Lecture Notes in Applied

and Computational Mechanics

Volume 36

Series Editors
Prof. Dr.-Ing. Friedrich Pfeiffer
Prof. Dr.-Ing. Peter Wriggers
Lecture Notes in Applied and Computational Mechanics
Edited by F. Pfeiffer and P. Wriggers
Further volumes of this series found on our homepage: springer.com

Vol. 36: Leine, R.I.; van de Wouw, N. in Civil Engineering

Stability and Convergence of Mechanical Systems 400 p. 2005 [978-3-540-25567-3]
with Unilateral Constraints Vol. 22: Chang C.H.
250 p. 2008 [978-3-540-76974-3] Mechanics of Elastic Structures with Inclined Members:
Vol. 35: Acary, V.; Brogliato, B. Analysis of Vibration, Buckling and Bending of X-Braced
Numerical Methods for Nonsmooth Dynamical Systems: Frames and Conical Shells
Applications in Mechanics and Electronics 190 p. 2004 [978-3-540-24384-7]
xxx p. 2008 [978-3-540-75391-9] Vol. 21: Hinkelmann R.
Vol. 34: Flores, P.; Ambrósio, J.; Pimenta Claro, J.C.; Efficient Numerical Methods and Information-Processing
Lankarani Hamid M. Techniques for Modeling Hydro- and Environmental
Kinematics and Dynamics of Multibody Systems Systems
with Imperfect Joints: Models and Case Studies 305 p. 2005 [978-3-540-24146-1]
186 p. 2008 [978-3-540-74359-0] Vol. 20: Zohdi T.I., Wriggers P.
Vol. 33: Niesłony, A.; Macha, E. Introduction to Computational Micromechanics
Spectral Method in Multiaxial Random Fatigue 196 p. 2005 [978-3-540-22820-2]
146 p. 2007 [978-3-540-73822-0] Vol. 19: McCallen R., Browand F., Ross J. (Eds.)
Vol. 32: Bardzokas, D.I.; Filshtinsky, M.L.; Filshtinsky, L.A. (Eds.) The Aerodynamics of Heavy Vehicles:
Mathematical Methods Trucks, Buses, and Trains
in Electro-Magneto-Elasticity 567 p. 2004 [978-3-540-22088-6]
530 p. 2007 [978-3-540-71030-1] Vol. 18: Leine, R.I., Nijmeijer, H.
Vol. 31: Lehmann, L. (Ed.) Dynamics and Bifurcations
Wave Propagation in Infinite Domains of Non-Smooth Mechanical Systems
186 p. 2007 [978-3-540-71108-7] 236 p. 2004 [978-3-540-21987-3]
Vol. 30: Stupkiewicz, S. (Ed.) Vol. 17: Hurtado, J.E.
Micromechanics of Contact and Interphase Layers Structural Reliability: Statistical Learning Perspectives
206 p. 2006 [978-3-540-49716-5] 257 p. 2004 [978-3-540-21963-7]
Vol. 29: Schanz, M.; Steinbach, O. (Eds.) Vol. 16: Kienzler R., Altenbach H., Ott I. (Eds.)
Boundary Element Analysis Theories of Plates and Shells:
571 p. 2006 [978-3-540-47465-4] Critical Review and New Applications
Vol. 28: Helmig, R.; Mielke, A.; Wohlmuth, B.I. (Eds.) 238 p. 2004 [978-3-540-20997-3]
Multifield Problems in Solid and Fluid Mechanics Vol. 15: Dyszlewicz, J.
571 p. 2006 [978-3-540-34959-4] Micropolar Theory of Elasticity
Vol. 27: Wriggers P., Nackenhorst U. (Eds.) 356 p. 2004 [978-3-540-41835-1]
Analysis and Simulation of Contact Problems Vol. 14: Frémond M., Maceri F. (Eds.)
395 p. 2006 [978-3-540-31760-9] Novel Approaches in Civil Engineering
Vol. 26: Nowacki, J.P. 400 p. 2003 [978-3-540-41836-8]
Static and Dynamic Coupled Fields in Bodies Vol. 13: Kolymbas D. (Eds.)
with Piezoeffects or Polarization Gradient Advanced Mathematical and Computational
209 p. 2006 [978-3-540-31668-8] Geomechanics
Vol. 25: Chen C.-N. 315 p. 2003 [978-3-540-40547-4]
Discrete Element Analysis Methods Vol. 12: Wendland W., Efendiev M. (Eds.)
of Generic Differential Quadratures Analysis and Simulation of Multifield Problems
282 p. 2006 [978-3-540-28947-0] 381 p. 2003 [978-3-540-00696-1]
Vol. 24: Schenk, C.A., Schuëller. G Vol. 11: Hutter K., Kirchner N. (Eds.)
Uncertainty Assessment of Large Dynamic Response of Granular and Porous Materials
Finite Element Systems under Large and Catastrophic Deformations
165 p. 2006 [978-3-540-25343-3] 426 p. 2003 [978-3-540-00849-1]
Vol. 23: Frémond M., Maceri F. (Eds.) Vol. 10: Hutter K., Baaser H. (Eds.)
Mechanical Modelling and Computational Issues Deformation and Failure in Metallic Materials
409 p. 2003 [978-3-540-00848-4]
Stability and Convergence
of Mechanical Systems with
Unilateral Constraints

Remco I. Leine Nathan van de Wouw

•

With 62 Figures
Dr.ir. habil. Remco I. Leine Dr.ir. Nathan van de Wouw
Institute of Mechanical Systems Dynamics and Control Group
Department of Mechanical Department of Mechanical
and Process Engineering Engineering
ETH Zurich Eindhoven University of Technology
CH-8092 Zurich PO Box 513
Switzerland 5600 MB Eindhoven
[email protected] The Netherlands
[email protected]

ISBN 978-3-540-76974-3 e-ISBN 978-3-540-76975-0

ISSN 1613-7736

Library of Congress Control Number: 2007941933

© 2008 Springer-Verlag Berlin Heidelberg

This work is subject to copyright. All rights are reserved, whether the whole or part of the material
is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broad-
casting, reproduction on microfilm or in any other ways, and storage in data banks. Duplication of
this publication or parts thereof is permitted only under the provisions of the German Copyright
Law of September 9, 1965, in its current version, and permission for use must always be obtained
from Springer. Violations are liable for prosecution under the German Copyright Law.

The use of general descriptive names, registered names, trademarks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.

Cover design: WMX Design GmbH, Heidelberg

Printed on acid-free paper

9 8 7 8 6 5 4 3 2 1 0

springer.com
Preface

During the last two decades a new research field has emerged: the field of
non-smooth dynamical systems and in particular non-smooth mechanics. Me-
chanics, being one of the oldest natural sciences, has played a forerunner role
in formalising a theory to describe evolution problems with some degree of
non-smoothness. The classical modelling approach used in engineering sci-
ences and physics is to express all relations within a system by equalities, a
dogma which has been reinforced by the emergence of computers. A crucial
point in dealing with non-smoothness is to leave the equality dogma and to
allow ourselves to think and work with (variational) inequalities and inclu-
sions. The key merit of modern non-smooth mechanics is the development
of a mathematical framework, based on convex analysis, which effectively de-
scribes non-smooth or set-valued relations within a system. This mathematical
framework has primarily been used to gain a proper understanding of unilat-
eral behaviour, to reveal the inherent structure of non-smooth mechanical
systems, as well as to set up numerical integration methods for the simulation
of non-smooth (mechanical) systems. These numerical methods fill the need
to obtain quantitative information about the motion starting from a particu-
lar initial state. Instead of the endeavour to obtain one or more approximate
solutions, one may strive to obtain qualitative information about all solutions.
The latter is referred to as a qualitative theory for dynamical systems and has
been initiated by H. Poincaré around the end of the 19th century. Of central
importance in the qualitative theory is the problem of the stability of motion.
The work of A. M. Lyapunov has been seminal in the formulation of a stability
theory. The aim of this monograph is to free the existing Lyapunov stability
theory from the omnipresent equality dogma by exploiting the structure of
non-smooth (mechanical) systems, opening the way to a qualitative analysis
of non-smooth dynamical systems. Furthermore, this monograph deals with
the concept of convergence as has been developed in the Russian literature in
the 1960’s. The convergence property reflects a stability property on system
level and has recently been shown to be highly instrumental in solving many
control problems, such as tracking, synchronisation and observer design. Here,
VI Preface

we investigate the convergence property for a class of non-smooth dynamical

systems, namely monotone measure differential inclusions.
This text is written to address primarily researchers interested in stability
properties of non-smooth mechanical systems. However, the text has been
arranged such that it is also of use for those interested in general non-smooth
dynamical systems. To this end, the exposition of the results for mechanical
systems is kept in separate chapters. First, the theory is explained for general
non-smooth dynamical systems. Subsequently, these results are specialised for
mechanical systems. Mechanical systems are taken as examples throughout
the work.
This monograph is the result of about four years active research of the
authors in the field of stability theory and non-smooth dynamical systems.
Part of the monograph is based on the habilitation thesis of the first author
and on joint-research papers of the authors.

Zürich, September 2007 Remco I. Leine

Eindhoven, September 2007 Nathan van de Wouw

Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V

Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . XI

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Historical Notes on the Theory of Stability . . . . . . . . . . . . . . . . . 3
1.3 Non-smooth Dynamical Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Stability and Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5 Literature Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.6 Objective and Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.7 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2 Non-smooth Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1 Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 Functions and Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3 Generalised Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4 Set-valued Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.5 Definitions from Convex Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.6 Subderivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3 Measure and Integration Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3.1 Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2 The Lebesgue Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3 Signed Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.4 Real Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.5 Differential Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.6 The Differential Measure of a Bilinear Form . . . . . . . . . . . . . . . . 55
3.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
VIII Contents

4 Non-smooth Dynamical Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4.1 Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2 Differential Inclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.3 Measure Differential Inclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

5 Mechanical Systems with Set-valued Force Laws . . . . . . . . . . . 79

5.1 Non-smooth Potential Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.2 Contact Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.3 Force Laws for Frictional Unilateral Contact . . . . . . . . . . . . . . . . 84
5.3.1 Signorini’s Contact Law . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.3.2 Coulomb’s Friction Law . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.3.3 Coulomb-Contensou Friction Law . . . . . . . . . . . . . . . . . . . 90
5.3.4 Rolling Friction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.3.5 Impact Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.4 Measure Newton-Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

6 Lyapunov Stability Theory for Measure Differential

Inclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.1 Mathematical Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6.2 Invariant Sets and Limit Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.3 Definitions of Stability Properties for Autonomous Systems . . . 119
6.4 Definitions of Stability Properties of Solutions
Non-autonomous Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.4.1 Differential Inclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.4.2 Measure Differential Inclusions . . . . . . . . . . . . . . . . . . . . . . 123
6.5 Basic Lyapunov Theorems of Autonomous Systems . . . . . . . . . . 126
6.5.1 Lyapunov Stability of Equilibrium Points . . . . . . . . . . . . . 126
6.5.2 Lyapunov Stability of Equilibrium Sets . . . . . . . . . . . . . . . 137
6.5.3 Stability through Maximal Monotonicity . . . . . . . . . . . . . 141
6.6 LaSalle’s Invariance Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.7 Instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

7 Stability Properties in Mechanical Systems . . . . . . . . . . . . . . . . 153

7.1 Total Mechanical Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
7.2 Stability Results for Mechanical Systems . . . . . . . . . . . . . . . . . . . 162
7.3 Attractivity of Equilibrium Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.3.1 Systems with Frictional Unilateral Constraints . . . . . . . . 168
7.3.2 Systems with Frictional Bilateral Constraints . . . . . . . . . 172
7.4 Instability of Equilibrium Positions and Sets . . . . . . . . . . . . . . . . 174
7.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.5.1 Falling Block . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.5.2 Rocking Bar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
 Contents IX

7.5.3 Constrained Bar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

7.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190

8 Convergence Properties of Monotone Measure Differential

Inclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
8.1 Convergent Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
8.2 Convergence of Maximal Monotone Systems . . . . . . . . . . . . . . . . 193
8.2.1 Maximal Monotone Systems . . . . . . . . . . . . . . . . . . . . . . . . 193
8.2.2 Existence of a Compact Positively Invariant Set . . . . . . . 195
8.2.3 Conditions for Convergence . . . . . . . . . . . . . . . . . . . . . . . . . 197
8.3 Tracking Control for Lur’e Type Systems . . . . . . . . . . . . . . . . . . . 199
8.4 Illustrative Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
8.4.1 One-way Clutch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
8.4.2 Tracking Control with Set-valued Friction . . . . . . . . . . . . 205
8.4.3 Tracking Control Using an Impulsive Input . . . . . . . . . . . 209
8.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214

9 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217

Sources and Translations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Notation

R set of all real numbers

R R ∪ {∞}, extended set of all real numbers
N set of all natural numbers
:= definition
inf infinum
sup supremum
∀ for all
∃ there exists
x scalar
x column-vector in Rn
X matrix in Rn×m
I identity matrix
det(A) determinant of A
cond(A) condition of A
xT transpose of x
X = diag x diagonal matrix with Xii = xi , Xij = 0, j = i
xi i-th element of x
|x| absolute value of x
x Euclidian norm of x
xM norm of x with respect to M
ẋ(t) differentiation w.r.t. time of an abs. cont. function x(t)
or density of dx with respect to the differential measure dt
a⊥b a is orthogonal to b
Br the n-dimensional open ball with radius r centred at the origin
Br the n-dimensional closed ball with radius r centred at the origin
[a, b] the closed interval {x ∈ R | a ≤ x ≤ b}
{a, b} the set comprising the elements {a} and {b}
XII Notation

C0 class of continuous functions

Ck class of functions differentiable up to order k
∅ empty set
C or cl C closure of C
co C closed convex hull of C
int C interior of C
bdry C boundary of C
x∈C x is an element of C (inclusion)
f (x) single-valued function Rn → R
f (x) single-valued function Rn → Rn
F (x) set-valued function Rn → Rn
f  (x) classical derivative of f with respect to x
∇f (x) classical gradient of f with respect to x
∂f (x) generalized differential of f with respect to x
(or subdifferential if f is a convex function)
df (x)(v) subderivative of f at x in the direction v
graph(f ) graph of f (2.2)
graph(F) graph of a set-valued function F (2.20)
epi(f ) epigraph of f (2.3)
f ∗ (x∗ ) conjugate function
ΨC (x) indicator function of C at x
ΨC∗ (x) support function of C at x
NC (x) normal cone of C at x
K◦ polar cone to the cone K
proxC (x) proximal point from x to the convex set C
distC (x) distance from x to the convex set C
sign(x) sign function with sign(0) = 0
Sign(x) set-valued sign function
var(f , I) variation of f over the interval I (Definition 2.11)
f ∈ lbv(I, X) f : I → X is a function of locally bounded variation on I
A a σ-algebra (Definition 3.1)
μ friction coefficient
μ(A) measure of the set A (Definition 3.2)
μf (I) Lebesgue-Stieltjes measure of f on an interval I (3.2)
χA (x) characteristic function of the set A (3.4)
ν(A) signed measure of the set A (Definition 3.3)
νf (I) signed Lebesgue-Stieltjes measure of f on an interval I (3.14)
ν =f μ the signed measure ν is has a density f with respect to μ
νμ the signed measure ν is μ-continuous (Definition 3.4)
ν⊥ρ the signed measures ν and ρ are orthogonal (Definition 3.5)
 Notation XIII

υ real measure
df differential measure of f
dt (differential) Lebesgue measure
dη atomic measure
ϕ(t, t0 , x0 ) a solution curve x(t) with initial condition x(t0 ) = x0
S(dΓ , t0 , x0 ) set of solution curves of dx ∈ dΓ (t, x)
with initial condition x(t0 ) = x0
Ωc level set (6.11)
Lc level surface (6.12)
F ◦G the composition of the two maps F and G
x∗ equilibrium point of a dynamical system
q∗ equilibrium position of a mechanical system
E equilibrium set (Definition 6.8)
Eq the set of equilibrium positions
A admissible set of a measure differential inclusion
Ac the complement of the set A
K the set of admissible generalised coordinates
Uε (M) the open ε-neighbourhood of the set M (6.23)
U ε (M) the closed ε-neighbourhood of the set M
1
Introduction

When looking at the title of this work, one may wonder why we need an-
other text on the stability of dynamical systems. Why do we need one more,
when the number of existing textbooks on the topic of stability is already
very large? This monograph is not just about stability, it deals in particular
with non-smooth systems. In this introductory chapter a motivation will be
given for a monograph on the stability of non-smooth (mechanical) systems
(see Section 1.1). Historical notes on the theory of stability in Section 1.2
might shed some light on the origin of terminology in stability theory. In
Section 1.3, we provide a brief, and unavoidably incomplete, discussion on
different mathematical formalisms used to model non-smooth dynamical sys-
tems. After introducing the reader to some basic stability notions and the
concept of convergent systems in Section 1.4, a concise literature survey is
presented in Section 1.5 on the topics of non-smooth analysis, (measure) dif-
ferential inclusions, non-smooth mechanics, stability of non-smooth systems
and convergent systems. The objective of this monograph will be put forth in
Section 1.6. Finally, a closing section containing the outline of the succeeding
chapters is included.

1.1 Motivation
Is the solar system stable? Under which load will a beam buckle? Is the
figure of equilibrium of a steady rotating fluid stable? These fundamental
questions were some of the major problems that motivated scientists such
as Euler, Lagrange, Poincaré and Lyapunov to think about the concept of
stability of motion and how to prove it. The origin of stability theory must
clearly be sought in mechanics. The interest in the stability of motion is now
greater than ever and is no longer confined to mechanics. Stability issues play a
role in economical models, numerical algorithms, quantum mechanics, nuclear
physics, and control theory as fruitfully applied in for example the fields of
mechanical and electrical engineering. The theory of stability is now studied
2 1 Introduction

by many branches of science and is well established as can be inferred from

the vast amount of textbooks, conferences and journal articles on the subject.
A much younger field of interest is the study of the dynamics of non-
smooth systems. A look at the programs of international conferences from the
past few years reveals a rapidly increasing interest in non-smooth systems.
It is perhaps not exaggerated to state that it is nowadays fashion to study
non-smooth systems. Non-smooth models appear in many different disciplines.
Mechanical engineers study stick-slip oscillations in systems with dry friction
and the dynamics of impact phenomena and unilateral constraints. Electrical
circuits contain diodes and transistors, which ideally behave in a non-smooth
way. Control theorists also have to deal with other sources of non-smoothness
such as switching control laws and actuator saturation affecting the resulting
closed-loop dynamics. Switching systems also arise in other fields such as air
traffic management, economic models of markets and scheduling of automated
railway systems. Although these examples come from very different fields, the
mathematical structure and the related questions of interest are very similar.
In particular, the time evolution is often described by non-smooth differential
equations or (measure) differential inclusions.
This research monograph puts itself squarely between stability theory and
the field of non-smooth dynamics. As opposed to stability theory for smooth
dynamical systems, the research on stability properties of non-smooth dyna-
mical systems is still in its infancy. Here, we will focus primarily on the devel-
opment of stability results for measure differential inclusions, which naturally
arise when considering (mechanical) systems with inequality (unilateral) con-
straints. The impulsive nature of the vectorfield and the resulting state jumps
(discontinuities in the state evolution) challenge the extension of the existing
stability results for smooth systems to measure differential inclusions. First of
all, it is not even trivial how to define stability, attractivity and other proper-
ties of an equilibrium if the dynamics of the system has unilateral (inequality)
constraints, which restrict the state of the system to an admissible domain
and may induce state jumps. Moreover, the loss of uniqueness and existence of
solutions (of the initial value problem) greatly complicates the unambiguous
definition of these stability properties. Typically, the terminology for stability
theory is designed for smooth dynamical systems and its direct application
to non-smooth systems does sometimes not make sense or leads to confusion.
The classical results on stability properties have been formulated for smooth
dynamical systems and intrinsically make use of the smoothness assumption.
Clearly, when trying to study stability properties in non-smooth dynamical
systems, one should proceed very carefully and can not rely on standard text-
books. The motivation of this monograph has now become apparent: there
exists a need for the development of stability and attractivity results for non-
smooth systems formulated as measure differential inclusions.
As one of the oldest natural sciences, mechanics occupies a certain pi-
oneering role in determining the development of exact sciences through its
interaction with mathematics (quote from [63]). The historical notes given
 1.2 Historical Notes on the Theory of Stability 3

below show that mechanics has also played a dominant role in the develop-
ment of the theory of stability. It is therefore only natural to reserve a special
place for non-smooth mechanical systems within this monograph and to adopt
it as the prime application field in the current work.

1.2 Historical Notes on the Theory of Stability

Many different stability concepts exist in literature and are often named after
an illustrious scientist of a long forgotten time. Some historical notes on how
these stability concepts came into being therefore seem appropriate. The fol-
lowing is based on original texts as well as numerous textbooks on the history
of mechanics, see for instance [45, 54, 69, 161, 164].
The problem of stability in dynamical systems has attracted much inter-
est of physicists, mathematicians, engineers and astronomers during the last
centuries. In 1644, E. Torricelli (1608-1647) postulated his axiom [163]:

E. Torricelli, Opera Geometrica, De Motu Gravium, p. 99.

which we might translate as
Two interconnected weights cannot start moving by themselves simul-
taneously if their common centre of gravity does not descend.
Torricelli simply uses the words ‘start moving by themselves’ without speaking
of an initial disturbance from an equilibrium. Although Torricelli did not use
the word ‘stability’, his axiom certainly preludes a stability concept based on
the (gravitational) potential energy.
Inspired by the work of Archimedes of Syracuse, the Flemish and Dutch sci-
entists Simon Stevin (1548-1620) and Christiaan Huygens (1629-1695) studied
the equilibrium of floating bodies. In his work Byvough der weeghconst [158]
of 1605 Stevin wrote (following Archimedes) that a floating body takes such
a position that its centre of gravity is on the vertical centre line of gravity of
the displaced fluid. Furthermore, he gives the following corollary:

S. Stevin, Van de vlietende topswaerheyt, Byvough der

weeghconst, p. 202.
4 1 Introduction

which reads in modern English as

It is obvious that if the body’s centre of gravity is above the centre of
gravity of the body of the displaced water, it has such top-heaviness
that everything turns over (provided, however, it be not supported)
until the body’s centre line of gravity is in the vertical centre line of
gravity of the body of displaced water, below the centre of gravity of
the body of displaced water.
Apparently, Stevin thought that a floating body is stable when the centre of
gravity is above the centre of buoyancy (‘the centre of gravity of the body
of the displaced water’), which is not generally true. Compared with Stevin,
the work of Huygens has a more mathematical style, but Huygens work still
relies upon geometrical methods as was common in the 17th century. In many
theorems, such as the following one, Huygens addresses stability problems of
floating bodies [79]:
Si corpus solidum liquido supernatans ultrò inclinetur et alium si-
tum acquirat; altitudo centri gravitatis totius corporis supra centrum
gravitatis partis mersae, minor erit positione corporis posteriori quam
priori. 1
Ch. Huygens, De iis liquido supernatant libri 3, Theo-
rema 6.
The theorems of Huygens have a strong similarity with the work of Stevin.
However, the novelty in the work of Huygens is, that he explicitly compares
two different positions of the system. In the 18th century, the study of the roll-
stability of ships was carried on by Daniel Bernoulli (1700-1782), Leonhard
Euler (1707-1783) and Pierre Bouguer (1698-1758). Daniel Bernoulli distin-
guishes between stable (which he calls ‘firm’) and unstable equilibria of float-
ing bodies and writes [18, 31]
... quo ambo aequilibrii situs ab invicem distinguuntur; minima quidem
vis quaevis corpora etiamsi in aequilibrio firmo posita aliquantillum
nutare facit, sublata autem vi corpus rursus ad situm naturalem tendit,
nisi nutatio certos quosdam terminos transgressa fuerit. 2
D. Bernoulli, Commentationes de statu aequilibrii corpo-
rum humido insidentium, 1738, p. 148.
1
Free translation: If a solid body, floating on a liquid, inclines and acquires another
position, then the height of the center of gravity of the total body over the center
of gravity of the submerged part will be smaller in the latter position than in the
former position.
2
Translation by F. Cerulus [31]: .. by this, both positions of equilibrium are dis-
tinguished one from the other; indeed, a minimal arbitrary force makes a body
- although put in firm equilibrium - nod a little, but when the force has been
undergone [i.e. ceases to act], the body tends again to its natural position, unless
the nodding would have exceeded certain bounds.
 1.2 Historical Notes on the Theory of Stability 5

Daniel Bernoulli speaks explicitly of the stability of an equilibrium and con-

siders the couple of restoring forces when the equilibrium of the floating body
is perturbed by a small amount. Similar to Huygens, Bernoulli implicitly con-
siders two different positions of the system but their distance is small. Euler
refines Daniel Bernoulli’s work in his two volume treatise Scientia Navalis [49]
and distinguishes between equilibria which are stable, unstable and indifferent:
Stabilitas, qua corpus aquae innatans in situ aequilibrii persuerat, aes-
timanda est ex momento potentiae restituentis, si corpus dato angulo
infinite paruo ex situ aequilibrii fuerit declinatum. 3
L. Euler, Scientia Navalis, Part 1, Chapter 3, Proposi-
tion 19, p. 86.
Euler uses the word ‘stability’ and associates stability with the response on
an infinitely small disturbance from the equilibrium position. The idea of
infinitely small disturbances will later play a role in the work of Lagrange.
Bouguer introduced the term ‘metacentric height’, which became the modern
expression to determine the roll-stability of ships, see [125].
The theory of elastic stability in statics began with the work of Euler on
the critical buckling load of columns. Daniel Bernoulli suggested in a letter to
Euler (Oct. 20th 1742) that the differential equation of the elastica could be
found by minimising the integral of the square of the curvature along the rod,
being proportional to the elastic strain energy. Euler acted on this suggestion
in his ‘Additamentum’ de curvas elasticis of his work Methodus inveniendi
lineas curvas maximi minive proprietate gaudentes [48] on the calculus of
variations. Euler found a certain length which a column must attain to be
bent by its own weight or an applied weight, and concluded that for shorter
lengths it will simply be compressed, while for greater lengths it will be bent,
i.e. buckle. Although Euler initiated the analysis of the elastic stability of
the static equilibrium, he tacitly left the notion of stability undefined. In this
context, we have to remark that stability is in essence a concept of dynamical
systems as time plays an essential role.
The development of a stability concept in dynamics was continued by
J. L. Lagrange (1736-1813), who formalised the axiom of Torricelli for conser-
vative dynamical systems employing the concept of potential energy. Lagrange
wrote in his work Méchanique Analytique [91]:
On vient de voir que la fonction Φ [the potential energy] est un min-
imum ou un maximum, lorsque la position du systême est celle de
l’équilibre; nous allons maintenant démontrer que si cette fonction est
un minimum, l’équilibre aura de stabilité; ensorte que le systême étant
d’abord supposé dans l’état de l’équilibre, & venant ensuite à être tant
3
Free translation: The stability of a floating body in equilibrium is determined by
the restoring moment arising when the body has been displaced from equilibrium
by an infinitesimally small angle.
6 1 Introduction

soit peu déplacé de cet état, il tendra de lui-même à s’y remettre, en

faisant des oscillations infiniment petites.4
J. L. Lagrange, Méchanique Analytique,
Part 1, Section 3, No. 16.
In other words, Lagrange posed the theorem that, if the system is conservative,
a state corresponding to zero kinetic energy and minimum potential energy is
a stable equilibrium point. Moreover, Lagrange gave a definition of stability
of an equilibrium. Clearly, Lagrange meant that an equilibrium is stable when
neighbouring solutions remain close to the equilibrium, which agrees with our
modern concept of stability in the sense of Lyapunov. Lagrange speaks of in-
finitely small oscillations around an equilibrium because a stable equilibrium
in a conservative system is necessarily a centre. Using a Taylor series approx-
imation of the potential energy up to second-order terms, Lagrange proved
that the equilibrium is indeed stable when the first-order terms vanish and
the second-order terms are positive, corresponding to a minimum of the po-
tential energy. J. P. G. Lejeune Dirichlet (1805-1859) added a note [105] to the
theorem of Lagrange, arguing that a minimum of the potential energy might
also be caused by fourth or higher order terms in the Taylor series, but a min-
imum of the potential energy is sufficient to prove stability. The theorem is
in literature therefore referred to as the Lagrange-Dirichlet stability theorem
and plays an important role in elasto-statics. An equilibrium in elasto-statics
is called stable when it corresponds to a minimum of the potential energy.
Hence, the Lagrange-Dirichlet stability theorem is used in elasto-statics as
the definition of stability (instead of a condition for stability). The reason
for this is that the notions of time, velocity and solution are non-existent in
statics and a definition of stability based on those notions can not be given.
Celestial mechanics has greatly influenced the terminology of modern sta-
bility theory. P. S. Laplace (1749-1827) studied the celestial three-body prob-
lem with a perturbation analysis neglecting terms in the mass of second-order
and higher and assuming small values of eccentricity [93]. He concluded that,
under these assumptions, the variation of the semi-major axis of the orbits
is periodic with a constant amplitude, i.e. of the form A sin(αt + β). Laplace
speaks of the stabilité du système du monde, the stability of the world system,
and uses the same notion of stability as employed by Lagrange. Following the
work of Laplace, Lagrange extended this result for arbitrary eccentricities, still
neglecting terms in the mass of second-order and higher [90]. S. D. Poisson
(1781-1840) extended the perturbation analysis to second as well as to third-
order terms in the mass. Poisson showed that the variation of the semi-major
4
Free translation: We have shown that the [potential energy] function Φ is in a
minimum or maximum, when the configuration of the system is one of equilibrium;
we are now going to demonstrate that if this function is in a minimum then the
equilibrium will be stable, such that the system, being assumed in equilibrium
and displaced by a small amount, will tend to return to it by itself while making
infinitely small oscillations.
 1.2 Historical Notes on the Theory of Stability 7

axis only contains terms of the form A sin(αt + β) if terms in the mass of
second-order are taken into account [145]. Based on this analysis, the motion
of the earth and moon with respect to the sun therefore remains bounded.
Similar to Laplace, Poisson associates the boundedness of the variation of the
semi-major axis with the concept of stability. The extension of Poisson to
third-order terms in the perturbation analysis reveals in addition terms of the
form At sin(αt + β) in the variation of the semi-major axis. The amplitude
of these oscillatory terms grows unboundedly with time. An analysis based
on third-order terms therefore implies that the motion of the planets does
not remain bounded but the planets come arbitrarily close to their original
position infinitely many times. Also C. G. J. Jacobi (1804-1851) contemplates
whether the results of Laplace, Lagrange and Poisson are the proof of the
Stabilität des Weltsystems, i.e. the stability of the world-system [80]. Jacobi,
like Laplace and Poisson, associates the word stability with periodic functions
which do not grow unboundedly. In 1887, King Oscar II of Sweden sponsored a
mathematical competition with a prize for a resolution of the question of how
stable the solar system is, a variation of the three-body problem. H. Poincaré
(1854-1912) won the prize which was the beginning of his work on the stability
of the solar system which accumulated in his work Les Méthodes Nouvelles de
la Mécanique Céleste [144]. Poincaré states that the term ‘stability’ is used in
different ways and he continues to discuss the differences between the results
of Lagrange and Poisson on the variation of the semi-major axis. According
to Poincaré, Lagrange found that the variation of the semi-major axis is gov-
erned by terms of the form A sin(αt + β), whereas Poisson found that there
are in addition terms of the form At sin(αt + β). He concludes that the word
stability does not have the same meaning for Lagrange and for Poisson. With
respect to the solar system, Poincaré speaks of the stability in the sense of
Lagrange, with which he means bounded behaviour of the planetary orbits,
and stability in the sense of Poisson, for which the planets come arbitrarily
close to their original position infinitely many times. Hence, Poincaré gives
the false impression that Lagrange associates his work on the semi-major axis
with the concept of stability. Moreover, Poincaré gives the false impression
that Poisson’s notion of stability is related to non-periodic terms which grow
unboundedly. These two errors have persisted in modern times and have led to
misnomers. Nowadays, the terms ‘Lagrange-stability’ and ‘Poisson-stability’,
inherited from Poincaré, are used to denote boundedness and recurrence of
solutions of arbitrary dynamical systems. The modern concept of Lagrange
stability is therefore very different from the concept of stability as expressed
by Lagrange himself in Méchanique Analytique. The work of Poincaré on pe-
riodic orbits led to the modern concept of Poincaré stability, which is also
called orbital stability.
In the nineteenth century, the development of regulators for steam engines
and water turbines led to the stability analysis of machines. J. C. Maxwell
(1831-1879) analysed in a paper of 1868 [112] the stability of Watt’s flyball
governor. His technique was to linearise the differential equations of motion
8 1 Introduction

to find the characteristic equation of the system. He studied the effect of the
system parameters on stability and showed that the system is stable if the
roots of the characteristic equation have negative real parts. In 1877, E. J.
Routh (1831-1907) provided an algorithm for determining when a character-
istic equation has stable roots. The Russian I. I. Vishnegradsky analysed in
1877 the stability of regulators using differential equations independently of
Maxwell and studied the stability of the Watt governor in more detail. In 1893,
A. B. Stodola studied the regulation of a water turbine using the techniques
of Vishnegradsky. Stodola modelled the actuator dynamics and included the
delay of the actuating mechanism in his analysis and was the first to men-
tion the notion of the system time constant. Unaware of the work of Maxwell
and Routh, Stodola posed the problem of determining the stability of the
characteristic equation to A. Hurwitz, who solved it independently.
An exact definition of stability for a dynamical system, as well as general
stability theorems for nonlinear systems, were first formulated in 1892 by the
Russian mathematician and engineer A.M. Lyapunov (1857-1918). Lyapunov
defined an equilibrium to be stable when for each ε-neighbourhood one can
find a δ-neighbourhood of initial conditions, such that their solutions remain
within the ε-neighbourhood. Loosely speaking, this means that neighbouring
solutions remain close to the equilibrium, which is essentially the same as what
Lagrange understood under the term stability (see the above citation). Lya-
punov proved stability using two distinct methods. In the first method, known
as Lyapunov’s first method or Lyapunov’s indirect method, the local stabil-
ity of an equilibrium is studied through linearisation. The second method,
also called the direct method of Lyapunov, is far more general. The funda-
mental idea behind the direct method of Lyapunov is the stability theorem
of Lagrange-Dirichlet, which is based on the mechanical energy. The direct
method of Lyapunov is able to prove the stability of equilibria of nonlinear
differential equations using a generalised notion of energy functions. Unfortu-
nately, though his work was applied and continued in Russia, the time was
not ripe in the West for his elegant theory, and it remained unknown there
until its French translation in 1907 [106] (reproduced in [109]). Its importance
was finally recognised in the 1960’s with the emergence of control theory.

1.3 Non-smooth Dynamical Systems

In this section we try to clarify the terminology around non-smooth dynamical

systems and briefly discuss various mathematical frameworks to describe non-
smooth dynamical systems.
Loosely speaking, a dynamical system is a system whose state evolves
in time. The time-evolution is governed by a set of rules (usually equations).
We distinguish between continuous-time dynamical systems and discrete-time
dynamical systems. The state of a continuous-time dynamical system may
change at every time-instance, either continuously or discontinuously. The
 1.3 Non-smooth Dynamical Systems 9

evolution of continuous-time systems is usually expressed by a set of or-

dinary differential equations (and sometimes algebraic equations if equality
constraints are involved). The state of a discrete-time system, for which the
evolution is governed by a discrete mapping, is only defined and can change at
certain discrete time-instances. In the following, we will mean with the term
‘system’ a dynamical system in continuous-time if not stated otherwise. Note
that the state of a continuous-time system can still vary discontinuously in
time, but time itself is a continuum.
Lagrangian mechanical systems are described by second-order differential
equations in the generalised positions q(t). A second-order differential equa-
tion can always be written as a system of first-order differential equations of
the form ẋ(t) = f (t, x(t)). We refer to the function f (t, x(t)) as the right-hand
side of the system which induces a vector field in the state-space.
The term ‘non-smooth dynamical system’ or ‘discontinuous dynamical sys-
tem’ is often used in literature without stating explicitly which properties of
the system are considered to be ‘non-smooth’. Non-smooth (continuous-time)
dynamical systems in first-order form can be divided in three types according
to their degree of non-smoothness:
1. Non-smooth continuous systems which are described by a Lipschitz con-
tinuous right-hand side f (t, x(t)) which is non-differentiable with respect
to x at certain hyper-surfaces in the state-space. For example, consider a
mechanical oscillator with a one-sided elastic support and external exci-
tation. If the position of the mass m is described by q, the spring stiffness
by k and the support stiffness ksup , then the system is described by the
second-order differential equation

0, q≥0
mq̈ + kq = f0 cos(ωt) − fsup (q), fsup (q) = , (1.1)
ksup q, q < 0

which can be put in the first-order form

 
q̇(t)
ẋ(t) = f (t, x(t)) = , (1.2)
−m
k
q(t) + fm0 cos(ωt) − 1
m fsup (q(t))
 T
where x(t) = q(t) q̇(t) and f (t, x(t)) is a right-hand side which is
continuous in x but non-smooth at the hyper-surface Σ = {x ∈ R2 | q =
0}.
2. Systems described by differential equations ẋ(t) = f (t, x(t)) with a right-
hand side f (t, x(t)) which is bounded but discontinuous on certain hyper-
surfaces Σi in the state-space. The solution of such a discontinuous differ-
ential equation has to be understood in the sense of Filippov [51] (i.e. the
differential equation is extended to a differential inclusion). The disconti-
nuities in f (t, x(t)) cause the time-derivative ẋ(t) not to be defined for all
t, but the state x(t) remains time-continuous. Examples are mechanical
systems with visco-elastic supports or dry friction.
10 1 Introduction

3. Systems which expose discontinuities (or jumps) in the state, such as

systems with impulsive effects. The state x(t) is not defined on such dis-
continuity points. Examples are mechanical systems with velocity jumps
due to impact.
Non-smooth dynamical systems are studied by various scientific communi-
ties using different mathematical frameworks. Examples of such mathematical
frameworks are:

1. Singular perturbations. The non-smooth system is replaced by a singularly

perturbed smooth system. The resulting ordinary differential equation is
extremely stiff and hardly suited for numerical integration. More impor-
tantly, possible stationary states, which exist in the non-smooth dynamical
system, may be lost due to the smoothening. For instance, the stationary
state of a block on a rough slope can not be described with a smooth
friction model for which the friction force vanishes at zero relative veloc-
ity (the block will therefore always slide). Therefore, this framework can
generally not be used to study stability properties of non-smooth systems.
2. Switched systems or Hybrid systems. A kind of switching is involved in
non-smooth systems and these systems are therefore often called ‘switched
systems’ or ‘differential equations with switching conditions’ [46, 107]. In
the field of systems and control theory, the term hybrid system is fre-
quently used for systems composed of continuous differential equations
and discrete event parts [23]. Nowadays, the term ‘hybrid system’ is
used for any system which exposes a mixed continuous and discrete na-
ture [71, 111, 174]. The switched/hybrid system concept switches between
differential equations with possible state re-initialisations and is not able
to describe accumulation points, e.g. infinitely many impacts which occur
in a finite time interval such as a bouncing ball coming to rest on a table.
3. Complementarity systems. Non-smooth systems can in some cases be con-
sidered as dynamical extensions of (non)linear complementarity problems,
which gives the complementarity systems concept [72].
4. Differential inclusions or Filippov systems. Systems described by differ-
ential equations with a discontinuous right-hand side, but with a time-
continuous state, can be extended to differential inclusions with a set-
valued right-hand side. The work of Filippov [51] proved to be seminal in
the extension of differential equations with discontinuous right-hand sides
to differential inclusions with set-valued right-hand sides and these sys-
tems are therefore often called Filippov systems (Section 4.2). The differ-
ential inclusion concept gives a simultaneous description of the dynamics
in terms of a single inclusion, which avoids the need to switch between dif-
ferent differential equations. Moreover, this framework is able to describe
accumulation points of time-instances at which the time-derivative of the
state jumps.
5. Measure differential inclusions. Systems which expose discontinuities in
the state and/or vector field can be described by measure differential
 1.4 Stability and Convergence 11

inclusions. The differential measure of the state does not only consist of a
part with a density with respect to the Lebesgue measure, but is also al-
lowed to contain an atomic part. The dynamics of the system is described
by an inclusion of the differential measure of the state to a state-dependent
set (similar to the concept of differential inclusions). Consequently, the
measure differential inclusion concept describes the continuous dynamics
as well as the impulse dynamics with a single statement in terms of an
inclusion and is able to describe accumulation phenomena. Moreover, the
framework of measure differential inclusions leads directly to a numerical
discretisation, called the time-stepping method, which is a robust algo-
rithm to simulate the dynamics of non-smooth systems.

The framework of measure differential inclusions allows us to describe systems

with state discontinuities and this framework is therefore more general than
differential inclusions. However, the great advantage of this framework over
other frameworks is, that physical interaction laws, such as friction and im-
pact in mechanics or diode characteristics in electronics, can be formulated as
set-valued force laws and be seamlessly incorporated in the formulation. We
will therefore use the framework of measure differential inclusions in this work
to study stability properties of non-smooth systems in general and mechan-
ical systems with unilateral constraints, such as Coulomb friction, unilateral
contact and impact, in particular.

1.4 Stability and Convergence

Stability theory plays today a dominant role in system and control theory.
Various stability concepts exist as has been already mentioned in the histori-
cal overview of Section 1.2. The most important stability concept is stability
in the sense of Lyapunov. An equilibrium point of a dynamical system is sta-
ble in the sense of Lyapunov if all solutions which start nearby remain in the
neighbourhood of the equilibrium point for all future times. An equilibrium
point which is not stable is called unstable. If all solutions starting within
some neighbourhood of the equilibrium point converge towards the equili-
brium point as time goes on, then the equilibrium point is called attractive.
Stability and attractivity are two different notions. An equilibrium point can
for instance be stable without being attractive, or it can be unstable and at
the same time attractive. An equilibrium point which is stable and attractive
is usually called asymptotically stable. This terminology will turn out to be
cumbersome when dealing with non-smooth systems. The notions of stability
and attractivity can be generalised to positively invariant sets such as periodic
solutions, quasi-periodic solutions etc.
One of the main issues in stability theory is the question of how to deter-
mine whether an equilibrium point is stable and/or attractive. The so-called
Lyapunov theorems give sufficient conditions for the stability (and sometimes
12 1 Introduction

attractivity) of equilibrium points (see e.g. [85] for a recent overview). Fur-
thermore, LaSalle’s Invariance Principle can give sufficient conditions for at-
tractivity [89].
The title of this monograph preludes that this work also deals with the
so-called convergence property. Lyapunov stability and attractivity concepts
are used to characterise stability properties of equilibrium points, or in the
more general case, of positively invariant sets. In contrast, convergence is a
property of the dynamic system. A system, which is excited by an input, is
called (uniformly) convergent if it has a unique solution that is bounded on
the whole time axis and this solution is globally attractive and stable. Obvi-
ously, if such a solution does exist, then all other solutions converge to this
solution, regardless of their initial conditions. Moreover, such a solution can
be considered as a steady-state solution. The property of convergence can be
beneficial from several points of view. Firstly, in many control problems it is
required that controllers are designed in such a way that all solutions of the
corresponding closed-loop system “forget” their initial conditions. Actually,
one of the main tasks of feedback is to eliminate the dependency of solutions
on initial conditions. In this case, all solutions converge to some steady-state
solution that is determined only by the input of the closed-loop system. This
input can be, for example, a command signal or a signal generated by a feed-
forward part of the controller or, as in the observer design problem, it can be
the measured signal from the observed system. Such a convergence property
of a system plays an important role in many nonlinear control problems in-
cluding tracking, synchronisation, observer design, and the output regulation
problem, see [131, 132, 169]. Secondly, from a dynamics point of view, conver-
gence is an interesting property because it excludes the possibility of different
coexisting steady-state solutions: namely, a convergent system excited by a
periodic input has a unique globally attractive and stable periodic solution
with the same period time. Moreover, the notion of convergence is a powerful
tool for the analysis of time-varying systems. This tool can be used, for exam-
ple, for performance analysis of nonlinear control systems [73, 171]. Namely,
as shown in [133], convergent systems allow for the definition of so-called gen-
eralised frequency response functions. For linear systems, frequency domain
analysis is based on frequency response functions and has been crucial for the
performance analysis of linear control systems. The recent extension to gen-
eralised frequency response functions for the class of (nonlinear) convergent
systems may open up a route towards more performance-based control design
techniques for nonlinear (non-smooth) control systems. Here, we aim to pro-
vide conditions for convergence for a class of measure differential inclusions,
making such systems accessible for analysis and control synthesis results for
convergent systems.
 1.5 Literature Survey 13

1.5 Literature Survey

In this section a short literature survey will be given, which may help the
reader to obtain some background knowledge or to direct the reader to further
references, but the survey is by no means complete.
First, a brief review will be given of the mathematical literature on non-
smooth analysis and (measure) differential inclusions, which forms a prerequi-
site for the study of non-smooth mechanical systems. Subsequently, the con-
tributions from the field of non-smooth mechanics are discussed. Finally, the
literature on stability in non-smooth dynamical systems and convergent dy-
namics is reviewed. Clearly, some publications contribute to more than one
topic.

Non-smooth Analysis

The book [147] by Rockafellar is regarded to be the classical reference on

Convex Analysis. Meanwhile, many other books on Convex Analysis have been
published, such as the books [77,78] by Hiriart-Urruty and Lemaréchal,
the book [149] by Rockafellar and Wets and the short overview in the
encyclopedia [56] of Gamkrelidze by Thikomirov.
The generalised gradient, which is used extensively here, is introduced in
the book [36] by Clarke and is also presented in the book [37] by Clarke
et al. The books [10, 12] by Aubin and co-workers are extensive references on
set-valued functions.

Differential Measures and (Measure) Differential Inclusions

The fundamental work of Filippov [50, 51, 96, 152] extends a discontinuous
differential equation to a differential inclusion (see Chapter 4). More results on
differential inclusions can be found in the book [10] by Aubin and Cellina,
the book [43] by Deimling and in the book [37] by Clarke et al. Utkin
introduced in [165–168] the concept of ‘equivalent control’, which is similar to
Filippov’s convex method, and is popular in control theory.
The classical theory of measures and integration can be found in the text-
books of Elstrodt [47] and Rudin [150]. Moreau [122, 124] directly intro-
duces the so-called differential measure, which can be related to real mea-
sures and (signed) measures known from classical measure theory. Discon-
tinuous evolution problems with state re-initialisations can be described by
measure differential inclusions, see Moreau [122, 124], Glocker [63], Mon-
teiro Marques [116], Ballard [14] and Stewart [159].
Stochastic measure differential inclusions are treated by Bernardin,
Schatzman and Lamarque [17].
14 1 Introduction

Non-smooth Mechanics

The number of publications on non-smooth systems in the field of Non-smooth

Mechanics is vast. The book [23] by Brogliato gives an extensive review as
well as the article of Brogliato et al. [28]. Only the main publications and
those which are of relevance for the next chapters will be briefly reviewed
here.
The work of Moreau and Jean [81, 121, 123] and of Panagiotopuo-
los [126, 127] has fulfilled a pioneering role in the Non-smooth Mechanics
community. The work of Panagiotopoulos focusses on variational formu-
lations and on the formulation of contact laws in elastostatics (see also the
work of Goeleven et al. [67,68]). Moreau and Jean developed a framework
to describe frictional impact and unilateral contact using Convex Analysis to
its full extent. The framework of Moreau treats the dynamics of systems
with impact on velocity–impulse level in terms of measure differential inclu-
sions. This framework is directly related to the framework of complementarity
systems, see Acary et al. [1, 25]. A school of researchers (mostly French) is
actively extending the work of Moreau.
Monteiro Marques [116] and Schatzman [153] prove the existence of
solutions of measure differential inclusions describing a mechanical system
with a single unilateral constraint. Uniqueness of solutions of frictionless im-
pact problems has been studied by Schatzman [154], Percivale [135, 136]
and Ballard [14]. Non-existence and non-uniqueness of solutions of me-
chanical systems with frictional unilateral contact is discussed in Génot and
Brogliato [23, 55] and Leine, Brogliato and Nijmeijer [98].
Much effort has been put in the formulation of frictional contact problems
by means of set-valued force laws. Glocker, Pfeiffer and co-workers ap-
plied the theory on linear complementarity problems and Convex Analysis on
rigid multibody systems with impact and friction. In Glocker [57] and the
book of Pfeiffer and Glocker [139], unilateral contact laws and Coulomb
friction are formulated as linear complementarity problems on acceleration
level and velocity level and simulation is performed using an event-driven in-
tegration method. The theory and methods have been applied to a variety
of engineering applications, see e.g. [142, 160, 177] and the overview in [139].
Later, spatial frictional contact problems have been formulated as solvable
nonlinear complementarity problems in Glocker [60, 63]. The theory of set-
valued force laws (i.e. set-valued constitutive laws) was put in a framework of
non-smooth potential theory in the book [63] by Glocker. The book [63] dis-
cusses non-smooth extensions to the classical variational principles in dynam-
ics, e.g. Gauß, Jourdain, d’Alembert/Lagrange (see also Glocker [58,59]). A
set-valued force law for spatial Coulomb-Contensou friction (combined sliding
and pivoting friction) can be found in Leine and Glocker [99].
Alart and Curnier [5] presented an augmented Lagrangian approach
to the solution of static frictional contact problems (see also the work of
Simo and Laursen [94, 155] and Klarbring [86]). The application of the
 1.5 Literature Survey 15

augmented Lagrangian method to dynamic frictional contact problems is pre-

sented in Leine and Glocker [99] and Leine and Nijmeijer [101].
The developed approach for the description and numerical treatment of
non-smooth Lagrangian mechanical systems can very well be equally ap-
plied to other types of engineering systems. For example, electrical circuits
are treated as measure differential inclusions with set-valued force laws by
Glocker [64].

Stability in Non-smooth Systems

The literature on stability theory for non-smooth systems is vast and rapidly
growing, providing dedicated results for specific mathematical formalisms. To
name a few, see e.g. [82] for results on piece-wise affine systems, [74, 107]
for results on switched systems, [29, 138, 151, 180] for results on hybrid sys-
tems, [13, 92] for results on impulsive dynamical systems, and [30] for results
on complementarity systems. For the sake of brevity, in the remainder of this
overview we will mainly focus on the existing literature on the stability of
(measure) differential inclusions (and, to a lesser extent, hybrid systems also
allowing for state jumps) and results for the specific case of non-smooth me-
chanical systems. Herein, it is notable that most results in literature focus on
the stability of isolated equilibria and few results for equilibrium sets exist (a
topic receiving considerable attention in the current work).
Yakubovich, Leonov and Gelig [179] discuss the stability of equili-
brium sets and the dichotomy property (meaning that solutions either con-
verge to an equilibrium or grow unboundedly) in differential inclusions. It is
explained how the construction of Lyapunov functions, which guarantee global
stability, leads to algebraic problems on the solution of matrix inequalities.
Adly, Goeleven and Brogliato [2, 3, 66] and Van de Wouw and
Leine [170] study stability properties of equilibrium sets of differential in-
clusions describing mechanical systems with friction. It is assumed that the
non-smoothness is stemming from a maximal monotone operator. Existence
and uniqueness of solutions is therefore always fulfilled. A basic Lyapunov
theorem for stability and attractivity is given in [3, 66] for first-order differ-
ential inclusions and in [2] for linear second-order differential inclusions with
maximal monotone operators. The results are applied to linear mechanical
systems with friction. It is assumed in [3,170] that the relative sliding velocity
of the frictional contacts depends linearly on the generalised velocities and
conditions for the attractivity of an equilibrium set are given. The results are
generalised in [3] to conservative systems with an arbitrary potential energy
function. The paper [170] gives many illustrating examples and a useful (and
intuitive) condition for attractivity of an equilibrium set for linear mechanical
systems. A numerical analysis of the stability properties of equilibrium sets in
a simple mechanical system with frictional unilateral contact is given in [16].
The stability of hybrid systems with state-discontinuities is addressed by
a vast number of researchers in the field of control theory. The books of
16 1 Introduction

Lakshmikantham, Bainov and Simeonov [13, 92] focus on systems with

impulsive effects and give many useful Russian references. Lyapunov stability
theorems, instability theorems and theorems for boundedness are given by Ye,
Michel and Hou [180]. Pettersson and Lennartson [138] propose stabil-
ity theorems using multiple Lyapunov functions. By using piecewise quadratic
Lyapunov function candidates and replacing the regions where the different
stability conditions have to be valid by quadratic inequality functions, the
problem of verifying stability is turned into a Linear Matrix Inequality (LMI)
problem. Pereira and Silva [137] study the stability of equilibrium points of
measure differential inclusions, but only using continuous Lyapunov functions.
See also the review article of Davrazos and Koussoulas [42].
Many publications focus on the control of mechanical systems with fric-
tionless unilateral contacts by means of Lyapunov functions. See for instance
Brogliato et al. [27] and Tornambè [162] and the book [23] for further
references.
The Lagrange-Dirichlet stability theorem is extended by Brogliato [24]
to measure differential inclusions describing mechanical systems with friction-
less impact. The idea to use Lyapunov functions involving indicator functions
associated with unilateral constraints is most probably due to [24]. More gen-
erally, the work of Chareyron and Wieber [32–34] is concerned with a
Lyapunov stability framework for measure differential inclusions describing
mechanical systems with frictionless impact. It is clearly explained in [34]
why the Lyapunov function in the presence of state-discontinuities has to be
globally positive definite, in order to prove local stability properties (when
no further assumptions on the system or the form of the Lyapunov function
are made). The importance of this condition has also been stated in [13, 180]
for hybrid systems and in [27, 162] for mechanical systems with frictionless
unilateral constraints. Moreover, LaSalle’s invariance principle is generalised
in [26] to differential inclusions and in [33,34] to measure differential inclusions
describing mechanical systems with frictionless impact. The proof of LaSalle’s
invariance principle strongly relies on the positive invariance of limit sets. It
is assumed in [33, 34] that the system enjoys continuity of the solution with
respect to the initial condition which is a sufficient condition for positive in-
variance of limit sets. In [34], an extension of LaSalle invariance principle to
systems with unilateral constraints is presented (more specifically it is applied
to mechanical systems with frictionless unilateral contacts). In [26] an exten-
sion of LaSalle invariance principle for a class of unilateral dynamical systems,
so-called evolution variational inequalities, is presented.
Instability results for finite dimensional variational inequalities can be
found in the work of Goeleven and Brogliato [65, 66], whereas Quit-
tner [146] gives instability results for a class of parabolic variational inequal-
ities in Hilbert space.
Closely related to stability problems are bifurcations of equilibria and pe-
riodic solutions. Bifurcations in non-smooth systems do recently receive much
 1.6 Objective and Scope 17

attention in literature (see the work of Leine [96, 97, 100, 101] for a literature
survey).

Convergent Dynamics

The first conditions for the convergence property have been formulated by
Pliss [143] and Demidovich [44] for smooth nonlinear systems, where Pliss
focussed on systems with periodic right-hand sides. Yakubovich [178] consid-
ered Lur’e-type systems, possibly with discontinuities, and proposed sufficient
convergence conditions based on the circle criterion. More recently, the work
of Pavlov, Van de Wouw and Nijmeijer [130] has given sufficient condi-
tions for continuous (though non-smooth) piece-wise affine (PWA) systems in
terms of the existence of a common quadratic Lyapunov function for all affine
systems constituting the PWA system. It is shown that the existence of such
a common quadratic Lyapunov function is by no means sufficient for conver-
gence of discontinuous PWA systems and sufficient conditions for convergence
of discontinuous PWA systems are proposed. Results on the convergence prop-
erty of a class of switched linear systems are proposed in [173].
Similar notions describing the property of solutions converging to each
other are studied in literature. The notion of contraction has been introduced
by Lohmiller and Slotine [108] (see also references therein). An operator-
based approach towards studying the property that all solutions of a system
converge to each other is pursued by Fromion and co-workers [52, 53]. In the
work of Angeli [7], a Lyapunov-based approach has been developed to study
the global uniform asymptotic stability of all solutions of a system (in [7], this
property is called incremental stability) as well as the so-called incremental
input-to-state stability property, which is compatible with the input-to-state
stability approach (see e.g. [157]). In the same spirit, the convergence property
has been extended to the so-called input-to-state convergence property [132].
Convergence properties of measure differential inclusions have been studied
by [104], see also Chapter 8.

1.6 Objective and Scope

As stated in Section 1.1, there exists a need to redefine the notion of stability
and related properties in the light of non-smooth systems as well as to for-
mulate stability results providing stability properties in the redefined sense.
The objective of this research monograph is, in one phrase, to partially ad-
dress this need within the setting of measure differential inclusions and with
an emphasis on mechanical (Lagrangian) systems with frictional unilateral
constraints. More specifically the aims are:
• to introduce the reader to the formulation of finite-dimensional Lagrangian
mechanical systems with unilateral constraints as measure differential
18 1 Introduction

inclusions, as well as to present various friction models as set-valued force

laws,
• to precisely define stability, attractivity and related terminology for mea-
sure differential inclusions,
• to present the Lyapunov stability theory (Lyapunov’s direct method,
LaSalle’s invariance principle, instability theorems) for measure differential
inclusions,
• to apply and specialise the Lyapunov stability theory to Lagrangian me-
chanical systems with frictional unilateral constraints,
• to study the convergence property for measure differential inclusions and
to give sufficient conditions for convergence for systems which enjoy certain
monotonicity properties,
• to show the use of the convergence property in solving the tracking control
problem for certain non-smooth mechanical systems.
Special attention will be paid to the stability/attractivity of equilibrium sets
of measure differential inclusions.

1.7 Outline
The first few chapters of this monograph (Chapter 2–5) introduce the mathe-
matical framework to describe non-smooth systems and deal with the formu-
lation of mechanical systems with frictional unilateral constraints within this
framework. Chapters 6 and 7 are devoted to the stability of non-smooth (me-
chanical) systems and form the core of the work. Finally, Chapter 8 revolves
around the convergence property for measure differential inclusions.
Chapter 2 presents some basic mathematical theory from non-smooth anal-
ysis and convex analysis. The notions of generalised derivatives, normal cones
and indicator functions will prove to be essential when dealing with non-
smooth systems and in particular with set-valued force laws derived from
non-smooth potentials for the description of frictional unilateral contact in
mechanical systems.
The theory of measures forms the basis for the modern integration theory
and both are quintessential for the understanding of measure differential in-
clusions. In Chapter 3 we give a short overview of measure and integration
theory and relate the differential measure (which is pivotal in the definition
of measure differential inclusions) to real measures and (signed) measures.
Chapter 4 is concerned with non-smooth dynamical systems and their
solution concepts. A brief introduction to differential equations is followed
by a discussion of differential equations with a discontinuous right-hand side.
The requirement for the existence of a solution leads to the need to fill in
the graph of the discontinuous right-hand side (Filippov’s convex method).
The resulting set-valued right-hand side brings forth a differential inclusion.
Subsequently, the differential measure of the state, which classically contains
 1.7 Outline 19

a density with respect to the Lebesgue measure, is extended with an atomic

part, which leads to a measure differential inclusion.
An important class of measure differential inclusions is formed by La-
grangian mechanical systems with frictional unilateral constraints. Chapter 5
deals with the mathematical formulation of Lagrangian mechanical systems
with unilateral contact and friction modelled with set-valued force laws. First,
it is shown how set-valued force laws can be derived from non-smooth poten-
tials. Subsequently we treat the contact laws for unilateral contact and various
types of friction within the setting of non-smooth potential theory. This leads
to a unified approach with which all set-valued forces can be formulated. Fi-
nally, the set-valued forces are incorporated as Lagrangian multipliers in the
Newton-Euler equations.
Chapter 6 is devoted to the Lyapunov stability theory for measure differ-
ential inclusions. Definitions of stability, attractivity and related terminology
are given, that make sense for non-smooth dynamical systems with possible
discontinuities in the state and loss of uniqueness of solutions. Generalisa-
tions of the direct method of Lyapunov, LaSalle’s invariance principle and
Chetaev’s instability theorem are given.
The stability results of Chapter 6 are applied in Chapter 7 to Lagrangian
mechanical systems with frictional unilateral constraints. The special struc-
ture of mechanical systems allows for a refinement of the stability results and
the formulation of intuitive stability conditions.
In Chapter 8 the convergence property of measure differential inclusions
is studied and results are presented which give sufficient conditions for the
convergence of measure differential inclusions with certain maximal mono-
tonicity properties. It is shown how these convergence results for measure
differential inclusions can be exploited to solve tracking control problems for
certain classes of non-smooth mechanical systems with friction and one-way
clutches. Illustrative examples of convergent mechanical systems are discussed
in detail.
Finally, some concluding remarks are given in Chapter 9, which put the
presented results in a broader perspective.
This page intentionally blank
2
Non-smooth Analysis

This chapter presents some basic mathematical theory from non-smooth anal-
ysis [10, 12, 37, 56, 77, 78, 147, 149]. The aim of this chapter is not to give a real
introduction to non-smooth analysis as the above textbooks are much bet-
ter suited for that task. Instead, the primary aim of the chapter is to make
the reader is familiar with the terminology and notation used in this mono-
graph. Moreover, it provides a compendium on non-smooth and convex anal-
ysis which is useful when reading the following chapters. The reader might
want to look up how a mathematical term is exactly defined, making use of
the index in combination with this chapter.
We begin with a brief introduction to sets (Section 2.1). The notion of
continuity of functions is relaxed in Section 2.2 to semi-continuity and the
notion of the classical derivative of smooth functions is extended to generalised
differentials for non-smooth functions in Section 2.3. Subsequently, we discuss
set-valued functions in Section 2.4. Topics from convex analysis are reviewed
in Section 2.5 and the subderivative is discussed in Section 2.6.

2.1 Sets
A number of properties of sets and set-valued functions will be briefly re-
viewed. Let C be a subset of the normed space Rn , equipped with the Eu-
clidean norm  · .
Definition 2.1 (Closed Set). A set C ⊂ Rn is closed if it contains all its
limit points. Every limit point of a set C is the limit of some sequence {xk }
with xk ∈ C for all k ∈ N.
The boundary of a set C, denoted by bdry C, is the set of points which can
be approached both from C and from the outside of C. We define the closure
of a set C as the smallest closed set containing C, i.e. C = C ∪ bdry C.
Furthermore, those points in C which are not on the boundary form the
interior of C, int C = C\ bdry C. We can uniquely decompose the closure of a
22 2 Non-smooth Analysis

set in its boundary and its interior: C = bdry C ∪ int C. A set is called open,
if it does not contain any of its boundary points, i.e. C ∩ bdry C = ∅. It holds
that int C is an open set.
Definition 2.2 (Bounded Set). A set C ⊂ Rn is bounded if there exists a
point y ∈ Rn and a finite number c > 0 such that x − y < c for all x ∈ C.
A set is bounded if it is contained in a ball of finite radius.
Definition 2.3 (Compact Set). A set C ⊂ Rn is compact if it is closed and
bounded.
An important property in Non-smooth Analysis is the convexity of sets.
Definition 2.4 (Convex Set). A set C ⊂ Rn is convex if for each x ∈ C
and y ∈ C also (1 − q)x + qy ∈ C for arbitrary q with 0 ≤ q ≤ 1.
It follows that a convex set contains all line segments between any two points
in the set.

