Thesis Ligterink

FUNCTIONAL SYSTEM DYNAMICS
Norbert E. Ligterink
University Twente, The Netherlands
E-mail address: n.e.ligterink@utwente.nl
Key words and phrases. Modelling, Control, System Theory, PDE, Dynamics
This research was financially supported by

the Dutch Technology Foundation STW (project grant TWI.6012).
Abstract. Functional system dynamics is the analysis, modelling, and simu-

lation of continuous systems usually described by partial differential equations.
From the infinite degrees of freedom of such systems only a finite number of
relevant variables have to be chosen for a practical model description. The
proper input and output of the system are an important part of the relevant
variables.
Proefschrift Universiteit Twente
graduation committee:
J. van Amerongen, A. J. van der Schaft, P. C. Breedveld, A. de Boer,
B. M. Maschke, J. J. W. van der Vegt, G. van Schothorst
university homepage: http://www.utwente.nl/

ISBN: 978-90-365-2538-1
Universiteit Twente
FUNCTIONAL SYSTEM DYNAMICS
Proefschrift
ter verkrijging van

de graad van doctor aan de Universiteit Twente,
op gezag van de rector magnificus,
prof.dr. W. H. M. Zijm,
volgens besluit van het College voor Promoties
in het openbaar te verdedigen
op vrijdag 31 augustus 2007 om 13.15 uur
door
Norbert Emiel Ligterink
geboren op 26 augustus 1968

te Enschede
Dit proefschrift is goedgekeurd door:
promotoren: Prof.dr. A. J. van der Schaft

Prof.dr.ir. J. van Amerongen
assistent-promotor: Dr.ir. P. C. Breedveld

Contents
Foreword ix
Chapter 1. Introduction 1
1.1. The sad history of operational calculus 1
1.2. Complexity 2
1.3. Port-based modelling 3
1.4. Stages of existence 5
1.5. Boundary impedance 6
1.6. Transmission versus vibration 7
1.7. Preliminary example 10
1.8. Outline 14
1.9. Nederlandse samenvatting 17
Chapter 2. Mathematical models of dynamical systems 19

2.1. Introduction 19
2.2. One-dimensional dynamical system 20
2.3. Finite-dimensional dynamical systems 26
2.4. A bridge problem 29
2.5. Infinite-dimensional systems 32
2.6. Notions of distance, approximation, and change 34
2.7. Variational methods 36
2.8. Differential operators 39
2.9. Complex coordinates 42
2.10. Bipolar coordinates 43
2.11. Infinite matrices 44
2.12. Krylov bases 46
2.13. Some remarks on state space and spatial reference space 48
Chapter 3. Evolution 49
3.1. Input-state-output, or A, B, C, D, model 50
3.2. Continuous spaces and the (A, B, C, D)-system 54
3.3. Observability and controllability 55
3.4. Integration 56
3.5. Second-order Hamiltonian integration 58
3.6. Higher order results 60
3.7. Integration and control 60
3.8. Time and time again 62
Chapter 4. Differential geometry 63

4.1. Surfaces 63
v
vi CONTENTS
4.2. Frenet-Serret coordinates 65

4.3. Integrability 66
4.4. Euler and Frobenius integrability 67
4.5. Conserved flow 68
4.6. Remarks on differential geometry and other differentials 70
Chapter 5. Partial differential equations 71

5.1. Uniqueness and existence 72
5.2. Sturm-Liouville theory 74
5.3. Initial conditions and boundary conditions 77
5.4. Energy flux and boundary conditions 84
5.5. A complex boundary-value problem 85
5.6. Lopatinski condition 87
5.7. Green function 88
5.8. Green function on a finite domain 91
5.9. Constrained minimization 91
5.10. Functional or variational derivative 92
5.11. Remarks on partial differential equations 95
Chapter 6. Models of infinite-dimensional systems 97

6.1. Beam models 98
6.2. Fluid models 101
6.3. Dimensional analysis 102
6.4. Laminar flow around a cylinder 105
6.5. Energy-based discrete fluid model 105
6.6. Control of the Voith-Schneider propeller 106
6.7. Elasticity 110
6.8. Viscoelasticity 113
6.9. Electrodynamics 114
6.10. Heat transfer 117
6.11. Multi-domain models 118
6.12. Some remarks on modelling 121
Chapter 7. Port-based modelling of infinite systems 123

7.1. Motivation 123
7.2. Beyond rigid body dynamics 124
7.3. Local theories 127
7.4. Boundary conditions 130
7.5. Flux 132
7.6. Core model 134
7.7. Internal dynamics 135
7.8. Pseudo-symplectic operators 137
7.9. Conclusion 140
Chapter 8. Linear finite element methods 141

8.1. Introduction 141
8.2. Spline as beam 142
8.3. Finite element method 145
8.4. Stencils 148
8.5. Energy stencils 150
CONTENTS vii
8.6. Boundary conditions for stencils 151

8.7. Finite volume versus finite difference 152
8.8. Model reduction 153
8.9. Proper orthogonal decomposition 154
8.10. Model dimension 155
8.11. Effective Hamiltonian 156
8.12. Plaquette 157
8.13. Simplex 159
8.14. Transverse motion of a table 160
8.15. Simulation of dynamics 164
8.16. Conclusion 165
Chapter 9. Analytic input 167

9.1. Stationary states 168
9.2. Serial network model through tridiagonalization 171
9.3. Analytic input 173
9.4. Delay systems 174
9.5. Finite time 175
9.6. Failing input richness 175
9.7. Smooth matching 176
9.8. Projection on the input space 176
9.9. Laguerre projection 178
9.10. Remarks on the limitations of linearity 180
Chapter 10. Real and virtual dynamics 181

10.1. Iterative methods 184
10.2. Adaptive damping and step size 186
10.3. Conjugate gradient method 189
10.4. Dissipative models 190
10.5. Diffusive-dynamics iteration using sum convergence 191
10.6. The string model 192
10.7. Constraints 194
10.8. Stewart platform 195
10.9. Virtual dynamics 197
10.10. Pendulum example 199
10.11. Mixing real and virtual parameters 201
10.12. General constraints 202
10.13. Remarks on lumped to finite 203
Chapter 11. Nonlinear FEM 205

11.1. The Euler-Bernoulli spline revisited 208
11.2. Invariant lines 211
11.3. Postponement and pre-work 213
11.4. Consistency and conservation laws 214
11.5. Element Hamiltonian 217
11.6. Boundary energy 218
11.7. Counting degrees of freedom 219
11.8. The element Hamiltonian as port Hamiltonian 222
11.9. From null-spaces to port variables 226
viii CONTENTS
11.10. Plate port-Hamiltonian element 228

11.11. Electromagnetism on a simplex 233
11.12. States from the null-space 236
11.13. Final remarks 236
Chapter 12. Force lines and velocity fields 237
12.1. Electrostatics 237
12.2. Force lines 239
12.3. Microscopic balance 241
12.4. Shape optimization 245
12.5. Following the force lines 246
Chapter 13. Conclusions 247
Appendix A. Literature 251
A.1. Functional analysis 251
A.2. Partial differential equations 251
A.3. Numerical methods 252
A.4. Geometry 252
A.5. Modelling 253
A.6. Dynamics 253
A.7. Classical field theories 254
A.8. Networks of dynamical systems 254
A.9. Control theory 254
A.10. History of science and engineering 255
Appendix. Bibliography 257
Appendix. Index 267
Foreword
... untutored mind

His soul proud Science never thaught to stray,
Far as the solar walk, or milky way;
Yet simple nature to his hope has given,
Behind the cloud-topped hill a humbler heaven.
[Alexander Pope, An essay on man]
In many respects this thesis is different from others. It does not treat a single
topic in the traditional sense. It does not give a concise mathematical theory from
a single perspective. This thesis is a collection of issues concerning the modelling
of infinite-dimensional, continuous, or distributed, systems. The writer is someone
who will first make the back-of-the-envelope calculation, before deciding on more
complex tools such as finite element methods (FEM). He hopes the thesis will
bridge some gaps between schools, which are based on tools rather than systems
and problems, of thought concerning modelling.
To deceive yourself, and me with you, you speak too much.

[Georges Bernanos, Sous le soleil de Satan]
In some cases, for example, where well-known material is covered for com-
pleteness, the style from this work is very concise. Facts essential to the story are
repeated three times in Hollywood movies, and every audience is triggered by dif-
ferent words and images. So the writer decided to use multiple views and different
approaches to the same problems and methods, in the hope that many angles are
covered. However, rather than repeating known and classical arguments, the thesis
is written in the hope to supply a new insight in familiar and less familiar topics.
Eventually the understanding of systems is the main goal of most system en-
gineers. Designing an optimal system for a particular purpose is not a process of
trying to hit it on the head. Something has to be learned, even from complete
failures. Insight in a system is not possible without having a language to think
and talk about a system. It is the breadth of this language the thesis hopes to
bring across. The ordering of topics and chapters is such that the physical exam-
ples such as elasticity and fluid dynamics appear and return throughout the thesis.
Moreover, also the mathematical tools, such as finite elements and modal analy-
sis, appear in several chapters. The language is that of variables, invariants and
physical principles. In the theory of system engineering they have be introduced in
different manners by different people. If this thesis follows any of these approaches
more closely, it is the theory of port-based, or bond graph, modelling, introduced
by Paynter, [Paynter, 1961] for discrete, or lumped, systems in the fifties.
ix
x FOREWORD
The less a science has advanced, the more its terminology to rest
on a uncritical assumption of mutual understanding.
[W.V. Quine, Truth by convention]
The principal concept is energy and energy conservation throughout the system.
Each component can be described by variables or states. Changing a variable will
result in a change in energy. The product of the enforced change and the system,
or component, response is the power. Hence pair of variables: the action and the
reaction, or the effort and the flow, are the power, or port, variables, common to
all physical systems. A network representation expressing the possible interactions
in terms of efforts and flows is called a bond graph.
From the other side energy also has played a crucial role in the underlying
mathematics of distributed systems. Distributed systems are described by partial
differential equations. In order to prove uniqueness and existence of a solution to
a particular problem definition, called the initial boundary value problem, mathe-
maticians have resorted often to defining an associated energy function. The energy
function is positive and it has a unique minimum: the zero-energy solution, or the
the null solution. The existence of the minimum is the basis of the existence of
the solution, and that the difference of two solutions has zero energy proves the
uniqueness of the solution.
In this thesis the opposite direction is followed. Rather than searching for an
energy function for a problem, we state the problem in terms of its energy func-
tion or Hamiltonian, and seek admissible interactions, or boundary conditions, for
this problem. The approach is similar to that of stability analysis using Lyapunov
theory. Finding a Lyapunov function for a problem expressed in terms of equa-
tions of motion is difficult. However, for systems given by Hamiltonians, both the
Lyapunov function and the equations of motion follow trivially, yielding essentially
stable dynamics.
I can’t apologize for the banality of my dreams
[Alice Munro, Bardon Bus]
If this thesis is to convey at least one message, this message is that physical
principles should be retained in the numerical analysis of open, or connected, sys-
tems described by partial differential equations. As a consequence the boundary
conditions are not the mathematical prerequisites for the analysis, but determine
the essence of the relation of a system with its surroundings, and thereby the func-
tion of the system.
CHAPTER 1
Introduction
I was in a printing house in Hell and saw the method in which

knowledge is transmitted from generation to generation.
[William Blake, Marriage of Heaven and Hell, 1790]
Functional system dynamics is a convenient name for the study of continu-
ous, distributed, or infinite-dimensional, systems from the perspective of effective
low-order approximations in terms of a small number of modes, which are chosen
appropriately and may depend on external parameters and conditions. Back-of-the-
envelope calculations have been a tool of experts to estimate unwanted effects such
as vibrations and deformations. However, in the cases these effects could not be
ignored, a large effort is required to get a quantitative handle. The resulting analy-
sis blocks fast development, and strains the relations between design, development,
and testing.
1.1. The sad history of operational calculus

In the first half of the twentieth century there has been an ambitious project
called operational calculus. [Heaviside, 1950, Churchill, 1944, van der Pol and
Bremmer, 1959, Wiener, 1949] It aimed to simplify and automate the calculations
of engineering mathematics, such that with limited understanding of the under-
lying theory, engineers could solve dynamical equations from tables and simple
procedures, called calculus grammar. These procedures should be as simple as the
rules used to differentiate functions. Not all derivatives are known. But with a
basic set, and elementary rules a high-school student is able to apply differentiation
rules to most functions at hand. The same set of procedures should have made it
possible to solve any dynamical problem from an electric circuit, or a mechanical
diagram, without much effort.
This program of operational calculus has died a silent death, unless one takes
the implementation of mathematics in economics, called operations research as its
successor and one takes it serious. Wiener considered Fourier-Laplace theory the
successor of operational calculus, [Wiener, 1949] however, its aim is less ambitious.
There are three main reasons for the failure of the program. First of all, the goal
of the engineer was not to produce a specific numerical answer for a specific model
in a specific case, but to investigate a system, determine its critical parts, and
optimize it. This investigation is part of the design process. By decoupling the
problem formulation from the problem solving, by lifting it and removing all design
parameters, which are fixed in the numerical analysis, the feedback into the design
process became impossible.
1
2 1. INTRODUCTION
Secondly, the lifted program, without references to actual systems was too ab-
stract and hard to understand for people who did not already apply the same meth-
ods in context to explicit cases. In lifting the problem all the irrelevant features for
solving it are removed, such that a general but abstract problem remains, which can
be treated with the generic methods of operational calculus. The abstraction made
it harder to apply and check the method; engineers like most scientists use intuition
and experience to guide them, which was not possible in the lifted approach.
Thirdly, computer modelling became the competition of operational calculus.
Engineers discovered direct numerical implementation as a reasonably fast, trouble
free, and accurate enough method to find results for their problems. It requires little
mathematics to turn a partial differential equation into a finite element code. The
ease of approximate numerical solutions plague all mathematicians, who require
some strictness and bounds. Engineers seem to communicate in terms of toolboxes
associated with software like Matlab. The inconsistencies and approximations re-
quired in such code become part of the engineering folklore. The situation could
have been very different if computers were developed twenty years later. We might
have had direct implementations of operational calculus, which now slowly arises
through algebraic programming, implemented in computer programs like Reduce,
Matlab, Maple, and Mathematica.
System theory may suffer from the same problems as operational calculus does.
The engineer who has the experience and intuition on a particular field is not likely
to resort to abstract methods to find answers. On the other hand his tools may
be limited to a set of basic and approximate methods, which might fail for higher
complexity or criticality of the system.
In order not to detach from the practice of the engineer one should formulate a
system in the typical physical parameters which lie at the basis of such systems, and
retain these parameters and variables throughout the formulation of problem, the
method of solving, and the results. The physical principles underlying the system
should be stressed by the method in order to pinpoint the critical elements of the
design.
1.2. Complexity
Even more important than the increasing complexity of the methods available
for analysis is the increasing complexity of the engineering systems. In the old days
machines had only few moving parts, but with the industrial revolution machines,
even simple ones like sewing machines and typewriters, have hundreds of moving
parts. The complexity has increased significantly. However, before the electrical
revolution the parts of a machine or system acted coherently, driven by a rotating
single power source. With the revolution of the axis all the phases or states of the
machine were determined. In the second half of the nineteenth century hydraulic
systems in ships were the first systems where lags, or delays, played an important
role. [Bennett, 1979] Effects of a single retardation were already known for gov-
ernors of steam engines.[Maxwell, 1868] The dynamics of these systems was still
limited. The single steam engine drove the single axis that powered the factory or
ship. Electrical power changed all that. The parts, separately powered, have their
own characteristics and timing.
Electrical engineers made things easy for themselves. The components they
used were designed to have a single characteristic, commonly known as resistive,
1.3. PORT-BASED MODELLING 3
capacitive, and inductive. In other words, they created a one-to-one mapping be-
tween ideal concept and physical component. It made them capable of designing
complex electrical networks with hundreds, thousands of elements. Furthermore,
these networks could again be separated into parts, each with its own function,
such as bridges, delays, amplifiers, and switches. With electrical engineering the
modular approach to engineering design was born. No longer a single engineer
needed to be able to build and understand the machine as a whole. The compo-
nents were designed and optimized independently and put together at a later stage.
In the same way mathematicians use functions and algorithms as building blocks
for more complex theory, and computer programmers write functions, subroutines,
and libraries.
This vision inspired the analysis in other engineering disciplines. Mechanical
parts of machines and chemical processes were decomposed in logical units. These
logical units required connectors similar to the wires in electrical circuits. Different
views led to different types of connectors. In the most general view an abstract
signal was transmitted from one component to another. The term transmission
implies a direction of transmission. In the behavioral approach to systems theory
this direction remains absent, [Polderman and Willems, 1998] the system is only
interconnected. Although an elegant formulation and an appropriate starting point
for modelling a system, [Paynter, 1961] it misses the power of physical principles
and heuristic insight. The structure of input and output; of cause and effect, and
the conservation of a quality, such as energy and mass, yield natural restrictions
within the framework of behavioral models, which are essential for consistency.
The general terms of input and output are reserved for such generic signals.
However, in practice more details are known about the system and the signals. In
the case of electrical networks, the Kirchhoff laws restrict the signals of currents
and voltages such that charge conservation and energy conservation are implicit in
the formulation of the system.
Physicists and chemists also have their own conservation laws: conservation of
mass, charge, energy, momentum, and so on. Not all mathematical models of these
systems satisfy all the conservation laws. In some cases it is extremely difficult,
or even impossible to satisfy all conservation laws that apply, in an approximate
model of the system. However, the restriction conservation laws impose on a model
of a system are important for the quality of the model. If a conservation law is not
built into the model of a system, it will be a good test for the validity of a model.
1.3. Port-based modelling

When you can measure what you are speaking about, and ex-
press it in numbers, you know something about it; but when you
cannot measure it, when you cannot express it in numbers, your
knowledge is of a meagre and unsatisfactory kind; it may be the
beginning of knowledge, but you have scarcely in your thoughts
advanced to the state of Science, whatever the matter may be.
[W. Thomson, Lord Kelvin]
In order to connect components their relation should be clear. Newton wrote
about action and reaction, where one component has the initiative, and the other
component responses. The relation of cause and effect. However, the cause and
effect can also be seen at more equal footing: if two ships are pulling, which is the
4 1. INTRODUCTION
cause, and which is the effect? There is a single force and a single motion of the
point on the line connecting the two ships. In the case of port-based modelling,
the connection is the port defined by force and velocity, and which is the cause and
which the effect are not determined. In the mental landscape of humans, we want
to distinguish between cause and effect. Even if components seem to be entangled
in an inseparable manner, they have been disentangled by picking a starting point
and following the thread from one to the other and back. The interdependence
of elements often gives rise to philosophical discussions, observational studies, and
statistical studies, which only seem to confirm and re-confirm this interdependence.
A pragmatic scientist or an engineer will choose, so to say, the x-axis and y-axis and
will try to disentangle the intrinsic properties of each element and their relations,
without deciding on the actor and the reactor among each two. Analysis is precisely
that; to decompose into elements.
However, for simulation one requires to decide on cause and effect, or causality.
In the case of the two ships tied together, we have to say, in order to resolve the
impasse, that for the first ship it yields a force given a certain velocity, while for the
other ship the force is given and the velocity is the result. In that way the two can
be combined and the result match to yield a consistent dynamical situation with
a force and a velocity of the whole system of two ships. For the model of the two
ships it is not required to determine which is the actor: the force, and which will
response with a velocity. However, for the simulation, and any numerical result
it is relevant. Two pulling ships is in this case also a cumbersome combination.
The common view is to see both as sources of force. The common causality is to
have sources of force as precisely that: a given source causing a velocity or velocity
change. Similarly, in electric networks it is common to have a potential source. The
network as a whole responses by a certain current.
A connection, or port, consists of a pair of variables. The product of the two
has the dimension of power, the rate of change of the energy. Such a general
physical definition makes is possible to connect all kinds of different components
together and still have a consistent model of the whole. This type of modelling
is called bond-graph modelling, or port-based modelling. For discrete or lumped
systems, such as networks or rigid-body dynamics it is a successful method. For
continuous systems, described by partial differential equations, the approach is
under development. There are no fundamental problems to generalize the approach.
However, the properties and phenomena of the dynamics of continuous systems are
much more involved than the dynamics of lumped components.
Part of the problem is to simplify, reduce, or restrict the continuous system to
such an extent it can be used in the multicomponent analysis for simulation and
control. So far the modelling of systems described by partial differential equations
are research projects in their own right. The advent of multiphysics FEM pack-
ages show that these tasks can be automated. On the other hand a lot software
deals with static solutions while dynamics requires time-dependent analysis and
simulation. Again energy can play an important role in this case. The stability of
solutions of partial differential equations are often analyzed using energy methods.
Therefore methods of simulation and control of continuous systems based on energy
methods have two advantages. First the simulations, based on energy principles, for
a energy preserving model are inherently stable in the sense that given the energy
the states are restricted to a subspace of the full state space. Secondly, the energy
1.4. STAGES OF EXISTENCE 5
methods yield automatically the boundary conditions which can be used to add the
component model of the continuous system as part to a bond-graph model.
The natural boundary conditions that arise in energy-based methods are di-
rectly related to particular cases of ports. Many standard boundary conditions
correspond to the absence of power flow through the boundary. These systems are
isolated. More general boundary conditions have to be defined for modelling all the
possible interactions, connections, or ports feasible. The definition of the change of
energy in the system is augmented with a flow of energy through the boundary. In
some respects such an approach resembles the systems of conservation laws, with
one fundamental difference. The conserved energy is not a degree of freedom, or
state, of the system. It can be expressed in terms of pairs of variables, like the
charge density and the potential, or the volume and the pressure. In nature these
pairs of variables are similar to the pairs of port variables. They are the bulk and
boundary pairs of the same concept. The linkage between the two is the geometry.
The domain and its boundary are not only aspects of the topology of an object.
Differential geometry links the two as well via Green identities and their general-
izations by Cartan, Hodge, and Whitney. [Hodge, 1941, Ivey and Landsberg, 2000,
Whitney, 1957] Another view onto the linkage of domain and boundary, between
conserved quantities and conserved currents is offered by functional analysis. The
energy is a norm on spaces of functions, i.e., the state variables of continuous sys-
tems. Combinations of operators on the domain have residual boundary operators.
The Green identities yield relations between norms on the space and the boundary;
the weak equivalent of the change of energy on a domain and its flow through the
boundary.
The functional analysis [Yosida, 1980] formulation has one great advantage over
the differential geometric formulation. Functional analysis allows us to restrict the
function spaces, such that the results carry over from the continuous theory to any
finite order discrete approximation we wish to use for numerical evaluation. In fact
many of the results of functional analysis are based on the convergence of a set of
approximations to the true solution in the continuous space. The typical conver-
gence is called weak convergence as no distinction between some solutions, differing
for example only on a finite set of positions, can be made. However, the results
of differential geometry are not spoilt. The approximations in function space are
made with the knowledge of the differential operators on elemental domains in the
back of our mind. The construction of the boundary operator in the approximate
space will be much simpler with the infinite continuous examples at hand.
1.4. Stages of existence

The numerical model of a system as a whole is the final stage before the nu-
merical simulation and analysis. Parameters need to acquire a value, components
need to be connected, and global boundary conditions need to be set to specify
the dynamical situation in which the system is to be considered. Once this final
stage is reached, the numerical model is finalized and few of the decisions in the
process can be reverted easily, in order to study the effects of these decisions in the
numerical results.
The power of simply analytical, often lumped, models is the easy study of
parametric dependence, such as the material choice, and design decisions, such as
6 1. INTRODUCTION
the manner in which components are interconnected. Furthermore, these intercon-

nections are usually much more easily understood than the boundary conditions,
internally and with the environment and the control in a complex distributed or
FEM model. Many of the choices made to arrive at a numerical model are inherent
in the approach and the tradition.
The models of the systems investigated in this thesis go through many stages
of existence. From a system first the relevant qualities are selected, such as con-
ductance in an electrical setting, or mass density in a mechanical setting. The
related quantities such as current and displacement are the relevant state variables.
A dynamical model of these continuous state variables are the partial differential
equations of a system. In many cases the boundary conditions, which determine
the interaction of the system with its surroundings, are used to select appropriate
state variables. In this thesis we try to move away from this practice, to give a
model of a component a status independent of its use, i.e., a particular interaction.
From the continuous, or distributed, model a, as large as possible, discrete,

or FEM, model is usually the first stage in numerical analysis. However, in many
cases the system is considered isolated for such models, and characteristics, such as
vibrational modes and frequencies, are the outcome of such analysis. We want to
postpone the specific choice of boundary conditions to a later stage, however, some
choices have to be made. In this thesis several options are given. Model reduction
techniques can focus on the recovery of lumped states of a system. For example, the
lumped mass is a property of a system which can be used in many different settings,
such as the total gravitational force it causes, the force it requires to set it in motion,
or the force under impact. The next stage would be to determine the elasticity of
the fixture, and express it in terms of a spring. The consistent decomposition of the
system in terms of lumped properties rather than dynamic responses allows one to
use the resulting model in several dynamical settings. Hence, keeping a modular
design in mind focuses on the qualities of the system.
1.5. Boundary impedance

In the mathematical theory of partial differential equations, boundary condi-
tions are set to generate a well-defined, closed problem. The system, described by
the standard theory of partial differential equations, has a very limited, preset in-
teraction with the environment, through the boundary conditions. Many choices of
boundary conditions, such as Dirichlet and Neumann boundary conditions isolate
the system completely from its environment, such that an energy function exists in
conjunction with the partial differential equations, which is instrumental to deter-
mine existence and uniqueness, the well-posedness of the problem, of the solution
to the partial differential equation.
Another type of boundary conditions is the transmission boundary condition.
A constant flow goes in at one end and out at the other. Although the matter is
in motion the velocity is stationary. This is one of the so-called quasi-stationary
states, considered in this thesis. However, in these cases the well-posedness is
conditional; it depends on the values at the boundary, not just the type of boundary
condition. The system may be driven by the flow, yielding oscillatory effects or even
turbulence.
1.6. TRANSMISSION VERSUS VIBRATION 7
In the most general case, under investigation here, we do not speak of boundary
conditions, but of boundary impedance. The boundary impedance is the response
of the surroundings on the behavior of the system at the common boundary. The
term impedance comes from the theory of transmission lines, but is now widely used
in scattering theory, electromagnetic and acoustic, to describe boundary conditions
for boundaries with nontrivial response, such as absorbing boundary conditions.
The onset of turbulence for a fluid flow in a tube depends on the choice of
boundary conditions. Experiments show, although the mean flow is the same, that
a constant pressure at the boundary yields different results than a constant flow.
The latter is difficult to achieve as most pumps to drive the flow have some response
to the variations in pressure at for example a driving piston. The pump must be
greatly overdimensioned to arrive at the constant-flow boundary condition. Even
though the experiment should be isolated from the surroundings, the engineering
solutions are not always able to do so; an ideal flow pump does not exists.
Boundary impedance models these non-ideal boundary conditions. For fluid
the pressure and the flow are related, for electrodynamics, the electric and mag-
netic fields are related at the boundary, for elastic bodies the pressure and the
displacement are related at the boundary. All of these boundary conditions can be
formulated as the work performed by, or on, the surroundings. Boundary impedance
may guarantee global energy conservation, or power continuity.
1.6. Transmission versus vibration

