)T. This reset can be modelled as
used for Filippov's example, there is well-posedness: the result of a (Dirac) pulse  exerted by the stop. In
existence and uniqueness of solutions given an fact,  ˆ  results in the state jump x(‡) x() ˆ
initial condition. However, if a general hybrid solution (0, 0, 1, 0)T. This motivates the use of distributional
concept (including the possibility of left-accumula- theory as a suitable mathematical framework for
tions in the event times set) is used then the system is describing physical phenomena such as collisions with
not well-posed as we have nondeterminism (non- discontinuities in the state vector as is done for
uniqueness). As one could have expected, the well- instance in [45].
posedness issue depends strongly on the notion of
Let us consider another Zeno example including
solution.
re-initialization maps.
Definition 5.3 requires that the state x of a solution
trajectory is continuous across events. For so-called Example 5.8. (Bouncing ball ). Consider a ball (height
``high-index'' systems (e.g. constrained mechanical of ball is x) with dynamics x ˆ g and constraint
systems as the two-carts example), this requirement is x > 0. To complete the model we include Newton's
too strong and, as noted before, one has to add jump _
restitution rule x…‡† ˆ ex…_ † when x( ) ˆ 0 and
rules that connect continuous states before and after _
x… † < 0 …0 < e < 1†. In case x… † ˆ x… _ † ˆ 0,
an event has taken place. The solution concept can the dynamics are equal to x ˆ 0 due to the constraint
then be generalized by interconnecting the continuous x > 0. The event times { i}i2N are related through (see
state x just before and after the event via these re- [17, p. 346])  0 ˆ 0 and
initialization rules. For linear CS a general jump rule
2ei x…0†
_
has been proposed in [45]. In this study the issue of ir- i‡1 ˆ i ‡ , i2N
regular initial states had to be tackled, i.e. the initial g
states for which there is no solution in the sense assuming that x(0) ˆ 0 and x…0† _ > 0. Hence, { i}i2N
defined so far for CS. A distributional framework was  _
has a ®nite limit equal to  ˆ 2x…0†=…g ge† < 1.
used to obtain a new solution concept for LCS [45]. In _
Since the continuous state …x…t†, x…t†† converges to
principle, this framework is based P on so-called Bohl (0, 0) when t "  ? a continuation beyond  ? can be
distributions of the form …t† ˆ liˆ0  i …i† ‡ reg …t†, _
de®ned by …x…t†, x…t†† ˆ …0, 0† for t >  ?. The physical
where  is the delta or Dirac distribution (supported at interpretation is that the ball is at rest within a ®nite
0), (i) is the i-th derivative of  and reg is a Bohl time span, but after in®nitely many bounces.
function. These distributions can equivalently be An alternative concept that foregoes explicit men-
characterized by the inverse Laplace transforms of tion of events is the following one, which is closer to
rational functions. A Bohl distribution (, x, w) is the solution concept used for differential inclusions.
called an initial solution for initial state x0, if it satisfies The notion turns out to be convenient for CS that
x_ ˆ Ax ‡ B ‡ x0 ; w ˆ Cx ‡ D as equalities of dis- satisfy a certain passivity condition.
tributions, there exists an I  {1, . . . , m} with wi ˆ 0,
i 2 I and i ˆ 0, i 2= I and finally, the Laplace trans- Definition 5.9. A triple …, x, w† 2 L2m‡n‡m [0, T ] is
^ > 0 and w…†
forms satisfy …† ^ > 0 for all sufficiently said to be an L2-solution of (60) on the interval [0, T ]
large . In case ((t), x(t), w(t)) is an ordinary function with initial condition x0 if for almost all t 2 [0, T ] the
these conditions mean that the system's equations (6) following conditions hold:
Z t
are satisfied on an interval of the form [0, ") for some
x…t† ˆ x0 ‡ f…x…s†, …s†† ds,
" > 0. In case the initial solution is not a function, the 0
impulsive part ofP (t) will result in a state jump from w…t† ˆ h…x…t†, …t††,
x0 to x ‡ :ˆ x0 ‡ iAiB i. Particularly, in [45] it is
shown that the above re-initialization procedure cor- 0 6 w…t† ? …t† > 0:
responds for linear mechanical systems with unilateral
constraints to the inelastic jump rule as proposed by 5.3. Well-posedness
Moreau [78]. Moreover, in some cases the jump of the
state variable can be made more explicit in terms of Well-posedness roughly means that solutions exist
the linear projection operator onto the consistent and are unique for any given initial condition. If
subspace of the new mode along a jump space [45]. solutions exist and are unique, a given system
description defines a mapping from initial conditions
Example 5.7. Reconsider the two-carts example. From to trajectories. In the theory of smooth dynamical
state x() ˆ (0, 1, 1, 0)T, we can enter the con- systems, it is usually taken as part of the definition of
strained mode by starting with an instantaneous reset well-posedness that this mapping is continuous with
Hybrid Dynamical Systems 339

respect to suitably chosen topologies. This may be a x1 x2

too strong requirement for hybrid systems and CS in
particular as we will see by Example 5.11.

5.4. Illustrative Examples

Fig. 14. Two-carts system with hook.
By adopting a hybrid solution concept as in
Definition 5.3 nonuniqueness of solution trajectories
can occur as shown by Filippov's example. Here some where 1, 2 denote the reaction forces exerted by the
other examples are in order. stop and hook, respectively. These equations are
completed by the complementarity conditions and the
5.4.1. Nonexistence and Nonuniqueness of inelastic impact rule. Taking
Trajectories    
1 0 0 0
Mˆ , Dˆ ,
Example 5.10. Consider the following simple LCS 0 1 0 0
    …65†
x_ ˆ x ‡ , …63a† 2 1 1 0
Kˆ , Eˆ
1 1 1 1
wˆx , …63b†
leads to a description of the form
0 6 w ?  > 0: …63c†
Note that this system is equivalent to M _ ‡ Kq…t† ˆ ET ,
q…t† ‡ Dq…t† …66a†

x, when x > 0,
x_ ˆ …64† w ˆ Eq…t†, …66b†
0, when x > 0,
which leads to nonexistence of solutions for x(0) ˆ 1 Moreau's impact rule, …66c†
and to nonuniqueness when x(0) ˆ 1.
where Moreau's inelastic impact rule is de®ned in
Other examples concerning mechanical systems can Eq. (93). This LCS displays the fact that the solutions
be found in [6], inspired by the well-known Bressan's oflinearcomplementaritysystemsdonotalwaysdepend
counter-example (see e.g. [17, Section 2.2.3]). continuously on the initial state. The discontinuous
dependence is caused by the sensitivity of solutions to
5.4.2. Discontinuous Dependence on the order in which constraints become active. Con-
Initial Conditions sider the initial states x0(") ˆ (", ", 2, 1)T, " > 0. For
" ˆ 0 the solution is a jump to (0, 0, 0, 0)T, after which
The requirement that solutions depend continuously the system stays at rest in its equilibrium position. For
on initial data, may be important for control. For " > 0, ®rst the hook becomes active, resulting in a
instance, the Krasovskii±LaSalle invariance lemma, jump to …", ", 12 , 12†T . This is followed by a regular
heavily relies on this property. It is therefore crucial to continuation in the hook-constrained mode until the
check, when dealing with hybrid systems, that solu- left cart hits the stop. The state just before the impact
tions are indeed continuous with respect to initial is …0, 0, 12 ‡ g…"†, 12 ‡ g…"††T for some continuous
conditions. This may, however, not always be the case, function g(") with g(0) ˆ 0. Re-initialization yields the
as is well-known for mechanical systems with impacts new state …0, 0, 0 12 ‡ g…"††T , which converges to
[6,17]. Let us consider a concrete example. …0, 0, 0 12†T if " # 0. Obviously, the system has a dis-
Example 5.11. [45]. Consider the two-carts system of continuity in …0, 0, 2, 1†T . One may also note that the
Section 5.1.1 extended with a hook. See Fig. 14. sequence of initial states x0(") ˆ (0, ", 2, 1), " > 0,
The complementarity description is given by leads after two re-initializations for " # 0 to the limit
state …0, 0, 12, 12†. This alternative limit corresponds to a
x_ 1 …t† ˆ x3 …t†, situation in which ®rst the stop-constrained and then
x_ 2 …t† ˆ x4 …t†, the hook-constrained mode is active.
x_ 3 …t† ˆ 2x1 …t† ‡ x2 …t† ‡ 1 …t† ‡ 2 …t†,
x_ 4 …t† ˆ x1 …t† x2 …t† 2 …t†, 5.5. Linear Passive Complementarity Systems
w1 …t† ˆ x1 …t†,
In [47] an overview is given of available well-posedness
w2 …t† ˆ x1 …t† x2 …t†, results for complementarity systems. Here we just
340 W.P.M.H. Heemels and B. Brogliato

recall a few interesting results just to give you the system (LPCS) [26,48]. Note that this case occurs
flavor of them. typically for electrical circuits containing ideal diodes
Consider an (input-free) LCS given by discussed in Section 3.1.

x_ ˆ Ax ‡ B, …67a† Theorem 5.14. [25]. Consider a LCS (A, B, C, D) with

(A, B, C, D) being passive, (A, B, C) being minimal and
w ˆ Cx ‡ D, …67b†  
B
06w ? >0 …67c† col…B, D ‡ DT † :ˆ
D ‡ DT
with (A, B, C, D) passive or dissipative with respect to of full column rank. Let QD ˆ {z j z solves LCP(0, D)}.
the supply rate Tw (in the sense of [101]). There exists a hybrid solution as de®ned in De®nition
Definition 5.12. [101]. A system (A, B, C, D) given by 5.3 of LCS(A, B, C, D) with the initial state x0 on
Eqs. (67a) and (67b) is called passive, or dissipative [0, 1) if and only if Cx0 2 Q?D :ˆ {v j vTz > 0 for all
with respect to the supply rate Tw, if there exists a z 2 QD}. Moreover, if a solution exists it is unique8
nonnegative function V : Rn ! R‡, (a storage func- and left Zeno free.
tion), such that for all t0 6 t1 and all time functions As outlined in [26,48] this theorem can be extended
(, x, w) 2 L2m‡n‡m (t0, t1) satisfying Eqs. (67a) and to include all initial states. The initial states for
(67b) the following inequality holds: which Cx0 2 = Q?D are ``inconsistent states'' as a re-
Z t1
V…x…t0 †† ‡ T …t†w…t† dt > V…x…t1 ††: initialization will happen from these points [26,48]
t0 (even with input signals) by following the discussion
The above inequality is called the dissipation inequal- on Section 5.2. It was shown in [26] that the jump rules
ity. The storage function represents a notion of are given as a cone complementarity problem, quite
``stored energy'' in the system. similar in their structure to what Moreau proposed for
mechanical systems [78], see Section 7.4 and (97).
Proposition 5.13. [101]. Consider a system (A, B, C, D)
in which (A, B, C) is a minimal6 representation. The Theorem 5.15. [26]. Consider a LCS(A, B, C, D) with
following statements are equivalent. (A, B, C, D) being passive, (A, B, C) being minimal and
 
 (A, B, C, D) is passive. B
col…B, D ‡ DT † :ˆ
 The transfer matrix G(s) :ˆ C(sI A) 1B ‡ D is D ‡ DT
positive real,7 i.e. x?[G() ‡ G?()]x > 0 for all
complex vectors x and all  2 C such that Re  > 0 of full column rank. Let QD ˆ {z j z solves LCP(0, D)}
and  is not an eigenvalue of A. and let if form with Q?D a pair of polar cones. Consider
 The linear matrix inequalities initial state x0 with Cx0 2= Q?D . Then a re-initializa-
  tion occurs to the state x ‡ :ˆ x0 ‡ B0 according to
AT K KA KB ‡ CT the following equivalent characterizations.
>0 …68a†
BT K ‡ C D ‡ DT (i) The multiplier 0 is the unique solution of the
cone complementarity problem
and

K ˆ KT > 0 …68b† Q3? Cx0 CB 2 Q?D : …69†

have a solution K. (ii) The cone QD is equal to {N j  > 0} and

Q?D ˆ fv j NT v 6 0g for some real matrix N.
Moreover, in case (A, B, C, D) is passive, all solutions The multiplier 0 ˆ N0 with 0 a solution to
to (68) are positive definite (i.e. (68b) holds with strict the LCP
inequality) and a symmetric K is a solution to (68) if
and only if V…x† ˆ 12xT Kx defines a storage function z ˆ NT Cx0 ‡ NT CK 1 CT N, …70a†
of the system (A, B, C, D)
Systems of the form (67) satisfying a passivity 0 6 z ?  > 0: …70b†
condition are called linear passive complementarity