Definition 2.5 (Convex Hull). The convex hull of a set C ⊂ Rn , denoted

by co(C), is the smallest convex set that contains all the points of C.

The convex hull of a set C is therefore the intersection of all the convex sets
containing C. Consequently, the closed convex hull of {x, y} ∈ Rn is the line
segment between x and y, i.e. the smallest closed convex set containing x
and y
co{x, y} = {(1 − q)x + qy, ∀ q ∈ [0, 1]}. (2.1)
Figure 2.1 illustrates the notion of convexity and the convex hull.

(a) (b)

Fig. 2.1. Convex set (a) and non-convex set with its convex hull (b).
 2.2 Functions and Continuity 23

Fig. 2.2. Upper and lower semi-continuity of functions.

2.2 Functions and Continuity

A function f : R → R associates to any element x of its domain a single func-
tion value f (x). A function is therefore single-valued which excludes vertical
lines and loops on its graph, where

graph(f ) = {(x, f (x)) | x ∈ R}. (2.2)

We define the epigraph of f to be the set above the graph of f

epi(f ) = {(x, y) | y ≥ f (x); x ∈ R}. (2.3)

Definition 2.6 (Convex Function). A function f : R → (−∞, ∞] is convex

when epi(f ) is a convex set.
For a differentiable function f : Rn → R, it follows that f (x) is convex if

f (x∗ ) ≥ f (x) + ∇f (x)T (x∗ − x), ∀x∗ and x. (2.4)

Let Bδ denote the open ball with radius δ centred at the origin. A function
f (x) : Rn → Rn is continuous at x ∈ X, X ⊂ Rn , provided that for all ε > 0,
there exists a δ > 0 such that y ∈ x + Bδ ⊂ X implies f (x) − f (y) < ε. For
single-valued functions this means that we can draw the graph of the function
without taking the pencil of the paper. In the following chapters we will often
use the word ‘smooth’.

Definition 2.7 (Smooth function). A function f (x) : Rn → R is called

smooth if it is continuously differentiable up to any order in x. A vector-valued
function f (x) : Rn → R is smooth if each element fi (x) is a smooth function.

For a function f : Rn → (−∞, +∞] we can define semi-continuity:

Definition 2.8 (Upper Semi-continuity). A function f (x) is upper semi-
continuous if
lim sup f (y) ≤ f (x) ∀x ∈ Rn .
y→x
24 2 Non-smooth Analysis

We say that f is upper semi-continuous at x if for every ε > 0 there exists a

neighbourhood U of x such that f (y) < f (x) + ε for all y ∈ U . Similarly, we
introduce lower semi-continuity of a real-valued function:
Definition 2.9 (Lower Semi-continuity). A function f (x) is lower semi-
continuous if
lim inf f (y) ≥ f (x) ∀x ∈ Rn .
y→x

A function f (x) is lower semi-continuous, when −f (x) is upper semi-contin-

uous. A function which is both lower and upper semi-continuous is called con-
tinuous. The function f in Figure 2.2 is continuous on (xA , xB )\{a, b, c, d, e}.
It is lower semi-continuous on (xA , xB )\{a, c, e} and upper semi-continuous
on (xA , xB )\{a, b, d}. Function values on discontinuities are depicted in Fig-
ure 2.2 as bullets (•). Left and right limits, which do not agree with the
function value, are depicted as circles (◦).
A stronger concept than continuity is absolute continuity:
Definition 2.10 (Absolute Continuity). A function f : I → Rn is ab-
solutely continuous on I ⊂ R if for every ε > 0 there exists a δ(ε) > 0 such
that
n
f (bi ) − f (ai ) < ε
i=1

for any n and any disjoint collection of intervals [ai , bi ] ∈ I satisfying


n
(bi − ai ) < δ.
i=1

Absolutely continuous functions have the property that they can be obtained
from integration of their derivative, which exists almost everywhere. A func-
tion f : Rn → Rn is said to satisfy a Lipschitz condition with constant K
provided that f is finite and satisfies

f (x) − f (y) ≤ Kx − y ∀x, y ∈ Rn . (2.5)

Every Lipschitz-continuous function f : I → Rn is absolutely continuous.

Let I be a real interval and X be a Euclidean space with the norm  · .
Definition 2.11 (Variation). Let f : I → X and let [a, b] be a subinterval
of I. The variation of f on [a, b] is the nonnegative extended real number


n
var(f , [a, b]) = sup f (xi ) − f (xi−1 ), (2.6)
i=1

where the supremum is taken over all strictly increasing finite sequences x1 <
x2 < . . . < xn of points on [a, b].
 2.2 Functions and Continuity 25

Fig. 2.3. The function f (x) and its associated variation function var(f, [a, x]) [63].

The variation of a constant function is therefore zero. Moreover, it holds that

A ⊂ B ⊂ I ⇒ var(f , A) ≤ var(f , B) (2.7)

and

var(f , [a, c]) = var(f , [a, b]) + var(f , [b, c]), ∀a ≤ b ≤ c ∈ I. (2.8)

Figure 2.3 illustrates the concept of the variation of a function f (x). The
associated variation function var(f, [a, x]) is an increasing function which has
kinks and discontinuities in its graph on the same location where f has kinks
and discontinuities.
Definition 2.12 (Locally Bounded Variation). The function f : I → X
is said to be of locally bounded variation, f ∈ lbv(I, X), if and only if

var(f , [a, b]) < ∞ (2.9)

for every compact subinterval [a, b] of I.

If f ∈ lbv(I, X), then we can define at each x ∈ I a right limit f + (x) and a
left limit f − (x) of f :

f + (x) = lim f (y), f − (x) = lim f (y). (2.10)

y↓x y↑x

If f (x) is locally continuous at x, then it holds that f (x) = f − (x) = f + (x).

The Cantor function [47,63] is an example of a function that is continuous and
of locally bounded variation, but is not absolutely continuous. It can be proven
that every absolutely continuous function is of locally bounded variation and
is differentiable almost everywhere (see [87], Theorem 2, p.337). To give an
example of a function which is continuous but not of locally bounded variation,
consider the function in Figure 2.4 (see [8])
26 2 Non-smooth Analysis

Fig. 2.4. Function (2.11) is continuous but not of bounded variation and not abso-
lutely continuous.

π
f (x) = x cos for x ∈ (0, 1] and f (0) = 0. (2.11)
2x
Clearly, the function f is continuous on [0, 1]. For the partition
1 1
x0 = 0, x1 = , x2 = , ..., x2n = 1
2n 2n − 1
one calculates

2n−1 
n
1
|f (xi+1 ) − f (xi )| = x1 + x1 + x3 + x3 + · · · + x2n−1 + x2n−1 =
i=0
k
k=1

∞
and hence the supremum over all possible partitions equals = ∞. The
1
k=1 k
function f (x) is therefore not of bounded variation on [0, 1] and therefore also
not absolutely continuous.

2.3 Generalised Derivatives

The classical derivative of smooth continuous functions will be extended in
this section to the generalised derivative (and differential) of Clarke for non-
smooth lower semi-continuous functions.
Consider a Lipschitz-continuous piecewise differentiable function f (x) with
a kink (i.e. non-smooth point) at one value of x, such as f (x) = |x| (Fig-
ure 2.5). The derivative f  (x) is defined by the tangent line to the graph of f
when the graph is smooth at x
 2.3 Generalised Derivatives 27

Fig. 2.5. Function (a), classical derivative (b) and generalised derivative (c).

df f (y) − f (x)
f  (x) = (x) = lim . (2.12)
dx y→x y−x
Although the function is not differentiable at every point x, it possesses at
each x a left and right derivative defined as

 f (y) − f (x)  f (y) − f (x)

f− (x) = lim , f+ (x) = lim . (2.13)
y↑x y−x y↓x y−x
The generalised derivative of f at x is declared as any value fq (x) included
between its left and right derivatives [37, 41]. Such an intermediate value can
be expressed as a convex combination of the left and right derivatives:

fq (x) = (1 − q)f−

 
(x) + qf+ (x), 0 ≤ q ≤ 1. (2.14)

Geometrically, a generalised derivative is the slope of any line which is tangent

to the graph at the point (x, f (x)). Such a line is necessarily between the left
and right tangent lines (drawn by dashed lines in Figure 2.5a). The set of all
the generalised derivatives of f at x, more generally the convex hull of the
derivative extremes, is called the generalised differential of f at x:
 
∂f (x) = co{f− (x), f+ (x)}
    (2.15)
= {fq (x) | fq (x) = (1 − q)f− (x) + qf+ (x), 0 ≤ q ≤ 1}.

The generalised differential of Clarke at x is the set of the slopes of all the lines
included in the cone bounded by the left and right tangent lines and is a closed
convex set (Figure 2.5b,c). In non-smooth analysis, the generalised differential
is for instance used to give a necessary condition for a local extremum of
f at x by 0 ∈ ∂f , which is the generalised form of f  (x) = 0 in smooth
analysis [37, 56].
Infinitely many directional derivatives exist for functions in Rn , whereas
only two directional derivatives exist for scalar functions (the left and right
derivative). For f : Rn → R, Lipschitz-continuous and differentiable almost
everywhere, the generalised differential of Clarke is defined as [37]

∂f (x) = co{ lim ∇f (xi ) : xi → x, ∇f (xi ) exists} ⊂ Rn , (2.16)

i→∞
28 2 Non-smooth Analysis

with the gradient

 T
∂f (x)
∇f (x) = ⊂ Rn . (2.17)
∂x
Equivalently, the generalised differential can be defined as [56]

∂f (x) = co{∇f (y) | y ∈ x + Bδ } ⊂ Rn . (2.18)
δ>0

The generalised differential (2.16) simplifies to (2.15) for the scalar case.
Note that f (x) can be convex or non-convex in the above definitions. For
continuous functions, the image of x under the generalised differential ∂f (·)
is always a closed convex set if it exists. If f (x) is a continuous convex function,
then the following relation holds

∂f (x) = {y | f (x∗ ) ≥ f (x) + y T (x∗ − x); ∀x∗ } ⊂ Rn . (2.19)

The generalised differential of a continuous convex function, defined by (2.19),

agrees with the subdifferential. The subdifferential is not restricted to con-
tinuous convex functions, but to lower semi-continuous convex functions
f (x) > −∞ with f (x) < ∞ for at least one x ∈ Rn .
Definition 2.13 (Subdifferential). For a lower semi-continuous convex
function f : Rn → (−∞, +∞], with the property

∃x ∈ Rn , f (x) < +∞,

we define the subdifferential of f at x ∈ Rn , denoted as ∂f (x), by

∂f (x) = {y | f (x∗ ) ≥ f (x) + y T (x∗ − x); ∀x∗ } ⊂ Rn .

2.4 Set-valued Functions

In this section we consider the main notions concerning set-valued functions
and their properties. We first define what we mean by a set-valued function.
Definition 2.14 (Set-valued Function). A set-valued function F : Rn →
Rn is a map that associates with any x ∈ Rn a set F (x) ⊂ Rn .
A set-valued function can therefore contain vertical segments on its graph

Graph(F ) = {(x, y) ∈ Rn × Rn | x ∈ Rn , y ∈ F (x)}. (2.20)

The image of x under F , being a set, is in general not closed or convex.

Convexity of a function should not be confused with convexity of the image
of x under F .
 2.4 Set-valued Functions 29

Definition 2.15 (Convex Image). A set-valued function F has a convex

image on X ⊂ Rn if the image of x under F is a convex set for all fixed
values x ∈ X.
We will reserve lowercase letters for single-valued functions, f (x) ∈ Rn , and
uppercase calligraphic letters or uppercase Greek letters for set-valued func-
tions, i.e. F (x) ⊂ Rn . The term multi-valued function or multifunction is
sometimes used instead of set-valued function [77].
We use the graph to define monotonicity of a set-valued function [12].
Definition 2.16 (Monotone Set-valued Function). A set-valued map
F (x) is called monotone if its graph is monotone in the sense that

∀(x, y) ∈ Graph(F ), ∀(x∗ , y ∗ ) ∈ Graph(F ), (y − y ∗ )T (x − x∗ ) ≥ 0.

In addition, if
(y − y ∗ )T (x − x∗ ) ≥ αx − x∗ 2
for some α > 0, then the set-valued map is strictly monotone.
For example, set-valued function
⎧
⎪
⎨−1 x<0
F(x) = {−1, +1} x=0 (2.21)
⎪
⎩
+1 x>0

is a monotone set-valued function, whereas F(x) + cx is strictly monotone for

c > 0. The subdifferential ∂f of a lower semi-continuous convex function f (x)
is monotone. Indeed, if y1 ∈ ∂f (x1 ) and y2 ∈ ∂f (x2 ), then the definition of
the subdifferential (Definition 2.13) gives f (x2 ) ≥ f (x1 ) + y1T (x2 − x1 ) and
f (x1 ) ≥ f (x2 ) + y2T (x1 − x2 ), which yields (y1 − y2 )T (x1 − x2 ) ≥ 0 after
substitution. The set-valued function
⎧
⎪
⎨−1 x<0
F(x) = ∂|x| = [−1, +1] x = 0 (2.22)
⎪
⎩
+1 x>0

is also monotone, but is even maximal monotone.

Definition 2.17 (Maximal Monotone Set-valued Function). A mono-
tone set-valued function F (x) is called maximal monotone if there exists no
other monotone set-valued function whose graph strictly contains the graph of
F . If F is strictly monotone and maximal, then it is called strictly maximal
monotone.
A necessary and sufficient condition for a set-valued function F to be maximal
monotone is that the property

∀(x, y) ∈ Graph(F ), (y − y ∗ )T (x − x∗ ) ≥ 0 (2.23)

30 2 Non-smooth Analysis

(a) Unilateral primitive. (b) Sign function.

Fig. 2.6. Maximal monotone set-valued functions.

is equivalent to y ∗ ∈ F (x∗ ), see [12].

Two maximal monotone set-valued functions, which are depicted in Fig-
ure 2.6, play a very dominant role in non-smooth analysis. The first one is the
unilateral primitive [63]
⎧
⎪
⎨0 x > 0,
Upr(x) = ∂ΨR+ = (−∞, 0] x = 0, (2.24)
⎪
⎩
∅ x < 0,

which is the subdifferential of the indicator function ΨR+ on the set R+ which
will be defined in Section 2.5. The second maximal monotone set-valued, which
is of prime importance, is the set-valued Sign-function
⎧
⎪
⎨−1 x < 0,
Sign(x) = ∂|x| = [−1, +1] x = 0, (2.25)
⎪
⎩
+1 x > 0,

which we already met in (2.22).

It follows from Definition 2.17 that if F is maximal monotone, then the
image of x under F is closed and convex for each x ∈ Rn . It can be shown,
that the subdifferential ∂f of a lower semi-continuous convex function f (x) is
not only monotone but also maximal monotone (see Corollary 31.5.1 in [147]).
Furthermore, the closure of the effective domain dom(F ) = {x | F (x) = ∅} of
a maximal monotone set-valued function is a convex set (Theorem 2.2 in [22]).

Proposition 2.18 (Convexity of the effective domain [22]). If F (x) is

a maximal monotone set-valued function, then dom(F ) is convex.

Proof: Consider x ∈ Rn and take xλ and yλ = x−xλ

λ such that yλ ∈ F (xλ )
for λ > 0. The monotonicity of F requires that
 2.4 Set-valued Functions 31

(yλ − y ∗ ) (xλ − x∗ ) ≥ 0
T

for all (x∗ , y ∗ ) ∈ Graph(F ). Hence, it must hold that


λ yλT (xλ − x∗ ) − y ∗ T (xλ − x∗ ) ≥ 0
x−xλ
or, using yλ = λ ,

−(xλ − x)T (xλ − x∗ ) − λy ∗ T (xλ − x∗ ) ≥ 0,

which simplifies to
xλ 2 ≤ xT (xλ − x∗ ) + xT ∗
λ x − λy
∗T
(xλ − x∗ ).
It holds that xλ remains bounded for λ → 0 and that dom F is dense in
Rn [22]. Consider a sequence λn → 0 such that xλn → x0 . Using this sequence,
we obtain the inequality
x0 2 ≤ xT (x0 − x∗ ) + xT
0x
∗
∀x∗ ∈ dom(F ).
In particular, for x∗1 , x∗2 ∈ dom(F ) it holds that
x0 2 ≤ xT (x0 − x∗1 ) + xT ∗
0 x1
∗ T ∗
x0  ≤ x (x0 − x2 ) + x0 x2
2 T

A convex combination x∗α = αx∗1 + (1 − α)x∗2 for all α ∈ [0, 1] gives

x0 2 ≤ xT (x0 − x∗α ) + xT ∗
0 xα

which is equivalent to x0 2 ≤ xT (x0 − x∗α ) + xT ∗ ∗

0 xα for all xα ∈ C :=
co dom(F ), in which we exploited the fact that the inequality contains its
boundary. A further reformulation gives (x−x0 )T (x∗α −x0 ) ≤ 0 for all x∗α ∈ C,
from which we deduce that x0 = proxC (x) (see Definition 2.31). It therefore
follows that xλ → proxC (x) for λ ↓ 0. Taking arbitrary values of x ∈ Rn , we
immediately see that xλ ∈ dom(F ) can attain every point in C. Consequently,
it must hold that C = dom(F ). 
The notion of semi-continuity of functions can be extended to set-valued
functions [37].
Definition 2.19 (Upper Semi-continuity of Set-valued Functions).
A set-valued function F (x) is upper semi-continuous in x if
 
lim sup inf a − b → 0.
y→x a∈F (y) b∈F (x)

This condition is equivalent to the condition that for all ε > 0 there exists
a δ > 0 such that x − y < δ ⇒ F (y) ⊂ F (x) + Bε . If a function F is
set-valued at a distinct x, then the graph of the function in a neighbourhood
around x is connected to the set F (x). Upper semi-continuity does not imply
convexity of the image.
32 2 Non-smooth Analysis

Fig. 2.7. Upper semi-continuity and lower semi-continuity of a set-valued func-

tion [12].

Fig. 2.8. Upper semi-continuity, convexity and closedness of a set-valued function.

Definition 2.20 (Lower Semi-continuity of Set-valued Functions).

A set-valued function F (x) is lower semi-continuous in x if
 
lim inf sup a − b → 0.
y→x a∈F (y) b∈F (x)

This condition is equivalent to the condition that for all ε > 0 there exists a
δ > 0 such that x − y < δ ⇒ F (x) ⊂ F (y) + Bε . A finite-valued function
F (x), that is both upper and lower semi-continuous is continuous. Upper and
lower semi-continuity of set-valued functions are sometimes called outer and
inner semi-continuity [77].
The difference between upper and lower semi-continuity is illustrated in
Figure 2.7 and the notions of upper semi-continuity, convexity and closedness
of (the images of) set-valued functions are illustrated in Figure 2.8. Left and
right limits, which do not agree with the function value, are depicted as circles
(◦) in Figure 2.8.
 2.5 Definitions from Convex Analysis 33

Fig. 2.9. Normal cone at different points of a convex set [57].

2.5 Definitions from Convex Analysis

In this section, the basic concepts of normal cone, indicator function, support
function, proximal point and distance are introduced. We will use the notion
of the subdifferential to reveal the relation between the indicator function and
the normal cone.
A cone is a subset of Rn consisting of rays (half lines emanating from the
origin).
Definition 2.21 (Cone). A subset C ⊂ Rn is called a cone if for any x ∈ C
and λ > 0 also λx ∈ C holds. A cone is convex if the subset C ⊂ Rn is convex.
A vector y is a normal vector of C at x if y does not make an acute angle
with every line segment in C starting from x.
Definition 2.22 (Normal Vector). Let C ⊂ Rn be a convex set and x ∈ C.
A vector y ∈ Rn is a normal vector of C at x ∈ Rn if

y T (x∗ − x) ≤ 0, x ∈ C, ∀x∗ ∈ C.

The normal cone of a set C at x is the set of rays that are normals of C at x
(Figure 2.9).
34 2 Non-smooth Analysis

Fig. 2.10. Tangent cone and contingent cone to a non-convex set C [63].

Definition 2.23 (Normal Cone). Let C ⊂ Rn be a convex set and x ∈ C.

The set of vectors y ∈ Rn that are normal vectors of C at x ∈ C form the
normal cone of C in x

NC (x) = {y | y T (x∗ − x) ≤ 0, x ∈ C, ∀x∗ ∈ C}.

If x is in the interior of C then NC (x) = 0. If x ∈

/ C then NC (x) = ∅.
Definition 2.24 (Polar Cone). Let K ⊂ Rn be a convex cone. The set of
normal vectors y to 0 with respect to K form the polar cone of K

K ◦ = {y | xT y ≤ 0, ∀x ∈ K}.

It therefore holds that K ◦ = NK (0).

There exist many ways in which one can express the concept of tangency
to a set C at a point x ∈ C. This leads to various cones such as the tangent
cone and contingent1 cone [12]. The contingent cone is the cone which touches
the boundary of C with its sides (see Figure 2.10).
Definition 2.25 (Contingent Cone [148]). Let C be a closed set and let
x be a point in C. The contingent cone is defined as

KC (x) = {y | ∃tk ↓ 0, yk → y, with x + tk yk ∈ C}.

Loosely speaking, the contingent cone is the cone which one obtains if one
zooms in on the point x after a shift of the set C such that x is at the origin.
The contingent cone, which is also called the ‘tangent cone’ of Bouligand, is
closed but not necessarily convex. If C is convex, then the contingent cone
agrees with the tangent cone.
1
from the Latin contingere, which means to touch on all sides.
 2.5 Definitions from Convex Analysis 35

Definition 2.26 (Tangent Cone [148]). Let C be a closed set and let x be
a point in C. The tangent cone is defined as

TC (x) = {y | ∀tk ↓ 0, xk → x with xk ∈ C, ∃yk → y, with x + tk yk ∈ C}.

The tangent cone, which is sometimes called the circatangent cone or tangent
cone of Clarke [12], is a closed convex cone and a subset of the contingent
cone, i.e. TC (x) ⊂ KC (x). The normal cone to an arbitrary set C, which is
not necessarily convex, can be defined as NC (x) = TC◦ (x).
One often needs to distinguish between points which belong to a set and
points which are outside. A useful tool to describe this distinction is the
indicator function.
Definition 2.27 (Indicator Function). Let C be a set. The indicator func-
tion of C is defined as

0, x ∈ C,
ΨC (x) =
+∞, x ∈ / C.

The indicator function is a convex function if C is convex. With the Defi-

nition 2.13 of the subdifferential and the indicator function it follows for a
convex set C that
∂ΨC (x) = {y | ΨC (x∗ ) ≥ ΨC (x) + y T (x∗ − x), x ∈ C, ∀x∗ }
(2.26)
= {y | 0 ≥ y T (x∗ − x), x ∈ C, ∀x∗ ∈ C}.

This is exactly the definition of the normal cone at C. The subdifferential of

the indicator function at x ∈ C is therefore the normal cone of C at x,

∂ΨC (x) = NC (x). (2.27)

Definition 2.28 (Conjugate Function). Let f be a convex function. The

function f ∗ is called the conjugate function of f and is defined as

f ∗ (x∗ ) = sup{xT x∗ − f (x)}.

x

∗
The conjugate f of a convex function is again convex. If we take the conjugate
of the conjugate function f ∗ , then we retrieve the original function.

Theorem 2.29 (Fenchel-Moreau). If f is a lower semi-continuous convex

function, then it holds that f ∗∗ = f .

Proof: See Theorem 11.1 in [149] and Theorem 4.4.2 in [11]. 

It therefore holds that

f (x) = sup{xT x∗ − f ∗ (x∗ )}. (2.28)

x∗
36 2 Non-smooth Analysis

Fig. 2.11. Indicator function and support function [57].

From (2.28) follows Fenchel’s inequality

xT x∗ ≤ f (x) + f ∗ (x∗ ). (2.29)

The equality holds when x∗ is a subgradient of f at x, i.e.

xT x∗ = f (x) + f ∗ (x∗ ) ⇐⇒ x∗ ∈ ∂f (x) ⇐⇒ x ∈ ∂f ∗ (x∗ ). (2.30)

Of special interest is the conjugate of the indicator function, which is called

the support function.
Definition 2.30 (Support Function). Let C be a convex set. The conju-
gate function of the indicator function ΨC is called support function,

ΨC∗ (x∗ ) = sup{xT x∗ − ΨC (x)}

x
= sup{xT x∗ | x ∈ C}.
x

The support function is positively homogeneous in the sense that

ΨC∗ (ax∗ ) = aΨC∗ (x∗ ) ∀a > 0. (2.31)

Using (2.30), we see that if x ∈ ∂ΨC∗ (x∗ ) then it holds that

xT x∗ = ΨC (x) + ΨC∗ (x∗ ). (2.32)

However, x ∈ ∂ΨC∗ (x∗ ) implies x ∈ C and ΨC (x) = 0. Hence, equation (2.32)

simplifies to
xT x∗ = ΨC∗ (x∗ ). (2.33)
Moreover, applying (2.30) on the function ΨC (x) with x∗ = 0, we obtain

0 ∈ ∂ΨC (x) ⇐⇒ x ∈ ∂ΨC∗ (0) (2.34)

 2.5 Definitions from Convex Analysis 37

Fig. 2.12. Normal cone, proximal point and distance to a convex set, z1 , z2 ∈ C,
z3 ∈ C.

or
0 ∈ NC (x) ⇐⇒ x ∈ ∂ΨC∗ (0) (2.35)
and therefore
x ∈ C ⇐⇒ x ∈ ∂ΨC∗ (0). (2.36)
Consequently, it follows that ∂ΨC∗ (0) = C. The support function ΨC∗ (x∗ ) is
a convex function with ΨC∗ (0) = 0. Hence, if 0 ∈ C, then 0 ∈ ∂ΨC∗ (0) from
which follows that ΨC∗ (x∗ ) attains a minimum at x∗ = 0, i.e.

0 ∈ C =⇒ ΨC∗ (x∗ ) ≥ 0. (2.37)

Moreover, if 0 ∈ int C then it follows that ΨC∗ (x∗ ) attains a global minimum
at x∗ = 0, i.e.
0 ∈ int C =⇒ ΨC∗ (x∗ ) > 0 ∀x∗ = 0. (2.38)

Definition 2.31 (Proximal Point). The proximal point of a closed convex

set C to a point z is the closest point in C to z:

proxC (z) = argmin z − x∗ , z ∈ Rn . (2.39)

x∗ ∈C
38 2 Non-smooth Analysis

Let x = proxC (z) then it holds that x ∈ C and


x = z, if z ∈ C,
(2.40)
x ∈ bdry C, if z ∈/ C,

where bdry C denotes the boundary of C. The vector z − proxC (z) is an

element of the normal cone of C at the proximal point:
1
x = proxC (z) ⇔ x = argmin z − x∗ 2
x∗ ∈C 2
1
⇔ x = argmin z − x∗ 2 + ΨC (x∗ ) (2.41)
x ∗ 2
⇔ z − x ∈ ∂ΨC (x)
⇔ z − x ∈ NC (x).

In the above derivation, the convex constrained minimisation problem is trans-

formed into a convex unconstrained minimisation problem by adding the
indicator function ΨC (x∗ ) to the cost function. Subsequently, the station-
arity condition is derived by making use of the subdifferential of the indi-
cator function which equals the normal cone. Note that the solution of a
convex minimisation problem is unique and the implication therefore holds
in both directions. Furthermore, by substituting z = x − ry, and using
−y ∈ NC (x) ⇔ −ry ∈ NC (x) for r > 0 we obtain

x = proxC (x − ry), r > 0 ⇐⇒ −y ∈ NC (x). (2.42)

Definition 2.32 (Distance). The distance from a point z ∈ Rn to a closed

convex set C is the Euclidian distance from the point to its proximal point in
C
distC (z) = z − proxC (z). (2.43)

It therefore holds that proxC (x) = x and distC (x) = 0 for all x ∈ C.

Proposition 2.33. Let C be a closed convex set. It holds that

1
∇ dist2C (z) = z − proxC (z). (2.44)
2
Proof: Consider the convex function f (z) = 12 dist2C (z) = 12 z − proxC (z)2 .
Definition 2.13 of the subdifferential applied on f (z) yields the condition

∂f (z)T (z ∗ − z) ≤ f (z ∗ ) − f (z), ∀z ∗ . (2.45)

Moreover, we deduce from Definition 2.31 the result

z ∗ − proxC (z ∗ ) = min z ∗ − z 0  ≤ z ∗ − z 0 , z 0 ∈ C. (2.46)

z ∈C
0
 2.6 Subderivative 39

We now take any point z such that z 0 = proxC (z) ∈ C in (2.46) and substitute
the result in (2.45)
1 ∗ 1
∂f (z)T (z ∗ − z) ≤ z − proxC (z ∗ )2 − z − proxC (z)2
2 2
1 ∗ 1
≤ z − proxC (z) − z − proxC (z)2
2
2 2 (2.47)
1 ∗ 1
= z − z + z − proxC (z)2 − z − proxC (z)2
2 2
1 ∗
= z − z + (z − z) (z − proxC (z)), ∀z ∗ .
2 ∗ T
2
Dividing by z ∗ − z for z ∗ = z yields
z∗ − z 1 ∗ ∗
T (z − z)
∂f (z)T ≤ z −z+(z−prox (z)) , ∀z ∗ = z. (2.48)
z ∗ − z 2 C
z ∗ − z
n→∞
Consider an arbitrary sequence zn∗ with zn∗ − z −−−−→ 0 and denote the
following limit as
(zn∗ − z)
e := lim ∗ − z
. (2.49)
n→∞ zn

Applying the sequence zn∗ to (2.48) and taking the limit n → ∞ yields
∂f (z)T e ≤ (z − proxC (z))T e, ∀e = 0. (2.50)
Consequently, it must hold that ∂f (z) = z − proxC (z). 
The concepts of proximal point, distance and normal cone are illustrated in
Figure 2.12.

2.6 Subderivative
The differentiability of a function f : Rn → R at a point x is con-
nected with the existence of a tangent hyperplane to the graph of f at the
point (x, f (x)) [148]. The concept of differentiability can be generalised by
considering the contingent cone (Definition 2.25) to the epigraph of f in-
stead. In this section, we consider a lower semi-continuous extended function
f : Rn → R ∪ {−∞, ∞} whose domain dom(f ) = {x ∈ Rn | f (x = ±∞} is
non-empty (i.e. the function is not trivial). The epigraph of the function f is
closed, because f is lower semi-continuous. Various generalised notions of gra-
dients exist, but the subderivative is the most natural object to focus on and is
often called the contingent epiderivative [12] or epicontingent derivative [11].
Definition 2.34 (Subderivative [149]). We define the function
f (x + tv  ) − f (x)
df (x)(v) = lim inf

t↓0, v →v t
as the subderivative of f at x in the direction v.
40 2 Non-smooth Analysis

Fig. 2.13. Subderivative of f and its relation to the contingent cone.

The epigraph of df (x)(·), as a function of the second argument, is the contin-

gent cone at the epigraph of f at (x, f (x))

epi df (x)(·) = Kepi f ((x, f (x)) . (2.51)

We observe that df (x)(0) = 0. If f is differentiable at x, then it holds that

df (x)(v) = −df (x)(−v) = (∇f (x))T v which agrees with the Lie derivative
Lv f (x). The epigraph of the subderivative is a cone (the contingent cone) and
the subderivative df (x)(·) is therefore positively homogeneous

df (x)(av) = a df (x)(v), ∀a ≥ 0. (2.52)

If the function f is convex, then we can express the subdifferential (Defini-

tion 2.13) as
∂f (x) = {y | df (x)(v) ≥ v T y}. (2.53)
Of special interest is the subderivative of an indicator function ΨC (x),

dΨC (x)(v) = ΨKC (x) (v), (2.54)

where KC (x) is the contingent cone to C at the point x.

2.7 Summary
An overwhelming amount of definitions and terminology has been introduced
in this chapter. Yet, virtually all concepts of non-smooth analysis which have
briefly been exposed here will be used in the following chapters. Set-valued
functions will be used in Chapter 4 to construct (measure) differential inclu-
sions. A maximal monotonicity property of (measure) differential inclusion
 2.7 Summary 41

will appear to lead to a contracting behaviour in those systems. Set-valued

force laws to describe frictional unilateral contact will be cast in Chapter 5
as an inclusion to a normal cone on a convex set of admissible contact forces.
The normal cone formulation of set-valued force laws brings the constitutive
description in relation with non-smooth potential theory, in which indicator
functions, conjugate functions and subdifferentials play an important role.
Moreover, non-smooth potentials and their associated mathematical appara-
tus will be used in Chapters 6 and 7 to set up Lyapunov functions to prove
stability properties of measure differential inclusions.
This page intentionally blank
3
Measure and Integration Theory

The theory of measures forms the basis for the modern integration theory and
both are quintessential for the understanding of measure differential inclusions
which will be dealt with in Chapter 4. Moreau, who introduced the notion
measure differential inclusions in [117–120, 122, 123], directly introduces the
so-called differential measure. In Chapter 4, it will be explained in more detail
that a measure differential inclusion actually describes how the differential
measure of the state relates to the state and time in analogy with the fact
that a (first-order) differential inclusion (or equation) describes how the time
derivative of the state depends on the state and time. In this chapter we give
a brief overview of measure and integration theory and relate the differential
measure used by Moreau to real measures and (signed) measures. The theory
presented in this chapter is based on standard textbooks about measure and
integration theory [47,150], various publications of Moreau and others [63,116,
122,124] as well as the work of P. Ballard communicated through lecture notes
of a summer course on Non-smooth Dynamical Systems (2003, Praz-sur-Arly).
This chapter may be very demanding and a reader on the verge of despair
should keep the key result of this chapter in mind: every function of locally
bounded variation can be decomposed in an absolutely continuous function, a
step function and a singular function. As will appear in Chapter 4, solutions of
measure differential inclusions are defined as functions of locally bounded vari-
ation. The latter fact clearly implies that the solutions of measure differential
inclusions may include the contribution of step functions, which can account
for jumps in the state evolution (such as e.g. velocity jumps in mechanical
systems with impacts).

3.1 Measures
The theory of measures associates with subsets A, B . . . of a set M nonnega-
tive numbers μ(A), μ(B), . . . We can think of μ(A) as the ‘volume’ or ‘mass’
44 3 Measure and Integration Theory

of A and μ(A) is called the measure of A. We now recall what we understand

by a σ-algebra and we define a measure μ on a σ-algebra A:
Definition 3.1 (σ-Algebra). Let M be a set. A non-empty system A of
subsets of M is called a σ-algebra if and only if:
1. M ∈ A and ∅ ∈ A,
2. A, B ∈ A implies A ∪ B ∈ A, A ∩ B ∈ A and A\B ∈ A,
∞ ∞
3. A1 , A2 · · · ∈ A implies ∪ Ak ∈ A and ∩ Ak ∈ A.
k=1 k=1

Definition 3.2 (Measure). Let A be a σ-algebra. A map μ : A → [0, ∞],

which maps each A ∈ A to a non-negative number μ(A), such that μ(∅) = 0,
is called a measure μ if

 ∪ B) 
1. μ(A = μ(A) + μ(B) if A and B are disjoint subsets, i.e. A ∩ B = ∅.
∞

∞
2. μ ∪ Ak = μ(Ak ) if A1 , A2 . . . are disjoint.
k=1 k=1

This definition allows μ(A) = ∞ for some A ∈ A. For a ∈ R we define: −∞ <

a < ∞, ∞ + a = ∞ and ∞ − a = ∞. An element {a} for which μ({a}) > 0
is called an atom1 of μ. A measure μ without any atoms is called diffuse.
The sets in A, for which we can define a measure μ, are called measurable.
Moreover, we call (μ, M, A) a measure space. A measurable set M is called
a (μ-)null-set if μ(M ) = 0. A measure μ is called finite if μ(M ) < ∞ for all
M ∈ A and is called σ-finite iffor all X ∈ A there exists a sequence of disjoint
∞
sets An ∈ A such that X = n=1 An and μ(An ) < ∞. In the following, we
will mostly consider measures on subsets of the real line R.
The most elementary measure is the Lebesgue measure on R, which assigns
the value b − a to the set (a, b] , i.e. μ((a, b]) = b − a for b ≥ a. The Lebesgue
measure is defined for half-open intervals, because every half-open interval
(a, c] can be split in two other half-open intervals (a, b] and (b, c], a ≤ b ≤ c,
such that
μ((a, c]) = μ((a, b]) + μ((b, c]), (3.1)
but the same holds of course for open and closed intervals because the
Lebesgue measure of a single element is zero, i.e. μ({a}) = 0 (i.e. the Lebesgue
measure is diffuse). A set M , with μ(M ) = 0 and therefore a null-set with
respect to the Lebesgue measure, is called Lebesgue negligible.
Of much importance are measures which measure the variation of functions
on intervals. Let f : R → R be a nondecreasing function of locally bounded
variation, i.e. f : lbv(I, R) with I ⊂ R. Then f has an at most countable2
number of discontinuities on I and for each t ∈ I we can define f + (t) and f − (t)
(see (2.10)). Let A be a σ-algebra on I. We now define the Lebesgue-Stieltjes
measure of f as the measure μf : A → [0, ∞] for which holds that
1
The word atom comes from the Greek άτ oμoν, which means ‘that which is indi-
visible/uncuttable’.
2
A set M is at most countable if there exists an injective map f : M → N.
 3.2 The Lebesgue Integral 45

Fig. 3.1. Step function.

μf ([a, b]) = f + (b) − f − (a),

μf ((a, b]) = f + (b) − f + (a),
(3.2)
μf ([a, b)) = f − (b) − f − (a),
μf ((a, b)) = f − (b) − f + (a),

with [a, b] ⊂ I. If f is discontinuous at a, then μf has an atom at {a}, i.e.

μf ({a}) = f + (a) − f − (a) > 0. (3.3)

The Lebesgue measure is a special case of the Lebesgue-Stieltjes measure μf

with f (x) = x and is, as mentioned above, a diffuse measure.

3.2 The Lebesgue Integral

The characteristic function χ : I → R of a set A ⊂ I is defined as

1 x ∈ A,
χA (x) = (3.4)
0 x ∈ I\A.

A function u : I → R, which takes a finite number of values on the domain I

is called a step function. Let u : I → R+ be a non-negative step function such
that

m
u(x) = αj χAj (x), (3.5)
j=1

with αj ≥ 0 and Aj ∈ A for j = 1 . . . m (see Figure 3.1).

The Lebesgue integral of u over I is defined as
 m
u dμ := αj μ(Aj ), (3.6)
I j=1

where μ(Aj ) denotes the Lebesgue measure of the set Aj . Every function
of bounded variation can be approximated by a step function. Let f + ∈
46 3 Measure and Integration Theory

lbv(I, R+ ) be a non-negative function of bounded variation, such that it is

the uniform limit of non-negative step functions, i.e. un ↑ f + for n → ∞. We
define the Lebesgue integral of f + over I as
 
+
f dμ := lim un dμ. (3.7)
I n→∞ I

Using the decomposition f = f + − f − with f + = max(f, 0) and f − =

max(−f, 0) we define the Lebesgue integral of an arbitrary function of lo-
cally bounded variation f ∈ lbv(I, R) over I as
  
f dμ := f dμ − f − dμ.
+
(3.8)
I I I

The function f is called quasi-integrable

 if f is measurable and if at least one
of the integrals I f + dμ and I f − dμ is finite. The Lebesgue integral (3.8) can
be applied to arbitrary measures μ. If we plug the Lebesgue-Stieltjes measure
μf in the integral (3.8), then we obtain the Lebesgue-Stieltjes integral.

3.3 Signed Measures

Signed measures, as opposed to measures, can also take negative values.

Definition 3.3 (Signed Measure). Let A be a σ-algebra. A map ν : A →
(−∞, ∞], which maps each A ∈ A to a real number ν(A), such that ν(∅) = 0,
is called a signed measure ν if

 ∪ B) 
1. ν(A = ν(A) + ν(B) if A and B are disjoint subsets, i.e. A ∩ B = ∅.
∞

∞
2. ν ∪ Ak = ν(Ak ) if A1 , A2 . . . are disjoint.
k=1 k=1

For instance, let ν : A → (−∞, ∞] be a signed measure, which maps intervals

to a real number and let [a, c] ∈ A with a ≤ b ≤ c. Then the application of
a signed measure ν on an interval can be decomposed in the application of ν
on disjoint subsets:

ν([a, b]) = ν({a} + ν((a, b)) + ν({b}),

ν((a, c)) = ν((a, b)) + ν({b}) + ν((b, c)),
(3.9)
ν((a, b]) = ν([a, b]) − ν({a}),
ν([a, b)) = ν([a, b]) − ν({b}).

If ν : A → (−∞, ∞] is a signed measure, then the set P ∈ A is called

ν-positive, if ν(A) ≥ 0 for all A ∈ A with A ⊂ P . Similarly, the set N ∈ A is
called ν-negative, if ν(A) ≤ 0 for all A ∈ A with A ⊂ N . The set Q ∈ A is
called a ν-null-set if ν(A) = 0 for all A ⊂ Q. The disjoint decomposition of ν
 3.3 Signed Measures 47

into a ν-positive and a ν-negative set (called Hahn decomposition) is unique

apart from a ν-null-set. The measures ν + , ν − : A → R, with R := R ∪ {+∞},
ν + (A) := ν(A ∩ P ), ν − (A) := −ν(A ∩ N ) (3.10)
are called the positive and negative variation of ν and lead to the measure
|ν| : A → R
|ν|(A) := ν + (A) + ν − (A), (3.11)
which is called the variation of ν. At least one of the measures ν + and ν − is
finite and it holds that
ν(A) = ν + (A) − ν − (A). (3.12)
Every function f ∈ lbv(I, R) of locally bounded variation equals the dif-
ference of two nondecreasing functions fp and fn such that
f = fp − fn . (3.13)
Using the Lebesgue-Stieltjes measures μfp and μfn , we can therefore define
the signed Lebesgue-Stieltjes measure νf as
νf (A) = μfp (A) − μfn (A), (3.14)

which gives the Hahn decomposition νf+ (A) = μfp (A) and νf− (A) = μfn (A).
Consequently, the relations (3.2) also hold for the signed Lebesgue-Stieltjes
measure:
νf ([a, b]) = f + (b) − f − (a),
νf ((a, b]) = f + (b) − f + (a),
(3.15)
νf ([a, b)) = f − (b) − f − (a),
νf ((a, b)) = f − (b) − f + (a).
If μ : A → [0, ∞] is a measure and f : X → R is quasiintegrable, then

ν(A) := f dμ (3.16)
A

is a signed measure, which we call the signed measure with density f with
respect to μ denoted by ν = f  μ. Signed measures which obey a density
with respect to some measure play an important role in integration theory.
We now introduce the notions of continuity and orthogonality (singularity) of
measures and come to the theorem of Radon-Nykodým:
Definition 3.4 (μ-Continuity). Let μ and ν be signed measures on A. The
signed measure ν is called μ-continuous, or absolutely continuous with respect
to μ, if every μ-null-set is a ν-null-set, which we denote by ν  μ.
Definition 3.5 (Orthogonality). Two (signed) measures ν, ρ : A → R
are called orthogonal (or mutually singular), denoted ν ⊥ ρ, if there exists a
decomposition X = A ∪ B, A ∩ B = ∅, A, B ∈ A, such that A is a ν-null-set
and B is a ρ-null-set.
48 3 Measure and Integration Theory

The theorem of Radon-Nykodým states that if ν is absolutely continuous with

respect to μ, then there must exist a density function f such that (3.16) holds.
Theorem 3.6 (Radon-Nykodým). Let μ be a σ-finite measure and ν  μ
be a signed measure on A. Then it holds that ν has a density function with
respect to μ, i.e. there exists a quasiintegrable function f : I → R such that
ν = f  μ.
Proof: See Elstrodt [47], p. 279. 
If μ and ν are two finite (signed) measures on A, then ν is not necessarily
μ-continuous, i.e. ν does not necessarily have a density with respect to μ. Now
we can try to separate from ν a part f  μ, such that ν − f  μ is as ‘small’
as possible.
Let μ be a finite measure and ν be a finite signed measure on A. Then
there exists a unique decomposition

ν = ν1 + ν2 , (3.17)

into two signed measures ν1 and ν2 , such that ν1  μ and ν2 ⊥ μ. This

decomposition is called the Lebesgue decomposition. The signed measure ν1
is therefore μ-continuous and admits a density function f with respect to μ,
i.e. ν1 = f  μ. We now use the Lebesgue decomposition to decompose the
signed Lebesgue-Stieltjes measure νf of a function f , being of locally bounded
variation.

Theorem 3.7 (Lebesgue Decomposition of the Lebesgue-Stieltjes

Measure). For ever function f ∈ lbv(I, R) there exists a decomposition

f = fabs + fs + fsing (3.18)

in functions of locally bounded variation and a decomposition of the signed

Lebesgue-Stieltjes measure

νf = νabs + νs + νsing (3.19)

in signed measures such that

1. fabs is absolutely continuous with respect to the Lebesgue measure μ:

νabs  μ, νabs = fabs  μ. (3.20)

The signed measure νabs is diffuse.

2. fs is a step function, which is constant almost everywhere and νs is a
purely atomic signed measure.
3. fsing is singular, i.e. νsing is a diffuse measure which is orthogonal to the
Lebesgue measure μ, i.e. νsing ⊥ μ. fsing is continuous and constant almost
everywhere.
 3.4 Real Measures 49

Fig. 3.2. Decomposition of f in fabs + fs if fsing = 0.

Proof: We decompose f in a continuous part fc = fabs + fsing and a step

function fs . Subsequently, we use the Lebesgue decomposition to decompose
fc in an absolutely continuous part and a singular part such that νabs  μ
and νsing ⊥ μ. 
The jumps in f are therefore solely described by the step function fs

νs ({a}) = f + (a) − f − (a) = νf ({a}) (3.21)

and f  = fabs

almost everywhere (see Figure 3.2). The Cantor function [47,63]
is an example of a function which is purely singular.

3.4 Real Measures

Measures and signed measures have been introduced above as mappings from
sets to the real numbers. We now introduce the so-called real measure which
maps a function to a real number.
We denote by C 0 (I, R) the vector space of continuous functions on the real
closed interval I, equipped with the norm of uniform convergence (supremum
norm)
gC 0 = sup |g(t)|. (3.22)
t∈I

Definition 3.8 (Real Measure). We call any linear form υ : C 0 (I, R) → R

which is continuous, i.e.

∃M ∈ R ∀g ∈ C 0 (I, R) |υ(g)| ≤ M sup |g(t)|, (3.23)

t∈I

a real measure on I.
The linearity of the linear form υ implies that υ(g − h) = υ(g) − υ(h) for all
g, h ∈ C 0 (I, R). Condition (3.23) states that

∃L ∈ R ∀g, h ∈ C 0 (I, R) |υ(g) − υ(h)| ≤ L sup |g(t) − h(t)|, (3.24)

t∈I
50 3 Measure and Integration Theory

which is a Lipschitz continuity condition. A real measure is therefore a Lip-

schitz continuous linear functional on the vector space C 0 (I, R). We use the
following notation 
υ(g) =: g υ, (3.25)
I
which involves artificially an integral sign and will be related in the next
section to a true Lebesgue integral if υ is a differential measure. A real measure
is said to be positive if υ(g) ≥ 0 for all g ∈ C 0 (I, R+ ). For any function
g ∈ C 0 (I, R) we define its positive part g + (t) = max(g(t), 0) and negative
part g − (t) = max(−g(t), 0), with which we obtain the relations

g = g+ − g− , |g| = g + + g − . (3.26)

Any function g(t) can therefore be decomposed into two nonnegative functions
g + (t) and g − (t). We now define the modulus measure |υ| for nonnegative
functions h(t) ∈ C 0 (I, R+ ):

|υ|(h) := sup f υ. (3.27)
f ∈C 0 (I,R) I
|f (t)|≤h(t)

It holds that |υ|(g + ) ≥ υ(g + ) and |υ|(g − ) ≥ υ(g − ). For instance, if υ = dt,
then |υ|(g ± ) = υ(g ± ) and if υ = −dt, then |υ|(g ± ) = υ(−g ± ). The modulus
measure |υ| for arbitrary functions g(t) ∈ C 0 (I, R) is defined as

|υ|(g) := |υ|(g + ) − |υ|(g − ). (3.28)

The modulus measure |υ| is a positive real measure. If υ = dt, then |υ|(g) =
υ(g) and if υ = −dt, then |υ|(g) = υ(−g). For a real measure υ we define the
positive real measures
1 1
υ+ = (|υ| + υ), υ− =
(|υ| − υ). (3.29)
2 2
 
We now define the linear operators (a,b) υ and [a,b] υ on a real measure
υ. Using the decomposition

υ = υ+ − υ− , (3.30)

which is known as the Jordan decomposition, we only need to define  these lin-
ear operators for positive real measures. The main idea is to relate (a,b) υ and
 
[a,b]
υ to the already defined integral υ(g) =: I g υ through the characteristic
functions χ(a,b) and χ[a,b] respectively. Unfortunately, we are not able to ap-
ply a real measure υ on a characteristic function χA , because a characteristic
function is not continuous. We therefore need to approach χA from below if
A is an open interval or from above if A is a closed interval (see Figure 3.3).
Let υ be a positive real measure; then, we define
 3.4 Real Measures 51

Fig. 3.3. Supremum for g(t) ≤ χ(a,b) (t) and infinum for g(t) ≥ χ[a,b] (t) with g ∈
C 0 (I, R).

 
υ := sup υ(g), υ := inf υ(g). (3.31)
(a,b) g∈C 0 (I,R) [a,b] g∈C 0 (I,R)
g(t)≤χ(a,b) (t) ∀t g(t)≥χ[a,b] (t) ∀t

If υ is an arbitrary real measure, then the Jordan decomposition (3.30) yields

  
υ= υ+ − υ− (3.32)
A A A

for any interval A ∈ A, which forms a signed measure ν : A → R


ν(A) := υ. (3.33)
A

Using the relations (3.9) for the decomposition of a signed measure on an

interval in the sum of signed measures on disjoint subsets it follows that
   
υ= υ+ υ+ υ
[a,b] {a} (a,b) {b}
   
υ= υ+ υ+ υ
(a,c) (a,b) {b} (b,c)
   (3.34)
υ= υ− υ
(a,b] [a,b] {a}
  
υ= υ− υ.
[a,b) [a,b] {b}

If in turn the signed measure ν, which is associated with the real measure υ,
has a density function f with respect to a measure μ, i.e. ν = f  μ, then the
real measure υ is associated with a measure μ
 
υ =: ν(A) := f dμ. (3.35)
A A
52 3 Measure and Integration Theory

Consequently, if ν = f  μ then we can write

υ = f dμ (3.36)

and we call such a real measure υ a differential measure with a density f with
respect to μ.

3.5 Differential Measures

Let f : I → Rn be a vector function of locally bounded variation, f ∈
lbv(I, Rn ) (i.e. each fi ∈ lbv(I, R)), and g : I → R be a continuous function,
g ∈ C 0 (I, R), with a compact support on the real interval I. Furthermore, let
[a, b] be a compact subinterval of I and S be the set of finite subsets of [a, b].
Consider a discretisation Sm ∈ S of [a, b] with m + 1 points

Sm : a = t0 < t1 < t2 . . . tm−1 < tm = b. (3.37)

Between every two neighbouring points ti−1 and ti (1 ≤ i ≤ m) of Sm

we choose an intercalator ϑSi m ∈ [ti−1 , ti ]. We now construct the Riemann-
Stieltjes sum

m
H(Sm , ϑSm , g, f ) = g(ϑSi m ) (f (ti ) − f (ti−1 )) . (3.38)
i=1

It has been proven in [122] that the mapping Sm → H(Sm , ϑSm , g, f ) con-
verges to a limit
lim H(Sm , ϑSm , g, f ), (3.39)
m→∞,Sm ∈S

which is independent on the choice of the intercalator ϑSm . This limit will be
denoted by 
g df := lim H(Sm , ϑSm , g, f ), (3.40)
[a,b] m→∞,Sm ∈S

which is the Riemann-Stieltjes integral of g with respect to df . The set of

discontinuity points of f is Lebesgue negligible. The Riemann-Stieltjes integral
on a bounded domain I agrees therefore with the Lebesgue-Stieltjes integral.
The linear mapping 
df : g → g df (3.41)

constitutes a real vector measure df : C 0 (I, R) → Rn on I. We call df the

differential measure (or Stieltjes measure) of f . For example, if f (t) = t,
then the differential measure df equals the (differential) Lebesgue measure
dt. Moreover, if f is the right-continuous unit step function

0 t < ti ,
f (t) = (3.42)
1 t ≥ ti ,
 3.5 Differential Measures 53

then the differential measure df equals the differential Dirac point measure
dδti , which is such that

γ(t)dδti = γ(ti ), for ti ∈ [a, b], (3.43)
[a,b]

or  
1 ti ∈ [a, b],
δti ([a, b]) = dδti = (3.44)
[a,b] 0 ti ∈
/ [a, b].
Let fs ∈ lbv(I, Rn ) be a right-continuous step function with p discontinu-
ities on the nodes τp with

a < τ1 < τ2 . . . τp < b. (3.45)

The step function fs (τ ) is constant for t = τj . Now we choose the points

ti of S ∈ S such that every open interval (ti−1 , ti ) contains at most one
node τj and every node falls in an open interval (ti−1 , ti ). As intercalator
we choose ϑ̄ = ti . The only non-zero terms in the Riemann-Stieltjes inte-
gral (3.38) originate from intervals (ti−1 , ti ) which contain a node τj and yield
g(ti ) (fs (ti ) − fs (ti−1 )). In the limit, when ti − ti−1 tends to zero, this term
tends to g(τj ) (fs+ (τj ) − fs− (τj )). The integral (3.40) is in this case
 
p
 
g dfs = g(τj ) fs+ (τj ) − fs− (τj ) . (3.46)
[a,b] j=1

Consequently, the differential measure dfs is the weighted sum of a number

of Dirac point measures

p
 
dfs = fs+ (τj ) − fs− (τj ) dδτj . (3.47)
j=1

If g(t) = 1, then the integral (3.40) is given by

  
p
 
dfs = fs+ (τj ) − fs− (τj ) dδτj
[a,b] [a,b] j=1


p 
 
= fs+ (τj ) − fs− (τj ) dδτj
j=1 [a,b] (3.48)
p
 
= fs+ (τj ) − fs− (τj )
j=1

= fs+ (τp ) − fs− (τ1 ) = fs+ (b) − fs− (a).

Every right-continuous locally bounded function f can be approximated by
a right-continuous local step function fs and their difference can be made
arbitrarily small by refining the partition of the local step function. The result
54 3 Measure and Integration Theory

df = f + (b) − f − (a) (3.49)
[a,b]

does therefore not only hold for local step functions, but holds for arbitrary
locally bounded functions f as has been proven rigourously in [122]. We can
shrink the interval [a, b] to a singleton {a} such that

df = f + (a) − f − (a), (3.50)
{a}

which yields a non-zero result if the differential measure df has an atom at

{a}. Note that there is no one-to-one relationship between f and df : the
differential measure df does not depend on the value f (a) but only on its
right and left value f ± (a). Using the integral relations (3.34) it follows for
f ∈ lbv(I, Rn ) that

df = f − (b) − f + (a),
(a,b)

df = f + (b) − f + (a), (3.51)
(a,b]

df = f − (b) − f − (a).
[a,b)

Hence, the integral of the differential measure df is precisely the signed

Lebesgue-Stieltjes vector measure νf of f

νf (A) = df . (3.52)
A

Using the Lebesgue decomposition of the Lebesgue-Stieltjes measure (The-

orem 3.7), we decompose the right-continuous function f ∈ lbv(I, Rn ) in
functions of locally bounded variation such that

f (t) = fabs (t) + fs (t) + fsing (t) (3.53)

and decompose the differential measure correspondingly

df = dfabs + dfs + dfsing . (3.54)

The function fabs is absolutely continuous with respect to the Lebesgue


measure and admits the density function fabs . Its differential measure dfabs

is the product of the density function fabs and the Lebesgue measure dt

dfabs = fabs dt, (3.55)

such that   b

dfabs = fabs dt = fabs (b) − fabs (a). (3.56)
[a,b] a
 3.6 The Differential Measure of a Bilinear Form 55

The right-continuous step function fs takes the discontinuities in f into ac-

count and is almost everywhere constant. The corresponding signed Lebesgue-
Stieltjes measure νs is purely atomic. The corresponding differential measure
dfs equals the weighted sum of Dirac point measures dδti , defined by (3.43),
with jump heights f + (ti ) − f − (ti )

dfs = (f + (ti ) − f − (ti ))dδti . (3.57)
i

For general step functions fs (t), we say that fs admits a density function fη
with respect to the differential measure dη (which we call the atomic measure)
and, consequently,  
dfs = fη dη. (3.58)

Now, using (3.57), the meaning of the atomic measure dη becomes clear (in
an integral sense) from
  
−
(f (ti ) − f (ti ))dδti =
+
fη dη. (3.59)
[a,b] i [a,b]

More conveniently, it is proposed in [63] to write (3.57) as


dfs = (f + − f − )dη, dη = dδti (3.60)
i

and dfs is called the atomic differential measure of f . Here, we will use the
notation (3.60), keeping in mind that differential measures always need to be
understood in an integral sense.
The singular term fsing is continuous (but not absolutely continuous) and
constant for almost all t. The corresponding signed Lebesgue-Stieltjes measure
νsing is non-atomic (diffuse) and is orthogonal to the Lebesgue measure. Its
differential measure dfsing is therefore said to be singular with respect to the
Lebesgue measure dt.
The Lebesgue decomposition (3.53) is possible for right-continuous func-
tions which are of locally bounded variation. A function f ∈ lbv(I, Rn ) which
is not defined on its discontinuity points can be decomposed in an abso-
lutely continuous function, a step function and a singular function for al-
most all t. The decomposition of measures (3.54) remains valid as measures
have to be understood in the sense of integration. Moreover, if the function
f ∈ lbv(I, Rn ) is not right-continuous, then we can still apply the Lebesgue
decomposition (3.53) on the function f + ∈ lbv(I, Rn ), which is again right-
continuous.

3.6 The Differential Measure of a Bilinear Form

In this section, which is based primarily on [122], we study the differential
measure of a bilinear form. For instance, in mechanics the kinetic energy
56 3 Measure and Integration Theory

T (u) = F (u, u), with F (u, v) = 12 uT M v, is a symmetric bilinear form in the

generalised velocities.
Consider the bounded variation functions x ∈ lbv(I, X) and y ∈ lbv(I, Y )
and the function t → F (x(t), y(t)), being a continuous bilinear form F :
X × Y → R, denoted by F (x, y) in short. First assume that x and y are
local step functions, each having their own set of discontinuity points. The
set of discontinuity points of F (x, y) is the union of the discontinuity points
of x and of y. Construct a sequence of nodes t1 < tt < . . . < tn on the
discontinuity points of F (x, y) on a subinterval [a, b] of I. The functions x(t),
y(t) and F (x, y) are therefore constant on each open subinterval (ti , ti+1 ).
The differential measure dF (x, y) equals the sum of a locally finite collection
of point measures placed at the discontinuity points of F
 
n
 
dF (x, y) = F (x+ (ti ), y + (ti )) − F (x− (ti ), y − (ti )) . (3.61)
[a,b] i=1

Similarly, dx equals the sum of point measures placed at the nodes ti with
values x+ (ti ) − x− (ti ) (and the same applies for dy). It therefore holds that
 
n
−
F (dx, y ) = F (x+ (ti ) − x− (ti ), y − (ti )),
[a,b]
 i=1

n (3.62)
−
+
F (x , dy) = F (x (ti ), y (ti ) − y (ti )).
+ +
[a,b] i=1

Exploiting the bilinearity of F yields

 
 
dF (x, y) = F (dx, y − ) + F (x+ , dy) . (3.63)
[a,b] [a,b]

Using again the fact that every locally bounded function can be approximated
by a local step function we conclude that equation (3.63) holds for arbitrary
locally bounded functions x and y (see [122]).
Consider now a symmetric quadratic form G(x) = F (x, x) = xT Ax, with
A = AT ∈ Rn×n . We deduce from (3.63) that
 
 
dG(x) = F (dx, x− ) + F (x+ , dx)
[a,b] [a,b]

= F (x+ + x− , dx) (3.64)
[a,b]

= (x+ + x− )T Adx
[a,b]

or simply dG = (x+ + x− )T Adx.

 3.7 Summary 57

3.7 Summary
An overview of measure and integration theory has been given in this chapter.
A measure assigns a non-negative real number to a set, while signed measures
can also take negative values. If a signed measure is absolutely continuous with
respect to a measure, then it admits a density function with respect to that
measure (Radon-Nykodým). A real measure is a continuous linear functional
which assigns a real number to a continuous function and is related to a signed
measure through the characteristic function. The differential measure is a real
measure, which assigns a real number to a continuous function by means of
the limit of Riemann-Stieltjes sums.
Of utmost importance for the next chapters is the fact that every (right-
continuous) function of locally bounded variation can be decomposed in an ab-
solutely continuous function, a (right-continuous) step function and a singular
function (3.53)– (3.60). The differential measures can be split correspondingly.
The singular function is assumed to vanish.
Differential measures will be used in Chapter 4 to formulate measure dif-
ferential inclusions, which are generalisations of differential inclusions and
equations. The solution concept of measure differential inclusions allows the
solution to be a function of locally bounded variation, and therefore to undergo
discontinuities in time (such as e.g. velocity jumps in mechanical systems with
impacts).
This page intentionally blank
4
Non-smooth Dynamical Systems

A dynamical system is a system whose state evolves with time. The evolution
is governed by a set of rules, and is usually put in the form of equations.
Continuous-time systems are dynamical systems of which the state is allowed
to change (continuously or discontinuously) at all times t. Continuous-time
systems are usually described by ordinary differential equations. Discrete-time
systems are dynamical systems of which the state can only change at discrete
time-instances t1 < t2 < t3 . . . . In this chapter we will consider continuous-
time dynamical systems with a non-smooth (and possibly discontinuous) time-
evolution of the state.
After a brief introduction to differential equations, we will consider dif-
ferential equations with a discontinuous right-hand side. The requirement for
the existence of a solution leads to the need to fill in the graph of the discon-
tinuous right-hand side (Filippov’s convex method). The resulting set-valued
right-hand side brings forth a differential inclusion, which will be discussed
in detail in Section 4.2. Subsequently, the differential measure of the state,
which classically contains a density with respect to the Lebesgue measure,
is extended in Section 4.3 with an atomic part. This leads to a measure dif-
ferential inclusion, being the mathematical framework used in this work to
describe dynamical systems with state evolutions with discontinuities (state
jumps). In Chapters 6 and 8, we will present stability and convergence results
for measure differential inclusions.

