SPRINGER BRIEFS IN MATHEMATICS

Vadim Utkin · Alex Poznyak ·

Yury V. Orlov · Andrey Polyakov

Road Map for

Sliding Mode
Control Design
SpringerBriefs in Mathematics

Series Editors
Nicola Bellomo, Torino, Italy
Michele Benzi, Pisa, Italy
Palle Jorgensen, Iowa, USA
Tatsien Li, Shanghai, China
Roderick Melnik, Waterloo, Canada
Otmar Scherzer, Linz, Austria
Benjamin Steinberg, New York, USA
Lothar Reichel, Kent, USA
Yuri Tschinkel, New York, USA
George Yin, Detroit, USA
Ping Zhang, Kalamazoo, USA

SpringerBriefs in Mathematics showcases expositions in all areas of mathematics

and applied mathematics. Manuscripts presenting new results or a single new result
in a classical field, new field, or an emerging topic, applications, or bridges between
new results and already published works, are encouraged. The series is intended for
mathematicians and applied mathematicians.
BCAM SpringerBriefs
Editorial Board
Enrique Zuazua
Deusto Tech
Universidad de Deusto
Bilbao, Spain
and
Departamento de Matemáticas
Universidad Autónoma de Madrid
Cantoblanco, Madrid, Spain

Irene Fonseca
Center for Nonlinear Analysis
Department of Mathematical Sciences
Carnegie Mellon University
Pittsburgh, USA

Juan J. Manfredi
Department of Mathematics
University of Pittsburgh
Pittsburgh, USA

Emmanuel Trélat
Laboratoire Jacques-Louis Lions
Institut Universitaire de France
Université Pierre et Marie Curie
CNRS, UMR, Paris

Xu Zhang
School of Mathematics
Sichuan University
Chengdu, China

BCAM SpringerBriefs aims to publish contributions in the following disciplines: Applied

Mathematics, Finance, Statistics and Computer Science. BCAM has appointed an Editorial Board,
who evaluate and review proposals.
Typical topics include: a timely report of state-of-the-art analytical techniques, bridge between new
research results published in journal articles and a contextual literature review, a snapshot of a hot
or emerging topic, a presentation of core concepts that students must understand in order to make
independent contributions.
Please submit your proposal to the Editorial Board or to Francesca Bonadei, Executive Editor
Mathematics, Statistics, and Engineering: [email protected].

More information about this series at http://www.springer.com/series/10030

Vadim Utkin Alex Poznyak
• •

Yury V. Orlov Andrey Polyakov

•

Road Map for Sliding Mode

Control Design

123
Vadim Utkin Alex Poznyak
Ohio State University Control Automatico Dpto
Columbus, OH, USA Cinvestav-Inst Politécnico Nacional
Mexico, Distrito Federal, Mexico
Yury V. Orlov
Electronics and Telecommunications Andrey Polyakov
CICESE Research Center NON-A team of Inria in Lille
Ensenada, Baja California, Mexico Inria Lille-Nord Europe
Villeneuve-d’Ascq, France

ISSN 2191-8198 ISSN 2191-8201 (electronic)

SpringerBriefs in Mathematics
ISBN 978-3-030-41708-6 ISBN 978-3-030-41709-3 (eBook)
https://doi.org/10.1007/978-3-030-41709-3
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, 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.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface

This book reviews the basic ideas of sliding mode control (SMC) theory and
demonstrates the deep interconnection of its different elements. The scope of the
book is broad, encompassing the main design principles applied to a wide range of
problems in control, observation, and information processing. The book is oriented
toward engineers and theoreticians working in any area of control theory and
applications.
In control systems, SMC is a nonlinear control method that alters the dynamics
of a nonlinear system by application of a discontinuous control signal that forces the
system to slide along manifolds with desired motions in the system state space. The
state feedback control law is not a continuous function of time. Instead, such a
system can switch from one continuous structure to another based on the current
position in the state space. This explains why SMC stems from the so-called
variable structure control method. In the context of modern control theory, any
variable structure system, like a system under SMC, may be viewed as a special
case of dynamical systems, supplied by switching logic.
Over the course of the development of the sliding modes theory during the past
half-century, the range of both control problems and design methods to solve them
has expanded significantly. Despite their great diversity, the basic principles of the
SMC theory have remained unchanged since the creation of the necessary mathe-
matical tools. The main control methods in sliding modes allow decomposition
of the original system into independent subsystems of lower dimension and iden-
tification of the class of systems with sliding modes that are insensitive to various
types of uncertainty in the description of the control object. These properties pre-
determine the success of application of the sliding modes theory for a wide class of
dynamic objects: finite-dimensional models with scalar and vector inputs and
outputs, models with delay, infinite-dimensional (described by partial differential
equations) models, discrete-time models, stochastic models, etc.
The synthesis control procedure within the SMS is divided into two stages. In
the first stage, it is necessary to choose the manifold in the state space of the system
with the desired trajectories. This problem is equivalent to the design of a control in
the systems of reduced dimension, for which any appropriate method of control

v
vi Preface

theory is applicable. At the second stage, the control is selected to ensure the
existence of a sliding mode on this preselected manifold. This is also a problem of
reduced order with the same dimensions of the state and control vectors.
The task of providing a sliding mode is equivalent to the problem of stability
of the zero solution of a nonlinear differential equation with a discontinuous
right-hand side. The book provides a solution to this problem for systems of a fairly
general form based on the Lyapunov function method. In particular, the described
method allows us to reduce the well-known problems of root placement and
quadratic optimization to solve similar problems of reduced dimension. All these
properties are preserved when designing state observers for dynamical systems.
The main obstacle to the implementation of sliding modes is a phenomenon
called chattering, which is caused by the mismatch between the ideal model and the
real process. Within the power electronics literature, it is frequently referred to as
ripple. In order to effectively suppress the chattering effect, the method of adapting
the amplitude of the discontinuous signal and the harmonic cancellation method are
used. These methods are applicable to both scalar and vector control systems. Note
that in systems with scalar control, the sliding manifold is usually a surface in the
state space. The high-order sliding mode method assumes that the sliding manifold
can have a smaller dimension. Accordingly, the sliding mode equation has a smaller
dimension. However, the problem of enforcing the sliding mode becomes more
difficult as the dimension of control is less than the dimension of the system to be
stabilized.
Methods developed for time-continuous systems have proved unsuitable for
discrete-time systems because discontinuous control always leads to chattering even
within the framework of an ideal model. It turns out that the main property
of the SMC, namely, the achievement of the desired manifold in a finite time and
the further motion in it, can be saved by using a control as a continuous function
of the system state vector. This property has made it possible to develop methods
for the synthesis of SMC for discrete systems.
SMC synthesis admits the presence of free parameters (such as the amplitude of
discontinuous control and parameters of sliding manifold equations), which can be
used to significantly improve the accuracy and dynamics of control processes. Such
methods for designing adaptive control laws are applicable in systems with sliding
modes.
The use of SMC for infinite-dimensional systems required a revision of the basic
concepts of the theory. Questions of the mathematical description of solutions of
partial differential equations with a discontinuous right-hand side had hardly been
considered in the literature. Component-wise synthesis methods, developed for
finite-dimensional systems, proved to be unacceptable. One of the chapters in this
book is devoted to the solution of these problems, where distributed thermal pro-
cesses and flexible mechanical systems are considered as applications.
An attempt to extend the SMC theory to the class of systems with random
perturbations requires the use of the modern theory of stochastic systems. These
studies are at an initial stage, and the first results are also reflected in this book.
Preface vii

Finally, further perspectives of the research are outlined in the concluding

chapter.
The book reflects the consensus view of the authors regarding the current status
of SMC theory, as it emerged from preceding discussions over a long period. We
would like to underline that the authors are affiliated with different research centers
and that their positions at the beginning of the discussions were far from being
identical, as they are now. Finally, we would also like to point out that the book
reflects the views of four generations of colleagues with rich experience in the area,
the age difference between these generations being 10 years or more in each
instance.

Columbus, USA Vadim Utkin

Mexico, Mexico Alex Poznyak
Ensenada, Mexico Yury V. Orlov
Villeneuve-d’Ascq, France Andrey Polyakov
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Mathematical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 Definition of Sliding Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Solution Existence and Uniqueness Problem . . . . . . . . . . . . . . . 12
2.3 Filippov Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 Equivalent Control Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5 Sliding Mode Equations in Control Affine Systems . . . . . . . . . . 21
2.6 Regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3 Design Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.1 Sketch of the Design Procedure . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 Regular Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Reaching and Existence Conditions . . . . . . . . . . . . . . . . . . . . . 32
3.4 Decoupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.5 Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.6 Output Sliding Mode Control . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.7 Integral SMC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.7.1 Main Idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.7.2 Design Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.8 Unit Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.9 Second-Order Sliding Mode Control . . . . . . . . . . . . . . . . . . . . 41
3.9.1 Twisting Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.9.2 Super-Twisting Algorithm . . . . . . . . . . . . . . . . . . . . . . 44
3.10 Suboptimal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

ix
x Contents

4 Lyapunov Stability Tools for Sliding Modes . . . . . . . . . . . . . . . . . . 49

4.1 Finite-Time Stabilization: What is This? . . . . . . . . . . . . . . . . . . 50
4.2 Strict Lyapunov Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3 Finite-Time Stability and Homogeneity . . . . . . . . . . . . . . . . . . 53
4.4 Fixed-Time Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5 SM Observers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.1 Observers for Linear Time-Invariant Systems . . . . . . . . . . . . . . 61
5.2 Observers for Uncertain Systems . . . . . . . . . . . . . . . . . . . . . . . 63
5.2.1 Second-Order System . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2.2 Uncertain System of an Arbitrary Order . . . . . . . . . . . . 64
5.2.3 Analysis of the State Estimation Error Convergence
for Matched Disturbances . . . . . . . . . . . . . . . . . . . . ... 66
5.2.4 Differentiators . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 68
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 70
6 Chattering Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.1 What is Chattering? Basic Ideas . . . . . . . . . . . . . . . . . . . . . . . 73
6.2 Power Converters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
7 High-Order Sliding Mode Control . . . . . . . . . . . . . . . . . ........ 83
7.1 HOSM Definition and Its Properties . . . . . . . . . . . . . ........ 83
7.2 Simple Examples of Systems with HOSM Control . . ........ 84
7.3 HOSM in SISO Affine Systems in Regulation
and Tracking Problems . . . . . . . . . . . . . . . . . . . . . . ........ 85
7.3.1 Uncertain Plant . . . . . . . . . . . . . . . . . . . . . . ........ 85
7.3.2 Lyapunov Function for HOSM Controllers’
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . ........ 86
7.3.3 HOSM Controllers . . . . . . . . . . . . . . . . . . . . ........ 87
7.3.4 Discontinuous and Quasi-Continuous HOSM
Controllers for SISO Systems . . . . . . . . . . . . ........ 87
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ........ 89
8 Discrete-Time Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
8.1 Discretization of Continuous-Time Models . . . . . . . . . . . . . . . . 91
8.2 Definition of SM for Discrete-Time Systems . . . . . . . . . . . . . . 93
8.3 Behavior Under Uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . 96
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
9 Adaptive SMC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
9.1 The r-Adaptation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
9.2 The Dynamic Adaptation Based on the Equivalent
Control Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Contents xi

9.2.1 The Simple Motivating Example . . . . . . . . . . . . . . . . . . 101

9.2.2 Multidimensional Case . . . . . . . . . . . . . . . . . . . . . . . . . 104
9.3 Super-Twisting Control with Adaptation . . . . . . . . . . . . . . . . . 107
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
10 SMC in Infinite-Dimensional Systems . . . . . . . . . . . . . . . . . . . . . . . 109
10.1 Infinite-Dimensional Setting . . . . . . . . . . . . . . . . . . . . . . . . . . 109
10.2 Unit Control for Infinite-Dimensional Systems . . . . . . . . . . . . . 111
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
11 Open Problems in SMC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
11.1 Stability Problem of the System x_ ¼ A signð xÞ . . . . . . . . . . . . . 115
11.2 SMC in the Presence of Noise in Measurements . . . . . . . . . . . . 117
11.3 Finite-Time Stability of SMs in Infinite-Dimensional
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
11.4 A Little Fantasy: SMC for Systems Governed by Shift
Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
11.5 On Stochastic SM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
12 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
About the Authors

Vadim Utkin graduated from Moscow Power Institute (Dipl. Eng.) and received
Ph.D. and Doctor of Science degrees from the Institute of Control Sciences
(Moscow, Russia). He worked at the Institute of Control Sciences from 1960 to
1994, and in 1973 was appointed as Head of the Discontinuous Control Systems
Laboratory. Currently, he is a Professor at Ohio State University. He is one of the
originators of the concepts of variable structure systems and sliding mode control.
His application interests are control of power converters and electric drives,
robotics, and automotive control. He is the author or co-author of five books and
350 papers.

Alex Poznyak graduated from Moscow Physical Technical Institute (MPhTI) in

1970. He earned Ph.D. and Doctor of Science degrees from the Institute of Control
Sciences of the Russian Academy of Sciences in 1978 and 1989, respectively. From
1973 to 1993, he served first as a researcher and then as leading researcher at this
institute, before accepting a post as Full Professor (3-F) at the Center for Research
and Advanced Studies of the National Polytechnic Institute (CINVESTAV-IPN) in
Mexico, where for 8 years he was Head of the Automatic Control Department. He
has been the supervisor for 43 Ph.D. theses. He has published more than 240 papers
in different international journals and 14 books. His areas of interest are robust
nonlinear deterministic and stochastic control, identification theory, Markov pro-
cesses, and game theory with economic applications.

Yury V. Orlov is a Professor in the Electronics and Telecommunication

Department, Scientific Research and Advanced Studies Center of Ensenada,
Mexico. His research interests lie in the analysis and synthesis of discontinuous as
well as time-delay and distributed-parameter systems. He has authored or
co-authored about 250 journal and conference papers in the above areas as well as
five monographs. He is an Associate Editor of IEEE Transactions on Control
Systems Technology, of the International Journal of Robust and Nonlinear Control,
and of the IMA Journal of Mathematical Control and Information.

xiii
xiv About the Authors

Andrey Polyakov received his Ph.D. in Systems Analysis and Control from
Voronezh State University in 2005. Until 2010, he was an Associate Professor with
this university. In 2007–2008, he worked at CINVESTAV-IPN in Mexico. From
2010 to 2013, he was a Lead Researcher of the Institute of Control Sciences of the
Russian Academy of Sciences, and he then joined INRIA, Lille, France. His main
research interests include robust and nonlinear control. He has authored or
co-authored more than 150 papers.
Chapter 1
Introduction

Abstract The initial ideas for designing systems with a sliding mode were formu-
lated half a century ago for dynamic models in canonical space with scalar input
and output. The interest in the behavior analysis in the space of the output variable
and its time derivatives was explained by the property of invariance with respect to
perturbations and parametric variations. This invariance property is lost for arbitrary
space in systems with vector input and output, which predetermined the need for
new mathematical methods for the analysis of such systems and design of control
methods with sliding modes. The set of these questions is the content of this book.

Keywords Sliding mode · Canonical space · Perturbations and parametric

variations

The multi-branched tree of Sliding Mode Control (SMC) theory was seeded more
than 50 years ago. At the very beginning, only SISO systems in the canonical space
were studied. The interest in the space of output signal and its time derivatives can
be explained easily: the sliding mode (SM) equation depends on neither parameter
variations nor external disturbances (see Fig. 1.1).
The example in Fig. 1.1 demonstrated that the motion in sliding mode is of a
reduced order and depends on the switching line equation only. It looks attractive
to utilize these properties for control at the presence of disturbances. We compare
the second-order systems with SMC and PID controller. As shown in Fig. 1.2, after
a finite time interval, the sliding mode occurs (since the distance to the switching
line s becomes equal to zero identically); the output does not depend on disturbance
and tends to zero exponentially. Parameters of PID controller are selected such that
the transient time for both processes are the same. The output of the system with
PID controller cannot be reduced to zero (Fig. 1.3). An integral component in PID
controller is able to reject constant disturbances only. The system with SMC is
equivalent to the system with a high feedback gain, implemented by finite control
actions. Therefore, an integral component is not needed for SMC, and steady-state
error is equal to zero identically in contrast to systems with a linear PID controller.
SMC design method can be generalized easily for systems of an arbitrary order
in canonical form
x (n) + an x (n−1) + · · · + a1 x = u + d,
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 1
V. Utkin et al., Road Map for Sliding Mode Control Design,
SpringerBriefs in Mathematics,
https://doi.org/10.1007/978-3-030-41709-3_1
2 1 Introduction

Fig. 1.1 Sliding mode in the second-order system

Fig. 1.2 ẍ = u + sin(3t), SMC: u = −3 sign (s), s = ẋ + x


n
u = −Msign(s), s = Ci x (i−1) , Cn = 1.
i=1

Similar to the second-order system, sliding mode is enforced on the switching plane
s = 0, governed by equation
n
Ci x (i−1) = 0.
i=1
1 Introduction 3

Fig. 1.3 ẍ = u + sin(3t), PID controller: u(s) = −(7 + 3.5/s + 4.5s)x(s)

This motion depends on neither parameters ai nor disturbance d. SMC in canonical

space was studied practically in all chapters of [1]. SMC in the canonical space needs
high-order time derivatives, and their ideal implementation is hardly possible. As a
rule, an output of any real differentiator is approximation of operator d/dt, which
leads to undesirable high-frequency oscillations, called the chattering. Chattering is
recognized as the main obstacle for implementation of SMC.
Sliding mode in the canonical space does not depend on uncertainties. Unfortu-
nately, this property is not inherent in sliding modes in an arbitrary state space. The
equation of a switching plane is not the equation of the sliding mode simultaneously.
The question of identifying a sliding mode equation in an arbitrary state space is not
trivial and needs special study. All above mentioned problems have been studied in
further (after [1]) publications and discussed in this book.
Since then the range of design methods and control problem statements have been
increased considerably. Now the scope of research embraces the design in an arbitrary
state space, MIMO systems, the observation problems, and mathematical methods
of analysis. Publication of survey papers in the 70–80s [2, 3], embracing the scope
of research in the area, stimulated proliferation of SMC methodology into research
programs of control theorists and engineers in many countries. As a result the sci-
entific arsenal, accumulated in the theory of SMC during more than 50 years, offers
a wide range of both problem statements and methods for their solutions. So wide
that any attempt to prepare a “State-of-the-Art” survey paper would hardly be suc-
cessful. Of course, excellent surveys were published about different aspects on SMC
(chattering problems, adaptive control, high-order sliding mode, applications). The
4 1 Introduction

list can be complemented by control in infinite-dimensional and stochastic systems,

generalization for discrete-time systems and integral control.
We are realistic and far from the idea of preparing a survey embracing the results
of numerous publications in those new areas. Our objective is more modest; we try
to describe the basic ideas of the SMC methodology and their development in a
historical perspective, not to describe only but to demonstrate deep interconnection
between so-called “conventional” and “new” methods.1 In the course of the SMC
history, the original ideas were modified in the context of new control problems, but
none of them was recognized as obsolete and replaced by a totally new idea. The
tracks of the main concepts formulated in “childhood” of sliding mode control can
be readily observed in any modern research. In conclusion, the further developments
are outlined. The recent results are further development of the “conventional” SMC
theory rather than an attempt to replace it. The accent in all sections is made on
revealing the reasons why new developments were needed and why their roots can
be found easily in the “conventional” methods. The purposes of the book are
• to inform the readers on a wide range of control problems successfully solved in
the framework of the SMC theory;
• to demonstrate that the background of all methods is the same: selection of a
reduced-order differential equation with desired solutions and implementation of
the selected motion by enforcing sliding mode;
• to demonstrate the potential and advantages of the developed control methods and
outline the areas of their efficient applications.
The book is oriented toward a wide range of engineers and theoreticians working
in any area of the control theory and applications. All chapters of the book are pre-
pared in this context. The scope embraces control of finite- and infinite-dimensional
processes with continuous- and discrete-time control, chattering suppression prob-
lems, and new methods of adaptation applicable for systems with sliding motions
only. Special mathematical methods are needed for all listed control tasks. The meth-
ods constitute the initial chapters and include the definition of the multidimensional
SM, the derivation of the differential equations of those motions, and the existing
conditions. The book also discusses areas of further research.
In our view, it is better to avoid separation of past and current times but establish
bridges between them.

1 We follow the terminology: “conventional” means the first-order sliding mode, “new” means the
high-order sliding mode. The term “conventional SMC” appeared in Shtessel et al. (2014), for the
first time. The details will be discussed in the subsequent chapters.
References 5

References

1. Emelyanov, S. (ed.): Theory of Variable Structure Control Systems (in Russian). Nauka (1970)
2. Utkin, V.: Variable structure systems with slidnng mode control. IEEE Trans. Autom. Control
22(2), 212–221 (1977)
3. DeCarlo, R., Zak, S., Matthews, G.: Variable structure control of nonlinear multivariable systems:
a tutorial. Proc. IEEE 76(3), 212–232 (1988)
Chapter 2
Mathematical Methods

Abstract This chapter presents the definition of sliding mode and discusses the con-
ditions of the existence and the uniqueness of the solution of the arising differential
inclusion. The definition of the Filippov solution is introduced and discussed. The
equivalent control method (ECM) is presented as an alternative procedure for deriv-
ing the sliding mode equation. Several types (Filippov, Utkin) of the regularization
methods are discussed and illustrated by the class of affine systems.

Keywords Sliding mode · Filippov solution · Regularization methods

2.1 Definition of Sliding Mode

State space representation is the conventional framework for analysis and design
of sliding mode control systems. A plant model in this case is described by some
ordinary differential equations (ODEs)

ẋ(t) = f (t, x(t), u(t, x(t))), t ∈ R, (2.1)

where x(t) ∈ R n is the vector of system states, u(t) ∈ R m is the vector of control
inputs, and f is a function. For simplicity, we assume that f is continuous, but the
control u, as a function of time and state, can be discontinuous in its arguments but
locally bounded. Later on, we discuss different definitions of a solution to the discon-
tinuous ODE (2.1). Here, we just assume that in a domain G ⊆ R n+1 the closed-loop
system (2.1) has absolutely continuous solutions, which are differentiable almost
everywhere and the derivatives are locally bounded.
The main idea of sliding mode control consists of enforcing the trajectories of
closed-loop system to belong to a prescribed manifold (sliding manifold). The man-
ifold has to be selected in advance in order to fulfill the required performance of the
closed-loop system. Next, the control law has to be designed to enforce the sliding
mode (see details in Chap. 4).

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 7

V. Utkin et al., Road Map for Sliding Mode Control Design,
SpringerBriefs in Mathematics,
https://doi.org/10.1007/978-3-030-41709-3_2
8 2 Mathematical Methods

To introduce a definition of sliding mode for a finite-dimensional model (2.1), we

try to find a minimal set of properties characterizing it. The finite-time convergence to
a sliding set is a specific feature of the finite-dimensional sliding mode systems [1–4].
That is impossible for ODEs with Lipschitzian right-hand sides. The discontinuity
of the sliding mode control provides some robustness properties for the closed-loop
system such as invariance with respect to the matched perturbations [5].

Definition 2.1 Let S be a r -dimensional manifold in the state space, r < n. An

open non-empty subset D ⊆ S is said to be sliding domain if it is locally uniformly
finite-time attractive, that is, for any y ∈ D there exists ε > 0 such that, if

x(t0 ) = x0 , x0 − y < ε,

then T (x0 ) can be found such that

x(T (x0 )) ∈ D and lim T (x0 ) = 0,

x0 →y

where x is a solution of the closed-loop system (2.1) with the initial condition x(t0 ) =
x0 .

If D = S, then D is called sliding manifold (Fig. 2.1). The conventional description

of sliding manifold is
S = {x ∈ R n : s(x) = 0},

where s is a smooth function that maps R n to R r. For ease of reference, it is often

written as s(x) = 0 or simply s = 0. Our definition based on finite-time convergence
cannot be called new. Similar definition was introduced 25 years ago in [1], see also
Chap. 4 in [3].
The definition does not imply discontinuities in control. As a result, the term “slid-
ing mode” proved to be applicable for discrete-time systems. For several decades,
starting from 50s the term “sliding mode” was associated with discontinuities in
motion equations and with trajectories in the discontinuities surface. Attempts to
extend it for other classes of systems based on this view were not successful. For
example, discontinuities in discrete-time systems resulted in chattering. An alterna-
tive approach is to make behavior of discrete-time systems similar to continuous-
time ones so that an integral manifold exists with finite reaching time. This approach
proved to be appropriate for discrete-time systems, when sliding mode can appear, if
right-hand side in motion equation is continuous (similar to dead-beat control [6]),
as well as for continuous-time systems with continuous non-Lipschitz functions. In
nutshell, sliding mode does not imply discontinuities always. Another interpretation
of the systems with sliding modes was offered in 1953 by Pontryagin. He said, a
system with “glued” trajectories does not have an inverse shift operator, and it is
even not a dynamic system. We follow this concept and in terms of our definition, a
sliding mode exists in the continuous-time system with a square root in the equation
(see [7]). The finite-time stable scalar system
2.1 Definition of Sliding Mode 9

Fig. 2.1 Sliding mode in discontinuity surfaces and their intersections


ẋ(t) = − |x(t)|sign(x(t))

is topologically equivalent to the system

ẏ(t) = −0.5sign(y(t)),

obtained by means of change of the coordinate


y= |x|sign(x).

After a finite time, y is identically nullified and as a result, the state x is as well in
spite of the continuous right-hand side in the original equation. Indeed, the coordinate
transformation is homeomorphism (continuously invertible) on R; moreover, it is
diffeomorphism (smooth) on R\{0}. No need to say, that continuous non-smooth
control resulting in SM is not applicable for rejection of additive disturbances, since
it is equal to zero on the desired sliding surface.
Minor remark: Excluding finite-time attractiveness property, we would certainly
admit that sliding mode exists in linear system. Widely used in literature inequality

s ṡ < 0 for s = 0

as the condition for sliding mode to exist does not imply finite-time attractiveness
property, as in the above definition. Accepting this property, we should admit that
sliding mode exists on the line s = 0 in the linear second-order system

ẍ = x, s = ẋ − x,
10 2 Mathematical Methods

Fig. 2.2 Illustration of the

situation with no sliding
mode in origin

since ṡ = −s and s ṡ < 0 if s = 0. One more example illustrating that condition

s ṡ < 0, if s = 0 is not sufficient for existence of sliding mode. Sliding mode occurs
on the switching line s = 0 with
s = ẋ + x

after finite time for any initial conditions from the open blue sector in the system (see
Fig. 2.2, where x1 = x and x2 = ẋ).

0.5x if x · s > 0
ẍ = .
1.5x if x · s < 0

Inequality s ṡ < 0 holds and s tends to zero for initial conditions on lines I and II;
however, the sliding mode does not exist in the origin; there are diverging trajectories
in any its vicinity.
For systems with scalar control undergoing discontinuities on surface s(x) = 0,
the sufficient conditions for sliding mode to exist [8, Chap. 2] are

lim ṡ < 0 and lim ṡ > 0. (2.2)

s→0+ s→0−

For example, the sector mn (Fig. 2.3) is a sliding domain for the second-order system

ẍ = u, u = −sign(s),
(2.3)
s = ẋ + cx, c > 0.

Note that x(t) tends to zero as a stable solution (see Fig. 2.3) to equation

s = ẋ + cx = 0,

and the state (x, ẋ) never reaches the boundary point m or n, should sliding mode
occurs. Sliding mode exists on segment mn of switching line

ẋ + cx = 0, c < 0
2.1 Definition of Sliding Mode 11

Fig. 2.3 Stable solution

Fig. 2.4 Unstable solution

(see Fig. 2.4), but the motion is unstable. It is important to note that the statements
about stability are based on heuristic arguments. The solution
∗
x(t) = e−c(t−t ) x(t ∗ )

is found from equation

ẋ + cx = 0.

Formally, it should satisfy equation

ẍ = u

or
∗
x(t ∗ )c2 e−c(t−t ) = −sign(0),

where t ∗ ∈ R is a time instant, when the trajectory reaches the sliding set. The
questions arise why the latter equality holds and what meaning is behind sign(0).
12 2 Mathematical Methods

The questions touch upon the fundamental problem—what the solution is, since the
conventional existence-uniqueness theorems [9] are not applicable for differential
equations with discontinuous right-hand sides. The above problem is treated in the
next section.
It is worth noticing that Definition 2.1 does not specify a class of dynamical
systems. It is applicable even for continuous non-Lipschitz ODEs and the so-called
high-order sliding mode (HOSM) systems [10], [4, Chap. 6] which are discussed in
Chap. 7 of this book.

2.2 Solution Existence and Uniqueness Problem

The simplest example of the sliding mode control system is

ẋ(t) = u(t) + d(t), u(t) = − sign(x(t)),

where the disturbance is bounded |d(t)| ≤ d0 < 1. Since

d
|x(t)| ≤ −(1 − d0 ) < 0
dt

for x(t) = 0, then there exists an instant of time

|x(t0 )|
T (x(t0 )) ≤ t0 +
1 − d0

such that
|x(t)| → 0 as t → T (x0 ) + 0− .

The mathematical obstruction appears if we try to define a solution for t > T (x(t0 ))
because classical concepts of solutions imply the existence of a Lipschitz constant
[9] and do not allow the right-hand side of ordinary differential equation to be discon-
tinuous. It is reasonable to assume that x(t) = 0 for t > T (x(t0 )). As in the example
of Sect. 2.1, the question arises again, why equality

−sign(0) = d(t)

holds. The unanswered question serves as an example, illustrating, why the conven-
tional theory is not applicable to discontinuous systems and why new mathematical
methods for discontinuous systems should be developed. The methods should answer
the fundamental questions of the classical theory of differential equations if a solu-
tion exists at points of discontinuities and if it exists, whether the solution is unique
or not.
The analysis is far from being trivial for multidimensional and nonlinear cases. It
may lead to solutions, which contradict engineering intuition.
2.2 Solution Existence and Uniqueness Problem 13

Fig. 2.5 Real sliding mode

for relay with hysteresis

Fig. 2.6 Phase shift for two

relay controls with hysteresis

Let us consider the system

⎫
ẋ1 (t) = u 1 (t), u 1 (t) = −sign(x1 (t)), ⎬
ẋ2 (t) = u 2 (t), u 2 (t) = −sign(x2 (t)), (2.4)
⎭
ẋ3 (t) = u 1 (t)u 2 (t).

