Nonlinear Control

Lecture Notes
Summer Semester 2025

April 24, 2025

© Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
Preface

The present manuscript is a typeset and edited version of lecture notes for the course “Nonlinear
Control” which I started giving in winter semester 2012/13 within the Master of Science in
Power Engineering (MSPE) programme at TUM. In the meantime, the audience has widened,
and “Nonlinear Control” appears in many study programmes.
The intention of the course is to give you insight into some of the basic concepts and methods
of modern nonlinear control theory. The central problem under consideration is stabilization
of equilibria of nonlinear dynamical systems. The main argument in many solutions is the
use of energy-type so-called Lyapunov functions to prove stability (Chapter 2) or to construct
stabilizing nonlinear controllers like in the Backstepping approach (Chapter 6). A different,
quite appealing idea is to make a given nonlinear system, or at least a part of it, linear by state
feedback (Chapter 4). After a stable feedback linearization, simple linear control theory can
be applied to the transformed system.
Not surprisingly, the application of such a promising approach is limited by more or less re-
strictive requirements. To understand the case of input-output linearization, the concepts
of relative degree and zero dynamics are necessary. To formulate conditions for input-state
linearization, some ideas from differential geometry need to be introduced. These concepts
give interesting insight into the dynamic behavior of nonlinear dynamical systems, and some
of them are introduced in Chapter 3. These concepts are also crucial to test controllability or
related properties of nonlinear systems, which is discussed in Chapter 4.
In Chapter 5, the fact is taken into account that stabilization is not the only relevant control
task. To achieve exact trajectory tracking of nonlinear systems within a two-degrees-of-freedom
controller structure, the flatness-based methods are powerful and widely applied tools.
Chapter 7 deals with the passivity formalism which has attracted a lot of attention, particularly
for system representations that arise from modeling of interconnected multi-physics systems
(see my course “Energy-Based Modeling of Complex Systems” for an extensive introduction
into the “port-Hamiltonian” modeling paradigm). After a discussion of the properties of
dissipative and passive systems and the statement of the Invariance Principle as an important
tool for stability analysis, passivity based state feedback control is introduced, with IDA-PBC
a systematic control design procedure exploiting the port-Hamiltonian formulation.
I would like to express my gratitude to my 2013/14 students Aastha Kanwar and Antonio
Blancato who did an excellent job in typesetting my hand written notes and producing the
computer drawings of the illustrations.
The lecture and the exercise will take place in presence at campus Garching. I will upload the
recordings on Moodle/Opencast after the course. Please make use of the optional offer of
revision exercises in presence or via Zoom (in a bi-weekly rhythm), which are an excellent
opportunity to clarify very early any doubts on lecture and exercises.

3
On the Moodle course you find all course documents: these lecture notes, the exercises, all
solutions (in the exercise course, I can only solve a part of them – the rest should be solved by
yourself as an excellent practice for the exam), the teaching recordings, additional material as
well as up-to-date information. Moreover, please use the discussion forum to connect with
your fellow students. You can exchange on the course topics there: post your questions and
doubts and respond to those of your fellow students if you know the answer or have a good
explanation by your own understanding.
I hope that you will enjoy our journey through nonlinear control systems and encourage you
to actively participate in our lessons. The last courses showed that active discussions in class
are extremely fruitful for a thorough understanding!

Garching, in April 2025

Paul Kotyczka

4 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
Contents

1 Introduction 9
1.1 Typical Controller Structures and Control Tasks . . . . . . . . . . . . . . . . . 9
1.2 Behavior of Linear and Nonlinear Systems . . . . . . . . . . . . . . . . . . . . 10
1.3 Motivation for Nonlinear Control . . . . . . . . . . . . . . . . . . . . . . . . . 13

2 Lyapunov Stability 19
2.1 The Linear Case: Lyapunov Equations . . . . . . . . . . . . . . . . . . . . . . 20
2.2 The Two Methods of Lyapunov for Nonlinear Systems . . . . . . . . . . . . . . 22
2.3 Control Lyapunov Functions and Sontag’s Formula . . . . . . . . . . . . . . . 25
2.4 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3 Some Concepts from Differential Geometry 29

3.1 Vector Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 Lie Derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3 Duality and Cotangent Space . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.3.1 Covectors or One-Forms . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.3.2 Bases of the Tangent and Cotangent Space . . . . . . . . . . . . . . . 33
3.4 Lie Bracket . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.5 Distributions and Involutivity . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.5.1 Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.5.2 Involutive Closure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.5.3 Annihilator of a Distribution . . . . . . . . . . . . . . . . . . . . . . . 38
3.6 Integrability and Frobenius’ Theorem . . . . . . . . . . . . . . . . . . . . . . . 39
3.6.1 Interpretation of Complete Integrability . . . . . . . . . . . . . . . . . 39
3.6.2 Straightening Out of Vector Fields . . . . . . . . . . . . . . . . . . . . 41
3.7 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

5
Contents

4 Controllability and Feedback Linearization 45

4.1 Controllability Definition and Driftless Systems . . . . . . . . . . . . . . . . . 45
4.2 Example: Satellite Attitude Control . . . . . . . . . . . . . . . . . . . . . . . . 46
4.2.1 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.2.2 Controllability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3 Accessibility of Nonlinear Systems . . . . . . . . . . . . . . . . . . . . . . . . 51
4.4 Input-State Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.5 Input-Output Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.5.1 Illustrative Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.5.2 General Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.6 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

5 Flatness Based Control 65

5.1 Two Degrees of Freedom Structure . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2 Differential Flatness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.3 Flatness Based Feedforward Control . . . . . . . . . . . . . . . . . . . . . . . 67
5.4 Flatness Based Feedback Control . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.4.1 Linear State Feedback Based on Linearization . . . . . . . . . . . . . . 68
5.4.2 Nonlinear State Feedback for SISO Systems in Normal Form . . . . . . 68
5.5 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

6 Backstepping 71
6.1 Derivation of the Backstepping Controller . . . . . . . . . . . . . . . . . . . . 71
6.2 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.3 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

7 The Passivity Approach 77

7.1 Dissipativity and Passivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
7.2 Invariance Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
7.3 Example: Passivity of Mechanical Multi-Body Systems . . . . . . . . . . . . . 81
7.4 Interconnections of Passive Systems . . . . . . . . . . . . . . . . . . . . . . . 82
7.4.1 Power-Preserving Interconnections . . . . . . . . . . . . . . . . . . . . 83
7.4.2 Output Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
7.5 Passivity-Based State Feedback Control . . . . . . . . . . . . . . . . . . . . . 85
7.5.1 Port-Hamiltonian Systems . . . . . . . . . . . . . . . . . . . . . . . . 85

6 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CONTENTS

7.5.2 Interconnection and Damping Assignment Passivity Based Control, IDA-

PBC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
7.6 Passivity of State Space Models∗ . . . . . . . . . . . . . . . . . . . . . . . . . 90
7.7 Feedback Passivity∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
7.8 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

A Handouts 97

© Prof. Dr.-Ing. habil. Paul Kotyczka 7

Chair of Automatic Control
Technical University of Munich
Contents

8 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
Chapter 1

Introduction

1.1 Typical Controller Structures and Control Tasks

This first section reviews in a very compact way the signals occuring in a control system and
the frequently used Two-Degrees-of-Freedom (2 DOF) controller structure, as well as typical
control goals (see also “Advanced Control”).

Figure 1.1: System and controller with all relevant signals

Figure 1.2: Frequently used Two-Degrees-of-Freedom (2 DOF) controller structure

The feedforward part does the control job in an ideal/perfect setting for stable systems without
disturbances and uncertainties. Feedforward control is based on the (dynamical) model of the

9
1.2. Behavior of Linear and Nonlinear Systems

system, which includes plant and actuators. The better the underlying model, the more precise
will the computed nominal control signal uF F be, and the less action must be taken by the
feedback control part.
The feedback is necessary to stabilize an unstable system and to compensate for errors that
originate in imperfect modeling and certain external disturbances. Nonlinear state feedback
control design, which is dealt with in this course, can improve the closed-loop performance,
e.g., it can enlarge the domain of attraction (see Definition 1.1) and robustify the control
system.

Typical Control Goals

1. Stability of an equilibrium state: lim x(t) = x∗ .

t→∞

2. Accuracy w.r.t. the reference: In the best case (exact tracking), lim y(t) → w(t) for
t→∞
arbitrary signals. Often only lim y(t) → w∗ will be achieved (steady state gain 1).
t→∞

3. Rejection/weakening of the effect of disturbance signals.

4. Dynamics assignment, satisfaction of optimality criteria.
5. Robustness of the above properties with respect to uncertainties.

In the course Nonlinear Control, the focus is on point 1: you will learn about stability analysis
for nonlinear systems and the design of state feedback controllers for nonlinear systems. (state
feedback methods for stabilization). Point 2 will be addressed in the section on flatness-based
control.

1.2 Behavior of Linear and Nonlinear Systems

On a few examples, we illustrate the following aspect, in which linear and nonlinear time-
invariant systems1 differ. In this study, we consider only the homogeneous case without input.

• A linear time-invariant system

ẋ(t) = Ax(t), x(0) = x0 (1.1)

has exactly one equilibrium if A has full rank. If A is rank deficient, there may be a
continuum of equilibria of the same type. The trajectories of an undamped linear oscil-
lator (where A has purely imaginary eigenvalues) are periodic. These periodic solutions,
however, do not attract other solutions. See Example 1.1.
• A nonlinear system
ẋ(t) = f (x(t)), x(0) = x0 (1.2)
can have a finite number of equilibria of different type, see Example 1.2. Moreover, in
nonlinear systems not only stable equilibria can attract solutions, but also (attractive)
limit cycles. The van der Pol oscillator in Example 1.3 is an example for such a nonlinear
system.
1
We do not consider time-varying systems in this course

10 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 1. INTRODUCTION

Example 1.1 (Equilibria of linear systems). The purpose of this example is to recall the different
possible behaviors of a linear time-invariant system near its equilibrium. Depending on the
location of the eigenvalues, their multiplicity and the direction of the (generalized) eigenvectors,
the trajectories of the system show different characteristic patterns. A phase portrait shows a
set of solutions (for different initial values) of a 2D LTI system

ẋ(t) = Ax(t), x(0) = x0 , x ∈ R2 (1.3)

in the (x1 , x2 )-plane. The eigenvalues and (generalized) eigenvectors for the depicted cases
are shown in Table 1.1. Figure 1.3 shows in gray the vector fields, in red the eigenvectors (only
in the case of the singular node v 2 is a generalized eigenvector) and in blue the trajectories.
Depicted are the stable versions of the considered equilibria. Their unstable counterparts can
be obtained by changing the sign of the eigenvalue real parts.
The knowledge of linear systems’ behavior is still very useful for the local analysis of nonlinear
systems near their equilibria. ◁

Table 1.1: Eigenvalues and (generalized) eigenvectors of the systems in Example 1.1.
λ1 λ2  v1   v2 
1 −1
Stable node −1 −2
0.25  1
1 −1
Stable dicritical node −1 −1
0.25  1 
1 −0.4
Stable singular node −1 −1
0.25 0.4
(v 2 generalized eigenvector)    
1 −1
Saddle point −1 2
0.25 1
1 −1
Continuum of stable nodes −1 0
 0.25   1 
−1 + j −1 − j
Center −j j
−0.25 − j  −0.25 + j 
−1 + j −1 − j
Stable Focus −0.1 − j −0.1 + j
−0.25 − j −0.25 + j

© Prof. Dr.-Ing. habil. Paul Kotyczka 11

Chair of Automatic Control
Technical University of Munich
1.2. Behavior of Linear and Nonlinear Systems

Figure 1.3: Phase portraits for different types of equilibria according to Table 1.1.

12 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 1. INTRODUCTION

ẋ

x
-1 1

-1

Figure 1.4: Equilibria and trajectories of ẋ = x3 − x in the (x, ẋ)-plane.

Example 1.2 (Multiple equilibria). The nonlinear system

ẋ = x3 − x, x(0) = x0 (1.4)

has the three equilibria x∗1 = 0 (stable) and x∗2/3 = ±1 (unstable). The three equilibria, as
well as the directions of the flow (the solution of the initial value problem) in the (x, ẋ)-plane
are depicted in Fig. 1.4. ◁

Example 1.3 (Van der Pol oscillator). The van der Pol oscillator
" # " #
ẋ1 (t) x2 (t)
= (1.5)
ẋ2 (t) −x1 (t) − ε(x21 (t) − 1)x2 (t)

is a nice example for a nonlinear system, which features an (unstable) equilibrium and an (at-
tractive) limit cycle. On the state differential equations (1.5) you recognize a linear oscillator,
which is perturbed by a nonlinear term, weighted with the factor ε. A van der Pol oscillator
can be realized as an LC circuit in parallel with the series connection of a voltage source and
a tunnel diode. Figure 1.5 shows three phase portraits for the values 0.1, 0.5 and 2 of ε. The
gray lines are isoclines, i.e., lines on which the solutions of (1.5) have the same slope (−5,
−1, 0 and 1). The three phase portraits nicely illustrate the transition from an almost linear
system behavior around the equilibrium to apparently nonlinear trajectories, which rapidly tend
to the limit cycle, with a periodic change of slower and faster phases. ◁

1.3 Motivation for Nonlinear Control

Linear state space methods are often the first choice in controller design. They are:

• Model-based, i.e., they respect/make use of the (approximated) physical model (physical
modelling → linearization).

© Prof. Dr.-Ing. habil. Paul Kotyczka 13

Chair of Automatic Control
Technical University of Munich
1.3. Motivation for Nonlinear Control

Figure 1.5: Phase portraits of the van der Pol oscillator with isoclines for −5, −1, 0 and 1.

ẋ ẋ ẋ

1 1 1

x x x
-1 1 -1 1 -1 1

-1 -1 -1

Figure 1.6: The non-smooth system in Example 1.4. Resulting dynamics with linear and non-
smooth controller.

• Conceptually simple. The inspection of system properties (controllability, stabilzability,

observability, etc.) and controller design (feedforward → feedback) is done using linear
algebra.

• Widely applicable in today’s technological systems due to the availability of different

types of sensors and computational devices.

• Highly performant and robust due to the use of measurements/estimation of the complete
system state – in contrast to simple PID output feedback.

However, the progress made since the 1980s in the theory of nonlinear control allows to consider
the detailed nonlinear description of the system’s dynamics which opens the door to derive
(state feedback) controllers that outperform the linear controllers. In some cases, equilibria of
dynamical systems can even not be stabilized by smooth state feedback2 .
A very simple example, which shows that a smooth (in this case linear) controller can be
inferior to a switching (i.e., non-smooth, nonlinear) controller is as follows.

Example 1.4. Global asymptotic stabilization of the equilibrium x∗ = 0 of the non-smooth

system
ẋ = x · sgn(x) + u (1.6)
2
See the famous Brockett’s condition, according to which for example the stabilization of a wheeled inverted
pendulum (a non-holonomic mechanical system) at an arbitrary point in configuration space is not possible by
smooth feedback. Switching between control laws is required.

14 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 1. INTRODUCTION

Figure 1.7: Illustration of the domain of attraction, the dashed line shows the boundary ∂D
of the domain of attraction of x∗ .

with a linear controller u = −kx works for k > 1. However, the closed-loop system has
still a non-smooth right hand side, with faster dynamics for x < 0 and slower dynamics for
x > 0. The middle plot in Fig. 1.6 shows illustrates the resulting dynamics under the feedback
u = −2x.
A nonlinear control law
u = −x · sgn(x) − kx, k>0 (1.7)
yields the linear (and therefore smooth) closed-loop dynamics ẋ = −kx.
Note that the latter cancellation of the nonlinear term (which is at the core of the feedback
linearization methods) can sometimes be problematic, if the cancellation is not exact, e.g. due
to model inaccuracy or measurement or other errors in the control loop.
In this course, we will only consider smooth, i.e., sufficiently often differentiable nonlinearities.
◁

Appropriately tuned (!) nonlinear control can be necessary for robust stabilization. It can also
improve the performance with respect to various criteria. A very important criterion is the
domain of attraction of the closed-loop equilibrium.

Definition 1.1 (Domain of Attraction). Given an autonomous dynamical system

ẋ = f (x), x(0) = x0 , (1.8)

x ∈ X ⊆ Rn , f : X → Rn , with equilibrium x∗ , i.e., f (x∗ ) = 0. The domain

of attraction D of the equilibrium x∗ is the set of all initial values x0 which tend to
equilibrium x∗ under the flow of f (x):

D := {x0 | lim x(t) = x∗ for x(0) = x0 } (1.9)

t→∞

Definition 1.2 (Flow of a vector field). We call the solution of the initial value problem
(1.8) also flow of the vector field f (starting at x0 ) and write

x(t) = Φft (x0 ) (1.10)

The following very simple example shows how the domain of attraction can be enlarged by
nonlinear control.

© Prof. Dr.-Ing. habil. Paul Kotyczka 15

Chair of Automatic Control
Technical University of Munich
1.3. Motivation for Nonlinear Control

Example 1.5. Consider the simple first order system

ẋ = x3 . (1.11)

At its equilibrium, 0 = (x∗ )3 holds ⇒ x∗ = 0. Stability of this equilibrium can easily be tested
by considering initial values x(0) ̸= 0 and observing the time derivative:

x(0) > 0 ⇒ ẋ(0) > 0,

x(0) < 0 ⇒ ẋ(0) < 0.

Upon an initial perturbation from the equilibrium, the state will move further avay from x∗ = 0.
Hence, x∗ is an unstable equilibrium.
Now assume, that a control input u is at our disposition to stabilize the equilibrium:

ẋ = x3 + u. (1.12)

1. Linear (state) feedback

Linearization of the nonlinear system at the equilibrium (x∗ , u∗ ) = (0, 0):

∂x3
∆ẋ = ∆x + ∆u
∂x x∗=0

The ∆ can be dropped because of (x∗ , u∗ ) = (0, 0). The control law

u = −x

yields the closed-loop linearized dynamics

ẋ = −x

with a stable eigenvalue at −1. Substituting the control law in the original nonlinear equation
yields
ẋ = x3 − x = − x · (1 − x2 ) .
|{z} | {z }
≶0 for x≷0 ≷0 for |x|≶0

From the definiteness of the terms on the right hand side, we can conclude that (see Fig. 1.8)

ẋ(0) < 0 for 0 < x(0) < 1 and x(0) < −1,
ẋ(0) > 0 for − 1 < x(0) < 0 and x(0) > 1.

Hence, x∗ = 0 is an asymptotically stable equilibrium and

D = {x0 | |x0 | < 1}

is its domain of attraction. The domain of attraction is bounded by the two unstable equilibria
at ±1.

16 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 1. INTRODUCTION

ẋ ẋ ẋ

1 1 1

x x x
-1 1 -1 1 -1 1

-1 -1 -1

Figure 1.8: Illustration of ẋ = x3 , and the closed-loop dynamics under the linear and the
nonlinear controller (gray: domain of attraction).

2. Feedback linearization

If we inspect
ẋ = x3 + u
and choose the control law
u = −x3 − x
to cancel the nonlinear (cubic) term and impose linear closed-loop behavior

ẋ = −x,

then x∗ = 0 is asymptotically stable (ẋ(0) ≶ 0, x(0) ≷ 0) and the domain of attraction is

D = R.
As an alternative, we can also choose asymptotically stable nonlinear behavior by an “over-
compensation” of x3 :
u = −2x3 ⇒ ẋ = −x3 .
◁

The observations from the simple example can be generalized: While for asymptotically stable
linear systems ẋ = Ax the domain of attraction3 of x∗ = 0 is the whole Rn , this is, in general,
not true for nonlinear systems. Moreover, nonlinear systems ẋ = f (x) – unlike linear systems
– can have multiple equilibria {x∗ | f (x∗ ) = 0} with different stability properties.
Stability of equilibria, as well as the extent of the domain of attraction in the nonlinear setting
are investigated with the help of so called Lyapunov functions, as presented in the following
chapter.

3
Note that we do not take into account saturation effects that limit the domain of attraction also for linear
systems.

© Prof. Dr.-Ing. habil. Paul Kotyczka 17

Chair of Automatic Control
Technical University of Munich
1.3. Motivation for Nonlinear Control

18 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
Chapter 2

Lyapunov Stability

Observation: Mechanical systems (and also systems from other domains) tend for t → ∞
to a state of minimum energy. This state – if it exists in the considered scale – is the
(asymptotically) stable equilibrium of the system. The energy (potential + kinetic energy)
decreases monotonically until the stable equilibrium is reached.
The key idea of Lyapunov stability analysis is to find a generalized (type of) energy function
which is non-increasing along the solutions of the state differential equations. This idea goes
back to the Russian mathematician A.M. Lyapunov (“The General Problem of Motion Stabil-
ity”, 1892 (!)). His work is the basis of most “modern” analysis and synthesis methods for
nonlinear dynamical systems, developed since the 1960s.
First recall the stability definitions from “Advanced Control”:

Definition 2.1 (Stability of an equilibrium). Given a dynamical system ẋ = f (x), x ∈

X ⊂ Rn with equilibrium x∗ , i.e. f (x∗ ) = 0.
The equilibrium is said to be:

• stable if for every ε > 0, there exists a δ(ε) > 0 such that

∥x(0) − x∗ ∥ < δ(ε) ⇒ ∥x(t) − x∗ ∥ < ε (2.1)

for all t > 0 (i.e. the solutions remain bounded).

• asymptotically stable if in addition δ can be chosen such that

lim x(t) = x∗ (2.2)

t→∞

(i.e. the equilibrium is attractive).

• unstable if not for every given ε > 0 there exists a δ = δ(ε) such that the solutions
remain bounded to the ε-neighborhood when ∥x(0) − x∗ ∥ < δ(ε).

