Hybrid Dynamic Systems Tutorial PDF
Hybrid Dynamic Systems Tutorial PDF
Hybrid Dynamic Systems Tutorial PDF
RICARDO G. SANFELICE,
and ANDREW R. TEEL
ROBUST STABILITY
AND CONTROL FOR
SYSTEMS THAT COMBINE
CONTINUOUS-TIME AND
DISCRETE-TIME DYNAMICS
M
any dynamical systems combine behaviors ally, systems involving both digital and analog components
that are typical of continuous-time dynamical form another class of examples. Finally, modern control al-
systems with behaviors that are typical of dis- gorithms often lead to both kinds of behavior, due to either
crete-time dynamical systems. For example, digital components used in implementation or logic and
in a switched electrical circuit, voltages and decision making encoded in the control algorithm. These
currents that change continuously according to classical examples fit into the class of hybrid dynamical systems, or
electrical network laws also change discontinuously due simply hybrid systems.
to switches opening or closing. Some biological systems This article is a tutorial on modeling the dynamics of
behave similarly, with continuous change during normal hybrid systems, on the elements of stability theory for
operation and discontinuous change due to an impul- hybrid systems, and on the basics of hybrid control. The
sive stimulus. Similarly, velocities in a multibody system presentation and selection of material is oriented toward
change continuously according to Newton’s second law but the analysis of asymptotic stability in hybrid systems
undergo instantaneous changes in velocity and momen- and the design of stabilizing hybrid controllers. Our em-
tum due to collisions. Embedded systems and, more gener- phasis on the robustness of asymptotic stability to data
perturbation, external disturbances, and measurement
Digital Object Identifier 10.1109/MCS.2008.931718 error distinguishes the approach taken here from other
NESTOR DUB
approaches to hybrid systems. While we make some con- We now present some background leading up to the
nections to alternative approaches, this article does not as- model of hybrid systems used in this article. A widely
pire to be a survey of the hybrid system literature, which is used model of a continuous-time dynamical system is the
first-order differential equation x 5 f 1 x 2 , with x belonging
#
vast and multifaceted.
The interaction of continuous- and discrete-time dynam- to an n-dimensional Euclidean space Rn. This model can
ics in a hybrid system leads to rich dynamical behavior and be expanded in two directions that are relevant for hybrid
phenomena not encountered in purely continuous-time systems. First, we can consider differential equations with
state constraints, that is, x 5 f 1 x 2 and x [ C, where C is a
#
systems. Consequently, several challenges are encountered
on the path to a stability theory for hybrid systems and to a subset of Rn. For example, the set C might indicate that
methodology for robust hybrid control design. The approach the force of gravity cannot push a ball through the floor.
outlined in this article addresses these challenges, by using Alternatively, the set C might indicate a set of physically
mathematical tools that go beyond classical analysis, and meaningful initial conditions of the system. Second, we
leads to a stability theory that unifies and extends the theo- can consider the situation where the right-hand side of the
ries developed for continuous- and discrete-time systems. In differential equation is replaced by a set that may depend
particular, we give necessary and sufficient Lyapunov con- on x. For example, when the force applied to a particle var-
ditions for asymptotic stability in hybrid systems, show uni- ies with time in an unknown way in the interval 3 a, b 4 ,
formity and robustness of asymptotic stability, generalize we can model the derivative of the velocity as belonging
the invariance principle to the hybrid setting and combine it to 3 a, b 4 . Another reason for considering set-valued right-
with Barabasin-Krasovskii techniques, and show the utility hand sides is to account for the effect of perturbations,
of such results for hybrid control design. Despite their neces- such as measurement error in a feedback control system,
sarily more technical appearance, these results parallel what on a differential equation. Both situations lead to the dif-
ferential inclusion x [ F 1 x 2 , where F is a set-valued mapping.
#
students of nonlinear systems are familiar with.
Combining the two generalizations leads to constrained dif- replace the set-valued mappings and the corresponding
ferential inclusions x [ F 1 x 2 , x [ C.
#
inclusions in (1), (2) with equations and proceed confi-
A typical model of a discrete-time dynamical system dently. It is often the geometry of sets C and D that pro-
is the first-order equation x 1 5 g 1 x 2 , with x [ Rn. The duces the rich dynamical phenomena in a hybrid system
notation x 1 indicates that the next value of the state is rather than the multivaluedness of the mappings F and G.
given as a function of the current state x through the However, this article does justify, beyond the sake of gen-
value g 1 x 2 . As for differential equations, it is a natural erality, the use of differential inclusions and difference
extension to consider constrained difference equations inclusions.
and difference inclusions, which leads to the model We provide examples of hybrid models in the follow-
x 1 [ G 1 x 2 , x [ D, where G is a set-valued mapping and ing section. Subsequently, we make precise the meaning
D is a subset of Rn. of a solution to a hybrid dynamical system and describe
Since a model of a hybrid dynamical system requires a basic mathematical properties of the space of solutions.
description of the continuous-time dynamics, the discrete- Afterward, we present results on asymptotic stability in
time dynamics, and the regions on which these dynamics hybrid systems, with an emphasis on robustness. Initially,
apply, we include both a constrained differential inclusion we focus on Lyapunov functions as the primary stability
and a constrained difference inclusion in a general model analysis tool and show how Lyapunov functions are used
of a hybrid system in the form in hybrid control design. Finally, we present tools for sta-
bility analysis based on limited events in hybrid systems
#
x [ F ( x ) , x [ C, (1) and show how these tools are related to hybrid feedback
x 1 [ G ( x ) , x [ D. (2) control algorithms.
The main developments of the article are comple-
The model (1), (2) captures a wide variety of dynamic mented by several supporting discussions. “Hybrid
phenomena including systems with logic-based state Automata” and “Switching Systems” relate systems
components, which take values in a discrete set, as well in the form (1), (2) to hybrid automata and switching
as timers, counters, and other components. Examples systems, respectively. “Related Mathematical Frame-
in this article demonstrate how to cast hybrid automata works” presents other mathematical descriptions of
and switched systems, as well as sampled-data and net- systems where features of both continuous- and discrete-
worked control systems, into the form (1), (2). We refer time dynamical systems are present. “Existence, Unique-
to a hybrid system in the form (1), (2) as H. We call C ness, and other Well-Posedness Issues” discusses basic
the flow set, F the flow map, D the jump set, and G the properties of solutions to hybrid systems in the form
jump map. (1), (2). “Set Convergence” and “Robustness and Gen-
For many systems, the generality provided by the in- eralized Solutions” introduce mathematical tools from
clusions in (1), (2) is not needed. Thus, the reader may beyond classical analysis and motivate the assumptions
placed on the data of a hybrid sys-
tem H, as given by 1 C, F, D, G 2 . “Mo-
tivating Stability of Sets” and “Why
ν10 ν20 ‘Pre’-Asymptotic Stability?” explain
distinct features of the asymptotic
stability concept used in the article.
0 0 p 0
0 p∗
p1 2 “Converse Lyapunov Theorems” and
(a) (b) “Invariance” state and discuss main
tools used in the stability analysis.
“Zeno Solutions” describes a phe-
nomenon unique to hybrid dynamical
(c) systems. “Simulation in Matlab/
Simulink” presents an approach to
FIGURE 1 Collision between two particles. (a) Two particles are initialized to positions p01 simulation of hybrid systems. The
and p02 and with velocities v 01 and v 02 . (b) An impact between the particles occurs at the notation used throughout this article
position p* . (c) The direction of the motion of each particle is reversed after the impact. is defined in “List of Symbols.”
R1f2 5 c
cosf 2 sinf
d
An impact occurs when the positions of the particles
are identical and their velocities satisfy v1 $ v2. These con- sinf cosf
ditions define the jump set D J 5 1 p, v 2 : p1 5 p2, v1 $ v2 6 . F(S) h x [SF ( x ) for the set-valued mapping
Letting v11 and v21 indicate the velocities after an impact, F : Rm S S Rn and a set S ( Rm
we have the conservation of momentum equation TS ( h ) The tangent cone to the set S ( Rn at h [ S .
TS ( h ) is the set of all vectors w [ Rn for
m1v11 1 m2v21 5 m1v1 1 m2v2 (3)
which there exist h i [ S , ti . 0 , for all
i 5 1, 2, c such that h i S h , ti R 0 , and
and the energy dissipation equation ( h i 2 h ) /ti S w as i S `
K` The class of functions from R $ 0 to R $ 0 that
v11 2 v21 5 2 r 1 v1 2 v2 2 , (4) are continuous, zero at zero, strictly increas-
ing, and unbounded
where m1 and m2 are the masses of the particles and the LV ( m ) The m -level set of the function V : d om V S R ,
constant r [ 1 0, 1 2 is a restitution coefficient. Solving (3) which is the set of points 5 x [ d om V :
and (4) for v11 and v21 and using (2) yields the jump map V ( x ) 5 m6
» a set of modes Q, which in most situations can be identi- When j is an element of two different guard sets Guard ( q, q r )
fied with a subset of the integers and Guard ( q, q s ) , Gq ( j ) is a set consisting of at least two
» a domain map Domain: Q S S Rn, which gives, for each points. Hence, Gq can be set valued. In fact, Gq is not neces-
q [ Q , the set Domain(q) in which the continuous state sarily a function even when every Reset ( q, q r, # ) is the identity
j evolves map. With Cq , Fq , Dq , and Gq defined above, we consider the
» a flow map f : Q 3 Rn S Rn , which describes, through hybrid system with state ( j, q ) [ Rn 3 R and representation
a differential equation, the continuous evolution of the
#
continuous state variable j j 5 Fq ( j ) , q [ Q, j [ Cq,
» a set of edges Edges ( Q 3 Q , which identifies the
pairs ( q, q r ) such that a transition from the mode q to
( j 1 , q 1 ) [ Gq ( j ) , q [ Q, j [ Dq.
the mode q r is possible
» a guard map Guard : Edges S S Rn, which identifies, for each Example S1: Reformulation of a Hybrid Automaton
edge q, q r [ Edges, the set Guard 1 q, q r 2 to which the
( ) Consider the hybrid automaton shown in figures S2 and S3,
continuous state j must belong so that a transition from q with the set of modes Q 5 5 1, 2 6 ; the domain map given by
to q r can occur
» a reset map Reset : Edges 3 Rn S Rn , which de- Domain ( 1 ) 5 R # 0 3 R, Domain ( 2 ) 5 5 0 6 3 R;
scribes, for each edge ( q, q r ) [ Edges, the value to
which the continuous state j [ Rn is set during a transi- the flow map, for all j [ R2 , given by
tion from mode q to mode q r. When the continuous state
variable j remains constant at a jump from q to q r, the f ( 1, j ) 5 ( 1, 1 ) , f ( 2, j ) 5 ( 0,21 ) ;
map Reset 1 q, q r, # 2 can be taken to be the identity.
Figure S1 depicts part of a state diagram for a hybrid automaton. the set of edges given by Edges 5 5 ( 1, 1 ) , ( 1, 2 ) , ( 2, 1 ) 6 ;
The continuous dynamics of two modes are shown, together the guard map given by
Guard ( 1, 1 ) 5 R $ 0 3 R # 0,
ξ ∈Guard(q, q′) ⇒ ξ+ ∈Reset(q, q′, ξ) Guard ( 1, 2 ) 5 R2$ 0,
Guard ( 2, 1 ) 5 5 0 6 3 R # 0;
q q′
. .
ξ = f (q, ξ ) and the reset map, for all j[ R2 , given
ξ = f (q′, ξ )
ξ ∈Domain(q) by
ξ ∈Domain(q′)
Reset ( 1, 1, j ) 5 ( 25, 0 ) ,
Reset ( 1, 2, j ) 5 j,
ξ+ ∈Reset(q′, q, ξ) ⇒
ξ ∈Guard(q′, q) Reset ( 2, 1, j ) 5 2j.
from mode q to mode qr can occur when, in mode q, j is in the set Guard 1 q, q r 2 .
overlap, indicating that, in mode 1, a re-
During the transition, j changes to a value j 1 in Reset 1 q, q r, j 2 . Transitions from set of the state j to (25, 0 ) or a switch
mode q to other modes, not shown in the figure, are governed by similar rules. of the mode to 2 is possible from points
# 4
j 5 fq ( j ) , (S1)
3
where, for each q [ Q 5 5 1, 2, c, qmax 6 , fq : Rn S Rn is a
continuous function. A complete solution to the system (S1) con- 2
sists of a locally absolutely continuous function j : R $ 0 S Rn
and a function q : R $ 0 S Q that is piecewise constant, has a 1
finite number of discontinuities in each compact time interval,
#
and satisfies j ( t ) 5 fq (t) ( j ( t ) ) for almost all t [ R $ 0 . In what
follows, given a complete solution ( j, q ) to (S1), let I be the 0 t1 t2 t3 t
number of discontinuities of q, with the possibility of I 5 ` , and
let t0 5 0 and 5 ti 6 Ii 5 1 be the increasing sequence of times at
(b)
which q is discontinuous. For simplicity, we discuss complete FIGURE S4 Hybrid time domain for a solution under dwell-time
solutions only. switching and average dwell-time switching. (a) Hybrid time do-
A solution ( j, q ) to (S1) is a dwell-time solution with dwell main for a dwell-time solution with dwell-time constant tD larger
time tD . 0 if ti 1 1 2 ti $ tD for all i 5 1, 2, c, I 2 1 . That is, or equal than min 5 t2 2 t1, t3 2 t2, c 6 . (b) Hybrid time domain for
an average dwell-time solution for parameters ( d, N ) satisfying
switches are separated by at least an amount of time tD . Each
the average dwell-time condition in (S2). For example, parameters
dwell-time solution can be generated as part of a solution to the 1 d, N 2 5 1 4 /min 5 t1, t2 2 t1 6 , 2 2 and ( d, N ) 5 ( 4 /t2, 4 ) satisfy
hybrid system with state x 5 ( j, q, t ) [ Rn 1 2 given by (S2). The domain repeats periodically, as denoted by the blue dot.
Consider a group of fireflies, each of which has an is continuous, and flow set C 5 3 0, 1 4 n, where 3 0, 1 4 n
internal clock state. Suppose each firefly’s clock state indicates the set of points x in Rn for which each com-
increases monotonically until it reaches a positive thresh- ponent xi belongs to the interval 3 0, 1 4 . The function f
old, assumed to be the same for each firefly. When a governs the rate at which each clock state evolves in the
firefly’s clock reaches its threshold, the clock resets to interval 3 0, 1 4 .
zero and the firefly flashes, which causes the other fire- Since jumps in the state of the system are allowed when
flies’ clocks to jump closer to their thresholds. In this any one of the fireflies’ clocks reaches its threshold, the
way, the flash of one firefly affects the internal clocks of jump set is D 5 5 x [ C : maxi x i 5 1 6 . One way to model
the other fireflies. the impulsive changes in the clock states is through a
Figure 2 depicts the evolution of the internal clocks of rule that instantaneously advances a clock state by a fac-
two fireflies coupled through flashes. The time units are tor 1 1 e, where e . 0, when this action does not take the
normalized so that each firefly’s internal clock state takes clock state past its threshold. Otherwise, the clock state
values in the interval 3 0, 1 4 , and thus the threshold for flash- is reset to zero, just as if it had reached its threshold. The
ing is one for each firefly. corresponding jump map G does not satisfy the regularity
A hybrid model for a system of n fireflies, each with the condition (A3) of the Basic Assumptions imposed in the
same clock characteristics, has the state x 5 1 x 1, c, x n 2 , section “Basic Mathematical Properties.” The algorithm
flow map F 1 x 2 5 1 f 1 x 1 2 , c, f 1 x n 2 2 , where f : 3 0, 1 4 S R .0 for defining generalized solutions in “Robustness and
t [ 3 0, d 4 , t [ 3 0, N 4 ,
# REFERENCES
(S3)
t 5 t 2 1, t [ 3 1, N 4 .
1
[S3] J. P. Hespanha, “Uniform stability of switched linear systems: Ex-
(S4) tensions of LaSalle’s invariance principle,” IEEE Trans. Automat. Contr.,
vol. 49, pp. 470–482, 2004.
The time domain for each solution of this hybrid system satisfies the [S4] D. Liberzon and A. S. Morse, “Basic problems in stability and design of
constraint (S2). Furthermore, for every hybrid time domain E sat- switched systems,” IEEE Control Syst. Mag., vol. 19, pp. 59–70, 1999.
[S5] S. Mitra and D. Liberzon, “Stability of hybrid automata with average
isfying (S2) there exists a solution of (S3), (S4), starting at t 5 N ,
dwell time: An invariant approach,” in Proc. 43rd IEEE Conf. Decision
and defined on E [S5], [14]. In turn, switching systems under an av- and Control, Bahamas, Dec. 2004, pp. 1394–1399.
Generalized Solutions” motivates the modified jump map are equal [55]. Synchronization analysis for the case n 5 2
G 1 x 2 5 1 g 1 x 1 2 , c, g 1 x n 2 2 , where is given in Example 25.
q
1
0 t
x2
1
t
–1
(a)
ξ2
0 t
fp 1 xp, kc 1 x 2 2
F1x2 5 c
Hybrid Controllers for Nonlinear Systems
d for all x [ C,
fc 1 x 2
(6)
Hybrid dynamical systems can model a variety of closed-
loop feedback control systems. In some hybrid control appli-
cations the plant itself is hybrid. Examples include juggling and jump map
[70], [73] and robot walking control [63]. In other applica-
tions, the plant is a continuous-time system that is con-
trolled by an algorithm employing discrete-valued states. u=z Nonlinear ξ
This type of control appears in a broad class of industrial System
applications, where programmable logic controllers and
microcontrollers are employed for automation. In these ap-
T
plications, discrete states, as well as other variables in soft- D/A A/D
ZOH T
ware, are used to implement control logic that incorporates
decision-making capabilities into the control system.
