A TUTORIAL ON SOLUTIONS,

NONSMOOTH ANALYSIS,
AND STABILITY

JORGE CORTÉS

© PHOTOCREDIT

iscontinuous dynamical systems arise in a large manipulation of objects by means of mechanical contact

D number of applications. In optimal control

problems, open-loop, bang-bang controllers
switch discontinuously between extreme values
of the inputs to generate minimum-time trajec-
tories from the initial to the final states [1]. Thermostats
implement on-off controllers to regulate room tempera-
ture [2]. When the temperature is above the desired
[5], discontinuities occur naturally from interaction with
the environment.
Discontinuities are also intentionally designed to
achieve regulation and stabilization. Sliding mode control
[6], [7] uses discontinuous feedback controllers for stabi-
lization. The design procedure for sliding mode control
begins by identifying a surface in the state space with the
value, the controller switches the cooling system on. property that the dynamics of the system restricted to this
Once the temperature reaches a preset value, the con- surface are easily stabilizable. Feedback controllers are
troller switches the cooling system off. The controller is then synthesized on each side of the surface to steer the
therefore a discontinuous function of the room tempera- solutions of the system toward the surface. The resulting
ture. In nonsmooth mechanics, the motion of rigid bod- closed-loop system is discontinuous on the surface. In
ies is subject to velocity jumps and force discontinuities robotics [8], it is of interest to induce emergent behavior in
as a result of friction and impact [3], [4]. In the robotic a swarm of robots by prescribing interaction rules among
individual agents. Simple laws such as “move away from
Digital Object Identifier 10.1109/MCS.2008.919306 the nearest other robot or environmental boundary” give

36 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008 1066-033X/08/$25.00©2008IEEE

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 rise to discontinuous dynamical systems. For example, Figure 2(a)]. During sliding, the Coulomb friction model
consider a robot placed in a corner of a room. On opposite states that the magnitude of the friction force is independent
sides of the bisector line between the two walls forming of the magnitude of the velocity and is equal to the normal
the corner, the “move away” law translates into different contact force times the coefficient of friction. The application
velocity vectors for the robot. The dynamical system is of this model to the sliding brick yields
thus discontinuous along the bisector line. In optimization
[9], [10], continuous-time algorithms that perform general- v̇(t) = g(sin θ) − νg(cos θ)sign(v(t)), (2)
ized gradient descent, which is a discontinuous function
of the state, are used when the objective function is not where v is the velocity of the brick, g is the acceleration
smooth. In adaptive control, [11] switching algorithms are due to gravity, θ > 0 is the inclination of the ramp, ν is
employed to select the most appropriate controller from a the coefficient of friction, and sign(0) = 0. The right-hand
given finite family to enhance robustness, ensure bound- side of (2) is a discontinuous function of v because of the
edness of the estimates, and prevent the system from step- presence of the sign function. Figure 2(b) shows the phase
ping into undesired regions of the state space. plot of this system for several values of ν.
Many control systems cannot be stabilized by continu- Depending on the magnitude of the friction force,
ous state-dependent feedback. As a consequence, it is nec- experiments show that the brick stops and stays stopped.
essary to consider either time-dependent or discontinuous In other words, the brick attains v = 0 in finite time and
feedback. As an illustration, consider the one-dimensional maintains v = 0. However, there is no continuously differ-
system ẋ = X(x, u) with an equilibrium at the origin for entiable solution of (2) that exhibits this type of behavior.
u = 0. When x > 0, we look for u such that X(x, u) < 0 (“go To see this, note that v = 0 and v̇ = 0 in (2) imply sin θ = 0,
left”). Likewise, when x < 0, we look for u such that which contradicts θ > 0. To explain this type of physical
X(x, u) > 0 (“go right”). Together, these conditions can be evolution, we need to understand the discontinuity in (2)
stated as xX(x, u) < 0. When the control u appears nonlin- and expand our notion of solution beyond continuously
early in X , it may be the case that no continuous state- differentiable solutions. 
dependent feedback controller x → u(x) exists such that
xX(x, u(x)) < 0 for all x ∈ R. The following example illus- Example 3: Nonsmooth Harmonic Oscillator
trates the above discussion. Consider a unit mass subject to a discontinuous spring
force. The spring does not exert any force when the mass is
Example 1: System Requiring
Discontinuous Stabilization
Consider the one-dimensional system [12]
u

ẋ = x[(u − 1)2 − (x − 1)][(u + 1)2 + (x − 2)]. (1)

2
The shaded areas in Figure 1 represent the regions in the
space (x, u) where xX(x, u) < 0. From the plot, it can be
seen that there exists no continuous function x → u(x) 1
defined on R whose graph belongs to the union of the
shaded areas. Even control systems whose inputs appear
linearly (see Example 21) may be subject to obstructions x
−2 −1 1 2 3
that preclude the existence of continuous state-dependent
stabilizing feedback [12]–[14]. 
Numerous fundamental questions arise when dealing −1
with discontinuous dynamical systems. The most basic
question is the notion of a solution. For a discontinuous
vector field, the existence of a continuously differentiable −2
solution, that is, a continuously differentiable curve
whose derivative follows the direction of the vector field,
is not guaranteed. The following examples illustrate the
−3
difficulties that arise in defining solutions of discontinu-
ous dynamical systems.
FIGURE 1 Obstruction to the existence of continuous state-dependent
feedback [12]. The shaded areas represent the values of the control u
Example 2: Brick on a Frictional Ramp that are needed to stabilize the equilibrium point 0 of system (1). There
Consider a brick sliding on a ramp [15]. As the brick moves, does not exist a function x → u(x) defined on R that at the same time
it experiences a friction force in the opposite direction [see is continuous and has a graph that belongs to the shaded areas.

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 37

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 at the reference position x = 0. When the mass is displaced which is a nonsmooth version of the classical harmonic
to the right, that is, x > 0, the spring exerts a constant neg- oscillator. The phase portrait of this system is plotted in
ative force that pulls it back to the reference position. Figure 3(a).
When the mass is displaced to the left, that is, x < 0, the It can be seen that (0, 0) is the unique equilibrium of
spring exerts a constant positive force that pulls it back to (4)–(5). By discretizing the equations of motion, we find
the reference position. According to Newton's second law, that the trajectories of the system approach, as the time
the system evolution is described by [16] step is made smaller and smaller, the set of curves in
Figure 3(b), which are the level sets of the function
ẍ + sign(x) = 0. (3) (x1 , x2 ) → |x1 | + x22 /2 . These level sets are analogous to the
level sets of the function (x1 , x2 ) → x21 + x22 , which are the
By defining the state variables x1 = x and x2 = ẋ, (3) can be trajectories of the classical harmonic oscillator ẍ + x = 0
rewritten as [17]. However, the trajectories in Figure 3(b) are not con-
tinuously differentiable along the x2 -coordinate axis since
ẋ1 (t) = x2 (t), (4) the limiting velocity vectors from the right of the axis and
from the left of the axis do not coincide. 
ẋ2 (t) = −sign(x1 (t)), (5)

Example 4: Move-Away-from-
Nearest-Neighbor Interaction Law
Consider n nodes p1 , . . . , pn evolving in a square Q
according to the interaction rule “move diametrically
away from the nearest neighbor.” Note that this rule is not
defined when two nodes are located at the same point,
that is, when the configuration is an element of S 
{(p1 , . . . , pn ) ∈ Qn : pi = pj for some i = j} . We thus
consider only configurations that belong to Qn \S . Next,
we define the map that assigns to each node its nearest
F υ
neighbor, where the nearest neighbor may be an element
of the boundary bndry(Q) of the square. Note that the
mg nearest neighbor of a node might not be unique, that is,
θ more than one node can be located at the same (nearest)
distance. Hence, we define the nearest-neighbor map
N = (N1 , . . . , Nn ) : Qn \ S → Qn by arbitrarily selecting,
(a)
for each i ∈ {1, . . . , n}, an element,
2
Ni(p1 , . . . , pn ) ∈ argmin{||pi − q||2
1.5 : q ∈ bndry(Q) ∪ {p1 , . . . , pn } \ {pi}},

ν 1
where argmin stands for the minimizing value of q, and
0.5
|| · ||2 denotes the Euclidean norm. By definition,
Ni(p1 , . . . , pn ) = pi . For i ∈ {1, . . . , n}, we thus define the
0 move-away-from-nearest-neighbor interaction law
−2 −1 0 1 2
υ pi(t) − Ni(p1 (t), . . . , pn (t))
(b) ṗi = . (6)
||pi(t) − Ni(p1 (t), . . . , pn (t))||2

FIGURE 2 Brick sliding on a frictional ramp with velocity ν. (a) The Changes in the value of the nearest-neighbor map N induce
physical quantities used to describe the example and (b) the one- discontinuities in the dynamical system. For instance, con-
dimensional phase portraits of (2) corresponding to values of the fric-
tion coefficient ν between zero and two, with a ramp inclination of
sider a node sufficiently close to a vertex of bndry(Q) so
30◦ . The phase portrait shows that, for sufficiently small values of ν, that the closest neighbor to the node is an element of the
every trajectory that starts with positive initial velocity (“moving to the boundary. Depending on how the node is positioned with
right”) never stops. However, for sufficiently large values of ν, every respect to the bisector line passing through the vertex, the
trajectory that starts with positive initial velocity eventually stops and node computes different directions of motion; see Figure 4
remains stopped. However, there is no continuously differentiable
solution of (2) that exhibits this type of behavior. We thus need to
for an illustration.
expand our notion of solution beyond continuously differentiable To analyze the dynamical system (6), we need to
solutions by understanding the effect of the discontinuity in (2). understand how the discontinuities affect its evolution.

38 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 1
1

.5 .5

x2
0
x2

−.5 −.5

−1
−1
−1 −.5 0 .5 1 −1 −.5 0 .5 1
x1 x1
(a) (b)

FIGURE 3 Nonsmooth harmonic oscillator. (a) The phase portrait on [−1, 1]2 of the vector field (x1 , x 2 ) → (x 2 , −sign (x1 )) and (b) the
contour plot on [−1, 1]2 of the function (x1 , x 2 ) → |x1 | + (x 22 /2). The discontinuity of the vector field along the x 2 -coordinate axis
makes it impossible to find continuously differentiable solutions of (10). On the other hand, the level sets in (b) match the phase por-
trait in (a) everywhere except for the x 2 -coordinate axis, which suggests that trajectories along the level sets are candidates for solu-
tions of (10) in a sense different from the classical one.

Since each node moves away from its nearest neighbors, The function γ : [a, b] → R is absolutely continuous if, for
it is reasonable to expect that the nodes never run into all  ∈ (0, ∞), there exists δ ∈ (0, ∞) such that, for each
each other, however, rigorous verification of this proper- finite collection {(a1 , b1 ), . . . , (an , bn )} of disjoint open

ty requires a proof. We would also like to characterize intervals contained in [a, b] with ni=1 (bi − ai) < δ, it fol-
the asymptotic behavior of the trajectories of the system lows that
(6). To study these questions, we need to extend our 
n
notion of solution.  |γ (bi) − γ (ai)| < .
i=1
Beyond Continuously Differentiable Solutions
Examples 2–4 can be described by a dynamical system of Equivalently [12], γ is absolutely continuous if there exists
the form a Lebesgue integrable function κ : [a, b] → R such that

ẋ(t) = X(x(t)), x(t0 ) = x0 , (7)

where x ∈ Rd , d is a positive integer, and X : Rd → Rd is

not necessarily continuous. We refer to a continuously dif-
ferentiable solution t → x(t) of (7) as classical. Clearly, if X
is continuous, then every solution is classical. Without loss
of generality, we take t0 = 0. We consider only solutions
that run forward in time.
Examples 2–4 illustrate the limitations of classical solu-
tions and confront us with the need to identify a suitable
notion for solutions of (7). Unfortunately, there is not a
unique answer to this question. Depending on the prob-
(a) (b)
lem and objective at hand, different notions are appropri-
ate. In this article, we restrict our attention to solutions
that are absolutely continuous. Although not treated in FIGURE 4 Move-away-from-nearest-neighbor interaction law. A node
detail in this article, it is also possible to consider solutions in (a) computes different directions of motion if placed slightly to the
right or to the left of the bisector line defined by two polygonal
that admit discontinuities and hence are not absolutely boundaries. A node in (b) computes different directions of motion if
continuous. These solutions are discussed in “Solutions placed slightly to the right or to the left of the half plane defined by
with Jumps.” two other nodes. In both cases, the vector field is discontinuous.

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 39

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
  t tion of the vector field at all times, that is, the differential
γ (t) = γ (a) + κ(s)ds, t ∈ [a, b].
a equation (7) need not be satisfied on a set of measure
zero. As shown in this article, Caratheodory solutions
Every absolutely continuous function is continuous. How- exist for Example 3 but do not exist for examples 2 and 4.
ever, the converse is not true, since the function Alternatively, Filippov solutions [18] replace the differ-
γ : [−1, 1] → R defined by γ (t) = t sin(1/ t) for t = 0 and ential equation (7) by a differential inclusion of the form
γ (0) = 0 is continuous but not absolutely continuous.
Moreover, every continuously differentiable function is ẋ(t) ∈ F(x(t)), (9)
absolutely continuous, but the converse is not true. For
instance, the function γ : [−1, 1] → R defined by γ (t) = |t| where F : Rd → B(Rd), and B(Rd) denotes the collection
is absolutely continuous but not continuously differen- of all subsets of Rd . Filippov solutions are absolutely contin-
tiable at zero. As this example suggests, every absolutely uous curves. At a given state x, instead of focusing on the
continuous function is differentiable almost everywhere. value of the vector field at x, the idea behind Filippov solu-
Finally, every locally Lipschitz function (see “Locally Lip- tions is to introduce a set of directions that are determined
schitz Functions”) is absolutely continuous, but the con- by the values of the vector field X in a neighborhood of x.
verse is not true. For instance, the function γ : [0, 1] → R Differential inclusions [19], [20] thus involve set-valued maps.
√
defined by γ (t) = t is absolutely continuous but not Just as a standard map or function takes a point in its
locally Lipschitz at zero. domain to a single point in another space, a set-valued map
Caratheodory solutions [18] are a generalization of clas- takes a point in its domain to a set of points in another space.
sical solutions. Roughly speaking, Caratheodory solutions A differential inclusion thus specifies that the state derivative
are absolutely continuous curves that satisfy the integral belongs to a set of directions, rather than being a specific
version of the differential equation (7), that is, direction. This flexibility is crucial for providing conditions
on the discontinuous vector field under which Filippov solu-
 t
tions exist. This solution notion plays a key role in many of
x(t) = x(t0 ) + X(x(s))ds, t > t0 , (8)
t0 the applications mentioned above, including sliding mode
control and mechanics with Coulomb-like friction.
where the integral is the Lebesgue integral. By using the Unfortunately, the obstructions to continuous stabiliza-
integral form (8), Caratheodory solutions relax the classi- tion illustrated in Example 1 also hold [21], [22] for Filip-
cal requirement that the solution must follow the direc- pov solutions. Sample-and-hold solutions [23], which are

Solutions with Jumps

I
n this article, we focus entirely on absolutely continuous solutions [S3] R.W. Cottle, J.-S. Pang, and R.E. Stone, The Linear Comple-
of ordinary differential equations. However, nonsmooth continu- mentarity Problem. New York: Academic, 1992.
ous-time systems can possess discontinuous solutions that admit [S4] G. Isac, Complementarity Problems. New York: Springer-Ver-
jumps in the state. Such notions are appropriate for dealing with lag, 1992.
mechanical systems, for instance, subject to unilateral constraints [S5] M.C. Ferris and J.-S. Pang, “Engineering and economic applica-
[15]. As an example, a bouncing ball hitting the ground experiences tions of complementarity problems,” SIAM Rev., vol. 39, no. 4, pp.
an instantaneous change of velocity. This change corresponds to a 669–713, 1997.
discontinuous jump in the trajectory describing the evolution of the [S6] B. Brogliato, “Some perspectives on the analysis and control of
velocity of the ball. Both measure differential inclusions [S1], [S2] complementarity systems,” IEEE Trans. Automat. Contr., vol. 48, no.
and linear complementarity systems [S3]–[S6] are approaches that 6, pp. 918–935, 2003.
specifically allow for discontinuous solutions. Within hybrid systems [S7] P.J. Antsaklis and A. Nerode, guest eds., “Special issue on
theory [S7]–[S10], discontinuous solutions arise in systems that hybrid control systems,” IEEE Trans. Automat. Contr., vol. 43, no. 4,
involve both continuous- and discrete-time evolutions. 1998.
[S8] A.S. Morse, C.C. Pantelides, S.S. Sastry, and J.M. Schu-
REFERENCES macher, Guest Eds., “Special issue on hybrid systems,” Automat-
[S1] J.-J. Moreau, “Unilateral contact and dry friction in finite free- ica, vol. 35, no. 3, 1999.
dom dynamics,” in Non-Smooth Mechanics and Applications, vol. [S9] A.J. van der Schaft and H. Schumacher, An Introduc-
302, J.-J. Moreau and P.D. Panagiotopoulos, Eds. New York: tion to Hybrid Dynamical Systems. New York: Springer-Ver-
Springer-Verlag, 1988, pp. 1–82. lag, 2000.
[S2] M.D. P.M. Marques, Differential Inclusions in Nonsmooth [S10] R. Goebel and A.R. Teel, “Solutions of hybrid inclusions via set
Mechanical Problems: Shocks and Dry Friction. Cambridge, MA: and graphical convergence with stability theory applications,” Auto-
Birkhäuser, 1993. matica, vol. 42, no. 4, pp. 573–587, 2006.

40 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 also absolutely continuous curves, turn out to be the tion. Therefore, for example, “weakly stable equilibrium
appropriate notion for circumventing these obstructions point” means that at least one solution starting close to the
[24]–[26]. “Additional Solution Notions for Discontinuous equilibrium point remains close to it, whereas “strongly
Systems” describes additional solution notions for discon- stable equilibrium point” means that all solutions starting
tinuous systems. In this article, we focus on Caratheodory, close to the equilibrium point remain close to it. For
Filippov, and sample-and-hold solutions. detailed definitions; see [14] and [18].
We present weak and strong stability results for discon-
Existence, Uniqueness, and Stability of Solutions tinuous dynamical systems and differential inclusions. As
In addition to the notion of solution, we consider existence suggested by Example 3, smooth Lyapunov functions do
and uniqueness of solutions as well as stability. For ordi- not suffice to analyze the stability of discontinuous sys-
nary differential equations, it is well known that continuity tems. This fact leads naturally to the subject of nonsmooth
of the vector field does not guarantee uniqueness. Not sur- analysis. In particular, we pay special attention to the gen-
prisingly, no matter what notion of solution is chosen for a eralized gradient of a locally Lipschitz function [9] and the
discontinuous vector field, nonuniqueness can occur. We proximal subdifferential of a lower semicontinuous func-
thus provide sufficient conditions for uniqueness. We also tion [24]. Building on these notions, weak and strong
present results specifically tailored to piecewise continu- monotonicity properties of candidate Lyapunov functions
ous vector fields and differential inclusions. can be established along the solutions of discontinuous
The lack of uniqueness of solutions must be considered dynamical systems. These results are used to provide gen-
when we try to establish properties such as local stability. eralizations of Lyapunov stability theorems and the invari-
This issue is reflected in the use of the adjectives weak and ance principle, which help us study the stability of
strong. The word “weak” is used when a property is satis- solutions. To illustrate the applicability of these results, we
fied by at least one solution starting from each initial condi- discuss in detail a class of nonsmooth gradient flows.
tion. On the other hand, “strong” is used when a property There are two ways to apply the stability results pre-
is satisfied by all solutions starting from each initial condi- sented here to control systems. The first way is to

Additional Solution Notions for Discontinuous Systems

S
everal solution notions are available in addition to Caratheodory, [S14] V.A. Yakubovich, G.A. Leonov, and A.K. Gelig, in Stability of
Filippov, and sample-and-hold solutions. These notions include Stationary Sets in Control Systems With Discontinuous Nonlineari-
the ones considered by Krasovskii [34], Hermes [S11], [S12], ties, vol. 14. Singapore: World Scientific, 2004.
Ambrosio [S13], Sentis [35], and Yakubovich-Leonov-Gelig [S14]; [S15] J.S. Spraker and D.C. Biles, “A comparison of the
see Table S1. As in the case in which the vector field is continuous, Caratheodory and Filippov solution sets,” J. Math. Anal. Appl., vol.
Euler solutions [18], [24] are useful for establishing existence and in 198, no. 2, pp. 571–580, 1996.
characterizing basic mathematical properties of the dynamical sys- [S16] S. Hu, “Differential equations with discontinuous right-hand
tem. Additional notions of solutions for discontinuous systems are sides,” J. Math. Anal. Appl., vol. 154, no. 2, pp. 377–390, 1991.
[S17] A. Bacciotti, “Some remarks on generalized solutions of dis-
provided in [S14, sect.1.1.3]. With so many notions of solution avail-
continuous differential equations,” Int. J. Pure Applied Math., vol. 10,
able, various works explore the relationships among them. For
no. 3, pp. 257–266, 2004.
example, Caratheodory and Filippov solutions are compared in
[S15]; Caratheodory and Krasovskii solutions are compared in
[S16]; Caratheodory, Euler, sample-and-hold, Filippov, and
TABLE S1 Several notions of solution for discontinuous
Krasovskii solutions are compared in [16]; Hermes, Filippov, and dynamics. Depending on the specific problem, some
Krasovskii solutions are compared in [S12]; and Caratheodory, notions give more physically meaningful solution
Euler, and Sentis solutions are compared in [S17]. trajectories than others.