The state variables in the first two equations and their time derivatives have opposite
signs; therefore, the sliding mode exists with x1 (t) = 0 and x2 (t) = 0. Assuming
u 1 (t) = 0 and u 2 (t) = 0 yields

ẋ3 (t) = 0 or x3 (t) = const.

If u 1 and u 2 are relay functions with hysteresis Δ (Fig. 2.5, v = u 1 or u 2 , s = x1

or x2 ), the average value of the product u 1 u 2 can take any value between +1 and
−1 depending on the phase shift ψ (Fig. 2.6). The high-frequency component in the
solution to the third equation is filtered out with Δ tending to zero. Therefore, x3 is
the solution to

ẋ3 (t) = A, −1 ≤ A ≤ 1, x3 = At + x3 (0)

(not constant as stated before).

If the equations for x1 and x2 are studied separately, we obtain the sliding motion on
the surfaces
S1 = {x = (x1 , x2 , x3 )T ∈ R2 : x1 = 0}
S2 = {x = (x1 , x2 , x3 )T ∈ R2 : x2 = 0},

respectively. One can define u i (t) = 0 for t > Ti (x(t0 )), where Ti (x(t0 )) is an instant
of time when the trajectory reaches the surface Si , i = 1, 2. In this case, ẋ3 (t) = 0
for t ≥ Ti (x(t0 )).
The proposed procedure provides the well-defined unique continuous solution to
the considered system. However, in practice, the relay control is realized with some
14 2 Mathematical Methods

defects, for example, delay and/or hysteresis. If the control inputs are replaced with

u ih (t) = −sign(xi (t − h)),

where h > 0, then the corresponding solution xih (t), i = 1, 2, 3 of the closed-loop
system with xih (t) = 0 for t ∈ [−h, 0) is well defined in the sense of Carathéodory.
Moreover, it can be shown that xih (t) → xi (t) as h → 0 uniformly on t for i = 1, 2,
but x3h (t) may not converge to x3 (t). Indeed, if

x1h (t0 ) = ±x2h (t0 ) = 0,

then
u 1h (t) = ±u 2h (t),
ẋ3h (t) = u 1h (t)u 2h (t) = ±1

for almost all t > 0. The last identity contradicts to

ẋ3 (t) = 0, t > T ∗ = max {T1 (x(t0 )), T2 (x(t0 ))}

derived in delay-free case. It can be shown that for any h > 0, the initial conditions can
1  t+4h h
be selected in such a way that the averaged value of time derivative ẋ3 (τ )dτ
∗
4h t
will take any value in (−1, 1) for t > T .

2.3 Filippov Solution

Filippov method [11, Chap. 2] is the classical tool to overcome the mathematical
obstructions of the discontinuous ODE. The main idea is to embed the ill-posed
ODE system into a well-posed differential inclusion, which coincides with ODE in
continuity domain. Let a time-varying feedback law u in (2.1) be a discontinuous
function of state variables. The Filippov method of solution continuation at any point
of discontinuity set N ⊆ R n of measure zero implies that the right-hand side of (2.1)
belongs to the convex hull of all vectors in its vicinity. It is important to note that
Filippov’s method has a very simple interpretation in the time domain. Let the right-
hand side of (2.1) take one of k possible values f˜1 ,…, f˜k in the vicinity of some
point (t0 , x0 ) in the state space, and a time interval Δt consists of non-overlapping k
subsets

k
Δt1 , Δt2 , . . . , Δtk : Δt = Δtk
i=1

with values of right-hand sides f˜1 , …, f˜k correspondingly. Then, the averaged speed
in this point satisfies
2.3 Filippov Solution 15

1
k k
Δti
ẋ(t) = lim Δti f˜i (t, x(t)), αi = , αi = 1.
Δt→0 Δt Δt
i=1 i=1

The right-hand side is nothing else but the convex hull of f˜1 ,…, f˜k . Therefore, αi can
be viewed as a specific time of the mode f˜i . The Filippov procedure considers the
following differential inclusion:

ẋ ∈ F(t, x(t)), t ∈ R,

provided that the set-valued function F : Rn+1 → 2R is defined as

n

F(t, x) = co f˜ (t, x + εB\N ) , (2.5)

ε>0 N : μ(N )=0

where
– f˜(t, x) = f (t, x, u(t, x)) is a right-hand side of the closed-loop system (2.1),
– B = {x ∈ R n : x ≤ 1} is the unit ball in R n ,
– co(M) is the convex hull of a set M,
– and the identity μ(N ) = 0 means that the set N is of measure zero.
The Filippov procedure is applicable for any locally measurable (possibly nowhere
continuous) function f˜. If it is known a priori that f˜ is discontinuous only on a set
N of the measure zero, then the second intersection can be excluded from (2.5).
Definition 2.2 An absolutely continuous function x is a solution of (2.1), if almost
everywhere it satisfies the inclusion

ẋ(t) ∈ F(t, x(t)), F(t, x) = lim Fε (t, x),
ε→0 (2.6)
Fε (t, x) = Conv{ f (t, x, u(t, {x + εB}\N ))},
where B is the ball centered in the origin of the unit radius, N ⊆ R n is the discon-
tinuity set of the measure zero, and the limit is defined by the Hausdorff distance,1
d(F(t, x), Fε (t, x)) → 0 as ε → 0.
Exclusion of a set of measure zero in (2.6) is the core condition when deriving the
sliding mode equation. Any attempts to assign the right-hand side on discontinuity

1 The Hausdorff distance between two non-empty sets S1 , S2 ⊆ R n is a number

d(S1 , S2 ) = max{d1 (S1 , S2 ), d2 (S1 , S2 )},

where
d1 (S1 , S2 ) = max min x1 − x2 
x 1 ∈S1 x 2 ∈S2
and
d2 (S1 , S2 ) = max min x1 − x2 
x 2 ∈S2 x 1 ∈S1
16 2 Mathematical Methods

manifold beyond the context of the system equation can meet contradictions. In many
publications ⎧
⎨ 1 if ρ > 0,
sign(ρ) = −1 if ρ < 0,
⎩
0 if ρ = 0,

or 
1 if ρ ≥ 0,
sign(ρ) =
−1 if ρ < 0,

which is not true for the value of sign(0) in real system. This fact is demonstrated in
the explanation for the system with friction. The friction force cannot be assigned,
if the external force is unknown. For example, in the mechanical system

m ẍ = g f r (t) − g(t), t > 0,

where x(t) ∈ R, the Coulomb friction

g f r (t) = −F0 sign(ẋ(t))

undergoes discontinuities on surface ẋ = 0, and sliding mode occurs on this surface if

F0 > |g(t)|. It is evident that g f r (t) = g(t), and this time function does not coincide
with an arbitrary assigned function (beyond the system context).
Let us write down sliding mode equation following the Filippov method for the
autonomous system (2.1) with scalar control, that is, discontinuous on a smooth
surface
S = {x ∈ R n : s(x) = 0, s ∈ C 1 },

which separates R n on two domains with

s(x) > 0, f (t, x, u(x)) = f + (t, x)

and with
s(x) < 0, f (t, x, u(x)) = f − (t, x),

where f + and f − (t, x) are some functions continuous on R n+1 .

For x ∈ S, the convex hull F(x) is the segment

μ f + (t, x) + (1 − μ) f − (t, x), 0 ≤ μ ≤ 1,

connecting vectors f + (t, x) and f − (t, x). The state velocity vector in sliding mode
should be in the plane tangential to the surface S at a point x ∈ S. If the existence con-
2.3 Filippov Solution 17

Fig. 2.7 Geometrical

illustration of Filippov
definition

ditions (2.2) hold, then this segment crosses the plane (see Fig. 2.7). The intersection
point defines the right-hand side in the sliding mode equation in surface s(x) = 0

ẋ(t) = f 0 (t, x(t)),

where
f 0 (t, x) = μ f + (t, x) + (1 − μ) f − (t, x)

and the parameter μ can be found from condition

∇s⊥[μ f + + (1 − μ) f − ].

Substituting the solution into the formula for f 0 results in

(∇ Ts f − ) f + + (∇ Ts f + ) f −
f0 = .
∇ Ts ( f + − f − )

If ∇s(x) is not orthogonal to μ f˜− (t, x) + (1 − μ) f˜+ (t, x) for every μ ∈ [0, 1], then
any trajectory of (2.6) comes through the surface resulting in an isolated “switching”
of the right-hand side of (2.1). Therefore, for the system with scalar control and two
possible vectors f + and f − , the right-hand side of the sliding mode equation should
belong to the convex hull of these two vectors.
Now let us consider again the example (2.4). From the first impression, u 1 (t) and
u 2 (t) should be equal to zero since x1 (t) and x2 (t) are equal to zero identically in
sliding mode and as a result

ẋ3 (t) = 0, x3 (t) = const.

However, both of them are the same pulse trains and depending on the phase shift
the average value of their product can take any value between −1 and +1. So, the
implementation of the control inputs by a relay with hysteresis showed that
18 2 Mathematical Methods

Fig. 2.8 Convex hull for the

system (2.4)

ẋ3 (t) = A, −1 ≤ A ≤ 1, x3 (t) = x3 (0) + At.

Validity of this equation can also be confirmed by Filippov’s method. Four possible
state speed vectors correspond to four combinations of two control functions. The
tetrahedron is their convex hull (Fig. 2.8), and its intersection with manifold x1 =
x2 = 0 (axis x3 ) is segment [−1, 1], which defines the same all possible sliding mode
equations
ẋ3 (t) = A, −1 ≤ A ≤ 1.

2.4 Equivalent Control Method

An alternative procedure for deriving the sliding mode equation ẋ(t) = f 0 (t, x(t))
is presented in this section for control system (2.1) under the assumption that f
is continuous on R n+m+1 , but each component u i of control u = (u 1 , . . . , u m )T is
discontinuous only on a smooth surface si (x) = 0. Assume that sliding mode occurs
in the intersection of k switching surfaces s k (x) = 0, where s k (x) is k-dimensional
vector consisting of k arbitrary components of the set {s1 (x), . . . , sm (x)}. Vector u k
consists of corresponding components (of course u k = u if k = m). Vector f 0 (t, x)
should belong to the space tangential to

s k (x) = 0, or ∇ Ts(x) f 0 (t, x) = 0,

where ∇ Ts(x) is k × n matrix consisting of gradients of components of s k as rows.

It looks natural to replace the discontinuous control in (2.1) by a continuous state
function, such that the condition ∇ Ts k (x) f 0 (t, x) = 0 holds. Consequently, a solution
to equation ∇ Ts k (x) f (t, x, u) = 0 with respect to u should exist. The solution u keq
called equivalent control [1, Chap. 2] is substituted into (2.1) to get the sliding mode
equation. The method of deriving the sliding mode equation is called equivalent
control method. Application of the equivalent control and Filippov’s methods to
Eq. (2.1) with the scalar control
2.4 Equivalent Control Method 19

Fig. 2.9 Different velocity

vectors: Filippov’s and
equivalent control methods

Fig. 2.10 Different velocity

vectors: Filippov’s and
equivalent control methods
(b)

 
u + (t, x) f (t, x, u + (t, x))
u(t, x) = and f (t, x) =
u − (t, x) f (t, x, u − (t, x))

demonstrates that, in general, they lead to different sliding mode equations (see
Figs. 2.9, 2.10). Indeed as follows from Filippov’s method, the right-hand side of the
sliding mode equation is defined by the intersection point of the tangential plane and
the straight line, connecting the ends of f + and f − (convex hull). The sliding mode
equation for the equivalent control method is defined by the intersection point of the
tangential plane and locus f (t, x, u) with respect to u (vector f (t, x, u eq ) should be
in the tangential plane).
Since both methods were postulated, therefore, the question “which of them is
correct” cannot be answered. Another important engineering question “which sliding
mode equation is in real-life systems” will be discussed in the next subsection.
It is interesting that for affine (linear with respect to control) systems of the form
⎫
ẋ(t) = f (t, x(t))+B(t, x(t))u(t, x(t)), x(t) ∈ R n , ⎪ ⎪
⎬
u i+ (t, x) if si (x) > 0,
u i (t, x) = − (2.7)
u i (t, x) if si (x) < 0, ⎪
⎪
⎭
u = (u 1 , . . . , u m ) , s = (s1 , . . . , sm ) ,

the both methods lead to the same sliding mode equation [1, p. 35]

ẋ(t) = f 0 (t, x(t)),
(2.8)
f 0 (t, x) = f (t, x) − B(t, x)(∇ Ts(x)B(t, x))−1 ∇ Ts(x) f (t, x).
20 2 Mathematical Methods

According to Filippov’s method, right-hand side f 0 of the sliding mode equation

should belong to the convex hull of all possible vectors of the original system (2.7).
Each control component can take two values; hence, the total number of all possible
m
control vectors is equal to 2m , {v 1 , . . . , v 2 }. The intersection of convex hull of the
right-hand side of (2.7)
 

m m
2 2
f (t, x) + B(t, x)z : z = μi v i , 0 ≤ μi ≤ 1, μi = 1
i=1 i=1

with the tangential manifold to s(x) = 0 defines sliding mode equation. It follows
that equation
∇ Ts(x) f (t, x) + ∇ Ts(x)B(t, x)z = 0

holds. Substitution of its solution

z = −(∇ Ts(x)B(t, x))−1 ∇ Ts(x) f (t, x)

into the equation for the convex hull results in the sliding mode equation. It coincides
with Eq. (2.8) resulting from equivalent control method.
An alternative procedure can be presented for control system (2.1) under assump-
tion that f is continuous on all arguments, but each component u i of a feedback
law
u : Rn+1 → Rm , u(t, x) = (u 1 (t, x), u 2 (t, x), . . . , u m (t, x))T

is discontinuous only on a surface

Si = {(t, x) ∈ Rn+1 : si (t, x) = 0},

where si : Rn+1 → R is a smooth function.

Let us construct the following multivalued function:

U (t, x) = co u (t, x + εB\N ) .

ε>0 μ(N )=0

Definition 2.3 An absolutely continuous function x : R → R n defined on some

interval or segment I is a solution to (2.1) if there exists a locally measurable function
u eq : R → R m such that
u eq (t) ∈ U (t, x(t))

and
ẋ(t) = f (t, x(t), u eq (t))

almost everywhere on I .
2.4 Equivalent Control Method 21

In [12, 13], the presented solution was called the Utkin solution, since it follows
the main idea of Equivalent Control Method proposed by Utkin [1]. Obviously, for
x∈ / Si we have u eq (t) = u(t, x(t)). So, the only question is, how to define u eq (t) on
a switching surface. The scheme presented in [1] is based on resolving the system
of algebraic equations:

∇ T si (x) f (t, x, u eq ) = 0 if x ∈ Si ,

(u eq )i = u i (t, x) if x ∈
/ Si ,

where i = 1, 2, . . . , m and u eq ∈ Rm is a vector to be defined. If the obtained solution

u eq (t, x) belongs to U (t, x), then it is called equivalent control [1].
The next lemma reduces the existence analysis of the Utkin solution to analysis
of a differential inclusion.
Lemma 1 ([14], p. 78) Let a function f : R n+m+1 → R n be continuous and a set-
valued function U : R n+1 → 2R be defined and upper semicontinuous on an open
m

set I × , where  ⊆ R n . Let U (t, x) be non-empty, compact, and convex for every
(t, x) ∈ I × . Let a function x : R → R n be absolutely continuous on I , x(t) ∈ 
for t ∈ I and
ẋ(t)∈ f(t, x(t), U(t, x(t))), (2.9)

almost everywhere on I. Then, there exists a measurable function u eq : R → R m

such that u eq (t) ∈ U (t, x(t)) and

ẋ(t) = f (t, x(t), u eq (t))

almost everywhere on I .
In general, the differential inclusion (2.9) may have non-convex right-hand side.
Unfortunately, the existence of solutions for non-convex differential inclusions is a
non-trivial problem [15]. Therefore, the algebraic approach introduced by equivalent
control method is still the most effective way to find the Utkin solutions.

2.5 Sliding Mode Equations in Control Affine Systems

It is interesting that the both methods lead to the same sliding mode equations for
affine (linear with respect to control) systems
⎫
ẋ (t) = f (t, x (t)) +  B(t, x (t))u(t, x (t)), x (t) ∈ R n ⎪
⎪
⎬
u i+ (t, x) if si (x) > 0
u i (t, x) = − (2.10)
u i (t, x) if si (x) < 0 ⎪
⎪
⎭
u = (u 1 , . . . , u m ) , s = (s1 , . . . , sm )
22 2 Mathematical Methods

where f : R n+k+1 → R n is a continuous vector-valued function, B : R n+1 → R n×m

is a continuous matrix-valued function, and u : R n+1 → R m is a piecewise continu-
ous function, such that u i has a unique time-invariant switching surface si (x) = 0,
where si : R n → R is a smooth function.

Theorem 1 Definitions of Filippov and Utkin are equivalent if and only if

 
det ∇ T s(x)B(t, x) = 0 for (t, x) ∈ S, (2.11)

where
s(x) = (s1 (x), s2 (x), . . . , sm (x))T , ∇s(x) ∈ R n×m

s(x) = (s1 (x), s2 (x), . . . , sm (x))T , ∇s(x) ∈ R n×m is the matrix of partial deriva-
∂s j
tives , and S is a discontinuity set of u(t, x).
∂xi
If the switching surface satisfies the condition of this theorem, then the equivalent
control and sliding motion equation can be constructed analytically. For instance,
the sliding motion condition s = 0, ṡ = 0 implies

∇ T s(x) f (t, x) + ∇ T s(x)B(t, x)u(t, x) = 0. (2.12)

The condition (2.12) allows us to obtain the equivalent control

 −1 T
u eq (t, x) = − ∇ T s(x)B(t, x) ∇ s(x) f (t, x) (2.13)

as well as the sliding mode equation

  −1 T 
ẋ(t) = f 0 (t, x(t)), f 0 = In + B ∇ T s B ∇ s f. (2.14)

According to Filippov’s method, right-hand side f 0 of the sliding mode equation

should belong to the convex hull of all possible vectors of the original system (2.10).
Each control component can take two values; hence, the total number of all possible
m
control vectors is equal to 2m , {v 1 , . . . , v 2 }. The intersection of convex hull of the
right-hand side of (2.10)
 

m m
2 2
f (t, x) + B(t, x)z : z = μi v i , 0 ≤ μi ≤ 1, μi = 1
i=1 i=1

with the tangent manifold s(x) = 0 defines the sliding mode equation. It follows
that the equation
∇ T s(x) f (t, x) + ∇ T s(x)B(t, x)z) = 0

holds. Substitution of its solution

2.5 Sliding Mode Equations in Control Affine Systems 23

 −1 T
z = − ∇ T s(x)B(t, x) ∇ s(x) f (t, x)

into the equation for the convex hall results in the sliding mode equation. It coincides
with Eq. (2.14) resulting from the equivalent control method.
Briefly describe the approaches to derive SM equation, different from Filippov
and equivalent control methods.
Input s(x) is equal to zero in sliding mode for the element implementing dis-
continuous control, while its output is different from zero. Based on this fact [16],
offered to replace it by a linear element with gain tending to infinity and take the
slow motion equation as a model of sliding mode. In [17], there was offered to find
the solution with respect to s in convolution form for LTI system and then the control
from the integral equation resulting from s(t) = 0. The sliding mode equation is
obtained after the substitution of this control into the original system. The authors
of [18] assumed that control in (2.7) takes all values between u + and u − , while the
authors of [19] offered to select the values of control from some set, which, on one
hand, may not include all values between u + and u − but, on the other hand, may
be beyond this interval. In both cases, the authors use differential inclusion form
ẋ ∈ F(x, t), where F(x, t) is a convex set thus introduced the right-hand side of the
original system.
Although the above methods may lead to different sliding mode equations, the
question “which one is correct?” cannot be asked, since all of them were postulated.
Variety of methodologies to analyze sliding mode equations can be found in
excellent survey papers [20, 21].

2.6 Regularization

As demonstrated in previous sections, different ways of solution continuation may

lead to different sliding mode equations except for non-singular affine systems. It
is of interest to answer the natural question, which motion equation is implemented
in a real-life system. Formally, the question cannot be answered. Systems with dis-
continuous control do not satisfy the conventional existence-uniqueness theorems
since the Lipschitz constant does not exist for discontinuous systems. Any solution
ambiguity may be excluded, if the system behavior is analyzed based on a more
accurate model. The accurate model takes into account the way of implementation
of discontinuous control such as hysteresis, time-delay, and high-gain amplifier with
output limitation. Then the solution of the corrected ideal model exists and unique.
If the limit of the solution exists with a small parameter, added to the ideal model
and tending to zero, then it can be taken as the solution to the ideal sliding mode
equations. This method is called Regularization [1, Chap. 2].
If different ways of regularization lead to different motion equations, we should
admit that the equations beyond discontinuity surfaces do not allow us to derive the
sliding mode equations unambiguously.
24 2 Mathematical Methods

An interesting example of the relay system


ẋ1 = 0.3x2 + x1 u, s = x1 + x2 ,
(2.15)
ẋ2 = −0.7x1 + 4x1 u, u = −sign(x1 s)

with different sliding mode equations depending on the way of regularization is given
in [1]. The solution exists for this system, if the ideal relay is replaced by a relay with
hysteresis or by a linear function in small vicinity of sliding surface (Figs. 2.11 and
2.12). This example corresponds to Fig. 2.10 and explains why it is happened. The
regularization method implies that the width of hysteresis loop or the vicinity with
the linear function tends to zero. The limit procedure results in Filippov’s equation
with unstable solution in the first case and in equivalent control method equation with
stable solution in the second case (Fig. 2.13). Further, the method by Filippov will be
modified slightly to get all possible solutions depending on the way of regularization.

Fig. 2.11 Relay with

hysteresis

Fig. 2.12 Relay with limiter

Fig. 2.13 Dependence of

stability of regularization
method: black line—solution
before reaching of the sliding
surface; green line—solution
derived by means of
equivalent control method;
red line—Filippov solution
2.6 Regularization 25

Filippov’s method was offered as a postulate. And it represents the following rather
evident fact. If the velocity vector can take several values in a vicinity of some point
in the state space, then due to switching, the right-hand side of the motion dynamics is
their minimal convex hull, and any other motions cannot appear. Generally speaking,
depending on a way of the regularization, control u can take additional values (not
u + or u − only) and all possible sliding mode equations are defined by the convex
hull, depending on all these values.
Regularization implies that the ideal control is replaced by a new one u ε ,

u(t, x) if x ∈
/ N + εB,
u ε (t, x) = ε > 0, (2.16)
u r eg (t, x, ε) if x ∈ N + εB,

where N = {x ∈ R n : s(x) = 0} is the discontinuity set of measure zero, B is the

unit ball, and u r eg depends on a regularization method and selected such that some
conventional uniqueness-existence theorem can be applied. Then, the set of all pos-
sible sliding mode Eq. (2.6), taking into account the regularization, can be found in
terms of u ε
F(t, x) = lim Conv{ f (t, x, u ε (t, {x + εB}\N ))}.
ε→0

For continuous approximation of a discontinuous function, the convex hull should

be found for all possible values of u r eg (dashed sector in Fig. 2.14). Indeed, consider
the set of sliding mode equations, resulting from the regularization approach, for the
system (2.15). Suppose that

u − if s < 0,
uε =
u + if s > 0.

As it can be seen in Fig. 2.14, sliding mode exists on s = 0. If the switching device
is implemented with hysteresis, this regularization leads to Filippov method [1].
The intersections of the straight line connecting f (x, t, u  ) and f (x, t, u  ) with the
tangential line determine the sliding mode equation. It is evident that any point on
the tangential line between f f ili ppov and f eq can be obtained by a proper choice of u −
and u + . It means that the set of all possible velocity vectors f 0 = Convu f (x, t, u)
can be found in the intersection of the convex hull of the arc between u − and u +
(dashed sector in Fig. 2.14) and tangential line. Note that the sliding mode equation
is not postulated but follows from the regularization method. Intersection of the hull
with the manifold, tangential to the sliding manifold, determines all possible sliding
mode vector fields, namely, all vector between f f ili ppov and f (x, u eq ) in Fig. 2.14.
Note that for this way of regularization they can be obtained from definition given
in [18], assuming that u − ≤ u r eg ≤ u + . But it is not the case always, for example,
for Coulomb friction [19].
Finally, it would be of interest to derive a class of systems with unique sliding mode
equation for any way of regularization, or for any function u r eg . The method (2.16) is
called “boundary layer regularization”. It was formulated in [1, Chap. 2] in a slightly
26 2 Mathematical Methods

Fig. 2.14 Regularization

method

different way. The solution of the original system with control (2.16) depends on u r eg
and ε. If for any function u r eg the following limit lim x(t, u r eg , ε) = x(t) exists and
ε→0
unique, then this function x(t) is taken as a solution to the motion equation in sliding
mode. As shown above, generally speaking it is not the case. Filippov method and
equivalent control method result in the same sliding mode equations (except for
singular cases) for affine systems. The validity of those equations was substantiated
for affine systems via the boundary layer regularization method [1]. It means that
the sliding mode is governed by (2.8) for any control implementation way. Note that
the system order can be reduced to n − m. Indeed, m state variables can be found
as functions of the remaining n − m ones from the algebraic equation s(x) = 0 (as
follows from [1] the solution exists). Then, m variables are substituted into sliding
mode equations, and finally the equations with respect to m variables should be
excluded from the system.

References

1. Utkin, V.: Sliding Modes in Control and Optimization. Springer, Berlin (1992)
2. Utkin, V., Guldner, J.: Sliding Mode Control in Electro-Mechanical Systems. CRC Press (2009)
3. Edwards, C., Spurgeon, S.: Sliding Mode Control: Theory and Applications. Taylor and Francis
(1998)
4. Shtessel, Y., Edwards, C., Fridman, L., Levant, A.: Sliding Mode Control and Observation.
Control Engineering Series, Birkhauser, NY (2014)
5. Drazenovic, B.: The invariance conditions in variable structure systems. Automatica 5(3),
287–295 (1969)
6. Glad, S.T.: Output dead-beat control for nonlinear systems with one zero at infinity. Syst.
Control Lett. 9, 249–255 (1987)
7. Utkin, V.I.: Variable structure and lyapunov control. In: Sliding Mode Control in Discrete-Time
and Difference Systems. Springer, Berlin (1993)
8. Barbashin, E.A.: Introduction to Stability Theory. Nauka, Moscow (1970) (in Russian)
9. Blanchard, P., Devaney, R.L., Hall, G.R.: Differential Equations. Thompson Brooks/Cole
(2006)
10. Levant, A.: Sliding order and sliding accuracy in sliding mode control. Int. J. Control 58(6),
1247–1263 (1993)
References 27

11. Filippov, A.F.: Differential Equations with Discontinuous Righthand Sides. Kluwer Academic
Publishers (1988)
12. Orlov, Y.: Discontinuous Systems: Lyapunov Analysis and Robust Synthesis Under Uncertainty
Conditions. Springer, Berlin (2008)
13. Polyakov, A., Fridman, L.: Stability notions and lyapunov functions for sliding mode control
systems. J. Frankin. Inst. 351(4), 1831–1865 (2014)
14. Filippov, A.F.: On certain questions in the theory of optimal control. J. SIAM Control 1(1),
76–84 (1962)
15. Bartolini, G., Zolezzi, T.: Variable structure systems nonlinear in the control law. IEEE Trans.
Autom. Control 30(7), 681–684 (1985)
16. Tsypkin, Ya.: Theory of Relay Automstic Regulations. State Publisher of Engineering Litera-
ture (1956)
17. Neimark, Y.I.: Note on A. Filippov Paper. In: Proceedings of 1st IFAC Congress II. Butter-
worths, London (1961)
18. Aizerman, M.A., Pyatnitskii, E.S.: Foundations of the theory of discontinuous systems. Autom.
Remote. Control. 8, 39–61 (1974)
19. Gelig, A.Kh., Leonov, G.A., Yakubovich, V.A.: Stability of Nonlinear Systems with Nonunique
Equilibrium Position. Nauka (1978) (in Russian)
20. Cortes, J.: Discontinuous dynamic systems. IEEE Control. Syst. Mag. 28(3), 36–73 (2008)
21. Leonov, G.A., Kuznetsov, N.V., Kiseleva, N.A., Mokaev, R.N.: Global problems for differential
inclusions. Differ. Equ. 53(13), 1671–1697 (2017)
Chapter 3
Design Principles

Abstract This chapter discusses the main principles of sliding mode design. Regular
form of the controlled system is introduced. Reaching and existence conditions are
presented. The decoupling procedure and invariance notion are discussed. The basic
principles of output, unit, and integral SMC are presented. Second-order sliding mode
(SOSM) controllers (such as twisting and super-twisting) are analyzed. Suboptimal
controllers are also discussed.

Keywords Regular form · Reaching and existence conditions · Decoupling

procedure · SOSM control · Twist and super-twist controllers

3.1 Sketch of the Design Procedure

The sliding mode equations based on an ideal model can be obtained unambiguously
for affine systems. Therefore, the methodology for designing feedback sliding mode
control was developed mainly for affine systems governed by Eq. (2.7). It is important
to note that

• the sliding mode equation is of a reduced order;

• it does not depend on the control values;
• it depends on discontinuity manifold equation (2.8).

These properties allow us to outline the design procedure for the sliding mode
control:
1. Select function s : R n → R m such that

| det(∇  s · B)| > Δ0

and the sliding mode on the manifold

s(x) = 0, x ∈ R n

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 29

V. Utkin et al., Road Map for Sliding Mode Control Design,
SpringerBriefs in Mathematics,
https://doi.org/10.1007/978-3-030-41709-3_3
30 3 Design Principles

has desired properties. Any design method of the control theory (such as eigen-
value placement and optimality) can be utilized for this reduced-order problem.

Remark 3.1 Step 1 implies that matrix ∇  s · B is non-singular. It means that the
relative degree between control as an input and vector s as an output is equal to 1. If
it is not the case, the control is selected as a function of s and its time derivatives to
make the system relative degree between control as an input and vector s as an output
equal to 1. It was done at the early stage of SMC, when the system behavior was
studied in the space of the output and its time derivatives. Later, the same method
was used in the publications on high-order sliding mode control [1].

Remark 3.2 “Select s such that the sliding mode has desired properties”. It is not
the conventional problem statement, since the right-hand side depends on not only
s, but also the gradients of its components.

Remark 3.3 Formally, sliding mode equations are equations of zero dynamics of
the system with control and s as input and output.