4.1 Differential Equations

We consider a dynamical system described by a set of ordinary differential
equations
ẋ(t) = f (t, x(t)), x(t) ∈ Rn , (4.1)
where x is the state vector and f (t, x(t)) is the right-hand side vector, called
the vector field, describing the time derivative of the state vector. We will
60 4 Non-smooth Dynamical Systems

assume that f (t, x) is linearly bounded [37], i.e. there exist positive constants
γ and c such that
f (t, x) ≤ γx + c, ∀ (t, x). (4.2)
If the vector field is smooth (see Definition 2.7), then a solution x(t) of the
system (4.1) exists for any given initial condition and is unique for all t ∈ R,
i.e. solutions of (4.1) are defined globally (in time). In fact, smoothness of
the vector field is not a necessary condition for existence and uniqueness
of the solution, as can be concluded from the following theorem (see [37],
theorem 1.1, page 178):
Theorem 4.1 (Existence and Uniqueness for Continuous Systems).
Suppose that f (t, x) is continuous, and let (t0 , x0 ) ∈ R × Rn be given. Then
the following holds:
1. There exists a solution of system (4.1) on an open interval (t0 − δ, t0 + δ),
for δ > 0, satisfying x(t0 ) = x0 .
2. If in addition we assume that f (t, x) is linearly bounded, so that (4.2)
holds, then there exists a solution of system (4.1) on (−∞, ∞) such that
x(t0 ) = x0 .
3. We now add the hypothesis that f (t, x) is locally Lipschitz at x, i.e. there
exists a constant L(x) > 0 and r > 0 such that

f (t, x1 ) − f (t, x2 ) ≤ L(x)x1 − x2 , ∀x1 , x2 ∈ Br + x.

Then there exists a unique solution of (4.1) on (−∞, ∞) such that x(t0 ) =
x0 .
Note that above theorem deals with systems with a continuous vector field
f (t, x) but the vector field is allowed to be non-smooth, i.e. the right-hand
side f may not be differentiable everywhere with respect to x.

4.2 Differential Inclusions

In this section we introduce the concept of a solution to a differential inclusion

ẋ(t) ∈ F (t, x(t)), (4.3)

where F : R × Rn → Rn is a set-valued map.

Existence and uniqueness of solutions of continuous dynamical systems
have been discussed in Section 4.1. However, differential equations stemming
from systems with switches may be discontinuous, i.e. the right-hand side
f can be piecewise continuous such that it is discontinuous in x on certain
hyper-surfaces in the state-space. The theory of Filippov [50, 51, 96, 152] gives
a generalised definition of the solution of differential equations which incorpo-
rates systems with a discontinuous right-hand side. The solution x(t) in the
sense of Filippov of a differential equation with a discontinuous right-hand side
 4.2 Differential Inclusions 61

(also called Filippov systems) is absolutely continuous in time, i.e. there are
no discontinuities in the state x(t). Systems with discontinuities in the state
x(t) can be described by measure differential inclusions (see Section 4.3).
Filippov’s theory will be briefly outlined in this section.
In order to make things as clear as possible, we first look at a very simple
one-dimensional example (see [88]). Consider the following differential equa-
tion with a discontinuous right-hand side
⎧
⎪
⎨ 3, x < 0,
ẋ = f (x) = 1 − 2 sign(x) = 1, x = 0, (4.4)
⎪
⎩
−1, x > 0,

with ⎧
⎪
⎨−1, x < 0,
sign(x) = 0, x = 0, (4.5)
⎪
⎩
1, x > 0.
For a given initial condition x(0) = 0 we can obtain a solution of the initial
value problem 
3t + C1 , x < 0,
x(t) = (4.6)
−t + C2 , x > 0,
with constants C1 and C2 being determined by the initial condition. Each
solution reaches x = 0 in finite time. If the solution arrives at x = 0, it can
not leave x = 0, because ẋ > 0 for x < 0 and ẋ < 0 for x > 0. The solution
will therefore stay at x = 0, which implies ẋ(t) = 0. Note that x(t) = 0 with
ẋ(t) = 0 is not a solution of the problem since 0 = 1 − 2 sign(0). Hence, a
natural idea to extend the notion of solution is to replace the right-hand side
f (x) by a set-valued function F(x) such that F(x) = f (x) for all x for which
f is continuous in x. At the points for which f is discontinuous in x a suitable
choice of F(x) is required. The differential equation is then replaced by the
differential inclusion [50, 51]
ẋ ∈ F(x), (4.7)
with the set-valued function

F(x) = 1 − 2 Sign(x). (4.8)

Here, Sign(x) denotes the set-valued Sign function (2.25), which is set-valued
at x = 0, and which is the generalised differential of |x|. With this definition
x(t) = 0 is a unique solution in forward time of the differential inclusion

ẋ ∈ 1 − 2 Sign(x), (4.9)

with initial condition x(0) = 0. Note that the solution of (4.9) is non-unique
in backward time. For instance, the solutions of (4.9) with initial condition
62 4 Non-smooth Dynamical Systems

x(−1) = 1 and initial condition x(−1) = −3 evolve both to x(0) = 0. In

fact, there exist infinitely many initial conditions at t = −1 which evolve to
x(0) = 0. Knowledge about the solution for t ∈ (−∞, 0) is therefore lost by
the non-uniqueness in backward time.
Let us also give a one-dimensional example of non-uniqueness in forward
time. Consider the one-dimensional differential inclusion
ẋ(t) ∈ Sign(x(t)), x(0) = 0. (4.10)
This initial value problem has three solutions on the domain [0, ∞):
x1 (t) = −t, x2 (t) = 0, x3 (t) = t. (4.11)
The solution of (4.10) is non-unique in forward time but unique in backward
time, i.e. x(t) = 0 for t ∈ (−∞, 0].
The above examples are one-dimensional. We now need to define a differ-
ential equation with a discontinuous right-hand side in a more general sense
for any dimension n. We restrict ourselves to differential equations with a
right-hand side that is discontinuous at one hyper-surface, but this restriction
is not essential. The state-space Rn is split into two subspaces V− and V+
by a hyper-surface Σ such that Rn = V− ∪ Σ ∪ V+ . The hyper-surface Σ is
called the switching boundary and is defined by a scalar switching boundary
function h(x). The state x is in Σ when
h(x) = 0 ⇐⇒ x ∈ Σ. (4.12)
The subspaces V− and V+ and switching boundary Σ can be formulated as
V− = {x ∈ Rn | h(x) < 0},
Σ = {x ∈ Rn | h(x) = 0}, (4.13)
V+ = {x ∈ Rn | h(x) > 0}.
The right-hand side of the dynamics ẋ = f (t, x) is assumed to be locally
continuous, smooth and linearly bounded for all x ∈ Σ. From this assumption
it follows that the solution x(t) within each subspace V− and V+ exists and
is unique (see Theorem 4.1).
The set-valued extension of f (t, x) of (4.1) for x ∈ Σ is given by the closed
convex hull of all the limits
F (t, x) = co{y ∈ Rn | y = lim f (t, x̃), x̃ ∈ Rn \Σ}, (4.14)
x̃→x

where the notation (2.1) has been used. All the limits exist because f (t, x) is
assumed to be locally continuous, smooth and linearly bounded for all x ∈ Σ.
We are now able to consider the following n-dimensional nonlinear system
with discontinuous right-hand side

f− (t, x(t)), x ∈ V− ,
ẋ(t) = f (t, x(t)) = (4.15)
f+ (t, x(t)), x ∈ V+ ,
 4.2 Differential Inclusions 63

with the initial condition x(0) = x0 . As mentioned before, the right-hand

side f (t, x) is assumed to be piecewise continuous such that it is smooth on
V− and V+ and not yet defined on Σ. We extend the function f− (t, x) in a
continuous way to V− ∪ Σ and extend f+ (t, x) continuously to V+ ∪ Σ. It is
not required that f− (t, x) and f+ (t, x) agree on Σ. Subsequently, the graph
is filled-in between the values of f− (t, x) and f− (t, x) on Σ which gives the
set-valued extension F (t, x)
⎧
⎪
⎨f− (t, x(t)), x ∈ V− ,
ẋ(t) ∈ F (t, x(t)) = co{f− (t, x(t)), f+ (t, x(t))}, x ∈ Σ, (4.16)
⎪
⎩
f+ (t, x(t)), x ∈ V+ .

The convex set with the two right-hand sides f− and f+ can be cast in the
form
co{f− , f+ } = {(1 − q)f− + qf+ , ∀ q ∈ [0, 1]}. (4.17)
The extension (or convexification) of a discontinuous system (4.15) into a
differential inclusion (4.16) with convex image is known as Filippov’s convex
method.
It was stated that the set-valued extension F of f should be suitable. If
the discontinuous system (4.15) is a mathematical model of a physical sys-
tem, then we are interested in a solution concept that guarantees existence of
solutions. Therefore, for practical reasons we demand that the choice for F
guarantees existence of solutions. Existence can be guaranteed with the notion
of upper semi-continuity of set-valued functions. We first define a solution of
the differential inclusion (4.3).
Definition 4.2 (Solution of a Differential Inclusion). An absolutely con-
tinuous function x : [0, τ ] → Rn is said to be a solution of the differential
inclusion (4.3) if it fulfils
ẋ(t) ∈ F (t, x(t))
for almost all 1
t ∈ [0, τ ].
The following existence theorem is proven in [10] (Theorem 3, page 98; see
also [156]):
Theorem 4.3 (Existence of Solution of a Differential Inclusion). Let
F be a set-valued function. We assume that F is upper semi-continuous and
that the image of (t, x) under F is closed, convex and bounded for all t ∈ R
and x ∈ Rn . Then, for each x0 ∈ Rn there exists a τ > 0 and an absolutely
continuous function x(t) defined on [0, τ ], which is a solution of the initial
value problem
ẋ ∈ F (t, x(t)), x(0) = x0 .
1
for almost all t means except for a set t of Lebesgue measure 0.
64 4 Non-smooth Dynamical Systems

The theorem clearly holds at values of x for which F (t, x) is locally single-
valued and continuous, because of the boundedness restriction (see Theo-
rem 4.1). To illustrate the theorem for set-valued F (t, x) we once more look
at the example of (4.9)

ẋ ∈ 1 − a Sign(x), with |a| > 1. (4.18)

To allow for the solution x(t) = 0 we must demand that 0 ∈ 1 − a Sign(0).

In order to guarantee the existence of a solution we therefore have to define
Sign(0) to be the set [−1, 1]. This set is upper semi-continuous with the values
of sign(x) = ±1 for x = 0, and the set is closed, convex and bounded.
Filippov’s convex method together with the above existence theorem de-
fines the solution in the sense of Filippov for a discontinuous differential equa-
tion.
Definition 4.4 (Solution in the Sense of Filippov of a Discontinuous
Differential Equation). An absolutely continuous function x : [0, τ ] → Rn
is said to be a solution of ẋ(t) = f (t, x) (4.15) in the sense of Filippov if for
almost all t ∈ [0, τ ] it holds that

ẋ(t) ∈ F (t, x(t)),

where F (t, x(t)) is the closed convex hull of all the limits of f (t, x(t)) as
in (4.14).
Remark: If x(t) is in a region where the vector field is continuous, x(t) ∈ V,
then of course F (t, x(t)) = f (t, x(t)) must hold. If the solution x(t) slides
along a switching boundary, x(t) ∈ Σ, then ẋ(t) ∈ F (t, x(t)). However, ẋ(t) is
not defined at time instances tΣ where the solution x(t) arrives at a switching
boundary Σ or leaves from Σ (the solution x(t) arrives at or leaves from
Σ if there exists an arbitrary small ε > 0 and t∗ ∈ tΣ + Bε \0 such that
x(t∗ ) ∈ Σ and x(tΣ ) ∈ Σ). The set of t for which the solution x(t) arrives at
or leaves from Σ is of Lebesgue measure zero. Hence, the set of time-instances
t for which ẋ(t) is not defined is Lebesgue negligible. Note that the absolutely
t
continuous solution x(t) = x(t0 )+ t0 ẋdt does not depend on the value (or the
lack of having a value) of ẋ(t) on a Lebesgue negligible set of time-instances
t.
It was assumed that f (t, x) is linearly bounded for x ∈ / Σ. In addition,
F (t, x(t)) is assumed to be bounded at values (t, x) for which F is set-valued.
Consequently, F (t, x(t)) is linearly bounded, i.e. there exist positive constants
γ and c such that for all t ∈ [0, ∞) and x ∈ Rn it holds that:

F (t, x) ≤ γx + c.

Solutions x(t) to (4.16) therefore exist on [0, ∞) (see [10, 37]) but uniqueness
is not guaranteed.
 4.2 Differential Inclusions 65

In the following, it will sometimes be useful to explicitly write the depen-

dence of the solution on the initial condition. We therefore denote a solution
x(t) of ẋ(t) ∈ F (t, x(t)) starting from the initial condition x(t0 ) = x0 by
ϕ(t, t0 , x0 ). The solution ϕ(t, t0 , x0 ) is in general non-unique. We denote by
S(F , t0 , x0 ) the set of solutions curves x(t), with t ≥ t0 , starting from the
initial condition x(t0 ) = x0 , i.e.

ϕ(·, t0 , x0 ) ∈ S(F , t0 , x0 ).

If the solution is unique, then S(F , t0 , x0 ) reduces to a singleton {ϕ(·, t0 , x0 )}.

A sufficient criterion for uniqueness of solutions can be gained from a
maximal monotonicity property (Definition 2.17) of the system [22].
Theorem 4.5 (Uniqueness of Solutions through Maximal Monotonic-
ity). Consider a differential inclusion of the form

ẋ ∈ −A(x).

If A(x) is a maximal monotone set-valued function with a linear boundednness

condition, then the differential inclusion has a unique solution x(t) for all
initial conditions x(0) = x0 .
Proof: The maximal monotonicity implies that A(x) is upper semi-continuous
and that the image of x under A is closed, convex and bounded for all x ∈ Rn .
Existence of solutions therefore follows from a linear boundedness condition.
Moreover, because A(x) is monotone, it holds that

(A(u) − A(v))T (u − v) ≥ 0 ∀u, v. (4.19)

Let u(·) ∈ S(−A, 0, u0 ) and v(·) ∈ S(−A, 0, v0 ) be solutions of the differ-

ential inclusion. The proof of uniqueness follows from reductio ad absurdum.
Assume that the system has non-uniqueness of solutions. Then there exists a
u0 = v0 such that u(t) = v(t). Consider the positive definite function V ∈ R
1
V (t) = u(t) − v(t)2 . (4.20)
2
The time-derivative of V yields
d
V̇ = u(t) − v(t) u(t) − v(t)
dt
= (u(t) − v(t))T (u̇(t) − v̇(t)) (4.21)
∈ −(u(t) − v(t))T (A(u(t)) − A(v(t)))
≤ 0.

Hence, the function V can not increase, meaning that the distance between
u(t) and v(t) can not increase, i.e.
66 4 Non-smooth Dynamical Systems

(a) (b)

Fig. 4.1. Stick-slip system and friction curve with Stribeck effect.

u(t) − v(t) ≤ u0 − v0 . (4.22)

Taking u0 = v0 it follows that u(t) = v(t), which is in contradiction with

u(t) = v(t). Consequently, the system has uniqueness of solutions. 
Although mechanical systems with set-valued force laws will be described
in a much more general setting in Chapter 5, we will already discuss here a
small example to illustrate Filippov systems. Consider the horizontal motion
of rigid mass m attached to a vertical wall by a spring k (Figure 4.1a). The
mass is riding on a belt, that is moving at a constant velocity vdr . We denote
the horizontal position of the block by x and its velocity by ẋ. The relative
velocity of the mass m with respect to the belt is denoted by γT = ẋ − vdr .
Dry friction occurs between the mass and the belt, with a friction force FT .
The friction force FT (γT ) is in the slip phase a function of the relative velocity
γT .
μ
FT (γT ) = −μ(γT )λN sign(γT ), μ(γT ) = , γT = 0, (4.23)
1 + δ|γT |

where μ(γT ) is the velocity dependent friction coefficient, λN = mg is the

normal contact force. In the stick phase, the magnitude of the friction force is
limited by the maximum static friction force Fs = μs λN ≥ |FT |. The friction
curve is drawn in Figure 4.1b for δ > 0 and shows a Stribeck effect, i.e. the
magnitude of friction force diminishes for increasing relative velocity γT . We
now combine the description of the friction force in the slip phase, which is
described by a function FT (γT ), and the restriction on the friction force in
the stick phase |FT | ≤ Fs in one set-valued force law

FT ∈ −μ(γT )λN Sign(γT ), (4.24)

where Sign(x) is the set-valued sign function (2.25). Newton’s second law gives
the equation of motion
mẍ(t) + kx(t) = FT . (4.25)
 4.2 Differential Inclusions 67

Fig. 4.2. Phase portrait of the stick-slip system.

The equation of motion together with the set-valued force law for the friction
force FT gives a second-order differential inclusion
mẍ(t) + kx(t) ∈ −μ(ẋ(t) − vdr )mg Sign(ẋ(t) − vdr ), (4.26)
which holds for almost all t. This second-order
  differential inclusion can be
cast in first-order form using xT = x ẋ :
 
ẋ(t)
ẋ(t) ∈ F (x(t)) = . (4.27)
−mk
x(t) − μ(ẋ(t) − vdr ) g Sign(ẋ(t) − vdr )
The differential inclusion (4.26) or (4.27) has been obtained by considering a
set-valued force law (4.24) for the friction force. Alternatively, one can also
derive the differential inclusion (4.27) using ‘Filippov’s convex method’, i.e.
along the lines of equations (4.15) and (4.16). The right-hand side of the
system in first-order form switches between f− (x) and f+ (x) on a hyper-
plane Σ = {x ∈ Rn | h(x) = 0} with h(x) = γT = ẋ − vdr (see (4.13))
and  
ẋ
f− (x) = k , (4.28)
− x + μ(ẋ(t) − vdr ) g
m
 
ẋ
f+ (x) = k , (4.29)
− x − μ(ẋ(t) − vdr ) g
m
in the spaces V− = {x ∈ Rn | h(x) < 0} and V+ = {x ∈ Rn | h(x) > 0},
respectively. The convexification of the right-hand side yields the differential
inclusion in the form (4.16)
⎧
⎪
⎨f− (x(t)), x ∈ V− ,
ẋ(t) ∈ F (x(t)) = co{f− (x(t)), f+ (x(t))}, x ∈ Σ, (4.30)
⎪
⎩
f+ (x(t)), x ∈ V+ ,
68 4 Non-smooth Dynamical Systems

which is equivalent to (4.27). The forward and backward slip phases are the
subspaces V+ and V− respectively, and the stick phase is contained in Σ.
The phase portrait of this system is depicted in Figure 4.2 for the parameter
2
values m = 1 kg, k = 1 N/m, vdr = 0.2 m/s, μs = 0.1, g = 10 m/s , Fs = 1 N
and δ = 3 s/m. The equilibrium point is an unstable focus surrounded by a
stable limit cycle, which alternates between the backward slip phase V− and
the stick phase. This is the so-called stick-slip motion. We can observe that
the solution (x, ẋ) shows a kink in the phase portrait when the solution goes
from forward to backward slip (or vice versa), or when the solution enters the
stick phase. At those time-instances, the friction force FT jumps to another
value. The acceleration ẍ(t) is not defined at those time-instances for which
such a change occurs. This is the reason why the differential inclusion (4.26)
 T
(or equivalently (4.27) or (4.30)) only describes ẋ = ẋ ẍ for almost all t,
i.e. not for time instances for which the acceleration is not defined.

4.3 Measure Differential Inclusions

In this section we introduce the measure differential inclusion

dx ∈ dΓ (t, x(t)) (4.31)

as has been proposed by Moreau [117–120, 122, 123].

Differential inclusions, as studied in Section 4.2, are able to describe an
absolutely continuous time-evolution x(t) which is non-differentiable on an
at most countable number of time-instances. Roughly speaking, we can say
that the function x(t) is allowed to have a ‘kink’. Discontinuities in the time-
evolution can however not be described by a differential inclusion. In the
following, we will extend the concept of differential inclusions to measure dif-
ferential inclusions in order to allow for discontinuities in x(t). The assumption
of absolute continuity of x(t) will be relaxed to locally bounded variation in
time.
Consider x ∈ lbv(I, X ), i.e. x(t) is a function of locally bounded vari-
ation in time. The function x(t) is therefore admitted to undergo jumps at
discontinuity points ti ∈ I, i = 1, 2, . . . , but the discontinuity points are either
1. separated in time such that t1 < t2 < t3 . . . , leading to a finite number of
discontinuities on a compact time-interval, or,
2. accumulate to an accumulation point limi→∞ ti = t∗ , leading to infinitely
many discontinuities on a compact time-interval. The jump x+ (ti )−x− (ti )
is assumed to decrease to zero for i → ∞.
In the first case, each discontinuity point ti is followed by a non-zero time-
interval of absolutely continuous evolution. In the case of an accumulation
point, the time-interval ti+1 − ti tends to zero as well as the discontinuities
(jumps) in x(t), which guarantees that x(t) is of locally bounded variation. At
 4.3 Measure Differential Inclusions 69

a time-instance t, including the discontinuity points t = ti and accumulation

points, we can therefore define a right limit x+ (t) and a left limit x− (t) of x
as one of the properties of functions of bounded variation:

x+ (t) = lim x(t + τ ), x− (t) = lim x(t + τ ). (4.32)

τ ↓0 τ ↑0

If x is locally continuous at t and ẋ+ (t) = ẋ− (t), then x is locally differentiable
(or locally smooth) at t. A function x : I → Rn is said to be smooth if it
is locally smooth for all t ∈ I. A function is said to be almost everywhere
continuous, if the set D ⊂ I of discontinuity points ti ∈ D, k = 1, 2, . . .
is of measure zero, i.e. μ(D) = 0. Similarly, a continuous function can be
differentiable almost everywhere.
We want to describe with x(t) an evolution in time and therefore consider
x(t) to be the result of an integration process
 t
x(t) = x(t0 ) + dx, t ≥ t0 , (4.33)
t0

where we call dx the differential measure of x [122]. If x(t) is a smooth

function, then dx admits a density function, say xt (t), with respect to the
Lebesgue measure dt
dx = xt (t)dt. (4.34)
We usually immediately associate the density function xt (t) with the deriva-
tive ẋ(t) and write
 t
x(t) = x(t0 ) + ẋ(t)dt, t ≥ t0 . (4.35)
t0

Subsequently, we consider x(t) to be an absolutely continuous function,

which is non-differentiable at the set D ⊂ I of points ti ∈ D. The derivative
ẋ(t) does therefore not exist for t = ti . Lebesgue integration over a singleton
{ti }, i.e. an interval with zero length, results in zero

dx = 0, with dx = xt (t)dt, (4.36)
{ti }

even if ẋ(t) does not exist for t = ti . The derivative ẋ(t) exists almost every-
where because x(t) is absolutely continuous. We say that the set D of points ti
for which ẋ(t) does not exist is Lebesgue negligible. Lebesgue integration over
a Lebesgue negligible set results in zero. Consequently, (4.33) also holds for
absolutely continuous functions, which are non-differentiable at a Lebesgue
negligible set D of time-instances ti .
Theorem 4.6 (Lebesgue-Vitali). A function x : [tl , tk ] → Rn is absolutely
continuous if and only if it admits a representation as
70 4 Non-smooth Dynamical Systems
 t
x(t) = x(tl ) + ẋ(t)dt, t l ≤ t ≤ tk ,
tl

where ẋ is the derivative of x, which exists at almost all points of [tl , tk ] and
is Lebesgue integrable on the interval.
The proof can be found in [47].
Finally, we consider x(t) to be a function of bounded variation on the
interval I, which is discontinuous at the set D ⊂ I of points ti ∈ D. Moreover,
we assume that x(t) does not contain any singular terms, i.e. fractal-like
functions such as the Cantor function. Although the function x(t) does not
exist at the discontinuity points t = ti , it admits a right limit x+ (t) and left
limit x− (t) at every time-instance t. Just as before, we consider x(t) to be
the result of an integration process

x+ (t) = x− (t0 ) + dx, t ≥ t0 , (4.37)
[t0 ,t]

where the integration process takes the left limit x− (t0 ) of the initial value to
the right limit x+ (t) of the final value over the closed time-interval [t0 , t] =
{τ ∈ I|t0 ≤ τ ≤ t}. The differential measure dx does therefore not only
contain a density xt with respect to the Lebesgue measure dt but also contains
a density xη with respect to the atomic measure dη, which gives a nonzero
result when integrated over a singleton, such that

dx = xt (t)dt + xη (t)dη, (4.38)

with 
dη = 1, ti ∈ D. (4.39)
{ti }

As discussed in Section 3.5, the atomic measure dη can be interpreted as the

sum of Dirac point measures dδti ,
 
 1 ti ∈ [tl , tk ],
dη = dδti , dδti = (4.40)
i [tl ,tk ] 0 ti ∈
/ [tl , tk ],

for any interval [tl , tk ] ⊂ I. Integration of the differential measure dx over a

singleton {tk } yields

x+ (tk ) − x− (tk ) = dx
{tk }

= xη (t)dη (4.41)
{tk }

= xη (tk ) dη.
{tk }
 4.3 Measure Differential Inclusions 71

It follows that x+ (tk ) = x− (tk ) when tk ∈

/ D, which obviously must hold for
a locally continuous function at t = tk . Moreover, if we choose tk = ti ∈ D,
then we can immediately associate the density xη (t) with respect to the atomic
measure dη as the jump in x(t) at the discontinuity point ti :

xη (ti ) = x+ (ti ) − x− (ti ), ti ∈ D. (4.42)

This reveals the fact that the interpretation of the atomic measure dη as
in (4.40) is becoming meaningful when the discontinuity points ti are known
and these discontinuity points ti are exactly the time instances at which xη (t)
is non-zero. Considering (4.42), we therefore usually write the differential mea-
sure (4.38) as
dx = ẋ(t)dt + (x+ (t) − x− (t))dη. (4.43)
Consequently, using the differential measure (4.43) we are able to describe
a locally absolutely continuously varying time-evolution (using the Lebesgue
measurable part of dx) together with discontinuities at time-instances ti ∈ D
(using the atomic part). Integration of the differential measure dx therefore
gives the total increment over the interval under consideration

x+ (tk ) − x− (tl ) = dx, (4.44)
[tl ,tk ]

which reduces to 
−
x (tk ) − x (tk ) =
+
dx (4.45)
{tk }

for a singleton {tk }. In fact, it is also possible to take the singular part of
a function of bounded variation into account. We will assume, however, that
all singular parts vanish, as fractal-like solutions are not of interest in the
evolution problems we would like to describe. Such functions are called special
functions of bounded variation (SBV), see [6].
With the differential inclusion (4.3)

ẋ(t) ∈ F (t, x(t)),

in which F (t, x(t)) is a set-valued mapping, we are able to describe a non-

smooth absolutely continuous time-evolution x(t). The solution x(t) fulfills
the differential inclusion almost everywhere, because ẋ(t) does not exist on
a Lebesgue negligible set D of non-smooth time-instances ti ∈ D. Instead of
using the density ẋ(t), we can also write the differential inclusion using the
differential measure
dx ∈ F (t, x(t))dt, (4.46)
which yields a measure differential inclusion. The solution x(t) fulfills the
measure differential inclusion (4.46) for all t ∈ I because of the underlying
integration process associated with measures. Moreover, writing the dynamics
in terms of a measure differential inclusion allows us to study a larger class
72 4 Non-smooth Dynamical Systems

of functions x(t), as we can let dx contain parts other than the Lebesgue
integrable part. In order to describe a time-evolution of bounded variation
but discontinuous at isolated time-instances, we let the differential measure
dx also have an atomic part such as in (4.43) and therefore extend the measure
differential inclusion (4.46) with an atomic part as well

dx ∈ F (t, x(t))dt + G(t, x(t))dη. (4.47)

Here, G(t, x(t)) is a single-valued or set-valued mapping, which is in general

dependent on t, x− (t) and x+ (t). Following [122], we simply write G(t, x(t))
although a dependence on t, x− (t) and x+ (t) is meant. If G is only a function
of x+ (t), then we write G(x+ (t)). Furthermore, if G(t, x(t)) is set-valued,
then it is often assumed to take only conic values: λG(t, x(t)) ⊂ G(t, x(t)) for
all λ ∈ [0, ∞) and (t, x) ∈ I × X. More conveniently, we write the measure
differential inclusion (4.47) as

dx ∈ dΓ (t, x(t)), (4.48)

where dΓ (t, x(t)) is a set-valued measure function defined as

dΓ (t, x(t)) = F (t, x(t))dt + G(t, x(t))dη. (4.49)

The measure differential inclusion (4.48) has to be understood in the sense of

integration. A solution x(t) of (4.48) is a function of locally bounded variation
which fulfills 
+ −
x (t) = x (t0 ) + f (t, x)dt + g(t, x)dη, (4.50)
I
for every compact interval I = [t0 , t], where the functions f (t, x) and g(t, x)
have to obey
f (t, x) ∈ F (t, x), g(t, x) ∈ G(t, x). (4.51)
Note that a function x(t) can be undefined on its discontinuity points and still
satisfy (4.50) for all t ≥ 0, i.e. the function x(t) does not have to be right-
continuous. Substitution of (4.43) in the measure differential inclusion (4.48)
gives

ẋ(t)dt + (x+ (t) − x− (t))dη ∈ F (t, x(t))dt + G(t, x(t))dη, (4.52)

which we can separate in the Lebesgue integrable part and atomic part

ẋ(t)dt ∈ F (t, x(t))dt, (x+ (t) − x− (t))dη ∈ G(t, x(t))dη (4.53)

from which we can retrieve (4.3) and the jump condition

x+ (t) − x− (t) ∈ G(t, x(t)). (4.54)

The latter clearly reveals that state jumps are induced by non-zero g(t, x) ∈
G(t, x). Moreover, by considering the limits t ↓ ti and t ↑ ti we obtain the
differential inclusions for post- and pre-jump times
 4.3 Measure Differential Inclusions 73

ẋ+ (t) ∈ F (t, x+ (t)), ẋ− (t) ∈ F (t, x− (t)), (4.55)

which we call the directional differential inclusions.

A special class of systems is described by set-valued measure functions
dΓ (t, x(t)) for which each density function is a conic subset of Rn . In par-
ticular, the set-valued functions F (t, x(t)) and G(t, x(t)) are often equal to
the same cone K(t, x(t)), i.e. F (t, x(t)) = G(t, x(t)) = K(t, x(t)). Following
Moreau [122], we write in this case

dx ∈ K(t, x(t)), (4.56)

and refrain from prescribing a measure in the right-hand side in advance. It

is to be understood from (4.56) that, if dx possesses a density function fμ
with respect to the non-negative differential measure dμ, then this density
function belongs to the cone K, i.e. fμ (t, x) ∈ K(t, x). In particular, this
applies for the Lebesgue measure as well as for the atomic measure. More
explicitly, dx ∈ K(t, x(t)) implies

ẋ± (t) ∈ K(t, x± (t)) and x+ (t) − x− (t) ∈ K(t, x(t)). (4.57)

In Lagrangian dynamics, the measure differential inclusion (4.48) typically

describes the time-evolution of absolutely continuous generalised coordinates
q(t) and generalised velocities u(t), which are of locally bounded variation.
The set-valued measure function dΓ (t, x(t)) typically contains indicator func-
tions, which impose constraints on the system. Unilateral constraints g(q) ≥ 0
and bilateral constraints g(q) = 0 restrict the generalised coordinates q(t) to
an admissible set. Consider for instance a ball with mass m of which we mea-
sure its height above the ground with the coordinate q and its velocity with
u. The ground imposes a unilateral constraint q(t) ≥ 0 which restricts the
coordinate q to the set R+ . If the ball hits the ground, i.e. q(t) = 0, then a
contact impulse can cause a velocity jump in u(t). We consider the impact
law u+ (t) = −eu− (t), 0 ≤ e ≤ 1, where e is the restitution coefficient. The
unilateral constraint q(t) ≥ 0 does not only restrict the generalised coordi-
nate, but if q(t) = 0 then also the post-impact velocity u+ (t) is restricted to
u+ (t) ≥ 0. Moreover, if we initialise our model for the ball with admissible
initial conditions

u+ (t0 ) > 0 q(t0 ) = 0
q(t0 ) ≥ 0, +
u (t0 ) :
u+ (t0 ) arbitrary q(t0 ) > 0,

then the time-evolution q(t) remains admissible for all t > t0 . In this sense,
the system is consistent. However, we can also specify the left-limit u− (t0 ) of
the velocity as initial condition. In the latter case there is no restriction on
u− (t0 ), because a possible impulse will force u+ (t0 ) to be admissible, because
e ≥ 0. Hence, the system is not consistent for e < 0. From the above example
it becomes clear that the states x(t) of the measure differential inclusion (4.48)
74 4 Non-smooth Dynamical Systems

are restricted to an admissible set A. What A is depends on how the initial

condition is understood.
Following [122], we imagine that the investigated process is already in
progress before t0 . While the evolution for t ≥ t0 is the time-domain of pre-
diction, the initial data are understood to convey the abridged information
on the system history before t0 . We therefore interpret the initial condition
as the left-limit, i.e. x− (t0 ) = x0 . Using this interpretation of the initial
condition, we find that the admissible set in the bouncing ball example is
A = {(q, u) ∈ R × R | q ≥ 0}, which does not involve a restriction on the ve-
locities. The condition u+ (t) ≥ 0 for q(t) = 0 with t ≥ t0 , being the restriction
on the velocity, is fulfilled by consistency of the impact law (e ≥ 0).
For the initial value problem, consisting of the measure differential inclu-
sion (4.48) with initial condition x− (t0 ) = x0 , we define a solution as follows:

Definition 4.7 (Solution of a Measure Differential Inclusion). A so-

lution x(t) = ϕ(t, t0 , x0 ) of the measure differential inclusion (4.48) with the
initial condition x− (t0 ) = x0 is a function x : R → Rn , being of locally
bounded variation, which fulfills (4.48) for all t ≥ t0 and which is not defined
at its discontinuity points.

Remark that a solution x(t) is not defined on the discontinuity points ti , but
still fulfills the measure differential inclusion for all t ≥ t0 in the sense of (4.50).
From the physical point of view (or modelling point of view), there exists no
value that we can meaningfully assign to x(t) on its discontinuity points. It
therefore makes sense to leave the function x(t) undefined at its discontinuity
points. For this reason, we will consider solutions of (4.48) which are not
defined at their discontinuity points.
We often prefer to write ϕ(t, t0 , x0 ) := x(t) to explicitly state its depen-
dence on the initial condition x− (t0 ) = x0 . The solution starting from a
specific initial condition (t0 , x0 ) is generally not unique, i.e. there exists a set
S(dΓ , t0 , x0 ) of solution curves such that

ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ). (4.58)

If S(dΓ , t0 , x0 ) contains a single element, then the solution is unique. If the

set of solutions is empty, S(dΓ , t0 , x0 ) = ∅, then a solution does not exist.
A solution curve ϕ(t, t0 , x0 ) with x0 ∈ A is called ‘viable’ if it remains
within the set A for all future times t ≥ t0 for which it is defined. If all solution
curves of a measure differential inclusion are viable, and at least one solution
curve exists for each initial condition, then the system is called consistent.
Definition 4.8 (Consistency of a Measure Differential Inclusion). A
measure differential inclusion (4.48) is called consistent if

x0 ∈ A ⇒ ϕ(t, t0 , x0 ) ∈ A for almost all t ≥ t0

for each solution curve ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) = ∅.

 4.3 Measure Differential Inclusions 75

The admissible set A of a consistent measure differential inclusion is there-

fore a positively invariant set (see Definition 6.4) of the system. This puts a
tangency restriction on the differential measure dx on the set A. If the sys-
tem (4.48) is consistent, and if the solution x(t) is absolutely continuous, then
it must hold that
dx = ẋdt ∈ KA (x), (4.59)
where KA (x) is the contingent cone of A at the point x (Definition 2.25).
In the following chapters, we will assume that the considered systems are
consistent.
A set-valued measure function dΓ (x) = F (x)dt+G(x+ )dη is called (max-
imal) monotone if the set-valued density functions F (x) and G(x+ ) are (max-
imal) monotone. A sufficient criterion for uniqueness of solutions can again
be gained from a monotonicity property of the system (see Section 4.2 for the
specific case of differential inclusions).
Theorem 4.9 (Uniqueness of solutions through monotonicity). Con-
sider a measure differential inclusion of the form

dx ∈ −dA(x+ ).

Let the system be consistent and have existence of solutions for all initial
conditions x0 ∈ A. If dA(x+ ) is a monotone set-valued measure function,
then the measure differential inclusion has a unique solution x(t) for all initial
conditions x(0) = x0 ∈ A.
Proof: The proof is similar to the proof of Theorem 4.5. Let x1 (·) ∈
S(−dA, 0, x10 ) and x2 (·) ∈ S(−dA, 0, x20 ) be solutions of the measure dif-
ferential inclusion with x10 , x20 ∈ A. For the differential measure of dx we
write
dx = −at dt − aη dη, (4.60)
where the single-valued densities obey the set-valued force laws

−ẋ = at ∈ At (x), −(x+ − x− ) = aη ∈ Aη (x+ ). (4.61)

The set-valued operators At and Aη are maximal monotone operators and

are related to the set-valued measure function by

dA(x+ ) = At (x)dt + Aη (x+ )dη. (4.62)

Consider the positive definite function V ∈ R

1
V (t) = x2 (t) − x1 (t)2 . (4.63)
2

The differential measure dV of V has a density V̇ with respect to the Lebesgue

measure and a density V + − V − with respect to the atomic measure dη
76 4 Non-smooth Dynamical Systems

dV = V̇ dt + (V + − V − )dη. (4.64)

The Lebesgue measurable part yields

V̇ = (x2 − x1 )T (ẋ2 − ẋ1 )

= −(x2 − x1 )T (at2 − at1 ) (4.65)
≤ 0,

with at1 ∈ At (x1 ), at2 ∈ At (x2 ), due to the monotonicity of At (·). Evalua-
tion of the atomic part gives
1 +
V+−V− = (x + x− − T − −
2 − x1 − x1 ) (x2 − x2 − x1 + x1 )
+ + +
2 2
1
= − (2x+ 2 − 2x1 + aη2 − aη1 ) (aη2 − aη1 )
+ T
2 (4.66)
1
= −(x+ 2 − x1 ) (aη2 − aη1 ) − aη2 − aη1 
+ T 2
2
≤ 0,

with aη1 ∈ Aη (x+ 1 ), aη2 ∈ Aη (x2 ), in which we used the monotonicity of

+

Aη (·). Hence, the function V can not increase, meaning that the distance
between x1 (t) and x2 (t) can not increase, i.e. x+ 1 (t) − x2 (t) ≤ x10 − x20 
+

for all t ≥ t0 . Taking x10 = x20 it follows that x1 (t) = x2 (t) for almost
all t. Using the fact that x1 (t) and x2 (t) are of locally bounded variation, it
follows that x1 (t) and x2 (t) have the same discontinuity points. The functions
x1 (t) and x2 (t) are not defined at their discontinuity points. Consequently,
the trajectories x1 (t) and x2 (t) are identical, i.e. the system has uniqueness
of solutions. 

4.4 Summary
Differential equations with a non-smooth continuous right-hand side have,
under mild conditions (Lipschitz constant), a unique solution in forward and
backward time. When dealing with differential equations with a discontinuous
right-hand side, one has to think about an appropriate solution concept. Filip-
pov’s solution concept consists of filling in the graph with a convex set, which
yields a differential inclusion. Differential inclusions can describe the (non-
smooth) time-evolution of systems with an absolutely continuous state. The
differential measure of the state, which usually consists of a Lebesgue measure-
able part, can be augmented with an atomic part. An inclusion based on this
differential measure yields a measure differential inclusion, which can describe
the time-evolution of systems with a discontinuous state (of locally bounded
variation). A sufficient condition for uniqueness of solutions of (measure) dif-
ferential inclusions follows from the monotonicity of the negative right-hand
 4.4 Summary 77

side. The framework of measure differential inclusions will be used in the

following chapters to set up a Lyapunov stability theory for non-smooth sys-
tems and to describe the non-smooth dynamics of mechanical systems with
frictional unilateral contacts.
This page intentionally blank
5
Mechanical Systems with Set-valued Force
Laws

Some theoretical background on non-smooth systems has been discussed in the

previous chapter. Mechanical multibody systems form a special and important
class of non-smooth systems, because they can be cast in an elegant structured
form. The special structure of mechanical systems is due to the fact that the
dynamics is described by the Lagrangian formalism, which links dynamics to
variational calculus. Moreover, contact forces are incorporated in the equation
of motion by using the Lagrange multiplier theorem. But most importantly,
contact forces are (mostly) derived from (pseudo-)potentials or dissipation
functions.
In this chapter we will discuss the mathematical formulation of Lagrangian
mechanical systems with unilateral contact and friction modelled with set-
valued force laws. It is important to note that (finite-dimensional) Lagrangian
mechanical systems encompass rigid multibody systems as well as discretized
continuous systems (e.g. through a Ritz approach or a finite-element dis-
cretization) with possible frictional unilateral contacts. First, we discuss how
set-valued force laws can be derived from non-smooth potentials. Subse-
quently, we treat the contact laws for unilateral contact and various types
of friction within the setting of non-smooth potential theory. This leads to a
unified approach with which all set-valued forces can be formulated. Finally,
we incorporate the set-valued forces as Lagrangian multipliers in the Newton-
Euler equations. The notation in this chapter is kept as close as possible to
the notation of Glocker [63].

5.1 Non-smooth Potential Theory

Mechanical constitutive laws are commonly expressed by derivatives of po-

tentials. Both conservative and non-conservative forces can be derived from
potentials. However, non-conservative potentials, also called dissipation func-
tions, are non-integrable, i.e. the value of the integral is path-dependent.
80 5 Mechanical Systems with Set-valued Force Laws

The potential of a linear mechanical spring is given by U (q) = 12 kq 2 , where

k is the spring stiffness and q is the elongation of the spring. The force law of
a linear spring can be expressed as

∂U (q)
−λ = = kq, (5.1)
∂q

where λ is the force exerted by the spring. Similarly, the potential (usually
called dissipation function) for a linear viscous damper is given by Φ(u) =
1 2
2 cu , where c is the damping constant and u is the time-derivative of the
elongation of the damper. The force law of a linear damper can be expressed
as
∂Φ(u)
−λ = = cu. (5.2)
∂u
The potentials U (q) and Φ(u) of a linear spring and damper are quadratic
smooth functions of the displacement q or velocity u. Non-quadratic smooth
potentials describe the energy or dissipation rate of nonlinear smooth force
laws. However, many force laws are expressed by non-smooth potentials [61,
63].
Force laws that can be derived from a potential or dissipation function are
generally of the form1
−λ ∈ ∂π(s), (5.3)
in which λ is the force in the force element, s is the flow variable and π(s) is the
potential. The potential π(s) is assumed to be differentiable in the generalised
sense of Clarke, i.e. the generalised differential ∂π(s) exists. The force law is
stated in the form of an inclusion because the generalised differential ∂π(s)
is set-valued if π(s) is non-smooth. Following [61, 63], our analysis of non-
smooth potential functions will be restricted to functions π : R → (−∞, ∞]
which can be decomposed into a differentiable function πD (s), an indicator
function πS (s) and a convex potential πP (s) with polyhedral epigraph:

π(s) = πD (s) + πS (s) + πP (s). (5.4)

Without loss of generality we set π(0) = 0 and therefore consider force

elements for which the energy or dissipation rate is zero for a vanishing flow
variable s. The indicator function πS is defined as πS (s) = ΨS (s) on some
closed interval S = [sA , sE ] and the convex potential πP (s) can be written in
the form

m
0 s ≤ si
πP (s) = πP i (s), πP i (s) = , (5.5)
i=1
a i (s − si ) s > si

with ai > 0. The decomposition (5.4) is illustrated in Figure 5.1.

1
The notation differs from [101] in which the force law is defined as λ ∈ ∂π(s).
 5.1 Non-smooth Potential Theory 81

Fig. 5.1. The decomposition of a potential π(s) = πD (s) + πS (s) + πP (s).

The generalised differential ∂ of πD (s) equals the gradient ∇, because

πD (s) is differentiable. The potentials πS (s) and πP (s) are convex and their
generalised differential therefore equals the subdifferential. Hence, we obtain

∂π(s) = ∇πD (s) + ∂πS (s) + ∂πP (s). (5.6)

We can therefore also decompose the force λ in the same fashion:

λ = λD + λ S + λ P , (5.7)

with the force laws

−λD = ∇πD (s), −λS ∈ ∂πS (s), −λP ∈ ∂πP (s). (5.8)

Note that only the differentiable part πD (s) is allowed to be non-convex.

We now study the case that the potential π(s) is convex, i.e. πD (s) is
convex or vanishes. A convex potential π(s) has a conjugate potential π ∗ (−λ)
(see Definition 2.28)

π ∗ (−λ) = sup{−sλ − π(s)}. (5.9)

s

The force law can now be also be stated in its conjugate form

−λ ∈ ∂π(s) ⇐⇒ s ∈ ∂π ∗ (−λ), s ∈ S, −λ ∈ C (5.10)

82 5 Mechanical Systems with Set-valued Force Laws

Fig. 5.2. Complementary behaviour of π(s) and π ∗ (−λ).

where C denotes the effective domain of π ∗ (−λ). If the force law −λ ∈ ∂π(s)
is fulfilled, then it holds that

−λs = π(s) + π ∗ (−λ), (5.11)

from which we see (Figure 5.2) that the potential π(s) and its conjugate
π ∗ (−λ) describe complementary areas in the first quadrant of the s–λ plane.
The conjugate potential π ∗ (−λ) is therefore known as ‘complementary energy
function’ in structural mechanics.
We will study the convex potentials πS (s) and πP (s) in more detail. The
subdifferential of the indicator function πS (s) = ΨS (s) is the normal cone to
S, i.e.
−λS ∈ NS (s). (5.12)
Similarly, taking the conjugate of the potential πP i (s)

πP∗ i (−λP i ) = sup{−sλP i − πP i (s)}

s
 
−sλP i if s ≤ si
= sup
s −sλP i − ai (s − si ) if s > si
 (5.13)
−si λP i if 0 ≤ −λP i ≤ ai
=
∞ else
= ΨAi (−λP i ) − si λP i , Ai = [0, ai ],

and using
s ∈ ∂πP∗ i (−λP i ) =⇒ s ∈ ∂ΨAi (−λP i ) + si , (5.14)
we deduce that it holds that

s − si ∈ NAi (−λP i ). (5.15)

The latter exposition shows how the force laws associated with the potentials
πS (s) and πP (s) can be written in a ‘normal cone’ formulation.
 5.2 Contact Geometry 83

Fig. 5.3. Contact distance gN and tangential velocity γT between two rigid bodies.

5.2 Contact Geometry

In the following section, we study a number of set-valued force laws that are
common in the description of contact problems of mechanical systems. We
therefore need to describe what we mean with contact points, their distance
and their relative velocity. Consider two convex bodies, say body 1 and 2,
which have a locally smooth surface in the region where they are close to each
other, as shown in Figure 5.3. We denote the point on body 1, which is closest
to body 2, by P1 . Similarly, the point P2 is the point on body 2, which is closest
to body 1. Note that the points Pi wander over the body while the bodies
move in time. Let the body-fixed point C1 on body 1 momentarily coincide
with the point P1 , i.e. rOP 1 = rOC1 . Similarly, let the body-fixed point C2
on body 2 momentarily coincide with the point P2 . For each point Cj on the
locally smooth surface of body j we can define an outward normal nj and
tangent plane Tj , j = 1, 2. Two points C1 and C2 , momentarily coinciding
with P1 and P2 , on the surface of body 1 and 2 respectively are said to form
a contact pair or contact point if their outward normals oppose each other:

n1 = −n2 , (5.16)

where nj is the outward normal on body j at Cj . The tangent planes T1 and

T2 are therefore necessarily parallel. The tangent plane Tj is spanned by the
vectors tj and sj , such that (sj , tj , nj ) is a body-fixed orthonormal frame
at the body-fixed point Cj . The points P1 and P2 are the closest extremal
points of the two bodies under consideration and are candidates to contact.
The contact distance

gN = (rOP 2 − rOP 1 )T n1 = (rOP 1 − rOP 2 )T n2 (5.17)

84 5 Mechanical Systems with Set-valued Force Laws

is the distance between the two bodies. The bodies are separated when gN > 0,
are in contact when gN = 0, and penetrate each other when gN < 0. The
bodies are approaching each other with the normal contact velocity

γN = (vC2 − vC1 )T n1 , (5.18)

where vCj are the absolute velocities of the body-fixed points Cj . The differ-
ence of velocities
vCjP j = vP j − vCj (5.19)
gives the velocity with which the point Pj wanders over body j. The points Cj
move also relative to each other along the tangent plane T1 with the tangential
contact velocity  
γT 1
γT = ∈ T1 , (5.20)
γT 2
with the components

γT 1 = (vC2 − vC1 )T t1 , γT 2 = (vC2 − vC1 )T s1 . (5.21)

If the bodies are in contact, i.e. gN = 0, then the tangential contact velocity
γT becomes the relative velocity with which the bodies slide over each other
expressed in the frame (t1 , s1 ), and is a two-dimensional vector in the tangent
plane T = T1 = T2 .

5.3 Force Laws for Frictional Unilateral Contact

In this section we will discuss various set-valued force laws for the constitutive
description of unilateral contact with friction. Bodies in contact naturally
deform under the influence of contact forces yielding a contact area. Usually,
the local deformations at the contact area, as well as the size of the contact
area, are negligibly small compared with the global motion of the contacting
bodies. We can therefore often assume that the bodies are deformable but
neglect the local indentation and idealise the contact area to be a contact
point [128]. A possible further modelling step, although not essential, is the
assumption of rigidity of the contacting bodies, in which also the deformation
of the bodies is neglected [139].

5.3.1 Signorini’s Contact Law

Contact in normal direction between contacting bodies is usually described by

Signorini’s law which is the most elementary set-valued force law. Signorini’s
law describes the impenetrability of the contact, i.e. the contact distance is
nonnegative (gN ≥ 0). A contact force λN along the line C1 –C2 can be positive
when contact is present (gN = 0), but must vanish when the contact is open
(gN > 0). The normal contact force λN between the bodies is nonnegative
 5.3 Force Laws for Frictional Unilateral Contact 85

Fig. 5.4. Potential, conjugate potential and Signorini’s force law.

because the bodies can only push on each other, i.e. the constraint is unilateral.
Only two situations may occur:

gN = 0 ∧ λN ≥ 0 contact
(5.22)
gN > 0 ∧ λN = 0 no contact

From (5.22) we see that the normal contact law shows a complementarity
behaviour: the product of the contact force and normal contact distance is
always zero:
gN λN = 0. (5.23)
Hence, Signorini’s law does not lead to energy creation or dissipation. The
relation between the normal contact force and the normal contact distance is
therefore given by
gN ≥ 0, λN ≥ 0, gN λN = 0, (5.24)
which is the inequality complementarity condition between gN and λN , also
known as the Signorini-Fichera condition. We now put the Signorini-Fichera
condition (5.24) in the framework of non-smooth potential theory. Consider
s = gN to be the flow variable with its dual force λ = λN . The Signorini-
Fichera condition (5.24) appears to be a maximal monotone set-valued force
86 5 Mechanical Systems with Set-valued Force Laws

law −λN (gN ) (see the lower right figure in Figure 5.4). The maximal mono-
tonicity of the force law implies that it can be derived from a convex potential
πN (gN ). We therefore write

−λN ∈ ∂πN (gN ) (5.25)

and find that the associated energy potential is πN (gN ):


∞, gN < 0,
πN (gN ) = (5.26)
0, gN ≥ 0.

We see that the potential πN (gN ) purely consists of an indicator function

∗
πN (gN ) = πN S (gN ) = ΨR+ (gN ). The conjugate potential πN (−λN ) can easily
be shown to be equal to the indicator function ΨCN (−λN ) on the convex set
CN = R− . We interpret the set

CN = {−λN ∈ R|λN ≥ 0} (5.27)

as the set of admissible (negative) normal contact forces. Equivalently, we

state the force law in its conjugate form
∗
gN ∈ ∂πN (−λN ). (5.28)

We now use the equivalence (2.27) between the normal cone and the subdif-
ferential of the indicator function and obtain NCN (−λN ) = ∂ΨCN (−λN ) =
∗
∂πN (−λN ) with which we cast the force law in a normal cone formulation

gN ∈ NCN (−λN ). (5.29)

If the system has a number of contact points i = 1, . . . , nC , then the force

law
gN i ∈ NCN i (−λN i ), CN i = {−λN i ∈ R|λN i ≥ 0}, (5.30)
holds for each contact point i. Note that all sets CN i are identical, CN i =
CN = R− . Using the short-hand notation

x≥0 ⇐⇒ xi ≥ 0 ∀i, (5.31)

the normal contact laws (5.30) can be written in vector form as

gN ∈ NCN (−λN ), CN = {−λN ∈ Rn |λN ≥ 0}, (5.32)

where λN is the vector containing the normal contact forces λN i and gN is

the vector of normal contact distances gN i .
 5.3 Force Laws for Frictional Unilateral Contact 87

Fig. 5.5. Potential, conjugate potential and Coulomb’s friction law.

5.3.2 Coulomb’s Friction Law

Coulomb’s friction law is another classical example of a force law that can
be described by a non-smooth potential. Before discussing Coulomb’s spatial
friction law, we first study the planar case for which the relative sliding velocity
γT is a scalar. If contact is present between the bodies, i.e. gN = 0, then the
friction between the bodies imposes a scalar force λT along the tangent line
of the contact point. When the bodies are sliding over each other, the friction
force λT equals to μλN and acts in the direction of −γT

−λT = μλN sign(γT ), γT = 0, (5.33)

where μ is the friction coefficient and λN is the normal contact force. The
classical Coulomb friction law considers the friction coefficient to be constant.
If the relative tangential velocity vanishes, i.e. γT = 0, then the bodies purely
roll over each other without slip. Pure rolling, or no slip for locally flat objects,
is denoted by stick. If the bodies stick, then the friction force must lie in the
interval −μλN ≤ λT ≤ μλN . For unidirectional friction, i.e. for planar contact
problems, the following three cases are possible:
88 5 Mechanical Systems with Set-valued Force Laws

γT = 0 ⇒ |λT | ≤ μλN sticking,

γT < 0 ⇒ λT = +μλN negative sliding, (5.34)
γT > 0 ⇒ λT = −μλN positive sliding.

We can express the friction force by a (pseudo-)potential πT (γT ), which we

mechanically interpret as a dissipation function,

−λT ∈ ∂πT (γT ), πT (γT ) = μλN |γT |, (5.35)

from which follows the maximal monotone set-valued force law

⎧
⎪
⎨μλN , γT > 0,
−λT ∈ [−1, 1]μλN , γT = 0, (5.36)
⎪
⎩
−μλN , γT < 0,
or
−λT ∈ μλN Sign(γT ). (5.37)
The admissible values of the (negative) tangential force −λT form a convex
set CT which is bounded by the values of the normal force [139]:

CT = {−λT | −μλN ≤ λT ≤ +μλN }. (5.38)

Coulomb’s law can be expressed with the aid of the indicator function of CT
as
γT ∈ ∂ΨCT (−λT ) ⇔ γT ∈ NCT (−λT ), (5.39)
where the indicator function ΨCT is the conjugate potential of πT , i.e.
πT (γT ) = ΨC∗ T (γT ), see Figure 5.5. We call πT (γT ) a pseudo-potential, be-
cause it is dependent on the normal contact force, which is in turn dependent
on the motion of the system, and is therefore not a true potential. In the
following, the word ‘pseudo’ will be omitted for the sake of brevity.
The classical Coulomb’s friction law for spatial contact formulates a two-
dimensional friction force  
λ
λT = T 1 , (5.40)
λT 2
with the components

λ T 1 = λT t 1 , λ T 2 = λT s1 , (5.41)

where λ = λN n1 + λT 1 t1 + λT 2 s1 is the total contact force exerted by body 1

on body 2. The set of admissible friction forces is a convex set CT ⊂ R2 which
is a disk for isotropic Coulomb friction:

CT = {−λT | λT  ≤ μλN }. (5.42)

Using the set CT , the spatial Coulomb friction law can be formulated as

γT ∈ ∂ΨCT (−λT ) ⇔ −λT ∈ ∂ΨC∗ T (γT ) ⇔ γT ∈ NCT (−λT ), (5.43)

 5.3 Force Laws for Frictional Unilateral Contact 89

Fig. 5.6. Graphical interpretation of Coulomb’s friction law.

in which γT ∈ R2 is the relative sliding velocity. Similarly, an elliptic choice

of CT would result in an orthotropic Coulomb friction law (see Figure 5.6).
Note that the dual variables −λT and γT are no longer aligned if CT is non-
circular. For a fixed vector γT , the vector −λT is that vector in CT which
causes the inner product −γTT λT to be maximal. In other words: a vector γT
defines a unique tangent plane on the boundary of CT such that n = γT /γT 
is the normal to the tangent plane. The projection of −λT on the normal n
to this tangent plane is maximal. The force law (5.43) corresponds therefore
to a principle of maximal dissipation −γTT λT .
Coulomb’s friction law in its classical setting considers the friction coeffi-
cient to be constant. In order to model the Stribeck effect, which can lead to
the occurrence of friction induced vibrations [101, 102], the friction coefficient
is usually considered to be dependent on the relative sliding velocity γT , i.e.

−λStribeck ∈ μ(γT )λN Sign(γT ), (5.44)

where μ : R → R+ is a continuous function. Such a type of law is as-

sociated with a non-convex potential πStribeck (−γT ) such that −λStribeck =
∂πStribeck (γT ). The friction force λStribeck can be decomposed in a force λT ,
which obeys the standard Coulomb set-valued friction law, and a force λsmooth
being a smooth function of γT :
−λStribeck ∈ μ(γT )λN Sign(γT )
= μ(0)λN Sign(γT ) + (μ(γT ) − μ(0))λN Sign(γT )
(5.45)
= μ(0)λN Sign(γT ) + (μ(γT ) − μ(0))λN sign(γT )
= −λT − λsmooth
with

−λT ∈ μ(0)λN Sign(γT ), −λsmooth = (μ(γT ) − μ(0))λN sign(γT ), (5.46)

 
μsmooth
90 5 Mechanical Systems with Set-valued Force Laws

and μsmooth (0) = 0. The potential πStribeck (γT ) can be decomposed accord-
ingly in a smooth potential πsmooth (γT ) and the convex potential πT (γT ) of
the classical Coulomb friction law:

πStribeck (γT ) = πsmooth (γT ) + πT (γT ), (5.47)

with
−λsmooth = ∇πsmooth (γT ), −λT ∈ ∂πT (γT ). (5.48)
The friction law of Coulomb, as defined above, assumes the friction forces
to be a function of the unilateral normal forces. Both the normal contact forces
and the friction forces have to be determined. However, in many applications
the situation is less complicated as the normal force is constant or at least a
given function of time. A known normal contact force allows for a simplified
contact law. The tangential friction forces are assumed to obey either one of
the following friction laws:
1. Non-associated Coulomb’s law for which the normal force is dependent on
the generalised coordinates q and/or generalised velocities u and therefore
not known in advance. The set of (negative) admissible contact forces is
given by
CT (λN ) = {−λT | λT  ≤ μλN }, (5.49)
which is dependent on the normal contact forces λN and friction coefficient
μ. The dissipation function is a pseudo-potential.
2. Associated Coulomb’s law for which the normal force is known in advance.
The set of (negative) admissible contact forces is given by

CT = {−λT | λT  ≤ μFN }, (5.50)

which is dependent on the known normal forces FN and friction coefficient

μ. The dissipation function is a true potential.