6
This means that (A, B) is controllable and (C, A) observable.
7
The  denotes conjugate transpose of vectors of matrices. 8
It can also be shown that this solution is unique in L2.
Hybrid Dynamical Systems 341

(iii) The re-initialized state x‡ is the unique on an interval of the form [0, ") for some " > 0 for
minimizer of all initial conditions and initial well-posedness mean-
minimize 1
‰p x0 ŠT K‰ p x0 Š …71a† ing the existence and uniqueness of an initial solution
2
(see Section 5.2) given arbitrary initial condition
subject to Cp 2 Q?D : …71b† x(0) ˆ x0. Loosely, speaking this means that for each
where K is any solution to (68) and thus initial state there is the unique possibility of either a
V…x† ˆ 12 xT Kx is a storage function for re-initialization or a smooth continuation. In the
(A, B, C, D). terminology of hybrid automata [54], initial well-
(iv) The jump multiplier 0 is the unique minimizer posedness is equivalent to the LCS being nonblocking
of and deterministic.
We present now three results on initial, local and
minimize 12 …x0 ‡ Bv†T K…x0 ‡ Bv† …72a† global well-posedness of LCS.
subject to v 2 QD : …72b†
5.6.1. Initial Well-posedness
Just as outlined in Section 5.2 these jump rules are For the LCS(A, B, C, D) as in [67] the rational matrices
based on a distributional framework in which Dirac G(s) and Q(s) are defined by C(sI A) 1B ‡ D and
pulses can occur in the currents and voltages (repre- Q(s) ˆ C(SI A) 1.
senting ``sparks''). For passive LCS without inputs
there are only jumps at the initial time t ˆ 0. After that Theorem 5.16. [44]. LCS(A, B, C, D) is initially well-
there is a hybrid solution with a continuous state posed if and only if for all x0 LCP(Q()x0, G()) is
trajectory meaning that Cx…t† 2 Q?D for all t > 0. uniquely solvable for sufficiently large  2 R
Note that this gives existence and uniqueness of The strength of this theorem is that dynamical
solutions on [0, 1) for all arbitrary initial conditions, properties of an LCS are coupled to properties of
which is called global well-posedness. families of static LCPs, for which a wealth of existence
There are some nice physical interpretations for and uniqueness are available [31]. For instance, a
the two optimization problems in Theorem 5.15. sufficient condition [31] for initial well-posedness is
Statement (iii) expresses the fact that among the G() being a P-matrix9 for sufficiently large .
admissible re-initialized states p (admissible in the Clearly, initial well-posedness does not imply local
sense that smooth continuation is possible after the existence of solutions as in principle, an infinite
reset, i.e. Cp 2 Q?D ) the nearest one is chosen in the number of re-initializations (jumps) may occur on one
sense of the metric defined by any arbitrary storage time-instance without ``time progressing.'' This phe-
function corresponding to (A, B, C, D). The quadratic nomenon is sometimes called ``live-lock.'' However,
program in (iv) states that the jump multiplier 0 sufficient conditions for local well-posedness have
satisfies the complementarity conditions (i.e. 0 2 QD) been provided for LCS [45,92], as presented next.
and minimizes the internal energy (expressed by the
storage function 12 xT Kx) after the jump. Note that
x0 ‡ B0 is the re-initialized state. Under the assump- 5.6.2. Local Well-posedness
tion of x ‡ x0 2 im B, it can be shown that the two
optimization problems are actually each other's dual Consider the LCS(A, B, C, D) as in Eq. (67) with
(see e.g. [31, p 117]). Markov parameters H0 ˆ D, H1 ˆ CB, H2 ˆ CAB,
We will see later that these conditions can be used to H3 ˆ CA2B, etc. and define the leading row and
establish stability of these types of systems. column indices by

j ˆ inffi j Hij ˆ
6 0g,
5.6. Initial, Local and Global Well-posedness i …73†
j ˆ inffi j Hj ˆ 6 0g,
In Section 5.5 we derived conditions that imply the so- where j 2 {1, . . . , m} and inf Æ :ˆ 1. The leading row
called global well-posedness, i.e. the existence and coef®cient matrix M and leading column coef®cient
uniqueness of a solution on the interval [0, 1) for any matrix N are then given for ®nite leading row and
initial condition. Depending on the interval on which
solutions exist, we can now distinguish between two
other types of well-posedness; local well-posedness as 9
A matrix M 2 Rmm is called a P-matrix, if all its principal minors
the existence and uniqueness of solution trajectories satisfy det MII > 0 for all I  {1, . . . , m}.
342 W.P.M.H. Heemels and B. Brogliato

column indices by mimics the acceleration with a left-accumulation of

0 1 1 impacts. Then two solutions are possible: detachment
H1 with a left accumulation of impacts, or rest on @. If
B .. C
M :ˆ @ . A and N :ˆ …H11


Hm
m †: this force is locally analytic, and if all derivatives of h
Hm
m in (14) satisfy h…i† …0 † ˆ 0, then it is impossible to get
h…i† …0‡ † > 0 for some i while h… j† …0‡ † ˆ 0 for all j < i
Theorem 5.17. [45]. If the leading column coef®cient (which is a necessary condition for take off ). For
matrix N and the leading row coef®cient matrix M detachment to occur, one has to force h(t) not to be
are both de®ned and P-matrices, then LCS(A, B, C, D) equal to its Taylor series, which means that it cannot
has a unique local left Zeno free solution on an be analytic. Analyticity forces the solution to remain
interval of the form [0, ") for some " > 0. Moreover, stuck on @.
after at most one state re-initialization a smooth The fact that velocities are of local bounded varia-
continuation exists. tion (LBV), is quite fundamental in view of analysis
for control, and control design. In particular quad-
5.6.3. Global Well-posedness for Bimodal LCS ratic functions of LBV functions are still LBV, deri-
vatives of LBV functions are particular measures
Theorem 5.18. Consider a bimodal LCS (A, B, C, D) whose primitive is the LBV function, LBV functions
with10 C 6ˆ 0. The following statements are equivalent. possess a countable set of discontinuities, LBV func-
tions whose measure-derivative has a negative density
1. The leading Markov parameter M ˆ N is de®ned with respect to some positive measure are decreasing
(i.e. 1 ˆ 1 < 1). and positive. functions, etc.
2. The LCS (6) is locally well-posed.
3. The LCS (6) is globally well-posed.
6. Simulation
For further work on well-posedness for com-
plementarity systems, see [47] and the references Simulation and numerical problems are not the topic
therein. of this paper. However, since any control design needs
some numerical simulation for verifying the synthesis,
5.7. Mechanical Systems it is worth recalling some basic facts about the
numerical simulation of complementarity systems.
Fundamental results have been obtained by Ballard The reader may have a look at the survey paper [21]
[6,7] and Stewart [97], see also [87]. The review paper which focuses on mechanical systems, or at [17,
[97] is worth reading. The most general result in the Section 5.6] for a brief summary. The work in [53] may
frictionless case is in [6]. It is proved that if all the data also be consulted with benefit. Mainly, the reader
are piecewise analytic, then uniqueness and existence should keep in mind that at the date of writing of this
of a solution with q(
) an absolutely continuous paper, there exist almost no commercially available
function, and q…
†_ a right-continuous function of software packages that incorporates all the specific
bounded variation (RCLBV), are assured. In [97] the features of complementarity systems (robust treat-
same is shown for the so-called Painleve example. The ment of accumulations of events like impacts, LCP
_ 2 RCLBV is fundamental in view of
fact that q…
† solvers, treatment of discontinuities with respect to
stability studies, see Section 7.4. The analyticity con- initial conditions, etc.). Simulations are therefore
ditions required in [6,7] also guarantee that solutions usually performed with self-developed codes, and only
are left Zeno free (see Definition 5.6). the simplest systems are simulated. Or, one is led to
Remark 5.19. Where does the analyticity condition adopt drastic simplifications. Another commonly
comes from? Intuitively, this is easy to understand. used trick is to penalize the model, i.e. to replace the
The counter-examples for uniqueness all consider the unilateral constraints by some ``stiff '' constraints. This
system at rest on @, and with an external force that is sometimes also called ``smoothing.'' However, such
a procedure often leads to bad results, though it looks
at first glance a convenient way to solve the problem
10
Note that C ˆ 0 is a degenerate and uninteresting case, since the since the penalized dynamics take the form of a pie-
complementarity conditions do not involve the state vector x. Any cewise smooth system. Moreover, it is often desirable
quadruple (, x, w) with (t) a solution to LCP(0, D) for all t and to keep the discontinuities in the model. For instance,
satisfying Eqs (67a) and (67b) is a solution to Eq. (67). It can easily
be seen that for a scalar D, LCP(0, D) has a unique solution if and the discontinuity of the relay characteristic that
only if D 6ˆ 0. models Coulomb friction, should often be kept since it
Hybrid Dynamical Systems 343

does represent the important physical effect of ``stick'' the nonsmoothness. This means that they may be
that is not taken into account if the model is much more reliable than event-driven schemes, if
regularized. ``hard'' events occur (like events accumulations,
multiple events), because they guarantee that the
global behavior is the right one. Also in the context
6.1. Event-driven Simulation of sampled-data control these methods might be a
promising way to go.
Event-driven algorithms of numerical simulation, are
based on the simple idea of integration between
``events'' and the application of algebraic conditions 6.3. Smoothing
when an event occurs. Hence, this complies directly
with the hybrid point of view as in Section 2.3. In Smoothing is a known technique in mechanics and
complementarity dynamics as in (4) one may distin- optimization. In the framework of (4) it consists of
guish between two types of events: those with a state replacing some complementarity conditions by a reg-
re-initialization, and those involving only a mode ularized version of it that approaches it when a reg-
transition: a change in the vector field f(
) structure ularization parameter goes to infinity. For instance, in
due to a new  (a bounded function of x, u, t computed case of a sign-characteristic x 7! sign x, a suitable
from some complementarity problem). The routine regularization might look like x 7! tanh nx for n large.
consists for (5) of repetitive cycles of: Though this appears to many as a miraculous solution
which solves in one shot all the complicated problems
1. DAE simulation: by standard integration routines
that arise with unilaterality and complementarity,
compute a solution to Eq. (7) for the currently
reality is more subtle as stiff differential equations
active mode I;
have to be solved which might yield unreliable results
2. event detection: determine the event time at which
or are computationally very demanding. To say
the inequalities !i(t) > 0, i 2
= I and i(t) > 0, i 2 I
nothing on physical parameters estimation which may
get violated, say at time  and state x( );
be a hard task in practice.
3. re-initialization: calculate the re-initialized state
just after the event x(‡) (note that it is not
excluded that multiple resets are required); 7. Analysis
4. mode selection: determine the new active set I that
will be active on the next time interval starting at For the analysis of controllability, observability and
time  with state x(‡) and go back to ``DAE- stability it has to be noted that all the ingredients in (4)
simulation'' again. play a role. The state jumps are important phenom-
ena, but the complementarity conditions and their
6.2. Time-stepping interconnection with the vector field f(
) are funda-
mental as well.
The time-stepping schemes are based on a time-
discretization of the whole dynamics, i.e. of the differ-
7.1. Controllability
ential equation and of the complementarity and state
jump rules. The first schemes developed along these
The controllability issue is crucial in control systems
lines, have been proposed by Moreau [80] in relation
for understanding to what extend we can influence the
with the numerical simulation of the sweeping process
system and hence, what can be achieved by (feedback)
[62]. Other contributions have been made in [87] and
control. It is important to observe that CS can in
[3]. The advantages of such schemes is that they do not
general not live at the whole state space Rn, but only
require the calculation of the event times. Evidently
on a subset of it. This subset is specified by the
this holds provided that a convergence proof has been
admissible states as introduced next.
derived. However, time-stepping schemes lend them-
To simplify the definition we consider a system (5)
selves for such theoretical proofs, much better than
that is time-invariant (i.e. the time dependence of the
event-driven ones. Other schemes have been devel-
functions f and g is absent).
oped for electrical circuits [26] with convergence
proofs. A drawback of time-stepping schemes may be Definition 7.1. A state x0 is called admissible for
their low order which precludes very accurate results system (5), if there exists an " > 0 and a control input
during free motion phases. However the low order function u 2 [0, ") 7! Rl such that there exists a solu-
seems to be a direct and unavoidable consequence of tion (, x, w) on the interval [0, ") for initial state
344 W.P.M.H. Heemels and B. Brogliato