Algorithm
Consider a plant described by the differential equation
x p 5 fp 1 x p, u 2 ,
#
(5)
FIGURE 5 Digital control of a continuous-time nonlinear system
where x p [ R , u [ R , and fp is continuous. A hybrid
n r with sample-and-hold devices performing the analog-to-digital
(A/D) and digital-to-analog (D/A) conversions. Samples of the
controller for this plant has state x c [ Rm , which can state j of the plant and updates of the control law k ( j ) com-
contain logic states, timers, counters, observer states, puted by the algorithm are taken after each amount of time T . The
and other continuous-valued and discrete-valued states. controller state z stores the values of k ( j ) .
G1x2 5 c d
xp commercial hard disk drives use mode-switching control
for all x [ D.
Gc 1 x 2
(7)
[27], [81], [87], which combines a track-seeking controller
and a track-following controller. The track-seeking con-
One way hybrid controllers arise is through supervisory troller rapidly steers the magnetic head to a neighborhood
control. A supervisor oversees a collection of controllers and of the desired track, while the track-following controller
makes decisions about which controller to insert into the regulates position and velocity, precisely and robustly, to
closed-loop system based on the state of the plant and the enable read/write operations. Mode-switching control
controllers. The supervisor associates to each controller a uses the track-seeking controller to steer the magnetic
region of operation and a region where switching to other head’s state to a point where the track-following control-
controllers is possible. These regions are subsets of the state ler is applicable, and then switches the control input to
space. In the region where changes between controllers are the track-following controller. The control strategy re-
allowed, the supervisor specifies the controllers to which sults in a hybrid closed-loop system.
authority can be switched. In supervisory control, it is pos- Let p [ R be the position and v [ R the velocity of the
sible for the individual controllers to be hybrid controllers. magnetic head in the disk drive. The dynamics can be ap-
# #
Through this degree of flexibility, it is possible to generate hy- proximated by the double integrator system p 5 v, v 5 u
brid control algorithms through a hierarchy of supervisors. [27], [87].
The following example features a supervisor for two The hybrid controller for the magnetic head super-
state-feedback control laws. vises both the track-seeking control law u 5 k1 1 p, v, p* 2
and the track-following control law u 5 k2 1 p, v, p* 2 ,
Example 1: Dual-Mode Control for Disk Drives where p* is the desired position. We assume that the
Control of read/write heads in hard disk drives requires track-seeking control law globally asymptotically stabi-
precise positioning on and rapid transitioning between lizes the point 1 p*, 0 2 , while the track-following control
tracks on a disk drive. To meet these dual objectives, some law locally asymptotically stabilizes the point 1 p*, 0 2 .
Let C2 be a compact neighborhood of 1 p*, 0 2 that is con-
tained in the basin of attraction for 1 p*, 0 2 when using
u = Kc (xp, xc) xp the track-following control law, and let D1 be a com-
pact neighborhood of 1 p*, 0 2 such that solutions using
Nonlinear
System
the track-following control law that start in D1 do not
Plant State: xp reach the boundary of C2. Also define C1 5 R2\D1 and
(Continuous State)
D2 5 R2\C2 . Figure 7 illustrates these sets.
Let the controller state q [ Q J 5 1, 2 6 denote the op-
Controller State: xc erating mode. The track-seeking mode corresponds to
(Such as Timers and q 5 1, while the track-following mode corresponds to
Discrete States) q 5 2. The mode-switching strategy uses the track-seek-
ing controller when 1 p, v 2 [ C1 and the track-following
controller when 1 p, v 2 [ C2. Figure 7 indicates the inter-
Hc
section of C1 and C2, where either controller can be used.
To prevent chattering between the two controllers in the
intersection of C1 and C2, the supervisor allows mode
switching when 1 p, v 2 [ Dq. In other words, a switch
FIGURE 6 Closed-loop system consisting of a continuous-time from the track-seeking mode to the track-following
nonlinear system and a hybrid controller. The nonlinear system mode can occur when 1 p, v 2 [ D1, while a switch from
has state xp , which is continuous, and input u . The hybrid control-
the track-following mode to the track-seeking mode can
occur when 1 p, v 2 [ D2.
ler has state xc , which has continuous state variables, such as
timer states, and discrete state variables, such as logic modes.
The control input u 5 kc ( xp, xc ) to the nonlinear system is a func- The hybrid controller executing this logic has the flow
tion of the plant state xp and the controller state xc . set Cc 5 h q[Q 1 Cq 3 5 q 6 2 , flow map fc 1 p, v, q 2 5 0, jump set
Flow Condition
For each j [ N such that Ij has nonempty interior,
x
Jump Condition
FIGURE 9 Evolution of a solution to a hybrid system. Flows and jumps For each 1 t, j 2 [ dom x such that 1 t, j 1 1 2 [ dom x,
of the solution x are allowed only on the flow set C and from the jump
set D , respectively. The solid blue curves indicate flow. The dashed
red arcs indicate jumps. The solid curves must belong to the flow set x ( t, j 1 1 ) [ G ( x ( t, j ) ) ,
C . The dashed arcs must originate from the jump set D .
x ( t, j ) [ D.
fixed j, t A x 1 t, j 2 is a locally absolutely continuous func-
tion on the interval If the flow set C is closed and Ij has nonempty interior,
then the requirement x 1 t, j 2 [ C for all t [ 3 min Ij, sup Ij 2
I j 5 5 t : 1 t, j 2 [ dom x 6 . in the flow condition is equivalent to x 1 t, j 2 [ C for
0 3 f ( s (t ) ) 2 f (s ) 4 2 f D (t ) 3 s (t ) 2 s 4 0 # e 0 s (t ) 2 s 0 .
Springer-Verlag, 1988, pp. 1–82.
[S14] G. N. Silva and R. B. Vinter, “Measure driven differential
i n c l u s i o n s ,” J. M at h. A n a l. A p p l i c at., vo l. 2 0 2 , n o. 3 , p p.
7 2 7–746, 1996.
The function f is differentiable if f D exists at every t [ T . A [ S15 ] T. Ya n g , I m p u l s ive C o n t ro l Th e o r y. B e r l i n : S p r i n g e r-
dynamical system on the time scale T has the form Verlag, 2001.
3
x d ( t, j ) 5 ( aj 2 ( t 2 tj ) , t 2 tj )
for j 5 0, 1, 2, c, t [ 3 tj, tj11 4 .
D
2
1
Finally, the maximal solution starting from 1 0, 0 2 , denoted
3 x1 by xe, never flows, has dom xe 5 5 0 6 3 N and is given by
xe 1 0, j 2 5 1 0, 0 2 for all j [ N.
The solutions xa, xb, xc, xd, and xe are complete. Note
that solutions xd and xe are complete even though dom xd
and dom xe do not contain points 1 t, j 2 with arbitrarily
large t. ■
We now consider the concept of graphical conver- which the jump occurs depends on the initial height.
gence of hybrid arcs along with the related concept of In fact, for t # 2, the hybrid time domain is given by
distance, which focus not just on the values of the hy- 3 0, "2 1 1 1 d 2 4 3 5 0 6 h 3 "2 1 1 1 d 2 , 2 4 3 5 1 6 .
brid arcs but on their graphs. One benefit of this ap- The graph of a hybrid arc x is the set
proach is that different hybrid time domains can be
handled easily. Note that bouncing balls dropped from gph x 5 5 1 t, j, j 2 : 1 t, j 2 [ dom x, j 5 x 1 t, j 2 6 .
different initial heights lead to different hybrid time
domains of the hybrid arcs representing their heights The sequence 5 xi 6 i51
`
of hybrid arcs converges graphically
and velocities. For example, the hybrid time domain of if the sequence 5 gph xi 6 i51
`
of graphs converges in the
the trajectory of the bouncing ball dropped from height sense of set convergence; see “Set Convergence.” The
1 1 d, where d [ 3 0, 1 2 , depends on d since the time at graphical limit of a graphically convergent sequence
The first conclusion of Theorem 4 is a property of set Given t $ 0 and e . 0, the hybrid arcs x and y are 1 t, e 2 -
convergence. The second conclusion is specific to the close if the following conditions are satisfied:
hybrid system setting. In summary, Theorem 4 states a) For each 1 t, j 2 [ dom x with t 1 j # t there exists
that, from each locally uniformly bounded sequence s [ R $0 such that 1 s, j 2 [ dom y, |t 2 s| , e, and
|x 1 t, j 2 2 y 1 s, j 2 | , e.
of solutions to a hybrid system that satisfies the Basic
Assumptions, we can extract a graphically convergent
subsequence whose graphical limit is a solution to the b) For each 1 t, j 2 [ dom y with t 1 j # t there exists
hybrid system. Consequences of Theorem 4 for asymp- s [ R $0 such that 1 s, j 2 [ dom x, |t 2 s| , e, and
|y 1 t, j 2 2 x 1 s, j 2 | , e.
totic stability in hybrid systems are discussed in the “As-
ymptotic Stability” section.
Graphical convergence of hybrid arcs to a hybrid arc The concept of 1 t, e 2 -closeness provides an equivalent
has an equivalent pointwise description. Consider a se- characterization of graphical convergence of hybrid arcs.
quence 5 xi 6 i51`
of hybrid arcs and a hybrid arc x. Then 5 xi 6 i51 `
Consider a locally uniformly bounded sequence of hybrid
converges graphically to x if and only if the following two arcs 5 xi 6 i51
`
and a hybrid arc x. Then the sequence 5 xi 6 i51
`
t [ 3 0, "2 1 1 1 d 2 4 , j 5 0,
x d2 1 t, j 2 5 e
Dependence of Solutions on Initial Conditions 2 t,
One of the consequences of Theorem 4 is the semicontinu- 2 t 1 3" 1 1 1 d 2 /2, t [ 3 "2 1 1 1 d 2 , 2 4 , j 5 1.
ous dependence of solutions to a hybrid system on initial
conditions. To state this result rigorously, in Theorem 5, Consider the hybrid arcs, x02 and xd2, where d [ 1 0, 1 2 .
we define a concept of distance between hybrid arcs that is These arcs are 1 t, e 2 -close, with any t $ 0 and
closely related to graphical convergence. e 5 3"2 1 "1 1 d 2 1 2 /2. To show this, it is sufficient to
consider t 5 3. Indeed, since we are considering arcs with is not a limit of any sequence of solutions from positive ini-
domains restricted to 1 t, j 2 such that t 1 j # 3, 1 3, e 2 -close- tial points. This simple setting already illustrates what outer
ness implies 1 t, e 2 -closeness for all t $ 0. semicontinuous dependence on initial conditions is, namely,
To verify condition a) of 1 t, e 2 -closeness, note that, for
each 1 t, 0 2 [ dom x02 with t # 3, we have 1 t, 0 2 [ dom xd2
and x02 1 t, 0 2 5 xd2 1 t, 0 2 . In practical terms, the velocities
of the balls are the same until the ball dropped from
the lower height bounces. For each 1 t, 1 2 [ dom x02 with
t # 2, that is, with t 1 1 # 3, there exists 1 s, 1 2 [ dom xd2
with |t 2 s| # "2 1 1 1 d 2 2 "2 and |x02 1 t, j 2 2 xd2 1 s, j 2 | # 10
3"2 1 "1 1 d 2 1 2 /2. In fact, for t # "2 1 1 1 d 2 , we can
take s 5 "2 1 1 1 d 2 , while, for the remaining t’s, we can
8
1 t, j 2 [ dom x, t 1 j # m 6
Velocities parameterized by t. The velocities are not close in the uniform distance
since their difference, at times after the first ball bounces and before the second ball
bounces, is large. (b) Velocities on hybrid time domains. The shaded neighborhoods
of their graphs indicate that the velocities are graphically close. is compact.
Cds J 5 x: 1 x 1 ds 1 x 2 B 2 d C 2 [ 6 , (8) S S
Fds 1 x 2 J con F 1 1 x 1 ds 1 x 2 B 2 d C 2 1 ds 1 x 2 B Spσ
for all x [ Cds, (9)
Dds J 5 x: 1 x 1 ds 1 x 2 B 2 d D 2 [ 6 , (10) FIGURE 15 Enlargement of a set S due to a state-dependent per-
Gds 1 x 2 J 5 v: v [ g 1 ds 1 g 2 B, g [ G 1 1 x 1 ds 1 x 2 B 2 d D 2 6 turbation of size ps . The perturbed set Sps contains all points x
for all x [ Dds. with Euclidean distance from the unperturbed set S no larger than
ps 1 x 2 .
(11)
p 5 v, v 5 k 1 p, v, p * 2 , 1 p, v 2 [ C.
# #
(12) According to the assumptions of Example 1 and the conclu-
sions drawn above about the system (12), for each i [ 5 1, 2 6
Assume that if C 5 R2 then the point 1 p *, 0 2 is locally asymp- the point 1 p *, 0 2 is globally pre-asymptotically stable for the
totically stable with basin of attraction B . Now let C ( B and system
notice that this choice eliminates each solution of (12) with
p 5 v, v 5 ki 1 p, v, p * 2 1 p, v 2 [ Ci.
# #
C 5 R2 that does not start in B . It follows, for the system (12) (15)
with C ( B , that the point 1 p *, 0 2 is globally pre-asymptoti-
cally stable. A generalization of this observation appears in Moreover, each solution of (15) with i 5 2 that starts in D1,
Theorem S10 of “Why ‘Pre’-Asymptotic Stability?” which contains a neighborhood of 1 p *, 0 2 and is contained
Now consider the hybrid system from Example 1, with in C2, does not reach the boundary of C2. Also, C1 5 R2\ D1
state x 5 1 p, v, q 2 satisfying and D2 5 R2\ C2 .
Global asymptotic stability of the point 1 p *, 0, 2 2 for the
p 5 v, v 5 kq 1 p, v, p * 2 , q 5 0, 1 p, v 2 [ Cq,
# # #
(13) system (13), (14) then follows from the global pre-asymptotic
p 1 5 p, v 1 5 v, q 1 5 3 2 q, 1 p, v 2 [ Dq. (14) stability of 1 p *, 0 2 for (15) with i [ 5 1, 2 6 together with the
3 G1 G2 4 J 1 I 2 exp 1 F2d 2 exp 1 2F1d 2 2 21 This theorem helps in developing some concepts used to
3 3 2 exp 1 F2d 2 exp 1 2F1d 2 I 4 . express the fact that pre-asymptotic stability is equivalent
to uniform pre-asymptotic stability on compact subsets of
the basin of pre-attraction. In that direction, the next result
This hybrid system contains # two different continuous-time states that excursions away from and convergence toward
observers, of the form j^i 5 1 F 2 L i H 2 j^i 1 L i Hj 1 v, the a pre-asymptotically stable compact set are uniform over
states of which make jumps every d seconds according to compact subsets of the basin of pre-attraction.
the rule specified by the jump map.
We show that the compact set A J 5 1 j, j^ 1, j^ 2 2 [ Theorem 13 [26, Prop. 6.3]
K1 3 Rn 3 Rn: j 5 j^ 1 5 j^ 2 6 3 3 0, d 4 is globally asymp- For a hybrid system H satisfying the Basic Assumptions, let
totically stable using Theorem 10. The set A is forward the compact set A be pre-asymptotically stable with basin of
invariant since G 1 A d D 2 ( A and, during flows, the pre-attraction BA. For each compact set K0 ( BA, the compact
#
errors ei J j 2 j^ i satisfy ei 5 Fiei. Moreover, the set set A is uniformly pre-attractive from K0, and there exists a
A is globally uniformly attractive. In particular, compact set K1 ( BA such that solutions starting in K0 satisfy
1 t, j 2 [ dom x and t $ 2d imply j $ 2 and x 1 t, j 2 [ A. x 1 t, j 2 [ K1 for all 1 t, j 2 [ dom x.
The condition on j follows from the nature of the data The properties established in Theorem 13, of uniform
of the hybrid system. The condition x 1 t, j 2 [ A follows overshoot and uniform convergence from compact sub-
from the fact that, when there exists w such that j, j^ 1 , sets of the basin of pre-attraction, can be expressed in
and j^ 2 satisfy terms of a single bound on solutions starting in the ba-
sin of pre-attraction. To that end, let A be compact and
j^ 1 5 j 1 exp 1 F1d 2 w, j^ 2 5 j 1 exp 1 F2d 2 w, let O be an open set containing A. A continuous func-
tion v : O S R $0 is called a proper indicator for A on O if
the jump map sends the state to A. The given relations v 1 x 2 5 0 if and only if x [ A, and also v 1 xi 2 tends to infin-
are satisfied after one jump followed by a flow interval of ity when xi tends to infinity or tends to the boundary of O.
length d. In this case, w is equal to the difference between Every open set O and compact set A ( O admit a proper
j and j^ 1 immediately after the jump. We conclude from indicator. Thus, using Theorem 12, for each pre-asymp-
Theorem 10 that the set A is globally asymptotically stable. totically stable set A there exists a proper indicator for A
Moreover, the analysis above shows that the convergence to on its basin of pre-attraction. The function v : Rn S R $0
A is in finite time t # 2d. ■ defined by v 1 x 2 J |x|A for all x [ Rn is a proper indi-
cator for A on Rn. For a general open set O, it is always
Equivalence with Uniform Asymptotic Stability possible to take v 1 x 2 5 |x|A for x sufficiently close to A.
In this section, we describe uniform pre-asymptotic sta- The concept of a proper indicator function first appears
bility on compact subsets of the basin of pre-attraction in conjunction with stability theory in [42]. A typical for-
and point out that this property is equivalent to pre- mula for a proper indicator is given in that work and in
asymptotic stability. The first step in stating this char- [39, (C.14)]. A sublevel set of a proper indicator on O is a
acterization is to make an observation about the compact subset of O.
basin of pre-attraction that extends classical results [42, A function b : R $0 3 R $0 S R $0 belongs to class-KL if
pp. 69–71] for differential and difference equations to the it is continuous; for each s $ 0, r A b ( r, s ) is nondecreas-
hybrid setting. This result, and all of the subsequent re- ing and zero at zero; and, for each r $ 0, s A b ( r, s ) is
sults in this section, depend on Theorem 4. nonincreasing and tends to zero when s tends to infinity.
