Professional Documents
Culture Documents
Given its utility, the LaSalle-Yoshizawa Theorem has been applied, for corollaries presented in this work can also be extended to Krasovskii
example, in adaptive control and in deriving stability from passivity solutions (see [3], for example). In the case of Krasovskii solutions,
properties such as feedback passivation and backstepping designs of one would get stronger conclusions (i.e., conclusions for a potentially
nonlinear systems [21]. Available proofs for the LaSalle-Yoshizawa larger set of solutions) at the cost of slightly stronger assumptions (e.g.,
Theorem exploit Barbalat’s Lemma, which is often invoked to show local boundedness rather than essentially local boundedness).
asymptotic convergence for general classes of nonlinear systems [33]. To facilitate the main results, four definitions are provided.
In general, adapting the LaSalle-Yoshizawa Theorem to systems where Definition 2. (Directional Derivative): [54] Given a function
the right-hand side is not locally Lipschitz has only recently been ex- , the right directional derivative of at in the
plored. The result in [49] presents three invariance-like semistability direction of is defined as
theorems that utilize similar arguments to the LaSalle-Yoshizawa The-
orem under the assumption that the system dynamics are uniformly
bounded. Alternatively, using Barbalat’s Lemma and the observation
that an absolutely continuous function that has a uniformly locally in- Additionally, the generalized directional derivative of at in the di-
tegrable derivative is uniformly continuous, the result in [50] proves rection of is defined as
asymptotic convergence of an output function for nonlinear systems
with disturbances. The result in [50] is developed for differential
equations with a continuous right-hand side, but [50, Facts 1–4] pro-
vide insights into the application of Barbalat’s Lemma to discontinuous Definition 3. (Regular Function): [34] A function
systems. is said to be regular at if for all , the right direc-
In this technical note, we present two corollaries to the LaSalle- tional derivative of at in the direction of exists and
Yoshizawa Theorem for nonautonomous systems with right-hand side 1
discontinuities that are essentially locally bounded, uniformly in , uti- Definition 4. (Clarke’s Generalized Gradient): [34] For a function
lizing Filippov solutions and Lipschitz continuous and regular Lya- that is locally Lipschitz in , define the
punov-like functions whose time derivatives can be upper bounded by generalized gradient of at by
negative semi-definite functions. Applicability of one of the corollaries
is illustrated for an example problem.
II. PRELIMINARIES where is the set of measure zero where the gradient of is not
Consider the system defined.
Definition 5. (Locally Bounded, Uniformly in ): Let
(1) . The map is locally bounded, uniformly
in , if for each compact set , there exists such that
where denotes the state vector, , .
is Lebesgue measurable and essentially locally bounded, uniformly in The following lemma provides a method for computing the time
and is an open and connected set. Existence and uniqueness of derivative of a regular function using Clarke’s generalized gradient
the continuous solution are provided under the condition that the [34] and , from (2), along the solution trajectories of (1).
function is Lipschitz continuous. However, if contains a discon- Lemma 1. (Chain Rule): [6], [56] Let be a Filippov solution of
tinuity at any point in , then a solution to (1) may not exist in the (1) and be a locally Lipschitz, regular function.
classical sense. Thus, it is necessary to redefine the concept of a solu- Then is absolutely continuous, exists almost
tion. Utilizing differential inclusions, the value of a generalized solu- everywhere (a.e.), i.e., for almost all , and
tion (e.g., Filippov [51] or Krasovskii [52] solutions) at a certain point
, where2
can be found by interpreting the behavior of its derivative at nearby
points. Generalized solutions will be close to the trajectories of the ac-
tual system since they are a limit of solutions of ordinary differential
equations with a continuous right-hand side [13]. While there exists a
Filippov solution for any arbitrary initial condition , the so- Remark 2: Throughout the subsequent discussion, for brevity of no-
lution is generally not unique [5], [51] . tation, let a.e. refer to almost all .
Definition 1. (Filippov Solution): [51] A function
is called a solution of (1) on the interval if is absolutely
III. MAIN RESULT
continuous and for almost all
For the system described in (1) with a continuous right-hand side,
existing Lyapunov theory can be used to examine the stability of
the closed-loop system using continuous techniques such as those
where is an upper semi-continuous, nonempty, compact described in [57]. However, these theorems must be altered for the
and convex valued map on , defined as set-valued map for systems with right-hand sides which
(2) are not Lipschitz continuous [6], [13], [14]. Lyapunov analysis for
nonsmooth systems is analogous to the analysis used for continuous
systems. The differences are that differential equations are replaced
denotes the intersection over sets of Lebesgue measure zero, with inclusions, gradients are replaced with generalized gradients,
denotes convex closure, and 1Note that any continuous function is regular and the sum of regular func-
. tions is regular [55].