If such energy transfer can be slowed down or spread over a
wider area, then the intensity of the force applied to the tissues
per unit time is less. This is part of the function of a seat belt in
a vehicle crash, where the stretching of the belt fabric extends
the time of energy exchange and the considerable area of the belt
is a preferable alternative to transferring all the kinetic energy
via a few square centimeters of forehead against the windscreen.
[Bernard Knight, Forensic Pathology]
A continuous piece of material is an abstract component in system without
form or function. Its purpose should dictate the analysis which leads to a particular
form and model, useful in the greater scheme of the system. Static descriptions are
easily made, but the dynamics of a component, i.e., the time-dependent properties
within the system as a whole, is conceptually more challenging. Three common
forms spring to mind. First, simple functions, described by ideal component such
as masses, springs, and dampers, in the mechanical domain. Secondly, the signal
lines with delays and distortions. Thirdly, the vibrations of an isolated component.
The function of a component is generally not to vibrate or oscillate. The

component is added to a larger system, and its function is to transmit or transform,
in the most general sense. Parts of an engine of a car are there to turn the expansion
of the exploded fuel mixture into forward motion of the car. They are all chains
in network, transmitting the motion or force. To be more precise the network
transmits motion under force or force producing motion. Hence a component must
have fixtures, or ports, to receive and send the mechanical energy along.
Many types of analysis of components are based on vibrational motion of the
isolated component. In this manner little information is gained on, for example, the
8 1. INTRODUCTION
rigidity of a component, useful for the transmission of a motion under force. Even
more, most forms of analysis do not have a language for the general transmission
of energy, such as the bond-graph language for lumped port-based modelling. The
typical boundary conditions for fixtures are invented to isolate a component from
its surroundings.
For continuous systems, described by partial differential equations, boundary
conditions are a necessary part to arrive at a well-defined mathematical problem.
The typical boundary conditions are those for isolated problems and will lead, quite
naturally, to vibrational problems in the dynamical regime, or static deformations
in the stationary regime. Interpreting the static vibrational results and combining
them with the design criterions for operation is a complex task, which may not
necessarily say much about the actual function, or transmission, the component
is designed to perform. Typically one would set as criterion that the lowest vi-
brational frequency must be higher than the highest operational frequency of the
system. Whether a vibrational mode is excited by the motion of the system is
often not determined. Furthermore, the actual required motion or operation of the
system has often very little to do with the analysis of the component on its own.
The mathematical boundary conditions do not allow for such global motion. For
example, the body forces due to rotation in a mechanical component are often not
taken into account in the vibration analysis, except for known problems related to
the rotational motion, such as in turbines. [Biezeno and Grammel, 1953]
In this thesis the question of transmission through boundary conditions and
the corresponding boundary variables is addressed. For the same component a
number of problems in terms of different boundary conditions exist, depending on
the particular setting. Take for example a ball. You can sit on it, which yields a
stationary situation with a position boundary condition, at the contact point with
the floor, free boundary conditions on the side, and a force boundary condition on
the top. The mass of the ball does not play a role. The same ball may be thrown.
In that case the mass will determine the acceleration given the force boundary
condition. The acceleration will lead to body forces in the ball. These two situations
correspond to two different problems in the standard mathematical language. In
this thesis we will bring these two problems together, since we are talking about
one and the same ball. The throwing of a ball, which falls and bounces, should be
seen as a single operation and not three.
However, unlike a single ball, with its simple purpose of bouncing, compo-
nents in a machine or system are connected together. The transmission means
components are part of a network. The connections can be distributed or lumped
interconnections or connections to actuators or sensors. The larger network should
be kept in mind, even when dealing with some peculiarities of a single part. The
component function is its reason for existence, in most cases. The relation to this
function is often formulated in terms of the vibrational properties, as described
above, or in terms of a transfer function, between input and output. Transfer func-
tions describe the transfer of abstract signals, rather than mechanical, electrical,
or fluidic transmission; the function of the component. In this thesis the focus is
very much on quantifying the physical transmission. In particular the transmis-
sion of energy or power is used, which is central to most system modelling and
stretches beyond particular physical domains, such as the electrical, mechanical, or
thermodynamical. Eventually, it should get a place within the existing framework
1.6. TRANSMISSION VERSUS VIBRATION 9
Figure 1.1. Most local variations in the boundary conditions are

not “transmitted” through the component, and in a global, net-
work view, despite the distributed nature, average or lumped vari-
ables can be selected.
of lumped energy-based bond-graph network modelling. [Paynter, 1961, Karnopp

et al., 2000]
Rather than starting from the network, and filling in the components, as parts,
we start from the continuous components, since the actual function of a component,
unlike in the lumped approaches, is not a priori clear. Whether a piece of structure
acts more like a spring or a mass, or a combination of the two like a resonator,
depends not only on the structure, but also on the operational setting. We try to
approximate the component by the best simple lumped model, in terms of a set of
modes. These modes are not the result the vibrational analysis, but in terms of the
transmission analysis.
System theory is the theory of cutting and pasting. It is the global view, without
loosing contact with the relevant details important for the function of the system as
a whole. The other details of each of de modules or components are of less interest.
Of interest are the qualities of the component which affect the other components
in the systems, and the nature and possibilities of this interaction. These qualities
of a component are very much related to a whole collection of simple dynamical
states one can bring this component in, due to simple, typically stationary input.
Hence the modular quality of a component is given by collecting the core states,
or modes, of all the possible interactions it can undergo. The limitations of the
fixtures, and thereby the stationary input, limits the network or global view and
the number of qualities of a component one cares to distinguish. A beam attached
only at both ends would have a core, or network, model of four degrees of freedom.
With another fixture in the middle, one or two degrees of freedom extra, depending
on the type of fixture, are required to describe the effects of interaction at one
fixture on another.
A global, or network view of a system does, however, limit the complexity of
the signals from one component to another. Fortunately, nature seems to have the
same limitations in most cases. For example, the precise details of the heat flux
between one component and another are irrelevant as the local temperature and flow
differences are smoothed out, and the total amount of heat flow mainly determines
state, or temperature, of the component. See Figure 1.1. The same principle in
elasticity is called Saint-Venant principle. [Timoshenko, 1953, Truesdell and Noll,
1965] It states that the precise details of the distributed force at the boundary are
irrelevant away from the boundary, only the force and the torsion are transmitted.
The force and the torque are associated with the boundary conditions arising from
the elastic energy in the system. These variables are therefore natural candidates
of network or graph variables in physical network models, the distributed analogues
of the bond-graph variables.
10 1. INTRODUCTION
F+
u+
F u(z)
u
z z=1
z=−1
Figure 1.2. A string with its boundary conditions. Either the
force or the position is given at the boundary, the other variable
can be seen as string response.
1.7. Preliminary example

In order to set the stage for the approach advocated here, we examine a simple,
well-known system of a string. [Courant and Hilbert, 1961, Fetter and Walecka,
1980, Sakamoto, 1982] The dynamics of the string is the wave equation:
(1.1) ρ∂t2 u − κ∂z2 u = 0 ,
where the string coordinate is z ∈ [−1, 1], and the displacement of the string is
u(z, t). The relevant material properties are the mass density ρ and the string
∂ ∂
stiffness κ. The time-derivative is ∂t = ∂t and the spatial derivative is ∂z = ∂z .
The wave equation, Eq. 1.1, may equivalently be formulated as a set of first-order
equations with two variables:
1
(1.2)
u̇
= ρ ∂z p ,
ṗ κ∂z u
where the momentum p = ρ∂z u̇ is an independent variable. An independent vari-
able, or a state, requires its own initial condition; the value a certain time, to
uniquely characterize or predict the systems behavior. Many other choices of the
second variable are possible. We will show that they lead to different boundary
conditions, and, thereby, to a variety of operational settings for the same string.
The energy in the string consists of kinetic and elastic energy:
1 1
Z
dz ρ(∂t u)2 + κ(∂z u)2 ≥ 0 .

(1.3) E=
2 −1
The integrand alone is the energy density; the energy per length. The mass density
ρ and the stiffness κ are the relevant material constantpof the string. The wave
velocity v is expressed in these material constants: v = κ/ρ.
Boundary conditions are introduced to complete the problem. However, for an
engineer the boundary conditions are the most essential part. Without boundary
conditions the string is just an isolated object, possibly of interest of scientists, but
not of engineers.
The boundary conditions come in a large variety. Each of the ends might fixed,
or a force may be applied to it. Furthermore, position and force may vary in time.
Eventually, the engineer cares little about the dynamics of the string. He cares
about the displacement positions u(±1) = u± and forces F± at both ends z = ±1.
See Figure 1.2.
1.7. PRELIMINARY EXAMPLE 11
Preferably, an instantaneous approximation is sought, for example, the force

F− at the left end and the position u+ at the right end as functions of the other
force and position:
(1.4) u− (F− (t), u+ (t)) and F+ (F− (t), u+ (t)) .
Instead of an instantaneous solution, the appropriate result would be the stationary
solutions for static input. Only in that case the internal dynamics do not have to
play a role. This is an important result, which might appear trivial, but is not
always easy to accept.
In general the result will be an integral over the history of the input F− (t) and
u+ (t):
Z t
(1.5) u− (t) = U− (F− (t0 ), u+ (t0 ))dt0 ,
−∞
and likewise for the force F+ .
Furthermore, we can speak easily of a force F , but it does not appear in the
equation of motion, the wave equation. Behind the existence of a force as a bound-
ary condition lies the theory of mechanics and energy. As it turns out, the force is
for the wave equation related to the derivative of the displacement at the boundary.
Force is not an arbitrary input to the system. Systems described by partial

differential equations may behave oddly. However, the energy in the system deter-
mines the bounds on this behavior. In the case that the energy is zero, the system
is at rest. The product of force and velocity is the power, or rate of change of the
energy, supplied to the system. In general, if no force is applied to the system the
energy remains the same. The state moves along lines with equal energy. A partic-
ular elegant way to describe the motion or dynamics is to rewrite the energy in such
a way that it forms the basis of equations of motion, the Hamilton equations. The
energy in this form is called the Hamiltonian. In the case of a Hamilton equations
of motion, in terms of the energy functional, or Hamiltonian H, the applied force
is additive to the change of momentum p:
(1.6)

q̇ δp H 0 0 1 δq H 0
= + = + ,
ṗ −δq H Fapplied −1 0 δp H Fapplied
which is the wave equation, Eq. 1.2, augmented with an input. The notation δp H
and δq H represent the variational derivative of the Hamiltonian with respect to
momentum p and position q.
The change of energy, due the additive force, follows from the time-derivative
of the Hamiltonian:
(1.7) Ḣ = δq H q̇ + δp H ṗ = δq Hδp H + δp H(−δq H + Fapplied) = q̇Fapplied .
The product of the generalized velocity q̇ = δp H with the applied force yields the
power, or change of energy, of the system.
The set of first order equations, with two initial conditions u(t = 0, z) and
p(t = 0, z), Eq. 1.2, can be derived from the wave equation by substituting p = ∂ z u̇.
However, in the Hamiltonian formulation, Eq. 1.6, the variables u and p do not
naturally arise. In the port-Hamiltonian formulation [Van der Schaft and Maschke,
2002] the elastic-energy variable ∂z u is used instead of the configuration variable
u. Likewise the momentum variable p changes. The choice of variables vary with
12 1. INTRODUCTION
the approach and principles, but the pair of generalized coordinate and generalized
momentum belong together. The explicit form of the port-Hamiltonian H 0 of the
wave equation is:
1 2 κ 2
(1.8) H0 = p + q ,
2ρ 2
where q = ∂z u and p = ρu̇. The equations of motion are therefore also not the
standard equations, Eq. 1.6. The symplectic matrix is replaced by the differential
operator:

0 1 0 ∂z
(1.9) → ,
−1 0 ∂z 0
yielding the same equations of motion, Eq. 1.2, for a different pair of variables.
Furthermore, spatially constant configuration solutions u0 (t) will not contribute
to the port-Hamiltonian, expect for the constant of motion, p0 , the momentum of
the string as a whole. In the port-Hamiltonian formulation these states can be
excluded, by the appropriate choice of the function space for the momentum p(z).
Often, we make no distinction between the Hamiltonian as the energy density

of a system, and the Hamiltonian as the total energy, or the integral over the energy
density. Their dimensions are different, but the difference is easily recognized and
understood.
The inclusion of a differential operator ∂z in the relation between velocity u̇(z)
and momentum p(z) is not the canonical choice, but related to the wave equa-
tion, Eq. 1.2. This, in turn, is related to the port-Hamiltonian form, discussed in
Chapter 7.
The coordinates q(z) and momenta p(z) are a collection of infinitely many
variables. For each string position z another value for q(z) and p(z) may occur.
Such systems are also called continuous or distributed. From the infinitely many
variables any subset may be selected to approximate the full system. However, care
must be taken. The canonical pair at a point z, i.e., q(z) and p(z), seem to suggest
that the same modes for q(z) and p(z) will lead to an appropriate discrete model.
This is not the case. For example, for Eq. 1.2, if q(z) is expanded in terms of cos k i z,
the corresponding expansion for the momentum p(z) is in terms of sin ki z.
The discretization of the local product p(z)p(z) may no longer be local, depend-
ing on the approximation of the displacement. The momentum p(z) is related to the
velocity u̇ in Eq. 1.1. This wave equation followed from the variational principles
associated with the Lagrangian, similar to the Hamiltonian for the two first-order
equations. [Goldstein, 1980] Hence in order to recover the proper momentum dis-
cretization we are forced to return to configuration-velocity space (u, u̇) since the
mode shapes of u̇(z) must be a subset of the mode shapes of u(z).
In order to make a coordinate transformation and reduction, to go from func-
tions u(z) and u̇(z) to a set of modes, where the dynamics is given by the coefficients
of the modes, the Lagrangian will serve for consistency. The Lagrangian is the in-
variant density which allows for contact coordinate transformations of u and u̇
simultaneously from a function of z to mode shapes u = qi φi (z), where qi is the
coefficient:
ρ κ
(1.10) L = u̇2 − (∂z u)2 .
2 2
1.7. PRELIMINARY EXAMPLE 13
The displacement u plays two roles: it is the dynamical variable associated with
the energy in the system, and it is the configuration and determines the possible,
or allowed, velocities. The evolution of the system should be consistent which the
configuration the system can attain.
Taking for example a set of modes φi (z) = z i :
(1.11) u(z) = q0 + q1 z + q2 z 2 + q3 z 3 ,
the velocities are directly given in terms of the time derivatives of the coefficients
qi :
(1.12) u̇(z) = q̇0 + q̇1 z + q̇2 z 2 + q̇3 z 3 ,
The same modal subspace is spanned for u and u̇. The canonical momenta p =
(p0 , p1 , p2 , p3 ) corresponding to the modes q = (q0 , q1 , q2 , q3 ) are the variational
derivatives:
Z 1
(1.13) pi = δq̇i dzL = [Mq̇]i ,
−1
where the mass matrix M is given by the integrals over products of modes as they
occur in the Lagrangian:
Z 1
(1.14) qi Mij q̇j = q̇i dzρφi (z)φj (z) q̇j .
−1
ρ 2
The mass matrix has only constant values for a kinetic density 2 u̇ :
 
2 0 1 0
 0 1 0 2 
(1.15) M = ρ 5 
 1 0 2 0  .
5
0 25 0 27
Hence, only for a particular choice of subspace, or modes, the Hamiltonian is prop-
erly defined. The momentum vector p, paired with the position vector q, is simply
p = Mq̇, with the given mass matrix. The local definition of momenta, p(z) for-
mally assigned, would arise from Fourier-Laplace theory [Hörmander, 1990] with
quasi-periodic modes.
The Hamiltonian is for a quadratic Lagrangian itself quadratic:
1 1
(1.16) H = pT M−1 p + qT Kq ,
2 2
where the stiffness matrix is recovered along the same lines as the mass matrix by
integrating the modes φi = z i :
 
0 0 0 0
 0 2 0 3 
(1.17) K = κ  0 0 4 0  .

0 3 0 18 5
The stiffness matrix has a null-space, corresponding to the constant displacement.
If, instead of the polynomial expansion, we would have used a trigonometric expan-
sion, the displacement would be associated with cos ki z while the momenta would
be associated with ki2 cos ki z, which vanishes for ki = 0.
The modes are no longer local. For example, all of the modes q together
determine the position of the ends z = ±1:
(1.18) u(±1) = (1, ±1, 1, ±1) · q = χT± q .
14 1. INTRODUCTION
For a given end position, for example u(−1) = 0, the Hamiltonian needs to be
projected perpendicular to the subspace spanned by χ− . The positions should
satisfy χT− q = 0, the momenta χT− M−1 p = 0. The second condition determines
the rate of change of the constraint χT− q̇ = 0. If the second condition is given, the
constraint χT− q = 0 only needs to be specified for the initial time t = 0.
It is easier to apply an external force to the system. The force applied to one
end, however, is distributed over all the modes.
We see that the collective modes give rise to some subtle results and effects
in the finite approximations of the infinite theory. Another approach would be to
retain as much of the local theory as possible. It is not the momentum p(z) which
is nonlocal, but the potential energy ∂z q(z) which couples the spatially separate
dynamics q(z) and q(z 0 ) for z 6= z 0 . By introducing the moment variable q 0 (z) =
∂z q(z) the Hamiltonian H(q, p) transforms to a port-Hamiltonian H 0 (q 0 , p), and the
equations of motion, rather than the port-Hamiltonian itself, is the source of the
nonlocal coupling of dynamics at distinct points z and z 0 .
1.8. Outline
This thesis addresses the question how continuous components can be analyzed
and described such that the resulting model can be used in a larger system for sys-
tem analysis and simulation. The question focusses very much on the treatment
of a single component and its means to interact with its surroundings. The inter-
action takes the form of the power variables of bond-graph modelling representing
the action-reaction pair which product is the energy flow in and out of the system.
The global question induces several more detailed questions. They fall in a
number of groups:
• Models of a system: variables, properties, transformation of variables,
reduction of variables, i.e., model order.
• Open, or modular, models: a model of a component, which can be inserted
in all possible orientations and situations in a larger system.
• Qualities of infinite-dimensional, or continuous, models: energy, continu-
ity relations, conservation laws, dimensionless constants.
• Dynamics of the model: time-behavior, input-output, stability, conserved
flow.
• Mathematical tools to describe continuous systems: partial differential
operators, differential geometry, variational principles, norms.
• Consistency of the model: constructing well-defined open mathematical
problems, retaining qualities of the system in the model.
• Numerical evaluation: approximation of continuous operators, time-inte-
gration as simulation, numerical stability, constraints implementation.
• Mismatch between variables for energy and for the state, or configuration:
null-spaces of operators, integrability, natural input variables.
These questions are addressed throughout the thesis. A number of preliminary
chapters treat some of the known material in this light. The later chapters are the
result of separate investigations.
For each chapter the purpose is described shortly.
1.8. OUTLINE 15
• Chapter 1 Introduction
Different concepts, such as complexity, model, port-based modelling, bound-
ary impedance, and transmission, related to the work are discussed in a
historical light. The chapter finishes with an illustrative example, this
outline, and a dutch summary.
• Chapter 2 Mathematical models of dynamical systems
The mathematical description of mainly isolated systems with increasing
complexity, in terms of the number of degrees of freedom. Some generic
methods associated with these models. The chapter begins with classical
mechanics, and ends with general geometric, variational, and algebraic
tools used in the description and analysis of dynamical models.
• Chapter 3 Evolution
Finite-dimensional linear models for simulation, simulation through time
integration, integration methods. The state-space, or A,B,C,D-model
to be used throughout the thesis for finite-dimensional linear models, is
defined, some of its properties are discussed, and the time-integration
methods, for simulation, are analyzed and improved.
• Chapter 4 Differential geometry
The spatial domain on which the states of the system are defined plays
an important role in the choice and transformation of the state variables.
Elements of differential geometry relevant for this choice and the possible
transformations are discussed in this chapter: invariant, or coordinate-
free, methods, integrability conditions, and conserved flow.
• Chapter 5 Partial differential equations
A system described by a partial differential equations as input-output
model, how to define energy, formulating it as an initial boundary value
problem, where the initial conditions are the states, the boundary condi-
tions the input. Results from the literature.
• Chapter 6 Models of infinite-dimensional systems
Physical systems described in continuous variables, their formulation and
intrinsic qualities. Elasticity, fluid dynamics, electromagnetism are dis-
cussed. Some models based on these theories illustrate modelling ap-
proaches. Dimensionless constants are discussed as an important tool for
analyzing such systems.
• Chapter 7 Port-based modelling of infinite systems
What are the equivalents of the lumped qualities, typical for bond-graph
modelling, in continuous systems. Can they be isolated and assigned
to boundary values? The port-Hamiltonian is constructed from varia-
tional principles with differential operators. The distributed equivalents
of lumped properties of components are constructed as quasi-stationary
states of the core model. A generalized view, which brings together a
variety of physical boundary conditions, is introduced.
• Chapter 8 Linear finite element methods
Short introduction into FEM, stencils, degrees of freedom, and their qual-
ities and reduction. Special attention is paid to the initial-boundary value
problem, dynamics, and the retention of conservation laws during FEM
approximation of the continuous model.
16 1. INTRODUCTION
• Chapter 9 Analytic input

Model order reduction of large state-space models based on the time-
dependence of the input, which is typical slow, and of low dimension,
compared to the full frequency spectrum of the isolated evolution. The re-
duction falls into two main parts: First, the construction of modes, based
on the input. Krylov method is used for this part. Secondly, the con-
struction of a dynamical model based on these modes through projection.
Different types of projections, time-independent and time-dependent, are
discussed.
• Chapter 10 Real and virtual dynamics
The connection between iterative numerical methods for systems of high
dimension and the simulation of such systems. The iterative methods
are expressed as dissipative systems, the time-step and the damping are
the relevant parameters. The extension to nonlinear problems. Several
lumped nonlinear models are used to test and illustrate the implementa-
tion of constraints.
• Chapter 11 Nonlinear FEM
The invariant description of finite elements in terms of a consistent set
of variables, for energy-conserving dynamics, decomposed in boundary
port variables and energy-based state variables. The geometric invariant,
constraint-free, finite-order approximation of an elastic curve is developed.
The counting of degrees of freedom, for potential energy given by differen-
tial operators such as the curl operator, on square and triangular elements
are discussed. The consistency conditions, arising from both the bound-
ary energy and the differential operators are discussed. The examples of a
triangular plate element and electrodynamics on the simplex are worked
out in some detail.
• Chapter 12 Force lines and velocity fields
A geometric view of the force balance in an extended object. Some aspects
of a global view of continuous components. The balance laws implemented
globally, to be adapted locally through material properties and dynamical
principles. Given a harmonic approximation of the force balance between
input and output, the material and dynamical aspects can be addressed
locally, with simpler methods.
• Chapter 13 Conclusions
A short summary and some conclusions concerning the conceptual devel-
opment proposed in this thesis.
• Chapter 14 Literature
Background reading.
Chapter 7 forms the conceptual core of this thesis. It is the juxtaposition be-
tween the earlier chapters which sketch the existing background within the frame-
work of the problems pointed out in Chapter 7. The later chapters are more de-
tailed answers to the questions posed. Each of these chapters contains stand-alone
results, covering the different aspects, such as input-based dynamics, finite-element
modelling, and global modelling.
1.9. NEDERLANDSE SAMENVATTING 17
1.9. Nederlandse samenvatting

Dit proefschrift “Functionele Systeem Dynamica” behandelt de vraag hoe con-
tinue componenten van een systeem goed beschreven kunnen worden met wiskundig
modellen voor analyse en simulatie van het systeem als geheel. Het systeem als
geheel wordt nauwelijks in beschouwing genomen, maar vooral de wisselwerking
van een enkele component met zijn omgeving wordt onderzocht. Het paar vari-
abelen, de kracht en de beweging, en het equivalent daarvan in andere fysische
domeinen, zijn de poort-variabelen van deze wisselwerking. In de simulatie één
van de variabelen is input, of actie, terwijl de andere de output, of de reactie,
representeert. Het product is de overgedragen energie, ofwel het vermogen. Voor
simpele modellen, met een beperkt aantal variabelen, dit modelleren is bekend als
bondgraafmodelleren.
Voor continue componenten, die typisch beschreven worden door partiële dif-
ferentiaalvergelijkingen, zijn een aantal aspecten van een dergelijke methode van
modelleren onderzocht in dit proefschrift:
• De mogelijke variaties in het model, keuzes van variabelen, transformaties
van het model, de relatie tussen de referentie-ruimte en de toestands-
ruimte, het reduceren van het aantal variabelen.
• Een open model, de mogelijkheden van wisselwerking met de omgeving, de
verschillende manieren en oriëntaties dat een model in een groter systeem
ingepast kan worden.
• De eigenschappen en fysische wetten impliciet in de continue modellen,
energie en andere behouden grootheden, dimensieloze constanten.
• De tijdsafhankelijkheid of dynamica van het model, tijdschalen, input-
output, stabiliteit, en de stroom van een behouden grootheid.
• De wiskundige technieken geassocieerd met continue systemen, differenti-
aal operatoren, differentiaal geometrie, variationele principes en de bijbe-
horende normen en functionalen.
• Het modelleren zelf, het behoud van de fysische principes in simpele mo-
dellen.
• De numerieke evaluatie, benaderingen, tijd-integratie als simulatie, nu-
merieke stabiliteit, en de implementatie van voorwaarden.
• Het verschil in configuratie toestanden en de variabelen relevant voor de
energie in het systeem, de nulruimte van differentiaal operatoren, en de
natuurlijke keuze van input-variabelen.
De resultaten leiden tot de volgende conclusies:
• De differentiaal operator in de moment- of energie-variabele maakt van
de Hamiltoniaan van het gesloten systeem de poort-Hamiltoniaan van het
open systeem.
• Het gebruik van basisfuncties, of modes, met exacte differentiatie en inte-
gratie in de eindige modelruimte, behouden de fysische principes van de
continue modellen.
• De fysische randwaarden, waarmee de energie-flux door de rand gedefini-
eerd wordt, worden uitgedrukt in termen van de potentiaal, of “effort”,
en niet de toestanden.
• In een goede systeembeschrijving van een continue component wordt er
een scheiding gemaakt tussen het kernmodel in termen van de mogelijke
18 1. INTRODUCTION
randwaarden, en het model voor de interne dynamica, waarvoor begin-

waarden nodig zijn.
• De randwaarden en randcondities zijn niet uniek maar afhankelijk van het
type wisselwerking met de omgeving.
• Voor het construeren van een dynamisch eindig element moeten de rand-
waarden en beginwaarden passen bij de bewegingsvergelijkingen. De nul-
ruimte van de bewegingsvergelijkingen speelt daarbij een belangrijke rol.
Het verband tussen de energie-variabelen en de configuratie-variabelen is
belangrijk voor consistentie.
• De quasi-stationaire toestand voor de stationaire input is de belangrijkste
toestand voor het beschrijven van dynamica. Deze toestand is de eerste in
een rij van toestanden, of modes, die de consequentie zijn van polynoom-
achtige tijdsafhankelijke input.
• Het aantal toestanden, of modes, van een FEM model overtreft het aan-
tal relevant of bruikbaar voor simulatie en regeling. De selectie van de
nuttige modes kan gedaan worden op grond van het inputsignaal en de
systeemrespons.
• De globale visie of systeemvisie staat op gespannen voet met de differ-
entiaalmethoden voor continue systemen. De gesloten stroomlijnen van
behouden grootheden en balanswetten vormen een mogelijke brug.
In het algemeen kan gesteld worden de modellen in termen van partiële differen-
tiaalvergelijkingen voor continue systemen leiden tot een naar binnen gerichte blik,
waarin de grote variëteit aan intrinsieke eigenschappen van de continue systemen
centraal staat. De randwaarden zijn daarin bijzaak. De systeemvisie is een naar
buiten gerichte blik, waarbij de randwaarden en de respons juist de hoofdzaak zijn.
De beperkte systeemeigenschappen van de component volgen uit de wisselwerking
met de omgeving.
CHAPTER 2
Mathematical models of dynamical systems
We perceive motion and time together. It sometimes happens

that we perceive the flow of time though we may not sense a
particular sensible motion.
[Thomas Aquinas, On Aristotle’s Physics]
Science is about quantifying and predicting. In order to test our understanding
of a system, we need to be able to predict its behavior, before it happens. Only if
we can predict a systems behavior in a wide variety of circumstances, we can say
we truly understand this system. In order to be able to predict appropriately, we
need to be able to describe; to separate different types of behaviors and states of a
system. Part of predicting is therefore also specifying the variables that describe the
system and its time-dependent behavior fully and uniquely. Some variables do not
vary, these are either irrelevant, or parameters or intrinsic material constants, or
constants of motion, in our language. The variables which evolve with time are the
initial conditions, or state variables, or kinematical variables, in, respective, mathe-
matical, system-theoretic, and physical language. Typically, the state variables are
the minimal, or independent, set of variables needed for the unique characterization
of the behavior of the system. Behavior is the generic name of state as function of
time. The greater part of this thesis is about separating and combining the input
from outside and the intrinsic behavior of an isolated system. In this chapter we
deal mainly with isolated systems, commonly referred to as dynamical systems.
Dynamical systems are the dressed down versions of the general models, where
only the essential variables for the description and evolution of a system are re-
tained. Dynamical systems is the study of the time-dependence of systems. In or-
der to have a well-defined evolution of a system, there is a minimal set of variables
required, these are the kinematical variables. The restricted study of kinematical
systems is, in this thesis, the unique characterization of a system at a certain time.
2.1. Introduction
In this chapter we treat the mathematical models from the perspective of
Hamiltonian dynamics and energy conservation. These models are classified by
the number of degrees of freedom. With the growing number new features arise in
dynamical systems. For a more complete introduction into Hamiltonian dynamics
the reader is referred to the classical literature, [Whittaker, 1917, Carathéodory,
1935] physics literature, [Landau and Lifschitz, 1976, Goldstein, 1980, Fetter and
Walecka, 1980], or the mathematical literature. [Gantmacher, 1970, Hirsch and
Smale, 1974, Arnold, 1989]
Energy conservation is a central principle in physical dynamical systems. Con-
servation means the energy remains equal with time. The notion of energy has
19
20 2. MATHEMATICAL MODELS OF DYNAMICAL SYSTEMS
been around since classical times, Huygens was the first person to equate two forms
of energy: kinetic and gravitational, and thereby quantifying the concept. [Huy-
gens, 1673, Theobald, 1966] But its central role, given to it by Hamilton, since the
eighteenth century, reversed its function. Rather than energy, E, being a quality
of a system, it became a dynamical principle. If energy is conserved, the state of
an isolated system must retain the same energy, which restricts the dynamics, i.e.,
the change of the state in the time. The resulting equations of motion are known
as the Hamilton, sometimes Hamilton-Jacobi, equations of motion. The energy E
in an appropriate choice of dynamical, or state variables x of the system is the
Hamiltonian: E = H(x).
The energy of a system has later been related to the translational invariance
in the time. For physical systems the results of experiments will typically not
depend on whether an experiment is done today or yesterday. This invariance is a
symmetry of the system and the resulting constant of motion is the energy. Noether
[Noether, 1918] and Weyl [Weyl, 1952] have been instrumental in the introduction
of this connection between time and energy. In that respect the Hamiltonian H was
identified as a single component of a space-time tensor, [Frankel, 2004] 4 × 4 matrix
that transforms with the transformation of space-time coordinates, T 00 = H. The
0-component indicates the time direction, while the other 1, 2, 3-components are the
spatial directions x, y, z. The full conserved energy-momentum tensor {T µν }3µ,ν=0
expresses all of the space-time symmetries; translational, rotational, and boost.
[Felsager, 1981] These aspects are further discussed in Chapters 5 and 7.
2.2. One-dimensional dynamical system