2. Discontinuous control is selected to enforce sliding mode in the preselected man-

ifold s(x) = 0. Again, we deal with the problem of a reduced order, since the state
trajectories should reach the manifold after a finite time interval in m-dimensional
subspace s in compliance with the equation

ṡ = ∇  s · f + ∇  s · Bu. (3.1)

The above two-step design methodology with sliding mode in (n − m)-

dimensional manifold was called “Conventional sliding mode control” in [2].
Further, this term will be used.

3.2 Regular Form

The design procedure becomes transparent, for the systems in the so-called “regular
form” [3] 
ẋ1 (t) = f 1 (t, x1 (t), x2 (t)),
(3.2)
ẋ2 (t) = f 2 (t, x1 (t), x2 (t) + B2 (t, x1 (t), x2 (t))u(t),

where
x1 (t) ∈ R n−m , x2 (t) ∈ R m , u(t) ∈ R m ,
f 1 : R n+1 → R n−m , f 2 : R n+1 → R m ,
B2 : R n+1 → R m×m and det(B2 ) = 0.

The state component x2 (t) is handled as a fictitious control input in the first equation
of (3.2). Select
x2 = −s0 (x1 )
3.2 Regular Form 31

to get the desired dynamics in the first subsystem. As a result, we deal with the
design problem in the system of the (n − m)th order with m-dimensional control
x2 (t) = −s0 (x1 (t)), where s0 is a smooth function. Then, the discontinuous control
should be designed to enforce sliding mode in the manifold

s(x1 , x2 ) = x2 + s0 (x1 ) = 0, x1 ∈ R n−m , x2 ∈ R m

(the design problem of the mth order with m-dimensional control).

If sliding mode in the manifold s(x1 , x2 ) = 0 starts after a finite time interval,
then the system exhibits the desired behavior governed by

ẋ1 (t) = f 1 (t, x1 (t), −s0 (x1 (t))).

Note that the motion is of a reduced order and depends neither on function f 2 nor on
function B2 in the second equation of the original system (3.2).
As mentioned in Remark 3.2, for a general case sliding mode equation depends
on not only function s, but also the gradients of its components. The first step of the
design procedure for systems in the regular form does not depend on gradients and
the conventional methods of the control theory can be applied directly. The necessary
and sufficient conditions for an affine system to be reduced to the regular form (3.2)
can be found in [3] and [4, Chap. 6].
To demonstrate the sliding mode control design methodology, consider the con-
ventional problem of linear control theory: eigenvalue placement in a linear time-
invariant system:
ẋ(t) = Ax(t) + Bu(t),

where x(t) and u(t) are n- and m-dimensional state and control vectors, respectively,
A and B are constant matrices, rank(B) = m. The system is assumed to be control-
lable. For any controllable system, there exists a linear feedback u = F x (F being
a constant matrix), such that the eigenvalues of the feedback system, i.e., of matrix
A + B F, take the desired values: as a result, the system exhibits desired dynamic
properties. Reducing system equations to the regular form will be performed as a pre-
liminary step in the sliding mode control design procedures. Since rank = m matrix
B may be partitioned (after reordering the state vector components) as
 
B1
B= ,
B2

where
B1 ∈ R n−m×m , B2 ∈ R m×m , det(B2 ) = 0.

The non-singular coordinate transformation

   
x1 In−m −B1 B2−1
= T x, T =
x2 0 B2−1
32 3 Design Principles

reduces the system equations to the regular form

ẋ1 (t) = A11 x1 (t) + A12 x2 (t),

ẋ2 (t) = A21 x1 (t) + A22 x2 (t) + u(t),

where x1 (t) ∈ R n−m , x2 (t) ∈ R m , and Ai j are constant matrices, i, j = 1, 2. It fol-

lows from the controllability of (A, B) that the pair (A12 , A22 ) is controllable as
well [4]. Handling x2 as an m-dimensional intermediate control in the controllable
(n − m)-dimensional first subsystem, all eigenvalues may be assigned arbitrarily by
a proper choice of matrix C∈ R m×(n−m) in

x2 = −C x1 .

To provide the desired dependence between components x2 and x1 of the state vector,
sliding mode should be enforced in the manifold

s(x1 , x2 ) = x2 + C x1 = 0,

where s = (s1 , s2 , . . . , sm ) is the difference between the real values of x2 and its
desired value −C x1 . After sliding mode starts, the motion is governed by the system
with the desired eigenvalues

ẋ1 (t) = (A12 − A12 C)x1 (t).

Now, eigenvalue problem can be solved easily for the system of a reduced order,
since the selection of matrix C is a conventional design problem of the linear control
theory.

3.3 Reaching and Existence Conditions

Although the design procedure in Sect. 3.1 consists of two problems of reduced
orders, it cannot be stated that the design problem is decoupled into two independent
subproblems. Indeed, Eq. (3.1) depends on function s selected at the first step as well.
The second step is selection of discontinuous control such that the sliding mode is
enforced in manifold s(x) = 0. It means that for any initial conditions the state should
reach s(x) = 0 and it should be a sliding manifold according to Definition 2.1.
The reaching and sliding mode existence conditions can be derived based on
Eq. (3.1). Let the ith component of control undergoes discontinuities on surface
si (x) = 0

u i+ (t, x) if si (x) > 0,
u i (t, x) = u i+ (t, x) = u i− (t, x),
u i− (t, x) if si (x) < 0,
3.3 Reaching and Existence Conditions 33

for example,
u i (t, x) = u i0 (t, x) + u i1 (t, x)sign(si (x)),
u i0 = 0.5(u i+ + u i− ), u i1 = 0.5(u i+ − u i− )

and ⎫
ṡ(t) = d(t, x(t)) − D(t, x(t))sign(s(x(t))), ⎬
d = ∇  s ( f +Bu 0 ), D = −∇  s Bdiag{u i1 }, (3.3)
⎭
u 0 = (u 01 , u 02 , . . . , u 0m ) .

Theorem 3.1 ([4]) If the matrix D + D  > 0 is positive definite, then there exists
scalar function M0 : R n+1 → R, M0 > 0 such that manifold s(x) = 0 is reached
after a finite time interval and s(x) = 0 is a sliding manifold for control

u = u 0 +M sign(s)

with M : R n+1 → R and M > M0 .

As mentioned above, the two steps of the control design are interconnected since
matrix ∇s depends on the preselected function s. Generally speaking, D + D  is
not positive definite.

3.4 Decoupling

Manifold s(x) = 0 is the intersection of m surfaces si (x) = 0 selected such that the
sliding mode has the desired properties. Manifold s(x) = 0 and as a result the sliding
mode equations are the same if vector s(x) is replaced by

s ∗ (t, x) = Q(t, x)s(x), det(Q) = 0,

and the control undergoes discontinuities on the surface s ∗ (t, x) = 0. However, the
motion in subspace s depends on matrix Q.
The idea of decoupling is in finding matrix Q(t, x) for any preselected vector s(x)
such that the state trajectories are converging to the origin in subspace s. Equation
(3.1) can be written in the form

ṡ = [∇  s · B](u − u eq ),

where u eq was found for the system (2.7). Calculate time derivative of Lyapunov
function V = 0.5s  s along the trajectories of the system (3.1):

V̇ = s  [∇  s · B](u − u eq ).

Then, replace the switching manifold by a new one

34 3 Design Principles

s ∗ (t, x) = 0, s ∗ = [∇  s · B] s.

For control u = −M sign(s ∗ ), M > 0, we have

V̇ = −M|s ∗ | − (s ∗ ) u eq ≤ −|s ∗ |(M − |u eq |).

Under a natural assumption that the upper estimate u max > |u eq | for the equivalent
control is known, V̇ < 0 and all trajectories are converging to manifold s(x) = 0 (or
to s ∗ (t, x)), if M > u max . Finally, it can be shown that the state reaches s(x) = 0
after a finite time interval provided that

V̇ ≤ −αs ∗ , α = min[M(t, x) − u max (t, x)].

(t,x)

Indeed, since
s ∗  = [∇  s · B] s ≥ λmin s,

where λmin : R n+1 → R is a minimal eigenvalue of the positive symmetric matrix

 = [∇  s · b] · [∇  s · B] : det() ≥ ξ > 0,

then √

V̇ ≤ −γ V , γ = α 2 inf λmin (t, x),

(t,x)

and the sliding manifold s(x) = 0 is reached in a finite time

√
2 s(x(0))
t r each ≤
γ

(see, for example, [5, Chap. 2]).

3.5 Invariance

Assume that for an arbitrary affine system

ẋ(t) = f (t, x(t)) + B(t, x(t))u(t) + h(t, x), (3.4)

the function h characterizes the disturbances and parameter variations which should
not affect the feedback system dynamics. In compliance with the equivalent control
method, the solution to
ṡ = ∇  s ( f + Bu + h) = 0

with respect to control,

3.5 Invariance 35

u eq = −[∇s  · B]−1 ∇s( f + h)

should be substituted into (3.4) to yield the sliding mode equation

  −1  
ẋ(t) = f 0 (t, x(t)), f 0 = In −B ∇  s B ∇ s ( f +h). (3.5)

Let range of B(t, x) be a subspace formed by the base vectors of matrix B(t, x) for
each point (t, x). The sliding mode is invariant with respect to function h if

h(t, x) ∈ range(B(t, x)), (t, x) ∈ R n+1 .

This condition means that there exists function γ such that h = Bγ. Direct substi-
tution of function h into (3.5) demonstrates that the sliding motion in any manifold
s(x) = 0 does not depend on perturbation function h. As it follows from the design
methods of this section, an upper estimate of this vector is needed to enforce the
sliding motion. Condition

h(t, x) ∈ range(B(t, x))

is called matching condition, and it is a natural generalization of the invariance

condition obtained in [6] for linear systems.
The invariant SMC can be designed, if a state vector is available; however, it is
not the case always, and an output can be measured only. The traditional approach
implies designing a state observer, but in our case it should be done under uncertainty
conditions (vector h(t, x) is unknown). The solution was offered in the set of pub-
lications by H. Khalil for the special case, when motion equations are in canonical
space, consisting of an output and its time derivative or a set of similar subspaces,
if an output is a vector [7, 8]. All state variables can be found by the observer with
the appropriate hierarchy of its input gains tending to infinity. It is important that
the offered method does need the unknown disturbance. The stability of the overall
system consisting of the observer and the plant is analyzed in [8].

3.6 Output Sliding Mode Control

The design methods of the previous section has been discussed under the assumption
that all state components are available. In most practical situations, an output vector
can be measured only. The objective of this section is to derive the class of systems,
which can be stabilized by sliding mode control utilizing a system output only. The
design methodology for a special case of systems with the same dimension of control
and output [9] will be considered. The objective of this subsection is to single out
the class of systems with matched uncertainties
36 3 Design Principles

ẋ (t) = Ax (t) + B [k (t, x) u (t) + h (t, x)] ,

y(t) = C x (t) ,
x (t) ∈ R n , u (t) ∈ R m , y (t) ∈ R m , rank(B) = rank(C) = m,
0 < k0 ≤ k (t, x) ∈ R, h (t, x) ≤ h 0 ,

assuming, that

• dimensions of the output and the control input are the same y(t) ∈ R m , u(t) ∈ R m ,
• det(C B) = 0,
• sliding mode is enforced in manifold y = 0 and sliding motion in this manifold
has some desired properties such as stability.

In contrast to the conventional two-step design procedure, the first step (selection
of sliding manifold) is not needed, and the second step (enforcing sliding mode)
should be solved using output y only. It is assumed that there exists a linear control
stabilizing the system with no disturbances. Finally, the real control consists of this
linear component and a discontinuous part to enforce sliding mode at the presence
of the bounded disturbance h(t, x), satisfying matching condition.
Now it is demonstrated, how this design idea is implemented. As in the previous
subsection, the system can be reduced to the regular form

ẋ1 (t) = A11 x1 (t) + A12 x2 (t),

ẋ2 (t) = A21 x1 (t) + A22 x2 (t) + k(t, x)u(t) + h(t, x)),
y(t) = C1 x1 (t) + C2 x2 (t), x1 (t) ∈ R n−m , x2 (t) ∈ R m .

Condition det(C B) = 0 means det(C2 ) = 0. Without loss of generality, sliding man-

ifold can be written as y = 0 with

y = C1 x1 + x2 .

As in previous subsections, the system can be reduced to the regular form

ẋ1 (t) = (A11 − A12 C1 )x1 (t) + A12 y(t)),

ẏ(t) = (C1 (A11 − A12 C1 ) + A21 − A22 C1 )x1 (t)
+(C1 A12 + A22 )y(t) + k(t, x)u(t) + h(t, x(t)).

Of course sliding mode equation

ẋ1 (t) = (A11 − A12 C)x1 (t)


in y = 0 is stable with a Lyapunov function defined as x1 P1 x1 , where P1 is a positive-
definite matrix. As shown in [9], using the Lyapunov function

V = x1 P x1 + y  y,

the system with control

3.6 Output Sliding Mode Control 37

u = −qy − M sign(y), M > h 0 /k0

and high enough q is asymptotically stable. It follows from Sylvester criterion directly
(see [10]). The final stage of the process is in sliding mode, since condition (2.2) for
this motion to exist holds in the vicinity of the origin.

3.7 Integral SMC

3.7.1 Main Idea

The design procedure of SMC implies that the motion consists of two phases: reach-
ing the preselected manifold in the state space and sliding along the manifold with
the desired properties. Integral sliding mode control (ISMC) excludes the reaching
phase and enforces the desired motion from an initial time instant [11], [5, Chap. 7].
Its design methodology can be demonstrated for the simple second-order system

ẋ1 (t) = x2 (t), x1 (t) ∈ R,
ẋ2 (t) = u + h(t, x), x2 (t) ∈ R,

where the upper limit of disturbance |h(t, x)| ≤ h 0 is known only. First, the control
is selected such that the system with h = 0 would be governed by the linear equation
with roots of characteristic equation equal to {−1, −1}. It happens if control is equal
to
u 0 = −x1 − 2x2 .

Complement the original system by the third equation

ż(t) = −u 0 (t), z(0) = 0,

and select the control

u(t) = u 0 (t) − M sign(s(t)), M > h 0 .

s(t) = x2 (t) − x2 (0) + z(t).

Since
ṡ(t) = −M sign(s(t)) + h(t, x)

and s(0) = 0, the sliding mode occurs at the initial instant of time and (M sign(s))eq =
h(t, x). Substitution of
u = u 0 − (M sign(s))eq

into the original system results in desired motion equation

38 3 Design Principles

ẋ1 (t) = x2 (t), ẋ2 (t) = −x1 (t) − 2x2 (t).

An affine system with matched uncertainties is considered


ẋ(t) = f (t, x(t)) + B(t, x(t))u(t) + h(t, x), t > 0,
(3.6)
x(t) ∈ R n , u(t) ∈ R m , h = Bγ, rank(B) = m.

Bounded function h (or γ) represents all uncertainties. Only upper estimate γ0 >
γ is known. It is assumed that control u 0 can be found for the nominal system
(system (3.6) with h = 0) in correspondence with some criterion. Let x0 (t), t > 0
be a solution to the nominal feedback system

ẋ0 (t) = f (t, x0 (t)) + B(t, x0 (t))u 0 (t, x0 (t)). (3.7)

Control u should be designed such that the solution to (3.6) coincides with the solution
to (3.7) starting from the initial time.

3.7.2 Design Method

Select the control in the form

u(t) = u 0 (t, x(t)) − M(t, x)(G(x(t)) B(t, x(t)))−1 sign(s(t)),

s(t) = s0 (x(t)) − s0 (x(0)) + z(t),
ż(t) = −[G(x(t)) f (t, x(t)) + G(x(t))B(t, x(t))u 0 (t, x(t))],
∂s0
G(x) = − , z(0) = 0.
∂x

At initial time, the state is in the manifold s = 0 in space (x, z), and function M
will be selected such that sliding mode is enforced in the manifold. Let s0 (x) be an
arbitrary function satisfying condition

det[G(x)B(t, x)] = 0, t ≥ 0, x ∈ R n .

Remark 3.4 In our second-order example,

G B = 1, f = 0, s0 = x2 .

As it follows from

ṡ = G B(u + γ) = −Msign(s) + G Bγ,

sliding mode in the manifold s = 0 exists if M > G Bγ0 . The equivalent control
3.7 Integral SMC 39

u eq (t, x(t)) = −γ(t, x(t))

is the solution to ṡ(t) = 0. Substitution of u eq into (3.6) results in sliding mode

equation
ẋ(t) = f (t, x(t)) + Bu 0 (t, x(t)).

It coincides with the desired equation for the system without uncertainties (3.7). The
desired motion occurs from initial time. Note that in contrast to the traditional design
methodology, the order of this motion is not reduced.
As an example, consider linear system

ẋ(t) = Ax(t) + Bu(t) + Bγ(t, x(t)),

A, B are constant matrices, rank(B) = m, γ0 > γ and select control such that the
system is governed by homogeneous time-invariant equation with desired eigenval-
ues starting from the initial time. If the system is controllable, then there exists matrix
F such that matrix A + B F has desired eigenvalues. Select control in the form

u(t, x(t)) = u 0 (x(t)) + u i (t, x(t)), u 0 (x) = F x,

u i (t, x) = −M sign(si (t, x)),
s(t, x(t)) = s0 (x(t)) − s0 (x(0)) + z(t), s0 (x) = B + x, B + B = I,
ż(t) = −B + Ax(t) − u 0 (x(t)).

Then
ṡ = −M sign(s) + γ

and sliding mode exists, if M > γ0 . Equivalent control

u ieq (t, x(t)) = −γ(t, x(t))

and sliding mode is governed by full-order equation starting from the initial time

ẋ(t) = (A + B F)x(t)

with the desired eigenvalues.

The application of ISM to LQ problem and specific algebraic observers may be
found in [12].

3.8 Unit Control

The unit feedback synthesis has been developed to simplify the stability analysis
and avoid extra computations required to bring the system into the regular form. The
controller is synthesized in such a manner that the time derivative of a Lyapunov
40 3 Design Principles

function, selected for a nominal, asymptotically stable system, remains negative

definite in spite of parameter variations and external disturbances, which affect the
system. The approach was developed in [13, 14], then was applied for designing
sliding mode control, and was referred to as a unit control [15]. Its norm is equal to
one everywhere except for the discontinuity manifold.
Similar to the component-wise control, the unit control with sufficiently high
magnitude can enforce asymptotically stable sliding mode, which is robust against
matched disturbances. The important point is that the trajectories of the closed-loop
system never pass through the discontinuity manifold. The system stability is thus
analyzed beyond the manifold. Once the trajectory is on the discontinuity manifold,
smooth dynamics restore, and the standard Lyapunov theory is in force.
The unit control methodology was developed for affine systems

ẋ(t) = f (t, x(t)) + B(t, x(t))u(t) + h(t, x) (3.8)

with matched disturbances such that h = Bγ, γ ≤ γ0 .

The equation
ẋ(t) = f (t, x(t)) (3.9)

represents an open-loop nominal system. For simplicity, the nominal system (3.9) is
assumed to be asymptotically stable with some a priori known positive-definite con-
tinuously differentiable Lyapunov function V , such that its time derivative computed
along the trajectories of (3.9) is negative definite,

W0 (t, x) = ∇  V (x) f (t, x) ≤ −W1 (x), (t, x) ∈ R n+1 ,

where W1 is a continuous, positive-definite function. The time derivative of V on the

trajectories of the perturbed system (3.8) is of the form

dV
W = = W0 + ∇  V · B(u + γ).
dt
Let the system (3.8) be driven by the control input
s
u = −ρ(t, x)U (s(t, x)), U = (3.10)
s

with a scalar function ρ such that

ρ > γ0 , (3.11)

and m-vector function

s(t, x) = B  (t, x)∇V (x).

Control (3.10) is called the unit control since U  = 1 everywhere beyond s = 0

where it undergoes a discontinuity.
3.8 Unit Control 41

Then due to (3.11), the time derivative of the Lyapunov function V (x), computed
on the trajectories of the closed-loop system (3.8), is negative definite:

W = W0 − ρ  B  ∇V  +∇  V · Bγ ≤
−W1 −  B  ∇V  [ρ − γ0 ] ≤ −W1 .

It means that the closed-loop system is asymptotically stable.

It is of interest to note that in contrast to the conventional sliding mode control
signals, which undergo discontinuities whenever a component of the sliding manifold
changes sign, the unit control action is a continuous state function until the manifold
s = 0 is reached. Due to this difference, the unit control method is an appropriate tool
of discontinuous control design not only in a finite-dimensional state space, but also
in an infinite-dimensional state space where control inputs are not (or even cannot
be) represented in a component-wise form.

3.9 Second-Order Sliding Mode Control

The second-order sliding mode control (SOSM) was originated for systems with
scalar input 
ẋ(t) = f (t, x(t)) + b(t, x(t))u(t),
(3.12)
x(t) ∈ R n , u(t) ∈ R.

The design method consists of two steps similar to the conventional one: selection
of a sliding manifold and selection of a discontinuous control, enforcing a sliding
mode. But in contrast to the conventional method, the dimension of a sliding manifold
is equal to n − 2, but not n − 1. It explains the title of the method. As before, a sliding
manifold is chosen such that the motion along it exhibits the desired properties and
the reaching problem should be solved. Since the order of sliding mode equation is
equal to n − 2, the reaching problem is a finite-time stability one of a second-order
nonlinear system.
We start with the second-order system with evident stability conditions (see Exam-
ple 3.8 in [16]). As seen later, they can be applied directly to SOSM. The mechanical
“mass-spring-damper” system (Fig. 3.1) with no Coulomb friction and with no exter-
nal forces is asymptotically stable because its energy dissipates and it may therefore
serve as a Lyapunov function. The right-hand side of the motion equation

Fig. 3.1 Mass-spring-

damper system
42 3 Design Principles

ẍ = −φ(x, ẋ) − F(x) (3.13)

consists of viscous friction φ satisfying

ẋφ(x, ẋ) > 0

for ẋ = 0 and spring force F such that

x F(x) > 0

if x = 0. Lyapunov function is sum of kinetic and potential energies

 x
m ẋ 2
V = + F(y)dy.
2 0

The time derivative

V̇ = m ẋ ẍ + F(x)ẋ = −ẋφ(x, ẋ) < 0

if ẋ = 0, which confirms our statement about stability. Note, V stops decaying only
if ẋ = 0, but, as follows from (3.13), ẋ ≡ 0 means x ≡ 0 and asymptotic stability.

3.9.1 Twisting Algorithm

Illustrate the design idea for twisting algorithm [17] for the case ∇  s b = 1

ṡ = ∇  s f + ∇  s b u. (3.14)

Certainly, we deal with the conventional sliding mode control if u is a discontinuous

function of s such that the sliding mode with the desired properties in s = 0 is
enforced. The range of sliding mode control application will be increased, if the same
effect will be obtained with continuous control actions, since not all actuators can
operate with discontinuous inputs. For example, high-frequency switching may result
in damage of valves in hydraulic actuators. A possible solution is evident: control
signal is generated by the output of an auxiliary integrator with a discontinuous
function of s and its time derivative as input. If this function is linear cs + ṡ, the
origin in plane (s, ṡ) is reached asymptotically in the sliding mode. Again this design
method is in a framework of the conventional sliding mode control.
Twisting algorithm

u(t) = v(t), v̇(t) = −M1 sign(s(t)) − M2 sign(ṡ(t)),
(3.15)
M1 , M2 are positive constants,
3.9 Second-Order Sliding Mode Control 43

Fig. 3.2 Part of the state

trajectory for twisting system

solves the same problem of reaching the origin in the plane (s, ṡ), but it is reached
after a finite time interval and there is no sliding mode in the reaching phase.
As follows from (3.14)

d
s̈(t) = d(t) − M1 sign(s(t)) − M2 sign(ṡ(t)), d(t) = (∇  s f ) . (3.16)
dt
Parameters should satisfy inequalities

M2 > |d|, M1 > M2 + |d|,

where d is a function of the states s(x) and u, u is a state variable like x or s and
it is common for relay system to assume that the domain of initial conditions is
bounded (otherwise, unstable plant cannot be stabilized). Then M2 sign(ṡ) − d and
M1 sign(s) may be considered as function φ and F in Eq. (3.12) of the mechanical
system; therefore, the motion in space (s, ṡ) is asymptotically stable. Condition
M1 > M2 + |d| means that sliding mode cannot appear on the line ṡ = 0, if s = 0,
because s̈ cannot change sign, which contradicts existence conditions (2.2).
And now about the finite-time convergence. Intuitively, the finite-time conver-
gence can readily be explained from the mechanical standpoint. The rate of energy
dissipation is proportional to friction. It is always not less than M2 − |d|, and there-
fore the energy is dissipated at a non-infinitesimal rate; and it will be equal zero after
final interval of time, and therefore both s and ṡ will be equal to zero as well.
Finite-time convergence can be proven easily for the system with no disturbance.
The states cannot remain sign constant (Fig. 3.2). Finding the solution to Eq. (3.16)
with d = 0 and a piecewise constant right-hand side (M1 > M2 ) is a trivial problem.
The results of the calculations:

|si+1 | = γ|si |,

where
M1 − M2 √
si = s(ti ), γ = , Δti = ti+1 − ti = δ |si | = δ γ i |s0 |,
+ M2
M 1 √
2 2(M1 − M2 )
δ= + .
M1 − M2 M1 + M2
44 3 Design Principles

Gain γ < 1, and therefore s converges to zero. Sequence Δti constitutes a geometric
√
progression also with gain γ < 1, which means finite-time convergence.
Finite-time convergence can be demonstrated based on the homogeneity condi-
tion. The right-hand side of the motion equation written in the form
   
y1 s
ẏ = f (y), y = = ,
 y 2 ṡ 
y2
f (y) =
−M1 sign(y1 ) − M2 sign(y2 )

is a particular case of functions with the homogeneity property

f (C y) = βC f (y), C = diag{ci }, i = 1, 2 β, ci − scalar parameters.

If the system is asymptotically stable and 0 < ci < 1, β > 1, the state converges to
the origin in a finite time [18]. Function f satisfies the homogeneity condition with

β = 2, c1 = 0.25, c2 = 0.5.

Asymptotic stability was shown above based on mechanical analogy. It proves

the finite-time convergence property of the twisting algorithm without disturbance.
Note that asymptotic stability takes place in the system with bounded disturbances
d(y1 , y2 , t),
|d| ≤ d0 = const,

if M2 > d0 . Homogeneity conditions, complemented by c2 β ≥ 1 (in our case c2 β =

1), mean finite-time convergence of the twisting algorithm with disturbances as well
[19].

3.9.2 Super-Twisting Algorithm

As before, the control input of the super-twisting algorithm [17] (Fig. 3.3) is a con-
tinuous state function
√ 
u(t) = v(t) − a |s(t)|sign(s(t)),
(3.17)
v̇(t) = M sign(s(t)), a, M are positive constants.

Formally, super-twisting control is two-dimensional, since real control and input of

artificially introduced integrator

u 1 = −M sign(s)
3.9 Second-Order Sliding Mode Control 45

Fig. 3.3 Super-twisting

control

are selected by a designer. At the same time, super-twisting control is the only
example when time derivative of the output s is not needed for SOSM.
Analyze stability of the motion in space (s, ṡ), governed by the second-order
equation
ṡ
s̈ = −a √ − M sign(s) + d. (3.18)
|s|

Mechanical model can be utilized to prove asymptotic stability if d = 0. Indeed,

ṡ
functions a √ and M sign(s) satisfy the conditions for φ and F in (3.13). As
|s|
shown in [17], the finite-time convergence takes place in the system in the presence
of bounded disturbances d.
The finite-time convergence of systems with super-twisting control and no distur-
bances follows from the homogeneity property as well. Asymptotic stability again
was demonstrated based on the mechanical analogy in this subsection. The motion
equation can be written in the form
   
y1 s
ẏ = f (y), y = = ,
y2 ṡ
 
y2
f (y) = −a √y2 − M sign(y ) .
|y | 1
1

Again function f satisfies the homogeneity condition with

β = 2, c1 = 0.25, c2 = 0.5,

which proves finite-time convergence. The general form of homogeneity conditions

will be discussed in the next section along with application for super-twisting systems
with non-zero disturbances.
Remark 3.5 As in the conventional sliding mode, enforcing the second-order sliding
mode in the twisting algorithm implies that the relative degree between the discontin-
uous control and the input of the element, implementing this control, is equal to one,
since ṡ is needed. The remarkable property of the super-twisting algorithm is that
the sliding mode is enforced with the relative degree between discontinuous input
−M sign(s) and the output s equal to two, and time derivative ṡ is not needed. The
property takes place in a special case only, when the√ continuous part of the equation
does not satisfy the Lipschitz condition (function |s|). Finally, as follows from
46 3 Design Principles

the motion equations, functions s and ṡ cannot remain sign constant, and they have
interlacing zeros. Therefore, the algorithms were called twisting and super-twisting.

Mechanical energy-based approach being applied to super-twisting Eq. (3.17)

leads to asymptotic stability with zero disturbance only, since the both conditions
ṡq1 (s, ṡ) > 0, sq0 (s) > 0 are satisfied for F(t) = 0 as well. Equations (3.15), (3.17)
with F(t) = 0 are particular cases of the above model of mechanical system with

q0 = M̄ sign(s), M̄ = M0 or M̄ = M.

Lyapunov function for these particular cases in the form

1 2
V = ṡ + M̄|s|
2
was offered in [20] or
1 2 k
V = ṡ + M̄|s| + s 2 ,
2 2
if control is complemented by a linear term ks.

3.10 Suboptimal Control

The so-called suboptimal controller [21, 22] can be classified as second-order sliding
control, since the origin in plane (s, ṡ) is reached after a finite time interval the with
no sliding modes at this interval. The design idea stems from time optimal control
for system
ẍ = u

with
|u| ≤ M = const.