Class-KL functions are featured prominently in [31] and
Theorem 12 [26, Prop. 6.4], [14, Thm. 3.14] have become familiar to the nonlinear systems and control
For a hybrid system H satisfying the Basic Assumptions, the community through their use in the input-to-state stability
basin of pre-attraction for a compact, pre-asymptotically property [76]. The next result is a generalization to hybrid
stable set A is an open, forward invariant set containing a systems of a result contained in [42] for continuous differ-
neighborhood of A. ential equations.
W hen designing control systems, engineers must ensure tions [S26], [S27], [S25]. The lack of robustness motivates
that the closed-loop behavior is robust to reasonable lev- generalized notions of solutions. For differential equations,
els of measurement noise, plant uncertainty, and environmental [S20] establishes the connection between the generalized
disturbances. notion of solution in [S26], expressed in terms of small state
For linear systems, eigenvalues of matrices change continu- perturbations, and the generalized notion of solution given in
ously with parameters, and asymptotic stability of an equilibrium [S28], expressed in terms of set-valued dynamics. The set-val-
point is an open-set condition on eigenvalues. It follows that suf- ued dynamics arise when a discontinuous right-hand side of a
ficiently small perturbations to the system matrices do not change differential, or difference, equation is converted into an inclu-
stability properties. Similarly, sufficiently small additive disturbanc- sion. The resulting system is the regularization of the original
es lead to small excursions from the equilibrium point. system, that is, it is a regularized system.
Classical results for ordinary differential and difference equa- The conversion of a differential, or difference, equation
tions with nonlinear but continuous right-hand sides establish that with a discontinuous right-hand side to an inclusion is done
small perturbations to the right-hand side result in small changes by considering a closure of the graph of the discontinuous
to the solutions on compact time intervals. This property is de- right-hand side and, in the differential equation case, by tak-
scribed for hybrid systems in theorems 5 and 8. If an equilibrium ing the convex hull of the values of the right-hand side. For
point or, more generally, a compact set is asymptotically stable an illustration, see Figure S8. The price paid for such a con-
for the nominal system, then these compact time intervals can be version is the introduction of additional solutions, some of
stitched together to establish that sufficiently small disturbances which may not behave well. However, these extra solutions
lead to small excursions from the asymptotically stable compact are meaningful since they arise from arbitrarily small state
set. This idea is behind the results in theorems 15 and 17. This perturbations that converge to zero asymptotically. More-
behavior is the essence of the total stability property described in over, asymptotic stability in the original, possibly discontinu-
[31, Sec. 56], and also contained in the local version of the input- ous, system is robust if and only if asymptotic stability holds
to-state stability property [76]. for the regularized system.
For differential equations and difference equations with a Anticipating that a similar result holds for hybrid systems
discontinuous right-hand side, asymptotic stability of a com- and desiring robustness in asymptotically stable hybrid sys-
pact set is not necessarily robust to arbitrary small perturba- tems, we consider the stability properties of regularized sys-
tems. Following the lead of discontinuous
continuous-time and discrete-time sys-
f (x) F1(x) tems, we convert discontinuous flow and
jump equations to inclusions. Interpreting
2 2 these inclusions as closure operations
on the discontinuous flow map and jump
1 1 map as Figure S8 indicates, we also take
the closures of the flow and jump sets.
As for continuous- and discrete-time sys-
0 tems, these operations may introduce
x x 0 x x
new solutions. However, the regularized
(a) (b)
hybrid system satisfies the Basic As-
F2(x)
sumptions, and therefore, the results of
2 the main text are applicable. In particular,
if the regularized system has an asymp-
totically stable compact set A, then A
1
is robustly asymptotically stable for the
regularized system, and consequently, A
is robustly asymptotically stable for the
0 x x original system.
(c)
Example S6: A Frictionless Ball
FIGURE S8 Regularization of a discontinuous function. (a) The function f : R S R has a and Two Rooms Separated by
discontinuity at x 5 x *. (b) The set-valued mapping F1 : R S S R is the mapping whose
graph is the closure of the graph of f . Therefore, F1 1 x* 2 is the set 5 1, 2 6 . (c) The set-
a Zero-Width Wall
valued mapping F2 : R S S R is the mapping obtained by taking the convex hull of the Consider a particle that moves, with no
values of the mapping in (b), at each x [ R. Therefore, F2 1 x* 2 is the interval 3 1, 2 4 . friction, in one of two rooms separated by
C J 5 ( j, q ) [ R 3 52 1, 1 6 : 2jq $ 2 1 6 ,
havior while using closed flow and jump sets. One natural ap-
proach can be interpreted as thickening the wall between the
two rooms. In particular, letting e [ 1 0, 1 2 , the data F ( j, q ) J c
2j1q
d for all ( j, q ) [ C,
0
results in solutions, as defined in the main text, rendering each This system can be associated with a hybrid control algorithm
##
of the sets 3 21, 2e 4 3 5 21, 1 6 and 3 e, 1 4 3 5 21, 1 6 robustly for the continuous-time, linear system j 5 u , which uses an
forward invariant. internal state q [ 5 21, 1 6 and aims to globally asymptotically
When we insist on a wall of zero width, an extra state vari- stabilize the two point set A J 5 1 21, 21 2 , 1 1, 1 2 6 . This set
able should be added that gives the system information about is globally asymptotically stable since A is contained in C ,
how it arrived at the point 1 0, q 2 , whether from the room on the the flow stabilizes the point j 5 q , A is disjoint from D, the
left or the one on the right. For example, we can consider the quantity 2jq increases along flows when it has the value 2 1,
hybrid system with state 1 j, q, r 2 [ R3 with data and jumps from D are mapped to points in C . Note, however,
that the jump set D is not closed. Consider the effect of clos-
C 5 5 1 j, q, r 2 [ 3 21, 1 4 3 5 21, 1 6 3 5 21, 1 6 : jr $ 0 6 , ing D . For the hybrid system with data 1 C, F, D, G 2 , solutions
F x 2 5 1 q, 0, 0 2 for all x [ C,
1 starting from points where 2x1x2 5 21 are no longer unique.
D 5 5 1 21, 21, 21 2 6 h 5 1 1, 1, 1 2 6 From such points, flowing is possible, as is a single jump fol-
h 5 1 0, 1, 21 2 6 h 5 1 0, 21, 1 2 6 , lowed by flow. Nevertheless, the set A is still globally asymp-
G 1 x 2 5 1 j, 2q, r 2 for all x [ D. totically stable for 1 C, F, D, G 2 . Thus, the asymptotic stability
of A in the original system is robust. The lack of unique solu-
The hybrid system with this data is such that the set C is for- tions in the regularized system can be associated with the fact
ward invariant. In particular, since r is constant and belongs to that very different solutions arise when starting from 2jq 5 2 1
the set 5 21, 1 6 , the state j cannot change sign. ■ and considering arbitrarily small measurement noise on the
The next example is discussed in [14]. state j in the closed-loop control system. ■
Theorem 14 [26, Thm. 6.5], [14, Prop. 7.3] and show that, when the perturbations are small enough,
For a hybrid system H satisfying the Basic Assumptions, the pre-asymptotic stability of A and the basin of pre-
if the compact set A is pre-asymptotically stable with attraction are preserved, as in Theorem 15. Typically,
basin of pre-attraction given by BA then, for each func- these state-dependent perturbations must decrease in size
tion v that is a proper indicator for A on BA, there exists as the state approaches the pre-asymptotically stable set
b [ KL such that each solution x starting in BA satisfies and also as the state approaches the boundary of the basin
v 1 x 1 t, j 2 2 # b 1 v 1 x 1 0, 0 2 2 , t 1 j 2 for all 1 t, j 2 [ dom x. of pre-attraction.
Theorem 14 contains the result in Theorem 13 that Another way to characterize robustness is to consider
excursions away from and convergence toward a pre- constant perturbation levels and show that these pertur-
asymptotically stable compact set are uniform over com- bations lead to “practical” pre-asymptotic stability from
pact subsets of the basin of pre-attraction. In other words, arbitrarily large subsets of the basin of attraction, as in
pre-asymptotic stability for compact sets is equivalent Theorem 17.
to uniform pre-asymptotic stability, which sometimes is In some cases the nominal system can tolerate state-
called KL-stability. Thus, when we provide sufficient con- dependent perturbations that grow without bound when
ditions for pre-asymptotic stability, we in fact give suffi- the state grows unbounded. These systems are closely re-
cient conditions for KL-stability. lated to hybrid systems having inputs and possessing the
When discussing global pre-asymptotic stability, where input-to-state stability (ISS) property [76]. ISS for hybrid dy-
BA 5 Rn, we can use v 1 x 2 5 |x|A. Thus, for global pre- namical systems is studied in [13].
asymptotic stability, the bound in Theorem 14 becomes For a given hybrid system H with data 1 C, F, D, G 2 and a
|x 1 t, j 2 |A # b 1 |x 1 0, 0 2 |A, t 1 j 2 for all 1 t, j 2 [ dom x. continuous function s : Rn S R $0 , we define the s-pertur-
When v 1 x 2 5 |x|A and b ( s, r ) 5 gs exp ( 2lr ) for some bation Hs of H through the data
positive real numbers g and l, the set A is globally pre-
exponentially stable. Cs J 5 x : ( x 1 s ( x ) B ) d C 2 [ 6 , (16)
Fs ( x ) J conF ( ( x1s ( x ) B ) d C ) 1 s ( x ) B
Robustness for all x [ Cs, (17)
Ds J 5 x : ( x 1 s ( x ) B ) d D 2 [ 6 ,
A feature of hybrid systems satisfying the Basic Assump-
(18)
tions is that pre-asymptotic stability is robust.
One way to characterize the robustness of pre-asymp- Gs 1 x 2 J 5 v : v [ g 1 s 1 g 2 B,
totic stability of a compact set A is to study the effect of g [ G1 1x 1 s1x2B2 x D2 6
state-dependent perturbations on the hybrid system data for all x [ Ds. (19)
b) there exist hybrid arcs xi and functions ei : dom xi S Rn, laws uc 5 kc 1 xi 1 ei 2 and ud 5 kd 1 xi 1 ei 2 , and poses stronger
i [ 5 1, 2, c 6 , such that limi S ` xi 1 0, 0 2 5 x 1 0, 0 2 , continuity assumptions on the functions fr and gr, but allows
the sequence 5 xi 6 i`5 1 converges graphically to x , the functions kc and kd to be discontinuous.
limi S ` sup1t, j2 [ dom xi |ei 1 t, j 2 | 5 0 , and, fo r eve r y
i [ 5 1, 2, c 6 the following hold:
– For each fixed j , t Aei 1 t, j 2 is measurable.
REFERENCES
– For all j [ N such that Ii, j J 5 t : 1 t, j 2 [ dom xi 6 has
[S25] G. Grimm, M. J. Messina, S. E. Tuna, and A. R. Teel, “Examples
when nonlinear model predictive control is nonrobust,” Automatica, vol.
nonempty interior, 40, no. 10, pp. 1729–1738, Oct. 2004.
[S26] H. Hermes, “Discontinuous vector fields and feedback control,”
xi 1 t, j 1 1 2 5 gi 1 xi 1 t, j 2 1 ei 1 t, j 2 2 ,
[S29] R. G. Sanfelice, R. Goebel, and A. R. Teel, “Generalized solutions
xi 1 t, j 2 1 ei 1 t, j 2 [ D.
to hybrid dynamical systems,” ESAIM: Control, Optim. Calculus Variat.,
vol. 14, no. 4, pp. 699–724, 2008.
These definitions match the definitions (8)–(11) with tinuous, there exists a continuous function s | such that
d 5 1. Figure 15 illustrates the idea behind perturbations of | 1 0, x 2 5 0 for all x [ Rn1m and, for all e [ Rn satisfying
s
the sets C and D. Note that C ( Cs, D ( Ds, F 1 x 2 ( Fs 1 x 2 for |e| # M and all x 5 1 xp, xc 2 [ Rn1m,
all x [ C and G 1 x 2 ( Gs 1 x 2 for all x [ D. Since the data of
Hs contain the data of H, the solutions of H are also solu- |fp 1 xp, kc 1 xp 1 e, xc 2 2 2 fp 1 xp 1 e, kc 1 xp 1 e, xc 2 2 |
tions to Hs. On the other hand, Hs typically exhibits solu- # s | ( M, x ) .
tions that are not solutions to H.
The extra solutions of Hs can be linked to solutions In this case, the hybrid control system (20)–(21) can be writ-
that arise due to parameter variations, measurement ten as
noise in control systems, and external disturbances. For
#
a link to solutions that arise from parameter variations, x 5 F ( x 1 d 1 ) 1 d 2, x 1 d1 [ C,
see Example 18. For the case of external disturbances x 1 [ G ( x 1 d1 ) , x 1 d1 [ D,
d that are bounded in norm by a value M . 0, observe
that the solutions of x 5 F 1 x 2 1 d are contained in the
#
where F and G are defined in (6) and (7), respectively,
set of solutions of x [ F 1 x 2 1 MB and, taking s 1 x 2 5 M | 1 M, x 2 . Therefore, the solutions of
#
|d1| # M, and |d2| # s
for all x, F 1 x 2 1 MB ( Fs 1 x 2 . The same link between ex- the hybrid control system (20)–(21) are contained in the solu-
ternal disturbances and system perturbations holds for tions of 1 Cs, Fs, Ds, Gs 2 where s ( x ) :5 max 5 M, s | ( M, x ) 6
x 1 5 G 1 x 2 1 d. for all x [ Rn1m.
Now consider the case of measurement noise in a hybrid The mappings Fs and Gs may be set valued at points
control system of the form discussed in the section “Hybrid x where s 1 x 2 . 0, even when F and G are single-valued
Controllers for Nonlinear Systems, ” given as mappings. Also, when x [ C d D and s 1 x 2 . 0, the point
x belongs to the interior of both Cs and Ds. Thus, at such
# points, the system Hs has solutions that initially flow and
xp 5 fp ( x p, kc ( x p 1 e, x c ) )
# r ( x p 1 e, x c ) [ C, (20) also solutions that initially jump.
xc 5 fc ( x p 1 e, x c )
For a hybrid system H having a compact set A that is pre-
x p1 5 x p asymptotically stable with basin of pre-attraction BA, Theo-
r ( x p 1 e, x c ) [ D, (21)
x c1 [ Gc ( x p 1 e, x c ) rem 15 below asserts the existence of a continuous function
s that is positive on BA\A so that, for the hybrid system Hs,
where e represents measurement noise, assumed to be the compact set A is pre-asymptotically stable with basin
bounded in norm by M . 0. When the function fp is con- of pre-attraction BA. In other words, pre-asymptotic stability
I ntuition built on the theory of linear systems conditions us to asymptotically stable. In particular, each solution to the hybrid
think of asymptotic stability as a property of an equilibrium system converges to the compact set A. This property cap-
point. However, it is also natural to consider the asymptotic tures the fact that the state j converges to zero.
stability of a set. Conceptually, set stability places many ap- As another example of set stability, consider a system switch-
parently different phenomena under one umbrella. Set stability ing among a family of asymptotically stable linear systems under
6
is needed for systems that include timers, counters, and other an average dwell-time constraint, as discussed in “Switching
discrete states that do not converge. In addition, set stability is Systems.” The corresponding hybrid system has the form
helpful for characterizing systems that exhibit complicated as- #
j 5 Fq j
C J Rn 3 Q 3 3 0, N 4 ,
ymptotic behavior. #
6
q 5 0
t [ 3 0, d 4
Asymptotic stability of a set is defined in the main text. #
Roughly speaking, a set is globally asymptotically stable if each
solution that starts close to the set remains close to the set, j1 5 j
each solution is bounded, and each solution with an unbounded q1 [ Q D J Rn 3 Q 3 3 1, N 4 ,
time domain converges to the set. t1 5 t 2 1
To illustrate set stability, we consider a linear, sampled-data
where Q ( R is compact and the eigenvalues of Fq have negative
control system. We group the plant state, hold state, and con-
real part for each q [ Q . When the average dwell-time param-
troller state into the state j [ Rn . Between sampling events,
eter d . 0 is small enough, the compact set 5 0 6 3 Q 3 3 0, N 4
the state j evolves according to a closed-loop, linear differen-
#. is asymptotically stable. In particular, the state j converges to
tial equation j 5 Fj for some matrix F . For each state com-
zero. The states q and t may or may not converge, depending
ponent that remains constant between sampling events, the
on the particular solution considered.
corresponding row of F is filled with zeros. At sampling events,
In some situations, hybrid systems admit complicated,
the state j is updated according to the equation j 1 5 Jj for
asymptotically stable compact sets that can be characterized
some matrix J . For each state component that does not change
through the concept of an V -limit set [S33]. V -limit sets are
at sampling events, the corresponding row of J is filled with
exploited for nonlinear control design, for example for nonlin-
zeros except for a one in the appropriate column. The interac-
#. ear output regulation [S30]. Results pertaining to V -limit sets
tion between the continuous evolution j 5 F j and the discrete
for hybrid systems are given in [S32] and [S31]. In [S32, Cor.
evolution j 1 5 J j is scheduled by a timer state t that evolves
# 3] and [S31, Thm. 1], conditions are given under which the V
according to t 5 1 between sampling events and is reset to zero
-limit set is an asymptotically stable compact set. In particular,
at sampling events. Sampling events occur whenever the timer
suppose there exist T . 0 and compact sets K0 and KT , with
state reaches the value T . 0, which denotes the sampling pe-
KT contained in the interior of K0 , such that each solution x
starting in K0 is bounded and, for all 1 t, j 2 [ dom x satisfying
riod of the system. Thus, the sampled-data control system can
be written as a hybrid system with state x 5 1 j, t 2 satisfying
t 1 j $ T , we have x 1 t, j 2 [ KT . Then, either the V -limit set
.#
x [ C J Rn 3 3 0, T 4 ,
j 5 Fj from K0 is empty or it is the smallest compact, asymptotically
# f stable set contained in KT .
t 5 1
j 1 5 Jj
f x [ D J Rn 3 5 T 6 . REFERENCES
t1 5 0
[S30] C. I. Byrnes and A. Isidori, “Limit sets, zero dynamics, and inter-
nal models in the problem of nonlinear output regulation,” IEEE Trans.