x(0) ˆ x0. The set of all admissible states is denoted or more compactly as x_ ˆ x ‡ max(0, x u). Note
by . that  ˆ R. It is easily seen that R ˆ R ‡ , C ˆ R ‡
and thus the system is uncontrollable. However, the
Just as for well-posedness, there is, of course, a
slightly modi®ed system given by
relationship between the nature of the considered
solutions and the controllability properties. Also the x_ ˆ x ‡  ‡ u,
set of admissible inputs should be sufficiently friendly. w ˆ x ‡ ,
Definition 7.2. (Controllability). The CS (5) is con- 0 6 w ?  > 0,
trollable, if for any pair of states x1, x2 2 , there exists or
an admissible input u defined on [0, T ], such that the 
x ‡ u, when x > 0,
corresponding state trajectory t 7! x (t, 0, x1, u) with x_ ˆ
2x ‡ u, when x 6 0
x(0) ˆ x1 satisfies x(T‡, 0, x1, u) ˆ x2.
is controllable, which can be most easily seen by
We assume for the moment that the origin lies in the feedback linearizing the system by applying the swit-
admissible set . ched feedback
Definition 7.3. Consider a CS given by (5). We say that 
a state x2 2  is reachable (from the origin), if there v, when x > 0,
uˆ
exists an admissible input u defined on [0, T], such that x ‡ v, when x 6 0,
the state trajectory t 7! x(t, 0, 0, u) with x(0) ˆ 0 satis-
which leads to the linear system x_ ˆ x ‡  that is
fies x(T ‡ , 0, 0, u) ˆ x2. The set of reachable states is
obviously controllable.
denoted by R. Similarly, we call a state x1 2  con-
Note that the feedback linearization of Example 7.5
trollable (to the origin), if there exists an admissible
can always be applied to LC systems of that form and
input u defined on [0, T ], such that the state trajectory
state dimension one, where the complementarity
t 7! (t, 0, x1, u) with x(0) ˆ x1 satisfies x(T‡,
variable w is not influenced by u, but the differential
0, x1, u) ˆ 0. The set of controllable states is denoted
equation is. However, for LC systems with state
by C.
dimension two this argument is only valid in parti-
To show the diversity of situations that arise in CS
cular cases. Fortunately, some work has been done for
we start off by some simple linear or affine com-
the case of state dimension 2 in [28]. Consider the LCS
plementarity systems with state dimension 1 to indi-
(55) with n ˆ 2, (cT, A) is observable, and d > 0, i.e. we
cate the problems that might occur.
look at the piecewise linear system (56) with n ˆ 2.
Example 7.4. Consider the example
Theorem 7.6. Consider the system (55) with n ˆ 2,
x_ ˆ u ‡ , …74a† (cT, A) is observable, and d > 0. It is controllable if
W ˆ x, …74b† and only if
06W ? >0 …74c†
f T Ae
f T …A bd 1 T
c †e > 0 …75†
of which the admissible set  is given by R ‡ . Since
 holds, where f is such that f Te ˆ 0 and f 6ˆ 0.
u, when x > 0,
x_ ˆ
0, when fx ˆ 0 and u 6 0g. Another result which applies to planar systems but
it can be easily be seen that the system is controllable. with d ˆ 0 in Eq. (55) can be found in [24]. It relies on
This example illustrates the need of the admissible the study of what happens on the boundary @K of the
states set  as it would not make sense to require for set K ˆ {x j cTx  0}. The complementarity conditions
controllability that there should be a smooth control are shown to play a major role for the controllability
input that steers the state from xa > 0 to xb < 0. properties in this setting. Some other examples are in
order now to show some tricky issues for systems with
Example 7.5. The system impact maps and affine CS.
x_ ˆ x ‡ , Example 7.7. Consider the following simple mechan-
w ˆ x ‡  ‡ u, ical system
06w ? >0 x_ 1 ˆ x2 , …76†
x_ 2 ˆ u ‡ , …77†
can be rewritten as
 W ˆ x1 , …78†
x, when x ‡ u > 0,
x_ ˆ
2x u, when x ‡ u 6 0, 06W ? >0 …79†
Hybrid Dynamical Systems 345

with reset map at x1 ˆ 0 given by x2(t‡) ˆ ex2(t ). linear, it is possible to derive a constrained equation of
A solution concept is used such that x1 is absolutely the form
continuous and discontinuous are possible in x2 A…w, zi †zf ‡ B…zi † ˆ 0,
according to the given jump rule. Note that the , …84†
C…w, zi †zf ‡ D…w; zi † > 0
admissible set  is equal to {x 2 R2 j x1 > 0}. For e ˆ 0
the system is not controllable as it is impossible to whose solvability is equivalent to having the ®nal state
reach states x 2 R2 with x1 ˆ 0 and x2 > 0 (unless you zf reachable from the initial one zi , and is w an inter-
start initially at these points). In case e 2 (0,1] this mediate input used in the analysis. The matrices, A(
),
system is controllable. This immediately shows the B(
), C(
) and D(
) have a strong structure which
influence of the rest map on these type of questions. allows one to derive some conditions for controll-
ability [23].
Example 7.8. Let " > 0 be given and formulate then the The results in [68] concern the characterization of
following affine complementarity system (ACS) the reachable subspaces for a planar juggler, and a
x ˆ u ‡ , …80† control for the stabilization of trajectories is proposed.
W ˆ x_ ", …81† Experimental results corroborate the theory.
The examples and theorems above indicate the
06W ?  > 0: …82†
diversity of the difficulties one might encounter when
_ T j x_ > "g.
The admissible set  is given by f…x, x† studying controllability for these type of hybrid sys-
 tems. This area is widely open and the answers seem to
u, when x > 0,
x_ ˆ be quite involved as is also evidenced by the proof of
0, when x ˆ 0 and u 6 0g:
Theorem 7.6 in [28]. In the case of discrete-time pie-
This system is accessible (cf. [85]), but not controllable cewise affine systems [9], or piecewise linear con-
as x_ > " prevents you from steering the state to other tinuous-time systems [63,99] some results might be
states with smaller x-values. found, but in general it is far from complete.

7.1.1. Controllability Results for Jugglers 7.2. Observability

Jugglers are a sub-class of complementarity Lagran- 7.2.1. Definition

gian mechanical systems which can be written as
follows: To define the concept of observability we slightly
adapted the observability definition in [85] which can
z_1 ˆ f1 …z1 , t, †, now be used for any rather arbitrary complementarity
z_2 ˆ f2 …z2 , u, †, system.
…83†
0 6 w ˆ h…q1 , q2 †? > 0, Definition 7.9 (Observability). Two states x1 and x2 in
 are said to be indistinguishable, if for every admis-
Collision rule
sible input function the output function t 7! y(t, t0,
with zTi ˆ …qTi , q_ Ti † 2 Rni . Examples of mechanical x1, u), t > t0 of the CS for initial state x(t0) ˆ x1 and the
jugglers are running biped robots, hoppers, controlled output function t 7! y(t, t0, x2, u) t > t0 of the CS for
structures, manipulators with dynamic passive envir- initial state x(t0) ˆ x2 are identical on their common
onment, systems with dynamic backlash or liquid domain of definition. The system is called observable,
slosh phenomena [69], tethered satellites [60], etc. The if there do not exist x1 2  and x2 2  with x1 6ˆ x2 and
analysis and control of jugglers have been investigated indistinguishable.
in [18,23,68,103,104]. It is apparent from Eq. (83) that Some approaches for observer design and analysis
the canonical form of jugglers is not controllable if of observability for piecewise linear systems are
 ˆ 0. The z1-dynamics can be controlled only through around in the literature (see e.g. [9,57] and the refer-
, either with collisions, or during periods of motion ences therein). For Lur'e type of systems (see e.g. [58])
where h(q1, q2) ˆ 0. However since  is signed and is observer design has been considered in [4]. Here we
not the available input signal, controllability and present some results for simple mechanical systems.
control problems are harder to solve. The reachable
subspaces may be de®ned in a natural manner, con- 7.2.2. Simple Mechanical Systems
sidering the paths that consist of the values of the
positions and velocities, at impact times [23]. Then Pioneering works have been published in [72,73]. The
when both the vector ®elds f1(
) and f2(
) in Eq. (83) are aim is not to derive conditions on observability, but to
346 W.P.M.H. Heemels and B. Brogliato

design asymptotically stable observers. A two-degree- common or multiple Lyapunov functions [14] and in
of-freedom system is considered in [73] (this is an the field of piecewise linear systems one has approa-
impacting pair that may model dynamic backlash [69], ches based on quadratic and piecewise quadratic
and whose dynamics fits within jugglers dynamics as Lyapunov functions [55]. However, in these approa-
in Eq. (83)). The measured output is assumed to be ches situations including state re-initializations or the
y(t) ˆ q2(t). Similarly as for controllability and stabi- fact that the dynamics associated with some modes
lization, observing the state z1 can be achieved only live on lower dimensional subspaces are hardly stu-
through the impacts. We will not elaborate on the died. Here we will present some preliminary work in
observer construction, but let us rather focus on a this direction within the context of complementarity
quite interesting comment in [73]. The dynamics is systems.
given by

m1 q1 …t† ˆ 1 2 , 7.3.1. Bimodal Planar Linear Complementarity

m2 q2 …t† ˆ 2 1 ‡ u…t†, Systems
y…t† ˆ q2 …t†, …85†
In this section, we will start by presenting some first
0 6 w ˆ h…q1 , q2 †? > 0, results dealing with linear complementarity systems
_ ‡
q…t _ k †,
k † ˆ E q…t (without external input u) of the form

where x_ ˆ Ax ‡ b, …87a†

     
q2 q1 1 q1
h…q1 , q2 † ˆ , ˆ ; qˆ w ˆ cT x ‡ d, …87b†
q2 q1 1 2 q2

and tk generically denotes the impact times. The res- 0 6  ? w > 0, …87c†
titution matrix E is function of the restitution coef®-
cient e 2 [0, 1] and of the masses. Consider that u(t)  1, where A 2 Rnn, b 2 Rn1, c 2 Rn1, and d 2 R. As
‡ usual, the system (87) is said to be asymptotically
‡
p data q2 …0† ˆ q1 …0† ˆ ‡1, 0 < q_ 1 …0 †
with initial
stable (with equilibrium 0) if all possible state trajec-
q_ 2 …0 † < 2. Then the following holds
tories x satisfy limt ! 1x(t) ˆ 0. A solution (, x, w) of
z…t‡ ‡
k‡1 † ˆ A1 …A2 z…tk † ‡ B2 † ‡ B1 ,
the system is called periodic if all three functions are
…86† periodic.
y…t‡ ‡
k † ˆ Cz…tk †,
Remark 7.10. Normally, one also includes Lyapunov
which can be thought of as a Poincare map stability in the definition of asymptotic stability. Due
with Poincare section ‡ ˆ f…q, q† _ j h…q1 , q2 † ˆ 0; to the structure of the system, we get Lyapunov sta-
rh…q1 , q2 †T q_ > 0g (this is called an impact Poincare bility for free in case we have asymptotic stability as
map). The matrices A1 and B1 take into account the defined above. Moreover, in that case we even have
impact conditions, while the matrices A2 and B2 come global exponential stability and asymptotic Lyapunov
from the continuous part of the dynamics. The pair stability (see, e.g. [58] for the exact definitions).
(C, A2) is not observable, but the pair (C, A1A2) is
observable. This nicely indicates that impacts may Note that (87) is replaced by
render the system observable from the output q2(t). 
Ax if cT x > 0,
_x ˆ …88†
…A bd c †x if cT x 6 0
1 T

7.3. Stability
if d > 0 and by11
Stability issues are of utmost importance for control 
Ax if …cT x, cT Ax†  0,
applications. It is in fact well-known that a com- x_ ˆ …89†
PAx if cT x ˆ 0 and cT Ax 6 0,
plementarity system can be stable (resp. unstable)
while the corresponding unconstrained system is where P ˆ I b(cTb) 1cT in case d ˆ 0 and cb > 0.
unstable (resp. stable), see e.g. [84, Example 3.2, and Compare this to Eqs (56) and (57)
41]. Hence, as is well-known also for piecewise smooth
systems [14], stability of the submodels (7) does in
general not say anything about stability of the overall 11
The notation  denotes lexicographic ordering, i.e. for this case
system. Some studies can be based on the use of either CTx > 0 or CTx ˆ 0 and CTAx > 0.
Hybrid Dynamical Systems 347