Notion of solution References

REFERENCES Caratheodory [18]
[S11] H. Hermes, “Discontinuous vector fields and feedback con- Filippov [18]
trol,” in Differential Equations and Dynamical Systems. New York: Krasovskii [34]
Academic, 1967, pp. 155–165. Euler [18], [24]
[S12] O. Hájek, “Discontinuous differential equations I,” J. Differ. Sample-and-hold [23]
Equ., vol. 32, pp. 149–170, 1979. Hermes [S11], [S12]
Sentis [35], [S17]
[S13] L. Ambrosio, “A lower closure theorem for autonomous orien-
Ambrosio [S13]
tor fields,” Proc. R. Soc. Edinb. A, Math, vol. A110, no. 3/4, pp. Yakubovich-Leonov-Gelig [S14]
249–254, 1988.

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 41

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 choose a specific input function and consider the result- we introduce the generalized gradient and proximal subdif-
ing dynamical system. The second way is to associate ferential from nonsmooth analysis and present various tools
with the control system the set-valued map that assigns for their explicit computation. Then, we develop analysis
each state to the set of all vectors generated by the results to characterize the stability and asymptotic conver-
allowable inputs and consider the resulting differential gence properties of the solutions of discontinuous dynami-
inclusion. Rather than focusing on a particular input, cal systems. We illustrate these nonsmooth stability results
this viewpoint allows us to consider the full range of by means of examples, paying special attention to gradient
trajectories of the control system. To analyze the stabili- systems. Throughout the discussion, we interchangeably
ty of the control system under this approach, we can use use “differential equation,” “dynamical system,” and “vec-
the nonsmooth tools developed for differential inclu- tor field.” For reference, “Index of Symbols” summarizes
sions. We explore this idea in detail. the notation used throughout.
Given the large body of work on discontinuous sys- To simplify the presentation, we have chosen to
tems, our aim is to provide a clear exposition of a few restrict our attention to time-invariant vector fields,
useful and central results. “Additional Topics on Dis- although most of the development can be adapted to the
continuous Systems and Differential Inclusions” briefly time-varying setting. We briefly discuss time-varying
discusses issues that are not considered in the main systems in “Caratheodory Conditions for Time-Varying
exposition. Vector Fields” and “Caratheodory Solutions of Differen-
tial Inclusions.” Likewise, for simplicity, we mostly con-
Organization of This Article sider vector fields defined over the whole Euclidean
We start by reviewing basic results on the existence and space, although the exposition can be carried out in more
uniqueness of classical, that is, continuously differentiable, general settings such as open and connected subsets of
solutions of ordinary differential equations. We also present the Euclidean space.
several examples in which the vector field fails to satisfy the
standard smoothness properties. We then introduce various EXISTENCE AND UNIQUENESS FOR
notions of solution for discontinuous systems, discuss exis- ORDINARY DIFFERENTIAL EQUATIONS
tence and uniqueness results, and present useful tools for In this section, we review basic results on existence and
analysis. In preparation for the statement of stability results, uniqueness of classical solutions for ordinary differential

Additional Topics on Discontinuous Systems and Differential Inclusions

B
eyond the topics discussed in this article, we briefly mention 37, pp. 813–840, 1999.
several that are relevant to systems and control. These top- [S19] F.H. Clarke, Y.S. Ledyaev, L. Rifford, and R.J. Stern, “Feed-
ics include continuous dependence of solutions with respect to back stabilization and Lyapunov functions,” SIAM J. Control Optim.,
initial conditions and parameters [18], [19], robustness proper- vol. 39, no. 1, pp. 25–48, 2000.
ties against external disturbances and state measurement [S20] A.R. Teel and L. Praly, “A smooth Lyapunov function from a class
errors [24], [S18], conditions for the existence of periodic solu- KL estimate involving two positive semidefinite functions,” ESAIM J. Con-
tions [18], [20], bifurcations [38], converse Lyapunov theorems trol, Optim. Calculus Variations, vol. 5, pp. 313–367, 2000.
[S19], [S20], controllability of differential inclusions [20], [24], [S21] J.P. Aubin, Viability Theory. Cambridge, MA: Birkhäuser, 1991.
output tracking in differential inclusions [43], systems subject to [S22] A. Nagurney and D. Zhang, Projected Dynamical Systems
gradient nonlinearities [S14], viability theory [S21], maximal and Variational Inequalities with Applications. Norwell, MA: Kluwer,
monotone inclusions [19], projected dynamical systems and 1996.
variational inequalities [S22], and numerical methods for dis- [S23] A. Dontchev and F. Lempio, “Difference methods for differential
continuous systems and differential inclusions [S23]–[S25]. inclusions: A survey,” SIAM Rev., vol. 34, pp. 263–294, 1992.
Several works report equivalence results among different [S24] F. Lempio and V. Veliov, “Discrete approximations of differential
approaches to nonsmooth systems, including [19] on the equiv- inclusions,” Bayreuth. Math. Schr., vol. 54, pp. 149–232, 1998.
alence between differential inclusions and projected dynamical [S25] V. Acary and B. Brogliato, Numerical Methods for Nonsmooth
systems and [S26] on the equivalence of these formalisms with Dynamical Systems: Applications in Mechanics and Electronics.
complementarity systems. New York: Springer-Verlag, 2008.
[S26] B. Brogliato, A. Daniilidis, C. Lemaréchal, and V. Acary, “On
REFERENCES the equivalence between complementarity systems, projected sys-
[S18] Y.S. Ledyaev and E.D. Sontag, “A Lyapunov characterization tems, and differential inclusions,” Syst. Control Lett., vol. 55, no. 1,
of robust stabilization,” Nonlinear Anal., Theory, Methods, Appl., vol. pp. 45–51, 2006.

42 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 equations. We also present examples that do not satisfy the Proposition 1
hypotheses of these results but nevertheless exhibit exis- Let X : Rd → Rd be continuous. Then, for all x0 ∈ Rd , there
tence and uniqueness of classical solutions as well as exists a classical solution of (10) with initial condition
examples that do not possess such desirable properties. x(0) = x0 .
The following example shows that, if the vector field is
Existence of Classical Solutions discontinuous, then classical solutions of (10) might not exist.
Consider the differential equation
Example 5: Discontinuous Vector Field
ẋ(t) = X(x(t)), (10) with Nonexistence of Classical Solutions
Consider the vector field X : R → R defined by
where X : Rd → Rd is a vector field. The point xe ∈ Rd is
an equilibrium of (10) if 0 = X(xe ). A classical solution of 
(10) on [0, t1 ] is a continuously differentiable map −1, x > 0,
X(x) = (11)
x : [0, t1 ] → Rd that satisfies (10). Note that, without loss 1, x ≤ 0,
of generality, we consider only solutions that start at
time zero. Usually, we refer to t → x(t) as a classical which is discontinuous at zero [see Figure 5(a)]. Suppose that
solution with initial condition x(0). We sometimes write there exists a continuously differentiable function
the initial condition as x0 instead of x(0). The solution x : [0, t1 ] → R such that ẋ(t) = X(x(t)) and x(0) = 0. Then
t → x(t) is maximal if it cannot be extended forward in ẋ(0) = X(x(0)) = X(0) = 1, which implies that, for all positive
time, that is, if t → x(t) is not the result of the truncation t sufficiently small, x(t) > 0 and hence ẋ(t) = X(x(t)) = −1,
of another solution with a larger interval of definition. which contradicts the fact that t → ẋ(t) is continuous. Hence,
Note that the interval of definition of a maximal solution no classical solution starting from zero exists. 
is either of the form [0, T), where T > 0, or [0, ∞). In contrast to Example 5, the following example shows
Continuity of the vector field suffices to guarantee the exis- that the lack of continuity of the vector field does not
tence of classical solutions, as stated by Peano’s theorem [27]. preclude the existence of classical solutions.

Index of Symbols
G [X ] Set-valued map associated with a control system ne Unit normal to the edge e of a polygon Q pointing
x : Rd × U → R d toward the interior of Q
co(S ) Convex closure of a set S ⊆ Rd f Set of points where the locally Lipschitz function
co(S ) Convex hull of a set S ⊆ Rd f : Rd → R fails to be differentiable
diam(π) Diameter of the partition π π Partition of a closed interval
f o(x; v ) Generalized directional derivative of the function B(S) Set whose elements are all of the possible sub-
f : Rd → R at x ∈ Rd in the direction of v ∈ Rd sets of S ⊆ Rd
f (x; v ) Right directional derivative of the function ∂P f Proximal subdifferential of the lower semicontinu-
f : Rd → R at x ∈ Rd in the direction of v ∈ Rd ous function f : Rd → R
SX Set of points where the vector field x : Rd → Rd is L̃F f Set-valued Lie derivative of the locally Lipschitz
discontinuous function f : Rd → R with respect to the set-val-
dist ( p, S ) Euclidean distance from the point p ∈ Rd to the ued map F : Rd → B( Rd )
set S ⊆ Rd L̃ X f Set-valued Lie derivative of the locally Lipschitz
F [X ] Filippov set-valued map associated with a vector function f : Rd → R with respect to the Filippov
field X : Rd → Rd set-valued map F [X ] : Rd → B( Rd )
∂f Generalized gradient of the locally Lipschitz func- LF f Lower set-valued Lie derivative of the lower
tion f : Rd → R semicontinuous function f : Rd → R with
∇f Gradient of the differentiable function f : Rd → respect to the set-valued map F : Rd → B( Rd )
R LF f Upper set-valued Lie derivative of the lower
Ln(S ) Least-norm elements in the closure of the set semicontinuous function f : Rd → R with
S ⊆ Rd respect to the set-valued map F : Rd → B( Rd )
(x ) Set of limit points of a curve t → x(t) F Set-valued map
N Nearest-neighbor map sm Q Minimum distance function from a point in a con-
vex polygon Q ⊂ Rd to the boundary of Q

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 43

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 Example 6: Discontinuous Vector Field which is discontinuous at zero [see Figure 5(b)]. If x(0) > 0,
with Existence of Classical Solutions then the maximal solution x : [0, x(0)) → R is
Consider the vector field X : R → R, x(t) = x(0) − t, whereas, if x(0) < 0, then the maximal solu-
tion x : [0, −x(0)) → R is x(t) = x(0) + t. Finally, if
⎧
⎨ −1, x > 0, x(0) = 0, then the maximal solution x : [0, ∞) → R is
X(x) = −sign(x) = 0, x = 0, (12) x(t) = 0. Hence the associated dynamical system
⎩
1, x < 0, ẋ(t) = X(x(t)) has a classical solution starting from every

Caratheodory Conditions for Time-Varying Vector Fields

ε ∈ (0, ∞) and an integrable function m : [t, t + δ] → (0, ∞) such
C
onsider the differential equation
that ||X(s, y )||2 ≤ m (s ) for almost all s ∈ [t, t + δ] and all
ẋ(t ) = X(t, x(t )), (S1) y ∈ B (x, ε). Then, for all (t0 , x0 ) ∈ [0, ∞)× Rd , there exists a
Caratheodory solution of (S1) with initial condition x(t0 ) = x0 .
where X : [0, ∞)× Rd → Rd is a time-varying vector field. The fol- The specialization of Proposition S1 to a time-invariant vector
lowing result is taken from [18]: a weaker version of the field requires that the vector field be continuous, which in turn
Caratheodory conditions is given, for instance, in [S27]. guarantees the existence of a classical solution.
PROPOSITION S1
Let X : [0, ∞)× Rd → Rd . Assume that i) for almost all t ∈ [0, ∞), REFERENCES
the map x → X(t, x ) is continuous; ii) for each x ∈ Rd , the map [S27] D.C. Biles and P.A. Binding, “On Caratheodory’s conditions for
t → X(t, x ) is measurable; and iii) X is locally essentially the initial value problem,” Proc. Amer. Math. Soc., vol. 125, no. 5, pp.
bounded, that is, for all (t, x) ∈ (0, ∞)× Rd , there exist 1371–1376, 1997.

Caratheodory Solutions of Differential Inclusions

A
differential inclusion [19], [20] is a generalization of a dif- [24]. The uniqueness of Caratheodory solutions is guaranteed by
ferential equation. At each state, a differential inclusion the following result.
specifies a set of possible evolutions, rather than a single
one. This object is defined by means of a set-valued map; PROPOSITION S3
see “Set-Valued Maps.” The differential inclusion associated In addition to the hypotheses of Proposition S2, assume that, for all
with a time-varying set-valued map F : [0, ∞)× Rd → B(Rd) x ∈ Rd , there exist ε ∈ (0, ∞) and an integrable function L x :
is an equation of the form R → (0, ∞) such that

ẋ(t) ∈ F(t, x(t)). (S2) (v − w)T (y − y ) ≤ Lx (t) y − y 22 , (S3)

The point xe ∈ Rd is an equilibrium of the differential inclusion if for almost every y, y ∈ B(x, ε) , every t ∈ [0, ∞) , every
0 ∈ F(t, xe) for all t ∈ [0, ∞). We now define the notion of solution v ∈ F(t, y), and every w ∈ F(t, y ). Then, for all (t0 , x0 ) ∈ [0, ∞)×
of a differential inclusion in the sense of Caratheodory. Rd , there exists a unique Caratheodory solution of (S2) with initial
A Caratheodory solution of (S2) defined on [t0 , t1 ] ⊂ [0, ∞) is condition x(t0 ) = x0 .
an absolutely continuous map x : [t0 , t1 ] → Rd such that The following example illustrates propositions S2 and S3.
ẋ(t) ∈ F(t, x(t)) for almost every t ∈ [t0 , t1 ]. The existence of at Following [43], consider the set-valued map F : R → B(R)
least one solution starting from each initial condition is guaran- defined by
teed by the following result (see, for instance, [14] and [19]). 
0, x = 0,
F(x) =
[−1, 1], x = 0.
PROPOSITION S2
Let F : [0, ∞)× Rd →B (Rd) be locally bounded and take non- Note that F is upper semicontinuous, but not lower semicontinu-
empty, compact, and convex values. Assume that, for each t ∈ R, ous, and thus it is not continuous. This set-valued map satisfies all
the set-valued map x → F(t, x) is upper semicontinuous, and, for of the hypotheses in Proposition S2, and therefore Caratheodory
each x ∈ Rd , the set-valued map t → F(t, x) is measurable. Then, solutions exist starting from all initial conditions. In addition, F sat-
for all (t0 , x 0 ) ∈ [0, ∞)× Rd , there exists a Caratheodory solution of isfies (S3) as long as y and y are nonzero. Therefore, Proposi-
(S1) with initial condition x(t0 ) = x 0 . tion S3 guarantees the uniqueness of Caratheodory solutions. In
This result is sufficient for our purposes. Additional existence fact, for every initial condition, the Caratheodory solution of
results based on alternative assumptions are given in [14] and ẋ(t ) ∈ F(x(t )) is just the equilibrium solution.

44 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 initial condition. Although the vector fields (11) and (12) This property is one sided because it imposes a require-
are identical except for the value at zero, the existence of ment on X only when the angle between the two vectors
classical solutions is surprisingly different.  on the left-hand side of (13) is between 0–180◦ . The follow-
ing result, given in [28], uses this property to provide a
Uniqueness of Classical Solutions sufficient condition for uniqueness.
Uniqueness of classical solutions of the differential equa-
tion (10) means that every pair of solutions with the Proposition 2
same initial condition coincide on the intersection of Let x : Rd → Rd be continuous. Assume that, for all
their intervals of existence. In other words, if x ∈ Rd , there exists ε > 0 such that X is one-sided Lip-
x1 : [0, t1 ] → Rd and x2 : [0, t2 ] → Rd are classical solu- schitz on B(x, ε). Then, for all x0 ∈ Rd , there exists a
tions of (10) with x1 (0) = x2 (0), then x1 (t) = x2 (t) for all unique classical solution of (10) with initial condition
t ∈ [0, t1 ] ∩ [0, t2 ] = [0, min{t1 , t2 }] . Equivalently, we say x(0) = x0 .
that there exists a unique maximal solution starting from Every vector field that is locally Lipschitz at x (see
each initial condition. “Locally Lipschitz Functions”) satisfies the one-sided Lip-
Uniqueness of classical solutions is guaranteed under a schitz condition on a neighborhood of x, but the converse is
wide variety of conditions. The book [28], for instance, is not true. For example, the vector field X : R → R defined
devoted to collecting various uniqueness criteria. Here, we by X(0) = 0 and X(x) = x log(|x|) for x = 0 is one-sided Lip-
focus on a uniqueness criterion based on one-sided Lip- schitz on a neighborhood of zero, but is not locally Lipschitz
schitzness. The vector field X : Rd → Rd is one-sided Lip- at zero. Furthermore, a one-sided Lipschitz vector field can
schitz on U ⊂ Rd if there exists L > 0 such that, for all be discontinuous. For example, the discontinuous vector
y, y ∈ U, field (12) is one-sided Lipschitz on a neighborhood of zero.
The following result is an immediate consequence of
[X(y) − X(y )]T (y − y ) ≤ L y − y 22 . (13) Proposition 2.

1 1

.5 .5

−1 −.5 .5 1 −1 −.5 .5 1
−.5 −.5

−1 −1
(a) (b)

1
.3
.8
.2
.6 .1
.4
−1 −.5 .5 1
.2 −.1
−.2
−1 −.5 .5 1
(c) (d)

FIGURE 5 Discontinuous and not-one-sided Lipschitz vector fields. The vector fields in (a) and (b), which differ only in their values at zero,
are discontinuous, and thus do not satisfy the hypotheses of Proposition 1. Therefore, the existence of solutions is not guaranteed. In fact,
the vector field in (a) has no solution starting from zero, whereas the vector field in (b) has a solution starting from all initial conditions. The
vector fields in (c) and (d) are neither locally Lipschitz nor one-sided Lipschitz, and thus do not satisfy the hypotheses of Proposition 2.
Therefore, uniqueness of solutions is not guaranteed. The vector field in (c) has two solutions starting from zero. However, the vector field
in (d) has a unique solution starting from all initial conditions.

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 45

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 
Corollary 1 0, 0 ≤ t ≤ a,
xa(t) =
Let X : Rd → Rd be locally Lipschitz. Then, for all x0 ∈ Rd , (t − a)2 /4, t ≥ a,
there exists a unique classical solution of (10) with initial
condition x(0) = x0 . and x : [0, ∞) → R , where x(t) = 0. 
Although local Lipschitzness is typically invoked as However, the following example shows that a differen-
in Corollary 1 to guarantee uniqueness, Proposition 2 tial equation can possess a unique classical solution even
shows that uniqueness is guaranteed under weaker when the hypotheses of Proposition 2 are not satisfied.
hypotheses. The following example shows that, if the
hypotheses of Proposition 2 are not satisfied, then solu- Example 8: Continuous, Not One-Sided Lipschitz
tions might not be unique. Vector Field with Unique Classical Solutions
Consider the vector field X : R → R defined by
Example 7: Continuous, Not One-Sided Lipschitz
Vector Field with Nonunique Classical Solutions ⎧
Consider the vector field X : R → R defined by ⎨ −x log x, x > 0,
X(x) = 0, x = 0, (15)
⎩
x log(−x), x < 0.
X(x) = |x| . (14)
This vector field is continuous everywhere and locally Lip-
This vector field is continuous everywhere, locally schitz on R \{0} [see Figure 5(d)]. However, X is not locally
Lipschitz on R \{0} [see Figure 5(c)] but is neither Lipschitz at zero nor one-sided Lipschitz on any neighbor-
locally Lipschitz at zero nor one-sided Lipschitz on hood of zero. Nevertheless, the associated dynamical sys-
any neighborhood of zero. The associated dynamical tem ẋ(t) = X(x(t)) has a unique solution starting from each
system ẋ(t) = X(x(t)) has infinitely many maximal solu- initial condition. If x(0) > 0, then the maximal solution
t i o n s s t a r t i n g f r o m z e r o , n a m e l y , f o r a l l a > 0, x : [0, ∞) → R is x(t) = exp(log x(0) exp(−t)), whereas, if
xa : [0, ∞) → R , where x(0) < 0, then the maximal solution x : [0, ∞) → R is

Locally Lipschitz Functions 1

function f : Rd → Rm is locally Lipschitz at x ∈ Rd if there exist
A Lx , ε ∈ (0, ∞) such that .8

.6
f (y) − f (y )2 ≤ Lx y − y 2 ,
.4

for all y, y ∈ B(x, ε). A function that is locally Lipschitz at x is .2

continuous at x , but the converse is not true. For example,
√
f : R → R, f (x) = |x|, is continuous at zero, but not locally Lip- −1 −.5 .5 1
schitz at zero; see Figure S1(a). A function is locally Lipschitz on (a)
S ⊆ Rd if it is locally Lipschitz at x for all x ∈ S . If f is locally Lip-
schitz on Rd , we simply say f is locally Lipschitz. Convex func-
1
tions are locally Lipschitz [S28], and hence concave functions
are also locally Lipschitz. Note that a function that is continuous- .8
ly differentiable at x is locally Lipschitz at x , but the converse is .6
not true. For example, f : R → R, f (x) = |x|, is locally Lipschitz
at zero, but not differentiable at zero; see Figure S1(b). A func- .4
tion f : R × Rd → Rm that depends explicitly on time is locally .2
Lipschitz at x ∈ Rd if there exists ε ∈ (0, ∞) and Lx : R → (0, ∞)
such that f (t, y) − f (t, y )2 ≤ Lx (t)y − y 2 for all t ∈ R
−1 −.5 .5 1
and y, y ∈ B (x, ε). (b)

FIGURE S1 Illustration of the difference among continuous, locally

REFERENCE Lipschitz, and differentiable functions. (a) The graph of
√
[S28] Dept. Mathematics, Wayne State Univ., “Every convex func- f : R → R, f (x) = |x|, which is continuous at zero but not
tion is locally Lipschitz,” Amer. Math. Monthly, vol. 79, no. 10, pp. locally Lipschitz at zero. (b) The graph of f : R → R, f (x) = |x|,
1121–1124, 1972. which is locally Lipschitz at zero, but not differentiable at zero.