The matrices F and J are designed so that j converges to Automat. Contr., vol. 48, no. 10, pp. 1712–1723, 2003.
zero. However, the timer state t does not converge. Letting [S31] C. Cai, R. Goebel, R. G. Sanfelice, and A. R. Teel, “Complex
t0 [ 3 0, T 4 denote the initial value of t, the solution satis- hybrid systems: Stability analysis for Omega limit sets,” in Proc. of the
26th Chinese Control Conf., Zhangjiajie, China, July 2007, pp. 766–769.
fies t ( t, j ) 5 t 2 ( jT 2 t0 ) for all positive integers j and all [S32] C. Cai, R. Goebel, R. G. Sanfelice, and A. R. Teel, “Hybrid sys-
t [ 3 jT 2 t0, ( j 1 1 ) T 2 t0 4 . Consequently, t revisits every point tems: Limit sets and zero dynamics with a view toward output regula-
in the interval 3 0, T 4 . For this reason, the hybrid system does tion,” in Analysis and Design of Nonlinear Control Systems, A. Astolfi
and L. Marconi, Eds. New York: Springer-Verlag, 2008, pp. 241–261.
not possess an asymptotically stable equilibrium point. On the
other hand, the compact set A J 5 0 6 3 3 0, T 4 is globally
[S33] J. K. Hale, L. T. Magalhães, and W. M. Oliva, Dynamics in Infinite
Dimensions, 2nd ed. New York: Springer-Verlag, 2002.
and the corresponding basin of pre-attraction are preserved Theorem 15 [14, Theorem 7.9]
under the perturbation s. This result provides a key step in For the hybrid system H satisfying the Basic Assumptions,
the proof of the existence of a smooth Lyapunov function for suppose that a compact set A ( Rn is pre-asymptotically
a hybrid system with a pre-asymptotically stable compact set. stable with basin of pre-attraction BA. Then there exists a
For more information, see “Converse Lyapunov Theorems.” continuous function s : Rn S R $0 satisfying s 1 x 2 . 0 for
all x [ BA\A such that, for the hybrid system Hs, the com- from sensing that the state deviates from the state used to
pact set A is pre-asymptotically stable with basin of pre- compute the control value by an amount that is significant
attraction BA. enough to warrant a change in the input value. This amount
The idea in the next example is related to Lebesgue sam- is determined by a continuous function s1 : Rn S R $0 re-
pling presented in [3] and to the ideas in [80] used for con- lated to the function s in Theorem 15.
trol in the presence of communication or power constraints. The closed-loop system can be modeled as a hybrid sys-
The idea exploits the robustness described in Theorem 15, tem with the data
but for the special case of classical systems, to explain the
asymptotic stability induced by a particular hybrid control C J 5 x [ R2n : |x 1 2 x 2| # s1 ( x 1 ) 6 ,
strategy.
F1x2 J c
f (x1, k(x2 ))
d for all x [ C,
Example 16: Control Through 0
Event-Based Hold Updates D J 5 x [ R2n :|x 1 2 x 2| $ s1 ( x 1 ) 6 ,
Consider the continuous-time, nonlinear control system
#
j 5 f ( j, u ) , 1 j, u 2 [ Rn 3 Rm, with the continuous state G1x2 J c
x1
d for all x [ D,
feedback k : Rn S Rm that stabilizes a compact set A with x1
basin of attraction BA ( Rn. Suppose the feedback is imple-
mented by keeping the control value constant until an event where s1 is specified as follows. Let s : Rn S R $0 satisfy
triggers a change in the control value. The trigger comes the conditions of Theorem 15 for the (classical) system
D espite the Lyapunov-based sufficient conditions for pre- logic-based input-to-state stabilization with respect to matched
asymptotic stability given by Theorem 32 and [6, 19, 90], it disturbances. Converse Lyapunov theorems can also be used
is a mistake to conclude that Lyapunov functions do not exist for to establish various forms of robustness, including robustness
asymptotically stable hybrid systems. Early results on the exis- to small, persistent perturbations.
tence of Lyapunov functions for hybrid systems can be found in The converse theorems for hybrid systems given in [14] draw
[90]. More recently, [14] establishes that a hybrid system with a heavily from the literature on converse theorems for continuous-
pre-asymptotically stable compact set admits a smooth ( C ` ) time and discrete-time systems. We recount some of the major
Lyapunov function as long as the hybrid system satisfies the milestones in the development of converse Lyapunov theorems
Basic Assumptions. Converse theorems of this type are typical- for time-invariant, finite-dimensional dynamical systems having
ly established theoretically, constructing smooth Lyapunov func- compact, asymptotically stable sets.
tions from the system’s solutions, which are not usually available In his 1892 Ph.D. dissertation, in particular [S38, Section 20,
explicitly. Thus, converse theorems are of limited help in con- Theorem II], Lyapunov provides the first contribution to converse
structing Lyapunov functions. Nevertheless, converse theorems theorems, where he addresses asymptotically stable linear sys-
justify searching for Lyapunov functions that can be constructed tems. Generalizations of this result to nonlinear systems did not
without knowledge of the system’s trajectories. appear until the 1940s and 1950s. For example, the 1949 paper
Converse theorems also play a role in establishing stabili- [S39] provides a converse Lyapunov theorem for time-invariant,
zation results for nonlinear control systems. For example, early continuously differentiable systems having a locally asymptoti-
results [S40, Theorem 4] and [S34, Theorem 2.1] on backstep- cally stable equilibrium.
ping, that is, proving that smooth stabilizability is not destroyed Essentially all of the converse Lyapunov theorems until [42]
by adding an integrator, construct a stabilizing feedback by pertain only to dynamical systems with unique solutions. In con-
using the gradient of a Lyapunov function whose existence is trast, [42] establishes the first converse Lyapunov theorems for
guaranteed by a converse theorem. Also, the result in [76] show- differential equations with continuous right-hand side without as-
ing that smooth stabilization implies coprime factorization, that suming uniqueness of solutions. This extension is significant in
is, input-to-state stabilization with respect to disturbances that anticipation of hybrid systems where nonuniqueness of solutions
add to the control variable, exploits a converse Lyapunov theo- is not uncommon. The contribution of [42] to the development of
rem. Similar results apply for hybrid systems. For example, in converse theorems is immense; it appears to be the first work
[14] converse theorems are developed and used to show that that relies explicitly on robustness and advanced smoothing tech-
smooth stabilization with logic-based feedback implies smooth, niques to establish the existence of smooth Lyapunov functions.
#
j 5 F 1 j 2 J f 1 j, k 1 j 2 2 . Take s1 : Rn S R $0 to be continu-
|
does x2. It thus follows for the closed-loop system that the
ous, positive on BA\A, zero on A, such that s1 1 x1 2 # s 1 x1 2 set A 3 A is asymptotically stable with basin of attraction
for all x1 [ BA, and BA 3 Rn. ■
For problems where the size of the perturbations does
|f 1 x1, k 1 x2 2 2 2 f 1 x2, k 1 x2 2 2 | # s 1 x1 2 not become arbitrarily small as the state approaches the set
for all x1 [ BA, x2 [ x1 1 s1 1 x1 2 B. A or the boundary of the basin of pre-attraction, the follow-
ing theorem is relevant.
For all x [ C, it follows that
Theorem 17 [26, Theorem 6.6]
f ( x 1, k ( x 2 ) ) 5 f ( x 2, k ( x 2 ) ) 1 f ( x 1, k ( x 2 ) ) 2 f ( x 2, k ( x 2 ) ) For a hybrid system H satisfying the Basic Assump-
tions, suppose that a compact set A is pre-asymptotical-
F 1 x 1 1 s 1 x 1 2 B 2 1 s 1 x 1 2 B.
[ con | ly stable with basin of pre-attraction BA . In particular,
suppose that there exist b [ KL and a proper indica-
The jumps do not change x1. Thus, according to Theorem tor function v for A on BA such that, for all solutions
15, x1 stays close to A and converges to A whenever the starting in BA ,
solution has a hybrid time domain that is unbounded in
the t direction. For such solutions, x2 also stays close to v 1 x 1 t, j 2 2 # b 1 v 1 x 1 0, 0 2 2 , t 1 j 2 for all 1 t, j 2 [ dom x.
A and converges to A. For solutions with a hybrid time
domain bounded in the t direction, notice that when x [ D Then, for each e . 0 and compact set K ( BA, there exists
but x1 o A, it follows that G 1 x 2 o D. Thus, the only way d . 0 such that, with the function s defined as s 1 x 2 5 d for
that a solution can have a domain bounded in the t direc- all x [ Rn, each solution to the hybrid system Hs starting
tion is if x1 converges to A. Whenever x1 converges to A, so in K satisfies
x 5 F 1 x 2 , x [ C 5 MB 3 MB, M . 0,
#
(24) Theorem 20
Consider the hybrid system H 5 1 C, F, D, G 2 satisfying the
it follows that the compact set A1 5 MB 3 5 0 6 is globally Basic Assumptions and the compact set A ( Rn satisfying
pre-asymptotically stable and that the system x 5 F 1 x 2 , G 1 A d D 2 ( A. If there exists a Lyapunov-function candi-
#
x [ C d A1 has the compact set A2 5 5 1 0, 0 2 6 globally pre- date V for 1 H, A 2 such that
asymptotically stable. We conclude from Corollary 19 that
the origin of the system (24) is globally pre-asymptotically 8=V 1 x 2 , f 9 , 0 for all x [ C\A, f [ F 1 x 2 ,
stable. This conclusion means that, for the original system V1g2 2 V1x2 , 0 for all x [ D\A, g [ G 1 x 2 \A,
(23), the origin is asymptotically stable. Moreover, since M
is arbitrary in (24), the basin of attraction for the origin for then the set A is pre-asymptotically stable and the basin
the system (23) contains the set of initial conditions from of pre-attraction contains every forward invariant, com-
which each solution of (23) is bounded. pact set.
A consequence of Theorem 20 is that the compact
STABILITY ANALYSIS USING LYAPUNOV FUNCTIONS set A is globally pre-asymptotically stable if C h D
Lyapunov functions are familiar to most control engineers. is compact or the sublevel sets of V|dom V d (C h D) are
These functions provide a way of establishing asymptotic compact. A sublevel set of V|dom V d (C h D) is the set
stability without having to construct the system’s solutions 5 x [ dom V d ( C h D ) : V ( x ) # c 6 , where c $ 0.
explicitly, a daunting task for almost anything other than a Theorem 20 encompasses classical Lyapunov theo-
linear system. Sufficient Lyapunov conditions for asymp- rems, both for continuous- and discrete-time systems. For
totic stability are well known for both continuous- and example, consider the case where F is a continuous func-
discrete-time systems. These conditions amount to finding tion, and suppose there exist a continuously differentiable,
a nonnegative-valued function that is strictly decreasing positive semidefinite function V and a compact neighbor-
along solutions. In the case of a smooth Lyapunov function hood C of the origin such that 8=V 1 x 2 , F 1 x 2 9 , 0 for all
x [ C\ 5 0 6 . According to Theorem 20, the origin of x 5 F 1 x 2 ,
#
for a continuous-time system, the decrease condition can be
verified from negative definiteness of the inner product of x [ C, is globally pre-asymptotically stable. Then, since C
the Lyapunov function’s gradient and the vector field that contains a neighborhood of the origin, it follows that the
origin is asymptotically stable for x 5 F 1 x 2 . This conclu-
#
generates solutions. The standard Lyapunov-based suffi-
cient conditions for asymptotic stability are covered in the sion parallels the conclusion of [39, Thm. 4.1]. Similarly, for
version below for hybrid dynamical systems; see also [72, the case where G is a continuous, single-valued mapping,
Cor. 7.7]. First, we give a definition. if G 1 0 2 5 0 and there exist a continuous, positive semidefi-
Given the hybrid system H with data 1 C, F, D, G 2 and nite function V and a compact neighborhood D of the ori-
the compact set A ( Rn, the function V : dom V S R is a gin such that V 1 G 1 x 2 2 , V 1 x 2 for all x [ D\ 5 0 6 , then the
Lyapunov-function candidate for 1 H, A 2 if i) V is continu- origin of x 1 5 G 1 x 2 is asymptotically stable.
ous and nonnegative on ( C h D ) \ A ( dom V , ii ) V is The following examples illustrate the use of Theorem 20.
V 1 g 1 x 2 2 5 2 rx2 1 k
1 2 2
Å2 2
rx 100 2 90 1
F 5 con • f1 J £ 2 90 § , f2 J £ 100 § , f3 J £ 2 90 § ,
b|x2|
k
# ra1 1 1 2 90 100
"2
2 90
b|x2| for all x [ D\ 5 0 6
1
, a211
k
"2 f4 J £ 50 § , f5 J £ 50 § s.
# V 1 x 2. 2 90 1
Hm, K. When this auxiliary function is always decreas- We illustrate Theorem 23 with several examples.
ing along solutions to Hm, K, the system Hm, K cannot have
complete solutions since 1 Cm h Dm 2 d K is compact for Example 3 Revisited: Lyapunov Analysis with
each m . 0 and each compact set K. More generally, Prop- the Invariance Principle.
osition 24 below can be applied iteratively to reduce the Consider the hybrid model of a bouncing ball in Ex-
size of the flow and jump sets in an attempt to make them ample 3. We establish global asymptotic stability of
empty eventually, thereby ruling out complete solutions. the origin. Let A be the origin in R2. The condition
This result is another consequence of the general invari- g 1 D d A 2 ( A holds. Consider a Lyapunov-function
ance principle discussed in “Invariance.” candidate V : R2 S R given by V ( x ) 5 gx 1 1 ( 1/2 ) x 22 .
The sublevel sets of V|C h D are compact, and V satisfies
Proposition 24
Consider the hybrid system H J 1 C, F, D, G 2 satisfying the 8=V 1 x 2 , f 1 x 2 9 5 gx2 2 gx2 5 0 for all x [ C
Basic Assumptions, and assume that C h D is compact. Let
W be continuously differentiable on an open set containing and
C and continuous on D. Suppose
V 1 g 1 x 2 2 5 r2x22 , x22 5 V 1 x 2
1 1
for all x [ D\A.
8=W 1 x 2 , f 9 # 0 f [ F1x2,
2 2
for all x [ C,
W g2 2 W1x2 # 0
1 for all x [ D, g [ G 1 x 2 d 1 C h D 2 . The set Dm defined in (25)–(28) is empty for each m . 0
since dom Gm is empty for each m . 0. Thus, A is globally
For each n [ W 1 dom W d 1 C hD 2 2 , define asymptotically stable as long as there are no complete solu-
tions that flow only and keep V equal to a positive constant.
Fn 1 x 2 J F 1 x 2 d =W 1 x 2 ' for all x [ C d LW 1 n 2 , It may be possible to see by inspection that no solutions
Gn 1 x 2 J G 1 x 2 d LW 1 n 2 for all x [ D d LW 1 n 2 ,
of this type exist. Otherwise, consider using the idea in
Proposition 24 to rule out complete solutions to Hm, m . 0,
Cn J dom Fn,
that only flow.
Dn J dom Gn d ( dom Fn h Gn ( dom Gn ) ) . Next, applying Proposition 24 with C J 5 x [ R2 : x1 $ 0,
V ( x ) 5 m 6 , D given by the empty set, and W1 1 x 2 5 x2,
Then the hybrid system H has complete solutions if we obtain
and only if there exists n [ W ( dom W d ( C h D ) ) such
that the hybrid system Hn J 1 Cn, Fn, Dn, Gn 2 has complete 8=W1 1 x 2 , f 1 x 2 9 5 2 g , 0 for all x [ C.
solutions.