For initial conditions x(0) > 0, ẋ(0) = 0, the optimal process consists of two inter-
vals. The state x decays from x(0) to x(0)/2 at the first interval with u = −M and
then from x(0)/2 to 0 at the second interval with u = M. Apply the design method-
ology to the equation

s̈(t) = u(t)+ d(t), |d(t)| ≤ d0 = const,

−M1 if s(t) − s ∗ /2 > 0, (3.19)
u(t) =
M2 if s(t) − s ∗ /2 ≤ 0,

where M1 > 0, M2 > 0 are constant to properly be chosen according to [21], s ∗ is

the value of s detected at the last time moment when ṡ(t) = 0. We define s ∗ = 0
at the initial instant of time. Note that the time instant when ṡ = 0 is needed for
3.10 Suboptimal Control 47

Fig. 3.4 Suboptimal control

implementation (3.19) rather than time function ṡ(t). It can be found estimating
difference between s(t) and s(t + δ) for small value of δ.
For initial conditions
s(0) > 0, ṡ(0) = 0

and Mi > d0 , the value of s is decaying until s(t1 ) = s(0)/2 (Fig. 3.4). Control u
is equal to M2 for t > t1 until ṡ(t2 ) = 0. The gains M1 and M2 are properly chosen
according to [22]. It is shown that for all t2i and Δti = t2i − t2i−2 , i = 1, 2, . . . the
conditions
s(t2i ) < γs(t2i−2 ),
δΔti < δΔti−1 , 0 < γ < 1 and 0 < δ < 1

s(t) = 0, ṡ(t) = 0 after a finite time interval, since sum of geo-

hold. It means that 
∞
metric progression i=1 Δti is bounded. As a result, the second-order sliding mode
is enforced in the system with an unknown disturbance d. Similar to the twisting
algorithm, implementation of the control needs a time derivative of s; therefore,
the relative degree is equal to one as for all types of the conventional sliding mode
control.

References

1. Levant, A.: Sliding order and sliding accuracy in sliding mode control. Int. J. Control. 58(6),
1247–1263 (1993)
2. Shtessel, Y., Edwards, C., Fridman, L., Levant, A.: Sliding Mode Control and Observation.
Control Engineering Series. Birkhauser, NY (2014)
3. Loukyanov, A., Utkin, V.: Methods of reducing equations of dynamic systems to a regular
form. Autom. Remote. Control. 42(4), 413–420 (1981)
4. Utkin, V.: Sliding Modes in Control and Optimization. Springer, Berlin (1992)
5. Utkin, V., Guldner, J., Shi, J.: Sliding Mode Control in Electro-Mechanical Systems. CRC
Press (2009)
6. Drazenovic, B.: The invariance conditions in variable structure systems. Automatica 5(3),
287–295 (1969)
48 3 Design Principles

7. Atassi, A.N., Khalil, H.K.: Separation results for stabilization of nonlinear systems using dif-
ferent high-gain observer designs. Syst. Control. Lett. 39, 183–191 (2000)
8. Oh, S., Khalil, H.K.: Nonlinear output feedback tracking using high-gain observer and variable
structure control. Automatica 33, 1845–1856 (1997)
9. Edwards, C., Spurgeon, S.: Sliding Mode Control: Theory and Applications. Taylor and Francis
(1998)
10. Poznyak, A.: Advanced Mathematical Tools for Automatic Control Engineers. Deterministic
Technique, vol. 1. Elsevier, Amsterdam, Netherlands (2008)
11. Utkin, V.I., Shi, J.: Integral sliding mode in systems operating under uncertainty conditions.
In: Proceedings of IEEE CDC, vol. 4, Kobe, Japan, pp. 4591–4596 (1996)
12. Fridman, L., Poznyak, A., Bejarano, F.J.: Robust Output LQ Optimal Control via Integral
Sliding Modes. Birkhäuser, Springer Science, Series Systems and Control: Foundations and
Applications, New York (2014)
13. Gutman, S.: Uncertain dynamic systems—a lyapunov min-max approach. IEEE Trans. Autom.
Control AC-24, 437–449 (1979)
14. Gutman, S. Leitmann, G.: Stabilizing feedback control for dynamic systems with bounded
uncertainties. In: Confrence on Decision and Control, pp. 94–99 (1976)
15. Orlov, Y., Utkin, V.: Unit sliding mode control in infinite-dimensional systems. J. Appl. Math.
Comput. Sci. 8, 7–20 (1998)
16. Khalil, H.: Nonlinear Systems. Prentice Hall (2002)
17. Levant, A.: Principles of 2-sliding mode control design. Automatica 43(4), 576–586 (2007)
18. Bhat, S.P., Bernstein, D.S.: Geometric homogeneity with applications to finite-time stability.
Math. Control Signals Syst. 17(2), 101–127 (2005)
19. Utkin, V.: Mechanical energy-based lyapunov function design for twisting and super-twisting
sliding mode control. IMA J. Math. Control Inf. 32(4), 675–688 (2015)
20. Orlov, Y.: Finite time stability and robust control synthesis of uncertain switched systems.
SIAM J. Control Optim. 43(4), 1253–1271 (2005)
21. Bartolini, G., Ferrara, A., Usai, E.: Chattering avoidance by second-order sliding mode control.
IEEE Trans. Autom. Control 43(2), 241–246 (1998)
22. Bartolini, G., Pisano, A., Punta, E., Usai, E.: A survey of application of second-order sliding
mode control to mechanical systems. Int. J. Control 76(9/10), 875–892 (2003)
Chapter 4
Lyapunov Stability Tools for Sliding
Modes

Abstract In this chapter, the Lyapunov stability tools for sliding modes analysis are
considered. The notion of finite-time stabilization is introduced, and the finite-time
suitability analysis is discussed based on the homogeneity property of the considered
controllable system. The realization of fixed-time stability is also presented.

Keywords Lyapunov stability tools · Finite-time suitability · Homogeneity ·

Fixed-time stability

The design methodology of sliding mode control implies the selection of sliding
manifold with the desired motion properties, and then selection of the discontinuous
control, enforcing the state to reach this manifold, and the existence of sliding mode.
Sliding mode is governed by equation with Lipschitz continuous right-hand side, and
any design method of the conventional control theory can be applied. The second
problem dealing with reaching and existence condition is non-trivial reduced-order
stability one due to discontinuities in control. It means that the design methodology
for sliding mode control is developed mainly in the context of reaching conditions.
The objective of this development is finite-time convergence to the sliding manifold
and estimation of the reaching time. For the conventional sliding mode with m-
dimensional control and (n − m)-dimensional sliding mode equations, the design
procedures based on Lyapunov function were demonstrated in the previous section.
The above problem becomes more complicated for SOSM, since it should be solved
using one-dimensional control input for the nonlinear two-dimensional system. As
mentioned above, the lower estimates of reaching time were found in [1, 4] and
they can be approached asymptotically. This section demonstrates how the reaching
problem with uncertainty conditions can be solved based on Lyapunov functions.
The Lyapunov function method, being initially proposed for smooth dynamic
systems, remains a classical analysis tool for Filippov’s dynamics in discontinu-
ous systems (see [3–5]), whose trajectories are viewed in the sense of Filippov or
according to the equivalent control method. It is well recognized to be useful not
only for stability analysis but also for the synthesis of robust sliding motions. In
order to analyze sliding mode systems, differentiable Lyapunov functions suffice in
many cases. For instance, the Lyapunov analysis of the conventional sliding mode
algorithms, demonstrated above, is first applied for the motion beyond the switching
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 49
V. Utkin et al., Road Map for Sliding Mode Control Design,
SpringerBriefs in Mathematics,
https://doi.org/10.1007/978-3-030-41709-3_4
50 4 Lyapunov Stability Tools for Sliding Modes

manifold to establish the attractiveness of this manifold. Once the trajectories are on
the switching manifold, smooth dynamics are restored, and the standard Lyapunov
theory is in force provided that the SM exists at any point of the manifold.
Recently, non-smooth Lyapunov functions were involved to analyze finite-time
stability of the so-called second-order sliding mode (SOSM) systems. In this kind
of systems, sliding motions are known to occur only on a discontinuity manifold
of co-dimension 2 while the system dynamics cross discontinuity manifolds of co-
dimension 1 countably many times within a finite time interval, preceding the SOSM.
Available methods of proving finite-time stability of such SM systems are reviewed
and illustrated next.

4.1 Finite-Time Stabilization: What is This?

Interest in finite-time stabilization dates back to the early work on optimal control
design, presenting the so-called Fuller effect [6]. The optimal control for the second-
order system

ẍ(t) = u(x(t), ẋ(t)), t > 0, x(t), u(t) ∈ R, |u| ≤ 1 (4.1)

∞
with the criterion 0 x 2 (t)dt is the discontinuous state function

u(x, ẋ) = −sign(s(x, ẋ)), s(x, ẋ) = x + c ẋ|ẋ|. (4.2)

The constant parameter c > 0 is such that the optimal trajectories cross the switching
line s = 0 at a countably infinite number of points, located within a finite time interval.
The time intervals between successive switches decrease as a geometric progression
and result in a finite accumulation point, known as Zeno behavior. The existence of
a finite accumulation point means finite settling time or finite-time convergence.
The system with SOSM depicted in Fig. 4.1 is identical to the above optimal
control system. Both systems are governed by similar second-order equations with
parabolic switching lines. For high values of c, the Fuller effect disappears because
the sliding mode occurs on a full measure time interval (see Fig. 4.2).
SOSM in twisting and super-twisting systems is different from the above system
with the Fuller effect: the conventional sliding mode in the reaching phase for them
cannot appear for any parameter combination. At the same time, twisting and super-
twisting systems exhibit the properties of the Fuller effect: infinite countable number
of switching points at a finite time interval and finite-time convergence.
4.2 Strict Lyapunov Functions 51

Fig. 4.1 Process without

sliding mode in the reaching
stage

Fig. 4.2 Fuller effect

disappearance

4.2 Strict Lyapunov Functions

Currently, the most popular mathematical technique of proving finite-time stability

of a n-vector system
ẋ(t) = φ(t, x(t)) (4.3)

with a piecewise continuous (or even locally measurable) right-hand side is finding a
Lyapunov composite function V (x(t)) such that its time derivative along the system
trajectories admits estimation

V̇ (x(t)) ≤ −q V ρ (x(t)) (4.4)

for almost all t and for some constants q > 0, ρ ∈ [0, 1). For such a function, hereafter
referred to as a strict Lyapunov function, the expression

V (x(t)) = 0

holds true (cf. [4, Lemma 4.3]) for all t ≥ Tste , where

Tste = [q(1 − ρ)]−1 V 1−ρ (x(0)) (4.5)

is a settling time estimate. A similar analysis can be provided for non-Lipschitz

Lyapunov functions using a generalized framework (see, e.g., [8–10]).
52 4 Lyapunov Stability Tools for Sliding Modes

A strict Lyapunov function is harder to design than an exponentially decaying

one, resulting from inequality (4.4) with ρ = 1. There are various examples (see
[7, 11]) in the sliding mode literature where a strict Lyapunov function is identified.
The methodology, proposed in [12], is based on partitioning the state space of (4.3)
into subdomains, in which a Lyapunov function V (x) is designed analytically, and
it relies on solutions of the first-order partial differential equation

∂V
φ̃(x) = −q V ρ . (4.6)
∂x
References [13, 14] identify a strict Lyapunov function for the super-twisting system
√ ⎫
ṡ1 = s2 − k1 |s1 |signs1 , ⎬
(4.7)
⎭
ṡ2 = −k2 signs1 + w(t, s1 , s2 ), k1 , k2 > 0,

written in terms of the variables s1 , s2 and affected by uniformly bounded distur-

bances which additionally admit the dependence on the state variable x and such that
the magnitude estimate
|w(t, s1 , s2 )| ≤ M = const (4.8)

holds almost everywhere. The Lyapunov function for (4.7) is selected as a quadratic
form
Vstw (s1 , s2 ) = ζ  Pζ (4.9)

with the positive-definite matrix

 
1 4k2 + k12 −k1
P= (4.10)
2 −k1 2

with respect to 

ζ = |s1 |signs1 , s2 . (4.11)

Taking into account that no sliding modes occur on the discontinuity surface x1 =
0 but the origin, the time derivative of the Lyapunov function Vstw (x(t)) is then
estimated along the trajectories of (4.7). Provided that

k2 > 3M + 2M 2 k1−2 , (4.12)

1 1/2
the desired upper estimate (4.4) is shown to hold true with ρ = , q = λmin
2
(P)λ−1
max (P)λmin (Q), and

 
1 2k2 + k12 − M −k1 − 2Mk1−1
Q= > 0. (4.13)
2 −k1 − 2Mk1−1 1
4.2 Strict Lyapunov Functions 53

Since the above relations are found in a computable form inequality (4.12) can be
viewed as a tuning parameter rule, guaranteeing the state convergence to zero in a
predetermined finite time given by (4.5).
As shown in [15], specifying the Lyapunov function of the super-twisting dynam-
ics (4.7), (4.8) in the form

1 1
2
Vstw (x1 , x2 ) = 2k2 |x1 | + x22 + e2 − k1 |x1 |sign x1 (4.14)
2 2
1
yields inequality (4.4) with ρ = and with the explicit decay rate
2
 
2(k1 k2 − M − Mk1 ) k1 − 2M
q = 2k2 min , . (4.15)
3k12 + 4k2 4

A less conservative tuning parameter rule

k1 k1 k2
min{ , }>M (4.16)
2 1 + k1

is thus obtained: for instance, given M >> 1, the tuning rule (4.16) requires that k2 >
M, whereas (4.12) would require k2 > 3M. Coupled to the settling time estimate
(4.5), relations (4.15), (4.16) allow one to tune the super-twisting gains k1 and k2
to have any desired settling time. An alternative settling time assignment, which is
based on integral HOSM design, can be found in [16].

4.3 Finite-Time Stability and Homogeneity

An alternative to proving finite-time stability through strict Lyapunov functions is

via first proving asymptotic stability by identifying a non-strict Lyapunov function
[17, 18]) and applying an extension of the invariance principle to discontinuous sys-
tems [19]). Once the asymptotic stability is established, the finite-time convergence
of trajectories is verified by using weighted homogeneity [20] of system dynamics.
The homogeneity concept for a discontinuous system (4.3) is introduced in [21].
Definition 4.1 A piecewise continuous vector function (vector field) φ = (φ1 , φ2 ,
. . . , φn ) of the variables t ∈ R and x = (x1 , x2 , . . . , xn ) ∈ Rn is said to be locally
homogeneous of degree κ ∈ R with respect to dilation r = (r1 , r2 , . . . , rn ) , where
ri > 0, i = 1, 2, . . . , n (as well as system (4.3) is) if there exists a constant c0 > 0
and a ball Bδ ∈ Rn such that

φi (c−κt, cr1 x1 , cr2 x2 , . . . , crn xn ) = cκ+ri φi (t, x1 , x2 , . . . , xn ) (4.17)

for all c ≥ c0 and almost all

54 4 Lyapunov Stability Tools for Sliding Modes

(x1 , x2 , . . . , xn ) ∈ Bδ , (cr1 x1 , cr2 x2 , . . . , crn xn ) ∈ Bδ ,

and almost all t ∈ R. Provided that relation (4.17) is additionally satisfied almost
everywhere in R × Rn , the function φ and system (2.1) are said to be globally
homogeneous of the same degree and dilation.
Recall that the homogeneity of a Lyapunov function V (x), x ∈ Rn is as follows.

Definition 4.2 A scalar continuous function V (x) is called homogeneous of degree

p ∈ R with respect to dilation r1 , r2 , . . . , rn , where ri > 0, i = 1, . . . , n, if

V cr1 x1 , cr2 x2 , . . . , crn xn = c p V (x1 , x2 , . . . , xn )

for all (x1 , x2 , . . . , xn ) ∈ Rn and for all c > 0.

Example 4.1
1. The function f (x) = x 4 is homogeneous of degree p = 4 and dilatation r = p = 4
since
f (cx) = (cx)4 = c4 x 4 = c4 f (x).

2. The function f (x1 , x2 ) = x12 + x24 is homogeneous of degree p and dilatation

(r1 = p/2, r2 = p/4). Indeed,

f (cr1 x1 , cr2 x2 ) = (cr1 x1 )2 + (cr1 x1 )4 =

c2r1 =c4r2 
c2r1 x12 + c4r2 x24 = c2r1 x12 + x24 = c p f (x1 , x2 )
for dilatations r1 = p/2, r2 = p/4 and any p > 0.

Both the disturbance-free twisting and super-twisting systems (3.16) and (4.7)
prove to be globally homogeneous of degree κ = −1 and dilation r = (2, 1) while
straightforwardly validating relation (4.17), properly specified for these systems. It
should be noted, however, that their perturbed versions (3.16) and (4.7) are no longer
homogeneous because of the presence of the disturbance term w(t, x1 , x2 ). These
systems may, however, be viewed as quasihomogeneous ones if embedded into the
framework of a differential inclusion

ẍ(t) ∈ (x(t), ẋ(t)) + Ψ. (4.18)

Then the homogeneity is defined in a similar manner to Definition 4.1, given above.
The right-hand side of (4.18) is a multivalued function, composed of the Filippov set
(x, ẋ) of the discontinuous function

φ(x, ẋ) = −αsignx − βsign ẋ

and the permanent set

Ψ = {x ∈ R : |x| ≤ M},
4.3 Finite-Time Stability and Homogeneity 55

formed by admissible disturbances w(t, x1 , x2 ) subject to the rectangular constraint

(4.8) on their magnitude. The differential inclusion (4.18), inspected with the embed-
ded twisting algorithm and with the captured class of external disturbances, is shown
[7] to be globally homogeneous of the same degree κ = −1 and dilation r = (2, 1)
as that of the disturbance-free twisting system (4.7). Just in case, such a generator
of a homogeneous differential inclusion (4.18) is referred to as a quasihomogeneous
system of the corresponding degree and dilation.
There is considerable literature employing the homogeneity technique for both
continuous and discontinuous vector fields. It is demonstrated in [22, 23] that a
globally asymptotically stable continuous homogeneous vector field admits a global
homogeneous Lyapunov function, and it is then extended in [24] to discontinuous
vector fields. Reference [25] employs the approach of [23] to establish finite-time
stability of continuous vector fields when the degree of homogeneity is negative by
showing the existence of strict Lyapunov functions. Locally (quasi-) homogeneous
functions are well recognized [21] to result in finite-time convergence of discontin-
uous systems as well when the (quasi-) homogeneity degree is negative [21]. The
homogeneous HOSM, proposed in [26], proves finite-time convergence of trajecto-
ries also using homogeneity of discontinuous controllers.
The majorant curve technique [27, 28] and homogeneous domination-based
approach [29] employ the design of a homogeneous output feedback controller such
that the controller stabilizes the nominal linear system in the presence of the uncer-
tainties. The principles of homogeneous domination at zero and at infinity have been
introduced originally in [22] [Theorems 6, 7]. Based on the philosophy of homo-
geneous domination, reference [30] gives a mathematical tool called homogeneity
in the bi-limit, which defines a homogeneous approximation. The connection with
the homogeneous domination approach is made clear in [30, Sect. 5]. The connec-
tion to uniform finite-time convergence of the trajectories of a class of nonlinear
autonomous systems is also obtained in [30, Corollary 2.24, Example 5.5] using the
concept of homogeneity in the bi-limit.
A philosophically similar concept of quasihomogeneity, concerning discontinu-
ous systems, leads to the uniform finite-time stability of the origin in the presence
of bounded persistent external disturbances, resulting in the equiuniform finite-time
stability [4, Sect. 4.4]. Reference [21] is the first article where such equiuniform
finite-time stability is investigated under uniformly bounded persistent external dis-
turbances. An equiuniform settling time estimate is obtained with respect to the entire
class of such disturbances. The proposed equiuniform finite-time stability analysis
is based on the premise of equiuniform asymptotic stability and is illustrated by the
PD-augmented perturbed twisting system

ṡ1 = s2 ,
(4.19)
ṡ2 = −αsign(s1 ) − βsign(s2 ) − hs1 − ps2 + ω

with the gains h, p > 0 [21].

First, equiuniform stability of (4.19) is verified with the non-strict Lyapunov
function
56 4 Lyapunov Stability Tools for Sliding Modes

1
Ṽ (s1 , s2 ) = α|s1 | + (s22 + hs12 ). (4.20)
2
Indeed, differentiating (4.20) along the perturbed system (4.19), initialized outside
the origin, yields the equiuniform estimate

Ṽ˙ (s1 (t), s2 (t)) ≤ −(β − M)|s2 (t)| (4.21)

for almost all t regardless of whichever disturbance (4.8) affects the system. While
deriving (4.21), it has been taken into account that no sliding modes occur along the
discontinuity lines x1 = 0 and x2 = 0 except the origin x1 = 0 and x2 = 0.
Next, a family of semiglobal Lyapunov functions

1
VR (s1 , s2 ) = Ṽ (s1 , s2 ) + μ R s1 s2 = α|x1 | + (s22 + hs12 ) + μ R s1 s2 , (4.22)
2
parameterized with R > 0, is involved with sufficiently small positive weight param-
eters such that each Lyapunov function (4.22) is positive definite on the corresponding
compact set
D R = {(s1 , s2 ) ∈ R2 : Ṽ (s1 , s2 ) ≤ R}. (4.23)

The parameter R defines the domain of the Lyapunov representative VR . The higher
the value of R, the smaller the value of the cross-term weight μ.
Since due to (4.21), all trajectories of the perturbed system (4.19), initialized
within compact (4.23), cannot escape from this compact, the estimate of the time
derivative of (4.22) along these trajectories is obtained in the form

V̇R (s1 (t), s2 (t)) ≤ −K R VR (s1 (t), s2 (t)) (4.24)

with some positive equiuniform exponential decay rate K R on compact (4.23). Cer-
tainly, the exponential decay (4.24) on compact (4.23) does not lead to the exponential
stability of the perturbed system (4.19) because the decay rate K R → 0 as R → ∞,
but it does guarantee the global equiuniform asymptotic stability of (4.19).
Finally, the equiuniform finite-time convergence is established based on the local
homogeneity of (4.19). Taking into account that the PD+disturbance term

−hx1 − px2 + ω(t, x1 , x2 )

remains locally uniformly bounded for admissible disturbances (4.8), system (4.19)
is embedded into the differential inclusion framework (4.18). This allows one to
ensure that the system is indeed locally (but not globally) quasihomogeneous of the
same degree κ = −1 and dilation r = (2, 1) as that of the disturbance-free twisting
system (4.19).
Thus, the PD-augmented perturbed twisting system (4.19) is globally equiuni-
formly asymptotically stable and it is locally quasihomogeneous of degree κ < 0.
4.3 Finite-Time Stability and Homogeneity 57

By applying the quasihomogeneity principle [4, Theorem 4.1], the global equiuni-
form finite-time stability of (4.19) is established.
In addition, the quasihomogeneity technique has successfully been applied in [31]
for tuning the twisting controller parameters to impose a predetermined settling time
estimate on the closed-loop system.
The homogeneous HOSM, proposed in [26], proves finite-time convergence of
trajectories also using homogeneity of discontinuous controllers.

4.4 Fixed-Time Stability

The notion of the so-called fixed-time stability refers to the case of finite-time stability
when the finite settling time can be specified independent of the initial conditions.
The fixed-time stability can be established in terms of a strict Lyapunov function V
that along the solutions of (2.1) satisfies a stronger version

V̇ (x(t)) ≤ −η(V (x(t))) (4.25)

of the Lyapunov derivative estimate (4.4) with a right-hand side such that the ODE

V̇ (x(t)) = −η(V (x(t)))

possesses fixed-time stable solutions. An example of such a function is

η(V ) = V + V3

Polyakov [32]. Indeed, the corresponding equation

V̇ (x(t)) = − V + V 3 (4.26)

possesses solutions with the convergence rate higher than that of

√
V̇ (x(t)) = − V 3 (4.27)

beyond the level set V = 1 and higher than that of

√
V̇ (x(t)) = − V (4.28)

within this level set. Since the solutions

V (t) = 4V (0)(2 + t V (0))−2

of (4.27) with initial conditions V (0) > 1 reach the level set V = 1 in
58 4 Lyapunov Stability Tools for Sliding Modes

T (V (0)) = 2( V (0) − 1)/ V (0) ≤ 2, (4.29)

whereas the solutions to (4.28), initialized within V (0) ≤ 1 and being a particular
case of (4.4), escape to zero according to (4.29) at most at T = 2, the solutions to
(4.26) escape to zero at most T = 4, regardless of initial conditions. Thus, (4.26) is
actually fixed-time stable.
For continuous vector fields, fixed-time stability has been derived in the context of
systems homogeneous in the bi-limit [30, Corollary 2.24]. The fixed-time stability
of a discontinuous super-twisting-based differentiator was proved in [33], where
it was shown that the predefined convergence time does not depend on the initial
condition. A more recent result in [32] proposes new nonlinear controllers for fixed-
time stabilization of linear control systems. The most recent reference in this area can
be found in [34], which studied the properties of the settling time function resulting
from the application of finite-time convergent sliding modes, giving rise to Filippov
inclusions. An important limitation of the fixed-time stable discontinuous systems
under forward Euler discretization was established in [34, Theorem 3]. It is shown
that there exist solutions of the inclusions whose norm escapes to infinity. Note
that the implicit Euler discretization does not have this limitation and preserves the
fixed-time convergence rate [35].

References

1. Utkin, V., Guldner, J., Shi, J.: Sliding Mode Control in Electro-Mechanical Systems. CRC
Press (2009)
2. Khalil, H.: Nonlinear Systems. Prentice Hall (2002)
3. Utkin, V.: Sliding Modes in Control and Optimization. Springer, Berlin (1992)
4. Orlov, Y.: Discontinuous Systems: Lyapunov Analysis and Robust Synthesis Under Uncertainty
Conditions. Springer, Berlin (2008)
5. Yakubovich, V., Leonov, G., Gelig, A.: Stability of Stationary Sets in Control Systems with
Discontinuous Nonlinearities. World Scientfic (2004)
6. Fuller, A.: Relay control systems optimized for various performance criteria. In: Proceedings
of the 1st IFAC Triennial World Congress, pp. 510–519 (1960)
7. Orlov, Y.: Finite time stability and robust control synthesis of uncertain switched systems.
SIAM J. Control Optim. 43(4), 1253–1271 (2005)
8. Roxin, E.: On finite stability in control systems. Rendiconti del Circolo Matematico di Palermo
15(3), 273–283 (1966)
9. Chellaboina, V., Leonessa, A., Haddad, W.: Generalized Lyapunov and invariant set theorems
for nonlinear dynamical systems. Syst. Control Lett. 38, 289–295 (1999)
10. Polyakov, A., Fridman, L.: Stability notions and Lyapunov functions for sliding mode control
systems. J. Frankin Inst. 351(4), 1831–1865 (2014)
11. Kamal, S., Moreno, J., Chalanga, A., Bandyopadhyay, B., Fridman, L.: Continuous terminal
sliding-mode controller. Automatica 69, 308–314 (2016)
12. Polyakov, A., Poznyak, A.: Reaching time estimation for super-twisting second order sliding
mode controller via lyapunov function designing. IEEE Transactions on Automatic Control
54(8), 1951–1955 (2009)
13. Moreno, J., Osorio, M.: A lyapunov approach to second-order sliding mode controllers and
observers. In: 47th IEEE Conference on Decision and Control (CDC), pp 2856–2861 (2008)
References 59

14. Moreno, J., Osorio, M.: Strict lyapunov functions for the super-twisting algorithm. IEEE Trans.
Autom. Control 57(4), 1035–1040 (2012)
15. Orlov, Y., Aoustin, Y., Chevellereau, C.: Finite time stabilization of a perturbed double integrator
- part i: continuous sliding mode-based output feedback synthesis. IEEE Trans. Autom. Control
56(3), 1035–1040 (2010)
16. Levant, A., Alelishvili, L.: Integral high order sliding modes. IEEE Trans. Autom. Control
52(7), 1278–1282 (2007)
17. Pyatnitskii, E.S.: Control of mechanical systems under uncertainty conditions in the absence of
quantitative information on the current state. Autom. Remote. Control 60(5), 739–743 (1999)
18. Orlov, Y.: Extended invariance principle for nonautonomous switched systems. IEEE Trans.
Autom. Control 48(5), 1448–1552 (2003a)
19. Alvarez, J., Orlov, Y., Acho, L.: An invariance principle for discontinuous dynamic systems
with application to a coulomb friction oscillator. ASME J. Dyn. Syst., Meas., Control 122(4),
687–690 (2000)
20. Zubov, V.I.: Methods of A.M. Lyapunov and Their Applications. Noordhoff, Leiden (1964)
21. Orlov, Y.: Finite time stability of homogeneous switched systems. In: 42nd IEEE Conference
on Decision and Control (CDC), pp. 4271–4276 (2003b)
22. Zubov, V.I.: On systems of ordinary differential equations with generalized homogenous right-
hand sides. Izvestia vuzov. Mathematica. 1, 80–88 (1958) (in Russian)
23. Rosier, L.: Homogeneous lyapunov function for homogeneous continuous vector field. Syst.
Control Lett. 19(6), 467–473 (1992a)
24. Rosier, L.: Inverse of lyapunov’s second theorem for measurable functions. In: Proceedings of
IFAC Symposium on Nonlinear Control Systems (NOLCOS), pp. 2856–2861 (1992b)
25. Bhat, S.P., Bernstein, D.S.: Geometric homogeneity with applications to finite-time stability.
Math. Control Signals Syst. 17(2), 101–127 (2005)
26. Levant, A.: Homogeneity approach to high-order sliding mode design. Automatica 41(5), 467–
473 (2005)
27. Levant, A.: Sliding order and sliding accuracy in sliding mode control. Int. J. Control 58(6),
1247–1263 (1993)
28. Orlov, Y., Aguilar, L., Cadiou, J.C.: Switched chattering control versus backlash/friction phe-
nomena in electrical servo-motors. Int. J. Control 76(9/10), 959–967 (2003)
29. Qian, C.: A homogeneous domination approach for global output feedback stabilization of
a class of nonlinear systems. In: Proceedings of American Control Conference (ACC), pp.
4708–4715 (2005)
30. Andrieu, V., Praly, L., Astolfi, A.: Homogeneous approximation, recursive observer design,
and output feedback. SIAM J. Control Optim. 47(4), 1814–1850 (2008)
31. Oza, H., Orlov, Y., Spurgeon, S.: Continuous uniform finite time stabilization of planar con-
trollable systems. SIAM J. Control Optim. 53(3), 1154–1181 (2015)
32. Polyakov, A.: Nonlinear feedback design for fixed-time stabilization of linear control systems.
IEEE Trans. Autom. Control 57(8), 2106–2110 (2012)
33. Cruz-Zavala, E., Moreno, J., Fridman, L.: Uniform robust exact differentiator. IEEE Trans.
Autom. Control 56(11), 2727–2733 (2011)
34. Levant, A.: On fixed and finite time stability in sliding mode control. In: 52nd IEEE Conference
on Decision and Control (CDC), pp. 4260–4265 (2013)
35. Polyakov, A., Efimov, D., Brogliato, B.: Consistent discretization of finite-time and fixed-time
stable systems. SIAM J. Control Optim. 57(1), 78–103 (2019)
Chapter 5
SM Observers

Abstract Most of the design methods considered above were developed under the
assumption that the state vector is available. In practice, however, only a part of
its components may be measured directly. The output feedback SMC method is
applicable to rather limited types of systems. An alternative approach is designing
asymptotic observers, which are dynamic systems for estimating all components
of the state vector using those measured directly. Sliding mode modifications for
state observation Chap. 14, [2, Chap. 6], and [3] are presented in this section. Time
derivatives are needed for design in canonical space and for high-order sliding mode
control. Differentiation design problem is discussed as well.