When the state space dimension (i.e. n) is 2, one can

derive necessary and sufficient conditions as in the G(s)
_
following theorem.
Theorem 7.11. [28]. Consider the LCS (87) with n ˆ 2
and (cT, A) is an observable pair. The following y
statements hold.
1. Suppose that d > 0. The LCS (87) is asymptotically
stable if and only if
(a) neither A nor A bd 1cT has a real non- yL
negative eigenvalue, and
(b) if both A and A bd 1cT have nonreal eigen-
values then 1/!1 ‡ 2/!2 < 0 where 1  i!1
(!1 > 0) are the eigenvalues of A and 2  i!2
(!2 > 0) are the eigenvalues of A bd 1cT. Fig. 15. Absolute stability with monotone multi-valued
2. Suppose that d > 0. The LCS (87) has a non- mappings.
constant periodic solution if and only if both A and
A bd 1cT have nonreal eigenvalues, and 1/!1 ‡ Section 4.2.1). As an example let us consider the linear
2/!2 ˆ 0 where 1i!1 are the eigenvalues of A passive complementarity system as in (67) with D ˆ 0
and 2i!2 are the eigenvalues of A bd 1cT. x_ ˆ Ax ‡ B,
Moreover, if there is one periodic solution, then
all other solutions are also periodic and /!1 ‡ /!2 z ˆ Cx, …90†
is the period of any solution (except the zero 0 6 z ?  > 0:
trajectory). In this case the feedback loop contains the so-called
3. Suppose that d ˆ 0. The LCS (87) is asymptotically corner law which is the graph of the operator z 7! .
stable if and only if A has no real nonnegative Basic convex analysis tells us that 0 6 z? > 0 is
eigenvalue and [I b(cTb) 1cT] A has a real nega- equivalent to  2 @ R ‡ (z), and the corner law is
tive eigenvalue (note that one eigenvalue is already precisely the graph of this subdifferential. Of course,
zero). any other operator can be used, provided it is mono-
In [28] one also presented some necessary conditions tone (maximality is related to well-posedness rather
for stability and controllability for higher order (n > 2) than to stability).
bimodal linear complementarity systems. In [41] the Remark 7.12. The work in [66] concentrated on the
Lyapunov stability of a class of evolution variational well-posedness issues of relay feedback (i.e.
inequalities as in (47) is studied and several stability yL 2 @jyj ˆ sign y in Fig. 15. The class of transfer
criteria are given. It is shown that adding constraints functions G(s) considered in [66] is larger than positive
on the ``free'' system may drastically modify its stability. real systems only (note that there one uses a left Zeno
free solution concept instead of a notion of solution
7.3.2. Linear Passive Complementarity Systems based on differential inclusions like in [33,38] allowing
and the Passivity Theorem left-accumulation points; the influence of this choice is
discussed in [90]).
It is known that passivity/dissipativity is a powerful Let us for the moment fix our attention on the
tool for stability analysis. It turns out that this is also dynamics that correspond to Fig. 15, and which can be
the case for nonsmooth complementarity systems. In rewritten as
this section we are going to deal with systems which a:e:
x_ ˆ Ax ByL ,
possess a structure as in Fig. 15. This class of systems
(a linear transfer function in negative feedback con- y ˆ Cx, …91†
nection with a multi-valued static nonlinearity) has yL 2 @'…y†,
been studied in [20,26,66]. Note also that we have
where y, yL 2 Rm, x 2 Rn and a.e. means almost
studied the well-posedness of linear passive com-
everywhere in the Lebesgue measure. Let us make the
plementarity systems in Section 5.5, which also fit in
following assumptions:
this context.
G(s) is a positive real transfer function, and (a) G(s) ˆ C(sI A) 1B, with (A, B, C) a minimal
the operator y 7! yL is maximal monotone [91] (cf. representation, is a strictly positive real (SPR)
348 W.P.M.H. Heemels and B. Brogliato

transfer matrix meaning that (68) leads to the ATK ‡ KA < 0 holds with strict inequality in (68),
existence of positive de®nite matrices P ˆ PT and then x ˆ 0 is the only equilibrium point, which is
Q ˆ QT such that PA ‡ ATP ˆ Q and BTP ˆ C asymptotically stable. The jumps as in Theorem 5.15
[67]. Moreover, CA 1B ‡ BTA TCT is negative satisfy V(x ‡ ) 6 V(x0) with V(x) ˆ xTKx an arbitrary
de®nite. storage function obtained by (68).
(b) ' : Rm ! R [ {‡1} is convex lower semicontin-
Note that the results in this section extend the
uous, so that @' is a maximal monotone multi-
passivity theorems that are formulated for the abso-
valued mapping (see e.g. [15, Example 2.3.4]).
lutely stability problems for Lur'e systems [58]. More
Then the following is true: general versions of the results presented here in the
direction of inclusion of pure switches and external
Lemma 7.13 [20]. Let assumptions (a) and (b) hold. If inputs (voltage and current sources) have been
Cx(0) 2 dom @', then the system in (91) has a unique obtained in [49].
absolutely continuous solution on [0, ‡ 1). Assume
also that the graph of @' contains (0, 0). Then: (i) x ˆ 0 7.4. Stabilization of Complementarity Lagrangian
is the unique solution of the generalized equilibrium Systems
equation Ax 2 B@'(Cx). (ii) The unique fixed point
x ˆ 0 of the system in (91) is exponentially stable. The so-called Lagrange±Dirichlet theorem in
mechanics, states that a mechanical system as in
In the particular case of the system in (90), the proof Eq. (99) whose potential energy is strictly convex, has
can also be found in [26]. In [26] one does not allow an equilibrium point that is Lyapunov stable. This
only continuous solutions, but also allows the possi- result has been used in control by suitably modifying
bility of state re-initializations. The corresponding the potential energy (shaping it) so that a desired
jump rules are derived in Section 5.5 and were based equilibrium is stabilized (see e.g. [67]). The question is:
on a distributional framework. Indeed it is apparent how does this extend to dynamics as in (14)? First of
from Lemma 7.13 that initial states which do not all let us add to (14) an impact law defined as [81]
satisfy the indicated constraint, have to jump instan-
taneously to some consistent value. Moreover, [26] _ ‡
q…tk† ˆ _ k†
eq…t
also included the case (67) with D 6ˆ 0 (and even allows _ k †, V…q…tk ††Š,
‡ …1 ‡ e†proxM…q…tk †† ‰q…t
external inputs in the study of well-posedness).
…93†
Lemma 7.14. A state x is an equilibrium point of (67),
if and only if there exist  and w
 2 Rm satisfying where e 2 [0, 1] is a restitution coef®cient. The set
 V(q(tk)) is a tangent cone and it is de®ned as [81]
0 ˆ Ax ‡ B, …92a†
 ˆ C
w x ‡ D,  …92b† V…x† ˆ fv 2 Rn j vT rhi …x† > 0, 8i 2 J…x†g …94†
06w  ?  > 0: …92c† with J(x) ˆ {i 2 {1, . . . , m} j hi(x) 6 0}, see Fig. 8 for
Moreover, if A is invertible, then we obtain the examples in the plane. The ``prox'' in Eq. (93) means
homogeneous LCP that characterizes all equilibria that the vector …q…t_ ‡ _ k ††=…1 ‡ e† has to be
k † ‡ eq…t
chosen as the closest vector to q…t _ k †, in V(q(tk)). The
x ˆ A 1 B and   > 0:
0 6 ‰ CA 1 B ‡ DŠ? distance is understood in the kinetic metric, i.e. it is
de®ned from the scalar product xTM(q(tk))y for two
Note that these equations comply with the gen-
vectors x and y. The link with quadratic programming
eralized equilibrium equation Ax 2 B@'(Cx) in
is easily done.
Lemma 7.13. From Lemma 7.14 it follows that x ˆ 0
It is necessary at this stage, to recall that the solu-
is an equilibrium.
tions of the dynamical system in (14)±(93) have right-
Theorem 7.15. [26]. Consider an LCS given by (67) continuous velocities of local bounded variation
such that (A, B, C, D) is passive, (A, B, C ) is minimal (RCLBV) [7]. This has important consequences, as
and derivatives of RCLBV functions may be complex
  objects, but are assured to be measures. In other
B
col…B, D ‡ DT † : ˆ words, the acceleration q is a measure, and quadratic
D ‡ DT
functions of q_ are also RCLBV. Thus their derivative
has full column rank. This LCS has only Lyapunov is also a measure. Since measures are signed (a Dirac
stable equilibrium points x. Moreover, if measure is signed ± it may be positive or negative ±), it
Hybrid Dynamical Systems 349

is meaningful to assign a sign to the derivative of a Lagrangian

RCLBV function and one knows that its primitive is dynamics
a function which is increasing or decreasing according –
to its measure-derivative sign. Consequently, having
q_ 2 RCLBV places us in a rather nice framework for
stability analysis with quadratic Lyapunov functions.
A fundamental result of [6] is that under piecewise
real-analycity of all data, q is the sum of two measures: 0
an atomic measure da (representing the acceleration
at impacts), and a measure f q…t†gdt where f
q…t†g is
Lebesgue integrable. One should be aware of the fact
that the measure da can be quite complex, in the
sense that its support may by far not merely consists of
a sequence {tk}k > of ordered times. As we shall see Fig. 16. Unilaterally constrained Lagrangian system.
considering measure derivatives and their densities
instead of the usual derivatives, paves the way toward The next result follows
a generalization of the second method of Lyapunov in
hybrid systems theory. Lemma 7.16. [20]. Consider a mechanical system as in
Before stating the lemma it is useful to recall that (14) and Eq. (93) and denote its smooth potential
the dynamics in (14) can be rewritten as a differential energy as U(q). Then If (q) ‡ U(q) has a strict
inclusion as follows minimum at q0, the equilibrium point (q0, 0)is
Lyapunov stable.
M…q† _ t† 2
q ‡ F…q, q, @  …q† …95† The proof is based on the use of the energy function

with  ˆ {q j h(q) > 0}. It becomes quite useful and _

V…q…t†, q…t†† _ T M…q…t††q…t†
ˆ 12 q…t† _ ‡  …q…t††
powerful now to introduce another inclusion, called
Moreau's measure differential inclusion [81], which ‡ U…q…t†† U…q0 † …98†
uses the cone @ V(q(t))(w(t)) instead of @ (q) in the
right-hand side of (95), with w(t) ˆ (v(t ‡ and calculating its derivative (characterized as the
R t ) ‡ ev(t ))/ density of a measure with respect to the measure
(1 ‡ e) 2 @V(q(t)) and q…t† ˆ q…0† ‡ 0 v…s† ds (@V(
)
denotes the boundary of the tangent cone in Eq. (94)). dt ‡ da as given above) along solutions of the
The interest for this manipulation which yields the inclusion (96).
inclusion
Remark 7.17. The reader may notice a direct similarity
M…q†dv ‡ F…q, v, t† 2 @ V…q…t†† …w…t†† …96† between this result, and what is presented in Section
7.3.2. This shows that the material presented by
is that (96) encompasses in one shot both continuous Moreau in the framework of nonsmooth mechanics in
(without impacts) and discontinuous motion. More- [81,78] relies on a solid analytical ground and lends
over there is a very nice dissipativity interpretation to itself to extensions to other types of dynamical
this. Indeed the operator  7! w(t) where  and w(t) systems. The key point is the design of state re-
satisfy the cone complementarity problem [31] initialization rules which can be expressed as polar
cones complementarity problems as in (97).
N …q†  @ V…q…t†† …w…t†† 3 ?w…t† 2 V…q† …97†
Remark 7.18. In a general setting, one should be aware
is a passive operator (in a convex analysis language, of the fact that the state re-initialization mapping in
this is a maximal monotone operator [91]), because the (4e) cannot be chosen arbitrarily. If the system is a
cones V(q) and N(q) are polar cones (the reader may physical system then it has to satisfy some physical
compare (97) with (69)). It follows that Moreau's coherence, like energy dissipation in Mechanics. But
measure inclusion has the interpretation in Fig. 16 this is not sufficient. Consider for instance two bodies
[20], which is well-known in control theory and exactly moving on a line and which collide. Newton's law tells
matches the absolute stability framework developed in us that linear momentum is conserved at the impact
the mid twentieth century by the Russian school. This time. A dissipative impact law implying that the
representation is valid for all times, including impact two bodies stop after the shock for any pre-impact
times. It has strong mathematical and mechanical velocity, surely is meaningles and will lead to
bases [74,75,78,81]. contradiction and bad-posed dynamics.
350 W.P.M.H. Heemels and B. Brogliato