Whenever 8=W ( x ) , f 9 , 0 for all x [ C, f [ F 1 x 2 , and The set Cn in Proposition 24 is empty for all n, and thus
W 1 g 2 2 W 1 x 2 , 0 for all x [ D, g [ G ( x ) d ( C h D ) , complete solutions that only flow are ruled out. It
it follows that Cn and Dn are empty and thus Hn has no follows from Theorem 23 that the set A is globally asymp-
complete solutions. Typically, Cn and Dn have fewer totically stable. ■
points than C and D, and so it may be easier to establish
that Hn has no complete solutions than to establish that H Example 25: Interacting Fireflies.
has no complete solutions. The combination of Theorem Consider the firefly model given earlier with two fireflies,
23 with a repeated application of Proposition 24 is the and suppose each flow map is equal to a constant f . 0.
idea behind Matrosov’s theorem for time-invariant hy- This choice results in a hybrid system H with x [ R2 and
brid systems, as presented in [74]. The original Matrosov the data
theorem, conceived for time-varying systems, appears
C J 3 0, 1 4 3 3 0, 1 4 , F 1 x 2 J c d for all x [ C,
in [53]. Matrosov theorems reach their full potency for f
time-varying systems, where invariance principles typi- f
cally do not hold. Refinements of Matrosov’s original D J 5 x [ C : max 5 x 1, x 2 6 5 1 6 ,
idea have appeared over the years. See [60] and [50] and
g 1 1 1 1 e 2 x1 2
the references therein. A version for time-varying hybrid G1x2 J c d for all x [ D,
systems appears in [52]. g 1 1 1 1 e 2 x2 2
T he invariance principle has played a fundamental role over Theorem S13 [72, Thm. 4.3]
the years as a tool for establishing asymptotic stability in Suppose that the hybrid system H 5 1 C, F, D, G 2 satisfies the
nonlinear control algorithms. Prime examples include the Jurd- Basic Assumptions. Let V : Rn S R be continuously differen-
jevic-Quinn approach to nonlinear control design [S45], [S47] tiable on an open set containing C , continuous on C c D , and
as well as general passivity-based control [S42], [S48]. satisfy
The invariance principle transcends stability analysis by
characterizing the nature of the sets to which a bounded so- uC ( x ) # 0 for all x [ C, where uC ( x ) J maxf[F(x) 8=V ( x ) , f9,
lution to a dynamical system converges. The basic invari- uD ( x ) # 0 for all x [ D, where uD ( x ) J maxg[G(x)V ( g ) 2V ( x ) .
ance principle for dynamical systems with unique solutions
is due to LaSalle [47], [S46]. It has been extended to sys- Then, there exists r [ R such that each complete and bounded
tems with possibly nonunique solutions, in particular, differ- solution x to H converges to the largest weakly invariant sub-
ential inclusions, as well as integral and differential versions set of the set
with nonsmooth functions [S49]. A version of the invariance
principle for hybrid systems with unique solutions and con- 5 z : V 1 z 2 5 r 6 d 1 uC21 1 0 2 h 1 uD21 1 0 2 d G 1 uD21 1 0 2 2 2 2 , (S8)
tinuous dependence on initial conditions appears in [51]. An
invariance principle for left-continuous dynamical systems where uC21 1 0 2 :5 5 z [ C : uC 1 z 2 5 0 6 and uD21 1 0 2 :5 5 z [ D :
having unique solutions and quasi-continuous dependence uD ( z ) 5 0 6 .
on initial conditions is given in [S43]. An invariance principle Further information on a particular complete and bounded
for switched systems is given in [S3] and [S44]. The results solution x may lead to more precise descriptions of the set
quoted below are from [72], which contains results analo- to which x converges. For example, if x is Zeno (see “Zeno
gous to those in [S49] but for hybrid dynamical systems. Solutions”) then both forward and backward weak invariance
of v 1 x 2 can be verified by a complete solution to H that never
Lemma S12 flows, and, moreover, we can limit our attention to weakly in-
Suppose that the hybrid system H satisfies the Basic Assump- variant subsets of
tions, and let x : dom x S Rn be a complete and bounded solu-
5 z [ D : uD 1 z 2 5 0 6 d G 1 5 z [ D : uD 1 z 2 5 0 6 2 .
tion to H. Then the omega-limit of x , that is, the set
On the other hand, if there exists t . 0 such that all jumps of x
v ( x ) 5 5 z [ Rn : there exists ( ti, ji ) [ dom x there exists are separated by at least an amount of time t, then we can limit
ti 1 ji S `, x ( ti, ji ) S z 6 , our attention to weakly invariant subsets of
5 z [ C : uC 1 z 2 5 0 6 .
is nonempty, compact, and weakly invariant in the sense that
the following conditions are satisfied: Example 14: Illustration of the Invariance Principle.
i) v 1 x 2 is weakly forward invariant, that is, for each z [ v 1 x 2 Consider the hybrid system in the plane with data
there exists a complete solution y to H such that
y 1 0, 0 2 5 z and y 1 t, j 2 [ v 1 x 2 for all 1 t, j 2 [ dom y . C J 5 x [ R2 : x2 $ 0 6 , F ( x ) J c
2x2
d for all x [ C,
ii) v 1 x 2 is weakly backward invariant, that is, for each x1
z [ v 1 x 2 and each m . 0, there exists a solution y to H D J 5 x [ R2 : x2 # 0 6 , G ( x ) J c
2x2
d for all x [ D.
such that y 1 t, j 2 [ v 1 x 2 for all 1 t, j 2 [ dom y and such that x1
y 1 tz, jz 2 5 z for some 1 tz, jz 2 [ dom y with tz 1 jz . m . Solutions flow in the counterclockwise direction when in
Furthermore, the distance from x 1 t, j 2 to v 1 x 2 decreases to the closed upper-half plane. From the closed lower-half
zero as t 1 j S ` , and, in fact, v 1 x 2 is the smallest closed set plane, which is where solutions are allowed to jump, solu-
with this property. tions undergo an instantaneous rotation of p/2 , also in
where e . 0, g 1 s 2 5 s when s , 1, g 1 s 2 5 0 when s . 1 and This function is continuously differentiable on the open set
g 1 s 2 5 5 0, 1 6 when s 5 1. X\A, where
Note that the compact set A J 5 x [ C : x1 5 x2 6 rep-
resents synchronized flashing. We establish asymptotic X J 5 x [ R2 : V ( x ) , ( 1 1 k ) /2 5 ( 1 1 e ) / ( 2 1 e ) 6
stability A and characterize the basin of attraction. 5 5 x [ R2 : |x 1 2 x 2| 2 ( 1 1 e ) / ( 2 1 e ) 6 .
Define k J e/ 1 2 1 e 2 and consider the Lyapunov-func-
tion candidate Let n* 5 1 1 1 e 2 / 1 2 1 e 2 , let n [ 1 0, n* 2 , Kn 5 5 x [ C h D:
V 1 x 2 # n 6, and define C( 5 C d Kn and D( 5 D d Kn. Since V
V ( x ) J min 5 |x 1 2 x 2|, 1 1 k 2 |x 1 2 x 2| 6 . is a function of only x1 2 x2 and F1 1 x 2 5 F2 1 x 2 , it follows that
S 5 5 z [ R2 : |z| 5 1, z1 $ 0 or z2 $ 0 6 .
Sinv 5 5 z [ R2 : |z| 5 1, z2 $ 0 6 h ( 0, 21 ) . FIGURE S10 A solution to the hybrid system in Example S14.
All solutions to this system flow in the counterclockwise di-
rection when in the closed upper-half plane. From the closed
As Figure S10 depicts, S inv is also exactly the range of the lower-half plane, all solutions jump and undergo a rotation by
periodic solution x and thus its omega-limit. Indeed, we have p/2, also in the counterclockwise direction. An application
x 1 t, j 2 5 1 cost, sint 2 for t [ 3 0, p 4 , x 1 p, 1 2 5 1 0, 21 2 , and of Theorem S13 indicates that all solutions converge to the
x 1 t, j 2 5 x 1 t 2 p, j 2 2 2 for t $ p, j $ 2. In particular, S inv is the set S inv 5 5 z [ R2 : |z| 5 1, z2 $ 0 6 c 1 0, 21 2 . This set co-
incides with the range of the solution that starts flowing from
x 1 0, 0 2 5 1 1, 0 2 .
smallest closed set to which x converges.
Note that asserting that the periodic solution x converges
to S inv is possible only by considering both weak forward and REFERENCES
backward invariance. Indeed, the largest weakly forward invari- [S42] C. I. Byrnes, A. Isidori, and J. C. Willems, “Passivity, feedback equiv-
alence, and the global stabilization of minimum phase nonlinear systems,”
ant subset of S is S itself. ■ IEEE Trans. Automat. Contr., vol. 36, no. 11, pp. 1217, 1228–1240, 1991.
In the conclusion of Theorem S13 we can equivalently con- [S43] V. Chellaboina, S. P. Bhat, and W. H. Haddad, “An invariance prin-
sider the largest weakly invariant subset of ciple for nonlinear hybrid and impulsive dynamical systems,” Nonlinear
Anal., vol. 53, pp. 527–550, 2003.
[S44] J. P. Hespanha, D. Liberzon, D. Angeli, and E. D. Sontag, “Nonlin-
5 z : V 1 z 2 5 r 6 d 1 uC21 1 0 2 huD21 1 0 2 2 . ear norm-observability notions and stability of switched systems,” IEEE
Trans. Automat. Contr., vol. 50, no. 2, pp. 154–168, 2005.
[S45] V. Jurdjevic and J. P. Quinn, “Controllability and stability,” J. Differ.
This set has a simpler description but is usually larger than the set Equ., vol. 28, no. 3, pp. 381–389, 1978.
used in Theorem S13. More extensive investigation using invariance [S46] J. P. LaSalle, The Stability of Dynamical Systems. Philadelphia,
properties is then needed to obtain the same set to which solutions PA: SIAM, 1976.
# 2j P 1 t 2 Bw 1 j M 1j 2 j ^ 1 P 1 t 2 B 2 M 3 2
pled-data systems. The second application covers nonlin- ^ ^
3 M221 1 B^P 1 t 2 2 M^
3 2j
ear networked control systems, which are becoming more
prevalent due to the ubiquity of computers and communi-
j ^ M M3 j
2 c d c ^1 d c d.
cation networks.
w M3 M2 w
Example 26: Absolute Stability for Sampled-Data Systems
In this example we study the absolute stability of periodic Completing squares leads to the conclusion that
jump linear systems, including sampled-data systems. Ab-
solute stability refers to asymptotic stability in the presence 8=V 1 x 2 , f 9 # 0 for all x [ C, f [ F 1 x 2 .
of arbitrary time-varying, sector-bounded nonlinearities
[39, Sec. 7.1]. It is possible to model sector-bounded nonlin- We now turn to the change in V during jumps. Asymp-
earities through a differential inclusion. totic stability of the set A follows from Theorem 23 when
Consider the class of hybrid systems with state x [ Rn11, there exists e . 0 such that
decomposed as x 5 1 j, t 2 , where j [ Rn, and data
1 j ^J ^ 1 w ^L^ 2 P 1 0 2 1 Jj 1 Lw 2 2 j ^P 1 T 2 j
C J 5 x : t [ 3 0, T 46 ,
j ^ N1 N3 j
Aj 1 Bw j ^ M1 M3 j 2c d c ^ d c d # 2 ej ^j.
F(x) J e c d , w: c d c ^ d c d # 0f
(30)
w N3 N2 w
1 w M3 M2 w
for all x [ C,
D J 5x : t 5 T6, To check this condition, we need to relate P 1 0 2 and P 1 T 2 .
The solution to the matrix differential equation (29) can be
Jj 1 Lw j ^ N1 N3 j
G(x) J e c d , w: c d c ^ d c d # 0f written explicitly by forming the Hamiltonian matrix
0 w N3 N2 w
for all x [ D, A 2 BM221M^ BM221B^
H5 c 3
^d ,
M1 2 M3M221M^3 1 3 2
2 A 2 BM221M^
where M1, M2, N1, and N2 are symmetric. We assume that
the eigenvalues of M1 and N1, which are real, are nonnega- forming the matrix exponential F 1 t 2 J exp 1 2 Ht 2 , parti-
tive; in other words, M1 and N1 are positive-semidefinite tioning F in the same way that H is partitioned, and verify-
matrices. The matrices M2 and N2 are positive definite, ing that, where P is defined,
meaning that their eigenvalues are positive. These condi-
tions guarantee that the Basic Assumptions hold. The state P 1 T 2 t 2 5 1 F 21 1 t 2 1 F 22 1 t 2 P 1 T 2 2 1 F 11 1 t 2 1 F 12 1 t 2 P 1 T 2 2 21.
t corresponds to a timer state that forces jumps every T sec-
onds. With w constrained to zero, the system is a simple When F 11 1 t 2 is invertible for all t [ 3 0, T 4 , the quantities
#
periodic jump linear system with flow equation j 5 Aj and 2F11 1 t 2 21F12 1 t 2 and F 21 1 t 2 F 11 1 t 2 21 are positive semidefi-
jump equation j 1 5 Jj. More generally, w is constrained to nite for all t [ 3 0, T 4 . Also, defining C J F 1 T 2 , X J P 1 T 2 ,
satisfy a quadratic constraint that is a function of the state and letting S be a matrix satisfying SS ^ 5 2C 11 21
C 12 ,
2 1 k 11 1 e 2 1 k 1 p 1 t 2 k 21 2
1 q
2 12 f 1x 2 l2 ar 2 1 b 2q lsq/c11
≥ ¥ c 1 1 d
£ § , 0,
c 211
f 2 1 x2 2 0 , q.
1 k 1 p 1 t 2 k 21 2 1 k 1 p 1 t 2 k 21 2 2
1 1
2 lsq/c11 2 1 1 s 2q
2 12 4k 11 12
ui 5 0 when xi , 1 1 2 k 2 /2 and ui 5 2k when xi . 1 1 2 k 2 /2 expressed directly in terms of a Lyapunov analysis and thus
results in global synchronization. This property is estab- avoids lifting, is reminiscent of the approach to sampled-
lished using the same Lyapunov-function candidate used data control taken in [79] and [85].
in Example 25.
Example 28: Resetting Nonlinear Control
Example 26 Revisited: Sampled-Data This example is based on [11], [29], and [30]. Consider
#
Feedback Control Design the nonlinear control system j 5 f 1 j, u 2 , where j [ Rn
The design of a linear sampled-data feedback controller and f is continuous, together with the dynamic control-
ler h 5 f 1 h, j 2 , u 5 k 1 h, j 2 , where h [ Rm and f and
#
can be carried out based on the analysis in Example 26.
The matrix M in Example 26 is used to indicate a desired k are continuous. Let x 5 1 j, h 2 and, for each x [ Rn1m,
dissipation inequality for the closed-loop system or a de- define F 1 x 2 J 1 f 1 j, k 1 h, j 2 2 , f 1 h, j 2 2 . We refer to the
system x 5 F 1 x 2 as Hc . Let V be a Lyapunov-function
#
sired stability robustness margin, while the matrices L
and N are typically zero. In this case, using the defini- candidate for 1 Hc, 5 0 6 2 that is positive on Rn1m\ 5 0 6 and
tions of S and C given in Example 26, the matrix inequal- for which the sublevel sets of V are compact. Moreover,
ity (31) reduces to suppose
R1) The condition 8=V 1 x 2 , F 1 x 2 9 # 0 holds for all
J ^ 1 C 11 2 J2X x [ Rn1m.
21 2^ 21
1 C 21C 11 J ^C 11
21
c d , 0. (36)
XC 11 XS
^ We do not assume that the function V satisfies the con-
2^
S XC 11 J 2 I 1 S ^XS
ditions of Theorem 23 for the system Hc. Instead, global
asymptotic stability of the origin is achieved by resetting
In turn, the matrix J decomposes as J 5 G 1 HUK, where the controller state h so that V decreases at jumps. To that
G, H, and K are fixed and U is a design parameter corre- end, suppose that
sponding to feedback gains. Then feasibility of the matrix R2) For each j [ Rn there exists h* [ Rm such that
inequality (36) in the parameters U and X 5 X^ . 0 can be V 1 j, h* 2 , V 1 j, h 2 for all h 2 h*.
identified with a synthesis problem for the discrete-time Since V is continuous and its sublevel sets are compact,
system the function jAh* 1 j 2 is continuous. Since V 1 x 2 is positive
when x 2 0, h* 1 0 2 5 0. As in [29] and [30], the hybrid con-
2^
j 1 5 C 11 1 G 1 HUK 2 j 1 Sw, troller resets h to h* 1 j 2 when 1 j, h 2 belong to the jump set
y 5 Y 1 G 1 HUK 2 j, D. We suppose that
R3) The set D ( Rn1m is closed, intersects the set
where Y^Y 5 C 21C 11 21
. Recall that the matrix C 21C 1121
is 5 1 j, h 2 : h 5 h * 1 j 2 6 only at the origin, and, for each
guaranteed to be positive semidefinite as long as F11 1 t 2 c . 0, there does not exist a complete solution to
is invertible for all t [ 3 0, T 4 . The synthesis problem cor-
responds to picking the parameter U to ensure that the
x 5 F1x2 x [ 1 Rn1m\D 2 d LV 1 c 2 ,
#
,2-gain from the disturbance w to the output y is less than
one, as certified by the energy function V 1 j 2 5 j^Xj through
the condition V 1 x 1 2 2 V 1 x 2 # 2 e|j|2 2 |y|2 1 |w|2 for where LV 1 c 2 denotes the c-level set of V.
some e . 0. Under conditions R1), R2), and R3), it follows from Theorem
When the parameters G, H, and K correspond to state 23 with the Lyapunov-function candidate V that the hybrid
feedback or full-order output feedback, the above synthe- resetting controller
sis problem can be cast as a convex optimization problem
in the form of a linear matrix inequality [22]. Generaliza-
u 5 k ( h, j )
tions of this sampled-data control approach are developed # f ( j, h ) [ Rn1m \ D,
for multirate sampled-data systems in [44] based on time- h 5 f ( h, j )
varying lifting. The approach discussed above, which is h 1 5 h*(j) 1 j, h 2 [ D
Z eno of Elea would have liked hybrid systems. According to start. When you reach that point, note where I am again. Keep do-
Aristotle [S59] with embellishment from Simplicius [S57] and ing this until you pass me (... if you ever do)!”
the authors of this article, Zeno was the ancient Greek philoso- To adhere to the Tortoise’s ground rules, Achilles needs a state
pher whose Tortoise bested the swift Achilles in a battle of wits variable r that keeps track of the Tortoise’s most recently observed
by appealing to hybrid systems theory. Here is our version of position. Specifically, each time Achilles reaches r, he must up-
the encounter, a modern paraphrase of Zeno’s paradox. date the value of r to be the Tortoise’s current position. Thus, we
Tortoise: Hey, Achilles, how about a race? have a hybrid system with the state x 5 1 a, t, r 2 and data
Achilles: Are you kidding me? I’ll trounce you!
2
C J 5 x [ R3 : a # r # t 6 , f ( x ) J £ 1 § for all x [ C,
Tortoise: Well, for sure, I need a head start...
Achilles: And how much of a head start would you like?
0
Tortoise: Doesn’t matter. Any will do.
a
D J 5 x [ R3 : a 5 r # t 6 , g ( x ) J £ t § for all x [ D.
Achilles: And why is that?