Keywords Observers for linear time-invariant systems · Uncertain systems ·

Differentiators

5.1 Observers for Linear Time-Invariant Systems

The conventional asymptotic state observer for LTI observable system


ẋ(t) = Ax(t) + Bu(t), y(t) = C x(t),
x(t) ∈ R n , u(t) ∈ R m , y(t) ∈ R l , l < n, rank(C) = l

has the same structure as the system under control with an additional input depending
on the mismatch between the output y(t) and its estimate

˙ = A x̂(t) + Bu(t) + L(y(t) − C x̂(t)), x̂(t) ∈ R n .

x̂(t) (5.1)

Matrix L can be selected such that the state estimate x̂(t) tends to the system state
x(t) at the desired rate. Let us proceed to the design of a state observer with inputs
as discontinuous functions of mismatches. Taking into account that the state vector
can be partitioned into two subvectors

x T = (x1T , x2T )T , x1 ∈ R n−l , x2 ∈ R l

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 61
V. Utkin et al., Road Map for Sliding Mode Control Design,
SpringerBriefs in Mathematics,
https://doi.org/10.1007/978-3-030-41709-3_5
62 5 SM Observers

such that
s = C1 x1 + C2 x2 , det(C2 ) = 0,

write down the observer equation with respect to estimates x̂1 and ŷ
⎫
x̂˙1 (t) = A11 x̂1 (t) + A12 ŷ(t) + B1 u(t) + Lv(t), ⎬
˙ = A21 x̂1 (t) + A22 ŷ(t) + B2 u(t) − v(t)
ŷ(t)
⎭
v(t) = M sign( ŷ(t) − y(t)), M > 0,

where all matrices are constant and are derived from the transformation of the
observer Eq. (5.1). The discontinuous vector function v(t) is chosen such that sliding
mode is enforced in the manifold

y(t) = ŷ(t) − y(t)

and the mismatch y(t) between the output y(t) and its estimate ŷ(t) is reduced to
zero. Matrix L must be found such that the mismatch x 1 (t) = x̂1 (t) − x1 (t) decays
at the desired rate. Sliding mode with y(t) = 0 occurs in the system with respect to
x 1 and y
ẋ 1 (t) = A11 x 1 (t) + A12 y(t) + Lv(t),
ẏ(t) = A21 x 1 (t) + A22 y(t) − v(t)

with high enough value of M > 0. Hence, for bounded initial conditions, sliding
mode can be enforced in manifold y = 0. To derive the sliding mode equation, the
solution
veq (t) = A21 x 1 (t)

to equation ẏ(t) = 0 with respect to v(t) should be substituted into the first equation

ẋ 1 (t) = (A11 + L A21 )x 1 (t).

The desired rate of convergence of x 1 (t) to zero and convergence of x̂1 (t) to x1 (t)
can be provided by a proper choice of matrix L [1] and then

x2 (t) = C2−1 (y(t) − C1 x1 (t)).

The observer with the input as a discontinuous function of the output in sliding mode
is equivalent to the linear reduced-order Luenberger observer [1]. However, if the
plant and observed signal are affected by noises, the nonlinear observer is preferable,
since it exhibits adaptivity properties with respect to varying parameters of the noises
[4].
5.2 Observers for Uncertain Systems 63

5.2 Observers for Uncertain Systems

5.2.1 Second-Order System

The main ideas of designing observers for nonlinear uncertain systems can be demon-
strated for the second-order system:

ẋ(t) = Ax(t) + d f (t, x(t), u(t)) + bu(t), y(t) = cx(t),

where x(t) ∈ R 2 , y(t) ∈ R, A, c, d, b are constant matrix and vectors, f is a scalar

bounded disturbance  f  < f 0 where f 0 is constant. The pair (A, c) is assumed to
be observable.
The observable system can be transformed into the form [1]

ẋ1 = a11 x1 + a12 y + d1 f + b1 u, a11 < 0.
ẏ = a21 x1 + a22 y + d2 f + b2 u.

The observer ⎫
d
x̂
dt 1
= a11 x̂1 + a12 y + b1 u, ⎬
d
ŷ = a21 x̂1 + a22 y + b2 u + v,
dt ⎭
v = −v0 sign( ŷ − y), v0 > |d2 | f 0

is designed, copying the parameters of the above system and assuming that d1 = 0.
The error equation has the form

ėx = a11 ex ,
ė y = a21 e y + d2 f + v.

The mismatch ex = x̂1 − x1 tends to zero (a11 < 0) and component x1 of the state
vector is found. Discontinuous input v enforces sliding mode with

e y = ŷ − y = 0

and
veq = d2 f,

which can be obtained using a low-pass filter [5, Chap. 2]. It means that the observer
can be applied for finding both the state vector and disturbance or for fault detection
if f models the effect of fault.
64 5 SM Observers

5.2.2 Uncertain System of an Arbitrary Order

The above method can be easily generalized for the uncertain system of an arbitrary
order
ẋ = Ax + Bu + D f (t, x, u), y = C x, (5.2)

where x ∈ R n , u ∈ R m y ∈ R p , A, C, D, B are constant matrices, f (t, x, u) ∈ R q

is a bounded disturbance. The pair (A, C) is assumed to be observable [6]. The
observable system can be transformed into the form
⎫
ẋ1 = A11 x1 + A12 y + D1 f + B1 u, ⎬
⎭
ẏ = A21 x1 + A22 y + D2 f + B2 u,

where all matrices are derived from (5.2) and A11 is a Hurwitz matrix (like a11 < 0).
Similar to the assumptions above D1 = 0 (like d1 = 0). Then the disturbance and
input v in the mismatch equations satisfy the matching conditions (see Sect. 3.5);
therefore, the motion in sliding mode does not depend on disturbance.
The equations of the observer and mismatches are of form
⎫
d
x̂1 = A11 x̂1 + A12 y + B1 u, ⎪
⎪
⎪
dt ⎬
d
ŷ = A21 x̂1 + A22 y + B2 u + v, ⎪
⎪
dt ⎪
⎭
v = −v0 sign(e y ),

and 
ėx = A11 ex ,
ė y = A21 e y + D2 f + v.

As for the second-order system vector x1 and the disturbance can be found. The state
and disturbance observer is designed in terms of the original system (A, B, C, D) in
[7, 8].
Further developments of the discussed methodology for state and fault estima-
tions can be found in [9, 10] for systems with noise in measurement [11], for full-
and reduced-order observers [12], for a set of interconnected systems [13], and for
systems with high-order sliding modes [14].
Of course the condition D1 = 0 is rather restrictive. It is not applicable for widely
used systems in canonical form

ẋi = xi+1 , i = 1, . . . , n − 1,
ẋn = f (t, x, u), y = x1 ).

The method does not work for our second-order example if

5.2 Observers for Uncertain Systems 65

d1 = 0, d2 = 0.

The design idea for the general case when condition D1 = 0 does not hold stems from
the so-called high-gain observer with specially selected hierarchy of gains tending
to infinity (discussed in section “Invariance”).
It is interesting that the canonical form for designing a high-gain observer [15]
is not obligatory, as shown in the paper, published in Russian [16]. The design
procedures for high-gain and sliding mode observers are outlined in this section.
Preliminarily demonstrate the idea of high-gain observer design for an arbitrary LTI
system. The first equation of the observer for the system in canonical form

ẋi = xi+1 , i = 1, . . . , n − 1,
ẋn = F (x1 , . . . , xn , t) + u,
y = x1

with respect to mismatch x̄i := x̂i − xi is

d
x̄1 := x̄2 − k1 x̄1 .
dt
If k1 → ∞, then k1 x̄1 tends to x̄2 . Similarly, known x2 can be used to find x3 with
corresponding hierarchy of high gains and so on. Now assume that y = x1 is not
scalar but output vector with equations ẋ1 = Ax, where x is a state vector and

d
x̄1 := A x̄ − k1 x̄1 .
dt
If scalar gain k1 tends to infinity, then k1 x̄1 tends to A x̄. Similarly, vector y1 = A x̄
can be used to find the next subvector of the state again with corresponding hierarchy
of the observer gains.
Now describe in details the high-gain observer design. Any LTI observable system

ẋ = Ax + Bu, y = C x, Cis a full rank matrix

can be represented in the form

ẏ = A11 y + A12 x1 + B1 u,
ẋ1 = A21 y + A22 x1 + B2 u,

where the pair (A22 , A12 ) is observable. If gain k0 → ∞ in the observer

ŷ˙ = A11 ŷ + A12 x̂1 + B1 u + v0 , v0 = k0 (y − ŷ),

x̂˙1 = A21 ŷ + A22 x̂1 + B2 u + vx1 ,
66 5 SM Observers

then ŷ → y and v0 → A12 (x̂1 − x1 ), rank A12 > 0 (otherwise, the system is not
observable), where vx1 to be defined next. Matrix R exists such that C1 = R A12
is a full rank matrix. Then

y1 = R(A12 x̂1 − v0 ) = R A12 x1 = C1 x1 .

If dim(y1 ) ≥ dim(x1 ), then x1 can be found from y1 and x can be found from y = C x
and x1 . It is important that the equation for x1 is not needed for this case. The equation
for which the output vector y1 = C1 x1 can be handled similar to the original system
if dim(y1 ) < dim(x1 ). It can be partitioned into two equations with respect to y1
and new variable x2 with new output vector y2 = C2 x and vx2 instead of vx1 . Again
vector y2 can be obtained, if v1 = k1 (y1 − ŷ1 ), k1 → +∞, k0 /k1 → ∞.
After a finite number of steps k with the hierarchy of gains kk−1 → ∞,
ki
→ ∞ the condition dim(yk ) ≥ dim(xk ) holds and xk can be found
ki+1 · . . . · kk−1
from the algebraic equation yk = Ck xk . All coordinate transformations are non-
singular, and the remaining components of the state vector can be calculated from
algebraic equations y = C x, yi = Ci xi , i = 1, . . . , k − 1. The last equations for xi
is not needed; therefore, its right-hand side may depend on any types of uncertainties
(unknown parameters, disturbances, nonlinear functions) similar to the systems in
canonical form with uncertainties in the last equation.
The high-gain design method is generalized for observers with finite discontinuous
inputs. Input v0 = k0 (y − ŷ) is replaced by v0 = M0 sign(y − ŷ) with finite gain
M0 . After sliding mode occurs, the average value of v0 is equal to A12 (x̂1 − x1 )
as well and can be obtained using a low-pass filter. Similarly, v1 is replaced by
v1 = M1 sign(y1 − ŷ1 ) and so on. All state components can be found in the observer
with appropriate hierarchy of low-pass filter time constants.
Similar multistep procedure is applicable for both linear and nonlinear systems
[17, 18]. This was demonstrated for time-varying linear systems in [5, Chap. 6].

5.2.3 Analysis of the State Estimation Error Convergence

for Matched Disturbances

The observer should be designed for the uncertain system


ẋ(t) = Ax(t) + Bu(t) + f (t, x(t), u(t)), y(t) = C x(t),
x(t) ∈ R n , u(t) ∈ R m , y(t) ∈ R p , rank(C) = p < n,

such that the mismatch e(t) = x̂(t) − x(t) between the estimate and system state is
reduced to zero. The disturbance is assumed to satisfy the matching condition

f (t, x(t), u(t)) = Bξ(t, x(t), u(t)),

ξ(t, x, u) ≤ k1 u + k2 y + α0 , y = C x,
5.2 Observers for Uncertain Systems 67

where k1 , k2 , α0 are constants and u ∈ R m , x ∈ R n , y ∈ R p . If the pair (A, C) is

observable, then there exists matrix L such that matrix A + LC has stable eigenval-
ues, i.e., for any Q > 0 the solution P > 0 to Lyapunov equation

(A + LC)T P + P(A + LC) = −Q

can be found.
Assumption. There exist matrices P, Q ∈ R n×n , L ∈ R n× p , F ∈ R m× p such that
C T F T = P B.
If this condition holds, then the amplitude of discontinuous control can be equal
to any upper estimate of the disturbance.
The observer equations [2] are selected with discontinuous input in the form of
unit control, introduced in Chapter IV
⎫
˙ = A x̂(t) + Bu(t) + L(y(t) − C x̂(t)) + v(t), ⎪
x̂(t) ⎪
⎬
−1 T T F(y(t) − C x̂(t))
v(t) = −ρ(u(t), y(t))P C F ,
F(y(t) − C x̂(t)) ⎪⎪
⎭
ρ(u, y) = k1 u + k2 y + k0 , k0 > α0 .

Calculate time derivative of

V (e(t)) = e T (t)Pe(t)

on the trajectories of the system

ė(t) = (A + LC)e(t) − Bξ(t, x(t), u(t)) + v(t).

We have ⎫
V̇ (e(t)) ≤ −e T (t)Qe(t) + 2e T C T F T ξ(t, x(t), u(t)) ⎪
⎪
FCe(t) ⎬
−2ρ (u(t), y(t)) e T (t)C T F T ≤
FCe(t) ⎪
⎪
⎭
−e T (t)Qe(t) − 2(k0 − α0 )FCe(t) < 0.

It means that the mismatch tends to zero, and the estimate tends to the state vector.
The design method is applicable under structural constraint

CT FT = P B

and for disturbance, depending on the control and output only. If the disturbance
depends on the state vector, then convergence of the mismatch to zero can be guaran-
teed for a bounded domain of initial conditions similar to the observer in the previous
section.
68 5 SM Observers

5.2.4 Differentiators

5.2.4.1 Single Differentiation

Implementation of sliding mode control both for systems in canonical space of 60s
and for systems with high-order sliding mode [19, Chap. 6] needs time derivatives
of different orders. During the entire history of sliding mode control theory, the
researchers tried to use sliding mode for designing differentiators. The simplest one
is shown in Fig. 5.1

ṡ(t) = f˙(t) − v(t), v(t) = M sign(s(t)),

where s(t) = f (t) − x(t). If M > | f˙|, then sliding mode occurs with x(t) ≡ f (t).
It means that the input of the integrator should be close to time derivative of f (t),
but the input v(t) is a discontinuous function. Explanation of this discrepancy is
simple: signal v(t) consists of high-frequency and average components. The high-
frequency one is filtered out by the integrator, while the average component is equal
to time derivative of f (t). The first-order low-pass filter is added to filter out the
high-frequency component and its output z(t) should be close to f˙(t). This heuristic
motivation can be confirmed analytically.
The block diagram of the two-loop differentiator for estimation of the first and
second time derivatives of function f (t) is shown in Fig. 5.2. The procedure of
selection of parameters τ , τ1 , μ and upper estimates of differentiation errors can be
found in [20].

5.2.4.2 Multiple Differentiation

The problem of multiple differentiation can be reformulated in terms of the dynamic

systems governed by equations

ẋi (t) = xi+1 (t), i = 1, . . . , n − 1

Functions xi , i = 2, . . . , n should be found if

f (t) = x1

Fig. 5.1 Block diagram of

the first-order sliding mode
differentiator
5.2 Observers for Uncertain Systems 69

Fig. 5.2 Block diagram of the second-order sliding mode differentiator

is measured. This problem is a particular case of designing an observer for a linear

system
ẋ(t) = Ax(t) + Bu(t), y = C x.

As discussed in the previous section, first the high-gain observer is designed with
appropriate hierarchy of gains. Then, high gains are replaced by discontinuous func-
tions with low-pass filters and appropriate hierarchy of the low-pass filter time con-
stants. The multistep methodology, offered in [16], is applied directly to the above-
formulated problem. After sliding mode occurs in the system
⎫
x̂˙1(t) = v1(t), ⎬
v1(t) = −M sign(x̂1(t)−x1(t)),
⎭
τ1 ż 1(t)+z 1(t) = v1(t),

z 1 is approximately equal to x2 the time derivative of x1 . The second derivative can

be estimated in the same way
⎫
x̂˙2(t) = v2(t), ⎬
v2(t) = −M sign(x̂2(t)−v1(t)),
⎭
τ2 ż 2(t)+z 2(t) = v2(t),

z 2 is approximately equal to x3 being the second time derivative of x1 . On the next

steps, the 3rd,…,(n − 1)th derivative can be estimated. The upper bounds of the
estimation errors can be found in [16].
The design method for super-twisting control can be utilized for the differentiation
problem as well [21]. To differentiate time function f (t), the super-twisting controller
is designed to reduce the error s(t) between its input f (t) and output x(t) to zero,
s = x − f: √ ⎫
ẋ(t) = −α |s(t)|sign(s(t)) + y(t), ⎬
⎭
ẏ(t) = −M sign(s(t))

with
| f¨(t)| < F0 = const
70 5 SM Observers

and M > F0 . As shown in [19], the error s(t) is reduced to zero after a finite time
interval and state component y(t) is equal to the first time derivative of function f (t).
If f (t) is corrupted by bounded noise

|s(t)| ≤ Δ = const,

then an upper bound of the differentiation error is estimated by inequality

√
|y(t) − f˙(t)| ≤ α1 Δ + α2 Δ,

where α1 and α2 are constant coefficients. The average value of

ẏ(t) = −M sign(s(t))

is close to second derivative of f (t) and it can be derived as output z(t) of low-pass
filter
τ ż(t) + z(t) = −M sign(s(t)).

References

1. Utkin, V.: Sliding Modes in Control and Optimization. Springer, Berlin (1992)
2. Edwards, C., Spurgeon, S.: Sliding Mode Control: Theory and Applications. Taylor and Francis
(1998)
3. Spurgeon, S.: Sliding mode observers: a survey. Int. J. Syst. Sci. 39(8), 751–764 (2008)
4. Drakunov, S.V.: On adaptive quasioptimal filter with discontinuous parameters. Autom. Remote
Control 44(9), 1167–1175 (1983)
5. Utkin, V., Guldner, J., Shi, J.: Sliding Mode Control in Electro-Mechanical Systems. CRC
Press (2009)
6. Edwards, Christopher, Spurgeon, Sarah K., Patton, Ron J.: Sliding mode observers for fault
detection and isolation. Automatica 36(4), 541–553 (2000)
7. Walcott, B.L., Zak, S.H.: Observation of dynamical systems in the presence of bounded non-
linearities/uncertainties. In: Confrence on Decision and Control, pp. 961–966 (1986)
8. Corless, M., Tu, J.: State and input estimation for a class of uncertain systems. Automatica
34(6), 757–764 (1998)
9. Edwards, C., Spurgeon, S.: On the development of discontinuous observers. Int. J. Control
59(5), 1211–1229 (1994)
10. Kalsi, K., Hui, S., Zak, S.H.: Unknown input and sensor fault estimation using sliding-mode
observers. In: American Control Conference, pp. 1364–1369 (2011)
11. Yang, J., Zhu, F., Wang, X., Xuhui, B.: Robust sliding-mode observer-based sensor fault esti-
mation, actuator fault detection and isolation for uncertain nonlinear systems. Int. J. Control
Autom. Syst. 13(5), 1037–1046 (2015)
12. Dhahri, S., Sellami, A., Hmida, F.B.: Fault detection and reconstruction via both full-order
and reduced-order sliding mode observers for uncertain nonlinear systems. In: International
Conference on Modelling, Identification and Contro (2015)
13. Menon, P.P., Edwards, C.: Fault estimation using relative information for a formation of dynam-
ical systems. In: Confrence on Decision and Control (2010)
References 71

14. Fridman, L., Levant, A., Davila, J.: Observation of linear systems with unknown inputs via
high-order sliding-modes. Int. J. Syst. Sci. 38(10), 773–791 (2008)
15. Oh, S., Khalil, H.K.: Nonlinear output feedback tracking using high-gain observer and variable
structure control. Automatica 33, 1845–1856 (1997)
16. Utkin, V.I.: Control problems for multi-connected plants. In: Motion Separation For Designing
State Observer, pp. 91–97. Nauka, Moscow (1983) (in Russian)
17. Drakunov, S.V.: Sliding-mode observers based on equivalent control method. In: Confrence on
Decision and Control, pp. 2368–2369 (1992)
18. Barbot, J.P., Boukhobza, T., Djemai, M.: Sliding mode observer for triangular input form. In:
Confrence on Decision and Control (1996)
19. Shtessel, Y., Edwards, C., Fridman, L., Levant, A.: Sliding Mode Control and Observation.
Control Engineering Series. Birkhauser, NY (2014)
20. Utkin, V.I.: Sliding Modes in Problems of Optimization and Control, Nauka, Moscow (1981)
(in Russian)
21. Levant, A.: Higher-order sliding modes, differentiation and output-feedback control. Int. J.
Control 76(9), 924–941 (2003)
Chapter 6
Chattering Problem

Abstract Due to different imperfections of switching devices, unmodeled dynam-

ics, and discrete-time implementation, the state trajectories are not confined to the
sliding manifold but run in its vicinity. Usually, it results in high-frequency oscil-
lation with a finite amplitude. This motion is referred to as chattering. The motion
in some vicinity of the sliding manifold with no high-frequency component can be
caused by continuous approximation of discontinuous control. When implementing a
continuous controller by power converters, a technique like pulse-width modulation
(PWM) is considered to be exploited to adapt the control law to the discontinuous
system inputs.

Keywords Chattering · Power converter

6.1 What is Chattering? Basic Ideas

Theoretical methods of analysis and design of sliding mode control constitute the
scope of the book, while the numerous applications proved to be beyond considera-
tion. We discuss the main obstacle for application only. Due to different imperfec-
tions of switching devices, unmodeled dynamics, and discrete-time implementation,
the state trajectories are not confined to the sliding manifold but run in its vicinity.
Usually, it results in high-frequency oscillation with a finite amplitude. This motion
is referred to as chattering. The motion in some vicinity of the sliding manifold
with no high-frequency component can be caused by continuous approximation of
discontinuous control. In all cases, the size of the vicinity determines accuracy of
the methods based on ideal models. Chattering is the main source of serious criti-
cism of sliding mode control. Sometimes too serious. Chattering is the common and
well-known phenomenon for several decades in the area of power converters, called
ripple. The experts understood that ripple or chattering cannot be avoided in dynamic
systems with discontinuous functions in motion equations and without dramatizing
their effect developed efficient methods for reducing it. In the framework of sliding
mode control theory, these methods were inherited and generalized along with the

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 73

V. Utkin et al., Road Map for Sliding Mode Control Design,
SpringerBriefs in Mathematics,
https://doi.org/10.1007/978-3-030-41709-3_6
74 6 Chattering Problem

development of new ones. Again following the selected style of the book, the basic
ideas of the methods are outlined in this section.
It is known that the systems with a continuous control input and feedback gain,
tending to infinity, sooner or later become unstable at the presence of unmodeled
dynamics. Similar situation takes place in the systems with discontinuous control:
the local gain in the vicinity of discontinuity point tends to infinity, which leads to
instability; in the course of motion from the discontinuity point, the local gain is
decreasing and the diverging process terminates. This picture explains the origin of
chattering—high-frequency oscillations in the vicinity of a switching surface.
In the first attempts [1] to suppress chattering, the authors offered to replace
discontinuous function  +
u if s > 0
u(s) = −
u if s < 0

for |s| ≤ Δ by continuous one

u+ − u−
u(s) = u − + (s + Δ).
2Δ
The new function tends to the original one with Δ tending to zero. However, decreas-
+
−u −
ing Δ leads to the high gain u 2Δ , the unmodeled dynamics will be excited, and
chattering will appear. The wide linear zone |s| ≤ Δ can eliminate high-frequency
+
−u −
oscillations, but low gain u 2Δ results in a low accuracy.
As follows from the above picture of chattering origination, at the intuitive level
we can state that chattering amplitude is increasing with the range of discontinuity
(see also the analysis in [2, Chap. 8]). Then keeping difference

|u + (x) − u − (x)|

at minimal level, preserving sliding mode leads to decreasing the chattering ampli-
tude. For this purpose, the state functions serving as upper and low estimates of func-
tions in right-hand sides of motion equations are needed [3, Chap. 3]. For instance,
the magnitude of discontinuous control should be proportional to the norm of the state
in linear systems. It is interesting that control was a piecewise linear state function in
the first publications on variable structure control [4], and the chattering amplitude
was decreasing in the course of approaching the origin in the state space.
A similar idea of decreasing the amplitude of discontinuous control can be imple-
mented in systems with high degree of uncertainty. As was shown in [2, Chap. 2],
the equivalent control is an average value of a real control with filtered out high-
frequency component, and it can be obtained with the help of a low-pass filter. Since

u − < u eq < u + ,

the extreme values of control can be reduced without violation of these inequalities.
6.1 What is Chattering? Basic Ideas 75

Assume that in the system

ẋ = f (t, x) + b(t, x)u

with the scalar control input

u = −M sign(s(x)),

the condition ∇s(x)b = 1 holds. Then

ṡ = F(t, x) − M sign(s), |F(t, x)| ≤ F0 = const

and the sliding mode exists on the surface s = 0, if M > F0 . The amplitude of
chattering is determined by the parameter M. If the function F(t, x) is known a priori
and far from the extreme value, the magnitude of discontinuity M can be decreased
and, as a result, the chattering amplitude can be decreased as well. Equivalent control

[sign(s)]eq = F/M

is the solution to
ṡ = 0.

It can be obtained by a low-pass filter

μż + z = sign(s), 0 < μ << 1, z ≈ (sign(s))eq .

Since |z| is decreasing function of |F(t, x)|, the chattering amplitude can be decreased
if the magnitude of discontinuity is selected as a function of |z|, for example,

u = −(M0 |z| + δ)sign(s).

If the sliding mode does not exist, then

sign(s) = ±1, |z| = 1

and control takes one of the extreme values needed for enforcing the sliding mode.
After the sliding mode occurs, the control magnitude is decreasing depending on the
level of F(t, x) along with the chattering amplitude.
Assume that the condition

|F(t, x)| < M0 , M0 = const

holds for the system with a scalar control u and state-dependent switching function
s
76 6 Chattering Problem

ṡ(t) = f (t, x(t))+u(t),
u(t) = −(M0 |z(t)|+δ) sign(s(t)),

where δ > 0 and the control u depends on the equivalent value of sign(s), obtained
by a low-pass filter

f
μż(t)+z(t) = sign(s(t, x(t))), μ  1, z ≈ [sign(s)]eq = .
M0 |z| + δ

If f > 0, then z > 0 and

δ f
z2 + z− = 0.
M0 M0

It follows from
 
δ δ2 f δ δ2
z=− + 2
+ ≤− + +1
2M0 4M0 M0 2M0 4M02

that z < 1. Similarly, it can be shown that z > −1 for z < 0. The condition

u − < u eq < u +
 
is the condition for sliding mode to exist. It is the case for our system: [sign(s)]eq  is
between the extreme values +1 and −1. This value tends to zero with f (t, x) tending
to zero and as a result the magnitude of the discontinuous control (M0 |z| + δ) is
decreased along with chattering amplitude.
Similar approach based on using equivalent control is considered in the next
chapter, titled “Adaptive SMC”, to minimize the discontinuity magnitude and
decrease chattering.
The above two methodologies, based on replacing discontinuous control by con-
tinuous one and varying amplitude of discontinuous control in the boundary layer of
switching surface, are not applicable for widely used power converters with outputs
taking one of the two extreme values only.
An asymptotic observer in the control loop can eliminate chattering despite dis-
continuous control laws. The key idea as proposed in [5] is to generate ideal sliding
mode in an auxiliary observer loop rather than in the main control loop. Ideal sliding
mode is possible in the observer loop since the control is designed as a discontinu-
ous function of the observer state, and it is entirely generated in the control software
which does not contain any unmodeled dynamics. The main loop follows the observer
loop according to the observer dynamics. Despite applying a discontinuous control
signal with switching as the plant input, no chattering occurs and the system behaves
as if an equivalent continuous control was applied. A block diagram for the system
with disregarded actuator dynamics in the ideal mode (μ2  1) is shown in Fig. 6.1.
The observer serves as a bypass for high-frequency component and the unmodeled
dynamics are not excited.
6.1 What is Chattering? Basic Ideas 77

Fig. 6.1 Simplest model with fast actuator

For the simplest example

ẋ(t) = u(t), u = −M sign(x)

the state x is reduced to zero in sliding mode. The state is known, but if it is corrupted
by noise, it is reasonable to use an observer for filtration:

d
x̂(t) = u(t) + l(x(t) − x̂(t)),
dt

where l > 0 is constant and x̂ is the estimate of x. If

1/(μs + 1)2 , 0 < μ  1

is a transfer function of the unmodeled dynamics of the actuator with output w = w1 ,

then ⎫
ẋ(t) = w1 (t), ⎪
⎪
⎬
μẇ1 (t) = w2 (t),
μẇ2 (t) = −2w2 (t) − w1 (t) + u(t), ⎪
⎪
⎭
u(t) = −M sign(x̂(t)).

d
Since x̂ and x̂(t) have opposite signs for a sufficiently high value M, sliding mode
occurs: dt
x̂ ≡ 0, u eq = −lx

and we have ⎫
ẋ(t) = w1 (t), ⎬
μẇ1 (t) = w2 (t),
⎭
μw2 (t) = −2w2 (t) − w1 (t) − lx.
78 6 Chattering Problem

According to the singular perturbation theory [6], the solution of this linear system
consists of fast and slow components. The first one decays rapidly such that after a
short time
|w1 + lx| = O(μ)

with no high-frequency oscillation while the slow motion (if μ → 0) is governed by

ẋ(t) = −lx(t)

and decays with the same rate as the observer state in the ideal system

d
x̄ = x̂ − x, x̄ = −l x̄.
dt
The control in the systems with observers is a discontinuous time function. It can
lead to high wear or even destroying actuators with moving mechanical parts (like
valves in pneumatic or hydraulic actuators). The destroying effect can be reduced
considerably by inserting a low-pass filter to make an actuator input smooth, as it
was demonstrated in Hsu [7] for model reference adaptive systems. The filter can be
handled as unmodeled dynamics, and it does not affect the ideal sliding mode in the
model loop. As a result, the system is free of chattering. It is important to note that
the output of the filter is close to the equivalent control with mismatch depending
on the filter time constant and the discontinuous control amplitude [3, Chap. 2].
If the equivalent control can be estimated, then the amplitude of the discontinuous
control can be decreased (recall that trajectories belong to the discontinuity surface
for control u = u eq ). Therefore, the effect of unmodeled dynamics in the plant model
can be reduced considerably.
Demonstrate the chattering suppression effect in systems with an observer for the
second-order system
ẍ = −ẋ − x + w

with disregarded dynamics of the actuator

1 ⎫
w(p) = u(p), 0 < μ << 1, ⎪
⎪
(μp + 1)2 ⎪
⎬
μ ẅ = −2w ẇ − w + u,
2

u = −u 0 sign(s), s = cx + ẋ, ⎪
⎪
⎪
⎭
u 0 and c are positive constants,p is Laplace variable.