Before we move on to the Lyapunov functions in the context of general nonlinear systems, we
consider the important special case of linear systems. For linear systems, finding a Lyapunov
function means solving a so-called Lyapunov equation – once again with the help of linear
algebra.

19
2.1. The Linear Case: Lyapunov Equations

Figure 2.1: Illustration of stable, asymptotically stable and unstable equilibrium points

2.1 The Linear Case: Lyapunov Equations

Given a linear time-invariant, autonomous dynamical system

ẋ(t) = Ax(t), x(0) = x0 , (2.3)
x ∈ X ⊆ Rn , A ∈ Rn×n . Of course, we can analyze the stability of an equilibrium x∗ of
this linear system using the well-known approach from linear systems’ theory based on the
properties of A. Nevertheless, we illustrate Lyapunov’s direct method for this particular case
first. We will reveal the connection to linear stability theory, which gives a solvability statement
for the associated Lyapunov equation.

Definition 2.2 (Positive (semi-)definiteness1 ). A symmetric matrix P = P T ∈ Rn×n is

called positive (semi-)definite if for all x ∈ Rn , x ̸= 0 the quadratic form xT P x is
positive/non-negative: xT P x > 0 (≥ 0). We write

P >0 ⇔ xT P x > 0 ∀x ̸= 0,
T
(2.4)
P ≥0 ⇔ x Px ≥ 0 ∀x ̸= 0.

Definiteness tests: A symmetric matrix P = P ⊤ is positive definite if and only if

1. all its eigenvalues (which are real-valued by the symmetry of P ⊤ ) are positive, or,
equivalently,
2. all its leading principal minors det ∆1 , . . . , det ∆n are positive, where ∆i = [P ]1...i,1...i
are the “northwestern” submatrices of P (Sylvester’s criterion).

Important: The eigenvalue criterion 1. can be applied accordingly for semi-definiteness (the
eigenvalues must be non-negative). To check semi-definiteness with Sylvester’s criterion, how-
ever, 2. is not sufficient; in this case, all principle minors must be non-negative, not only the
leading ones.
Suppose there exists a positive definite function
1
V = xT P x, P = PT > 0 (2.5)
2
1
We will later use a more generalized definition of positive (semi-)definiteness and call a function V (x)
positive (semi-)definite if it has a minimum V ∗ at x∗ and V (x) > V ∗ (≥ V ∗ ) everywhere else. This includes
the case of quadratic forms in shifted coordinates (x − x∗ )⊤ P (x − x∗ ).

20 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 2. LYAPUNOV STABILITY

which continuously decreases along the solutions of Eq. (2.3), i.e. (we differentiate V w.r.t.
x and xT separately, and apply the chain rule),
∂V ∂V
V̇ = ẋ + ẋT
∂x ∂xT
1 T 1
= x P Ax + xT AT P x
2 2
1 T
= x (P A + AT P )x
2
1
= − xT Qx < 0, Q = QT > 0 (2.6)
2
for a certain, given Q > 0. Then clearly

lim V (x(t)) = min V (x) = 0 (2.7)

t→∞ x

We integrate Eq. (2.6) in time and let t → ∞:

Z ∞
lim V (x(t)) = V (x(0)) + V̇ (x(t)) dt =
t→∞ 0
1 ∞ T
Z
= V (x(0)) − x (t)Qx(t) dt. (2.8)
2 0

We know that the solution of Eq. (2.3) is (verify this)

x(t) = eAt x(0). (2.9)

Replacing this and Eq. (2.7) in Eq. (2.8) yields

Z ∞
1 T
0 = V (x(0)) − xT (0)eA t Q eAt x(0) dt. (2.10)
2 0 | {z }
x(t)

Due to positive definiteness of Q = QT > 0 the integrand is > 0 for x(t) ̸= 0. The equation
is true if and only if the second term takes the finite value V (x(0)), ∞ > V (x(0)) > 0. This
in turn is possible if and only if the exponential functions under the integral decay to zero:

A positive definite function (2.5) which decreases continuously according to (2.6) under
the flow of (2.3) exists if and only if

eAt → 0 for t→∞ ⇔ σ(A) ∈ C− .

Notation: σ(A) denotes the spectrum, i.e., the set of all eigenvalues of the matrix A. C−
represents the open left complex half plane.
Recall (2.6): The matrices P and Q must be related according to

AT P + P A = −Q, Q = QT > 0. (2.11)

With the above argumentation, it becomes clear that this so-called Lyapunov equation has,
for a given positive definite matrix Q = QT > 0, a positive definite solution P = P T > 0 if
and only if σ(A) ∈ C− .

© Prof. Dr.-Ing. habil. Paul Kotyczka 21

Chair of Automatic Control
Technical University of Munich
2.2. The Two Methods of Lyapunov for Nonlinear Systems

Theorem 2.1 (Lyapunov equations and stability). Given the autonomous linear system
(2.3).
(i) A (positive definite) Lyapunov function
1
V = xT P x, P = PT > 0 (2.12)
2
which strictly decreases under the dynamics (2.3) according to
1
V̇ = − xT Qx, Q = QT > 0 (2.13)
2
exists if and only if the Lyapunov equation

AT P + P A = −Q (2.14)

has a solution P = P T > 0 for given Q = QT > 0.

(ii) The Lyapunov equation (2.14) has a solution P = P T > 0 for given Q > 0 if and
only if the matrix A is Hurwitz i.e.

σ(A) ∈ C− . (2.15)

Point (ii) allows the following reasoning, which is identically used in nonlinear stability analysis:
Suppose one can find a quadratic function V = 21 xT P x, P > 0, which is positive definite
with its strict minimum in x∗ = 0 (x∗ = 0 without loss of generality) and which strictly
decreases along the solutions of the system of ODEs, V̇ = − 12 xT Qx, Q > 0, i.e., a Lyapunov
function, then the equilibrium x∗ is asymptotically stable.
Note however, that the theorem states “A Lyapunov function [. . . ] exists if and only if [. . . ]”. If
a specific Lyapunov candidate function V = 12 xT P x does not satisfy the Lyapunov equation,
i.e. no Q = QT > 0 can be found to satisfy (2.11) for the given P , then this does not mean
that the considered system is necessarily unstable. It rather means that this choice of V (x)
is not suitable to show stability of the equilibrium. Another function V̂ (x) = 12 xT P̂ x may do
the job. If indeed no such function can be found (which is in general hard to verify), then x∗
is unstable.

2.2 The Two Methods of Lyapunov for Nonlinear Systems

The two methods of Lyapunov are of central importance in nonlinear stability analysis and
controller design.
Let us first formulate Lyapunov’s indirect (first) method which we have already applied implic-
itly when checking the stability of an equilibrium of a linearized system in Advanced Control.

22 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 2. LYAPUNOV STABILITY

Theorem 2.2 (First/indirect method). Let x∗ be an equilibrium of

ẋ = f (x), x ∈ X ⊆ Rn

and f : X → Rn twice continuously differentiable in an environment of x∗ . The lineariza-

tion
∂f (x)
∆ẋ = A∆x, A = (2.16)
∂x x∗
with ∆x = x − x∗ , is a local approximation of the dynamics of the nonlinear system.
σ(A) denotes the spectrum of the matrix A (the entirety of all eigenvalues) and λi (A)
denotes a single eigenvalue.
The equilibrium x∗ is
• asymptotically stable if σ(A) ∈ C− ,

• unstable if at least one eigenvalue λi (A) ∈ C+ .

• If at least one eigenvalue lies on the imaginary axis (while the others are in C− ), the
equilibrium is called “non-hyperbolic” and no direct stability statement is possible.

For the non-hyperbolic case, consider the system with cubic nonlinearity as in Example 1.1.
For both ẋ = x3 and ẋ = −x3 , the linearization in x∗ = 0 yields ∆ẋ = 0 · ∆x, i.e., an
eigenvalue in zero. However, in the former case, the equilibrium is unstable, in the latter it is
globally asymptotically stable.
Lyapunov’s linearization method provides a purely local stability statement. The direct or
second method, which generalizes Theorem 2.1 to nonlinear systems, allows for an estimate of
the domain of attraction of the equilibrium by the shape of the obtained Lyapunov function:

Theorem 2.3 (Second/direct method). Let x∗ be an equilibrium of

ẋ = f (x), x ∈ X ⊆ Rn .

Let V : D → R, D ⊆ X , be a continuously differentiable function with V (x∗ ) = V ∗ such

that

(i) V (x) > V ∗ on D\{x∗ }, (2.17)

∂V (x)
(ii) V̇ (x) = f (x) ≤ 0 on D. (2.18)
∂x
Then V (x) is called a Lyapunov function and the equilibrium x∗ is stable (in the sense
of Lyapunov).
If, in addition,
V̇ (x) < 0 on D\{x∗ }, (2.19)
then x∗ is asymptotically stable.

Proof : see Handout 1.

Remarks

• Lyapunov functions are not unique.

© Prof. Dr.-Ing. habil. Paul Kotyczka 23

Chair of Automatic Control
Technical University of Munich
2.2. The Two Methods of Lyapunov for Nonlinear Systems

• A single Lyapunov function candidate can prove stability but not instability.

• Equilibria may be asymptotically stable despite V̇ ≤ 0 (e.g. pendulum with friction). ⇒

Asymptotic stability can be proven using LaSalle’s Invariance Principle. We will discuss
this in the chapter on passive systems.

Domain of attraction A Lyapunov function V (x) provides an estimate of the domain of

attraction of x∗ . Let
XV = {x ∈ X | V (x) > V ∗ } ∪ x∗ (2.20)
and
XV̇ = {x ∈ X | V̇ (x) < 0} ∪ x∗ . (2.21)
We consider the level sets of V (x)

Ωc := {x ∈ X | V ∗ ≤ V (x) ≤ c}. (2.22)

Let c̄ be the value of V (x) on the largest closed and bounded (i.e., compact) level set Ωc̄ ,
which is completely contained in XV ∩ XV̇ . Then all trajectories that start in Ωc̄ remain in Ωc̄
and approach x∗ asymptotically for t → ∞. Ωc̄ is an estimate of the domain of attraction of
x∗ .
Explanation:

• x∗ is an isolated equilibrium, there is no other point s.t. ẋ = 0 in Ωc̄ .

• In Ωc̄ , except for x∗ , V̇ (x) = ∂V∂x(x) ẋ < 0 holds, and hence ∂V (x)
∂x ̸= 0T . ⇒ There are
no further stationary points of V (x).

• ⇒ Every point in Ωc̄ is “transient” (ẋ ̸= 0) and V (x) decreases strictly: V̇ (x) < 0.

• ⇒ As V (x) is bounded from below by V ∗ = V (x∗ ), every trajectory which starts in Ωc̄
must end asymptotically in x∗ .

Global asymptotic stability If V̇ < 0 holds on the complete Rn and if V (x) has a particular
shape, it can prove global asymptotic stability:

Corollary 2.1. Let V (x) be a positive definite function, which is strictly increasing with
the distance ∥x − x∗ ∥ from x∗ (such a function is radially unbounded, i.e. V (x) → ∞
for ∥x − x∗ ∥ → ∞). Let in addition V̇ (x) < 0 on Rn \{x∗ } . Then the equilibrium x∗ is
globally asymptotically stable and the domain of attraction is Rn .

We can summarize the characteristics of the two Lyapunov methods as follows.

First (indirect) method Second (direct) method

Easily applicable, based on Generalization of Theorem 2.1 to
linearization and eigenvalues. nonlinear systems.
Only local information about stability of the Provides information about the
equilibrium. domain of attraction.

24 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 2. LYAPUNOV STABILITY

Figure 2.2: Illustration of the estimate of the domain of attraction provided by a Lyapunov
function.

Remark 2.1. The quadratic Lyapunov function V (∆x) = 12 ∆xT P ∆x, where P = P T > 0 is
the solution of a Lyapunov equation (2.14) and the state matrix A represents the linearization
according to (2.16), serves also as a Lyapunov function for the nonlinear system ẋ = f (x) – in
a region around the equilibrium x∗ , where the linearization is a sufficiently good approximation
of the nonlinear system. This can be seen as follows. We consider w.l.o.g. x∗ = 0 and write
the differential equation as
ẋ = f (x∗ ) +Ax + r(x),
| {z }
=0

where r(x) represents a residual term of order O(∥x∥2 ). Express now the time derivative of
the quadratic Lyapunov function:
1 1
V̇ = xT P (Ax + r(x)) + (xT AT + r T (x))P x
2 2
1 T
= x (P A + A P )x + xT P r(x) .
T

|2 {z } | {z }
perturbation
=− 12 xT Qx, Q>0

As long as the quadratic first term dominates the second non-quadratic perturbation term,
V (x) is a Lyapunov function for the nonlinear system.

2.3 Control Lyapunov Functions and Sontag’s Formula

Lyapunov stability analysis is the basis of Lyapunov-based controller design techniques for
nonlinear systems. Most of the time we consider so-called input-affine systems, where the
vector-valued input function u(t) acts linearly on the system dynamics,

ẋ = f (x) + G(x)u, x ∈ X ⊂ Rn , u ∈ U ⊂ Rp . (2.23)

© Prof. Dr.-Ing. habil. Paul Kotyczka 25

Chair of Automatic Control
Technical University of Munich
2.3. Control Lyapunov Functions and Sontag’s Formula

The goal is to find a state feedback

u = β(x), β:X →U (2.24)

such that the closed-loop system

ẋ = f (x) + G(x)β(x) := f cl (x) (2.25)

has the desired equilibrium x∗ , i.e., f cl (x∗ ) = 0, and x∗ is asymptotically stable. To prove
stability and to give an estimate of the domain of attraction, we want to determine a closed-loop
Lyapunov function V (x) such that for all x ̸= x∗ in some neighborhood of the equilibrium

∂V (x)
V (x) > V (x∗ ) and V̇ (x) = f cl (x) < 0. (2.26)
∂x

Control Lyapunov Functions While Lyapunov functions where defined for autonomous sys-
tems, the concept of Control Lyapunov Functions (CLFs) for input-affine systems takes into
account rendering the time derivative V̇ (x) negative by means of the control input. For
the discussion of CLFs, we assume that the input-affine system (2.23) has an equilibrium
(x∗ , u∗ ) = (0, 0), which implies f (0) = 0. Note that this can always be achieved by a simple
coordinate shift (x, u) 7→ (x − x∗ , u − u∗ ).

Definition 2.3 (Control Lyapunov Function, CLF). A scalar function V : X → R is a

Control Lyapunov Function (CLF) for the system (2.23) with equilibrium (x∗ , u∗ ) = (0, 0)
if the following implication holds for x ̸= 0:
∂V ∂V
G(x) = 0 ⇒ f (x) < 0. (2.27)
∂x ∂x
In words: Whenever the decrease of V (x) cannot be induced by the control input (through
the vector field Gu), V (x) must decrease due to the drift term f (x).

Sontag’s Formula Once a CLF V (x) according to the previous definition is found, an asymp-
totically stabilizing feedback control law can be constructed, the so-called Sontag’s formula,
see [1] or [2], Section 3.5.3. Define
⊤
∂V (x) ∂V (x)

a(x) := f (x) and b(x) := G(x) . (2.28)
∂x ∂x

Recall that a(x) < 0 must hold whenever b(x) = 0 for x ̸= 0. Sontag’s formula is then given
by  q

 a(x) + a2 (x) + (b⊤ (x)b(x))2
−b(x) , b(x) ̸= 0,

uS (x) = b⊤ (x)b(x) (2.29)



0, b(x) = 0.
The control law has an interpretation in terms of optimal control, as it minimizes a certain
cost functional, see also [3] for a generalized version.
Before we learn about Lyapunov-based control methods, where (closed-loop) Lyapunov func-
tions and the corresponding control laws are constructed – like Backstepping or passivity-based

26 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 BIBLIOGRAPHY

control with IDA-PBC – we take a closer look at the dynamical behavior of nonlinear systems.
We study the local effects of nonlinear vector fields on the flow they induce. At the end of
the next chapter, the famous Frobenius theorem makes the link between local properties of
the vector fields and the global behavior of the solutions of the nonlinear state differential
equations.

2.4 References

Classical textbooks that cover Lyapunov-based analysis and control design for nonlinear systems
are [4], [5] and [6].

Bibliography

[1] E. D. Sontag. A ‘universal’ construction of artstein’s theorem on nonlinear stabilization.

Systems & Control Letters, 13(2):117–123, 1989.

[2] R. Sepulchre, M. Janković, and P. V. Kokotović. Constructive Nonlinear Control. Springer-

Verlag London, 1997.

[3] M. Sackmann. Modifizierte Optimale Regelung–Stabilitätsorientierter nichtlinearer Re-

glerentwurf (Modified Optimal Control–Stability-Oriented Nonlinear Control). at-
Automatisierungstechnik, 53(8):367–377, 2005.

[4] H. Khalil. Nonlinear Systems. Prentice Hall, 3rd edition, 2002.

[5] J.-J. E. Slotine and W. Li. Applied Nonlinear Control, volume 199. Prentice Hall Englewood
Cliffs, NJ, 1991.

[6] M. Vidyasagar. Nonlinear Systems Analysis, volume 42. SIAM, 2002.

© Prof. Dr.-Ing. habil. Paul Kotyczka 27

Chair of Automatic Control
Technical University of Munich
Bibliography

28 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
Chapter 3

Some Concepts from Differential

Geometry

We present some fundamental concepts from differential geometry, which help to characterize
the evolution of the solutions of nonlinear control systems of the form
m
(3.1)
X
ẋ(t) = f (x(t)) + g i (x(t))ui (t).
i=1

which are defined on a (differentiable) manifold X ⊂ Rn . We simply consider X as a possibly

curved subspace of Rn , without going into detail concerning the precise definition in terms of
differential geometry (which contains coordinate charts, atlases, diffeomorphisms, etc.). We
assume the values ui (t), i = 1, . . . , m of the control inputs to be from an admissible set, which
is expressed by1 u : [0, ∞) → U ⊂ Rm . The flow of (3.1) is induced by the drift vector field
f and the control vector fields g i , i = 1, . . . , m.

3.1 Vector Fields

First of all, we clarify our notion of vector fields, which are the objects on the right hand side
of (3.1). They can clearly be understood as mappings from X to Rn , i.e.,

f : X → Rn , g i : X → Rn , i = 1, . . . , m. (3.2)

The image spaces of these mappings can be, however, endowed with a more geometric meaning
if we think of what these vector fields do: They induce the solutions of the state differential
equation (3.1), depending on the initial value, i.e., the flow x(t) = Φ(x0 , t). If we for the
moment consider the unforced system, i.e., zero input u1 = . . . = um = 0, then the drift
vector field f can be understood in the following sense:

∂Φ(x0 , t + ε)
f (x(t)) = lim , (3.3)
ε→0 ∂ε
see Fig. 3.1 for illustration.
1
For brevity, we will simply write u ∈ U ⊂ Rm , which tacitly contains the fact that we deal with functions
in time.

29
3.1. Vector Fields

Figure 3.1: The (representative of a) vector field is tangent to the flow of the differential
equation.

With this relation to the solution of the differential equation in mind, we can understand the
vector field as a mapping from the space X , in which the solution is defined and evolves, to a
space, which is tangent to this solution.

If, as in the case of the control system (3.1), several vector fields induce (depending on the
choice of the inputs) different possible solutions, we can imagine that for every state x ∈ X the
possible time derivatives ẋ lie in a (hyper-)plane, which is tangent to all the possible solutions
through x. The space X ⊂ Rn can be a curved subspace, a (solution) manifold, and we can
nicely illustrate the tangent space Tx X at X in x, see Fig. 3.2. Whenever we want to stress
this geometric notion of a vector field, we write

f : X → Tx X , g i : X → Tx X , i = 1, . . . , m. (3.4)

Note that there is a tangent space in every point x ∈ X . Each tangent space has the structure
of a linear vector space 2 . The union of the tangent spaces in all points of X is called the
tangent bundle T X = ∪x∈X Tx X .

Figure 3.2: Flow and control vector field of a system ẋ = f (x) + g(x)u as elements of the
tangent space.

2
See Def. 3.2 further below.

30 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 3. SOME CONCEPTS FROM DIFFERENTIAL GEOMETRY

3.2 Lie Derivative

We introduce this directional derivative in an algebraic manner and give it a more geometric
meaning in the following section.

Definition 3.1 (Lie derivative). Given a scalar function h : Rn → R and a vector field
f : Rn → Rn . The Lie derivative or directional derivative of h(x) in direction of the
vector field f (x) is given by

∂h(x)
Lf h(x) = f (x). (3.5)
∂x

Example 3.1. Given the scalar function (which could be a Lyapunov candidate)
1 1
V (x) = x21 + x22
2 2
and the vector field (here constant)
" #
1
f (x) = .
0
The Lie derivative of V in direction of f , which is nothing else than the rate of change of
V (x) along the flow induced by the vector field f , is given by
" #
h i 1
Lf V (x) = x1 x2 = x1 . (3.6)
0
See Fig. 3.3 for an illustration. ◁

Figure 3.3: Level sets of the function V (x) and vector field f (x) in Example 3.1. For x1 = 0,
there is no differential change of V (x) in the direction of f , the vector field f is tangential to
the level sets V (x) = const.

Homework: Sketch the same with the vector field representing the dynamics of a damped
oscillator (with unit parameter values)
" # " #
0 1 x2
f (x) = x= .
−1 −1 −x1 − x2

© Prof. Dr.-Ing. habil. Paul Kotyczka 31

Chair of Automatic Control
Technical University of Munich
3.3. Duality and Cotangent Space

3.3 Duality and Cotangent Space

Recall the definition of a (finite-dimensional) vector space:

Definition 3.2 (Vector space). A linear vector space V over the real numbers R (or an-
other field) is a nonempty set with the operations addition + : V × V → V and scalar
multiplication · : R × V → V . These two operations satisfy a set of properties: (i)
commutativity, (ii) associativity, existence of (iii) a zero element and (iv) an inverse with
respect to addition, (v) associativity and existence of (vi) a unit element with respect to
scalar multiplication, as well as two distributivity properties (vii and viii), which combine
addition and scalar multiplication.