46 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 x(t) = − exp(log(−x(0)) exp(t)) . Finally, if x(0) = 0, then zero. However, this system has two Caratheodory solutions
the maximal solution x : [0, ∞) → R is x(t) = 0.  starting from zero, namely, x1 : [0, ∞) → R , where x1 (t) = t,
Note that Proposition 2 assumes that the vector field is con- and x2 : [0, ∞) → R , where x2 (t) = −t. Note that both x1
tinuous. However, the discontinuous vector field (12) in and x2 violate the differential equation only at t = 0, that is,
Example 6 is one-sided Lipschitz in a neighborhood of zero ẋ1 (0) = X(x1 (0)) and ẋ2 (0) = X(x2 (0)). 
and, indeed, has a unique classical solution starting from each
initial condition. This observation suggests that discontinuous Example 3 Revisited: Existence of Caratheodory
systems are not necessarily more complicated or less “well Solutions for the Nonsmooth Harmonic Oscillator
behaved” than continuous systems. The continuous system The nonsmooth harmonic oscillator in Example 3 does
(14) in Example 7 does not have a unique classical solution not possess a classical solution starting from any initial
starting from each initial condition, whereas the discontinuous condition on the x2 -axis. However, the closed level sets
system (12) in Example 6 does. A natural question to ask is depicted in Figure 3(b), when traversed clockwise, are
under what conditions does a discontinuous vector field have Caratheodory solutions. 
a unique solution starting from each initial condition. Of Unfortunately, the good news is quickly over since it
course, the answer to this question relies on the notion of solu- is easy to find examples of discontinuous dynamical sys-
tion itself. We explore these questions in the next section. tems that do not admit Caratheodory solutions. For
example, the physical motions observed in Example 2,
NOTIONS OF SOLUTION FOR where the brick slides for a while and then remains
DISCONTINUOUS DYNAMICAL SYSTEMS stopped, are not Caratheodory solutions. Furthermore,
The above discussion shows that the classical notion of the discontinuous vector field (11) does not admit a
solution is too restrictive when considering a discontin- Caratheodory solution starting from zero. Finally, the
uous vector field. We thus explore alternative notions move-away-from-nearest-neighbor interaction law in
of solution to reconcile this mismatch. To address the Example (4) is yet another example where Caratheodory
discontinuities of the differential equation (10), we first solutions do not exist, as we show next.
relax the requirement that solutions follow the direction
specified by the vector field at all times. The precise Example 4 Revisited: Nonexistence
mathematical notion corresponding to this idea is that of Caratheodory Solutions for the
of Caratheodory solutions, which we introduce next. Move-Away-from-Nearest-Neighbor Interaction Law
Consider one agent moving in the square [−1, 1]2 ⊂ R2
Caratheodory Solutions under the move-away-from-nearest-neighbor interaction
A Caratheodory solution of (10) defined on [0, t1 ] ⊂ R is an law described in Example 4. Since no other agent is pre-
absolutely continuous map x : [0, t1 ] → Rd that satisfies sent in the square, the agent moves away from the near-
(10) for almost all t ∈ [0, t1 ] (in the sense of Lebesgue mea- est polygonal boundary, according to the vector field
sure). In other words, a Caratheodory solution follows the
⎧
direction specified by the vector field except for a set of ⎪ (−1, 0), −x1 < x2 ≤ x1 ,
⎪
⎨ (0, 1),
time instants that has measure zero. Equivalently, x2 < x1 ≤ −x2 ,
X(x1 , x2 ) = (17)
Caratheodory solutions are absolutely continuous func- ⎪
⎪ (1, 0), x1 ≤ x2 < −x1 ,
⎩
tions that solve the integral version of (10), that is, (0, −1), −x2 ≤ x1 < x2 .
 t
Since the move-away-from-nearest-neighbor interaction
x(t) = x(0) + X(x(s))ds. (16)
0 law takes multiple values on the diagonals
{(a, ±a) ∈ [−1, 1]2 : a ∈ [−1, 1]} of the square, we choose
Of course, every classical solution is also a Caratheodory one of these values in the definition (17) of X. The phase
solution. portrait in Figure 6(a) shows that the vector field X is dis-
continuous on the diagonals.
Example 9: System with Caratheodory The dynamical system ẋ(t) = X(x(t)) has no
Solutions and No Classical Solutions Caratheodory solution if and only if the initial condition
The vector field x : R → R defined by belongs to the diagonals. This fact can be justified as follows.
Consider the four open regions of the square separated by
⎧
⎨ 1, x > 0, the diagonals; see Figure 6(a). On the one hand, if the initial
X(x) = 12 , x = 0, condition belongs to one of these regions, then the dynami-
⎩
−1, x < 0, cal system has a classical solution; depending on which
region the initial condition belongs to, the agent moves
is discontinuous at zero. The associated dynamical system either vertically or horizontally toward the diagonals. On the
ẋ(t) = X(x(t)) does not have a classical solution starting from other hand, on the diagonals of the square, X pushes

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 47

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 trajectories outward, whereas, outside the diagonals, X Sufficient Conditions for the Existence
pushes trajectories inward. Therefore, if the initial condition of Caratheodory Solutions
belongs to the diagonals, then the only candidate trajectory Conditions under which Caratheodory solutions exist are
for a Caratheodory solution is a trajectory that moves along discussed in “Caratheodory Conditions for Time-Varying
the diagonals. From the definition (16) of Caratheodory solu- Vector Fields.” For time-invariant vector fields, the
tion, it is clear that a trajectory that moves along the diago- Caratheodory conditions given by Proposition S1 specialize
nals of the square is not a Caratheodory solution. We show to continuity of the vector field. This requirement provides
later that this trajectory is instead a Filippov solution.  no improvement over Proposition 1, which guarantees the
existence of classical solutions under continuity.
Therefore, it is of interest to determine conditions for
the existence of Caratheodory solutions specifically tai-
lored to time-invariant vector fields. For example, direc-
1 tionally continuous vector fields are considered in [26]. A
vector field X : Rd → Rd is directionally continuous if there
exists δ ∈ (0, ∞) such that, for every x ∈ Rd with X(x) = 0
and every sequence {xn }n ∈N ⊂ Rd with xn → x and
.5

xn − x X(x)
− < δ, n ∈ N, (18)
xn − x2 X(x)2 2
x2

0
it follows that X(xn ) → X(x). If the vector field X is direction-
ally continuous, then, for all x0 ∈ Rd , there exists a
−.5 Caratheodory solution of (10) with initial condition x(0) = x0 .
Patchy vector fields [30] are another class of time-invariant,
discontinuous vector fields that have Caratheodory solutions.
Additional conditions for the existence and uniqueness of
−1
−1 −.5 0 .5 1
Caratheodory conditions can be found in [31]–[33]. Beyond dif-
x1 ferential equations, Caratheodory solutions can also be defined
(a) for differential inclusions, as explained in “Set-Valued Maps”
and “Caratheodory Solutions of Differential Inclusions.”

1 Filippov Solutions
The above discussion shows that the relaxation of the value
of the vector field on a set of times of measure zero in the def-
.5 inition of Caratheodory solution is not always sufficient to
guarantee that such solutions exist. Due to the discontinuity
of the vector field, its value can exhibit significant variations
arbitrarily close to a given point, and this mismatch might
0
x2

make it impossible to construct a Caratheodory solution.

What if, instead of focusing on the value of the vector
field at individual points, we consider how the vector field
−.5 looks like around each point? The idea of looking at a
neighborhood of each point is at the core of the notion of
Filippov solution [18]. Closely related notions are those of
Krasovskii [34] and Sentis [35] solutions.
−1
−1 −.5 0 .5 1 The mathematical framework for formalizing this
x1 neighborhood idea uses set-valued maps. The idea is to
(b) associate a set-valued map to X : Rd → Rd by looking at
the neighboring values of X around each point. Specifical-
FIGURE 6 Move-away-from-nearest-neighbor interaction law for one ly, for x ∈ Rd , the vector field X is evaluated at the points
agent moving in the square [−1, 1]2 ⊂ R2 . (a) shows the phase por- belonging to B(x, δ), which is the open ball centered at x
trait, while (b) shows several Filippov solutions. On the diagonals, with radius δ > 0. We examine the effect of δ approaching
the vector field pushes trajectories out, whereas outside the diago-
nals, the vector field pushes trajectories in. This fact makes it impos-
zero by performing this evaluation for smaller and smaller
sible to construct a Caratheodory solution starting from any initial δ. For additional flexibility, we exclude an arbitrary set of
condition on the diagonals of the square. measure zero in B(x, δ) when evaluating X, so that the

48 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 outcome is the same for two vector fields that differ on a
set of measure zero. Set-Valued Maps
Mathematically, the above procedure can be summa-
A
set-valued map, as its name suggests, is a map that
rized as follows. Let B(Rd) denote the collection of subsets assigns sets to points. We consider time-varying, set-valued
of Rd . For X : Rd → Rd , define the Filippov set-valued map maps of the form F : [0, ∞)× Rd →B(Rd ). Recall that B(Rd)
F[X] : Rd→ B(Rd) by denotes the collection of all subsets of Rd . The map F assigns
to each point (t, x) ∈ [0, ∞)× Rd the set F(t, x) ⊆ Rd . Note
F [X](x)  co{X(B(x, δ)\S)}, x ∈ Rd. (19) that a standard map f : [0, ∞)× Rd → Rd can be interpreted as
δ>0 μ(S) =0 a singleton-valued map. A complete analysis for set-valued
In (19), co denotes convex closure, and μ denotes maps can be developed, as in the case of standard maps [43].
Lebesgue measure. Because of the way the Filippov set- Here, we are mainly interested in concepts related to bounded-
valued map is defined, the value of F[X] at a point x is ness and continuity, which we define next.
independent of the value of the vector field X at x. The set-valued map F : [0, ∞)× R d →B(Rd ) is locally
bounded (respectively, locally essentially bounded ) at
Examples 5 and 6 Revisited: Filippov (t, x) ∈ [0, ∞)× Rd if there exist ε, δ ∈ (0, ∞) and an integrable
Set-Valued Map of the Sign Function function m : [t, t + δ] → (0, ∞) such that ||z ||2 ≤ m(s) for all
Let us compute the Filippov set-valued map for the vec- z ∈ F(s, y), all s ∈ [t, t + δ], and all y ∈ B(x, ε) (respectively,
tor fields (11) and (12). Since both vector fields differ only almost all y ∈ B(x, ε) in the sense of Lebesgue measure).
at zero, which is a set of measure zero, their associated The time-invariant set-valued map F : Rd →B(Rd ) is upper
Filippov set-valued maps are identical and equal to semicontinuous (respectively, lower semicontinuous) at x ∈ Rd
F[X] : R→ B(R), where if, for all ε ∈ (0, ∞) , there exists δ ∈ (0, ∞) such that
F(y) ⊆ F(x) + B(0, ε) (respectively, F(x) ⊆ F(y) + B(0, ε) )
⎧
⎨ −1, x > 0, for all y ∈ B(x, δ). The set-valued map F : Rd →B(Rd ) is con-
F[X](x) = [−1, 1], x = 0, (20) tinuous at x ∈ Rd if it is both upper and lower semicontinuous at
⎩
1, x < 0. x ∈ Rd . Finally, the set-valued map F : Rd →B(Rd ) is locally
Lipschitz at x ∈ Rd if there exist Lx , ε ∈ (0, ∞) such that
Note that this Filippov set-valued map is multiple valued
only at the point of discontinuity of the vector field. This F(y ) ⊆ F(y) + Lx ||y − y ||2 B(0, 1),
observation is valid for all vector fields. 
We are now ready to handle the discontinuities of the for all y, y ∈ B(x, ε). A locally Lipschitz set-valued map at x is
vector field X by using the Filippov set-valued map of X. upper semicontinuous at x , but the converse is not true.
We replace the differential equation ẋ(t) = X(x(t)) by the The notion of upper semicontinuity of a map f : Rd → R,
differential inclusion which is defined in the section “The Proximal Subdifferential of a
Lower Semicontinuous Function,” is weaker than the notion of
ẋ(t) ∈ F [X](x(t)). (21) upper semicontinuity of f when viewed as a (singleton-valued)
set-valued map from Rd to B(R). Indeed, the latter is equivalent
A Filippov solution of (10) on [0, t1 ] ⊂ R is an absolutely con- to the condition that f : Rd → R is continuous.
tinuous map x : [0, t1 ] → Rd that satisfies (21) for almost all
t ∈ [0, t1 ]. Equivalently, a Filippov solution of (10) is a
Caratheodory solution of the differential inclusion (21); see
“Caratheodory Solutions of Differential Inclusions.” map satisfies all of the hypotheses of Proposition S2
Because of the way the Filippov set-valued map is (see “Caratheodory Solutions of Differential Inclu-
defined, a vector field that differs from X on a set of mea- sions”), which in turn guarantees the existence of Fil-
sure zero has the same Filippov set-valued map, and hence ippov solutions.
the same set of solutions. The next result establishes mild
conditions under which Filippov solutions exist [18], [19]. Examples 5 and 6 Revisited:
Existence of Filippov Solutions for the Sign Function
Proposition 3 The application of Proposition 3 to the bounded vector
Let X : Rd → Rd be measurable and locally essentially fields in examples 5 and 6 guarantees that a Filippov
bounded, that is, bounded on a bounded neighborhood of solution of (10) exists for both examples starting from
every point, excluding sets of measure zero. Then, for all each initial condition. Furthermore, since the vector fields
x0 ∈ Rd , there exists a Filippov solution of (10) with initial (11) and (12) have the same Filippov set-valued map (20),
condition x(0) = x0 . examples 5 and 6 have the same maximal Filippov
In Proposition 3, the hypotheses on the vector solutions. If x(0) > 0, then the maximal solution
field imply that the associated Filippov set-valued x : [0, ∞) → R is

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 49

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
  exist but Filippov solutions are not Caratheodory solu-
x(0) − t, t ≤ x(0),
x(t) = tions is given in [30]. The following is an example of the
0, t ≥ x(0),
opposite case.
whereas, if x(0) < 0, then the maximal solution x : [0, ∞) → R is
Example 10: Vector Field with a Caratheodory Solution

x(0) + t, t ≤ −x(0), That Is Not a Filippov Solution
x(t) =
0, t ≥ −x(0). Consider the vector field X : R → R given by

1, x = 0,
Finally, if x(0) = 0, then the maximal solution x : X(x) = (24)
0, x = 0.
[0, ∞) → R is x(t) = 0. 
Similar computations can be made for the move-away- The dynamical system (10) has two Caratheodory solu-
from-nearest-neighbor interaction law in Example 4 to tions starting from zero, namely, x1 : [0, ∞) → R , where
show that Filippov solutions exist starting from each initial x1 (t) = 0, and x2 : [0, ∞) → R , where x2 (t) = t. However,
condition, as we show next. the associated Filippov set-valued map F[X] : R→ B(R) is
F[X](x) = {1}, and hence t → x2 (t) is the unique Filippov
Example 4 Revisited: Filippov Solutions for the solution starting from zero. 
Move-Away-from-Nearest-Neighbor Interaction Law On a related note, Caratheodory solutions are always
Consider again the discontinuous vector field for one agent Krasovskii solutions (see “Additional Solution Notions for
moving in the square [−1, 1]2 ⊂ R2 under the move-away- Discontinuous Systems”) but the converse is not true [16].
from-nearest-neighbor interaction law described in Example 4.
The corresponding set-valued map F [X] : [−1, 1]2 → B(R2 ) Computing the Filippov Set-Valued Map
is given by (22), shown at the bottom of the page. Since X is Computing the Filippov set-valued map can be a daunting
bounded, it follows from Proposition 3 that a Filippov solution task. A calculus is developed in [36] to simplify this calcu-
exists starting from each initial condition; see Figure 6(b). In lation. We summarize some useful facts below. Note that
particular, each solution starting from a point on a diagonal is a the Filippov set-valued map can also be constructed for
straight line flowing along the diagonal itself and reaching maps of the form X : Rd → Rm , where d and m are not nec-
(0, 0). For example, the maximal solution x : [0, ∞) → R2 essarily equal.
starting from (a, a) ∈ R2 is given by Consistency: If X : Rd → Rm is continuous at x ∈ Rd , then

 
a − 12 sign(a)t, a − 12 sign(a)t , t ≤ |a|, F[X](x) = {X(x)}. (25)
t → x(t) =
(0, 0), t ≥ |a|.
Sum Rule: If X1 , X2 : Rd → Rm are locally bounded at
(23) x ∈ Rd , then

Note that the solution slides along the diagonals, following a F[X1 + X2 ](x) ⊆ F[X1 ](x) + F[X2 ](x). (26)
convex combination of the limiting values of X around the
diagonals, rather than the direction specified by X itself. We Moreover, if either X1 or X2 is continuous at x, then equal-
study this type of behavior in more detail in the section “Piece- ity holds.
wise Continuous Vector Fields and Sliding Motions.”  Product Rule: If X1 : Rd → Rm and X2 : Rd → Rn are
locally bounded at x ∈ Rd , then
Relationship Between Caratheodory and Filippov Solutions
In general, Caratheodory and Filippov solutions are not
related. A vector field for which both notions of solution F[(X1 , X2 )T ](x) ⊆ F[X1 ](x) × F[X2 ](x). (27)

⎧
⎪ {(y1 , y2 ) ∈ R2 : |y1 + y2 | ≤ 1, |y1 − y2 | ≤ 1}, (x1 , x2 ) = (0, 0),
⎪
⎪
⎪
⎪ {(−1, 0)}, −x1 < x2 < x1 ,
⎪
⎪
⎪ {(y1 , y2 ) ∈ R2
⎪ : y1 + y2 = −1, y1 ∈ [−1, 0]}, 0 < x2 = x1 ,
⎪
⎪
⎪
⎪
⎨ {(0, 1)}, x2 < x1 < −x2 ,
F [X](x1 , x2 ) = {(y1 , y2 ) ∈ R2 : y1 − y2 = −1, y1 ∈ [−1, 0], 0 < −x1 = x2 , (22)
⎪
⎪
⎪
⎪ {(1, 0)}, x1 < x2 < −x1 ,
⎪
⎪
⎪
⎪ 1 , y2 ) ∈ R2
{(y : y1 + y2 = 1, y1 ∈ [0, 1]}, x2 = x1 < 0,
⎪
⎪
⎪
⎪ {(0, −1)}, −x2 < x1 < x2 ,
⎪
⎩
{(y1 , y2 ) ∈ R2 : y1 − y2 = 1, y1 ∈ [0, 1]}, 0 < x1 = −x2 .