5.3.3 Coulomb-Contensou Friction Law

Bodies in contact do not only slide over each other, but also pivot with respect
to each other. Their relative pivoting velocity

ωN = (Ω2 − Ω1 )T n1 (5.51)

depends on the angular velocity vectors Ωj of each body j. In the following,

we refer to ωN as the normal spin. A combined friction law, which takes
into account sliding friction as well a pivoting (or drilling) friction, can be
formulated using a three-dimensional set of admissible (generalised) friction
forces and is called the spatial Coulomb-Contensou friction law [99].
The set-valued force law for spatial Coulomb friction describes the friction
 T
force λT of a single contact point with two components λT = λT 1 λT 2 . The
 5.3 Force Laws for Frictional Unilateral Contact 91

Fig. 5.7. The Contsensou-effect explained on an electric polishing machine.

friction force λT lies within a disk λT  ≤ μλN on the tangent plane of the
contact point. The friction disk constitutes a friction cone in the (λT 1 , λT 2 , λN )
space. The pair of contacting bodies are assumed to be rigid and impenetrable
within the framework of rigid multibody dynamics and the contact is therefore
idealised to be a point. The contact point can not transmit a friction torque
and the influence of normal spin and pivoting friction on the sliding friction
λT is therefore usually neglected. In reality, the stiff (but still deformable)
bodies deform and touch each other on a small contact surface, being more or
less circular. The small deformations of stiff bodies are negligible compared to
the geometry of the bodies and the global rigid body motion they undergo, but
lead to contacting areas which can influence the dynamics of the system. A
contact surface can not only transmit a sliding force λT tangent to the contact
surface, but also transmit a friction torque τN normal to the contact surface.
The effective radius of the contact surface is influenced by the normal contact
force λN , the elasticity of the contacting bodies, the surface roughness and
the pressure distribution. The friction torque τN is, in the absence of sliding
and tangential contact force, more or less proportional to the normal contact
force, the friction coefficient and the effective radius, i.e τN ∝ μλN Reff . The
influence of the friction torque is in most applications neglected because the
effective radius is very small in practice. If, however, an object is spinning
fastly, then the influence of normal spin and pivoting friction on the dynamics
becomes large and can no longer be neglected.
Contensou [39] realised that pivoting friction and especially normal spin
are important for the dynamics of fastly spinning tops. The pivoting friction
is obviously necessary to describe the gradual loss of energy of the top which
brings it back to rest. More of interest is the influence of the spinning veloc-
ity on the sliding friction force. A fastly spinning top experiences very little
resistance in sliding direction and easily wanders over the floor. The same phe-
nomenon occurs in an electric polishing machine with turning brushes used to
92 5 Mechanical Systems with Set-valued Force Laws

clean floors [110]. The machine (Figure 5.7) is hard to move when the brushes
are not rotating (Coulomb friction), but the machine can easily be pushed
over the floor with rotating brushes. This is called the Contensou effect. Note
that the opposite holds for the torque which the motor delivers. If the machine
is not moved then the motor has to deliver a high torque and is experiencing
pure Coulomb friction in a rotational sense. If the sliding speed is much higher
than the normal spin, then the torque is lowered.
Consider a contact surface, which may be elliptic or even non-convex, with
a normal pressure distribution σ. The relative sliding velocity of the contact
is denoted by γT and the normal spin by ωN . The normal spin ωN is scaled
to a velocity
γP = ωN R, (5.52)
where R is some characteristic length of the contact surface (e.g. the radius if
the contact surface is circular). The vector
⎡ ⎤
γT 1
γF = ⎣γT 2 ⎦ (5.53)
γP

expresses the generalised sliding velocity. Coulomb’s spatial friction law is

assumed to hold on an arbitrary surface element dA with sliding velocity w.
The sliding velocity w is a function of the generalised sliding velocity γF .
The magnitude of the friction force on dA is according to Coulomb’s law
dλT = −μσdA w w
for w = 0. The velocity potential (or dissipation function)
for dA is dπF = −wT dλT = μσwdA. The potential for the total contact
surface is obtained by integrating over the contact surface

πF (γF ; σ) = μσw(γF )dA. (5.54)
A

It can be proven that the potential πF (γF ; σ) is a support function on a convex

set BF (σ) which we call the friction ball:

πF (γF ; σ) = ΨB∗ F (σ) (γF ). (5.55)

We can therefore express the combined Coulomb-Contensou friction law as

γF ∈ NBF (σ) (−λF ), (5.56)

where NBF (σ) is the normal cone on the friction ball BF (σ). The generalised
force vector −λF is the dual variable to γF and can be interpreted as
⎡ ⎤
λT 1
λF = ⎣λT 2 ⎦ , (5.57)
λP
 5.3 Force Laws for Frictional Unilateral Contact 93

Fig. 5.8. Coulomb-Contensou friction law with the friction ball BF .

in which λP = τN /R is the scaled pivoting torque. In [99], the potential πF and

the friction ball BF have been derived for a circular contact with a uniform or
parabolic pressure distribution σ. The friction ball BF (see Figure 5.8) is in
these cases squashed but axisymmetric around the −λP axis. The intersection
of BF with the (−λT 1 , −λT 2 ) plane is formed by the friction disk CT .
Denote the magnitude of the sliding velocity by v = γT . The friction
characteristic λT (v) for a fixed value of γP can be derived from (5.56) and
is shown in Figure 5.9a for different fixed values of γP . A uniform pressure
distribution and a circular contact surface have been assumed. The curves
for γP > 0 are all single-valued functions of v. Apparently, a superimposed
normal spin ωN on a sliding velocity v causes a smoothing effect of the
friction characteristic λT (v). The friction characteristic for γP = 0 is (for
λP = 0) the classical set-valued friction characteristic of Coulomb. Similarly,
the dependence of λP on γP , for different fixed values of v, is shown in Fig-
ure 5.9b. The same smoothing effect occurs in the λP (γP ) relationship due
to a superimposed velocity v > 0. Again, a set-valued relationship is ob-
tained for v = 0, which corresponds (for λT = 0) to the classical set-valued
friction law of a purely spinning contact. The boundaries λT  = μλN and
|λP | = 23 μλN of the sets for v = γP = 0 are the extreme values of the friction
ball BF (λN ) along its principal axes under the assumption of a uniform pres-
sure distribution and a circular contact surface. For the classical set-valued
friction characteristic of Coulomb (i.e. a purely translational contact λP = 0,
γP = 0), the magnitude friction force λT  rises up to the value μλN when
the contact changes from a sticking state to sliding. This is in general not the
case for Coulomb-Contensou friction as soon as we have some spin. A stick-
ing contact obeying Coulomb-Contensou’s friction law will start slipping for
λT  < μλN if λP = 0. The slip-criterion for Coulomb-Contensou friction is
given by −λF ∈ bdry BF (λN ).
94 5 Mechanical Systems with Set-valued Force Laws

(a) λT as function of v for fixed values of γP .

(b) λP as function of γP for fixed values of v.

Fig. 5.9. Friction curves for a uniform pressure distribution and circular contact
surface.

5.3.4 Rolling Friction

A resistance against sliding and pivoting of bodies in contact is modelled

by the Coulomb-Contensou friction model of the previous section. However,
the bodies may also experience a resistance against rolling over each other.
At this point we have to ask ourselves what we exactly mean when we say
that bodies ‘roll’ over each other. We may call ‘rolling’ the movement of
the contact point Pi over the surface of body i. A resistance against such
 5.3 Force Laws for Frictional Unilateral Contact 95

a type of movement will be called contour friction [95]. Usually, the term
rolling is associated with resistance against a difference in angular velocity
of the contacting bodies, i.e. Ω2 − Ω1 , tangential to the tangent plane of
contact (see for instance [83]). This will be called classical rolling friction.
Contour friction and classical rolling friction may be identical to each other
or be essentially different, depending on the type of system. For instance,
if a planar wheel rolling over a flat floor is considered, then the two types
of rolling friction yield the same kind of dissipation mechanism, because the
velocity of the contact point over the contour of the wheel is directly related
to the angular velocity of the wheel. However, the two types of rolling friction
are essentially different if we consider two wheels in contact. If we let the two
wheels move over each other with no relative angular velocity, i.e. Ω2 = Ω1 ,
then the contact point still moves over the contour of both wheels. The fact
that the classical rolling friction model yields no dissipation for this case shows
that this rolling friction model is cumbersome. The classical rolling friction
model becomes also questionable when applied to two wheels in contact with
different diameters. The main problem is that it is not clear which dissipation
mechanism the classical rolling friction model tries to model. A mathematical
formulation for contour and classical rolling friction will be discussed in the
following.
Contour friction [95] is defined as the resistance against the contour veloc-
ity on body j  
γ
γC,j = C1,j ∈ R2 , (5.58)
γC2,j
which has the components
T T
γC1,j = vDjP j tj , γC2,j = vDjP j sj , (5.59)

where vDjP j is the velocity with which contact point Pj moves with respect to
an arbitrary body-fixed point Dj on body j. The velocity vDjP j is necessarily
T
a vector in the tangent plane Tj of the contact point Pj , because vDjP j nj = 0.
Each body j ∈ {1, 2} has its own contour velocity γC,j for which we can set
up a contour friction law. In the following, we will suppress the subscript ,j .
In order to set up the contact law, we choose a potential πC (γC ) which acts
as dissipation function and which leads to the force law

−λC ∈ ∂πC (γC ), (5.60)

where λC is the contour friction force, being the dual variable to γC . If we

choose the πC (γC ) to be a support function on a convex set CC

πC (γC ) = ΨC∗ C (γC ), (5.61)

then we obtain a Coulomb-like friction law which can be expressed as a normal

cone inclusion
γC ∈ NCC (−λC ). (5.62)
96 5 Mechanical Systems with Set-valued Force Laws

The set CC of admissible contour friction forces can be chosen to be dependent

on the normal contact force λN

CC = {−λC | λC  ≤ μC λN }, (5.63)

where μC is the contour friction coefficient.

Although the classical rolling friction model is cumbersome (see the above
discussion), we will still briefly discuss how it can be mathematically de-
scribed. Classical rolling friction [83] is similar to pivoting friction, which de-
scribes dissipation due to a normal spin ωN . Classical rolling friction defines
a resistance against the tangential spin of the two contacting bodies
 
ωT 1
ωT = , (5.64)
ωT 2

which has the components

ωT 1 = (Ω2 − Ω1 )T t1 , ωT 2 = (Ω2 − Ω1 )T s1 . (5.65)

The tangential spin ωT is a vector in the tangent plane of the contact point,
which can be normalised to a rolling velocity

γR = ωT L, (5.66)

where L is some characteristic length of body 1 or 2. Just as before, we choose

a potential πR (γR ) = ∂ΨC∗ R (γR ) in the form of a support function which acts
as dissipation function and which leads to the force law

−λR ∈ ∂πR (γR ) ⇐⇒ γR ∈ NCR (−λR ), (5.67)

where λR is the classical rolling friction force, being the dual variable to γR
and CR is the admissible set of −λR . The classical rolling friction force λR
can be interpreted as a normalised rolling friction moment τT :

τT = λR L, (5.68)

with components τT 1 and τT 2 dual to ωT 1 and ωT 1 . The total torque trans-

mitted by the contact is therefore the vector

τ = τN n1 + τT 1 t1 + τT 2 s1 . (5.69)

Coulomb friction, Coulomb-Contensou and two types of rolling friction

have been discussed in the previous sections. If dry friction laws are considered,
then all these friction laws can be formulated as a normal cone inclusion

γD ∈ NCD (−λD ), γD , λD ∈ Rp , CD ⊂ Rp . (5.70)

 5.3 Force Laws for Frictional Unilateral Contact 97

(a) (b)

Fig. 5.10. Multi-contact collision.

The dimension p is 1 for planar Coulomb friction and 2 for spatial Coulomb
friction. Coulomb friction and pivoting friction were considered to be coupled
to each other which has been described by the Coulomb-Contensou friction
law for which p = 3. Contour and classical rolling friction have dimension
p = 2. However, it might also be possible to set up a dissipation function
for combined sliding–pivoting–rolling, which would result in p = 5. In the
following, we will just write (5.70) to denote any frictional dissipation law,
∗
being derived from a velocity pseudo-potential πD(λN ) (γD ) = ΨD(λ N)
(γD ).
This notation allows us to describe the mechanical system in a general way
and to prove the stability results for arbitrary friction laws.

5.3.5 Impact Laws

Signorini’s law and Coulomb’s friction law are set-valued force laws for non-
impulsive forces. In order to describe impact, we need to introduce impact
laws for the contact impulses. We will consider a Newton-type of restitution
law,
−
+
γN = −eN γN , gN = 0, 0 ≤ eN ≤ 1, (5.71)
+
which relates the post-impact velocity γN of a contact point to the pre-impact
−
velocity γN by Newton’s coefficient of restitution eN . The case eN = 1 cor-
responds to a completely elastic contact, whereas eN = 0 corresponds to a
completely inelastic contact. The impact, which causes the sudden change in
relative velocity, is accompanied by a normal contact impulse ΛN > 0. Follow-
ing [62], suppose that, for any reason, the contact does not participate in the
impact, i.e. that the value of the normal contact impulse ΛN is zero, although
the contact is closed. This happens normally for multi-contact situations. Con-
sider for example a rod, depicted in Figure 5.10a, which is supported on two
supports, somewhat offset to the right. A ball is hitting the rod on its right
end. Three contact points are closed when the ball hits the rod: 1. the rod–left
support contact, 2. the rod right-support contact, and 3. the ball–rod contact.
Contact 2 and 3 will transmit a contact impulse which ensures the the impen-
etrability of these contacts. Contact 1, however, contact 1 will not transmit
a contact impulse. Such a contact will be called superfluous. The pre-impact
velocity of contact 1 is zero and its post-impact velocity is positive due to
the transmitted impacts at contact points 2 and 3. Another example shows
98 5 Mechanical Systems with Set-valued Force Laws

Figure 5.10b. A horizontal rod is fall on two supports, both located on the left
side of the rod’s center of mass. Again, contact 1 is superfluous. This time,
however, its pre-impact relative velocity is negative. For superfluous contacts
we allow post-impact relative velocities higher than prescribed by Newton’s
−
impact law in the case of a non-vanishing impulse, γN +
> −eN γN , in order to
express that the contact is superfluous and could be removed without chang-
ing the contact-impact process. Summarising, two cases can occur at a closed
contact
1. The contact is actively participating in the impact process, i.e. ΛN > 0
−
+
and γN = −eN γN ,
−
2. The contact is superfluous, i.e. ΛN = 0 and γN +
≥ −eN γN .
We can combine these two cases in an impact law formulated as an inequality
complementarity condition on velocity–impulse level:

ΛN ≥ 0, ξN ≥ 0, ΛN ξN = 0, (5.72)
+ −
with ξN = γN + eN γN (see [62]). Similarly to Signorini’s law on velocity level
we can write the impact law in normal direction as

ξN ∈ NCN (−ΛN ), gN = 0, CN = R− , (5.73)

or by using the support function

−ΛN ∈ ∂ΨC∗ N (ξN ), gN = 0. (5.74)

A normal contact impulse ΛN at a frictional contact leads to a frictional

contact impulse ΛD ∈ Rp with −ΛD ∈ CD (ΛN ). We therefore have to specify
a frictional impact law as well, taking into account a possible frictional resti-
tution. Restitution in tangential direction occurs for instance in the motion of
the Super Ball, being a very elastic ball used on play grounds. The frictional
impact law can be formulated in a similar way as has been done for the normal
impact law:
−ΛD ∈ ∂ΨC∗ D (ΛN ) (ξD ), gN = 0, (5.75)
−
with ξD = γD +
+ eD γD and |eD | ≤ 1.
Let us analyse the frictional impact law (5.75) for planar Coulomb friction,
i.e.
−ΛT ∈ ∂ΨC∗ T (ΛN ) (ξT ), gN = 0, (5.76)
with ξT = γT+ + eT γT− and CT (ΛN ) = {ΛT | −μΛN ≤ ΛT ≤ μΛN }. This
friction law is equivalent to

−ΛT ∈ μΛN Sign(γT+ + eT γT− ), (5.77)

which is similar to (5.37). We distinguish three cases:

• impact with stick
|ΛT | ≤ μΛN , γT+ = −eT γT− ,
 5.4 Measure Newton-Euler Equations 99

• impact with forward slip

ΛT = −μΛN , γT+ > −eT γT− ,

• impact with backward slip

ΛT = μΛN , γT+ < −eT γT− .

In this section we have presented a generalisation of Newton’s impact law.

Such a generalisation is also possible for Poisson-type of impact laws [139],
which distinguish between a compression and an expansion phase. Poisson’s
impact law relates the impulse in the expansion phase to the impulse in the
compression phase through a restitution coefficient [62]. A generalisation of
Poisson’s impact law involves two complementarity inequalities, one for each
phase.
A restitution coefficient must be regarded as an impact process parameter,
not as a material constant. Consider for instance a ball which is falling on a
flexible bar. The dynamic response will greatly depend on the location of
impact. If the ball falls on a node of the first bending eigenmode of the bar,
then this will cause another response than if the ball hits an antinode. Hence,
the impact process is governed by wave effects and the dissipation caused
by wave effects. In a rigid-body approach, we are not interested in the wave
effects themselves, but we only want to capture the macroscopic motion of
the bodies. The wave and damping effects are therefore represented by the
restitution coefficients of the impact law.
The class of impact laws (5.74) and (5.75) does not parameterise the
whole set of possible post-impact velocities that are kinematically admissi-
ble (see [4]). The impact laws presented in this section are local, i.e. the
post-impact velocity of a contact point is related to the pre-impact velocity of
the same contact point. Non-local effects, which are due to wave effects, play
for instance a role in Newton’s cradle [4]. Non-local effects can be taken into
account by specifying restitution coefficients between different contact points.

5.4 Measure Newton-Euler Equations

In this section, we set up the measure Newton-Euler equations which form
together with the set-valued force laws for the contact forces a measure differ-
ential inclusion which describes the temporal dynamics of mechanical systems
with discontinuities in the velocity. Subsequently, we study the equilibrium set
of the measure differential inclusion. For simplicity, we will restrict the study
to time-autonomous systems, although the measure Newton-Euler equations
can also be set up for the non-autonomous case.
We assume that the time-autonomous mechanical systems under consid-
eration exhibit only bilateral holonomic frictionless constraints and unilateral
100 5 Mechanical Systems with Set-valued Force Laws

constraints in which dry friction can be present. This assumption is not es-
sential. We can very well model bilateral constraints with friction as we will
do in Section 7.3.2, but to simplify the notation we will start with bilateral
frictionless constraints and frictional unilateral constraints. Furthermore, we
assume that a set of independent generalised coordinates, q ∈ Rn , for which
these bilateral constraints are eliminated from the formulation of the dynam-
ics of the system, is known. The generalised coordinates q(t) are assumed to
be absolutely continuous functions of time t. Also, we assume the generalised
velocities, u(t) = q̇(t) for almost all t, to be functions of locally bounded vari-
ation. At each time-instance it is therefore possible to define a left limit u−
and a right limit u+ of the velocity. The generalised accelerations u̇ are there-
fore not for all t defined. The set of discontinuity points {tj } for which u̇ is
not defined is assumed to be Lebesgue negligible. We formulate the dynamics
of the system using a Lagrangian approach, resulting in2
 T 
m
d
(T,u ) − T,q + U,q =f nc
+ (wN i (q)λN i + WDi (q)λDi ) . (5.78)
dt i=1

The scalar T represents the kinetic energy and U denotes the potential en-
ergy. The column-vector f nc in (5.78) represents all smooth generalised non-
conservative forces. The vector wN i ∈ Rn and matrices WDi ∈ Rn×p define
the force directions of the normal contact forces λN i and frictional contact
forces λDi ∈ Rp .
Alternatively we write the dynamics as

M (q)u̇ − h(q, u) = WN (q)λN + WD (q)λD , (5.79)

which is a differential equation for the non-impulsive part of the motion.

Herein, M (q) = M T (q) > 0 is the mass-matrix, and we assume that the
kinetic energy for an autonomous system can be written as T = 12 uT M (q)u.
The state-dependent column-vector h(q, u) in (5.79) contains all differentiable
forces (both conservative and non-conservative), such as spring forces, gravita-
tional forces, smooth damper forces and gyroscopic terms. The contact forces
are assembled in vectors λN and λD and their corresponding force directions
in matrices WN and WD :

λN = {λN i }, λD = {λDi }, WN = {wN i }, WD = {WDi }. (5.80)

Due to the fact that the kinetic energy can be described by

1 T 1
T = u M (q)u = Mrs ur us , (5.81)
2 2 r,s

with M (q) = M T (q), we can write in tensorial language

2
Note that the subscript ,x indicates a partial derivative operation ∂/∂x.
 5.4 Measure Newton-Euler Equations 101
 
∂T 1  ∂Mrs ∂T 
= ur us , = Mkr ur ,
∂q k 2 r,s ∂q k ∂uk r
     ∂Mkr 
d ∂T
= Mkr u̇r + ur us
dt ∂uk r r,s
∂q s
(5.82)
 ∂T   ∂Mkr ∂Mrs

= r
Mkr u̇ + 2 k + s
− k
ur us
r
∂q r,s
∂q ∂q
d
(T,u ) = u̇T M (q) + 2T,q − (f gyr )T for almost all t,
dt
with the gyroscopic forces [129]
  ∂Mkr ∂Mrs

f gyr = {fkgyr }, fkgyr = − − ur us . (5.83)
r,s
∂q s ∂q k

The gyroscopic forces fgyr have zero power [129]

   ∂Mkr ∂Mrs

k gyr
T gyr
u f = u fk = − − ur us uk = 0. (5.84)
∂q s ∂q k
k k,r,s

In the same way as before, we can write the differential measure of T,u as

d (T,u ) = duT M (q) + 2T,q dt − (f gyr )T dt ∀t. (5.85)

Comparison with (5.79) and (5.78) yields

T
h = f nc + f gyr − (T,q + U,q ) , (5.86)

or in index notation
∂U ∂T
hk = fknc − k
− k + fkgyr
∂q ∂q
∂U ∂T   ∂Mkr ∂Mrs

= fk − k − k −
nc
s
− k
ur us
∂q ∂q r,s
∂q ∂q
∂U 1   ∂Mkr ∂Mrs

= fk − k −
nc
2 − ur us (5.87)
∂q 2 r,s ∂q s ∂q k
 
∂U 1  ∂Mkr ∂Mks ∂Mrs
= fk − k −
nc
+ − ur us
∂q 2 r,s ∂q s ∂q r ∂q k
∂U 
= fknc − − Γk,rs ur us
∂q k r,s

in which we recognise the holonomic Christoffel symbols of the first kind [129]
 
1 ∂Mkr ∂Mks ∂Mrs
Γk,rs = Γk,sr := + − . (5.88)
2 ∂q s ∂q r ∂q k
102 5 Mechanical Systems with Set-valued Force Laws

We introduce the following index sets:

IG = {1, . . . , m} the set of all contacts,

(5.89)
IN = {i ∈ IG | gN i (q) = 0} the set of all closed contacts,

and set up the force laws and impact laws of each contact as has been elab-
orated in the previous sections. The normal contact distances gN i (q) depend
on the generalised coordinates q and are gathered in a vector gN (q).
During a non-impulsive part of the motion, the normal contact force
−λN i ∈ CN and friction force −λDi ∈ CDi ⊂ Rp of each closed contact
i ∈ IN , are assumed to be associated with a non-smooth potential, being the
support function of a convex set, i.e.

−λN i ∈ ∂ΨC∗ N (γN i ), −λDi ∈ ∂ΨC∗ Di (γDi ), (5.90)

where CN = R− and the set CDi can be dependent on the normal contact
force λN i . The normal and frictional relative velocities are gathered in columns
γN = {γN i } and γD = {γDi }, for i = 1, . . . , m. We assume that these contact
velocities are related to the generalised velocities through:

γN (q, u) = WNT (q)u, γD (q, u) = WDT (q)u. (5.91)

This assumption is very important as it excludes rheonomic contacts3 . Note

also that WXT (q) = ∂γ∂u for X = N, D, which also holds for rheonomic con-
X

tacts.
Equation (5.79) together with the set-valued force laws (5.90) form a dif-
ferential inclusion
 
M (q)u̇ − h(q, u) ∈ −wN i (q)∂ΨC∗ N (γN i ) − WDi (q)∂ΨC∗ Di (γDi ) ,
i∈IN
(5.92)
for almost all t. Differential inclusions of this type are called Filippov systems,
which obey the solution concept given by Definition 4.2. The differential in-
clusion (5.92) only holds for impact free motion.
Subsequently, we define for each contact point the constitutive impact laws

−ΛN i ∈ ∂ΨC∗ N (ξN i ), −ΛDi ∈ ∂ΨC∗ Di (ΛN i ) (ξDi ), i ∈ IN , (5.93)

with
+ − + −
ξN i = γN i + eN i γN i , ξDi = γDi + eDi γDi , (5.94)
in which 0 ≤ eN i ≤ 1 and |eDi | ≤ 1 are the normal and frictional restitu-
tion coefficients respectively. The impact laws (5.93) can be generalised by
replacing the restitution coefficient eDi by a matrix.
3
A rheonomic contact is characterised by a contact distance g(q, t) being explicitly
dependent on time, or a contact velocity γ(q, u, t) being an affine function in
u, i.e. γ = W T (q, t)u + w(q, t). A contact which is not rheonomic is called
scleronomic.
 5.4 Measure Newton-Euler Equations 103

The force laws for non-impulsive motion can be put in the same form
as (5.93)
−λN i ∈ ∂ΨC∗ N (ξN i ), −λDi ∈ ∂ΨC∗ Di (λN i ) (ξDi ). (5.95)
because u+ = u− holds (for non-impulsive motion) and because of the positive
homogeneity of the support function (see Section 2.5):

ΨC∗ (ξ) = ΨC∗ (γ + eγ) = ΨC∗ ((1 + e)γ) = (1 + e)ΨC∗ (γ)

⇒ ∂ΨC∗ (ξ) = ∂ΨC∗ (γ).

We now replace the differential inclusion (5.92), which holds for almost
all t, by an equality of measures

M (q)du − h(q, u)dt = WN (q)dPN + WD (q)dPD ∀t, (5.96)

which holds for all time-instances t. The differential measure of the contact
percussions dPN and dPD contains a Lebesgue measurable part λdt and an
atomic part Λdη

dPN = λN dt + ΛN dη, dPD = λD dt + ΛD dη, (5.97)

which can be expressed as inclusions:

−dPN i ∈ ∂ΨC∗ N (ξN i )dt + ∂ΨC∗ N (ξN i )dη,

(5.98)
−dPDi ∈ ∂ΨC∗ Di (λN i ) (ξDi )dt + ∂ΨC∗ Di (ΛN i ) (ξDi )dη.

The set ∂ΨC∗ N (ξN i ) is a cone and the measure constitutive law for normal con-
tact (5.98) and positive measures dt and dη can therefore be written as (4.56)

−dPN i ∈ ∂ΨC∗ N (ξN i ), (5.99)

which means that the density functions −λN i and −ΛN i both belong to this
cone (see Section 4.3). The fundamental idea behind the equality of mea-
sures (5.96) is that it describes the differential equation for impact-free mo-
tion and the impact equations by a single equation. Indeed, if we integrate the
equation of measures over an arbitrary impact-free time-interval I for which
ΛN = ΛD = 0, then we obtain the integral equation
   
M (q)du = h(q, u)dt + WN (q)λN dt + WD (q)λD dt.
I I I I

Hence, the differential measure du can only have a density with respect to the
Lebesgue measure dt, i.e. du = u̇dt, from which we retrieve the differential
equation of motion (5.79)

M (q)u̇ = h(q, u) + WN (q)λN + WD (q)λD .

104 5 Mechanical Systems with Set-valued Force Laws

Similarly, if we consider an integration interval consisting of a singleton {ti } for

which the impulses ΛN and/or ΛD are non-zero, then we obtain the integral
equation
  
M (q)du = WN (q)ΛN dη + WD (q)ΛD dη,
{ti } {ti } {ti }

where we already used that the integral of the Lebesgue measure dt over the
singleton {ti } vanishes. Hence, the differential measure du must have a density
with respect to the atomic measure dη, i.e. du = (u+ − u− )dη, from which
we retrieve the impact equation which descibes the velocity jump at t = ti :

M (q)(u+ − u− ) = WN (q)ΛN + WD (q)ΛD . (5.100)

As an abbreviation of the equality of measures (5.96) we write

M (q)du − h(q, u)dt = W (q)dP ∀t, (5.101)

using    
dPN   γ
dP = , W = W N WD , γ= N . (5.102)
dPD γD
Furthermore we introduce the variables ξ and δ

ξ = γ + + Eγ − , δ = γ+ − γ−, (5.103)

with E = diag({eN i , eDi }) from which we deduce

γ + = (I + E)−1 (ξ + Eδ),
(5.104)
γ − = (I + E)−1 (ξ − δ).

The equality of measures (5.101) together with the set-valued force laws (5.98)
form a measure differential inclusion which describes the time-evolution of a
mechanical system with discontinuities in the generalised velocities.
Here, we introduced the restitution coefficient matrix E, which has the
diagonal elements eii . In Chapter 7, we will make use of the dissipation index
matrix Δ defined by
Δ := (I − E)(I + E)−1 , (5.105)
which is a diagonal matrix with dissipation indices Δii = 1−e 1+eii . If |eii | < 1,
ii

then it holds that Δ > 0. Note that Δmax = 1+emin and Δmin = 1−e
1−emin
1+emax .
max

Moreover, if all dissipation indices Δii are equal, then the global dissipation
index δ of Moreau [123] is related to the dissipation index matrix by Δ = δI.
The measure differential inclusion described by (5.101) and (5.98) may
exhibit equilibrium sets, i.e. simply connected sets of equilibrium points. Note
that u = 0 implies γD = 0, see (5.91). This means that every equilibrium
point implies sticking in all closed contact points. We will reserve the word
equilibrium point x∗ for an equilibrium of a measure differential inclusion in
 5.4 Measure Newton-Euler Equations 105

Fig. 5.11. Bouncing ball example.

first-order form and use the term equilibrium position q ∗ to denote an equili-
brium configuration of a Lagrangian mechanical system. Hence, if the measure
differential inclusion (6.1) in first-order form represents the Lagrangian me-
chanical system (5.101) then it holds that
 ∗
∗ q
x = . (5.106)
0

Every equilibrium position q ∗ has to obey the equilibrium inclusion:

 
h(q ∗ , 0) − wN i (q ∗ )∂ΨC∗ N (0) + WDi (q ∗ )∂ΨC∗ Di (0)  0 (5.107)
i∈IN

or using C = ∂ΨC∗ (0)

 
h(q ∗ , 0) − wN i (q ∗ )R− + WDi (q ∗ )CDi  0. (5.108)
i∈IN

An equilibrium set E, being a simply connected set of equilibrium points

(Definition 6.8), therefore obeys
 &
%  
%
E ⊂ (q, u) (u = 0) ∧ h(q, 0) + wN i (q)R − WDi (q)CDi  0
+

i∈IN
(5.109)
and is positively invariant (see Definition 6.4 in Chapter 6) if we assume
uniqueness of the solutions in forward time. It should be noted that due to
the fact that nonlinear mechanical systems, without dry friction, can exhibit
multiple equilibrium points, the system with dry friction may exhibit multi-
ple equilibrium sets. A formal definition of an equilibrium set in a measure
differential inclusion will be given in Definition 6.8.

In order to illustrate the modelling techniques introduced in the chapter,

we study the dynamics of the bouncing ball example shown in Figure 5.11. A
106 5 Mechanical Systems with Set-valued Force Laws

rigid ball with height q and mass m is falling on a rigid floor. Only vertical
motion is considered. The restitution coefficient is eN . The position q(t) is an
absolutely continuous function in time with differential measure

dq = udt, (5.110)

where u is the velocity of the ball. The velocity u is assumed to be of locally

bounded variation with vanishing singular part and its differential measure
therefore contains a Lebesque measurable part and an atomic part

du = u̇dt + (u+ − u− )dη, (5.111)

where u̇ is the acceleration and u+ −u− is a jump in the velocity. The equation
of motion in terms an equality of measures (5.101) gives for this system

mdu + mgdt = dPN , q ≥ 0, (5.112)

where the contact effort dPN obeys the set-valued force law

dPN = 0 if q > 0,
(5.113)
−dPN ∈ ∂ΨC∗ N (ξN ) if q = 0,

with CN = R− and ξN = u+ + eN u− . Simplifying the term ΨC∗ N (ξN ) with

ΨR∗− = ΨR+ and ∂ΨC = NC gives

dPN = 0 if q > 0,
(5.114)
−dPN ∈ NK (ξN ) if q = 0,

where K = {q | q ≥ 0} = R+ is the admissible set of positions q. We can

further abbreviate the set-valued force law to (see [63])

−dPN ∈ NKK (q) (ξN ), (5.115)

where KK is the contingent cone to K. Hence, the system is completely de-

scribed by the measure differential inclusion

dq = udt,
mdu + mgdt = dPN , (5.116)
−dPN ∈ NKK (q) (ξN ), ξN = u+ + eN u− K = {q | q ≥ 0}.

5.5 Summary
The description of Lagrangian mechanical systems with frictional unilateral
contact in the form of measure differential inclusions has been discussed in
this chapter. Set-valued force laws, which are used to describe constitutive
behaviour, can be derived from non-smooth potentials. The non-smooth po-
tential is decomposed in a differentiable function, an indicator function and a
 5.5 Summary 107

convex potential with polyhedral epigraph. The smooth part of the set-valued
force is derived from the differentiable function while the non-smooth set-
valued part stems from the subdifferential of the convex remainder. Set-valued
force laws for sliding and pivoting friction, rolling friction as well as impact
have been set up as inclusions to normal cones on convex sets of admissible
contact forces. The contact forces can be incorporated in the Newton-Euler
equations as Lagrangian multipliers. The Newton-Euler equations in terms
of differential measures, together with set-valued force laws for the contact
forces, form a measure differential inclusion in second-order form. The sta-
bility of equilibrium sets of measure differential inclusions, and in particular
of those stemming from Lagrangian mechanical systems, will be discussed in
Chapters 6 and 7.
This page intentionally blank
6
Lyapunov Stability Theory for Measure
Differential Inclusions

Lyapunov stability theory has originally been developed for smooth ordinary
differential equations. In this chapter, we generalise the stability theory of
Lyapunov to measure differential inclusions of the form

dx ∈ dΓ (x), (6.1)

being time-autonomous, or non-autonomous

dx ∈ dΓ (t, x). (6.2)

We assume that the measure differential inclusion is consistent (see Defini-

tion 4.8) and that it has at least one solution ϕ(t, t0 , x0 ) for all admissible
initial conditions x0 ∈ A, i.e. it holds that S(·, t0 , x0 ) = ∅ ∀t0 , ∀x0 ∈ A(t0 ).
In other words: existence is assumed for all admissible conditions. First, some
mathematical prerequisites are presented in Section 6.1. Subsequently, the
invariance of sets and limit sets of the autonomous measure differential in-
clusion (6.1) are studied in Section 6.2. Definitions of stability properties of
equilibria and positively invariant sets of autonomous measure differential in-
clusions are given in Section 6.3. In Sections 6.2 and 6.3, we restrict ourselves,
for the sake of simplicity, to the case of autonomous measure differential inclu-
sions, because the stability results in Sections 6.5, 6.6 and 6.7 are formulated
for autonomous measure differential inclusions. Moreover, in Chapter 7, we
apply these results to autonomous mechanical systems.
In Section 6.4, however, we propose stability definitions of solutions of non-
autonomous measure differential inclusions, from which stability definitions
of, for example, equilibria of non-autonomous measure differential inclusions
can be directly deducted as non-autonomous variants of the definitions in
Section 6.3. Moreover, these stability notions for solutions of non-autonomous
measure differential inclusions are needed in defining the notion of convergence
for such systems in Chapter 8.
In Section 6.5, we give generalisations of the direct method of Lyapunov.
Subsequently, a generalisation is given of LaSalle’s invariance principle in
110 6 Lyapunov Stability Theory for Measure Differential Inclusions

Section 6.6 and of Chetaev’s instability theorem in Section 6.7 for measure dif-
ferential inclusions (6.1). Although in the latter sections only the autonomous
case is considered, the results can be readily extended to the non-autonomous
case.

6.1 Mathematical Preliminaries

Classically, the direct method of Lyapunov considers a Lyapunov function
V (x), being a (locally) positive definite function, and concludes stability prop-
erties by analysing the time-derivative V̇ (t) = ∇V T ẋ(t) along solution curves
of the system. In this chapter, we will generalise the direct method of Lya-
punov to measure differential inclusions. The inequality constraints on the
state x(t), which are implicitly defined by the measure differential inclusion,
need to be taken into account in the Lyapunov function. For this reason,
we will consider a Lyapunov function V : Rn → R ∪ {∞} which can take
infinite values on its domain. Such a function is called an extended func-
tion. Furthermore, we will need to generalise the notion of the derivative of
the Lyapunov function V . By considering a lower-semi-continuous Lyapunov
function V (x(t), we are able to define a subderivative of V if x(t) is absolutely
continuous in time.
An extended lower semi-continuous function V (x), where V : Rn → R ∪
{∞}, is called a
• radially unbounded function if

V (x) → ∞ as x → ∞ , (6.3)

• locally positive definite function (LPDF) if there exists an h > 0 such

that
V (0) = 0 and V (x) > 0 ∀x ∈ Bh \{0}, (6.4)
• positive definite function (PDF) if V (x) is radially unbounded and

V (0) = 0 and V (x) > 0 ∀x = 0. (6.5)

Definition 6.1 (Class K function and class KR function). A function

α : R+ → R+ , with α(0) = 0, which is continuous and strictly increasing is
called a class K function. If in addition α(x) → ∞ as x → ∞, then function
α is called a class KR function.

Hence, V (x) is PDF if and only if it is bounded from below by a class KR

function α(x) [152], i.e.

V (x) ≥ α(x) ∀x ∈ Rn . (6.6)

In the above definition, a PDF function is by definition radially unbounded.

Other authors, e.g. [85], do not immediately associate positive definiteness
 6.1 Mathematical Preliminaries 111

with radially unboundedness and have to state radially unboundedness ex-

plicitly when required.
The function V (x) is called quadratic if it is of the form V (x) = xT Bx
with B ∈ Rn×n . It is always possible to express a quadratic form with a
symmetric matrix A = 12 (B + B T ) as

V (x) = xT Ax, A = AT ∈ Rn×n . (6.7)

If the symmetric matrix A ∈ Rn×n is positive definite (PD), i.e.

xT Ax > 0 ∀x = 0, (6.8)

then it holds that

λmin (A) x2 ≤ xT Ax ≤ λmax (A) x2 , ∀x (6.9)

where λmin (A), λmax (A) > 0 are the minimal and maximal eigenvalues of A
and the quadratic function (6.7) is therefore radially unbounded. By definition
it holds that V (x) = xT Ax > 0 for all x = 0. A quadratic function V with
positive definite matrix A is therefore PDF. For a quadratic positive definite
function V it holds that

xT ∇V (x) = 2xT Ax > 0 ∀x = 0, (6.10)

which means that the vector x and the gradient always form an acute angle.
For an extended lower semi-continuous function V : Rn → R ∪ {∞} and
c ≥ 0 we define the level set Ωc as

Ωc = {x ∈ Rn | V (x) ≤ c}, (6.11)

with the corresponding level surface Lc

Lc = bdry Ωc = {x ∈ Rn | V (x) = c}. (6.12)

The level sets of quadratic positive definite functions are ellipsoids centred
at the origin (see Figure 6.1). Level sets of (extended) lower semi-continuous
PDF functions are closed and bounded.
Positive definite functions have important properties.
Proposition 6.2. If an extended lower semi-continuous function V is a PDF,
then there exists a c∗ > 0 such that each level set Ωc = {x ∈ Rn | V (x) ≤ c}
with c ≤ c∗ is simply connected.
Proof: Reductio ad absurdum: if Ωc is not simply connected then there exists
a local minimum of V (x) at xmin = 0 with V (xmin ) ≤ c. If V (x) is PDF then
V (x) > 0 for all x = 0, and the origin is therefore the global minimum and
the only minimum of V (x) in some neighbourhood Bε of the origin. For this
Bε , there exists a c∗ such that Ωc ⊂ Bε for all c ≤ c∗ . Hence, if Ωc ⊂ Bε ,
112 6 Lyapunov Stability Theory for Measure Differential Inclusions

Fig. 6.1. Level surfaces of a quadratic positive definite function.

then the origin is the only minimum in Ωc , which implies that Ωc is simply
connected if c ≤ c∗ . 
Quadratic PDF’s have a unique minimum at x = 0 and each level set is
therefore simply connected.
Let the state x(t) of a system be governed by the autonomous ordinary
differential equation ẋ(t) = f (x(t)). A function V (x(t)) depends therefore
implicitly on time t. If V (x) is continuous in x and x(t) is continuous in t,
then V (x(t)) varies continuously in time. A smooth function V decreases if it
holds that
∂V dx
V̇ (x(t)) = = ∇V (x(t))T f (x(t)) < 0 (6.13)
∂x dt
along trajectories x(t) of the system.
If the state x(t) of a system is governed by a measure differential inclu-
sion (6.1), then V (x(t)) is in general not continuous in t because x(t) might
have discontinuities in t (even if V is continuous in x). Moreover, if V is dis-
continuous in x, then V (x(t)) is in general discontinuous in t. When studying
measure differential inclusions, we relax the continuity with respect to time
of functions V (x(t)) to locally bounded variation in time.

Proposition 6.3 (Lyapunov functions of locally bounded variation).

Let V : X → R be a function which is Lipschitz continuous on the closed
domain D ⊂ X with Lipschitz constant K. If it holds that x ∈ lbv(I, D), and
consequently x(t) ∈ D ∀t ∈ I, then it holds that the function Ṽ (t) = V ◦ x =
V (x(t)) is of locally bounded variation on I.

Proof: Because V : X → R is Lipschitz continuous on the closed domain

D ⊂ X, it holds that

∃K < ∞, V (x) − V (y) ≤ Kx − y ∀x, y ∈ D. (6.14)

The variation of Ṽ on a compact interval [a, b] ⊂ I gives

 6.1 Mathematical Preliminaries 113

n
var(Ṽ , [a, b]) = sup Ṽ (ti ) − Ṽ (ti−1 )
i=1
n
= sup V (x(ti )) − V (x(ti−1 ))
i=1
(6.15)

n
≤ K sup x(ti ) − x(ti−1 )
i=1
≤ K var(x, [a, b])
< ∞,

since x(t) is of locally bounded variation. This proves that Ṽ (t) = V ◦ x =

V (x(t)) is of locally bounded variation on I. 
In particular, if V (x) = v(x) + ΨD (x), where v : X → R is a Lipschitz
continuous function on X and x ∈ lbv(I, D), then it follows that Ṽ = V ◦ x ∈
lbv(I, R).
We consider the differential measure of V to be composed of an absolutely
continuous part and an atomic part:
dV := d(V ◦ x) = V̇ (x(t))dt + (V + (x(t)) − V − (x(t)))dη. (6.16)
The function V (x) will be assumed to be an extended lower semi-continuous
function V : Rn → R ∪ {∞}. For a number of special cases we can express the
differential measure dV in more detail:
1. If V (x) is a locally continuous function in x, then it holds that V + (x(t)) =
V (x+ (t)) and V − (x(t)) = V (x− (t)).
2. If x(t) is an absolutely continuous function in time and V is an extended
lower semi-continuous function, then V admits a subderivative dV (x)(v)
at x in the direction v (see Section 2.6). Using the chain rule we express
the differential measure of V in the differential measure dx = ẋdt as
dV = dV (x)(dx), dx ∈ dΓ (x) = F (x)dt. (6.17)
Note that dV denotes the differential measure of V and dV is used to
denote the subderivative of V .
3. If x(t) is of locally bounded variation and if V is a quadratic form V (x) =
1 T T
2 x Ax, with A = A, then it follows from (3.64) that

dV = (x+ + x− )TAdx. (6.18)

Substitution of dx = ẋdt + (x+ − x− )dη gives
dV = (x+ + x− )TA(ẋdt + (x+ − x− )dη)
= xTAẋdt + (x+ + x− )T A(x+ − x− )dη
T T (6.19)
= xTAẋdt + (x+ Ax+ − x− Ax− )dη
= (∇V (x)) ẋdt + (V (x+ ) − V (x− ))dη.
T
114 6 Lyapunov Stability Theory for Measure Differential Inclusions

In many practical applications (e.g. Lagrangian mechanical systems with uni-

lateral constraints), the system is such that some of the states are absolutely
continuous (e.g. generalised positions) while the other states are of locally
bounded variation (e.g. generalised velocities). If the function V consists of
a lower semi-continuous function, dependent on the absolutely continuous
states, and a quadratic function, dependent on the other states, then the
differential measure of V can be obtained using the above special cases.

6.2 Invariant Sets and Limit Sets

The state-space of a dynamical system may contain equilibrium points and

sets, periodic solutions and other types of attractors, repellors or saddles such
as quasi-periodic solutions and chaotic attractors. For smooth dynamical sys-
tems, these kinds of motions have one thing in common: each of these motions
forms a point or set in the state-space from which the solution curves remain
within the set. We denote these sets as invariant sets. In this section, we con-
sider in particular positively invariant sets of autonomous measure differential
inclusions of the form (6.1).
Definition 6.4 (Positively Invariant Set). A set M ⊂ Rn is said to be
positively invariant with respect to the measure differential inclusion (6.1) if
for every initial condition x0 ∈ M it holds that each forward solution curve
ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) = ∅ satisfies

ϕ(t, t0 , x0 ) ∈ M for almost all t > t0 .

The solution ϕ(t, t0 , x0 ) is not required to stay in M on its discontinuity

points ti , because ϕ(ti , t0 , x0 ) is not defined. However, due to the fact that
ϕ(t, t0 , x0 ) is of locally bounded variation, it still holds on a discontinuity
point ti > t0 that

ϕ(ti , t0 , x0 )+ ∈ M, ϕ(ti , t0 , x0 )− ∈ M,

when M is a positively invariant set of (6.1). In other words, the solution

before and after the discontinuity are in M. Moreover, every positively in-
variant set M of (6.1) is a subset of the admissible set A, which is the largest
positively invariant set of the system. The solution curve ϕ(·, t0 , x0 ) is not
necessarily unique. Positive invariance of a set M means that each solution
curve of (6.1) starting from (t0 , x0 ) remains within M for all x0 ∈ M and for
almost all succeeding times t > t0 . This is sometimes referred to as strongly
invariant, in contrast to weakly invariant which only demands that at least
one solution curve remains within M [37, 156]. A negatively invariant set M
consists of initial conditions from which all backward solution curves remain
within M, i.e. ϕ(t, t0 , x0 ) ∈ M for almost all t < t0 . A set which is both
positively and negatively invariant is called an invariant set.
 6.2 Invariant Sets and Limit Sets 115

Positive invariance of a level set Ωc (6.11) of a function V (x) with re-

spect to a measure differential inclusion can be proven using the following
proposition.
Proposition 6.5. Let Ωc = {x ∈ Rn | V (x) ≤ c} be a bounded set. Consider
the differential measure dV (6.16) and the measure differential inclusion (6.1).
If dV ≤ 0 for all x ∈ Ωc , then the set Ωc is positively invariant with respect
to (6.1).
Proof: The function V can not increase in Ωc because dV ≤ 0 for all x ∈ Ωc .
For x0 ∈ Ωc it holds that

V (x+ (t)) = V (x0 ) + dV
[t0 ,t]

≤ V (x0 )
≤c

for all solution curves x(t) = ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ). Consequently, all

solution curves x(t) remain in Ωc for almost all t (except at time-instances
for which x(t) is not defined). 
Note that if V is a PDF, and therefore by definition radially unbounded,
then Ωc is bounded for all bounded values c. Moreover, the level set Ωc is
in general not simply connected and may consist of connected components
Ωci , i = 1, 2, . . . . For example, consider the level set Ωc in Figure 6.2 which
consists of Ωc1 , Ωc2 and Ωc3 . We assume dV ≤ 0 in Ωc . If a solution x(t)
of (6.1) starts in Ωc2 , then the function V (x) will not increase along the
solution curve, forcing the solution to stay within the level set Ωc . Hence, the
level set is positively invariant. However, the solution x(t) might jump during
the time-evolution to another connected component of Ωc , e.g. to Ωc3 as is
depicted in Figure 6.2. Consequently, the connected components of a positively
invariant level set Ωc are therefore generally not positively invariant.
The most elementary invariant set in a smooth dynamical system is an
invariant equilibrium point.
Definition 6.6 (Equilibrium Point). A point x∗ is called an equilibrium
point of (6.1) if there exists a forward solution curve ϕ(·, t0 , x∗ ) ∈ S(dΓ , t0 , x∗ )
such that
ϕ(t, t0 , x∗ ) = x∗ , ∀t > t0 .

Corollary 6.7. It holds that 0 ∈ dΓ (x∗ ) if and only if x∗ is an equilibrium

point of (6.1).

Proof: If 0 ∈ dΓ (x∗ ), then there exists a forward solution curve ϕ(t, t0 , x∗ ) =

x∗ for all t > t0 , i.e. x∗ is an equilibrium point. Moreover, because an equili-
brium point x∗ is a constant solution of (6.1), it must hold that dx = 0 along
a constant solution curve ϕ(t, t0 , x∗ ) = x∗ , i.e. 0 ∈ dΓ (x∗ ). 
116 6 Lyapunov Stability Theory for Measure Differential Inclusions

Fig. 6.2. Non-invariance of connected components Ωci of an invariant level set Ωc .

Non-uniqueness of solutions in non-smooth dynamical systems may destroy

the invariance of an equilibrium point. If the forward solution from an equili-
brium point is non-unique, then there exists a solution curve which does not
remain on the equilibrium point. The equilibrium point is in this case not a
positively invariant set. Non-smooth dynamical systems may also exhibit sim-
ply connected sets of equilibrium points which may or may not be positively
invariant.
Definition 6.8 (Equilibrium Set). A simply connected set E is called an
equilibrium set of (6.1) if every point x∗ ∈ E is an equilibrium point and if
there exists no equilibrium point x∗ ∈
/ E infinitely close to E.
Periodic solutions in smooth dynamical systems form invariant closed orbits1
in the state-space. Periodic solutions in non-smooth systems still have a peri-
odicity property

ϕ(t, t0 , x0 ) = ϕ(t + kT, t0 , x0 ), k = 1, 2, 3, . . . , ∀t > t0 , t = ti , (6.20)

but the solution ϕ(t, t0 , x0 ) is not defined on the discontinuity points {ti }.
The image of the periodic solution is in general therefore not a closed orbit.
We therefore will use the term periodic orbit for the set O defined by

O = cl{x ∈ Rn | x = ϕ(t, t0 , x0 ), t0 ≤ t ≤ t0 + T }, (6.21)

which is the closure of the image of the periodic solution ϕ(t, t0 , x0 ). The set O
is closed, in the sense that it contains its boundary, but O is not necessarily
a closed orbit (Figure 6.3). If the periodic solution ϕ(t, t0 , x0 ) is positively
invariant, then its corresponding periodic orbit is a closed positively invariant
set.
1
The term orbit shows the influence of astronomy on the terminology in stability
theory.
 6.2 Invariant Sets and Limit Sets 117

Fig. 6.3. Periodic solution ϕ(t, t0 , x0 ) and periodic orbit O within the admissible
set A.

Closely related to the invariance of a set is the behaviour of solution curves

when time approaches infinity. After Birkhoff [20], we introduce the terms
positive limit point and set (see also [70, 85, 89]).
Definition 6.9 (Positive Limit Point). A point p ∈ Rn is a positive limit
point of a solution curve ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) of (6.1) if there exists
a sequence {tj } with tj → +∞ as j → ∞, such that ϕ(tj , t0 , x0 ) → p for
j → ∞.

Definition 6.10 (Positive Limit Set). A positive limit set L+ of a solution

curve ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) is the set of all its positive limit points.

A positive limit set L+ is a set which is approached for t → ∞. Similarly, a

negative limit set L− is a set which is approached for t → −∞. For smooth
dynamical systems it holds, under the assumption of bounded state behaviour,
that every positive limit set is an invariant set. This assertion is in general not
true for measure differential inclusions (6.1). Consider for instance the scalar
measure differential inclusion

1, x− (t) = 0,
dx = −x(t)dt + a(x(t))dη, with a(x(t)) = (6.22)
0, x− (t) = 0.

The system is described by the differential equation ẋ(t) = −x(t), with the
solution x(t) = x0 e−(t−t0 ) for x0 = 0. If x− (t) = 0, then an impulse occurs
which brings the state to x+ (t) = 1. A solution with x0 = 0 will asymptotic-
ally approach the origin p = 0, but will never reach it. The point p = 0 is
therefore a positive limit point, but is not a positively invariant point because
ϕ(t, t0 , p) = p for t > t0 . The non-invariance of the limit point p is caused
by a discontinuous dependence of the solution on the initial condition. Con-
tinuous dependence with respect to the initial condition is defined in [50] for
differential inclusions. Here, we generalise the definition to measure differential
inclusions.
118 6 Lyapunov Stability Theory for Measure Differential Inclusions

Definition 6.11 (Continuous Dependence on the Initial Condition).

Consider an autonomous measure differential inclusion dx ∈ dΓ (x(t)) with
the admissible set A. The system has a continuous dependence on the initial
condition if for any finite time-interval [t0 , t1 ] and any solution ϕ(t, t0 , x01 ) ∈
S(Γ , t0 , x01 ) with x01 ∈ A it holds that for each ε there exists a δ(ε) such that
x02 − x01  < δ ⇒ ϕ(t, t0 , x02 ) − ϕ(t, t0 , x01 ) < ε for almost all t ∈ [t0 , t1 ],
for all ϕ(t, t0 , x02 ) ∈ S(Γ , t0 , x02 ) with x02 ∈ A.
Continuous dependence on the initial condition implies that if two initial
points are infinitely close to each other, then their solutions curves remain
infinitely close. Non-uniqueness of solutions destroys this property. Hence, if
the measure differential inclusion has a continuous dependence on the initial
condition, then the solution is unique, i.e. S(dΓ , t0 , x0 ) = {ϕ(·, t0 , x0 )}. The
converse is not true: a system can have uniqueness of solutions but have a
discontinuous dependence of solutions on the initial conditions. The following
proposition (taken from [33, 34]) states a sufficient condition for a positive
limit set to be invariant:
Proposition 6.12 (Positive Invariance of a Positive Limit Set [33]).
If (6.1) is consistent and the solutions have a continuous dependence on the
initial condition, then each positive limit set L+ is positively invariant.
Proof: Let p ∈ L+ be a positive limit point of a solution curve ϕ(·, t0 , x0 ).
Then there exists a sequence {tj } with tj → +∞ as j → ∞, such that
ϕ(tj , t0 , x0 ) → p for j → ∞. Uniqueness of solutions in forward time is
guaranteed because of the continuous dependence on initial conditions, and
allows us to speak of the solution curve. Moreover, if there is a continuous
dependence on the initial condition, then the solution curve starting from p
is identical to the solution curve starting from ϕ(tj , t0 , x0 ), i.e.
ϕ(t, t0 , p) = lim ϕ(t, t0 , ϕ(tj , t0 , x0 )),
j→∞

for almost all t ≥ t0 . The concatenation of solution curves ϕ(t, t0 , ϕ(tj , t0 , x0 ))

is equivalent to ϕ(t + tj − t0 , t0 , x0 ) (where we again used uniqueness of solu-
tions in forward time as well as the fact that (6.1) is time-autonomous), which
gives after substitution in the above equation
ϕ(t, t0 , p) = lim ϕ(t + tj − t0 , t0 , x0 ),
j→∞

for almost all t ≥ t0 . Consequently, for almost all t ∈ [t0 , ∞) there exists a
sequence {tj } such that the solution curve ϕ(t + tj − t0 , t0 , x0 ) converges to
ϕ(t, t0 , p). This means that ϕ(t, t0 , p) is a limit point of the solution curve
ϕ(·, t0 , x0 ) for almost all t ∈ [t0 , ∞). Using the definition of L+ , it holds for
each p ∈ L+ that
ϕ(t, t0 , p) ∈ L+ , for almost all t ≥ t0 ,
which shows the positive invariance of the positive limit set L+ . 
 6.3 Definitions of Stability Properties for Autonomous Systems 119

Fig. 6.4. The concept of Lyapunov stability of an equilibrium point.

6.3 Definitions of Stability Properties for Autonomous

Systems
In this section we will give definitions of stability properties of equilibria and
positively invariant sets of autonomous measure differential inclusions of the
form (6.1) which are autonomous in time. Here, we restrict ourselves to the
case of autonomous measure differential inclusions, because the stability re-
sults in Sections 6.5, 6.6 and 6.7 are formulated for autonomous measure
differential inclusions.
The stability definition of Lyapunov defines an equilibrium point to be
stable if all neighbouring solutions remain in the neighbourhood of the equi-
librium point.
Definition 6.13 (Lyapunov Stability of an Equilibrium Point).
An equilibrium point x∗ of the consistent time-autonomous measure differen-
tial inclusion (6.1) is Lyapunov stable if for each ε > 0 there exists a δ(ε) > 0
such that for any x0 ∈ A with

x0 − x∗  < δ

each solution curve ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) = ∅ satisfies

ϕ(t, t0 , x0 ) − x∗  < ε for almost all t ≥ t0 .

The definition of Lyapunov involves an ε-δ argumentation. We consider a

tube with radius ε around the equilibrium point x∗ . The stability concept
demands that neighbouring solutions of ϕ(t, t0 , x∗ ) = x∗ have to remain in
the ε-tube, when the admissible initial condition x0 is chosen close enough to
120 6 Lyapunov Stability Theory for Measure Differential Inclusions

x∗ (in a δ-sphere around x∗ ). An initial condition is admissible if x0 ∈ A and

each solution curve ϕ(t, t0 , x0 ) remains within the admissible set for almost
all t > t0 , because the system is assumed to be consistent. For every ε-tube
we have to seek a sphere with radius δ centred at x∗ for which holds that
all solutions with admissible initial conditions in this δ-sphere remain in the
ε-tube (see Figure 6.4). If we can indeed find such a δ for every ε, then the
equilibrium point is said to be Lyapunov stable. Obviously, the δ we choose
depends on ε, which we express by δ(ε) and it obviously holds that δ(ε) ≤ ε.
An essential point in the definition is the fact that we have to find such a δ-
sphere for arbitrary small values of ε, which leads to the following proposition.
Proposition 6.14. A stable equilibrium point is positively invariant.
Proof: Using the definition of stability and taking the limit ε ↓ 0 gives
ϕ(t, t0 , x0 ) − x∗  < ε ↓ 0 with x0 − x∗  < δ ≤ ε ↓ 0. We deduce that
ϕ(t, t0 , x∗ ) = x∗ for almost all t ≥ t0 . The equilibrium point x∗ is therefore
positively invariant. 
Stability therefore implies positive invariance. The neighbouring solutions
which start in the admissible part of the δ-sphere are required to remain
in the ε-tube, that is to say they stay in the neighbourhood. The neighbour-
ing solutions around a stable equilibrium point do not necessarily converge
towards the equilibrium point. This additional property is called attractivity.
Definition 6.15 (Local Attractivity of an Equilibrium Point). An
equilibrium point x∗ of the measure differential inclusion (6.1) is locally at-
tractive if there exists a δ > 0 such that for any x0 ∈ A with

x0 − x∗  < δ

each solution curve ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) converges to x∗ in positive

time:
lim ϕ(t, t0 , x0 ) − x∗  = 0.
t→∞

An attractive equilibrium point has a domain of attraction. The admissible

part of the δ-sphere in the above definition, i.e. Bδ ∩ A, forms a conservative
estimate for the domain of attraction. If the domain of attraction extends to
the whole admissible set A, then the equilibrium point is said to be globally
attractive.
Definition 6.16 (Global Attractivity of an Equilibrium Point). An
equilibrium point x∗ of the measure differential inclusion (6.1) is globally
attractive if each solution curve ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) converges to x∗ :

lim ϕ(t, t0 , x0 ) − x∗  = 0 ∀x0 ∈ A.

t→∞

An equilibrium point of a smooth dynamical system which is both stable and

locally/globally attractive is called in literature locally/globally ‘asymptotic-
ally stable’. The words ‘local’ or ‘global’ refer to the asymptotic behaviour of
 6.3 Definitions of Stability Properties for Autonomous Systems 121

the solution curves in the neighbourhood of the equilibrium point, i.e. to the
attractivity property and not to the stability (note that Lyapunov stability
is in essence a local requirement, see Definition 6.13). Solution curves of a
smooth dynamical system can never meet each other because of the unique-
ness of solutions in forward and backward time. The attractivity in smooth
systems is therefore always asymptotic 2 in the sense that neighbouring solu-
tion curves approach but never reach the equilibrium point when t → ∞. The
attractivity of an equilibrium point of a non-smooth system is not necessarily
asymptotic as it might be reached in a finite time. For instance, if we consider
the differential inclusion (4.9), ẋ ∈ 1 − 2 Sign(x) with x(0) = x0 , then we
immediately see that the equilibrium position x = 0 is reached when t = x0
if x0 > 0 or t = − 13 x0 if x0 < 0. We therefore refrain from the terminology
‘asymptotic stability’ to denote an equilibrium point which is both attractive
and stable. Instead, we will use the terminology attractive stability. If the so-
lution curves converge asymptotically to the equilibrium point, then we speak
of asymptotic attractivity. If the solution curves converge to the equilibrium
point in a finite time, then we speak of symptotic attractivity.
Definition 6.17 (Symptotic Attractivity of an Equilibrium Point).
An equilibrium point x∗ of the measure differential inclusion (6.1) is symp-
totically attractive if there exists a δ > 0 such that for any bounded x0 ∈ A
with
x0 − x∗  < δ
each solution curve ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) reaches x∗ in a finite time, i.e.

ϕ(T + t0 , t0 , x0 ) = x∗ ,

with T < ∞.
Symptotic attractive stability is sometimes called finite-time stability in lit-
erature [19].
The notion of stability and attractivity of an equilibrium point can be
extended to the stability and attractivity of a set (e.g. an equilibrium set,
periodic orbit or chaotic attractor). A stable set is necessarily positively in-
variant. Let M be a closed positively invariant set of (6.1). Define the open
ε-neighbourhood of M by

Uε (M) = {x ∈ Rn | distM (x) < ε}, (6.23)

where distM (x) is the minimal distance from x to a point in M (see Defini-
tion 2.32).
2
The word asymptote stems from the Greek words a (not) and sympiptein (to
meet) and therefore means ‘not meeting’. Proclus Diadochus (411-485 A.D.)
writes in his Commentary on Euclid’s Elements about asymptotic lines as well
as symptotic lines (those that do meet).
122 6 Lyapunov Stability Theory for Measure Differential Inclusions

Definition 6.18 (Stability of a Closed Positively Invariant Set). A

closed positively invariant set M of (6.1) is stable if for each ε > 0 there
exists a δ(ε) > 0 such that for any x0 ∈ A with

x0 ∈ Uδ (M)

each solution curve ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) satisfies

ϕ(t, t0 , x0 ) ∈ Uε (M) for almost all t ≥ t0 .

Definition 6.19 (Local/Global Attractivity of a Closed Positively In-

variant Set). A closed positively invariant set M of (6.1) is locally attractive
if there exists a δ > 0 such that for any x0 ∈ A with

x0 ∈ Uδ (M)

each solution curve ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) converges to M:

lim distM (ϕ(t, t0 , x0 )) = 0.

t→∞

If convergence occurs for all x0 ∈ A, then M is called globally attractive.

If a positively invariant equilibrium set E is stable in the sense of Defini-

tion 6.18, then each equilibrium point x∗ ∈ E is Lyapunov stable in the sense
of Definition 6.13. However, if a positively invariant orbit O is stable in the
sense of Definition 6.18, then neighbouring solutions of the periodic solution
ϕ(t, t0 , x0 ) stay in the neighbourhood of O, but do not necessarily stay close
to ϕ(t, t0 , x0 ) because they generally encircle the orbit O in a time differ-
ent from the period time of the periodic solution. A periodic solution with a
‘stable’ behaviour, i.e. it is associated with a stable closed orbit, is therefore
called orbitally stable or Poincaré stable.