8. Control Lyapunov function V…~ q, q~_ , t†, where q~ ˆ q qd is the

tracking error, qd(
) is the desired trajectory. It is
In the previous section we have already seen a stabi- shown that the closed-loop ®xed point …~ q, q~_ † ˆ …0, 0† is
lization problem. In this section we will present some globally uniformly asymptotically stable. The central
other control approaches to CS, which lie in the question which arises now in the body of this paper is:
mechanical domain. Of course, one might think of can we extend tracking control to complementarity
other approaches. One of them is the use of passifi- Lagrangian systems as in (101)?
cation techniques and relying on the stability theory
derived in Section 7.3.2. Indeed, if a controller can be M…q† _ q_ ‡ g…q† ˆ u ‡ rh…q†,
q ‡ C…q, q†
found that makes the transfer matrix between the 0 6 h…q† ?  > 0, …101†
maximally monotone nonlinearity (e.g. the com-
plementarity conditions) (strictly) positive real, stabi- _ ‡
Collision rule: q…t _ k ††:
k † ˆ F k …q…t
lity has been realized. Another approach can be based
on using the time-stepping methods (see Section 6) to In fact, bringing an answer to such a question, implies
convert a continuous-time LCS into a discrete-time ®rst to be more accurate and rigorous on the objective
version, which turns out to be equivalent to a mixed of the control task. Indeed a task involving impacts,
logical dynamical system as in Section 4.1.3. As var- ¯ight phases and contact phases, may take various
ious control algorithms have been proposed for this forms. This may be a hammer-like task (a succession
class of hybrid models [8], these become applicable. of impacts with no permanent contact phase as we
Studies are still necessary to get better insight in the shall see in Section 8.2), this can be a bouncing-ball-
real potential of the latter idea. like task (the system stabilizes on the constraint after
a sequence of impacts, and stays there). A com-
8.1. Tracking Control for Complementarity plementarity system as in (101) may evolve in three
Lagrangian Systems different phases of motion:
(i) A free motion phase, where the mechanical
The problem of tracking control for unconstrained system is not subject to any constraints (i.e.
and bilaterally constrained Lagrangian systems, has h(q) > 0, h(
) 2 Rm). The system is represented
been solved [67,70]. The extension toward com- by an Ordinary Differential Equation.
plementarity Lagrangian systems is not trivial, and a (ii) A permanently constraint phase where the dyna-
solution is described now. mical system is subject to holonomic constraints
(hi(q) ˆ 0 during a nonzero time interval and for
8.1.1. Problem Statement some indexes i 2 {1, . . . , m}). The system is re-
presented by a DAE.
The tracking control problem for unconstrained fully (iii) A transition phase whose goal is to stabilize the
actuated Lagrangian systems system on some surface I ˆ \i2I i , where I is
some subset of {1, . . . , m} and i ˆ {q j hi(q) ˆ 0}.
M…q† _ q_ ‡ g…q† ˆ u
q ‡ C…q, q† …99† In other words a transition control has to assure
that hi(q(t)) ˆ 0 and rhi …t††q…t _ ‡ † ˆ 0 for all
has been given various solutions (among them let us i 2 I , where t is a finite time for obvious practi-
cite the computed torque algorithm, the Paden and cal reasons. The system is a Measure Differen-
Panja, Slotine and Li schemes [67]). The extension tial Equation, that may be represented through
toward bilaterally constrained systems an impact Poincare map.
M…q† _ q_ ‡ g…q† ˆ u ‡ rh…q†,
q ‡ C…q, q†
…100†
h…q† ˆ 0, 8.1.2. Stability Framework
where  is the Lagrange multiplier representing the Let us place ourselves in the perspective of a cyclic
contact force, has been achieved by McClamroch and task. In the time domain one gets a representation as:
Wang [70], and Yoshikawa [102]. It essentially relies
on a suitable generalized coordinate transformation, R‡ ˆ
0 [ I0 [
1 [
2 [ I1 [



which allows one to apply the ``free-motion'' schemes |‚‚‚‚‚‚‚‚‚{z‚‚‚‚‚‚‚‚‚}
cycle 0
on a reduced-order subsystem (the ``tangent dynam-
[
2k 1 [
2k [ Ik [
2k‡1


, …102†
ics'' to the constraint surface {q j h(q)=0}). In both |‚‚‚‚‚‚‚‚‚‚‚‚‚{z‚‚‚‚‚‚‚‚‚‚‚‚‚}
cases, the stability analysis relies on the choice of a cycle k
Hybrid Dynamical Systems 351

with state xT ˆ …~q, q~_ † is stable in the sense of Defini-

tions 8.1 or 8.2.
There are two important features in this objective:
 We wish to use a controller that assures asymptotic
Impacts LCP( )
convergence when it is applied on a system without
Fig. 17. The control tasks and transitions between them. constraints as in Eq. (99). Finite-time convergent
controllers are excluded from our study. This is
likely to complicate the control design, since
where
2k denotes the time intervals associated to impacts are not controllable (only the pre-impact
free-motion phases and
2k‡1 those for constrained- velocities can be controlled)
motion phases. The k-th cycle is
2k [ Ik [
2k‡1.  The design of the desired trajectory qd(
) during the
Figure 17 gives a further insight on the dynamics and transition phase, is a crucial step for the overall
the need for considering such cycles (i.e. such discrete- stability.
event paths). In terms of hybrid dynamical systems,
Let us emphasize that the framework we choose, in
Eq. (102) is a discrete-event system trajectory. Let us
particular the cyclic decomposition in Eq. (102), is
denote
ˆ [ k>0
k.
somewhat stringent. Indeed the impacts, which have a
Since we need a stability framework, let us propose
considerable influence on the velocity variation, have
the following [16,19,13], where x(
) denotes the state of
effects which cannot be expected to be compensated
the closed-loop system:
for easily, except in very special cases. An asymptotic
Definition 8.1. (
-weakly stable system). The closed- convergent controller will, in general, not enable one
loop system is
-weakly stable if for each  > 0, there to ``erase'' impacts effects in finite time. Consequently
exists () > 0 such that kx(0)k 6 () ) k x(t)k 6  for great care will be needed to analyze the system from
all t > 0, t 2
ˆ [ k>0
k. Asymptotic
-weak stability one cycle to the next.
holds if in addition x(t) ! 0 as t ! ‡1, t 2
. Prac-
tical
-weak stability holds if there is a ball centered at
x ˆ 0, with bounded radius R > 0, and such that 8.1.3. Lyapunov's Second Method (Suitable
x(t) 2 B(0, R) for all t > T, T < ‡1, t 2
. Extensions)

In view of item (iii) let us define the closed-loop The basic idea is to use a positive function V(x, t) such
impact Poincare map that corresponds to the section that the above stability notions can be proved. As a
I ˆ fx j hi …q† ˆ 0, q_ T rhi …q† < 0, i 2 I g, which is a consequence the design will be based on the choice of a
hypersurface of codimension  ˆ card…I †. The re- nonlinear tracking controller which is known to
impact velocities are chosen to define PI and the render the closed-loop error system globally asymp-
reason for this choice will be made clear after Claim totically stable when the system is unconstrained.
8.5. We define:
Why choosing a single Lyapunov function (conse-
PI : I ! I quently a single controller structure) and not one
function per phase?
xI …k†7!xI …k ‡ 1†, …103†
In fact it is known that for phases
2k and
2k‡1,
where xI is the state of PI . the same controller structure can be used. Thus using
two different controllers for these phases, is useless.
Definition 8.2. (Strongly stable system). The system is
Moreover this would bring serious complications in
said strongly stable if: (i) it is
-weakly stable, (ii) on
the stability analysis, since the Lyapunov function for
phases Ik, PI is Lyapunov stable, and (iii) the
phases
2k has to be monitored during phases
2k‡1,
sequence ftk gk2N has a finite accumulation point
and vice versa. To say nothing of the third phase Ik. It
t1 < ‡1.
therefore seems that there would be very little (if none)
Clearly PI has a fixed point xI 2 @. Now that advantage in using a multiple-Lyapunov functions
we are equipped with a stabilization framework, let us approach for the tracking control of the hybrid
state the main control objective: dynamics (101)±(102).
Given a global asymptotic tracking controller for Let us define the jump function f(t) ˆ f(t‡) f(t )
free-motion tasks, with Lyapunov function V…~ q, q~_ , t†, and [.] is the Lebesgue measure. Let V(
) satisfy
design a feedback control u in (101) such that given a …kXk† > V(x, t) > (kxk), …0† ˆ (0) ˆ 0, (
) and
desired trajectory qd(
), the closed-loop error system …
† strictly increasing. Let Ik ˆ ‰0k , tkf Š.
352 W.P.M.H. Heemels and B. Brogliato

Claim 8.3. (
-Weak Stability [16]). Assume that the difficult to master such jumps. The central question
task is as in Eq. (102), and that then is: can the decrease of V(t) during the impactless
phases, compensate for its variation during Ik? In the
(a) [
] ˆ ‡ 1,
ideal situation, we should manage so that V(t)
(b) for each k 2 N, [Ik] < ‡ 1,
decreases at impacts (this is the meaning of the strong
(c) V…x…tkf †, tkf † 6 V…x…0k †, 0k †,
stability concept in Definition 8.2). But this will not
(d) V(x(.), .) uniformly bounded on each Ik.
always be easy to get. On the other hand, it is crucial
to assure that the variation of V(
) during Ik, does not
_
If on
, V…x…t†, t† 6 0 and V(tk) 6 0 for all k > 0, then amplify from one cycle to the next when the phases
k
the closed-loop system is
-weakly stable. If are of bounded duration.
_
V…x…t†, t† 6 …k x k†, (0) ˆ 0, (
) strictly increas-
Claim 8.5. (Strong Stability [16]). The system is
ing, then the system is asymptotically
-weakly stable.
strongly stable if in addition to the conditions in
Claim 8.3 one has:
This accommodates for other types of motions than
the one as in Eq. (102), see [16]. Let us assume that V…tk‡1 † 6 V…tk †;
t1 < ‡1. It is noteworthy that from [7, Proposition V is uniformly bounded and time continuous on
4.11] this precludes elastic impacts (because if there is Ik [ k{tk}.
no loss of kinetic energy atP impacts, impact times
Sufficient conditions are that V(tk) 6 0 and
satisfy tk‡1 tk >  k > 0 with k>0 k unbounded, so
V…tk‡1 † 6 V…t‡ k †, but this framework permits
that t1 ˆ ‡ 1). On the contrary the results presented
V(tk) > 0 provided V…tk‡1 † < V…t‡ k†  for some
in Section 8.2 hold only if impacts are elastic. Let us _ needs not be
large enough  > 0. Notice also that V…t†
now state a result which incorporates the relation
6 0 along the system's trajectories on the whole of
between the speed at which cycles in Eq. (102) are
covered, and the decrease of the Lyapunov function (tk, tk ‡ 1). The reason why we have chosen I and not
‡I in (103) is that it allows us to take into account the
after stabilization on I has been obtained:12
value V…t0 † in the stability analysis. Notice that
Claim 8.4. (
-Weak Stability [13]). Let us assume that _ ‡
q…t _ 1†
1 † ˆ q…t
(a) and (b) in Claim (8.3) hold and that
(a) outside phase Ik one has V…t† _ 6 V…t† for some 8.1.4. What is the Difference Compared to the
> 0, Unconstrained Case?
(b) inside phase Ik one has V…tk‡1 † V…t‡ k † 6 0,
0
(c) the
P system is initialized on
0 with V…0 † 6 1, Consider the dynamics in Eq. (99). A globally tracking
 k
(d)  …t
k>0 V k † j 6 KV …0 † for some  > 0 and controller guarantees that the closed-loop system
some K > 0. possesses the desired trajectory qd(
) as its unique
Then there exists a constant N < ‡ 1 such that isolated invariant set, with a global basin of attrac-
‰tk1 , tkf Š 6 N, for all k > 0 (the cycle index), and such tion. The control input is a function of qd(
), and the
that: Lyapunov function satisfies V…~ q, q~_ , t† ˆ 0 when
q(
) ˆ qd(
). In fact the closed-loop system that we are
(i) If  > 1 and N ˆ 1= ln……1 ‡ K†=† for some going to design, differs a lot. There are some funda-
0 <  < 1, then V…0k‡1 † 6 V…0k †. The system is mental reasons for such discrepancies, which we try
asymptotically weakly stable. now to summarize. Let us start from a simple one
(ii) If  < 1, then V…0k † 6 … †, where ( ) can be degree-of-freedom example:
made arbitrarily small by increasing . The
system is practically weakly stable with m…
q qd † ‡ 1 …q_ q_ d † ‡ 2 …q qd † 2 @  …q†
R ˆ  1(( )). …104†
The spirit of Claim 8.4 is really ``hybrid'' since it
for some (possibly time-varying though we dropped
merges the continuous and the discrete-event dyna-
arguments in (104)) signal qd …
†. The feedback gains
mical features of the system. The upper-bound on the
are 1 > 0 and 2 > 0. Assume now that qd …t† > 0 for
sum of the Lyapunov function jumps in (d), is the key
all t > 0 and qd …t† ˆ 0 during nonzero time intervals.
of Claim 8.4. As said above it is in general quite
Clearly if  ˆ R then the right-hand side of (104) is
equal to 0, and the unique globally stable invariant
12
In the following we sometimes denote with some abuse of trajectory is q…
† ˆ qd …
†. If  ˆ ( 1, a], a > 0, then
notation V(t) ˆ V(x(t), t). this trajectory is still the unique invariant one,
Hybrid Dynamical Systems 353