Remark 1: One can also formulate the solutions of (1) in other ways 2Equivalently, almost everywhere, for all and
[53]; for instance, using Krasovskii’s definition of solutions [52]. The some .
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 58, NO. 9, SEPTEMBER 2013 2335
and points are replaced with sets throughout the analysis. The fol- solution starting in stays in , and con-
lowing presentation and subsequent proofs demonstrate how the sequently in , for all future time. Hence, the
LaSalle-Yoshizawa Theorem can be adapted for such systems. Filippov solution is bounded such that , .
The following auxiliary lemma from [56] and Barbalat’s Lemma are From Lemma 2, is also bounded such that
provided to facilitate the proofs of the nonsmooth LaSalle-Yoshizawa . Since is Lebesgue measurable from (4),
Corollaries.
Lemma 2: [56] Let be any Filippov solution to the system in
(1) and be a locally Lipschitz, regular function.
If , then .
Proof: For the sake of contradiction, let there exist some (6)
such that . Then
Therefore, is bounded . Existence of
is guaranteed since the left-hand side of (6)
is monotonically nondecreasing (based on the definition of )
and bounded above. Since is locally absolutely continuous and
is essentially locally bounded, uniformly in , is uniformly
It follows that on a set of positive measure, which continuous. Because is continuous in , and is on the compact
contradicts that , a.e. set , is uniformly continuous in on . Therefore,
Lemma 3. (Barbalat’s Lemma): [57] Let be a uni- by Lemma 3, .
formly continuous function. Suppose that exists and Remark 3: From Def. 1, is an upper semi-continuous,
is finite. Then nonempty, compact and convex valued map. While existence of a Fil-
ippov solution for any arbitrary initial condition is provided
by the definition, generally speaking, the solution is non-unique [5],
[51].
Based on Lemmas 2 and 3, nonsmooth corollaries to the LaSalle-
Note that Corollary 1 establishes (5) for a specific . Under the
Yoshizawa Theorem (c.f., [33, Theorem 8.4] and [46, Theorem A.8])
are provided in Corollary 1 and 2. stronger condition that3 , it is possible to
Corollary 1: For the system given in (1), let be an open and show that (5) holds for all Filippov solutions of (1). The next corollary
connected set containing and suppose is Lebesgue measurable is presented to illustrate this point.
and is essentially locally bounded, uniformly in . Let Corollary 2: For the system given in (1), let be an open and
be locally Lipschitz and regular such that connected set containing and suppose is Lebesgue measurable
and is essentially locally bounded, uniformly in . Let
(3) be locally Lipschitz and regular such that
(4) (7)
is used in the proof for Corollary 1); however, the use of Barbalat’s After using (13), the expression in (16) can be written as
Lemma would require the identification of the region of attraction for
which convergence holds and does not provide boundedness of the (17)
trajectories. For illustrative purposes, the following design example
targets the regulation of a first order nonlinear system. where such that if ,
Consider a first-order nonlinear differential equation given by [ 1, 1] if , and if .
Remark 4: One could also consider the discontinuous function in-
(9) stead of the differential inclusion (i.e., the function can alterna-
tively be defined as ) using Caratheodory solutions; how-
where is an unknown, linear-parameteriz-
ever, this method would not be an indicator for what happens when
able, essentially locally bounded, uniformly in function that can be
measurement noise is present in the system. As described in results
expressed as where is a vector of un-
such as [58]–[60], Filippov and Krasovskii solutions for discontin-
known constant parameters and
uous differential equations are appropriate for capturing the possible
is the regression matrix for . In addition,
closed-loop system behavior in the presence of arbitrarily small mea-
is the control input, is the measurable system state, and
surement noise. By utilizing the set valued map in the anal-
is an essentially locally bounded disturbance
ysis, we account for the possibility that when the true state satisfies
that satisfies
, (of the measured state) falls within the set .