The most generic dynamical system is the one-dimensional harmonic oscilla-
tor. In many cases, dynamical systems are reduced or approximated by harmonic
oscillators, or independent copies of multiple harmonic oscillators, in a linear ap-
proximation. In the section we will discuss the harmonic oscillator in some detail
and introduce most of the concepts we will use later for finite-dimensional and
infinite-dimensional systems. The harmonic oscillator can be described by a second
order differential equation:
(2.1) mẍ(t) + kx(t) = 0 ,
where the position x is a function of the time t. The time derivatives are denoted
by the dots ẋ = ∂t x = ∂x 2
∂t , ẍ = ∂t ∂t x = ∂t x, etc. Two parameters play a role:
the mass m and the stiffness k. There exists a continuous set of solutions to this
problem characterized by two parameters: the amplitude x0 and the angle φ:
r ! r r
k k k
(2.2) x(t) = x0 sin t + φ = x0 cos φ sin t + x0 sin φ cos t .
m m m
q
k
The constant m is the radial frequency usually denoted by ω. In order to define
uniquely the evolution of the system x(t) for t > 0, we require the knowledge of two
variables at t = 0, the position x(0) and the velocity ẋ(0), these are the kinemat-
ical variables. The system, at all times, is uniquely described by the kinematical
variables at a given time. The point x(0) = 0 and ẋ(0) = 0, is however unique.
This point is a critical point, in the sense that the state remains the constant over
time. For no other point (x, ẋ) this is the case.
2.2. ONE-DIMENSIONAL DYNAMICAL SYSTEM 21
ẋ
-
-
-
x

Figure 2.1. Some trajectories of the harmonic oscillator in

configuration-velocity space, the vectors indicate the velocities.
The closed orbits in the form of ellipses are the contours of equal
energy.
The critical point is also singular point in the polar coordinates (x0 , φ). The
angle variable is irrelevant at x0 = 0, as the system will remain stuck at x(t) = 0,
hence it is characterized by its zero amplitude x0 = 0. If we draw the trajectories of
the kinematical variables as a function of time we find they are ellipses around the
critical point. See Figure 2.1. The closed orbits of the harmonic oscillator are very
interesting. It tells us that the time evolution is described by the angle φ(t) only;
the amplitude x0 remains constant. The amplitude is associated with the energy
E in the system. Since two variables are needed to describe a system, we might as
well pick two variables which brings the system to the general canonical form for
arbitrary vectors x, which coefficients are the kinematical or state variables:
(2.3) ẋ = Ax ,
which means using the momentum p = mẋ instead of the velocity ẋ, since the
leading order term in the differential equation mẍ must be on the left hand side of
the canonical evolution equation. The matrix A is called the evolution operator.
The ellipses in the configuration-velocity space (x, ẋ) can be mapped to ellipses
in the phase space (x, p) = (x, mẋ). The energy, yielding the families of trajectories,
is given by:
1 2 k 2
(2.4) E= p + x .
2m 2
The motion is perpendicular to the lines with constant E. Hence it can be treated
geometrically rather than as a differential equation of time. The two variables, the
position x1 = x and the momentum x2 = p form a vector x:

x1 x
(2.5) x= = .
x2 p
The direction in which the system moves in time: ẋ is a vector tangent to the line
E(x) = constant, which is most easily constructed as the vector perpendicular the
xE
dxdp
xE
E=E 3
dxdp
x
E=E 1 E=E 2
Figure 2.2. The gradient, in x = (x, p), of the energy is a vector
perpendicular to the lines of constant energy. The time evolution
is perpendicular to the gradient, and therefore along the lines of
constant energy. If the lines are closer together, the gradient ∇x E
is larger, and so is the evolution ẋ. The area of the phase space
dxdp is preserved.
gradient of the energy:

0 1
(2.6) ẋ = ∇x E(x) .
−1 0
The choice of factors m−1 and k and the variables q and p for the energy E is
now clear, as it yields the correct time dependence. Even if the trajectories are not
simple ellipses but some complicated closed orbit in phase space, the time evolution
of the system ẋ would still be given by the gradient of the energy, for the simple
reason that if the energy is conserved; it does not vary in time:
(2.7) 0 = Ė = ∇x E(x)T ẋ ,
the vectors ∇x E(x) and ẋ should be perpendicular. It is an application of the chain
rule of differentiation, [Rudin, 1966] and the equations of motion. For example, if
the symplectic matrix (10 −1
0 ), guaranteeing the orthogonality, would not be constant,
but a function of x, it would have led to a diagonal terms under differentiation with
as consequence Ė 6= 0.
An interesting consequence of the symplectic rotation, i.e., the matrix (10 −10 )
in front of the gradient, and the gradient itself is the conservation of an area of
the phase space. See Figure 2.2. This is known as Liouville theorem. [Arnold,
1989, Thirring, 1997] The configuration-velocity space does not have this property.
If we consider an area x ∈ S spanned in phase space, the area remains constant
under time evolution. Heisenberg’s uncertainty principle ∆p∆x ≥ ~ can therefore
be a fundamental consequence of quantum mechanics independent of coordinate
and time. In simple terms, no quantum system can be squeezed into an area of
the phase space smaller than ~, Planck’s constant. Furthermore, the phase space
of a single degree of freedom can be mapped to the problem of two-dimensional
incompressible flows, which is equivalent to conformal mappings of the complex

plane. [Spiegel, 1964] See also Section 4.5. indexconformal
If the phase space consists of simple orbits around the critical point. The
position and momentum variables can be mapped to action and angle variables.
[Fetter and Walecka, 1980] The action indicates the orbit, or the amplitude, and
the angle the position along the orbit. The time evolution of the angle variable
φ is therefore given by a function of the amplitude. However, it would be more
convenient to use a function of the energy as the second variable in the kinematical
pair. Such a set of kinematical variables are called action-angle variables. The
action W (E) is given as the first kinematical variable, hence, in order to have a
dynamical system with a conserved energy in the canonical form:

Ẇ 0 1 ∂W E(W ) 0
(2.8) = = ,
φ̇ −1 0 ∂φ E(W ) ∂W E(W )
which would yield a linear change of the angle φ(t) = ∂W E(W )t + φ(0). In the case
of a harmonic oscillator the action variable W is the energy: W = E. Consequently,
for the harmonic oscillator, the frequency, the time it takes to traverse each orbit,
is constant and independent of the energy or amplitude. In more general cases
of dynamical systems this is no longer true: the frequency will depend on the
amplitude. It will not always be simple to construct the action variable. The reason
why the harmonic oscillator still is the cornerstone of dynamical systems is that
around most critical points, for small amplitudes, the system can be approximated
by a harmonic oscillator.
The mathematical pendulum is a point mass on a massless stick which can ro-
tate in the plane under the influence of gravity. The differential equation governing
its motion is:
(2.9) mlẍ + mlg cos x = 0 ,
where x is the angle, g the gravitational constant, and l the length of the stick. It
can be derived easily from the energy considerations. The kinetic energy T is the
same as before:
p2 ml2 ẋ2
(2.10) T = = .
2m 2
The potential energy V is the gravitational energy with a constant force downward
or vertical, such that the potential depends on the vertical position of the mass:
(2.11) V = mgl cos x .
Inserting the total energy E = T + V into the canonical dynamical equation and
using p = mlẋ, the differential equation for the mathematical pendulum arises,
again described by two kinematical variables, in this case x and p. A closer look
at the phase space we see it is more complicated than the elliptic orbits of the
harmonic oscillator. See Figure 2.3. The numerical evaluation of the mathematical
pendulum, as a constrained system in two dimensions, is discussed in Section 10.10.
This is the first of many nonlinear systems we will encounter. In this case
the equations of motions are nonlinear, because the potential V is not a quadratic
function of the energy. Another nonlinearity arises if the mass m(x) is a function
of the position. In most cases we will consider the energy to be only quadratic
in the momentum, yielding always a linear dependence on the momentum in the
Figure 2.3. Some trajectories of the mathematical pendulum in

phase space. The crossing lines are homoclinic orbits, reaching the
saddle points only after infinite time. The angle variable leads to
a 2π-periodic phase space.
equations of motion. While investigating dynamical systems in higher dimensions,

more complex nonlinearities and higher order time-derivatives will arise. However,
often these systems can be mapped to products of one-dimensional systems (xi , pi ),
at least in a small part of the multi-dimensional phase space. There are also many
different ways to characterize the nonlinearity of a system. The phase space picture
is generally a good one, for second-order systems, or independent copies thereof.
However, care must be taken. Transformations of the system, e.g., x → x0 (x) might
change V and m beyond recognition, but leave the dynamics invariant.
For the global motion and distinguishing different types of behavior the critical
points are essential. For the mathematical pendulum, the critical points lie on the
p = 0 axis and are alternating stable critical points, with harmonic motion in its
vicinity, and unstable critical points of the saddle type, stable in the p-direction
and unstable in the x-direction. Such points can never be reached in finite time,
as the system will slow down as it comes closer to the critical point. Such orbits
are called homoclinic orbits and separate the phase space in regions with typically
different global behavior, like periodic and non-periodic. See Figure 2.3.
So far, we have ignored damping or external forces, we have looked at the
system as an isolated system. The energy was conserved for this reason. It allowed
us to bring the equations of motion in the convenient Hamilton form. Energy is
conserved in general, however, it might take other forms than we care to consider.
For example, friction or damping will turn kinetic energy into heat or vibrations.
Hence we will put the friction force, and external forces, on the right-hand side
of the dynamical equation as input or output of the dynamical system x = (x, p)
under consideration: (p = mẋ)
γ
(2.12) ṗ + kx = mẍ + kx = − p + F = −γ ẋ + F .
m
where F is the applied force and −γ ẋ is the induced friction. In terms of the
Hamilton equations of motion, this is an augmentation of the second, or bottom-
row equation, Eq. 2.6.
The change of energy Ė is a function of the right-hand side. The definition of
energy will not change, however, it is no longer conserved. The change of energy is
given by the equations of motion:
(2.13) Ė = ∇x E T ẋ = −γ ẋ2 + ẋF .
The friction can only lower the energy of the system, since −γ ẋ2 < 0. Friction is
a dissipative force. The applied force F can both lower and increase the energy of
the system. The product of velocity ẋ and force F is called the power product, as it
indicates the rate of change of energy put in or taken out of the system. It will play
an essential role in constructing connections between separate systems and defining
boundary variables. Traditionally, in order to define an isolated system in a general
setting, either ẋ = 0 is set, or F = 0. The latter corresponds to free motion, the
first to fixed position.
Treating in friction on the same footing as the energy-conserving dynamics
of the harmonic oscillator can be a formidable task. [Soper, 1976] The Rayleigh
and the Gossick terms [Gossick, 1967] allow for the inclusion of damping in the
equations of motion. However, friction or damping is not the loss of energy. It
is the conversion of energy, to the thermal or vibrational domain. [Breedveld,
1984] Friction represents a port, or connection, between domains, just as applied
forces are. Damping and friction are input and output between components. These
components are not spatially separated, but conceptually.
There is another form of input, which is the constrained motion, which means
specifying x(t) or ẋ(t). For a particular initial condition, x(0), ẋ(0) at time t = 0,
specifying x(t) or ẋ(t) is equivalent, since ẋ(t) can be integrated with the initial
condition x(0) to yield x(t). They specify a fixed motion in phase space, hence a
trivial dynamics. However, in higher dimensions, fixing one position xi (t), would
yield time-dependent equations of motion. In practice it is important to keep track
of the force fi (t) associated with the constrained coordinate xi (t), as the pair tells
how the energy changes in the system:
∂E
(2.14) Ė = ẋi = fi ẋi .
∂xi
At this point it makes no difference whether we constrain the motion xi (t), or

whether we applied the correct force fi (t) which yields this particular motion. In
practice it will make a difference, physical models of motors which yield the applied
force have a particular response which is a function of input u(t), xi (t) and fi (t):
χ(xi , fi , u) = 0.
Apart from identifying a dynamical system by its phase space, one could also
identify the system by its amplitude x0 and phase φ, as specified in Eq. 2.2. The
response is given in terms of amplitude and phase of an oscillatory x(t) to an
external oscillatory force F (t) = F cos(ωt), or usually in frequency space F (ω). The
plot of amplitude and phase against the frequency of the applied force F (t) is called
a Bode plot. For a linear system this identification suffices, the function H(ω) =
x0 (ω)
F (ω) is called the transfer function from applied force to amplitude. However, in
the nonlinear case this will not do. If the phase is included, one generally has much
more insight in the system. However, global knowledge, as depicted in the phase
space, is often lost. To reach certain amplitudes for, for example, the mathematical
pendulum with a limited force max F < F0 , the frequency ω has to decrease with
time to drive the system close to resonance with the proper phase.
2.3. Finite-dimensional dynamical systems

From the one-dimensional dynamical systems there are many routes along
which we can extend to higher-dimensional systems. The simplest approach would
be to consider n copies of a one-dimensional system, with 2n variables in canonical
pairs. The Hamiltonian for n copies of a one-dimensional system is the sum of n
copies of the one-dimensional Hamiltonian:
n
X
(2.15) H(q, p) = Hi (qi , pi ) .
i=1
The Hamilton equation of motion is identical to the one-dimensional case, even for
an interacting system:

q̇ 0 1 ∇q H(q, p)
(2.16) = .
ṗ −1 0 ∇p H(q, p)
The mass and stiffness matrices are given for a value of the position vector q and
the momentum vector p:
(2.17) Mij−1 (q, p) = ∂pi ∂pj H(q, p) ,
and
(2.18) Kij (q, p) = ∂qi ∂qj H(q, p) .
In the mathematical literature, [Arnold, 1989] however, a finite-dimensional Hamil-
tonian is more general than the n copies with 2n canonical variables. It might have
an odd number of variables x = (q, p), where there is no longer a clear distinction
between position and momentum variables.
Considering just geometrical arguments would be one approach to arrive at such
general Hamiltonians. The link with rotation may clarify some of the peculiarities.
In that case only the energy, and the symplectic structure are of interest. If we still
want to satisfy energy conservation, the only requirement is that the evolution of
the system ẋ is perpendicular lies in the E =constant surfaces, which will be of n−1
dimensions, if the space of x is n dimensional. Hence, ẋ has to be perpendicular to
the gradient of the energy. A vector perpendicular to the gradient of the energy is
necessarily the generator of rotation. Given the evolution equation:
(2.19) ẋ = J∇x E(x) .
The symplectic matrix J should satisfy:
(2.20) J T = −J .
In a simplified case, where the energy surfaces, E = contant, are surfaces of a
sphere the canonical energy is E = 21 xT x. The symplectic matrix corresponds to a
rotation X of the vector x:
Z
(2.21) eJt dt = X(t) with XT (t)X(t) = 1 = X(t)X(t)T .
Rotations leave the sphere invariant: a point x with energy E is mapped unto a
point x0 with the same energy. In dynamical systems, from a configuration space,
each position xi has an associated velocity ẋi , yielding a canonical pair (xi , pi ). On
the other hand, the symplectic matrix J can be odd-dimensional. For example, for
2.3. FINITE-DIMENSIONAL DYNAMICAL SYSTEMS 27
the rotation of a free rigid body, no potential appears and the symplectic matrix J
is the three-dimensional generator of rotations:
      
ẋ1 x1 0 ω12 −ω13 x1
(2.22)  ẋ2  = J  x2  ≡  −ω12 0 ω23   x2  ,
ẋ3 x3 ω13 −ω23 0 x3
where ωij are the components of the rotation vector. The three-dimensional gen-
erator is build up of three basic generators J12 :
     
0 1 0 0 0 −1 0 0 0
J(ω) =  −1 0 0  ω12 +  0 0 0  ω13 +  0 0 1  ω23
0 0 0 1 0 0 0 −1 0
(2.23) = J12 ω12 + J13 ω13 + J23 ω23 = J · ω .
The rotation dynamics of the rigid body is a well-known example of an odd-

dimensional dynamical systems. Since the rotation of a rigid body has no associ-
ated potential energy, only the three angular velocities xi are of interest, the actual
orientation angles not. Hence only three of six degrees of freedom are kept in the
dynamics. There are many other examples of reduced dynamics, they are the result
of constraint relations, conserved quantities such as momentum, or approximations.
One could argue that such reduced dynamics are the fundamental dynamics. How-
ever, the connection between position and velocity, useful for consistency, is often
severed. For the high-dimensional and infinite-dimensional systems considered later
the reduction will play a role in the index of the operators at hand. In practice,
reducing a high-dimensional system by a couple of orders is not worth the trouble.
Only for consistency, it is important to single out irrelevant variables.
Since a symplectic approach to dynamics is more general than a local product
of one-dimensional systems, we will generally stick to the geometrical, symplectic
approach. However, Darboux theorem yields an important connection between
the two approaches. Locally, that is: in a small region of phase space, a symplectic
structure {J, E} can be mapped, through the right coordinate transformation x0 (x),
to a set of canonical pairs of the canonical energies, with remainder coordinates
mapped to the evolution null-space, i.e., fixed x0i :
 
0 −1
 1 0 
 
 .. 

 . 

0
 0 −1  0
(2.24) ẋ = 
 x .
 1 0 


 0 


 . ..


0
Hence, from the general mathematical description, in terms of the symplectic ma-
trix, it is again possible to arrive at a local coordinate system, in some small neigh-
borhood of the phase space, where the system corresponds again to n copies of the
one-dimensional oscillator.
For the case of the three-dimensional rotation of a rigid body this theorem has
a simple geometrical interpretation. The surface of constant energy is a sphere.
 
0
 0 
x - 1

j 0 
E = constant    1 
E 0
 0 
0 E = constant
Figure 2.4. On the energy surface in three-dimensional space, the

two directions perpendicular to the surface normal form a canonical
pair.
At a particular point of the sphere, there are only two tangential directions in the
energy surface. To first order, the third direction remains constant. For example,
if x2 = x3 = 0, the generator J23 will change the system. For example, if we map
the position on the sphere to x = (1, 0, 0)T , the two tangential directions, forming a
canonical pair, undergoing a symplectic evolution, are (0, 1, 0)T and (0, 0, 1)T . See
Figure 2.4.
The energy sphere, invariant under rotations, has another interesting feature.
The evolution of a system can be seen as a continuous transformation, mapping
the phase space unto itself. Such a transformation is a group, and for many groups
the structure is analyzed in depth. Hence it might be possible to map a dynamical
system to a group. However, only for very simple linear dynamical systems such a
transformation is interesting as it simplifies the formulation, and yields global prop-
erties, such as spectra. The one-dimensional harmonic oscillator can be mapped to
the SO(2) (all orthogonal 2 × 2 matrices with unit determinant) or the U (1) (all
unitary 1 × 1 matrices) group. The U (1) group is the multiplication group of all
complex phases:
(2.25) eiφ eiψ = ei(φ+ψ) ,
where the generator of rotations is the multiplication with i:
(2.26) ż = iz ,
where the complex variable represents the canonical pair: z = x1 + ix2 , with the
radial frequency is ω = 2π. In this case the action-angle variables correspond to
the modulus-argument variables (|z|, Im ln z).
The equivalent SO(2) formulation yield group elements: X ∈ so(2):

cos φ sin φ
(2.27) X(φ) = ,
− sin φ cos φ
with the properties:
(2.28) X(−φ)X(φ) = X T (φ)X(φ) = 1 = X(0) = X(φ)X T (φ) ,
(2.29) X(φ)X(ψ) = X(φ + ψ) ,

0 1
(2.30) ∂φ X(φ) = X(φ) .
−1 0
2.4. A BRIDGE PROBLEM 29
The nicest aspect of a group is that only a single point of the phase space has to be
analyzed, all other points are related to that single point by a group transformation
X. Furthermore, the coordinate system of the phase space, and the tangential
vectors ∇x E(x) and ẋ, are naturally given by the group generators and the group
orbits. These principles can be used to improve the numerical simulation of such
system. See Section 3.4.
Another class of groups are the symplectic groups, which are the coordinate
transformations x0 (x) which leave the dynamics invariant. Not all coordinate trans-
formations are allowed, as is to be expected. This can most easily be seen from
considering the configuration space (x, ẋ). Given a transformation of the position
x0 (x) the transformations of the velocity ẋ0 is given:
∂x0
(2.31) ẋ0 = ẋ .
∂x
0
The matrix F = ∂x ∂x is the transformation Jacobian, sometimes called the defor-
mation gradient.[Truesdell and Toupin, 1960]
2.4. A bridge problem

Before we arrive at dynamical systems with an infinite number of degrees of
freedom, we will first discuss a typical example of a finite-order approximation
of such systems. All macroscopic objects have in principle an infinite number of
degrees of freedom. A stone, subject to enough force will shatter in many pieces,
but for all practical purposes it can be treated as a single extended object, with six
degrees of freedom; translations and rotations.
Hence the number of degrees of freedom depends on the functionality; the type
of interaction with its environment. One step beyond a completely undeformable
object would be a simple deformation based on static applied force. [Cohen and
Muncaster, 1988] The greater part of this thesis is devoted to put such approxi-
mations on a solid footing, with systematic extensions. Furthermore, if interaction
defines the functionality of an object, and thereby its degrees of freedom, it is im-
portant to clearly define this interaction. Interaction for structural elements is often
the result from identification of points or nodes. A single position can be part of
two or more elements, thus joining them together. Implicitly, the mutual forces
should cancel each other, according to Newton’s action and reaction. In chemical
and fluid systems different types of interactions exist. However, they can be seen
as the natural partner of the structural collocation interaction. In such systems
matter, charge, or another quantity is conserved. The transmission of such con-
served quantity can be seen as the analogue of the action-reaction pair, the loss of
one component equals the gain of the other. The notion of energy comes into the
picture from the effort or potential at which this transmission occurs.
The modular approach consists of two major elements. First, the dynamical
model of each part. Secondly, the interaction between two parts. The collocation
interaction, identifying configuration points, leads to the simplest modular model;
a dynamical model, given by the Hamiltonian, of the whole.
As an example of the modular approach to structural modelling we examine a
structure like a bridge, which has horizontal and vertical segments. Each segment
and each junction can be analyzed on its own, and the whole, with reduced order

x 1 x 2 x 3

Figure 2.5. The analysis of the bridge combines all components.