however it may not be globally stable (there may be qd …
† ˆ qd …
†. On
2k‡1, qd(
) ˆ 0. In order to converge
impacts with the boundary of ). Consider now toward qic(
), the signal qd …
† has to evolve from cycle k
 ˆ ( 1, a], a > 0. Then clearly qd …
† can no longer be to cycle k ‡ 1 such that, in proportion as tracking
the invariant trajectory, since it ``penetrates'' outside errors decrease on
k, qd …
† approaches an impactless
. Therefore the invariant and stable trajectory (if trajectory. There will consequently be two distinct
any) must be different from qd …
†. Let us denote it notions of convergence: one in time t, the other one
as qic(
). Now let us think about stability with a Lya- along the cycle index k.
punov-like function V(t). One requirement is that
Remark 8.7. An idea (which will be discussed again in
V ˆ 0 when the tracking error is 0, i.e. when the system
Section 8.2) is to design qd(
) such that its first deri-
evolves on the closed-loop invariant. Consequently let
vative has discontinuities at times tk, i.e. simultaneous
us admit for the moment that V(
) has to be a function
_
to jumps in q…
†. This trick would allow one to get
of q qic and q_ q_ ic instead of q_ q_ d and q_ q_ d .
V…t‡0 † ˆ 0 when perfect tracking is assured. However
Remark 8.6. Notice that the right-hand side of the one has also to keep in mind that the variation of V(
)
inclusion in (104) can also be written in a closed-loop has to be characterized for nonzero tracking errors as
form as @ …t†~ …~q† where  ~ q, t† > 0, where
~ ˆ q~ j h…~ well, and studying such variations when impacts do
~ q, t† ˆ h…~
h…~ q ‡ qd …t††.This closed-loop formalism has not match with discontinuities in q_ d …
† adds useless
not yet been used for tracking design purpose (convex difficulties to the problem. The choice for qd (
) in
analysis could be at the core of such a design, as a [19,13] is made to simplify as much as possible all
direct extension of the stabilization framework in calculations for the variations of V(
) during transition
Section 7.4). phases Ik.
Let us think now of the transition phase of motion. In conclusion, asymptotic tracking along a discrete
Let us take  ˆ R ‡ and qd a negative constant. Then trajectory as in Eq. (102) implies in general that [13]:
(104) is the bouncing-ball dynamics. The only invar-
 the closed-loop invariant trajectory is impactless,
iant trajectory is …q, q†_ ˆ …0, 0†. If one wants that the
but robustness of the stabilization on @ implies
same function V(
) serves for this impacting phase as
that impacts do occur during transition phases,
well (see Definition 8.2), then during phases Ik one
 the signals ``desired trajectory'' which enter the
must have V…q ˆ 0, q_ ˆ 0† ˆ 0. In this case it is even
controller and the Lyapunov function, are not
more appropriate to use an impact Poincare mapping
the same.
to represent the dynamics. Once again one sees that it
is not qd which enters the function V(
). What do we mean by ``in general?'' In fact there are
A first conclusion is that the signal which enters the special cases where the inertia matrix is such that the
control input (i.e. qd …
†) is not the signal which enters dynamics tangential to the boundary @ and normal
the Lyapunov function. This is due to the fact that the to @, are decoupled (i.e. in a specific set of general-
system evolves along phases of motion, whose ized coordinates, M(q) is block diagonal). This
underlying dynamics are of different natures. There is decoupling allows one to compensate for the jump
another subtlety in the control design which is a V…t‡0† V…t0 † in a finite time, at each cycle (see [16]).
consequence of asymptotic stability. Indeed assume But as soon as couplings exist, this nice feature van-
that the system is initialized on
0 and with tracking ishes. Let us expand a little on this by showing equa-
errors q qd ˆ 0, q_ q_ d ˆ 0. Then V(t) ˆ 0 for all tions. Assume that h(q) ˆ q1 2 R in (101), where q1 is
t > 0, until there is a first impact at time t0. Then if the first component of q. Let us write the inertia
there is a jump in V(
) at t0, one must have matrix as
V…t‡ 0† V…t0 † ˆ V…t‡0 † > 0. Since we restrict our-
 
M11 …q† M12 …q†
selves to using only controllers which assure asymp- M…q† ˆ :
totic convergence of the tracking errors toward zero, it MT12 …q† M22 …q†
will generally be impossible to compensate for such Then, choosing for instance Moreau's impact law with
positive jumps. Therefore the invariant closed-loop restitution e 2 [0, 1] [81], at an impact one has
trajectory qic will generally have to be impactless,
otherwise its asymptotic stabilization is impossible. q_ 1 …t‡
k† q_ 1 …tk † ˆ …1 ‡ e†q_ 1 …tk †,
Let us recapitulate: V(
) cannot be a function of q_ 2 …t‡
k† q_ 2 …tk † ˆ M221 …q†MT12 …q†…1 ‡ e†q_ 1 …tk †:
qd …
†, qic(
) must be impactless, and during Ik, …105†
V…q ˆ 0, q_ ˆ 0† ˆ 0. This yields us to define a third
signal qd (
) that will enter V(
). On Ik and still dealing It clearly appears that if M12 ˆ 0, then the dynamics in
with (104), qd ˆ 0 to cope with the third item. On
2k, the (q1, q2) coordinates is decoupled, and there is no
354 W.P.M.H. Heemels and B. Brogliato

jump in the generalized velocity q_ 2 . But if M12 6ˆ 0, this reason the Paden and Panja scheme, which is the
then jumps occur in q_ 2 and consequently in the most direct extension of PD control, is chosen in a first
Lyapunov function V(
). If V(
) exactly matches the instance. As we shall see, the closeness of the Lyapu-
kinetic energy at times tk, its variation will auto- nov function to the system's total energy, does not
matically be negative at tk. However making V(
) allow one to overcome all the obstacles of this control
match T…q, q† _ and at the same time assuring the sta- problem. Other parameters, like the ability of char-
bility as in De®nitions 8.1 or 8.2, is not an easy design acterizing the decreasing of V(
) along continuous
task. When M12 ˆ 0 it is easy to split the control motion phases t 2
, may be important.
problem into two subproblems with V(
) ˆ V1(
) ‡ In order to cope with the robustness issue (stabili-
V2(
) where V1(
) and V2(
) evolve independently. The zation on I ) and the asymptotic convergence of the
reader is referred to [13] for a description of this tracking error toward zero (which implies an impact-
control problem in terms of invariant trajectories of less closed-loop invariant trajectory), a specific signal
the closed-loop dynamical system. qd …
† has to be designed during transition phases Ik.
This signal should be such that when perfect tracking
is obtained on
k, then the systems tracks an
8.1.5. Control Design impactless trajectory, whereas for nonzero tracking
errors a bouncing-ball-like motion occurs during Ik.
Let us briefly summarize a solution for tracking con- Such a trajectory is depicted in Fig. 19, where only the
trol, in a general setting (no dynamic decoupling). As normal direction q1 is considered.
said above, the underlying strategy is to use a single The transition phase starts at a time 0k 2
2k (this
controller structure (consequently a unique Lyapunov k is a cycle
2k [ Ik [
2k ‡ 1 index), which is chosen
function) which, when applied to an unconstrained by the designer when stabilization on @ is desired. If
system, guarantees asymptotic global tracking. The V…0k † ˆ 0 (perfect tracking) then there is no impact
input that we are going to present, has some switches and @ is attained tangentially. If V…0k † 6ˆ 0 then
in it, but the switches are at the level of the desired q1d …t† decreases toward V…0k † < 0. The time 0k and
trajectory (see Fig. 18). Therefore it is a feedback the value V…0k † reflect the convergence of qd …
† toward
control of the form u ˆ u…q, q, _ t† that has a fixed qic(
) as explained in the foregoing subsection. The
_ but is discontinuous
structure with respect to q and q, Paden and Panja scheme has the Lyapunov function
in its second argument.
We do not have an unbounded panoply of non- V…t, q~, q~_ † ˆ 12 q~_ T M…q†q~_ ‡ 12 1 q~T q~ ‡  …q†,
linear controllers for tracking control at our disposal.
Among them: feedback linearization (known as the …106†
computed torque method in Robotics), the Paden and
Panja scheme, the Slotine and Li scheme (see e.g. [67]). where, following the above discussion, q~ ˆ q qd . We
Since impacts dissipate kinetic energy it is natural to can set qd …t† ˆ qd …t† until a ®rst impact occurs at t0.
consider those controllers whose Lyapunov function Then qd(t) ˆ 0. Thus for t > t0 the Lyapunov function
is as close as possible the system's total energy. For in Eq. (106) resembles the system's energy, and in

Switching
strategy
Force control
Desired trajectories
generator

Free motion phase trajectory

Lagrangian
Transition phase trajectory Nonlinear controller complementarity
system
structure

Constraint phase trajectory

Fig. 18. The overall structure of the control input signal.

Hybrid Dynamical Systems 355

next idea is to use the decrease of V(
) on phases

k to
compensate for possible V (t0) > 0. But since V(
) can
at best decrease exponentially on
k (the Paden and
Panja scheme does not even guarantee this), one
should take great care that the phases
k duration
and/or the feedback gains, remain uniformly bounded
over the discrete motion in Eq. (102). In other words,
once the cycle frequency and feedback gains have
been chosen, one has to ®nd a control strategy
which prevents jumps V (t0) from amplifying from
one cycle k to the next cycle k ‡ 1. Here Claim 8.4 may
be useful.

Fig. 19. A specific transition trajectory.

8.1.6. Detachment Conditions (Controlled LCP)
particular V…t‡
k† V…tk † ˆ T…t‡
k†
T…tk † 6 0 for
We have essentially focused on the stabilization on @
k > 1, where T…q, q† _ is the kinetic energy. One sees that
and its consequences on stability. However the system
at this stage, an impact model is not necessary. In fact
also has to take off @ when desired. When h(q) ˆ q1,
one major dif®culty lies in the position of the ®rst
such an event is ruled by the complementarity
impact time t0 with respect to the time 1k at which
conditions
qd …
† becomes a constant signal, see Fig. 19. For t < t0
one can set qd …
† ˆ qd …
† and one has V…t†_ 6 0 along 0 6 q1 …t† ? 1 > 0, …108†
k
closed-loop trajectories. If t0 > 1 then V(
) in
Eq. (106) can serve as a Lyapunov function for the which implies on phases
2k‡1 where q1 …t† ˆ q_ 1 …t† ˆ 0
impact Poincare map in (103). But if t0 6 1k (which is the following
the most general case), it is dif®cult to get
V…t‡ 0 6 q1 …t† ? 1 > 0: …109†
0† V…t0 † 6 0, except, as stated above, in special
dynamically decoupled cases13. This ®rst positive
Replacing q1 …t† by its value calculated from the
jump in the Lyapunov function, creates some dif®-
dynamics, allows one to transform (109) into a LCP
culty in applying Claims 8.3 and 8.5. More concretely,
with unknown 1. This LCP(1) is the tool that allows
if the impact rule in Eq. (93) is used, and if q1d …t† is set
one to detect detachment as summarized in Fig. 17. It
constant just after the ®rst shock (which can be
can be shown that q1 …t† ˆ A…q†1 ‡ b…q, q, _ u†. There-
interpreted as a sort of plastic impact in the signal
fore if b…q…t†, q…t†,_ u…t†† > 0 then 1(t) ˆ 0 and
q1d …t† at t0) whereas qd is set to zero, then one gets
necessary conditions for detachment are ful®lled.
e2n 1 It then suf®ces to check whether or not
V …t0 † ˆ ‰M11 M12 M221 MT12 Šq_ 21 …t0 † _
q1 …t† ˆ b…q…t†, q…t†, u…t†† > 0 to conclude on actual
2
1 detachment. Thus the conditions in (109) can be
M11 q_ 21d …t0 † ‡ M11 q_ 1 …t0 †q_ 1d …t0 † considered as a controlled LCP.
2
1 Remark 8.8. Clearly the complementarity conditions
‡ q_ 2 …t0 †T M21 q_ 1d …t0 † 1 q21d …t0 †:
2 cannot be ignored in the control design. If there are
…107† m constraints then one has to check the solution of a
LCP with higher dimension.
Rendering the expression in Eq. (107) nonpositive
thanks to a suitable de®nition of the signals q1d …
† and
q1d (
), is the cornerstone of this tracking control
design. It is also assumed in Eq. (107) that the ``tan- 8.2. Tracking Control Inside a Disk
gential'' part of qd …
†, namely q2d …
† if we adopt the (Billiard Control)
notation in the foregoing section, is set to a constant
value during Ik. Since making V(t0) 6 0 is dif®cult, the A different approach has been taken by Menini
and Tornambe [71]. They consider a planar
system (q 2 R2) that evolves inside a disc, i.e.
13
This result stands also because of the choice we made of the  ˆ f…q1 , q2 † j q21 ‡ q22 1 6 0g. The impacts are
signal qd (
) that is set to zero on (t0, tkf , where tkf is the end of Ik. perfectly elastic and frictionless, which, in particular,
356 W.P.M.H. Heemels and B. Brogliato