Tortoise: Because this will be a hybrid race. t
Achilles: What do you mean by that?
Tortoise: Well, Achilles, as you try to catch me, I want you to We call this system the Achilles. Achilles’ initial goal is to reach
keep track of where I have been. This will help you see that you the set A J 5 x [ R3 : a 5 r 5 t 6 , but he is observant enough to
are catching up to me. Note where I start. When you reach that see that he cannot reach this goal. Indeed, Achilles cannot reach
point, note where I am again. Keep doing this until you pass me! A by flowing, since then it would be possible to follow his flowing
Surely, your brain can handle this, can’t it Achilles? trajectory backward from a point in A to a point outside of A while
remaining in C . However, no solution of the equation x 5 2 f 1 x 2,
#
Achilles: So, the race is hybrid because, while my feet are
x [ C can start in A and leave A. In fact, satisfying x 5 2 f 1 x 2
#
running a race I can win, my mind is racing without end? You
know, I just felt a twinge in a tendon near my ankle. Do you think from any initial condition in A immediately would force t , r, that
we can postpone this race for a couple of millennia? is, the Tortoise is behind Achilles’ most recent observation of the
Achilles would overtake the Tortoise and win if the race de- Tortoise’s location, which is outside of C . Also, Achilles cannot
pended only on Achilles’ feet. However, the Tortoise’s ground reach A by jumping to it. Indeed, being in D but not in A means
rules require Achilles’ mind to be involved as well. The mental that a 5 r , t. The value of the state after a jump would satisfy
task that the Tortoise gives to Achilles prevents Achilles from a , r 5 t, which is neither in A nor D .
catching the Tortoise. Indeed, no matter how many steps of the It is no consolation for Achilles that the set A is glob-
mental assignment that Achilles completes, he will still trail the ally asymptotically stable for the Achilles system. This fact
Tortoise physically. can be established using the Lyapunov-function candidate
In modern terms, here is the hybrid model that the Tortoise V 1 x 2 J 2 1 x2 2 x1 2 1 x2 2 x3, which is zero on A, positive oth-
suggests. Let t [ R denote the position of the Tortoise, which is erwise, and such that the sublevel sets of V |dom V d 1C h D2 are
initialized to zero as a reference point. Let a [ R denote the posi- compact. The function V satisfies 8=V 1 x 2 , f 1 x 2 9 5 2 1 for all
tion of Achilles. Due to the Tortoise’s head start, a is initialized to x [ C , and V 1 g 1 x 2 2 # 1 2 / 3 2 V 1 x 2 for all x [ D .
a negative value. The speed of the Tortoise is normalized to one, Neither is it a consolation for Achilles that the ordinary time t
whereas the speed of Achilles, for sake of convenience, is taken to used to approach A is bounded, a property that can be established
be two. If this description were the complete model, Achilles would by integrating the derivative of V along flows of the solution to the
reach the Tortoise at a time proportional to the initial separation be- Achilles system. Figure S11 shows typical hybrid time domains for
tween the Tortoise and Achilles, and then Achilles would continue solutions to the Achilles system. The hybrid time domain in Fig-
past the Tortoise to victory. But this point is where the Tortoise gets ure S11(a) occurs for solutions starting in 1 C c D 2 \ A . The hybrid
clever. He gives Achilles the following instructions: “Note where I time domain in Figure S11(b) occurs for solutions starting in A.
R3’) For each c . 0 and c2 . 0, there are no complete solu- 5 x 5 1 j, h 2 :V2 1 x 2 $ r 1 V1 1 j 2 2 6 is forward invariant for
tions to the systems the system
x 5 F 1 x 2 , x [ 5 x5 1 j, h 2 [ LV 1 c 2 : V2 1 x 2 # r 1 V1 1 j 2 2 6
#
x 5 F1x2, x [ 5 x [ L V 1 c 2 : 8=V2 1 x 2 , F 1 x 2 9 $ 0 6 .
#
and
# Finally, note that when the set D satisfies R3), condition
x 5 F(x),
x [ 5 x 5 1 j, h 2 [ LV 1 c 2 d LV2 1 c2 2 : V2 1 x 2 $ r 1 V1 1 j 2 2 6 . R3) is also satisfied for any closed set that contains D and
does not intersect 5 1 j, h 2 : h 5 h * 1 j 2 6 except at the origin.
The fact that R3’) implies R3) follows from Propo- So, for example, jumps might be enabled when x belongs
sition 24 with W J 2V2 and the fact that the set to the set D defined in (37) and also when x belongs to
G0 1 x 2 J G 1 x 2 d 5 v : E 1 v, x 2 5 5 0 6 6 for all x [ D,
We discuss below when pre-asymptotic stability of
(38)
D 0 J D d 5x : G0 ( x ) 2 [6.
a compact set for a hybrid system can be deduced from
pre-asymptotic stability of the set for the system with its (39)
events removed.
Events are defined by an event indicator, which is an The data of H0 satisfies the Basic Assumptions and the
outer semicontinuous set-valued mapping E : Rn 3 Rn S S solutions to H0 experience no events.
5 0 6 . An event is a pair 1 g, x 2 such that E 1 g, x 2 is empty. The
number of events experienced by a hybrid arc x is given Example 30: Jumps Out of C h D
by the cardinality of the set For a hybrid system H 5 1 C, F, D, G 2 , it is possible that
g o C h D for some x [ D and g [ G 1 x 2 . These jumps
5 j : for some t $ 0, 1 t, j 2 , 1 t, j 1 1 2 [ dom x, can be inhibited with an event indicator E satisfying
E 1 x 1 t, j 1 1 2 , x 1 t, j 2 2 5 [ 6 . E 1 g, x 2 5 0 if g [ C h D and E 1 g, x 2 5 [ otherwise. In this
case, the definition (38) gives G 0 ( x ) 5 G ( x ) d ( C h D )
Example 27 Revisited: Networked Control for each x [ D. Each solution of H experiences no more
Systems with Transmission Dropouts than one event, since a solution cannot be extended from
Consider a networked control system with the pos- a point g o C h D.
sibility of communication dropouts, modeled by the According to Theorem 31, the compact set A is globally
hybrid system with data C J Rn1 1n2 3 3 0, T 4 3 5 0, 1 6 , pre-asymptotically stable for the hybrid system 1 C, F, D, G 2
D J Rn1 1n2 3 5 T 6 3 5 0, 1 6 , whenever G 1 D d A 2 ( A, and A is globally pre-asymptot-
ically stable for the hybrid system 1 C, F, D0, G0 2 , where
F1 1 x1, x2 2 G 0 ( x ) 5 G ( x ) d ( C h D ) for each x [ D and D0 is the set
F2 1 x1, x2 2 of points x [ D for which G0 1 x 2 2 [. ■
F1x2 J ≥ ¥ for all x [ C, The stability properties of H0 5 1 C, F, D0, G0 2 , with D0
1
and G0 defined in (38)–(39), can be used to derive stabil-
ity properties for H 5 1 C, F, D, G 2 under a variety of condi-
0
that combine dwell-time conditions and Lyapunov condi- equations in [1], extend to hybrid systems. The stability con-
tions, as in [33]. ditions are expressed in terms of an event indicator E , the
eventless system with events inhibited, as in (38)–(39), and
Finite Number of Events a decomposition of the event indicator into a finite number
An approach used in adaptive supervisory control to estab- of event types.
lish convergence properties is to establish that the number
of switches experienced by a solution is finite [58]. A related Theorem 32
result for asymptotic stability in hybrid systems is given in For a hybrid system H 5 1 C, F, D, G 2 satisfying the Basic
the next theorem. Assumptions with an event indicator E , the compact set A
is globally pre-asymptotically stable if and only if the fol-
Theorem 31 lowing conditions are satisfied:
Suppose that the hybrid system H 5 1 C, F, D, G 2 with state 1) G ( D d A ) ( A,
x [ Rn satisfies the Basic Assumptions. Let the compact 2) A is globally pre-asymptotically stable for the system
set A ( Rn satisfy G 1 D d A 2 ( A, and assume that A is glob- without events H0 J 1 C, F, D0, G0 2 .
ally pre-asymptotically stable for the eventless hybrid system 3) There exist a function W : C h D S R $0, and class-
H0 5 1 C, F, D0, G0 2 , where D0 and G0 are defined in (38)–(39). K` functions a1, a2, a3, such that
Also suppose that, for the hybrid system H and each compact
set K ( Rn, there exists N . 0 such that each solution starting a1 1 |x|A 2 # W 1 x 2 # a2 1 |x|A 2 for all x [ C h D,
in K experiences no more than N events. Then the set A is W 1 g 2 # a3 1 W 1 x 2 2 for all x [ D, g [ G 1 x 2 .
globally pre-asymptotically stable for the system H.
Theorem 31 is behind the proof of asymptotic stability 4) There exists a positive integer , and outer semicontin-
for the mode-switching controller for the disk drive sys- uous mappings Ei : Rn 3 Rn S S 5 0 6 , i 5 1, c, ,, such
tem in Example 1. As shown for “Example 1 Revisited,” that, for each 1 v, x 2 satisfying E 1 v, x 2 5 [, there exists
the origin of the closed-loop system is pre-asymptotical- i [ 5 1, c, , 6 satisfying Ei 1 v, x 2 5 [, and there exists
ly stable when jumps are eliminated, and when jumps a continuous function a4 satisfying 0 , a4 1 s 2 , s for
are allowed the maximum number of jumps is two. all s . 0 and such that, for each i [ 5 1, c, , 6 and
An outline of the proof of Theorem 31 is as follows. each solution x,
W 1 x 1 t2, j2 2 2 # a4 1 W 1 x 1 t1, j1 2 2 2 ,
Boundedness of maximal solutions and global pre-
attractivity of A for H follow from the fact that the num-
ber of events is finite and that the maximal solutions of where 1 tk, jk 2 [ dom x, k [ 5 1, 2 6 , t2 1 j2 . t1 1 j1,
H0 are bounded and the complete ones converge to A . are the two smallest elements of dom x satisfying
Stability can be established by concatenating trajecto- Ei 1 x 1 tk, jk 2 , x 1 tk, jk 2 1 2 2 5 [.
ries of the system H0, which can include jumps but not Figure 17 indicates a possible evolution of W 1 x 1 t, j 2 2 .
events, with jumps from G that correspond to events, and Some jumps, such as the one from 1 t2, 1 2 to 1 t2, 2 2 , do not
repeating this process up to N times. In particular, let correspond to events. When a particular type of event
e 5: e0 . 0 be given. Since G 1 D d A 2 ( A and G is outer occurs twice, the value of W at the second occurrence is
semicontinuous and locally bounded, there exists d0 . 0 required to be strictly less than the value of W at the first
such that G 1 D d 1 A 1 d0B 2 2 ( A 1 e0B. Since A is stable occurrence of the event. In the figure, one type of event is in-
for H0, there exists e1 . 0 such that, for each solution dicated in blue while the other is indicated in red. Blue events
x1 to H0, |x 1 ( 0, 0 ) |A # e1 implies |x1 1 t, j 2 |A # d0 for all occur at 1 t1, 1 2 and 1 t4, 4 2 . Thus, the value of W 1 x 1 t4, 4 2 2 is
1 t, j 2 [ dom x1. Repeating this construction N times and required to be less than the value of W 1 x 1 t1, 1 2 2 .
letting ej generate ej11 for j [ 5 0, c, N 2 1 6 , it follows Consider Theorem 32 in the context of multiple Lyapunov
that |x1 1 0, 0 2 |A # eN implies |x1 1 t, j 2 |A # e0 5 e for all functions Vq, where q [ 5 1, c, , 6 , for a switching sys-
solutions x of HN and all 1 t, j 2 [ dom x. tem with state z. We take W 1 x 2 J Vq 1 z 2 , where x 5 1 z, q 2 .
Sometimes it is assumed that Vq does not increase during
Multiple Lyapunov Functions flows [6], although this property is not necessary [90], and
Stability results involving multiple Lyapunov functions [6], is not assumed in Theorem 32. For i [ 5 1, c, , 6 , we take
[19] for switched systems, as well as related results for dis- Ei 1 v, x 2 5 0 for switches that turn the state q into a value j 2 i,
continuous dynamical systems in [90] and for differential otherwise we take Ei 1 v, x 2 5 [. Then the third assumption
d , Cc 5 h q[Q ( Cq 3 5 q 6 ) ,
fq ( x p, h )
fc ( x p, x c ) 5 c
0 H q 5 F q h U \ ( Cq h D q h F q ) (41)
G q ( x p, h )
G c ( x p, x c ) 5 c d d Jq ( x p , h ) , and in (42), shown at the bottom of the page. By construc-
q
tion, the mapping Jq is outer semicontinuous.
Dc 5 h q[Q ( ( Dq h Hq ) 3 5 q 6 ) , (40) For the closed-loop system, events are defined to be
jumps from D \ 1 A 3 Q 2 that change the value of q. In
and kc ( x p, x c ) 5 kq ( x p, h ) . other words, events are jumps that are due to Jq. Due to
The goal of the supervisor is to achieve global asymptotic the definition of the map Jq, if a solution experiences one
stability of the set A 3 Q, where A ( Rn1m is compact, while event then there is a time where the state ( x p, h, q ) satis-
satisfying Cc h Dc 5 U 3 Q, where U ( Rn1m is closed, and fies 1 x p, h 2 [ C q. Thereafter, due to the “Supervisor As-
to have that all maximal solutions starting in Cc h Dc are sumption,” until the state reaches the set A 3 Q, the index
complete. To guarantee that a supervisor can be constructed q jumps monotonically. Therefore, the number of events
to achieve the goal, as in [83] we assume the following. experienced by a solution will be bounded by k 1 1. More-
over, by the “Supervisor Assumption,” the corresponding
(Supervisor Assumption) eventless system has the set A 3 Q globally pre-asymptoti-
There exist closed sets Cq ( C | d U and D ( D | d U so cally stable. These observations lead to the following corol-
q q q
that the following conditions hold: lary of Theorem 31.
1) There exist closed sets C q ( Cq h Dq such that
h q[Q C q 5 U. Corollary 33
2) For each q [ Q, the interconnection of the plant (5) Under the “Supervisor Assumption” on the family of con-
and the controller Hc, q 5 1 Cq, fq, Dq, Gq, kq 2 , having trollers Hc, q, q [ Q, the closed-loop interconnection of the
closed-loop data given in (6), (7), is such that, using plant (5) and the hybrid supervisor Hc, with data given
the definition Fq J h 5i[Q, i.q6C i, the following condi- in (40)–(42), satisfies the Basic Assumptions and has the
tions are satisfied: compact set A 3 Q globally asymptotically stable.
a) The set A is globally pre-asymptotically stable. For the disk drive in Example 1 and for additional exam-
b) No solution starting in C q reaches ples reported below, the “Supervisor Assumption” holds
U \ ( Cq h Dq h F q ) \ A. and the supervisory control algorithm indicated in Corol-
c) Each maximal solution is either complete or ends lary 33 is used.
in F q h U \ ( Cq h Dq h F q ) . Some asymptotically controllable nonlinear control sys-
Condition 2c) is a type of local existence of solutions con- tems cannot be robustly stabilized to a point using classical,
dition. It is assumed to guarantee that a supervisor can be time-invariant state feedback. Examples include systems
constructed to achieve not only global pre-asymptotic sta- that fail Brockett’s necessary condition for time-invariant,
bility but also that all maximal solutions are complete. continuous stabilization [7], like the nonholonomic mobile
The combination of conditions 2b) and 2c) imply that robot, and other systems that satisfy Brockett’s condition,
solutions starting in C q are either complete or end in A or like the system known as Artstein’s circles [2]. The topologi-
Fq. This property enables building a hybrid supervisor cal obstruction to robust stabilization in Artstein’s circles is
to guarantee that each solution experiences no more than the same one encountered when trying to globally stabilize
k 1 1 events, where events are defined to be jumps in the a point on a circle and is related to the issues motivating
value of q when ( x p, h ) o A. the development of topological complexity in the context of
The sets involved in the “Supervisor Assumption” for motion planning in [21]. We now illustrate how hybrid feed-
the problem of stabilizing the inverted position of a pendu- back, and supervisory control in particular, makes robust
lum on a cart are depicted in Figure 20. stabilization possible.
5 xc 6
c d, ( x p , h ) [ U \ ( Cq h D q h F q ) ,
5 i [ Q : 1 x p, h 2 [ C i 6
Jq 1 x p, h 2 J e
5 xc 6
c d, ( x p , h ) [ H q \ ( U \ ( Cq h D q h F q ) ) .
5 i [ Q : i . q, 1 x p, h 2 [ C i 6
(42)
4
x′
FIGURE S13 Internal diagram of the Integrator System. The CT dynamics subsystem computes the flow of the hybrid system and
updates the parameter t . The Jump Logic subsystem triggers the jumps when the state x belongs to the jump set D . The Update
Logic subsystem updates the value of the simulation as well as the pair 1 t, j 2 at jumps. The Stop Simulation subsystem stops the
simulation when either the flow time is larger than or equal to the maximum flow time specified, the jump time is larger than or equal
to the maximum number of jumps specified, or the simulation is in neither C nor D .