At each junction the position xi and the angle between the com-
ponents are relevant.
segments can be combined and analyzed. Apart from the computational advan-
tage of not having to perform a very large structural calculation, the advantage of
decomposing the large structure of a bridge into logical units is the fact that each
unit, such as a bridge column, a connection, of a surface element, can be analyzed
on its own, exposing its weaknesses.
The vibrational modes of a bridge or a comb arise in many engineering prob-
lems. The usual approach is a large structural calculation, using all the elements
combined. However, such problems may be regular, having a large amount of
repetitive structure, such a row of identical bridge columns and road segments. See
Figure 2.5.
Each of the columns we may investigate separately, yielding a response to a
force f :
(2.32) x(f ) ≈ Cf .
For a composite column calculating the response to a stationary force is the easiest
approach since for each segment for the column the same force applies, and the sum
of all displacements is the total displacements in the linear approximation:
X
(2.33) C= Csegment .
segments
See Figure 2.6 The elastic energy of a column is:

1
(2.34) Eelastic = xT C−1 x .
2
However, this does not complete the analysis of the column. Possibly, the columns
may possess sufficient inertia, that they yield an effect in lowering the vibrational
modes of the bridge as a whole. The displacement x(z) as function of the distance
from the top z = 0, to the bottom of the column z = 1 is inverse proportional to
the integrated stiffness:
Z z
(2.35) x(z) − x(0) = c(z)f dz .
0
For a homogeneous column the displacement is linear. In that case only a third of
the mass of the column contributes to the inertia of the bridge:
1
(2.36) Ekinetic = Mcolumnẋ2 .
6
In most practical cases the bottom of the column is heavier and stiffer that the top,
reducing the inertial column effects further.
2.4. A BRIDGE PROBLEM 31
x = x1 + x2+ x 3
3
x = x1 + x2
z
2
x = x1
1
Figure 2.6. A single column can be analyzed. Its characteristic

is the sum displacement as function of the applied force. The
displacements of all segments can be summed to the total displace-
ment.
The inertia of the surface of the bridge contributes in full to the vibrational
inertia. The stiffness is however more complicated. Between two columns a surface
segment has stretch and bending stiffnesses. The stretch stiffness depends only on
the distance between the ends:
1
(2.37) Estretch = k(xi − xi+1 − l)2 ,
2
where l is the length of a segment.
The bending stiffnesses depend both on the angle the bridge surface makes
between two neighboring segments, and the angle each segment makes with each of
the two support columns. Hence there are two tensorial elastic energy contributions
depending on the vectors (xi − xi+1 ) and (xi − xi−1 ) for the surface bending, and
(xi − xi+1 ) and (xi − xbase ) for the column-surface angle:
(2.38) Ebending = K(yT z, y2 , z2 ) ,
where the variables of the function K are the invariants, like y T z, y2 , and z2 ,
which can be made of the two relative vectors y and z. The parameters appearing
in the function, will depend on whether it represents the surface bending or the
surface-column shear. See Figure 2.5 The position and orientation of the column
or a surface element will not influence the elastic energy. The invariants single out
the orientation-independent deformations.
Each of these terms can be determined in a limited and more detailed static
analysis. Afterward all the elements are combined to a single energy, from which
the Hamiltonian is constructed by inverting the mass matrix.
In Chapters 8 and 11 the elements of the finite element methods are constructed
along the same lines, where in the latter chapter, Chapter 11 the energy is indeed
constructed in terms of invariants. The simplest linear elements are most easily
constructed, but elements satisfying some physical laws, such as orientation invari-
ance, may predict the correct behavior much more accurately. The numerical code
Spacar, to predict and control the behavior of industrial robots using the orientation
of different points on a robot arm, is a good example, although the implementation
is not in terms of invariants. [Waiboer et al., 2005]
2.5. Infinite-dimensional systems

Infinite dimensional systems come in four main categories: First, the countable
infinite systems, which are made up of a countable set of finite systems. Second,
the compact infinite systems, defined by a finite number of functions on a compact
domain. For example the complex functions f (z) with on the unit circle |z| = 1 is
such a compact infinite system, which may be represented by a trigonometric sum
{einz }n=−∞
n=∞
. Third, the infinite systems on an infinite domain. Such systems will
cause all kind of mathematical and physical problems, if the so-called “behavior
at infinity” is not specified. The domain z ∈ Ω on which the functions f (z) are
defined in any of the latter two categories is called the reference space.
Finally, there are ensemble systems, which arise by taking some finite-dimensional
system given by the state x(t) and evolving it for an infinite number of initial states
x(0) ∈ S, where S is a set from the infinite-dimensional state space. The properties
of the domain S make little difference in the treatment of ensemble systems, as
usually, for t → ∞, the domain {x(t)|x(0) ∈ S} will not be compact and often very
complicated. Chaotic systems are usually analyzed as ensemble systems. The spe-
cial domain S of a lower dimension than the original space in which x resides, such
that {x(t) ∈ S|x(0) ∈ S ∧ t ∈ R+ } is called an attractor, which, in nonlinear dy-
namics, can have a very complicated form. It corresponds to the subspace in which
0
a system will end up in many cases. The Poincaré map S∆ = {x(t + ∆t)|x(t) ∈ S}
is an important tool to recover attractors. Mathematicians like Arnold [Arnold,
1973] and Smale [Hirsch and Smale, 1974] have investigated dynamics as mappings
of a domain onto a space.
The first three systems: countable, compact infinite, and infinite are the sys-
tems we will consider here. Using mathematical tools and restrictions one can map
systems downward, from infinite, to compact infinite, to countable. In the appro-
priate setting the numerical results of the countable system will converge fast at
finite truncation.
The work of Weierstrass, Schauder, Leray, and others has established a clear
link between compact infinite and countable. [Gohberg and Goldberg, 1981] For
linear systems, Hilbert or Banach spaces, a basis for a function space, on a compact
domain, would link a large class of functions to a countable basis, the Schauder
basis. Typically a set of eigenmodes of vibration would be a useful countable basis
for the displacements of a compact object. However, such countable basis is the
result of proper conditions on the operator underlying the vibration dynamics,
which are not always satisfied. In the case of nonlinear operators there are few
guarantees for a connection to a countable basis.
Functions on an infinite domain suffer from more mathematical difficulties.
Intrinsically complicated physical problems, such as the irreversibility in thermo-
dynamics, the decay of quantum states, or radiation into open space, will manifest
itself in the mathematical description, typical for infinite-dimensional systems on
an infinite domain. For example, radiation will radiate off to infinity. Any measure-
ment, corresponding to a compact operator, will yield eventually zero, but without
energy or mass loss in the system. Irreversible systems connected to a heat bath
suffer from the same complications in connection with mathematics. These prob-
lems are also related to the problems around the direction of time. [Rosenfeld, 1961]
The resulting irreversible systems require different modelling principles. [Breedveld,
1984]
2.5. INFINITE-DIMENSIONAL SYSTEMS 33
Apart from these, more or less, fundamental distinctions among infinite sys-
tems, there are many mathematical distinctions among infinite systems, character-
ized by the treatment of the system, rather than its nature. It is important to note
that, for example for a partial differential equation, there is little to say about the
solution, if one does not specify the space in which this solution lies. This space will
be considered in this thesis more freely than most approaches in functional analysis
prescribe.
In functional analysis, one starts with the notion of “size” in the form of a
norm k · k of a function. Several norms exist, and different norms may be used in
conjunction. For compact infinite systems the norm of a function is usually the
integral over this function:
Z n1
n

(2.39) kf kn = |f (x)| dx
,
Ω
which gives rise to Hölder’s inequality:

Z

(2.40) f (x)g(x)dx ≤ kf kp kgkq ,

Ω
for any p, q ≥ 0 : p−1 + q −1 = 1. Therefore the norm k · k2 is special, as one can

associate a bilinear inner-product, or dot-product, to it:
Z
(2.41) hf, gi = f ∗ (x)g(x)dx ,
Ω
where f is the complex conjugate of f (if f is complex), such that hf, f i = kf k22 =
∗
kf ∗ k22 . This space is called the Lebesgue space; any function in this domain is
f ∈ L2 (Ω). [Rudin, 1966]
Important to the notion of norm is the fact that zero norm means the zero-
function:
(2.42) kf k = 0 ⇒ f = 0 .
Rather than a result, this implication should be read as a definition of the zero-
function in a particular space. For integrals many functions are equivalent, in the
sense that their difference is a zero function: kf − gk = 0.
The norm turns out to be very useful, as most global properties of a system,
such as the mass and energy, can be expressed as norms. However, without any
sense of “distance” on the reference space Ω, a system will just be a very large
collection of independent systems. In Table 2.1 different notions of distance are
listed. The notion of distance gives rise to the notion of collective behavior: the
behavior f (z) at two points z1 and z2 will be similar if these points are close
together. The behavior is correlated. Note that this distance is in the reference
space Ω, and not the infinite function space. The spatial derivative gives rise to
this notion. Higher values for the derivative means less correlation. The notion of
collective behavior means that ∂z f (z) is bounded in some sense. Hence, it might be
more appropriate to start with a mathematical function space which is tied closer
to a definition of a spatial derivative ∂z .
Instead of using norms, other mathematical views of functions are a more natu-
ral means to consider differentiability. One could consider analytic functions, which
are infinitely differentiable on its domain Ω. If the domain is simply connected, such
name topology
point set subsets x ∈ S0 ⊂ S
metric distance ρ(x, y) ≥ 0
normed norm kxk ≥ 0 with ρ(x, 0) = kxk
Banach linear norm kλxk = |λ|kxk
Hilbert inner product hx, yi with hx, xi = kxk2
linear algebra mass matrix M with hx, yi = xT My
Table 2.1. Closed and complete spaces in analysis. Completeness
means that a converging series kxi − xi+1 k → 0 has a limit kxi −
xk → 0 within the space. Dynamical systems fall naturally in
the category of normed spaces, where the energy E = kxk ≥ 0.
For mathematical analysis more structure, like a Banach space or
Hilbert space is more useful. For numerical analysis, a finite linear
algebra, at least locally, is necessary.
that is has no holes, a small neighborhood of a single point on the domain is re-
quired for the evaluation. The value at all other points z ∈ Ω result from analytic
continuation. Complex functions have many other advantages, since integration
and differentiation are naturally defined on complex functions. The consistency re-
lations for complex analyticity, the Cauchy-Riemann relations, are even defined as
differential equations. See, for example, Section 5.5, where the boundary conditions
for the Cauchy-Riemann equations and the Laplace equation are investigated.
However, analyticity is a very strict condition, many functions will not satisfy
this property, and, for example, no function on a compact support, such that it is
only nonzero on a compact domain Ω, can be analytic. There are means around this
problem, by considering only finite domains, which lead to, e.g., Fourier series and
Hardy spaces, however, the class of finite differentiable functions C n (Ω) will be in
practice the function space of choice. It only requires the necessary condition needed
to perform the calculations. It is a strictly local property: in every neighborhood
of a point z ∈ Ω, the n-th derivative of a function f ∈ C n (Ω) is continuous. The
derivative of a function can be seen as an algorithmic or operational property.
Looking at differentiation is such an abstract sense, defining “∂z f (z)” rather than
calculating it, avoids some of the pitfalls of real analysis. The use of differential
operators in finite function spaces is a possible alternative. See Sections 2.8 and
11.4.
Real functions of real variables are a possible beehive of stings. Weierstrass has
learned us as much. It is possible to define the most horrible functions, which are
nowhere continuous, or differentiable. [Gelbaum and Olmsted, 1964] However, the
functions in this thesis are the mathematical representations of physical systems,
therefore, we assume they behave nicely, at least, at the scale we look at them.
Furthermore, they are the result of some external input, and eventually we will
discover that the quality of the input, we are free to choose, will determine the
quality of the functions representing the system under investigation.
2.6. Notions of distance, approximation, and change

Continuous dynamical systems are based on three notions, related to three
forms of continuity: the continuity of the underlying space z ∈ Ω ⊂ Rn , the
2.6. NOTIONS OF DISTANCE, APPROXIMATION, AND CHANGE 35
continuity of the states, functions or fields, on that space, f (z) ∈ Rm , and the
continuity in time; the dynamics A. These three notions are intimately related.
Without the underlying space Ω the dynamics consists of infinitely many copies
of finite-dimensional dynamics for each of the positions z. The differential equa-
tions couple these systems through the notion of distance: kz − z 0 k ≈ 0, however,
regularity, and collective motion arise only if the states change only slowly with
position:
kf (z) − f (z 0 )k
(2.43) lim0 = finite ,
z→z kz − z 0 k
which is typical for linear, Banach spaces. However, more closely related to the
notion of energy, E = kf k is the finite variation in the form:
kf (z)k − kf (z 0 )k
(2.44) lim0 = finite .
z→z kz − z 0 k
From the global norm, kf k, over the whole domain, for example kf k = maxz∈Ω kf (z)k,
the local norm kf (z)k can be reached through the use of either some test, or mea-
sure, functions, or a variational principle:
(2.45) g(z) = f (z) + h(z) ,
where h(z) is some localized variation, around z and z 0 , and an infinitesimal
quantity. It uses the Taylor expansion in and thereby the tools of linear analysis.
However, generally, the energy E = kf k, and the fields f (z) on the domain
Ω have a strained relation. Many approaches are based on some form of regu-
larity. A priori finite approximations are maybe the most successful approaches.
[Kantorovich and Krylov, 1958, Funaro, 1992, Creutz, 1983] They form the basis of
Galerkin methods and finite element methods.
Sobolev spaces are another method to introduce regularity in the functions. The
modern theory of existence and uniqueness of elliptic partial differential equations
is based upon the theory of Sobolev spaces. Not only the values of the functions
are bounded, but also the variations, given by the derivatives up to order l. The
typical Sobolev norm, defining the Sobolev space Wpl (Ω) is:
l Z
! p1
X
(2.46) kf kWpl = |∂zi f (z)|p dz <∞ .
i=0 Ω
In particular for p = 2 the Banach space has an inner product and defines a Hilbert
space, with all its nice consequences. For formal results such as existence and
uniqueness [Hörmander, 1990] the Sobolev space has been instrumental. In order
to find some finite approximation fd (z) of a solution f (z), such that kfd − f k < ,
these methods have limited use. Eventually, Weierstrass approximation theorem,
[Yosida, 1980] which proves that a continuous function on a finite domain can be
approximated by a polynomial, in the local sense kfd (z) − f (z)k < , remains one
of the strongest results for those who view functions in the classical sense, i.e., as a
graph (z, f (z)). The lifted approaches of jet spaces, and fiber and vector bundles
can be seen as a geometrical extension of this view. [Abraham et al., 1991] The lifted
approach is particularly useful in the study of integrability, if the state variables do
not correspond to the configuration variables. [Ivey and Landsberg, 2000] See also
Section 4.3. indexintegrability
If a linear normed space, or Banach space, is separable, i.e., can be expanded in

a countable set, it can be mapped, to a continuous function C[0, 1] on the domain
[0, 1], endowed with the standard norm: kf k = maxz kfd (z) − f (z)k. [Liusternik
and Sobolev, 1961] For Banach spaces, Schwartz spaces and other distribution, or
generalized function, spaces, have proved also useful in combination with Fourier-
Laplace transformations.[Gel’fand and Shilov, 1964, Shilov, 1968, Schwartz, 1966]
With the change, or dynamics, of a system there seems little formal problem.
The integrability condition of the kernel, or evolution operator, A, is very simple,
and a wide variety of operators may be integrated without problem, formally or
explicitly, considering the correct measure and integration theory, typically Stieltjes
integration, sometimes Lebesgue integration for stationary and variational problems
which are more easily dealt with in Hilbert space:
(2.47) x(t) = eAt x0 = Tt x0 .
The complication arises from the interference between the approximation of the
space Ω and, in particular, the approximation of the fields x = f (z). They interfere
with the desired features of the time-evolution. For example, in the case of energy
conservation kxk = E = constant, the time-translation operator Tt must be a norm-
preserving unitary operator, kx0 k = kTt x0 k, in short, kTt k = 1 in operator norm.
[Gohberg and Goldberg, 1981]
Besides regularity, homogeneity can help to establish results. One of the most
successful treatments of linear dynamical systems is based on the time-translational
invariance of the system. The system can therefore be described by a semi-group:
0
0
(2.48) Tt+t0 x = Tt Tt0 x = eA(t+t ) x = eAt eAt x .
The continuity in time follows from the basic properties of the semi-group, e.g.,
T0 = 1. [Yosida, 1980, Curtain and Zwart, 1995]
On the other hand the dynamics of the system dictates a relation between
the space discretization and the time discretization of a system. The well-known
Courant-Friedrichs-Lewy relation determines that the step size must be smaller
than the mesh size divided by the characteristic velocity for stability. [Lax and
Wendroff, 1960, Courant and Hilbert, 1961, Morton and Mayers, 1994] Hence at
different levels the continuities are intertwined.
From the abstract continuous systems there are many routes to a variety of
finite-order numerical models. Not only the approximations, but also the order in
which they are made, will determine the results. It is important to structure these
approximations, and the different notions of continuity.
2.7. Variational methods

In general, variational methods are a powerful method to find functional so-
lutions. In dynamics, they yield an alternative formulation to the equations of
motion. In optimization they yield an explicit implementation of multiple criteria
in a single cost function, possibly augmented with Lagrange multipliers. Together
with a variational principle, an optimal solution can thus be found. [Carathéodory,
1935, Lanczos, 1966, Mikhlin, 1964, Troutman, 1983]
In this thesis extensive use is made of the energy functional, i.e., the energy as
function of the state functions or fields. To link this global property to equations of
2.7. VARIATIONAL METHODS 37
motion, variational principles and variational derivatives are required. See, among
others, Sections 5.10, 7.5, and 5.3. In this section we will give some background
to these methods. Variational methods were introduced by Johann Bernoulli and
Euler. They stated that any physical problem can be formulated as an extremal
problem of some kind.
Historically, with variational principles the modern analysis of static and dy-
namical systems started. The variational form, such as the minimum or maximum
of a function, allowed one to transform the problem into a convenient form, in order
to find the solution. Coordinate transformations, and model reduction, are there-
fore closely related to the variational form of problems described by differential,
and partial differential, equations. Still such formulations prevail in many of the
descriptions of physical systems. Some notable exceptions are, for example, Fara-
day and Gauss, who, each in their own way had a more visual and geometrical view
of the physical systems they studied. In Chapter 12 we establish some connections
between the variational and the geometric views. In some respect is the variational
principle more fundamental than the formulation in terms of equations of motion
or equations of stability.
We start with a simple example of a line x(z) as a function of z ∈ [−1, 1]. A
variational principle would be the shortest path between x(−1) = x− and x(1) =
x+ . It is common knowledge that the shortest path is a straight line:
1 1
(2.49) x0 (z) = x− (1 − z) + x+ (1 + z) .
2 2
However, to derive this result is less than straightforward.
The path length is the integral:
Z 1 p Z 1 p
(2.50) J[x] = 2
dz (∂z z) + (∂z x) = 2 dz 1 + (∂z x)2 .
−1 −1
If a path x(z) is the minimum, any variation x(z) → x(z)+δx(z) means an increase
of length. The minimum path satisfies that an infinitesimal variation δx, with
δx(−1) = δx(1) = 0, vanishes:
(2.51) J[x + δx] − J[x] = 0 ,
which yields, by expanding the integral and partial integration, the differential
equation:
Z 1
δx 1
(2.52) J[x + δx] − J[x] = dz p ∂2x = p ∂2x = 0 .
1 + (∂ x) 2 z 1 + (∂ x) 2 z
−1 z z
Since is for any δx(z) the integral should vanish, the factor must vanish for all z.
Hence, the straight path x0 (z) is a solution to this equation. Such a solution
is called a geodesic, and commonly the, rather unseeming, square root is removed
to yield the standard action principle x(z) ∈ Rn :
1
Z
(2.53) J[x] = dz ẋT ẋ .
2
The z coordinate is interpreted as time coordinate, with ∂z x = ẋ.
Many different variational principles exist. Commonly, not the initial and fi-
nal positions are given, but just the initial position x(−1) and the initial velocity
x+
x−
x(z)
x−
x’−
Figure 2.7. The equivalent descriptions of a line: the values x−

and x+ of both ends, the value x− and the derivative x0− at one
end, and the linear function x(z).
∂z x(−1). However, the same variational principle yields the same differential equa-
tion, with a different set of boundary conditions:
(2.54) ∂z2 x(z) = 0, x(−1) = x− , ∂z x(−1) = x0− ,
yielding the same straight line, in terms of different conditions:
(2.55) x0 (z) = x− + (z + 1)x0− .
The mapping between different sets of boundary conditions, such as between x− , x+
and x− , x0− , is an important aspect of this work. In this case the map is linear and
straightforward:

x− 1 0 x−
(2.56) = .
x+ 1 2 x0−
See Figure 2.7.
The geodesic is not always a straight line. The metric tensor g is the geometric
equivalent of the mass matrix, which may depend on the position x:
1
Z
(2.57) J[x] = dzgij ẋi ẋj .
2
For example, the geometric theory of light only takes into account the refraction in-
dex n(x), which yields a change of velocity, due to varying densities of the traversing
medium. The functional for a geometric light ray is:
1 1
Z
(2.58) J[x] = dz ẋT ẋ .
2 n(x)2
For example, in the atmosphere light rays are deflected by warm air with a lower
refraction index n.
A solution x(t) which minimizes the time, or in this case the approximate
functional J[x], is called a geodesic. It is a special case n−2 δij = gij of the general
metric tensor gij . The corresponding geodesic differential equation is:
(2.59) ∂t2 xi + Γijk ∂t xj ∂t xk = 0 ,
2.8. DIFFERENTIAL OPERATORS 39
ct+z ct−z
ct−z ct+z
u(z,t=1)
u(z,t=1) time
u(z,t=0)
z
u(z,t=0)
Figure 2.8. The two extreme cases of initial conditions, and the
corresponding time-evolution along the characteristic directions,
given by the arrows. The two functions represent the initial value
u(z, t = 0) and the initial time-derivative u̇(z, t = 0). The left
case corresponds to only a non-zero value, the right case to only
a non-zero derivative. Both correspond to the superposition of
two triangular wave fronts moving away in opposite directions.
Separating the boundary conditions into u± (z, t = 0) = c∂z u(z, t =
0)± u̇(z, t = 0) leads to left-moving and right-moving parts, related
to scattering variables.[John, 1982]
where the Christoffel symbols Γijk are given by:

1 im
(2.60) Γijk = g (∂xk gmj + ∂xi gmk − ∂xm gjk ) ,
2
The geometric theory of light, using only the direction of the ray given by geodesics,
is an approximation of the wave theory of light. [Born et al., 1980] In the theory of
hyperbolic partial differential equations, or wave equations, the dynamics of the ray
correspond to the dynamics of shock waves or singularities, which retain their form
during propagation. The path of propagation is described by the characteristics of
the wave equation: [Hörmander, 1990]
(2.61) ∂t2 y(t, z) − c2 ∂z2 y(t, z) = 0 ⇒ characteristics : ct = ±z ,

where c is the wave velocity. Characteristics have extensive use in the theory of
partial differential equations. In the case of linear equations with constant coef-
ficients, the characteristics are simple. See Figure 2.8. In many other cases the
construction of characteristics is as complicated as solving the hyperbolic equations
themselves. [Carathéodory, 1935, Kamke, 1962]
For a long time it was thought that wave equations, of the form above, were
fully characterized by tracing back the rays along the characteristics, but the reality
turned out to be more complicated, depending on the dimension of the space.
[Gårding, 1998]
2.8. Differential operators

A number of near-equivalent formulations of classical electrodynamics exists.
[Barut, 1981] The Lagrangians differ only by a total divergence, a term which can
be written as:
(2.62) Ltotal divergence =∇·S .

In practice, it means that the action integral will only have contributions from such
terms at the boundary, through the divergence theorem:
Z Z
(2.63) ∇ · Sdn z = n · Sdn−1 z → 0 .
Ω ∂Ω
In electrodynamics the boundary is usually set at infinity, where the term is ex-
pected to vanish, giving rise to the equivalence. Rather than dealing with cleverly
chosen boundary conditions, we study here boundary control; the effects of bound-
ary conditions and the variations of them. Hence such near-equivalent formulations
in terms of Lagrangians differing only a total divergence Ltotal divergence are rather
ambiguous. The energy of the field will lead to a consistent definition to which the
boundary conditions can be related.
The connecting elements are the differential operators. [Lanczos, 1961, Gilles,
1973] Adding a total divergence to a Lagrangian allows one to rewrite the equations,
ignoring the boundary terms:
(2.64) (∂z φ)2 − ∂z (φ∂z φ) = −(φ∂z2 φ) .
The second term on left-hand side is a total divergence, which allows us to rewrite
the weak formulation, in terms of the energy or Hamiltonian density, to the strong
formulation, in terms of the differential equation, using variational principles.
Another manner to express the relation between the weak and the strong for-
mulation is the use of formal adjoint differential operators. The adjoint operator D ∗
arises partial integration under the integral ignoring the boundary terms, previously
represented by the total divergence:
Z Z
(2.65) dzφ(z)∂z ψ(z) = − dz(∂z φ(z))ψ(z) + boundary terms .
Hence D∗ = −∂z is the formal adjoint of D = ∂z .

The boundary terms can be seen as a functional on the boundary, the trace
operator, or the boundary operator θ. In Table 2.2 we list a number of well-known
differential operators and their boundary or trace operators. The difference between
the operator D and its formal adjoint D ∗ is the boundary operator:
Z Z
(2.66) hf, Dgi − hg, D ∗ f i = (f Dg − gD∗ f )dn z = f θgdn−1 z = (f, θg) .
Ω ∂Ω
For example, the curl operator D = ∇× maps a vector-valued function f back onto
a vector-valued function. A bi-linear functional J[f , g] may be constructed with
the curl operator:
Z Z Z
T
(2.67) J[f , g] = g (∇ × f ) = ijk gi ∂j fk = ijk (∂j (gi fk ) − fk ∂j gi ) .
Ω Ω Ω
Z Z
(2.68) = nT (f × g) + f T (∇ × g) .
∂Ω Ω
Hence, for D = ∇× the formal adjoint is D ∗ = D = ∇× and the boundary
operator is θ = n×. The Poynting vector, expressing the electromagnetic energy
flow through the boundary is an example of such a boundary operator. [Jackson,
1975]
Combining adjoint operators with variational equations, the differential opera-
tor D can be inserted to the variation, leading to the relation:
(2.69) δφ H(φ) = D∗ δDφ H ,
2.8. DIFFERENTIAL OPERATORS 41
D D∗ θ
∂z −∂z 1
∂z2 ∂z2 ∂z − ∂z†
∇ ∇· n
∆ ∆ n T ∇ − n T ∇†
∇× ∇× n×
Table 2.2. Differential operators, D, their formal adjoints, D ∗
and the boundary operators θ, yielding together the Green identi-
ties. The dagger † indicates the differential operator acts on the
left: f ∂z† g = (∂z f )g
which is equivalent for fields φ which do not vanish under differentiation Dφ 6=

0. This statement can be made more explicit if we use a particular basis j =
{0, 1, 2, · · · }:
1 j
(2.70) φj (z) = z ,
j!
and a particular differential operator:
(2.71) D = ∂zn ⇒ D∗ = (−1)n ∂zn ,
such that Dφj = φj−n for j ≥ n and Dφj = 0 for j < n. Hence, the variational
derivative with respect to φj includes φj with j < n, while the variational derivative
with respect to Dφj does not:

j<n 0
(2.72) [δφj H]i ⇒ [D∗ δDφj H]i = ,
j ≥ n [δφj H]i
where the projection [f (z)]i of f (z) onto the state φi is achieved through the use
on a dual basis {φ∗i }:
(2.73) [φj ]i = hφ∗i , φj i = δij ,
which must be based on the exact integral inner-product, on a fixed domain, for
example z ∈ [0, 1]:
Z 1
1
(2.74) Nij = hφi , φj i = dzφi φj = ,
0 (i + j + 1)i!j!
such that:
X
(2.75) φ∗i = Nij−1 φj .
j
Simply said, the use of the differential operator in the variational derivative corre-
sponds to the removal of the corresponding terms of the Taylor expansion of the
original function f (z) = δφ H:
1
(2.76) f (z) → f (z) − z∂z f (z)|z=z0 − z 2 ∂z2 f (z)|z=z0 · · · .
2
The arbitrariness in the polynomial subtraction, apparent through the indetermi-
nate z0 , are precisely the degrees of freedom not determined by the variational
derivative of δDφ , but determined by other means, usually boundary conditions or
as constants of motion.
Later, in Chapter 7 we will use these methods to single out the part of the
Hamiltonian which should be determined by the input through boundary condi-
tions. The differential operator D is set to be the same as the leading differential
operator in the Hamiltonian. In this manner, the arbitrariness, i.e., the null-space,
of the Hamiltonian is determining the input and the corresponding boundary con-
ditions, rather than boundary conditions being the starting point of a proper de-
scription of continuous dynamical system. For example, this allows us to use the
same Hamiltonian for a component in the case it is freely moving, and in the case
it is fixed. In the first case, the arbitrariness is a conserved momentum, set by the
initial conditions at the starting time. In the second case, the arbitrariness is fixed
by the boundary conditions for every time.
2.9. Complex coordinates

The use of complex coordinates can be very practical in a number of cases, due
to the inherent structure of analytical functions. The complex variable z = x + iy
is two dimensional. [Conway, 1978] An analytical function f (z), is a function that
can be differentiated with respect to z infinitely many times, in most of the complex
space, excluding possible singular points or singular lines. The function f (z) is also
two dimensional:
(2.77) f (z) ≡ f (x, y) = f1 (z) + if2 (z) ,
where fi is a real function. Analyticity requires that the f (x, y) : R2 → R2 satisfies
the Cauchy-Riemann equations:
∂f1 ∂f2 ∂f2 ∂f1
(2.78) = and =− .
∂x ∂y ∂x ∂y
In most cases only one of the two functions, f1 or f2 , is of interest. Therefore the
reduction of the analytical function to this quantity consists of projecting onto the
real (Re) or the imaginary (Im) part. In the case of time-dependent results, the
time dependence is given by the complex phase eiωt . In this case the amplitude is
given by the modulus of a complex function |f |.
It has some interesting consequences, which makes analytical functions useful
for describing two dimensional, and 2n dimensional, systems. First, each of the
fi (x, y) satisfies the Laplace equation:
∂ 2 fi ∂ 2 fi ∂ 2 fi
(2.79) ∆fi = = 2
+ =0 ,
∂z∂z ∂x ∂y 2
for the very simple reason that fi (z) is not a function of the complex conjugate
coordinate z = x − iy.
Second, any analytical function f (z) can serve as a conformal mapping, or sym-
plectic coordinate transformation from z → z 0 = f (z). Such coordinate transfor-
mations are angle preserving, and an analytical function g(z) will also be a solution
of the Laplace equation in the new coordinate: ∆z0 g(z 0 ) = 0. Since x and iy are
perpendicular, and can be seen as the potential surfaces and the streamlines of a
dynamical system, so will x0 and iy 0 be. The presence of a singularity in the trans-
formation z → z 0 , e.g., z 0 = z + z −1 , will cause the presence of a circulating flow
in the new coordinate, if none was present in the original coordinates. The princi-
ples of conformal mapping are used the solve the Laplace equation and potential
problems in two dimensions in all kind of geometries, such as wedges, cylindrical
2.10. BIPOLAR COORDINATES 43
cavities. Furthermore, Joukowsky (Zhukovsky) and Kutta derived independently

an important result for incompressible flows, without friction or viscosity, for wing
sections, which basically states that any wing section will do for such limited model.
[Tokaty, 1971]
Third, the Cauchy theorem is Stokes theorem in two dimensions, and gives
us the means to invert the Laplace operator and solve inhomogeneous differential
equations in two dimensions. The Cauchy theorem, also known as the residue
theorem, tells us the relation between the line integral clockwise over a closed
contour Γ = ∂Ω, or the boundary of a domain Ω, and the values at poles in the
domain:
Z X
(2.80) f (z)dz = 2πi ∂z ((z − zj )f (z))|z=zj .
Γ poles zj
For example, for f (z) < ∞ for z lying inside the contour z ∈ Ω:
f (z 0 ) z 0 ∈ Ω

1 f (z)dz
Z
(2.81) = .
0
2πi Γ (z − z ) 0 z 0 6∈ Ω
Hence it tells us, that if we know a function f (z) on the boundary Γ, we can
construct a solution f (z) for z ∈ Ω, which satisfies the Laplace equation: ∆f = 0,
using the integral above. It is known as the Dirichlet problem. The function (z −
z 0 )−1 is also known as the Green function for this problem, converting a differential
equation into an integral equation. Care must be taken to understand the form
f (z)dz correctly. This is the complex product, and written out in terms it is the
exterior product for the real part of the expression:
(2.82) f (z)dz = (f1 + if2 )(dx + idy) = f1 dx − f2 dy + i(f1 dy + f2 dx) .
In the chapter of partial differential equations, Chapter 5, we will discuss the con-
straints on the function f (z) on the boundary z ∈ Γ = ∂Ω, as and example of
problems with boundary conditions.
In quantum mechanics the complex functions are more than a practical tool.
The wave function in quantum mechanics is complex. Only our observations are
real, and therefore a measurement is a projection on the real axis in the complex
plane.
In higher dimensions there is no unique extension of complex spaces. Three
important extensions exist. First, the quarternions, which is a three dimensional
extension of the skew symmetry, similar to the generators of the rotation group.
Second, the projection of a 2n dimensional space on n complex planes. In the sense
of Darboux theorem such a projection exists for nonsingular symplectic geometry,
however, in general such projections are an active field of research. Third, the
mapping on an Euclidean norm for the modulus, with some winding number as
complex argument, or phase.
2.10. Bipolar coordinates

A typical nontrivial example of a conformal mapping of complex coordinates
is the mapping which produces bipolar coordinates. The solution of the Laplace
equation ∆f = 0 in the case of a two-center problem, or a problem between eccentric
cylinders yield easy solutions in terms of bipolar coordinates.
Bipolar coordinates ξ, η give the Cartesian x, y-position in the plane: [Korn

and Korn, 1968]
α sinh ξ
(2.83) x= ,
cosh ξ − cos η
α sin η
(2.84) y= .
cosh ξ − cos η
It can be seen as the conformal mapping:

u+α u + α y y
(2.85) v = log = log
+ i arctan − arctan ,
u−α u − α x+α x−α
with u = x + iy and v = ξ + iη. [Spiegel, 1964, Magnus et al., 1966] This is also
known as a bilinear conformal mapping. The linear conformal map is the Möbius
transform which maps lines and circles to lines and circles. See also Figure 12.4.
The Laplacian is
1
(2.86) ∆x,y f = (cosh ξ − cos η)2 ∆ξ,η f .
α2
Hence the coordinates ξ and η themselves are solutions of the Laplacian, except at
the singularities ξ = ±∞.
The bipolar coordinates have several uses. For example, the electrostatic field
between two charged cylinders is most easily described in terms of bipolar coordi-
nates, and the Newtonian flow between two rotating eccentric cylinders is given by
the velocity potential in bipolar coordinates
In Section 12.3 the geometric aspects of bipolar coordinates are used the analyze
the material dependency of the transmission of force through a plane.
2.11. Infinite matrices

From continuous systems, through some set of basis functions, one arrives at
infinite vectors. These are the common form of pre-numerical representations of
continuous systems. Such vectors usually lie in the Hilbert space called l2 . [Gohberg
and Goldberg, 1981] The infinite vectors are truncated in the numerical evaluation,
where the length of the truncated vector can be varied to test convergence of the
result. However, as we shall see in this section, the theory of infinite vectors and
the linear operations thereon, in the form of infinite matrices, is not completely
trivial.
The simplest class of operators on the infinite space are infinite matrices, which
can be infinite in two senses, half-infinite and double-infinite:
.. .. ..
 