The system with the observer practically is free of chattering and similar to the ideal
system with μ = 0 (Fig. 6.2).
6.2 Power Converters 79

Fig. 6.2 Chattering with unmodeled dynamics (upper plots) and zoomed processes with and without
unmodeled dynamics

6.2 Power Converters

The main disadvantage of the observer-based chattering suppression method is need

of the plant equations. The influence of uncertainties can be reduced by increasing the
observer gains, which is not dangerous from the point of stability, since an observer
is implemented in software part of control and free of unmodeled dynamics. The
natural way of reducing the chattering effect is increasing the oscillation frequency
to be filtered out by plant dynamics. When implementing a continuous controller
by power converters, a technique like pulse-width modulation (PWM) has to be
exploited to adapt the control law to the discontinuous system inputs. In the light of
recent advances of high-speed circuitry and insufficient linear control methodologies
for internally nonlinear high-order plants such as AC motors, sliding mode control
has become increasingly popular, since this control methodology implies discontin-
uous control actions and on/off is the only admissible mode for power converters.
Therefore, outputs of power converters can be used as control action directly without
PWM. Due to the switching element (Fig. 6.3), the control is discontinuous and can
be selected as a function of the mismatch between real and desired values of the
output voltage. The tracking problem is solved if sliding mode is enforced.

Fig. 6.3 Power converter

80 6 Chattering Problem

Commercially available electronic converters enable handling switching frequen-

cies in the range of hundreds of kHz, and the chattering component will be filtered
out by plant dynamics. On the other hand, heat losses in a converter are proportional
to switching frequency and they should be limited. Seemingly, the dead-end situation
is encountered: the desired accuracy, depending on chattering amplitude, cannot be
reached at maximal admissible frequency. The challenge is to design a converter
based on a given, fixed switching frequency, reducing chattering to a desired level.
The so-called “harmonic cancellation” principle can be used to find a way out of the
situation. The converter with the desired output u 0 is replaced by m parallel identical
converters called phases with outputs u 0 /m and summation block. Amplitudes and
frequencies of periodic oscillations with period T or chattering in all converters are
the same, but phases depend on initial conditions and may be different. For the worst
case, the chattering amplitude after summation takes maximum value, if all phases
are equal. Representing the chattering signal in each phase as Fourier series, calculate
the amplitude of kth harmonic in the output, if phase shift between ith and (i + 1)th
phases is equal to T /m:
2π
ω= , ωk = kω,
T

k−1   
m−1  
j (ωk t− 2πk
m i)
i=0 sin ω k t − 2π
ωm
i = i=0 I m e
 
m−1 − j 2πk i
= I m e jωk t Z , Z = i=0 e m .

Function Z satisfies condition Z = Z e− j m and is different from zero if e− j m = 1

2πk 2πk

or for k = m, 2m, 3m, . . . . . .. It means that at least the first m − 1 harmonics are
canceled and increasing number of phases will decrease the chattering amplitude
essentially.
To get this effect, called harmonic cancellation, the switching commands should
be selected such that phase shifts between two neighbor phases are equal to T /m.
Switching elements of power converters are implemented with state-dependent width
of hysteresis loops and the same magnitudes. For the first phase, it is selected such
that the switching frequency is equal to maximal admissible value. If the switching
function of the second phase depends not only on the tracking error (differences
between real and desirable output values) but also on the error of the first phase,
then chattering frequency of the second phase coincides with that of the first phase.
The width of hysteresis loop in the second phase is selected such that the phase shift
is equal to T /m. The subsequent phases are designed in similar ways. Theoretical
substantiation and experimental results may be found in [3, Chap. 8], [8]. The results
of simulation of the four-phase power converter with parameters in Table 6.1 are
shown in Fig. 6.4.
In spite of high level of chattering in each phase, it is practically deleted in the
converter output after summation.
The objective of all the above methods is estimation of chattering amplitude
and its suppression. In addition, the harmonic cancellation method implies control
of a chattering frequency. Since as a rule chattering is far from being harmonic
6.2 Power Converters 81

Table 6.1 Parameters of Parameter Value

power converter
L(H ) 1
C(F) 1
Ra (Ω) 1
R L (Ω) 1
Vs (V) 12

Fig. 6.4 Four-phase

converter, u 0r e f = 4.2 V

oscillations, it is difficult to find the solution of the corresponding differential equation

as a time function. Frequency analysis proved to be an efficient approximate tool to
solve the problem under two main assumptions: chattering is a periodic time function
and high-order harmonics can be disregarded. In the framework of this approach,
describing function method and its further development can be found in [9].

References

1. Slotine, J.J., Li, W.: Applied Nonlinear Control. Prentice Hall (1991)
2. Utkin, V.: Sliding Modes in Control and Optimization. Springer, Berlin (1992)
3. Utkin, V., Guldner, J., Shi, J.: Sliding Mode Control in Electro-Mechanical Systems. CRC Press
(2009)
4. Emelyanov, S. (ed.): Theory of Variable Structure Control Systems. Nauka (1970) (in Russian)
5. Bondarev, A., Bondarev, S., Kostyleva, N., Utkin, V.: Sliding mode in systems with asymptotic
state observers. Autom. Remote Control 46, 49–64 (1985)
82 6 Chattering Problem

6. Kokotovic, P., Khalil, H., O’Relly, J.: Singular Perturbation Method in Control: Analysis and
Design. Academic Press (1986)
7. Hsu, L.: Smooth sliding mode control of uncertain systems based on prediction error. Int. J.
Robust Nonlinear Control 7(4), 353–372 (1997)
8. Biel, D., Fossas, E.: Some experiments on chattering suppression in power converters. In: 18th
IEEE International Conference on Control Applications, pp. 1523–1529 (2009)
9. Boiko, I.: Discontinuous Control Systems : Frequency-domain Analysis and Design. Birkhauser,
Boston (2009)
Chapter 7
High-Order Sliding Mode Control

Abstract The high-order sliding mode (HOSM) is discussed here as an alternative

methodology to the conventional sliding mode control to design feedback control,
observers, and to handle the chattering problem. Here, the principle ideas of HOSM
methodology are presented. Similar to the second-order sliding mode, HOSM implies
that sliding mode occurs in the manifold of a smaller order after a finite time interval.

Keywords High-order sliding mode · Finite-time convergence

7.1 HOSM Definition and Its Properties

The core idea of HOSM control design implies existence of sliding mode in the
manifold on n − r order (1 < r < n − 1) after a finite time interval similar to systems
in Sect. 3.9 with sliding mode of the second order.
The next definition of HOSM is taken from [1].

Definition 7.1 Let σ be an output of the closed-loop system (1), and there exists
r ∈ N such that

• for p = 1, 2, . . . , r − 1 total time derivatives σ ( p) along the trajectories of the

closed-loop system (2.1) are continuous functions of time and state variables;
• the set  
S = σ ( p) = 0 for any p = 0, 1, 2, . . . , r − 1

is the sliding manifold of the closed-loop system (2.1);

• σ (r ) is discontinuous in a vicinity of S.

Then S is the r th-order sliding manifold.

The efficiency of SMC has been demonstrated for a set of applications [2, Chap. 8],
[3, Chaps. 10–13].
From the application standpoint, as stated in [4], it is important for the conventional
sliding mode that
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 83
V. Utkin et al., Road Map for Sliding Mode Control Design,
SpringerBriefs in Mathematics,
https://doi.org/10.1007/978-3-030-41709-3_7
84 7 High-Order Sliding Mode Control

– the implementation of SM implies that the relative degree (R D) of the system

should be equal to one which is considered as a serious constraint for SM appli-
cations;
– the discontinuities in control lead to chattering—an unacceptable phenomenon in
some applications.
High-order sliding mode (HOSM) was offered as an alternative methodology to the
conventional sliding mode control. According to [4, Chap. 8], “the intrinsic difficul-
ties of conventional SMC are mitigated by HOSM controllers and this approach is
effective for arbitrary relative degree, and the well-known chattering is considerably
reduced”. The principle ideas of HOSM methodology and references can be found
also in [4, Chap. 8]. Similar to the second-order sliding mode, HOSM implies that
sliding mode occurs in the manifold of the (n − k)th order (1 < k < n ) after a finite
time interval.

7.2 Simple Examples of Systems with HOSM Control

First, consider the system (3.12) with sliding mode control of the second order. Let
the system output σ be governed by equation

σ̈ = u + F, (7.1)

where F is a bounded function of the state and time.

As shown in Sect. 3.9, the state can be reduced to the origin in the plane (σ, σ̇)
after a finite time interval by twisting and super-twisting algorithms and the second-
order sliding mode arises. The authors of HOSM control offered two more versions
to enforce sliding modes of the second order:
⎫
u = −Msign(s), M > 0, ⎬
√ (7.2)
⎭
s := σ̇ + α |σ|sign (σ) , α > 0,

or s
u = −ρ √ . (7.3)
|σ̇| + |σ|

For the first version (7.2), the origin is reached in sliding mode after finite time. The
state plane is similar to that in Fig. 4.2. For the second version (7.3), the origin is
reached after finite time [4, Chap. 4] but without intermediate sliding mode (Fig. 7.1).
Both versions were generalized for the systems with sliding modes of an arbitrary
order r and finite time of reaching the sliding manifold.
Here are control algorithms for the third-order sliding modes:
 1/6  
u = −αsign(σ̈ + 2 |σ̇|3 + |σ|2 sign σ̇ + |σ|2/3 sign (σ) ) (7.4)
7.2 Simple Examples of Systems with HOSM Control 85

Fig. 7.1 HOSM without

sliding mode in the reaching
interval

and
 −1/2  
σ̈ + 2 |σ̇| + |σ|2/3 σ̇ + |σ|2/3 sign (σ)
u=α   . (7.5)
|σ̈| + 2 |σ̇| + |σ|2/3

It is interesting that the discontinuous controls similar to (7.3) and (7.5) have one
discontinuity point (σ, σ̇, . . . , σ (r −1) ) only. Implementation of all HOSM controls
(except for super-twisting) needs all time derivative (σ̇, . . . , σ (r −1) ) of the output σ.

7.3 HOSM in SISO Affine Systems in Regulation

and Tracking Problems

7.3.1 Uncertain Plant

Following [5], consider the nonlinear system

⎫
ż = f (t, z) + g(t, z)u, ⎬
(7.6)
⎭
y = h(t, z),

where z ∈ R n defines the state vector, u ∈ R is the control input, y ∈ R is the output,
and h(t, z) : R × R n → R is a smooth output function. The functions f (t, z) and
g(t, z) are uncertain smooth vector fields on R n.
A standard problem of control is the output tracking problem, consisting in forc-
ing the output y to track a (time-varying) signal y R (t), namely, in making the output
σ = y − y R vanishing in finite time and to keep σ ≡ 0 exactly by a bounded (discon-
tinuous) feedback control. All differential equations are understood in the Filippov’s
sense.
When the relative degree ρ with respect to σ is known, well defined and constant
this is equivalent to designing a controller for the differential inclusion (DI)
86 7 High-Order Sliding Mode Control
⎫
ẋi = xi+1 , i = 1, . . . , ρ − 1 ⎬

(7.7)
⎭ DI
ẋρ ∈ [−C, C] + [K m , K M ]u

where

x = (x1 , . . . , xρ ) = (σ, σ̇, . . . , σ (ρ−1) ) ,

di di
σ (i) = h(z, t) − yR
dt i dt i

(if such σ (i) (i = 1, . . . , ρ − 1) exist). Below based on smooth control and Lyapunov
functions, we derive a family of homogeneous HOSM controllers.

7.3.2 Lyapunov Function for HOSM Controllers’ Analysis

Given the relative degree ρ ≥ 2, assign the homogeneity weights ri = ρ − i + 1 to

the variables xi , obtaining the dilatation vector r = (ρ, ρ − 1, . . . , 1).1 Define also
an arbitrary non-decreasing sequence of positive real numbers αi, so that

0 ≤ α1 ≤ · · · ≤ αρ−1 ≤ αρ .

Furthermore, we define recursively, for i = 2, . . . , ρ, the r -homogeneous functions

(ki > 0) from C 1:
ρ+αi
⎫
ρ+α1 ρ+αi ρ+αi
⎪
σ1 (x̄1 ) = x1 ρ , . . . , σi (x̄i ) = xi ρ−i+1 + ki−1 σi−1 xi−1
ρ−i+1 ρ+αi+1
,⎬
⎪
⎭
x̄i = (x1 , . . . , xi ) , xi s := |xi |s sign (xi ) .

For any constant m ≥ max1≤i≤ρ {2ρ + 1 + αi−1 − i} we also define recursively, for
i = 2, . . . , ρ, the C 1 r -homogeneous functions

ρ m ⎫
V1 (x1 ) = |x | ρ , . . . , Vi (x̄i ) = γi−1 Vi−1 (x̄i−1 ) + Wi (x̄i ) ⎪
⎪
m 1 ⎪
⎪
⎪
⎪
m m−ri  ri m ⎬
Wi (x̄i ) = ri
m i
|x | ri − νi−1 (x̄i−1 ) ri
xi + 1 − |νi−1 (x̄i−1 )| ri
m ⎪
⎪
⎪
⎪
⎪
⎪
r2 ρ−1 ri +1
⎭
ν1 (x1 ) = −k1 σ1 ρ+α1
= −k1 x1 ρ , . . . , νi (x̄i ) = −ki σi (x̄i ) ρ+αi

1 Definitions of homogeneity and dilation are given in Chap. 4 (see Definitions 4.1 and 4.2).
7.3 HOSM in SISO Affine Systems in Regulation and Tracking Problems 87

with (arbitrary) constants γi > 0. Here, σi (x̄i ), νi (x̄i ), and Vi (x̄i ) are r -homogeneous

of degrees ρ + αi , ri + 1 and m, respectively. So, Vc (x) = Vρ x̄ρ is a smooth
Lyapunov function for the considered uncertain plant (7.6).

7.3.3 HOSM Controllers

From Vc (x), we can derive

 different
 controllers. In particular,
  we obtain the following
family of discontinuous u ρD and quasi-continuous u ρQ controllers
 0 ⎫
u ρD = −kρ ϕ D (x) = −kρ σρ (x) ⎪ ⎪
⎬
σρ (x) ⎪
⎪
u ρQ = −kρ ϕ Q (x) = −kρ ⎭
M(x),

where M(x) is any continuous r -homogeneous positive-definite function

  of degree
 σρ (x) 
ρ + αρ , and (for simplicity) we assume that it is scaled so that   ≤ 1. The
M(x) 
homogeneous controllers are derived by imposing the condition

∂Vc (x)
ϕ(x) > 0
∂xρ

∂Vc (x)
at all points where
= 0, resulting from the condition V̇c (x) < 0 and in view
∂xρ
of the relation
∂Vc (x) ∂Vc (x)  
V̇c (x) = ẋρ ∈ [−C, C] + [K m , K M ]u ρ
∂xρ ∂xρ
 
 ∂Vc (x) 

≤ 2C   − kρ K m ∂Vc (x) ϕ(x) =⇒ ∂Vc (x) ϕ(x) > 0.
∂xρ  ∂xρ ∂xρ

The values of ki , for i = 1, . . . , ρ − 1 can be fixed depending only on ρ and αi .

7.3.4 Discontinuous and Quasi-Continuous HOSM

Controllers for SISO Systems

Varying the selection of the free parameters 0 ≤ α1 ≤ · · · ≤ αρ , we obtain different

families of controllers (for the orders of the relative degrees ρ = 2, 3).
88 7 High-Order Sliding Mode Control

Discontinuous Controllers

– Nested Sliding Controllers (when some of the αi are different)

 2+α2
0
N est
u 2D = −k2 x2 2+α2 + k12+α2 x1 2 ,

  3+α3 0
 3+α
3+α2 3+α2 3+α2 2
N est
u 3D = −k3 x3 3+α3 + k23+α3 x2 2 + k1 2 x1 3 .

– Relay Polynomial Controllers (when αρ = αρ−1 =···= α1 = α ≥ 0)

 2+α

Rel
u 2D = −k2 sign x2 2+α + k̄1 x1 2 ,
 3+α 3+α

Rel
u 3D = −k3 sign x3 3+α + k̄2 x2 2 + k1 x1 3 .

Quasi-Continuous Controllers

– Nested Sliding Controllers (when some of the αi are different, βi > 0)

2+α2
x2 2+α2 + k12+α2 x1 2
N est
u 2Q = −k 2 2+α2 ,
|x2 |2+α2 + β1 |x1 | 2

 3+α2
 3+α
3+α3
3+α2 3+α2 2
x3
3+α3
+ k23+α3 x2 2 + k1 x1 3
2

N est
u 3Q = −k 3 3+α3 3+α3 .
|x3 |3+α3 + β2 |x2 | 2 + β1 |x1 | 3

– Relay Polynomial Controllers (when αi = α ≥ 0)

2+α
x2 2+α + k12+α x1 2
Rel
u 2Q = −k 2 2+α ,
|x2 |2+α + β1 |x1 | 2

3+α 3+α
x3 3+α + k̄2 x2 2 + k̄1 x1 3
Rel
u 3Q = −k 3 3+α 3+α .
|x3 |3+α + β2 |x2 | 2 + β1 |x1 | 3

All these controllers solve the problem posed before.

Theorem 7.1 ([5]) For any relative degree ρ ≥ 2 between x1 and control, each con-
troller of the families of discontinuous or quasi-continuous controllers with arbitrary
parameters 0 ≤ α1 ≤ ···≤ αρ is ρ-sliding homogeneous, and for kρ sufficiently large
the ρth-order sliding mode

x = (x1 , . . . , xρ ) = (σ, σ̇, . . . , σ (ρ−1) ) = 0

7.3 HOSM in SISO Affine Systems in Regulation and Tracking Problems 89

is established in finite time for the uncertain system (7.6)–(7.7) under properly
chosen gains k1 , . . . , kρ−1 and β1 , . . . , βρ−1 .

Conclusion
In the framework of HOSM study, several interesting research directions were ini-
tiated by A. Levant (basic concepts, feedback design, observers [1, 6]), J. Moreno
(stability of systems with HOSM [7, 8]), and I. Boiko (frequency methods [9]).
HOSM gave born to many mathematical problems. All authors of this book pub-
lished solutions to some of them (see, for example, [10–14]). The interested reader
can find the arguments confirming the stated advantages of HOSM control in
[4, Chap. 8].

References

1. Levant, A.: Higher-order sliding modes, differentiation and output-feedback control. Int. J.
Control 76(9), 924–941 (2003)
2. Edwards, C., Spurgeon, S.: Sliding Mode Control: Theory and Applications. Taylor and Francis
(1998)
3. Utkin, V., Guldner, J., Shi, J.: Sliding Mode Control in Electro-Mechanical Systems. CRC
Press (2009)
4. Shtessel, Y., Edwards, C., Fridman, L., Levant, A.: Sliding Mode Control and Observation.
Control Engineering Series. Birkhauser, NY (2014)
5. Cruz-Zavala, E., Moreno, J.A.: Homogeneous high order sliding mode design: a lyapunov
approach. Automatica 80, 232–238 (2017)
6. Levant, A.: Sliding order and sliding accuracy in sliding mode control. Int. J. Control 58(6),
1247–1263 (1993)
7. Moreno, J., Osorio, M.: A lyapunov approach to second-order sliding mode controllers and
observers. In: 47th IEEE Conference on Decision and Control (CDC), pp. 2856–2861 (2008)
8. Moreno, J.: A lyapunov approach to output feedback control using second-order sliding modes.
IMA J. Math. Control Inf. 29(3), 291–308 (2012)
9. Boiko, I.: Discontinuous Control Systems : Frequency-domain Analysis and Design.
Birkhauser, Boston (2009)
10. Utkin, V.: Mechanical energy-based lyapunov function design for twisting and super-twisting
sliding mode control. IMA J. Math. Control Inf. 32(4), 675–688 (2015)
11. Utkin, V.I.: Discussion aspects of high-order sliding mode control. IEEE Trans. Autom. Control
61(3), 829–833 (2016)
12. Polyakov, A., Poznyak, A.: a). Lyapunov function design for finite-time convergence analysis:
“twisting” controller for second-order sliding mode realization. Automatica 45, 444–448 (2009)
13. Orlov, Y.: Finite time stability and robust control synthesis of uncertain switched systems.
SIAM J. Control Optim. 43(4), 1253–1271 (2005)
14. Orlov, Y.: Discontinuous Systems: Lyapunov Analysis and Robust Synthesis Under Uncertainty
Conditions. Springer, Berlin (2008)
Chapter 8
Discrete-Time Systems

Abstract Discretization of continuous-time models is considered. Definition of SM

for discrete-time systems is presented. The behavior under uncertainties of discrete-
time systems, controlled by SM, is briefly discussed.

Keywords Discretization of continuous-time models · The definition of SM for

discrete-time systems · The behavior under uncertainties

8.1 Discretization of Continuous-Time Models

The design principles of sliding mode control were developed mainly for finite-
dimensional continuous-time systems. At the same time, more and more systems are
implemented based on digital microcontrollers. In both discrete- and continuous-time
cases, the control is assumed to take one of the finite number of values. The ideal
siding mode in a continuous-time system can be obtained with switching frequency
tending to infinity. But it cannot be higher than a sampling frequency in discrete-time
systems. As a result, the state oscillates in a vicinity of the switching surface and the
oscillations are referred to as chattering as well, although they are of different nature
than those caused by unmodeled dynamics. Natural way to increase the chattering
frequency and decrease its amplitude is decreasing a sampling time, but it may prove
to be too short for calculation of a control signal. Sliding mode with x(t) ≡ 0 occurs
in the simple first-order system

ẋ(t) = f (t, x(t)) + u(t), t > 0,
(8.1)
| f (t, x)| < f 0 ∈ R, u(t) = −u 0 sign(x(t)), f 0 < u 0 ∈ R

after a finite time interval. Figure 8.1 illustrates chattering caused by discrete-time
implementation of the control for explicit Euler approximation of the above system

xk+1 = xk + δ( f (kδ, xk ) − u 0 sign(xk )),

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 91

V. Utkin et al., Road Map for Sliding Mode Control Design,
SpringerBriefs in Mathematics,
https://doi.org/10.1007/978-3-030-41709-3_8
92 8 Discrete-Time Systems

Fig. 8.1 Direct discrete implementation of sliding mode control

where δ is a sampling time. Since the control is constant within the sampling interval,
chattering appears always starting from ksm , if the control is a discontinuous state
function (except for unlikely cases f = u 0 or f = −u 0 ), even if function f is known.
This motion was called the “zigzag regime” in [1] (see Fig. 8.1).
Let us now discuss the methods, offered in the literature, which intend to
reduce chattering amplitude for n-dimensional linear discrete-time systems with
m-dimensional control
xk+1 = Axk + Bu k + dk , (8.2)

where A and B are matrices with bounded elements, and dk is a disturbance vector
with bounded elements. Control is selected as a piecewise linear state function in
[2], u k = K xk , with elements of K undergoing discontinuities on planes

si (x) = ciT x = 0, ci ∈ R n , i = 1, . . . , m.

If the disturbance is equal to zero and the motion in manifold

s(x) = (s1 (x), . . . , sm (x)) = 0

8.1 Discretization of Continuous-Time Models 93

is stable with the state tending to zero, then the system can be stabilized by a proper
choice of matrix K . The state reaches s(x) = 0, and its trajectories are converging to
the origin with chattering amplitude tending to zero [2]. Of course the stabilization
problem cannot be solved at the presence of disturbance dk , since xk = 0 cannot be
a solution to (8.2).
The chattering amplitude (Fig. 8.1) for system (8.1) is increasing with increasing
magnitude u 0 of the discontinuous control. For control as sum of linear and discon-
tinuous functions of s, the magnitude of the second term can be decreased and as a
result it decreases the chattering amplitude [3].
Replacing discontinuous function by some continuous state function in a boundary
layer of manifold s(x) = 0 is also the way to reduce chattering caused by discretiza-
tion of the control [4]. The control is selected such that s(xk+1 ) = s(xk ) in some
vicinity of s(x) = 0. It is called the equivalent control u eq similar to continuous-
time systems with ṡ(x(t)) = 0 for s(x(t)) = 0 and u(t) = u eq (t). Having reached
the vicinity, the state remains in it without chattering. Implementation of such selected
control implies that the disturbance is known; otherwise, s(xk+1 ) cannot be found.
The upper estimate of the boundary layer with state trajectories of so-called quasi-
sliding mode is found in [4] for systems with known bounds of disturbances only.
The width of the boundary layer or the amplitude of chattering can be reduced if
compared with those for direct implementation of discontinuous control.
As follows from the titles of cited above paper, analysis and design methods are
studied for discrete-time systems with sliding modes. At the same time, the concept
“discrete-time sliding mode” needs a strict definition as well as the conditions for this
motion to exist are needed for developing the theory of this class of control systems.

8.2 Definition of SM for Discrete-Time Systems

For continuous-time systems, state trajectories belong to manifold s(x) = 0 and the
manifold is reached after a finite time interval. These properties are formalized in
Definition 2.1 for continuous-time systems. The definition in terms of discrete-time
systems
xk+1 = F(xk ), F : R n → R n (8.3)

is formulated in [5].
Definition 8.1 Discrete-time sliding mode takes place on subset  of manifold

s(x) = 0, s : R n → R m , m < n,

if there exists an open neighborhood U of this subset such that s(F(x)) ∈  for any
x ∈ U (see Fig. 8.2).
94 8 Discrete-Time Systems

Fig. 8.2 Discrete-time

sliding manifold

It makes sense to note that the idea to give a definition in terms of finite-time
convergence works for both continuous- and discrete-time systems. In contrast to
continuous-time systems, sliding mode may arise in discrete-time systems with con-
tinuous right-hand side in (8.3) satisfying Lipschitz condition even for a linear func-
tion F(x) = Ax, where A is a constant matrix. Sliding mode exists on linear manifold

s(x) = C x = 0

in an affine system

xk+1 = F(k, xk ) + B(k, xk )u k , xk ∈ R n , u k ∈ R m , (8.4)

if s(xk+1 ) = 0 for any xk . This condition holds for control

u k = u k = − (C B(k, xk ))−1 C F(k, xk )

eq
(8.5)

assuming that matrix C B(k, xk ) is non-singular. Control (8.5) is called equivalent

control, but it is different from that selected in [4] to fulfill the condition s(xk+1 ) =
s(xk ). Control (8.5) makes s(x) equal to zero after one step [5–8]. It is not realistic
to implement control (8.5) for small values of a sampling time δ, if Eq. (8.4) is an
approximation of a continuous-time system

ẋ(t) = f (t, x(t)) + b(t, x(t))u(t), t > 0.

Then
F(k, xk ) = xk + δ f (kδ, xk ),
B(k, xk ) = δb(kδ, xk ),
s(xk+1 ) = s(xk ) + δ(C f (k, xk ) + C B(k, xk )u k )

and
1
u k = − (C B(k, xk ))−1 s(xk ) − (C B(k, xk ))−1 C F(k, xk ).
eq
(8.6)
δ
For any s(xk ) = 0, the control tends to infinity with δ tending to zero. It means that
control (8.6) should be modified taking into account control constraints existing in
8.2 Definition of SM for Discrete-Time Systems 95

Fig. 8.3 Ideal discrete-time sliding mode control in system with known parameters

any real-life system. For control bounded by norm u k  ≤ ũ = const, this constraint
is taken into account in [6]
⎧ eq eq
⎨ uk if u k  ≤ ũ,
eq
uk = uk eq (8.7)
⎩ ũ eq if u k  > ũ.
u k 

Control (8.7) is collinear to u eq , and its norm does not exceed the admissible value.
Control should be equal to the equivalent control on the manifold s (x) = 0; hence,
the condition  
(C B(k, xk ))−1 C F(k, xk ) < ũ (8.8)

should hold. As follows from (8.6), the control is equal to u eq in δ—vicinity of

manifold s(x) = 0. As shown in Fig. 8.3, xk for the discrete-time version of system
(8.1) is equal to zero identically after a finite number of steps.
96 8 Discrete-Time Systems

8.3 Behavior Under Uncertainties

Complete information on plant parameters and disturbance is needed for imple-

mentation of the equivalent control. Suppose now that the system operates under
uncertainty conditions: matrices A and B and disturbance d in (8.2) may vary in
some ranges. Suppose that the input matrix can be represented in the form

B = B0 + B (k, xk )

with known matrix B0 . Then the condition similar to (8.8) can be written in the form
    
ũ 1 − (C B0 )−1 CB (k, xk ) > (C B(k, xk ))−1 C F(k, xk ) .

The range of admissible variations B (k, xk ) of matrix B can be found from this
inequality.
Similar to (8.7), control
⎧
⎨ −(C B0 )−1 s(xk ) if (C B0 )−1 s(xk ) ≤ ũ,
uk = (C B0 )−1 s(xk ) (8.9)
⎩ −ũ if (C B0 )−1 s(xk ) > ũ
(C B0 )−1 s(xk )

is in the admissible domain but it depends on neither plant parameters nor distur-
bances. Again, s(xk ) is monotonously decreasing and after finite number of steps
control
u k = −(C B0 )−1 s(xk )

will belong to the admissible domain and the subsequent motion will be in δ-vicinity
of manifold s(x) = 0, while the control values remain in the admissible domain.
This result should be expected for the systems operating under uncertainty condi-
tions, because control is constant within a sampling time, while the parameters and
disturbance can take any value from some ranges. In contrast to discrete-time sys-
tems with discontinuous control, the systems with control (8.9) are free of chattering,
because the control is a continuous state function for the motion within delta vicinity
of s = 0.