0
−2 % state
−4 x1 = u(1);
−6 x2 = u(2);
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
j % jump map
x1plus = –0.8*x1;
(b)
x2plus = –0.8*x2;
out = [x1plus; x2plus];
FIGURE S15 Simulation of the bouncing ball system in Matlab/Simulink for an initial condi-
tion with height x1 5 1 and vertical velocity x2 5 0. The maximum flow time is T = 10 s,
the maximum number of jumps is J 5 20, and the flag priority is set to one, which indi- function out = D(u)
cates that priority is given to the jump condition coded in D.m. (a) shows the height of % state
the ball, which is denoted by x1 , while (b) shows the velocity of the ball, which is denoted x1 = u(1);
by x2 , projected to the t axis and to the j axis of its hybrid time domain. The plots indicate
x2 = u(2);
that the solution jumps more frequently as the height of the ball approaches zero, which
occurs when the flow time approaches the Zeno time of the bouncing ball system. For the % jump condition
given initial condition and parameters, the Zeno time is given by tZ 5 4.066 s. if (x1 <= 0 && x2 <= 0)
Example 34: Robust Global Asymptotic tive angular velocity corresponds to counterclockwise
Stabilization of a Point on a Circle rotation. Moving on the unit circle constrains the ve-
Consider a plant with state j that is constrained to evolve locity vector to be perpendicular to the position vector,
on the unit circle with angular velocity u, where posi- as generated by rotating the position vector using the
0.5
menting the condition x1 5 0 in 18 between jumps approach zero.
0
the jump set 5 x : x1 5 0, x2 # 0 6 0 1 2 3 4
20 Finally, Figure S16 depicts the
with a Simulink block that is ca- t [s] hybrid arc and hybrid time do-
pable of detecting zero crossings main for the height component
FIGURE S16 Hybrid arc and hybrid time domain corre-
of x1 as shown in Figure S14. sponding to the height component of the simulation to the of the solution obtained from the
This implementation uses three bouncing ball system in Matlab/Simulink depicted in Fig- numerical simulation.
Simulink blocks, namely, a “Hit ure S15. The hybrid arc is shown in blue, while the hybrid The Matlab/Simulink files
Crossing” block, a “Compare to time domain is shown in red. for this implementation and the
Zero” block, and an “AND Logi- functions used for plotting solu-
cal Operator” block. The last block combines the outputs of the tions as well as several examples are available online [S65].
first two blocks. The “Hit Crossing” block is used to detect when
the state component x1 crosses zero. For this purpose, this block REFERENCES
is required to have enabled Simulink’s zero-crossing detection [S60] R. Alur, T. Dang, J. Esposito, Y. Hur, F. Ivancic, V. Kumar, I. Lee, P.
algorithm for numerical computations. Then, when x1 crosses Mishra, G. J. Pappas, and O. Sokolsky, “Hierarchical modeling and analy-
sis of embedded systems,” Proc. IEEE, vol. 91, pp. 11–28, 2003.
zero, this algorithm sets back the simulation step to a time instant [S61] H. Elmqvist, S. E. Mattsson, and M. Otter, “Modelica: The new
close enough to the time at which x1 is zero, recomputes the object-oriented modeling language,” in Proc. 12th European Simula-
state, and sets its output to one. The “Compare to Zero” block is tion Multiconf., Manchester, U.K., 1998.
[S62] E. A. Lee and H. Zheng, “Operational semantics for hybrid
used to report when the state component x2 satisfies the condi- systems,” in Proc. 8th Int. Workshop Hybrid Systems: Computation
tion x2 # 0 by setting its output to one. Then, when the outputs and Control, Zurich, Switzerland, 2005, pp. 25–53.
of the “Hit Crossing” and “Compare to Zero” blocks are equal to [S63] J. Liu and E. A. Lee, “A component-based approach to model-
ing and simu lating mixed-signal and hybrid systems,” ACM Trans.
one, which indicates that x1 5 0 and x2 # 0 are satisfied, a jump Model. Comput. Simul., vol. 12, no. 4, pp. 343–368, 2002.
is reported to the Integrator System. [S64] The MathWorks. (2008). Matlab Simulink R2008 HelpDesk
Figure S15 depicts the height and velocity of a solution to [Online]. Available: http://www.mathworks.com/access/helpdesk/
help/toolbox/simulink/simulink.shtml
the bouncing ball system projected to the t axis and j axis of [S65] R. G. Sanfelice. (2008). Simulating hybrid systems in Matlab/
its hybrid time domain. The solution is generated from the initial Simulink [Online]. Available: http://www.u.arizona.edu/,sricardo/
1 2
condition 1, 0 with maximum flow time T = 10 s, maximum software.html
[S66] F. D. Torrisi and A. Bemporad, “HYSDEL—A tool for generating
number of jumps J 5 20, and highest priority for jumps, that is, computational hybrid models for analysis and synthesis problems,”
priority 5 1. It is easy to verify that the solutions to the bouncing IEEE Trans. Contr. Syst. Technol., vol. 12, pp. 235–249, 2004.
rotation matrix R 1 2p/2 2 . The constrained equations of For this system, global asymptotic stabilization of the
motion are point (1, 0) with a continuous state-feedback law is impos-
sible. Global asymptotic stabilization of this point by dis-
#
j 5 uR 12p/2 2 j j [ S 1. continuous feedback is possible, although the resulting
global closed-loop behavior is not robust to arbitrarily 5 j [ S1 : j 1 # 22/3 6 and Jq 5 3 2 q. Then, using the defini-
small measurement noise. For an explanation, see “Ro- tion of a supervisory controller in (40), we get a hybrid con-
bustness and Generalized Solutions.” We now describe a troller with state x c 5 q, data Cc 5 ( C1 3 5 1 6 ) h ( C2 3 5 2 6 ) ,
hybrid controller that achieves robust, global asymptotic fc 5 0, Dc 5 ( H1 3 5 1 6 ) h ( H2 3 5 2 6 ) , Gc 5 3 2 q, and
stabilization of the point (1, 0). kc 1 j, q 2 5 j q. See Figure 18.
Consider a hybrid supervisor for classical control- According to Corollary 33, this controller renders the set
lers H | , q [ Q 5 5 1, 2 6 , that are static and are given by 5 1 1, 0 2 6 3 Q globally asymptotically stable. In fact, it can
k q 1 j 2 5 j q . The “Supervisor Assumption” is satisfied for be established that the point 1 1, 0, 2 2 is globally asymptoti-
q
|
A 5 1 1, 0 2 and U 5 S1 by taking D1 5 D2 5 [ and cally stable. ■
C 1 5 C1 J 5 j [ S1 : j 1 # 2 1/3 6 ,
Example 35: Mobile Robot Control
C2 J 5 j [ S1 : j 1 $ 2 2/3 6 ,
Consider a stabilization problem for the control system
C 2 J 5 j [ S1 : j 1 $ 2 1/3 6 . #
r 1 z, j 2 [ R2 3 S1,
z 5 jv
#
j 5 uR 1 2p/2 2 j
Following the prescription of the components of a hybrid
supervisor in (41) and (42), we get H1 5 C 2, H2 5 S1 \ C2 5 where z [ R2 corresponds to the position of a nonholo-
nomic vehicle, v [ R denotes its ve-
locity, which is treated as a control
q=1 q=2 variable, j [ S1 denotes the orientation
of the vehicle, and u [ R denotes angu-
ξ2 ξ2
lar velocity, which is the other control
variable. The plant state is x p 5 1 z, j 2 .
Cp Cp
As in Example 34, positive angular ve-
locity corresponds to rotation of j in the
counterclockwise direction. The control
(1, 0)
objective is to asymptotically stabilize
1 ξ1 2 1 ξ 1 the point x *p 5 1 0, j * 2 , where j * [ S1.
3
We start by picking v 5 2r ( z ^j ) ,
where r is a continuous function satis-
fying sr 1 s 2 . 0 for all s 2 0. A discon-
Ψ1 = C1 (0, –1) H2 tinuous or set-valued choice for r is also
Ψ2 = H1 C 2 possible. The condition sr 1 s 2 . 0 for all
s 2 0 ensures that, during flows, the size
FIGURE 18 Illustration of the hybrid controller, for robust global stabilization of the point of z decreases except when z ^j 5 0 at
11, 0 2 on a circle S1, in Example 34. The sets C1 , C 1 , H1 and C2 , C 2 , H2 are subsets which points z# 5 0.
of the circle, contained in the half-planes as indicated by the arrows. These sets define With the vehicle velocity specified,
the flow and jump sets in modes q 5 1 and q 5 2 , respectively. In mode q 5 1 , the
controller drives the state toward the point 10, 2 12 from every point in C1 , as indicated
what remains is a control system with
by the arrows, and the switch to mode q 5 2 happens when the state is in H1 . In mode a single input u. Following the ideas
1 2
q 5 2 , the controller drives the state toward the point 1, 0 from every point in C2 , as in Example 34, a hybrid feedback algo-
indicated by the arrows, and the switch to mode q 5 1 happens when the state is in H2 . rithm for u exists that drives j to a given
This controller results in robust asymptotic stability of 11, 0 2 on the circle. Note that the point on the unit circle. More generally,
solutions starting from j 1 5 21 / 3 , j 2 . 0 , q 5 1 are not unique. This initialization
results in one solution that jumps immediately to q 5 2 and then flows to 11, 0 2 and
a comparable hybrid algorithm can be
another solution that flows to j 1 5 21 / 3 , j 2 , 0 , q 5 1 before jumping. Similarly, the developed to track any continuously
solutions starting from j 1 5 2 2 / 3 , j 2 . 0 , q 5 2 are not unique. differentiable function, of the plant
| hD
C | 5 ( R2 \ 5 0 6 ) 3 S1 3 V.
1 1
(b)
This specification avoids the singularity in 2z/|z|
at z 5 0. FIGURE 19 Position z 5 1 z1 , z2 2 and orientation j 5 1 j 1 , j 2 2 of the mo-
The second hybrid controller drives j to bile robot with hybrid controller in Example 35. (a) The position, orienta-
R 1 s 1 z 2 x1 2 j *, where s is a continuously differentia- tion, and ratio 2z / |z| with the first controller are shown on the z and j
ble function that is zero at zero and positive other- planes. The position z is in red, the orientation j is in blue, and 2z / |z|
is in black. The first controller drives z to a neighborhood of z 5 0 and
wise, and x1 is a state of the controller satisfying j to 2z /|z|. (b) The second controller drives the orientation to a cone
# centered at j * with aperture depending on z . Cones associated to two
x 5 vR 1 p/2 2 x, x [ S 1, points in a neighborhood of z 5 0 are depicted. The two points, indi-
cated by a square and a circle, along with the neighborhood around
where v . 0. This controller satisfies C | hD| 5 z 5 0 , indicated with dashed line, are depicted in the z plane, while
2 2
their associated cones are denoted on the j plane. As z approaches
R2 3 S1 3 V. This second controller is closely the origin, the aperture of the cone approaches zero.
related to time-varying stabilization algorithms that
have appeared previously in the literature. See, for
example, [64] and [71]. Example 36: Global Asymptotic Stabilization of
To satisfy the “Supervisor Assumption” for A 5 5 0 6 3 the Inverted Position for a Pendulum on a Cart
5 j* 6 3 V and U 5 R2 3 S1 3 V, we let 0 , e1 , e2 and take Consider the task of robustly, globally asymptotically sta-
bilizing the point x*p J 1 0, 1, 0 2 for the system with state
C1 J C| d 1 1R2 \ e B 2 3 S1 3 V 2 , x p 5 1 j, z 2 [ R3 and dynamics
1 1
D1 J D| d 1 1 R2 \ e B 2 3 S1 3 V 2 ,
#
zR 1 2p/2 2 j
1 1
C 1 J C1 h D1 5 1 R2 \ e1B 2 3 S1 3 V,
d 5: fp 1 xp, u 2
j
c#d5c
z j 1 1 j 2u
and xp [ S1 3 R.
CJ |C d 1 e B 3 S1 3 V 2 ,
2 2
D2 J D2 d 1 e2B 3 S1 3 V 2 ,
| This model corresponds to a pendulum on a cart. The model
includes an input transformation from force to acceleration
C2 J e1 B 3 S1 3 V.
of the cart, which is the control input u in the system above.
The components of the supervisor are then given by Moreover, the ratio of the gravitational constant to the pendu-
lum length has been normalized to one. The cart dynamics are
H1 5 C 2 5 e1B 3 S1 3 V, ignored to simplify the presentation; however, global asymp-
H2 5 1 R2 \ e2B 2 3 S1 3 V, totic stabilization of the full cart-pendulum system can be ad-
Jq 5 3 2 q. dressed with the same tools used below. The state j denotes
the angle of the pendulum. The point j 5 1 0, 1 2 corresponds
With these specifications, according to Corollary 33 the to the inverted position while j 5 1 0, 2 1 2 corresponds to the
hybrid controller (40)–(42) renders the set A 3 Q globally down position. The state z corresponds to the angular velocity,
asymptotically stable. ■ with positive velocity in the clockwise direction.
We construct a hybrid feedback stabilizer that supervis- Next, we check that A is globally pre-asymptotically
stable for x p 5 fp 1 x p, kq 1 x p 2 2 , x p [ Cq for each q [ Q. This
#
es three classical controllers. The first controller steers the
state x p out of a neighborhood of the point 2x *p . The second property holds for q 5 3 since C3 is a subset of the basin of
attraction for x *p for the system x p 5 fp 1 x p, k3 1 x p 2 2 .
#
controller moves the system to a neighborhood of the point
x *p. The third controller locally asymptotically stabilizes the For q 5 1, we note that x p [ C1 implies that W 1 x p 2 # d.
point x *p. For an illustration of the data of these controllers, In particular, |z| # "2d , "2 and j 2 # d 2 1 , 0. We
see Figure 20. use the Lyapunov function V1 1 x p 2 J 2 1 z, which is posi-
The third controller, a local asymptotic stabilizer for x *p, tive on C1, and we obtain
is simple to synthesize. For example, the idea of partial
feedback linearization with “output” j 1 can be used since j 2
is positive and bounded away from zero in a neighbor- 8=V1 1 x p 2 , fp 1 x p, k1 1 x p 2 2 9 5 j 1 1 j 2k
hood of x *p. Let k3 : S1 3 R S R denote this local asymp- # "1 2 j 22 2 |j 2|k
totic stabilizer, let C3 be a compact neighborhood of the # "d 1 2 2 d 2 2 1 1 2 d 2 k
point x *p that is also a subset of the basin of attraction , 0
for x *p for the system x p 5 fp 1 x p, k3 1 x p 2 2 , x p [ S1 3 R, and
#
let C 3 be a compact neighborhood of the point x *p with
the property that solutions of x p 5 fp 1 x p, k3 1 x p 2 2 starting
#
for all x p [ Cq. It follows from Theorem 20 that A is glob-
ally pre-asymptotically stable for x p 5 fp 1 x p, k1 1 x p 2 2 , x p [ C1.
#
in C 3 do not reach the boundary of C3. Then, redefine C3
and C 3 by intersecting the original choices with S1 3 R. This conclusion is equivalent to saying that there are
The set C 3 is indicated in green in Figure 20(d) while the no complete solutions that remain in the green region in
set C3 is the union of the green and yellow regions in the Figure 20(a).
same figure. For q 5 2, we use the Lyapunov function V2 1 xp 2 J
For the second controller, let 0 , d , e , 1 and 1 1 1 W 1 xp 2 2 2 2 2, which is positive on C2. For all x p [ C2,
define
8=V2 1 xp 2 , fp 1 xp, k2 1 xp 2 2 9 5 21 zj2 2 2 1 W 1 xp 2 2 2 2 2 # 0.
W 1 xp 2 J
1 2
z 1 1 1 j 2, It follows from Theorem 23 that A is globally pre-
2
asymptotically stable for x p 5 fp 1 x p, k2 1 x p 2 2 , x p [ C2. This
#
C 2 J 5 1 j, z 2 [ S1 3 R : W 1 xp 2 $ e 6 \ C 3,
C2 J 5 1 j, z 2 [ S1 3 R : W 1 xp 2 $ d 6 \ C 3, is seen by noting that V2 1 x p 2 5 m $ 1 implies
k2 xp 2 J 2 zj2 1 W 1 xp 2 2 2 2 for all xp [ C2.
1
8=V2 1 x p 2 , fp 1 x p, k2 1 x p 2 2 9 5 2 1 zj 2 2 2 1 m 2 1 2 .
2 2
1.5 1.5
1 1
0.5 0.5
z 0 z 0
−0.5 −0.5
−1 −1
−1.5 −1.5
−1 −1
−0.5 −1 −0.5 −1
0 −0.5 0 −0.5
ξ2 0.5 0 ξ2 0.5 0
0.5 ξ1 0.5 ξ1
1 1 1 1
(c) (d)
FIGURE 20 The sets used in the construction of the hybrid controller in Example 36 for q 5 1, q 5 2 , and q 5 3 are shown in (a),
(b), and (d), respectively. In (c), the sets for q 5 2 are depicted with perspective rotated by 180˚. The state x 5 1 j, z 2 evolves
on the cylinder S1 3 R ( R3 , while q [ 5 1, 2, 3 6 . The vertical position 1 j, z 2 5 1 1 0, 1 2 , 0 2 is indicated by a black x . The black
curve represents the set of points where W 1 x 2 J 0.5z2 1 1 1 j2 5 2. To meet the “Supervisor Assumption,” the sets Dq can be
taken to be empty. The sets F q : 5 h i . qC q are indicated in blue, the sets C q ( Cq are indicated in green, the sets Cq\C q are
indicated in yellow, and the sets 1 S1 3 R 2 \ 1 Cq h 1 h i . qC i 2 2 are indicated in red. The sets Hq constructed in the supervisory control
algorithm are the unions of the blue and red regions. The key properties are that, for each q [ 5 1, 2, 3 6 , the vertical position is
globally pre-asymptotically stable for the system xp 5 fp 1 xp, kq 1 xp 2 2 , xp [ Cq , and each solution starting in a green region does
#
not reach a red region. Taken together, these assumptions guarantee that solutions starting in a green region reach a blue
region, for q [ 5 1, 2 6 .
This deduction is equivalent to saying that there are no converge to the set where W 1 x p 2 5 2, which is indicated by
complete solutions that remain in the union of the green the black curve, and moving along this set leads to points
near x *p where the flow x p 5 fp 1 x p, k2 1 x p 2 2 cannot be contin-
#
and yellow regions in Figure 20(b) or, alternatively, Figure
20(c). This behavior is because trajectories in those regions ued in the union of the green and yellow regions.