  . . .
a00 a01 · · ·
 
 .. 
 a10 a11  . a 00 a 01 · · · 
(2.87)  .
 

.. .. ..
  . .

. . .  · · · a10 a11
 . 

.. .. ..
. . .
This infinite dimension has a number of unexpected consequences, in the sense that
cutting a couple of dimensions of an infinite dimensional space, yields an infinite
dimensional space which can be mapped onto the original space.
2.11. INFINITE MATRICES 45
Considering the half-infinite matrix, the shift operator:

 
0 1
 0 1 
(2.88) S=  ,
 
 0 1 
.. .
. ..
transforms a row (a0 , a1 , a2 , · · · ), into a row (0, a0 , a1 , · · · ). The adjoint operator
S ∗ , does the opposite: S ∗ (a0 , a1 , a2 , · · · ) = (a1 , a2 , a3 , · · · ). Hence, the kernel of
S ∗ is the space (a0 , 0, 0, 0, · · · ), and the index associated with the operator S is:
[Gohberg and Goldberg, 1981]
(2.89) indexS = dim kerS − dim kerS ∗ = −1 .
For the reduction to a finite dimensional space it is appropriate to split the spaces
such that the infinite dimensional operator has index zero.
For example, the differential ∂z operator can be written as a shift operator S
on the basis:
     
1 1 0
 z   z   1 
 1 2   1 2   
 z   z   z 
 21 3  ∗ 2
(2.90) ∂z  z  = S  1 z 3  =  1 z 2  .
  
 61 4   61 4   21 3 
 24 z   24 z   6 z 
     
.. .. ..
. . .
Note that the S operator on the coefficients vector corresponds to the adjoint
operator S ∗ on the basis elements.
However, if the monomials z n /n! are a basis of L2 ([−π, π]), the appropriate
representation would be the trigonometric series:
(2.91) ∂z {sin(n + 1)z, cos nz}∞ ∞
n=0 = {(n + 1) cos(n + 1)z, −n sin nz}n=0 ,
such that the null space cos nz = 1 can be separated from the infinite-dimensional
subspace on which the operator ∂z has an index zero. In the study of boundary
conditions, Section 5.5, the index will arise as well.
Another feature of operators on infinite dimensional spaces the more restrictive
norm. For example, the interchanging of summations, no problem for finite sums,
leads to different answers for the matrix: [Gelbaum and Olmsted, 1964]
1 1 1
 
0 2 4 8 ···
 −1 0 1 1
··· 
 21 1
2 4
1

A =  41 − 21
 − 0 ··· 
(2.92) 1
2  ,
 −
 8 −4 −2 0

.. .. ..

..
. . . .
lead to two distinct double sums:
  !
X X X X
(2.93)  Aij = 2 6= −2 =
 Aij .
i j j i
As can be seen from the differential operator ∂z above, many operators associated
with partial differential equations (PDE’s) are unbounded, i.e., might yield very
large results, in some norm k · k, for the traditional functions and function spaces.
[Gohberg and Goldberg, 1981] For example the function f (z) = |z|α for 0 < α < 1
is a well-behaved continuous function of z, but its derivative is singular. A large
part of research on PDE’s is devoted of constructing function spaces S appropriate
for functions f (z), such that the image Df (z) of the differential operator D still
yields well-behaved functions, in the sense kf kS < ∞ ⇒ kDf kD < ∞. Sobolev
spaces and Schwartz spaces are examples of such restricted function spaces, with
well-behaved images. [Hörmander, 1990]
Quite often, in numerical evaluations and simulations, one has no other choice
than to truncate the space, first from continuous to countable, and subsequently to
finite. See, for example, Chapter 8. The solution space is thereby regularized. The
functions are restricted to some well-behaved set. In nonlinear partial differential
equations, discontinuities and singularities may arise, which can no longer be found
in the numerical simulation. [Hörmander, 1990] Another problem are spurious
solutions, or intruder states. The name arises from the discovery of this phenomena
in large matrix calculations in nuclear shell theory. The collective behavior of a
number of functions, each separately irregular enough to be discarded from the finite
approximation of the system, may yield some smooth and global dynamics. Such
states may arise from systems described either by matrices which large off-diagonal
coefficients in A, like in Eq. 2.92, or, similar, from systems with small differences
between the diagonal coefficients. In the mathematical theory of matrices, such
systems are ill-conditioned. [Wilkinson, 1965]
The high density of states, or modes, of the eigenspectrum of systems, which
plagues the conditioning of the resulting matrix, is common to all system of two
or more spatial dimensions. The density of states in the eigenspectrum for a given
frequency grows with the power of dimension: ∼ ω dim . Even for small coefficients
in the off-diagonal part of the matrix, like in Eq. 2.92, the sheer number of off-
diagonal coefficients can yield a large truncation error. The local form of the partial
differential equation, makes a finite difference approximation near diagonal, which
has the advantage of a reduced number of off-diagonal coefficients, and therefore
better conditioning.
2.12. Krylov bases

From the large matrices of FEM, Chapter 8, or the infinite matrices from an
expansion in basis functions, in Section 2.11, there are many different ways to arrive
at a small lumped model, which contains the dominant modes. [Antoulas, 2005] In
Chapter 9, the physical and modelling aspects of this reduction are discussed. From
the numerical perspective, the Krylov bases are an important tool in the reduction.
[Freund, 2003]
The Krylov basis Kn (A, X) is a basis generated through the successive op-
erations of the operator A on the vector x, or on the collection of vectors X =
{x1 , x2 , · · · , xm }:
(2.94) Kn (A, X) = {xi , Axi , A2 xi , · · · , An xi } ,
which will be denoted by:
(2.95) Kn (A, X) = {xji }j=n,i=m
j=0,i=1 .
It has the main advantage that the operation A on the basis is near trivial:
(2.96) Axij = x(j+1)i .
2.12. KRYLOV BASES 47
However, the basis is not orthogonal, except in special cases.

The Krylov basis is closely related to the moment problem for integral opera-
tors. An operator or function, for example in L2 can be expanded into moments:
Z
(2.97) Ii (f, g) = µ(z)dzf (z)z i g(z) .
The question is how much information of can be recovered of f and g, or f , given

g, using only the moments Ii . [Widder, 1941, Akhiezer, 1965] The original question
by Stieltjes was whether for f (z) = g(z) = 1, the measure µ(z) > 0 could be
reconstructed from the sequence Ii . [Stieltjes, 1894-1895]
Many function transformations are based, in some sense, on the moment ex-
pansion. These transformations can be inverted in a particular function space. Well
known are the class of almost everywhere (a.e.) functions for transformations, like
Fourier, Laplace, and Hankel transformations, based on the Lebesgue integral on
a compact domain. Such functions are considered identical if they differ only on a
discrete set of points.
Krylov space, with a single vector X = v is defined as
(2.98) KN (A, v) = {v, Av, A2 v, · · · AN v} ,
which lies at the basis of Lanczos or Arnoldi tridiagonalization. In the case of
a symmetric matrix, AT = A, it can lie at the basis of tridiagonalization. The
corresponding algorithm picks a vector v0 , applies it to the matrix A, and projects
back on v0 , the remainder is a new vector v1 , which applied to A, yields three
parts: parallel to v0 , and to v1 and perpendicular to both vectors: v2 . After the
next step, we find that Av2 is perpendicular to v0 , in the case of a symmetric
matrix.
v0T Av0
(2.99) v1 = Av0 − v0 ,
kv0 k2
such that v0 ⊥ v1 . Similarly for v2 , we subtract the projections onto v0 and v1
from Av1 :
v1T Av1 v0T Av1
(2.100) v2 = Av1 − v 1 − v0 .
kv1 k2 kv0 k2
From a lengthy algebra it can be determined that v0 Av2 = 0. However, since v2 is
perpendicular to both v0 and v1 , by construction, it follows for a symmetric matrix
that:
v0T Av0 T
(2.101) v0T Av2 = v2T Av0 = v2T v1 + v v0 = 0 .
kv0 k2 2
which is zero by construction, since v2T v1 = v2T v0 = 0. Hence the tridiagonal
form of A. The vectors V = {v0 , v1 , · · · vn } form an orthogonal basis for the
Krylov space Kn (A, v), for any n. The conjugate gradient method, described in
Section 10.3, uses the same principles. In that case the inner-product is viT Avj = 0
for i 6= j. The actual matrix under investigation, in that case, is A−1 . [Stoer and
Bulirsch, 1980]
In a given basis {v0 , v1 , v2 , · · · , vn } the matrix A is represented as:
(2.102) Avij = viT Avj .
The matrix constructed by successive orthogonalization has a symmetric tridiagonal

form:
 
α 0 β0 0
 β0 α 1 β1 0 
(2.103) VT AV = Av =  0 β1 α2 β2 0  .
 
 
.. .. ..
. . .
This is the Lanczos tridiagonal form of a symmetric matrix.
2.13. Some remarks on state space and spatial reference space

In this chapter we were concerned mainly with the variables, the states, of the
dynamical model. The properties of continuous, or infinite-dimensional, systems are
also determined by the properties of the underlying, or reference, space. The two
are linked through differential geometry and partial differential equations, which
is discussed in the Chapters 4 and 5. The existence of a state for a given local
energy density is, in the language of differential geometry called the integrability.
The state of a conserved current is a central concept from differential geometry.
Furthermore, the state variables of a component in this chapter are, as yet, not
invariant under coordinate transformations, which is also discussed the Chapter 4.
The dynamics of the state space is discussed in more general terms in the
following chapter. Were the models of dynamical systems in this chapter mainly
isolated, in the following chapter the input and output will play a central role.
Furthermore, we will take a closer look at the simulation, or time-integration, of
dynamics, i.e., the evolution in time.
CHAPTER 3
Evolution
The sciences do not try to explain, they hardly ever try to inter-
pret, they mainly make models.
[J. von Neumann]
Linear models of dynamical systems, with input and output, are often brought
to a canonical form, before numerical analysis and simulation. In system engineer-
ing, this canonical form is often the state-space model, or input-state-output model,
or the A, B, C, D-model. [Polderman and Willems, 1998, Curtain and Zwart, 1995,
Nijmeijer and van der Schaft, 1990] Physical dynamical systems are only a subset
of all systems that can be brought in the input-state-output form. In the more
mathematical literature, the output is deemed less relevant, and the same model is
referred to as the model for system evolution, i.e., the motion in time. [Hirsch and
Smale, 1974]
The states of input-state-output model are the necessary internal degrees of
freedom, which can be the charge on a capacitor in an electric system, or the position
of a mass in a mechanical system. The state is often seen as the configuration,
independent of any input. For example, the input force may cause an acceleration,
but does not determine the start position or velocity. These are states, however,
only if the position or velocity is relevant for the dynamics. Some parts of the
configuration may not be relevant for the dynamics. For example, in the language
of Einstein’s relativity, experiments carried out on a moving train will have the
same outcome, as experiments carried out on solid ground. Hence both the global
position and the global velocity of the experimental setup are irrelevant, and they
not required as states.
The generic input-state-output system we will call an evolution system. The
dynamical systems of the previous chapter are a subclass of the broader set of evo-
lution systems. However, the absence of certain physical principles, such as energy
and momentum conservation, will lead to poorly defined models, which might de-
scribe certain behavior well, but it yields less insight in the internal dynamics of a
system. For example, dissipation is an integral part of such models. Prying away
the dissipation part from the oscillatory part is often done a posteriori in order
to be able to analyze stability.[Willems, 1972a] Furthermore, the physical basis of,
for example, dissipation is usually absent as A, B, C, D models do not start with
a detailed analysis of the system at hand, but are postulated and later adjusted
to fit the dynamics. In the extreme case of this limited insight in the model, the
model is treated as a black box, where the number of states and the parameters of
the input-state-output model is fitted to experiments. This process is called system
identification. [Ljung, 1999] The identification is usually with a linear A, B, C, D-
model, since nonlinearities will make identification much more difficult.
49
50 3. EVOLUTION
3.1. Input-state-output, or A, B, C, D, model

A general, but linear, model is usually called cryptically the A, B, C, D model.
[Brockett, 1970] The matrices A, B, C, D represent the model with the state vector
x and the input and output u and y:
(3.1) ẋ = Ax + Bu ,
with the output:
(3.2) y = Cx + Du ,
where D is the direct input-to-output part. The matrix A determines the evolution
of the homogeneous system. The input matrix is B with the input vector u, while
the output matrix is C. One essential ingredient of dynamical systems is retained,
which is the invariance under time-translation. If an experiment is done today or
tomorrow, under invariant conditions, the outcome should be the same, hence A
does not depend on the time t. Hence, we can still use t = 0 as the initial time at
which the system has to be characterized for a unique evolution. For a set of first
order equations, such as the A, B, C, D model, only x(0) needs to be given.
Linearity is a severe restriction on a dynamical system. In many cases it can be
solved, yielding simply the time-dependence of the state x(t). In the case of numer-
ical solutions of very large systems it is usually the only option. Nonlinearity makes
it very hard to solve such systems. Furthermore, it is difficult to select the right
solution, from the multiple solutions that arise from nonlinearities. Later, once
physical principles are introduced, more may be said about nonlinearities, however,
for many numerical implementations the system is analyzed in a small neighbor-
hood around the state x(t) under consideration, such that one still resorts to an
appropriately linearized A, B, C, D form, i.e., the local decomposition. [Isodori,
1985]
The state x(t), in Eq. 3.1, given the initial conditions x(0), is easily integrated:
Z t
At −Aτ
(3.3) x(t) = e e Bu(τ )dτ + x(0) .
0
However, in practice, if A is a large, or infinite, matrix or some abstract operator,
it might be difficult to recover x(t) from the equation above. Numerical integration
methods, described in Section 3.4, are often used to determine the evolution of x(t).
In terms of the control vector u several operators can be distinguished. The
steady-state operator A−1 B yields, if it exists, the state corresponding to a sta-
tionary input:
(3.4) x = A−1 Bu .
In the more general case of time-dependent control, using the space L2 ([0, t]) the
end-point adjustment operator yields the optimal control u(t0 ) for the change of
the free evolution of the state: xfree (t) = eAt x(0) to the desired end-point x(t):
T
(t−t0 )
(3.5) u(t0 ) = BT eA W −1 (t)((x(t) − eAt x(0)) ,
where the controllability Gramian W(t) is given by: [Skogestad and Postlethwaite,
1996]
Z t
0 T 0
(3.6) W(t) = eA(t−t ) BBT eA (t−t ) dt0 .
0
3.1. INPUT-STATE-OUTPUT, OR A, B, C, D, MODEL 51
The input is restricted to the same time-dependence as the states themselves, which
made the self-dual formulation possible.
Another common operator associated with the state space representation is the
Laplace transformation [Widder, 1941] of the system evolution:
Z ∞
At
(3.7) G(s) = L[Ce B] + D = e−st CeAt Bdt + D = C(s1 − A)−1 B + D .
0
In practical applications of control engineering, the right-hand side is commonly
referred to as the transfer matrix from input u to output y. [Ogata, 1990]
Although, the dynamical systems under consideration are more general in the
sense that nonlinearities occur, and more restrictive, in the sense that the matrix
A must be Hamiltonian, such that the energy is conserved, it is useful to keep the
A, B, C, D description in mind. In a small neighborhood of a particular state x,
the system can be approximated by a A, B, C, D. The state space can therefore be
characterized by properties of the matrices A, B, C, D, rather than by “dynamics”
in the abstract sense. Regions can be indicated by the ranks of the observability
and controllabity matrices, as the typical invariants of the linear system.
Occasionally the matrix E is introduced:
(3.8) Eẋ = Ax + Bu .
In the case the matrix E is singular, every independent vector g in the left null-
space of E would corresponds to a constraint g T Ax + gT Bu = 0 on the state
vector x. Since the system is linear, it should be no problem to solve the constraint
explicitly, and use a lower dimensional state. Conceptually, E is unwanted too.
A state x should define the system. Every component of the state is an initial
condition required for the evolution. With a singular E, initial conditions x(0) with
gT Ax(0) + gT Bu 6= 0 will not lead to a well-defined evolution. Initial conditions
for partial differential equations are further discussed in Section 5.3.
On the other hand, the singular E or A will arise from the internal dynamics of
distributed, or continuous systems, described by partial differential equations. The
null-space corresponds to degrees of freedom not fixed by the internal dynamics.
In the case of a discrete approximation of such systems it is important to retain
the null-space. Depending on the view the left null-space of A corresponds to the
constants of motion, such as the momentum of an isolated system or the conserved
charge, while the right null-space corresponds to boundary conditions, which act
as constraints on the internal dynamics. Hence the interaction of the system with
its environment determines the view, and it may not always be proper to reduce
the dimension such that A or E is nonsingular, if the setting is not specified. In
the port-Hamiltonian approach, discussed in Chapter 7, the null-space is isolated
but retained, such that, depending on the interaction of the system, the variables
in the null-space can act as input, i.e., constraints, or output, i.e., constants of
motion, or a more complicated interactive and dynamical response, which we call
port-variables.
Linear systems can be brought to the A, B, C, D form. Most physical, mechan-
ical dynamical systems are in their native form described by the Euler-Lagrange
equations, and are consequently written as second-order equations. They depend
on the the position q and the velocity q̇, to determine the evolution in terms of the
acceleration q̈. To turn the second-order equations into a set of first-order equations
the velocity is used as an independent variable. In the simplest approach v = q̇,
52 3. EVOLUTION
however, other choices of independent variables exist, such as the momentum p(q̇).
The state vector is therefore a combination of position and a function of velocity
x = (q, f (q̇)).
In mechanical systems E usually incorporates the mass of the system. However,
the Hamiltonian formulation, or the Lancaster scheme, makes it possible to bring
the mass to the left-hand side. The Lancaster scheme turns the Newton equation:
(3.9) Mq̈ + Dq̇ + Kq = F ,
into a first-order equations with x = (q, q̇):

D M K 0 F
(3.10) ẋ + x= .
M 0 0 −M 0
In this case the matrix E is:

D M
(3.11) E= .
M 0
To bring the system into the A, B, C, D form, the matrix E in front of ẋ must be
inverted, which is possible as long as the damping D is much smaller than the mass
M. The Lancaster scheme E matrix is a special case of the E matrix that arises in
constrained minimization, encountered later.
The Hamilton equations are a similar reduction of the second-order equations
to a coupled set of first-order equations, however, in a more systematic manner,
such that it can be generalized to nonlinear systems. See also Chapter 2. The
Hamilton system is automatically the core of the A, B, C, D system:

ṗ 0 −K F
(3.12) = ẋ = x+ .
q̇ M−1 0 0
In this case the augmented x vector does not contain the velocity, but the momen-
tum p = Mq̇.
Hamilton formulated these equations of motion based on the conserved en-
ergy, or Hamiltonian. The contribution of Jacobi has been to remove superfluous
variables from these equations using coordinate transformations and generalized
coordinates. A similar procedure was carried out by Lagrange to arrive at gener-
alized coordinates for Newtonian mechanics. The names of Hamilton and Jacobi
are also connected to the equation of the generating functional S for the canonical
coordinate transformation: [Goldstein, 1980]
(3.13) H(q, ∇q S, t) + ∂t S = 0 .
The conserved energy in the Hamiltonian dynamics will lead to a natural def-
inition of the output, corresponding to the input force F. In the absence of an
applied force, the Hamiltonian H is constant in time:
1 1
(3.14) H = pT M−1 p + qT Kq .
2 2
In the case of an applied force F, the energy varies. The equations of motion can
be used to express the change in energy in terms of the state variables q and p:
(3.15)
Ḣ = ∇q H q̇ + ∇p H ṗ = qT KM−1 p − qT KM−1 p + pT M−1 F = pT M−1 F = q̇F ,
where is used that both the stiffness matrix K = KT and the mass matrix M = MT
are symmetric. Hence the change of energy, or power, supplied to the system
3.1. INPUT-STATE-OUTPUT, OR A, B, C, D, MODEL 53
corresponds to the product of the applied force and the velocity. The natural
output matrix C is therefore:

0
(3.16) C= .
M−1
The spectral method is a common approach to simplify the evolution of a

system. It consists of decomposing the system into independent modes, through
diagonalization of the matrix A:
(3.17) A = V† ΩU ,
where, V and U are respectively the left and right eigenvectors and Ω the diagonal
matrix consisting of the eigenfrequencies. The dagger † represents the Hermitian
conjugate of the matrix: A† = AT . For example, the 2 × 2 symplectic matrix has
complex eigenvectors U and eigenvalues:

0 1 1 1 1 i 0 1 1 −i
(3.18) =√ · ·√ .
−1 0 2 i −i 0 −i 2 1 i
In the absence of degenerate eigenfrequencies the normalized left and the right
eigenvectors are trivially related:
(3.19) V† = U−1 .
Furthermore, if A is symmetric, the matrix U is orthonormal, representing a rota-
tion. The exponentiation of a symmetric matrix A boils down to the exponentiation
of the eigenvalues of Ω:
(3.20) eAt = V† eΩt U .
The eigenvalues ω for stable dynamical systems lie in the complex half plane Reω ≤
0, while for an energy conserving system, they consist of pairs ω = ±iλ.
This spectral composition does not always exist, which is a common problem
for analysis of linear systems. The general matrix A can be diagonalized up to
Jordan blocks: [Wilkinson, 1965]
 
J1
(3.21) A=
 ..  ,

.
Jn
with a single block
 
λ 1 0 0
 0 λ 1 0 
(3.22) Ji = 
 0
 ,
0 λ 1 
0 0 0 λ
where the block can be of arbitrary length. Instead of an evolution given by the
sum of exponentials, the evolution of the Jordan block evolution matrix Ji will be
given by a polynomial P (t) times the exponential:
(3.23) x(t) = Vi P (t)eλt U†i x(0) ,

54 3. EVOLUTION
where Vi and Ui correspond to the projection on the eigenvectors of Jordan block

Ji . For example, the four-dimensional Jordan-block arises from the differentiation
of the polynomial vector {ti eλt }i=3
i=0 :
    
1 λ 1 0 0 1
∂  1t 2  eλt =  0 λ 1 0   1t 2  eλt .
   
(3.24)
∂t 
2 t   0 0 λ 1  
2 t 
1 3 1 3
6 t 0 0 0 λ 6 t
The exponentiation of a Jordan block, in order to explicitly solve the time evolu-
tion of an input-state-output system, is less trivial than the exponentiation of the
diagonal matrix, Eq. 3.20. The exponential of J ∈ R5×5 is given by:
2 3
λeλ λ2 eλ λ6 eλ
 λ 
e
2
(3.25)
 0
eJ =  eλ λeλ λ2 eλ   .
 0 0 eλ λeλ 
0 0 0 eλ
The typical off-diagonal coefficient, indexed (i, j), is given by: λj−i eλ /(j − i)!.
In the case of infinite-dimensional systems, such a decomposition is not possible.
There might not be a finite set of eigenvalues of the operator A, although in many
cases, given certain conditions of A, certain conclusions might be drawn. For the
analysis of infinite-dimensional systems the analyticity of the resolvent RA is often
examined:
1
I
(3.26) RA = dzA ,
Γ z−A
which appears in all kind of scalar forms, such as in the Gramian above, Eq. 3.6.
The complex variable z is more general than the variable of the Laplace transform,
Eq. 3.7, since particular singularities of the spectrum of A can be singled out by an
integration contour around them. See also Section 5.7. The Cauchy theorem can
be applied, as often is for the inverse Laplace transform, for a closed contour Γ:
1
I
(3.27) WA,B = BT Bdz .
Γ z−A
If a singularity is encircled, or a cut is crossed, by the contour Γ, the resolvent
yields the spectral properties of these singular points. [Kato, 1966]
3.2. Continuous spaces and the (A, B, C, D)-system

In the later chapters continuous systems described by partial differential equa-
tions are discussed. The state space x and the operators A, B, C, and D are no
longer finite vectors in Rn and matrices in Rm×n . Instead the state space is a func-
tion space, with the operators acting on them. [Riesz and Nagy, 1955] Eventually,
for numerical evaluation using the methods of linear algebra, the model must be
reduced to, or approximated by, a simple form such as the (A, B, C, D)-form.
The partial differential equations have initial conditions and boundary con-
ditions. See Section 5.1. In the finite-dimensional state-space representation, the
initial conditions correspond to the state space x, for which an initial state x(t = 0)
is required. The boundary conditions become the input matrix B and the input-to-
output matrix D, where the actual value of u is the finite-dimensional representa-
tion of the boundary value. In some cases there might be a part of the state vector
3.3. OBSERVABILITY AND CONTROLLABILITY 55
associated with an input. For example, a force acting on an inertia, chances the ve-
locity and the position, which are the two states associated with a force input. For
consistency these states should be brought into the core model; the model directly,
algebraically, related to the boundary conditions. See Section 7.6. Depending on
the actual choice of boundary conditions; the causality, these states are either al-
gebraically related to the boundary conditions, or state variables of integrating or
buffering elements, consequences of the boundary conditions.
The (A, B, C, D)-form of the core model of the finite approximation of a partial
differential equation is therefore not a priori fixed. It depends on the type of
boundary condition is used. It is the analogue of a lumped mass, which can be
controlled either by single position, or velocity, or more naturally by force control.
The number of states are respectively: zero; one; or two, i.e., none; the position; or
the position and the momentum. For the lumped case, position control is rather
unnatural. The continuous system static analogues are for position the Dirichlet
boundary condition and for force Neumann boundary condition, which are both
quite natural. The velocity control does not correspond to any static situation, it
lies in the more complicated domain of hyperbolic system, discussed in Section 5.3.
The aspect of the number of states, or equivalently the constraints on states, of
continuous systems was already raised in the Introduction, in Section 1.7. In the
port-based modelling [Paynter, 1961, Breedveld, 1984, Karnopp et al., 2000] it is
more natural to retain these states and associate them with dynamical states, with
initial conditions, or algebraically relate them to input, depending on the situation.
3.3. Observability and controllability

The observability and controllabity criteria are the rank conditions on: [Kalman
et al., 1969]
h i
(3.28) n = rank CT , AT CT , AT 2 CT , · · · AT (n−1) CT ,
n = rank B, AB, A2 B, · · · An−1 B ,

(3.29)
where n is the dimension of the state space.