50 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 Moreover, if either X1 or X2 is continuous at x, then equal- of the vector field to D2 can be continuously extended to
ity holds. D2 by setting X|D (0) = g(sin θ − ν cos θ). Therefore, the
2
Chain Rule: If Y : Rd → Rn is continuously differen- associated Filippov set-valued map F[X] : R→ B(R) is
tiable at x ∈ Rd with Jacobian rank n, and X : Rn → Rm is given by
locally bounded at Y(x) ∈ Rn , then
⎧
⎨ {g(sin θ − ν cos θ)}, v > 0,
F[X ◦ Y](x) = F[X](Y(x)). (28) F [X](v) = {g(sin θ − d ν cos θ) : d ∈ [−1, 1]}, v = 0,
⎩
{g(sin θ + ν cos θ)}, v < 0,
Matrix Transformation Rule: If X : Rd → Rm is locally
bounded at x ∈ Rd and Z : Rd → Rd×m is continuous at which is singleton valued for all v ∈ SX = {0}, and a closed
x ∈ Rd , then segment at v = 0. If the friction coefficient ν is large enough,
that is, satisfies ν > tan θ , then F [x](v) < 0 if v > 0 and
F[Z X](x) = Z(x) F[X](x). (29) F [x](v) > 0 if v < 0. Therefore, the Filippov solutions that
start with an initial positive velocity v eventually reach v = 0
Piecewise Continuous Vector Fields and Sliding Motions and remain at zero. This fact precisely corresponds to the
In this section we consider vector fields that are continu- observed physical motions for Example 2, where the brick
ous everywhere except on a surface of the state space. slides for a while and then remains stopped. This example
Consider, for instance, two continuous dynamical sys- shows that Filippov solutions have physical significance.
tems, one per side of the surface, glued together to give The vector field X : R2 → R2 for the nonsmooth har-
rise to a discontinuous dynamical system. Here, we ana- monic oscillator in Example 3 is continuous on each of the
lyze the properties of the Filippov solutions of this type of half planes {D1 , D2 }, with
systems.
The vector field X : Rd → Rd is piecewise continuous if D1 = {(x1 , x2 ) ∈ R2 : x1 < 0},
there exists a finite collection of disjoint, open, and con- D2 = {(x1 , x2 ) ∈ R2 : x1 > 0},
nected sets D1 , . . . , Dm ⊂ Rd whose closures cover Rd , that
is, Rd = ∪mk =1 Dk , such that, for all k = 1, . . . , m, the vector and discontinuous on SX = {(0, x2 ) : x2 ∈ R}. Therefore, X
field X is continuous on Dk . We further assume that the is piecewise continuous. Its Filippov set-valued map
restriction of X to Dk admits a continuous extension to the F[X] : R2 → B(R2 )} is given by
closure Dk , which we denote by X|D . Every point of dis- 
k {(x2 , −sign(x1 ))}, x1 = 0,
continuity of X must therefore belong to the union of the F[X](x1 , x2 ) =
{x2 } × [−1, 1], x1 = 0.
boundaries of the sets D1 , . . . , Dm . Let us denote by
SX ⊆ bndry(D1 ) ∪ · · · ∪ bndry(Dm ) the set of points where Therefore, the closed level sets depicted in Figure 3(b),
X is discontinuous. Note that SX has measure zero. when traversed clockwise, are Filippov solutions.
The Filippov set-valued map of a piecewise continuous The discontinuous vector field X : [−1, 1]2 → R2 for
vector field X is given by the expression one agent moving in the square [−1, 1]2 ⊂ R2 under
  the move-away-from-nearest-neighbor interaction law
F[X](x) = co lim X(xi) : xi → x , xi ∈ / SX . (30) described in Example 4 is piecewise continuous, with
i →∞

This set-valued map can be computed as follows. At points D1 = {(x1 , x2 ) ∈ [−1, 1]2 : −x1 < x2 < x1 },
of continuity of X, that is, for x ∈
/ SX , the consistency prop-
D2 = {(x1 , x2 ) ∈ [−1, 1]2 : x2 < x1 < −x2 },
erty (25) implies that F[X](x) = {X(x)}. At points of discon-
tinuity of X, that is, for x ∈ SX , F[X](x) is a convex D3 = {(x1 , x2 ) ∈ [−1, 1]2 : x1 < x2 < −x1 },
polyhedron in Rd of the form D4 = {(x1 , x2 ) ∈ [−1, 1]2 : −x2 < x1 < x2 }.
 
F [X](x) = co XD (x) : x ∈ bndry(Dk ) . (31) Its Filippov set-valued map, described in (22), maps points
k
outside the diagonals SX = {(a, ±a) ∈ [−1, 1]2 : a ∈ [−1, 1]}
As an illustration, let us revisit examples 2–4. to singletons, maps points in SX \{(0, 0)} to closed seg-
ments, and maps (0, 0) to a square polygon. Several Filip-
Examples 2–4 Revisited: Computation of the Filippov pov solutions starting from various initial conditions are
Set-Valued Map and Filippov Solutions plotted in Figure 6(b). 
The vector field of the sliding brick in Example 2 is Let us now discuss what happens on the points of dis-
piecewise continuous with D1 = {v ∈ R : v < 0} and continuity of the vector field X : Rd → Rd . Suppose that
D2 = {v ∈ R : v > 0}. Note that the restriction of the vec- x ∈ SX belongs to just two boundary sets, that is,
tor field to D1 can be continuously extended to D1 by set- x ∈ bndry (Di) ∩ bndry (D j) , for some distinct i, j ∈
ting X|D (0) = g(sin θ + ν cos θ). Likewise, the restriction {1, . . . , m}, but x ∈
/ bndry (Dk ), for all k ∈ {1, . . . , m}\{i, j}.
1

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 51

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 In this case, Uniqueness of Filippov Solutions
  A discontinuous dynamical system does not necessarily
F[X](x) = co X|D (x), X|D (x) . have a unique Filippov solution starting from each initial
i j
condition. The situation depicted in Figure 7(c) is a qualita-
We consider three possibilities. First, if all of the vectors tive example in which multiple Filippov solutions exist for
belonging to F[X](x) point into Di , then a Filippov solution the same initial condition. The following is another exam-
that reaches SX at x continues its motion in Di [see ple of lack of uniqueness.
Figure 7(a)]. Likewise, if all of the vectors belonging to
F[X](x) point into D j, then a Filippov solution that reaches Example 11: Vector Field with Nonunique Filippov Solutions
SX at x continues its motion in D j [see Figure 7(b)]. Finally, Consider the vector field X : R → R defined by
if a vector belonging to F[X](x) is tangent to SX , then either X(x) = sign (x). For all x0 ∈ R \{0}, the system (10) has a
all Filippov solutions that start at x leave SX immediately unique Filippov solution starting from x0 . However, the
[see Figure 7(c)], or there exist Filippov solutions that reach system (10) has three maximal solutions x1 , x2 , x3 :
the set SX at x, and remain in SX afterward [see Figure 7(d)]. [0, ∞) → R starting from x0 = 0 given by x1 (t) = −t,
The last kind of trajectories are called sliding motions, x2 (t) = 0, and x3 (t) = t. 
since they slide along the boundaries of the sets where the We now provide two complementary uniqueness
vector field is continuous. This type of behavior is illustrat- results for Filippov solutions. The first result [18] considers
ed for Example 4 in (23). Sliding motions can also occur the Filippov set-valued map associated with a discontinu-
along points belonging to the intersection of more than ous vector field and identifies conditions under which
two sets in D1 , . . . , Dm . The theory of sliding mode control Proposition S3 in “Caratheodory Solutions of Differential
builds on the existence of this type of trajectories to design Inclusions” can be applied to the resulting differential
stabilizing feedback controllers. These controllers induce inclusion. To state this result, we need to introduce the fol-
sliding surfaces with the desired properties so that the lowing definition. The vector field X : Rd → Rd is essential-
closed-loop system is stable [6], [7]. ly one-sided Lipschitz on U ⊂ Rd if there exists L > 0 such
The solutions of piecewise continuous vector fields that, for almost all y, y ∈ U,
appear frequently in state-dependent switching dynamical
systems [37], [38]. Consider, for instance, the case of two [X(y) − X(y )]T (y − y ) ≤ L y − y |22 . (32)
dynamical systems with the same unstable equilibrium. By
identifying an appropriate switching surface between the The first uniqueness result for Filippov solutions is
two systems, it is often possible to synthesize a discontinu- stated next.
ous dynamical system for which the equilibrium is stable.
Proposition 4
Let X : Rd → Rd be measurable and locally essentially
bounded. Assume that, for all x ∈ Rd , there exists ε > 0
Di Di such that X is essentially one-sided Lipschitz on B(x, ε).
Then, for all x0 ∈ Rd , there exists a unique Filippov solu-
tion of (10) with initial condition x(0) = x0 .
Note the parallelism of this result with Proposition 2 for
Dj Dj ordinary differential equations with a continuous vector
(a) (b) field X. Let us apply Proposition 4 to an example.

Di Example 12: Vector Field with Unique Filippov Solutions

Di
Let Q denote the set of rational numbers, and define the
vector field X : R → R by

Dj
1, x ∈ Q,
Dj X(x) =
−1, x ∈ Q,
(c) (d)
which is discontinuous everywhere on R. Since Q has mea-
FIGURE 7 Piecewise continuous vector fields. These dynamical sys- sure zero in R, the value of the vector field at rational points
tems are continuous on Di and Dj , and discontinuous on SX . In (a) plays no role in the computation of F[X]. Hence, the associ-
and (b), Filippov solutions, known as transversally crossing trajecto-
ated Filippov set-valued map F[X] : R → B(R) is
ries, cross SX . In (c), there are two Filippov solutions, known as
repulsive trajectories, starting from each point in SX . Finally, in (d), F[X](x) = {−1}. Since (32) holds for all y, y ∈/ Q, there exists
the Filippov solutions that reach SX , known as attractive trajectories, a unique solution starting from each initial condition, more
continue sliding along SX . precisely, x : [0, ∞) → R , where x(t) = x(0) − t. 

52 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 The hypotheses of Proposition 4 are somewhat restric- v = 0, the vector X|D (0) points into D2 ; see Figure 2(b).
1
tive. This assertion is justified by the observation that, Proposition 5 then ensures that there exists a unique Filip-
for d > 1, piecewise continuous vector fields on Rd are pov solution starting from each initial condition.
not essentially one-sided Lipschitz. We justify this asser- For Example 3, the vector X|D points into D2 at every
1
tion in “Uniqueness of Filippov Solutions of Piecewise point in SX ∩ {x2 > 0}, while the vector X|D points into D1
2
Continuous Vector Fields.” Figure
S2 shows an example of a piece-
wise continuous vector field with
unique solutions starting from Uniqueness of Filippov Solutions
each initial condition. However, of Piecewise Continuous Vector Fields
this uniqueness cannot be guaran-
H
ere we justify why, to guarantee the uniqueness of Filippov solutions for a piecewise
teed by means of Proposition 4. continuous vector field, we cannot resort to Proposition 4 and instead must use
The following result [18] identi- Proposition 5. To see this, consider a piecewise continuous vector field x : Rd → Rd ,
fies sufficient conditions for d ≥ 2, and let x ∈ SX be a point of discontinuity. Let us show that X is not essentially
uniqueness specifically tailored for one-sided Lipschitz on every neighborhood of x. For simplicity, assume x belongs to the
piecewise continuous vector fields. boundaries of just two sets, that is, x ∈ bndry(Di ) ∩ bndry(Dj ) (the argument proceeds
similarly for the general case). For ε > 0, we show that (32) is violated on a set of
Proposition 5 nonzero measure contained in B(x, ε). Notice that
Let X : Rd → Rd be a piecewise
continuous vector field, with Rd = (X (y) − X (y ))T (y − y ) = X (y) − X (y )2 y − y 2 cos α(y, y ),
D1 ∪ D2 . Let SX = bndry(D1 ) =
bndry(D2 ) be the set of points at where α(y, y ) =  (X(y) − X(y ), y − y ) is the angle between the vectors X(y) − X(y ) and
which X is discontinuous, and y − y . Therefore, (32) is equivalent to
assume that SX is a C2 -manifold.
Furthermore, assume that, for
X (y) − X (y )2 cos α(y, y ) ≤ Ly − y 2 . (S4)
i ∈ {1, 2}, X|D is continuously dif-
i
ferentiable on Di and X|D − X|D
1 2
is continuously differentiable on Consider the vectors X|Di (x) and
SX . If, for each x ∈ SX , either X|D j (x). Since X is discontinuous
X|D (x) points into D2 or X|D (x) at x, we have X|Di (x) = X|D j (x). R X D (x) – X D (x)
1 2 Sx i j
points into D1 , then there exists a Take y ∈ Di ∩ B(x, ε) and
β
unique Filippov solution of (10) y ∈ D j ∩ B(x, ε). Note that, as y
starting from each initial condition. and y tend to x , the vector
Note that the continuous differ- X(y) − X(y ) tends to x
entiability hypothesis on X already X|Di (x) − X|D j (x). Consider then
guarantees uniqueness of solutions the straight line that crosses SX ,
on each of the sets D1 and D2 . passes through x, and forms a R Di
β>0 Dj
Roughly speaking, the additional small angle with l
assumptions on X along SX in X|Di (x) − X|D j (x); see Figure S2.
Proposition 5 guarantee that Let R be the set enclosed by
FIGURE S2 Piecewise continuous vector field. The
uniqueness on Rd is not disrupted the line and the line in the vector field has a unique Filippov solution starting
by the discontinuities. Under the direction of the vector from all initial conditions. Note that solutions that
stated assumptions, when reach- X|Di (x) − X|D j (x) . If y ∈ Di ∩ R reach SX coming from D j cross SX , and then contin-
ing SX , Filippov solutions can and y ∈ D j ∩ R tend to x , we ue in Di . However, the vector field is not essentially
one-sided Lipschitz, and hence Proposition 4 cannot
cross it or slide along it. The situa- deduce that y − y 2 → 0 while,
be invoked to conclude uniqueness.
tion depicted in Figure 7(c) is thus at the same time,
ruled out.

Examples 2–4 Revisited: Unique X (y) − X (y )2 | cos α(y, y )|

Filippov Solutions for the Sliding ≥ X (y) − X (y )2 cos β −→ X|Di (x) − X|Dj (x)2 cos β > 0.
Brick, the Nonsmooth Harmonic
Oscillator, and the Move-Away-from- Therefore, there does not exist L ∈ (0, ∞) such that (S4) is satisfied for y ∈ R ∩ Di ∩ B(x, ε)
Nearest-Neighbor Interaction Law and y ∈ R ∩ D j ∩ B(x, ε). Thus, X is not essentially one-sided Lipschitz on any neighbor-
As an application of Proposition 5, hood of x, and the hypotheses of Proposition 4 do not hold.
let us reconsider Example 2. At

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 53

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 Discontinuities are intentionally designed to achieve regulation and stabilization.

at every point in SX ∩ {x2 < 0}; see Figure 3(a). Moreover, Solutions by Means of Differential Inclusions
there is only one solution (the equilibrium solution) start- A first alternative to defining a solution notion consists of
ing from (0, 0). Therefore, using Proposition 5, we con- associating a differential inclusion with the control equa-
clude that this system has a unique Filippov solution tion (33). In this approach, the set-valued map
starting from each initial condition. G[X] : Rd → B (Rd) is defined by
For Example 4, it is convenient to define D5 = D1 . Then,
at (x1 , x2 ) ∈ bndry (Di) ∩ bndry (Di+1 ) \{(0, 0)}, with G[X](x)  {X(x, u) : u ∈ U}. (34)
i ∈ {1, . . ., 4}, the vector X|D (x1 , x2 ) points into Di+1 , and
i
the vector X|D (x1 , x2 ) points into Di ; see Figure 6(a). In other words, the set G[X](x) captures all of the direc-
i+1
Moreover, there is only one solution (the equilibrium solu- tions in Rd that can be generated at x with controls belong-
tion) starting from (0, 0). Therefore, using Proposition 5, ing to U . Consider now the differential inclusion
we conclude that Example 4 has a unique Filippov solution
starting from each initial condition.  ẋ(t) ∈ G[X](x(t)). (35)
Proposition 5 can also be applied to piecewise contin-
uous vector fields with an arbitrary number of partition- A solution of (33) on [0, t1 ] ⊂ R is defined to be a
ing domains, provided that the set where the vector Caratheodory solution of the differential inclusion (35),
field is discontinuous is a disjoint union of surfaces that is, an absolutely continuous map x : [0, t1 ] → Rd such
resulting from pairwise intersections of the boundaries that ẋ(t) ∈ G[X](x(t)) for almost all t ∈ [0, t1 ].
of pairs of domains. Alternative versions of this result If we choose an open-loop input u : [0, ∞) → U in
can also be stated for time-varying piecewise continuous (33), then a Caratheodory solution of the resulting
vector fields, as well as for situations in which more dynamical system is also a Caratheodory solution of the
than two domains intersect at a point of discontinuity differential inclusion (35). Alternatively, it can be
[18, thrm 4, p. 115]. shown [18] that, if X is continuous and U is compact,
The literature contains additional results guarantee- then the converse is also true. The differential inclusion
ing uniqueness of Filippov solutions tailored to specific (35) has the advantage of not focusing attention on a
classes of dynamical systems. For instance, [39] studies particular control input but rather allows us to compre-
uniqueness for relay linear systems, [11] investigates hensively study and understand the properties of the
uniqueness for adaptive control systems, while [40] control system.
establishes uniqueness for a class of discontinuous dif-
ferential equations whose vector field depends on the Sample-and-Hold Solutions
solution of a scalar conservation law. A second alternative to defining a solution notion for
the control equation (33) uses the notion of sample-
Solutions of Control Systems and-hold solution [41]. As discussed in the section
with Discontinuous Inputs “Stabilization of Control Systems,” this notion plays a
Let X : Rd × U → Rd , where U ⊆ Rm is the set of allowable key role in the stabilization of asymptotically control-
control-function values, and consider the control equation lable systems.
on Rd given by A partition of the interval [t0 , t1 ] is an increasing
sequence π = {si}N i =0 with s0 = t0 and sN = t1 . The
ẋ(t) = X(x(t), u(t)). (33) partition need not be finite. The notion of a partition of
[t0 , ∞) is defined similarly. The diameter of π is
At first sight, a natural way to identify a notion of solu- diam(π)  sup{si − si−1 : i ∈ {1, . . . , N}} . Given a control
tion for (33) is to select a control input, either an open- input u : [0, ∞) × Rd → U , an initial condition x0 , and a
loop u : [0, ∞) → U , a closed-loop u : Rd → U , or a partition π of [0, t1 ], a π -solution of (33) defined on
combination u : [0, ∞) × Rd → U , and then consider the [0, t1 ] ⊂ R is the map x : [0, t1 ] → Rd , with x(0) = x0 ,
resulting differential equation. When the selected con- recursively defined by requiring the curve
trol input u is a discontinuous function of x ∈ Rd , then t ∈ [si−1 , si] → x(t) , for i ∈ {1, . . . , N − 1} , to be a
we can consider the solution notions of Caratheodory Caratheodory solution of the differential equation
or Filippov. At least two alternatives are considered in
the literature. We discuss them next. ẋ(t) = X(x(t), u(si−1 , x(si−1 ))). (36)

54 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 Many control systems cannot be stabilized by continuous state-dependent feedback.

π -solutions are also referred to as sample-and-hold solutions The above example illustrates the need to consider non-
because the control is held fixed throughout each interval smooth analysis. Similar examples can be found in [14, sect.
of the partition at the value according to the state at the 2.2.2]. It is also worth noting that the presentation in “Stabili-
beginning of the interval. Figure 8 shows an example of a ty Analysis by Means of the Generalized Gradient of a Non-
π -solution. The existence of π -solutions is guaranteed [24] smooth Lyapunov Function” boils down to classical stability
if, for all u ∈ U ⊆ Rm , the map x → X(x, u) is continuous. analysis when the candidate Lyapunov function is smooth.
In this section we discuss two tools from nonsmooth
NONSMOOTH ANALYSIS analysis, namely, the generalized gradient and the proximal
We now consider candidate nonsmooth Lyapunov functions subdifferential [9], [24]. As with the notions of solution of
for discontinuous differential equations. The level of general- discontinuous differential equations, multiple generalized
ity provided by nonsmooth analysis is not always necessary. derivative notions are available in the literature when a func-
The stability properties of some discontinuous dynamical tion fails to be differentiable. These notions include, in addi-
systems and differential inclusions can be analyzed with tion to the two considered in this section, the generalized
smooth functions, as the following example shows. (super or sub) differential, the (upper or lower, right or left)
Dini derivative, and the contingent derivative [19], [24], [42],
Examples 5 and 6 Revisited: Stability [43]. Here, we focus on the notions of the generalized gradi-
of the Origin for the Sign Function ent and proximal subdifferential because of their role in pro-
We have already established that the vector fields (11) and viding stability tools for discontinuous differential equations.
(12) in examples 5 and 6, respectively, have unique Filip-
pov solutions starting from each initial condition. Now, The Generalized Gradient
consider the smooth Lyapunov function f : R → R, where of a Locally Lipschitz Function
f (x) = x2 /2. Now, for all x = 0, we have Rademacher’s theorem [9] states that every locally Lipschitz
function is differentiable almost everywhere in the sense of
∇ f (x) · X(x) = −|x| < 0. Lebesgue measure. When considering a locally Lipschitz
function as a candidate Lyapunov function, this statement
Since, according to (20), F[X](x) = {X(x)} on R \{0}, we may raise the question of whether to disregard those points
deduce that the function f is decreasing along every
Filippov solution of (11) and (12) that starts on R \{0}.
Therefore, we conclude that the equilibrium x = 0 is glob- x
ally asymptotically stable.  2.75
However, nonsmooth Lyapunov functions may be
needed if dealing with discontinuous dynamics, as the fol- 2.5
lowing example shows.
2.25
Example 3 Revisited: The Nonsmooth Harmonic Oscillator
2
Does Not Admit a Smooth Lyapunov Function
Consider the vector field for the nonsmooth harmonic
1.75
oscillator in Example 3. We reason as in [16]. As stated in
the section “Piecewise Continuous Vector Fields and Slid-
1.5
ing Motions,” all of the Filippov solutions of (3) are period-
ic and correspond to the curves plotted in Figure 3(b).
1.25
Therefore, if a smooth Lyapunov function f : R2 → R
exists, then it must be constant on each Filippov solution.
t
Since the level sets of f must be one dimensional, it fol- .2 .4 .6 .8 1
lows that each curve must be a level set. This property con-
FIGURE 8 Illustration of the notion of sample-and-hold solution. For
tradicts the fact that the function is smooth, since the level
the control system ẋ = u, we choose the control input u(x) = x. The
sets of a smooth function are also smooth, and the Filippov upper curve is the classical solution starting from x0 = 1, while the
solutions plotted in Figure 3(b) are not smooth along the lower curve is the sample-and-hold solution starting from x0 = 1
x2 -coordinate axis.  corresponding to the π -partition {0, 1/4, 1/2, 3/4, 1} of [0, 1].