Next we verify that, for each q [ Q, each solution a collection of functions Vq and sets V q, V qr for each q [ Q,
of x p 5 fp 1 x p, kq 1 x p 2 2 , x p [ Cq is either complete or ends in
#
such that:
F q h 1 S1 3 R 2 \ Cq, where F q 5 h i[Q, i.qC i. This condition i) 5 V q 6 q[Q and 5 V qr 6 q[Q are families of nonempty open
follows immediately from the fact that the second condi- subsets of Rn such that
tion in Proposition 2 holds for each x p [ Cq \ 1 S1 3 R 2 \ Cq.
Thus, maximal solutions that are not complete must end Rn 5 d V q 5 d V qr,
in 1 S1 3 R 2 \ Cq. With respect to Figure 20, this conclu- q[Q q[Q
C2 J 5 j [ S3 : j 4 $ 2 2/3 6 ,
C 2 J 5 j [ S3 : j 4 $ 2 1/3 6 .
independently of how kq is extended from Cq \ A to Cq .
To preserve outer semicontinuity and local bounded-
ness, set kq 1 j 2 5 mq 1 0 2 B for j [ Cq d A . The “Supervisor
Assumption” holds. In particular, conditions 2a) and Following the prescription of the components of
2b) are addressed by the inequalities displayed above, a hybrid supervisor in (41) and (42), we get H1 5 C 2,
while 2c) holds since U 5 Rn . By taking kq 1 xp, xc 2 5 kq 1 xp 2 H2 5 S3 \ C2 5 5 j [ S3 : j 4 # 22/3 6 , and Jq 5 3 2 q. Then,
for each q [ Q, the feedbacks kq can be combined in using the definition of a supervisory controller in (40), we
a hybrid supervisor Hc with data given in (40). Al- get a hybrid controller with state xc 5 q, data Cc 5 1 C1 3 5 1 6 2
though kq are set-valued on A , the arguments used still h 1 C2 3 5 2 6 2 , fc 5 0, Dc 5 1 H1 3 5 1 6 2 h 1 H2 3 5 2 6 2 ,
apply. Furthermore, robustness results like Theorem Gc 5 3 2 q, and kc 1 j, q 2 5 kq 1 j 2 .
15 hold. According to Corollary 33, this controller renders the
set 5 j* 6 3 Q globally asymptotically stable. In fact, it can
Example 37: Global Asymptotic Stabilization be established that the point 1 j*, 2 2 is globally asymptoti-
of a Point on the Three-Sphere cally stable.
Consider the problem of globally asymptotically stabiliz- This example can also be interpreted in terms of the con-
ing the point j* J 1 0, 0, 0, 1 2 [ S3 J 5 x [ R4 : xTx 5 1 6 for cept of patchy control-Lyapunov functions. Indeed, it can
the system be verified that, with Q 5 5 1, 2 6 ,
## V1 1 j 2 5 1 2 j 3,
j 5 L1u2j
V 1 r 5 V 1 J S 3,
j [ S 3,
V2 1 j 2 5 1 2 j 4, V 2 r J 5 j [ S3 : j 4 $ 2 1/3 6 ,
V 2 J 5 j [ S3 : j 4 $ 2 2/3 6 ,
where
constitutes a patchy control-Lyapunov function. ■
0 u3 2u2 u1
Example 38: Nonholonomic Integrator
L1u2 5 ≥ ¥.
2u3 0 u1 u2
Consider the nonlinear control system
u2 2u1 0 u3
2u1 2u2 2u3 0 #
j1 5 u
s 5: fp 1 j, u 2
#
j2 5 u2
These equations can model the orientation kinematics of a #
j3 5 j 1u2 2 j 2u1
rigid body expressed in terms of unit quaternions. In this
case, j* corresponds to a desired orientation and the inputs known as the nonholonomic integrator. According to
ui correspond to angular velocities. Brockett’s necessary condition, the origin is not stabiliz-
First we study the effect of the feedback controls ui J 2j i, able by continuous, static-state feedback, nor it is robustly
i [ 5 1, 2, 3 6 . Denote this feedback k2. Using the Lyapunov- stabilizable by discontinuous, static state feedback. It is
function candidate V 1 j 2 5 1 2 j4, and noting that robustly stabilizable by time-varying feedback and by
use ui 5 %si 1 1 2 . For q 5 5, we use ui 5 %si 1 2 2 . For each q [ Q, 8=V 1 j 2 , L 1 k4 1 x 2 2 j9 # |j1| # "2"V 1 j 2 ,
the resulting feedback law is denoted u 5 kq 1 x 2 .
For q [ 5 1, 2 6 , we define Cq J 5 x [ K : j2q $ m 1 j232q 1 j23 2 6 , for all x [ C4 \ A it follows that
where m [ 1 0, 1/2 2 . For q [ 5 3, 6 6 , we define Cq J
5 x [ K : 2%j 1 $ 0 6 . We also define C5 J 5 x [ K : 2%j2 $ 0 6 8=W 1 x 2 , F4 1 x 2 9 #
r
2 1 , 0,
and C4 J 5 x [ K : t [ 3 0, min 5 "1 2 j4, w 6 46. For q [ Q, "2
we define Dq 5 K \ Cq. where F4 is the closed-loop flow map when q 5 4.
# # #
We pick the flow map so that q 5 % 5 w 5 0. In addition, To assess pre-asymptotic stability of the flow dynamics
on C3, C5 and C6 we use W 1 x 2 5 V 1 j 2 . We obtain
# #
t 5 1 for q 5 4 while t 5 0 for q 2 4.
We pick the jump map G so that x 1 [ G 1 x 2 gives the
following relationships: j 1 5 j; t 1 5 0; for q [ 5 3, 4, 5 6 , 8=V 1 j 2 , L 1 kq 1 x 2 2 j9 5 %jj # 0,
q 1 5 q 1 1; for q [ 5 1, 2, 6 6 , q 1 5 3 when j21 1 j22 , m^ j23,
where m^ . 2m/ 1 1 2 m 2 , q 1 [ I 1 j 2 when j21 1 j22 . m^ j23, where j 5 2 for q 5 5 and j 5 1 for q [ 5 3, 6 6 . This
where I 1 j 2 :5 5 i [ 5 1, 2 6 : j 2i 5 max 5 j 21, j 22 6 6 , and q 1 5 property establishes stability of A , and since j4 is
5 3 6 h I 1 j 2 when j21 1 j22 5 m^ j23. The possible mode transi- bounded away from zero, there are no complete flowing
tions are indicated in Figure 21. solutions on C3 , C5 or C6 , which implies that A is pre-
When q [ 5 1, 2, 5, 6 6 , % 1 [ 5 s [ 5 2 1, 1 6 : 2 sj 1 $ 0 6 , asymptotically stable.
when q 5 3, % 1 5 sgn 1 j2j3 2 for j2j3 2 0 and % 1 5 5 2 1, 1 6 To assess pre-asymptotic stability of the combined flow
for j2j3 5 0; when q 5 4, % 1 [ 5 s [ 5 2 1, 1 6 : sj 2 $ 0 6 , and jump dynamics Cq, Dq, q [ 5 1, 2 6 , we use the Lyapunov
When q 2 3, w 1 5 w; when q 5 3, w 1 5 0.5 cot 21 1 |j2 /j3| 2 function W 1 x 2 5 V 1 j 2 . We obtain
for j3 2 0, w 1 5 0 for j3 5 0 and j2 2 0, and w 1 [ 3 0, p/4 4
jump to q 5 3 has just occurred. By the time a jump to 1 2 x4, 6 # r2 1 x4, 3 2 1 1 2 x4, 3 2 .
q 5 4 occurs, W 1 x 2 has not increased and we have j1 5 0
and j22 # m^ j23 . When q 5 4, W 1 x 2 increases, but we argue Finally, we establish |x3, 5| # r 1 x4, 3 2 |x3, 3|. Let tq,
that the sequence of modes q 5 4, q 5 5, q 5 6 results in a q [ 5 4, 5 6 , denote the time spent in mode q. Let %q,
decrease in W 1 x 2 . Since W 1 x 2 is also strictly decreasing q [ 5 4, 5 6 , denote the value of % in mode q. A routine calcu-
for q [ 5 1, 2 6 , this implies that when q 5 3 is revisited, lation involving the solution of a linear, two-dimensional
W 1 x 2 has decreased. The key to showing this proper- oscillator and using that x1, 3 5 x2, 5 5 0 gives
ty is to establish that, with xi, q denoting ji at the end
of mode q, there exists a continuous function r that is cos 1 %4t4 2 x3, 5 5 cos 1 %5t5 2 3 1 cos 1 %4t4 2 2 2 sin 1 %4t4 2 2 2 x3, 3
less than one except when its argument is one, so that 2 2cos 1 %4t4 2 sin 1 %4t4 2 x2, 3 4 .
|x3, 5| # r 1 x4, 3 2 |x3, 3|. Indeed, using that x1, 6 5 x2, 5 5 0
and x22, 6 1 x23, 6 5 x22, 5 1 x22, 6 , this gives Due to the construction of the jump map from mode q 5 3,
%4 5 sgn 1 x2, 3x3, 3 2 whenx2, 3x3, 3 2 0;otherwise%4 [ 521, 1 6 .
Then, due to the value of w in mode q 5 4, which lim-
1 2 x 24, 6 5 x 21, 6 1 x 22, 6 1 x 23, 6 its t4, it follows that the sign of sin 1 %4t4 2 x2, 3 is the same
5 x 22, 6 1 x 23, 6 as the sign of x3, 3. Using t4 # p/4 and %4 [ 5 2 1, 1 6 ,
5 x 22, 5 1 x 23, 5 it follows that
5 x 23, 5
# r 2 1 x4, 3 2 x 23, 3
cos 1 t4 2 2 2 sin 1 t4 2 2
# r 2 1 x4, 3 2 1 1 2 x 24, 3 2 .
cos 1 t4 2
|x3, 5| # |x3, 3|.
Since x4, 6 and x4, 3 are positive, this implies The function involving t4 on the right-hand side takes
values in the interval 3 0, 1 2 for all
t4 [ 1 0, p/4 4 . Now, by the definition of
C4 , k4 , and the fact that x1, 3 5 0, t4 can
be expressed as a continuous function
q=1 of x4, 3 that is zero when x4, 3 5 1 and is
positive otherwise. In turn, this estab-
lishes the bound |x3, 5| # r 1 x4, 3 2 |x3, 3|
q=3 q=4 q=5 q=6 for a continuous, nonnegative-valued
function r that is less than one except
when x4, 3 5 1. ■
q=2
CONCLUSIONS
Hybrid dynamical systems combine
flows and jumps. They can be modeled in
FIGURE 21 Possible mode transitions when stabilizing a point on the three-sphere by a compact form, and they cover a fasci-
hybrid feedback that uses only two angular velocities, one at a time. Normal opera-
tion corresponds to jumps between modes q 5 1 and q 5 2 . However, at times the
nating variety of dynamic phenomena.
sequence of modes 3 S 4 S 5 S 6 is required. The initiation of such a sequence is With the use of hybrid time domains
associated with an event in the hybrid system. Transitions associated with events are and the notion of graphical conver-
indicated by red arrows in the diagram. The quantity V 1 j 2 5 1 2 j 4 does not increase gence, sequential compactness of the
along solutions except when q 5 4 . When events are inhibited, so that the mode q 5 4 space of solutions and semicontinuous
is reached no more than once, the resulting hybrid system has the correct pre-as-
ymptotically stable set. With events included and V 1 j 2 considered at event times, the
dependence of solutions on initial con-
resulting values of V 1 j 2 constitute a strictly decreasing sequence. Thus, Theorem 32 ditions and perturbations can be estab-
can be used to establish pre-asymptotic stability for the hybrid system. lished under mild conditions.
The properties of the space of solutions to a hybrid Buenos Aires, Argentina, in 2001. He joined the Center
system have important consequences for stability theory. for Control, Dynamical Systems, and Computation at the
For example, these properties imply that asymptotic sta- University of California, Santa Barbara in 2002, where he
bility of a compact set is uniform and robust to pertur- received the M.S. and Ph.D. degrees in 2004 and 2007, re-
bations. The properties also facilitate extensions of the spectively. In 2007 and 2008, he held postdoctoral positions
classical invariance principle and converse Lyapunov at the Laboratory for Information and Decision Systems at
theorems to the hybrid setting. The hybrid invariance the Massachusetts Institute of Technology and at the Centre
principle and Lyapunov functions lend themselves to Automatique et Systèmes at the École de Mines de Paris. In
natural sufficient conditions for asymptotic stability in a 2009, he joined the faculty of the Department of Aerospace
hybrid system. Additional stability analysis tools, relat- and Mechanical Engineering at the University of Arizona,
ed to identifying and limiting events in a hybrid system, where he is currently an assistant professor. His research
can also be developed. interests are in modeling, stability, robust control, and simu-
The stability analysis tools can be used to predict the be- lation of nonlinear, hybrid, and embedded systems with ap-
havior of hybrid systems and to design hybrid control algo- plications to robotics, aerospace, and biology.
rithms. The examples of hybrid control systems provided Andrew R. Teel ([email protected]) received the A.B.
in this article only scratch the surface of what is possible degree in engineering sciences from Dartmouth College
using hybrid feedback control. The framework and tools in 1987 and the M.S. and Ph.D. degrees in electrical en-
presented in this article may help in the process of discov- gineering from the University of California, Berkeley, in
ering new hybrid feedback control ideas. 1989 and 1992, respectively. After receiving the Ph.D., he
was a postdoctoral fellow at the École des Mines de Paris
ACKNOWLEDGMENTS in Fontainebleau, France. From 1992 to 1997 he was a fac-
Many of the developments reported in this article are a ulty member in the Electrical Engineering Department at
result of research supported by the Air Force Office of the University of Minnesota. In 1997, he joined the faculty
Scientific Research under grant number F9550-06-1-0134, of the Electrical and Computer Engineering Department
the National Science Foundation under grant numbers at the University of California, Santa Barbara, where he
ECS-0622253 and CNS-0720842, and the Army Research is currently a professor. He has received NSF Research
Office. Initiation and CAREER Awards, the 1998 IEEE Leon K.
Kirchmayer Prize Paper Award, the 1998 George S. Axel-
AUTHOR INFORMATION by Outstanding Paper Award, and the SIAM Control and
Rafal Goebel received the M.Sc. degree in mathematics Systems Theory Prize in 1998. He was the recipient of
in 1994 from the University of Maria Curie Sklodowska the 1999 Donald P. Eckman Award and the 2001 O. Hugo
in Lublin, Poland, and the Ph.D. degree in mathematics Schuck Best Paper Award, both given by the American Au-
in 2000 from the University of Washington, Seattle. He tomatic Control Council. He is a Fellow of the IEEE. He
held a postdoctoral position at the Departments of Math- can be contacted at the University of California, Electrical
ematics at the University of British Columbia and Simon and Computer Engineering Department, Santa Barbara,
Fraser University in Vancouver, Canada, 2000–2002; a CA 93106-9560 USA.
postdoctoral and part-time research positions at the
Electrical and Computer Engineering Department at the REFERENCES
University of California, Santa Barbara, 2002–2005; and a [1] D. Aeyels and J. Peuteman, “A new asymptotic stability criterion for non-
part-time teaching position at the Department of Math- linear time-variant differential equations,” IEEE Trans. Automat. Contr., vol.
43, no. 7, pp. 968–971, July 1998.
ematics at the University of Washington, 2004–2007. In [2] Z. Artstein, “Stabilization with relaxed controls,” Nonlinear Anal., vol.
2008, he joined the Department of Mathematics and 11, pp. 1163–1173, 1983.
Statistics at Loyola University Chicago. His interests in- [3] K. J. Astrom and B. Bernhardsson, “Systems with Lebesgue sampling,”
in Directions in Mathematical Systems Theory and Optimization, A. Rantzer
clude convex, nonsmooth, and set-valued analysis; con- and C. I. Byrnes, Eds. New York: Springer-Verlag, 2003, pp. 1–13.
trol, including optimal control; hybrid dynamical sys- [4] J.-P. Aubin and G. Haddad, “Cadenced runs of impulse and hy-
tems; and optimization. brid control systems,” Int. J. Robust Nonlinear Control, vol. 11, no. 5, pp.
401–415, 2001.
Ricardo G. Sanfelice received the B.S. degree in elec- [5] E. A. Barbašin and N. N. Krasovskiı̆, “On stability of motion in the large,”
tronics engineering from the Universidad de Mar del Plata, Doklady Akad. Nauk SSSR, vol. 86, pp. 453–456, 1952.
W ith no other choice but to live one day at a time and to keep intellectually alive, Tsien contin-
ued to work and teach and await the next step, whatever it might be, whenever it might be.
He turned to other fields of research, such as the study of games and economic behavior. In 1954,
he published a textbook titled Engineering Cybernetics, a book on systems of communication and
control. It too would be well received.
Years later, Wallace Vander Velde, an MIT professor and renowned expert in cybernetics, would
describe the book as “remarkable” and “an extraordinary achievement in its time.” Wrote Vander
Velde of the book:
In 1954, a decent theory of feedback control for linear, time invariant systems existed and
servomechanism design was an established practice. But Tsien was looking ahead to more
complex control and guidance problems—notably the guidance of rocket-propelled vehi-
cles. This stimulated his interest in the systems with time-varying coefficients, time lag and
nonlinear behavior. All these topics are treated in this book.
But Tsien went further to deal with optimal control via the variational calculus, optimalizing
control and fault-tolerant control systems among other topics! He visualized a theory of guidance
and control which would be distinct from, and would support, the practice of these disciplines.
This has certainly come to be, and his pioneering effort may be thought of as a major foundation
stone of that effort which continues to this day.
—Thread of the Silkworm, by Iris Chang, BasicBooks, New York, 1994, pp. 175–176.