The conditions imply that the projections, through the left and right eigenvec-
tors, of the input matrix B and the output vectors C, should not be perpendicular
to any of the Jordan blocks, for the controllability and observability.
The dynamics of a block is given by an arbitrary polynomial times the leading
exponential P (t)e−λt , Eq. 3.22. The coefficients can be determined by the same
number of time-derivatives at a time t, as the order of the polynomial. Hence, the
size of the Jordan blocks determines the order of the differential equation.
The observability and controllability conditions are also closely related to the
Krylov expansion. In a sense, for a finite system, the Krylov expansion, using A
and the input, or start, matrix B, will only span the controllable subspace of the
full dynamics. If the full dynamical subspace is much smaller that the full state
space, one can use it to one’s advantage to reduce the system. These principles are
explored in Chapter 9.
56 3. EVOLUTION
3.4. Integration
Simulation boils down to the time-integration of dynamical equations. For
numerical integration many schemes have been developed. From the known schemes
we will treat only three very briefly, afterwards we discuss more general problems
and approaches to integration.
For Hamiltonian systems, integration methods which conserve energy, i.e., geo-
metric integration, [Hairer et al., 2002] also guarantee stability, as the energy norm
of the system serves as the measure of stability. Apart from general integration
methods, we will discuss mainly the use of energy conservation to improve the
numerical integration schemes. [Krenk, 2006]
For time-integration a differential equation is mapped to a set of coupled first-
order equations:
(3.30) ẋ(t) = f (x(t)) ,
which immediately gives rise to the Euler integration scheme, nowadays known as
the forward Euler scheme:
(3.31) x(t + δ) = x(t) + δf (x(t)) forward .
The backward Euler scheme is implicit, in the sense that it uses the function eval-
uation at the forward time: f (x(t + δ)):
(3.32) x(t + δ) = x(t) + δf (x(t + δ)) backward .
The problem of the implicit scheme is that is requires knowledge of x at t + δ in
order to determine x at that point in time.
The error of an integration scheme depends on the application. Say f (x) is a
constant vector, independent of x, the result is exact for any step size δ. However,
typically one would compare the error to that of a Taylor expansion:
(3.33) x(t + δ) = x(t) + x1 δ + x2 δ 2 + · · · .
Hence, both the Euler integration schemes yield a δ 2 error at each step.
The most used, and best-known, integration scheme is the Runge-Kutta [Stuart
and Humphries, 1998, Press et al., 1992] fourth-order integration. It requires a
number of function evaluations, which map out the integral over the finite domain
[t, t + δ], which is integrated as accurate as possible:
(3.34)
δ
x(t + δ) = x(t) + (f (x(t)) + 2f (x(t) + ∆1 ) + 2f (x(t) + ∆2 ) + f (x(t) + ∆3 )) ,
6
where the predictions of the state at intermediate times are generated recursively:

δ δ
(3.35) ∆1 = f x(t) + f (x(t)) ,
2 2

δ δ
(3.36) ∆2 = f x(t) + f (x(t) + ∆1 ) ,
2 2
with the final step:
(3.37) ∆3 = δf (x(t) + δf (x(t) + 2∆2 )) .
Dynamical systems do not explicitly depend upon time, since they are usually time-
invariant. Therefore we have ignored the typical explicit time-dependence which
appears in many ordinary differential equations and their integration schemes. The
3.4. INTEGRATION 57
q(p) q(p)
q(p)
p
q(p)
q q
p q
p(q) p(q)
p(q)
q
Figure 3.1. The leap-frog integration uses successively the values

of the momentum p to determine the change in position q, and the
new position q to determine the change in the momentum p.
fourth-order Runge-Kutta integration requires only four function evaluations to

determine both the intermediate steps ∆i and the values f at these positions, while
higher orders require more function evaluations than the particular order.
Besides basic analytic principles to improve the accuracy of the integration,
there are other concepts which are used in time-integration. For the integration
of the Hamilton equations of motion, which preserve the energy of the system,
symplectic integration schemes has been developed to preserve the energy as well
as possible. One simple, but powerful method is the leap-frog integration. The
symplectic system consists of pairs of canonical variables: positions and momenta
q and p, or magnetic and electric variables. The leap-frog method consists of
alternating Euler integrations of each set with the new values of the other set:
δ δ δ
(3.38) p(t + ) = p(t − ) − δhq (q(t), p(t − )) ,
2 2 2
δ
(3.39) q(t + δ) = q(t) + δhp (q(t), p(t + )) .
2
The staggered values t ± 2δ for the momenta p are to make apparent the effect
of the alternated integration. One integration uses the middle value of the other
integration. See Figure 3.1.
The leap-frog integration is a large improvement upon the standard Euler in-
tegration for Hamiltonian systems, since it uses, in some variance for the implicit
form, a midpoint scheme, also used for symplectic systems: [Hairer et al., 2002]

x(t + δ) + x(t)
(3.40) x(t + δ) = x(t) + δf ,
2
which corresponds to the use of a tangential direction x(t + δ) − x(t) at the point
of evaluation x(t+δ)+x(t)
2 .
The leap-frog integration uses the additional information known about the sys-
tem. In many cases of higher dimension more information is known. One could, for
example alternate a number of leap-frog steps with a single corrector step which
corrects the state xold such that the energy E corresponds again to the starting
energy E0 :
(E − E0 )∇x H(xold )
(3.41) xnew = xold − .
|∇x H(xold )|2
This would be the smallest correction step xnew − xold due to the steepest descent
∇x H direction. Although, no formal justification for such a correction step exists,
58 3. EVOLUTION
the results in the section below, on second-order Hamiltonian integration will lead
to the similar results.
3.5. Second-order Hamiltonian integration

For very large systems determining higher-order derivatives is often prohibi-
tively expensive. In the chapter on finite element methods we will discuss iterative
approaches for sparse matrices. In many cases it is computationally favorable to
restrict the integration to a second-order method. Furthermore, in the case of FEM,
the time-evolution does not only depend on the history of a single node xi (t), but
also, and maybe even more so, on the history of the neighboring nodes xi−1 (t)
and xi+1 (t). All the nice properties, which hold for one-dimensional results, i.e., a
single function y(t), do not hold for high-dimensional dynamical systems. Gener-
ally, the behavior of a component xi (t), treated as a single function y(t), depends
more on the rest of the state vector x, than on the history t0 < t of that particular
component. In the one-dimensional case, on the contrary, all depends of xi (t0 ) for
t0 < t.
The integration of Hamiltonian systems is given by the Hamiltonian H, and
the symplectic matrix J:
(3.42) ẋ = J∇x H ,
we can integrate these equations of a short time δ, assuming J∇x H is constant for
the short time-space [t, t + δ]:
Z t+δ
(3.43) x(t + δ) − x(t) = J∇x Hdt ≈ δJ∇x H(x(t)) .
t
However, we can improve the result by using a Taylor expansion of the variational
derivative around x(t), i.e., by using the equations of motion to approximate the
change in J∇x H(x(t)):
Z t+δ Z t+δ
(J∇x H) dt0 ≈ δ (J∇x H) + J∇2x H ẋ (t − t0 )dt0
t t
δ2
(3.44) = δJ∇x H + J∇2x HJ∇x H .
2
Hence, for a finite time the Euler integration x(t + δ) = δJδ ∇H(x(t)) + x(t), the
symplectic structure is adjusted, to first order by:
δ −1
δ −2M I
(3.45) Jδ = J + J∇2x HJ = ,
2 −I − 2δ K
where M and K are the linearized mass and stiffness matrices. The diagonal part
corresponds to a model dependent correction in the direction of the steepest descent:
∇x H. For a finite step size δ in the numerical integration scheme, the symplectic
0 1
matrix J = (−1 0 ) should be replaced with Jδ in order to attain only second-order
errors in integration.
The matrix Jδ is the first correction to change the generator J of symplectic
rotations into the rotation matrix Jθ over a finite angle θ, based on the local cur-
vatures of the energy surface, in the neighborhood of a point (q, p) of the phase
space.
3.5. SECOND-ORDER HAMILTONIAN INTEGRATION 59
The Hamiltonian in the neighborhood of (q, p) can be written as a quadratic

Hamiltonian, up to an additive constant which can be ignored:
1 T 1
(3.46) H= q Kq + pT M−1 p .
2 2
Therefore, a local coordinate system y can map this surface to the unit sphere:
√ !
K 0
(3.47) y = Ωx = √1
x ,
0 M
yielding a normalized Hamiltonian:

1 T
(3.48) H= y y .
2
Since KT = K and MT = M are both positive we can take the square root of these
matrices. In this coordinate system, any time evolution is a rotation R(θ(δ)):
(3.49) y(t + δ) = R(θ)y(t) .
Hence, inserting an explicit rotation, based on the standard symplectic structure,
being the generator of rotation, we find:

−1 cos θ − I sin θ
(3.50) x(t + δ) − x(t) = Ω Ω−1 Ω2 x .
− sin θ cos θ − I
Locally, by definition:
(3.51) Ω 2 x = ∇x H ,
such that the symplectic structure for the finite time-step reduces to Ω−1 (R−1)Ω−1 .
In the limit δ → 0, we find:
r r
K K
(3.52) θ = arcsin δ≈ δ .
M M
The finite-step symplectic structure δJδ , Eq. 3.45, also reduces to the same result
if expanded as a polynomial in δ to the second order:
 q q √ 
√1 (cos K
δ − 1) √1 √1 sin K
δ M
M K q M
(3.53) δJδ + O(δ 2 ) =  K√ q K
√ √  ,
K √1 K
− M sin M δ K M(cos M δ − 1) M
In practice, calculating the squareroot of a matrix, in Eq. 3.53, may be much more
cumbersome than the second order correction of Eq. 3.45.
Hence for quadratic Hamiltonians, the energy conservation can be made exact
by using the finite rotation θ, although the evaluation of the analytic functions
of the matrices is more expensive than the diagonalization, which allows for an
exact solution in terms of independent eigenmodes. This formulation is more useful
for general nonlinear Hamiltonians, which already require the evaluation of the
mass and stiffness matrix. These evaluations can be used to yield a second-order
integration method. Higher-order effects of the nonlinear Hamiltonians can be
expressed in the variations along the path of the mass and stiffness matrices, which
is discussed in the next section.
60 3. EVOLUTION
3.6. Higher order results

From the previous section it is clear there are several ways to proceed to higher
order results. In the case of a quadratic Hamiltonian, the exact rotation can be
implemented for diagonal mass and stiffness matrices. For non-diagonal parts, the
rotation still yields high-order corrections. In particular for FEM Hamiltonians we
can improve upon the result, as we will see later.
For nonlinear Hamiltonians, the mass and stiffness matrices are local approxi-
mations. A linear dependence can be fitted to both these matrices A = M−1 and
A = K:
(3.54) A(t + δ) ≈ (1 + δA0 (t))A(t) ,
and
(3.55) A−1 (t + δ) ≈ A−1 (t)(1 − δA0T (t)) .
We no longer rely on the state x dependence. Generally, we would expect the mass
and stiffness matrices, generically denoted by A + A0 t, to vary more in time slowly
and weakly than the state x(t). The linear approximation for these matrices would
in most cases yield a higher-order accuracy for the state.
In the simulation the matrices are extrapolated from the previous iteration.
Inserting this back into the time integral we find:
0T 0
!
− δM(1−M2
t)
I
(3.56) Jδ = −1 0T 0 .
−I − δK (1−K 2
t)
with the variational derivative contributing:

(1 + K0 t0 )∇q H

(3.57) δx H =
(1 + M0 t0 )∇p H .
yielding to second order:
0 0T 2
!
3 M(K −M ) 0
(2) − δM
2 +δ 4 1 + δ 2M
(3.58) Jδ = 2 −1 0 0T .
+ δ 3 K (M4 −K )
0 −1
1 + δ 2K − δK2
This result is already closer to that of methods such as Runge-Kutta, where the
state x is estimated at later times, to yield an improved mean derivative H(x(t)+∆)
over the step size domain [t, t + δ].
The expense, however, of determining and storing the linear dependence M0
and K0 of the mass and stiffness matrices, might not to be worth the trouble if
large systems are studied. Although the mass and stiffness matrices themselves
are usually sparse, no guarantee exists that along the path x(t) these matrices are
sparse too. The computational effort increases with the number of nonzero elements
of these matrices, if efficient iterative, sparse-matrix routines are used.
3.7. Integration and control

Simulation is about predicting a result based on a model expressed as a differ-
ential equation. Control is to generate a desired output of a system. Hence with
some knowledge of the system, expressed in a plant model, integration methods can
be used to predict the output and adjust the input such that the desired output
is achieved. A more competent model of the system will yield a better prediction.
The purpose of this simple set-point control, is to give an input, like an applied
3.7. INTEGRATION AND CONTROL 61
force, to change the actual output to a desired output. Even with little knowledge
of the plant model, such as inaccurate values for masses and springs, the output
can be brought to a desired value. The connection between feedback control theory
and numerical integration schemes has been fruitful in many different respects, for
example, in audio compression and signal transmission.
In many cases in control the dynamical behavior of a system or plant is approx-
imated by a second-order system. It corresponds to a mass-spring-damper model
for system. In discrete time two values x(ti ) and x(ti−1 ) are used to predict the
subsequent value x(ti+1 ). Consider the force-actuated, damped oscillator:
(3.59) mẍ + dẋ + kx = Fcontrol .
In the case of integration we can take small time steps ti+1 − ti = δ, and the
equation reduces to:
x(ti+1 ) − 2x(ti ) + x(ti−1 ) x(ti ) − x(ti−1 )
(3.60) m +d + kx(ti ) = F (ti ) .
2δ 2 δ
For a more advanced discretization and integration, the control can evolve with it.
In many cases the physical constants: the mass m, the damping d, and the stiffness
k are unknown. System identification tries to recover these constants from responses
to particular input F (t), and output x(t), and in the case that identification and
control are combined one speaks of adaptive, or learning, control. [Åström and
Wittenmark, 1989]
If we wish to control the state variable x(t) it means we wish to determine
x(ti+1 ) by adjusting F (ti ) such that |x(ti+1 ) − xdesired (ti+1 )| is as small as possible.
However, to do so, we should first predict what the result would be without applied
force:
(3.61)
x(ti ) − x(ti−1 ) x(ti )
x(ti+1 ) ≡ xδ (x(ti ), x(ti−1 )) = 2x(ti ) − x(ti−1 ) − dδ − kδ 2 .
m m
The simplest control is proportional control. The force proportional to the differ-
ence between the desired and the actual output:
(3.62) Fi = KP (xdesired (ti ) − x(ti )) .
In many cases this will drive x(ti ) to xdesired (ti ). However, a more advanced ap-
proach takes into account the velocity which will change x(ti+1 ) with respect to the
control and measure point x(ti ):
(3.63) Fi = KP (xdesired (ti+1 ) − (1 + δ)x(ti ) + δx(ti−1 )) .
= KP ((xdesired (ti+1 ) − x(ti )) − KD (x(ti ) − x(ti−1 )) .
The additional term is called the derivative action.
Finally, the stationary deviation of x(t∞ ) and xdesired (t∞ ) will remain finite,
even for a constant set point xdesired , due to the reaction force −kx(ti ):
(3.64) k(xdesired (t∞ ) − x(t∞ )) = F = KP (xdesired (t∞ ) − x(t∞ )) .
This can be corrected by adding an increasing, integrating, the gain KI , as the
difference persists: [Föllinger, 1974]
X
(3.65) Fi = KP (xdesired (ti+1 ) − x(ti ) + KI (xdesired (j + i) − x(j + i)) .
j=−∞
62 3. EVOLUTION
which allows one to increase the total gain substantially without the fast oscilla-
tions of the system blowing up. Without the derivative action, this constitutes a
proportional-integrating (PI) controller.
However, the time-dependent behavior will always trail behind the set-point
trajectory xdesired (t). These are the typical elements which make up a proportional
integrating differential (PID) controller. [Ogata, 1990] The integrating gain K I and
the differential gain KD are often expressed as their typical time constants τI and
τD times the proportional gain:
KP
(3.66) KI = , KD = K P τ D .
τI
If we would have had exact values for the system parameters m, d, and k, we
could have integrated the equation exactly, yielding:
m(xdesired (ti+1 ) − xδ (x(ti ), x(ti−1 )))
(3.67) Fi = ,
δ2
which would yield typically very large forces, if x 6= xdesired , since the time-step δ is
small. The integration method, which yields the result xδ , however, does not need
to be specified for the control purpose. The control based on the exact model, which
therefore requires no feedback, if x = xdesired at the start, is called feed-forward
control.
3.8. Time and time again

In this chapter, we established a simple model description of dynamical systems,
for the purpose to study the dynamical, or time-dependent behavior. Simulation
plays an important role in the study of complex interacting, or open systems. Their
behavior is not easily characterized, and control of dynamical systems are based on
the time-dependent behavior of these systems. For continuous systems, a proper
selection of modes, which lie at the basis of the A, B, C, D-model, is determined
by the behavior as the result of input. This problem is investigated in Chapter 9.
The simulation, or time integration, itself, and the implementation together with
constraints, linear and nonlinear, is investigated in Chapter 10. The particular
structure of the evolution operator A in the case of FEM, and how it can be used
for physical consistency is part of Chapter 8, in particular in Sections 8.5 and 8.7.
Eventually, we perceive time through the dynamical systems that surround us.
Therefore, dynamical models, time integration, and simulation are closely related,
and none of them can exist without the others.
The next two chapters on differential geometry and partial differential equations
take a closer look at the mathematics of continuous models. First, aspects of the
underlying reference space are examined, in Chapter 4. In the second chapter,
Chapter 5, the state space as function with boundary values is studied within the
theory of partial differential equations.
CHAPTER 4
Differential geometry
Infinite-dimensional systems only come to life as collective objects through the

fact they are continuous systems, i.e., the state variables are defined on some spatial
domain, the reference space. The consequences are manifold, and some of them are
directly related to the geometric aspects of the domain. They are expressed in
terms of differential geometry, linking the properties of the reference space, to the
properties of the state space. Furthermore, a port-based approach tries to distill the
invariant and conserved global aspects from a system, these aspects are intricately
related to transformations and conserved flows, based on the geometry.
In this chapter we aim to introduce some physical concepts related to differential
geometry, [Guggenheimer, 1963, Spivak, 1970, Felsager, 1981, Dubrovin et al., 1984,
Abraham et al., 1991, Ivey and Landsberg, 2000, Morita, 2001, Frankel, 2004] which
will be used later. Differential geometry is a large field of mathematical research. In
principle differential geometry deals with differential functions in higher dimensions.
Global differential geometry deals with the topology of objects and invariants such
as the winding or linking numbers. [Chern, 1989]
In this thesis we deal mainly with local differential geometry; the local, often
analytic properties of functions in higher dimensional space. Differential geome-
try arises in many different aspects of modelling. The integrability condition is
probably the most important aspect. For example the elastic energy is given in
terms of curvatures and bending. These are the natural variables for the energy,
but not for the configuration. Whether for all values of elastic-energy variables a
configuration actually exists, is an example of the question of integrability. See for
example Section 11.10. The actual construction of appropriate geometric energy
variables and Hamiltonians which are invariant, for example, under rotations and
translations, are the general study of Chapter 11, where it should be noted that we
seek invariant elements and invariant boundary variables. [Kron, 1934, 1965]
From the physical perspective the conservation of energy and the power conti-
nuity is an essential part of the modelling strategy. Furthermore, other conservation
laws will play an important part to construct consistent models. For that purpose,
the energy conservation is expressed as a local law, i.e., the energy continuity, which
requires the construction of an energy flux. See Sections 5.3 and 7.5.
4.1. Surfaces
A surface in three dimensional space can be seen as a set of three continuous
functions of two variables:
 
x(u, v)
(4.1)  y(u, v)  = x(u, v) ,
z(u, v)
63
64 4. DIFFERENTIAL GEOMETRY
touching circle with radius=1

2
y = x /2
0
-2 -1 0 1 2
Figure 4.1. The curvature of the parabola y = x2 /2 at x = 0, y =

0 is one. The circle touching the parabola has a radius r = 1/k = 1.
which assign to each point u, v of the surface a three-dimensional coordinate x =

(x, y, z).
Likewise, if the surface is not perpendicular to either the x or the y coordinate,
the surface coordinates u, v can be replaced by x, y, leaving a single relation for the
height z(x, y) as function of the projected surface position.
The invariants associated with a surface are the quantities which do not depend
on the orientation of the surface or the choice of local coordinates, like u, v or x, y.
The three invariants are the area and two curvatures. The area is given by the
integral:
Z
(4.2) A = |∂u x(u, v) × ∂v x(u, v)|dudv .
The cross product, in more modern mathematical language, is replace by the ex-
terior product dx ∧ dy(u, v), where the sign is given by the orientation of x, y in
terms of u, v. [Flanders, 1963]
At each point of the surface it also has two curvatures: the maximal curvature
and the minimal curvature. The curvature k of a line can be defined as the inverse
radius of a circle which touches a curve tangentially. For example, a parabola
y = x2 /(2r) has a curvature at x = y = 0 of k = 1/r. Take a circle of radius
r = 1/k, touching the x-axis, it can be approximated near x ≈ 0, y ≈ 0 by:
p x2
(4.3) 0=r− x2 + (y − r)2 ≈ y − .
2r
See Figure 4.1.
The two curvatures of a surface are related to the largest and smallest eigenvalue
of the Hessian, which describes the parabolic approximation of the surface around
a point x0 = x(u0 , v0 ) with a tangent surface V:
(4.4)
x = x0 +Vu (u−u0)+Vv (v−v0 )+Huu (u−u0)2 +Huv (u−u0)(v−v0 )+Hvv (v−v0 )2 .
4.2. FRENET-SERRET COORDINATES 65
The tangential directions V are

(4.5) Vu = ∂u x(u0 , v0 ) .
The Hessian consists of the second-order terms in the double Taylor expansion:
1
(4.6) Huu = ∂u2 x(u0 , v0 ) .
2
We choose rectified coordinates with the tangent direction of the surface in the
x − y-plane such that the Vu = ex and Vv = ey . The Hessian is a vector in the
z-direction, Hij = ez zij . The eigenvalues of the Hessian, or rather the invariants
of the Hessian H = detzij = λ1 λ2 and K = 21 Trzij = 12 (λ1 + λ2 ) are not invariant
under coordinate transformations, they are defined in this coordinate frame, how-
ever, other values for Vi = ei + ez zi with zi 6= 0 are easily accounted for. The
invariants are the Gauss curvature:
2
zxx zyy − zxy
(4.7) H= ,
(1 + zx2 + zy2 )2
and the mean curvature:
(1 + zy2 )zxx + (1 + zx2 )zyy − 2zx zy zxy
(4.8) K= 3 .
(1 + zx2 + zy2 ) 2
A surface with a zero mean curvature everywhere is a minimal surface. Films
of soap are the most popular examples of minimal surfaces, however, any system
where the energy is dominated by the surface tension has a minimal surface solution
as the static solution. [Chern, 1989] There is therefore a connection between the
variational problem of minimizing the area A of a surface Ω, given the boundary
∂Ω, and the mean curvature:
Z
(4.9) δA = δ dxdy|∂x x × ∂y x| = 0 ⇔ K|Ω = 0 .
Ω
The surface tension γ relates the area of the surface to its energy E = γA.
4.2. Frenet-Serret coordinates

For curves, a natural, local coordinate system arises from the curvature. These
are called the Frenet-Serret coordinates[Lipschutz, 1969, Guggenheimer, 1963], and
are normally denoted by the basis vectors (n, b, r). A curve is given by a single
variable s, the reference coordinate, which denotes the position x(s) along the curve.
The direction of the curve is given by the change of position as a function of the
change of reference position:
∂s x(s)
(4.10) n(s) = .
|∂s x(s)|
The direction of the curvature is given by the second derivative:
∂s n(s)
(4.11) r= ,
|∂s n(s)|
which is perpendicular to the curve direction n since:
(4.12) 0 = ∂ s 1 = ∂ s n2 = n T r .
The third direction is given by:
(4.13) t=n×r .
See also Figure 6.1. The triple (n, b, r) form an orthonormal basis, and can be
represented as a rotation matrix R: RT R = 1
(4.14) R = (n, b, r) .
The angular representations of the rotation matrix R can be used in analytic theory
of curves. The matrix is represented by a rotation axis ω/|ω| and angle |ω|:
(4.15) R = eJ·ω(s) ,
where the generators are:
 
0 ω3 −ω2
(4.16) J · ω =  −ω3 0 ω1  .
ω2 −ω1 0
The direction vector n = (1, 0, 0) yields the curvature ∂s ω3 while ∂s ω2 = 0 through
the definition, and the twist ∂s ω1 .
A curve is uniquely defined by the curvature ∂s ω3 and the twist ∂s ω1 . Fur-
thermore, the elastic energy of a curve can only be a function of these invariants.
On the other hand, it is difficult to recuperate the actual form x(s) of a curve
from these differential quantities. Some simple examples are the straight line for
∂s ω1 = 0, ∂s ω3 = 0, circle for ∂s ω1 = 0, ∂s ω3 = constant, and the circular helix for
∂s ω1 = 0, ∂s ω3 = constant.
4.3. Integrability
From the local properties of a function, usually expressed in terms of differential
equations, it might be possible to reconstruct the function. However, such result
might not always exists. In some cases the differential equations do not allow for
a solution, in other cases a solution may exist, but might be difficult to recover
in practice. The condition whether the curve or surface may be reconstructed
from a differential equation is called the integrability condition. It is a large and
diverse field of research, which received a boost after the discovery of integrability
of nonlinear systems, which yield an infinite number of conserved quantities. These
results are closely related to the inverse scattering problem. [Ablowitz and Clarkson,
1991]
We will discuss only some elementary, classical results, in integrability. A
general theory of integrability for nonlinear systems is absent. Furthermore, we take
the constructive approach. Rather than defining equations of motion, and asking
ourselves the question whether a generating Hamiltonian exists, we construct the
Hamiltonian. The state of the system, without dissipation, will stay in surfaces
with constant value of the Hamiltonian.
For a curve in three dimensions there is a unique one-to-one mapping from
differential invariants, curvature and torsion, to the shape of the curve x(s). For
surfaces in three-dimensional space and in general for differential formulations a
global, or even a local, solution may not exists. The Maxwell-Pfaff-Cartan consis-
tency relations are the best known example of a test of the existence of a solution.
Given the two differential equations:
(4.17) ∂x z(x, y) = p(x, y, z) ,
and
(4.18) ∂y z(x, y) = q(x, y, z) ,
4.4. EULER AND FROBENIUS INTEGRABILITY 67
they will define a surface z(x, y) if they satisfy the hidden, or compatibility equation:
(4.19) ∂y p + q∂z p = ∂x q + p∂z q .
In the absence of z dependence, this relation is the Maxwell reciprocity relation
in thermodynamics. The local condition at some point x, y, z implies the global
condition as well.
Another problem would be if the surface normal n at each point x, y defines
a surface uniquely [Chirgwin and Plumpton, 1964]. Lagrange has studied this
question and related the answer to the existence of total level surfaces. Lagrange
partial differential equation uses the surface normal n = (nx , ny , nz ) to define the
surface implicitly:
(4.20) nx (x, y)∂x z(x, y) + ny (x, y)∂y z(x, y) − nz (x, y) = 0 .
The surface defined by the differential equation can be constructed with the use of
two constraint surfaces:
(4.21) u(x, y, z) = constant, v(x, y, z) = constant ,
which satisfy the equations:
dx dy dz
(4.22) = = .
nx ny nz
For such surfaces it follows that:
(4.23) 0 = du = ∂x udx + ∂y udy + ∂z udz = ∂x unx + ∂y uny + ∂z unz .
A general function F (u, v) = 0 of the two independent streamline functions u and
v satisfies the differential equation including the Lagrange differential equation as
a factor:
(4.24) (∂z u∂u F + ∂z v∂v F )(nx (x, y)∂x z(x, y) + ny (x, y)∂y z(x, y) − nz (x, y)) = 0 ,
which follows from combining the ∂x F and ∂y F . Hence defining a surface from
its normal is related to the integrability condition of a vector fields defined by
infinitesimals:
(4.25) ny nz dx = nx nz dy = nx ny dz .
4.4. Euler and Frobenius integrability