JUNE 2008 «
IEEE CONTROL SYSTEMS MAGAZINE 55
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 where the gradient does not exist. Conceivably, the solu- everywhere except at zero, where it is only differentiable. It
tions of the dynamical system stay for almost all time away can be shown that ∇ f (0) = 0, while ∇ f (x) =
from “bad” points where no gradient of the function exists. 2x sin(1/x) − cos(1/x) for x = 0. From (37), we deduce that
However, this assumption is not always valid. As we show ∂ f (0) = [−1, 1], which is different from {∇ f (0)} = {0}. This
in Example 16 below, the solutions of a dynamical system example illustrates that Proposition 6-iii) may not be valid if
may insist on staying on the “bad” points forever. In that f is not continuously differentiable at the point x. 
case, having some gradient-like information is helpful.
Let f : Rd → R be a locally Lipschitz function, and let Computing the Generalized Gradient
f ⊂ R denote the set of points where f fails to be
d Computation of the generalized gradient of a locally
differentiable. The generalized gradient ∂ f : Rd → B(Rd) of Lipschitz function is often a difficult task. In addition to
f is defined by the brute force approach of working from the definition
(37), various results are available to facilitate this compu-
 
tation. Many results that are valid for ordinary deriva-
∂ f (x)  co lim ∇ f (xi) : xi → x , xi ∈
/ S∪ f , (37)
i→∞ tives have a counterpart in this setting. We summarize
some of these results here, and refer the reader to [9] and
where co denotes convex hull. In this definition, S ⊂ Rd is [24] for a complete exposition. In the statements of these
a set of measure zero that can be arbitrarily chosen to sim- results, the notion of a regular function plays a prominent
plify the computation. The resulting set ∂ f (x) is indepen- role; see “Regular Functions” for a precise definition.
dent of the choice of S. From the definition, the generalized Dilation Rule: If f : Rd → R is locally Lipschitz at x ∈ Rd
gradient of f at x consists of all convex combinations of all and s ∈ R, then the function s1 f1 + s2 f2 is locally Lipschitz
of the possible limits of the gradient at neighboring points at x, and
where f is differentiable. Equivalent definitions of the gen-
eralized gradient are given in [9]. ∂(sf )(x) = s∂ f (x). (38)
Some useful properties of the generalized gradient are
summarized in the following result [24]. Sum Rule: If f1 , f2 : Rd → R are locally Lipschitz at
x ∈ Rd and s1 , s2 ∈ R, then the function s1 f1 + s2 f2 is local-
Proposition 6 ly Lipschitz at x, and
If f : Rd → R is locally Lipschitz at x ∈ Rd , then the fol-
lowing statements hold: ∂(s1 f1 + s2 f2 )(x) ⊆ s1 ∂ f1 (x) + s2 ∂ f2 (x), (39)
i) ∂ f (x) is nonempty, compact, and convex.
ii) The set-valued map ∂ f : Rd → B(Rd), x → ∂ f (x), is where the sum of two sets A1 and A2 is defined as
upper semicontinuous and locally bounded at x. A1 + A2 = {a1 + a2 : a1 ∈ A1 , a2 ∈ A2 }. Moreover, if f1 and
iii) If f is continuously differentiable at x, then f2 are regular at x, and s1 , s2 ∈ [0, ∞), then equality holds
∂ f (x) = {∇ f (x)}. and s1 f1 + s2 f2 is regular at x.
Let us compute the generalized gradient for an illustra- Product Rule: If f1 , f2 : Rd → R are locally Lipschitz at
tive example. x ∈ Rd , then the function f1 f2 is locally Lipschitz at x, and

Example 13: Generalized Gradient

∂( f1 f2 )(x) ⊆ f2 (x)∂ f1 (x) + f1 (x)∂ f2 (x). (40)
of the Absolute Value Function
Consider the locally Lipschitz function f : R → R, where
f (x) = |x|, which is continuously differentiable every- Moreover, if f1 and f2 are regular at x, and f1 (x), f2 (x) ≥ 0,
where except at zero. Since ∇ f (x) = 1 for x > 0 and then equality holds and f1 f2 is regular at x.
∇ f (x) = −1 for x < 0, from (37), we deduce Quotient Rule: If f1 , f2 : Rd → R are locally Lipschitz at
x ∈ Rd , and f2 (x) = 0, then the function f1 /f2 is locally Lip-
∂ f (0) = co{1, −1} = [−1, 1], schitz at x, and
 
while, by Proposition 6-iii), we deduce ∂ f (x) = {1} for f1 1
∂ (x) ⊆ 2 ( f2 (x)∂ f1 (x) − f1 (x)∂ f2 (x)). (41)
x > 0 and ∂ f (x) = {−1} for x < 0.  f2 f2 (x)
The next example shows that Proposition 6-iii) may not
be true if f is not continuously differentiable. Moreover, if f1 and − f2 are regular at x, and f1 (x) ≥ 0 and
f2 (x) > 0, then equality holds and f1 /f2 is regular at x.
Example 14: Generalized Gradient of a Differentiable Chain Rule: If each component of h : Rd → Rm is local-
But Not Continuously Differentiable Function ly Lipschitz at x ∈ Rd and g : Rm → R is locally Lipschitz
Consider the locally Lipschitz function f : R → R , where at h(x) ∈ Rm , then the function g ◦ h is locally Lipschitz
f (x) = x2 sin(1/x), which is continuously differentiable at x, and

56 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 
m
∂(g ◦ h)(x) ⊆ co αk ζk : Regular Functions
k =1 o introduce the notion of regular function, we need to first
(α1 , . . . , αm ) ∈ ∂g(h(x)),

T define the right directional derivative and the generalized right
directional derivative. Given f : Rd → R, the right directional
(ζ1 , . . . , ζm ) ∈ ∂h1 (x) × · · · × ∂hm (x) . derivative of f at x in the direction v ∈ Rd is defined as
(42) f (x + hv) − f (x)
f (x; v) = lim+ ,
h→0 h
Moreover, if g is regular at h(x), each component of h is when this limit exists. On the other hand, the generalized direc-
regular at x, and ∂g(h(x)) ⊂ [0, ∞)m , then equality holds tional derivative of f at x in the direction v ∈ Rd is defined as
and g ◦ h is regular at x.
The following useful result from [9, Prop. 2.3.12] con-
cerns the generalized gradient of the maximum and the f (y + hv ) − f (y)
f o (x; v ) = lim sup
minimum of a finite set of functions. y →x h
h →0+

f (y + hv ) − f (y)
= lim+ sup .
Proposition 7 δ →0 y∈B(xδ)
ε →0+ h∈[0,ε)
h
For k ∈ {1, . . . , m}, let fk : Rd → R be locally Lipschitz at
x ∈ Rd , and define the functions fmax , fmin : Rd → R by The advantage of the generalized directional derivative compared
to the right directional derivative is that the limit always exists.
fmax (y)  max{ fk (y) : k ∈ {1, . . . , m}}, When the right directional derivative exists, these quantities may
fmin (y)  min{ fk (y) : k ∈ {1, . . . , m}}. be different. When they are equal, we call the function regular.
More formally, a function f : Rd → R is regular at x ∈ Rd if, for all
v ∈ Rd , the right directional derivative of f at x in the direction of v
Then, the following statements hold:
exists, and f (x; v) = f o(x; v). A function that is continuously dif-
i) fmax and fmin are locally Lipschitz at x.
ii) Let Imax (x) denote the set of indices k for which ferentiable at x is regular at x. Also, a convex function is regular
fk (x) = fmax (x). Then (cf. [9, Prop. 2.3.6]).
The function g : R → R, g (x ) = −|x |, is not regular. Since
 g is continuously differentiable everywhere except for zero, it is
∂ fmax (x) ⊆ co {∂ fi(x) : i ∈ Imax (x)}. (43)
regular on R\{0}. However, its directional derivatives
 
Furthermore, if fi is regular at x for all i ∈ Imax (x), −v, v > 0, v, v > 0,
g (0; v) = go(0; v) =
v, v < 0, −v, v < 0,
then equality holds and fmax is regular at x.
iii) Let Imin (x) denote the set of indices k for which do not coincide. Hence, g is not regular at zero.
fk (x) = fmin (x). Then


∂ fmin (x) ⊆ co {∂ fi(x) : i ∈ Imin (x)}. (44)
⎧
⎨ {1}, x > 0,
Furthermore, if − fi is regular at x for all i ∈ Imin (x),
∂ f (x) = [−1, 1], x = 0, (45)
then equality holds and − fmin is regular at x. ⎩
{−1}, x < 0.
It follows from Proposition 7 that the maximum of a
finite set of continuously differentiable functions is a
locally Lipschitz and regular function whose generalized This result is obtained in Example 13 by direct
gradient at each point x is easily computable as the con- computation. 
vex hull of the gradients of the functions that attain the
maximum at x.
Example 15: Generalized Gradient
Example 13 Revisited: Generalized Gradient of the Minus Absolute Value Function
of the Absolute Value Function The minimum of a finite set of regular functions is not
The absolute value function f (x) = |x| can be rewritten as always regular. A simple example is given by
f (x) = max{x, −x}. Both x → x and x → −x are continuous- g(x) = min{x, −x} = −|x|, which is not regular at zero, as
ly differentiable and hence locally Lipschitz and regular. we show in “Regular Functions.” However, according to
Therefore, according to Proposition 7-i) and -ii), f is locally Proposition 7-iii), this fact does not mean that its general-
Lipschitz and regular, and its generalized gradient is ized gradient cannot be computed. Indeed,

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 57

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 Just as a standard map or function takes a point in its domain to a
single point in another space, a set-valued map takes a point
in its domain to a set of points in another space.

⎧
⎨ {−1}, x > 0, Example 16: Minimum-Distance-
∂g(x) = [−1, 1], x = 0, (46) to-Polygonal-Boundary Function
⎩
{1}, x < 0. Let Q ⊂ R2 be a convex polygon. Consider the minimum
distance function smQ : Q → R from a point within the
This result can also be obtained by combining (45) with the polygon to its boundary defined by
application of the dilation rule (38) to the function
f (x) = |x| with the parameter s = −1.  smQ (p)  min{p − q : q ∈ bndry(Q)}.

Critical Points and Directions of Descent Note that the value of smQ at p is the radius of the largest
A critical point of f : Rd → R is a point x ∈ Rd such that disk with center p contained in the polygon. Moreover, the
0 ∈ ∂ f (x). According to this definition, the maximizers and function smQ is locally Lipschitz on Q. To show this,
minimizers of a locally Lipschitz function are critical rewrite smQ as
points. As an example, x = 0 is the unique minimizer of
f (x) = |x|, and, indeed, we see that 0 ∈ ∂ f (0) in (45).
smQ (p)  min{dist(p, e) : e is an edge of Q},
If a function f is continuously differentiable, then the gra-
dient ∇ f provides the direction of maximum ascent of f
(respectively, −∇ f provides the direction of maximum where dist(p, e) denotes the Euclidean distance from the
descent of f ). When we consider a locally Lipschitz function, point p to the edge e. Indeed, the function smQ is concave
however, the question of choosing the directions of descent on Q.
among all of the available directions in the generalized gra- Let us consider the generalized gradient vector field
dient arises. Without loss of generality, we restrict our dis- corresponding to smQ , where the definition of smQ is
cussion to directions of descent, since a direction of descent extended outside Q by setting smQ (p) = − min{p − q2 :
of − f corresponds to a direction of ascent of f , while f is q ∈ bndry (Q)} for p ∈
/ Q. Applying Proposition 7-iii), we
locally Lipschitz if and only if − f is locally Lipschitz. deduce that −smQ is regular on Q and its generalized
Let Ln : B(Rd) → B(Rd) be the set-valued map that gradient at p ∈ Q is
associates to each subset S of Rd the set of least-norm ele-
ments of its closure S. If the set S is convex and closed, ∂smQ (p) = co{ne : e edge of Q
then the set Ln(S) is a singleton, which consists of the such that smQ (p) = dist(p, e)},
orthogonal projection of zero onto S. For a locally Lipschitz
function f , consider the generalized gradient vector field where ne denotes the unit normal to the edge e pointing
Ln(∂ f ) : Rd → Rd defined by toward the interior of Q. Therefore, at points p in Q for
which there exists a unique edge e of Q nearest to p, the
function smQ is differentiable, and its generalized gradient
x → Ln(∂ f )(x)  Ln(∂ f (x)).
vector field is given by Ln(smQ )(p) = ne . Note that this
vector field corresponds to the move-away-from-nearest-
It turns out that −Ln(∂ f )(x) is a direction of descent of f at neighbor interaction law for one agent moving in the poly-
x ∈ Rd . More precisely [9], if 0 ∈ / ∂ f (x), then there exists gon introduced in Example 4!
T > 0 such that At points p of Q where various edges {e1 , . . . , em } are at
the same minimum distance to p, the function smQ is not
t differentiable, and its generalized gradient vector field is
f (x − tLn(∂ f )(x)) ≤ f (x) − Ln(∂ f )(x)22 < f (x),
2 given by the least-norm element in co{ne1 , . . . , nem }. If p is
0 < t < T, (47) not a critical point, zero does not belong to co{ne1 , . . . , nem },
and the least-norm element points in the direction of the
that is, by taking a small step in the direction −Ln(∂ f )(x), bisector line between two of the edges in {e1 , . . . , em }.
the function f is guaranteed to decrease by an amount that Figure 9 shows a plot of the generalized gradient vector
scales linearly with the stepsize. field of smQ on the square Q = [−1, 1]2 . Note the similarity

58 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 with the plot in Figure 6(a). The critical points of smQ are
characterized [44] by the statement that 0 ∈ ∂smQ (p) if and
only if p belongs to the incenter set of Q. 1
The incenter set of Q is composed of the centers of the
largest-radius disks contained in Q. In general, the incenter
set is a segment (consider the example of a rectangle).
.5
However, if 0 ∈ interior(∂smQ (p)), then the incenter set of
Q is the singleton {p}. 

Nonsmooth Gradient Flows

x2
0
Given a locally Lipschitz function f : Rd → R , the non-
smooth analog of the classical gradient descent flow of a
differentiable function is defined by
−.5
ẋ(t) = −Ln(∂ f )(x(t)). (48)

According to (47), unless the flow is already at a critical

−1
point, −Ln(∂ f )(x) is a direction of descent of f at x. Since −1 −.5 0 .5 1
this nonsmooth gradient vector field is discontinuous, we x1
have to specify a notion of solution for (48). In this case, we
use the notion of Filippov solution due largely to the FIGURE 9 Generalized gradient vector field. This plot shows the
remarkable fact [36] that the Filippov set-valued map asso- generalized gradient vector field of the minimum-distance-to-
ciated with the nonsmooth gradient flow of f given by (48) polygonal-boundary function sm Q : Q → R on the square [−1, 1]2 .
The vector field is discontinuous on the diagonals of the square.
is precisely the generalized gradient of the function, as the Note the similarity with the phase portrait of the move-away-from-
next result states. nearest-neighbor interaction law for one agent moving in the square
[−1, 1]2 ⊂ R2 plotted in Figure 6(a).
Proposition 8: Filippov Set-Valued
Map of Nonsmooth Gradient case. These notions allow us to study the asymptotic
If f : Rd → R is locally Lipschitz, then the Filippov set-val- convergence properties of the trajectories of nonsmooth
ued map F[Ln(∂ f )] : Rd → B(Rd) of the nonsmooth gradi- gradient flows.
ent of f is equal to the generalized gradient
∂ f : Rd → B(Rd) of f , that is, for x ∈ Rd , The Proximal Subdifferential of a
Lower Semicontinuous Function
F[Ln(∂ f )](x) = ∂ f (x). (49) A complementary set of nonsmooth analysis tools for deal-
ing with Lyapunov functions arises from the concept of
As a consequence of Proposition 8, the discontinuous proximal subdifferential. This concept has the advantage of
system (48) is equivalent to the differential inclusion being defined for the class of lower semicontinuous func-
tions, which is larger than the class of locally Lipschitz
functions. The generalized gradient provides us with direc-
ẋ(t) ∈ −∂ f (x(t)). (50)
tional descent information, that is, directions along which
the function decreases. The price we pay for using the prox-
Solutions of (48) are sometimes called “slow motions” of imal subdifferential is that explicit descent directions are
the differential inclusion because of the use of the least- not generally known to us. Nevertheless, the proximal sub-
norm operator, which selects the vector in the generalized differential allows us to reason about the monotonic prop-
gradient of the function with the smallest magnitude. How erties of the function, which, as we show in the section
can we analyze the asymptotic behavior of the solutions of “Stability Analysis by Means of the Proximal Subdifferen-
this system? If the function f is differentiable, then the tial of a Nonsmooth Lyapunov Function,” turns out to be
invariance principle allows us to deduce that, for functions sufficient to provide stability results. The proximal subdif-
with bounded level sets, the solutions of the gradient flow ferential is particularly appropriate when dealing with con-
converge to the set of critical points. The key tool behind vex functions. Here, we briefly touch on the topic of convex
this result is to establish that the function decreases along analysis. For further information, see [45]–[47].
solutions of the system. This behavior is formally A function f : Rd → R is lower semicontinuous [24] at
expressed through the notion of the Lie derivative. In the x ∈ Rd if, for all ε ∈ (0, ∞), there exists δ ∈ (0, ∞) such that,
section “Nonsmooth Stability Analysis,” we discuss gener- for y ∈ B(x, δ), f (y) ≥ f (x) − ε. The epigraph of f is the set of
alizations of the notion of Lie derivative to the nonsmooth points lying on or above its graph, that is,

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 59

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 A Caratheodory solution follows the direction specified by the vector field
except for a set of time instants that has measure zero.

epi( f ) = {(x, μ) ∈ Rd × R : f (x) ≤ μ} ⊂ Rd+1 . The function y ∈ Rd such that x is the nearest point in S to x + λy, for
f is lower semicontinuous if and only if its epigraph is λ > 0 sufficiently small. Then, ζ ∈ ∂P f (x) if and only if
closed. The function f : Rd → R is upper semicontinuous at
x ∈ Rd if − f is lower semicontinuous at x. Note that f is (ζ, −1) ∈ Nepi(f) (x, f (x)). (52)
continuous at x if and only if f is both upper semicontinu-
ous and lower semicontinuous at x. Figure 10(b) illustrates this geometric interpretation.
For a lower semicontinuous function f : Rd → R , the
vector ζ ∈ Rd is a proximal subgradient of f at x ∈ Rd if there Example 17: Proximal Subdifferentials of the
exist σ, δ ∈ (0, ∞) such that, for all y ∈ B(x, δ), Absolute Value Function and Its Negative
Consider the locally Lipschitz functions f, g : R → R ,
f (y) ≥ f (x) + ζ(y − x) − σ 2 ||y − x||22 . (51) f (x) = |x| and g(x) = −|x|. Using the geometric interpreta-
tion of (51), it can be seen that
The set ∂P f (x) of all proximal subgradients of f at x is the
proximal subdifferential of f at x. The proximal subdifferen- ⎧
⎨ {1}, x < 0,
tial at x, which may be empty, is convex but not necessari- ∂P f (x) = [−1, 1], x = 0, (53)
ly open, closed, or bounded. ⎩
{−1}, x > 0,
Geometrically, the definition of a proximal subgradient ⎧
can be interpreted as follows. Equation (51) is equivalent to ⎨ {−1}, x < 0,
∂P g(x) = ∅, x = 0, (54)
saying that, in a neighborhood of x, the function y → f (y) ⎩
{1}, x > 0.
majorizes the quadratic function y → f (x)+ ζ(y − x)
−σ 2 y − x22 . In other words, there exists a parabola that
locally fits under the epigraph of f at (x, f (x)). This geo- Note that (53)–(54) is an example of the fact that ∂P (− f )
metric interpretation is useful for explicitly computing the and −∂P f are not necessarily equal. The generalized gradi-
proximal subdifferential; see Figure 10(a) for an illustra- ent of g in (46) is different from the proximal subdifferen-
tion. An additional geometric characterization of the proxi- tial of g in (54). 
mal subdifferential of f can be given in terms of normal Unlike the case of the generalized gradient, the prox-
cones and the epigraph of f . Given a closed set S ⊆ Rd and imal subdifferential may not coincide with ∇ f (x) when
x ∈ S, the proximal normal cone NS (x) to S is the set of all f is continuously differentiable. The function f : R → R,
where f (x) = −|x|3/2 , is continu-
ously differentiable with
∇ f (0) = 0, but ∂P f (0) = ∅ . In
fact, a continuously differen-
tiable function on R with an
empty proximal subdifferential
1
almost everywhere is provided
in [48]. However, the density
–1 1
theorem (cf. [24, thrm. 3.1]) states
that the proximal subdifferential
–1 of a lower semicontinuous func-
tion is nonempty on a dense set
of its domain of definition,
(a) (b) although a dense set can have
zero Lebesgue measure.
FIGURE 10 Geometric interpretations of the proximal subdifferential of the function x → |x| at On the other hand, the proxi-
x = 0 computed in Example 17. The epigraph of the function is shaded, while the proximal nor-
mal subdifferential can be more
mal cone to the epigraph at zero is striped. In (a), according to (51) each proximal subgradient
corresponds to a direction tangent to a parabola that fits under the epigraph of the function. In useful than the generalized gradi-
(b), according to (52), each proximal subgradient can be uniquely associated with an element of ent in some situations, as the fol-
the proximal normal cone to the epigraph of the function. lowing example shows.