We introduce duality on the example of Eq. (3.6), where we call

" #
h i 1
w(x) = 2x1 2x2 , f (x) = (3.7)
0

dual objects, which is in accordance with the following definition:

Definition 3.3 (Dual vector space). Given a vector space V . Its dual space V ∗ is the space
of all linear functionals φ : V → R on V .

Apparently, if we understand f (x) as column vector, i.e., an element of R2 , and w(x) as a

row vector, which then is an element of the dual space3 (R2 )∗ , the linear functional φ can be
written as a duality pairing, which is here the standard scalar product between row and column
vectors; see (3.9) below.

3.3.1 Covectors or One-Forms

Consider Eq. (3.6). Lf h(x) can be understood as a linear functional on the tangent space
Tx X : A vector f (x) ∈ Tx X is mapped (by means of w(x)) to the real numbers. Equivalently,
this functional can be represented as a duality product or duality pairing

⟨·, ·⟩ : Tx∗ X × Tx X → R (3.8)

as follows:
2
∂h(x)
with (3.9)
X
Lf h(x) = ⟨w(x), f (x)⟩ = wi (x)fi (x) w(x) = .
i=1
∂x

In the duality product, the vector f (x) is paired with its dual object w(x), which we call a
covector or (differential) one-form. While we think of f (x) “living” in the tangent space Tx X ,
we can consider its dual4 object w(x) to live on the cotangent space Tx∗ X .
3
The dual spaces (Rn )∗ are isomorphic to Rn . Roughly speaking, they can roughly be parameterized by the
same (number of) coordinates.
4
A naive illustration of the different character of vectors and covectors stems from writing the duality product
as scalar product, with f (x) a column vector and wT (x) a row vector: ⟨w(x), f (x)⟩ = wT (x)f (x). Note
however that this is only notation and the definition of a covector by duality is much more general and can be
extended to other vector spaces than the Euclidean space.

32 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 3. SOME CONCEPTS FROM DIFFERENTIAL GEOMETRY

In the above example, w(x) is a special one-form, a so-called exact differential. The compo-
nents of w(x) are obtained from differentiation of a scalar function:
h i h i
∂h(x) ∂h(x)
w(x) = w1 (x) w2 (x) = ∂x1 ∂x2
=: dh(x). (3.10)

Equation (3.9), and therefore the definition of the Lie derivative, can be rewritten as the duality
pairing of a vector field with the exact differential of a scalar function:

Lf h(x) = ⟨dh(x), f (x)⟩. (3.11)

3.3.2 Bases of the Tangent and Cotangent Space

Consider again the definition of the Lie derivative. By writing

i f1 (x)
 
∂h(x)  .. 
h
∂h(x)
∂xn  . 
Lf h(x) = ∂x1 ...
fn (x) (3.12)
∂ ∂
 
= f1 (x) + . . . + fn (x) (h(x)),
∂x1 ∂xn

a vector field f (x) can be understood as a first order partial differential operator, which acts
on a real-valued function h(x). Hence, it makes sense to define { ∂x∂ 1 , . . . , ∂x∂n } as the basis
of the tangent space. In this basis, a vector f (x) is written

∂ ∂
f (x) = f1 (x) + · · · + fn (x) . (3.13)
∂x1 ∂xn

Consider the rate of change of a scalar function h : X → R along the solutions of a system
ẋ = f (x) + g(x)u. Assume that, as indicated in Fig. 3.2, the solutions evolve for arbitrary
inputs u(t) on a 2-dimensional integral (solution) manifold of X ⊂ R3 , on which h(x0 ) =
h(x(t)) = const. holds. Equivalently, ḣ = 0 along the flow Φ(x0 , t). We now write the
difference h(x(ε)) − h(x(0)) = 0 on a small time interval [0, ε], and thereby highlight the basis
elements of the cotangent space and justify the notion and notation of an exact differential:
Z ε
0= ḣ(x(t))dt
0
dx1
 
Z εh i dt
∂h ∂h ∂h  dx
=  dt2  dt

∂x1 ∂x2 ∂x3
0 dx3
dt
 
Z x(ε) h i dx1
= ∂h ∂h ∂h 
∂x3 dx2  dxi : differentials of coordinate functions (3.14)

∂x1 ∂x2
x(0)
dx3
Z h(x(ε))
= dh dh: differential/increment of h along the path
h(x(0))
= h(x(ε)) − h(x(0))

The third line nicely illustrates that {dx1 , . . . , dxn } is a natural basis of the cotangent space.
The exact differential dh, as a particular element of the cotangent space Tx∗ X with partial

© Prof. Dr.-Ing. habil. Paul Kotyczka 33

Chair of Automatic Control
Technical University of Munich
3.4. Lie Bracket

derivatives of h as coefficient functions, is represented in this basis. dh can be exactly inte-

grated5 , which is illustrated by the notation of the step from the 4th to the 5th line.
With the property – duality pairing of basis vectors –
(
∂ 1, i=j
⟨dxi , ⟩ = δij = (3.15)
∂xj 0, i ̸= j
where δij denotes the Kronecker delta, we verify that
n n
X X ∂
⟨w(x), f (x)⟩ = ⟨ wi (x)dx, fi (x) ⟩
i=1 i=1
∂xi
n (3.16)
X
= wi (x)fi (x)
i=1

is the result of the duality pairing between a covector field w(x) ∈ Tx∗ X and a vector field
f (x) ∈ Tx X , i.e., we obtain the “intuitive” result by doing the calculus in the appropriate
bases of tangent and cotangent space.

Remark 3.1. An interpretation why {dx1 , . . . , dxn } is the natural basis for the space of one-
forms on Rn is as follows. Consider the interval I = [a, b] ⊂ R with the single spatial coordinate
x ∈ I ⊂ R and the exact differential6
∂Q(x)
dQ = dx.
∂x
This exact differential can describe the distribution of charge on an electric transmission line,
with ∂Q(x)
∂x = ρ(x) the charge density per length.
Z b
∂Q(x)
Z
Qab = dQ = dx
I a ∂x
then represents the total charge, which is obtained by integration over I. How is the total
charge on a transmission line influenced (if we neglect distributed currents to ground)? Only
by the currents injected at the two terminals:
d
Qab = Ia − Ib .
dt
This is an integral conservation law for the charge and a 1D example for a class of systems
that can be modelled in a physically intuitive and natural way in the language of differential
forms (of degree k in general – k-forms), see e.g., [1]. Maxwell’s equations, heat transfer and
fluid dynamical problems are further examples on higher-dimensional spatial domains.

3.4 Lie Bracket

With the new understanding of vector fields and their dual objects in the differential geometric
setting, we now introduce the Lie bracket. While the Lie derivative returns the rate of change
5
This is not the case for an arbitrary one-form whose coefficient functions are not partial derivatives of a
scalar function.
6
Note that in this example, x is a spatial coordinate and not a state, which does only show another context
where differential forms naturally arise, i.e., integration in space.

34 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 3. SOME CONCEPTS FROM DIFFERENTIAL GEOMETRY

Figure 3.4: Effect of the Lie bracket in car parking

of a scalar function h(x) when transported along a vector field f (x), the Lie bracket gives an
answer to what occurs, when two vector fields interact. We will introduce the idea behind the
Lie bracket at an everyday life situation:

Example 3.2 (Parking a car in a parallel parking lot). Assume you want to bring your car from
the initial position A to the final position C as depicted in Fig. 3.4, knowing that your car
doesn’t allow a direct parallel shift by the available controls due to the kinematic constraints.
What you do, is the following sequence of actions.

1. Go back, steer right,

2. go back, steer left,
3. go forward, steer right (= do the “inverse” of 2.),
4. go forward, steer left (= do the “inverse” of 1.).
The result of this maneuver is the desired parallel shift of the vehicle, a motion which can not
be realized directly. ◁

The example illustrates an important point in nonlinear systems:

In nonlinear systems, in general, the order of subsequently applying (control) vector fields
plays an important role and can be used to generate new vector fields / directions of
motion.

We would like to abstract from the car parking problem and investigate the solution of the
following switched system.



 f (x), 0 ≤ t < ∆t,

 g(x), ∆t ≤ t < 2∆t,

ẋ = (3.17)


 −f (x), 2∆t ≤ t < 3∆t,

−g(x), 3∆t ≤ t < 4∆t.



For intervals of time ∆t, two vector fields, with positive and negative sign each, are subse-
quently applied and induce the solution x(t). If one lets ∆t → 0, then an interesting question
is whether the final value x(4∆t) equals the initial value x(0) or not.

© Prof. Dr.-Ing. habil. Paul Kotyczka 35

Chair of Automatic Control
Technical University of Munich
3.4. Lie Bracket

Figure 3.5: Constant vector fields

Example 3.3 (Constant vector fields). Consider the two constant vector fields
" # " #
1 1
f= , g= ,
1 −1
and let ∆t = 1. As an alternative to the sequence of vector fields as described in Eq. (3.17),
one could also do as follows: Start from an initial value and apply first f , then g. Then, start
from the same initial value and permute the sequence of the vector fields, i.e., start with g
and then apply f . Figure 3.5 depicts the solutions of the differential equations for the two
consant vector fields in this scenario. Both solutions end at the same point, they commute.
In the case of constant vectors f and g, we have
Φgt2 (Φft1 (x0 )) = Φft1 (Φgt2 (x0 )), (3.18)
which can be read as follows: Φft1 (x0 ) is the solution of ẋ = f (x), starting at x0 and evaluated
at time t1 , or alternatively, the flow induced by the vector field f . ◁

What we observed for constant vector fields, is not true in general for non-constant vector
fields. The situation then is sketched in Fig. 3.6.

Definition 3.4 (Lie bracket). The Lie bracket

∂g(x) ∂f (x)
[f (x), g(x)] := f (x) − g(x). (3.19)
∂x ∂x
is a measure for the non-commutativity of vector fields. If the Lie bracket of two vector
field is zero, we say that their flows commute.

For the derivation, see Handout 2.

The Lie bracket can point into a new direction that can not be represented by the original
vector fields. This is the property which allows parking a car in a parallel parking lot!
If the Lie bracket of two vector fields f (x) and g(x) is zero, then the flows of the two vector
fields commute. This is a very rare situation, which occurs for example if the two constant
vector fields are constant, f = const., g = const.
For two linear vector fields f (x) = Ax and g(x) = Bx, one might think that the Lie bracket
[Ax, Bx] (3.20)
lies in span{Ax, Bx}. This is, however not true in general, see the exercise.

36 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 3. SOME CONCEPTS FROM DIFFERENTIAL GEOMETRY

Figure 3.6: Non-commutativity of two vector-fields

3.5 Distributions and Involutivity

A weaker property than commutativity is involutivity, which is a property of a collection of

vector fields. Let us first define such objects, which are called distributions (of vector fields).

3.5.1 Distribution

Given a set of k vector fields f 1 (x), . . . , f k (x) ∈ Tx X . In every point x ∈ X , the vector fields
span a subspace of the tangent space

∆(x) := span{f 1 (x), f 2 (x), . . . , f k (x)} ⊂ Tx X . (3.21)

Definition 3.5 (Distribution). The assignment of a set of vector fields to the corresponding
subspace of Tx X in every point of X is called a distribution and denoted

∆ = span{f 1 , . . . f k }. (3.22)

The dimension of ∆ in a point x is defined as

dim(∆(x)) := rank[f 1 (x), . . . , f k (x)]. (3.23)

The dimension of ∆ may vary on X . If, however, for all x ∈ X , the vector fields f 1 (x), . . . , f k (x)
are linearly independent, then the matrix on the right hand side of (3.23) has full rank k in
every x ∈ X and
dim(∆(x)) = k = const. on X (3.24)
and we call ∆ nonsingular.
When f i (x) are smooth (i.e., sufficiently often differentiable) vector fields of a nonsingular
distribution ∆, then every g ∈ ∆ can be expressed as a linear combination of the vector fields
f i,
k
(3.25)
X
g(x) = ci (x)f i (x)
i=1

© Prof. Dr.-Ing. habil. Paul Kotyczka 37

Chair of Automatic Control
Technical University of Munich
3.5. Distributions and Involutivity

with smooth functions ci (x).

In the previous section, we noted that the Lie bracket of two vector fields generates – except
in the case of commutativity – a new vector field. Distributions, which are closed under the
Lie bracket7 are called involutive.

Definition 3.6 (Involutive distribution). A k-dimensional distribution ∆ is called involutive

if the following implication holds for all i, j ∈ 1, . . . , k:

f i, f j ∈ ∆ ⇒ [f i , f j ] ∈ ∆. (3.26)

3.5.2 Involutive Closure

If a distribution ∆ = span{f 1 , . . . , f k } turns out to be not involutive, then an involutive

distribution of higher dimension can be constructed in the following way: The involutive
closure of ∆ is denoted ∆ ¯ and is obtained by adding those (repeated) Lie brackets to ∆,
which are not contained in ∆:
¯ = span{f , . . . , f , [f , f ], . . . }.
∆ (3.27)
1 k i j
| {z }
∈∆
/

The iteration to construct the involutive closure can be written in a compact, algorithmic way
as follows:

¯ 0 := ∆
∆
i := 0
repeat
¯ i+1 := ∆
∆ ¯ i ∪ span{[f , g] | f , g ∈ ∆
¯ i}
i := i + 1
until ∆¯i = ∆¯ i−1
¯ := ∆
∆ ¯i

Example 3.4. The involutive closure of the distribution ∆ = span {g 1 , g 2 } in the unicycle
¯ = span {g , g , [g , g ]}
example (see exercise) with g 1 , g 2 ∈ Tx R3 vector fields over R3 is ∆ 1 2 1 2
¯
and has dimension dim ∆ = 3 = n. ◁

3.5.3 Annihilator of a Distribution

In the same way as we collect vector fields for a distribution, a set of covector fields (one-forms)
spans a codistribution. A special codistribution is the annihilator ∆⊥ of a distribution ∆:

7
More precisely, the vector spaces (tangent spaces) Tx X spanned by the vector fields in every point x ∈ X
are closed under the Lie bracket.

38 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 3. SOME CONCEPTS FROM DIFFERENTIAL GEOMETRY

Figure 3.7: Sketch of two vector fields and a “perpendicular” covector.

Definition 3.7 (Annihilator). The annihilator ∆⊥ of a distribution ∆ is defined as

∆⊥ = span{w ∈ T ∗ X | ⟨w, f i ⟩ = 0, f i ∈ ∆}. (3.28)

If ∆ has constant dimension k, then dim(∆⊥ ) = n − k.

Figure 3.7 shows the illustration of ∆ = span{f 1 , f 2 } with f 1 , f 2 ∈ R3 vectors in Euclidean

space. The annihilator ∆⊥ = span{w} consists of a single covector which can be drawn
perpendicular to the plane spanned by f 1 and f 2 . (The duality products of f 1 and f 2 with
w, which correspond to scalar products in R3 in the illustration, are zero.)

3.6 Integrability and Frobenius’ Theorem

We introduce a theorem, which is of fundamental importance in differential geometric methods

for control and observer design.

Definition 3.8 (Complete integrability). If the annihilator of a distribution ∆ is spanned

by exact differentials, then the distribution is called completely integrable.

Recall that we defined an exact differential as a covector whose components are the partial
derivatives of a scalar function λ(x):

∂λ ∂λ
dλ(x) = dx1 + · · · + dxn . (3.29)
∂x1 ∂xn

3.6.1 Interpretation of Complete Integrability

Consider Fig. 3.8, where the distribution ∆ comprises the two vector fields f 1 and f 2 . The
condition for complete integrability can be formulated as follows: There exists a scalar function
λ : R3 → R such that
∂λ(x)
⟨dλ(x), f i (x)⟩ = f (x)
∂x i
= Lfi λ(x) (3.30)
=0 for i = 1, 2.

© Prof. Dr.-Ing. habil. Paul Kotyczka 39

Chair of Automatic Control
Technical University of Munich
3.6. Integrability and Frobenius’ Theorem

Figure 3.8: Two leaves of an integral manifold (solution manifold) to illustrate the idea of
integrability of a distribution.

This means that in every point x the gradient of a function λ(x) shall be “perpendicular” to
f 1 (x) and f 2 (x). In this case, the possible solutions of

ẋ = ±f i (x), i = 1, 2,

starting from some initial point x(0), evolve on an object where dt

d ∂λ
λ = ∂x ẋ = 0 holds, or
respectively λ(x) = c = const. Such an object in 3-dimensional space, which is characterized
by one algebraic constraint, will be a curved plane. Such a curved plane is called a two-
dimensional integral manifold8 in R3 . The integral manifold on which the solutions actually
evolve, depends on the value of λ(x) = λ(x0 ) = c which is determined by the initial value
x(0) = x0 .

Remark 3.2. In our context, we always understand (complete) integrability of a k-dimensional

distribution as the existence of a k-dimensional integral manifold.

Looking at the equation (3.30) from a “computational” point of view, the following interpre-
tation of complete integrability can be given (for some general dimension of the state space
n = dim(X ) and dimension of the distribution k = dim(∆)):
The distribution ∆ = span{f 1 , . . . , f k } is completely integrable, if there are n − k real-valued
functions λj which each at the same time satisfy the k partial differential equations
∂λj (x)
f i (x) = 0, i = 1, . . . , k j = 1, . . . , n − k. (3.31)
∂x
A necessary and sufficient condition for this to be true is given by the famous Frobenius’
Theorem:

Theorem 3.1 (Frobenius). A distribution ∆ = span{f 1 , . . . , f k } is completely integrable

if and only if it is involutive.

Frobenius’ Theorem is an important result for differential geometric control theory. It allows
an easy check of complete integrability of a distribution: If the Lie brackets of all vector fields
8
The term integral manifold stems from the fact that it can be characterized by the solutions (integrals) of
differential equations.

40 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 3. SOME CONCEPTS FROM DIFFERENTIAL GEOMETRY

Figure 3.9: Sketch of the bases of the tangent space in straightened out coordinates. The
vector fields that induce the flow have zero component in the transformed ∂z∂ 3 direction.

in ∆ again lie in ∆, then the condition is satisfied and ∆ is integrable. If ∆ is not involutive,
¯ will be an integrable distribution if dim(∆)
then its involutive closure ∆ ¯ < n.

Note that if dim(∆)¯ = n, there exists no function λ(x), which remains constant along the
solutions of the underlying dynamical system, and which could define an integral manifold of
dimension less than n. In fact, the “integral manifold” is the state space itself in this case.
Frobenius’ Theorem gives us an idea how the solutions of a system, which is driven by a
number of vector fields, evolve in state space, without explicitly computing its solutions. An
important implication is the following.

3.6.2 Straightening Out of Vector Fields

If and only if a k-dimensional distribution

∆ = span{f 1 , . . . , f k }

is involutive, then the k vector fields can be represented in new (local) coordinates as vector
fields with zero components in n − k directions. Under a smooth coordinate transformation
(diffeomorphism)
z = s(x) (3.32)
the vector fields, i = 1, . . . , k
n
∂
(3.33)
X
f i (x) = fij (x)
j=1
∂xj

transform to
k
∂
f˜ij (z) (3.34)
X
f̃ i (z) = .
j=1
∂z j

fij denotes the j-th component of the vector f i .

© Prof. Dr.-Ing. habil. Paul Kotyczka 41

Chair of Automatic Control
Technical University of Munich
3.7. References

Proof (only necessity): Vector fields are transformed using the Jacobian9 of the transformation,
i.e.,
∂s(x)
f̃ i (z) = f (x) ◦ s−1 (z).
∂x i
The condition for the required shape of the transformed vector fields is then

∗ ··· ∗
  
 .. .. .. 


 .
 . . 
 k
 ∗ ··· 
∗ 
∂s(x) 
! 

[f 1 (x), . . . , f k (x)] = − − − − − − (3.35)

∂x  0 ···
  
0 
 .. .. .. 
  
 . . .  n−k


0 ··· 0

Therein, ∗ in the upper block denotes an arbitrary element, as opposed to the necessary 0
elements in the lower block. The n − k partial dimensional equations represented by this lower
block are
∂si (x)
f j (x) = 0, i = k + 1, . . . , n j = 1, . . . , k, (3.36)
∂x
or in different notation

⟨dsi , f j ⟩ = 0, i = k + 1, . . . , n j = 1, . . . , k. (3.37)

The latter equation obviously reveals that the condition for this straightening out of the vector
fields of ∆ is equivalent to the requirement that the annihilator of ∆ is spanned by exact
differentials. This is equivalent to ∆ being completely integrable, for which involutivity is a
necessary and sufficient condition according to Frobenius.

3.7 References

For a more detailed and complete introduction to the sketched concepts of differential geometry
for dynamical systems, see for example Sections 1.3 and 1.5 of [2] and Sections 34 and 39 of
[3]. Section 12.4 of [4] contains an introduction from a control perspective. The books [5]
and [6] deal with control theory from a differential geometric point of view.

Bibliography

[1] H. Flanders. Differential Forms with Applications to the Physical Sciences. Number 3.
Academic Press, New York, 1963.

[2] P. J. Olver. Applications of Lie Groups to Differential Equations. Springer, New-York,

1993.

[3] V. I. Arnold. Mathematical Methods of Classical Mechanics, volume 60. Springer Science
& Business Media, 1989.
9 ∂z
Think of ż = ∂x ẋ and replace the vector fields in x- and z-coordinates
The ◦s−1 (z) indicates, that in the result of the operation on the right hand side, the argument, i.e., x, has to
be replaced by s−1 (z) in order to have the result in new coordinates z.