References

1. Milosavlevic, C.: General conditions for the existence of a quasi-sliding mode on the switching
hyperplane in discrete variable structure systems. Autom. Remote Control 46(3), 307–314 (1985)
2. Sarpturk, S.Z., Istefanopulos, Y., Kaynak, O.: On the stability of discrete-time sliding mode
control systems. IEEE Trans. Autom. Control 32(10), 930–932 (1987)
3. Gao, W., Wang, Y., Homaifa, A.: Discrete-time variable structure control systems. IEEE Trans.
Ind. Electron. 42(2), 117–122 (1995)
4. Furuta, K.: Sliding mode control of discrete system. Syst. Control Lett. 14(2), 145–152 (1990)
References 97

5. Drakunov, S.V., Utkin, V.I.: On discrete-time sliding modes. In: IFAC Conference on Nonlinear
Control System Design. Capri, Italy (1989)
6. Utkin, V.I.: Variable Structure and Lyapunov Control. In: Sliding Mode Control in Discrete-Time
and Difference Systems. Springer, Berlin (1993)
7. Hui, S., Zak, S.H.: On discrete-time variable structure sliding mode control. Syst. Control Lett.
38(4), 283–288 (1999)
8. Brogliato, B., Acary, V., Orlov, Y.: Chattering-free digital sliding-mode control with state
observer and disturbance rejection. IEEE Trans. Autom. Control 57(5), 1087–1101 (2012)
Chapter 9
Adaptive SMC

Abstract In this chapter, the basic ideas of adaptive SMC are presented. They are
intended to minimize an undesired chattering effect during motion in sliding mode.
Two main approaches for designing adaptive SMC are discussed: the σ-adaptation
and the dynamic adaptation based on the equivalent control, measured by a first-order
low-pass filter with the discontinuous control as an input. Super-twisting control with
adaptation is also considered.

Keywords Adaptive SMC · σ-adaptation · Dynamic adaptation based on the

equivalent control · Adaptive super-twisting controller

The main obstacles for application of sliding mode control are two interconnected
phenomena: chattering and high activity of control action. It is well known that the
amplitude of chattering is proportional to the magnitude of a discontinuous control.
These two problems can be handled simultaneously if the magnitude is reduced to
a minimal admissible level defined by the conditions for sliding mode to exist. The
adaptation methodology for obtaining the minimum possible value of control is based
on two approaches developed in recent publications [1]:
• The σ-adaptation, providing the adequate adjustments of the magnitude of a dis-
continuous control within the “reaching phase”, that is, when state trajectories are
out of a sliding surface [2];
• Dynamic adaptation or the adaptation within a sliding mode (on a sliding surface),
based on the equivalent control obtained by the direct measurements of the output
signals of a first-order low-pass filter containing the discontinuous control in the
input with the specially adapted magnitude value [1, 3].

9.1 The σ-Adaptation Method

Consider the nonlinear uncertain system


ẋ (t) = f (x (t)) + g (x (t)) u,
(9.1)
x (0) = x0 , t ≥ 0,
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 99
V. Utkin et al., Road Map for Sliding Mode Control Design,
SpringerBriefs in Mathematics,
https://doi.org/10.1007/978-3-030-41709-3_9
100 9 Adaptive SMC

where x (t) ∈ X ⊂ R n the state vector and u ∈ R the control input to be designed.
Functions f (x) and g (x) are supposed to be smooth uncertain functions and are
bounded for all x ∈ X ; furthermore, f (x) contains unmeasured perturbations term
and g (x) = 0 for all x ∈ X .
The control objective consists in forcing the continuous function σ(t, x), named
sliding variable, to 0. Supposing that σ admits the relative degree equal to 1 with
respect to u, one gets

σ̇(t, x) = Ψ (t, x)+ Γ (t, x)u,

∂σ(t, x) ∂σ(t, x) 
Ψ (t, x) := + f (x) ,
∂t  ∂x
 (9.2)
∂σ(t, x)
Γ (t, x) := g (x) .
∂x

Functions Ψ (t, x) and Γ (t, x) are supposed to be bounded such that for all x ∈ X
and all t ≥ 0
|Ψ (t, x)| ≤ Ψ M , 0 < Γm ≤ Γ (t, x) ≤ Γ M . (9.3)

It is assumed that Ψ M , Γm , and Γ M exist but are not known. The objective for a designer
is to propose a sliding mode controller u (t, σ) with the same features as classical
SMC, namely, robustness and finite-time convergence to some zone, closed to an
equilibrium point, but without any information on uncertainties and perturbations
(appearing in f (x)).
In the sequel, the definitions of ideal and real sliding mode are recalled.

Definition 9.1 We say that an ideal sliding mode exists if

lim σ̇ < 0 and lim σ̇ < 0.

σ→ +0 σ→ −0

If, due to some small positive parameter μ0 , the state trajectories belong to domain

|σ (t)| ≤ Δ (μ0 ) , lim Δ (μ0 ) = 0,

μ0 →0

then the motion is called a real sliding mode.

In real applications, an “ideal” sliding mode, as in Definition 9.1, cannot be estab-

lished. But the concept of a “real” sliding mode seems to be acceptable.
The control is selected as a discontinuous state function with adaptation of its
magnitude:

 u (t, σ) = −K (t) sign (σ(t, x (t))) ,

K̄ |σ(t, x (t))| sign (|σ(t, x (t))| − ε) if K > μ0 (9.4)
K̇ (t) =
0 if K ≤ μ0
9.1 The σ-Adaptation Method 101

with K̄ > 0, ε > 0 and a small enough positive μ0 . The parameter μ0 is introduced
in order to get only positive values for K .
Once sliding mode with respect to σ(t, x (t)) is established, the proposed gain
adaptation law (9.4) allows the gain K declining (while |σ(x (t) , t)| < ε). In other
words, the gain K will be kept at the smallest level that allows a given accuracy of
σ-stabilization. Of course, as described in the sequel, this adaptation law allows to
get an adequate gain with respect to uncertainties’/perturbations’ magnitude.

Theorem 9.1 ([2]) Given the nonlinear uncertain system (9.1) with the sliding vari-
able σ(x (t) , t) dynamics (9.2) controlled by (9.4), there exists a finite time t f so that
a real sliding mode is established for all t ≥ t F , i.e.,

|σ(x (t) , t)| < δ

for all t ≥ t f with   

δ= ε2 +Ψ 2M / K̄ Γ m . (9.5)

So, the convergence to the domain |σ| ≤ ε is in a finite time but could be sustained
in the bigger domain |σ| ≤ δ. Therefore, the real sliding mode exists in the domain
|σ| ≤ δ. The choice of parameter ε has to be made by an adequate way because a
“bad” tuning could provide either instability and control gain increasing to infinity
or bad accuracy for closed-loop system. [2] suggested to select ε adjusted in time.

9.2 The Dynamic Adaptation Based on the Equivalent

Control Method

9.2.1 The Simple Motivating Example

Consider the first-order system


ẋ (t) = a (t) + u,
(9.6)
u = −ksign (x (t)) , k > 0.

The ranges of a time-varying parameter

0 < |a (t)| ≤ a+

and the upper bound A for its time derivative

|ȧ (t)| ≤ A
102 9 Adaptive SMC

are known only. The sliding mode with x (t) ≡ 0 exists for all values of unknown
parameter a (t) if k > a+ . However, if parameter a (t) is varying, the gain k can
be decreased and, as a result, chattering amplitude can be reduced. The objective of
adaptation is decreasing k to the minimal value preserving sliding mode, if parameter
a is unknown. If the condition k > a+ holds, then sliding mode with x (t) ≡ 0 occurs
and control in (9.6) should be replaced by the equivalent control u eq for which the
right-hand side in (9.6) is equal to zero:

ẋ (t) = 0 = a (t) + u eq , (9.7)

that leads to

k (t) sign (x (t)) eq = a (t) . (9.8)

If k < a, the set x (t)

≡ 0 is of zero measure in time and can be disregarded. The
function sign (x (t)) eq is an average value, or a slow component of discontinuous
function sign (x (t)) switching at high frequency and can be easily obtained by a
low-pass filter filtering out the high-frequency component [4]. Of course, the average
value is in the range (−1, 1). Then the design idea of adaptation

looks natural: after
sliding mode occurs the control parameter sign (x (t)) eq should be increased until
it becomes close to 1. On one hand, the condition k (t) > a (t) should hold. But
the chattering amplitude is proportional to k (t). The objective of adaptation process
looks now transparent:
The gain k (t) should tend to a (t) /α with α ∈ (0, 1) which is very close to 1.
As a result, the minimal possible value of discontinuity magnitude is found for
the current value of parameter a (t) to reduce the amplitude of chattering. For that
purpose, select the adaptation algorithm in the form

⎫
k̇ (t) = ρk(t)sign (δ
 (t)) − M k(t)  − k + + + M [μ0 − k(t)]+ ⎪
⎪

 ⎪
⎬
δ (t) :=  sign (x (t)) eq  − α, α ∈ (0, 1)
 ⎪
(9.9)
1 if z ≥ 0 A ⎪
[z]+ := , M > ρk + , k + > a + , ρ > > 0. ⎪⎭
0 if z < 0 αμ0

The gain k can vary in the range μ0 , k + , μ0 > 0 is a preselected minimal value
of k. For the adaptation algorithm (9.9), sliding mode will occur after a finite time
interval. Indeed, if it does not exist, then



 sign (x (t)) eq  = 1

that leads to δ > 0, and the increasing gain k (t) will reach the value k + which
is sufficient for enforcing sliding mode for any value of parameter a (t). In sliding
mode, the adaptation process (9.9) with δ (t) = 0 is over after a finite time t f . Indeed,
calculate the time derivative of the Lyapunov function V (δ) = δ 2 /2 assuming that
9.2 The Dynamic Adaptation Based on the Equivalent Control Method 103

during the adaptation process k (t) ∈ μ0 , k + which means that |a (t)| /α > μ0


or (|a (t)| > αμ0 ). The time derivatives of  sign (x (t)) eq  (9.8) and |a (t)| exist,
and the terms depending on M in the adaptation algorithm (9.9) are equal to zero.
Calculate the time derivative of the Lyapunov function V (δ) by virtue of (9.8) and
(9.9):
⎫
d 
 d
V̇ (δ) = δ δ̇ = δ  sign (x) eq  = δ (|a| /k) = ⎪ ⎪
dt
dt ⎪⎪
⎬
− |a| δk −1 ρsign δ − M k − k + + + M [μ0 − k]+ (9.10)
⎪
⎪
−1 −1
+ δk ȧsign (a) ≤ − |a| δk ρsign (δ) + |δ| k A ⎪ −1 ⎪
⎭
≤ −αμ0 ρk −1 |δ| + |δ| k −1 A = − |δ| k −1 (αμ0 ρ − A)

and if ρ > A/αμ0 it follows

√ (αμ0 ρ − A) 
V̇ (δ) ≤ − 2 V (δ),
k+
√
implying that V (δ) = 0 at least after

k+  k+
tf= 2V (δ (0) )= |δ (0)|
(αμ0 ρ − A) (αμ0 ρ − A)

and, as a result, δ (t) becomes equal to zero identically after the finite time t f .
After the adaptation process is over (t > t f ), we have

 |a|

 sign (x (t)) eq  = = α.
k
So, k = |a| /α. If in the course of motion |a (t)| /α < μ0 , then the gain k (t) decreases
until k (t) = μ0 and, as it follows from (9.9), it will be maintained at this level. Since
the gain a(t) is time varying, its increase can result in |a (t)| /α = μ0 and δ (t) = 0 at
a time t f . As it
follows from the above analysis,
 for the further motion in the domain
k (t) ∈ μ0 , k + with the initial condition δ t f = 0 the time function δ (t) will be
equal to zero with α = |a (t)| /k (t).

Remark 9.1 The function sign (x (t)) eq is needed here for the implementation of
the adaptation algorithm (9.9). It can be derived by filtering out a high-frequency
component of the discontinuous function sign (x (t)) by a low-pass filter

τ ż + z = sign (x (t)) , z (0) = 0

with
constant τ > 0 and the output z (t, τ ) which is, in fact, an estimate
a small time
of sign (x (t)) eq satisfying



z (t, τ ) − sign (x (t)) eq  ≤ O (τ ) → 0.
τ →0
104 9 Adaptive SMC

Then the convergence analysis of (9.9) and (9.10) with δ (t, τ ) = z (t, τ ) − α is
valid beyond the domain |δ (t, τ )| ≤ O (τ ). This inequality defines the accuracy of
adaptation. Note that the switching frequencies of the modern power converters are
of order dozens of kHz, and very small time constant τ can be selected to get a high
accuracy of adaptation.

Notice also that, as follows from [4],

z(t, τ ) = sign (x (t)) eq + ψ(t) + O(sup|x (t) | + τ ) + O(τ −1 sup|x (t) |)

where ψ(t) is the fast rate exponentially decreasing function. The term sup|x (t) | is
inverse proportional to the sliding mode frequency f . It is of order of dozen k H z in
the modern switching devices. Therefore, it is not a problem to make the term

O(sup|x| + τ ) + O(τ −1 sup|x|)

negligible. Of course, this engineering language can be translated into mathematical

one, for example, as follows: for any ε > 0, there exists a switching frequency f 0
such that
|z(t, τ ) − u eq | < ε,

if f > f 0 implying



sign[ sign (x (t)) eq  − α] = sign[|z(t, τ )| − α].

9.2.2 Multidimensional Case

9.2.2.1 Main Assumptions

Here, we consider an arbitrary order system

⎫
ẋ (t) = f (t, x (t)) + b (t, x (t)) u (t, x (t)) ⎬
x (t) ∈ R n , f : R + × R n → R n (9.11)
⎭
u : R + × R n → R, b : R + × R n → R n

for which we assume that

A1 the control u = u (t, x) enforces siding mode on some surface σ (x) = 0 (σ ∈
C 1 ) and is in the following form:
   
u (t, x) = −k (t) 1 + λ
x
2 +ε sign (σ (x))

(9.12)
λ ≥ 0, ε > 0, k (t) ∈ μ0 , k + , μ0 > 0
9.2 The Dynamic Adaptation Based on the Equivalent Control Method 105

Similar to example (9.6), the control gain k (t) is a time-varying function gov-
erned by the adaptation procedure described below.
A2 the uncertain functions f (t, x) and b (t, x) satisfy the commonly accepted con-
ditions (which are much more general than in (9.3)):


f (t, x)
≤ f 0 + f 1
x
, 0 < b0 ≤ ∇  σ (x) b (t, x) ,
(9.13)

b (t, x)
≤ b+ ,
∇σ (x)
≤ σ +

∇  σ (x) f (t, x)
Φ(t, x) := ,
∇  σ(x) b (t, x) (9.14)
 ∂ 

∇  Φ(t, x)
≤ Φ 0 +Φ 1
x
,  Φ(t, x) ≤ ϕ0 +ϕ1
x
.
∂t

All coefficients in the right-hand sides of these inequalities are constant and positive.
The function σ (x) and its time derivative

σ̇ (x) = ∇  σ (x) f (t, x) −
 (9.15)

∇ σ (x) b (t, x) k (t) 1 + λ
x
2 + ε sign (σ (x))

should have opposite signs (σ (x) σ̇ (x) < 0 if σ (x) = 0), which is the condition
for sliding mode to exist on the surface σ (x) = 0. The sufficient condition for this
follows from (9.13), (9.14), and (9.15):

σ (x) σ̇ (x) = σ(x) ∇  σ (x) f (t,x) −



∇ σ (x) b (t, x) k (t) 1 + λ
x
2 +ε |σ (x)| ≤
   
[∇  σ (x) b (t, x)] |σ (x)| |Φ(t, x)| − k (t) 1 + λ
x
2 +ε < 0,

if   
|Φ(t, x)| − k (t) 1 + λ
x
2 +ε < 0 (9.16)

which always holds when



λ ≥ f 1 / f 0 , μ0 > f 0 σ + /b0 , k (t) ∈ μ, k + (9.17)

in view of the relation

  
σ + (1 +
x
f 1 / f 0 )
|Φ(t, x)| − k (t) 1 + λ
x
2 + ε ≤ f 0 − μ0 (1 + λ
x
) .
b0

To derive the sliding mode equation, the function sign (σ (x)) should be replaced by
the solution of the equation σ̇ (x) = 0 with respect to the term sign (σ (x)), called
the equivalent control:
106 9 Adaptive SMC

sign (σ (x)) eq :=
⎧
⎪ Φ(t, x) if
⎪
⎪   
⎨ σ (x (t)) =0
k (t) 1 + λ
x
+ ε2 (9.18)
⎪
⎪ if
⎪
⎩ sign (σ (x (t)))
σ (x (t)) = 0

satisfying (in view of (9.16)) in the sliding mode (σ (x (t)) = 0)




 sign (σ (x)) eq  < 1. (9.19)

9.2.2.2 Description of the Adaptation Procedure

The idea of the adaptation law for the control gain k (t) is similar to that for our
first-order system in the previous subsection:

(γ 0 + γ1
x
)

k(t)sign (δ (t))
k̇(t)= (9.20)
− M k(t) − k + + + M [μ0 − k(t)]+

where 


δ(t) :=  sign (σ (x)) eq  − α
(9.21)
α ∈ (0, 1) , λ > 0, γ0 , γ1 > 0.

f0
Select in (9.20) k + > σ + . If sliding mode does not exist, then
b0



 sign (σ (x)) eq  = 1

and the gain k(t) will be equal to k + which results in the occurrence of this motion
in the surface σ (x (t)) = 0.
Theorem 9.2 ([3]) For the dynamic system (9.11) closed by the control (9.12) with
the gain adaptation law (9.20) and (9.21) with the parameters satisfying

f0 ⎫
k + > σ + , μ0 > f 0 σ + /b0 , 0 < ε  1 ⎪
⎪
⎪
b0  ⎪
⎪
⎬
f 0 ϕ 0 +
γ0 > α−1 +b+ Φ0 + + f 0 +b k + (9.22)
μ0  μ0 ⎪
⎪
⎪
⎪
γ1 ≥ α −1 f0 +
+b Φ1 , M > γ0 k + ⎪
⎭
μ0

there exist
9.2 The Dynamic Adaptation Based on the Equivalent Control Method 107
  
f0 ϕ0
θ := αγ 0 − + b+ Φ0 + + f 0 +b+ k + > 0 (9.23)
μ0 μ0

and t f = θ−1 |δ (0)| (where δ (0) is defined by (9.21)) such that for all t ≥ t f the
condition 


 sign (σ (x (t))) eq  = α (9.24)

holds. It means that the sliding surface σ (x) = 0 is attained in a finite time t f ,
and for small enough positive number ε0 the suggested adaptation procedure with
α = 1 − ε0 provides k (t) tending to a vicinity of the minimum possible value kmin (t),
that is, as it follows from (9.18), in sliding mode
⎧
⎨ 1
kmin (t) i f kmin (t) ≥ μ
k(t) = 1 − ε0
⎩ μ i f kmin (t) < μ (9.25)
|Φ(t, x (t))|
kmin (t) :=  .
1 + λ
x (t)
2 + ε

9.3 Super-Twisting Control with Adaptation

The dynamic system with super-twist controller can be represented as

√ 
ẋ1 = x2 − ᾱ |x1 |sign(x1 ),
(9.26)
ẋ2 = φ(t) − ksign(x1 ).

Take σ(x) = x1 and permit for the gain parameter to be time-varying, i.e., k (t) =
β (t).

Theorem 9.3 (on adaptive super-twisting [3])

The system (9.26) with disturbances φ(t) having a bounded derivative (fulfilling
d
dt
|φ(t)| ≤ L), and with the parameter β̄(t) = k(t) adapted online according to the
adaptation law
⎧

⎪
⎪ γ0 k(t)sign (δ (t)) −M k(t) − k + + +M [μ0 −k(t)]+
⎨
if 0 < μ0 ≤ k (t) ≤ k +
k̇ (t) =
⎪
⎪
⎩
0 otherwise

where γ0 > L/μ converges in the finite time

t f = |δ (0)| k + / (μ0 γ0 − L)

to the sliding mode regime σ(x) = x1 = 0 maintaining within the relation

108 9 Adaptive SMC

φ(t)/k(t) = α = 1 − ε0

for small enough ε0 > 0. It means that the magnitude of the discontinuous control is
close to its minimal value as well as the amplitude of chattering.

References

1. Utkin, V.I., Poznyak, A.S.: Advances in Sliding Mode Control: Concept, Theory and Implemen-
tation. In: Adaptive Sliding Mode Control. Springer, Berlin (2013a)
2. Plestan, F., Shtessel, Y., Bregeault, V., Poznyak, A.: New methodologies for adaptive sliding
mode control. Int. J. Control 83(9), 1907–1919 (2010)
3. Utkin, V.I., Poznyak, A.S.: Adaptive sliding mode control with application to super-twist algo-
rithm: Equivalent control method. Automatica 49(1), 39–47 (2013a)
4. Utkin, V.: Sliding Modes in Control and Optimization. Springer, Berlin (1992)
Chapter 10
SMC in Infinite-Dimensional Systems

Abstract This brief chapter introduces a reader in the problem of SMC designing the
class of infinite-dimensional systems governed by a linear parabolic PDE containing
uncertainties. The distributed unit signal is considered as a robust controller. The
Lyapunov-Krasovskii functional is suggested for the analysis of the corresponding
closed-loop dynamics.

Keywords Infinite-dimensional systems · Parabolic PDE · Distributed unit

signal · Lyapunov-Krasovskii functional

Many important applications are governed by functional and partial differential equa-
tions or, more general, differential equations in a Hilbert space. Relevant examples
are flexible manipulators and structures, heat transfer processes, combustion, and
fluid mechanical systems, as well as retarded systems. As these systems are often
described by models with a significant degree of uncertainty, an interest has emerged
to develop consistent control methods that would be capable of utilizing distributed-
parameter and time-delay models and providing the desired system performance in
spite of significant model uncertainties.
Thus motivated, SMC algorithms have extensively been generalized in the infinite-
dimensional setting starting from early 1980s. First infinite-dimensional generaliza-
tions of SMC algorithms, made in [1–3] for some parabolic and hyperbolic PDEs,
not only corroborated their utility but also faced with challenging problems calling
for further investigation.

10.1 Infinite-Dimensional Setting

SM challenges, arising in the infinite-dimensional setting, can readily be illustrated

by a benchmark example of the heat diffusion along a 1-D bar with isolated ends.
Such a diffusion process is governed by the linear parabolic PDE

xt (r, t) = xrr (r, t) + f (r, t) + u(r, t) (10.1)

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 109
V. Utkin et al., Road Map for Sliding Mode Control Design,
SpringerBriefs in Mathematics,
https://doi.org/10.1007/978-3-030-41709-3_10
110 10 SMC in Infinite-Dimensional Systems

with the following Neumann boundary conditions:

xr (0, t) = 0, xr (1, t) = 0, (10.2)

where x(·, t) is the temperature distribution along the bar at a time instant t > 0, and
the control input u and external disturbance f affect the bar temperature through the
entire spatial bar location r ∈ (0, 1). Hereinafter, the symbols

∂x(r, t) ∂x(r, t) ∂ 2 x(r, t)

xt (r, t) = , xr (r, t) = , xrr (r, t) =
∂t ∂r ∂r 2

stand for the corresponding partial state derivatives, and the symbol x(·, t)2 is
1 2
reserved for the L 2 -norm 0 x (r, t) dr .
Let the distributed unit signal

x(r, t)
u(r, t) = −M , (10.3)
x(·, t)2

magnified by the gain M > 0, is applied to the heat process (10.1) for enforcing the
plant dynamics to slide the plant dynamics along the discontinuity set x = 0. By
inspection, the L 2 -norm of the unit signal x(r, t)x(·, t)−1
2 does equal one every-
where but the origin x ≡ 0. Fundamental questions then arise:
• Under which conditions a sliding mode occurs on the discontinuity set?
• What equivalent value of the unit signal x(r, t)x(·, t)−1
2 should be utilized in the
boundary-value problem (10.1), (10.2) to correctly describe the sliding mode on
the discontinuity set x = 0?
Being inspired by the finite-dimensional counterpart, the conditions

lim x↑0 xt (r, t) = lim x↑0 [xrr (r, t) + f (r, t) + u(r, t)] > 0,
(10.4)
lim x↓0 xt (r, t) = lim x↓0 [xrr (r, t) + f (r, t) + u(r, t)] < 0

for sliding dynamics to locally exist along the discontinuity set x(·, t) = 0 require
the controller magnitude M to exceed (locally in t) not only the uniform norm
 f (·, t)C(0,1) of the external disturbance, but also the uniform norm xrr (·, t)C(0,1)
of the second-order spatial derivative xrr . The latter requirement is, however, impossi-
∂2
ble to meet because the spatial double differentiation operator A = 2 is unbounded
∂r
(indeed, A cos (nr ) C(0,1) → ∞ as n → ∞). Thus, the SM existence condition
(10.4) should be revised in the infinite-dimensional setting.
The Lyapunov-Krasovskii (Lyapunov-Razumikhin) approach is presently well
recognized for revealing SMs in the PDEs (respectively, delayed functional differen-
tial equations) setting [4]. In a particular case of the closed-loop system (10.1)–(10.3),
a Lyapunov-Krasovskii functional can be chosen in the form
10.1 Infinite-Dimensional Setting 111
 1
1
V (x) = x 2 (r, t)dr. (10.5)
2 0

Indeed, differentiating (10.5) along the solutions of (10.1)–(10.3) yields

 1
V̇ = [xrr + u + f ]xdr ≤ −(M −  f (·, t) L 2 )x(·, t) L 2
0

= −(M −  f (·, t) L 2 ) 2V (t). (10.6)

Thus, provided that

M >  f (·, t) L 2 ,

the finite-time stability of the closed-loop system (10.1)–(10.3) is established and

the existence of sliding modes on the discontinuity set x = 0 is guaranteed in finite
time √ 
T = 2(M −  f (·, t) L 2 )−1 V (0).

10.2 Unit Control for Infinite-Dimensional Systems

Generalization of the unit control design for arbitrary systems in Hilbert space is
further presented. For infinite-dimensional dynamics

ẋ = Ax + Bu + f (t, x),

evolving in a Hilbert space and affected by unknown disturbances, the typical control
objective is their asymptotical stabilization. The state and control are elements of
Hilbert spaces x ∈ X , u ∈ U , f : R × X → X is a disturbance,

A : D(A) ⊂ X → X

is a closed linear operator with the domain D(A) dense in X and B : U → X is a

linear bounded operator, the unknown disturbance f (t, x) is assumed to be bounded
with an a priori known norm bound. Of course, the problem can be solved if the
matching condition

f (t, x) = Bγ(t, x), γ(t, x) ≤ γ0 , γ : R × X → U

holds.
Without touching details on strong solutions of the closed-loop system to exist (the
interested reader can refer, e.g., to [4, 5]), select the design procedure, following the
methodology, offered in Sect. 3.8 and in [6, 7]. Such a procedure becomes applicable
112 10 SMC in Infinite-Dimensional Systems

if the nominal system ẋ = Ax is exponentially stable with a known positive-definite

continuous Lyapunov functional

v : X → R, v(0) = 0, v(x) > 0 for x = 0

such that
v̇ = ∇x {v}Ax < 0,

where ∇x {v} : X → R is a linear functional similar to ∇  V : R n → R for finite-

dimensional space with the property

v(x + x) − v(x) − ∇x {v}(x)

lim = 0, x ∈ X.
x→0 x

Calculating the time derivative of v(x) on trajectories of the original systems yields

v̇ = ∇x {v}Ax + ∇x {v}B(u + γ).

Let
(∇x {v}B)∗ ∈ U

be an adjoint operator to ∇x {v}B [8]. By property of an adjoint operator [9],

∇x {v}B (∇x {v}B)∗ = ∇x {v}B2 .

For control in the form

(∇x {v}B)∗
u = −ρ , ρ > γ0
∇x {v}B

(with the same motivation as before in Sect. 3.8 it can be called unit control), we
derive
v̇ ≤ ∇x {v}Ax + ∇x {v}B(−ρ + γ0 ) < 0.

Hence, the state x = 0 is asymptotically stable provided that strong solutions of the
closed-loop system exist.
Unit control approach in the infinite-dimensional setting has successfully been
illustrated for the parabolic boundary-value problem (10.1), (10.2). Note that along
with the unit control input (10.3), the distributed relay control u = −ρsign(x) of the
diffusion process results in the asymptotic stabilization as well, but an upper estimate
of the decay of x(t, r ) has not been found until now. Even a solution concept for
the closed-loop diffusion system, driven by such a distributed relay input, should be
revisited. In addition, the unit control is a continuous function during the transient
process and becomes discontinuous only when the process is over and x becomes
equal to zero for any point r . It is not the case for the distributed relay control,
undergoing discontinuity at the points, where x(t, r ) = 0.
10.2 Unit Control for Infinite-Dimensional Systems 113

It is worth noticing that in the above example, the sliding mode set x = 0 manifests
the trivial sliding mode dynamics ẋ = 0 itself. However, in order to describe sliding
mode dynamics in less trivial examples where sliding modes appear to be not con-
fined to the origin, the equivalent control method should properly be revalidated in the
infinite-dimensional setting. Infinite-dimensional extensions of the equivalent con-
trol method, capturing a broad class of semilinear PDEs, were developed in [10, 11].
These extensions, coupled to the Lyapunov-Krasovskii and Lyapunov-Razumikhin
approaches, resulted in numerous SMC algorithms of the infinite-dimensional sys-
tems. Relevant references may be found in [4] for Distributed-Parameter Systems
(DPS) and in [12] for time-delay systems.
Recently, a powerful backstepping tool was extended to the PDEs setting [13]
through a non-trivial integral state transformation for the purpose of stabilizing lin-
ear DPS of infinite relative degree with anticollocated boundary sensing and actua-
tion. This technique was then combined with SMC algorithms in order to robustify
the existing PDEs-oriented boundary controllers. Such combined SM-backstepping
algorithms were proposed in [14–16] to name a few.
The modern trend in the DPS design is twofold. The closed-loop DPS is required
to possess a good performance not only in the L 2 state space, but also in a Sobolev
space. Having a good performance in the former would admit destroying peaks
within small spatial subdomains of the operation. Apart from this, having a good
performance in a Sobolev space would result in the uniform point-wise state decay,
thereby avoiding extremely undesired peaking phenomena within the spatial oper-
ation domain. Motivated by this trend, twisting and super-twisting algorithms were
recently re-worked for uncertain heat and wave propagation processes (see [17] and
references therein) to address their point-wise stabilization and observation where the
available sensing and actuation were either distributed over the entire state space or
located at the boundary. Certainly, the SM theory of the infinite-dimensional systems
is in progress, and many SMC challenges remain open in the infinite-dimensional
setting and promising directions in their further investigation are announced in the
next section.