60 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 Example 18: Proximal Subdifferential locally Lipschitz at h(x) ∈ Rm . If ζ ∈ ∂P (g ◦ h)(x) and ε > 0,
of the Square Root of the Absolute Value then there exist x̃ ∈ Rd , ỹ ∈ Rm , and γ ∈ ∂P g( ỹ) with
√
Consider the function f : R → R , where f (x) = |x| , max{||x̃ − x||2 , || ỹ − h(x)||2 } < ε such that ||h(x̃) − h(x)||2
which is continuous at zero, but not locally Lipschitz at < ε and
zero, which is the global minimizer of f . The generalized
gradient does not exist at zero, and hence we cannot char- ζ ∈ ∂P (γ T h)(x̃) + εB(0, 1), (59)
acterize this point as a critical point. However, the function
f is lower semicontinuous with proximal subdifferential
where γ T h : Rd → R is defined by (γ T h)(x) = γ T h(x).
 
⎧ √ 1 The statement of the chain rule (59) shows a characteris-
⎪ , x > 0,
⎪
⎪ 2 x tic feature of the proximal subdifferential. That is, rather
⎪
⎨
than at a specific point of interest, the proximal subdiffer-
∂P f (x) = R, x = 0, (55)
⎪
⎪ ential can only be expressed with objects evaluated at
⎪
⎪ 
⎩ neighboring points.
− 2√1−x , x < 0.
Computation of the proximal subdifferential of a twice
continuously differentiable function is particularly simple.
As we discuss later, it follows from (55) that zero is the If f : Rd → R is twice continuously differentiable on the
unique global minimizer of f .  open set U ⊆ Rd , then, for all x ∈ U,
If f : R d → R is locally Lipschitz at x ∈ R d , then the
proximal subdifferential of f at x is bounded. In general, ∂P f (x) = {∇ f (x)}. (60)
the relationship between the generalized gradient and the
proximal subdifferential of a function f that is locally This simplicity also works for continuously differentiable
Lipschitz at x ∈ Rd is expressed by convex functions, as the following result [20], [24], [47]
  states. Note that every convex function on Rd is locally
∂ f (x) = co lim ζn ∈ Rd : ζn ∈ ∂P f (xn ) and lim xn = x . Lipschitz, and hence continuous.
n→∞ n→∞

Computing the Proximal Subdifferential Proposition 9

As with the generalized gradient, computation of the prox- If f : Rd → R is convex, then the following statements
imal subdifferential of a lower semicontinuous function is hold:
often far from straightforward. Here we provide some use- i) For x ∈ Rd , ζ ∈ ∂P f (x) if and only if, for all y ∈ Rd ,
ful results following the exposition in [24]. f (y) ≥ f (x) + ζ(y − x).
Dilation Rule: If f : Rd → R is lower semicontinuous at ii) The set-valued map ∂P f : Rd→ B(Rd), x → ∂P f (x),
x ∈ Rd and s ∈ (0, ∞), then the function sf is lower semi- takes nonempty, compact, and convex values and is
continuous at x, and upper semicontinuous and locally bounded.
iii) If, in addition, f is continuously differentiable, then,
∂P (sf )(x) = s ∂P f (x). (56) for all x ∈ Rd , ∂P f (x) = {∇ f (x)}.

Sum Rule: If f1 , f2 : Rd → R are lower semicontinuous Example 17 Revisited: Proximal

at x ∈ Rd , then the function f1 + f2 is lower semicontinu- Subdifferential of the Absolute Value Function
ous at x, and The computation (53) of the proximal subdifferential of the
function f (x) = |x| can be performed using Proposition 9
∂P f1 (x) + ∂P f2 (x) ⊆ ∂P ( f1 + f2 )(x). (57) by observing that f is convex. By Proposition 9-i), for
x = 0, ζ ∈ ∂P f (0) if and only if, for all y ∈ R,
Moreover, if either f1 or f2 are twice continuously differ-
entiable, then equality holds. |y| ≥ |0| + ζ(y − 0) = ζ y. (61)
Fuzzy Sum Rule: If f1 , f2 : Rd → R is lower semicontin-
uous at x ∈ Rd , ζ ∈ ∂P ( f1 + f2 )(x), and ε > 0, then there Hence, for y ≥ 0, (61) implies ζ ∈ (−∞, 1], while, for y ≤ 0,
exist x1 , x2 ∈ B(x, ε) with | fi(xi) − fi(x)| < ε, i ∈ {1, 2}, such (61) implies that ζ ∈ [−1, +∞). Therefore, ζ ∈ (−∞, 1] ∩
that [−1, +∞), and ∂P f (0) = [−1, 1]. By Proposition 9-iii), for
x > 0, ∂P f (x) = {1}, and, for x < 0, ∂P f (x) = {−1}. 
ζ ∈ ∂P f1 (x1 ) + ∂P f2 (x2 ) + εB(0, 1). (58) x
Regarding critical points, if is a local minimizer of the
lower semicontinuous function f : Rd → R , then
Chain Rule: Assume that either h : Rd → Rm linear and 0 ∈ ∂P f (x). Conversely, if f is convex, and 0 ∈ ∂P f (x), then
g : Rm → R lower semicontinuous at h(x) ∈ Rm , or x is a global minimizer of f . For the study of maximizers,
h : Rd → Rm locally Lipschitz at x ∈ Rd and g : Rm → R instead of lower semicontinuous function, convex function,

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 61

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 Every absolutely continuous function is differentiable almost everywhere.

and proximal subdifferential, the relevant notions are to Filippov solutions by taking F = F[X], and to control
upper semicontinuous function, concave function, and systems by taking F = G[X].
proximal super differential, respectively [24]. Before proceeding with our exposition, we recall that
solutions of a discontinuous system are not necessarily
Gradient Differential Inclusion of a Convex Function unique. Therefore, when considering a property such as
It is not always possible to associate a nonsmooth gradient Lyapunov stability, we must specify whether attention is
flow with a lower semicontinuous function because the being paid to a particular solution starting from an initial
proximal subdifferential might be empty almost every- condition (“weak”) or to all the solutions starting from an
where. However, following Proposition 9-ii), we can asso- initial condition (“strong”). As an example, the set M ⊆ Rd
ciate a nonsmooth gradient flow with each convex is weakly invariant for (63) if, for each x0 ∈ M, M contains at
function, as we briefly discuss next [49]. least one maximal solution of (63) with initial condition x0 .
Let f : Rd → R be convex, and consider the gradient Similarly, M ⊆ Rd is strongly invariant for (63) if, for each
differential inclusion x0 ∈ M, M contains every maximal solution of (63) with
initial condition x0 . Rather than reintroducing weak and
ẋ(t) ∈ −∂P f (x(t)). (62) strong versions of standard stability notions, we rely on
the guidance provided above and the context for the spe-
Using the properties of the proximal subdifferential given cific meaning in each case. Detailed definitions are given in
by Proposition 9-ii), the existence of a Caratheodory solu- [14] and [18].
tion of (62) starting from every initial condition is guaran- The following discussion requires the notion of a limit
teed by Proposition S2. Moreover, the uniqueness of point of a solution of a differential inclusion. The point
Caratheodory solutions can be established as follows. Let x∗ ∈ Rd is a limit point of a solution t → x(t) of (63) if there
x, y ∈ Rd , and take ζ1 ∈ −∂P f (x) and ζ2 ∈ −∂P f (y). Using exists a sequence {tn }n∈N such that x(tn ) → x∗ as n → ∞.
Proposition 9-i), we have We denote by (x) the set of limit points of t → x(t).
Under the hypotheses of Proposition S2, (x) is a weakly
f (y) ≥ f (x) − ζ1 (y − x), f (x) ≥ f (y) − ζ2 (x − y). invariant set. Moreover, if the solution t → x(t) lies in a
bounded, open, and connected set, then (x) is nonempty,
From here, we deduce −ζ1 (y − x) ≤ f (y) − f (x) ≤ bounded, connected, and x(t) → (x) as t → ∞; see [18].
−ζ2 (y − x), and therefore (ζ2 − ζ1 )(y − x) ≤ 0, which, in
particular, implies that the set-valued map x → −∂P f (x) Stability Analysis by Means of the Generalized
satisfies the one-sided Lipschitz condition (S3). Gradient of a Nonsmooth Lyapunov Function
Proposition S3 then guarantees that there exists a unique In this section, we discuss nonsmooth stability analysis
Caratheodory solution of (62) starting from every initial results that use locally Lipschitz functions and generalized
condition. gradients. We present results taken from various sources
in the literature. This discussion is not intended to be a
NONSMOOTH STABILITY ANALYSIS comprehensive account of such a vast topic but rather
In this section, we study the stability of discontinuous serves as a motivation for further exploration. The books
dynamical systems. Unless explicitly mentioned otherwise, [14], [18] and journal articles [51]–[54] provide additional
the stability notions employed here correspond to the usual information.
ones for differential equations; see, for instance, [50]. The pre-
sentation focuses on the time-invariant differential inclusion Lie Derivative and Monotonicity
A common theme in stability analysis is establishing the
ẋ(t) ∈ F(x(t)), (63) monotonic evolution of a candidate Lyapunov function
along the trajectories of the system. Mathematically, the
where F : Rd → B(Rd) . Throughout this section, we evolution of a function along trajectories is captured by the
assume that the set-valued map F satisfies the hypotheses notion of Lie derivative. Our first task is to generalize this
of Proposition S2, so that the existence of solutions of (63) notion to the setting of differential inclusions following
is guaranteed. The scenario (63) has a direct application to [53]; see also [51] and [52].
discontinuous differential equations and control systems. Given a locally Lipschitz function f : Rd → R and a set-
For instance, the results presented here apply to valued map F : Rd→ B(Rd), the set-valued Lie derivative
Caratheodory solutions by taking a singleton-valued map, L̃F f : Rd → B(R) of f with respect to F at x is defined as

62 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 In this article, we focus on Caratheodory, Filippov, and sample-and-hold solutions.


{(sign(x1 ), x2 )}, x1 =
 0,
L̃F f (x) = {a ∈ R : there exists v ∈ F(x) such that ∂ f (x1 , x2 ) =
[−1, 1] × {x2 }, x1 = 0.
ζ T v = a for all ζ ∈ ∂ f (x)}. (64)

With this information, we can compute the set-valued Lie

If F takes convex and compact values, then, for each derivative L̃X f : R2 → B(R) as
x ∈ Rd , L̃F f (x) is a closed and bounded interval in R, pos- ⎧
⎨ {0}, x1 = 0,
sibly empty. If f is continuously differentiable at x, then
L̃X f (x1 , x2 ) = ∅, x1 = 0 and x2 = 0, (66)
⎩
{0}, x1 = 0 and x2 = 0.
L̃F f (x) = {(∇ f (x))T v : v ∈ F(x)}.
From (65) and (66) we conclude that f is constant along
the Filippov solutions of the discontinuous dynamical
The usefulness of the set-valued Lie derivative stems system. Indeed, the level sets of f are exactly the curves
from the fact that it allows us to study how the function described by the solutions of the system in Figure 3(b).
f evolves along the solutions of a differential inclusion
without having to explicitly obtain the solutions. Specifi- Stability Results
cally, we have the following result; see [53]. The above discussion on monotonicity is the stepping
stone to stability results using locally Lipschitz functions
Proposition 10 and generalized gradient information. Proposition 10 pro-
Let x : [0, t1 ] → Rd be a solution of the differential inclu- vides a criterion for determining the monotonic behavior
sion (63), and let f : Rd → R be locally Lipschitz and regu- of a locally Lipschitz function along the solutions of dis-
lar. Then, the following statements hold: continuous dynamics. This result, together with the
i) The composition t → f (x(t)) is differentiable at almost appropriate positive definite assumptions on the candi-
all t ∈ [0, t1 ]. date Lyapunov function, allows us to synthesize check-
ii) The derivative of t → f (x(t)) satisfies able stability tests. We start by formulating [53] the
natural extension of Lyapunov’s stability theorem for
d ordinary differential equations. Recall that, at each x ∈ Rd ,
( f (x(t))) ∈ L̃F f (x(t)) for almost every t ∈ [0, t1 ]. the Lie derivative L̃F f (x) is a set contained in R. For the
dt
(65) empty set, we adopt the convention max ∅ = −∞.

A similar result can be established for a larger class of func- Theorem 1

tions when the set-valued map F is singleton valued [54], [55]. Let F : Rd → B(Rd) be a set-valued map satisfying the
Given the discontinuous vector field X : Rd → Rd , con- hypotheses of Proposition S2, let xe be an equilibrium of
sider the Filippov solutions of (10). In this case, with a the differential inclusion (63), and let D ⊆ Rd be an open
slight abuse of notation, we denote L̃X f = L̃F[X] f . Note and connected set with xe ∈ D . Furthermore, let
that if X is continuous at x, then F[X](x) = {X(x)}, and f : Rd → R be such that the following conditions hold:
therefore, L̃X f (x) corresponds to the singleton {LX f (x)}, i) f is locally Lipschitz and regular on D.
whose sole element is the usual Lie derivative of f in the ii) f (xe ) = 0, and f (x) > 0 for x ∈ D\{xe }.
direction of X at x. iii) max L̃F f (x) ≤ 0 for each x ∈ D.
Then, xe is a strongly stable equilibrium of (63). In
Example 3 Revisited: Monotonicity in addition, if iii) above is replaced by
the Nonsmooth Harmonic Oscillator iii)’ max L̃F f (x) < 0 for each x ∈ D\{xe },
For the nonsmooth harmonic oscillator in Example 3, con- then xe is a strongly asymptotically stable equilibrium
sider the locally Lipschitz and regular map f : R2 → R , of (63).
f (x1 , x2 ) = |x1 | + x22 /2 . Recall that Figure 3(b) shows the Let us apply this result to the nonsmooth harmonic
contour plot of f . Let us determine how f evolves along oscillator.
the Filippov solutions of the dynamical system by looking
at the set-valued Lie derivative. First, we compute the gen- Example 3 Revisited: Stability Analysis of
eralized gradient of f . To do so, we rewrite f as f (x1 , x2 ) = the Nonsmooth Harmonic Oscillator
max{x1 , −x1 } + x22 /2, and apply Proposition 7-ii) and the The function (x1 , x2 ) → |x1 | + x22 /2 satisfies hypotheses
sum rule to find i)–iii) of Theorem 1 on D = Rd . Therefore, we conclude

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 63

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 that zero is a strongly stable equilibrium. The phase portrait in Another useful result in the theory of differential equa-
Figure 3(a) indicates that zero is not strongly asymptotically tions is the invariance principle [50]. In many situations,
stable. Using Theorem 1, it can be shown that the nonsmooth this principle allows us to determine the asymptotic con-
harmonic oscillator under dissipation, with vector field vergence properties of the solutions of a differential equa-
(x1 , x2 ) → (x2 , −sign (x1 ) − k sign (x2 )) , where k > 0, has tion. Here, we build on the above discussion to present a
zero as a strongly asymptotically stable equilibrium. The generalization to differential inclusions (63) and non-
phase portrait of this system is plotted in Figure 11(a), while smooth Lyapunov functions. This principle is thus applica-
several Filippov solutions are plotted in Figure 11(b).  ble to discontinuous differential equations. The
formulation is taken from [53], which slightly generalizes
the presentation in [51].
1
Theorem 2
Let F : Rd → B(Rd) be a set-valued map satisfying the
hypotheses of Proposition S2, and let f : Rd → R be a
.5 locally Lipschitz and regular function. Let S ⊂ Rd be com-
pact and strongly invariant for (63), and assume that
max L̃F f (y) ≤ 0 for each y ∈ S. Then, all solutions
x : [0, ∞) → Rd of (63) starting at S converge to the largest
x2

0
weakly invariant set M contained in

−.5 S ∩ {y ∈ Rd : 0 ∈ L̃F f (y)}.

Moreover, if the set M consists of a finite number of points,

−1
then the limit of each solution starting in S exists and is an
element of M.
−1 −.5 0 .5 1
x1 We now apply Theorem 2 to nonsmooth gradient flows.
(a)
Stability of Nonsmooth Gradient Flows
1 Consider the nonsmooth gradient flow (48) of a locally
Lipschitz function f : Rd → R . Assume further that f is
regular. Let us examine how f evolves along the solutions
of the flow using the set-valued Lie derivative. Given
.5
x ∈ Rd , let a ∈ L̃−Ln(∂ f ) f (x). By definition, there exists
v ∈ F[−Ln(∂ f )](x) such that

a = ζ T v for all ζ ∈ ∂ f (x).

x2

0 (67)

Recall from (49) that F [−Ln(∂ f )](x) = −∂ f (x). Since (67)

−.5 holds for all elements in the generalized gradient of f at x,
we can choose in particular ζ = −v ∈ ∂ f (x). Therefore,

a = (−v)T v = −||v||22 ≤ 0. (68)

−1
−1 −.5 0 .5 1
From (68), we conclude that all of the elements of
x1
(b)
L̃−Ln(∂ f ) f (x) belong to (−∞, 0], and therefore, from (65),
f monotonically decreases along the solutions of its
FIGURE 11 Nonsmooth harmonic oscillator with dissipation. In nonsmooth gradient flow (48). Moreover, we deduce
(a), the phase portrait on [−1, 1]2 of the vector field that 0 ∈ L̃−Ln(∂ f ) f (x) if and only if 0 ∈ ∂ f (x), that is, if x is
(x1 , x 2 ) → (x 2 , − sign (x1 ) − k sign (x 2 )), where k = 0.75, while (b) a critical point of f . The application of the Lyapunov
shows some Filippov solutions of the associated dynamical system
stability theorem and the invariance principle in theo-
(10). In the nonsmooth harmonic oscillator in Example 3, the equilib-
rium at the origin is strongly stable, but not strongly asymptotically rems 1 and 2, respectively, gives rise to the following
stable; see Figure 3. The addition of dissipation renders the equilib- nonsmooth counterpart of the classical smooth result for
rium at the origin strongly asymptotically stable. a gradient flow [56].

64 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 (a) (b) (c) (c) (e)

FIGURE 12 From left to right, evolution of the nonsmooth gradient flow of the function −sm Q in a convex polygon. At each snapshot, the value
of sm Q is the radius of the largest disk (plotted in gray) contained in the polygon with center at the current location. The flow converges in
finite time to the incenter set, which, for this polygon, is a singleton whose only element is the center of the disk in the rightmost snapshot.

Proposition 11 flow monotonically maximizes the radius of the largest

Let f : Rd → R be locally Lipschitz and regular. Then, the disk contained in the polygon (that is, smQ !) until it reaches
strict minimizers of f are strongly stable equilibria of the an incenter point. This fact is illustrated in Figure 12.
nonsmooth gradient flow (48) of f . Furthermore, if the What if, instead, we want to pack an arbitrary number n
level sets of f are bounded, then the solutions of the non- of spheres within the polygon? It turns out that the move-
smooth gradient flow asymptotically converge to the set of away-from-nearest-neighbor interaction law is a discontin-
critical points of f . uous dynamical system that solves this problem, where the
The following example illustrates the above discussion. solutions are understood in the Filippov sense. Figure 13
illustrates the evolution of this dynamical system. The sta-
Example 16 Revisited: Stability Analysis of bility properties of this law can be determined using the
the Nonsmooth Gradient Flow of −smQ Lyapunov function HSP : Qn → R defined by
Consider the nonsmooth gradient flow of −smQ , which is
the minimum-distance-to-polygonal-boundary function 1
introduced in Example 16. Proposition 5 guarantees HSP (p1 , . . . , pn ) = min |pi − pj||2 , dist(pi, e) :
2 
uniqueness of solutions for this flow. Regarding conver-
i = j ∈ {1, . . . n}, e edge of Q .
gence, Proposition 11 guarantees that solutions converge
asymptotically to the incenter set. Indeed, the incenter set is (69)
attained in finite time, and hence each solution converges
to a point of the incenter set [44]. The nonsmooth gradient The value of HSP corresponds to the largest radius such
flow can be interpreted as a sphere-packing algorithm in that spheres with centers p1 , . . . , pn and radius
the sense that, starting from an arbitrary initial point, the HSP (p1 , . . . , pn ) fit within the environment and do not

(a) (b) (c)

FIGURE 13 The move-away-from-nearest-neighbor interaction law for solving a sphere-packing problem within the polygon Q. (a) The initial
configuration, (b) the evolution, and (c) the final configuration of a Filippov solution. In (c), the minimum radius of the shaded spheres corre-
sponds to the value of the locally Lipschitz function HSP , defined in (69). The Filippov solutions of the move-away-from-nearest-neighbor
dynamical system monotonically increase the value of HSP . In (c), every node is at equilibrium since the infinitesimal motion of a node in any
direction would place it closer to at least another node. This discontinuous dynamical system [44] is an example of how simple local interac-
tions can achieve a global objective.