6.4 Definitions of Stability Properties of Solutions

Non-autonomous Systems

The previous section dealt with the definition of stability properties of posi-
tively invariant sets of time-autonomous systems. In this section we will take
a more general perspective and define stability properties for (time-varying)
solutions of non-autonomous systems. The definitions for (uniform) stability
and attractivity of solutions of differential equations have been well-defined,
see [38,132,176]. Here, we generalise these definitions to differential inclusions
and measure differential inclusions.
6.4 Definitions of Stability Properties of Solutions Non-autonomous Systems 123

6.4.1 Differential Inclusions

Consider the differential inclusion

ẋ ∈ F (t, x), x ∈ Rn , t ∈ R, (6.24)

where the function F (t, x) is a set-valued function. A solution x(t) of (6.24)

is an absolutely continuous function, defined for all t (at least locally), which
fulfills (6.24) for almost all t. The set of forward solutions of (6.24) with
x(t0 ) = x0 is denoted by S(F , t0 , x0 ).

Definition 6.20 (Stability). A solution x̄(t) of system (6.24), with x̄(t0 ) =

x̄0 and which is defined on t ∈ (t∗ , +∞), is said to be
• stable if for any t0 ∈ (t∗ , +∞) and ε > 0 there exists a δ = δ(ε, t0 ) > 0 such
that x0 −x̄(t0 ) < δ implies that each forward solution x(t) ∈ S(F , t0 , x0 )
satisfies x(t) − x̄(t) < ε for all t ≥ t0 .
• uniformly stable if it is stable and the number δ in the definition of stability
is independent of t0 .
• attractively stable if it is stable and for any t0 ∈ (t∗ , +∞) there exists
δ̄ = δ̄(t0 ) > 0 such that x0 − x̄0  < δ̄ implies that each forward solution
x(t) ∈ S(F , t0 , x0 ) satisfies limt→+∞ x(t) − x̄(t) = 0 for all t ≥ t0 .
• uniformly attractively stable if it is uniformly stable and there exists δ̄ > 0
(independent of t0 ) such that for any ε > 0 there exists T = T (ε) > 0 such
that x0 − x̄0 | < δ̄ for t0 ∈ (t∗ , +∞) implies that each forward solution
x(t) ∈ S(F , t0 , x0 ) satisfies x(t) − x̄(t)| < ε for all t ≥ t0 + T .

Note that if x̄(t) is a stable solution, then it must be the unique forward so-
lution from (t0 , x̄0 ), i.e. S(F , t0 , x̄0 ) = {x̄(·)}. Definitions of global attractive
stability and global uniform attractive stability of a solution can be given in
a similar way.

6.4.2 Measure Differential Inclusions

Consider the measure differential inclusion (6.2). As has been discussed in

Section 4.3, a solution of (6.2) is a function of locally bounded variation that
fulfills (6.2) in a measure sense for all t. A solution x(t) of (6.2) is defined for
almost all t, i.e. not for a Lebesgue negligible set of time-instance for which
the solution x(t) jumps. A measure differential inclusion usually describes a
physical process. Set-valued force-laws restrict the state x in (6.2) to some
admissible set A(t). For instance, contact laws and restitution laws prohibit
penetration of a unilateral contact in a mechanical system and therefore re-
strict the position to some admissible set.

Definition 6.21 (Stability). A solution x̄(t) of system (6.2), with x̄(t0 ) =

x̄0 ∈ A(t0 ) and which is defined on t ∈ (t∗ , +∞), is said to be
124 6 Lyapunov Stability Theory for Measure Differential Inclusions

• stable if for any t0 ∈ (t∗ , +∞) and ε > 0 there exists a δ = δ(ε, t0 ) > 0
such that x0 − x̄(t0 ) < δ with x0 ∈ A(t0 ) implies that each forward
solution x(t) ∈ S(dΓ , t0 , x0 ) satisfies x(t) − x̄(t) < ε for almost all
t ≥ t0 .
• uniformly stable if it is stable and the number δ in the definition of stability
is independent of t0 .
• attractively stable if it is stable and for any t0 ∈ (t∗ , +∞) there exists
δ̄ = δ̄(t0 ) > 0 such that x0 − x̄0  < δ̄ with x0 ∈ A(t0 ) implies that each
forward solution x(t) ∈ S(dΓ , t0 , x0 ) satisfies limt→+∞ x(t) − x̄(t) = 0
for almost all t ≥ t0 .
• uniformly attractively stable if it is uniformly stable and there exists δ̄ > 0
(independent of t0 ) such that for any ε > 0 there exists T = T (ε) > 0 such
that x0 − x̄0 | < δ̄ with x0 ∈ A(t0 ) for t0 ∈ (t∗ , +∞) implies that each
forward solution x(t) ∈ S(dΓ , t0 , x0 ) satisfies x(t) − x̄(t)| < ε for almost
all t ≥ t0 + T .

So far, stability properties have been defined for equilibrium points, posi-
tively invariant sets and, in the current section, of any (time-varying) solution
of the system under study. Let us now discuss a stability property on system
level, called incremental stability, which implies the stability of all solutions of
a system (with respect to each other) [7]. Incremental stability reflect a kind
of contraction property of the system. We will define incremental stability for
non-autonomous measure differential inclusions.
Definition 6.22 (Incremental Stability of a Measure Differential In-
clusion). The measure differential inclusion (6.2) is
• incrementally stable if for all t0 , any x1 , x2 ∈ A(t0 ) and all corresponding
solution curves ϕ(·, t0 , xi ) ∈ S(dΓ , t0 , xi ) (i = 1, 2), it holds that for any
ε > 0 there exists a δ = δ(ε, t0 ) such that x1 − x2  < δ implies that
ϕ(t, t0 , x1 ) − ϕ(t, t0 , x2 ) < ε, for almost all t ≥ t0 .
• In addition, the system (6.2) is called attractively incrementally stable if
it is incrementally stable and

lim ϕ(t, t0 , x1 ) − ϕ(t, t0 , x2 ) = 0, ∀t0 , ∀x1 , x2 ∈ A(t0 ).

t→∞

If the system is attractively incrementally stable, then all solutions converge

to one another. In Chapter 8, we will discuss the property of convergence.
Many Lyapunov characterisations of convergence (based on quadratic Lya-
punov functions) also imply incremental stability.
Remark: Note that the definition of stability as given in Definition 6.21,
which is based on comparing different solutions on the same time instant, may,
in a certain sense, be rather restrictive for solutions with jumps. Namely, it
excludes certain situations that we may intuitively consider to be stable. Let
us, for example, consider a particular solution x̄(t) (where x is scalar), of which
we study the stability properties, see Figure 6.5. If all other solutions, such as
6.4 Definitions of Stability Properties of Solutions Non-autonomous Systems 125

Fig. 6.5. Solution x̄(t) is a stable solution in the sense of Definition 6.21.

e.g. x1 (t), x2 (t) and x3 (t), behave as depicted in Figure 6.5, then we would
call x̄(t) stable (or even attractively stable) in the sense of Definition 6.21.
More specifically, on all time instances t̄i , tki corresponding to jump times of
x̄(t) and xk (t), k = 1, 2, 3, . . ., respectively, the distance between x̄(t) and
xk (t) does not increase.
Next, we consider the situation as depicted in Figure 6.6. Clearly, x̄(t) is
not stable in the sense of Definition 6.21 if solutions xk (t) starting arbitrarily
close to x̄(t) at t = t0 do not jump at the same times as x̄(t). However, if all
other solutions xk (t) behave qualitatively similar to x1 (t) in Figure 6.6, then
one would be inclined to call x̄(t) stable (even attractively stable), because
x̄(t) and x1 (t) remain close and converge to each other (both) in a graphical
sense. Therefore, we believe that in order to include such situations, new sta-
bility definitions for time-varying solutions with jumps should be developed,
possibly based on certain graph-closeness properties of solutions with respect
to each other, rather than comparing different solutions on the exact same
time instant.

Fig. 6.6. Solution x̄(t) is not a stable solution in the sense of Definition 6.21.
126 6 Lyapunov Stability Theory for Measure Differential Inclusions

6.5 Basic Lyapunov Theorems of Autonomous Systems

In this section we present generalised versions of basic Lyapunov stability
results, that can be used to prove the (attractive) stability of equilibrium
points and positively invariant sets of time-autonomous measure differential
inclusions. Section 6.5.1 gives an introduction to Lyapunov-type theorems and
presents several theorems that can be used to prove the (attractive) stability
of equilibrium points of measure differential inclusions. These theorems are
generalised in Section 6.5.2 to prove the (attractive) stability of equilibrium
sets. The stability properties of systems with a monotonicity property are dis-
cussed in Section 6.5.3. A sufficient condition is given for incremental stability
of an autonomous measure differential inclusion.

6.5.1 Lyapunov Stability of Equilibrium Points

The fundamental idea behind the original Lyapunov stability result is the
stability theorem of Lagrange-Dirichlet. Consider the mathematical pendulum
equation
ml2 θ̈ + mgl sin θ = 0, (6.25)
in which m is the mass, l is the length of the pendulum, g is the gravita-
tional acceleration and θ designates the angle of the pendulum relative to
its downward hanging equilibrium position. The total energy E of the sys-
tem consists of the kinetic energy T (θ̇) = 12 ml2 θ̇2 and the potential energy
U (θ) = mgl(1 − cos θ):

E := E(θ, θ̇) = T (θ̇) + U (θ). (6.26)

The kinetic energy function is a positive definite function. The potential en-
ergy U (θ) is a locally positive definite function, which implies local positive
definiteness of the total energy E. The downward hanging equilibrium po-
sition (θ, θ̇) = (0, 0) is therefore a local minimum of the total energy. Let
(θ(t), θ̇(t)) be the time-evolution of the system starting from the initial state
(θ0 , θ̇0 ) at t = t0 . Because the system is conservative, the total energy is
constant. The time-evolution therefore has to remain on the level surface
E(θ(t), θ̇(t)) = E(θ0 , θ̇0 ) for all t > t0 . In the neighbourhood of the origin, the
level surfaces of E are concentric ellipses and the trajectories in the neigh-
bourhood of the origin are therefore closed orbits. Neighbouring trajectories
of the origin therefore stay in this neighbourhood. Consequently, the equili-
brium at the origin is stable in the sense of Definition 6.13. Of course, if we
add damping to the system then the system is no longer conservative and the
energy can decrease, i.e. Ė ≤ 0 along trajectories of the system. Hence, the
trajectories remain within the level set E(θ(t), θ̇(t)) ≤ E(θ0 , θ̇0 ) for all t > t0 ,
which leads again to stability of the origin. Summarising, we can prove stabil-
ity of an equilibrium with the Lagrange-Dirichlet theorem if the equilibrium
 6.5 Basic Lyapunov Theorems of Autonomous Systems 127

is a local minimum of the total mechanical energy and if the total mechanical
energy is not increasing along trajectories of the system.
Lyapunov showed that certain other functions can be used to prove the sta-
bility of an equilibrium point. This leads to what is called the direct method of
Lyapunov or method of Lyapunov functions. The theorems of Lyapunov can
therefore be seen as a generalisation of the stability theorem of Lagrange-
Dirichlet. These functions, which we call Lyapunov functions, are charac-
terised by (local) positive definiteness and by the property that they do not
increase, or even strictly decrease, along trajectories of the system. Classi-
cally, smooth Lyapunov functions are used to study stability properties of
time-autonomous ordinary differential equations (4.1). A Lyapunov function
V (x(t)) of a system (4.1) decreases if V̇ (x(t)) = (∇V (x(t)))T f (x(t)) < 0
along trajectories x(t) of the system (see (6.13)). When studying measure dif-
ferential inclusions, we choose Lyapunov functions V (x) which are bounded
for x ∈ A and unbounded for x ∈ / A, where A is the admissible set of the
measure differential inclusion (see Section 4.3). A Lyapunov function V (x)
can therefore be decomposed into
V (x) = v(x) + ΨA (x), (6.27)
where v(x) is a locally bounded function and ΨA (x) is the indicator function
(Definition 2.27) on the admissible set A. We will assume that V is of the
form (6.27) and that v : Rn → R is a smooth continuous function. Lyapunov
functions of the form (6.27) are therefore extended lower semi-continuous
functions. Moreover, the gradient ∇v as well as the subderivative of V exist.
Note that if v(x) is PDF, then also V (x) is PDF, because V (x) ≥ v(x).
If x(t) is absolutely continuous, then we can express the differential mea-
sure dV using the subderivative as in (6.17)
dV = dV (x)(dx)
T
(6.28)
= (∇v(x)) dx + dΨA (x)(dx),
with dx = ẋdt ∈ dΓ (x(t)). The subderivative of the indicator function is
equal to the indicator function of the associated contingent cone (see (2.54))
dΨA (x)(dx) = ΨKA (x) (dx). (6.29)
Moreover, the system is assumed to be consistent (Definition 4.8) and the
differential measure dx = ẋdt lies therefore in the contingent cone of the
admissible set A, i.e. dx ∈ KA (x). Hence, the subderivative of the indicator
function vanishes, dΨA (x)(dx) = 0, and the differential measure of V yields
T
dV = (∇v(x)) dx (6.30)
where x(t) is an absolutely continuous trajectory.
We now present generalised versions of basic theorems of Lyapunov, that
can be used to prove the (attractive) stability of equilibrium points of measure
differential inclusions.
128 6 Lyapunov Stability Theory for Measure Differential Inclusions

Theorem 6.23 (Lyapunov Stability of an Equilibrium Point). Let

x∗ = 0 ∈ A be an equilibrium point of the consistent measure differen-
tial inclusion (6.1), which admits solutions ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) for
all x0 ∈ A. Let V : Rn → R ∪ {∞} be an extended lower semi-continuous
function, which is PDF and of the form (6.27).
a. If dV (x(t)) ≤ 0 ∀x(t) ∈ Bh ∩ A for some h > 0, then the equilibrium
point is stable.
b. If there exists a function β of class K and a constant h > 0 such that
dV (x(t)) ≤ −β(x)dt ∀x(t) ∈ Bh ∩ A, then the equilibrium point is
locally attractively stable.
c. If there exists a function β of class K such that dV (x(t)) ≤ −β(x)dt
∀x(t) ∈ A, then the equilibrium point is globally attractively stable.

Proof:
Theorem 6.23a: The function V (x) is a PDF, so there exists a class KR
function α(·) (see Definition 6.1) such that

V (x) ≥ α(x) ∀x ∈ Rn . (6.31)

Take ε > 0 and define c as c = α(min(ε, h)). The level set Ωc = {x ∈ Rn |

V (x) ≤ c} is a subset of Bε as well as Bh (the proof is illustrated in Figure 6.7).
Necessarily, it holds that Ωc ⊂ A, because V is unbounded outside A. Because
V (x) is a PDF, there exists a δ > 0 such that Bδ is the largest ball which lies
in Ωc ∪ Ac (where Ac denotes the complement of A), i.e.

sup V (x) ≤ c. (6.32)

x∈Bδ ∩A

Hence, it holds that (Bδ ∩ A) ⊂ Ωc ⊂ Bε . Using dV (x(t)) ≤ 0 for all x(t) ∈

Bh ∩ A, the fact that Ωc ⊂ Bh ∩ A, and Proposition 6.5, it follows that Ωc
is positively invariant. Since ε > 0 can be arbitrarily chosen in the latter
exposition, we can find for each ε > 0 a δ(ε) such that for each x0 ∈ A it
holds that
x0  < δ(ε) ⇒ ϕ(t, t0 , x0 ) < ε ∀t > t0 ,
with ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ). This proves stability of the equilibrium point
x∗ = 0 in the sense of Definition 6.13.
Theorem 6.23b: The conditions of b are stronger than those of a and the
equilibrium point x∗ = 0 is therefore stable. Let c∗ = α(h) and Bδ∗ be the
largest ball which lies in Ωc∗ ∪Ac . Note that V is bounded from below. Because
dV (x(t)) < 0 ∀x(t) ∈ (Bh ∩ A)\{0} for some h > 0, it holds that if x0 ∈ A
then
x0  < δ ∗ ⇒ lim V (ϕ(t, t0 , x0 )) = a ≥ 0,
t→∞

for each solution curve ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ), where a is some nonnega-

tive constant. We now have to show that a = 0 for which we use a contradiction
argument as in [85]. Suppose that a > 0 and choose d such that Bd ⊂ Ωa .
 6.5 Basic Lyapunov Theorems of Autonomous Systems 129

Fig. 6.7. Definition of the sets in the proof of Theorem 6.23 with ε < h, for the
case that x ∈ R2 .
130 6 Lyapunov Stability Theory for Measure Differential Inclusions

The limit V (x(t)) → a for t → ∞ implies that x(t) lies outside the ball Bd
for all t ≥ t0 . Moreover, because −dV is LPDF in the sense that there exists
a function β of class K such that dV (x(t)) ≤ −β(x)dt ∀x(t) ∈ Bh ∩ A we
have

V (x(t)) = V (x(t0 )) + dV
[t0 ,t]
 t
≤ V (x(t0 )) − β(x(t))dt
t0
≤ V (x(t0 )) − β(d)(t − t0 ) .

Since the right-hand side will eventually become negative, the inequality con-
tradicts the assumption that a > 0. Hence, it holds that V (x(t)) → a = 0 for
t → ∞. The function V (x) is PDF and V (x) = 0 therefore implies x = 0.
Each solution curve ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ), which starts in the admissible
part of the ball Bδ∗ , is therefore attracted to the origin, i.e. for x0 ∈ A it holds
that
x0  < δ ∗ ⇒ lim ϕ(t, t0 , x0 ) = 0.
t→∞

Consequently, the equilibrium point x∗ = 0 is locally attractively stable.

Theorem 6.23c: The conditions of c are global versions of the conditions of b.
The condition dV (x) ≤ −β(x)dt ∀x ∈ A implies that ∇v(x) = 0 for x = 0.
The level sets of V are therefore simply connected. Hence, using the positive
definiteness of V , we conclude that the origin x = 0 is the unique minimum of
V and it therefore holds that we can take h = ∞. Moreover, V (x) is radially
unbounded because it is PDF. Every level set Ωc of V is therefore bounded,
simply connected and positively invariant. Following the proof of b but with
h = ∞, it follows that the equilibrium point x∗ = 0 is globally attractively
stable. 
The proof of Theorem 6.23a is illustrated in Figure 6.7. The function V (x1 , x2 ),
shown in the figure, is PDF in x1 and quadratic in x2 . To the chosen value of
ε < h corresponds a value c = α(ε). The level set Ωc consists of the connected
components Ωc1 and Ωc2 . The value of δ is such that Bδ is the largest ball
in Ωc . If a solution starts in Bδ , then it might jump to Ωc2 , but it can never
escape from Ωc and therefore remains within Bε .
Lyapunov stability theorems give sufficient conditions for stability and at-
tractive stability. They do not say whether these conditions are also necessary.
When applying the Lyapunov theorems, we look for a suitable function that
can serve as Lyapunov function. Such functions are referred to as Lyapunov
function candidates. When it finally turns out that a Lyapunov function can-
didate can indeed be used to prove stability, then we say it is a Lyapunov
function.
Theorem 6.23c is a generalisation of the Barbashin-Krasovskii theorem [15].
Radially unboundedness is essential in this theorem to ensure that each level
set of the Lyapunov candidate function is bounded, so that solution curves
 6.5 Basic Lyapunov Theorems of Autonomous Systems 131

can not escape to infinity. The set Bδ∗ is a conservative estimate for the region
of attraction in the proof of Theorem 6.23b. A less conservative estimate is
the level set Ωc∗ .
Classically, the Lyapunov function V (x) is only required to be LPDF in
Theorem 6.23a and b [85, 152]. Indeed, local positive definiteness is enough
when the time-evolution x(t) of the system is continuous. This is not the
case for systems with a discontinuous state x(t). The level sets of V (x) are
generally not simply connected and a discontinuous solution may jump from
a connected component Ωc1  0 to another connected component Ωc2 (see
Figure 6.2). If V (x) is PDF then we can find an r > 0 such that Ωc is simply
connected for all c with 0 ≤ c < r. Such an r > 0 can in general not be found if
the function V (x) is only LPDF. The condition that V (x) is PDF is therefore
essential for measure differential inclusions, when no further assumptions on
the system or the form of V (x) are made. The importance of this condition
has been stated in [180] for hybrid systems, in [34] for measure differential
inclusions and in [27,34,162] for mechanical systems with frictionless unilateral
constraints.
For special classes of systems and Lyapunov functions, it is possible to relax
the condition of positive definiteness to local positive definiteness. Consider
the class of systems for which the state vector x(t) ∈ Rn consists of time-
continuous states x1 (t) ∈ Rm and states x2 (t) ∈ Rp which are of locally
bounded variation:
 
x (t)
x(t) = 1 , x1 ∈ C0 (I, Rm ), x2 ∈ lbv(I, Rp ), (6.33)
x2 (t)

with n = m + p and I = [t0 , ∞). Moreover, we assume that the admissible set
A = A1 × A2 is such that A2 = Rp , i.e. only the states x1 are restricted to
an admissible set A1 . Let V (x) be of the form

V (x) = V1 (x1 ) + V2 (x1 , x2 ), V2 (x1 , x2 ) = xT

2 P (x1 )x2 , (6.34)

in which V (x1 ) is LPDF and P (x1 ) ∈ Rp×p is a symmetric positive definite

matrix. The differential measure dV1 can be expressed in a subderivative,
because x1 (t) is absolutely continuous and V1 is an extended lower semi-
continuous function. Using (6.28)-(6.30) we obtain the differential measure of
V1 (x1 ) = v1 (x1 ) + ΨA1 (x1 )

dV1 = dV1 (x1 )(dx1 )

T
= (∇v1 (x1 )) dx1 + dΨA1 (x1 )(dx1 )
T (6.35)
= (∇v1 (x1 )) dx1 + ΨKA1 (x1 ) (dx1 )
T
= (∇v1 (x1 )) ẋ1 dt.

Moreover, V2 is a quadratic form and dV2 can be expressed using (6.19) and
a partial derivative with respect to x1
132 6 Lyapunov Stability Theory for Measure Differential Inclusions

− T ∂V2
dV2 = (x+
2 + x2 ) P (x1 )dx2 + ẋ1 dt. (6.36)
∂x1
In Chapter 7 we will see that mechanical systems with impact belong to
this class of systems for which the generalised coordinates q(t) are absolutely
continuous and the generalised velocities u(t) are functions of locally bounded
variation. Before stating a Lyapunov theorem for this class of systems, we
need a proposition on the invariance of level sets of V (x) and their connected
components.
Proposition 6.24. Consider the function V (x) of the form (6.34) and the
measure differential inclusion (6.1) with the property (6.33). Let Ωci be a
bounded connected component of the level set Ωc = {x ∈ Rn | V (x) ≤ c}. If
dV ≤ 0 for all x ∈ Ωci , then the set Ωci is positively invariant with respect
to (6.1).
Proof: The condition dV ≤ 0 gives V̇ ≤ 0 and V + ≤ V − , see (6.16). Us-
ing (6.33) and the fact that the system is consistent, the latter can be written
as
−
2 ) ≤ V1 (x1 ) + V2 (x1 , x2 ),
V1 (x1 ) + V2 (x1 , x+ (6.37)
with V1 (x1 ) < ∞ for x1 ∈ A1 and hence
−
2 ) ≤ V2 (x1 , x2 ).
V2 (x1 , x+ (6.38)
Let Ωci = Ωci1 × Ωci2 with Ωci1 ⊂ Rm and Ωci2 ⊂ Rp . Consider the function
V1∗ (x1 ) to be an extension of V1 (x1 ):

V1 (x1 ), x1 ∈ Ωci1 ,
V1∗ (x1 ) = (6.39)
+∞, x1 ∈
/ Ωci1 .
It therefore holds that Ωci1 is a level set of V1∗ (x1 ). Moreover, the function
V2 (x1 , x2 ) is a quadratic positive definite form in x2 for fixed x1 . The set Ωci
is therefore a level set of V ∗ (x) = V1∗ (x1 ) + V2 (x1 , x2 ). It holds that V̇ ∗ = V̇
and
V ∗ + = V ∗ (x+ )
= V1∗ (x1 ) + V2 (x1 , x+
2)
(6.40)
≤ V1∗ (x1 ) + V2 (x1 , x−
2)

≤ V ∗−,
which gives dV ∗ ≤ 0. Using Proposition 6.5, we conclude that each level set
of V ∗ (x) is positively invariant. Consequently, Ωci is positively invariant. 
The idea behind Proposition 6.24 is illustrated in Figure 6.8. The function
V (x) is a potential with two wells, such that the level set Ωc consists of two
connected components Ωc1 and Ωc2 . Consider a solution curve with x0 ∈ Ωc1
and dV ≤ 0. Discontinuities in x(t) can only occur in the x2 -direction. The
solution curve can therefore not leave the connected component, which makes
Ωc1 positively invariant.
 6.5 Basic Lyapunov Theorems of Autonomous Systems 133

Fig. 6.8. Positive invariance of the connected component Ωc1 .

Theorem 6.25 (Lyapunov Stability of an Equilibrium with (6.33)).

Let x∗ = 0 ∈ A be an equilibrium point of dx ∈ dΓ (x(t)), which has the
property (6.33) and admits solutions ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) for all x0 ∈
A. Let V (x) ∈ R ∪ {∞} be LPDF and of the form (6.34). The following
statements can be made:
a. If dV (x(t)) ≤ 0 ∀x(t) ∈ Bh ∩ A for some h > 0, then the equilibrium
point is stable.
b. If there exists a function β of class K and a constant h > 0 such that
dV (x(t)) ≤ −β(x)dt ∀x(t) ∈ Bh ∩ A, then the equilibrium point is
locally attractively stable.

Proof: If V (x) is LPDF, then there must exist a bounded ball Br such that
the function V (x) has no minima, maxima or saddle-points for x ∈ Br other
than the origin. The function V (x) therefore has the form of a cup within Br
and x = 0 is the unique minimum in Br . Let ε∗ be the minimum of r and h.
Using Proposition 6.24, together with (6.33) and the special form of V (6.34),
it follows that each connected component Ωci ⊂ Bε∗ is a positively invariant
set. Moreover, the sets Ωci ⊂ Bε∗ are concentric in the sense that Ωdi ⊂ Ωci if
d ≤ c ≤ c∗ , where Ωc∗ i is the largest connected component in Bε∗ . The proofs
of Theorem 6.25a and b can be continued in the same way as in the proof of
Theorem 6.23a and b. 
134 6 Lyapunov Stability Theory for Measure Differential Inclusions

In order to prove global attractive stability of an equilibrium point, a positive

definite function V (x) is required. This theorem has already been given in
Theorem 6.23c.
In order to show the use of Theorem 6.23, we study the stability of the
bouncing ball example shown in Figure 5.11 and studied in Section 5.4. Using
 T
the state-vector x(t) = q(t) u(t) we write the measure differential inclusion
for the bouncing ball example as
 
udt
dx ∈ (6.41)
−gdt + m 1
dPN

with the set-valued force law for the measure of the contact percussion dPN

−dPN ∈ NKK (q) (ξN ), ξN = u+ + eN u− , (6.42)

and the set K = {q | q ≥ 0} of admissible positions q. The admissible set

of x(t) is the half-space A = {x ∈ R2 | q ∈ K}. We are now interested in
the stability of the equilibrium point x∗ = 0. This is the unique equilibrium
point of the system and is therefore isolated. As Lyapunov candidate function,
we choose the sum of kinetic and potential energy together with the energy
potential πN (q) = ΨC∗ N (q) of the contact force −λN ∈ CN = {−λN | λN ≥ 0}:
1
V (x) = mu2 + mgq + πN (q). (6.43)
2
The energy potential πN (q) equals the indicator function on the admissible
set A
πN (q) = ΨC∗ N (q) = ΨA (x) = ΨK (q). (6.44)
The function V (x) is PDF, because f (x) = 12 mu2 +mg|q| is PDF and V (x) ≥
f (x). We see that the term πN (q) = ΨK (q) is essential in (6.43) in order
to make V a positive definite function. Each level set of V is bordered by
a parabola q = mg c
− 2g
1 2
u , c > 0, and the line q = 0 (see Figure 6.9).
Using (3.64), the subderivative of the indicator function (2.54) and the chain
rule, we obtain the differential measure
1
dV = m(u+ + u− )du + mgdq + dΨK (q)(dq)
2 (6.45)
1
= m(u+ + u− )du + mgdq + ΨKK (q) (dq),
2
 T
which we evaluate along solution curves x(t) = q(t) u(t) of the system
giving
1 +
dV = (u + u− )(dPN − mgdt) + mgudt + ΨKK (q) (udt). (6.46)
2
The measure of the contact percussion dPN obeys the set-valued force
law (6.42) with
 6.5 Basic Lyapunov Theorems of Autonomous Systems 135

Fig. 6.9. Stability of the bouncing ball example.

dPN = λN dt + ΛN dη, (6.47)

in which λN is the contact force and ΛN the contact impulse. Substitution of
the decomposition (6.47) in (6.46) and using the fact that 12 (u+ +u− )dt = udt
holds in front of all Lebesgue measurable terms yields
1 +
dV = (u + u− )(λN dt + ΛN dη − mgdt) + mgudt + ΨKK (q) (udt)
2
1
= (u+ + u− )ΛN dη + uλN dt − mgudt + mgudt + ΨKK (q) (udt), (6.48)
2
1
= (u+ + u− )ΛN dη + uλN dt + ΨKK (q) (udt),
2
with
−ΛN ∈ NKK (q) (u+ + eN u− ),
(6.49)
−λN ∈ NKK (q) (u).

If q(t) > 0, then it holds that u(t) is locally continuous and KK (q) = R. As
a consequence, the contact force and impulse vanish for an open contact, i.e.
NR (·) = 0 ⇒ λN = ΛN = 0. Moreover, the indicator term in the Lyapunov
function gives ΨR (·) = 0. Hence, the differential measure of the Lyapunov
function vanishes for q > 0

q > 0 ⇒ dV = 0, (6.50)
136 6 Lyapunov Stability Theory for Measure Differential Inclusions

because the system is conservative in this part of the state-space.

If q(t) = 0, then also the contact force λN and the impulse ΛN come into
play. It holds that KK (0) = R+ . The set-valued force laws for the contact force
and impulse are such that no penetration can occur. The unilaterality of the
contact is preserved. In other words, the system is consistent and it holds that
dq = udt = u+ dt ∈ KK (q). Hence, it holds that ΨKK (q) (dq) = ΨKK (q) (udt) = 0
and necessarily u ≥ 0 for q = 0. The contact force therefore obeys

u > 0 ⇒ −λN = 0
−λN ∈ NR+ (u) ⇒ (6.51)
u = 0 ⇒ −λN ≤ 0

We conclude that the contact force λN has no power, i.e. uλN = 0. Similarly,
from the set-valued force law for ΛN follows

u+ > −eN u− ⇒ −ΛN = 0
−ΛN ∈ NR+ (u+ + eN u− ) ⇒ (6.52)
u+ = −eN u− ⇒ −ΛN ≤ 0

and it therefore holds that dV = 0 if u+ > −eN u− . On the other hand, if

u+ = −eN u− , then an impulse ΛN = m(u+ (t)−u− (t)) ≥ 0 causes the velocity
to jump to u+ = −eN u− ≥ 0 and we obtain the differential measure dV :
1 +
q = 0 ⇒ dV = (u + u− )ΛN dη
2
1
= m(u+ + u− )(u+ − u− )dη
2 (6.53)
1  2
= − m(1 − eN )(1 + eN ) u− dη
2
1  2
= − m(1 − e2N ) u− dη.
2
Hence, it holds that dV ≤ 0 if |eN | ≤ 1, and because of the consistency
requirement eN ≥ 0 this condition reduces to 0 ≤ eN ≤ 1. Comparison of the
decomposition (6.16)

dV = V̇ dt + (V + − V − )dη

of the differential measure dV with (6.53) gives

1  2
V̇ = 0, V + − V − = − m(1 − e2N ) u− . (6.54)
2
From the above, it becomes apparent that energy is only dissipated when a
not completely elastic impact (i.e. eN < 1) occurs, which agrees with our
physical interpretation of the system.
Consequently, V (x) is PDF and dV (x(t)) ≤ 0 ∀x(t) ∈ A. These conditions
are stronger than the conditions required by Theorem 6.23a, but weaker than
those in Theorem 6.23b or c. Using Theorem 6.23, we can therefore only prove
 6.5 Basic Lyapunov Theorems of Autonomous Systems 137

stability. For each ε > 0 we find the largest level set Ωc of V (x) that fits in
the ball Bε (see Figure 6.9). It holds that Ωc ⊂ A, because V (x) = +∞ for
x∈ / A. Each level set Ωc is a positively invariant set, because dV ≤ 0 for all
x ∈ A (Proposition 6.5). We now choose δ such, that Bδ is the largest ball
in Ωc . Consequently, for each ε > 0 we can find a δ > 0 such that for each
x0 ∈ A it holds that

x0  < δ(ε), ⇒ ϕ(t, t0 , x0 ) < ε ∀t > t0

with ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ), which proves stability of the equilibrium

point x∗ = 0 in the sense of Definition 6.13. For eN < 1, all solutions con-
verge to the origin in a finite time and the equilibrium point is even globally
symptotically attractively stable, but this can not be proven using Theo-
rem 6.23. To prove attractivity, we need the invariance principle of LaSalle,
which will be discussed in Section 6.6. The solution curve drawn in Figure 6.9
corresponds to the time-histories q(t) and u(t) of Figure 5.11. We see that
the equilibrium point is symptotically attractive, as it is reached in a finite
time. Moreover, the equilibrium point is an accumulation point, because it is
reached in a finite time after an infinite number of impacts.
In order to gain more insight in the problem, we replace the rigid uni-
lateral contact with a unilateral spring with spring stiffness k. The to-
tal mechanical energy is then composed of the kinetic energy, the poten-
tial energy due to gravitation and the potential energy of the unilateral
spring: E(x(t)) = 12 mu(t)2 + mgq(t) + 12 kd(t)2 with the penetration depth
d(t) = max(−q(t), 0). Clearly, when k tends to ∞, then E(x) → V (x). The
dissipation in a compliant contact model can be modelled by a viscous damper.
For the one-dimensional case, the existence of a sequence of compliant dis-
sipative models which converge to the rigid limiting model with restitution
is shown in [23]. This example illustrates that the decomposition of the Lya-
punov function in (6.27) (or (6.43) in the example) is a natural choice for
mechanical systems with unilateral constraints and will be used in Chapter 7.

6.5.2 Lyapunov Stability of Equilibrium Sets

Theorem 6.23 provides a means to prove the (attractive) stability of equili-

brium points. However, (measure) differential inclusions generally have equi-
librium sets. Without much effort, we can extend Theorems 6.23 and 6.25 to
prove the stability of an equilibrium set. Instead of using a Lyapunov function
which is PDF, as has been done in Theorem 6.23 to prove the stability of an
equilibrium point, we have to use a Lyapunov function which is zero within
the equilibrium set and positive outside.
Theorem 6.26 (Stability of an Equilibrium Set). Let E ⊂ A be an
equilibrium set of dx ∈ dΓ (x(t)), which admits solutions ϕ(·, t0 , x0 ) ∈
S(dΓ , t0 , x0 ) for all x0 ∈ A. Let V (x) ∈ R ∪ {∞} be a lower semi-continuous
function, such that
138 6 Lyapunov Stability Theory for Measure Differential Inclusions

i. V (x) = 0 for all x ∈ E,

ii. V (x) > 0 for all x ∈ A\E,
iii. V (x) = +∞ for x ∈ / A and V (x) is radially unbounded.
Then the following statements hold:
a. If dV (x(t)) ≤ 0 ∀x(t) ∈ Uh (E) ∩ A for some h > 0, then the equilibrium
set is stable.
b. If there exists a function β of class K and a constant h > 0 such that

dV (x(t)) ≤ −β(distE (x))dt ∀x(t) ∈ Uh (E) ∩ A,

then the equilibrium set is locally attractively stable.

c. If there exists a function β of class K such that dV (x(t)) ≤ −β(distE (x))dt
∀x(t) ∈ A, then the equilibrium set is globally attractively stable.

Proof:
The proof of Theorem 6.26 is analogous to the proof of Theorem 6.23. Con-
ditions i – iii replace the positive definiteness requirement whereas the neigh-
bourhood Uh replaces the ball Bh of Theorem 6.23. In this way, stability of
the equilibrium set can be proven in the sense of Definition 6.18. 
Clearly, if we let the equilibrium set shrink to an equilibrium point, then the
conditions on the Lyapunov function reduce to positive definiteness and that
it is of the form (6.27). Theorem 6.26 is therefore a direct generalisation of
Theorem 6.23. In the same way, we generalise Theorem 6.25.

Theorem 6.27 (Stability of an Equilibrium Set with (6.33)). Let E ⊂

A be an equilibrium set of dx ∈ dΓ (x(t)), which has the property (6.33)
and admits solutions ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) for all x0 ∈ A. Let V (x) ∈
R ∪ {∞} be a lower semi-continuous function of the form (6.34), such that
i. V (x) = 0 for all x ∈ E,
ii. V (x) > 0 for all x ∈ (Ur (E) ∩ A)\E for some r > 0,
iii. V (x) = +∞ for x ∈ / A.
Then the following statements hold:
a. If dV (x(t)) ≤ 0 ∀x(t) ∈ Uh (E) ∩ A for some h > 0, then the equilibrium
set is stable.
b. If there exists a function β of class K and a constant h > 0 such that

dV (x(t)) ≤ −β(distE (x))dt ∀x(t) ∈ Uh (E) ∩ A,

then the equilibrium set is locally attractively stable.

Proof:
The proof of Theorem 6.27 is analogous to the proof of Theorem 6.25. Condi-
tions i and ii replace the local positive definiteness requirement whereas the
 6.5 Basic Lyapunov Theorems of Autonomous Systems 139

neighbourhood Uh replaces the ball Bh of Theorem 6.25. In this way, stability

of the equilibrium set can be proven in the sense of Definition 6.18. 
The main difficulty with Theorem 6.26 and 6.27 is the fact that a suitable
Lyapunov function is hard to find. The Lyapunov function V has to be zero
within the equilibrium set and be positive outside and still fulfill the conditions
i to iii. It might prove to be easier to prove the stability of each equilibrium
point x∗ ∈ E separately using Theorem 6.23a using a different Lyapunov
function Vx∗ (x) for each equilibrium point. Subsequently, we can prove the
stability of the equilibrium set as a whole [114, 115]:
Theorem 6.28 (Stability of an Equilibrium Set as Union of Equili-
brium Points). Let E ⊂ A be an equilibrium set of dx ∈ dΓ (x(t)), which
admits solutions ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) for all x0 ∈ A. If it can be proven
that each equilibrium point x∗ ∈ E is Lyapunov stable using a PDF function
Vx∗ (x), then the equilibrium set is stable in the sense of Definition 6.18.
Proof: Let Ωx∗ ,r be the r-level set of the Lyapunov function Vx∗ (x)

Ωx∗ ,r = {x ∈ Rn | Vx∗ (x) ≤ r}.

Define r∗ > 0 as the largest value of r such that Ωx∗ ,r∗ is a positively invariant
set for all x∗ ∈ E. Note that such an r∗ exists by virtue of the fact that each
equilibrium point x∗ ∈ E is Lyapunov stable. For each r ∈ (0, r∗ ], we introduce
the set Ωr as the union of all r-level sets of Vx∗ (x):
'
Ωr = Ωx∗ ,r , r ∈ (0, r∗ ].
∀x∗ ∈E

Note that each set Ωr is the union of positively invariant sets, which makes
it a positively invariant set itself and E ⊂ int Ωr . For each r ∈ (0, r∗ ], we now
choose ε(r) > 0 as the smallest value such that Ωr ⊂ Uε (E). We can find for
each ε ∈ (0, ε∗ ], with ε∗ = ε(r∗ ), a δ(ε) > 0 such that for any x0 ∈ A with

x0 ∈ Uδ (E) ⊂ Ωr ,

each solution curve ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) satisfies

ϕ(t, t0 , x0 ) ∈ Uε (E).

For each ε > ε∗ we take δ(ε) = δ(ε∗ ). Consequently, we can find for each ε > 0
a δ(ε) > 0 such that for any x0 ∈ A with x0 ∈ Uδ (E) ⊂ Ωr each solution
curve ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) satisfies ϕ(t, t0 , x0 ) ∈ Uε (E), which proves
stability of the equilibrium set E. 
The different sets used in the proof of Theorem 6.28 are depicted in Fig-
ure 6.10. Theorem 6.28 gives an alternative way to prove the stability of an
equilibrium set which does not require to find a Lyapunov function for the
whole set. However, the proof of Theorem 6.28 introduces a set Ωr , being
140 6 Lyapunov Stability Theory for Measure Differential Inclusions

Fig. 6.10. Construction of Ωr in the proof of Theorem 6.26 as the union of level
sets Ωx ∗ ,r .

positively invariant and enclosing the equilibrium set. The question now rises
which function V is associated with Ωr , if we interpret Ωr as a level set:

Ωr = {x ∈ Rn | V (x) ≤ r}. (6.55)

The set Ωr is the union of all r-level sets of each individual Lyapunov function
Vx∗ (x) and it therefore must hold that
'
epi V = epi Vx∗ . (6.56)
∀x∗ ∈E

The function V (x) is therefore the pointwise minimum of all functions Vx∗ (x)

V (x) = min
∗
Vx∗ (x). (6.57)
x ∈E

The function V defined by (6.56) (or equivalently (6.57)) has the properties

V (x) = 0 ∀x ∈ E, V (x) > 0 ∀x ∈

/ E, V (x) = +∞ ∀x ∈
/A (6.58)

and V (x) is radially unbounded since each Vx∗ (x) is a PDF. Consequently,
the function V as in (6.57) is a suitable Lyapunov function candidate to
prove stability of an equilibrium set using Theorem 6.26. The search for a
Lyapunov function V in Theorem 6.26 has therefore been reduced to the
search of Lyapunov functions Vx∗ (x) for all equilibrium points x∗ in the set
E.
The Lyapunov function V in (6.57) can be simplified if the Lyapunov
functions Vx∗ (x) have the generic quadratic form
1
Vx∗ (x) = (x − x∗ )T P (x − x∗ ) + ΨA (x), P = GGT > 0. (6.59)
2
with G being the square root matrix of P = P T . Substitution of this special
form of Vx∗ (x) in (6.57) gives
 6.5 Basic Lyapunov Theorems of Autonomous Systems 141
1
V (x) = min
∗
(x − x∗ )T GGT (x − x∗ ) + ΨA (x)
x ∈E 2
1
= min GT (x − x∗ )2 + ΨA (x)
x∗ ∈E 2
1
= min z − z ∗ 2 + ΨA (x), (6.60)
z ∗ ∈Ez 2
1
= z − proxEz (z)2 + ΨA (x)
2
1
= dist2Ez (z) + ΨA (x),
2
where z = GT x and Ez = {z | z = GT x, x ∈ E}. Hence, a generic quadratic
form of Vx∗ (x) allows to express V (x) as the distance to the equilibrium set
(see Definition 2.32), but with a metric determined by P . If the state x(t)
is absolutely continuous, dx = ẋdt, then the differential measure dV can be
obtained from (2.44)

dV = dV (x)(dx)
= (z − proxEz (z))T dz + dΨA (x)(dx)
(6.61)
= (z − proxEz (z))T dz + ΨKA (x) (dx), dx = ẋdt ∈ KA (x)
= xT P dx − proxT T T
Ez (G x)G dx.

6.5.3 Stability through Maximal Monotonicity

A special class of systems is formed by measure differential inclusions of the

form
dx ∈ −dA(x+ ), (6.62)
where dA is a maximal monotone (set-valued) measure function. This class
of systems has already been encountered in Section 4.3. The maximal mono-
tonicity of dA is a sufficient condition for the stability of an equilibrium set
of (6.62).
Theorem 6.29 (Stability of the Equilibrium Set through Monotonic-
ity). Consider a measure differential inclusion of the form

dx ∈ −dA(x+ ) = −At (x)dt − Aη (x+ )dη,

with a non-empty equilibrium set E = {x ∈ Rn | 0 ∈ dA(x)}. Let the system

be consistent and have existence of solutions for all initial conditions x0 ∈
A, where the admissible domain A is a closed set. If dA(x+ ) is a maximal
monotone set-valued measure function, then the equilibrium set E is stable.
Moreover, if At (x) is a strictly maximal monotone set-valued function, then
the equilibrium set3 is globally attractively stable.
3
If At (x) is strictly maximal monotone, then the equilibrium ‘set’ is an equilibrium
point.
142 6 Lyapunov Stability Theory for Measure Differential Inclusions

Proof: The set E = dA−1 (0) is a closed convex set, because dA−1 is maximal
monotone if dA is maximal monotone and the images of a maximal monotone
operator are closed convex sets [12]. For the differential measure of dx we write

dx = −at dt − aη dη, (6.63)

where the single-valued densities obey the set-valued force laws

−ẋ = at ∈ At (x), −(x+ − x−) = aη ∈ Aη (x+ ). (6.64)

The set-valued operators At and Aη are maximal monotone operators and

are related to the set-valued measure function by

dA(x+ ) = At (x)dt + Aη (x+ )dη. (6.65)

Consider the Lyapunov function

1
V (x) = dist2E (x) + ΨA (x), (6.66)
2
which is a positive definite function with respect to E. Evaluation of the
differential measure V̇ gives

V̇ = (x − proxE (x))T ẋ
(6.67)
= −(x − proxE (x))T at .

Using At (proxE (x))  at0 = 0 together with the monotonicity of At (x) yields
T
V̇ = − (x − proxE (x)) (at − at0 ) ≤ 0. (6.68)

We now consider the atomic part

1 1
V+−V− = dist2E (x+ ) − dist2E (x− )
2 2
1( (2 1 ( (2
= (x+ − proxE (x+ )( − (x− − proxE (x− )( . (6.69)
2 2

Due to the convexity of E it holds that

( + ( ( (
(x − proxE (x+ )( ≤ (x+ − proxE (x− )( , (6.70)

which gives
1( ( ( (
(x+ − proxE (x− )(2 − 1 (x− − proxE (x− )(2 .
V+−V− ≤ (6.71)
2 2
Using x = xT x and rearranging terms gives
 6.5 Basic Lyapunov Theorems of Autonomous Systems 143
1 + T  + 
V+−V− ≤ x − proxE (x− ) x − proxE (x− )
2
1 − T  − 
− x − proxE (x− ) x − proxE (x− )
2
1 + 2 
≤ x  − 2 proxE (x− )T x+ − x− 2 + 2 proxE (x− )T x−
2
1 + T
≤ x + x− − 2 proxE (x− ) (x+ − x− ).
2
(6.72)

Furthermore, because of the invariance of the equilibrium set E it holds that

0 ∈ Aη (proxE (x− )). Substitution of aη = −(x+ − x− ) and aη0 = 0 yields

1 + T
V+−V− ≤− 2x + aη − 2 proxE (x− ) aη
2 (6.73)
1  T
≤ − aη 2 − x+ − proxE (x− ) (aη − aη0 ) ,
2
Clearly, using the monotonicity of Aη , it follows that V + − V − ≤ 0. Stability
of the equilibrium set follows from Theorem 6.26. Moreover, if At (x) is strictly
maximal monotone, then it holds that there exists a KR-function β such that
V̇ ≤ −β(distE (x)). Furthermore, if At (x) is strictly maximal monotone, then
the closed equilibrium set E is an equilibrium point and therefore compact.
The Lyapunov function (6.66) is therefore a radially unbounded positive def-
inite function with respect to E and has to decrease as long as the solution
x(t) ∈/ E, which proves global attractivity. 
Consider for instance the measure differential inclusion

dx ∈ −(Sign(x) + cx)dt, (6.74)

which has the unique equilibrium point x∗ = 0. The operator dA(x) =

(Sign(x)+cx)dt is maximal monotone if c ≥ 0 and strictly maximal monotone
if c > 0. Hence, Theorem 6.29 immediately proves stability of x∗ = 0 for c ≥ 0
and attractive stability of x∗ = 0 for c > 0. If c < 0, then the equilibrium
point is still locally attractively stable but this is not proven by Theorem 6.29,
because the measure function dA(x) is no longer maximal monotone for c < 0.
We conclude this section with a sufficient condition for incremental sta-
bility of a system (see Definition 6.22), which follows directly from the mono-
tonicity property of (6.62).
Theorem 6.30 (Incremental Stability through Monotonicity). Con-
sider a measure differential inclusion of the form

dx ∈ −dA(x+ ) = −At (x)dt − Aη (x+ )dη.

Let the system be consistent and have existence of solutions for all initial con-
ditions x0 ∈ A, where the admissible domain A is a closed set. If dA(x+ ) is
144 6 Lyapunov Stability Theory for Measure Differential Inclusions

a maximal monotone set-valued measure function, then the system is incre-

mentally stable. Moreover, if At (x) is a strictly maximal monotone set-valued
function, then the system is attractively incrementally stable.
Proof: Consider two arbitrary solutions x1 (t) and x2 (t) of the measure dif-
ferential inclusion with xi (t0 ) ∈ A, i = 1, 2. In order to prove incremental
stability, we measure the distance between these two arbitrary solutions and,
therefore, consider the Lyapunov function candidate
1
V = x1 − x2 2 ,
2
which has the differential measure

dV = V̇ dt + (V + − V − )dη, (6.75)

with
V̇ = (x2 − x1 )T (ẋ2 − ẋ1 )
= −(x2 − x1 )T (at2 − at1 ) at1 ∈ At (x1 ), at2 ∈ At (x2 ) (6.76)
≤0

and
1 +
V+−V− = (x + x− − T − −
2 − x1 − x1 ) (x2 − x2 − x1 + x1 )
+ + +
2 2
1
= − (2x+ 2 − 2x1 + aη2 − aη1 ) (aη2 − aη1 )
+ T
2 (6.77)
1
= −(x+ 2 − x1 ) (aη2 − aη1 ) − aη2 − aη1 
+ T 2
2
≤ 0,

where aη1 ∈ Aη (x+ 1 ) and aη2 ∈ Aη (x2 ), see the proof of Theorem 4.9.
+

Clearly, if At (x) and Aη (x ) are monotone, then it holds that dV ≤ 0, from

+

which we deduce that V + (t) ≤ V (t0 ) or

1 (t) − x2 (t) ≤ x1 (t0 ) − x2 (t0 )

x+ +
∀t ≥ t0 .

Consequently, the distance between any two solutions can not increase, which
proves incremental stability of the system. Moreover, if At (x) is strictly mono-
tone then it holds that
dV ≤ αx1 − x2 2 dt
< 0, for x1 = x2 ,

for some α > 0. The solutions therefore have to approach each other as long
as they are different from each other, i.e.

lim x1 (t) − x2 (t) = 0,

t→∞
 6.6 LaSalle’s Invariance Principle 145

which proves attractive incremental stability. 

Incremental stability of the system implies that solutions can not diverge
from each other and stability of all its equilibrium points is therefore assured.
Consider an equilibrium x∗ ∈ E and a ball x∗ + Bε around that equilibrium
point. Solutions starting in the ε-ball can not leave this ball, because their
distance to x∗ can not increase due to incremental stability. For each ε > 0 we
can take δ = ε and prove Lyapunov stability in the usual way. Theorem 6.29
therefore follows from Theorem 6.30.
Attractive stability excludes the possibility of more than one equilibrium
because all solutions have to converge to each other. Hence, if dA(x+ ) is a
strictly maximal monotone operator, then the system dx ∈ −dA(x+ ) does
not posses an equilibrium set and at most a unique equilibrium point. The-
orems 6.29 and 6.30 will be used in Section 7.2 to prove the stability of the
equilibrium set of a mechanical system with friction. In Chapter 8, we will
further study the convergence property of maximal monotone measure differ-
ential inclusions.

6.6 LaSalle’s Invariance Principle

Using the basic Lyapunov theorems, proposed in the previous section, we are
able to prove attractive stability if the Lyapunov function is strictly decreasing
outside the equilibrium point/set, i.e. dV < 0. In many situations, however,
we are not able to find such a Lyapunov function (e.g. the bouncing ball
example). If the Lyapunov function is not strictly decreasing, then the solution
can remain on a level surface of the Lyapunov function and not be attracted
towards the equilibrium set. LaSalle introduced a theorem with which we are
able to prove the conditional attractivity even when dV ≤ 0. The main idea
of LaSalle is, that when the level surfaces of the Lyapunov function outside
the equilibrium point/set are not positively invariant, then the equilibrium
point/set must be attractive. In this section, we give a generalisation of the
invariance principle of LaSalle [89], which has originally been stated for smooth
dynamical systems, and formulate it for measure differential inclusions. In the
following, we consider the extended lower semi-continuous function V : Rn →
R ∪ {+∞} to be of the form

V (x) = v(x) + ΨA (x), (6.78)

in which v(x) is a smooth, continuous, and locally bounded function. The

following theorem and proof are adaptations of LaSalle’s theorem as given
in [85]:

Theorem 6.31 (LaSalle’s Invariance Principle). Let Ω ⊂ A be a com-

pact set that is positively invariant with respect to the measure differential
inclusion (6.1). Let every limit set L+ in Ω be positively invariant. Let V (x)
146 6 Lyapunov Stability Theory for Measure Differential Inclusions

be a function of the form (6.78), with v(x) being continuous and bounded from
below, such that dV ≤ 0 along solution curves in Ω. Let Z be the set of all
points in Ω where dV = 0 can hold, i.e.
Z = {x ∈ Ω | 0 ∈ dV (x), dx ∈ dΓ (x)}.
Let M be the largest positively invariant set in Z. Then every solution curve
ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) with x0 ∈ Ω approaches M as t → ∞.
Proof: Let x(t) = ϕ(t, t0 , x0 ) be a solution of (6.1) starting in Ω, i.e. x0 ∈ Ω.
The solution x(t) remains in Ω for all t ≥ t0 , because Ω is positively invariant.
The function V (x(t)) is non-increasing because dV (x) ≤ 0 in Ω. Since V (x)
is bounded from below, V (x(t)) has a limit a as t → ∞. The positive limit set
L+ of ϕ(·, t0 , x0 ) is in Ω because Ω is a closed set. According to Definition 6.9,
for any p ∈ L+ there exists a sequence {tj } with tj → ∞ as j → ∞, such
that x(tj ) → p for j → ∞. It holds that x(tj ) ∈ A since Ω ⊂ A is positively
invariant. By continuity of v(x) it holds that v(p) = limj→∞ v(x(tj )) = a.
Moreover, using that x(tj ) ∈ A and p ∈ A it follows with ΨA (p) = 0 that
V (x) = a for all x ∈ L+ . The positive limit set L+ is assumed to be positively
invariant. Consequently, dV = 0 on L+ and
L+ ⊂ M ⊂ Z ⊂ Ω
The compactness of Ω implies boundedness of x(t) (from which we can deduce
that L+ is non-empty [85]) and x(t) approaches L+ as t → ∞. Consequently,
x(t) approaches M as t → ∞. 
Note that a sufficient condition for the positive invariance of limit sets
is a continuous dependence of solutions x(t) ∈ Ω on initial data (Proposi-
tion 6.12). The set Ω in Theorem 6.31 has to be positively invariant which
can, for instance, be proven with Proposition 6.5. LaSalle’s invariance princi-
ple does not require the function V to be positive definite, but the function
V is usually chosen to be positive definite in order to prove the boundedness
(and positive invariance) of the set Ω, which is used in Proposition 6.5.
We briefly discuss an example which shows why the condition that ev-
ery limit set is positively invariant is essential for LaSalle’s theorem to hold.
Consider the discontinuous differential equation

−x + 1 x > 1,
ẋ = (6.79)
−x x ≤ 1,

which has a unique equilibrium point at the origin x = 0. Note that this system
is not a Filippov system. Consider the positive definite function V (y) = 12 x2 .
The function V does not increase along solution curves, which can be seen
from ⎧
⎪
⎨−x + x < 0 x > 1,
2

V̇ = xẋ = −x2 < 0 x ≤ 1, x = 0, (6.80)

⎪
⎩
0 x = 0.
 6.7 Instability 147

Each level set of V is positively invariant and the origin is the largest invariant
set within the set for which V̇ = 0. Still, the origin is not attractive within
level sets V > 12 . A solution x(t) with x(0) > 1 converges asymptotically to
the limit point x = 1, i.e. limt→∞ x(t) = 1, but it never reaches the limit
point. The limit point itself is not invariant because ẋ < 0 for x = 1. A
solution with x(0) > 1 will therefore not be attracted to the origin. Note also
that the conditions for global attractivity in Theorem 6.23 are not fulfilled
because there does exist a function β of class K such that V̇ ≤ −β(|x|) for all
x ∈ R. In other words, the function −V̇ is not positive definite.
We now return to the bouncing ball example which was studied in the
previous section. Each level set Ωc of V , given by (6.43), is bounded because
V is PDF. Proposition 6.5 therefore proves that each level set Ωc is positively
invariant. Following Theorem 6.31, let Z be the set of all points in Ωc where
dV = 0 can hold, i.e. Z is the union of the set {x ∈ Ω | q = 0, u ≥ 0} and
the set {x ∈ Ω | q > 0}:
'
Z = {x ∈ Ω | q = 0, u ≥ 0} {x ∈ Ω | q > 0}.

The equilibrium x∗ = 0 is an element of Z. We now seek the largest positively

invariant set in Z, which we call M. Because of the gravitation, if the ball
is in the air, then it will hit the floor in a finite time. In other words: the
solution curve will leave int A in a finite time and the set int A is therefore
not positively invariant. The largest positively invariant set in Z is therefore
M = x∗ = 0. This implies that the equilibrium point x∗ is globally attractive.
Moreover, stability has been proven in the previous section. Consequently,
the equilibrium point x∗ is globally attractively stable. To prove symptotic
attractivity is much more difficult. Symptotic stability of the bouncing ball
system for e < 1 has been proven in [27,175] using an event-type Poincaré map
method. Such an analysis can only be carried out explicitly in very special
cases.

6.7 Instability

Lyapunov-type theorems for the stability and attractivity of equilibrium

points and sets have been presented in the previous sections. It is also possible
to prove instability and/or non-attractivity using a Lyapunov-type argument.
We give a generalisation of Chetaev’s instability theorem [35,85] for an equili-
brium point or set of a measure differential inclusion using an extended lower
semi-continuous function V : Rn → R ∪ {∞} of the form (6.27).

Theorem 6.32 (Instability and Non-Attractivity of an Equilibrium

Point). Let x∗ = 0 ∈ A be an equilibrium point of the consistent measure dif-
ferential inclusions (6.1), which admits solutions ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 )
for all x0 ∈ A. Let V (x) be an extended lower semi-continuous function of
148 6 Lyapunov Stability Theory for Measure Differential Inclusions

the form (6.27), such that V (0) = 0 and V (xa ) = a < 0 for some xa with
arbitrarily small xa . Moreover, V is such that the set U defined by

U = {x ∈ B r ∩ A | V (x) ≤ 0}

for some r > 0 is a nonempty set. Let every limit set L+ in U be positively
invariant. Three statements can be made
a. If dV < 0 for all x ∈ U\{0}, then it holds that x∗ = 0 is unstable.
b. If dV ≤ 0 for all x ∈ U and the solution can not stay in Z = {x ∈ U\{0} |
dV = 0}, then x∗ = 0 is unstable.
c. If dV ≤ 0 for all x ∈ U and 0 ∈ U\(bdry U\ bdry A), then it holds that
x∗ = 0 is not attractive.
Proof: Theorem 6.32a: First of all, remark that U ⊂ A is compact and all
solution curves ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) remain in A for x0 ∈ A due to the
consistency assumption (Definition 4.8). From (6.27) we see that V (x) = v(x)
for all x ∈ U and V (x) is therefore bounded on U. It holds that xa ∈ int U
because V (xa ) = a < 0 and xa  < r is arbitrarily small. Moreover, for
every solution curve ϕ(·, t0 , xa ) ∈ S(dΓ , t0 , xa ) it holds that the function
V (ϕ(t, t0 , xa )) with

V (ϕ(t, t0 , xa )) = V (xa ) + dV (6.81)
[t0 ,t]

is strictly decreasing as long as ϕ(t, t0 , xa ) ∈ U since dV < 0 in U\{0}.

This means that if a solution curve ϕ(t, t0 , xa ) would stay forever in U
then V (ϕ(t, t0 , xa )) would tend to −∞. This is not possible because V (x)
is bounded from below on U and all solution curves ϕ(t, t0 , xa ) must there-
fore leave the set U. The solution curves ϕ(·, t0 , xa ) ∈ S(dΓ , t0 , xa ) can
not leave through bdry A because of the viability of solution curves. The
solution curves can also not leave through the surface V (x) = 0 because
V (ϕ(t, t0 , xa )) ≤ a < 0 for almost all t ≥ t0 . Consequently, all solution curves
ϕ(·, t0 , xa ) ∈ S(dΓ , t0 , xa ) leave through the sphere x = r. Notice that
bdry Br ∩ bdry U = ∅ because the solution curves have to escape from U. We
can take xa arbitrarily close to the origin and the origin is therefore unstable.
Theorem 6.32b: We extend Theorem 6.32a with a LaSalle-type of argumen-
tation. The function V is non-increasing along solution curves of the system
within U because dV ≤ 0. No solution curve ϕ(t, t0 , xa ) ∈ U therefore exists
which is attracted to the origin. Solution curves might remain temporarily on
a level surface of V if dV = 0, i.e. if x ∈ Z. However, the solution can not
stay in Z, i.e. there exists no positively invariant subset of Z. This means that
solution curves ultimately will leave this set and enter U\Z for which dV < 0.
So, the function V will decrease again. The rest of the reasoning is similar
to the proof of Theorem 6.32a. All solution curves may enter and leave Z a
couple of times, but finally leave U through the sphere x = r which proves
instability.
 6.7 Instability 149

Theorem 6.32c: As in b it holds that the function V is non-increasing along

solution curves of the system within U because dV ≤ 0. No solution curve
ϕ(t, t0 , xa ) ∈ U therefore exists which is attracted to the origin. Moreover,
since 0 ∈ U\(bdry U\ bdry A) it holds that either 0 ∈ int U or the origin
is on the boundary between U and Ac , i.e. it is surrounded by U ∪ Ac (see
Figure 6.12). If 0 ∈ int U then the set U is a neighbourhood of the origin.
Consequently, there exists no solution which is attracted to the origin. Simi-
larly, if 0 is on the boundary between U and Ac , then it can not be attracted
through U nor through Ac . 

Fig. 6.11. Sets used in the proof of Theorem 6.32a.