implies that solutions (velocities) are piecewise

continuous [7]. More precisely, there exists  > 0 such
that tk ‡ 1 tk >  for all k > 0. The desired trajectory
qd (
) to be tracked is chosen as a polygon with N
vertices inside the disc. Therefore q_ d …
† has dis-
continuities at some known instants. The stability
notion that is introduced takes into account the fact
that two trajectories which do not jump simulta-
neously cannot be arbitrarily close one to each other
in the neighborhood of an impact time, as is well- Fig. 20. Closeness of trajectories around impacts.
known (see e.g. [17, Sections 1.3.2 and 7.1.1]). Here
again the use of dead-beat controllers which assure
some finite-time convergence and would enable one to exists a constant N > 0 such that, if 1 ˆ 2 and
_
get simultaneous discontinuities in both q_ d …
† and q…
†, 2 ˆ 2 one has
is disregarded because it has little chance to work in
q…k ‡ 12†, q~_ …k ‡ 12††
V…~ V…~
q…k 1 _
†, q~…k 1
††
practice. 2 2
1 _
6 … † k …~
q…k 2
†, q~…k 1
2
†† k2 , …111†
Remark 8.9. One sees that the tracking problem that is
attacked in [71] and the one described in Section 8.1, where ( ) > 0 for > N. Thus the Poincare map is
are quite different one from each other. It is not exponentially stable. It follows that:
question in [71] to stabilize the system on @. The task
is closer to a hammer-like motion, or to stabilization  limt!‡1 k q~…t† kˆ 0, limk!‡1 k q~_ …k ‡ † kˆ 0,
by feedback of some billiard trajectory (see [17, limk!‡1 k q~_ …k ‡ †‡ kˆ 0 for all  2 (0, 1),
Definition 7.4]). The central question then is: can  for all  > 0, for all  2 …0, 12†, there exists  > 0
tracking be assured despite the impacts? The type of such that if k q~…0† k<  and k q~_ …0† k< , then
dynamics considered here is therefore close to that of k q~…t† k< , k q~_ …t† k<  for all t such that
an ODE with impulsive input as considered in [5], with jt kj > .
state dependent jump times tk(x). However one should
keep in mind that impact times are defined only The very last condition means that all neighborhoods
implicitly in mechanical systems subject to unilateral of times k where q_ d …
† jumps, are eliminated. See
constraints, not explicitly. Whether or not this com- Fig. 20 for an illustration: between points a and b both
plicates trajectory tracking control design, is another velocity trajectories cannot be arbitrarily close one to
issue. each other, despite they may be when t > tk and t < k.
Evidently this result holds only if the system is initi-
The main result of [71] is that a PD-like control
alized far enough from any impact so that the trajec-
u ˆ 1 …q qd † 2 …q_ q_ d † is sufficient to assure
tory has sufficient time to converge close enough
tracking control of the polygon trajectory, provided
toward the desired one. This is why the result holds
the feedback gains are large enough. We retrieve here
only if impacts are well separated on the time axis.
one characteristic of the controllers described in the
previous section (see Fig. 18), i.e. they have a fixed Remark 8.10. Another common feature of both con-
structure with respect to q and q, _ but are dis- trollers in Sections 8.1 and 8.2, is that they assure local
continuous in their time argument (because q_ d …
† has _ has to be initialized
stability only, i.e. the state (q, q)
jumps). The stability is proved using a discrete-time inside some set.
Lyapunov function that corresponds to a strobo-
scopic Poincare map. More precisely, assume that
q_ d …
† jumps at integer times tk ˆ k. The system is 8.3. Summary
sampled at times t ˆ k 12. The Lyapunov function is
simply chosen as In summary, the extension of trajectory tracking to
complementarity Lagrangian systems, is not trivial.
We have described two approaches in the foregoing
q, q~_ † ˆ 12 q~_ T q~_ ‡ 12 1 q~T q~:
V…~ …110†
Sections 8.1 and 8.2.
One problem that has not been treated yet, is that of
Under the condition that there is exactly one impact in orbital stabilization. In other words, given a desired
the interval (k 12 , k ‡ 12), the following is true: there orbit d in the configuration space C, design a
Hybrid Dynamical Systems 357

controller such that the trajectory q(
) converges evolutional variational inequalities, projected dyna-
towards d in the sense that minp2 d kq(t) pk ! 0 mical systems and so on. However, one has observed
for all t > 0. Some constraints on the velocity should that the class of complementarity systems has its own
also be added to avoid triviality of this stability peculiarities in the sense of mode dynamics living on
concept. lower dimensional subspaces and the possibility of
discontinuities in the state variables. This complicates
8.4. Control of Biped Robots the analysis on one hand, but makes it a huge chal-
lenge on the other. In this paper we have tentatively
The control of biped robots is a very specific and shown that complementarity systems, despite the fact
challenging control problem. It is not our goal here to that they constitute only a subclass of hybrid systems,
survey this wide area, see [52] for references and provide a wealth of potential theoretical studies
discussions on modeling, stability and control. The (controllability, observability, stabilization, etc.) that
stability of a biped robot has itself been the subject of are widely open. Some preliminary results have been
some controversies, and still remains open (see [100] included in the paper that could serve as a starting
where it is shown that the widely used ZMP criterion, point on one hand and indicate the difficulties one has
has severe drawbacks and should not be used for to face on the other. We hope that we stimulated many
walking stability). The complementarity framework is readers in going into this appealing research field and
the suitable one for the design of stable feedback contributing to the analysis and synthesis problems,
controllers. It retains the most important dynamical that are interesting to overcome in view of the broad
features (impacts, Coulomb friction), and at the same range of applications.
time is simple enough.
The stability of a biped robot that walks, can be
seen from two points of view. The first one is that a Acknowledgements
biped is stable, as long as its foot does not slip on the
ground and as long as it does not fall down: this can be The first author would like to thank Kanat CËamllbel,
called the discrete-event stability. The second one is Hans Schumacher, Arjan van der Schaft, Bart De
trajectory tracking: the legs have to track some desired Schutter, Alberto Bemporad, Henk Nijmeijer and
trajectory. This is a low-level, or continuous part Sasha Pogromsky for the fruitful and pleasant coop-
stability. Both are related since the discrete-event eration on complementarity systems.
stability restricts the class of trajectories that can be
tracked. The feedback control of biped robots is, to a
large extent, still an open problem. References

1. Alur R, Courcoubetis C, Halbwachs N, Henzinger TA,

9. Conclusions Ho PH, Nicollin X, Olivero A, Sifakis J, Yovine S. The
algorithmic analysis of hybrid systems. Theoretical
Computer Science 1995; 138: 3±34
In this paper we have presented the complementarity 2. Anderson BDO, Vongpanitlerd S. Network analysis
class of hybrid dynamical systems. Although relatively and synthesis. A modern systems theory approach.
new in the systems and control field, such systems Pentice-Hall, Englewood Cliffs, NJ, 1973
possess a long history in other fields of science like 3. Anistescu M, Potra FA, Stewart D. Time-stepping for
mechanics, optimization and circuit theory. The aim three-dimensional rigid-body dynamics. Computer
Methods in Applied Mechanics Eng 1999; 177: 183±
of the paper was to make the reader familiar with the 197
main features of these highly nonsmooth and highly 4. Arcak M, Kokotovic P. Observer based control of
nonlinear systems and to provide many examples systems with slope-restricted nonlinearities. IEEE
illustrating the developments. This class of hybrid Trans Autom Control 2001; 46(7): 1146±1150
systems is particularly interesting because it has a 5. Bainov DD, Simeonov PS. Systems with impulse
effects; stability, theory and applications. Ellis
strong structure related to convex analysis and Horwood series in Mathematics and its applications,
mathematical programming, which allows one to Wiley, New York 1989
investigate deeply its properties. At the same time 6. Ballard P. The dynamics of discrete mechanical systems
there are many applications domains, motivated with perfect unilateral constraints. Arch Rational Mech
either by physics or by abstract problems. Moreover, Anal 2000; 154(3): 199±274
7. Ballard P. Formulation and well-posedness of the
various links exists to many other interesting model- dynamics of rigid-body systems with perfect unilateral
ing formalism like piecewise affine systems, mixed constraints. Phil Trans Royal Soc., Mathematical,
logical dynamical systems, differential inclusions, Physical and Engineering Siences, special issue on
358 W.P.M.H. Heemels and B. Brogliato

Nonsmooth Mechanics series A 2001; 359 1789: 25. CËamllbel MK. Complementarity methods in the analy-
2327±2346 sis of piecewise linear dynamical systems. Tilburg
8. Bemporad A, Morari M. Control of systems integrating University, The Netherlands, 2001
logic, dynamics, and constraints. Automatica 1999; 26. CËamllbel MK, Heemels WPMH, Schumacher JM. On
35(3): 407±427 linear passive complementarity systems. Eur J Control
9. Bemporad A, Ferrari-Trecate G, Morari M. Observa- 2002; 8(3): 220±237
bility and controllability of piecewise affine and hybrid 27. CËamllbel MK, Heemels WPMH, Schumacher JM.
systems. IEEE Trans Autom Control 2000; 45 10: Consistency of a time-stepping method for a class of
1864±1876 piecewise-linear networks. IEEE Trans circuits and
10. Bemporad A, Morari M, Dua V, Pistikopoulos EN. systems I. Fundamental Theory and Applications 2002;
The explicit linear quadratic regulator for constrained 49(3): 349±357
systems. Automatica 2002; 38(1): 3±20 28. CËamllbel MK, Heemels WPMH, Schumacher JM.
11. Bemporad A, Heemels WPMH, De Schutter, B. On Stability and controllability of planar bimodal linear
hybrid systems and closed-loop MPC systems. IEEE complementarity systems. submitted for publication,
Trans Autom Control 2002; 47(5): 863±869 2003
12. van Bokhoven, WMG. Piecewise linear modelling and 29. Clarke FH, Ledyaev YuS, Sontag ED, Subbotin AI.
analysis. Kluwer, Deventer, The Netherlands, 1981 Asymptotic controllability implies feedback stabilisa-
13. Bourgeot JM, Brogliato B. Tracking control of tion. IEEE Trans Autom Control 1997; 42(10): 1394±
complementarity Lagrangian systems. Submitted to 1407
the International Journal of Bifurcation and Chaos, 30. Clarke FH, Ledyaev YuS, Stern RJ, Wolenski PR.
special issue Non-smooth Dynamical Systems: Recent Nonsmooth analysis and control theory. Springer
Trends and Perspectives. INRIA Research Report, GTM 178, New York, 1998
2003 available at http://www.inria.fr/rrrt/index.en.html 31. Cottle RW, Pang JS, Stone RE. The linear comple-
14. Branicky MS. Stability theory for hybrid dynamical mentarity problem. Academic Press, New York, 1992
systems. IEEE Trans Autom Control 1998; 43(4): 32. Ten Dam AA, Dwarshuis E, Willems JC. The contact
475±482 problem for linear continuous-time dynamical systems:
15. Brezis H. OpeÂrateurs maximaux monotones. Mathe- a geometric approach. IEEE Trans Autom Control
matics Studies, North-Holland, 1973 1997; 42(4): 458±472
16. Brogliato B, Niculescu S, Orhant P. On the control of 33. Deimling K. Multivalued differential equations. De
finite dimensional mechanical systems with unilateral Gruyter series in nonlinear analysis and applications,
constraints. IEEE Trans Autom Control 1997; 42(2): Berlin 1992
200±215 34. De Schutter B, van den Boom T. 2000; Model
17. Brogliato B. Nonsmooth mechanics. 2nd ed. Springer predictive control for max-min-plus-scaling systems.
CCES, London 1999. Erratum and addenda available Technical report of Technical University of Delft,
at http://www.inrialpes.fr/bip/people/brogliato/bro- Electrical Engineering, Control Lab. No. bds:00-03.
gli.html 35. Dupuis P, Nagurney A. Dynamical systems and
18. Brogliato B, Zavala-Rio A. On the control of variational inequalities. Ann Oper Res 1993; 44: 9±42
complementary-slackness juggling mechanical systems. 36. Eaves BC, Lemke CE. Equivalence of LCP and PLS.
IEEE Trans Autom Control 2000; 45(2) 235±246. Math Oper Res 1981; 6(4): 475±484
19. Brogliato B, Niculescu SI, Monteiro Marques MDP. 37. Ferris MC, Pang JS. Engineering and economic
On tracking control of a class of complementary- applications of complementarity problems. SIAM Rev
slackness mechanical systems. Systems Control Lett 1997; 39(4): 669±713
2000; 39(4) 255±266 38. Filippov AF. Differential equations with discontinuous
20. Brogliato B. Absolute stability and the Lagrange- righthand sides. In the series Mathematics and Its
Dirichlet theorem with monotone multivalued map- Applications, Kluwer, Dordrecht, The Netherlands,
pings. 40th IEEE Conference on Decision and Control, 1998
Orlando, USA, December 4±7 2001. Submitted to 39. FreÂmond M. Non-smooth thermomechanics. Springer
Systems and Control Letters. INRIA Research Report Verlag, Berlin-Heidelberg, 2002
4794, available at http://www.inria.fr/rrrt/index.en.html 40. GeÂnot F, Brogliato B. New results on painleve para-
21. Brogliato B, ten Dam AA, Paoli L, GeÂnot F, Abadie M. doxes. Eur J Mechanics A/Solids, 1999; 18(4): 653±678
Numerical simulation of finite dimensional multibody 41. Goeleven D, Brogliato B. 2003; Stability and unstabil-
nonsmooth mechanical systems. ASME Applied ity matrices for linear evolution variational inequalities.
Mechanics Rev 2002; 55(2): 107±150 INRIA Research Report 4768 (http://www.inria.fr/rrrt/
22. Brogliato B. Some perspectives on the analysis and index.en.html). Submitted
control of complementarity systems. IEEE Trans 42. Hartl RF, Sethi SP, Vickson RG. A survey of the
Autom Control June 2003; 48(6) maximum principles for optimal control problems with
23. Brogliato B, Mabrouk M, Zavala-Rio A. On the state constraints. SIAM Rev 1995; 37(2): 181±218
controllability of linear juggling mechanical systems. 43. Heemels WPMH, Schumacher JM, Weiland S. 1999;
INRIA Research Report 4772, 2003. available at Applications of complementarity systems. Proceedings
http://www.inria.fr/rrrt/index.en.html. Submitted to of the European control conference in Karlsruhe
Automatica. (Germany)
24. Brogliato B. Some results on the controllability of 44. Heemels WPMH, Schumacher JM, Weiland S. The
planar variational inequalities. INRIA Report 4788, rational complementarity problem. Linear Algebra and
2003. http://www.inria.fr/rrrt/index.en.html its Appli 1999; 294: 93±135
Hybrid Dynamical Systems 359