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 65

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 intersect each other (except at most at the boundary). The The proof of Proposition 12 builds on the stability tools
function HSP is locally Lipschitz, but not concave, on Q, presented in this section. Specifically, the invariance prin-
and its evolution is monotonically nondecreasing along the ciple can be used to establish convergence toward the set
move-away-from-nearest-neighbor dynamical system. The of critical points of f . Finite-time convergence can be
reader is referred to [44] for details on the relationship established by deriving bounds on the evolution of f along
between HSP and the minimum distance function smQ , as the solutions of the discontinuous dynamics using the set-
well as additional discontinuous dynamical systems that valued Lie derivative. This analysis also allows us to pro-
solve this sphere-packing problem and other exciting geo- vide upper bounds on the convergence time. A more
metric optimization problems.  comprehensive exposition of results that guarantee finite-
time convergence of general discontinuous dynamics is
Finite-Time-Convergent Nonsmooth Gradient Flows given in [57].
Convergence of the solutions of a dynamical system in
finite time is a desirable property in various settings. For Example 19: Finite-Time Consensus
instance, a motion planning algorithm is effective if it can The ability to reach consensus, or agreement, on some (a
steer a robot from the initial to the final configuration in priori unknown) quantity is critical for multiagent
finite time rather than asymptotically. Another example is systems. Network coordination problems require that
given by a robotic network trying to agree on the exact individual agents agree on the identity of a leader, jointly
value of a quantity sensed by each agent. Reaching agree- synchronize their operation, decide which specific pat-
ment in finite time allows the network to use precise infor- tern to form, balance the computational load, or fuse con-
mation in the completion of other tasks. sistently the information gathered on some spatial
Broadly applicable results on finite-time convergence process. Here, we briefly comment on two discontinuous
for discontinuous dynamical systems can be found in algorithms that achieve consensus in finite time, follow-
[36] and [57]–[59]. Here, we briefly discuss the finite-con- ing [57].
vergence properties of a class of nonsmooth gradient Consider a network of n agents with states
flows. Let f : Rd → R be a continuously differentiable p1 , . . . , pn ∈ R . Let G = ({1, . . . , n}, E) be an undirected
function and assume that all of the level sets of f are graph with n vertices, describing the topology of the net-
bounded. It is well known [56] that all solutions of the work. Two agents i and j agree if and only if pi = pj. The
gradient flow ẋ(t) = −∇ f (x(t)) converge asymptotically disagreement function G : Rn → [0, ∞) quantifies the
toward the set of critical points of f , although the con- group disagreement
vergence does not occur in finite time. Here, we slightly
modify the gradient flow into two different nonsmooth 1 
G (p1 , . . . , pn ) = (pj − pi)2 .
flows that achieve finite-time convergence. Consider the 2 (i, j) ∈E
discontinuous differential equations
It is known [60] that, if the graph is connected, the gra-
dient flow of G achieves consensus with an exponen-
∇ f (x(t))
ẋ(t) = − , (70) tial rate of convergence. Actually, agents agree on the
||∇ f (x(t))||2
average value of their initial states (this consensus is
ẋ(t) = −sign(∇ f (x(t))), (71)
called average consensus). If G is connected, the non-
smooth gradient flow (70) of G achieves average con-
where ·2 denotes the Euclidean distance, and sign (x) = sensus in finite time, and the nonsmooth gradient flow
(sign (x1 ), . . . , sign (xd)) ∈ Rd . The nonsmooth vector field (71) of G achieves consensus on the average of the
(70) always points in the direction of the gradient with unit maximum and the minimum of the initial states in finite
speed. Alternatively, the nonsmooth vector field (71) speci- time; see [57]. 
fies the direction of motion by means of a binary quantiza-
tion of the direction of the gradient. The following result Stability Analysis by Means of the Proximal
(57) characterizes the properties of the Filippov solutions Subdifferential of a Nonsmooth Lyapunov Function
of (70) and (71). This section presents stability tools for differential
inclusions using a lower semicontinuous function as a
Proposition 12 candidate Lyapunov function. We use the proximal
Let f : Rd → R be a twice continuously differentiable subdifferential to study the monotonic evolution of the
function. Let S ⊂ Rd be compact and strongly invariant for candidate Lyapunov function along the solutions of the
(70) [respectively, for (71)]. If the Hessian of f is positive differential inclusion. As in the previous section, we
definite at every critical point of f in S, then every Filippov present a few representative and useful results. We
solution of (70) [respectively, (71)] starting from S con- refer the interested reader to [24] and [25] for a more
verges in finite time to a minimizer of f . detailed exposition.

66 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 Weak and strong monotonicity properties of candidate Lyapunov functions can be
established along the solutions of discontinuous dynamical systems.

Lie Derivative and Monotonicity the role played by the set-valued Lie derivative L̃F f for a
Let D ⊆ Rd be an open and connected set. A lower semi- locally Lipschitz function. These objects allow us to study
continuous function f : Rd → R is weakly nonincreasing on how f evolves along the solutions of a differential inclu-
D for a set-valued map F : Rd → B(Rd) if, for all y ∈ D, sion without having to obtain the solutions in closed form.
there exists a solution x : [0, t1 ] → Rd of the differential Specifically, we have the following result [24]. In this and
inclusion (63) starting at y and contained in D that satisfies in forthcoming statements, it is convenient to adopt the
convention sup ∅ = −∞.
f (x(t)) ≤ f (x(0))
Proposition 13
= f (y) for all t ∈ [0, t1 ].
Let F : Rd → B(Rd) be a set-valued map satisfying the
hypotheses of Proposition S2, and consider the associated
If, in addition, f is continuous, then being weakly non- differential inclusion (63). Let f : Rd → R be a lower semi-
increasing is equivalent to the property of having a solu- continuous function, and let D ⊆ Rd be open. Then, the
tion starting at y such that t → f (x(t)) is monotonically following statements hold:
nonincreasing on [0, t1 ]. i) The function f is weakly nonincreasing on D if and
Similarly, a lower semicontinuous function f : Rd → R only if
is strongly nonincreasing on D for a set-valued map
F : Rd → B(Rd) if, for all y ∈ D , all solutions sup LF f (y) ≤ 0, for all y ∈ D.
x : [0, t1 ] → Rd of the differential inclusion (63) starting at
y and contained in D satisfy ii) If, in addition, either F is locally Lipschitz on D, or F
is continuous on D and f is locally Lipschitz on D,
then f is strongly nonincreasing on D if and only if
f (x(t)) ≤ f (x(0))
= f (y) for all t ∈ [0, t1 ].
sup LF f (y) ≤ 0, for all y ∈ D.

Note that f is strongly nonincreasing if and only if

t → f (x(t)) is monotonically nonincreasing on [0, t1 ] for all Let us illustrate this result in a particular example.
solutions t → x(t) of the differential inclusion.
Given the set-valued map F : Rd → B(Rd) , which Example 20: Cart on a Circle
takes nonempty, compact values, and the lower semicon- Consider, following [25] and [61], the control system on R2
tinuous function f : Rd → R, the lower and upper set-valued
Lie derivatives LF f, LF f : Rd → B(R) of f with respect to 
ẋ1 = x21 − x22 u, (72)
F at y are defined by, respectively,
ẋ2 = 2x1 x2 u, (73)
LF f (y)  {a ∈ R : there exists ζ ∈ ∂P f (y) such that
with u ∈ R . Equations (72)–(73) are named after the
a = min{ζ T v : v ∈ F(y)}},
f a c t that the integral curves of the vector field
LF f (y)  {a ∈ R : there exists ζ ∈ ∂P f (y) such that (x1 , x2 ) → (x21 − x22 , 2x1 x2 ) are circles whose centers lie on
a = max{ζ T v : v ∈ F(y)}}. the x2 -axis and are tangent to the x1 -axis; see the phase
portrait in Figure 14(a).
If, in addition, F takes convex values, then for each Let X : R2 × R → R2 be defined by X((x1 , x2 ), u) =
ζ ∈ ∂P f (y), the set {ζ T v : v ∈ F(y)} is a closed interval of (x21 − x22 , 2x1 x2 ) u. Following (34), consider the associated
the form [min {ζ T v : v ∈ F(y)}, max {ζ T v : v ∈ F(y)}]. Note set-valued map G[X] : R2 → B(R2 ) defined by
that the lower and upper set-valued Lie derivatives at a G[X](x1 , x2 ) = {X((x1 , x2 ), u) : u ∈ R}. Since G[X] does not
point y might be empty. take compact values, we instead take a nondecreasing
The lower and upper set-valued Lie derivatives play a map σ : [0, ∞) → [0, ∞), and consider the set-valued
role for a lower semicontinuous function that is similar to map Fσ : R2 → B(R2 ) given by

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 67

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 Fσ (x1 , x2 ) = {X((x1 , x2 ), u) ∈ R2 : |u| ≤ σ ((x1 , x2 )2 )}. With this information, we compute the set
(74)
{ζ T v : ζ ∈ ∂P f (x1 , x2 ), v ∈ Fσ (x1 , x2 )}
Consider the locally Lipschitz function f : R2 → R ⎧ 
(x21 +x22 )
2
defined by ⎪
⎪ u : |u| ≤ σ ((x , x ) ) , x1 > 0,
⎪
⎪ x21 +x22 +x1 x21 +x22
1 2 2
⎪
⎪
⎧ x21 +x22 ⎨
⎨ , x = 0, = ∅, x1 = 0,
f (x1 , x2 ) =
x21 +x22 +|x1 | ⎪
⎪
⎩ ⎪
⎪ 
⎪
⎪ (x21 +x22 )
2
0, x = 0. ⎩ −u : |u| ≤ σ ((x , x
1 2 2 ) ) , x1 < 0.
2
x1 +x2 −x1
2 2
x1 +x2
2

The level set curves of f are depicted in Figure 14(c). Let

us determine how f evolves along the solutions of the We are now ready to compute the lower and upper set-
control system by using the lower and upper set-valued valued Lie derivatives as
Lie derivatives. First, we compute the proximal subdiffer-
ential of f . Using the fact that f is twice continuously dif- ⎧ (x2 +x2 )
3/2

ferentiable on the open right and left half-planes, together ⎨ −σ ((x1 , x2 )2 ) 21 22 , x1 = 0,
x1 +x2 +|x1 |
LFσ f (x1 , x2 ) =
with the geometric interpretation of proximal subgradi- ⎩
ents, we obtain −∞, x1 = 0,
(75)
⎧ (x2 +x2 )
3/2
⎨ σ ((x1 , x2 )2 ) 21 22 , x1 = 0,
∂P f (x⎧
1 , x2 )
⎧ ⎪⎛  ⎞⎫ LFσ f (x1 , x2 ) = x1 +x2 +|x1 |
⎨ ⎪
⎬ ⎩
⎪
⎪ ⎜ x1 +x2 −2x1 x1 +x2 2 1
2 2 2 2 x 2x + x 2
+x2
2 ⎟ −∞, x1 = 0.
⎪
⎪ ⎝− 2 2 ,
1
 ⎠ , x1 > 0,
⎪
⎪ ⎪ 2
⎪
⎪
⎪ ⎩ x1 +x2 +x1 x21 +x22 x1 + x21 +x22 ⎭ (76)
⎪
⎪
⎪
⎨
= ∅, x1 = 0, Therefore, sup LFσ f (x1 , x2 ) ≤ 0 for all (x1 , x2 ) ∈ R2 . Now
⎪
⎪
⎪
⎪ ⎧⎛  ⎞⎫ using Proposition 13-i), we deduce that f is weakly
⎪
⎪ ⎪ ⎪
⎪ ⎨
⎪ ⎜ x1 +x2 +2x1 x1 +x2 x −2x + x2
+x 2
2 ⎟
⎬ nonincreasing on R2 . Since f is continuous, this fact is
⎪
⎪
2 2 2 2
,
2 1 1
⎪ ⎝ 2 2
⎩ 2 ⎠ , x1 < 0. equivalent to saying that there exists a control input u
⎪
⎩ x1 +x2 −x1 x1 +x2
2 2
x1 − x21 +x22
⎪
⎭
such that the solution t → x(t) of the dynamical system

x2

2
2
2

1
1 1

0
x2

0 x1
x2

−1.5 −1 −0.5 0.5 1 1.5

−1 −1
−1

−2
−2 −2
−2 −1 0 1 2
−2 −1 0 1 2 −3 x1
x1 (c)
(a) (b)

 the input vector field (x1 , x 2 ) → (x1 − x 2 , 2x1 x 2 ), (b) its integral curves, and (c) the con-
2 2
FIGURE 14 Cart on a circle. (a) The phase portrait of
tour plot of the function 0 = (x1 , x 2 ) → (x1 + x 2 /( x1 + x 2 + |x1 |)), (0, 0) → 0. The origin cannot be asymptotically stabilized by means of
2 2 2 2

continuous feedback in the cart dynamics (72)–(73). However, the origin can be asymptotically stabilized by means of discontinuous feedback
when solutions are understood in the sample-and-hold sense.

68 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 The notion of sample-and-hold solution plays a key role in the
stabilization of asymptotically controllable systems.

resulting from substituting u in (72)–(73) has the prop- f, g : Rd → R are a Lyapunov pair for an equilibrium
erty that t → f (x(t)) is monotonically nonincreasing.  xe ∈ Rd if they satisfy f (x), g(x) ≥ 0 for all x ∈ Rd , and
g(x) = 0 if and only if x = xe ; f is radially unbounded,
Stability Results and, moreover,
The results presented in the previous section establishing
the monotonic behavior of a lower semicontinuous func- sup LF f (x) ≤ −g(x) for all x ∈ Rd.
tion allow us to provide tools for stability analysis. We pre-
sent here an exposition parallel to the discussion for the If an equilibrium xe of (63) admits a Lyapunov pair, then there
locally Lipschitz function and generalized gradient case. exists at least one solution starting from every initial condition
We start by presenting a result on Lyapunov stability that that asymptotically converges to the equilibrium; see [24].
follows from the exposition in [24].
Example 20 Revisited: Asymptotic Stability
Theorem 3 of the Origin in the Cart Example
Let F : Rd → B(Rd) be a set-valued map satisfying the As an application of the above discussion and the version
hypotheses of Proposition S2. Let xe be an equilibrium of of Theorem 3 for weak stability, consider the cart on a circle
the differential inclusion (63), and let D ⊆ Rd be a domain introduced in Example 20. Setting xe = (0, 0) and D = R2 ,
with xe ∈ D. Let f : Rd → R and assume that the following and taking into account the computation (75) of the lower
conditions hold: set-valued Lie derivative, we conclude that (0, 0) is a glob-
i) F is continuous on D and f is locally Lipschitz on D, ally weakly asymptotically stable equilibrium. 
or F is locally Lipschitz on D and f is lower semicon- We now turn our attention to the extension of the
tinuous on D, and f is continuous at xe . invariance principle for differential inclusions using the
ii) f (xe ) = 0, and f (x) > 0 for x ∈ D\{xe }. proximal subdifferential of a lower semicontinuous func-
iii) sup LF f (x) ≤ 0 for all x ∈ D. tion. The following result can be derived from the exposi-
Then, xe is a strongly stable equilibrium of (63). In tion in [24].
addition, if iii) above is replaced by
iii)’ sup LF f (x) ≤ 0 for all x ∈ D\{xe }, Theorem 4
then xe is a strongly asymptotically stable equilibrium Let F : Rd → B(Rd) be a set-valued map satisfying the
of (63). hypotheses of Proposition S2, and let f : Rd → R. Assume
A similar result can be stated for weakly stable that either F is continuous and f is locally Lipschitz, or F
equilibria substituting i) by “i)’ f is continuous on D,” is locally Lipschitz and f is continuous. Let S ⊂ Rd be
and the upper set-valued Lie derivative by the lower compact and strongly invariant for (63), and assume that
set-valued Lie derivative in iii) and iii)’. Note that, if sup LF f (y) ≤ 0 for all y ∈ S. Then, every solution
the differential inclusion (63) has a unique solution x : [0, ∞) → Rd of (63) starting at S converges to the
starting from every initial condition, then the notions largest weakly invariant set M contained in
of strong and weak stability coincide, and it is suffi-
cient to satisfy the simpler requirements i)’ and iii)’ for S ∩ {y ∈ Rd : 0 ∈ LF f (y)}.
weak stability.
Similarly to the case of a continuous differential equa- Moreover, if the set M consists of a finite number of points,
tion, global asymptotic stability can be established by then the limit of each solution starting in S exists and is an
requiring the Lyapunov function f to be continuous and element of M.
radially unbounded. The equivalence between global Next, we apply Theorem 4 to gradient differential
strong asymptotic stability and the existence of infinitely inclusions.
differentiable Lyapunov functions is discussed in [62].
This type of global result is commonly invoked when Stability of Gradient Differential
dealing with the stabilization of control systems by refer- Inclusions for Convex Functions
ring to control Lyapunov functions [25] or Lyapunov Consider the gradient differential inclusion (62) associat-
pairs [24]. Two lower semicontinuous functions ed with the convex function f : Rd → R. We study here

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 69

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 Filippov solutions are used in problems involving electrical circuits
with switches, relay control, friction, and sliding.

the asymptotic behavior of the solutions of (62). From the Theorem 5

section “Gradient Differential Inclusion of a Convex Let X : Rd × Rm → Rd be continuous and X(0, 0) = 0. If
Function,” we know that solutions exist and are unique. there exists a continuous stabilizer of the control system
Consequently, the notions of weakly nonincreasing and (77), then there exists a neighborhood of the origin in
strongly nonincreasing function coincide. Therefore, it Rd × Rm whose image by X is a neighborhood of the ori-
suffices to show that f is weakly nonincreasing on Rd for gin in Rd .
the gradient differential inclusion (62). In particular, Theorem 5 implies that control systems of
For all ζ ∈ ∂P f (x), there exists v = −ζ ∈ −∂P f (x) such the form
that ζ T v = −ζ 22 ≤ 0. Hence,
ẋ = u1 X1 (x) + · · · + um Xm (x), (78)
L−∂P f f (x) ≤ 0 for all x ∈ Rd.
with m < n and Xi : Rd → Rd , i ∈ {1, . . . , m}, continuous
Proposition 13-i) now guarantees that f is weakly nonin- with rank(X1 (0), . . . , Xm (0)) = m, cannot be stabilized by a
creasing on Rd . Since the solutions of the gradient differen- continuous feedback controller.
tial inclusion (62) are unique, t → f (x(t)) is monotonically
nonincreasing for all solutions t → x(t). Example 21: Nonholonomic Integrator
The application of the Lyapunov stability theorem and Consider the nonholonomic integrator [64]
the invariance principle in theorems 3 and 4, respectively,
gives rise to the following nonsmooth counterpart of the ẋ1 = u1 , (79)
classical smooth result for a gradient flow [56]. ẋ2 = u2 , (80)
ẋ3 = x1 u2 − x2 u1 , (81)
Proposition 14
Let f : Rd → R be convex. Then, the strict minimizers of f
are strongly stable equilibria of the gradient differential where (u1 , u2 ) ∈ R2 . The nonholonomic integrator is a con-
inclusion (62) associated with f . Furthermore, if the level trol system of the form (78), with m = 2, X1 (x1 , x2 , x3 ) =
sets of f are bounded, then the solutions of the gradient (1, 0, −x2 ), and X2 (x1 , x2 , x3 ) = (0, 1, x1 ). The nonholonomic
differential inclusion asymptotically converge to the set of integrator is controllable, that is, for each pair of states
minimizers of f . x, y ∈ R3 , there exists an input t → u(t) such that the corre-
Convergence rate estimates of (62) can be found in [63]. sponding solution of (79)–(81) starting at x reaches y. How-
ever, the nonholonomic integrator does not satisfy the
Stabilization of Control Systems condition in Theorem 5 and thus is not continuously stabi-
Consider the control system on Rd given by lizable. Indeed, no point of the form (0, 0, ε), where ε = 0,
belongs to the image of the map X : R3 × R2 → R3 ,
ẋ = X(x, u), (77) X(x1 , x2 , x3 , u1 , u2 ) = (u1 , u2 , x1 u2 − x2 u1 ). 
The condition in Theorem 5 is necessary but not suffi-
where X : Rd × Rm → Rd . The system (77) is locally cient. There exist control systems that satisfy the condition
(respectively, globally) continuously stabilizable if there and still cannot be stabilized by means of a continuous sta-
exists a continuous map k : Rd → Rm such that the closed- bilizer. The cart on a circle in Example 20 is one of these
loop system systems. The map ((x1 , x2 ), u) → X(x1 , x2 , u) satisfies the
necessary condition in Theorem 5 but cannot be stabilized
ẋ = X(x, k(x)) with a continuous k : R2 → R; see [25].
Another obstruction to the existence of continuous
is locally (respectively, globally) asymptotically stable stabilizing controllers is given by Milnor’s theorem [12],
at the origin. The following result by Brockett [13] which states that the domain of attraction of an asymp-
(see also [12] and [14]) states that a large class of con- totically stable equilibrium of a locally Lipschitz vector
trol systems are not stabilizable by a continuous feed- field must be diffeomorphic to Euclidean space. Since
back controller. environments with obstacles are not diffeomorphic to

70 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 Caratheodory solutions are employed for time-dependent vector fields that
depend discontinuously on time such as dynamical systems involving impulses
as well as control systems with discontinuous open-loop inputs.