(10) Therefore, the presented analysis is more robust to measurement noise
than an analysis that depends on to be defined as a known
where is a positive constant, and is a positive, singleton.
globally invertible, state-dependent function. A regulation controller Substituting for the adaptive update law in (12), canceling terms and
for (9) can be designed as utilizing the bound for in (10), the expression in (16) can be upper
bounded as
(11)
(18)
where is an estimate of , are gain constants,
and is defined as The set in (17) reduces to the scalar inequality in (18) since in the case
. Based on the subsequent stability when is defined as a set, it is multiplied by , i.e., when
analysis, an adaptive update law can be defined as , . Regrouping similar terms, the expression
in (18) can be written as
(12)
(19)
where is a positive gain matrix. The closed-loop system is
given by
Provided and , the expression in (19) can be
(13) upper bounded as where
is a positive semi-definite function defined on the domain
where denotes the mismatch . In (13), it is . The inequalities in (15)
apparent that the RHS contains a discontinuity in , and the use of can be used to show that in ; hence, and
differential inclusions provides for existence of solutions. in . Since contains the constant unknown
Let be defined as and choose a posi- system parameters and in , the definition for
tive-definite, locally Lipschitz, regular candidate Lyapunov function as can be used to show that in . Given
that in , in . Since , ,
(14) and in , the control is bounded from (11) and the
adaption law in (12). The closed-loop dynamics in (10) and (13) can
The candidate Lyapunov function in (14) satisfies the following in-
be used to conclude that in ; hence, is uniformly
equalities:
continuous in .
(15) Choose such that denotes a closed ball,
and let denote the set defined as
where the continuous positive-definite functions
are defined as , and , where (20)
are known constants. Then,
and Invoking Corollary 2,
, thus, . The
region of attraction in (20) can be made arbitrarily large to include all
initial conditions (a semi-global type result) by increasing the gain .
Remark 5: For some systems (e.g., closed-loop error systems with
Since is in , 4 sliding mode control laws), it may be possible to show that Corollary
2 is more easily applied, as is the focus of the example in Section IV.
(16) However, in other cases, it may be difficult to satisfy the inequality
in (8). The usefulness of Corollary 1 is demonstrated in those cases
4For continuously differentiable Lyapunov candidate functions, the general- where it is difficult or impossible to show that the inequality in (8) can
ized gradient reduces to the standard gradient. However, this is not required by be satisfied, but it is possible to show that (4) can be satisfied for almost
the Corollary itself and only assists in evaluation. all time.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 58, NO. 9, SEPTEMBER 2013 2337
[44] J. Alvarez, Y. Orlov, and L. Acho, “An invariance principle for discon- Structural Analysis of Laplacian Spectral
tinuous dynamic systems with applications to a Coulomb friction os- Properties of Large-Scale Networks
cillator,” ASME J. Dynam. Syst., Meas., Control, vol. 74, pp. 190–198,
2000. Victor M. Preciado, Member, IEEE,
[45] Y. Orlov, “Extended invariance principle for nonautonomous switched
Ali Jadbabaie, Senior Member, IEEE, and
systems,” IEEE Trans. Autom. Control, vol. 48, no. 8, pp. 1448–1452,
Aug. 2003.
George C. Verghese, Fellow, IEEE
[46] M. Krstic, P. V. Kokotovic, and I. Kanellakopoulos, Nonlinear and
Adaptive Control Design. New York: Wiley, 1995.
[47] J. P. LaSalle, “Some extensions of Liapunov’s second method,” IRE Abstract—Using methods from algebraic graph theory and convex opti-
Trans. Circuit Theory, vol. CT-7, pp. 520–527, 1960. mization, we study the relationship between local structural features of a
[48] T. Yoshizawa, “Asymptotic behavior of solutions of a system of network and the eigenvalues of its Laplacian matrix. In particular, we pro-
differential equations,” Contrib. Differ. Equat., vol. 1, pp. 371–387, pose a series of semidefinite programs to find new bounds on the spectral
1963. radius and the spectral gap of the Laplacian matrix in terms of a collec-
[49] Q. Hui, W. M. Haddad, and S. P. Bhat, “Semistability for time-varying tion of local structural features of the network. Our analysis shows that the
discontinuous dynamical systems with application to agreement prob- Laplacian spectral radius is strongly constrained by local structural fea-
lems in switching networks,” in IEEE Proc. Conf. Decis. Control, 2008, tures. On the other hand, we illustrate how local structural features are
p. 29852990. usually insufficient to accurately estimate the Laplacian spectral gap. As a
[50] A. Teel, “Asymptotic convergence from Lp stability,” IEEE Trans. consequence, random graph models in which only local structural features
Autom. Control, vol. 44, no. 11, pp. 2169–2170, Nov. 1999. are prescribed are, in general, inadequate to faithfully model Laplacian
[51] A. F. Filippov, Differential Equations with Discontinuous Right-hand spectral properties of a network.