Another integrability condition is associated with the construction of a potential
from a vector field. In electrostatics there are two equivalent descriptions of the
electrostatic field. One in terms of the electric field E, and one in terms of the
potential V . The two are related through:
(4.26) E = −∇V .
The question is, if given an arbitrary electric field E, there always is a potential V .
Since the electric field is rotation free, i.e., ∇ × E = 0, this is indeed the case. A
slightly more general condition is Euler’s condition: [Frankel, 2004]
(4.27) nT (∇ × n) = 0 ,
where the surface normal n = |E|−1 E is based on the vector field E.
The general setting of the Frobenius integrability theorem has many uses in
control.[Nijmeijer and van der Schaft, 1990, Slotine and Li, 1991, Pappas et al.,
1998] It is stated in terms of Lie derivatives, and states that a potential V , or a

solution of lower dimension of V = constant for multiple vectors fields can only be
found if the fields are involutive. A Lie derivative associated with a vector field E
is the differential operator:
∂ ∂ ∂
(4.28) L E = ET ∇ = Ex + Ey + Ez .
∂x ∂y ∂z
A set of differential equations based on the vector fields Ei :
(4.29) L E1 V = L E2 V = · · · = L Er V = 0 ,
only has a solution V the Lie-derivatives of the vector fields Ei lie in the subspace
spanned by these vector fields. The anti-symmetric product, Lie product or com-
mutator, of two differential operators LEi and LEj is again a differential operator:
(4.30) [LEi , LEj ] = LEi LEj − LEj LEi = ETij ∇ .
If for all i, j at every point of space independently holds:
(4.31) Eij ∈ span{E1 , E2 , · · · Er } ,
the vector fields are called involutive, and an integral solution V will exists.
4.5. Conserved flow

The conserved flows are central to stable dynamics. There exist many different
notions of stability. [Lyapunov, 1992, Willems, 1972a, Nijmeijer and van der Schaft,
1990, Worden and Tomlinson, 2001] In this thesis the stability is related to the en-
ergy norm. Therefore, stability, even in the case of nonlinear systems is guaranteed
by the conservation of energy. If the energy flux, the differential geometric version
of power in the continuous case, can be defined, it gives one a handle on treating
boundary conditions and initial conditions in a systematic manner.
However, conserved flow is more prevalent than only in energy conservation.
Incompressible flow is described by a velocity field v(x) which is divergence free:
(4.32) ∇ · v(x) = 0 ,
indicating that fluid cannot build up in any region of space; whatever flows in must
flow out. The integral version is called the divergence theorem:
Z Z
(4.33) ∇ · v(x)d3 x = n · v(x)d2 x ,
Ω ∂Ω
where, n is the surface normal. The theorem states that the total divergence of a
flow v(x) in a domain Ω is equal to the total flow through the surface ∂Ω of the
same domain. The former has little meaning in our macroscopic world, however,
the latter is nothing but the bookkeeping of the amount of fluid, associated with
the flow, in a domain Ω.
For example, the stress tensor σ is constructed, such that the divergence theo-
rem can be applied. Each of the force components Fx , Fy , and Fz of a stationary
object is in balance, which means that the sum of all the forces Fi applied to the
surface must add up to zero:
Z
(4.34) 0= Fi (x)d2 x .
∂Ω
4.5. CONSERVED FLOW 69
Hence, in order to describe the forces inside the object it makes sense to use a
divergence of a tensor:
X
(4.35) ∂j σji = fi ,
j
where fi is the body force, such as gravity f = −gρ, or inertial force: f = ρv̇.
The stress tensor σ is the microscopic implementation of the force-balance law.
Its properties and consistency conditions of the stress tensor have been a matter of
debate till far in the twentieth century.
The force balance is just one of the continuity equations. In their generic form
they are written as the equation for rate of change of the charge Q equals the
divergence of the flow J:
(4.36) Q̇ = −∇ · J .
The continuity equation is not affected by rotational flows. Any term ∇ × A, where
A is an arbitrary vector potential yields the same continuity relation:
(4.37) Q̇ = −∇ · (J + ∇ × A) = −∇ · J .
The flow represented by ∇ × A are typical whirls and eddy currents. On the other
hand, such whirly flow is conserved by itself. One has to think only of large weather
depressions and hurricanes. In the literature a localized whirl is called a Helmholtz
flow:
(4.38) v = n × ∇ψ(x − (nT x)n) ,
which yields planar flow perpendicular to the normal vector n, where the stream-
lines follow the levels, or contours of the potential ψ.
In general a vector field can be decomposed into a rotational part, a gradient
part, and a harmonic part. This is called the Helmholtz-Hodge decomposition:
[Flanders, 1963]
(4.39) v = ∇ × A + ∇ψ + w ,
where the harmonic function satisfies the Laplace equation ∆w = 0. The Green
function, or fundamental solution, of the Laplace operator is essential to construct
this decomposition:
(4.40) ∆G(x) = δ n (0) ,
The vector potential generating the rotational part is given by:
Z
(4.41) A(x) = dn x0 G(x0 − x)(∇x0 × v(x0 )) .
The scalar potential is given by:

Z
(4.42) ψ(x) = dn x0 G(x0 − x)(∇x0 · v(x0 )) .
The harmonic part is constructed from the remainder, which can be written as a
projection operator onto the harmonic part:
Z
(4.43) w = v − dn x0 G(x0 − x)(∆x0 v(x0 )) .
The Green function in n dimensions is given by: [Gel’fand and Shilov, 1964, Shilov,
1968]
−1 1
(4.44) G(x) = ,
(n − 2)Ωn |x|n−2
where Ωn is the surface area of the n-dimensional sphere, where n > 2. In the
two-dimensional case n = 2 the Green function of the Laplacian is:
1
(4.45) G(x) = log |x| .
2π
See also Section 12.3. Using the Green functions, the problem of finding a solution
to a partial differential equation can be mapped to the, often much simpler, problem
of determining an integral.
4.6. Remarks on differential geometry and other differentials

Differential geometry has been useful to distinguish two aspects of a state de-
fined on a domain. First, the possibility to change the state variables and the
invariant state variables which transform with the coordinate transformations on
the reference space. Second, the relation between the domain and its boundary,
which is an essential concept to be able to construct interfaces between interacting,
open systems.
For the study of the bulk, i.e., the behavior on the domain, the theory of partial
differential equations is more appropriate. This is investigated in the following chap-
ter. The aspects of differential geometry arise in these partial differential equations
in two ways. First, the study of boundary conditions, since differential geometry
determines the relation between the domain and its boundary. Secondly, the differ-
ential operators which appear as building block is partial differential equations and
map between different geometric function spaces. As it turns out, the two aspects
are intimately related.
CHAPTER 5
Partial differential equations
In the mid-twentieth century the theory of partial differential

equations was considered the summit of mathematics, both be-
cause of the difficulty and the significance of the problems it
solved and because it came into existence later than most areas
of mathematics.
Nowadays many are inclined to look disparagingly at this re-
markable area of mathematics as an old-fashioned art of juggling
inequalities or as a testing ground of functional analysis.
[V. I. Arnold, Lectures on partial differential equa-
tions]
We will not attempt to give an introduction into the theory of partial differential
equations. There are many good textbooks, which, together, give an insight in the
full breadth of the theory of partial differential equations. The older literature deals
with explicit solutions and both linear and nonlinear theory. [Carathéodory, 1935,
Courant and Hilbert, 1961, John, 1982, Kamke, 1962, Hadamard, 1923, Arnold,
2004] In intermediate years, the middle of the twentieth century, many new devel-
opments arose, [Bers et al., 1964, Gårding, 1998, Garabedian, 1964, Shilov, 1968]
but shifted the focus towards operational and weak formulations of the theory of
linear partial differential equations, [Hörmander, 1990, Lions, 1970] because the
general theory of nonlinear partial differential equations in more than two dimen-
sions, without appropriate structure, is considered prohibitively complicated.[Serre,
2001]
Since our main interest is interacting systems described by partial differential
equations, nontrivial boundary conditions are our special interest. The theory of
initial and boundary conditions of hyperbolic differential equations is another small
niche in the study of partial differential equations. [Sakamoto, 1982, Ladyzhenskaya,
1985] In particular the weak formulations, in terms of norms, are generally set up
with vanishing boundary terms. Hence these approaches need to be adapted to
make them suitable for nontrivial boundary conditions. The interaction of systems
is through the boundary conditions. The corresponding volume norm can be ex-
pressed in terms boundary norms, the energy flux, and initial-state and final-state
norm, the energy of the respective state.
In this chapter we review and develop some theory concerning boundary condi-
tions in, mainly, hyperbolic partial differential equation. The starting point is the
partial differential equation itself, and the energy norm and energy flux, essential
for the correct physical interpretation of the boundary condition, are derived. The
energy norm is useful for both the mathematical and the physical consistency of
the model. Uniqueness and existence follows from the proper energy norm, and
energy continuity is an important criterion in modelling.
71
72 5. PARTIAL DIFFERENTIAL EQUATIONS
In this chapter we set up the conventions and tools which we will use in later
chapters. The Sturm-Liouville theory (Section 5.2) is considered the paradigm
of the modern approach to partial differential equations. It has a clear link to a
functional formulation, of which the energy functional is a particular useful physical
case.
Partial differential equations are equations in higher order spaces which depend
on derivatives with respect to the different variables:
(5.1) P (z, ∇z )u(z) = 0 ,
where u is the unknown function to be determined, and the differential operator P
is a function of the variables z = (z0 , z1 , · · · , zn ) and a polynomial with respect to
the different derivatives, or the gradient:
   
∂ z0 ∂0
 ∂ z1   ∂ 1 
   
(5.2) ∇ z =  ∂ z2  ≡  ∂ 2  .
   
 ..   .. 
 .   . 
∂ zn ∂n
The standard linear hyperbolic differential equation is the wave equation, which is
a second-order polynomial in terms of the derivatives:
(5.3) (∂02 − ∂12 − ∂22 − ∂32 )u(z) = (∂t2 − ∆)u(z) = u(z) = 0 ,
where z0 = t is the time, and z1 , z2 and z3 the spatial coordinates. The Laplace
operator ∆ and the D’Alembertian or wave operator are the standard differential
operators for isotropic spaces. [Magnus et al., 1966] This is a hyperbolic differential
equation since both signs occur in the quadratic polynomial.
Many types of analysis of PDE’s are based on looking at the polynomial P (x, ξ),
where the vector ξ replaces the gradient operator. It is closely related to taking the
Fourier transform of the equation as a whole:
Z
T
(5.4) dzP (z, ∇z )u(z)eξ z = P (∇ξ , ξ)ũ(ξ) .
If the dependence on z of P is limited, say P (z, ξ) does not vanish sign for any z
given ξ 6= 0.
A number of well-known constant differential operators P are listed in Table 5.1.
They are used in physical models which are isotropic, which means that the space
is homogeneous and the equations do not depend upon orientation and translation
of the coordinates.
5.1. Uniqueness and existence

For a well-defined mathematical problem uniqueness and existence are essential.
For the linear wave equation, Eq. 5.3, the energy norm will lead directly to these
results. [Courant and Hilbert, 1961, Garabedian, 1964, Sakamoto, 1982]
The uniqueness is the simplest condition. Given the partial differential equa-
tion, Eq. 5.3:
(5.5) u(t, z) = 0, z ∈ Ω, t ≥ 0
5.1. UNIQUENESS AND EXISTENCE 73
name equation dimension phenomena

gradient ∇ = (∂1 , ∂2 , ∂3 )T n×1 force
curl ∇× = ijk ∂k 3×3 rotation
divergence ∇· = (∂1 , ∂2 , ∂3 ) 1×n flux
Laplacian ∆ = ∂12 + ∂22 + ∂32 1 electrostatics
biharmonic ∆∆ 1 elasticity, fluid
P √ √
Laplace-Beltrami ∆ = ij g−1 ∂i ( gg ij ∂j ) 1 curved spaces
D’Alembertian = v −2 ∂t2 − ∆ 1 wave equation
parabolic ∂t + ∆ 1 heat equation
Helmholtz ∆ + k2 1 scattering equation
√ p
Table 5.1. Some named differential operators ( g = det|gij |).
for an unknown solution u, with initial conditions:

u(0, z) = u0 (z)
(5.6) , z∈Ω
∂t u(0, z) = u1 (z)
where u0 and u1 are the given initial values, and boundary conditions, for boundary
value f :
(5.7) u(t, z) = f (t, z), z ∈ ∂Ω, t ≥ 0 ,
for the boundary ∂Ω of the domain Ω. The uniqueness means that if two solutions
u = v and u = w satisfy the initial and boundary values, u0 , u1 , and f , they are
identical v = w. Considering their difference v − w, the corresponding initial and
boundary conditions vanish: u0 = 0, u1 = 0, and f = 0. Hence, if u = 0, for
vanishing boundary conditions, the solution u for arbitrary boundary conditions is
indeed unique. This relies on the linearity of the partial differential equation.
The energy norm is a positive functional of the function u at a particular time
t:
1
Z
(5.8) E[u] = dz(∂t u)2 + (∇z u)T (∇z u) ≥ 0 .
2 Ω
The vanishing initial conditions also means that at t = 0 the energy is zero. The
time-derivative of the integral over the energy can be formulated in terms of the
wave equation and boundary terms:
(5.9) Z Z Z
∂t E[u] = dz∂t u∂t2 u + (∇z u)T (∇z ∂t u) = dz∂t u(u) + dz∂t unT (∇z u) .
Ω Ω ∂Ω
The last integral; the boundary integral, vanishes if ∂t u = 0, or nT (∇z u) = 0 on

the boundary. These lead to the Dirichlet boundary condition, described here, and
the Neumann boundary conditions. The first term vanishes if u is a solution to the
wave equation u = 0. Hence E = 0 for t ≥ 0, and thereby u = 0, which proves
uniqueness.
Existence of a solution u in combination with boundary conditions is a rather
complex problem. Often it is solved by construction. If one is able to construct a
solution u(t, z), given u0 (z), u1 (z) and f (t, z) the existence follows. The simplest
approach is finite forward integration, using the analyticity. Given u = 0 the
evolution given by the Euler finite-step δ integration is:

(5.10)

∂t u ∆u0 u1 (t = δ) u1 (t = 0) + δ∆u0 (t = 0)
∂t = → = .
u u1 u0 (t = δ) u0 (t = 0) + δu1 (t = 0)
The solution at an infinitesimal time t = δ satisfies the wave equation u = ∂t u =
0, but not necessary the boundary conditions u|∂Ω = f . This can be corrected
by adding at every time t a solution ∆ub = 0 to the Laplace equation such that
u|∂Ω +ub |∂Ω = f . The Green function on a finite domain, Section 5.8, is constructed
along similar lines. The solution ub does not contribute to the forward evolution
in the next Euler step, hence for static boundary conditions the problem reduces
to finding boundary conditions for the Laplace equation or its elliptic equivalent,
which may still be formidable. [Bitsadze, 1968, Wendland, 1979] In a more modern
language, the existence question concerns the stability of the evolution operator, on
the right-hand side of Eq. 5.10. In the dynamic case, with time-dependent bound-
ary conditions we use the Hamiltonian dynamics to construct allowed boundary
conditions.
A particular interesting feature of hyperbolic differential equations is the time-
reversal invariance. If t is replaced by −t, the same equation arises. In some cases,
such as electrodynamics the dual equation arises, which correspond to a simultane-
ous spatial reflection z → −z. Time reversal links existence and uniqueness. The
forward time evolution yielded the existence, as we saw above. The same, or dual,
evolution applied in the backward time direction will give the admissible initial and
boundary conditions for that state.
5.2. Sturm-Liouville theory

The description of dynamical system in terms of partial differential equations,
such as the wave equation, Eq. 5.3, is only a particular choice of description. In
many cases, for analysis of the solution and stability, it is more appropriate to use
a weak formulation in terms of norms, such as Sobolev spaces. [Hörmander, 1990]
One choice is particular relevant for physical system. This choice is the energy
norm. It can also be applied to nonlinear systems. The mapping from the partial
differential equation to a minimization, or optimization, problem stems from the
days of Bernoulli. However, a practical method, also for solving inhomogeneous
equations was developed by Sturm and Liouville.
Sturm and Liouville published in the mid-nineteenth century a series of papers
on differential equations in the journal founded by Liouville, [Sturm, 1836, Liouville,
1836] which, together with the work of Heaviside and Laplace-Fourier, modernized
the approach to differential equations, in the sense that they constituted a method
of solving such equations. One could argue the work was the forerunner of modern
functional analysis.
Instead of solving a single differential equation at the time, they studied an as-
sociated eigenvalue problem. The expansion into eigenfunctions can be used to solve
both inhomogeneous equations and time-dependent problems. The Sturm-Liouville
problems are self-adjoint, such that the eigenvalues are real. The boundary condi-
tions must be chosen such that the formally self-adjoint problem, from inspecting
the differential equation, remains self-adjoint for a particular choice of boundary
conditions.
5.2. STURM-LIOUVILLE THEORY 75
Many well-known special functions, such as Bessel and Hankel functions, Le-
gendre polynomials, etc., can be expressed as solutions of the Sturm-Liouville prob-
lem. [Sommerfeld, 1992, Kamke, 1962, Courant and Hilbert, 1961] Modern theory
of functional analysis of differential equations is still very much based on the Sturm-
Liouville theory. The boundary conditions are restricted such that the problem is
self-adjoint. In practice it means no energy can be transferred in or out of the
system. If the self-adjoint operator is positive, a quadratic functional can be con-
structed which is the energy function of the system. Allowing the energy to change
by the boundary conditions breaks the positivity and the self-adjointness. At worse
a boundary integral may be added to the total energy. [Sakamoto, 1982] This
boundary integral is the power transferred in or out of the system. In the frame-
work of an action integral over space and time, consistent formulations can be made.
We will seek extensions of the Sturm-Liouville theory in a different manner. The
solution space will be separated in a boundary part, and a Sturm-Liouville part
with vanishing boundary conditions.
The Sturm-Liouville differential equation is:

∂ ∂
(5.11) − p(z) f (z) + q(z)f (z) = λf (z) .
∂z ∂z
The functions p(z) and q(z) can be interpreted as weights or measures, if they are
positive definite. The parameter λ is the eigenvalue. The associated operator is:
∂ ∂
(5.12) L = − p(z) + q(z) ,
∂z ∂z
and given an interval z ∈ [a, b] and boundary conditions, such as, for example,
f (a) = f (b) = 0, the corresponding weak formulation is:
2
1 b

∂
Z
2
(5.13) J[f ] = hf, Lf i = dzp(z) f (z) + q(z) (f (z)) .
2 a ∂z
In the standard Sturm-Liouville theory q(z) can be negative, we, however, want to
use Sturm-Liouville theory for energy functionals and require a positive q(z). The
smallest eigenvalue λ0 in that case is larger than zero:
Z b
2
(5.14) 0 < min J[f ] = λ0 dz (f0 (z)) = λ0 kf0 k2 ,
a
where f0 (z) is the eigenfunction of the Sturm-Liouville operator, corresponding to
the eigenvalue λ.
The Sturm-Liouville operator is self-adjoint and generates a spectrum:
(5.15) Lfi (z) = λi fi (z) ,
where λi < λj for i < j. The set of eigenfunctions {fi } span the full function
space, in the sense that any continuous function can be approximated by a sum of
eigenfunctions, with arbitrary accuracy. Therefore in the case of an inhomogeneous
equation with the inhomogeneous term g(z) the solution f (z) is easily found:
∂ ∂
(5.16) − p(z) f (z) + q(z)f (z) = g(z) .
∂z ∂z
The spectral decomposition yields the solution:
∞
X hfi , gi
(5.17) f (z) = fi (z) .
i=0
λi
The series generally converges due to the increasing eigenvalues λi .

This result can also be expressed as the integral over a Green function G:
Z
(5.18) f (z) = dzG(z, z 0 )g(z 0 ) .
where the Green function, or fundamental solution is given in terms of the spectral
decomposition consisting of the pairs of eigenfunctions fi and eigenvalues λi :
∞
X fi (z)fi (z 0 )
(5.19) G(z, z 0 ) = .
i=0
λi
The Green function G(z, z 0 ) is the inverse of the Sturm-Liouville differential oper-
ator L. See also Sections 5.7 and 5.8.
The Sturm-Liouville theory forms also the basis for time-dependent analysis.
The eigenvalue λ may be replaced by time-derivatives:

∂t parabolic
(5.20) λ→ .
−∂t2 hyperbolic
These systems may be analyzed in space-time, but also in frequency ω or Laplace

s variables. The eigensystem corresponds to the the decay modes and vibrational
modes for parabolic and hyperbolic systems respectively.
The best known eigenfunctions of a Sturm-Liouville problem are the expansion
in term of trigonometric functions, in the case p = constant and q = constant. On
the domain z ∈ [0, L] the eigenfunctions of the differential operators ∂z2j are:
(5.21)
 2j
 sin π(2i+1)z

 2L , f (0) = 0, ∂z f (L) = 0, λi = (−1)j π(2i+1)2L

 2j
cos π(2i+1)z j π(2i+1)

fi (z) = 2L , ∂ z f (0) = 0, f (L) = 0, λ i = (−1) 2L .
 πiz j πi
2j


 cos L , f (0) = 0, f (L) = 0, λi = (−1) L

πiz
2j
sin L , ∂z f (0) = 0, ∂z f (L) = 0, λi = (−1)j πi

L
The different boundary conditions yield different eigenfunctions fi . Each set is,
however, orthogonal on the domain [0, L]:
Z L
1 L L
Z
(5.22) dzfi (z)fj (z) = dzfi (z)∂z2j fj (z) = δij .
0 λ i 0 2
The vanishing boundary conditions allowed us to map the strong problem of

finding a solution to a second-order differential equation, to the weak problem of
finding a minimizing function f0 . The boundary terms vanish under the partial in-
tegration. However, more general boundary conditions exist for the Sturm-Liouville
theory. The boundary terms from partial integration are:
(5.23)
θ(h, f ) = h∂z h, p(z)∂z f i − hh, −∂z p(z)∂z f i = p(b)h(b)∂z f (b) − p(a)h(a)∂z f (a) .
Hence, in the case of f (a) = 0 or ∂z f (a) = 0 and the same for the other boundary
condition at b, the boundary terms vanish and the weak and the strong formulation
are equivalent on a finite domain. Furthermore, in the case of boundary conditions
5.3. INITIAL CONDITIONS AND BOUNDARY CONDITIONS 77
of the form ca f (a) = ∂z f (a), the weak formulation can be symmetrized:

(5.24)
Z b 2
2 2 ∂ 2
J[f ] = hf, Lf i = −ca p(a)f (a) +cb p(b)f (b) + dzp(z) f (z) +q(z) (f (z)) ,
a ∂z
which yields a positive definite functional J[f ] if ca ≤ 0 and cb ≥ 0.
Hence, in the general theory of systems described by partial differential equa-
tions, those which satisfy a positive energy norm J[f ] ≥ 0 form a restricted set.
However, the restriction is quite natural for physical systems, and allows one to
handle nonlinear systems where the stability is governed be the energy conserva-
tion. If the domain of solutions f ∈ S for J[f ] < E0 is restricted, the system
remains in this domain due to energy conservation under evolution.
5.3. Initial conditions and boundary conditions

The study of partial differential equations of a dynamical systems has a long and
slightly troubled history as a mathematical problem. [Hadamard, 1923, Bernstein,
1950, Kreiss, 1970, Hörmander, 1990, Sakamoto, 1982, Gårding, 1998, Garabedian,
1964, Highdon, 1986, Lebeau, 1992, Majda and Osher, 1975] The recent results are
based mainly on the approaches with linear operators. The evolution of a system in
infinite space is called the Cauchy problem. The well-posedness of the Cauchy prob-
lem refers to the existence and uniqueness of the forward evolution of a systems for
some specified initial conditions, typically at t = 0. In a finite space, with bound-
aries, and boundary conditions one typically speaks of the initial boundary value
problem, or the mixed boundary value problem, or even the hyperbolic boundary
value problem.
Our particular interest is in the initial boundary value problem, since we want
our system to interact with its surroundings. The standard formulation fixes the
boundary condition from the start, which leaves a very limited interaction with the
surroundings. In the more general case we will not speak of boundary conditions
but boundary impedances, so indicate the two-way interaction of the dynamical,
evolving system with its surrounding, through the boundary. [Hogan, 1985] The na-
ture of the well-posedness changes drastically, because of the boundary impedance.
The energy norm forms the basis of the well-posedness. Originally energy arose
as a tool to prove well-posedness, in this case it defines the appropriate forms for
boundary impedance.
Given the hyperbolic differential equation, Eq. 5.3:
(5.25) φ̈ = ∆φ ,
the evolution is not uniquely specified by φ(z, t = 0). A second kinematical variable
ψ(z, t = 0) is required to specify sufficient initial conditions for unique evolution.
e.g.:
(5.26) ψ(z, t) = φ̇(z, t) .
The system is now given by two first-order equations:

φ̇ 0 1 φ
(5.27) = .
ψ̇ ∆ 0 ψ
However, likewise we might have chosen as second variable:
(5.28) ∆ψ(z, t) = φ̇(z, t) ,
such that the first order equations of motion are:

φ̇ 0 ∆ φ
(5.29) = .
ψ̇ 1 0 ψ
The two types of differential equations, Eq. 5.27 and Eq. 5.27, corresponding to
different choices of ψ are related to the typical choice of boundary conditions. The
Laplacian ∆ acts on one of the two fields φ or ψ, which should be complemented
with boundary conditions in the same fields. The choice of boundary conditions
determines the form of the equations. In modelling this choice is associated with
causality, and therefore not an intrinsic property of the system, or component, but
of its interaction. [Paynter, 1961, Karnopp et al., 2000]
In the port-Hamiltonian approach, [Van der Schaft and Maschke, 2002] re-
lated to the systems of conservation laws, [Serre, 2001, Dafermos, 1999] usually a
symmetric form for the first order equations is chosen:
(5.30) ∇ · ψ(z, t) = φ̇(z, t) ,
such that the equations of motion are similar in structure to the Hamilton equations
of motion:

φ̇ 0 ∇· φ
(5.31) = .
ψ̇ ∇ 0 ψ
The differences in the choice of second kinematical variable ψ(z, t) will play mainly
a role in the boundary conditions on ∂Ω if the functions are specified on a domain
z ∈ Ω. The spatial differential operators, like the Laplacian ∆, the gradient ∇ and
the divergence ∇·, on the right-hand side of the equations of motion have non-trivial
null-spaces. Hence an arbitrary function in the null-space will not affect the forward
evolution, and the backward evolution is specified up to solutions in the null-space.
In order to arrive at a well-defined set of equations of motion with an existing and
unique solution for the initial conditions φ(z, t = 0) and ψ(z, t = 0), and boundary
conditions B(φ(z ∈ ∂Ω, t), ψ(z ∈ ∂Ω, t)) = 0, must yield a non-singular evolution
operator.
In the case of the Laplace operator, Dirichlet and Neumann boundary condi-
tions yield a well-posed problem for the time-independent case:
∆f (z) = 0 z∈Ω
(5.32) ,
Bf (z) = h(z) z ∈ ∂Ω
where B is the boundary operator, and h is the boundary value or function, in
the case of Dirichlet boundary conditions B = 1, and in the case of Neumann:
B = nT ∇ = ∂ν , where n is the surface normal.
In Section 5.1 we saw that existence could be proven by constructing an explicit
forward integration, using for example the Euler integration. If both the forward
and the backward evolution must exist, it means the Euler integration yields a bi-
jective mapping. In the Section 5.1 we supplemented the evolution on the domain
defined by the Laplacian, by a solution of the Laplace equation ∆ub = 0, which cor-
rected deviations from the desired boundary conditions. Hence the null-space of the
Laplacian supplied additional degrees of freedom in the function space. However,
the natural course of action based on evolution operators is to attempt to construct
a nonsingular operator. This construction is well-defined for finite dimensions, but
rather ambiguous in the infinite-dimensional case. Take for example the domain
z ∈ [−1, 1] and the basis {z j }N 2
j=0 . The differential operator ∂z lowers the power by
Legendre polynomial P50(z)