42 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 BIBLIOGRAPHY

[4] H. Khalil. Nonlinear Systems. Prentice Hall, 3rd edition, 2002.

[5] A. Isidori. Nonlinear Control Systems. Springer Science & Business Media, 2013.

[6] H. Nijmeijer and A. van der Schaft. Nonlinear Dynamical Control Systems. Springer, 2017.

© Prof. Dr.-Ing. habil. Paul Kotyczka 43

Chair of Automatic Control
Technical University of Munich
Bibliography

44 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
Chapter 4

Controllability and Feedback

Linearization

In this chapter, we start with discussion of controllability and different related notions for
nonlinear input-affine systems. Based thereon, we present a condition, under which a nonlinear
system can be made completely linear by state feedback and a state transformation. The
condition for input-state-linearizability being quite restrictive, we end the chapter with the
presentation of input-output linearization and the definition of internal/zero dynamics.

4.1 Controllability Definition and Driftless Systems

For linear systems ẋ = Ax + Bu, controllability is defined as the property that for every initial
state x(0) ∈ Rn , there exists a control input u(t) such that the equilibrium x∗ = 0 is reached
in finite time. By a simple coordinate shift, we can also argue that for a controllable linear
system, there exist finite time control inputs, which steer the state from x∗ = 0 to any other
final state xf ∈ Rn . For nonlinear systems, these cases are not necessarily equivalent, and
various notions like reachability and accessibility describe the possibilities to steer the nonlinear
system between points and sets in finite time. The most general controllability definition is as
follows:

Definition 4.1 (Controllability, [1], Def. 3.2). The nonlinear system

m
(4.1)
X
ẋ = f (x) + g i (x)ui ,
i=1
h iT
with x ∈ X ⊂ Rn , u = u1 . . . um , u ∈ U ⊂ Rm , is controllable if for any two states
x1 , x2 ∈ X , there exists a finite time control input u : [0, T ] → U, T < ∞ such that
x(T ) = x2 for the initial value x(0) = x1 .

Example 4.1 (Controllability of the unicycle). Consider the unicycle model from the exercise,
for which we computed the Lie bracket of the two input vector fields g 1 (x) and g 2 (x). We
argued that by switching the control inputs u1 and u2 in an alternating manner, we can
move (indirectly) the unicycle also in the sideways direction [g 1 (x), g 2 (x)]. We can reach (as

45
4.2. Example: Satellite Attitude Control

we experience with a car or a bicycle) from any initial state an arbitrary final configuration
(position plus orientation) in finite time. The unicycle is, hence, controllable. ◁

The statement that the unicycle is controllable can, however, not be deduced from its lin-
earization (see exercise). In general this means that information on the controllability of a
nonlinear system can get lost under linearization.
The unicycle is an example for a driftless system, i.e., the drift vector field in the system
representation (4.1) is zero, f (x) = 0. Before we enter the discussion on controllability and
related properties for nonlinear systems of the form (4.1), we concentrate on driftless systems

ẋ = g 1 (x)u1 + · · · + g m (x)um . (4.2)

We first recall how the solutions of such systems with multiple control inputs evolve in state
space, by using our insights from the previous chapter. These insights will make plausible the
statement of the famous Chow–Rashewskii Theorem for controllability of driftless systems.
Let ∆ = span{g 1 , . . . , g m } be the distribution of control vector fields of dimension m and
∆¯ = inv(∆) its involutive closure with dim ∆ ¯ = l < n, where l ≥ m. Then, by alternating the
inputs u1 , . . . , um , the flow generated by the control vector fields evolves on an l-dimensional
subspace (manifold) of X ⊂ Rn . As discussed earlier, the state space X is foliated into l-
dimensional submanifolds, depending on the initial values x(0). These submanifolds can be
imagined as leaves 1 , see Fig. 4.4.
¯ = n, then the flow is not confined to a lower-dimensional subspace, which is necessary
If dim ∆
for controllability of (4.2) according to Definition (4.1). If this condition holds for all x ∈ X ,
then it is necessary and sufficient for controllability. The theorem goes back to the 1930s,
when it was proven independently by Chow in China [2] and Rashevskii in the USSR [3]. Here
we reproduce the control systems oriented formulation as in [4] (Chapter 10, Theorem 7):

Theorem 4.1 (Controllability of driftless systems). The driftless system (4.2) is controllable
¯
if and only if dim ∆(x) = n at each x ∈ X .

4.2 Example: Satellite Attitude Control

Figure 4.1: Sketch of a symmetric rigid with body fixed frame and torques applied in the
principal axes. The discs could be reaction wheels on a satellite.
1
Think of the coordinate transformation which straightens out the vector fields such that they have n − l
zero components.

46 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 4. CONTROLLABILITY AND FEEDBACK LINEARIZATION

This subsection is inspired by Example 10.11 in [4], where Chow’s theorem is applied to the
controllability analysis for the rotation of a rigid body in R3 . One might think of an object
as depicted in Fig. 4.1. Before analysing controllability in the space of angular velocities,
and then in the space of configurations / the manifold of spherical rotations, we present the
modeling equation.

4.2.1 Modeling

We first illustrate that 3 × 3 rotation matrices R(t), which depend on a parameterization in

terms of angles around principal axes, are a natural choice to describe the attitude of the
satellite. We then represent the rate of change of the configuration Ṙ(t) in terms of a vector
of angular velocities. Thereafter, we introduce input torques, which brings us to the Euler
differential equations in terms of angular velocities for the rotating body.

Kinematics An arbitrary configuration of the rigid body (the satellite) can be described as
the composition of three rotations around its principal axes. For example can we use Euler
angles to express the sequence of rotations in the body frame starting with the body and
inertial frame aligned:

Figure 4.2: Illustration of the rotations with the three Euler angles according to Eq. (4.3).

First turn by an angle ϕ about the z-axis, then by θ about the y-axis, and finally by ψ about
the z-axis. The rotation matrix, see [4], Subsection 2.5.1,

R = Rz,ϕ Ry,θ Rz,ψ

   
cϕ −sϕ 0 cθ 0 sθ cψ −sψ 0 (4.3)
= sϕ cϕ 0  0 1 0  sψ cψ 0
   
0 0 1 −sθ 0 cθ 0 0 1

takes a vector r b written in the body fixed frame (representing a point on the satellite) to its
inertial frame representation r 0 (t). If R(t) is varying in time, then we write

r 0 (t) = R(t)r b . (4.4)

Rotation matrices are orthogonal matrices with determinant 1, they belong to the so-called
special orthogonal group SO(3), which is a particular matrix Lie group 2 . The description of the
velocity kinematics below follows Section 4.1 of [5]. With R(t) ∈ SO(3) a description of the
2
This is only to mention the term, we will not use this framework in the rest of the course.

© Prof. Dr.-Ing. habil. Paul Kotyczka 47

Chair of Automatic Control
Technical University of Munich
4.2. Example: Satellite Attitude Control

time-varying orientation of the satellite3 , we are now interested in the velocities. Differentiating
(4.4), we obtain the velocity of r b in the inertial frame:
v 0 (t) = Ṙ(t)r b
= Ṙ(t)R−1 (t) R(t)r b . (4.5)
| {z } | {z }
ω̂ 0 (t) r 0 (t)

The role of the instantaneous spatial (or inertial) angular velocity matrix ω̂ 0 (t) will be clarified
in a moment. First note that from orthogonality of R(t), we know that
R(t)RT (t) = I (and also RT (t)R(t) = I) (4.6)
with I ∈ R3×3 the identity matrix, which implies R−1 (t) = RT (t). Differentiating the left
equation with respect to time (the same works for the right one), we get
T
Ṙ(t)RT (t) + R(t)Ṙ (t) = 0 ⇔ Ṙ(t)RT (t) = −(Ṙ(t)RT (t))T , (4.7)
i.e., the matrix
ω̂ 0 (t) = Ṙ(t)R−1 (t) (4.8)
is skew-symmetric. Multiplication with R(t) from the right gives the velocity kinematics
equation for the rotating rigid body
Ṙ(t) = ω̂ 0 (t)R(t), (4.9)
with (we omit the time argument)
 
0 −ωz ωy
ω̂ 0 =  ωz 0 −ωx 
 
−ωy ωx 0 (4.10)

= ωx êx + ωy êy + ωz êz .

h iT
The vector ω 0 = ωx ωy ωz ∈ R3 contains the angular velocities of the rotating frame
with respect to the axes of the inertial frame and
     
0 0 0 0 0 1 0 −1 0
êx = 0 0 −1 , êy =  0 0 0 , êz = 1 0 0 , (4.11)
     
0 1 0 −1 0 0 0 0 0
with ex , ey and ez the unit vectors of R3 , are the matrices of infinitesimal rotations around
the respective axes. These skew-symmetric matrices, which represent infinitesimal generators
of rotation matrices, are elements of the Lie algebra so(3) of SO(3).
In a dual manner, one obtains the velocity kinematics with respect to the instantaneous body
velocities as
Ṙ(t) = R(t)ω̂ b (t). (4.12)
From this equation and (4.9) it follows that
ω̂ 0 = Rω̂ b R−1 , (4.13)
and one can show the relation
ω 0 = Rω b (4.14)
between the velocity vectors ω b , ω 0 ∈ R3 in the body and the inertial frame.
3
 
Indeed, it is sufficient to have two columns of R(t) = r x (t) r y (t) r z (t) to parameterize the attitude,
as the third axis completes the coordinate frame according to the right hand rule.

48 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 4. CONTROLLABILITY AND FEEDBACK LINEARIZATION

Dynamics We shall now express the dynamics of the rotating body in terms of the angular
velocities with respect to the body frame ω = ω b . To obtain the forced equations of motion
h iT
with τ = τ = τx τy τz the body-fixed input torques and J = diag{Jx , Jy , Jz } the
diagonal matrix of the moments of inertia around the principal axes, we start with Newton’s
second law for the vector of angular momenta L0 ∈ R3 in the inertial frame:

L̇0 = τ 0 . (4.15)

The torque vector4 expressed in the inertial frame is5

τ 0 = Rτ . (4.16)

The body frame momentum vector Lb = J ω transforms according to (4.14),

L0 = RLb , (4.17)

and its time derivative becomes

L̇0 = ṘLb + RL̇b

(4.18)
= R(ω̂Lb + L̇b ).

Substitution of L̇0 and τ 0 according to (4.16), (4.18) in Newton’s law (4.15) yields, after
multiplication with RT ,

L̇b = −ω̂Lb + τ . (4.19)

With the property6 ω̂Lb = ω × Lb , the definition of Lb and anti-commutativity of the cross
product, we can write
J ω̇ = (J ω) × ω + τ . (4.20)
In what follows, we first assume a velocity feedback of the form

τ = ω × (J ω) + J (ω × u) (4.21)

with u ∈ R3 the new input vector with units of angular velocity, which renders the Euler
equations “velocity controlled”:
ω̇ = ω × u. (4.22)

4.2.2 Controllability Analysis

We now analyse the controllability of

• the velocity controlled dynamics equation (4.22) on R3 and

4
Geometrically speaking, it is a co-vector: It can be paired with the collocated velocity vector to get the
scalar supplied power.
5
That this relation holds, can be proven by the requirement that the power τ T0 ω 0 = τ Tb RT Rω b = τ Tb ω b is
invariant with respect to the considered coordinate frame.
6
Prove this property “for fun” using (4.10).

© Prof. Dr.-Ing. habil. Paul Kotyczka 49

Chair of Automatic Control
Technical University of Munich
4.2. Example: Satellite Attitude Control

• the kinematics equation (4.9) on SO(3), where we assume ω ≈ ω ∗ as input, realized by

a fast underlying velocity control loop. You can imagine the torque control law

τ = ω × (J ω) + J T −1 (ω ∗ − ω) (4.23)

with T = diag {Tx , Ty , Tz } a matrix of small time constants, which renders (4.20) the
controlled velocity dynamics
T ω̇ = −ω + ω ∗ . (4.24)

By neglecting this dynamics of the underlying velocity control loop, we only consider the
velocity kinematics
Ṙ(t) = R(t)ω̂ ∗ (t). (4.25)

Controllability on R3 Equation (4.22) can be written in the form of a driftless control system
(for convenience we write 1, 2, 3 instead of x, y, z as indices) as

ω̇ = ω × u
     
0 −ω3 ω2
=  ω3  u1 +  0  u2 + −ω1  u3 (4.26)
     
−ω2 ω1 0
= g 1 (ω)u1 + g 2 (ω)u2 + g 3 (ω)u3 .

For Chow’s theorem we try to construct the involutive closure of the distribution

∆ = span {g 1 , g 2 , g 3 } (4.27)

on R3 \{0}. We compute the Lie brackets

[g i (ω), g j (ω)] = S j g i (ω) − S i g j (ω) (4.28)

where S i = ∂g∂ωi (ω)

gives skew-symmetric (basis) matrices, linear combinations of the ones in
(4.11). It turns out that

[g 1 , g 2 ] = g 3 , [g 2 , g 3 ] = g 1 , [g 3 , g 1 ] = g 2 , (4.29)

¯ = ∆, i.e., the distribution ∆ is already involutive. By

which shows that ∆

ω1 g 1 (ω) + ω2 g 2 (ω) + ω3 g 3 (ω) = 0 (4.30)

for all ω ∈ R3 , which can easily be verified, we see that each vector fields in ∆ depends linearly
on the others, which shows that

¯ = dim ∆ = 2 < 3.
dim ∆ (4.31)

Therefore, the driftless system (4.27) is not controllable according to Chow’s theorem.

50 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 4. CONTROLLABILITY AND FEEDBACK LINEARIZATION

Controllability on SO(3) We now analyse controllability on the space of rotations (or rota-
tion matrices) SO(3) for the velocity-controlled kinematics equation (4.25) (omitting the “*”).
We decompose the rotation matrix R into its column vectors
 
h i rx1 ry1 rz1
R = rx ry rz = rx2 ry2 rz2  , (4.32)
 
rx3 ry3 rz3

each of which gives the direction of the corresponding body axes (x, y and z) in the inertial
frame. First consider the dynamics of the x-axis of the body-fixed frame,

ṙ x = Rω̂ b ex
 
h i 0
= rx ry rz  ωz  (4.33)
 
−ωy
= r y ωz − r z ωy .

By orthogonality of the vectors r x , r y and r z that form the body-fixed frame, the two vector
fields in this driftless system span the tangent plane to the unit sphere attached at the tip
of r x , see Fig. 4.3. This means, that the vector r x can be rotated in all possible directions
on the sphere by the two control inputs ωz and ωy . To also arbitrarily align the body y- and
z-axes, another pure rotation around the x-axis, realized by ωx is necessary.

Figure 4.3: The vector r x represents an element of the sphere S 2 . Together with the ori-
entation of r y (or r z ), a complete rotation R(t) is parameterized. We illustrate the above
argumentation that arbitrary directions in the tangent space to the sphere plus orientations
can be achieved by controlling the angular velocities.

The argumentation makes plausible that we can join two arbitrary attitudes R(t0 ) and R(t1 ) of
the body fixed frame in finite time by the velocity controls, which corresponds to controllability
of (4.25) on SO(3).

4.3 Accessibility of Nonlinear Systems

In control it is desirable to reach all states in the neighborhood of an initial value in finite time,
this means in particular in every direction from the initial value. This is basically the notion
of (strong) accessibility, which we will have defined and for which we will have stated criteria
at the end of this section.

© Prof. Dr.-Ing. habil. Paul Kotyczka 51

Chair of Automatic Control
Technical University of Munich
4.3. Accessibility of Nonlinear Systems

Figure 4.4: Foliation of the state space.

First recall Def. 3.2 of a linear vector space. The Lie algebra, as a mathematical structure
which will appear in the definitions and criteria below, is a linear vector space, whose elements
satisfy additional conditions:

Definition 4.2 (Lie algebra, [1], Def. 2.28). A vector space V is a Lie algebra, if in addition
to the above operations and properties, there is an operation [·, ·] : V × V → V such that
for all f , g, h ∈ V the following properties hold:

a) Bilinearity (k, l ∈ R):

[kf + lg, h] = k[f , h] + l[g, h]

b) Skew-symmetry:
[f , g] = −[g, f ]

c) Jacobi identity:
[f , [g, h]] + [h, [f , g]] + [g, [h, f ]] = 0.

We recognize that the set of (smooth) vector fields over an n-dimensional subspace of Rn ,
together with the Lie bracket, is a Lie algebra: The Lie bracket of a smooth vector field is again
a smooth vector field, and the Lie bracket has the properties given in the above definition (do
Exercise 3.2 to verify them).

Definition 4.3 (Subalgebra, [1], below Def. 2.28). A subalgebra of a Lie algebra (V, [·, ·])
is a linear subspace V ′ ⊂ V such that [f ′ , g ′ ] ∈ V ′ for all f ′ , g ′ ∈ V ′ , i.e., a subspace
that is closed under the Lie bracket.

In other words, a subalgebra is a subspace of a Lie algebra, which is closed under the (Lie)
bracket. Coming back to our nonlinear dynamical system (4.1), we can define a particular
subalgebra.

Definition 4.4 (Accessibility algebra, [1], Def. 3.7). The accessibility algebra C is the
smallest subalgebra of the set of smooth vector fields on X , which contains the drift
vector field f and the control vector fields g 1 , . . . , g m .

To construct the accessibility algebra, we start with the set {f , g 1 , . . . , g m }, and we succes-

52 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 4. CONTROLLABILITY AND FEEDBACK LINEARIZATION

sively add Lie brackets of the contained vector fields, until the set is closed under the Lie
bracket, i.e., until the corresponding Lie bracket can be expressed as a linear combination of
the elements of the (extended) set7 .

Remark 4.1. Note that a subalgebra, like the accessibility algebra, is by definition a subspace
of a vector space. It can therefore be represented by a set of (basis) vectors, which span the
according subspace. With the same vector fields, we can define a corresponding distribution,
which is understood (see previous chapter) as the mapping from a point x ∈ X to a subspace
of the tangent space Tx X .

Definition 4.5 (Accessibility distribution). The accessibility distribution C is the distribu-

tion generated by the vector fields in the accessibility algebra C. In a point x ∈ X , we
have
C(x) = span {v(x) | v ∈ C}. (4.34)

By the fact that C is a subalgebra (a set of vector fields, which is closed under the Lie bracket),
the involutivity of the accessibility distribution C immediately follows. The following iterative
procedure to construct C is presented in Appendix D.2 of [6].
C 0 := span{f , g 1 , . . . , g m }
i := 0
repeat
C i+1 := C i ∪ span{[v, w] | v ∈ C i , w ∈ C0 }
i := i + 1
until C i = C i−1
C := C i

Example 4.2. For the linear system SISO system

ẋ = f (x) + g(x)u
(4.35)
= Ax + bu,

we start with C 0 (x) = span{Ax, b}, obtain C 1 (x) = span{Ax, b, Ab}, and the algorithm
terminates with

C(x) = C n+1 (x) = C n (x) = span{Ax, b, Ab, . . . , An−1 b}. (4.36)

The distribution is spanned by n + 1 elements (one linear and n constant vector fields), if the
system is controllable according to Kalman. The reason that the last element is necessary, is
that An−1 b cannot be expressed as a linear combination of the first n vector fields for all x,
because the first one is linear in the state x. ◁

Now how does all this help us in checking controllability-related properties of nonlinear systems
(4.1)? We first define a property, which is weaker than controllability. To this end, we introduce
reachable sets 8 .
7
See [1], Prop. 3.8: Every element of C is a linear combination of repeated Lie brackets.
8
We use the same symbol V for the neighborhood x0 and a (generic) vector space. Due to specific context,
this should not cause confusion.

© Prof. Dr.-Ing. habil. Paul Kotyczka 53

Chair of Automatic Control
Technical University of Munich
4.3. Accessibility of Nonlinear Systems

Reachable "tube"
for small T

System locally accessible from x0 System NOT loc. access. from x0

Figure 4.5: Illustration of two sets RTV (x0 ), one with non-empty (left), the other with empty
interior.

Definition 4.6 (Reachable set). The set of states x = x(τ ), which can be reached from
an initial value x(0) = x0 by an admissible input u : [0, τ ] → U in finite time τ , while the
trajectory x(t) remains in the neighborhood V of x0 for 0 ≤ t ≤ τ , is called a reachable
set and denoted RV (x0 , τ ).

The union of reachable sets up to a time T > 0 is denoted RTV (x0 ) = ∪τ ≤T RV (x0 , τ ) and is
sometimes referred to as reachable tube.

Definition 4.7 (Local accessibility, [1], Def. 3.10). The system (4.1) is called locally
accessible from x0 if RTV (x0 ) contains a non-empty open set for all neighborhoods V and
all times T > 0. If the system is locally accessible from all x0 ∈ X , it is called locally
accessible.

The openness of the mentioned subset of RTV (x0 ) means that points in any direction in the
neighborhood can be reached, and the solutions are not confined to a subspace of lower
dimension. Note however that the definition of local accessibility does not require the single
reachable sets RV (x0 , τ ), τ ≤ T to contain open sets! The following theorem gives a criterion
for local accessibility.

Theorem 4.2 (Local accessibility, [1], Cor. 3.11 of Thm. 3.9). If dim C(x) = n for all
x ∈ X , with C the accessibility distribution, then the system (4.1) is locally accessible.

A stronger property, which requires all reachable sets to contain open sets is local strong
accessibility.

Definition 4.8 (Local strong accessibility, [1], Def. 3.18). The system (4.1) is locally
strongly accessible from x0 if for any neighborhood V and every T > 0 sufficiently small,
the reachable sets RV (x0 , τ ), τ ≤ T are non-empty open sets. The system is called
locally strongly accessible if it is locally strongly accessible from every x0 ∈ X .