Sides. Norwell, MA: Kluwer, 1988. Index Terms—Laplacian matrix, spectral.
[52] N. N. Krasovskii, Stability of motion. Stanford, CA: Stanford Uni-
versity Press, 1963.
[53] O. Hájek, “Discontinuous differential equations,” J. Diff. Equat., vol.
I. INTRODUCTION
32, pp. 149–170, 1979.
[54] W. Kaplan, Advanced Calculus, 4th ed. Reading, MA: Ad- Understanding the relationship between the structure and dynamics
dison-Wesley, 1991, Reading. of a network is a central question in the field of network science [1].
[55] F. Clarke, Y. Ledyaev, R. Stern, and P. Wolenski, Nonsmooth Analysis Since the behavior of many networked dynamical processes is closely
and Control Theory, 178th ed. New York: Springer, 1998, Graduate
Texts in Mathematics.
related with the Laplacian eigenvalues (see [2], [3] and references
[56] B. Paden and S. Sastry, “A calculus for computing Filippov’s differen- therein), it is of relevance to study the relationship between structural
tial inclusion with application to the variable structure control of robot features of the network and its spectral properties.
manipulators,” IEEE Trans. Circuits Syst., vol. 34, no. 1, pp. 73–82, In this technical note, we study this relationship, focusing on the role
Jan. 1987. played by structural features that can be extracted from localized sam-
[57] H. K. Khalil, Nonlinear Systems, 2nd ed. Upper Saddle River, NJ: ples of the network structure. We show how structural information ex-
Prentice-Hall, 1996.
tracted from these local samples can be efficiently aggregated to infer
[58] H. Hermes, “Discontinuous vector fields and feedback control,” in Dif-
ferential Equations and Dynamical Systems. New York: Academic global properties of the Laplacian spectrum. In particular, we propose
Press, 1967. a graph-theoretical approach to relate structural features of a network
[59] J.-M. Coron and L. Rosier, “A relation between continuous with algebraic properties of its Laplacian matrix. Our analysis reveals
time-varying and discontinuous feedback stabilization,” J. Math. that there are certain spectral properties, such as the so-called spec-
Systems. Estim. Control, vol. 4, no. 1, pp. 67–84, 1994. tral moments, that can be efficiently computed from these structural
[60] R. Goebel, R. Sanfelice, and A. Teel, “Hybrid dynamical systems,”
features. Furthermore, applying a recent result by Lasserre [4], we pro-
IEEE Control Syst. Mag., vol. 29, no. 2, pp. 28–93, 2009.
pose a series of semidefinite programs to compute new bounds on the
Laplacian spectral radius and spectral gap from a truncated sequence
of spectral moments.
The technical note is organized as follows. In the next subsection,
we define terminology needed in our derivations. In Section II, we in-
troduce a graph-theoretical methodology to derive closed-form expres-
sions for the so-called Laplacian spectral moments in terms of struc-
tural features of the network. In Section III, we use semidefinite pro-
gramming to derive optimal bounds on the Laplacian spectral radius
Manuscript received May 23, 2012; revised November 01, 2012; accepted
April 22, 2013. Date of publication May 01, 2013; date of current version Au-
gust 15, 2013. This work was supported by ONR MURI “Next Generation Net-
work Science” and AFOSR “Topological and Geometric Tools for Analysis of
Complex Networks.” Recommended by Associate Editor F. Paganini.
V. M. Preciado and A. Jadbabaie are with the Department of Electrical
and Systems Engineering, University of Pennsylvania, Philadelphia, PA
19104 USA (e-mail: victormpreciado@gmail.com; preciado@seas.upenn.edu;
jadbabai@seas.upenn.edu).
G. C. Verghese is with the Department of Electrical Engineering and Com-
puter Science, Massachusetts Institute of Technology, Cambridge, MA 02139
USA (e-mail: verghese@mit.edu).
Color versions of one or more of the figures in this technical note are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TAC.2013.2261187