The various sets used in the proof of Theorem 6.32a have been elucidated in
Figure 6.11. A solution curve is drawn which starts in U at point xa , slides
along the boundary of A and finally leaves U through the sphere x = r.
Notice that U can consist of more than one part (e.g. two, as drawn in the
Figure 6.11), but has to contain the origin in its interior or boundary.
Theorem 6.32a and b prove the conditional instability of an equilibrium
point. All solution curves ϕ(t, t0 , xa ) with xa ∈ U have to leave U through the
sphere x = r > 0. So there exists at least one solution curve which starts
arbitrarily close to the origin and which arrives at a distance r > 0 from the
origin. Non-attractivity is however not proven in a and b because solution
curves which have left U might still be attracted to the origin in A\U . Such
a ‘return-loop’ of solution curves can of course not exist when x∗ ∈ int U and
this is the main idea behind Theorem 6.32c. Also, when x∗ is on the boundary
of A but within A surrounded by U, then such a return-loop can not exist,
see Figure 6.12. Therefore, a return-loop can not exist if x∗ is in the relative
150 6 Lyapunov Stability Theory for Measure Differential Inclusions

interior of U with respect to A:

∃r > 0 Br (x∗ ) ⊂ (U ∪ Ac ) ⇐⇒ x∗ ∈ U\(bdry U\ bdry A). (6.82)

An equilibrium set E is called stable when all equilibrium points x∗ ∈ E

are stable. To prove instability of an equilibrium set, it therefore suffices to
prove the instability of one equilibrium point x∗ ∈ E, for instance by using
Theorem 6.32.

Fig. 6.12. Meaning of U\(bdry U\ bdry A) in Theorem 6.32c.

An equilibrium set E is called attractive if all solution curves in a neigh-

bourhood of E are attracted to the equilibrium set. To prove non-attractivity
of an equilibrium set it does therefore not suffice to prove the non-attractivity
of one equilibrium point x∗1 ∈ E, because repelling solution curves of x∗1 might
still be attracted to x∗2 ∈ E. If we want to prove non-attractivity of E, then
we have to prove the non-attractivity of the set E as a whole. We can do this
for instance with a function V (x) with the following properties

V (x) = 0 ∀x ∈ E, V (x) < 0 ∀x ∈ Uδ (E)\E (6.83)

for some δ > 0. If dV ≤ 0 for all x ∈ Uδ (E) then the equilibrium set is non-
attractive and even unstable if dV < 0 for all x ∈ Uδ (E)\E. However, in the
following theorem we will propose less stringent conditions for instability and
non-attractivity of an equilibrium set.

Theorem 6.33 (Instability and Non-Attractivity of an Equilibrium

Set). Let E ∈ A be a compact equilibrium set of the consistent measure differ-
ential inclusion (6.1), which admits solutions ϕ(·, t0 , x0 ) ∈ S(dΓ , t0 , x0 ) for
all x0 ∈ A. Let V (x) be an extended lower semi-continuous function of the
form (6.27), such that V (xa ) < VE ≤ 0 for some xa , for which distE (xa ) > 0
is arbitrarily small, with VE = minx∈E V (x). Moreover, let V be such that the
set U defined by
 6.8 Summary 151

U = {x ∈ U r (E) ∩ A | V (x) ≤ 0}
for some r > 0 is a nonempty set. Let every limit set L+ in U be positively
invariant. Three statements can be made:
a. If dV < 0 for all x ∈ U\E, then E is unstable.
b. If dV ≤ 0 for all x ∈ U and the solution can not stay in Z = {x ∈ U\E |
dV = 0}, then it holds that E is unstable.
c. If dV ≤ 0 for all x ∈ U and E ⊂ U\(bdry U\ bdry A), then E is not
attractive.

Proof: Theorem 6.33a: The set U is compact as it is the intersection of com-

pact sets. The point xa ∈ / E is in the interior of U and can be taken arbitrarily
close to an equilibrium point x∗ ∈ bdry E. The local continuity of V (x) for
x ∈ int A implies that the function V (x) attains its minimum VE in E for a
point x∗ ∈ bdry E, because V (xa ) < VE with xa taken arbitrarily close to E.
Figure 6.13 shows, similarly to Figure 6.11, the various sets used in the proof
of Theorem 6.33a. Using a very similar reasoning as in the proof of Theo-
rem 6.32a, we can prove that x∗ is an unstable equilibrium point from which
follows the instability of E. Each solution curve ϕ(·, t0 , xa ) ∈ S(dΓ , t0 , xa ) can
not be attracted to E ∩ U because V (ϕ(t, t0 , xa )) < VE will decrease as long as
ϕ(t, t0 , xa ) ∈ U. Moreover, each solution curve can not remain in U because
V would then tend to −∞, which is not possible due to the boundedness of
U. Hence, x∗ is unstable, and so is E.
Theorem 6.33b: We use again the reasoning of LaSalle. A solution curve
ϕ(t, t0 , xa ) might temporarily remain on a level surface V (ϕ(t, t0 , xa )) = c
when ϕ(t, t0 , xa ) ∈ Z, but finally will have to leave Z because there exists no
positively invariant subset of Z.
Theorem 6.33c: Each solution curve ϕ(t, t0 , xa ) ∈ U can not be attracted
to E because V (ϕ(t, t0 , xa )) < VE will not increase as long as ϕ(t, t0 , xa ) ∈ U.
If a solution curve ϕ(t, t0 , xa ) has left U it can not be attracted to E because
E ⊂ int U or E is surrounded by U and Ac . Consequently, the equilibrium
point is not only not attractive, but its repelling neighbouring solution curves
can not return to E. The equilibrium set E is therefore not attractive in the
sense of Definition 6.18. 

6.8 Summary

Lyapunov’s direct method, LaSalle’s invariance principle and Chetaev’s in-

stability theorem have been generalised to prove stability properties of au-
tonomous measure differential inclusions of the form (6.1). All theorems rely
on a suitable choice of a Lyapunov-type function V (x). This is exactly the
weak spot of theorems of this type. How do we find a suitable Lyapunov-type
function V (x)? For mechanical systems with unilateral constraints, however,
152 6 Lyapunov Stability Theory for Measure Differential Inclusions

Fig. 6.13. Sets used in the proof of Theorem 6.33a.

there is a natural choice for the Lyapunov function: the total mechanical en-
ergy which includes the support function of the unilateral constraints.
The Lyapunov-type theorems, which have been presented in this chapter,
will be used in Chapter 7 to derive theorems for the stability and attractivity
of equilibrium sets in mechanical systems with frictional unilateral contact.
7
Stability Properties in
Mechanical Systems

In this chapter, we apply the stability results of Chapter 6 for measure differ-
ential inclusions to Lagrangian mechanical systems with set-valued force laws,
which have been formulated in Chapter 5. The special structure of Lagrangian
mechanical systems allows for a natural choice of the Lyapunov function, a
systematic derivation of the proof for this large class of systems as well as a
physical interpretation of the results.
In the following, we study stability properties of equilibrium sets of the
measure differential inclusion (5.96)

M (q)du − h(q, u)dt = WN (q)dPN + WD (q)dPD ∀t,

where dPN and dPD obey the set-valued constitutive laws (5.98). We assume
existence of solutions of (5.96) for all admissible initial conditions. We denote
an equilibrium position of (5.96) by q ∗ . The generalised velocities u vanish
at an equilibrium point as we assume all contacts to be scleronomic (i.e. the
normal contact distances gN (q) as well as all relative velocities γD (q, u) do
not explicitly depend on time). With E we denote an equilibrium set of the
measure differential inclusion in first-order form, while Eq is reserved for the
union of equilibrium positions q ∗ , i.e. E = {(q, u) ∈ Rn × Rn | q ∈ Eq , u = 0}.

7.1 Total Mechanical Energy

The basic Lyapunov theorems stated in Theorem 6.23 for the (attractive)
stability of an equilibrium point demand the choice of a suitable Lyapunov
function V . The function V should be of the form (6.27)

V (x) = v(x) + ΨA (x), (7.1)

which contains an indicator function on the set A, which is the set of ad-
missible states. When studying mechanical systems it is natural to choose
the total mechanical energy of the system as a Lyapunov candidate function.
154 7 Stability Properties in Mechanical Systems

Motivated by the bouncing ball example of Section 6.5, it will sometimes be

useful to choose the sum of the kinetic energy T (q, u), the potential energy of
the smooth conservative forces U (q) and the energy potential of the normal
contact forces as Lyapunov candidate function.
Let the potential Q(q) be the total potential energy of the system

Q(q) = U (q) − U (0) + πN i (gN i (q))
i∈IG
 (7.2)
= U (q) − U (0) + ΨR+ (gN i (q)).
i∈IG

which is the sum of the potential energy U (q) of all smooth potential forces
and the support functions πN i (gN i (q)) of the normal contact forces. The
support function πN i (gN i (q)) equal indicator functions ΨR+ (gN i (q)) which
express the unilaterality of the constraints gN i (q) ≥ 0, which restrict the
generalised coordinates q to the set K

K = {q ∈ Rn | gN i (q) ≥ 0 ∀i}, (7.3)

which we call the set of admissible generalised coordinates. We can therefore

write the total potential energy function as

Q(q) = U (q) − U (0) + ΨK (q). (7.4)

Moreover, in the stability proofs of Section 7.2 we will assume the total po-
tential energy Q(q) to be a locally positive definite function such that

Q(0) = 0, Q(q) > 0 ∀q ∈ U\{0}, (7.5)

where U is some neighbourhood of the origin.

The kinetic energy T (q, u) is assumed to be a symmetric positive definite
quadratic form in u
1
T (q, u) = uT M (q)u (7.6)
2
with M (q) = M (q)T > 0.
In the following, we will use the total mechanical energy function

Etot = T (q, u) + Q(q) = T (q, u) + U (q) − U (0) + ΨK (q), (7.7)

being the sum of kinetic and total potential energy. The total mechanical
energy function Etot (7.7) is locally positive definite if the total potential
energy is locally positive definite and Etot has the property

Etot (q, u) = ∞ ∀q ∈
/ K, (7.8)

where K is defined by (7.3). If we gather the generalised positions q and ve-

 T
locities u in the state-vector x = q T uT , then the set K induces a set
 7.1 Total Mechanical Energy 155

A = {x | q ∈ K} of admissible states. Clearly, the total mechanical en-

ergy function Etot (q, u) is an extended lower semi-continuous function of the
form (6.27). Moreover, the states (q, u) of a mechanical system consist of gen-
eralised positions q(t), which are assumed to be absolutely continuous, and
generalised velocities u(t), which are of locally bounded variation. Only the
generalised positions are restricted to an admissible set K, whereas no restric-
tion exists on the velocities u. The total mechanical energy function Etot (q, u)
consists of a (possibly) locally positive definite function Q(q) which is solely
dependent on q (being absolutely continuous) and a function T (q, u) which
is a PDF in u for fixed values of q. Consequently, a mechanical system (5.96)
is a measure differential inclusion with the property (6.33) and the total me-
chanical energy Etot (q, u) is a function of the form (6.34) if Q(q) is a LPDF.
We compute the differential measure of Etot along solution curves of (5.96):

dEtot = dT + dQ. (7.9)

The total potential energy is only a function of the generalised displacements

and it holds that
dQ = dQ(q)(dq)
(7.10)
= U,q dq + dΨK (q)(dq)

where dQ(q)(dq) is the subderivative (see Section 2.6) of Q at q in the direc-

tion dq. The subderivative dΨK (q)(dq) of the indicator function ΨK (q) equals
the indicator function on the associated contingent cone KK (q) (see (2.54))

dΨK (q)(dq) = ΨKK (q) (dq). (7.11)

It holds that dq = udt with u = u+ ∈ KK (q) due to the consistency of the

system and the indicator function on the contingent cone therefore vanishes,
i.e. ΨKK (q) (udt) = 0. The differential measure of Q simplifies to

dQ = U,q dq + ΨKK (q) (dq)

= U,q udt + ΨKK (q) (udt), u ∈ KK (q) (7.12)
= U,q udt.

Recalling that the kinetic energy T (q, u) (7.6) is a quadratic form and
using the results of Section 3.6, we deduce that the differential measure of T
is
1
dT = (u+ + u− )T M (q)du + T,q dq. (7.13)
2
The differential measure of Etot becomes by (7.9), (7.12) and (7.13)
1 +
dEtot = (u + u− )T M (q)du + (T,q + U,q ) dq
2 (7.14)
(5.101) 1
= (u+ + u− )T (h(q, u)dt + W dP ) + (T,q + U,q ) udt
2
156 7 Stability Properties in Mechanical Systems

A term 12 (u+ + u− )T dt in front of a Lebesgue measurable term equals uT dt.

T
Together with (5.86), yielding h = f nc + f gyr − (T,q + U,q ) and (5.97)
with (5.102) we obtain
1
dEtot = uT f nc dt + uT f gyr dt + uT W λdt + (u+ + u− )T W Λdη (7.15)
2
The gyroscopic forces have zero power, i.e. uT f gyr = 0 as motivated by (5.84).
Moreover, the constraints are assumed to be scleronomic and it therefore holds
according to (5.91) that γ = W T u, which gives
1
dEtot = uT f nc dt + γ T λdt + (γ + + γ − )T Λdη
2
(5.104) T nc 1 T
= u f dt + γ T λdt + (I + E)−1 (2ξ − (I − E)δ) Λdη
2
1
= uT f nc dt + γ T λdt + ξ T (I + E)−1 Λdη − δ T (I − E)(I + E)−1 Λdη
2
(5.97)+(5.103) T nc 1
= u f dt + ξ T (I + E)−1 dP − δ T (I − E)(I + E)−1 Λdη
2
  ξN i dPN i ξ T
dPDi

1
= uT f nc dt + + Di − δ T Δ Λdη,
1 + eN i 1 + eDi 2
i∈IN
(7.16)

where Δ := (I − E)(I + E)−1 is the dissipation index matrix (5.105). Us-

ing (5.98) and (2.33), we obtain

ξN i dPN i = −ΨC∗ N (ξN i )(dt + dη) = 0

(7.17)
T
ξDi dPDi = −ΨC∗ Di (λN i ) (ξDi )dt − ΨC∗ Di (ΛN i ) (ξDi )dη ≤ 0,

because of (2.37) and ΨC∗ N (ξN i ) = ΨR+ (ξN i ) = 0 for admissible ξN i ≥ 0.

Moreover, applying (5.91) to (5.103) and using the impact equation (5.100)
gives

δ := γ + − γ − = W T (u+ − u− ) = W T M −1 W Λ = GΛ, (7.18)

in which we used the abbreviation

G := W T M −1 W , (7.19)

which is known as the Delassus matrix. The matrix G is positive definite

when W has full rank, because M > 0. The matrix G is only positive semi-
definite if the matrix W does not have full rank, meaning that the generalised
force directions of the contact forces are linearly dependent. However, we
assume that the matrix W only contains the generalised force directions of
unilateral constraints, and that these unilateral constraints do not constitute
a bilateral constraint. It therefore holds that there exists no ΛN = 0 such
 7.1 Total Mechanical Energy 157

that WN ΛN = 0. The impact law requires that ΛN ≥ 0. Hence, it holds that

−1
ΛT T
N WN M WN ΛN > 0 for all ΛN such that ΛN = 0 and ΛN ≥ 0, even if
the unilateral constraints are linearly dependent. Moreover, ΛT = 0 implies
ΛN = 0. The inequality ΛT GΛ > 0 therefore holds for all Λ = 0 which obey
the impact law (5.74), even if dependent unilateral constraints are considered.
Using (7.18), we can put the last term in (7.16) in the quadratic form
1 T 1
δ ΔΛdη = ΛT GΔΛdη. (7.20)
2 2
in which GΔ is a square matrix. The dissipation index matrix Δ (5.105)
is a diagonal matrix which is positive definite if the contacts are not purely
elastic, i.e. 0 ≤ eN i < 1 and |eDi | < 1 for all i. However, the matrix GΔ
is in general not a positive definite matrix. We first present two propositions
which guarantee the positive definiteness of the product of two square positive
definite matrices.

Proposition 7.1. Let A = AT ∈ Rn×n and B = B T ∈ Rn×n be two square

symmetric matrices which commute, i.e. AB = BA. If A > 0 and B > 0,
then it holds that AB > 0. If A > 0 and B ≥ 0, then it holds that AB ≥ 0.

Proof: The matrix AB is symmetric, because AB = (AB)T . The matrix

A is non-singular because it is positive definite and symmetric. Let λ be an
eigenvalue of AB with eigenvector v,

ABv = λv,

which we can write as Bv = λA−1 v. The Rayleigh coefficient shows that the
matrix AB has real eigenvalues,

v T Bv
λ= ,
v T A−1 v
which are positive if B > 0 and non-negative if B ≥ 0. Moreover, because
AB is symmetric and has positive or non-negative real eigenvalues it holds
that
xT ABx ≥ λmin (AB) x2 ,
where λmin (AB) is the smallest eigenvalue of AB. Hence, if B > 0, then
xT ABx > 0 for all x = 0 and if B ≥ 0, then xT ABx ≥ 0 for all x. 

Proposition 7.2. Let A ∈ Rn×n be a symmetric positive definite matrix and

B ∈ Rn×n be a diagonal positive definite matrix with the diagonal elements
bii which fulfil 1 ≥ bii ≥ bmin > 0, i = 1, . . . , n. If 1 − bmin < cond(A)
1
then it
holds that the matrix AB is positive definite.

Proof: The matrix A = AT > 0 has real positive eigenvalues and it therefore
holds that
158 7 Stability Properties in Mechanical Systems

xT Ax ≥ λmin (A) x2 , (7.21)

where λmin (A) is the smallest eigenvalue of A. Moreover, it holds that

xT A(I − B)x ≤ |xT A(I − B)x|

≤ |A| |I − B| x2 (7.22)
≤ λmax (A) (1 − bmin )x ,2

where λmax (A) is the largest eigenvalue of A and bmin ≤ 1 is the smallest
diagonal element of B. Using the above inequalities, we deduce that
xT ABx = xT (A − A(I − B))x
≥ (λmin (A) − λmax (A) (1 − bmin )) x2 .
Hence, if it holds that
λmin (A) 1
1 − bmin < =: ,
λmax (A) cond(A)
then it follows that xT ABx > 0 holds for all x = 0. 

Proposition 7.3. Let A ∈ Rn×n be a symmetric positive definite matrix and

B ∈ Rn×n be a diagonal positive definite matrix with the diagonal elements
bii which fulfil 1 ≥ bmax ≥ bii ≥ bmin > 0, i = 1, . . . , n. If it holds that
cond(B) − 1 1
< ,
cond(B) + 1 cond(A)
then the matrix AB is positive definite.
Proof: The matrix A = AT > 0 has real positive eigenvalues and it therefore
holds that
xT Ab̄ x ≥ b̄λmin (A) x2 ,
where λmin (A) is the smallest eigenvalue of A and b̄ = 12 (bmax + bmin ) is the
mean of the largest and smallest diagonal element of B. Moreover, it holds
that
xT A(b̄I − B)x ≤ |xT A(b̄I − B)x|
≤ |A| |b̄I − B| x2
≤ λmax (A) (bmax − b̄)x2 ,

where λmax (A) is the largest eigenvalue of A. Using the above inequalities,
we deduce that
xT ABx = xT (Ab̄ − A(b̄I − B))x
 
≥ b̄λmin (A) − λmax (A) (bmax − b̄) x2 .
1
= ((bmax + bmin )λmin (A) − (bmax − bmin )λmax (A)) x2 .
2
 7.1 Total Mechanical Energy 159

Hence, if it holds that

bmax − bmin λmin (A)

< ,
bmax + bmin λmax (A)

which is equivalent to

cond(B) − 1 1
< ,
cond(B) + 1 cond(A)

then it follows that xT ABx > 0 holds for all x = 0. 

In the theorems which are to be presented in the remainder of this work,
the assumption will be made that the system is energetically consistent in the
sense that the contact impulses Λ are dissipative, i.e.
1 +
(γ + γ − )T Λ ≤ 0. (7.23)
2
If no conditions on the restitution coefficients exist (other than 0 ≤ eN i <
1 and |eDi | < 1∀i) and if friction is present, then the impact laws (5.74)
and (5.75) can, under circumstances, lead to an energy increase. Such an
energetic inconsistency has been reported by Kane and Levinson [84] (see
also [23]). We will give sufficient conditions for the system parameters such
that energetic consistency (7.23) can be guaranteed:
1. Frictionless contacts. If friction is absent then there exist only normal
contact impulses. Let IA = {i ∈ IN | ΛN i > 0} be the set of contact
points which participate in the contact. The dissipation due to the contact
impulses becomes
1 + 1  + −
(γ + γ − )T Λ = (γN i + γN i )ΛN i
2 2
i∈IA
(7.24)
1  −
= (1 − eN i )γN i ΛN i
2
i∈IA

Unilaterality of the contact implies ΛN i ≥ 0, whereas kinematic compat-

ibility requires γN i ≤ 0. Hence, if eN i < 1 for all i, then each term in
the above sum is negative, implying that each normal contact dissipates
energy.
2. Commuting matrices G and Δ. If the Delassus matrix G > 0 is pos-
itive definite (no dependent constraints) and if Δ > 0 (it holds that
0 < eN i < 1 and |eDi | < 1), then the matrix GΔ is positive definite (see
Proposition 7.1) and energetic consistency is guaranteed. Moreover, if one
of these matrices is positive definite and the other is positive semi-definite
while they still commute, then energetic consistency is still guaranteed.
There are some special cases for which these matrices commute:
160 7 Stability Properties in Mechanical Systems

a) Equal restitution coefficients. If all restitution coefficients are equal,

i.e. eN i = eDi = e for all i, then the matrix Δ simplifies to 1−e1+e I,
which commutes with G. This case is equivalent to Moreau’s global
dissipation index and includes the special case of completely inelastic
contacts e = 0.
b) Decoupling. If G is diagonal, then the contacts are decoupled and G
and Δ commute as they are both diagonal.
3. Small restitution coefficients. The smallest diagonal element of Δ is (1 −
emax )/(1 + emax ) where emax is the largest diagonal element of E. Using
Proposition 7.1 we derive that if the restitution coefficients are small in
the sense that
2emax 1
1 − Δmin = < ∀q ∈ K, (7.25)
1 + emax cond(G(q))

then the matrix GΔ is positive definite and energetic consistency is there-

fore guaranteed.
4. Almost equal restitution coefficients. Using Proposition 7.3 we derive that
if the restitution coefficients are almost equal in the sense that

cond(Δ) − 1 1
< ∀q ∈ K, (7.26)
cond(Δ) + 1 cond(G(q))

then the matrix GΔ is positive definite and energetic consistency is there-

fore guaranteed. This case includes the special case 2(a) of equal restitu-
tion coefficients.
Looking again at the differential measure of the total energy (7.16), we
realise that if the above conditions are met then all terms related to the
contact forces and impulses are dissipative or passive. Moreover, if we consider
not purely elastic contacts, then nonzero contact impulses Λ strictly dissipate
energy.
We define the following nonlinear functionals Rn → R on u ∈ Rn :
• Dqnc (u) := −uT f nc (q, u) is the (rate of) dissipation function1 of the
smooth non-conservative

forces,
• DqλD (u) := i∈IN 1+e1 Di ΨC∗ Di (λN i ) (ξDi (q, u)) ≥ 0 is the dissipation func-
tion of the frictional

contact forces,
• DqΛD (u) := i∈IN 1+e1 Di ΨC∗ Di (ΛN i ) (ξDi (q, u)) ≥ 0 is the dissipation func-
tion of the frictional contact impulses.
The dissipation functions DqλD (u) and DqΛD (u) are non-negative because they
consist of convex support functions whereas a priori nothing can be said of
the dissipation function Dqnc (u) of the smooth non-conservative forces. For
1
Usually, the term ‘dissipation function’ is used to denote the pseudo-potential of
the non-conservative forces. Here, we use the term dissipation function for the
function Dqnc (u), which gives the rate of dissipation.
 7.1 Total Mechanical Energy 161
+ −
non-impulsive motion, it holds that γD = γD = γD and ξD = (1 + eD )γD .
Due to the fact that the support function is positively homogeneous it follows
that
 
DqλD (u) = ΨC∗ Di (λN i ) (γDi (q, u)) = −λDi γDi (q, u), (7.27)
i∈IN i∈IN

from which we see that the dissipation function of the frictional contact forces
does not depend on the restitution coefficient eD . The above dissipation func-
tions are of course functions of (q, u), but we write them as nonlinear func-
tionals on u so that we can speak of the zero set of the functional Dq (u):

Dq−1 (0) = {u ∈ Rn | Dq (u) = 0}. (7.28)

We can now decompose the differential measure dEtot in a Lebesgue part

and an atomic part
−
+
dEtot = Ėtot dt + (Etot − Etot )dη, (7.29)

with
 1
Ėtot = uT f nc − Ψ∗ (ξDi )
1 + eDi CDi (λN i ) (7.30)
i∈IN

= −Dqnc (u) − DqλD (u)

and
 1

1
−
+
Etot − Etot =− ΨC∗ Di (ΛN i ) (ξDi ) − ΛT GΔΛ
1 + eDi 2
i∈IN (7.31)
1
= −DqΛD (u) − ΛT GΔΛ.
2
We see that the dissipation rate during non-impulsive motion is governed by
the dissipation of the non-conservative smooth forces and the dissipation of the
(Lebesgue measurable) friction forces, whereas the jumps in the total energy
are due to the dissipation of the frictional impulses and the dissipation caused
−
by impact. Energetic consistency of the system requires that Etot +
− Etot ≤ 0.
A sufficient condition for energetic consistency is the positiveness of the term
2 Λ GΔΛ. Sufficient conditions for 2 Λ GΔΛ ≥ 0 have been derived in the
1 T 1 T

conditions 1-4 on page 159.

Remark: In Section 5.4 (Equation (5.81) or (7.6)), we assumed that the
the kinetic energy is a purely quadratic form in the generalised velocities:
T = 12 uT M (q)u. Subsequently, the analysis in Section 5.4 has been based
on this assumption, as are the expressions for the differential measure of the
total mechanical energy function in (7.29)-(7.31). The latter expression will
prove to be of central importance in exploiting Etot as a Lyapunov function
162 7 Stability Properties in Mechanical Systems

candidate in the next section. One could consider the more general case of a
kinetic energy of the form

T (q, u) = T2 (q, u) + T1 (q, u) + T0 (q), (7.32)

with
1 T
T2 (q, u) := u M (q)u,
2 (7.33)
T1 (q, u) := mT (q)u.

In this more general case, one should use an adapted form of the Lyapunov
function candidate, which involves the function Ētot = T2 (q, u)−T0 (q)+Q(q)
(instead of Etot = T2 (q, u) + Q(q)). The function Ētot is called the Jacobi-
Painlevé generalised energy [129]. Under the assumption that T2 , T1 and T0 do
not depend explicitly on time, similar expressions for dĒtot can be derived (as
in (7.29)-(7.31) for dEtot ). For the sake of simplicity, we restrict ourselves to
the case of a purely quadratic kinetic energy in the next section. However, we
mention that, using the Jacobi-Painlevé generalised energy function, the more
general case (i.e. for systems characterised by a kinetic energy as in (7.32))
can be treated in an analogous fashion.

7.2 Stability Results for Mechanical Systems

The stability of the equilibrium position of the bouncing ball example has been
proven in Section 6.5 by using Theorem 6.23a. Mechanical systems, for which
the generalised position q(t) are absolutely continuous and the generalised
velocities u(t) are of locally bounded variation, are of the form (6.33). We
can therefore use Theorem 6.25a to prove the stability of an equilibrium point
of a general mechanical system (5.96), taking the total mechanical energy
function (7.7) as a Lyapunov candidate function.

Theorem 7.4 (Lyapunov Stability of an Equilibrium Position at the

Origin). Let q ∗ = 0 ∈ K be an equilibrium position of (5.96) and let the
total potential energy Q as in (7.4) be a locally positive definite function. If
Dqnc (u) + DqλD (u) ≥ 0 ∀(q, u) ∈ Bh ∩ (K × Rn ) for some h > 0, and if the
system is energetically consistent, then the equilibrium position is stable.

Proof: Consider the Lyapunov function candidate V = Etot given by (7.7)

which is a locally positive definite function with V (0, 0) = 0 and V (q, u) >
0 ∀(q, u) ∈ (U ×Rn )\(0, 0). The differential measure dV consists of a Lebesgue
part and an atomic part

dV = V̇ dt + (V + − V − )dη. (7.34)
 7.2 Stability Results for Mechanical Systems 163

Using Dqnc (u) + DqλD (u) ≥ 0 and (7.29) it follows that V̇ ≤ 0 ∀(q, u) ∈
Bh ∩ (K × Rn ) for some h > 0. Moreover, it holds that V + − V − ≤ 0 because
of the system is assumed to be energetically consistent. The total mechanical
energy V does therefore not increase along solution curves of the system in the
neighbourhood of the origin. Lyapunov stability in the sense of Definition 6.13
follows from Theorem 6.25a. 
Notice that it is not useful to state Theorem 6.25b or Theorem 6.23c for
mechanical systems using the total energy function as Lyapunov function V ,
because the dissipation is absent for u = 0 and dV (q, 0) therefore vanishes for
all q ∈ K. Hence, attractivity has to be proven separately using LaSalle-type
of arguments.
Up to now, we only considered the stability of an equilibrium position
q ∗ = 0. If we are interested in the stability of an equilibrium position q ∗
which is not located at the origin, then we we have to consider an altered
potential energy function Ua (q) such that the altered total potential energy
function 
Qa (q) = Ua (q) + πN i (gN i (q)), (7.35)
i∈IG

fulfills the conditions

Qa (q ∗ ) = 0, Qa (q) > 0 ∀q ∈ U\{q ∗ } (7.36)

where U is some neighbourhood of q ∗ . The latter condition is a local positive

definiteness requirement for Qa (q) with respect to the point q ∗ . Using Qa (q),
we can construct the Lyapunov function Vq∗ (q, u) = T (q, u) + Qa (q) which
can be used to prove the stability of q ∗ . The big question is: how do we find
an altered total potential energy function Ua (q) which fulfills (7.36)? We are
not able to give one function Ua (q) that can be used to prove the stability of
equilibrium positions of a general mechanical system (5.96), but we will study
some special cases.
If the total potential energy function Q(q) (7.4) has a local minimum at q ∗
and is locally strictly convex, then the altered total potential energy function
Qa (q) based on the shifted potential energy function Ua (q), with

Ua (q) = U (q) − U (q ∗ ), (7.37)

fulfills the conditions (7.36). Using dUa = dU it follows that the differential
measure of Vq∗ (q, u) = T (q, u) + Qa (q) agrees with the differential measure
of the total mechanical energy function V , i.e. dVq∗ = dV = dEtot , and
the stability therefore depends on the sign of the term Dqnc (u) + DqλD (u).
This special case may occur for instance when only frictionless contacts are
considered [33]. An example is the bouncing ball example, but with a floor on
a non-zero height.
Another special case occurs when the system only contains frictional bilat-
eral contacts with an associated Coulomb friction law (5.50) with a constant
164 7 Stability Properties in Mechanical Systems

normal contact force FN , i.e. the system is described by the differential inclu-
sion

M (q)u̇ − h(q, u) = WD (q)λD , −λD ∈ ΨC∗ D (γD ), (7.38)

where CD is a constant set. Generally, an equilibrium q ∗ is accompanied by

a nonzero static friction force λ∗D which makes equilibrium with the vector h
of smooth forces:
−h(q ∗ , 0) = WD (q ∗ )λ∗D . (7.39)
The vector h (5.86) consists of non-conservative forces f nc (q, u), which are
assumed to vanish for u = 0, forces f gyr (q, u) + T,q (q, u) derived from the
kinetic energy which also vanish at an equilibrium point and forces U,q (q)
derived from the potential energy. The force equilibrium (7.39) therefore sim-
plifies to
U,q (q ∗ )T = WD (q ∗ )λ∗D . (7.40)
In order to prove the (conditional) stability of an equilibrium position q ∗ = 0
of (7.38), we consider the altered potential energy function

Ua (q) = U (q) − (q − q ∗ )T WD (q ∗ )λ∗D − U (q ∗ ). (7.41)

If U (q) consists of a linear and quadratic form (e.g. the system is linear)
1 T
U (q) = q Kq − q T f0 (7.42)
2
then it holds that U,q (q ∗ )T = Kq ∗ − f0 which gives
1 T 1
Ua (q) = q Kq − q T f0 − (q − q ∗ )T WD (q ∗ )λ∗D − q ∗ T Kq ∗ + q ∗ T f0
2 2
1 1
= q T Kq − q T f0 − (q − q ∗ )T (Kq ∗ − f0 ) − q ∗ T Kq ∗ + q ∗ T f0
2 2
1 1
= q T Kq − q T Kq ∗ + q ∗ T Kq ∗
2 2
1
= (q − q ∗ )T K(q − q ∗ ).
2
(7.43)

This shows that the altered potential energy function of a linear system with
K > 0 is positive definite with respect to q ∗ , i.e. (7.36) is satisfied. Moreover,
it holds that Ua (q) = U (q − q ∗ ) if f0 = 0. If we now consider the differential
measure of the Lyapunov function candidate

Vq∗ (q, u) = T (q, u) + Ua (q)

(7.44)
= Etot (q, u) − (q − q ∗ )T WD (q ∗ )λ∗D − U (q ∗ ),

then the additional terms only affect the Lebesgue part V̇q∗ of the differential
measure dVq∗ :
 7.2 Stability Results for Mechanical Systems 165

Fig. 7.1. One-DOF system with friction.

V̇q∗ = Ėtot − uT WD (q ∗ )λ∗D

(7.45)
= −Dqnc (u) − DqλD (u) − uT WD (q ∗ )λ∗D

If WD is constant, then V̇q∗ simplifies to

T ∗
V̇q∗ = −Dqnc (u) − DqλD (u) − γD λD . (7.46)

It holds that γD ∈ ∂ΨCD (−λD ). Using (2.26) it immediately follows that

0 ≥ −γDT
(λ∗D − λD ) for all −λ∗D ∈ CD . The dissipation of the friction forces
T ∗
DqλD (u) + γD T
λD = γD (λ∗D − λD ) ≥ 0 (7.47)

is therefore always non-negative.

Theorem 7.5 (Lyapunov Stability of an Equilibrium Position). Let

q ∗ ∈ K be an equilibrium position of the differential inclusion (7.38) and let
the altered potential energy

Ua (q) = U (q) − (q − q ∗ )T WD (q ∗ )λ∗D − U (q ∗ )

be a locally positive definite function with respect to q ∗ , i.e. (7.36) is satisfied.

If Dqnc (u)+DqλD (u)+uT WD (q ∗ )λ∗D ≥ 0 ∀(q, u) ∈ Bh ∩(K×Rn ) for some h >
0, and if the system is energetically consistent, then the equilibrium position
q ∗ is stable.

Proof: The proof is identical to the proof of Theorem 7.4, but we now use
the Lyapunov function Vq∗ (q, u) given by (7.44). 
If we are able to prove the stability of each equilibrium position in an equi-
librium set Eq , then stability of the equilibrium set as a whole follows from
Theorem 6.28, or alternatively from Theorem 6.27 using a global Lyapunov
function (6.57).
We now discuss a one-degree-of-freedom example which illustrates the use
of Theorem 7.5. Consider a block with mass m on a floor which is attached to
a wall by a spring k and a dashpot c (see Figure 7.1). The spring is unstressed
for q = 0. Associated Coulomb friction with friction coefficient μ is present
166 7 Stability Properties in Mechanical Systems

between the block and the floor. The contact force FN has the constant value
mg. The dynamics of the system is described by the differential inclusion

q̇ = u, mu̇ + cu + kq = λT , −λT ∈ μFN Sign(u) a.e. (7.48)

and has the equilibrium set

) μmg μmg *
Eq = {q ∈ R | kq ∈ −μFN Sign(0)} = − , . (7.49)
k k
An equilibrium position q ∗ is accompanied by a friction force λ∗T = kq ∗ ∈ −CT
with CT = [−μmg, μmg]. We consider the stability of an equilibrium position
q ∗ ∈ Eq using the Lyapunov function

Vq∗ (q, u) = T (u) + Ua (q)

1 1 (7.50)
= mu̇2 + k(q − q ∗ )2
2 2
where Ua (q) is the altered potential energy function (7.41) which can be sim-
plified as in (7.43). It holds that Vq∗ (q ∗ , 0) = 0 and Vq∗ (q, u) > 0 for all
(q, u) = (q ∗ , 0) and the Lyapunov function is therefore positive definite with
respect to the equilibrium point (q ∗ , 0). The time derivative of Vq∗ along so-
lution curves of the system is given by

V̇q∗ = muu̇ + k(q − q ∗ )u

= −cu2 + λT u − kq ∗ u
= −cu2 + (λT − λ∗T )u (7.51)
= −cu − μmg|u| −
2
λ∗T u, |λ∗T | ≤ μmg
≤0

and is non-positive for c ≥ 0 for all equilibrium positions q ∗ ∈ Eq . Conse-

quently, each equilibrium position q ∗ ∈ Eq is stable. Stability of Eq as a whole
follows from Theorem 6.28. Alternatively, we can use Theorem 6.26 to prove
the stability of the equilibrium set taking a Lyapunov function V which is
zero on the whole equilibrium set (see [170]). To this end, we introduce a
 T √ √ T
coordinate transformation z = z1 z2 = kq mu . The equilibrium set
E in z-coordinates is given by the convex set
 
μFN
E = z ∈ R | |z1 | ≤ √ , z2 = 0 .
2
(7.52)
k
Consider the Lyapunov candidate function V , as defined by
1
V (z) = dist2E (z), (7.53)
2
where Definition 2.32 has been used. This Lyapunov function V has the de-
sirable property that V (z) = 0 for z ∈ E, which allows for the stability
 7.2 Stability Results for Mechanical Systems 167
T
analysis of the equilibrium set as a whole. Using V̇ = (z − proxE (z)) ż it
is shown in [170] that V̇ ≤ 0 along solution curves of the system for c ≥ 0.
Consequently, Lyapunov stability of the equilibrium set E follows from The-
orem 6.26 for c ≥ 0. Moreover, if we evaluate the distance function using
proxE (z) = [qp 0]T , with qp = proxEq (q), as

distE (z) = z − proxE (z)

+ (7.54)
= k(q − qp )2 + mu2 ,

then we see that the Lyapunov function V (7.53) has the form
1
V (z) = dist2E (z)
2
1 1
= mu2 + k(q − qp )2 (7.55)
2  2 
1 1
= min mu + k(q − qp ) ,
2 2
qp ∈Eq 2 2
which we recognise to be a Lyapunov function of the form (6.57)

V (q, u) = min
∗
Vq∗ (q, u), (7.56)
q ∈Eq

where Vq∗ (q, u) is given by (7.50). It is also of interest to consider the system
in z-coordinates in the form

ż ∈ −F (z), (7.57)

with  √ 
− ku
F (z) = √1 . (7.58)
m
(cu + kq + μFN Sign(u))
It holds that
(z − z ∗ )T (F (z) − F (z ∗ ))

= −k(q − q ∗ )(u − u∗ ) + (u − u∗ ) cu + kq + μFN Sign(u)

− cu∗ − kq ∗ − μFN Sign(u∗ ) (7.59)
= (u − u∗ ) (cu + μFN Sign(u) − cu∗ − μFN Sign(u∗ ))
≥0

because μFN Sign(u) + cu is maximal monotone in u for c ≥ 0 and μFN ≥ 0.

Consequently, the operator F (z) is maximal monotone, but not strictly max-
imal monotone because of its dependence on q. Stability of the equilibrium
set E immediately follows from Theorem 6.29. Moreover, the system is incre-
mentally stable (see Theorem 6.30). Similar stability results for Lagrangian
systems described by differential inclusions with maximal monotone operators
can be found in [3].
168 7 Stability Properties in Mechanical Systems

7.3 Attractivity of Equilibrium Sets

In this section, we will investigate the attractivity properties of equilibrium
sets of Lagrangian mechanical systems with impact and/or friction using a
generalisation of LaSalle’s invariance principle (Theorem 6.31). We first dis-
cuss attractivity of equilibrium sets in systems with frictional unilateral con-
straints in Subsection 7.3.1 and subsequently discuss systems with frictional
bilateral constraints as a special case in Subsection 7.3.2.

7.3.1 Systems with Frictional Unilateral Constraints

As stated before, the type of systems under investigation described by (5.96),

(5.98) may exhibit multiple equilibrium sets. Here, we will study the attrac-
tivity properties of a specific given equilibrium set E satisfying (5.109). The
set E may be the total equilibrium set of the system or just only a simply
connected subset of the total equilibrium set. By qe we denote an equilibrium
point of the system with unilateral frictionless contacts

M (q)du − h(q, u)dt − WN (q)dPN = 0, (7.60)

from which follows that the equilibrium is determined by the inclusion


h(qe , 0) − wN i (qe )∂ΨC∗ N (gN i (qe ))  0 (7.61)
i∈IG

or 
h(qe , 0) − wN i (qe )∂ΨC∗ N (γN i (qe , 0))  0, (7.62)
 
i∈IN
=0
which is equivalent to

h(qe , 0) + WN (qe )R+  0, WN = {wN i }, i ∈ IN . (7.63)

We assume that the equilibrium position qe is a local minimum of the total

potential energy function Q(q) (7.4), i.e.

0 q = qe
Q(q) − Q(qe ) = , 0∈/ ∂Q(q), ∀q ∈ U\{qe }. (7.64)
> 0 ∀q ∈ U\{qe }

The subset U is assumed to enclose the equilibrium set Eq of the equilibrium

positions under investigation, where Eq = {q | (q, 0) ∈ E}. Notice that the
equilibrium point qe of the system without friction is also an equilibrium point
of the system with friction, qe ∈ Eq . In case the system does exhibit multiple
equilibrium sets, the attractivity of E will be only local for obvious reasons.
In the following, we will use the total mechanical energy (7.7) shifted with
respect to qe as Lyapunov candidate function V (q, u) = Etot (q, u) − Q(qe ).
We now formulate a theorem which states sufficient conditions under which
an equilibrium set E can be proven to be (locally) attractive.
 7.3 Attractivity of Equilibrium Sets 169

Theorem 7.6 (Attractivity of the equilibrium set).

Consider system (5.96), with constitutive laws (5.98) and with M (q) =
M T (q) > 0. If
1. the equilibrium position qe is a local minimum of the total potential en-
ergy Q(q) and Q(q) has a non-vanishing generalised gradient for all
q ∈ U\{qe }, i.e. 0 ∈ / ∂Q(q) ∀q ∈ U\{qe }, and the equilibrium set Eq
is contained in U, i.e. Eq ⊂ U,
2. Dqnc (u) = −uT f nc ≥ 0, i.e. the smooth non-conservative forces are dissi-
pative, and f nc = 0 for u = 0,
3. there exists a non-empty set IC ⊂ IG and an open neighbourhood V ⊂ Rn ×
Rn of the equilibrium set, such that γ̇N i (q, u) < 0 (a.e.) for ∀i ∈ IC \IN
and (q, u) ∈ V (see (5.89) for the definition of the index sets),
−1
4. Dqnc −1 (0) ∩ DqλDC (0) ∩ ker WNTC (q) = {0} ∀q ∈ C with
gN C = {gN i }, WN C = {wN i } for i ∈ IC as defined in 3.,

C = {q | gN C (q) = 0}, DqλDC (u) = ΨC∗ Di (λN i ) (γDi (q, u)),
i∈IC ∩IN

5. 0 ≤ eN i < 1, |eDi | < 1 ∀i ∈ IG ,

6. E ⊂ Ωρ∗ in which the set Ωρ∗ is the largest level set of V (q, u) =
Etot (q, u) − Q(qe ), that is contained in V and Q = {(q, u) ∈ Rn × Rn |
q ∈ U}, i.e.
ρ∗ = max ρ,
{ρ:Ωρ ⊂(V∩Q)}

7. the system is energetically consistent,

8. E is the largest set of equilibrium points in Ωρ∗ ,
9. each limit set in Ωρ∗ is positively invariant,
then the equilibrium set E ⊃ (qe , 0) is locally attractive and Ωρ∗ is a conser-
vative estimate for the region of attraction.
Proof: Note that V is positive definite around the equilibrium point (q, u) =
(qe , 0) due to condition 1 in the theorem. The differential measure dV = dEtot
has been computed in Section 7.1. We can decompose the differential measure
dV in a Lebesgue part and an atomic part:
dV = V̇ dt + (V + − V − )dη, (7.65)
with (7.30)
V̇ = −Dqnc (u) − DqλD (u), (7.66)
which is non-positive due to condition 2, and (7.31) gives
1
V + − V − = −DqΛD (u) − ΛT GΔΛ, (7.67)
2
which is non-positive due to condition 5 and 7. For positive differential mea-
sures dt and dη we deduce that the differential measure of V (7.65) is non-
positive, dV ≤ 0. There are a number of cases for dV to distinguish:
170 7 Stability Properties in Mechanical Systems

• Case u = 0: It directly follows that dV = 0.

−
• Case gN i = 0 and γN i < 0 for some i ∈ IN : One or more contacts are
closing, i.e. there are impacts. It follows from (7.67) that V + − V − < 0
and dV < 0.
• Case gN C = 0, u ∈ ker WNTC and u = u− = u+ with gN C = {gN i } for
i ∈ IC : It then holds that all contacts in IC are closed and remain closed,
IC ⊂ IN . We now consider V̇ as a nonlinear functional on u and write

V̇ (q, u) = 0, u ∈ V̇q−1 (0),

(7.68)
/ V̇q−1 (0),
V̇ (q, u) < 0, u ∈

with
−1
V̇q−1 (0) = Dqnc −1 (0) ∩ DqλD (0)
−1
(7.69)
⊂ Dqnc −1 (0) ∩ DqλDC (0).

Condition 4 states that, if the contacts in IC are persistent (WNTC u = 0),

−1
then dissipation can only vanish if u = 0, i.e. Dqnc −1 (0)∩DqλDC (0) = {0}.
In other words, if all contacts in IC are closed and remain closed and u = 0
then dissipation is present. Using condition 4 and u ∈ ker WNTC \ {0}, it
follows that V̇q−1 (0) = {0} and hence

V̇ = 0, u = 0,
(7.70)
 0.
V̇ < 0, u =

Impulsive motion for this case is excluded. For a strictly positive differen-
tial measure dt we obtain the differential measure of V as given in (7.65)

dV = 0, u = 0,
(7.71)
 0.
dV < 0, u =

• Case gN C = 0, u ∈ / ker WNTC : It then holds that wN T

i u > 0 for some
i ∈ IC and one or more contacts will open. All we can say is that dV ≤ 0.
• Case gN i > 0 for some i ∈ IC : One or more contacts are open. All we
can say is that dV ≤ 0.
We conclude that
dV = 0 for u = 0,
dV ≤ 0 for gN C = 0, (7.72)
dV < 0 for gN C = 0, u− = 0.

We now apply LaSalle’s invariance principle (Theorem 6.31), which is valid

when every limit set is a positively invariant set as stated in Condition 9 (see
the remark below). Let us hereto consider the set Ωρ∗ where ρ∗ is chosen such
that Ωρ∗ ⊂ (V ∩ Q). Note that Ωρ∗ is a positively invariant set due to the
choice of V . Moreover, the set Z is defined as
 7.3 Attractivity of Equilibrium Sets 171

Z = {(q, u) | dV = 0}, (7.73)

which generally has a nonzero intersection with P = {(q, u) | gN C = 0, gN C ≥

0}.
Consider a solution curve with an arbitrary initial condition in P for t = t0 .
Due to condition 3, which requires that γ̇N i < 0 (a.e.) for ∀i ∈ IC \IN , at least
one impact will occur for some t > t0 . The impact does not necessarily occur
at a contact in IC . In any case, the impact will cause dV < 0 at the impact
time. Therefore, there exists no solution curve with initial condition in P that
remains in the intersection P ∩ Z. Hence, it holds that the intersection P ∩ Z
does not contain any invariant subset. Therefore, we seek the largest invariant
set in T = {(q, u) | gN C (q) = 0, u = 0}. Using the fact that u should be
zero, and that this implies that no impulsive forces can occur in the measure
differential inclusion describing the dynamics of the system, yields:

M (q)du − h(q, 0)dt = WN (q)dPN + WD (q)dPD

⇒ h(q, 0)dt + WN (q)λN dt + WD (q)λD dt = 0
⇒ h(q, 0) + WN (q)λN + WD (q)λD = 0
  (7.74)
⇒ h(q, 0) − WNi (q)∂ΨC∗ N (0) − WDi (q)∂ΨC∗ D (0)  0
i i
i i
 
⇒ h(q, 0) + WNi (q)R −
+
WDi (q)CDi  0.
i i

The equilibrium set E is the largest set of equilibria in Ωρ∗ (Condition 8).
Consequently, we can conclude that the largest invariant set in Z is the equi-
librium set E. Therefore, it can be concluded from Theorem 6.31 that E is an
attractive set. 
Remark: Theorem 6.31, being a generalisation of LaSalle’s invariance princi-
ple, is valid when every limit set is a positively invariant set [33]. A sufficient
condition for the latter is continuity of the solution with respect to the initial
condition (Proposition 6.12). Non-smooth mechanical systems with multiple
simultaneous impacts do generally not possess continuity with respect to the
initial condition. Hence, the positive invariance of limit sets has been explic-
itly assumed in Condition 9 of Theorem 7.6. However, the assumptions of the
above theorem make it likely that all contacts close in finite time (just as
in the bouncing ball system). When all contacts are closed, then the system
behaves after this time-instance like a Filippov system (a differential inclusion
of Filippov-type), for which continuity with respect to the initial condition is
guaranteed. It then would hold that that every limit set is a positively invari-
ant set and the generalisation of LaSalle’s invariance principle can be applied.

In the following propositions, we derive some sufficient conditions for con-

dition 2 and 4 of Theorem 7.6 if the system obeys some additional properties.
172 7 Stability Properties in Mechanical Systems

Proposition 7.7. If f nc = −Cu, then it holds that Dqnc −1 (0) = ker C, i.e.
the zero set of Dqnc (u) is the nullspace of C.

Proof: Substitution gives Dqnc (u) = uT Cu. The proof is immediate. 

The forces λDi (and impulses ΛDi ), which are derived from a support function
on the set CDi , have in the above almost always been associated with friction
forces, but can also be forces from a one-way clutch. Friction and the one-
way clutch are described by the same inclusion on velocity level, but they
are different in the sense that 0 ∈ bdry CDi holds for the one-way clutch and
0 ∈ int CDi holds for friction. The dissipation function of friction is a PDF,
meaning that friction is dissipative when a relative sliding velocity is present,
whereas no dissipation occurs in the one-way clutch. This insight leads to the
following proposition:
−1
Proposition 7.8. If 0 ∈ int CDi ∀i ∈ IG , then it holds that DqλD (0) =
ker WDT (q), i.e. the zero set of DqλD (u) is the nullspace of WDT (q).
Proof: Because of 0 ∈ int CDi ∀i ∈ IG , it follows from (2.38) that ΨC∗ Di (γDi ) >
0 for γDi = 0, i.e. ΨC∗ Di (γDi (q, u)) = 0 ⇔ γDi (q, u) = 0. Moreover, it follows
from assumption (5.91) that γDi (q, u) = 0 ⇔ u ∈ ker WDi T
(q). The proof
follows from the definition (7.27) of Dq (u).λD

If the above propositions are fulfilled then we can simplify condition 2 and 4
of Theorem 7.6.
Corollary 7.9. If f nc = −Cu and 0 ∈ int CDi ∀i ∈ IG , then condition 2 is
equivalent to C > 0 and condition 4 is equivalent to ker C ∩ ker WDT (q) ∩
ker WNT (q) = {0}.

7.3.2 Systems with Frictional Bilateral Constraints

In this section, we focus on systems with bilateral constraints with Coulomb

friction (frictional sliders). The restriction to bilateral constraints excludes
unilateral contact phenomena such as impact and detachment. This kind of
systems is very common in engineering practice; think for example of industrial
robots with play-free joints. We assume that a set of independent generalised
coordinates is known (denoted by q ∈ Rn in this section), for which these
bilateral constraints are eliminated from the formulation of the dynamics of
the system. We formulate the dynamics of the system using a Lagrangian
approach, resulting in
 T
d
(T,u ) − T,q + U,q = f nc + WD (q)λD , (7.75)
dt

or, alternatively,
M (q)u̇ − h(q, u) = WD (q)λD . (7.76)
 7.3 Attractivity of Equilibrium Sets 173

The friction forces are assumed to obey a set-valued force law (5.70). Note that
no unilateral contact forces are present in this formulation. Since impacts are
excluded, there is no need to formulate the dynamics on momentum level, as
no impulsive forces occur. Consequently, the equation (7.75) or (7.76) together
with the set-valued force law (5.70) represent a differential inclusion on force
level. An equilibrium set of (7.76) obeys
 &

E ⊂ (q, u) ∈ R × R | (u = 0) ∧ h(q, 0) −
n n
WDi (q)CDi  0 , (7.77)
i∈IG

where IG is the set of all frictional bilateral contact points (frictional sliders).
An equilibrium set is positively invariant if we assume uniqueness of solutions
in forward time. By qe we denote an equilibrium position of the system without
friction, i.e. h(qe , 0) = 0. We consider the attractivity of an equilibrium set
E ⊃ (qe , 0) in the following theorem. Theorem 7.10 is almost a corollary
of Theorem 7.6, but the smooth non-conservative forces f nc (q, u) are not
necessarily dissipative.
Theorem 7.10 (Attractivity of the equilibrium set). Consider sys-
tem (7.76) with friction law (5.70). If
1. the equilibrium position qe is a local minimum of the potential energy U (q)
and U (q) has a non-vanishing generalised gradient for all q ∈ U\{qe }, i.e.
0∈ / ∂U (q) ∀q ∈ U\{qe }, and the equilibrium set Eq is contained in U, i.e.
Eq ⊂ U,
2. Dqnc (u) = −uT f nc and f nc = 0 for u = 0,
3. there exists a set D ⊂ Rn × Rn , with E ⊂ int D, such that Dqnc (u) +
DqλD (u) > 0 ∀(q, u) ∈ D\N , where N = {(q, u) ∈ Rn × Rn | u = 0},
4. E ⊂ Ωρ∗ in which the set Ωρ∗ is the largest level set of V (q, u) = T (q, u)+
U (q) − U (qe ), that is contained in D and Q = {(q, u) ∈ Rn × Rn | q ∈ U},
i.e.
ρ∗ = max ρ, (7.78)
{ρ:Ωρ ⊂(D∩Q)}

5. E is the largest set of equilibrium points in Ωρ∗ ,

then the equilibrium set E ⊃ (qe , 0) is locally attractive and Ωρ∗ is a conser-
vative estimate for the region of attraction.
Proof: The Lyapunov function V (q, u) = T (q, u) + U (q) − U (qe ) is positive
definite around the equilibrium point (q, u) = (qe , 0) due to condition 1 in
the theorem. The function V is absolutely continuous along solutions curves
of the system because the system is of Filippov-type. The differential measure
dV only consists of a Lebesgue part V̇ dt with

V̇ = −Dqnc (u) − DqλD (u). (7.79)

Conditions 2 and 3 assure that the Lyapunov function does not increase along
solution curves within the set D:
174 7 Stability Properties in Mechanical Systems

V̇ = 0 u = 0
(7.80)
V̇ < 0 (q, u) ∈ D\N

The proof now continues in the same way as the proof of Theorem 7.6. 
The above theorem shows that the equilibrium set E of a mechanical sys-
tem with dry friction can be (locally) attractive even when the equilibrium
position qe of the mechanical system without dry friction is unstable due to
negative damping, i.e. the smooth non-conservative forces are non-dissipative
in certain generalised force directions. For instance, fluid, aeroelastic, control
and gyroscopic forces can lead to a non-positive definite damping matrix of
the linearised system and can therefore pump energy in the system and desta-
bilise an equilibrium through a Hopf bifurcation. The fact that the presence
of dry friction can have a ‘stabilising’ effect can be explained by pointing out
that the dry friction forces are of zero-th order (in terms of generalised veloc-
ities) whereas the ‘destabilising’ damping forces are of first order or higher.
Consequently, the ‘stabilising’ effect of the dry friction forces can locally dom-
inate the ‘destabilising’ smooth non-conservative forces leading to the local
attractivity of the equilibrium set. In [170], these facts have been proved rig-
orously for linear mechanical systems with bilateral frictional constraints. The
requirement Dqnc (u) + DqλD (u) > 0 of condition 3 simplifies for a linear me-
chanical system with a symmetric damping matrix C ≯ 0 and a constant
matrix WD to
Uci ∈ span WD , i = 1 . . . nq (7.81)
where Uc = {Uci } is a matrix containing the nq eigencolumns of the eigenval-
ues of C which lie in the left-half complex plane [170]. This condition can be
interpreted as follows: the space spanned by the eigendirections of the damp-
ing matrix, related to negative eigenvalues, lies in the space spanned by the
generalised force directions of the friction force.
Resuming, we can conclude that, in this subsection, we have formulated
sufficient conditions for the (local) attractivity of equilibrium sets of a rather
wide class of nonlinear mechanical systems with bilateral frictional sliders.
The nonlinearities may involve: nonlinearities in the mass-matrix, both non-
linear conservative forces and non-conservative forces (possibly even non-
dissipative). Moreover, the generalised force directions of the dry friction
forces may depend on the generalised coordinates and the normal forces in
the friction sliders may depend on both the generalised coordinates and the
generalised velocities.

7.4 Instability of Equilibrium Positions and Sets

Instability of equilibrium points and sets of measure differential inclusions has

been discussed in Section 6.7. When speaking of mechanical systems, at least
two kinds of instability of equilibrium positions and sets can be distinguished.
 7.4 Instability of Equilibrium Positions and Sets 175

The first kind is due to the fact that the total potential energy does not
have a local minimum at the equilibrium point (or at an equilibrium point in
an equilibrium set). Such an equilibrium point is a saddle-point in the phase-
space when the system is described by a smooth differential equation. For
a linear mechanical system this would mean that the stiffness matrix is not
positive semi-definite. Examples of such equilibrium points can be found in
buckling phenomena caused by static bifurcations (pitchfork, transcritical or
saddle-node bifurcations) of an equilibrium point.
The second kind of instability is caused by non-dissipating non-conservative
forces. Roughly speaking, we could say that the system has ‘negative damp-
ing’, i.e. some non-conservative forces are pumping energy into the system. A
linear mechanical system would in this case have a damping matrix which is
not positive semi-definite. This phenomenon occurs in nonlinear smooth dyna-
mical systems when a Hopf bifurcation occurs, i.e. a stable focus transforms
into an unstable focus and periodic solutions are created.
If we want to prove the instability of an equilibrium set of a mechanical
system with frictional unilateral contacts using Theorem 6.33, then the choice
of the V has to be based on the kind of instability in hand. We first present a
theorem that can be used to prove instability of the first kind, i.e. when the
total potential energy is not a positive definite function.

Theorem 7.11 (Instability Type I of an Isolated Equilibrium at the

Origin). Let q ∗ = 0 ∈ K be an isolated equilibrium position of (5.96), (5.98)
and let the total potential energy Q(q) be such that Q(qa ) < 0 for some qa with
arbitrary small qa . We assume the system to be energetically consistent. If
it holds that Dqnc (u) + DqλD (u) > 0 ∀q ∈ K, u = 0, then the equilibrium
position is unstable.

Proof: It holds that qa ∈ K because Q(qa ) < 0 and Q(q) = +∞ for all q ∈ / K.
Consider the function V to be the total energy function

V (q, u) = Etot (q, u) = T (q, u) + U (q) − U (0) + πN i (gN i (q)),
i∈IG

which has the properties

V (0, 0) = 0, V (qa , 0) < 0

for some qa with arbitrary small qa . The set

U = {(q, u) ∈ B r | V (q, u) ≤ 0}

is a nonempty set because (qa , 0) ∈ U and it holds that U ⊂ A = {(q, u) ∈

Rn × Rn | q ∈ K} because V (q, u) = ∞ for q ∈ / K. Because q ∗ = 0 is an
isolated equilibrium position, we can take r such that q ∗ = 0 is the unique
equilibrium position in U. The differential measure dV consists of a Lebesgue
part and an atomic part:
176 7 Stability Properties in Mechanical Systems

dV = V̇ dt + (V + − V − )dη.

Using Dqnc (u) + DqλD (u) ≥ 0, (7.29) and (7.30) it follows that V̇ ≤ 0 ∀(q, u) ∈
U ⊂ K × Rn . Moreover, it holds that V + − V − ≤ 0 because of the energetic
consistency of the system. The function V does therefore not increase along
solution curves of the system in U. The set

Z = {(q, u) ∈ U\{0} | dV = 0}

can be written as Z = {(q, u) ∈ U\{0} | u = 0}. An invariant set in Z

necessarily consists of equilibrium points. However, the origin is the unique
equilibrium point in U and Z ⊂ U\{0} does therefore not contain an equili-
brium point. Consequently, Z is not positively invariant. Instability of q ∗ = 0
follows from Theorem 6.32b. 
Theorem 7.11 can be applied when we can indeed find an arbitrary small qa for
which Q(qa ) < 0. This means that the total potential energy does not have
a local minimum at the origin and corresponds therefore to an instability
of the first kind. We now generalise this theorem to prove instability of an
equilibrium set.

Theorem 7.12 (Instability Type I of an Equilibrium Set). Consider

an equilibrium set E of (5.96), (5.98) which is assumed to be energetically
∗
consistent. Let qm ∈ bdry Eq be an equilibrium position for which the total
potential energy Q(q) is minimal on the boundary, i.e.
∗
Q(qm ) ≤ Q(q ∗ ) ∀q ∗ ∈ bdry Eq . (7.82)
∗
Assume that Q(qa ) < Q(qm ) for some qa ∈/ Eq with arbitrarily small qa −
∗
qm . Moreover, if Dq (u) + DqλD (u) > 0 ∀q ∈ K, u = 0, then the equilibrium
nc

set E is unstable.
∗
Proof: It holds that qa ∈ K because Q(qa ) < Q(qm ) < ∞ and Q(q) = +∞
for all q ∈
/ K. Consider the function V to be the total energy function

V (q, u) = Etot (q, u) = T (q, u) + U (q) − U (0) + πN i (gN i (q)),
i∈IG

which has the properties

∗
V (0, 0) = 0, V (qa , 0) < V (qm , 0) =⇒ V (qa , 0) < V (q ∗ , 0) ∀q ∗ ∈ bdry E
∗
for some qa ∈
/ Eq with arbitrary small qa − qm . The set
∗ ∗
U = {(q, u) ∈ (qm , 0) + B r | V (q, u) ≤ V (qm , 0)}

is a nonempty set because (qa , 0) ∈ U and it holds that U ⊂ A = {(q, u) ∈

Rn × Rn | q ∈ K} because V (q, u) = ∞ for q ∈ / K. We can take r such
that U does not contain equilibrium points other than those in E, being the
 7.4 Instability of Equilibrium Positions and Sets 177

equilibrium set under investigation. The differential measure dV consists of a

Lebesgue part and an atomic part

dV = V̇ dt + (V + − V − )dη.