Fig. 4.5 illustrates the distinction between the cases of an open and a closed set RTV (x0 ). We
close this section with a criterion for local strong accessibility, for which we need the following
definition:

54 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 4. CONTROLLABILITY AND FEEDBACK LINEARIZATION

Definition 4.9 (Strong accessibility algebra/distribution, [1], Def. 3.19). Let C be the
accessibility algebra of (4.1) according to Def. 4.4. C0 is the smallest subalgebra of C,
which contains g 1 , . . . , g m and satisfies

[f , v] ∈ C0 for all v ∈ C0 . (4.37)

The corresponding involutive distribution is

C0 (x) = span {v(x) | v vectorfield in C0 }. (4.38)

C0 and C0 are called strong accessibility algebra and strong accessibility distribution,
respectively.

The property (4.37) that Lie brackets of the elements of C0 with f provide no “new directions”
is called f -invariance of the strong accessibility algebra/distribution.

Theorem 4.3 (Local strong accessibility, [1], Thm. 3.21). If

dim C0 (x) = n, (4.39)

then the system (4.1) is locally strongly accessible from x.

Example 4.3. Consider the local accessibility distribution C(x) for the linear SISO system in
Example 4.2, Eq. (4.36). We are now looking for the smallest distribution C0 (x) contained in
C(x), which (i) contains g(x) = b, which (ii) is involutive and (iii) which is f -invariant, this
means the Lie bracket of each element with the drift vector field f (x) = Ax lies in C0 (x).
You can verify that for the linear SISO system, the local strong accessibility distribution is

C0 (x) = span{b, Ab, . . . , An−1 b}, (4.40)

which is C(x) without the drift vector field. You directly recognize the column vectors of
Kalman’s controllability matrix, and dim C0 (x) = n is equivalent to Kalman’s controllability
criterion – this is also true in the MIMO case ẋ = Ax + Bu. See also the exercise. ◁
Example 4.4 (See [1], Example 3.24). For the attitude control of a satellite using control
torques as inputs, we use the Euler equations (4.20). We assume that
   
τ1 u1
τ2  = u2  , (4.41)
   
τ3 0

i.e., two axes are actuated. The equations can be rewritten in the form
     
a1 ω2 ω3 b1 0
ω̇ = a2 ω1 ω3  +  0  u1 + b2  u2 (4.42)
     
a3 ω1 ω2 0 0

with a1 = (J2 − J3 )/J1 , a2 = (J3 − J1 )/J2 , a3 = (J1 − J2 )/J3 and b1 = 1/J1 , b2 = 1/J2 .
We want to investigate local strong accessibility from ω = 0, i.e., from rest. To set up
the strong accessibility distribution C0 (ω), we start with the two input vector fields g 1 , g 2

© Prof. Dr.-Ing. habil. Paul Kotyczka 55

Chair of Automatic Control
Technical University of Munich
4.4. Input-State Linearization

and add the Lie brackets with the drift vector field [g 1 , f (ω)], [g 2 , f (ω)]. The resulting
distribution has only dimension 2 at ω = 0 (see below), therefore (and to ensure involutivity
or closedness under the Lie bracket) we add another Lie bracket of the contained vector fields,
[g 2 [g 1 , f (ω)]] = [g 1 [g 2 , f (ω)]]. The result is finally

C0 (ω) = span {g 1 , g 2 , [g 1 , f (ω)], [g 2 , f (ω)], [g 2 [g 1 , f (ω)]]}

         
b1 0 0 a1 b2 ω3 0 (4.43)
= span { 0  , b2  , a2 b1 ω3  ,  0  ,  0 },
         
0 0 a3 b1 ω2 a3 b2 ω1 a3 b1 b2

which is an involutive distribution of dimension 3 at ω = 0 if and only if a3 = (J1 −J2 )/J3 ̸= 0.

Local strong accessibility at ω = 0 of the satellite attitude control with two torque actuators
therefore requires different moments of inertia J1 and J2 of the actuated axes. ◁

We will re-use the notions that appear in the definitions of subalgebras and (strong) acces-
siblity distributions in the following section in the formulation of a criterion for input-state-
linearizability.

4.4 Input-State Linearization

A straightforward idea for nonlinear controller design is the following. Take the nonlinear
system (x ∈ Rn , we restrict ourselves to the SISO case u, y ∈ R)9

ẋ = f (x) + g(x)u (4.44)

and find a feedback control law

u = α(x) + β(x)v (4.45)
as well as a smoothly invertible coordinate transformation

z = t(x) ⇔ x = t−1 (z), (4.46)

such that the closed-loop system with the new input v behaves like a linear one:

ż = Az + bv. (4.47)

The advantage of such complete feedback linearization is very clear: All known linear tech-
niques for control design can be directly applied to the transformed system10 (4.47). It will
become very clear in the next subsection on input-output linearization that such input-to-state
linearizability is equivalent to the existence of a (ficticious or virtual) output y = h(x), for
which the system (4.44) has full relative degree n. You will see that in this case, it is very
easy to endow the system with completely linear dynamics.

9
We are aware that the state space and the set of possible inputs are usually restricted, i.e. x ∈ X ⊂ Rn ,
and correspondingly u ∈ U ⊂ R. For brevity we write Rn and R.
10
As in the whole course we assume the state x to be measurable and therefore also z to be known.

56 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 4. CONTROLLABILITY AND FEEDBACK LINEARIZATION

Definition 4.10 (Relative degree). The relative degree r ≤ n of the SISO system

ẋ = f (x) + g(x)u
(4.48)
y = h(x)

is the number of times the output y has to be differentiated with respect to time until
the input appears for the first time.

Applying this definition to (4.48) gives the sequence of computations

y = h(x)
∂h ∂h ∂h
ẏ = ẋ = f (x) + g(x) u
∂x ∂x ∂x
| {z } | {z }
Lf h(x) Lg h(x)=0

∂Lf h ∂Lf h ∂Lf h

ÿ = ẋ = f (x) + g(x) u
∂x | ∂x{z } | ∂x{z }
L2f h(x) Lg Lf h(x)=0

..
.

(r) ∂Lr−1
f h
y = ẋ = Lrf h(x) + Lg Lr−1
f h(x) u. (4.49)
∂x | {z }
̸=0

We recognize that for relative degree n the following repeated Lie derivatives must be zero,

Lg h(x) = Lg Lf h(x) = . . . = Lg Lfn−2 h(x) = 0, (4.50)

while Lg Ln−1
f h(x) ̸= 0. Notation of repeated Lie derivatives:

Lf (Lf h(x)) = L2f h(x),

(4.51)
Lg (Lf h(x)) = Lg Lf h(x) etc.

We now state the conditions for a system to be input-state linearizable.

Theorem 4.4 (Input-state linearizability, [7], Thm. 12.2). The SISO system (4.44) is
input-state linearizable if and only if

1. The matrix
G(x) = [g(x), adf g(x), . . . , adfn−1 g(x)] (4.52)
has full rank n for all x ∈ X (i.e., dim C0 = n with C0 the strong accessibility
distribution) and

2. the distribution
D = span{g, adf g, . . . , adn−2
f g} (4.53)

is involutive on X .

© Prof. Dr.-Ing. habil. Paul Kotyczka 57

Chair of Automatic Control
Technical University of Munich
4.5. Input-Output Linearization

The proof for the existence of a (virtual) output y = h(x) = t1 (x) of full relative degree n,
based on which the invertible coordinate transformation (4.46) can be constructed, can be
found on the corresponding handout. Some comments:

• The notation adkf g is a shortcut for repeated Lie brackets with f :

adf g = [f , g], ad2f g = [f , adf g], ad3f g = [f , ad2f g], etc.

• G(x) is nothing else than the matrix containing the vector fields, which span the strong
accessibility distribution C0 according to Definition 4.9 for the SISO case. The rank
condition is nothing else than dim C0 (x) = n in Theorem 4.3.

• The requirement that D be an involutive distribution of dimension n − 1 is not at odds

with the fact that rank G(x) = n (or equivalently dim C0 = n). C0 is constructed by
adding Lie brackets of the elements with f , while the involutivity condition on D means
that there is no other linearly independent vector field that results from bracketing with
the elements contained in D. The distribution D, however, does not contain f .

4.5 Input-Output Linearization

We consider a SISO system with given output

ẋ = f (x) + g(x)u
(4.54)
y = h(x).

Instead of a complete linearization of the dynamics (i.e., a completely linear state representation
in new coordinates) by feedback, we aim this time at linear input-output behavior, which can
be desribed by a target transfer function G(s):

Y (s) = G(s)V (s). (4.55)

As above, v(t) and its Laplace transform V (s) represent the input to the closed-loop system.

Remark 4.2. We consider the input-output linearization for the SISO case. The MIMO case
can be considered a nonlinear generalization of the decoupling control for linear MIMO systems
(see e. g. Advanced Control).

One advantage of making the I/O behavior of the system (4.54) linear (with new input) is
that trajectory tracking can be easily achieved for the system in the form (4.55). To make the
output exactly follow a reference signal w(t) c s W (s), we demand
!
Y (s) = W (s). (4.56)

Obviously, if 1
G(s) is a stable transfer function (this means G(s) has no right hand zeros), the
input signal v(t) c s V (s) with

1
V (s) = W (s) (4.57)
G(s)

58 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 4. CONTROLLABILITY AND FEEDBACK LINEARIZATION

does the job. If G(s) is stable, initial deviations from the desired trajectory (due to unknown
initial values) disappear, i.e., y(t) → w(t) for t → ∞. If G(s) is unstable or the dynamics in
closed loop shall be modified, feedback is necessary to stabilize the trajectory tracking error at
0.
I/O linearization is easy to derive. However, in certain cases the approach is inapplicable due
to the possible instability of the so-called zero dynamics 11 . We start with an example, where
the key aspects of I/O linearization become clear and we then generalize the approach.

4.5.1 Illustrative Example

Given the nonlinear, input-affine second order system

ẋ1 = sin x2 + u
ẋ2 = x1 + x2 (4.58)
y = x1 .
We take the output y = x1 , differentiate it with respect to time,
ẏ = sin x2 + u, (4.59)
and observe that the control input u appears in this equation. It directly affects the first order
time derivative ẏ of the output. We now can use u in order to impose desired first order
dynamics for the system with new input v, e. g.
ẏ = −y + v. (4.60)
By comparison of equations (4.59) and (4.60) we obtain the control law
u = −y − sin x2 + v. (4.61)
The resulting dynamics (4.60) is represented in the frequency domain by
1
Y (s) = G(s)V (s) = V (s), (4.62)
s+1
where G(s) is the stable closed-loop transfer function.
Now consider the following situation: The output y(t) as well as its first time derivative ẏ(t)
are supposed to be kept zero for all times t. From the output equation we get the following
condition on x1 (t):
y(t) = 0 ⇒ x1 (t) = 0 ∀ t ≥ 0. (4.63)
From equations (4.60) and (4.61), the control input can be computed that keeps y(t) = 0 for
all t:
y(t) = 0, ẏ(t) = 0 ⇒ v(t) = 0, u(t) = − sin x2 (t). (4.64)
Restricted to these conditions, the differential equation for x2 is
ẋ2 (t) = x2 (t), (4.65)
representing unstable zero dynamics. This means that x2 grows unboundedly, although, the
output y(t) is kept at the constant value 0. The block diagram in Fig. 4.6 illustrates what
happens when I/O linearization is applied to system (4.58).
11
In controllable linear systems, unstable zero dynamics corresponds to unobservable eigenvalues in the right
complex half plane.

© Prof. Dr.-Ing. habil. Paul Kotyczka 59

Chair of Automatic Control
Technical University of Munich
4.5. Input-Output Linearization

Figure 4.6: Block diagram for the I/O linearization example

Comments on Figure 4.6

1. Note that it takes one integration from the input u to the output y (one differentiation in
opposite direction). This means, the first order time derivative of y is directly influenced
by u and we say the the output y has relative degree r = 1.

2. By choosing the control law u = −y − sin x2 + v, the first order input-output behaviour
ẏ = −y + v is imposed.

3. The dynamics of x2 disappears from the I/O behaviour. However, it is still present in
the system, and the I/O linearizing control law also depends on x2 . The differential
equation
ẋ2 = x1 + x2 (4.66)
represents the “hidden” or internal dynamics.
In case of internal instability,

• the internal states grow unboundedly, which on the one hand may cause failure or
damage of the system.
• On the other hand, the commanded control u, that in general depends on the
internal states, can go into saturation. If this occurs, the system can be considered
open-loop and I/O linearization fails.

To prove internal stability it is sufficient to show the stability of the zero dynamics, i.e.,
(r)
the internal dynamics, restricted to y = ẏ = . . . = y = 0 (in the example, r = 1).

60 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 4. CONTROLLABILITY AND FEEDBACK LINEARIZATION

4.5.2 General Procedure

The approach can be divided into three steps.

1. Relative degree

Determine the relative degree r according to Definition (4.10) by differentiating the output y
with respect to time, until the input u appears.

2. Desired dynamics

Choose asymptotically stable linear I/O behaviour of order r, represented by the r-th order
differential equation
(r) (r−1)
y + ar−1 y + · · · + a2 ÿ + a1 ẏ + a0 y = b0 v (4.67)
or the corresponding transfer function
Y (s) b0
= r r−1
. (4.68)
V (s) s + ar−1 s + · · · + a2 s 2 + a1 s + a0
The difference of the orders of the denominator and the numerator polynomial is the relative
degree r of the transfer function. For steady-state gain equal to 1, b0 = a0 must be chosen.
Comparison of (4.67) with (4.49), i.e.,
(r)
y = Lrf h(x) + Lg Lr−1
f h(x)u,

yields the following result.

Theorem 4.5 (Input-output linearization). The input-output linearizing control law

r−1
!
1 (i)
(4.69)
X
u= −Lrf h(x) + ai y + b0 v
Lg Lr−1
f h(x) i=0

imposes the linear closed-loop input-output behavior according to (4.67) to the SISO
system (4.54).

(r)
The simplest (yet unstable) transfer function from v to y is an integrator chain y = v, which
is obtained by setting b0 = 1, ai = 0, i = 0, . . . , r − 1.

3. Zero dynamics

Differentiating the output according to Step 1 defines the following (local) coordinate change
(diffeomorphism12 , if the vector field f (x) and the function h(x) are sufficiently smooth):
" #
ξ
z = t(x), z= , ξ ∈ Rr , η ∈ Rn−r (4.70)
η
12
An invertible (i.e., regular) coordinate transformation z = t(x) is a diffeomorphism, if the mapping t(x)
as well as its inverse t−1 (z) are differentiable.

© Prof. Dr.-Ing. habil. Paul Kotyczka 61

Chair of Automatic Control
Technical University of Munich
4.5. Input-Output Linearization

with
ξ1 = t1 (x) = h(x)
ξ2 = t2 (x) = Lf h(x)
.. (4.71)
.
ξr = tr (x) = Lr−1
f h(x).

To complete the diffeomorphism, n − r coordinate functions tr+1 (x), . . . , tn (x) are missing.
They can be chosen arbitrarily, with the only restriction that z = t(x) is invertible, i.e., the
inverse coordinate transformation x = t−1 (z) exists (and is also smooth):
η1 = tr+1 (x)
.. (4.72)
.
ηn−r = tn (x).

The closed-loop system, resulting from the control law (4.69) can be written in new coordinates:
ξ˙1 = ξ2



ξ˙2 = ξ3



.. ξ̇ = Aξ ξ + bξ v (4.73)
. 


ξ˙r = −a0 ξ1 − · · · − ar−1 ξr + b0 v




η̇1 = fη,1 (ξ, η)
..


. η̇ = f η (ξ, η) (4.74)

η̇n−r = fη,n−r (ξ, η)


Equation (4.73) is the state space representation of the asymptotically stable13 input-output
transfer function (4.68), equation (4.74) represents the internal dynamics which is excited by
the ξ-subsystem.
Setting all “external states” ξ1 = · · · = ξr = 0, the unforced zero dynamics
η̇ = f n (0, η) (4.75)
remains. If the equilibrium η ∗ of (4.75) is (locally) asymptotically stable, then (local) asymp-
totic stability of the overall I/O linearized system is guaranteed.
In practice, it is not always necessary to determine η(x) explicitely to obtain an expression for
the zero dynamics: Take the state differential equation
ẋ = f (x) + g(x)u. (4.76)
Substitute the state variables according to the zero output conditions
y = ξ1 (x) = 0
..
. (4.77)
(r−1)
y = ξr (x) = 0
and the output-zeroing control law, i.e., (4.69) with v = 0 and the zero output conditions.
The resulting set of differential equations will contain redundancies. The set of n−r remaining
independent differential equations describes the zero dynamics, see the exercise for an example
13
by choice of the coefficients a0 , . . . , ar−1

62 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 BIBLIOGRAPHY

Remark 4.3. The procedure to check stability of the zero dynamics holds in complete analogy,
if another constant (non-zero) output y(t) = const. for all t ≥ 0 is considered. Only v(t) has
to be changed to the corresponding constant stationary value.

4.6 References

The discussion on controllability is based on Section 3.1 of [1], where more details and the
proofs can be found. Input-output-linearization and input-state linearization are for example
treated in Chapter 6 of [8] and Chapter 12 of [7]. A nice slide set on nonlinear controllability
and geometric control is [9]14 , where particularly Chow’s Theorem is discussed. Chapter 10 of
[4] gives a very readable overview over geometric control in the context of robotics.

Bibliography

[1] H. Nijmeijer and A. van der Schaft. Nonlinear Dynamical Control Systems. Springer, 2017.

[2] W.-L. Chow. Über Systeme von linearen partiellen Differentialgleichungen erster Ordnung.
Mathematische Annalen, 117-117(1):98–105, dec 1940.

[3] P. K. Rashevskii. About connecting two points of complete non-holonomic space by ad-
missible curve (in Russian). Uch. Zapiski Ped. Inst. Libknechta, 2:83–94, 1938.

[4] M. W. Spong, S. Hutchinson, and M. Vidyasagar. Robot Modeling and Control. Wiley,
2006.

[5] R. M. Murray, S. S. Sastry, and L. Zexiang. A Mathematical Introduction to Robotic

Manipulation. CRC Press, Inc., 1994.

[6] B. Siciliano, L. Sciavicco, L. Villani, and G. Oriolo. Robotics: Modelling, Planning and
Control. Springer, 2010.

[7] H. Khalil. Nonlinear Systems. Prentice Hall, 3rd edition, 2002.

[8] J.-J. E. Slotine and W. Li. Applied Nonlinear Control, volume 199. Prentice Hall Englewood
Cliffs, NJ, 1991.

[9] L. Rifford. Geometric control and applications. Course at the Fields Institute, 2014.

14
https://math.unice.fr/~rifford/Papiers_en_ligne/COURSE_FIELDS_trans.pdf

© Prof. Dr.-Ing. habil. Paul Kotyczka 63

Chair of Automatic Control
Technical University of Munich
Bibliography

64 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
Chapter 5

Flatness Based Control

As a very useful tool for nonlinear control, we now introduce the differential flatness approach.
The flatness approach proves useful both in the derivation of highly accurate feedforward
control laws, and the design of nonlinear feedback controls.

5.1 Two Degrees of Freedom Structure

So far we only discussed the problem of feedback stabilization. However, modern control
systems, particularly if the control task is to follow or track a desired reference y d (t) of the
output, have a so-called two-degrees-of-freedom (2 DOF) structure, see Fig. 5.1.

Some comments on the 2 DOF structure:

• If a precise model of the system is known, it can be used in the feedforward block: A
nominal input ud (t) = uF F (t) is generated, which – in the ideal (nominal) case without
disturbances z(t) and with the knowledge of the initial value x(0) – would result in exact
tracking y(t) = y d (t).
• Knowing the initial state x(0) is an utopistic situation. The feedback controller needs
to achieve x(t) → xd (t) (and therefore y(t) → y d (t)). The feedback mechanism is also
necessary to cope with model inaccuracies.

Figure 5.1: Block diagram of a system in 2 DOF controller structure

65
5.2. Differential Flatness

• To cope with disturbances z(t) ̸= 0, additional measures have to be taken, like dis-
turbance feedback or integral action (the latter to compensate for piecewise constant
disturbances).

• The better the model fits reality,

∥uF B (t)∥
– the lower should be the ratio ∥uF F (t)∥ between the feedback and feedforward control
efforts1 , and consequently
– the simpler can the feedback controller be.

In many cases, the combination of a “sophisticated” feedforward controller (e.g., deter-

mined based on flatness) and a simple linear state feedback does an excellent job.

5.2 Differential Flatness

Definition 5.1 (Flat system). A system

ẋ = f (x, u) (5.1a)
y = h(x, u) (5.1b)

where x ∈ Rn , u, y ∈ Rm , is called (differentially) flat, if the states x = [x1 , . . . , xn ]T

and the inputs u = [u1 , . . . , um ]T can be uniquely expressed by a finite number of time
(k)
derivatives yi of the output y = [y1 , . . . , yn ]T :
(β1 ) (βm )
x = ψ x (y1 , . . . , y1 , . . . , ym , . . . , ym )
(β)
= ψ x (y, ẏ, . . . , y ), (5.2a)
(β1 +1) (βm +1)
u = ψ u (y1 , . . . , y1 , . . . , ym , . . . , ym )
(β+1)
= ψ u (y, ẏ, . . . , y ). (5.2b)

Equations (5.2a) and (5.2b) are called the differential parametrization of the system (5.1).

Remarks:

• If (5.1) is a flat system, then y according to (5.1b) is called a flat output.

• Flat outputs are not unique. There might be several flat outputs for (5.1a), which, e.g.,

– have a (direct) physical interpretation,

– allow for a simple differential parametrization (5.2),
– or, for example, depend only on the state x.