Euclidean space, one can use Milnor’s theorem to justi- Theorem 7

fy [25] the nonexistence of continuous globally stabiliz- Let X : Rd × Rm → Rd be continuous and X(0, 0) = 0 .
ing controllers in environments with obstacles. Then, the control system (77) is globally asymptotically
The obstruction to the existence of continuous stabi- controllable if and only if it admits a measurable, locally
lizers has motivated the search for time-varying and bounded stabilizer in the sample-and-hold sense.
discontinuous feedback stabilizers. Regarding the lat- The “only if” implication is clear. The converse implica-
ter, given a discontinuous k : Rd → Rm , the immediate tion is proved by explicit construction of the stabilizer and
question that arises is how to understand the solutions is based on the fact that the control system (77) is globally
of the resulting discontinuous dynamical system asymptotically controllable if and only if it admits a con-
ẋ = X(x, k(x)). From the previous discussion, we know tinuous Lyapunov pair; see [65]. Using the continuous
that Caratheodory solutions are not a good candidate, Lyapunov function provided by this characterization, the
since in many situations they fail to exist. The following discontinuous feedback for the control system (77) can be
result [21], [22] shows that Filippov solutions are also constructed explicitly [25], [41].
not a good candidate. The existence of a Lyapunov pair in the sense of gener-
alized gradients (that is, when the set-valued Lie deriva-
Theorem 6 tive involving the generalized gradient is used instead of
Let X : Rd × Rm → Rd be continuous and X(0, 0) = 0 . the lower set-valued Lie derivative involving the proxi-
Assume that, for each U ⊆ Rm and each x ∈ Rd , mal subdifferential), however, turns out to be equivalent
X(x, co U) = co X(x, U) holds. With solutions understood to the existence of a stabilizing feedback in the sense of
in the Filippov sense, if there exists a measurable, locally Filippov; see [66].
bounded stabilizer of the control system (77), then there
exists a neighborhood of the origin in Rd × Rm whose Example 20 Revisited: Cart Stabilization
image by X is a neighborhood of the origin in Rd . by Discontinuous Feedback
In particular, control systems of the form (78) cannot As an illustration, consider Example 20. We have
be stabilized by means of a discontinuous feedback if already shown that (0, 0) is a globally weakly asymptoti-
solutions are understood in the Filippov sense. This cally stable equilibrium of the differential inclusion (74)
impossibility result, however, can be overcome if solu- associated with the control system. Therefore, the con-
tions are understood in the sample-and-hold sense, as trol system is globally asymptotically controllable and
shown in [41]. Let us briefly discuss this result in light of can be stabilized in the sample-and-hold sense by means
the above exposition. For more details, see [24]–[26]. of a discontinuous feedback. The stabilizing feedback
Consider the differential inclusion (35) associated with that results from the proof of Theorem 7 can be
the control system (77). The system (77) is (open loop) described as follows; see [25] and [67]. If placed to the
globally asymptotically controllable (to the origin) if zero is left of the x2 axis, move in the direction of the control
a Lyapunov weakly stable equilibrium of the differential vector field (x1 , x2 ) → (x21 − x22 , 2x1 x2 ). If placed to the
inclusion (35), and every point y ∈ Rd has the property right of the x2 axis, move in the opposite direction of the
that there exists a solution of (35) satisfying x(0) = y and contol vector field. Finally, on the x2 -axis, choose an arbi-
lim t→∞ x(t) = 0. On the other hand, a map k : Rd → Rm trary direction. The stabilizing nature of this feedback
stabilizes the system (77) in the sample-and-hold sense if, for can be graphically checked in Figure 14(a) and (b). 
all x0 ∈ Rd and all ε ∈ (0, ∞), there exist δ, T ∈ (0, ∞) Remarkably, for a system that is affine in the con-
such that, for all partitions π of [0, t1 ] with diam(π) < δ, trols, it is possible to show [68] that there exists a stabi-
the corresponding π -solution t → x(t) of (77) starting at lizing feedback controller whose discontinuities form a
x0 satisfies x(t)2 ≤ ε for all t ≥ T. set of measure zero, and, moreover, the discontinuity
The following result [41] states that both notions, global set is repulsive for the solutions of the closed-loop sys-
asymptotic controllability and the existence of a feedback tem. In particular, this fact means that, for the closed-
stabilizer in the sample-and-hold sense, are equivalent. loop system, the solutions can be understood in the

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 71

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 We state analysis results to characterize the stability and asymptotic convergence
properties of the solutions of discontinuous dynamical systems.

Caratheodory sense. This situation exactly corresponds ACKNOWLEDGMENTS

to the situation in Example 20. This research was supported by NSF CAREER Award
ECS-0546871. The author wishes to thank Francesco Bullo
CONCLUSIONS and Anurag Ganguli for countless hours of fun with
This article has presented an introductory tutorial on dis- Filippov solutions and Bernard Brogliato, Rafal Goebel,
continuous dynamical systems. Various examples illus- Rishi Graham, and three anonymous reviewers for numer-
trate the pertinence of the continuity and Lipschitzness ous comments that improved the presentation.
properties that guarantee the existence and uniqueness of
classical solutions to ordinary differential equations. The REFERENCES
[1] A.A. Agrachev and Y. Sachkov, Control Theory from the Geometric View-
lack of these properties in examples drawn from various point. New York: Springer-Verlag, 2004.
disciplines motivates the need for more general notions [2] S. Bennett, A History of Control Engineering 1930–1955. London, U.K.: Inst.
than the classical one. This observation is the starting Elect. Eng., 1993.
[3] B. Brogliato, Nonsmooth Mechanics: Models, Dynamics, and Control, 2nd ed.
point into the three main themes of our discussion. First, New York: Springer-Verlag, 1999.
we introduced notions of solution for discontinuous sys- [4] F. Pfeiffer and C. Glocker, Multibody Dynamics with Unilateral Contacts.
tems. Second, we reviewed the available tools from non- New York: Wiley, 1996.
[5] G.A.S. Pereira, M.F.M. Campos, and V. Kumar, “Decentralized algo-
smooth analysis to study the gradient information of rithms for multi-robot manipulation via caging,” Int. J. Robotics Res., vol. 23,
candidate Lyapunov functions. And, third, we presented no. 7–8, pp. 783–795, 2004.
nonsmooth stability tools to characterize the asymptotic [6] V.I. Utkin, Sliding Modes in Control and Optimization. New York: Springer-
Verlag, 1992.
behavior of solutions. [7] C. Edwards and S.K. Spurgeon, Sliding Mode Control: Theory and Applica-
The physical significance of the solution notions dis- tions. New York: Taylor & Francis, 1998.
cussed in this article depend on the specific setting. [8] R.C. Arkin, Behavior-Based Robotics. Cambridge, MA: MIT Press, 1998.
[9] F.H. Clarke, Optimization and Nonsmooth Analysis (Canadian Math. Soc.
Caratheodory solutions are employed for time-depen- Series of Monographs and Advanced Texts). New York: Wiley, 1983.
dent vector fields that depend discontinuously on time [10] A. Bhaya and E. Kaszkurewicz, Control Perspectives on Numerical Algo-
such as dynamical systems involving impulses as well as rithms and Matrix Problems. Philadelphia, PA: SIAM, 2006.
[11] M.M. Polycarpou and P.A. Ioannou, “On the existence and uniqueness
control systems with discontinuous open-loop inputs. of solutions in adaptive control systems,” IEEE Trans. Automat. Contr., vol.
Filippov solutions are used in problems involving elec- 38, no. 3, pp. 474–479, 1993.
trical circuits with switches, relay control, friction, and [12] E.D. Sontag, Mathematical Control Theory: Deterministic Finite Dimensional
Systems, 2nd ed. New York: Springer-Verlag, 1998.
sliding. This observation is also valid for solution [13] R.W. Brockett, “Asymptotic stability and feedback stabilization,” in Geo-
notions similar to Filippov’s, such as Krasovskii and Sen- metric Control Theory, R.W. Brockett, R.S. Millman, and H.J. Sussmann, Eds.
tis solutions; see “Additional Solution Notions for Dis- Cambridge, MA: Birkhäuser, 1983, pp. 181–191.
[14] A. Bacciotti and L. Rosier, Liapunov Functions and Stability in Control The-
continuous Systems.” The notion of a π -solution for ory, 2nd ed. New York: Springer-Verlag, 2005.
control systems has the physical interpretation of itera- [15] D.E. Stewart, “Rigid-body dynamics with friction and impact,” SIAM
tively evaluating the input at the current state and hold- Rev., vol. 42, no. 1, pp. 3–39, 2000.
[16] F. Ceragioli, “Discontinuous ordinary differential equations and stabi-
ing it steady for some time while the closed-loop lization,” Ph.D. dissertation, Univ. Firenze, Italy, 1999 [Online]. Avail:
dynamical system evolves. As illustrated above, this http://calvino.polito.it/~ceragioli
solution notion plays a pivotal role in the stabilization of [17] H. Goldstein, Classical Mechanics, 2nd ed. Reading, MA: Addison-Wesley,
1980.
asymptotically controllable systems. [18] A.F. Filippov, Differential Equations with Discontinuous Righthand Sides.
There are numerous important issues that are not treat- Norwell, MA: Kluwer, 1988.
ed here; “Additional Topics on Discontinuous Systems and [19] J.P. Aubin and A. Cellina, Differential Inclusions. New York: Springer-
Verlag, 1994.
Differential Inclusions” lists some of them. The topic of [20] G.V. Smirnov, Introduction to the Theory of Differential Inclusions. Provi-
discontinuous dynamical systems is vast, and our focus on dence, RI: American Mathematical Society, 2001.
the above-mentioned themes is aimed at providing a [21] E.P. Ryan, “On Brockett’s condition for smooth stabilizability and its
necessity in a context of nonsmooth feedback,” SIAM J. Control Optim., vol.
coherent exposition. We hope that this tutorial serves as a 32, no. 6, pp. 1597–1604, 1994.
guided motivation for the reader to further explore the [22] J.M. Coron and L. Rosier, “A relation between continuous time-varying
exciting topic of discontinuous systems. The list of refer- and discontinuous feedback stabilization,” J. Mathematics Syst., Estimation
Contr., vol. 4, no. 1, pp. 67–84, 1994.
ences of this manuscript provides a good starting point to [23] N.N. Krasovskii and A.I. Subbotin, Game-Theoretical Control Problems.
undertake this endeavor. New York: Springer-Verlag, 1988.

72 IEEE CONTROL SYSTEMS MAGAZINE » JUNE 2008

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.
 [24] F.H. Clarke, Y. Ledyaev, R.J. Stern, and P.R. Wolenski, Nonsmooth Analy- [55] M. Valadier, “Entraînement unilatéral, lignes de descente, fonctions
sis and Control Theory. New York: Springer-Verlag, 1998. lipschitziennes non pathologiques,” C.R. Acad. Sci. Paris Sér. I Math, vol. 8,
[25] E.D. Sontag, “Stability and stabilization: Discontinuities and the effect of pp. 241–244, 1989.
disturbances,” in Nonlinear Analysis, Differential Equations, and Control, F.H. [56] W.M. Hirsch and S. Smale, Differential Equations, Dynamical Systems and
Clarke and R.J. Stern, Eds. Norwell, MA: Kluwer, 1999, pp. 551–598. Linear Algebra. New York: Academic, 1974.
[26] F.H. Clarke, “Lyapunov functions and feedback in nonlinear control,” in [57] J. Cortés, “Finite-time convergent gradient flows with applications to
Optimal Control, Stabilization and Nonsmooth Analysis, M.S. de Queiroz, M. network consensus,” Automatica, vol. 42, no. 11, pp. 1993–2000, 2006.
Malisoff, and P. Wolenski, Eds. New York: Springer-Verlag, 2004, pp. 267–282. [58] S. Adly, H. Attouch, and A. Cabot, “Finite time stabilization of nonlinear
[27] E.A. Coddington and N. Levinson, Theory of Ordinary Differential Equa- oscillators subject to dry friction,” in Progresses in Nonsmooth Mechanics and
tions. New York: McGraw-Hill, 1955. Analysis, P. Alart, O. Maisonneuve, and R.T. Rockafellar, Eds. Norwell, MA:
[28] R.P. Agarwal and V. Lakshmikantham, Uniqueness and Nonuniqueness Kluwer, 2006, pp. 289–304.
Criteria for Ordinary Differential Equations. Singapore: World Scientific, 1993. [59] A. Cabot, “Stabilization of oscillators subject to dry friction: Finite time
[29] A. Pucci, “Traiettorie di campi di vettori discontinui,” Rend. Ist. Mat. convergence versus exponential decay results,” Trans. Amer. Mathemat. Soc.,
Univ. Trieste, vol. 8, pp. 84–93, 1976. vol. 360, pp. 103–121, 2008.
[30] F. Ancona and A. Bressan, “Patchy vector fields and asymptotic sta- [60] R. Olfati-Saber and R.M. Murray, “Consensus problems in networks of
bilization,” ESAIM. Control, Optimisation Calculus of Variations, vol. 4, pp. agents with switching topology and time-delays,” IEEE Trans. Automat.
419–444, 1999. Contr., vol. 49, no. 9, pp. 1520–1533, 2004.
[31] A. Bressan, “Unique solutions for a class of discontinuous differential [61] Z. Artstein, “Stabilization with relaxed controls,” Nonlinear Analysis,
equations,” Proc. Amer. Math. Soc., vol. 104, no. 3, 1988, pp. 772–778. vol. 7, pp. 1163–1173, 1983.
[32] A. Bressan and W. Shen, “Uniqueness for discontinuous o.d.e. and con- [62] F.H. Clarke, Y.S. Ledyaev, and R.J. Stern, “Asymptotic stability and
servation laws,” Nonlinear Anal., vol. 34, pp. 637–652, 1998. smooth Lyapunov functions,” J. Differ. Equ., vol. 149, pp. 69–114, 1998.
[33] J.-I. Imura and A.J. van der Schaft, “Characterization of well-posedness [63] O. Güler, “Convergence rate estimates for the gradient differential inclu-
of piecewise linear systems,” IEEE Trans. Automat. Contr., vol. 45, no. 9, pp. sion,” Optim. Methods Software, vol. 20, no. 6, pp. 729–735, 2005.
1600–1619, 2000. [64] R.W. Brockett, “Control theory and singular Riemannian geometry,” in
[34] N.N. Krasovskii, Stability of Motion. Applications of Lyapunov's Second New Directions in Applied Mathematics, P. Hilton and G. Young, Eds. New
Method to Differential Systems and Equations with Delay. Stanford, CA: Stan- York: Springer-Verlag, 1982, pp. 11–27.
ford Univ. Press, 1963. (Transl. from Russian by J.L. Brenner). [65] E.D. Sontag, “A Lyapunov-like characterization of asymptotic controlla-
[35] R. Sentis, “Equations differentielles à second membre mesurable,” Boll. bility,” SIAM J. Control Optim., vol. 21, pp. 462–471, 1983.
Unione Matematica Italiana, vol. 5, no. 15-B, pp. 724–742, 1978. [66] L. Rifford, “On the existence of nonsmooth control-Lyapunov functions
[36] B. Paden and S.S. Sastry, “A calculus for computing Filippov’s differ- in the sense of generalized gradients,” ESAIM. Control, Optimiz. Calculus
ential inclusion with application to the variable structure control of robot Variations, vol. 6, pp. 593–611, 2001.
manipulators,” IEEE Trans. Circuits Syst., vol. 34, no. 1, pp. 73–82, 1987. [67] G.A. Lafferriere and E.D. Sontag, “Remarks on control Lyapunov func-
[37] D. Liberzon, Switching in Systems and Control. Cambridge, MA: tions for discontinuous stabilizing feedback,” in Proc. IEEE Conf. Decision and
Birkhäuser, 2003. Control, San Antonio, TX, 1993, pp. 306–308.
[38] M. di Bernardo, C.J. Budd, A.R. Champneys, and P. Kowalczyk, Piecewise- [68] L. Rifford, “Semiconcave control-Lyapunov functions and stabilizing
smooth Dynamical Systems: Theory and Applications. New York: Springer-Verlag, 2007. feedbacks,” SIAM J. Control Optimiz., vol. 41, no. 3, pp. 659–681, 2002.
[39] A.Y. Pogromsky, W.P.M.H. Heemels, and H. Nijmeijer, “On solution
concepts and well-posedness of linear relay systems,” Automatica, vol. 39,
no. 12, pp. 2139–2147, 2003. AUTHOR INFORMATION
[40] R.M. Colombo and A. Marson, “Hölder continuous o.d.e. related to traf- Jorge Cortés ([email protected]) received the Licenciatura
fic flow,” Proc. R. Soc. Edinb.. A., Math, vol. 133A, pp. 759–772, 2003.
[41] F.H. Clarke, Y.S. Ledyaev, E.D. Sontag, and A.I. Subbotin, “Asymptotic degree in mathematics from the Universidad de
controllability implies feedback stabilization,” IEEE Trans. Automat. Contr., Zaragoza, Spain, in 1997, and his Ph.D. degree in engi-
vol. 42, no. 10, pp. 1394–1407, 1997. neering mathematics from the Universidad Carlos III de
[42] R.T. Rockafellar and R.J.B. Wets, Variational Analysis. New York:
Springer-Verlag, 1998.
Madrid, Spain, in 2001. He held postdoctoral positions at
[43] J.P. Aubin and H. Frankowska, Set-Valued Analysis. Cambridge, MA: the Systems, Signals, and Control Department of the
Birkhäuser, 1990. University of Twente and at the Coordinated Science
[44] J. Cortés and F. Bullo, “Coordination and geometric optimization via
distributed dynamical systems,” SIAM J. Control Optim., vol. 44, no. 5, pp.
Laboratory of the University of Illinois at Urbana-Cham-
1543–1574, 2005. paign. From 2004 to 2007, he was an assistant professor
[45] J.M. Borwein and A.S. Lewis, Convex Analysis and Nonlinear Optimization: with the Department of Applied Mathematics and Statis-
Theory and Examples. New York: Springer-Verlag, 2000.
[46] J.-B. Hiriart-Urruty and C. Lemaréchal, Fundamentals of Convex Analysis,
tics, University of California, Santa Cruz. He is currently
2nd ed. New York: Springer-Verlag, 2004. an assistant professor in the Department of Mechanical
[47] R.T. Rockafellar, Convex Analysis. Princeton, NJ: Princeton Univ. Press, and Aerospace Engineering, University of California,
1970 (reprint 1997).
San Diego. His research interests focus on mathematical
[48] F.H. Clarke, Y.S. Ledyaev, and P.R. Wolenski, “Proximal analysis and min-
imization principles,” J. Math. Anal. Appl., vol. 196, no. 2, pp. 722–735, 1995. control theory, distributed motion coordination for
[49] J.P. Aubin and I. Ekeland, Applied Nonlinear Analysis. New York: Wiley, 1984. groups of autonomous agents, and geometric mechanics
[50] H.K. Khalil, Nonlinear Systems, 3rd ed. Englewood Cliffs, NJ: Prentice-
and geometric integration. He is the author of Geometric,
Hall, 2002.
[51] D. Shevitz and B. Paden, “Lyapunov stability theory of nonsmooth sys- Control and Numerical Aspects of Nonholonomic Systems
tems,” IEEE Trans. Automat. Contr., vol. 39, no. 9, pp. 1910–1914, 1994. (Springer Verlag, 2002) and the recipient of the 2006
[52] E.P. Ryan, “An integral invariance principle for differential inclusions Spanish Society of Applied Mathematics Young
with applications in adaptive control,” SIAM J. Control Optim., vol. 36, no. 3,
pp. 960–980, 1998. Researcher Prize. He is currently an associate editor for
[53] A. Bacciotti and F. Ceragioli, “Stability and stabilization of discontinu- the European Journal of Control. He can be contacted at the
ous systems and nonsmooth Lyapunov functions,” ESAIM. Control, Optim. Department of Mechanical and Aerospace Engineering,
Calculus Variations, vol. 4, pp. 361–376, 1999.
[54] A. Bacciotti and F. Ceragioli, “Nonpathological Lyapunov functions and dis- University of California at San Diego, 9500 Gilman Dr,
continuous Caratheodory systems,” Automatica, vol. 42, no. 3, pp. 453–458, 2006. La Jolla, CA 92093 USA.

JUNE 2008 « IEEE CONTROL SYSTEMS MAGAZINE 73

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on April 21,2025 at 03:11:55 UTC from IEEE Xplore. Restrictions apply.