Using Dqnc (u) + DqλD (u) ≥ 0, (7.29) and (7.30) it follows that V̇ ≤ 0 ∀(q, u) ∈
U ⊂ K × Rn . Moreover, it holds that V + − V − ≤ 0 because of the energetic
consistency requirement. The function V does therefore not increase along
solution curves of the system in U. This means that the forward solution
ϕ(t, t0 , qa ) can not cross the boundary of E and therefore not enter E. The
set
Z = {(q, u) ∈ U\E | dV = 0}
can be written as Z = {(q, u) ∈ U\E | u = 0}. An invariant set in Z
necessarily consists of equilibrium points, but the value of r is chosen such
that the set U\E does not contain an equilibrium point. The set Z ⊂ U\E is
∗
therefore not positively invariant. Instability of the equilibrium point (qm , 0)
follows from Theorem 6.32b. Consequently, the equilibrium set E is unstable.

We now focus on instability of the second type, i.e. Hopf bifucation in-
stability caused by energy creating non-conservative forces. The dissipation
Dqnc (u) due to the non-conservative forces is therefore by definition negative
for non-zero generalised velocities. As Lyapunov function we choose again the
total mechanical energy. In order to prove instability using Theorem 6.32,
we have to pose conditions which force the total dissipation rate to be non-
positive. The ‘atomic’ dissipation change due to impacts can only be non-
positive if the contacts are assumed to be completely elastic (all restitution
coefficients equal unity) or if impacts are absent in the system. The dissipation
rate during impact-free motion is governed by the non-conservative forces as
well as the friction forces. This implies that we have to state the condition
Dqnc (u) + DqλD (u) ≤ 0. However, the friction forces have a set-valued nature
and dominate over the smooth non-conservative forces for small generalised
velocities. Hence, we have to exclude friction in order to prove instability of
the second type using Theorem 6.32.
Theorem 7.13 (Instability Type II of an Isolated Equilibrium at the
Origin). Let q ∗ = 0 ∈ K be an isolated equilibrium position of (5.96), (5.98)
and let the total potential energy Q(q) (7.4) be locally positive definite. Let all
contacts be frictionless and be completely elastic or bilateral. If it holds that
Dqnc (u) := −uT f nc (q, u) < 0 ∀q ∈ K, u = 0, then the equilibrium position is
unstable.
Proof: Consider the function V to be the total mechanical energy function

V (q, u) = Etot (q, u) = T (q, u) + U (q) − U (0) + πN i (gN i (q)),
i∈IG
178 7 Stability Properties in Mechanical Systems

which is a locally positive definite function. Because q ∗ = 0 is an isolated

equilibrium position, we find an r > 0 such that q ∗ = 0 is the unique equili-
brium position in U = B r ∩ A. The function V is bounded on U because U is
compact and a subset of A. The differential measure dV consists of a Lebesgue
part and an atomic part: dV = V̇ dt + (V + − V − )dη. Using Dqnc (u) < 0 for
u = 0, DqλD (u) = 0 and (7.29), (7.30) it follows that V̇ > 0 when u = 0
and V̇ = 0 for u = 0. Moreover, it holds that V + − V − = 0 because all
contacts are assumed to be completely elastic or bilateral, which implies that
no dissipation can occur due to impacts. Hence, it holds that dV = 0 for
(q, u) ∈ U0 = {(q, u) ∈ U | u = 0} and dV = V̇ dt > 0 for (q, u) ∈ U \U0 .
The set U0 \{0} does not contain any positively invariant subset because the
origin is the unique equilibrium in U. The equilibrium is not attractive be-
cause V is locally positive definite and can not decrease in U. We will prove
that U\{0} does not contain any positively invariant subset by a reductio ad
absurdum. Imagine that U\{0} has the positively invariant subset M. Con-
sider the solution curve (q(t), u(t)) starting from (q0 , u0 ) ∈ M. It holds that
(q(t), u(t)) can only stay for a Lebesgue negligible set of time-instances on
U0 \{0}. As a consequence, the mean value of V̇ (t) will be strictly positive for
any non-empty time-interval. Consequently, if the solution remains within M
for all future times, then V will grow unboundedly. This is in contradiction
with the boundedness of V on U. Instability of (q ∗ , 0) = (0, 0) follows from
M = ∅, i.e. there does not exist a δ > 0 such that all solutions starting in the
δ-neighbourhood of the origin remain in U. 
The exclusion of friction and inelastic impact makes the above theorem rather
weak. An instability theorem for equilibrium sets for this kind of instability
can be pursued in the same way, but the limited use of the theorem remains.
More elaborate instability theorems have to be considered when dealing with
instability of the second kind.

7.5 Examples
In this section we show how the results of the previous sections can be used
to prove the stability, attractivity or instability of an equilibrium set of a
number of mechanical systems. Sections 7.5.1 and 7.5.2 involve examples of
mechanical systems with unilateral contact, impact and friction and are of
increasing complexity. Section 7.5.3 treats an example of a mechanical system
with bilateral frictional constraints.

7.5.1 Falling Block

Consider a planar rigid block (see Figure 7.2) with mass m under the action of
gravity (gravitational acceleration g), which is attached to a vertical wall with
a spring. The block can freely move in the vertical direction but is not able
 7.5 Examples 179

Fig. 7.2. Falling block.

to undergo a rotation. The coordinates x and y describe the position of the

block. The spring, with stiffness k, is unstressed for x = 0. The block comes
into contact with a horizontal floor when the contact distance gN = y becomes
zero. The constitutive properties of the contact are the friction coefficient μ
and the restitution coefficients 0 ≤ eN < 1 and eT = 0. The equations of
motion for impact free motion read as

mẍ + kx = λT ,
(7.83)
mÿ = −mg + λN .
 T
Using the generalised coordinates q = x y , we can describe the system in
the form (5.96) with
       
m 0 −kx 0 1
M= , h= , WN = , WT = . (7.84)
0 m −mg 1 0

The system for μ = 0 admits a unique equilibrium position qe = 0. For μ > 0

there exists an equilibrium set E = {(x, y, ẋ, ẏ) | k|x| ≤ μmg, y = 0, ẋ = ẏ =
0} and it holds that (qe , 0) ∈ E.
The total potential energy function consists of the potential energy of the
spring, the gravitation and the support function of the unilateral contact

Q(q) = U (q) + ΨC∗ N (gN (q))

1
= kx2 + mgy + ΨR∗− (y) (7.85)
2
1
= kx2 + mgy + ΨR+ (y).
2
Notice that the term mgy + ΨR+ (y) is a positive definite term in y. It holds
that Q is a positive definite function in q, because it is bounded from below by
another positive definite function Q(q) ≥ 12 kx2 + mg|y|. Moreover, the global
minimum of Q is located at the equilibrium point qe = 0, because ∂Q(qe )  0
and is unique because of the strict convexity of Q. In order to study the altered
potential energy of an equilibrium point (q ∗ , 0) = ((x∗ , 0), 0) ∈ E, we use the
altered potential energy function (7.43)
180 7 Stability Properties in Mechanical Systems
Qa (q) = U (q − q ∗ ) + ΨC∗ N (gN (q))
1 (7.86)
= k(x − x∗ )2 + mgy + ΨR+ (y),
2
which is positive definite with respect to the equilibrium position q ∗ . The
system does not contain smooth non-conservative forces, i.e. f nc = 0. Using
the fact that the dissipation due to friction (7.47), DqλT (u) + γTT λ∗T , is always
non-negative, it follows from Theorem 7.5 that each equilibrium position q ∗ ∈
Eq is stable. Henceforth, we are able to prove that the equilibrium set E is
stable.
Attractivity of the equilibrium set E can be proven with Theorem 7.6. The
total potential energy function Q(q) (7.85) fulfills condition 1 of Theorem 7.6
for all q ∈ R2 . The system does not contain smooth non-conservative forces,
i.e. f nc = 0, which fulfills condition 2 of Theorem 7.6. Denote the contact
between block and floor as contact 1 and take IC = IG = {1}. It holds that
γ̇N = −g for gN = y > 0, which guarantees that if the block is not in contact
with the floor, then it will come again in contact with the floor in a finite
time. Condition 3 of Theorem 7.6 is therefore satisfied. Furthermore, it holds
−1
that Dqnc −1 (0) = Rn and DqλT C (0) = ker WTT . Because the vectors WN and
WT are linearly independent it holds that ker WTT ∩ ker WNT = {0} and con-
dition 4 of Theorem 7.6 is therefore fulfilled. The restitution coefficients fulfill
condition 5. Condition 6 is fulfilled because V = U = A = {(q, u) | y ≥ 0}.
Condition 7 is fulfilled because the system has only one contact point and
eN < 1. Condition 8 is fulfilled by definition. The only limit set is the equi-
librium set E, which is positively invariant by definition. Hence, condition 9
is fulfilled. Consequently, Theorem 7.6 proves that the equilibrium set E is
attractive. Moreover, the Lyapunov function V (q, u) = Etot (q, u) − Q(qe ) is
radially unbounded and the equilibrium set is therefore globally attractively
stable. The equilibrium set is clearly also symptotically attractively stable,
but this can not be proven with the presented theorems.

7.5.2 Rocking Bar

Consider a planar rigid bar with mass m and inertia JS with respect to the
centre of mass S, which is attached to a vertical wall with a spring (Figure 7.3).
The gravitational acceleration is denoted by g. The position and orientation
of the bar are described by the generalised coordinates
 T
q= xyϕ , (7.87)

where x and y are the displacements of the centre of mass S with respect
to the coordinate frame (eIx , eIy ) and ϕ is the inclination angle. The spring
is unstressed for x = 0. The bar has length 2a. The two endpoints 1 and 2
can come into contact with the floor. The contact between bar and floor is
described by a friction coefficient μ > 0 and a normal restitution coefficient
 7.5 Examples 181

Fig. 7.3. Rocking bar.

0 ≤ eN < 1 which is equal to the tangential restitution eT = eN . The contact

distances, indicated in Figure 7.3, are

gN 1 = y − a sin ϕ,
(7.88)
gN 2 = y + a sin ϕ.

The relative velocities of contact points 1 and 2 with respect to the floor read
as
γT 1 = ẋ + aϕ̇ sin ϕ,
(7.89)
γT 2 = ẋ − aϕ̇ sin ϕ.
We can describe the system in the form (5.96) with
⎡ ⎤ ⎡ ⎤
m 0 0 −kx
M = ⎣ 0 m 0 ⎦ , h = ⎣−mg ⎦ , (7.90)
0 0 JS 0
   
T 0 1 −a cos ϕ T 1 0 a sin ϕ
WN = , WT = . (7.91)
0 1 a cos ϕ 1 0 −a sin ϕ
The system contains a number of equilibrium sets. We will consider the equi-
librium set

E = {(x, y, ϕ, ẋ, ẏ, ϕ̇) | k|x| ≤ μmg, y = 0, ϕ = 0, ẋ = ẏ = ϕ̇ = 0}, (7.92)

for which gN 1 = gN 2 = 0. The total potential energy function

Q(q) = U (q) + ΨC∗ N (gN 1 (q)) + ΨC∗ N (gN 2 (q))

1
= kx2 + mgy + ΨR∗− (gN 1 ) + ΨR∗− (gN 2 ) (7.93)
2
1
= kx2 + mgy + ΨR+ (gN 1 ) + ΨR+ (gN 2 )
2
182 7 Stability Properties in Mechanical Systems

contains a quadratic term in x, a linear term in y and two indicator functions

on the contact distances. Notice that Q(q) = 0 for q = qe = 0, where (qe , 0) 
E is the equilibrium point for μ = 0. Moreover, it holds that if gN 1 ≥ 0 and
gN 2 ≥ 0 then y ≥ 0 and a| sin ϕ| ≤ y. We therefore deduce that
1 2
gN 1 ≥ 0 ∧ gN 2 ≥ 0 =⇒ Q(q) = kx + mgy
2
1 mg
Q(q) = kx2 + (|y| + y) (7.94)
2 2
1 mg
Q(q) ≥ kx2 + (|y| + a| sin ϕ|)
2 2
and
gN 1 < 0 ∨ gN 2 < 0 =⇒ Q(q) = +∞
1 mg (7.95)
Q(q) > kx2 + (|y| + a| sin ϕ|).
2 2
The function f (q) = 12 kx2 + mg2 (|y| + a| sin ϕ|) is locally positive definite in
the set U = {q ∈ Rn | |ϕ| < π2 }. Consequently, the total potential energy
function Q(q) ≥ f (q) is locally positive definite in the set U as well. It can be
easily checked that the generalised gradient
⎡ ⎤
kx
∂Q(q) = ⎣ mg + ∂ΨR+ (gN 1 ) + ∂ΨR+ (gN 2 ) ⎦ (7.96)
−∂ΨR+ (gN 1 )a cos ϕ + ∂ΨR+ (gN 2 )a cos ϕ

can only vanish in the set U for q = qe , i.e. 0 ∈

/ ∂Q(q) ∀q ∈ U\{qe } and
0 ∈ ∂Q(qe ). Similarly, we can prove that the altered total potential energy
function
Qa (q) = Q(q − q ∗ ) = U (q − q ∗ ) + ΨC∗ N (−gN 1 (q)) + ΨC∗ N (−gN 2 (q))
1 (7.97)
= k(x − x∗ )2 + mgy + ΨR+ (gN 1 ) + ΨR+ (gN 2 ),
2
with q ∗ ∈ Eq , is locally positive definite with respect to q ∗ .
Smooth non-conservative forces are absent in this system, i.e. f nc = 0
and Dqnc (u) = 0. Using the fact that the dissipation due to friction (7.47),
DqλT (u) + γTT λ∗T , is always non-negative, it follows from Theorem 7.5 that
each equilibrium position q ∗ ∈ Eq is stable. Henceforth, we are able to prove
that the equilibrium set E is stable.
Proving that the equilibrium set E is also attractive, is much more com-
plicated. We now have to prove that condition 3 of Theorem 7.6 holds with
IC = {1, 2}. Consider the open subset V = {(q, u) ∈ Rn × Rn | μ| tan ϕ| <
1, aϕ̇2 < g} which contains the equilibrium set, i.e. E ⊂ V. We consider the
following cases with (q, u) ∈ V:
 7.5 Examples 183

• IN = ∅: both contacts are open, i.e. gN 1 > 0 and gN 2 > 0. It holds for
(q, u) ∈ V that
γ̇N 1 = ÿ − aϕ̈ cos ϕ + aϕ̇2 sin ϕ
= −g + aϕ̇2 sin ϕ
<0
(7.98)
γ̇N 2 = ÿ + aϕ̈ cos ϕ − aϕ̇2 sin ϕ
= −g − aϕ̇2 sin ϕ
< 0.
• IN = {1}: contact 1 is closed and contact 2 is open, i.e. gN 1 = 0 and
gN 2 > 0. We consider contact 1 to be closed for a nonzero time-interval.
The normal contact acceleration of the closed contact 1 must vanish:
γ̇N 1 = ÿ − aϕ̈ cos ϕ + aϕ̇2 sin ϕ
1 a2 a2
0 = −g + λN 1 + cos2 ϕ λN 1 − cos ϕ sin ϕ λT 1 + aϕ̇2 sin ϕ
m JS JS
 
1 a2
0 = −g + + cos ϕ(cos ϕ − μ̄ sin ϕ) λN 1 + aϕ̇2 sin ϕ,
m JS
(7.99)
with λT 1 = μ̄λN 1 , i.e. μ̄ ∈ −μ Sign(γT 1 ). It follows from (7.99) that the
normal contact force λN 1 is a function of ϕ and ϕ̇. The contact acceleration
of contact 2 therefore becomes
γ̇N 2 = ÿ + aϕ̈ cos ϕ − aϕ̇2 sin ϕ
1 a2 a2
= −g + λN 1 − cos2 ϕ λN 1 + cos ϕ sin ϕ λT 1 − aϕ̇2 sin ϕ
m JS JS
 
1 a2
= −g + − cos ϕ(cos ϕ − μ̄ sin ϕ) λN 1 − aϕ̇2 sin ϕ
m JS
a2
1
m −
JS cos ϕ(cos ϕ − μ̄ sin ϕ)
= (g − aϕ̇2 sin ϕ) − g − aϕ̇2 sin ϕ
+ JaS cos ϕ(cos ϕ − μ̄ sin ϕ)
1 2
m
−2ga2 JmS cos ϕ(cos ϕ − μ̄ sin ϕ) − 2aϕ̇2 sin ϕ
= .
1 + a2 JmS cos ϕ(cos ϕ − μ̄ sin ϕ)
(7.100)
Using |μ̄| ≤ μ and (q, u) ∈ V it follows that γ̇N 2 < 0.
• IN = {2}: contact 1 is open and contact 2 is closed, i.e. gN 1 > 0 and
gN 2 = 0. Similar to the previous case we can prove that γ̇N 1 < 0.
Hence, there exists a non-empty set IC = {1, 2}, such that γ̇N i (q, u) < 0
(a.e.) for ∀i ∈ IC \IN and ∀(q, u) ∈ V. Condition 3 of Theorem 7.6 is therefore
fulfilled.
−1
It holds that Dqnc = 0 and using Proposition 7.8 it follows that DqλT (0) =
ker WTT (q). Furthermore, for q ∈ C = {q ∈ Rn | gN 1 = gN 2 = 0} follows the
184 7 Stability Properties in Mechanical Systems

implication WNT (q)u = 0 =⇒ ẏ = 0 ∧ ϕ̇ = 0 and similarly WTT (q)u = 0 =⇒

ẋ = 0. We conclude that there is always dissipation when both contacts are
closed and u = 0 because

ker WTT (q) ∩ ker WNT (q) = {0} ∀q ∈ C, (7.101)

and condition 4 of Theorem 7.6 is therefore fulfilled. The largest level set of
V = T (q, u) + Q(q) which lies entirely in Q = {(q, u) ∈ Rn × Rn | q ∈ U} is
given by V (q, u) < mga. The largest level set of V which lies entirely in V is
determined by V (q, u) < 12 JS ag and V (q, u) < √mga 2 . We therefore choose
1+μ
the set Ωρ∗ as
 
∗ ∗ 1 g mga
Ωρ∗ = {(q, u) ∈ R ×R | V (q, u) < ρ }, with ρ = min
n n
JS , , .
2 a 1 + μ2
(7.102)
If additionally
1 (μmg)2
< ρ∗ , (7.103)
2 k
then it holds that E ⊂ Ωρ∗ , which is condition 6 of Theorem 7.6. Conditions 5
and 7 are fulfilled because all normal restitution coefficients are equal and
strictly smaller than one. Condition 8 is fulfilled because of (7.102). The only
limit set in Ωρ∗ is the equilibrium set E because eN < 1. The equilibrium set
is positively invariant and condition 9 is therefore fulfilled. We conclude that
Theorem 7.6 proves conditionally the local attractivity of the equilibrium set E
and that Ωρ∗ is a conservative estimate of the region of attraction. Naturally,
the attractivity is only local, because the system has also other attractive
equilibrium sets for ϕ = nπ with n ∈ Z and unstable equilibrium sets around
ϕ = π2 + nπ. As in the previous example of the falling block, the equilibrium
set is symptotically attractive.
In a similar way we can study the stability properties of equilibrium sets of
a rocking block (Figure 7.4). Although being a slight modification of the rock-
ing bar, the verification of condition 3 of Theorem 7.6 becomes very elaborate
for the rocking block example [103].

7.5.3 Constrained Bar

We now study an example with bilateral constraints. Consider a bar with

mass m, length 2l and moment of inertia JS around its centre of mass S, see
Figure 7.5. The gravitational acceleration is denoted by g. The bar is subject
to two holonomic constraints: Point 1 of the bar is constrained to the vertical
slider and Point 2 of the bar is constrained to the horizontal slider. Coulomb
friction is present in the contact between these endpoints of the bar and the
grooves (friction coefficient μ1 in the vertical slider and friction coefficient μ2
in the horizontal slider). It should be noted that the realised friction forces
 7.5 Examples 185

Fig. 7.4. Rocking block.

Fig. 7.5. Constrained bar.

depend on the constraint forces in the grooves (i.e. the friction is described by
the non-associated Coulomb’s law (5.49)). The dynamics of the system will
be described in terms of the (independent) coordinate θ, see Figure 7.5. The
corresponding equation of motion is given by
 2 
ml + JS θ̈ + mgl sin θ = 2l sin θλT1 − 2l cos θλT2 , (7.104)
where λT1 and λT2 are the friction forces in the vertical and horizontal sliders,
respectively. Equation (7.104) can be written in the form (7.76), with
186 7 Stability Properties in Mechanical Systems

Fig. 7.6. Attainable friction forces in equilibrium.

 
M (q) = ml2 + JS , h(q, u) = −mgl sin θ, WT (q) = 2l sin θ −2l cos θ .
(7.105)
The equilibrium set of (7.104) is given by (7.77), with CTi = {−λTi |
−μi |λNi | ≤ λTi ≤ +μi |λNi |}, i = 1, 2. Note that CTi depends on the nor-
mal force λNi , which in turn may depend on the friction forces. The static
equilibrium equations of the bar yield:

λN1 + λT2 = 0,
λN2 + λT1 − mg = 0, (7.106)
l cos θλN1 − l sin θλN2 + l sin θλT1 − l cos θλT2 = 0.

Based on the first two equations in (7.106) and the non-associated Coulomb’s
law (5.49), the following algebraic inclusions for the friction forces in equili-
brium can be derived:
λT1 ∈ [−μ1 |λT2 |, μ1 |λT2 |] ,
(7.107)
λT2 ∈ [−μ2 |λT1 − mg|, μ2 |λT1 − mg|] .

The resulting set of friction forces in equilibrium is depicted schematically

in Figure 7.6. The equilibrium set E in terms of the independent generalised
coordinate θ now follows from the equation of motion (7.104):

mgl sin θ = 2l sin θλT1 − 2l cos θλT2 . (7.108)

 7.5 Examples 187

Fig. 7.7. Phase plane and the set in which V = T + U < c∗ .

For values of θ such that cos θ = 0 (we assume that, for given values for m, g
and l, the friction coefficients μ1 and μ2 are small enough to guarantee that
this assumption is satisfied) we obtain:
 
λT2
θ = arctan + (k − 1)π, k = 1, 2, (7.109)
− mg
2 + λT1

for values of λT1 and λT2 taken from (7.107). Equation (7.109) describes the
fact that there exist two isolated equilibrium sets (an equilibrium set E1 around
θ = 0 and E2 around θ = π) for small values of the friction coefficients. The
equilibrium sets are given by
 %  
% 2μ2
%
Ek = (θ, θ̇) % θ̇ = 0, |θ − (k − 1)π| ≤ arctan , (7.110)
1 − μ1 μ2
for k = 1, 2 and μ1 μ2 < 1. Note that for μ1 μ2 ≥ 1 these isolated equilibrium
sets merge into one large equilibrium set, such that any value of θ can be
attained in this equilibrium set. We will consider the case of two isolated
equilibrium sets here.
The potential energy of the constrained bar system

U (q) = −mgl(cos(θ) − 1) (7.111)

188 7 Stability Properties in Mechanical Systems

is locally positive definite. The altered potential energy function can be written
as
Ua (q) = U (q) − (q − q ∗ )T WT (q ∗ )λ∗T − U (q ∗ )
= −mgl(cos θ + (θ − θ∗ ) sin θ∗ − cos θ∗ ) (7.112)
∗ 2
≈ mgl(θ − θ )
which is locally positive definite with respect to q ∗ = θ∗ ∈ Eq .
We first study the stability properties of the equilibrium set E1 around θ =
0. The system does not contain smooth non-conservative forces, i.e. f nc = 0.
Using the fact that the dissipation due to friction (7.47), DqλT (u)+γTT λ∗T , is al-
ways non-negative, it follows from Theorem 7.5 that each equilibrium position
q ∗ ∈ E1q is stable. Henceforth, we are able to prove that the equilibrium set E1
is stable. Attractivity of E1 can be studied with Theorem 7.10 and we check
the conditions stated therein. Condition 1 of this theorem is clearly satisfied.
Namely, take the set U = {θ | |θ| < π} and realise that indeed the poten-
tial energy U (7.111) is locally positive definite in U and ∂U/∂θ = mgl sin θ
satisfies the demand ∂U/∂θ = 0, ∀θ ∈ U\{0}. Since there are no smooth non-
conservative forces Dqnc (u) = 0, condition 2 is satisfied. Finally, we note that
DqλT (u) > 0 for (q, u) ∈ D = Rn × Rn , which implies that condition 3 is sat-
isfied. The set U contains the equilibrium set E1 and part of the equilibrium
set E2 (see Figure 7.8). We now consider the largest level set V < c∗ for which
the set E1 is the only equilibrium set within the level set of V = T + U . This
level set is an open set and the value
 
1 − μ μ
c∗ = mgl 1 − , 2
1 2
(7.113)
4μ2 + (1 − μ1 μ2 )2

is chosen such that its closure touches the equilibrium set E2 . Consequently, we
can conclude that the equilibrium set E1 is locally attractive. The phase plane
of the constrained bar system is depicted in Figure 7.7 for the parameter values
m = 1 kg, JS = 13 kg m2 , l = 1 m, εN = εT = 0, μ1 = μ2 = 0.3, g = 10 m/s2 .
The trajectories in Figure 7.7 have been obtained numerically using the time-
stepping method (see [101] and references therein). The equilibrium sets E1
and E2 are indicated by thick lines on the axis θ̇ = 0. It can be seen in
Figure 7.7 that the level set V < c∗ is a fairly good (though conservative)
estimate for the region of attraction of the equilibrium set E1 . Moreover, we
see that the equilibrium set E1 is symptotically attractive.
Subsequently, we study the stability properties of the equilibrium set E2
around θ = π, under the condition that E1 and E2 are two distinct equilibrium
sets. We apply Theorem 7.12 and check the conditions stated therein. Consider
the most left boundary point θm ∈ bdry E q2 for which
U (θ) becomes minimal
on the boundary, i.e. θm = π − arctan 1−μ 2μ2
1 μ2
. We can find a position
θa < θm , arbitrarily close to θm , such that U (θa ) < U (θm ), due to the fact that
the potential energy is locally negative definite with respect to the position
 7.5 Examples 189

Fig. 7.8. Schematic representation of the set U in which V ≥ 0 (V as in (7.114)).

θ = π. The position θa is not an equilibrium position, because the equilibrium

sets E1 and E2 are distinct. Using the fact that the dissipation in the system
is nonnegative it follows from Theorem 7.12, that the equilibrium set E2 is
unstable.
Theorem 7.12 follows from Theorem 6.33. If we would like to apply The-
orem 6.33 directly, then the function V has to be chosen as follows:
1 
V =− JS + ml2 θ̇2 + mgl (1 + cos θ) , (7.114)
2
where V ≥ 0 ∈ U with the set U depicted schematically in Figure 7.8. The
time-derivative of V obeys

V̇ = −θ̇WT (q)λT = −γTT λT ,

 
with the generalised force direction WT (q) = 2l sin θ −2l cos θ , friction
 
forces λT T = λT1 λT2 and sliding relative velocities γT = WTT (q)θ̇ =
 T
2lθ̇ sin θ −2lθ̇ cos θ in the two frictional sliders. Note that V̇ ≥ 0 for all
(θ, θ̇) ∈ U and V̇ = 0 if and only if θ̇ = 0. We can easily show that solutions
can not stay in Z\E2 , with Z = {(θ, θ̇) | θ̇ = 0}, using the equation of mo-
tion (7.104). The conditions of statement 3 of Theorem 6.33 are satisfied and
it can be concluded that the equilibrium set E2 is unstable.
The equilibrium set E2 becomes a saddle point for μ1,2 = 0. This saddle
structure in the phase plane (see Figure 7.7) remains for μ1,2 > 0, but E2 is
a set instead of a point. Interestingly, the stable manifold of E2 is ‘thick’, i.e.
there exists a bundle of solutions (depicted in dark grey) which are attracted
to the unstable equilibrium set E2 . Put differently: the equilibrium set E2 has
a region of attraction, where the region is a set with a non-empty interior.
The unstable half-manifolds of E2 originate at the tips of the set E2 and are
heteroclinic orbits to the stable equilibrium set E1 .
190 7 Stability Properties in Mechanical Systems

7.6 Summary
The stability theorems of Chapter 6 for measure differential inclusions have
been applied in this chapter to Lagrangian mechanical systems with frictional
unilateral and bilateral constraints. The total mechanical energy has been de-
fined to be the mechanical energy of the system together with the potential
of the normal contact forces of the unilateral constraints. This potential is an
indicator function on the admissible domain of the generalised coordinates,
which are minimal coordinates with respect to the bilateral constraints. The
total mechanical energy is a natural choice for a Lyapunov (candidate) func-
tion and has been used to formulate Lyapunov-type theorems for stability,
attractivity and instability of equilibrium points and sets. The stability and
attractivity theorems rely on the energetic consistency, i.e. dissipativity of im-
pacts. Sufficient conditions for energetic consistency have been given in terms
of conditions on the system parameters. A number of examples have been
presented which show the use of the aforementioned theorems, but also show
their limitations. Notably, condition 3 of Theorem 7.6 might be extremely
hard to prove and limits the use of this theorem considerably.
8
Convergence Properties of Monotone Measure
Differential Inclusions

In this chapter, we present theorems which give sufficient conditions for the
convergence of measure differential inclusions with certain maximal mono-
tonicity properties. The framework of measure differential inclusions allows
us to describe systems with state discontinuities, as has been shown in the
previous chapters. The material presented in this chapter is based on the
paper [104].
The chapter is organised as follows. First, we define the convergence prop-
erty of dynamical systems in Section 8.1 and state the associated properties
of convergent systems. Theorems are presented in Section 8.2 which give suf-
ficient conditions for the convergence of measure differential inclusions with
certain maximal monotonicity properties. Furthermore, we illustrate in Sec-
tion 8.3 how these convergence results for measure differential inclusions can
be exploited to solve tracking problems for certain classes of non-smooth me-
chanical systems with friction and one-way clutches. Illustrative examples of
convergent mechanical systems are discussed in detail in Section 8.4. Finally,
Section 8.5 presents concluding remarks.

8.1 Convergent Systems

In this section, we will briefly discuss the definition of convergence and certain
properties of convergent systems. In the definition of convergence, the Lya-
punov stability of solutions of (8.1) plays a central role. Definitions of (uni-
form) stability and attractivity of measure differential inclusions are given in
the Section 6.4.
The definitions of convergence properties presented here are based on and
extend the definition given in [44] (see also [132]). Consider a system described
by the measure differential inclusion
dx ∈ dΓ (t, x), (8.1)
where x ∈ Rn , t ∈ R.
192 8 Convergence Properties of Monotone Measure Differential Inclusions

Let us formally define the property of convergence.

Definition 8.1. System (8.1) is said to be
• convergent if there exists a solution x̄(t) satisfying the following conditions:
(i) x̄(t) is defined for almost all t ∈ R,
(ii) x̄(t) is bounded for all t ∈ R for which x̄(t) exists,
(iii) x̄(t) is globally attractively stable.
• uniformly convergent if it is convergent and x̄(t) is globally uniformly at-
tractively stable.
• exponentially convergent if it is convergent and x̄(t) is globally exponen-
tially stable.
The wording ‘attractively stable’ has been used instead of the usual term
‘asymptotically stable’, because attractivity of solutions in (measure) differ-
ential inclusions can be asymptotic or symptotic (finite-time attractivity), see
page 121.
The solution x̄(t) is called a steady-state solution. As follows from the
definition of convergence, any solution of a convergent system “forgets” its
initial condition and converges to some steady-state solution. In general, the
steady-state solution x̄(t) may be non-unique. But for any two steady-state
solutions x̄1 (t) and x̄2 (t) it holds that x̄1 (t) − x̄2 (t) → 0 as t → +∞. At
the same time, for uniformly convergent systems the steady-state solution is
unique, as formulated below.
Property 8.2 ( [131, 132]). If system (8.1) is uniformly convergent, then the
steady-state solution x̄(t) is the only solution defined and bounded for almost
all t ∈ R.
In many engineering problems, dynamical systems excited by time-varying
perturbations are encountered. Therefore, we will consider convergence prop-
erties for systems with time-varying inputs. So, instead of systems of the
form (8.1), we consider systems of the form

dx ∈ dΓ (x, w(t)), (8.2)

with state x ∈ Rn and input w ∈ Rd . The right-hand side of (8.2) is assumed

to be continuous in w. In the following, we will consider the class PCd of
piecewise continuous inputs w(t) : R → Rd which are bounded on R. Below
we define the convergence property for systems with inputs.
Definition 8.3. System (8.2) is said to be (uniformly, exponentially) conver-
gent if it is (uniformly, exponentially) convergent for every input w ∈ PCd . In
order to emphasize the dependency on the input w(t), the steady-state solution
is denoted by x̄w (t).
Uniformly convergent systems excited by periodic or constant inputs ex-
hibit the following property, that is particularly useful in, for example, bifur-
cation analyses of periodically perturbed systems.
 8.2 Convergence of Maximal Monotone Systems 193

Property 8.4 ( [44,132]). Suppose system (8.2) with a given input w(t) is uni-
formly convergent. If the input w(t) is constant, the corresponding steady-
state solution x̄w (t) is also constant; if the input w(t) is periodic with pe-
riod T , then the corresponding steady-state solution x̄w (t) is also periodic
with the same period T .

8.2 Convergence of Maximal Monotone Systems

In this section we will consider the dynamics of measure differential inclu-
sions (8.2) with certain maximal monotonicity conditions on Γ (x, w(t)). In
particular, we study systems for which Γ (x, w(t)) can be split in a state-
dependent part and an input-dependent part. The state-dependent part is,
with the help of a maximal monotonicity requirement, assumed to be strictly
passive with respect to the Lebesgue measure and passive with respect to the
atomic measure. Such kind of systems will be simply referred to as ‘maximal
monotone systems’ in the following.
We first formalise maximal monotone systems in Section 8.2.1, subse-
quently give sufficient conditions for the existence of a compact positively
invariant set in Section 8.2.2 and finally give sufficient conditions for conver-
gence in Section 8.2.3.

8.2.1 Maximal Monotone Systems

Let x ∈ Rn be the state-vector of the system and w ∈ Rm be the input vector.

Consider the time-evolution of x to be governed by a measure differential
equation of the form

dx = −da − c(x)dt + db(w), (8.3)

where c : Rn → Rn is a single-valued function and da and db(w) are differ-

ential measures with densities with respect to dt and dη, i.e.

da = at dt + aη dη, (8.4)

and
db(w) = bt (w)dt + bη (w)dη. (8.5)
In the following, we will assume xT bη (w) to be bounded from above by a
constant β. Basically, this gives an upper-bound on the energy input of the
impulsive inputs. Such an assumption makes sense from the physical point of
view, see the example in Section 8.4.1. The quantities at and aη , which are
functions of time, obey the set-valued laws

at ∈ A(x), (8.6)

194 8 Convergence Properties of Monotone Measure Differential Inclusions

aη ∈ A(x+ ), (8.7)

where A is a set-valued mapping. The dynamics can be decomposed in a
Lebesgue measurable part and an atomic part. The Lebesgue measurable part
gives the differential equation

ẋ(t) := xt = −at (x(t)) − c(x(t)) + bt (w(t)), (8.8)

which forms with the set-valued law (8.6) a differential inclusion

ẋ ∈ −A(x) − c(x) + bt (w) a.e. (8.9)

The atomic part gives the state-reset rule

x+ − x− := xη = −aη + bη (w). (8.10)

In mechanics, the state-reset rule is called the impact equation. The above
impact law (8.7), for which A is only a function of x+ , corresponds to a
completely inelastic impact equation. Because of the similarity between the
laws (8.6) and (8.7), we can combine these laws into the measure law

da ∈ dA(x+ ) = A(x+ )(dt + dη). (8.11)

The equality of measures (8.3) together with the measure law (8.11) consti-
tutes a measure differential inclusion

dx ∈ −dA(x+ ) − c(x)dt + db(w) := dΓ (x, w). (8.12)

The set-valued operator A(x) models the non-smooth dissipative elements in

the system. We assume that A(x) is a maximal monotone set-valued mapping,
i.e. A(x) satisfies

(x2 − x1 )T (A(x2 ) − A(x1 )) ≥ 0, (8.13)

for any two states x1 , x2 ∈ X . Moreover, we assume that 0 ∈ A(0). This last
assumption together with the monotonicity assumption implies the condition

xT A(x) ≥ 0 (8.14)

for any x ∈ X , i.e. the action of A is passive. Furthermore, we assume that

A(x) + c(x) is a strictly maximal monotone set-valued mapping, i.e. there
exists an α > 0 such that

(x2 − x1 )T (A(x2 ) + c(x2 ) − A(x1 ) − c(x1 )) ≥ αx2 − x1 2 , (8.15)

for any two states x1 , x2 ∈ X .

 8.2 Convergence of Maximal Monotone Systems 195

8.2.2 Existence of a Compact Positively Invariant Set

The existence of a compact positively invariant set is useful in the proof of

convergence as will become clear in Section 8.2.3. Clearly, if the impulsive
inputs are passive in the sense that (x+ )T bη (w(t)) ≤ 0 for all t, then the
system is dissipative for large x and all solutions must be bounded. In
the following theorem, we give a less stringent sufficient condition for the
existence of a compact positively invariant set of (8.12) based on a dwell-time
condition [75, 76].
Theorem 8.5. A measure differential inclusions of the form (8.12) has a
compact positively invariant set if
1. A(x) is a maximal monotone set-valued mapping with 0 ∈ A(0),
2. A(x)+c(x) is a strictly maximal monotone set-valued mapping, i.e. there
exists an α > 0 such that (8.15) is satisfied,
3. there exists a scalar β ∈ R such that (x+ )T bη (w) ≤ β for all x ∈ X , i.e.
the energy input of the impulsive inputs is bounded from above,
4. the time-instances ti for which the input is impulsive are separated by
the dwell-time τ ≤ ti+1 − ti , with τ = 2(δ−1)α δ
ln(1 + δ2β
2 γ 2 ) and γ :=
 
α supt∈R,at (0)∈A(0) {−at (0) − c(0) + bt (w(t))} for some δ > 1.
1

Proof: Consider the Lyapunov candidate function W = 12 xT x. The differ-

ential measure of W has a density Ẇ with respect to the Lebesgue mea-
sure dt and a density W + − W − with respect to the atomic measure dη, i.e.
dW = Ẇ dt + (W + − W − )dη. We first evaluate the density Ẇ :

Ẇ = xT (−at − c(x) + bt (w))

(8.16)
= xT (−at − c(x) + at (0) + c(0)) + xT (−at (0) − c(0) + bt (w)),

with at ∈ A(x) and at (0) ∈ A(0). Due to strict monotonicity of A(x)+c(x),
there exists a constant α > 0 such that
Ẇ ≤ −αx2 + xT (−at (0) − c(0) + bt (w)),
 (8.17)
≤ −x αx − supt∈R,at (0)∈A(0) {−at (0) − c(0) + bt (w(t))} .

Note that Ẇ < 0 for x satisfying

1
x > supt∈R,at (0)∈A(0) {−at (0) − c(0) + bt (w(t))} . (8.18)
α
Let γ be
 
1  
γ = max 0, supt∈R,at (0)∈A(0) {−at (0) − c(0) + bt (w(t))} . (8.19)
α
For x > γ we can prove an exponential decay of W (in between state jumps
at t = ti ). The function f (x) = −(1− 1δ )αx2 is greater than g(x) = −αx2 +γαx
196 8 Convergence Properties of Monotone Measure Differential Inclusions

for x > δγ, where δ > 1 is an arbitrary constant and γ > 0. It therefore holds
that Ẇ ≤ −(1 − 1δ )αx2 for x ≥ δγ, i.e.
 
1
Ẇ ≤ −2 1 − αW for x ≥ δγ. (8.20)
δ

Subsequently, we consider the jump W + − W − of W :

1 +  
W+ − W− = (x + x− )T x+ − x− (8.21)
2
with x+ − x− = −aη + bη (w) and aη ∈ A(x+ ). Elimination of x− and
exploiting the monotonicity of A(x) gives
1  
W+ − W− = (2x+ + aη − bη (w))T −aη + bη (w)
2
  1( (2
= (x+ )T −aη + bη (w) − (aη − bη (w)( (8.22)
2
≤ β,

in which we used the assumption that the energy input of the impulsive inputs
bη (w) is bounded from above by β (see condition 3 in the theorem) and the
monotonicity and passivity of A. Then, due to (8.17) and (8.18), for the non-
impulsive part of the motion it holds that if x(t0 ) ≤ γ then x(t) ≤ γ
for all t ∈ [t0 , t∗ ] (if no state resets occur in this time interval). Moreover, as
far as the state resets are concerned, (8.22) shows that a state reset from a
state x− (ti ) ∈ V with V = {x ∈ X | x ≤ δγ} can only occur to x+ (ti ) such
that W (x+ (ti )) := 12 x+ (ti )2 ≤ W (x− (ti )) + β ≤ 12 δ 2 γ 2 + β (note hereto the
specific form of W = 12 xT x). During the following open time-interval (ti , ti+1 )
for which bη (w(t)) = 0, the function W evolves as

− +
W (x (ti+1 )) = W (x (ti )) + dW, (8.23)
(ti ,ti+1 )

which may involve impulsive motion due to dissipative impulses aη . Let tV ∈
(ti , ti+1 ) be the time-instance for which x− (tV ) = δγ. The function W
will necessarily decrease during the time-interval (ti , tV ) due to (8.20) and
( (2
W + − W − = −(x+ )T aη − 12 (aη ( ≤ 0 (the state-dependent impulses are
passive). It therefore holds that

W (x− (tV )) ≤ e−2(1− δ )α(tV −ti ) W (x+ (ti )),

1
(8.24)

because dW ≤ −2(1 − 1δ )αW dt + (W + − W − )dη ≤ −2(1 − 1δ )αW dt for

positive measures. Using W (x− (tV )) = 12 δ 2 γ 2 and W (x+ (ti )) ≤ 12 δ 2 γ 2 + β in
the exponential decrease (8.24) gives
1 2 2 1
δ γ ≤ e−2(1− δ )α(tV −ti ) ( δ 2 γ 2 + β)
1
(8.25)
2 2
 8.2 Convergence of Maximal Monotone Systems 197

or
δ 2β
tV − ti ≤ ln(1 + 2 2 ). (8.26)
2(δ − 1)α δ γ
Consequently, if the next impulsive time-instance ti+1 of the input is after tV ,
then the solution x(t) has enough time to reach V. Hence, if the impulsive
time-instance of the input are separated by the dwell-time τ , i.e. ti+1 − ti ≥ τ ,
with
δ 2β
τ= ln(1 + 2 2 ), (8.27)
2(δ − 1)α δ γ
then the set  % 
%1 1 2 2
W= %
x ∈ X % x ≤ δ γ + β
2
(8.28)
2 2
is a compact positively invariant set. 
Typically, we would like the invariant set W to be as small as possible, as it
gives an upper-bound for the trajectories of the system. On the other hand,
we also want the dwell-time to be as small as possible. The constant δ > 1
plays in interesting role in the above theorem. By increasing δ, we allow the
invariant set W to be larger, thereby decreasing the dwell-time τ . So, there
is a kind of trade-off between the size of the invariant set and the dwell-time.
Any finite value of δ is sufficient to prove the existence of a compact positively
invariant set. We therefore can take the dwell-time τ to be an arbitrary small
value, but not infinitely small. This brings us to the following corollary:

Corollary 8.6. If the size of the compact positively invariant set is not of
interest, then Condition 4 in Theorem 8.5 can be replaced by an arbitrary
small dwell-time τ > 0.

Proof: Taking the limit of δ → ∞ gives the condition τ > 0 for arbitrary γ
and β. 
It therefore suffices to assume that the impulsive inputs are separated in time
(which is not a strange assumption from a physical point of view) and simply
put τ equal to the (unknown) minimal time-lapse between the impulsive in-
puts. Then, we calculate the corresponding value of δ and obtain the size of
the compact positively invariant set.
In this section, we presented a sufficient condition for the existence of a
compact positively invariant set, but the attractivity of solutions outside W
to W is not guaranteed. If in addition the system is incrementally attractively
stable, for which we will give a sufficient condition in Section 8.2.3, then it is
also assured that all solutions outside W converge to W.

8.2.3 Conditions for Convergence

In the following theorem, it is stated that strictly maximal monotone measure

differential inclusions are uniformly convergent.
198 8 Convergence Properties of Monotone Measure Differential Inclusions

Theorem 8.7. A measure differential inclusion of the form (8.12) is expo-

nentially convergent if
1. A(x) is a maximal monotone set-valued mapping, with 0 ∈ A(0),
2. A(x) + c(x) is a strictly maximal monotone set-valued mapping,
3. system (8.12) exhibits a compact positively invariant set.

Proof:
Let us first show that system (8.12) is incrementally attractively stable, i.e. all
solutions of (8.12) converge to each other for positive time. Consider hereto
two solutions x1 (t) and x2 (t) of (8.12) and a Lyapunov candidate function
V = 12 x2 − x1 2 . Consequently, the differential measure of V satisfies:

1 +
dV = (x + x− + − T
2 − x1 − x1 ) (dx2 − dx1 ) , (8.29)
2 2
with

dx1 = −da1 − c(x1 )dt + db(w), dx2 = −da2 − c(x2 )dt + db(w), (8.30)

where da1 ∈ A(x+ 1 ) and da2 ∈ A(x2 ). The differential measure of V has a
+

density V̇ with respect to the Lebesgue measure dt and a density V + − V −

with respect to the atomic measure dη, i.e. dV = V̇ dt + (V + − V − )dη. We
first evaluate the density V̇ :

V̇ = −(x2 − x1 )T (at (x2 ) + c(x2 ) − bt (w) − at (x1 ) − c(x1 ) + bt (w))
= −(x2 − x1 )T (at (x2 ) + c(x2 ) − at (x1 ) − c(x1 )),
(8.31)

where at (x1 ) ∈ A(x1 ) and at (x2 ) ∈ A(x2 ), since both solutions x1 and
x2 correspond to the same perturbation w. Due to strict monotonicity of
A(x) + c(x), there exists a constant α > 0 such that

V̇ ≤ −αx2 − x1 2 . (8.32)

Subsequently, we consider the jump V + − V − of V :

1 +  + 
V+−V− = (x + x− − T
2 − x1 − x1 )
+
x2 − x− −
2 − x1 + x1 ,
+
(8.33)
2 2
with
−  
1 − x1 = −aη (x1 ) + bη (w),
x+ aη (x1 ) ∈ A(x+
1 ),
−  
(8.34)
2 − x2 = −aη (x2 ) + bη (w),
x+ aη (x2 ) ∈ A(x+
2 ).

Elimination of x− −
1 and x2 and exploiting the monotonicity of A(x) gives
 8.3 Tracking Control for Lur’e Type Systems 199
1   
V+−V− = (2x+  
2 + aη (x2 ) − 2x1 − aη (x1 ))
+ T
−aη (x2 ) + aη (x1 )
2
   1( (2
= −(x+
2 − x1 )
+ T
aη (x2 ) − aη (x1 ) − (aη (x2 ) − aη (x1 )(
2
≤ 0.
(8.35)

It therefore holds that V strictly decreases over every non-empty compact

time-interval as long as x2 = x1 . In turn, this implies that all solutions
of (8.12) converge to each other exponentially (and therefore uniformly).
Finally we use Lemma 2 in [178], which formulates that if a system exhibits
a compact positively invariant set, then the existence of a solution that is
bounded for t ∈ R is guaranteed. We will denote this ‘steady-state’ solution
by x̄w (t). The original lemma is formulated for differential equations (possibly
with discontinuities, therewith including differential inclusions, with bounded
right-hand sides). Here, we use this lemma for measure differential inclusions
and would like to note that the proof of the lemma allows for such extensions
if we only require continuous dependence on initial conditions. The latter is
guaranteed for monotone measure differential inclusions, because incremental
stability implies a continuous dependence on initial conditions.
Since all solutions of (8.12) are globally exponentially stable, also x̄w (t)
is a globally exponentially stable solution. This concludes the proof that the
measure differential inclusions (8.12) is exponentially convergent. 

8.3 Tracking Control for Lur’e Type Systems

An important application of convergence theory is the tracking control of
dynamical systems, i.e. the design of a controller, such that a desired trajec-
tory xd (t) of the system exists and is globally attractively stable. Tracking
control of measure differential inclusions has received very little attention in
literature [21, 27, 113].
In this section, we consider the tracking control problem of a nonlinear
measure differential inclusion, which can be decomposed into a linear mea-
sure differential inclusion with a nonlinear maximal monotone operator in the
feedback path. We allow the desired trajectory xd (t) to have discontinuities
in time (but assume it to be of locally bounded variation). The open-loop
dynamics is described by an equality of measures:

dx = Axdt + Bdp + Dds,

y = Cx (8.36)
−ds ∈ H(y),

with A ∈ Rn×n , B ∈ Rn×np , C ∈ Rm×n and D ∈ Rn×m . Herein, x ∈ Rn

is the system state (of locally bounded variation), dp = wdt + W dη is the
200 8 Convergence Properties of Monotone Measure Differential Inclusions

differential measure of the control action and ds = λdt+Λdη is the differential

measure of the nonlinearity in the feedback loop that is characterised by the
set-valued maximal monotone mapping H(y). The problem that we consider
here is the design of a control law dp such that the tracking of the desired
trajectory xd (t) is assured. We propose to tackle the tracking problem by
means of a combination of Lebesgue measurable linear tracking error-feedback
and a possibly impulsive feedforward control:

dp = wf b dt + dpff (t) = K (x − xd (t)) dt + wff (t)dt + Wff (t)dη, (8.37)

with

wf b = K (x − xd (t)) , dpff (t) = wff (t)dt + Wff (t)dη, (8.38)

where K ∈ Rnp ×n is the feedback gain matrix and xd (t) the desired state
trajectory. We restrict the energy input of the impulsive control action Wff (t)
to be bounded from above

(x+ )T BWff ≤ β. (8.39)

Note that this condition puts a bound on the jumps in the desired trajectory
xd (t) which can be realised. Combining the control law (8.37) with the system
dynamics (8.36) yields the closed-loop dynamics:

dx = Acl xdt + Dds + B(−Kxd (t)dt + dpff (t)),

y = Cx (8.40)
−ds ∈ H(y),

with
Acl = A + BK. (8.41)
We now propose a convergence-based control design. The main idea of this
convergence-based control design is to find a controller of the form (8.37) that
guarantees two properties:
a. the closed-loop system has a trajectory which is bounded for all t and along
which the tracking error x − xd (t) is identically zero. In other words, the
feedforward wff (t) and Wff (t) has to be designed such that it induces
the desired solution xd (t);
b. the closed-loop system is uniformly convergent. Hereto, the control gain
matrix K should be designed appropriately.
Condition b) guarantees that the closed-loop system has a unique bounded
UGAS steady-state solution, while condition a) guarantees that, by Prop-
erty 8.2, this steady-state solution equals the bounded solution of the closed-
loop system with zero tracking error. For other types of systems, the con-
vergence property has been exploited to solve the output regulation problem,
tracking problems and the synchronisation problem, see e.g. [132,134,169,172].
 8.3 Tracking Control for Lur’e Type Systems 201

For the design of the feedback gain (to ensure that condition b is met), we
employ the following strategy. First, we design K such that the linear part of
system (8.40), (8.41) is strictly passive. Subsequently, using the fact that H(y)
is maximal monotone, we show that this implies that the measure differential
inclusion (8.40), (8.41) is (after a coordinate transformation) maximal mono-
tone. Hence, exponential convergence for measure differential inclusions of the
form (8.40) can be proven using Theorem 8.7. A similar result was found for
a class of differential inclusions by Yakubovich [178]. In [178] it is shown that
strict passivity of the linear part of the system is sufficient for exponential
convergence for Lur’e-type systems with monotone set-valued nonlinearities
and absolutely continuous state (i.e. for a class of differential inclusions).
Here, we will show that for measure differential inclusions (8.40), (8.41)
that, if the triple (Acl , D, C) is strictly positive real (i.e. the linear part of
the system (8.40) is strictly passive) and the nonlinearity H(y) is a monotone
nonlinearity, then the system is uniformly convergent. Therefore, the feedback
gain matrix K should be designed such that the triple (Acl , D, C) is strictly
positive real.
Note that the triple (Acl , D, C) is rendered strictly passive by means of
the feedback design. In other words we design K such that there exists a
positive definite matrix P = P T > 0 for which the following conditions are
satisfied:
AT
cl P + P Acl < 0,
(8.42)
D T P = C.

Let us now introduce a linear coordinate transformation x̃ = Sx, where P =

S T S, i.e. S is the square root of P . Using these transformed coordinates, the
closed-loop dynamics can then be formulated in the form (8.12):

dx̃ ∈ −dA(x̃+ ) − c(x̃)dt + db(w) (8.43)

with
dA(x̃+ ) = SDH(CS −1 x̃+ )(dt + dη), (8.44)
c(x̃) = −SAcl S −1 x̃, (8.45)
db(w) = SB(−Kxd (t)dt + dpff (t)). (8.46)
We will now show that condition (8.42) together with the monotonicity of
the set-valued mapping H(y) implies strict monotonicity of the differential
inclusion (8.43). Hereto, we prove the strict monotonicity of the set-valued
operator −SAcl S −1 x̃ + SDH(CS −1 x̃). Using λi ∈ −H(CS −1 x̃i ), i = 1, 2,
we can verify that it holds that
202 8 Convergence Properties of Monotone Measure Differential Inclusions
 
−SAcl S −1 x̃1 − SDλ1 + SAcl S −1 x̃2 + SDλ2
T
(x̃1 − x̃2 )
T 
= (x̃1 − x̃2 ) −SAcl S −1 (x̃1 − x̃2 ) + (x̃1 − x̃2 ) SD (λ2 − λ1 )
T

T   T
= − (x1 − x2 ) S T SAcl (x1 − x2 ) + (x1 − x2 ) S T SD (λ2 − λ1 )
1 T  T
= − (x1 − x2 ) P Acl + AT cl P (x1 − x2 ) + (x1 − x2 ) P D (λ2 − λ1 ) .
2
(8.47)

cl P + P Acl )x ≥ αx for

Using the conditions (8.42), we can write −xT (AT 2
T T T T
some α > 0 and x P D = x C = y and Eq. (8.47) becomes
T 
(x̃1 − x̃2 ) −SAcl S −1 x̃1 − SDλ1 + SAcl S −1 x̃2 + SDλ2
α T (8.48)
≥ x1 − x2 2 + (y1 − y2 ) (λ2 − λ1 ) .
2
Finally, we use the fact that λi ∈ −H(yi ), i = 1, 2, and the monotonicity of
the set-valued nonlinearity H(y) to conclude that
  α
−SAcl S −1 x̃1 − SDλ1 + SAcl S −1 x̃2 + SDλ2 ≥ x1 − x2 2 .
T
(x̃1 − x̃2 )
2
(8.49)
In other words, strict monotonicity of the x̃-dynamics is guaranteed. Ear-
lier in the chapter we have shown that strict monotonicity implies uniform
convergence. Moreover, the convergence property is conserved under smooth
coordinate transformations (see [132]). Consequently, if the x̃-dynamics is
uniformly convergent, then also the x-dynamics is uniformly convergent.

8.4 Illustrative Examples

In the next sections, examples concerning models for the control of mechanical
systems with set-valued friction and one-way clutches illustrate the power of
the result in Theorem 8.7. Moreover, the results of Section 8.3 on tracking
control are applied to mechanical systems with friction and a one-way clutch
in Sections 8.4.2 and 8.4.3.

8.4.1 One-way Clutch

The time-evolution of the velocity of a mass m (Figure 8.1) subjected to a one-

way clutch, a dashpot b > 0 and an external input (considering both bounded
and impulsive contributions) can be described by the equality of measures

mdu = dp + ds − budt. (8.50)

We can decompose the differential measure ds of the one-way clutch in

ds = λdt + Λdη, (8.51)

 8.4 Illustrative Examples 203

Fig. 8.1. Mass with one-way clutch and impulsive actuation.

where λ := st is the contact force and Λ = sη is the contact impulse. The
differential impulse measure ds of the one-way clutch obeys the set-valued
force law
−ds ∈ Upr(u+ ), (8.52)
where Upr(x) is the unilateral primitive (2.24).

−y ∈ Upr(x) ⇐⇒ 0 ≤ x ⊥ y ≥ 0 ⇐⇒ x ≥ 0, y ≥ 0, xy = 0, (8.53)

being a maximal monotone operator.

The input consists of a bounded force f and an impulse F

dp = f dt + F dη. (8.54)

We assume that an impulsive input F > 0 is transmitted by firing bullets

with mass m0 and constant speed v ≤ vmax on the left side of the mass m.
We assume a completely inelastic impact. If u ≥ v, then the bullet is not able
to hit the mass m and then the impulse F equals zero. If u+ < v, then the
impulse F equals the mass of the bullet multiplied with its velocity jump:

F = m0 (v − u+ ). (8.55)

Similarly, we assume that an impulsive input F < 0 is transmitted by firing

on the right side of the mass m with a speed v < 0, bounded by |v| ≤ vmax .
The energy input u+ F = m0 u+ (v − u+ ) of the impulse F is maximal when
u+ = 12 v and is therefore bounded from above by β := 14 m0 vmax
2
≥ |u+ F |.
We first prove the existence of a compact positively invariant set with
Theorem 8.5. Theorem 8.5 uses the Lyapunov function W (u) = 12 u2 , which we
recognise to be the kinetic energy divided by the mass m. The time-derivative
Ẇ gives, using uλ = 0,
b 2
Ẇ ≤ − u + u sup(f (t)), (8.56)
m t∈R

b
and it therefore holds that α = m and γ = 1b supt∈R (f (t)) with α and γ
defined in Theorem 8.5. Theorem 8.5 states that if the time-instances ti of
the impulses F are separated by the dwell-time
204 8 Convergence Properties of Monotone Measure Differential Inclusions

δ 2β
τ= ln(1 + 2 2 ), (8.57)
2(δ − 1)α δ γ
then the set  % 
%
+% 1 2 1 2 2
W = u∈R % u ≤ δ γ +β (8.58)
2 2
is a compact positively invariant set for arbitrary δ > 1. Following Corol-
lory 8.6, we conclude that the dwell-time can be made arbitrary small by
increasing δ. We therefore take τ to be smaller than the minimal time-lapse
between two succeeding impulsive time-instances, which gives a lower bound
for δ.
Just as in the proof of Theorem 8.7, we prove incremental stability using
the Lyapunov function
1
V = (u2 − u1 )2 . (8.59)
2
First, we consider the time-derivative V̇ :

V̇ = (u2 − u1 )(u̇2 − u̇1 )

1
= (u2 − u1 ) (λ2 − bu2 − λ1 + bu1 )
m
1 b
= (u2 − u1 ) (λ2 − λ1 ) − (u2 − u1 )2 , −λ1 ∈ Upr(u1 ), −λ2 ∈ Upr(u2 )
m m
b
≤ − (u2 − u1 )2 .
m
(8.60)

Subsequently, we consider a jump in V :

− −
V + − V − = V (u+
1 , u2 ) − V (u1 , u2 )
+

1 1 − − 2
= (u+ − u+1 ) − (u2 − u1 )
2
(8.61)
2 2 2
1
= (u+ + u− − − −
2 − u1 − u1 )(u2 − u2 − u1 + u1 ).
+ + +
2 2
Following the proof of Theorem 8.7, we eliminate u− −
1 and u2 by substituting
−
the impact equation m(uj − uj ) = Λj + F , j = 1, 2:
+

1 1 1 1
V+−V− = 2 −
(2u+ Λ2 − 2u+
1 + Λ1 ) (Λ2 − Λ1 )
2 m m m
+ 1 1 (8.62)
= (u+2 − u1 ) (Λ2 − Λ1 ) − (Λ2 − Λ1 )2
m 2m2
≤ 0.

Hence, it holds for the differential measure dV that

b
dV = V̇ dt + (V + − V − )dη ≤ −α(u2 − u1 )2 dt, α= . (8.63)
m
 8.4 Illustrative Examples 205

Fig. 8.2. Typical motor-load configuration with non-collocated friction and actua-
tion.

Integration of dV over a non-empty time-interval therefore leads to a strict de-

crease of the function V as long as u2 = u1 . This proves incremental stability.
Consequently, the system is exponentially convergent (see Theorem 8.7).

8.4.2 Tracking Control with Set-valued Friction

In this section, we consider the tracking control problem for mechanical sys-
tems with set-valued friction. Hereto, we study a common motor-load con-
figuration as depicted in Figure 8.2. The essential problem here is the fact
that the friction and the actuation are non-collocated (i.e. the motor, mass
m1 , is actuated and the load, mass m2 , is subject to friction). Note that the
spring-damper combination, with stiffness c and viscous damping constant
b, reflects a finite-stiffness coupling between the motor and load common in
many motion systems. A common approach in tackling control problems for
systems with friction is that of friction compensation. This angle of attack is
clearly not feasible here since the actuation cannot compensate directly for the
friction. Another common approach in compensating for nonlinearities can be
recognised in the backstepping control schemes [85]. However, these generally
require differentiability of the nonlinearity, which is not the case here due to
the set-valued nature of the friction law.
In many applications, mainly the velocity of the load is of interest. In
this context, one can think of controlling a printhead in a printer, where the
printhead is to achieve a constant velocity when moving across the paper
or drilling systems where the bottom-hole-assemble (including the drill bit)
should achieve a constant cutting speed. From this perspective, the following
third-order differential inclusion describes the dynamics of the system under
study:

ẋ = Āx + Bw + D λ̄,
y = Cx (8.64)
λ̄ ∈ −H̄(y),

with
206 8 Convergence Properties of Monotone Measure Differential Inclusions

Fig. 8.3. Friction law H(y) and transformed friction law H̄(y).

⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
0 −1 1 0 0 0
Ā = ⎣ mc − mb1 b ⎦
m1 , B = ⎣ m11 ⎦ , D = ⎣ 0 ⎦ , C T = ⎣0⎦ . (8.65)
1
1
− mc2 b
m2 − mb2 0 m2 1
 T
Herein, x = q2 − q1 q̇1 q̇2 is the absolutely continuous system state, w ∈ R
is the control action and λ̄ ∈ R is the friction force that is characterised by the
set-valued mapping H̄(·) : R → R. The set-valued friction law adopted here
includes a combination of Coulomb friction, viscous friction and the Stribeck
effect:  
μ2 y
H̄(y) = m2 g μ0 Sign(y) + μ1 y − , (8.66)
1 + μ2 |y|
where g is the gravitational acceleration, μ0 > 0 is the Coulomb friction
coefficient, μ1 > 0 is the viscous friction coefficient and μ2 is an additional
coefficient characterising the modelling of the Stribeck effect. It is well known
that exactly such a Stribeck effect can induce instabilities, complicating the
design of stabilising controllers, see e.g. [9]. In Figure 8.3, such type of set-
valued static friction is depicted schematically. At this point, we will transform
the friction law H̄(y) to a strictly maximal monotone operator H(y)

H(y) = H̄(y) + κy, with κ = m2 g(μ2 − μ1 ), (8.67)

where the choice of κ ensures that the set-valued mapping H(y) is a maxi-
mal monotone mapping. System (8.64) can therefore be transformed into the
form (8.36)
 8.4 Illustrative Examples 207

Fig. 8.4. Feedback and feedforward control.

dx = Axdt + Bdp + Dds,

y = Cx (8.68)
−ds ∈ H(y),

where A = Ā + κDC, dp = wdt, ds = λdt and H(y) is a maximal monotone

mapping. We now use the convergence-based tracking control strategy pro-
posed in Section 8.3 to solve the tracking problem of this mechanical system
with friction. Hereto, we us a combination  of linear  error-feedback and feedfor-
ward control as in (8.37), where K = k1 k2 k3 is the feedback gain matrix
which has to be chosen such that the triple (Acl , D, C) is strictly positive
real.
We adopt the following system parameters: g = 9.81, m1 = 1, m2 = 1, c =
100, b = 1, μ0 = 0.2, μ1 = 0.1 and μ2 = 0.2. The resulting friction map and
the transformed (monotone) friction map are shown in Figure 8.3. Moreover,
we aim at tracking a constant velocity solution (with desired velocity vd )
for both the motor and the load; i.e. the desired state-trajectory is given by
)   *T
xd (t) = 1c −m2 g μ0 + μ1 vd − 1+μ μ2 vd
2 |vd |
v d v d , where vd = 1. Note that
this velocity lies in the range in which the friction law exhibits a pronounced
Stribeck effect. Let us design the controller in the form (8.37). Firstly, the
feedforward which induces the desired solution is given by
 
μ2 vd
wff = m2 g μ0 + μ1 vd − . (8.69)
1 + μ2 |vd |
Secondly, by checking appropriate LMI conditions or frequency-domain con-
ditions (see e.g. [85]) for the strictly positive realness of the triple (Acl , D, C),
208 8 Convergence Properties of Monotone Measure Differential Inclusions

Fig. 8.5. Only feedforward control.

suitable controller gains can be selected: k1 = −30, k2 = −150 and k3 = −150.

The strict positive realness of the triple (Acl , D, C) can be proven using the
following symmetric, positive definite matrix P
⎡ ⎤
50.98 −0.33 0
⎢ ⎥
P = ⎣−0.33 0.005 0⎦ > 0. (8.70)
0 0 1

satisfying the LMIs (8.42).

Next, we implement control law (8.37) on system (8.64), with these control
gains and feedforward (8.69) and use numerical time-stepping schemes [1,101,
123] to numerically compute solution of the closed-loop system. In Figure 8.4,
the velocities of both masses are depicted when the controller is active and
asymptotic tracking of the constant velocity solution is achieved. Note that
when only the feedforward is applied, the desired solution is still a solution of
the system; however, no asymptotic tracking achieved, see Figure 8.5. In this
figure, it is shown that both masses ultimately come to a standstill. Clearly, the
system now exhibits at least two steady-state solutions; the desired solution
and the solution on which x1 = −uff /c, x2 = 0 and x3 = 0, as depicted in
Figure 8.5. Consequently, the system without feedback is not convergent. For
 T
both cases the initial condition x(0) = 0 0.8 0.8 was used.
Note that we solve a stabilisation problem in this example. However, using
the strategy discussed here, we can make any bounded feasible time-varying
desired solution xd (t) attractively stable using the same feedback gain matrix
K and an appropriate feedforward.
 8.4 Illustrative Examples 209

Fig. 8.6. Motor-load configuration with one-way clutch and impulsive actuation.

8.4.3 Tracking Control Using an Impulsive Input

In the example of Section 8.4.2 we solved the tracking problem mechanical

motion system with set-valued friction. In the current example, we consider
a similar system; however, the set-valued friction is replaced by a one-way
clutch and impulsive control action is needed to achieve tracking of a periodic
trajectory.
More specifically, we study a variant of the previous problem and replace
the friction element by a one-way clutch and add an additional damper b2 .
Moreover, we allow for impulsive inputs on the first mass. The open-loop
dynamics is now described by an equality of measures:
dx = Axdt + Bdp + Dds,
y = Cx (8.71)
−ds ∈ H(y),
with
⎡ ⎤ ⎡ ⎤ ⎤
⎡ ⎡ ⎤
0 −1 1 0 0 0
A = ⎣ mc1 − 1m1 2 m11 ⎦ , B = ⎣ 1 ⎦ , D = ⎣ 0 ⎦ , C T = ⎣0⎦ . (8.72)
b +b b
m1
1
− m2 m2 − m
c b1 b1
2
0 m2 1
 T
The evolution x(t) of the state vector x = q2 − q1 u1 u2 is of locally
bounded variation. The differential measure of the control action dp =
wdt + W dη now also contains an impulsive part W . The differential measure
ds of the force in the one-way clutch is characterised by the scalar set-valued
maximal monotone mapping H(x) = Upr(x).
In this example, we try to let the velocity x3 (t) = q̇2 (t) approach the de-
sired trajectory xd3 (t). Hereto, we design trajectories xd1 (t) and xd2 (t) which
generate the desired xd3 (t). Subsequently, we aim at tracking of the desired
state trajectory xd (t). The state-tracking problem is solved by making the sys-
tem uniformly convergent with a feedback K(x − xd (t)). As in Section 8.4.2,
we can design K such that the triple (Acl , D, C) is rendered strictly passive,
which, given the monotonicity of H(y), makes the system uniformly conver-
gent.
210 8 Convergence Properties of Monotone Measure Differential Inclusions

Fig. 8.7. Desired trajectory xd3 (t).

Fig. 8.8. Feedforward dpff /dt.

 8.4 Illustrative Examples 211

We adopt the following system parameters: g > 0, m1 = 1, m2 = 1, c = 10,

b1 = 1 and b2 = −1.4. The negative damping b2 < 0 causes the system matrix
A to have a positive real eigenvalue. The desired velocity of the second mass
is characterised by a periodic sawtooth wave with period time T :
⎧
⎪
⎪mod(t, T ) for 0 ≤ mod(t, T ) ≤ T4 (ramp-up)
⎨
xd3 (t) = − mod(t, T ) + T for T ≤ mod(t, T ) ≤ T (ramp-down) (8.73)
⎪
⎪ 2 4 2
⎩
0 for 2 ≤ mod(t, T ) ≤ T (deadband)
T

The signal xd3 (t) for T = 1 s is shown in Figure 8.7. The desired trajectory
xd3 (t) is a periodic signal which is time-continuous but has three kinks in each
period. Kinks in xd3 (t) can be achieved by applying an impulsive force on the
first mass which causes an instantaneous change in the velocity x2 = q̇1 and
therefore a discontinuous force in the damper b1 . The one-way clutch on the
second mass prevents negative values of xd3 and no impulsive force on the
first mass is therefore necessary for the change from ramp-down to deadband.
In a first step, the signals xd1 (t), xd2 (t) and ds(t) are designed such that

ẋd1 (t) = −xd2 (t) + xd3 (t)

dxd3 (t) = (− mc2 xd1 (t) − m b1
2
(−xd2 (t) + xd3 (t)))dt + 1
m2 ds(t)
(8.74)
−ds(t) ∈ Upr(xd3 (t)),

for the given periodic trajectory xd3 (t). The solution of this problem is not
unique as we are free to chose ds(t) ≥ 0 for xd3 (t) = 0. By fixing ds(t) = ṡ0 dt
to a constant value for xd3 (t) = 0 (i.e. ṡ0 is a constant), we obtain the following
discontinuous differential equation for xd1 (t):
⎧
⎨ m2 (−ẋ (t) − c x ) xd3 (t) > 0,
d3 m2 1d
ẋd1 = b1 (8.75)
⎩ m2 (−ẋd3 (t) − c x1d + 1 ṡ0 ) xd3 (t) = 0.
b1 m2 m2

The numerical solution of xd1 (t) gives (after a transient) a periodic signal
xd1 (t) and xd2 (t) = −ẋd1 (t) + xd3 (t) (see the dotted lines in Figures 8.11
and 8.12 which are mostly below the solid lines). We have taken ṡ0 = 1.
Subsequently, the feedforward input dpff = wff dt + Wff dη is designed such
that
dpff = m1 dxd2 − (cxd1 + b1 (−xd2 + xd3 ) − b2 xd2 ) dt (8.76)
and it therefore holds that x(t) = xd (t) for t ≥ 0 if x(0) = xd (0), where
x(t) is a solution of (8.71), (8.72), with dp = dpff . The feedforward input
dpff /dt is shown in Figure 8.8 and is equal to wff (t) almost everywhere.
Two impulsive inputs Wff (t) per period can be seen at the time-instances for
which there is a ‘change ramp-up to ramp-down’ and ‘ramp-down to dead-
band’. Next, we implement the control law (8.37)) on system
* (8.71) with the
feedforward dpff as in (8.76). We choose K = 0 −4 0 which renders the
212 8 Convergence Properties of Monotone Measure Differential Inclusions

Fig. 8.9. Trajectories x3 (t) (solid) and xd3 (t) (dotted) for the case of feedback and
feedforward control.

Fig. 8.10. Trajectories x3 (t) (solid) and xd3 (t) (dotted) for the case of only feed-
forward control.
 8.4 Illustrative Examples 213

triple (Acl , D, C) strictly positive real and, therefore, the closed-loop sys-
tem (8.71), (8.72), (8.37), (8.76) has convergent dynamics. The strict positive
realness of the triple (Acl , D, C) can be proven using the following matrix P
⎡ ⎤
1 3
5 5 0
⎢ ⎥
P = ⎣ 35 4 0⎦ > 0, Q = −(Acl P + P AT cl ) > 0, D T P = C, (8.77)
0 01

where Acl = A+BK. Figure 8.9 shows the closed-loop dynamics for which the
desired periodic solution xd (t) is globally attractively stable. The attraction
to the periodic solution from an arbitrary initial condition occurs in finite time
(symptotic attraction). Figure 8.10 shows the open-loop dynamics for which
there is no state-feedback. The desired periodic solution xd (t) is not globally
attractive, not even locally, and the solution from the chosen initial condition
is attracted to a stable period-2 solution. Clearly, the system without feedback
) *T
is not convergent. For both cases the initial condition x(0) = 0.16 2.17 0
was used. Figures 8.11 and 8.12 show the time-histories of x1 (t) and xd1 (t),
respectively x2 (t) and xd2 (t), in solid and dotted lines. Jumps in the state
x2 (t) and desired state xd2 (t) can be seen on time-instances for which the
input is impulsive.

Fig. 8.11. Trajectories x1 (t) (solid) and xd1 (t) (dotted) for the case of feedback
and feedforward control.
214 8 Convergence Properties of Monotone Measure Differential Inclusions

Fig. 8.12. Trajectories x2 (t) (solid) and xd2 (t) (dotted) for the case of feedback
and feedforward control.

8.5 Summary

In the previous sections, sufficient conditions have been derived for the uniform
convergence of a class of measure differential inclusions with certain maximal
monotonicity properties. We will summarise the main ideas of the chapter.
First, sufficient conditions have been presented in Theorem 8.5 for the exis-
tence of a compact positively invariant set. Theorem 8.5 relies on a Lyapunov-
based argument with the squared magnitude of the state as Lyapunov func-
tion, which acts as a kind of energy function. The assumption of strict mono-
tonicity of the Lebesgue part of the state dependent right-hand side equals
a strict passivity requirement with a quadratic dissipation. The quadratic
dissipation can always outperform the linear energy input of bounded non-
impulsive forces. Hence, during non-impulsive motion, the system dissipates
energy for large enough magnitudes of the state. The assumption of mono-
tonicity of the atomic part of the state dependent right-hand side equals a
passivity requirement. Moreover, the energy input of the impulsive inputs is
assumed to be bounded. This means that, for a given size of the compact
positively invariant set of which we like to prove existence, we can find a
dwell-time for the impulsive inputs. If the time-lapse between subsequent im-
pulsive inputs is larger than this dwell-time, then the Lebesgue measurable
dissipative forces have enough time to ‘eat’ the energy input of the impulsive
input. This reasoning works also in the opposite direction. Given a certain
dwell-time, there exists a certain compact positively invariant set of the sys-
tem. This means that the dwell-time is not really a condition for the existence
of a compact positively invariant set, but is merely a constant which relates
 8.5 Summary 215

to the size of such a set. The existence of such a set guarantees the existence
of a solution that is bounded for all times.
Subsequently, sufficient conditions for incremental attractive stability have
been derived in Theorem 8.7 using again a Lyapunov-based approach. The
decrease of the Lyapunov function, which measures the distance between
two arbitrary solutions, follows from a monotonicity condition. Incremental
attractive stability implies that all solutions converge to one another. The
aforementioned bounded (steady-state) solution must therefore be globally
asymptotically stable for all bounded inputs, which rigourously proves uni-
form convergence of the system (Theorem 8.7).
The above theorems hinge on a few important assumptions, which we can
give the following interpretations in the context of mechanical systems with
impulsive right-hand sides:
1. Separation of state-dependent forces and inputs. In other words: no cross-
talk between state-dependent forces and inputs. This excludes mixed
terms in state and input, which for instance arise if the generalised force
directions of the input forces are state-dependent.
2. A strict monotonicity condition on the Lebesgue measurable right-hand
side. This implies that the state-dependent forces in the system are strictly
passive.
3. A monotonicity condition on the atomic (impulsive) right-hand side. This
implies that the state-dependent impulses in the system are passive.
4. Bounded energy input of the impulsive inputs. The physical meaning of
this assumption has been elucidated in Section 8.4.1.
5. A dwell-time condition. The dwell-time can be chosen to be arbitrary
small. In practice, there always exist a minimal time between two impul-
sive inputs which can be exerted on the system.
Condition 2 is the condition which may limit most of all the use of Theo-
rem 8.7, simply because many systems are not dissipative. However, systems
can be made dissipative using an appropriate control. In other words, the pre-
sented theorems give us the knowledge how to design controllers, such that
the closed loop system is uniformly convergent. The uniform convergence can
then be used for tracking control purposes, synchronisation etc. In Section 8.3
we presented such a convergence-based tracking control design for a class of
measure differential inclusions in Lur’e form. Finally, we presented examples
of mechanical systems with set-valued force laws. In these examples, it has
been demonstrated that the tracking problem for a class of systems with non-
collocated actuation and set-valued friction can be solved using the results
presented in this chapter.
This page intentionally blank
9
Concluding Remarks

An introductory part on non-smooth analysis and measure theory followed

by a discussion of mechanical systems with frictional unilateral constraints
within a setting of measure differential inclusions, led finally to a Lyapunov
stability theory for non-smooth (Lagrangian) systems, being the very heart
of this monograph. Which conclusions can we draw? The most remarkable
and fundamental conclusion which we can give is, that we can indeed formu-
late a stability theory for non-smooth Lagrangian systems. It is sometimes
believed that the mathematical apparatus is not yet available to rigorously
deal with stability issues of non-smooth systems. The converse is true. Funda-
mental mathematical tools are non-smooth and convex analysis as well as the
extension of the differential measure to include an atomic part, which leads
to measure differential inclusions. Quintessential, however, is also the clear
formulation of the mechanical modelling in terms of set-valued constitutive
laws embedded in non-smooth potential theory. This last step requires in-
depth knowledge in analytical mechanics as well as the aforementioned math-
ematical concepts. The generalisation of the Lyapunov stability framework to
non-smooth Lagrangian systems follows in a natural way, but is certainly not
trivial.
The previous chapters demonstrate that considerable progress has been
made in the field of stability theory since the year 1644, in which Torricelli pos-
tulated his axiom which gave the onset to stability theory. In fact, with respect
to the stability theory for smooth dynamical systems, this monograph only
reveals the tip of the iceberg. Modern systems and control theory deals with
stability-related topics such as input-output stability, input-to-state stability,
absolute stability and convergent dynamics to name a few. Only a limited
number of these stability-related issues have been studied in this monograph.
The aim of this monograph is in a different direction. The previous chapters
try to formulate a mathematical framework for the stability analysis of non-
smooth systems within a setting of measure differential inclusions. The discus-
sion has therefore been confined to the most fundamental issues: the definition
of stability properties and stability analysis of equilibrium points and sets in
218 9 Concluding Remarks

autonomous measure differential inclusions using generalised Lyapunov-type

theorems. The chosen Lyapunov approach results in constructive proofs of
fundamental stability theorems, which paves the road towards the analysis of
other stability properties.
The chosen focus of this work has on the one hand given the opportunity to
discuss in-depth the basic philosophy behind non-smooth systems and stability
theory, but opened on the other hand a Pandora’s box of unresolved issues
that may lead to new research directions. Clearly, further research has to deal
with the stability analysis of other invariant sets such as periodic solutions in
autonomous as well as non-autonomous systems.
The scientific merit of this monograph can be sought in the assistance to
other theoretical or more practical fields of science. For instance, the theory of
bifurcations might benefit from the stability framework for non-smooth sys-
tems, developed in the previous chapters. Although great interest is currently
shown in the bifurcation analysis of non-smooth systems (mostly differen-
tial inclusions), no complete theory and classification of bifurcations in non-
smooth systems exists to date. Bifurcations of equilibrium points and sets,
as well as bifurcations of periodic solutions, are closely related to the gain or
loss of stability under variation of a system parameter. A clear definition of
stability, together with theorems to prove stability, is clearly of use for the
further study on bifurcations in non-smooth systems.
Lagrangian mechanical systems, being an important class of physical sys-
tems that can be modelled using measure differential inclusions, have been
given much attention in Chapters 5 and 7. Applications of the developed
stability framework in mechanics may vary from the stability analysis of his-
torical monuments to industrial applications such as brake systems, drilling
systems and railway vehicles. The special structure of Lagrangian systems
allows for a natural choice of the Lyapunov function and gives a physical in-
terpretation of the theoretical results. Notably, the set-valued force laws in
mechanics for unilateral frictional contact can be derived from non-smooth
(pseudo-)potentials. Electrical circuits with diodes and similar switching ele-
ments, as well as hydraulic networks with one-sided valves, are other examples
of engineering systems that can be modelled in the framework of non-smooth
potentials and measure differential inclusions. The presented stability results
therefore apply to a wide class of engineering systems. Naturally, specialised
theorems for specific classes of engineering systems can be deducted from
the results of Chapter 6 in the same way as has been done for Lagrangian
mechanical systems in Chapter 7.
Clearly, stability theory is not only of importance from the point of view of
dynamical systems analysis, but also plays a central role in control theory. The
design of stabilising controllers for non-smooth systems, such as mechanical
systems with unilateral contact, impact and friction (think of walking robots,
drilling systems etc.), is currently receiving wide attention. We believe that
the results in this monograph will prove to be useful in this respect. Moreover,
in Chapter 8, on the convergence properties of the class of monotone measure
 9 Concluding Remarks 219

differential inclusions, it has been shown how a symbiosis between non-smooth

dynamics and recent developments from systems and control theory can be
used to solve tracking control problems of mechanical systems with friction
or impulsive inputs.
The above applications of the presented theory to other fields in science
puts the work in a broader perspective and further justifies this monograph.
However, of equal importance is the theory of stability and non-smooth dy-
namics in its own right. Both research fields, that of stability theory and
non-smooth dynamics, put problems from different applied fields on an ab-
stract mathematical level, but both fields also play their own role in science.
Instead of considering quantitatively specific time-evolutions of reality, we are
urged by stability theory to consider the qualitative behaviour of models of
reality. Stability theory therefore leads science to a qualitative analysis of
systems. Stability theory might be seen as a branch of nonlinear dynamics,
which classically deals with the dynamics and bifurcations of differential and
difference equations. The possibility of inequalities is blatantly and obstinately
neglected in nonlinear dynamics. The relatively young field of non-smooth dy-
namics frees nonlinear dynamics from an established equality dogma and puts
(variational) inequalities on the foreground. Stability theory and non-smooth
dynamics therefore do not only sharpen the intellectual mind and catalyse
mathematical development, but have a profound academic value. As a closing
remark we would like to state that the preservation of academic values and
continuous development of theoretical knowledge must be of prime impor-
tance to contemporary university politics when striving for further scientific
success.
Sources and Translations

Various original texts have been studied while writing Section 1.2 of this
book in which some historical notes are given on the theory of stability. We
would like to thank the following libraries and institutions for their courtesy.
Furthermore, the authors would like to thank S. Pleines (ETH Zurich) and
Prof. em. F. Cerulus (K.U. Leuven) for their help with the translation of Latin
texts.

p.3 E. Torricelli, Opera Geometrica: Max Planck Institute for the History of
Science (digital library)
p.3 S. Stevin, Byvough der Weeghconst: Digital Library of the Royal Nether-
lands Academy of Arts and Sciences (KNAW).
Translation: The principal works of Simon Stevin, edited by E. Crone,
E. J. Dijksterhuis, R. J. Forbes, M. G. J. Minnaert and A. Pannekoek
p.4 C. Huygens, Œuvres Complètes de Christiaan Huygens: La Bibliothèque
Nationale de France/Gallica
p.4 D. Bernoulli, Commentationes de statu aequilibrii corporum humido insi-
dentium: ETH-Bibliothek Zurich (Alte Drucke).
Translation: F. Cerulus [31]
p.5 L. Euler, Scientia Navalis: The Euler Archive (digital library). Translation:
see [125]
p.5 J. L. Lagrange, Méchanique Analytique: ETH-Bibliothek Zurich (Alte
Drucke)
References

1. Acary, V., Brogliato, B., and Goeleven, D. Higher order Moreau’s

sweeping process: mathematical formulation and numerical simulation. Math-
ematical Programming, Series A (2006), pp. 85. submitted.
2. Adly, S. Attractivity theory for second order non-smooth dynamical systems
with application to dry friction. Journal of Mathematical Analysis and Appli-
cations 322 (2006), 1055–1070.
3. Adly, S., and Goeleven, D. A stability theory for second-order non-
smooth dynamical systems with application to friction problems. Journal de
Mathématiques Pures et Appliquées 83 (2004), 17–51.
4. Aeberhard, U., and Glocker, Ch. Energy considerations for excited perfect
collisions. Proceedings of the ENOC-2005 Conference, 7–12 August, Eindhoven
(2005), 422–431.
5. Alart, P., and Curnier, A. A mixed formulation for frictional contact
problems prone to Newton like solution methods. Computer Methods in Applied
Mechanics and Engineering 92 (1991), 353–375.
6. Ambrosio, L. Variational problems in SBV and image segmentation. Acta Ap-
plicandae Mathematicae: An International Survey Journal on Applying Math-
ematics and Mathematical Applications 17, 1 (1989), 1–40.
7. Angeli, D. A Lyapunov approach to incremental stability properties. IEEE
Transactions on Automatic Control 47, 3 (2002), 410–421.
8. Apostol, T. Mathematical Analysis. Addison Wessley, Reading MA, 1974.
9. Armstrong-Hélouvry, B., Dupont, P., and Canudas de Wit, C. A
survey of models, analysis tools and compensation methods for the control of
machines with friction. Automatica 30, 7 (1994), 1083–1138.
10. Aubin, J.-P., and Cellina, A. Differential Inclusions, vol. 264 of Grund-
lehren der mathematischen Wissenschaften. Springer-Verlag, Berlin, 1984.
11. Aubin, J.-P., and Ekeland, I. Applied Nonlinear Analysis. Pure and Applied
Mathematics. Wiley, Berlin, 1984.
12. Aubin, J.-P., and Frankowska, H. Set-valued Analysis, vol. 2 of Systems
and Control: Foundations and Applications. Birkhäuser, Boston, 1990.
13. Bainov, D. D., and Simeonov, P. S. Systems with Impulse Effect: Stability,
Theory and Applications. Wiley, New York, 1989.
14. Ballard, P. The dynamics of discrete mechanical systems with perfect uni-
lateral constraints. Archive for Rational Mechanics and Analysis 154 (2000),
199–274.
224 References

15. Barbashin, Y. A., and Krasovskii, N. N. The stability of motion as a

whole. Dokl. Akad. Nauk SSSR 86, 3 (1952), 453–456.
16. Basseville, S., Leger, A., and Pratt, E. Investigation of the equili-
brium states and their stability for a simple model with unilateral contact
and Coulomb friction. Archive of Applied Mechanics 73 (2003), 409–420.
17. Bernardin, F., Schatzman, M., and Lamarque, C. H. Second-order multi-
valued stochastic differential equations on Riemannian manifolds. Proceedings
of the Royal Society of London A 460 (2004), 3095–3121.
18. Bernoulli, D. Commentationes de statu aequilibrii corporum humido insid-
entium. Comment. Acad. Scient. Imp. Petrop. X (1738 (1747)), 147–163.
19. Bhat, S. J., and Bernstein, D. S. Finite-time stability of continuous au-
tonomous systems. SIAM Journal on Control and Optimization 38, 3 (2000),
751–766.
20. Birkhoff, G. D. Dynamical Systems. A.M.S. Publications, Providence, 1927.
21. Bourgeot, J. M., and Brogliato, B. Tracking control of complementarity
Lagrangian systems. International Journal of Bifurcation and Chaos 15, 6
(2005), 1839–1866.
22. Brézis, H. Operateurs Maximaux Monotones et Semi-groupes de Contrac-
tions dans les Espaces de Hilbert, vol. 5 of North-Holland Mathematics Studies.
North-Holland Publishing Company, Amsterdam, 1973.
23. Brogliato, B. Nonsmooth Mechanics, 2 ed. Springer, London, 1999.
24. Brogliato, B. Absolute stability and the Lagrange-Dirichlet theorem with
monotone multivalued mappings. Systems & Control Letters 51 (2004), 343–
353.
25. Brogliato, B., Daniilidis, A., Lemaréchal, C., and Acary, V. On the
equivalence between complementarity systems, projected systems and unilat-
eral differential inclusions. Systems and Control Letters 55 (2006), 45–51.
26. Brogliato, B., and Goeleven, D. The Krakovskii-LaSalle invariance prin-
ciple for a class of unilateral dynamical systems. Mathematics of Control,
Signals, and Systems 17, 1 (2005), 57–76.
27. Brogliato, B., Niculescu, S.-I., and Orhant, P. On the control of finite-
dimensional mechanical systems with unilateral constraints. IEEE Transac-
tions On Automatic Control 42, 2 (1997), 200–215.
28. Brogliato, B., Ten Dam, A. A., Paoli, L., Génot, F., and Abadie, M.
Numerical simulation of finite dimensional multibody nonsmooth mechanical
systems. ASME Applied Mechanics Reviews 55, 2 (2002), 107–150.
29. Cai, C., Teel, A. R., and Goebel, R. Smooth Lyapunov functions for hybrid
systems Part I: Existence is equivalent to robustness. IEEE Transactions on
Automatic Control (2007). Accepted.
30. Camlibel, M., Pang, J., and Shen, J. Lyapunov stability of complemen-
tarity and extended systems. SIAM Journal on Optimization 17, 4 (2006),
1056–1101.
31. Cerulus, F. private communication, to be published in vol. 6 of the ‘Werke
von Daniel Bernoulli’, 2007.
32. Chareyron, S. Stabilité des Systèmes Dynamiques Non-réguliers, Applica-
tion aux Robot Marcheurs. Ph.D. thesis, Institut National Polytechnique de
Grenoble, Grenoble, France, 2005.
33. Chareyron, S., and Wieber, P. B. A LaSalle’s invariance theorem for
nonsmooth Lagrangian dynamical systems. Proceedings of the ENOC-2005
Conference, 7–12 August, Eindhoven (2005).
 References 225

34. Chareyron, S., and Wieber, P. B. Stability and regulation of nonsmooth

dynamical systems. IEEE Transactions on Automatic Control (2006). submit-
ted.
35. Chetaev, N. G. The Stability of Motion. Pergamon, New York, 1961.
36. Clarke, F. H. Optimization and Nonsmooth Analysis. Wiley, New York,
1983.
37. Clarke, F. H., Ledyaev, Y. S., Stern, R. J., and Wolenski, P. R. Non-
smooth Analysis and Control Theory, vol. 178 of Graduate Texts in Mathemat-
ics. Springer, New York, 1998.
38. Coddington, E. A., and Levinson, N. Theory of Ordinary Differential
Equations. McGraw-Hill, New York, 1955.
39. Contensou, P. Couplage entre frottement de glissement et frottement de
pivotement dans la théorie de la toupie. In Kreiselprobleme; Gyrodynamics
(Berlin, 1963), H. Ziegler, Ed., Springer-Verlag, pp. 201–216. IUTAM Sympo-
sium Celerina 1962.
40. Curnier, A., Ed. Proc. Contact Mechanics International Symposium, Lau-
sanne, Switzerland, October 7-9, 1992 (Lausanne, 1992), Presses Polytech-
niques et Universitaires Romandes.
41. Curnier, A. Unilateral contact. In Pfeiffer and Glocker [141], pp. 1–68.
42. Davrazos, G. N., and Koussoulas, N. T. A review of stability results for
switched and hybrid systems. 9th Mediterranean Conference on Control and
Automation June 27-29, Dubrovnik, Croatia (2001).
43. Deimling, K. Multivalued Differential Equations. De Gruyter, Berlin, 1992.
44. Demidovich, B. P. Lectures on Stability Theory (in Russian). Nauka, Moscow,
1967.
45. Dijksterhuis, E. J. De Mechanisering van het Wereldbeeld. Meulenhoff,
Amsterdam, 2000, originally printed in 1950.
46. Eich-Soellner, E., and Führer, C. Numerical Methods in Multibody Dy-
namics. Teubner, Stuttgart, 1998.
47. Elstrodt, J. Maß- und Integrationstheorie. Springer, Berlin Heidelberg New
York, 1996.
48. Euler, L. Methodus inveniendi lineas curvas maximi minive proprietate gau-
dentes. Bousquet, Lausanne & Geneva, 1744.
49. Euler, L. Scientia navalis seu tractatus de construendis ac dirigendis, vol. 1
and 2. Academiae Scientarum, St. Petersburg, 1749.
50. Filippov, A. F. Differential equations with discontinuous right-hand side.
American Mathematical Society Translations, Series 2 42 (1964), 199–123.
51. Filippov, A. F. Differential Equations with Discontinuous Right-hand Sides.
Kluwer Academic, Dordrecht, 1988.
52. Fromion, V., Monaco, S., and Normand-Cyrot, D. Asymptotic properties
of incrementally stable systems. IEEE Transactions on Automatic Control 41
(1996), 721–723.
53. Fromion, V., Scorletti, G., and Ferreres, G. Nonlinear performance of
a PI controlled missile: an explanation. International Journal of Robust and
Nonlinear Control 9 (1999), 485–518.
54. Fuss, P. H. Correspondance mathématique et physique de quelques célèbres
géomètres du XVIIIème siècle, Tome II. No. 35 in The Sources of Science.
Johnson reprint corporation, New York London, 1968, originally printed in
St.-Pétersbourg 1843.
226 References

55. Génot, F., and Brogliato, B. New results on Painlevé paradoxes. European
Journal of Mechanics – A/Solids 18 (1999), 653–677.
56. Gamkrelidze, R. V. Analysis II, Convex Analysis and Approximation The-
ory, vol. 14 of Encyclopaedia of Mathematical Sciences. Springer-Verlag, New
York, 1989.
57. Glocker, Ch. Dynamik von Starrkörpersystemen mit Reibung und Stößen,
vol. 18, no. 182 of Fortschr.-Ber. VDI. VDI Verlag, Düsseldorf, 1995.
58. Glocker, Ch. The principles of d’Alembert, Jourdain and Gauß in non-
smooth mechanics, Part I: scleronomic multibody systems. Zeitschrift für
angewandte Mathematik und Mechanik 78, 1 (1998), 21–37.
59. Glocker, Ch. Discussion of d’Alembert’s principle for non-smooth unilateral
constraints. Zeitschrift für angewandte Mathematik und Mechanik 79, 1 (1999),
S91–S94.
60. Glocker, Ch. Formulation of spatial contact situations in rigid multibody
systems. Computer Methods in Applied Mechanics and Engineering 177 (1999),
199–214.
61. Glocker, Ch. Scalar force potentials in rigid multibody systems. In Pfeiffer
and Glocker [141], pp. 69–146.
62. Glocker, Ch. On frictionless impact models in rigid-body systems. Philo-
sophical Transactions of the Royal Society of London A 359 (2001), 2385–2404.
63. Glocker, Ch. Set-Valued Force Laws, Dynamics of Non-Smooth Systems,
vol. 1 of Lecture Notes in Applied Mechanics. Springer-Verlag, Berlin, 2001.
64. Glocker, Ch. Models of non-smooth switches in electrical systems. Interna-
tional Journal of Circuit Theory and Applications 33, 3 (2005), 205–234.
65. Goeleven, D., and Brogliato, B. Stability and instability matrices for lin-
ear evolution variational inequalities. IEEE Transactions on Automatic Control
49 (2004), 521–534.
66. Goeleven, D., and Brogliato, B. Necessary conditions of asymptotic sta-
bility for unilateral dynamical systems. Nonlinear Analysis 61, 6 (2005), 961–
1004.
67. Goeleven, D., Motreanu, D., Dumont, Y., and Rochdi, M. Variational
and Hemivariational Inequalities: Theory, Methods and Applications, Volume I:
Unilateral Analysis and Unilateral Mechanics, vol. 69 of Nonconvex Optimiza-
tion and its Applications. Kluwer Academic Publishers, Dordrecht, 2003.
68. Goeleven, D., Motreanu, D., Dumont, Y., and Rochdi, M. Varia-
tional and Hemivariational Inequalities: Theory, Methods and Applications,
Volume II: Unilateral Problems, vol. 70 of Nonconvex Optimization and its
Applications. Kluwer Academic Publishers, Dordrecht, 2003.
69. Goldstine, H. H. A History of the Calculus of Variations from the 17th
through the 19th Century, vol. 5 of Studies in the History of Mathematics and
Physical Sciences. Springer, New York, 1980.
70. Guckenheimer, J., and Holmes, P. Nonlinear Oscillations, Dynamical Sys-
tems, and Bifurcations of Vector Fields, vol. 42 of Applied Mathematical Sci-
ences. Springer-Verlag, New York, 1983.
71. Haddad, W., Chellaboina, V., and Nersesov, S. Impulsive and Hybrid
Dynamical Systems: Stability, Dissipativity, and Control. Princeton Series in
Applied Mathematics. Princeton University Press, Princeton, 2006.
72. Heemels, M. Linear Complementarity Systems: a Study in Hybrid Dynamics.
Ph.D. thesis, Eindhoven University of Technology, The Netherlands, 1999.
 References 227

73. Heertjes, M. F., Pastink, H. A., van de Wouw, N., and Nijmeijer, H.
Experimental frequency-domain analysis of nonlinear controlled optical storage
drives. IEEE Transactions on Control Systems Technology 14, 3 (2006), 389–
397.
74. Hespanha, J. P. Uniform stability of switched linear systems: Extensions of
LaSalle’s invariance principle. IEEE Transactions on Automatic Control 49, 4
(2004), 470–482.
75. Hespanha, J. P., Liberzon, D., and Teel, A. R. On input-to-state stability
of impulsive systems. In Proc. of the 44th IEEE Conf. on Decision and Control
and the European Control Conf. 2005 (Sevilla, 2005).
76. Hespanha, J. P., and Morse, A. S. Stability of switched systems with
average dwell-time. Proc. of the 38th Conf. on Decision and Control (1999),
2655–2660.
77. Hiriart-Urruty, J.-P., and Lemaréchal, C. Convex Analysis and Min-
imization Algorithms I, vol. 305 of Grundlehren der mathematischen Wis-
senschaften. Springer-Verlag, Berlin, 1993.
78. Hiriart-Urruty, J.-P., and Lemaréchal, C. Convex Analysis and Min-
imization Algorithms II, vol. 306 of Grundlehren der mathematischen Wis-
senschaften. Springer-Verlag, Berlin, 1993.
79. Huygens, C. De iis quae liquido supernatant libri tres, 1650, vol. 11 of Œuvres
Complètes de Christiaan Huygens. Martinus Nijhoff, The Hague, 1908.
80. Jacobi, C. G. J. Vorlesungen über Dynamik. (edited by A. Clebsch, reprinted
by Chelsea, New York, 1969), Berlin, 1866.
81. Jean, M., and Moreau, J. J. Unilaterality and dry friction in the dynamics
of rigid body collections. In Curnier [40], pp. 31–48.
82. Johansson, M. Piecewise Linear Control Systems - A Computational Ap-
proach, vol. 284 of Lecture Notes in Control and Information Sciences.
Springer-Verlag, Heidelberg, 2002.
83. Johnson, K. L. Contact Mechanics. Cambridge University Press, Cambridge,
1985.
84. Kane, T. R., and Levinson, D. A. Dynamics, theory and applications.
McGraw-Hill, New York, 1985.
85. Khalil, H. K. Nonlinear Systems. Prentice-Hall, New Jersey, 1996.
86. Klarbring, A. Mathematical Programming and Augmented Lagrangian
Methods for Frictional Contact Problems. In Curnier [40], pp. 409–422.
87. Kolmogorov, A., and Fomin, S. V. Introductory Real Analysis. Dover
Publications, New York, 1975.
88. Kunze, M., and Küpper, T. Qualitative bifurcation analysis of a non-smooth
friction oscillator model. Zeitschrift für angewandte Mathematik und Physik
48, 1 (1997), 87–101.
89. La Salle, J., and Lefschetz, S. Stability by Liapunov’s Direct Method.
Academic Press, New York, 1961.
90. Lagrange, J. L. Sur le mouvement des nœuds des orbites planétaires.
Mémoires de l’Académie Royale des Sciences et Belles-Lettres, Année 1774
(1776), 276–307.
91. Lagrange, J. L. Méchanique Analytique. La Veuve Desaint, Paris, 1788.
92. Lakshmikantham, V., Bainov, D. D., and Simeonov, P. S. Theory of Im-
pulsive Differential Equations, vol. 6 of Series in Modern Applied Mathematics.
World Scientific, Singapore, 1989.
228 References

93. Laplace, P. S. Sur l’équation séculaire de la lune. Mémoires de l’Académie

Royale des Sciences de Paris, Année 1786 (1788), 243–271.
94. Laursen, T. A., and Simo, J. C. Algorithmic symmetrization of Coulomb
frictional problems using augmented Lagrangians. Computer Methods in Ap-
plied Mechanics and Engineering 108 (1993), 133–146.
95. Le Saux, C., Leine, R. I., and Glocker, Ch. Dynamics of a rolling disk in
the presence of dry friction. Journal of Nonlinear Science 15, 1 (2005), 27–61.
96. Leine, R. I. Bifurcations in Discontinuous Mechanical Systems of Filippov-
Type. Ph.D. thesis, Eindhoven University of Technology, The Netherlands,
2000.
97. Leine, R. I. Bifurcations of equilibria in non-smooth continuous systems.
Physica D 223 (2006), 121–137.
98. Leine, R. I., Brogliato, B., and Nijmeijer, H. Periodic motion and bi-
furcations induced by the Painlevé paradox. European Journal of Mechanics –
A/Solids 21, 5 (2002), 869–896.
99. Leine, R. I., and Glocker, Ch. A set-valued force law for spatial Coulomb–
Contensou friction. European Journal of Mechanics – A/Solids 22 (2003),
193–216.
100. Leine, R. I., Glocker, Ch., and Van Campen, D. H. Nonlinear dynamics
and modeling of various wooden toys with impact and friction. Journal of
Vibration and Control 9, 1 (2003), 25–78.
101. Leine, R. I., and Nijmeijer, H. Dynamics and Bifurcations of Non-Smooth
Mechanical Systems, vol. 18 of Lecture Notes in Applied and Computational
Mechanics. Springer Verlag, Berlin, 2004.
102. Leine, R. I., Van Campen, D. H., De Kraker, A., and Van den Steen, L.
Stick-slip vibrations induced by alternate friction models. International Jour-
nal of Nonlinear Dynamics and Chaos in Engineering Systems 16, 1 (1998),
41–54.
103. Leine, R. I., and van de Wouw, N. Stability properties of equilibrium sets
of nonlinear mechanical systems with dry friction and impact. International
Journal of Nonlinear Dynamics and Chaos in Engineering Systems (2007).
104. Leine, R. I., and van de Wouw, N. Uniform convergence of monotone
measure differential inclusions: with application to the control of mechanical
systems with unilateral constraints. International Journal of Bifurcation and
Chaos (2007). accepted.
105. Lejeune-Dirichlet, J. P. G. Über die Stabilität des Gleichgewichts.
CRELLE, Journal für die reine und angewandte Mathematik 32 (1846), 85–88.
106. Liapounoff, A. Problème général de la stabilité du mouvement. Annales de
la faculté des sciences de Toulouse, 2e série 9 (1907), 203–474.
107. Liberzon, D. Switching in Systems and Control. Birkhäuser, Boston, 2003.
108. Lohmiller, W., and Slotine, J.-J. E. On contraction analysis for nonlinear
systems. Automatica 34 (1998), 683–696.
109. Lyapunov, A. M. Problème général de la stabilité du mouvement. No. 17 in
Annals of Mathematics Studies. Princeton University Press, Princeton, 1947.
110. Magnus, K. Kreisel; Theorie und Anwendungen. Springer-Verlag, Berlin,
1971.
111. Matveev, A. S., and Savkin, A. V. Qualitative Theory of Hybrid Dynamical
Systems. Control Engineering Series. Birkhäuser, Boston, 2000.
112. Maxwell, J. C. On governors. Proceedings of the Royal Society of London
16 (1868), 270–283.
 References 229

113. Menini, L., and Tornambè, A. Asymptotic tracking of periodic trajectories

for a simple mechanical system subject to nonsmooth impacts. IEEE Trans-
actions on Automatic Control 46, 7 (2001), 1122–1126.
114. Mihajlović, N. Torsional and Lateral Vibrations in Flexible Rotor Systems
with Friction. Ph.D. thesis, Eindhoven University of Technology, The Nether-
lands, 2005.
115. Mihajlović, N., van Veggel, A. A., van de Wouw, N., and Nijmeijer,
H. Analysis of friction-induced limit cycling in an experimental drill-string
system. ASME Journal of Dynamic Systems, Measurements and Control 126,
4 (2004), 706–720.
116. Monteiro Marques, M. Differential Inclusions in Nonsmooth Mechanical
Systems. Birkhäuser, Basel, 1993.
117. Moreau, J. J. Rafle par un convexe variable (première partie). Séminaire
d’Analyse Convexe, exposé no 15, University of Montpellier (1971), 43 pp.
118. Moreau, J. J. Rafle par un convexe variable (deuxième partie). Séminaire
d’Analyse Convexe, exposé no 3, University of Montpellier (1972), 36 pp.
119. Moreau, J. J. Problème d’évolution associé à un convexe mobile d’un es-
pace hilbertien. Comptes Rendus de l’Academie des Sciences, Séries A-B 276
(1973), 791–794.
120. Moreau, J. J. Evolution problem associated with a moving convex set in a
Hilbert space. Journal of Differential Equations 26 (1977), 347–374.
121. Moreau, J. J. Dynamique de systèmes à liaisons unilatérales avec frottement
secéventuel; essais numériques. Tech. Rep. 85-1, LMGC, Montpellier, France,
1986.
122. Moreau, J. J. Bounded variation in time. In Topics in Nonsmooth Mechanics,
J. J. Moreau, P. D. Panagiotopoulos, and G. Strang, Eds. Birkhäuser Verlag,
Basel, Boston, Berlin, 1988, pp. 1–74.
123. Moreau, J. J. Unilateral contact and dry friction in finite freedom dynamics.
In Non-Smooth Mechanics and Applications, J. J. Moreau and P. D. Pana-
giotopoulos, Eds., vol. 302 of CISM Courses and Lectures. Springer, Wien,
1988, pp. 1–82.
124. Moreau, J. J. Jump functions of a real interval to a Banach space. Annales
de la faculté des sciences de Toulouse 5e série, (suppl.) S10 (1989), 77–91.
125. Nowacki, H. Leonhard Euler and the Theory of Ships. Max Planck Society -
eDocument Server, http://edoc.mpg.de/ ID: 314833.0, 2007.
126. Panagiotopoulos, P. D. A nonlinear programming approach to the unilat-
eral contact-, and friction-boundary value problem in the theory of elasticity.
Ingenieur Archiv 44 (1975), 421–432.
127. Panagiotopoulos, P. D. Hemivariational Inequalities. Springer, Berlin,
1993.
128. Panagiotopoulos, P. D., and Glocker, Ch. Inequality constraints with
elastic impacts in deformable bodies. The convex case. Archive of Applied
Mechanics 70 (2000), 349–365.
129. Papastavridis, J. G. Analytical Mechanics: A Comprehensive Treatise on the
Dynamics of Constrained Systems: for Engineers, Physicists and Mathemati-
cians. Oxford University Press, New York, 2002.
130. Pavlov, A., Pogromsky, A., van de Wouw, N., and Nijmeijer, H. On
convergence properties of piecewise affine systems. International Journal of
Control (2007). Accepted.
230 References

131. Pavlov, A., van de Wouw, N., and Nijmeijer, H. Convergent systems:
Analysis and design. In Control and Observer Design for Nonlinear Finite and
Infinite Dimensional Systems (Stuttgart, 2005), T. Meurer, K. Graichen, and
D. Gilles, Eds., vol. 332 of Lecture Notes in Control and Information Sciences,
pp. 131–146.
132. Pavlov, A., van de Wouw, N., and Nijmeijer, H. Uniform Output Reg-
ulation of Nonlinear Systems: A Convergent Dynamics Approach. Systems &
Control: Foundations and Applications (SC) Series. Birkhäuser, Boston, 2005.
133. Pavlov, A., van de Wouw, N., and Nijmeijer, H. Frequency response
functions for nonlinear convergent systems. IEEE Transactions on Automatic
Control 52, 6 (2007), 1159–1165.
134. Pavlov, A., van de Wouw, N., and Nijmeijer, H. Global nonlinear output
regulation: convergence-based controller design. Automatica 43, 3 (2007), 456–
463.
135. Percivale, D. Uniqueness in the elastic bounce problem I. Journal of Differ-
ential Equations 56 (1985), 206–215.
136. Percivale, D. Uniqueness in the elastic bounce problem II. Journal of Dif-
ferential Equations 90 (1991), 304–315.
137. Pereira, F. L., and Silva, G. N. Lyapunov stability of measure driven
impulsive systems. Differential Equations 40, 8 (2004), 1122–1130.
138. Pettersson, S., and Lennartson, B. Hybrid system stability and robust-
ness verification using linear matrix inequalities. International Journal of Con-
trol 75, 16–17 (2002), 1335–1355.
139. Pfeiffer, F., and Glocker, Ch. Multibody Dynamics with Unilateral Con-
tacts. Wiley, New York, 1996.
140. Pfeiffer, F., and Glocker, Ch., Eds. Proc. IUTAM Symposium on Uni-
lateral Multibody Contacts, Munich, Germany, August 3–7, 1998 (Dordrecht,
1998), Kluwer Academic Pulishers.
141. Pfeiffer, F., and Glocker, Ch., Eds. Multibody Dynamics with Unilateral
Contacts, vol. 421 of CISM Courses and Lectures. Springer, Wien, 2000.
142. Pfeiffer, F., and Hajek, M. Stick-slip motion of turbine blade dampers.
Philosophical Transactions of the Royal Society A 338 (1992), 503–517.
143. Pliss, V. A. Nonlocal Problems of the Theory of Oscillations. Academic Press,
London, 1966.
144. Poincaré, H. Les Méthodes Nouvelles de la Mécanique Céleste, Tome I, II,
III. Gautier-Villars, Paris, 1892, 1893, 1899.
145. Poisson, S. D. Mémoire sur les inégalités séculaires des moyens mouvements
des planètes. Journal de l’École Polytechnique XV, 1 (1808).
146. Quittner, P. An instability criterion for variational inequalities. Nonlinear
Analysis 15, 12 (1990), 1167–1180.
147. Rockafellar, R. T. Convex Analysis. Princeton Landmarks in Mathematics.
Princeton University Press, Princeton, New Jersey, 1970.
148. Rockafellar, R. T. The Theory of Subgradients and its Applications to
Problems of Optimization. Convex and Nonconvex Functions, vol. 1 of Research
and Education in Mathematics. Heldermann Verlag, Berlin, 1970.
149. Rockafellar, R. T., and Wets, R.-B. Variational Analysis. Springer,
Berlin, 1998.
150. Rudin, W. Real and Complex Analysis. McGraw-Hill, New York, 1987.
 References 231

151. Sanfelice, R. G., Goebel, R., and Teel, A. R. Invariance principles for hy-
brid systems with connections to detectability and asymptotic stability. IEEE
Transactions on Automatic Control (2007). Accepted.
152. Sastry, S. Nonlinear Systems: Analysis, Stability and Control, vol. 10 of
Interdisciplinary Applied Mathematics. Springer, New York, 1999.
153. Schatzman, M. A class of nonlinear differential equations of second order
in time. Nonlinear Analysis, Theory, Methods and Applications 2, 3 (1978),
355–373.
154. Schatzman, M. Uniqueness and continuous dependence on data for one di-
mensional impact problems. Mathematical and Computational Modelling 28
(1998), 1–18.
155. Simo, J. C., and Laursen, T. A. An Augmented Lagrangian treatment
of contact problems involving friction. Computers & Structures 42, 1 (1992),
97–116.
156. Smirnov, G. V. Introduction to the Theory of Differential Inclusions, vol. 41
of Graduate Studies in Mathematics. American Mathematical Society, Provi-
dence, 2002.
157. Sontag, E. D. On the input-to-state stability property. European Journal of
Control 1 (1995), 24–36.
158. Stevin, S. Wisconstighe Ghedachtenissen. Bouwensz, Leiden, 1608.
159. Stewart, D. E. Formulating measure differential inclusions in infinite dimen-
sions. Set-valued Analysis 8 (2000), 273–293.
160. Stiegelmeyr, A. Chimney dampers. In Pfeiffer and Glocker [140], pp. 299–
308.
161. Szábo, I. Geschichte der mechanischen Prinzipien und ihrer wichtigsten An-
wendungen, vol. 32 of Wissenschaft und Kultur. Birkhäuser, Basel, 1996.
162. Tornambè, A. Modeling and control of impact in mechanical systems: Theory
and experimental results. IEEE Transactions on Automatic Control 44, 2
(1999), 294–309.
163. Torricelli, E. Opera Geometrica. Florentiæ typis Amatoris Masse & Lau-
rentij de Landis, Firenze, 1644.
164. Truesdell, C. A. Essays in the History of Mechanics. Springer, Berlin, 1968.
165. Utkin, V. I. Equations of slipping regimes in discontinuous systems I. Au-
tomation and Remote Control 32 (1971), 1897–1907.
166. Utkin, V. I. Equations of slipping regimes in discontinuous systems II. Au-
tomation and Remote Control 33 (1972), 211–219.
167. Utkin, V. I. Sliding Modes in Control and Optimization. Springer-Verlag,
Berlin, 1992.
168. Utkin, V. I., Guldner, J., and Shi, J. Sliding Mode Control in Electrome-
chanical Systems. Taylor & Francis, London, 1999.
169. Van de Wouw, A., Pavlov, A., and Nijmeijer, H. Controlled synchroni-
sation of continuous PWA systems. In Group Coordination and Cooperative
Control (Trondheim, Norway, 2006), K. Y. Pettersen, J. T. Gravdahl, and
H. Nijmeijer, Eds., vol. 336 of Lecture Notes in Control and Information Sci-
ences, pp. 271–289.
170. Van de Wouw, N., and Leine, R. I. Attractivity of equilibrium sets of
systems with dry friction. International Journal of Nonlinear Dynamics and
Chaos in Engineering Systems 35, 1 (2004), 19–39.
232 References

171. Van de Wouw, N., Pastink, H. A., Heertjes, M. F., Pavlov, A. V.,
and Nijmeijer, H. Performance of convergence-based variable-gain control of
optical storage drives. Automatica (2007). Accepted.
172. Van de Wouw, N., and Pavlov, A. Output tracking control of PWA systems.
In Proc. of the 44th IEEE Conf. on Decision and Control and the European
Control Conf. 2006 (San Diego, 2006), pp. 2637–2642.
173. van den Berg, R. A., Pogromsky, A. Y., Leonov, G. A., and Rooda,
J. E. Design of convergent switched systems. In Group Coordination and
Cooperative Control, K. Y. Pettersen, H. Nijmeijer, and T. Gravdahl, Eds.,
vol. 336 of Lecture Notes in Control and Information Sciences. Springer, 2006.
174. Van der Schaft, A. J., and Schumacher, J. M. An Introduction to Hy-
brid Dynamical Systems, vol. 251 of Lecture Notes in Control and Information
Sciences, vol. 251. Springer, London, 2000.
175. Wang, Y. Dynamic modeling and stability analysis of mechanical systems
with time-varying topologies. ASME Journal of Mechanical Design 115 (1993),
808–816.
176. Willems, J. L. Stability Theory of Dynamical Systems. Thomas Nelson and
Sons Ltd., London, 1970.
177. Wolfsteiner, P. Dynamik von Vibrationsfördern, vol. 2, no. 511 of Fortschr.-
Ber. VDI. VDI Verlag, Düsseldorf, 1999.
178. Yakubovich, V. Matrix inequalities method in stability theory for nonlinear
control systems: I. absolute stability of forced vibrations. Automation and
Remote Control 7 (1964), 905–917.
179. Yakubovich, V. A., Leonov, G. A., and Gelig, A. K. Stability of Sta-
tionary Sets in Control Systems with Discontinuous Nonlinearities, vol. 14 of
Series on Stability, Vibration and Control of Systems. Series A. World Scien-
tific, Singapore, 2004.
180. Ye, H., Michel, A., and Hou, L. Stability theory for hybrid dynamical
systems. IEEE Transactions on Automatic Control 43, 4 (1998), 461–474.
Index

σ-algebra, 44 classical rolling friction, 95

closed orbit, 116
absolute continuity, 24, 48, 69, 100 closed set, 21
accumulation point, 10, 137 closure, 21
admissible set, 74, 127, 153 compact set, 22
admissible set of generalised coordi- complementarity system, 10
nates, 154 complementary energy function, 82
almost everywhere, 24, 25, 48, 69 compliant contact model, 137
altered (total) potential energy function, cone, 33
163 conjugate function, 35
anisotropic friction, 89 conjugate potential, 81, 88
asymptotic attractivity, 121 consistency, 74, 127
asymptotic stability, 120 constrained bar example, 184
atom, 44 contact geometry, 83
attractive stability, 121
contact pair, 83
attractivity, 120
contact point, 83
closed positively invariant set, 121
contingent cone, 34, 39, 75, 127
equilibrium point, 121
continuity, 47
augmented Lagrangian, 14
continuous dependence on initial
Barbashin-Krasovskii theorem, 130 condition, 118, 171, 199
bifurcations, 175, 218 contour friction, 95
bilateral constraint, 73, 99, 172, 184 control theory, 15, 199
bilinear form, 55 convergence, 124
block-on-floor example, 165 convergent dynamics, 12, 17, 191
bouncing ball example, 105, 134, 147, convex analysis, 33
163 convex function, 23
bounded set, 22 convex hull, 22, 27, 62
buckling instability, 175 convex image, 29
convex set, 22
characteristic function, 45 Coulomb friction, 87
Chetaev’s instability theorem, 147 anisotropic, 89
Christoffel symbols, 101 associated, 90, 163
class KR function, 110, 128 non-associated, 90
234 Index

spatial, 88 Fenchel-Moreau theorem, 35

unidirectional (planar), 87 Filippov system, 10, 61, 171
Coulomb-Contensou friction, 14, 90 Filippov’s convex method, 63
finite-time stability, 121
Delassus matrix, 156 force laws, 80
density function, 51, 54 friction, 66, 84, 100, 205, 206
differential equation, 59 friction ball, 92
differential inclusion, 10, 13, 60, 102, friction coefficient, 87, 89, 91, 96
164 friction disk, 91, 93
definition of solution, 63
differential measure, 113 generalised coordinates, 100
directional differential inclusion, 73 generalised differential (of Clarke), 26,
dissipation function, 79, 88, 160 80
dissipation index, 104, 157 generalised velocities, 100
distance to a convex set, 38 global attractivity, 120
drilling friction, see Coulomb-Contensou global dissipation index, 104, 160
friction graph, 23
dwell-time, 195, 197, 203, 204, 214 gyroscopic forces, 101

energetic consistency, 159, 176, 177, Hahn decomposition, 47

190 heteroclinic orbit, 189
sufficient conditions, 159 Hopf bifurcation instability, 175, 177
epigraph, 23, 39 Huygens, Christiaan, 3
equilibrium point, 104, 115 hybrid system, 10
attractivity, 120, 127
impact, 97, 102
instability, 147
impulsive actuation, 193, 195, 197, 202
non-attractivity, 147
incremental stability, 124, 143, 145, 167,
stability, 119, 127
199, 204
equilibrium position, 153
index sets, 102
instability, 174
indicator function, 35, 80, 86, 88, 127
stability, 162
inequality complementarity condition,
equilibrium set, 104, 116, 153
85, 98
attractivity, 122
instability, 16, 147, 174
instability, 150
invariant set, 114
non-attractivity, 150
stability, 122, 137 Jordan decomposition, 50
equilibrium set (Lagrangian), 105, 153
attractivity, 168, 173 kinetic energy, 154
bilateral constraints, 173
instability, 174, 176, 189 Lagrange stability, 7
stability, 166 Lagrange, Joseph-Louis, 5
Euler, Leonhard, 4 Lagrange-Dirichlet stability theorem, 6,
existence of solutions, 63 16, 126
exponential convergence, 192 LaSalle’s invariance principle, 16, 145,
168, 171
falling block example, 178 Lebesgue decomposition, 48, 54
feedback control, 200 Lebesgue integral, 45
feedforward control, 200 Lebesgue measure, see measure,
Fenchel’s inequality, 36 Lebesgue
 Index 235

Lebesgue negligible, 44, 52, 69, 71, 100 negative damping, 175
Lebesgue-Stieltjes integral, 46, 52 negative limit set, 117
Lebesgue-Vitali theorem, 69 negatively invariant set, 114
level set, 111, 115 non-autonomous systems, 109, 122
level surface, 111 non-conservative forces, 100, 164, 174,
limit set, 171 175, 177
linear complementarity problem, 10, 14 non-smooth analysis, 13, 21
linear matrix inequality, 16 non-smooth dynamical system, 8
linearly bounded, 60 non-smooth mechanics, 14
Lipschitz continuity, 50 non-smooth potential theory, 79, 86, 87
local attractivity, 120 non-uniqueness of solutions, 116
locally bounded variation, 25, 45, 47, normal cone, 34, 38, 82, 86, 92
100, 112
locally positive definite function, 110, one-way clutch, 172, 202, 209
154 orbital stability, 122
lower semi-continuity, 24, 32 orthogonality, 47
Lur’e type systems, 199
Lyapunov function, 130, 139, 153 periodic orbit, 116
Lyapunov stability theorems, 126 periodic solution, 116
Lyapunov theorems, 162 pivoting friction, see Coulomb-
Lyapunov’s direct method, 8, 127 Contensou friction
Lyapunov’s indirect method, 8 Poincaré stability, 7, 122
Lyapunov, Aleksandr Mikhailovich, 8 Poisson stability, 7
polar cone, 34
positive definite function, 110
maximal monotone, 15, 29, 65, 85, 88, positive definite matrix, 111
141, 167, 193 positive limit point, 117
measure, 44 positive limit set, 117
atomic, 48, 70 positively homogeneous, 36, 40, 103,
differential, 52, 71 161
diffuse, 44 positively invariant set, 114, 116, 171,
Dirac point, 53, 55, 70 195
Lebesgue, 44, 48 potential, see non-smooth potential
Lebesgue (differential), 52, 54 theory
Lebesgue-Stieltjes, 44, 47, 54 principle of maximal dissipation, 89
modulus, 50 proximal point, 37
real, 49 pure rolling, 87
signed, 46, 51
singular, 48 quadratic function, 111, 140, 154
Stieltjes, 52
measure differential inclusion, 10, 13, radially unbounded function, 110, 130
68, 99, 104, 112 Radon-Nykodým theorem, 48
solution of, 74 region of attraction, 131, 173
consistency, 74 restitution, 97, 102, 137, 161, 177
measure Newton-Euler equations, 99 rheonomic contact, 102
method of Lyapunov functions, 127 Riemann-Stieltjes integral, 52
monotone, 29, 75, 141 Riemann-Stieltjes sum, 52
multi-valued function, 29 rocking bar example, 180
multifunction, 29 rocking block example, 184
236 Index

rolling friction, 94 Stribeck effect, 66, 89, 206, 207

classical rolling friction, 96 strictly positive real, 201, 207
contour friction, 95 subderivative, 39, 113, 127, 134, 155
subdifferential, 28, 35, 40, 81
scleronomic contact, 102 support function, 36, 102, 103, 154
set-valued force law, 14, 80, 84, 90 switched system, 10
set-valued function, 28 symptotic attractivity, 121, 137, 147,
Sign function, 30 180, 188, 192
Signorini’s contact law, 84
Signorini-Fichera condition, 85 tangent cone, 35
single-valued function, 23 time-stepping method, 11, 188, 208
singular perturbations, 10 Torricelli, Evangelista, 3
smooth function, 23 total mechanical energy, 153, 162
solution in the sense of Filippov, 64 total potential energy, 154, 163
special functions of bounded variation, tracking control, 199, 202
71
stability uniform convergence, 192
history of, 3 unilateral constraint, 73, 100, 168
incremental, 124, 143, 167 unilateral contact, 84
Lagrange, 7 unilateral primitive, 30
Lyapunov, 119 uniqueness of solutions, 65, 75
non-smooth systems, 15 upper semi-continuity, 23, 31, 63
Poincaré, 7, 122
Poisson, 7 variation, 24, 47
step function, 45, 48, 49, 53
stick-slip motion, 68 zero set, 161