• The mathematical analysis whether there exist flat outputs for a given system of differ-
ential equations (5.1a) is a difficult task. A sufficient – but not necessary – condition
is that the system is input-state linearizable, see [1], Subsection 3.1.2 “Known Results”.
1
∥u(t)∥ denotes an appropriate norm on the signal u(t).

66 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 5. FLATNESS BASED CONTROL

Figure 5.2: The integrator chain from the example with n = 3.

The question of a constructive procedure to find flat outputs (or not) has been actively
discussed in research with different contributions for particular cases. To the best of my
knowledge a general procedure to determine whether a given system has a flat output
or not has not been reported, yet.
• In practice, flat outputs are often found in a rather heuristic way, exploiting the knowledge
of the system dynamics: Good candidates are outputs which are “far away” from the
input. If the output of a SISO system has full relative degree n, then the system can be
feedback-transformed into a linear integrator chain, and it is straightforward to find the
differential parametrization.

Example 5.1 (Integrator chain). Consider the integrator chain with n = 3 in Fig. 5.2. The
differential parametrization is obvious from the block diagram. The equations
x1 = y
x2 = ẏ
x3 = ÿ
(3)
u= y
represent the differential parametrizations of the state and the input. ◁

Once a flat output of a general MIMO system (and not a simple integrator chain) is found,
the differential parametrization is derived as follows:

• Differentiate the outputs until the resulting set of equations can be solved for the un-
knowns x and u.
• If on the way time derivatives of u appear, then they are also treated as unknowns. The
number of equations obtained from differentiation, which is necessary to solve for x and
u and the derivatives of u, increases correspondingly.

5.3 Flatness Based Feedforward Control

Given a differential parametrization (5.2) and a sufficiently smooth reference trajectory y d (t)
(such that all necessary derivatives exist), the nominal feedforward control for a system (5.1)
is immediately given by
(β+1)
ud (t) = ψ u (y d (t), ẏ d (t), . . . , y d (t)). (5.3)
For the feedback controller presented in the next subsection, also the nominal state trajectory
(β)
xd (t) = ψ x (y d (t), ẏ d (t), . . . , y d (t)) (5.4)
is required.

© Prof. Dr.-Ing. habil. Paul Kotyczka 67

Chair of Automatic Control
Technical University of Munich
5.4. Flatness Based Feedback Control

5.4 Flatness Based Feedback Control

The task of the feedback controller in a 2 DOF structure is to stabilize the dynamics of the
state error e(t) = x(t) − xd (t) at its equilibrium e∗ = 0. We present how this can be achieved
1. locally and 2. globally. Note that in the sequel, we will write the argument (t) only
for those quantities that have a time dependence, which comes explicitly from the desired
reference trajectory y d (t).

5.4.1 Linear State Feedback Based on Linearization

The error system

ė = ẋ − ẋd (t)
(5.5)
= f (x) + g(x)(ud (t) + uc ) − ẋd (t)

is linearized along the state trajectory x = xd (t). This yields in general a time-varying linear
dynamics
ė = A(t)e + b(t)uc . (5.6)
For this system, a time-varying or constant state feedback

uc = −R(t)e or uc = −Re (5.7)

can be designed to locally stabilize the equilibrium e∗ = 0

Remark 5.1. Note that to rigorously prove the equilibrium stability of a linear time-varying
system, it is not sufficient that the eigenvalues of the time-varying state matrix are in the
open left half plane for all times t. However, if the time-variance is sufficiently slow and the
eigenvalues are “far enough left”, a feedback controller based on this argument works.

5.4.2 Nonlinear State Feedback for SISO Systems in Normal Form

We consider a SISO system with full relative degree r = n. For such a system, the n-th time
derivative of the output is
(n)
y = Lnf h(x) + Lg Lfn−1 h(x) (ud (t) + uc ) . (5.8)
| {z } | {z }| {z }
a(x) b(x) u

The nominal input ud (t) is given from the feedforward design, as well as the nominal evolution
of the state vector xd (t). The highest order time derivative of the desired output clearly
satisfies the equation
(n)
yd (t) = a(xd (t)) + b(xd (t))ud (t). (5.9)
Subtracting (5.9) from (5.8), the n-th order differential equation for the output error ϵ =
y − yd (t) is obtained:
(n) (n)
y − yd (t) = a(x) − a(xd (t)) + (b(x) − b(xd (t)))ud (t) + b(x)uc . (5.10)
| {z }
(n)
ϵ

68 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 5. FLATNESS BASED CONTROL

Let
(n) (n) (n−1) (n−1)
y − yd (t) = −pn−1 ( y − yd (t)) − · · · − p1 (ẏ − y˙d (t)) −p0 (y − yd (t)), (5.11)
| {z } | {z } | {z } | {z }
(n) (n−1) ϵ̇ ϵ
ϵ ϵ
wherein
(n−1)
y = h(x), ẏ = Lf h(x), ... y = Lfn−1 h(x), (5.12)
be the desired trajectory error dynamics with p0 , . . . , pn−1 coefficients of a Hurwitz polynomial
sn + pn−1 sn−1 + · · · + p1 s + p0 . The control signal to (globally) asymptotically stabilize the
equilibrium 0 of the state error e = x − xd (t) follows from comparison of the right hand sides
of (5.10) and (5.11).

The nonlinear tracking control law then reads

n−1
!
1 X (i)
uc = a(xd (t)) − a(x) + (b(xd (t)) − b(x))ud (t) − pi ϵ (t) (5.13)
b(x) i=0

with
(i) (i)
ϵ = Lif h(x) − y d (t), i = 0, . . . , n − 1. (5.14)

Remark 5.2. We come to the same result, if we write the original dynamics (5.8) in the state
space normal form
ξ˙1 = ξ2
..
.
(5.15)
ξ˙n−1 = ξn
ξ˙n = a(t−1 (ξ)) + b(t−1 (ξ))(ud (t) + uc )
where ξ = t(x) with
(n−1)
ξ1 = y = h(x), ξ2 = ẏ = Lf h(x), ... ξn = y = Ln−1
f h(x), (5.16)
denotes the nonlinear transformation to normal form. Defining the error states
(i−1)
εi = ξi − y d (t), i = 1, . . . , n, (5.17)
we can compare the state differential equations (the last line)
ε̇1 = ε2
..
. (5.18)
ε̇n−1 = εn
ε̇n = a(x) + b(x)(ud (t) + uc ) − a(xd (t)) − b(xd (t))ud (t)
with the desired error dynamics in linear controller canonical form
ε̇1 = ε2
..
. (5.19)
ε̇n−1 = εn
ε̇n = −p0 ε1 − p1 ε2 − . . . − pn−1 εn .

© Prof. Dr.-Ing. habil. Paul Kotyczka 69

Chair of Automatic Control
Technical University of Munich
5.5. References

The result is again the tracking control law (5.13).

5.5 References

The present chapter gave only a very short introduction to the idea of flatness and its use for
nonlinear control. Much more can be found for example in the textbooks [2] (with a lot of
examples, for linear and nonlinear systems) and [3] (with a stronger emphasis on the differential
geometric perspective).

Bibliography

[1] Ph. Martin, R. M. Murray, and P. Rouchon. Flat systems. Technical report, 2002.

[2] H. Sira-Ramírez and S. K. Agrawal. Differentially Flat Systems. Marcel Dekker, Inc., 2004.

[3] J. Lévine. Analysis and Control of Nonlinear Systems. Springer-Verlag, 2009.

70 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
Chapter 6

Backstepping

Backstepping is a Lyapunov based design technique for nonlinear systems. It exploits the
structure of a system which can be split into two (or more) subsystems. In the simplest case
this splitting looks as follows:

η̇ = f 0 (η) + g 0 (η)ξ (6.1)

ξ˙ = u. (6.2)

The η-subsystem is fed by the output of the integrator ξ˙ = u, see Fig. 6.1. While deriving the
Backstepping controller for a more general state representation in the following subsection,
to illustrate the method graphically, the Figures 6.2 and 6.3 will refer to this basic case of
Integrator Backstepping.
The idea of backstepping is to start with the design of a ficticious controller ξf = ϕ(η) to
stabilize the equilibrium of the η-subsystem, understanding ξ as the input to this subsystem.
Then, the stabilizing controller for the overall system can be derived by construction. If there
is more than one integrator (or ξ-subsystem) “left” of the η-subsystem, the controller design
is performed recursively.
The big advantage of Backstepping is that, exploiting the system structure, the complexity of
the control design task is reduced, as only for the lower-dimensional η-subsystem creativity is
needed. Moreover, in contrast to linearizing techniques (as seen in Chapter 4), nonlinearities
are not cancelled by force – these cancellations are in general very sensitive with respect to
parameter uncertainties.

6.1 Derivation of the Backstepping Controller

We consider the following class of systems

η̇ = f 0 (η) + g 0 (η)ξ (6.3)

ξ˙ = f1 (η, ξ) + g1 (η, ξ)u, (6.4)

with η ∈ Rn and ξ ∈ R. Observe that the right hand side of the η-subsystem – except for
ξ which acts as the input – only depends on η, while the ξ-subsystem can depend on all the
states. Such a system structure (see in particular the next section) is called strict feedback
form.

71
6.1. Derivation of the Backstepping Controller

Figure 6.1: Simplest system structure for which Backstepping is applicable (f1 = 0, g1 = 1).

1. Ficticious control for the η-subsystem

In a first step, we determine – no matter by which method – a fictitious controller ξf = ϕ(η)

that, replaced for ξ in (6.3), stabilizes asymptotically the equilibrium η ∗ = 0 of this subsystem1 .
Moreover, a Lyapunov function V0 (η) for this subsystem – under the fictitious controller –
needs to be determined. V0 (η) must have a (local) minimum at η ∗ = 0 and satisfy
∂V0 (η)
(f 0 (η) + g 0 (η)ϕ(η)) < 0 for η ̸= 0. (6.5)
∂η

2. Reformulation of the system – input and state transformation

Exploiting the fact, that the ficticious controller asymptotically stabilizes the η-subsystem, we
reformulate it as follows2 :
η̇ = f 0 (η) + g 0 (η)ϕ(η) + g 0 (η) (ξ − ϕ(η)) . (6.6)
| {z } | {z }
Drift vector field :=z
of as. stable system

z = ξ − ϕ(η) plays now the role of the input to the asymptotically stabilized subsystem. The
situation for the case of f1 = 0, g1 = 1 is depicted in Fig. 6.2.
With the coordinate change ξ → z = ξ − ϕ(η), we obtain
η̇ = (f 0 (η) + g 0 (η)ϕ(η)) + g 0 (η)z (6.7)
ż = ξ˙ − ϕ̇(η)
= f1 (η, ξ) − ϕ̇(η) + g1 (η, ξ)u . (6.8)
| {z }
:=v

The resulting, transformed system now has an asymptotically stable η-subsystem (for z = 0).
With the input transformation u → v such that
v = f1 (η, ξ) − ϕ̇(η) + g1 (η, ξ)u, (6.9)
now again the z-subsystem can be represented as a simple integrator, see Fig. 6.3. In this
figure, also the name Backstepping is illustrated: The ficticious control ϕ(η) is stepped back
to the left of the integrator and now appears as ϕ̇(η).
1
η ∗ = 0 is chosen for simplicity and without loss of generality.
2
g 0 (η)ϕ(η) − g 0 (η)ϕ(η) = 0 is added to the right hand side.

72 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 6. BACKSTEPPING

Figure 6.2: With the help of the ficitcious control ϕ(η) an asymptotically stable η-subsystem
can be formulated (f1 = 0, g1 = 1). The new state variable z = ξ − ϕ(η) now appears as
input to the stabilized η-subsystem.

Figure 6.3: Illustration of the input transformation (f1 = 0, g1 = 1).

3. Construction of the overall stabilizing controller

With V0 (η), a Lyapunov function is known for the η-subsystem (when stabilized by the ficticious
control). The total system according to Fig. 6.3 consists in addition of a simple integrator
subsystem ż = v. For this simplest linear first order system, a simple quadratic function is an
appropriate energy-like function, hence we set up the following Lyapunov candidate function
for the total system:
1
V (η, z) = V0 (η) + z 2 . (6.10)
2
V (η, z) is obviously positive definite, as both parts are positive definite in their corresponding
arguments. Now we want to construct a control law (this time a real and not a ficticious one)

v = β(η, z), (6.11)

under which V is indeed a closed-loop Lyapunov function. In order to show this, the condition
" # " #
η 0
V̇ (η, z) < 0 for ̸= . (6.12)
z 0

must be satisfied. The time derivative of V can be written as follows, where in particular
the reformulated state representation of the system (6.7), (6.8) is used to replace the time

© Prof. Dr.-Ing. habil. Paul Kotyczka 73

Chair of Automatic Control
Technical University of Munich
6.1. Derivation of the Backstepping Controller

derivatives η̇ and ż:

∂V (η, z) ∂V (η, z)
V̇ (η, z) = η̇ + ż
∂η ∂z
∂V0 (η, z) ∂V0 (η, z) ! (6.13)
= (f 0 (η) + g 0 (η)ϕ(η)) + g0 (η)z + zv < 0.
∂η ∂η
| {z } | {z }
<0 for η̸=0, see (6.5) !
=−kz 2 , k>0

The first underbraced term is negative definite in η, as it describes the evolution of V0 in the
ficticiously controlled η-subsystem. In order to make the whole expression negative definite,
the second term is made negative definite by requiring it to be a negative quadratic function
in z. The latter leads directly to the following expression for the transformed control input v:

∂V0 (η)
v=− g0 (η) − kz. (6.14)
∂η

Figure 6.4: Illustration of the overall stabilizing control law v = β(η, z) (f1 = 0, g1 = 1).

Expressing now the original input by inversion of (6.9),

1
u= (v + ϕ̇(η) − f1 (η, ξ)), (6.15)
g1 (η, ξ)

and replacing z = ξ − ϕ(η) where necessary, finally results in the Backstepping control law:

Theorem 6.1 (Backstepping). The Backstepping controller

!
1 ∂ϕ ∂V0
u= (f + g 0 ξ) − g0 − k(ξ − ϕ) − f1 (6.16)
g1 ∂η 0 ∂η
" # " #
η∗ 0
stabilizes the equilibrium ∗ = of (6.3), (6.4) asymptotically.
ξ 0

74 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 6. BACKSTEPPING

6.2 Summary

To sum up, Backstepping control design consists of three steps:

1. Find a stabilizing fictitious input ϕ(η) for the η-subsystem and the corresponding Lya-
punov function V0 (η).
2. Transform the state and the input in an appropriate way to obtain an asymptotically
stable η-subsystem (ξ → z, u → v).
3. Set up a total Lyapunov function V (η, z) = V0 (η)+ 12 z 2 and construct the Backstepping
controller by demanding V̇ < −kz 2 . Transform the condition back to original states and
input (z → ξ, v → u) to obtain the Backstepping control law.

Remark 6.1. A popular error which must be avoided (in the exercise or the exam), is to simply
replace ξ by ϕ(η) in the right hand side of η̇ during the derivation of the Backstepping control
law, in particular in Eq. (6.13). This is wrong. It is correct to replace η̇ by the reformulated
version (6.6) or (6.7) of the state differential equation.

Some features/advantages of Backstepping:

• The approach does not rely on linearization by force which is not very robust to parameter
uncertainties.
• The difficult/creative control design part is reduced in its complexity: Rather than to
design a controller for the overall system, a fictitious controller for the lower dimensional
η-subsystem must be derived. The rest of the design procedure follows by construction.
• The presented approach can be extended iteratively to systems in a strict feedback form:
η̇ = f0 (η) + g0 (η)ξ1
ξ˙1 = f1 (η, ξ1 ) + g1 (η, ξ1 )ξ2
.. (6.17)
.
˙ξk = fk (η, ξ1 , . . . , ξk ) + gk (η, ξ1 , . . . , ξk )u.

• For systems in strict forward form, there exists a complementary technique: Forwarding.

6.3 References

A textbook which deals with different Lyapunov and passivity based control design techniques
for nonlinear systems is [1]. Backstepping and forwarding are addressed in Chapter 6. See also
Section 14.3 of [2], which is devoted to Backstepping. “Backstepping” has also become a very
popular approach for the boundary control and observation of distributed parameter systems,
see e.g., the course slides [3]3 for an overview.
3
http://www.eeci-institute.eu/GSC2012/Photos-EECI/EECI-GSC-2012-M7/
Krstic-slides-Supelec-2012.pdf

© Prof. Dr.-Ing. habil. Paul Kotyczka 75

Chair of Automatic Control
Technical University of Munich
Bibliography

Bibliography

[1] R. Sepulchre, M. Janković, and P. V. Kokotović. Constructive Nonlinear Control. Springer-

Verlag London, 1997.

[2] H. Khalil. Nonlinear Systems. Prentice Hall, 3rd edition, 2002.

[3] M. Krstic. Boundary control of pdes: A course on backstepping designs. Course at EECI,
2012.

76 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
Chapter 7

The Passivity Approach

This chapter guides you from the definitions of dissipative and passive systems and their prop-
erties over the Invariance Principle to prove asymptotic stability to passivity-based control
using the so-called IDA-PBC approach, which exploits the port-Hamiltonian system represen-
tation. If you want to learn more about energy-based modeling and order reduction, you are
invited to follow our course Modeling and Reduction of Complex Systems in summer semester.
The last two Sections 7.6 and 7.7 are marked with a star, which means that they are not
taught this year, and therefore they are not relevant for the exam (yet interesting).

7.1 Dissipativity and Passivity

The existence of certain (potential, energy, storage, . . .) scalar functions, whose rate of change
is limited by the supply from outside the system, is the characteristic feature of dissipative
systems. If the supply rate has the form of physical power (in the sense that it is a product
of conjugated variables), the system is called passive. We now give the formal definitions of
dissipativity and passivity for the following general class of systems:
ẋ(t) = f (x(t), u(t)) (7.1a)
y(t) = h(x(t), u(t)) (7.1b)
with x ∈ Rn , u, y∈ Rm ,
equilibrium values1
= 0 andx∗ = 0 of states and in-
u∗
puts, i.e., f (0, 0) = 0. Moreover, the output at the equilibrium is supposed to be zero:
y ∗ = h(0, 0) = 0.

Definition 7.1 (Dissipativity). The system (7.1) is called dissipative on X ⊆ Rn if there ex-
ists a locally integrable2 supply rate s(u, y) and a positive semi-definite3 storage function
H(x) such that the dissipation inequality
Z t
H(x(t)) − H(x(t0 )) ≤ s(u(τ ), y(τ )) dτ (7.2)
t0

holds for all admissible inputs u ∈ U ⊂ Rm and for all t0 , t such that t0 ≤ t. (The state
x(τ ) is supposed to remain in X for all τ ∈ [t0 , t]).
1
The zero values of states and in-/outputs at the equilibrium are chosen for simplicity of presentation. By
a simple linear coordinate shift, all statements can be accordingly reformulated for the nonzero case.

77
7.1. Dissipativity and Passivity

The supply rate s(u, y) is an arbitrary real valued function. There are several choices for
s(u, y) which are interesting and useful in the context of system theory (also for linear systems),
e.g.,
1 1
s(u, y) = uT u + y T y. (7.3)
2 2
By the fact that the dissipation inequality (7.2) must hold for all t0 and t, we can deduce that
the (total) integral dissipation
Z t
D(t, t0 ) := s(u(τ ), y(τ )) dτ − (H(x(t)) − H(x(t0 ))) (7.4)
t0

must monotonically increase in t for arbitrary (given) t0 . Therefore, Ḋ(t, t0 ) ≥ 0, and accord-
ingly from
d t
Z
0≤ s(u(τ ), y(τ )) dτ − Ḣ(x(t)) (7.5)
dt t0
we obtain the dissipation inequality (7.2) in its differential form:

Ḣ(x(t)) ≤ s(u(t), y(t)) ∀t ≥ 0. (7.6)

The supply rate

m
(7.7)
X
s(u, y) = y T u = yi ui
i=1

is of particular importance in physical modeling and energy-based control. If the pairs (ui , yi )
consist of conjugate power variables (like current and voltage or force and velocity), then y T u
represents the power which is supplied to the system through the power ports (u1 , y1 ), . . . , (um , ym ).

Definition 7.2 (Passivity). The system (7.1) is passive if it is dissipative with a supply rate
s(u, y) = y T u.

Observe that the definition of passivity (dissipativity) connects the evolution of the states to
inputs and outputs, by the passivity (dissipativity) inequality.

The increase of energy stored in a passive system is bounded from above by the power
supplied through the system boundaries.

From the passivity inequality Ḣ(x) ≤ y T u, the following immediately follows:

Theorem 7.1 (Stability of a passive system). An unforced passive system (u ≡ 0) with a

positive definite storage function H(x) has a Lyapunov stable equilibrium at the (local)
minimum x∗ of H(x).

Proof. For u ≡ 0, the passivity inequality becomes Ḣ(x) ≤ 0, which together with positive
definiteness of H(x) shows (local) Lyapunov stability of x∗ = arg minx H(x). □
2
The integral over every compact domain has a finite value.
3
Remember that we call a function H(x), with a slight extension of the mathematical definition, positive
definite or semi-definite, if it has a (local) minimum at x∗ and H(x) > H(x∗ ) or H(x) ≥ H(x∗ ) if x ̸= x∗ .