References

1. Orlov, Y., Utkin, V.I.: Use of sliding modes in distributed system control problems. Autom.
Remote Control 43, 1127–1135 (1982)
2. Orlov, Y.: Application of Lyapunov method in distributed systems. Autom. Remote Control
44, 426–430 (1983)
3. Zolezzi, T.: Differential inclusions and sliding mode control. In: Sliding Mode Control in
Engineering (W. Perruquetti and J.P. Barbot, Eds.). Marcel Dekker, pp. 29–52 (2002)
4. Orlov, Y.: Discontinuous Systems: Lyapunov Analysis and Robust Synthesis Under Uncertainty
Conditions. Springer, Berlin (2008)
5. Curtain, R., Zwart, H.: An Introduction to Infinite-dimensional Linear Systems. Springer, New
York (1995)
6. Gutman, S.: Uncertain dynamic systems—a lyapunov min-max approach. IEEE Trans. Autom.
Control AC-24, 437–449 (1979)
114 10 SMC in Infinite-Dimensional Systems

7. Gutman, S., Leitmann, G.: Stabilizing feedback control for dynamic systems with bounded
uncertainties. In: Confrence on Decision and Control, pp. 94–99 (1976)
8. Orlov, Y., Utkin, V.: Unit sliding mode control in infinite-dimensional systems. J. Appl. Math.
Comput. Sci. 8, 7–20 (1998)
9. Rudin, W.: Functional Analysis, 2nd edn. McGraw-Hill, New York (1991)
10. Orlov, Y., Dochain, D.: Discontinuous feedback stabilization of minimum-phase semilinear
infinite-dimensional systems with application to chemical tubular reactor. IEEE Trans. Autom.
Control 47, 1293–1304 (2002)
11. Levaggi, L.: Infinite dimensional systems sliding motions. Eur. J. Control 8, 508–518 (2002)
12. Yan, X., Spurgeon, S., Edwards, C.: Variable-Structure Control of Complex Systems. Springer,
Berlin (2017)
13. Krstic, M., Smyshlyaev, A.: Boundary Control of PDEs: A Course on Backstepping Designs.
SIAM, Philadelphia PA (2008)
14. Guo, B.-Z., Jin, F.-F.: Sliding mode and active disturbance rejection control to stabilization
of one-dimensional anti-stable wave equations subject to disturbance in boundary input. IEEE
Trans. Autom. Control 58(13), 1269–1274 (2013)
15. Cheng, M.B., Radisavljevic, V., Su, W.C.: Sliding mode boundary control of a parabolic pde
system with parameter variations and boundary uncertainties. Automatica 47(2), 381–387
(2011)
16. Guo, B.Z., Jin, F.-F.: Output feedback stabilization for one-dimensional wave equation subject
to boundary disturbance. IEEE Trans. Autom. Control 60, 824–830 (2015)
17. Pisano, A., Orlov, Y.: Boundary second-order sliding-mode control of an uncertain heat process
with unbounded matched perturbation. Automatica 48, 1768–1775 (2012)
Chapter 11
Open Problems in SMC

Abstract This chapter describes several open problems that can be formulated in
the frame of SMC concept. They are
– stability problem of the system ẋ = A sign(x) ,
– SMC in the presence of noise in measurements,
– finite-time stability of SMs in infinite-dimensional systems,
– SMC for systems governed by shift operators, and
– stochastic sliding mode.

Keywords Stability problem · Noise in measurements · Finite-time stability of

SMs in infinite-dimensional systems · Shift operators · Stochastic SM

There are many open challenging problems in each of the above chapters. Certainly,
it does not exclude totally new problems which are formulated as well.

11.1 Stability Problem of the System ẋ = A sign(x)

We hope that the stability conditions for the system

ẋ(t) = Asign(x(t)), (11.1)

where
x(t) = (x1 (t), . . . , xn (t))T ∈ R n ,

and A ∈ R n×n being a constant matrix, can be used for finding sliding domains in
the system (2.7). Note that necessary and sufficient stability conditions for (11.1)
were established for two-dimensional systems in [1]. The existence of sliding mode
on the surface s = 0 (see, (3.3)) is determined by stability of the equation

ṡ(t) = −M D(t)sign(s) + d(t),

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 115
V. Utkin et al., Road Map for Sliding Mode Control Design,
SpringerBriefs in Mathematics,
https://doi.org/10.1007/978-3-030-41709-3_11
116 11 Open Problems in SMC

where M > 0 is a tuning parameter, d(t) is a disturbance, and D(t) is a time-varying

matrix. We hope that stability conditions of the time-varying system can be formu-
lated in terms of stability of time-invariant system (11.1). Then, an optimal (in some
sense) selection of a sliding surface may improve significantly the control perfor-
mance.
To the moment, we have only one argument in favor of our hope. Let us assume
that we can find Lyapunov function as a quadratic form V = x T P x for any constant
matrix A with matrix P > 0, depending on A, such that

V̇ = x T P Asign(x) + sign T (x)A T P x < 0.

Let now A(t) be a time-varying matrix with bounded time derivative, and let the
system be “skillfully” stable, that is, the real part of the spectrum of matrix A is equal
or less than −δ, δ > 0 for any fixed time t. We can find Lyapunov function

V (t) = x T P(t)x > 0

 
with P(t) > 0 depending on time. It can be shown that P(t) as well as  Ṗ(t)
are bounded. Calculate the derivative of V (t):

V̇ (t) = x T P(t)A(t)sign(x) + sign T (x)A T (t)P(t)x + x T Ṗ(t)x.

The first two terms tend to zero with x tending to zero as a linear function, whereas
the third term tends to zero as a quadratic function. It means that the origin in space
x is asymptotically stable for the time- varying system as well (of course it is not
always the case for linear systems ẋ = A(t)x). Then, the sliding domain can be found
from equation
ṡ = G(x, t) f (x, t) + G(x, t)B(x, t)u,

if the matrix G(x, t)B(x, t) satisfies stability condition for any constant x and t.
For linear systems ẋ(t) = Ax(t), the necessary and sufficient condition of asymp-
totic stability of the origin is well known: all eigenvalues of the matrix A possess
negative real parts. However, this is not the case for the relay system (11.1). For
instance, the matrix ⎛ ⎞
−3 1 1
A = ⎝4.5 −1 2 ⎠
4.5 −1.5 −1

is Hurwitz, but the corresponding relay system (11.1) is unstable. In this particular
case, instability can be easily proven by means of the equivalent control method.
Indeed, x1 = 0 is the globally attractive sliding surface and one has

[sign(x1 )]eq = 1/3(sign(x2 ) + sign(x3 ))

11.1 Stability Problem of the System ẋ = A sign(x) 117

as soon as sliding mode occurs. It becomes evident that the system is unstable after
substitution of the equivalent control into the third equation, since

ẋ3 = 0.5sign(x3 ).

Several sufficient stability conditions for (11.1) have been presented in [2–4]. The
simplest one can be formulated in the form of linear matrix inequality:

A T Λ + ΛA < 0,

where Λ is a diagonal positive-definite matrix. The necessary and sufficient condition

of asymptotic stability of the relay system (11.1) is still not discovered. We consider
this as the open problem, which is important for tuning of relay sliding mode control
algorithms.

11.2 SMC in the Presence of Noise in Measurements

The output-based sliding mode design is usually studied without any assumption
about noise in the measurement. However, in practice, the measurements are usually
corrupted by noise, and an appropriate modification of the sliding mode control
methodology is required.
Consider the system

ẋ(t) = Ax(t) + Bu(t) + Dg(t), t > 0,

y(t) = C x(t) + ω(t), x(0) ∈ x0 ,

where x(t) ∈ R n is the system state, u(t) is the control input, y(t) ∈ R k is the mea-
sured output, the functions ω : R → R k , and g : R → R p describe measurement
noise and exogenous disturbance, respectively, the matrices

A ∈ R n×n , B ∈ R n×m , C ∈ R k×n , D ∈ R n× p

are assumed to be known. Note that for Dg(t) ∈ range(B) the exogenous disturbance
becomes matched.
As usual, we deal with the classical sliding mode control problem: to design a
control algorithm, which realizes finite-time reaching of a given linear manifold

F x = 0, F ∈ R m×n , det(F B) = 0

and further sliding on this manifold.

Due to measurement noises and system disturbances, the sliding mode cannot
appear on the given manifold F x = 0. For practice, it is important to know that
118 11 Open Problems in SMC

which sort of feedback control is optimal to provide the system motion as close as
possible to the preselected manifold.
The case of L 2 bounded noises and disturbances has been studied using the mini-
max observer design in [5]. It was proven that a possible optimal control law realizing
required motion (F x(t) → min) is a linear feedback. Notice that in the presence
of the observation noise ω(t), this minimum never can be done equal to zero (inde-
pendently whether the noise is from L 2 or from L ∞ ). In the case of a measurement
noise, the convergence of both the state estimates and the distance to the desired
manifold to some zone can be guaranteed only. The following question arises: under
which restrictions to measurement noises and disturbances the SM controllers and
observers provide the convergence of the system trajectories to a smaller zone than
linear algorithms? It can be considered as an interesting open problem.
Another important problem is the optimal selection of the sliding manifold
F x = 0 to minimize the effects of mismatched perturbations. For the noise-free
state feedback sliding mode control with y = x, the sliding mode control problem
has been studied in [6] for exogenous disturbances bounded in L ∞ . The background
for possible future developments in this context is presented in [7], where the general
framework for L 2 -gain analysis of sliding mode controllers is introduced.

11.3 Finite-Time Stability of SMs in Infinite-Dimensional

Systems

The extension of the SM analysis tools to the infinite-dimensional setting is far from
being trivial. Although the finite-time extinction in the infinite-dimensional setting
has been recognized in the literature for (linear, in particular) evolution equations
(see [8] and references therein), however, the finite-time stability analysis of SMs
remains a challenging problem for infinite-dimensional systems. The strict Lyapunov
functional (10.5), resulting in the finite-time stability of the closed-loop heat process
(10.1), (10.2), enforced by the unit distributed control signal (10.3), remains the only
one which is available in the PDE setting [9].
It should be pointed out that the conventional homogeneity approach proves to be
useless in the finite-time stability analysis of the above closed-loop system (10.1),
(10.2), (10.3). Indeed, the right-hand side of the PDE (10.1) contains linear and unit
terms, which are of distinct homogeneity degree on one hand, so that it cannot be
homogeneous in the standard sense. On the other hand, the unit term, magnified
with a finite gain, cannot locally exceed the linear unbounded operator of the double
differentiation so that it cannot reject (even locally) such a disturbing term. Such a
situation is opposed to that in the finite-dimensional setting.
Remarkably, the extension of the homogeneity concept to infinite-dimensional
systems, proposed in [10], is neither suitable for the finite-time stability analysis of
the rather trivial distributed relay control example, given above, and it calls for an
appropriate modification in the infinite-dimensional setting. In addition, the existing
counterparts of the twisting and super-twisting algorithms, developed in the heat
11.3 Finite-Time Stability of SMs in Infinite-Dimensional Systems 119

and wave PDEs settings [11], manifest the need for the deeper finite-time stability
insight on these algorithms in the PDE setting as they have been established to yield
the closed-loop asymptotic stability only.

11.4 A Little Fantasy: SMC for Systems Governed by Shift

Operators

Let us consider the evolution equation [12] of the form

ẋ(t) = f (t, x(t)), t > t0 , x(t0 ) = x0 , (11.2)

where x(t) is an element of an appropriate finite or infinite-dimensional linear space

X and f : R × X → X is an appropriate operator (function). We assume that the
Cauchy problem (11.2) has a solution p(t, t0 , x0 ) for any x0 ∈ X.
A model of dynamic system usually considers its state x(t) (or x(k)) as an element
(point) of an appropriate finite- or infinite-dimensional normed linear space X , where
t denotes a continuous time (and k denotes a discrete time). The evolution of the
dynamic system in time defines a trajectory (curve) in the space X . A point x(t0 ) = x0
starting from an instant of time t0 moves along the trajectory and reaches a point
x(t1 ) = x1 at a time instant t1 . The operator

U (t1 , t0 ) : X → X

of translation of the point x(t0 ) = x0 to the point x(t1 ) = x1 is called shift operator
[13] (or “solution operator”). The operator U (t1 , t0 ) is single-valued if the trajectory
is unique. The classical example of the shift operator is the matrix exponent

U (t1 , t0 ) = e A(t1 −t0 )

for a linear ODE:

ẋ(t) = Ax(t), x ∈ Rn , A ∈ Rn×n .

We confine ourself to an autonomous case only. In this case, a dynamic system

has an evolution law, which is invariant with respect to time, so the shift operator
depends only on t1 − t0 (i.e., U (t1 , t0 ) = U (t1 − t0 )), so U (s) is a semigroup on X ,
indeed, U (0) = I is an identity operator and

U (t + s) = U (t)U (s), t, s > 0

is a semigroup property. The shift operator may be unbounded if trajectories blow up

(i.e., x(t) X → +∞ as t → T∗ < +∞). Note that if the trajectories are not unique
in the backward time, then the shift operator is not invertible or, at least, its inverse
120 11 Open Problems in SMC

is not single-valued. The latter case is the most interesting for us, since the sliding
mode control principle does not allow the system to have a unique solution in the
backward time. Any point of sliding manifold can be reached from the outside or
along a trajectory in the manifold. For this general framework, the sliding manifold
can be introduced by the same Definition 2.1 (or Definition 8.1 for discrete-time
systems) using a shift operator U , a properly adapted notion of a manifold in X and,
for example, finite-time extinction property [14–16].
For strongly continuous semigroups of linear bounded operators with U (t) ∈ L,
where L is a space of linear bounded operators, finite-time extinction can be defined
as follows: [16, 17]: U (t)L = 0 for t > t0 , where t0 > 0 is a sort of settling time.
It is worth stressing that finite-time extinction in the infinite-dimensional framework
may demonstrate even linear evolution equations [8, 16, 18]. Development of sliding
mode control principle for abstract evolution system based on semigroup approach
appears to be an interesting theoretical problem. Some attempts in this field were
already done in [19–22].

11.5 On Stochastic SM

The publications on SMC have been devoted to the stability analysis of a class of
nonlinear discontinuous feedback systems containing bounded deterministic uncer-
tainties. Since stochastic systems contain unbounded perturbations typically having
Gaussian distributions, the direct extension of SM ideas, successfully applied for
deterministic systems, should be considerably changed or at least constructively
reconsidered. After earlier 80s, some publications (see [23], [3, Chap. 14]), [24–28])
dealing with nonlinear Itô-type stochastic systems appeared in different scientific
journals. They studied such new effects as actuator nonlinearities, time-delay terms,
control input with sector nonlinearities and dead zones, Markovian jump parameters,
polynomial system over linear observations, and so on. It is important to notice that
majority of the cited references dealt with the so-called multiplicative noise where
its power was proportional to the norm a sliding variable decreasing to zero when
process approached the desired sliding surface.
To describe arising problems in stochastic SM systems, consider the simplest
controllable stochastic model given in the Itô form in some filtered probability space
⎫
d x1,t = x2,t dt, ⎬
d x2,t = |u (xt , t) + f (xt , t)] dt + σ1 (xt , t) dW1,t , (11.3)
⎭
dyt = Cd xt + σ2 (xt , t) dW2,t ,

where
  
xt := x1,t , x2,t ∈ R 2n , yt ∈ R m

and
σ1 (xt , t) ∈ R n×n , σ2 (xt , t) ∈ R m×n
11.5 On Stochastic SM 121

are diffusion matrices of the standard n-dimensional Wiener processes, i.e.,

a.s.   a.s.
E {Wit | Ft−0 } = 0, E Wit Wit | Ft−0 = t In×n , i = 1, 2.

Function f (xt , t) is supposed to be bounded ( f (x, t) ≤ L ) but unknown. Let

the sliding surface s (x) = 0 is defined as

s (x) := x2 + x1 = ẋ1 + x1 = 0 (11.4)

and the sliding mode control has the structure


u t = −K xt − k (x, t) sign (s (xt )) , 0 < k (xt , t) ∈ R1 ,
(11.5)
sign (s) := (signs1 , . . . , signsn ) , 0 < K ∈ Rn×2n .

The strong solutions of the closed-loop system are considered, see [29]. We intend to
design K and k (x, t) in (11.5) which guarantee the stabilization of the system (11.3)
in a μ-neighborhood of the sliding surface s (x) = 0 (11.4) in some probabilistic
sense.
The partial case when m = n , C = In×n (all states are observable) and σ2 (xt , t) =
0 is considered in [30]. As shown in [30], selection
⎫
u t = −x2,t − k (xt , t) SIGN (s (xt )) ⎪

  k0  ⎪⎪
⎬

k (xt , t) = Lϕε x1,t + x2,t +   x1,t + x2,t 
 2 ⎪
(11.6)
1 if s > ε ⎪
⎪
ϕε (s) := −1 , ε>0 ⎭
ε s if s ≤ ε

tr {σσ  } + 2Lε
guarantees the exponential μ-zone (with μ = ) convergence for the
2k0
second moment of the state, i.e.,
  
E s (xt )2 − μ + ≤ Ce−k0 t → 0 when t → ∞.

The stochastic analog of the deterministic super-twisting control may be expressed

as
  √ 
d xt =  −α |xt |signx(t) +
 yt dt
dyt = f (xt , t) − βsignxt dt + σdW (t)
a.s.
| f (xt , t)| ≤ L , σ > 0, α = α (x, y, t) , β = β (x, y, t) .

As it is shown in [31], if we select

α = α0 , β = β0 ϕε (x) + βad
122 11 Open Problems in SMC

with α0 , β0 such that

 
α0 α0 β0
min , > L , α0 (β0 − L) > L , α0 > 4L
α0 >0, β0 >0 2 1 + α0

and st
βad = − sign (vt ) signx,
|vt | + εt

where
√ √
vt = 2yt − α |xt |sign (xt ) , st = k Vt − θVt ,
 √ 2
Vt = 2β |xt | + 21 yt2 + 21 yt − α ⎧|xt |sign (xt ) ,
 ⎨ any > 0 if st +
≥ 0,
1 if |x| > ε > 0, |vt |
ϕε (|x|) := −1 , εt =
ε |x| if |x| ≤ ε, ⎩
if st +
< 0,
|st +
|

we can guarantee the mean-square exponential convergence of Vt in the prespecified

σ2 +

zone μ = , namely,
θ


−2θt  z if z ≥ 0,
[E {Vt } − μ]+ = O e
2
→ 0, [z]+ :=
t→∞ 0 if z < 0.

Possible applications of other Lyapunov’s functions Vt are also discussed in [31].

As it was expected, sliding mode cannot exist in the traditional deterministic sense,
governed by stochastic differential equations similar to (11.3). For the proposed con-
trol algorithm (11.6), we can speak about asymptotic convergence of the mean square
value to vicinity of manifold s = 0. One more interesting aspect is the substantiation
of validity of sliding mode equations using stochastic differential equations based
on the regularization method (Sect. 2.6). It implies that due to small imperfections
solution exists in the conventional sense and state trajectories are not confined to
manifold s = 0, but run in a boundary layer, tending to zero if imperfections tend
to zero. This limiting procedure results in equivalent control method equations for
deterministic affine systems. Stochastic noise added to function s may be handled
as small imperfection, if the gain multiplying noise tends to zero. At the intuitive
level, this new regularization method should result in the equivalent control method
equation as well. However, the condition “the trajectories run in a boundary layer”
is not true for stochastic systems. Open problem: show the validity of the equivalent
control method equations by this new regularization method.
Summarizing the results above, we can state that for the additive stochastic noise
the special design (11.6) of the sliding mode gain parameters (in fact, linearly depend-
ing on the norm of the sliding variable s) may guarantee the exponential convergence
of the averaged squared norm of the state vector to a μ-zone (μ-neighborhood) around
the sliding surface s(x) = 0. This zone turns out to be exactly proportional to the
11.5 On Stochastic SM 123

diffusion parameter σ in the model description. Certainly, stochastic systems with

SM controllers require very profound further investigation. As one can see, the sys-
tem (11.3) with controller (11.5) should be considered as a stochastic differential
inclusion. The existence of the solution for such system is also discussed in [30].
Several versions, such as
– the equivalent control method for stochastic SM,
– stochastic twisting and nested controllers,
– stochastic integral SM, stochastic SM observers,
– output-based SM with stochastic perturbations, and
– adaptive stochastic SM,
studied in [3, 32–34], require the extensions for the class of stochastic systems. In
particular, the system with super-twisting control is discussed in ([31]).

References

1. Hsu, L., Kaszkurewicz, E., Bhaya, A.: Matrix-theoretic conditions for the realizability of sliding
manifolds. Syst. Control Lett. (2000)
2. Filippov, A.F.: Differential Equations with Discontinuous Righthand Sides. Kluwer Academic
Publishers (1988)
3. Utkin, V.: Sliding Modes in Control and Optimization. Springer, Berlin (1992)
4. Polyakov, A.: On settling time function and stability of vector relay systems. In: 12th IEEE
International Workshop on Variable Structure Systems, pp. 149–154 (2012)
5. Zhuk, S., Polyakov, A., Nakonechnyi, O.: Note on minimax sliding mode control design for
linear systems. IEEE Trans. Autom. Control 62(7), 3395–3400 (2017)
6. Polyakov, A., Poznyak, A.: Invariant ellipsoid method for minimization of unmatched distur-
bance effects in sliding mode control. Automatica 47(7), 1450–1454 (2011)
7. Osuna, T., Orlov, Y.: l2 -gain analysis of sliding mode controllers. In: Workshop on Variable
Structure Systems. Nantes, France (2014)
8. Perrollaz, V., Rosier, L.: Finite-time stabilization of 2x2 hyperbolic systems on tree-shaped
networks. SIAM J. Control Optim. 52(1), 143–163 (2014)
9. Orlov, Y., Utkin, V.: Unit sliding mode control in infinite-dimensional systems. J. Appl. Math.
Comput Sci 8, 7–20 (1998)
10. Polyakov, A., Efimov, D., Fridman, E., Perruquetti, W.: On homogeneous distributed parameter
systems. IEEE Trans. Autom. Control 61(11), 3657–3662 (2016)
11. Orlov, Y., Pisano, A., Sodina, S., Usai, E.: On the lyapunov-based second order sm for some
classes of distributed parameter systems. IMA J. Math. Control Inf. 29(4), 437–457 (2012)
12. Pazy, A.: Semigroups of Linear Operators and Applications to Partial Differential Equations.
Springer, Berlin (1983)
13. Krasnoselskii, M.A.: Shift Operator for Differential Equations. Nauka, Moscow (1966)
14. Sabinina, E.S.: A class of non-linear degenerating parabolic equations. Soviet Methematics
Doklady 148, 495–498 (1962)
15. Galaktionov, V.A., Vazquez, J.L.: Necessary and sufficient conditions for complete blow-up
and extinction for one-dimensional quasilinear heat equations. Arch. Rat. Mech. Anal. 129(3),
225–244 (1995)
16. Balakrishnan, A.V.: Superstability of systems. Appl. Math. Comput. 164, 321–326 (2005)
17. Gohberg, I.C., Krein, M.G.: Theory and Applications of Volterra Operators in Hilber Space.
American Mathematical Society, Providence, Rhode Island (1970)
124 11 Open Problems in SMC

18. Creutz, D., Mazo, M. Jr., Preda, C.: Superstability and finite time extinction for c0 -semigroups
(2013). arXiv:0907.4812v4
19. Levaggi, L.: a). Infinite dimensional systems sliding motions. Eur. J. Control 8, 508–518 (2002)
20. Levaggi, L.: b). Sliding modes in banach spaces. Differ. Integr. Equ. 15, 167–189 (2002)
21. Orlov, Y., Dochain, D.: Discontinuous feedback stabilization of minimum-phase semilinear
infinite-dimensional systems with application to chemical tubular reactor. IEEE Trans. Autom.
Control 47, 1293–1304 (2002)
22. Yan, X., Spurgeon, S., Edwards, C.: Variable-Structure Control of Complex Systems. Springer,
Berlin (2017)
23. Drakunov, S.V.: On adaptive quasioptimal filter with discontinuous parameters. Autom. Remote
Control 44(9), 1167–1175 (1983)
24. Niu, Y., Ho, D.W.C.: Design of sliding mode control for nonlinear stochastic systems subject
to actuator nonlinearity. Control. Theory Appl., IEE Proc. 153(6), 737–744 (2006)
25. Niu, Y., Liu, Y., Jia, T.: Reliable control of stochastic systems via sliding mode technique.
Optim. Control. Appl. Methods 34(6), 712–727 (2013)
26. Wu, L., Ho, D.W.C.: Sliding mode control of singular stochastic hybrid systems. Automatica
46(4), 779–783 (2010)
27. Gao, L., Wu, Y.: Control for stochastic systems with markovian switching and time-varying
delay via sliding mode design. Math. Problems Eng. (2013)
28. Basin, M., Rodriguez-Ramirez, P.: Sliding mode controller design for stochastic polynomial
systems with unmeasured states. IEEE Trans. Ind. Electron. 61(1), 387–396 (2013)
29. Da Prato, G., Frankowska, H.: A stochastic Filippov theorem. Stoch. Anal. Appl. 12(4), 409–
426 (1994)
30. Poznyak, A.: Sliding mode control in stochastic continuos-time systems: μ-zone ms-
convergence. IEEE Trans. Autom. Control 62(3), 863–868 (2017)
31. Poznyak, A.: Stochastic super-twist sliding mode controller. IEEE Trans. Autom. Control 63(5),
1538–1544 (2018)
32. Polyakov, A., Poznyak, A.: Reaching time estimation for super-twisting second order sliding
mode controller via lyapunov function designing. IEEE Trans. Autom. Control 54(8), 1951–
1955 (2009)
33. Plestan, F., Shtessel, Y., Bregeault, V., Poznyak, A.: New methodologies for adaptive sliding
mode control. Int. J. Control 83(9), 1907–1919 (2010)
34. Utkin, V.I., Poznyak, A.S.: Adaptive sliding mode control with application to super-twist algo-
rithm: Equivalent control method. Automatica 49(1), 39–47 (2013)
Chapter 12
Conclusions

1. Mathematical methods
All publications in this area since 50s until now can be separated into two groups:
– Sliding mode equations are postulated.
– Validity of equations is substantiated by regularization methods.
The universal definitions of sliding manifold and sliding domain are discussed
in the book; it is demonstrated that generally speaking sliding mode equations
cannot be derived unambiguously and the cases when it is possible are singled
out.
2. Design procedures
The design methodology for SMC is the same in majority of publications:
– A discontinuity manifold is selected such that the sliding motion of a reduced
order exhibits the desired properties.
– Discontinuous control is selected to enforce the sliding mode.
Although dozens of publications can be found in the framework of this
methodology, all of them are under the umbrella of the following modifi-
cations: component-wise discontinuous control, unit control, integral control,
and high-order sliding mode control. The design ideas of each of them are
explained in our paper. Selection of discontinuity manifold is a reduced-order
conventional problem of the control theory. Enforcing sliding mode is equiv-
alent to a stability problem also of a reduced order. The ideas of designing
Lyapunov functions for each modification are explained.
3. Lyapunov Stability Tools
Different forms of Lyapunov functions along with the homogeneity concept are
offered to guarantee finite-time convergence, estimate the convergence time, and
provide the desired convergence time by a proper choice of control parameters.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 125
V. Utkin et al., Road Map for Sliding Mode Control Design,
SpringerBriefs in Mathematics,
https://doi.org/10.1007/978-3-030-41709-3_12
126 12 Conclusions

4. High-order sliding mode control

The new phenomenon was discovered: (n − k)-dimensional manifold (k > 1) is
reached after a finite time interval, and then k-dimensional sliding mode, called
HOSM, occurs. HOSM is intended to overcome the problems related to relative
degree and chattering.
5. SM observers
All sliding mode observers in numerous publications are based on the same
idea: enforcing the sliding mode along the observer input signal, specified as
a discontinuous function of the observation error. As shown in our book, the
majority of publications demonstrate that
– order of the system is reduced;
– the state vector can be found for time-varying systems;
– the class of systems can be found such that the state vector is appropriately
estimated in the presence of unknown disturbances.
6. Discrete-time systems
Direct implementation of discontinuous control by discrete-time controller leads
to chattering in some vicinity of the discontinuity manifold. The main attention
was paid to analysis of this motion in majority publications in this area within last
three decades. The recently introduced new concept “discrete-time sliding mode
control” establishes the bridge between discrete- and continuous-time systems,
since both of them have the same property: the state reaches a manifold in the
state space and then the trajectories do not leave this manifold.
7. SMC in infinite-dimensional systems
The chapter explains the reasons why the mathematical methods related to slid-
ing mode equations for ODE should be newly developed. The new methods of
deriving sliding mode equations and reaching conditions, based on Lyapunov’s
functional, for equations in Hilbert space are demonstrated. Practically, they
embrace all particular cases.
8. Chattering analysis
Chattering is the main obstacle for implementation of sliding mode control. It
explains high intensity of research and publications on development of chatter-
ing suppression methods. Again each of them belongs to one of the following
directions:
– Systems with observers preserving a high-frequency component in control.
– State-dependent amplitude of discontinuity in control.
– Filtering out high frequencies based on the harmonic cancellation principle.
9. Adaptive SMC
The objective of designing adaptive SMC is to reduce chattering following the
second direction of the previous section. One of the two approaches can be found
in each publication on adaptive SMC:
12 Conclusions 127

– Amplitude of control depends on the distance to discontinuity manifold.

– Amplitude of control depends on the equivalent control found easily by a
low-pass filter.
The transparent explanations for both design ideas are given.
10. Open Problems in SMC theory
Interesting research directions are listed only, although not for all of them, even
a solution idea can be outlined.