45. Heemels WPMH, Schumacher JM, Weiland S. 63. Lee KK, Arapostathis A. On the controllability of
Linear complementarity systems. SIAM J Appl Math piecewise-linear hypersurface systems. Systems &
2000; 60(4): 1234±1269 Control Lett 1987; 9: 89±96
46. Heemels WPMH, Schumacher JM, Weiland S. Pro- 64. Leenaerts DMW, van Bokhoven WMG. Piecewise
jected dynamical systems in a complementarity form- linear modelling and analysis. Kluwer Academic
alism. Oper Res Lett 2000; 27(2): 83±91 Publishers, Dordrecht, The Netherlands, 1998
47. Heemels WPMH, Camlibel MK, van der Schaft AJ, 65. LoÈtstedt P. Mechanical systems of rigid bodies subject
Schumacher JM. Well-posedness of the complementar- to unilateral constraints. SIAM J of Applied Mathe-
ity class of hybrid systems. Proceedings of the IFAC matics, 1998; 42(2): 281±296
World Congress, Barcelona, 2002 66. Lootsma YJ, van der Schaft AJ, Camlibel MK.
48. Heemels WPMH, CËamlibel MK, Schumacher JM. Uniqueness of solutions of linear relay systems.
On the dynamic analysis of piecewise-linear networks. Automatica 1999; 35(3): 467±478
IEEE Trans Circuits Systems ± I 2002; 49(3): 67. Lozano R, Brogliato B, Egeland O, Maschke B.
315±327 Dissipative Systems Analysis and Control. Theory
49. Heemels WPMH, CËamlibel MK, van der Schaft AJ, and Applications. Springer CCES, London, 2000
Schumacher JM. Modelling, well-posedness and stabi- 68. Lynch KM, Black CK. Recurrence, controllability, and
lity of switched electrical networks. In: Proceedings of stabilization of juggling. IEEE Trans Robotics Auto-
Hybrid Systems: Computation and Control, Prague, mation 2001; 17(21): 113±124
2003, 249±266 69. Mata Jimenez M, Brogliato B. Analysis of PD and
50. Heemels, WPMH, De Schutter B, Bemporad A. nonlinear control of mechanical systems with dynamic
Equivalence of hybrid dynamical models. Automatica backlash. J Vibrations Control 2003; 9(1): 119±156
2001; 37(Iss.): 1085±1091 70. McClamroch NH, Wang D. Feedback stabilization and
51. Hiriart-Urruty JB, LemareÂchal C. Fundamentals of tracking of constrained robots. IEEE Trans Autom
Convex Analysis. Springer Grundlehren Text Edition, Control 1988; 33(5): 419±426
Berlin, 2001 71. Menini L, Tornambe A. Asymptotic tracking of
52. Hurmuzlu Y, GeÂnot F, Brogliato B. 2001; Modeling, periodic trajectories for a simple mechanical system
stability and control of biped robots ± a general frame- subject to nonsmooth impacts. IEEE Trans Autom
work. INRIA Research Report 4290, http://www. Control 2001; 46(7): 1122±1126
inria.fr/rrrt/index.en.html. Submitted to Automatica. 72. Menini L, Tornambe A. Velocity observers for non-
53. Janin O, Lamarque CH. Comparison of several numer- linear mechanical systems subject to non-smooth
ical methods for mechanical systems with impacts. Int J impacts. Automatica 2002; 38(12): 2169±2176
Numerical Methods Eng 2001; 51(9): 1101±1132 73. Menini L, Tornambe A. State estimation of (otherwise
54. Johansson KJ, Egerstedt M, Lygeros J, Sastry S. On unobservable) linear mechanical systems through the
the regularization of zeno hybrid automata. Systems use of non-smooth impacts: the case of two mating
Control Lett 1999; 38: 141±150 gears. Automatica 2002; 38(10): 1823±1826
55. Johansson M, Rantzer A. Computation of piecewise 74. Moreau JJ. Les liaisons unilateÂrales et le principe de
quadratic lyapunov functions for hybrid systems. IEEE Gauss. CRAS Paris 1963; 256: 871±874
Trans Autom Control 1998; 43(4): 555±559 75. Moreau JJ. Quadratic programming in mechanics. J
56. John F. Extremum problems with inequalities as SIAM Control 1966; 4(1): 153±158
subsidiary conditions. In: Studies and Essays, Courant 76. Moreau JJ. Fonctionnelles convexes. SeÂminaire sur les
Anniver-sary Volume, Interscience, New York, 1948; eÂquations aux deÂriveÂes partielles, ColleÁge de France,
148±204 Paris, 1966±1967
57. Juloski A, Heemels WPMH, Weiland S. Observer 77. Moreau JJ. Evolution problem associated with a
design for a class of piecewise affine systems. In: moving convex set in a Hilbert space. J Differential
Proceedings of the conference on decision and control Equations 1977; 26: 347±374
2002, Las Vegas, USA, pp. 2606±2611 78. Moreau JJ. Liaisons unilateÂrales sans frottement et
58. Khalil HK. Nonlinear Systems. MacMillan, New York chocs ineÂlastiques. C.R. Acad Sc Paris 1983; t.296: seÂrie
1992 II 1473±1476
59. Klarbring A. Mathematical programming in contact 79. Moreau JJ. Standard inelastic shocks and the dynamics
problems. Computational Mechanics Publications, of unilateral constraints. Springer Verlag, CISM
Southamton, 1993, 233±264 Courses and Lectures 1985; 288: 173±221
60. Krupa M, Khun A, Poth W, Schagerl M, Steiner W, 80. Moreau JJ. Dynamique des systeÁmes a liaisons uni-
Steindl A, Troger H, Wiedermann G. Tethered lateÂrales avec frottement sec eÂventuel; essais numeÂr-
satellite systems: A new concept of space flight. iques. Technical note 85-1, LMGC, University of
Euro J Mechanics A/Solids 2000; 19(special issue): Montpellier 2, France 1986
149±164 81. Moreau JJ. Unilateral contact and dry friction in finite
61. Kuhn HW, Tucker AW. Nonlinear programming. freedom dynamics. In Nonsmooth Mechanics and
In: Proceedings of 2nd Berkeley Symposium on Applications, CISM Courses and Lectures no 302,
Mathematical Statistics and Probability, 1951, Moreau JJ, P.D. Panagiotopoulos (eds). International
481±492 Centre for Mechanical Sciences, Springer-Verlag, 1988;
62. Kunze M, Monteiro Marques M. An introduction to pp. 1±82
Moreau's sweeping process. In: Brogliato B (ed). 82. Moreau JJ. An introduction to unilateral dynamics. In:
Impacts in Mechanical Systems, Springer Lecture FreÂmond M, Macceri F (eds). Novel approaches in civil
Notes in Physics LNP 551, 2000, pp. 1±60. engineering, Springer Verlag, 2003; pp. 1±45
360 W.P.M.H. Heemels and B. Brogliato

83. Murty KG. Linear complementarity, linear and non- 93. van der Schaft AJ, Schumacher JM. Complementarity
linear programming. 1997 available at http://www- Modelling of Hybrid Systems. IEEE Trans Autom
personal. engin.umich.edu/Ämurty/book/LCPbook/ Control 1998; 43(4): 483±490
84. Nagurney A, Zhang D. Projected dynamical systems 94. van der Schaft AJ, Schumacher JM. An introduction
and variational inequalities with applications. Kluwer's to hybrid dynamical systems. Springer, London, Lect-
International Series in Operations research and ure Notes in Control and Information Sci 1999; 251
Management Science, Boston, 1996 95. Schumacher JM. Complementarity systems in optimi-
85. Nijmeijer H, van der Shaft AJ. Nonlinear dynamical zation. submitted for publication 2003
control systems. Springer Verlag, Berlin, 1990 96. Sontag ED. Nonlinear regulation: the piecewise linear
86. Paoli L. 2002 A numerical scheme for impact problems approach. IEEE Trans Automatic Control 1981;
with inelastic shocks: a convergence result in the multi- 26(2): 346±357
constraint case. In: Proc. of the international con- 97. Stewart DE. Rigid-body dynamics with friction and
ference on nonsmooth/nonconvex mechanics with impacts. SIAM Rev 2000; 42(1): 3±39
applications in Engineering, 5±6 July, Thessaloniki, 98. Stronge WJ. Impact Mechanics. Cambridge Univer-
Greece, C.C. Baniotopoulos (ed), Ziti editions, pp. 269± sity Press, Cambridge 2000
274 99. Veliov VM, Krastanov MI. Controllability of piece-
87. Paoli L, Schatzman M. A numerical scheme for impact wise linear systems. Systems & Control Lett 1986; 7:
problems. I The one-dimensional case; II The multi- 335±341
dimensional case. SIAM J Numerical Analysis 2002; 100. Wieber PB. On the Stability of walking systems. 3rd
40(2): 702±768 IARP international workshop on Humanoid
88. Pfeiffer F, Glocker C, Multibody Dynamics with and Human Friendly Robotics, December 11±12,
Unilateral Contacts. Wiley interscience, series in non- Tsukuba Research Center, Ibaraki, Japan, 2002
linear Dynamics 1996 101. Willems JC. Dissipative dynamical systems. Arch
89. Pontryagin LS, Boltyanskii VG, Gamkrelidze RV, Rational Mech Anal 1972; 45: 321±393
Mishchenko EF. The mathematical theory of optimal 102. Yoshikawa T, Dynamic hybrid position/force control
pro-cesses. John Wiley & Sons, 1962 of robot manipulatros ± Description of hand con-
90. Yu Pogromsky A, Heemels WPMH, Nijmeijer H. On straints and calculation of joint driving force. IEEE
well-posedness of relay systems. Proc. NOLCOS 2001, Trans Robotics and Autom 1987; 3(5): 386±392
St. Petersburg (Russia) 2001 103. Zavala-Rio A, Brogliato B. On the control of a one
91. Rockafellar RT. Convex Analysis. Princeton Univer- degree-of-freedom juggling robot. Dynamics and
sity Press, Princeton, NJ, 1972 Control 1999; 9: 67±90
92. van der Schaft AJ, Schumacher JM. The complemen- 104. Zavala-Rio A, Brogliato B. Direct adaptive control
tary-slackness class of hybrid systems. Mathematics of design for one-degree-of-freedom complementary-
Control, Signals and Systems 1996; 9: 266±301 slackness jugglers. Automatica, 2001; 37: 1117±1123