78 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 7. THE PASSIVITY APPROACH

Roughly speaking, passive systems in physical modeling can be characterized by the lack of
internal energy sources. Passive systems have nice properties which can be exploited for
analysis and controller design, some of which shall be mentioned in the next section.
If we choose the physical energy H(x) as a Lyapunov candidate function to prove stability of
the equilibrium x∗ of an autonomous system

ẋ(t) = f (x(t)), (7.8)

quite often we can only show

Ḣ(x(t)) ≤ 0. (7.9)
This is in particular the case for systems with a mechanical structure, where only the differential
equations for the velocities/momenta are affected by damping, see e.g., the exercise on the
pendulum with friction.
To prove asymptotic stability of the equilibrium, one possibility is to look for an alternative,
slightly modified Lyapunov function (see the exercise). However, this can be quite cumbersome.
A direct way to prove asymptotic stability of an equilibrium, or asymptotic convergence to a
certain limit set, is the application of the Invariance Principle of Krasovskii / LaSalle.

7.2 Invariance Principle

Theorem 7.2 (LaSalle, 1960). Given the system

ẋ = f (x), x ∈ Rn . (7.10)

• Let Ω ⊂ Rn be a compact set, which is positively invariant with respect to (7.10).

• Let V (x) be a continuously differentiable function on Ω which is non-increasing,

i.e., V̇ (x) ≤ 0 on Ω.

• Let E = {x ∈ Ω | V̇ (x) = 0} bet the subset of Ω where V̇ (x) is zero, and M the
largest invariant subset in E.

• Then every solution x(t) which starts in Ω tends to M for t → ∞.

Let us clarify the following terms which appear in the formulation of the Invariance Principle:

Compact. A set which is closed (i.e., it contains its boundary) and bounded is called compact.

Positively invariant. A set Ω is called positively invariant (w.r.t. some dynamics) if the
following holds:
x(t0 ) ∈ Ω ⇒ x(t) ∈ Ω ∀ t ≥ t0 , (7.11)
i.e., the trajectory x(t) must remain in Ω if it starts in Ω.

Invariance. In contrast, a set Ω is called invariant if

x(t0 ) ∈ Ω ⇒ x(t) ∈ Ω ∀ t ∈ R. (7.12)

Every trajectory which passes through a point x(t0 ) in Ω must remain in Ω for all t ≥ t0
and must have been in Ω for all t < t0 .

© Prof. Dr.-Ing. habil. Paul Kotyczka 79

Chair of Automatic Control
Technical University of Munich
7.2. Invariance Principle

Figure 7.1: Illustration of the sets used in the Invariance Principle. In the example, x∗ is an
unstable equilibrium, and the limit set of the system is the depicted limit cycle. (The dashed
green line shall coincide with the blue line – their distance is for better visibility in the sketch.)

Example: (Positive) invariance can be illustrated with the help of the simple: first order dynamics
ẋ = −ax, a > 0. (7.13)

Every set {x, |x| ≤ c} is positively invariant as the trajectories tend asymptotically to zero without
ẋ(t) changing sign. The equilibrium x = 0 in contrast is the only invariant set. It is certainly positive
invariant as ẋ|x=0 = 0. Moreover, x = 0 is only approached asymptotically by any other solution. Hence,
any solution which at some time is in 0 must have been there for all past times (negative invariance).
Together with positive invariance, the invariance of x = 0 is shown.

Largest invariant subset. As largest invariant subset we understand the union of all invariant
subsets.

With the help of the Invariance Principle, the limit set (the set to which all trajectories
converge) of a dynamical system (7.8) can be determined.
If one is only interested in the stability property of an equilibrium, the following theorem is
easier to apply:

Theorem 7.3 (Barbashin, 1952). Given the system (7.10) with an equilibrium at x∗ .

• Let V (x) be a positive definite, continuously differentiable function on a neighbour-

hood Ω of x∗ , which is non-increasing: V̇ (x) ≤ 0.

• Let E = {x ∈ Ω | V̇ (x) = 0}.

• If x∗ is the only solution of (7.10) which remains in E, then x∗ is (locally) asymp-

totically stable.

Note that in this theorem, the function V (x) has to be positive definite. However, only positive
invariance of x∗ needs to be checked.
Sketch of the proof :

1. From positive definiteness of V , i.e., V > V ∗ , x ̸= x∗ , and the fact that V is non-
increasing, V̇ ≤ 0, it follows that for any initial state x0 (t0 ), the solutions x(t), t > t0
stay in the level set {x | V (x) ≤ V (x0 ) = c0 } (Lyapunov stability).

80 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 7. THE PASSIVITY APPROACH

2. If no point other than x∗ can stay in {x | V̇ (x) = 0}:

a) There are no limit cycles on which ∂V

∂x f (x) ≡ 0, f (x) ̸= 0.
(On a limit cycle, V̇ = 0 must hold, when V̇ ≤ 0 is given. Otherwise, a given state with a certain
value of V could not be reached a second time/multiple times by the trajectory.)
b) There are no other equilibria with ∂V
∂x f (x) ≡ 0, f (x) = 0.

3. This means, that after some time t1 > t0 , the value of V must have decreased, ∃ t1 > t0 ,
V (x(t1 )) < V (x(t0 )). Then the trajectory x(t), t > t1 stays in the level set {x | V (x) ≤
V (x, (t1 )) = c1 }, with c1 < c0 .

4. With the same arguments as in Step 2 and 3, it can be shown that there exist also
times t2 , t3 , . . . with t0 < t1 < t2 < t3 < . . . and a decreasing series of level sets
with c0 > c1 > c2 > c3 > . . . which asymptotically encircle the equilibrium x∗ to show
limt→∞ x(t) = x∗ . □

7.3 Example: Passivity of Mechanical Multi-Body Systems

We consider simple mechanical systems4 , whose Hamiltonian, i.e., the sum of kinetic and po-
tential energy, serves as a storage function to show their passivity. The second order equations
of motion
M (q)q̈ + C(q, q̇)q̇ + g(q) = f e (7.14)
with symmetric, positive definite mass matrix M (q), the Coriolis matrix C(q, q̇) and the
gradient g(q) = ∇V (q) of the potential energy follow from the Euler-Lagrange equations

d ∂L(q, q̇) ∂L(q, q̇)

 
− = fie , i = 1, . . . , n (7.15)
dt ∂ q̇i ∂qi
for each of the n system degrees of freedom. qi and q̇i are the respective generalized coordinates
and momenta, while fie are external forces/torques. The Lagrangian function
1
L(q, q̇) = q̇ T M (q)q̇ − V (q)
2
is the difference of kinetic energy5 and potential energy. The generalized momenta
T
∂L(q, q̇)

p= = M (q)q̇
∂ q̇
are defined via the Legendre transformation, and the Hamiltonian (total energy) is
1
H(q, p) = pT q̇ − L(q, q̇) = pT M −1 (q)p + V (q).
2
The Legendre transformation is symmetric, therefore also
T
∂H(q, p)

q̇ =
∂p
4
In a simple mechanical system, the total energy is the sum of kinetic and potential energy.
5
To be precise: co-energy. However, this only plays a role at relativistic velocities.

© Prof. Dr.-Ing. habil. Paul Kotyczka 81

Chair of Automatic Control
Technical University of Munich
7.4. Interconnections of Passive Systems

holds. The Euler-Lagrange equations can then be written in the alternative Hamiltonian
representation
q̇ = ∇p H(q, p)
ṗ = −∇q H(q, p) + f e
or in matrix form " # " #" # " #
q̇ 0 I ∇q H(q, p) 0 e
= + f . (7.17)
ṗ −I 0 ∇p H(q, p) I
We denote with ∇q H and ∇p H the column vectors of partial derivatives of the energy with
respect to q and p. The collocated outputs, with respect to the external generalized forces,
are the velocities " #
h i ∇ H(q, p)
q
q̇ = 0 I . (7.18)
∇p H(q, p)
Equations (7.17) and (7.18) build the Hamiltonian state space representation of mechanical
systems, which is dual to the Lagrangian representation (7.14). The passivity of mechanical
systems (or conservativeness in the absence of friction) becomes evident from the balance
equation
Ḣ = (∇q H)T q̇ + (∇p H)T ṗ =
= (∇q H)T ∇p H + (∇p H)T (−∇q H) + (∇p H)T f e =
| {z }
q̇ T

= q̇ T f e .
Splitting the generalized forces into a part u corresponding to control inputs, and velocity-
proportional6 friction forces/torques Rq̇,
f e = u − Rq̇, R = RT ≥ 0,
then Eq. (7.17) becomes
" # " #" # " #
q̇ 0 I ∇q H(q, p) 0
= + u. (7.19)
ṗ −I −R ∇p H(q, p) I
In this case, we obtain the differential energy balance
Ḣ = q̇ T u − q̇ T Rq̇ ≤ y T u
with the definition of the output vector
y = q̇.

7.4 Interconnections of Passive Systems

Passivity is a property which is particularly favorable for the modeling and analysis of large
scale systems, like circuits and networks. In this section, we first consider two typical intercon-
nections of passive systems and check passivity of the resulting system in both cases. Feedback
and parallel interconnection, see Fig. 7.2, are prototypical examples of power-preserving in-
terconnections.
We will also examine what happens if the output of a passive system is fed back, and we obtain
the simplest possible passivity-based (output) controller.
6
Of course, this is a simplifying assumption if we think of the different types of existing friction models

82 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 7. THE PASSIVITY APPROACH

Figure 7.2: Parallel and feedback interconnection of two (passive) systems

7.4.1 Power-Preserving Interconnections

Let us first consider two systems Σ1 and Σ2 with their corresponding state, input and output
vectors x1,2 , u1,2 , and y 1,2 . Both systems obey a passivity inequality, where H1,2 denotes
their energy/storage function:

Σ1 : H1 (x1 ) = y T1 u1 − d1 (x1 ) ≤ y T1 u1
(7.20)
Σ2 : H2 (x2 ) = y T2 u2 − d2 (x2 ) ≤ y T2 u2 .

d1,2 (x1,2 ) ≥ 0 are the non-negative dissipation terms. In the interconnected system, we denote
the composed state vector and the total energy function
" #
x
x= 1 , H(x) = H1 (x1 ) + H2 (x2 ).
x2

Parallel interconnection In this case, the in- and output vectors are related according to

u = u1 = u2
y = y1 + y2.

The energy balance for the composed system is

Ḣ(x) = Ḣ1 (x1 ) + Ḣ2 (x2 ) =

= y T1 u1 + y T2 u2 −d1 (x1 ) − d2 (x2 )
| {z }
≤0

≤ (y T1 + y T2 ) u
| {z }
yT

⇒ Ḣ(x) ≤ y T u

Feedback interconnection The interconnection conditions here are

u − y 2 = u1
y = y 1 = u2 ,

© Prof. Dr.-Ing. habil. Paul Kotyczka 83

Chair of Automatic Control
Technical University of Munich
7.4. Interconnections of Passive Systems

and the energy balance becomes

Ḣ(x) = y T1 u1 + y T2 u2 −d1 (x1 ) − d2 (x2 )

| {z }
≤0

≤ y T1 (u − y 2 ) + y T2 u2
|{z} |{z}
yT y

= y T u −y T y 2 + y T2 y
| {z }
=0
⇒ Ḣ(x) ≤ y T u.

Observe that in both (special) cases, we can write

y T u = y T1 u1 + y T2 u2 , (7.21)

i.e., the power supplied to the subsystems through the ports (u1 , y 1 ), (u2 , y 2 ) equals the power
supplied to the interconnected system via the port (u, y).

Multiple subsystems An interconnection of N subsystems is power-preserving if

N
(7.22)
X
yT u = y Ti ui ,
i=1

where (ui , y i ), i = 1, . . . , N , are the power ports of the subsystems and (u, y) is the power
port for the interconnected system.

Systems which result from a power-preserving interconnection of passive systems are pas-
sive as well.

7.4.2 Output Feedback

This is the simplest form of a feedback interconnection. The system Σ is interconnected with
a static relation as depicted in Fig. 7.3. Via output feedback – under a certain condition –
the equilibrium of a passive system can be asymptotically stabilized.

Theorem 7.4 (Asymptotic stabilization by output feedback). Given a passive system with
positive definite storage function H(x). Let the system be (locally) zero state detectable.
Then every output feedback
u = −ϕ(y), (7.23)
with ϕ(0) = 0, y T ϕ(y) > 0 for y ̸= 0, stabilizes the equilibrium x∗ = 0 (locally)
asymptotically.

Let’s clarify the term zero state detectability (ZSD), starting with zero state observability
(ZSO):

84 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 7. THE PASSIVITY APPROACH

Figure 7.3: Output feedback, feedback function in the first and third quadrant

ZSO A system is zero state observable if a zero output implies a zero state7 :

y(t) = 0 ⇒ x(t) = 0.

ZSD A system is zero state detectable if a constant zero output implies the asymptotic
convergence of the state to zero:

y(t) = 0 ⇒ lim x(t) = 0.

t→∞

ZSD is hence a weaker notion than ZSO.

Sketch of the proof: Passivity inequality with u = −ϕ(y) and a positive definite storage
function H(x) (with minimum in x∗ and H(x∗ ) = 0):

Ḣ(x) ≤ −y T ϕ(t) ≤ 0.

The equilibrium x∗ of the system is Lyapunov stable, the state remains within closed contour
surfaces of H(x). The state will tend to a limit set {x̄ | Ḣ(x̄) = 0}. Assume H(x̄) = a ≥ 0.
Remaining on this set from a time t′ on means that
Z t
0 = H(x̄(t)) − H(x̄(t′ )) ≤ − y T (τ )ϕ(y(τ ))dτ ≤ 0 (7.24)
t′

for all t > t′ . From the zero on the left hand side, it follows that y(t) = 0 for t > t′ . From
zero state detectability finally lim x(t) = 0 follows and also lim H(x(t)) = 0. □
t→∞ t→∞

7.5 Passivity-Based State Feedback Control

Our goal in this section is to find a way how to determine state feedback controllers which
render a given system passive. In particular, we want to bring the system by state feedback
into a special, physically inspired form which is passive by construction.

7.5.1 Port-Hamiltonian Systems

We consider a particular class of8 explicit port-Hamiltonian state space models.

7
Again, by a coordinate shift, this notion can be extended to nonzero (equilibrium) states and outputs.
8
Generalizations include implicit systems and more general definitions of input/output port variables.

© Prof. Dr.-Ing. habil. Paul Kotyczka 85

Chair of Automatic Control
Technical University of Munich
7.5. Passivity-Based State Feedback Control

Definition 7.3 (Port-Hamiltonian system). A system with a state space representation

(x ∈ Rn , u, y ∈ Rm )

ẋ = (J (x) − R(x))∇H(x) + G(x)u

(7.25)
y = GT (x)∇H(x),

with the subsequently listed properties, is called a port-Hamiltonian system.

• H(x) is a positive definite energy function, i.e., it has a minimum at the (desired)
equilibrium x∗ :
arg min H(x) = x∗ .
x

In general H(x) is at least C2 (twice continuously differentiable).

• J (x) = −J T (x) is a skew-symmetric interconnection or structure matrix, which

gives an idea about the structure of the internal energy flows.

• R(x) = RT (x) ≥ 0 is a symmetric, at least positive semi-definite, dissipation or

damping matrix which captures dissipative effects.

The most important property of a port-Hamiltonian system (7.25) is that the following energy
balance equation follows directly from the structure of the equations:

∂H
Ḣ(x) = ẋ
∂x
= (∇H(x))T (J (x) − R(x))∇H(x) + (∇H(x))T G(x) u
| {z }
yT (7.26)
T T T
= (∇H(x)) J (x)∇H(x) − (∇H(x)) R(x)∇H(x) + y u
| {z } | {z }
=0 ≥0

≤ y T u.

Hence, a port-Hamiltonian system with a positive definite energy H(x) and R(x) ≥ 0 is
passive, and y = GT (x)∇H(x) is the so-called collocated passive or power-conjugated output.
The port-Hamiltonian state representation unites ideas from geometric mechanics and network
theory and proves particularly useful in modeling and nonlinear control.

Some remarks:

• The port-Hamiltonian (PH) state representation can be understood as a generalization

of the Hamiltonian representation of mechanical systems:
" # " #" # " #
q̇ 0 I ∇q H(q, p) 0
= + u, (7.27)
ṗ −I −R ∇p H(q, p) G

where q ∈ Rn are the configuration coordinates and p ∈ Rn the generalized momenta.

The total energy (Hamiltonian) is given by the sum of kinetic and potential energy
1
H(q, p) = pT M −1 (q)p + V (q), M (q) = M T (q) > 0. (7.28)
2

86 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 7. THE PASSIVITY APPROACH

u ∈ Rm contains the available generalized forces acting on the corresponding mechanical

degrees of freedom over the matrix G. If m = n, the system is called fully actuated, if
m < n, it is an underactuated mechanical system.

• In the mechanical system, the skew-symmetric part of the matrix in (7.28) represents the
continuous exchange between potential and kinetic energy. The structure matrix J (x)
is a generalization of this and captures the exchange between different energy forms in
an interconnected system.
Note, however, that the exchange between different energy forms can be also coded by
a coupling in the energy function, and then not be visible in J (x). An example is the
magnetic floating ball, where the inductance, which appears in the expression for the
magnetic energy, depends on the distance.

• Power-preserving interconnections of PH systems are again PH systems. As a conse-

quence, large scale models of networks of subsystems can be iteratively constructed by
interconnection. The power preserving interconnection of different subsystems within
a network can be formulated on a mathematically abstract level with the concept of a
so-called Dirac structure.

• The definition of the passive output of a port-Hamiltonian system is not unique. Besides
the canonical passive output y = GT (x)∇H(x), there exist other passive outputs.

Show that y = GT (x)∇H(x) + D(x)u with D T (x) = −D(x) is also a passive

output of the port-Hamiltonian system (7.25).

© Prof. Dr.-Ing. habil. Paul Kotyczka 87

Chair of Automatic Control
Technical University of Munich
7.5. Passivity-Based State Feedback Control

7.5.2 Interconnection and Damping Assignment Passivity Based Control, IDA-

PBC

The simple idea of IDA-PBC is the following:

For the given system

ẋ = f (x) + G(x)u (7.29)
design a state feedback
u = α(x) + v (7.30)
such that the closed-loop system has port-Hamiltonian structure

ẋ = (J d (x) − Rd (x)) ∇Hd (x) + G(x)v, (7.31)

| {z }
=F d (x)

where J Td (x) = −J d (x), and the following definiteness conditions hold:

x∗ = arg min Hd (x), (7.32a)

x
Rd (x) = RTd (x) ≥ 0 (7.32b)

(the latter at least in the neighborhood of x∗ ).

In the rare case that the number n of states equals the number m of inputs and the square
input matrix G(x) is invertible, the matching equation

f (x) + G(x)u = F d (x)∇Hd (x) + G(x)v (7.33)

can be directly solved for u:

u = G−1 (x)(F d (x)∇Hd (x) − f (x)) + v. (7.34)

In this case the designer can choose the design matrix F d (x) = J d (x) − Rd (x) and the
closed-loop desired energy function Hd (x) completely freely. However, m < n is the usual
case, which comes along with restrictions on the possible/admissible design parameters.
Basically, the differential equations (7.29) may not be changed in directions in state space
which are not affected by the control input.

Definition 7.4 (Left hand annihilator). A matrix G⊥ (x) ∈ R(n−m)×n of rank n−m, which
satisfies G⊥ (x)G(x) = 0 for a given input matrix G(x) ∈ Rn×m , is called (full rank) left
hand annihilator of G(x).

Multiplying the matching equation (7.33) with G⊥ (x) eliminates u and v and we get the
so-called projected matching equation or matching PDE

G⊥ (x)f (x) = G⊥ (x)F d (x)∇Hd (x). (7.35)

The vector valued matching PDE consists of n − m scalar first order linear partial differential
equations of the type

bi (x) = aTi (x)∇Hd (x), i = 1, . . . , n − m, (7.36)

88 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 7. THE PASSIVITY APPROACH

bi (x) ∈ R, aTi (x) ∈ R1×n , which must be satisfied simultaneously under the definiteness
conditions (7.32b) and (7.32a). This is the main difficulty in the application of IDA-PBC.
There are two different approaches to attack the problem, depending on which of the free design
quantities are prespecified at the beginning. Both approaches contain as a main argument the
suitable choice or modification of the closed-loop energy (energy shaping), hence, IDA-PBC
is a typical representative of energy-based control schemes.

A Non-algebraic approach

At the beginning of the design process, desired interconnection and damping matrices are
chosen. The design steps are:

1. Specify the design matrix F d (x) = J d (x) − Rd (x).

2. Solve the set of n − m scalar matching PDEs (7.35) for the closed-loop energy Hd (x).

In order to achieve a minimum of Hd (x) in the second step, it is exploited that a solution of
(7.35) is composed of two parts Ψ(x) and Φ(ξ):

Hd (x) = Ψ(x) + Φ(ξ). (7.37)

While the particular solution Ψ(x) solves the inhomogeneous PDE (7.35), the homogeneous
version of the PDE
0 = G⊥ (x)F d (x)∇ξi (x) (7.38)
has a number (in general m) of solutions ξi (x), which are summarized in the vector of char-
actistic coordinates ξ ∈ Rm . Arbitrary functions Φ(ξ) of the characteristic coordinates can be
added to the particular solution without changing its validity.
Besides the solution of the vector-valued matching PDE itself, the main difficulty is to choose
design matrices F d (x) such that the inhomogeneous PDE is indeed solvable. This means that
in Step 1, the admissible parameterizations of F d (x) have to be found based on a solvability
test of the resulting matching PDE. A solvability test, which is based on the requirement that
the solutions must have permutable second order partial derivatives (Poincaré’s Lemma or
Schwarz’ Theorem)
∂ 2 Hd (x) ∂ 2 Hd (x)
= , i, j = 1, . . . , n, (7.39)
∂xi ∂xj ∂xj ∂xi
can be rather easily derived if the coefficient matrix G⊥ (x)F d (x) of the matching PDE is
constant.
Another difficulty is the large number (of order n2 ) of available and interdepending design
quantities. However, different design steps, like the solvability check, the solution of the PDE,
the definiteness check of Rd (x) can be to a certain extent performed automatically with a
computer algebra system.

B Algebraic approach

In the algebraic approach, to solve the projected matching equation, the following design steps
are performed:

© Prof. Dr.-Ing. habil. Paul Kotyczka 89

Chair of Automatic Control
Technical University of Munich
7.6. Passivity of State Space Models∗

1. Prespecify a desired, positive definite closed-loop energy function Hd (x).

2. Solve the projected matching equation (7.35) for the unknown elements of F d (x) such
that
1
Rd (x) = − (F d (x) + F Td (x)) ≥ 0 (7.40)
2
in a neighborhood of x∗ .

The advantage of this approach is that the projected matching equation is not a PDE but a
rather easy to solve algebraic equation.
The disadvantage is that it depends on the choice of the closed-loop energy Hd (x) whether
the definiteness condition (7.40) can be satisfied or not. The reason is that Hd (x) defines
the “canonical” passive (or collocated) output y = GT (x)∇Hd (x) of the desired closed-loop
system. Then the problem can be formulated as follows:
Find a state feedback such that the system (original system + new output)

ẋ = f (x) + G(x)u
(7.41)
y = GT (x)∇Hd (x)

becomes the passive closed-loop port-Hamiltonian system

ẋ = (J d (x) − Rd (x))∇Hd (x) + G(x)v

(7.42)
y = GT (x)∇Hd (x).

This question on feedback equivalence to passive systems is addressed in Section 7.7. Feed-
back passivity of a system requires relative degree {1, 1, . . . , 1} of the output and stable zero
dynamics. If one of these conditions is violated for system (7.41) with prespecified output
y = GT (x)∇Hd (x), then the system (7.42), which results from state feedback, cannot be
passive. In such a case, the projected matching equation (7.35) can not be solved such that
the dissipation matrix is positive semi-definite: Rd (x) ≱ 0.

7.6 Passivity of State Space Models∗

In this section, first we formulate conditions for passivity of a given nonlinear input-affine system
and once more recognize the strong relation between passivity and stability of a system. We
already know that a positive definite storage function H(x) of a passive system serves as a
Lyapunov function for the stable equilibrium at x∗ = arg minx H(x), due to Ḣ ≤ 0 for u = 0.
Consider the system

ẋ = f (x) + G(x)u
(7.43)
y = h(x).

The question at hand is: “What are the conditions for passivity of the system (7.43)?” In
other words, using the definition of passivity: “What are the conditions for the existence
of a positive (semi-)definite storage function H(x) which satisfies the differential passivity
inequality Ḣ(x) ≤ y T u?”

90 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 7. THE PASSIVITY APPROACH

The answer follows from expressing the balance equation for the unknown storage function
and replacing the dynamics of (7.43):

∂H
Ḣ = ẋ
∂x
∂H ∂H
= f (x) + G(x) u
∂x
| {z } |∂x {z } (7.44)
! !
≤0 = yT
!
≤ y T u.

The required boundedness of Ḣ(x) can only be guaranteed for arbitrary inputs u if the two
underbraced terms meet the indicated (in-)equalities:

Theorem 7.5 (Passivity of a nonlinear system). System (7.43) is passive if and only if
there exists a C 1 storage function H(x) ≥ 0 with the two properties
∂H
a) f (x) ≤ 0,
Lf H(x) = (7.45)
∂x
∂H
b) LG H(x) = G(x) = hT (x). (7.46)
∂x

In most cases, including many physical examples and passivity in the context of controller
design, it makes sense to assume H(x) to be positive definite, i.e., H(x) > 0, with a minimum
at an actual or desired equilibrium state x∗ . Then condition a) means that indeed H(x) is a
Lyapunov function for the equilibrium x∗ of the unforced system. Condition b) establishes a
necessary relation between in- and output vector by fixing the output map h : Rn → Rm .
Let us now look at conditions a) and b) for two interesting special cases:

Output strict passive systems

If there exists a positive definite function H(x) such that, besides condition b), the stronger
version of condition a)

Lf H(x) ≤ −k · hT (x)h(x), k>0 (7.47)

holds, then the system (7.43) is called output strict passive. For such a system, the stronger
passivity inequality
Ḣ(x) ≤ y T u − k · y T y < 0 for y = ̸ 0 (7.48)

holds. For u = 0, H(x) decreases strictly as long as y ̸= 0. Consequently, Ḣ(x) = 0 is

true only if y = 0. If the system is zero state observable, then x = x∗ can be concluded
from y = 0 and we have that Ḣ(x) < 0 for x ̸= x∗ which proves asymptotic stability of the
equilibrium of the unforced system:

Output strict passive systems which are zero state observable have an asymptotically
stable equilibrium for u = 0.

© Prof. Dr.-Ing. habil. Paul Kotyczka 91

Chair of Automatic Control
Technical University of Munich
7.7. Feedback Passivity∗

Linear systems

For linear systems

ẋ = Ax + Bu
(7.49)
y = Cx,

when a positive definite storage function H(x) = 1 T

2 x P x, P = P T > 0 is required, the
above conditions become
1 T
Lf H(x) ≤ 0 ⇒ x (P A + AT P )x ≤ 0, (7.50)
2
LG H(x) = hT (x) ⇒ xT P B = xT C T . (7.51)

The conditions have to hold for all x, hence we can state the following:

Theorem 7.6 (Passivity of a linear system). The linear system (7.49) is passive with a
positive definite, quadratic storage function H(x) = 12 xT P x if an only if there exists a
positive definite solution P = P T > 0 of the Lyapunov equation a) and in- and output
are related by condition b):

a) P A + AT P = −Q, Q ≥ 0, (7.52)
b) PB = C . T
(7.53)

Once more, the Lyapunov equation a) highlights the strong relation between passivity and
stability. As known from Chapter 2, it possesses a solution P = P T > 0 for Q = QT > 0
(or Q = QT ≥ 0 if an additional observability condition holds) only if A is a Hurwitz
matrix. Equations a) and b) are known from the famous Kalman-Yakubovich-Popov (KYP)
Lemma which was formulated to express conditions for positive realness of a transfer function
G(s) = C(sI −A)−1 B. Positive realness is a frequency domain concept which roughly means
that the phase shift of a system remains within ±90◦ . It is strongly related to the time domain
concept of passivity.

7.7 Feedback Passivity∗

We learned so far: The drift part of a given input-affine control system (7.43) (i.e., ẋ = f (x))
must be stable if the overall system (with input and output) is supposed to be passive.
What if ẋ = f (x) is unstable? Is it possible to make the closed-loop system passive by state
feedback? This would be a nice property, as the storage function in this case would serve as
a Lyapunov function to prove stability of the closed-loop equilibrium and to give an estimate
of its domain of attraction.

92 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 CHAPTER 7. THE PASSIVITY APPROACH

Definition 7.5 (Feedback passivity). The system (7.43) is called feedback passive or feed-
back equivalent to a passive system if we can find a state feedback

u = α(x) + β(x)v, (7.54)

with α : Rn → Rm and β : Rn → Rn×m invertible such that the resulting closed-loop

system

ẋ = f (x) + G(x)α(x) + G(x)β(x) v

| {z } | {z }
f cl (x) Gcl (x) (Σcl )
y = h(x)

with new input v ∈ Rm is passive.

The conditions for feedback passivity are given and proven in [1]. We state the result and
show its validity for the special (but insightful) case of linear systems.

Theorem 7.7 (Feedback passivity). Assume that x∗ is a so-called regular point of the
system (7.43), i.e., rank(LG h(x)) = m in a neighborhood of x∗ . Then (7.43) is feedback
passive with a C 2 positive definite storage function if and only if

a) it has relative degree {1, 1, . . . , 1} at x = x∗ and

b) its zero dynamics is Lyapunov stable.

Condition a) means that every scalar output y1 , . . . , ym has relative degree one. A sys-
tem with asymptotically stable/Lyapunov stable zero dynamics is frequently called minimum-
phase/weakly minimum-phase.

Sketch of the proof for linear systems. The key idea is to show that a passive linear system
has relative degree {1, 1, . . . , 1} and stable zero dynamics. Then the fact is used that these
system properties are invariant under state feedback of the form (7.54): The relative degree
of a scalar output is the number of times it has to be derived w.r.t. time until an arbitrary
component of the input u appears. Hence, if the mapping β is regular (invertible), the relative
degree is the same, also with the new input v. Concerning the stability of the zero dynamics,
the invariance becomes plausible if one thinks about the (invariant) zeros of a linear system.
Their location, which cannot be changed via a regular feedback of type (7.54) determines the
stability of the internal/zero dynamics in the linear case.
Consider the linear passive system with a positive definite storage function H(x) = 21 xT P x,

ẋ = Ax + Bu
(Σlin )
y = Cx.

Assume that both B and C have full rank m. According to Theorem 7.6, the matrix P =
P T > 0 satisfies the conditions

AT P + P A ≤ 0 (7.55)
T
B P = C. (7.56)

© Prof. Dr.-Ing. habil. Paul Kotyczka 93

Chair of Automatic Control
Technical University of Munich
7.7. Feedback Passivity∗

As P has full rank n, also the square matrix B T P B = CB has full rank m and is invertible.
Then
ẏ = C ẋ
= CAx + CB (7.57)
|{z} u
invertible

holds, which shows that y has relative degree {1, 1, . . . , 1}.

There exist n − m row vectors, perpendicular to the columns of B (in the kernel of B),
which can be summarized in a matrix T ∈ R(n−m)×n which satisfies T B = 0. Together with
C ∈ Rm×n , the regular coordinate transformation
" # " #
η T
= x (7.58)
y C

can be defined and the transformed state differential equations are

η̇ = T Ax + T Bu} = A
| {z
e 11 η + A
e 12 y
=0 (7.59)
ẏ = CAx + CBu = A
e 21 η + A
e 22 y + CBu,

where only the second is affected by the input u. Applying the feedback

u = (CB)−1 (v − A
e 21 η + A
e 22 y) (7.60)

brings the system into the form

η̇ = A
e 11 η + A
e 12 y
(7.61)
ẏ = v,

where the zero dynamics (i.e., (7.49) restricted to y = 0, ẏ = 0) can be easily identified as

η̇ = A
e 11 η. (7.62)

In transformed coordinates, we have the input and output matrices

" #
0 h i
B
f= , e = 0 I .
C (7.63)
CB

fT P
The second passivity condition (7.56) in new coordinates is B e or written out,
e = C,
" #
h i Pe 11 P e 12 h i
0 (CB)T T = 0 I . (7.64)
P
e
12 P 22
e
| {z }
P
e >0

Two submatrices are immediately defined as

P 12 T = P 12 = 0, e 22 = (CB)−T = (CB)−1
P (7.65)

and the Lyapunov inequality (7.55) becomes

 "
eT eT
# " #" #
A11 A21  P 11 P A
e 0 e 11 0 e 11 A
e 12
T T + e 22 ≤ 0. (7.66)
A12 A22
e e 0 P 22
e 0 P 22
e A
e 21 A

94 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
 BIBLIOGRAPHY

In particular, the following inequality must have a positive definite solution P

e 11 :

T T
A 11 11 + P 11 A11 ≤ 0.
e P e e e (7.67)

From the converse Lyapunov theorems for linear systems, it is known that A
e 11 must be stable,
i.e., the zero dynamics (7.62) must not be unstable.
For completeness, observe from the transformed state differential equations that A
e 21 and A
e 22
can be arbitrarily altered by state feedback, and hence, can be used to satisfy the Lyapunov
inequality (7.66).

We have seen that in nonlinear feedback controller design, different methods are at our dispo-
sition. Each of these methods has advantages and drawbacks. Certain methods can be applied
to certain types of systems only. There is not the one preferrable procedure for nonlinear con-
troller design, rather in many cases it makes sense to decompose a problem into subproblems
and to attack the subproblems with the suitable methods. For example, in systems with an ap-
propriate structure, feedback linearization or IDA-PBC could be used to determine a ficticious
controller within the Backstepping procedure.

7.8 References

For invariance theorems, see the corresponding Chapter 4.3 of [2]. The textbook [3] deals
with dissipative systems theory and passivity-based feedback design. Chapters 6 and 7 are
devoted to port-Hamiltonian systems and their control. [4] gives a wider overview over port-
Hamiltonian systems’ modeling and control. The article [5] presents a solution to the problem
of transparent dynamics assignment in IDA-PBC. There is quite a large amount of research
articles on different aspects and approaches in the field of port-Hamiltonian systems, which
have been published in the past almost 30 years (search in Google Scholar for example).

Bibliography

[1] C. I. Byrnes, A. Isidori, and J. C. Willems. Passivity, feedback equivalence, and the
global stabilization of minimum phase nonlinear systems. IEEE Transactions on Automatic
Control, 36:1228–1240, 1991.

[2] H. Khalil. Nonlinear Systems. Prentice Hall, 3rd edition, 2002.

[3] A. J. van der Schaft. L2-Gain and Passivity Techniques in Nonlinear Control. Springer,
3rd edition, 2017.

[4] A. J. van der Schaft, D. Jeltsema, et al. Port-Hamiltonian Systems Theory: An Introductory
Overview. Foundations and Trends in Systems and Control, 1(2-3):173–378, 2014.

[5] P. Kotyczka. Local linear dynamics assignment in IDA-PBC. Automatica, 49(4):1037–1044,

2013.

© Prof. Dr.-Ing. habil. Paul Kotyczka 95

Chair of Automatic Control
Technical University of Munich
Bibliography

96 © Prof. Dr.-Ing. habil. Paul Kotyczka

Chair of Automatic Control
Technical University of Munich
Appendix A

Handouts

Find here three handouts on proofs of Lyapunov’s theorems, the Lie bracket, and the derivation
of the conditions for input-state linearization.

97
Technische Universität München
Lehrstuhl für Regelungstechnik
PD Dr.-Ing. habil. Paul Kotyczka

Handout 1 – Proofs of Lyapunov's Theorems

Proof of the first (direct) method:

(see e.g. Slotine, Li: Applied Nonlinear Control)

Given a positive definite function with a minimum value in the

equilibrium (i.e. , ) and along the solutions of
the state differential equations.

Alexander Michailovich
Lyapunov (1857-1918)
1) Lyapunov stability Source: Wikipedia

Positive definiteness of the function means that – at least locally around the equilibrium – there
are closed contour lines of the energy function where

Let be an „open ball“ in state space around the equilibrium with radius , i.e. inside this ball
and on the boundary we have .

Let be the minimum value of on the boundary . Then the contour line lies
completely in the ball . Let be another ball which lies completely inside the contour , i.e.
holds.

Starting in , where obviously , and because of , remains smaller than and

therefore remains in .

2) Asymptotic stability

Now the case is considered. Since is lower bounded by and strictly decreases, it
must tend asymptotically towards a limit: .

Assumption:

Under this assumption there exists a ball

around which the trajectory never enters.

is a negative definite / is a
positive definite function around . Remember
that .

If the trajectory never enters the ball , then

there is a lower bound for , i.e.
, or conversely .

Nonlinear Control – WiSe 2019/20 Handout 1 1

Technische Universität München
Lehrstuhl für Regelungstechnik
PD Dr.-Ing. habil. Paul Kotyczka

This means that the function decreases with a minimum rate of . Consequently the trajectory
would enter the interior of the ball in a finite time less than , which contradicts the
above assumption.

The trajectory approaches asymptotically for .

Proof of the second (indirect) method:

(see e.g. Khalil, Nonlinear Systems)

Consider the nonlinear system . The mean value theorem states that

where is a point on the line between and . In the equilibrium the first term
is zero and rearranging the expression yields

The so-defined functions satisfy Schwarz' inequality

Due to the continuity of the first term on the right hand side tends to for
and therefore also (now the are summarized in a vector)

as .

The nonlinear system can now be written (observe that )

, with .

For sufficiently small the second term tends to and the linear approximation of the
nonlinear system is

Nonlinear Control – WiSe 2019/20 Handout 1 2

Technische Universität München
Lehrstuhl für Regelungstechnik
PD Dr.-Ing. habil. Paul Kotyczka

We prove only the statement that asymptotic stability of the nonlinear system can be
immediately deduced from asymptotic stability of the linearized system.

Let be Hurwitz, i.e. all eigenvalues are in the open left half plane: . From Theorem
2.1 we know that there exists a positive definite solution of the Lyapunov equation

for given .

We use the quadratic function as a Lyapunov function candidate. The

time derivative along the solution trajectories of the system is

The underbraced term is negative definite, and an upper bound can be constructed taking the
minimum eigenvalue (which is real and positive) of the symmetric matrix :

Due to the properties of for (see Schwarz' inequality above) there exists for
every a such that

for all .

Then with the upper bound

the time derivative of can be estimated:

Taking it is ensured that for , i.e. in a neighborhood of .

is a strictly decreasing Lyapunov function for the nonlinear system which proves
local asymptotic stability of the equilibrium.

Nonlinear Control – WiSe 2019/20 Handout 1 3

Technische Universität München
Lehrstuhl für Regelungstechnik
PD Dr.-Ing. habil. Paul Kotyczka

Handout 2 – Lie Bracket

We investigate the non-commutativity of vector fields, a property which is of special importance in the
analysis and control of nonlinear systems.

Starting at an initial state , and applying the vector field and then the vector field for an
infinitesimal time each, what is the difference to the opposite case ( and then )?

(If and are constant, the answer is zero, see the example from the lecture).

We represent the solutions and for the two cases

and

as Taylor series in time.

Start with case (a):

Accordingly:

Herein the following holds (we only need terms which appear with order less or equal in ):

Nonlinear Control – WiSe 2019/20 Handout 2 1

Technische Universität München
Lehrstuhl für Regelungstechnik
PD Dr.-Ing. habil. Paul Kotyczka

Therefore:

For case (b) we obviously get (permute and ):

Subtracting both expressions yields

which contains exactly what we have defined as the Lie bracket.

Nonlinear Control – WiSe 2019/20 Handout 2 2

Technische Universität München
Lehrstuhl für Regelungstechnik
PD Dr.-Ing. habil. Paul Kotyczka

Handout 3 – Input-State Linearization

Given the single input system

(1)

Find the state feedback

and the state transformation

(2)

such that

where the right hand side linear system is supposed to be controllable. Replace

and sort the terms with and without . This yields the two conditions

(3)

and

. (4)

As we want a closed loop controllable system, we can assume without loss of generality that the system
has controller canonical structure (the coordinate change then contains the corresponding linear
transformation), i.e.,

The first condition (3) means (evaluate the first rows; the elements of the last row are arbitrary)

Nonlinear Control – WiSe 2020/21 Handout 3 1

Technische Universität München
Lehrstuhl für Regelungstechnik
PD Dr.-Ing. habil. Paul Kotyczka

or, respectively,

(5)

The second condition (3) is then

or, rewritten,

(6)

This is exactly the condition that is an output function of relative degree . The functions
according to (5) must be such that the transformation (2) is invertible. The theorem
stated on the next page gives a necessary and sufficient condition for such an inverible state
transformation.

Nonlinear Control – WiSe 2020/21 Handout 3 2

Technische Universität München
Lehrstuhl für Regelungstechnik
PD Dr.-Ing. habil. Paul Kotyczka

Theorem: Necessary and sufficient conditions for the existence of a (virtual) output with
maximum relative degree , such that the coordinate transformation (2) with the remaining
coordinate functions (5) is invertible on , are:

1. The matrix

has full rank for all and

2. the distribution

is involutive on .

This is equivalent to input state linearizability on a domain of the nonlinear system.

Proof: (Necessity:) If (we use for better readability) has relative degree , then (see
the Lemma 12.2 below), the matrix has full rank and, consequently, is nonsingular and has
dimension on .

From Lemma 12.1 we know that (with and ):

, (7)

which can be rewritten as

. (8)

This means that the distribution is annihilated by an exact differential, i.e., . Hence,
is completely integrable and therefore involutive on (Frobenius).

(Sufficiency:) If 1. and 2. hold, then is a nonsingular distribution of dimension and possesses an

annihilator of dimension , which is the gradient covector of some real-valued function .
From Frobenius' theorem, it is clear that satisfies (7). Applying the Jacobi identity (see below) to
the Lie derivatives in (7) it turns out that

The vector then has the form , and because , it

follows that . Another application of the Jacobi identity yields , which
proves that is an output of relative degree .

Nonlinear Control – WiSe 2020/21 Handout 3 3

Technische Universität München
Lehrstuhl für Regelungstechnik
PD Dr.-Ing. habil. Paul Kotyczka

The following property and Lemmas 12.1 and 12.2 (from H. Khalil: Nonlinear Systems) are used in the
proof of the above theorem.

With the help of the Jacobi identity of the Lie bracket (see exercise) and the fact that the Lie bracket
of two vector fields can be understood as , it can be derived that

With this identity one can prove the following lemma.

Lemma 12.1: The following holds ( is the relative degree of ):

Proof by induction: In the case , the following holds by the definition of the relative degree:

Now show that, if it holds for a given , then also for .

The underbraced term is zero because we have , or equivalently, , which,

by the assumption, results in zero. Moreover, we can investigate the case . Then again, by
assumption

Exploiting the last equation finally gives

which shows that the statement also holds for .

This lemma is used to prove the following one.

Nonlinear Control – WiSe 2020/21 Handout 3 4

Technische Universität München
Lehrstuhl für Regelungstechnik
PD Dr.-Ing. habil. Paul Kotyczka

Lemma 12.2:

• The row vectors („gradient covectors“)

are linearly independent.

• The column vectors

are linearly independent.

Proof:

The red elements are non-zero (see Lemma 12.1), therefore the right hand matrix is non-singular. Then,
also both matrices on the left hand side are non-singular.

Nonlinear Control – WiSe 2020/21 Handout 3 5