Legendre polynomial P49(z)
0.5
-0.5
-1
-1 -0.5 0 0.5 1
z
Figure 5.1. The two polynomials perpendicular, in Lebesgue
inner-product, to the image of the Laplace operator, ∂z2 , on the
polynomial space {z j }50
j=0 . They localize at the boundaries z = ±1,
where they are normalized Pn (+1) = 1.
j N −2
two: ∂z2 {z j }N 2
j=0 = {z }j=0 . Based on the Lebesgue L norm, the boundary func-
tions, the extensions of the boundary values, are the Legendre polynomials LN (z)
and LN −1 (z), which do tend to localize at the boundaries for high values of N . See
Figure 5.1.
Hence in the infinite-dimensional function space one can consider a bijective
mapping:
(5.33) (∆f (z), h(z)) ↔bijection f (z) ,
where h(z) is defined on the boundary, which may possibly be extended to the
whole domain.
This bijection is not a priori evident, and may depend, as noted, on subtleties of
the function space we consider. The boundary conditions h(z) augment the range
space ∆f (z), such that for every function f (z) there is a unique decomposition in
these functions, which lies at the basis of dynamics.
In the case of dynamical equations with the two initial conditions φ(z, t = 0)
and ψ(z, t = 0), and the appropriate differential equations:

φ̇ 0 D φ
(5.34) = ,
ψ̇ −Dc 0 ψ
where D is the spatial differential operator, arising from the problem at hand, such
as D = ∇ as gradient operator and D c = ∇· as divergence operator in Eq. 5.31.
The two right-hand sides could be mapped back to the full space, with the use of
boundary conditions h(z) and g(z):
(Dc φ(z), h(z))

ψ(z)
(5.35) ↔bijection .
(Dψ(z), g(z)) φ(z)
The problem is complete in the sense that the kinematical variables φ(z) and ψ(z)
can be expressed in terms of their evolution: Dψ(z) and D c φ(z) together with the
boundary conditions, g(z) and h(z).
The wave equation on a string and the transmission line are typical one-
dimensional (z ∈ [0, 1]) examples of systems where two boundary conditions, say
h(0) and g(1), are used to complete the system. In that case D = D c = ∂z , the
one-dimensional null-spaces φ(z) = constant and ψ(z) = constant each require a
single boundary condition. For the transmission line, an applied potential h(0) at
the end z = 0 will cause a current g(1) at the other end z = 1, not only depending
on the properties of the transmission line, but also on the impedance at z = 1. The
two extreme cases of an open end and a short-circuited end at z = 1 will both lead
to complete reflection at z = 1 of the signal from z = 0, in the second case with a
phase shift: cos π = −1.
These mixed boundary conditions, with both nontrivial h(z) and g(z), are
typical for the analysis in an operational setting. If the component is part of a
greater network, it has a certain functionality, which leads to mixed boundary
conditions associated with a signal flow, or Green function or source-and-sink view.
[Hirsch and Smale, 1974] Apart from a response view, with general impedance,
the actual choice of mixed boundary conditions are not the standard mathematical
boundary conditions, which are more likely be associated with either a h(z) or a
g(z), not a combination. Of course it is possible to put a potential on each end of
the transmission line, h(0) and h(1), it determines its resistance and capacitance,
but there is often little purpose to do so. It determines the resistance, and not the
transmission properties. For a string these are the typical boundary conditions to
determine vibrational modes and frequencies.
The completeness of the stationary problem in terms of the kinematical vari-
ables is required for a well-defined dynamical problem, an initial boundary value
problem. However, it does not necessarily lead to a well-defined problem. [Serre,
2001, Sakamoto, 1982, Gårding, 1998] Especially the nonlinear hyperbolic problem
in more than one space dimension is an open problem, where existence and unique-
ness are not a priori clear. On the other hand stability analysis based on contracting
maps and Lyapunov functions, typically for discrete systems, yields a well-defined
problem in the sense of an energy norm. We will augment the equations of motion
with a compatible energy norm. The boundary conditions express the appropriate
energy flux, or power flow, for this norm.
Given the energy norm, the possibility to find the associated power-flow bound-
ary impedance depends on the possibility to formulate it as an action integral
with a space-time divergence. This approach is similar to the E-well-posedness
of [Sakamoto, 1982]. The construction of the energy-momentum tensor goes along
similar lines. [Felsager, 1981, Landau and Lifschitz, 1987a] Although, in the general
argument of the conserved energy-momentum tensor, one relies on the translational
invariance, which, in the case of a domain Ω with boundaries ∂Ω is not possible, as
Ω t=T
S
−E Γ
t
Ω
t=0
Figure 5.2. The space-time tube which combines the initial and
boundary conditions. The bottom and top cap t = 0 and t =
T correspond to the energy in the system, while the boundary
Γ = ∂Ω, corresponds to the power flow. The hyperbolic partial
differential equation inside the tube is the Euler-Lagrange equa-
tion.
translation will move states in and out of the domain under consideration. Simply
said, the divergence terms and the corresponding boundary terms are ignored under
partial integration to construct the energy-momentum tensor.
The zero-energy state E = 0 corresponds to the null-state:
(5.36) E[φ] = 0 ⇔ φ(z) = 0 .
Hence, if we can associate to the partial differential equation the continuity of
power, i.e., the conservation of energy, the existence and uniqueness follow from
the energy norm E ≥ 0.[Friedrichs, 1954, Courant and Hilbert, 1961, Garabedian,
1964, Friedrichs, 1973]
On a finite domain Ω, the power, or energy flux though the boundary plays an
important role:
(5.37) Efinal − Einitial = Pboundary .
We seek an explicit form of Pboundary in terms of the fields φ, given the energy E.
Such a form already exists in the energy-momentum tensor.
The difference in energy between two times t = 0 and t = T can be written as
the power integral P over the time-derivative of the energy density:
Z T Z Z Z

(5.38) P = ∂t Edz dt = Edz − Edz .
0 Ω Ω t=T Ω t=0
The energy density itself does not depend on time, only via its field variables φ(z, t).
If the power is conserved, the energy density is part of a space-time total divergence,
the result of the continuity of energy, [Heaviside, 1891] which can be extended to:
(5.39) ∂t E + ∇ · S = 0 ,
where S is the spatial power flow, or energy flux. Eq. 5.39 is the local energy
continuity, for a conserved energy density. Therefore, given the total space-time
divergence (E, S), the full power integral is:

Z Z Z T Z
nT Sdzdt .

(5.40) P = Edz − Edz +
Ω t=T Ω t=0 0 ∂Ω
See Fig. 5.2.

All the boundaries of the space-time region Ω × [0, T ] are now included in the
power integral. If the energy is conserved, the integral is zero; P = 0. Hence, the
pair (E, S) must be a total divergence in the lifted space. [Ivey and Landsberg,
2000] The lifted space is the space not only consisting of the fields φ(z, t), but also
of its derivatives φ̇ and ∇φ. For the Euler-Lagrange equations of motion it was
recognized that the initial conditions do not only depend on φ but also on its time-
derivative φ̇, in the space-time generalization, one includes ∇φ in order to yield the
boundary conditions. The generating functional, the Lagrangian L(φ, φ̇, ∇φ), for
the Euler-Lagrange equations of motion, can be generalized to the energy-energy-
flux total divergence:
!
δL
φ̇ − L

E δ φ̇
(5.41) = δL .
S δ∇φ φ̇
It is easy to show that the corresponding energy density E and energy flux S satisfy
the energy continuity relation, Eq. 5.39. Taking the time derivative of the energy
will yield a total divergence −∇ · S, with the use of the Euler-Lagrange equations
of motion. The explicit time-derivative of the energy E is calculated with the use
of the differentiation chain rule:
δL δL δL δL δL
(5.42) ∂t E = ∂t φ̇ − φ̇ + φ̈ − φ̈ − ∇φ̇ ,
δ φ̇ δφ δ φ̇ δ φ̇ δ∇φ
with the use of the Euler-Lagrange equations of motion, Eq. 5.43:
δL δL δL
(5.43) − ∂t −∇ =0 ,
δφ δ φ̇ δ∇φ
the first two terms on the right-hand side of Eq. 5.42 are replaced by the gradient
term in Eq.5.43. For the wave equation with unit material constants ρ = κ = 1, the
first term is absent, and the last term reduces to the Laplacian: ∆φ. The Laplacian
represents the net force of the left and the right force, i.e, the change in the moment
variable.
Furthermore, the third and the fourth term in Eq. 5.42 cancel, and we recover
the energy continuity relation, Eq. 5.39, with the explicit form of the energy flux,
S:

δL
(5.44) ∂t E = −∇ φ̇ = −∇ · S .
δ∇φ
If we choose the straightforward generalization of the momentum π of finite dy-
namical systems:
δL
(5.45) π(z) = ,
δ φ̇
the energy flux reduces to:
δH δH
(5.46) S=− .
δ∇φ δπ
name PDE (z ∈ Ω) boundary (z ∈ ∂Ω)

Dirichlet ∆u(z) = 0 u(z) = f (z)
Neumann ∆u(z) = 0 ∂ν u(z) = g(z)
Robin ∆u(z) = 0 a(z)∂ν u(z) + u(z) = h(z)
Poincaré ∆u(z) = 0 a(z) · ∇u(z) + u(z) = h(z)
Table 5.2. Named elliptic boundary conditions
The Lagrange formulation of dynamics naturally yields the energy flux in this form.
However, given the Hamilton equations of motion, more general formulations of
the energy flux exists. The differential operator associated with continuity is the
divergence operator, with the gradient operator ∇φ as adjoint, but more differential
operators may be used. This problem is pursued further in Chapter 7.
The initial boundary value problem is the problem of finding an energy norm
E and boundary conditions B, like in Eq. 5.32, such that the boundary conditions
uniquely specify the energy flux S. In the traditional setting, the boundary con-
ditions yield nT S = 0 in the linear case with non-vanishing mass density ρ and
stiffness κ:
ρ(z) δH

φ(z)|∂Ω = 0 ⇒ δπ ∂Ω = 0

(5.47)
δH
∂ν φ(z)|∂Ω = 0 ⇒ κ(z)−1 nT δ∇φ =0 .
∂Ω
However, more general boundary conditions exist. Especially in a control setting,

one likes to specify the boundary condition as relation among the fields φ(z) and
π(z) at the boundary ∂Ω.
The stationary Dirichlet boundary condition, in terms of energy conservation,
can be rewritten as the rate-of-change condition: φ̇(z) = 0 for z ∈ ∂Ω, indicating
that a prestrained state:
(5.48) φ(z, t) = h(z) for z ∈ ∂Ω and t ∈ [0, T ] ,
will lead to S = 0. The energy E 6= 0 for h(z) 6= 0. In the case that the boundary
conditions do not change in time, the energy flux S is automatically zero. The sta-
tionary problem can be mapped to an elliptic problem, [Bitsadze, 1968, Wendland,
1979, Krantz, 1992] without a reference to time. The time-dependent problem may
be solved by solving the elliptic eigenvalue problem first, and using the eigenstates,
which all satisfy the stationary boundary-value problem, to generate dynamics,
similar to the Sturm-Liouville approach of Section 5.2. Typical elliptic boundary
conditions are listed in Table 5.2.
For a system in a subspace H(φ, π) < E0 of the full state space can be kept in
that subspace by the appropriate choice of boundary values, for particular boundary
conditions, consistent with the energy flux. The initial boundary value problem, in
terms of the energy H, and the energy flux S is always conditionally stable. The
energy can be extracted from the system through the boundary, but not all forms
of boundary impedance will keep the energy in the system within bounds. Take,
for example, the flow through a tube. The product of pressure and volume flow
is the power at the boundary. For a given flow the pressure at beginning and end
may vary, yielding a variation of the energy in the system, which may be stored
E=0
E=0 S
Figure 5.3. A simple case where the boundary flux S may vanish
dynamically, although there remains energy on the domain, in an
isolated part.
as kinetic energy in the flow, possibly giving rise to turbulence beyond certain
boundary values.
5.4. Energy flux and boundary conditions

The energy flux equal zero S = 0 condition, which corresponds to an energy-
isolated system, reduces to a variety of well-known boundary conditions in the linear
case, as we have seen in the previous section. The more general case, specifying a
boundary condition in terms of one of the two boundary efforts:
δH δH
(5.49) − = h or =g ,
δ∇φ δπ
will also suffice for a well-defined hyperbolic initial boundary value problem, based
on Hamiltonian dynamics. The differential form of the energy continuity relation
of the space-time volume is the power continuity:
(5.50) Efinal − Einitial = Pboundary ⇒ ∂t E = nT S .
The boundary conditions, in terms of the energy which in turn yields the well-
posedness, are proper if either the energy flux S = 0, or if any other value of S 6= 0,
for any state (φ, ψ) can be achieved by an appropriate boundary value:
(5.51) ∀φ, ψ, ∀S 6= 0 : ∃g, h : S = S(φ, ψ, g, h) .
This definition allows one to always extract energy from the system if the energy
flux is nonzero, driving the energy E ≥ 0 to zero. If the energy flux is zero S = 0
for any value of f or g, two situations can occur. Either the energy E = 0 and
consequently the fields φ = 0 and ψ = 0, or part of the boundary is isolated from
the fields φ 6= 0 and ψ 6= 0 in the system causing E 6= 0. In the latter case the
initial conditions determine the energy retained in the system. See Figure 5.3.
For the case of a string or beam the well-definedness would correspond to the
possibility to use boundary values, for the particular boundary conditions, in terms
of (g, h), which are absorbing. For example, for position or velocity boundary
conditions, the ability to choose the motion of the boundary in the direction of
the force exerted by the object on the boundary. Likewise for force boundary
conditions to choose the force in the direction of the motion. In both cases, if there
is force or motion at the boundary, it is used to extract energy from the system. The
extraction of energy means the boundary conditions may yield a contracting map on
the state space. This may be the most general boundary-condition well-posedness
on an abstract nonlinear energy-norm system.
The product form of the energy flux, Eq. 5.46, yields a set of simple solutions,
Eq. 5.49, for the power condition, Eq. 5.51. As long as the nonspecified effort
5.5. A COMPLEX BOUNDARY-VALUE PROBLEM 85
is nonzero, the boundary condition can be adjusted to extract energy out of the
system. The linear results, in the proper formulation as variational derivatives of
the Hamiltonian, carry over to the nonlinear theory.
In the case the nonlinear Hamiltonian has several pointlike minima Emin =
H(φmin , ψmin ) ≥ 0, the uniqueness is much more complex. In principle, one can
define a new local Hamiltonian in the neighborhood of (φmin , ψmin ) and rely on the
local properties:
(5.52) H 0 (φ0 , ψ 0 ) = H(φmin + φ0 , ψmin + ψ 0 ) − H(φmin , ψmin ) .
In the case of domain-like minima, like a line or an area, a whole new field of
dynamics arises. In the traditional sense, such problems are not well-posed initial
boundary value problems.
5.5. A complex boundary-value problem

Not all the problems concerning boundary conditions are discussed in the previ-
ous sections. However, physically based boundary condition automatically restrict
the problems that may occur. In this section we discuss two problems where the
boundary value itself is constrained.
The connection with analytic functions, of complex variables, has been im-
portant for the development of the theory of partial differential equations, in par-
ticular of elliptic equations, in the nineteenth and the first half of the twentieth
century.[Bers et al., 1964] Analytic functions automatically satisfy the Laplace equa-
tion, due to the relations between the real and imaginary parts.
The issues concerning boundary-value problems are easily illustrated with a
simple elliptic problem in two-dimensions. [Lanczos, 1961] The Cauchy integral
equation reconstructs an analytic function f (z) = u(x + iy) + iv(x + iy) from values
at the boundary:
1 f (z 0 )dz 0
Z
(5.53) f (z) = ,
2πi ∂Ω z 0 − z
where each component u and v of f satisfies the Laplace equation: ∆u = ∆v = 0,
since they satisfy the Cauchy-Riemann equations:
∂u ∂v ∂v ∂u
(5.54) − = + =0 .
∂x ∂y ∂x ∂y
The Cauchy-Riemann equations can be seen as a linear operator A on a two-
dimensional space:

u ∂x −∂y u
(5.55) A = =0 .
v ∂y ∂x v
However, the Cauchy-Riemann equations imply analyticity and therefore integra-
bility of the equations along the boundary ∂Ω, which means the closed integral over
the boundary vanishes, for an analytic function f :
Z
(5.56) f (z 0 )dz 0 = 0 .
∂Ω
Furthermore, for any analytic function g(z 0 ) the product f (z)g(z) is again analytic
and subsequently these integrals should also vanish:
Z
(5.57) g(z 0 )f (z 0 )dz 0 = 0 .
∂Ω
Hence the boundary condition f (z)|z∈∂Ω must satisfy a condition, Eq. 5.56, itself,
actually infinitely many conditions, Eq. 5.57. Part of the equations did not only
determine a unique solution on the domain Ω, but also restricted the vector field
(u, v) on the boundary ∂Ω. Lanczos, [Lanczos, 1961] in the spirit of linear oper-
ators, calls such problems overdetermined. However, it is an overdetermined set
of conditions on the boundary, rather than an overdetermined set on the domain.
Nowadays, the problem is known as the trace problem. This shows that not all
boundary conditions will lead to well-defined problems.
From a single component u or v on the boundary, the other component can be
reconstructed, using for example the Hilbert transform. Hence, up to a constant
term for the other field, the analytic boundary conditions can be reconstructed,
making well-defined boundary condition for the Laplace equation, or the Cauchy-
Riemann equations.
A simpler case with a single condition, instead of the infinitely many in Eq. 5.57
above, is the Neumann boundary condition on a compact simply-connected domain,
like a disc. [Arnold, 2004] The Dirichlet boundary condition yield a unique solution
u for the real component of a harmonic function:
(5.58) ∆u = 0|Ω u = f |∂Ω .
A unique solution u is a point in a function space and therefore zero-dimensional.
The Neumann boundary condition, on the other hand, yields a one-dimensional
solution space since any constant function u0 , with ∂ν u0 = 0, can be added to the
solution u:
(5.59) ∆u = 0|Ω ∂ν u = g|∂Ω .
However, there exists a mapping between the real and imaginary part, u and v,
of the harmonic function, through the Cauchy-Riemann equation, such that the
normal derivative ∂ν of u corresponds to the derivative along the boundary for v.
Hence the Neumann problem can be mapped to the following Dirichlet problem:
(5.60) ∆v = 0|Ω ∂s v = g|∂Ω ,
where s is the coordinate along the boundary. Hence, a single integrability, or
constraint, exists for the boundary condition, if the boundary is simply connected,
such that:
I
(5.61) gds = 0 ,
∂Ω
and a unique boundary condition f exists.
The Laplace operator maps harmonic functions unto harmonic functions: ∆h =
g. The index of the Laplace operator on a compact simply connected domain is zero.
The index, introduced in Section 2.11, is the difference between the dimensions of
the domain and image space of the operator. The Neumann boundary condition
led to an one-dimensional null-space, consequently, the boundary condition g has a
one-dimensional constraint, to retain the index = 0 of the Laplacian.
For a domain with a hole, the same mapping between the real and imaginary
part is no longer possible. For example, there is no single coordinate s describing
the distance along the boundary. The index of a differential operator on a domain
is related to the topology of the domain. [Grosche et al., 1995] The consequences of
boundary conditions on, for example, an annulus, in the theory of elasticity, have
made it impossible, as yet, to formulate general boundary conditions with a unique
5.6. LOPATINSKI CONDITION 87
Figure 5.4. A simple example where the same boundary condi-

tions may be fulfilled by infinitely many deformation states. For
these symmetric boundary conditions, associated with a left and
a right-hand turn of 180 degrees, two of the elastic deformation
states, with a fixed center and a rotated edge, have the same, low-
est energy.
solution for static elasticity problems on domain which is not simply connected.
[Truesdell and Noll, 1965, Synge, 1957] See Figure 5.4.
5.6. Lopatinski condition

The best known condition for initial-boundary value problems arises from the
theory of elliptic partial differential equations. It has been generalized to the hy-
perbolic case, but, as much of the older theory on partial differential equations, it
relies on a local analytic expansion of the field in terms of the coordinates around
a point at the boundary. It is therefore not a global condition, and in most cases
the boundaries are rectified to a rectangular grid.
Mixed initial-boundary value problems are hyperbolic partial differential equa-
tions with initial data at t = 0 on the kinematical domain Ω and boundary data for
all times t ≥ 0 at the domain boundary: ∂Ω. For an n-th order partial differential
equation, which means the highest time derivative is ∂tn the initial data requires to
specify all time derivatives to the order n − 1.
The initial-boundary value problem is the problem of finding f (t, z) satisfying
the domain condition:
(5.62) L[f ] = 0, (t, z) ∈ R+ ⊗ Ω ,
and the initial value conditions:
(5.63) ∂tj f (0, z) = φj (z), j ∈ {0, 1, · · · , n − 1} ,
and the boundary value conditions:
(5.64) Bj [f ] = 0, (t, z) ∈ R+ ⊗ ∂Ω, j ∈ {0, 1, · · · , k} .
The number of conditions k the boundary value should satisfy is determined by
the Lopatinski condition. For example, for the D’Alembertian; the wave equation
in three dimensions, there is a single boundary condition required, for example,
f (t, z 0 ) = 0 for z 0 ∈ ∂Ω. For higher-order differential operators this is no longer

trivial. The spatial derivatives should be separated into normal and parallel compo-
nents, with respect to the boundary ∂Ω. In most cases only the half-space z1 ≥ 0,
with ∂Ω = {z|z1 = 0}, is treated. [Hörmander, 1983, Krantz, 1992] Sakamoto
[Sakamoto, 1982] treats the unit circle as the standard domain, with the radial co-
ordinate r = kzk as normal coordinate. In general, the boundary should be smooth.
The leading parts, or principal parts, of the differential equation are selected.
The highest order derivatives to the time, in each of the terms determine the condi-
tions, just as the n-th order of the operator L, determined the n−1 initial conditions.
We can express it locally as an analytical function of λ, ξ, κ the Fourier-Laplace
transform of ∂t , ∇⊥ , ∂k :
(5.65) Ln (t, z⊥ , zk , ∂t , ∇⊥ , ∂k ) → Ln (t, z⊥ , zk , λ, ξ, κ) .
The boundary operators can be mapped to analytical functions likewise:
(5.66) Bj0 (t, z⊥ , zk , ∂t , ∇⊥ , ∂k ) → Bj0 (t, z⊥ , zk , λ, ξ, κ) .
The existence of a solution f to Ln [f ] = 0 is determined by the solutions of
Ln (t, z⊥ , zk , λ, ξ, κ) = 0. These roots can be separated in two parts, corresponding
to characteristics moving away from the boundary, and toward the boundary. The
solutions associated to the roots of characteristics moving away from the boundary
are related to solutions which need to be determined by the boundary conditions
Bj . Hence the number of conditions k should equal the roots κ(t, z⊥ , zk , λ, ξ) with
Imκ > 0.
Furthermore, the conditions themselves should not give rise to zeros coinciding
with the characteristic roots. For that purpose the function R is constructed which
is the determinant of the k × k matrix with coefficients Rj,l :
!
Bj0 (t, z⊥ , zk , λ, ξ, κ)κl−1
I
(5.67) R(t, z⊥ , zk , λ, ξ) = det dκ ,
Γ L+ (t, z⊥ , zk , λ, ξ, κ)
where L+ is the factor containing only the k positive roots of L, and the contour
encircles all these poles. The Lopatinski condition for existence and uniqueness of
the solution f to the initial boundary value problem is:
(5.68) inf |R(t, z⊥ , λ, ξ)| = c > 0 .
t, z⊥ , Imλ < 0
|ξ| + |λ| = 1
The function R for the Lopatinski condition is close to the construction of a local
Green function.
5.7. Green function

For a partial differential equation with constant coefficients the Green function
approach is often used to solve such equations. In the most appropriate setting
distribution theory, or generalized functions are used, together with Fourier-Laplace
transform. [Hörmander, 1990, Shilov, 1968, Gel’fand and Shilov, 1964, Schwartz,
1966] Fourier-Laplace transform is, simply said, an extension of ordinary Fourier
transform where functions are extended into the complex domain: (x ∈ Rn , ξ ∈ Cn )
1
Z
(5.69) F[u](ξ) = n dxeihξ,xi u(x) .
(2π) 2
5.7. GREEN FUNCTION 89
In the case of time and space coordinates the conjugate variable of the time
component is often denoted by τ . In the case of the wave equation:
(5.70) (∂t2 − ∆)u(x, t) = 0 ,
with the initial condition:
(5.71) u(x, 0) = g(x), ∂t u(x, 0) = h(x) ,
the corresponding transformed problem is:
(5.72) (τ 2 − ξ 2 )u(ξ, τ ) = 0 ,
with
(5.73) u(ξ) = g(ξ), τ u(ξ) = h(ξ) .
The Green functions are the solutions to the fundamental problems:
(5.74) (∂t2 − ∆)G(x, t) = δ(x)δ(t) ,
and
(5.75) (∂t2 − ∆)H(x, t) = δ(x)θ(t) ,
where θ is the step function: θ(t) = 1 for t ≥ 0 and θ(t) = 0 for t < 0. The second
Green function is also referred to as the Neumann function, or Green function of
the second kind, satisfying:
(5.76) (∂t2 − ∆)∂t H = δ(x)∂t θ(t) = δ(x)δ(t) .
Instead of treating the initial conditions a part of the solution u(x, t), they are
treated as source terms to a homogeneous differential equation in the whole space-
time. At t = 0 some disturbances occur such that the solution for t < 0 is zero, and
at t = 0 satisfies the initial conditions. These disturbances are therefore singular
in nature, but eventually easy to work with as they can be regularized under the
integral.
The Cauchy problem can be written as an integral:
Z
(5.77) u(x, t) = dx0 G(x − x0 , t)g(x0 ) + H(x − x0 , t)h(x0 ) ,
since from standard Heaviside calculus it follows that: ∂t δ(x)θ(t) = δ(x)δ(t).

The Green function is easily written down formally:
1 1
Z
(5.78) G(x, t) = dξdτ e−ihξ,xi−iτ t 2 .
(2π)n τ − ξ2
These functions are defined sloppy, especially on the initial surface t = 0. Further-
more, the integrals, defining the Green functions, are singular; the poles τ = ±|ξ| lie
on the real axis. As it turns out, the sloppy definitions hardly hinder the construc-
tion of a Green function, and the treatment of the poles determines the choice of
homogeneous part, since any function p(ξ) gives a homogeneous solution f , whence
integrated with the delta function containing the characteristic equation:
Z
(5.79) f (x, t) = dτ dξδ(τ 2 − ξ 2 )p(ξ)e−ihξ,xi−iτ t ,
which is a solution to the wave equation in the whole space. The difference between
different treatments of the poles are homogeneous functions in the whole space-time.
The particular treatment we seek yield solutions only for t ≥ 0; so-called causal
solutions. In the relativistic theory, causality has a different meaning; causality
requires the Green function to be zero outside the light-cone, as no signal can
travel faster than the speed of light.[Streater and Wightman, 1964]
The time-derivative part together with causality dictates the treatment of the
poles. The simplest way to look at it, is to separate the homogeneous solution into
odd and even parts in time. The even part is dictated by g(x), the odd part by
h(x). However, odd functions have equal but opposite forward and backward in
time, which will requires us to turn the solution causal.
Hence, the odd Green function would given by:
1 1
Z
0
(5.80) dξdτ e−ihξ,x−x i−iτ t ,
(2π)n τ (τ 2 − ξ 2 )
where the singularity at τ = 0 is treated as a principal value integral:

Z − Z ∞
f (τ ) f (τ ) f (τ )
Z
(5.81) dτ = lim dτ + dτ ,
τ ↓0 −∞ τ τ
which is often more easily expressed as:
f (τ ) 1 f (τ ) f (τ )
Z Z
(5.82) dτ = lim dτ + .
τ ↓0 2 τ + i τ − i
Such a small imaginary component ±i to a pole is called a pole prescription.

The pole prescription of the poles τ = ±|ξ| follows from the causality criterion:
The solution exists for t ≥ 0. Hence the poles should contribute only to t ≥ 0
values. For t ≥ 0, the contour integration must be closed in the Imτ < 0 plane.
Hence, the poles must lie in the lower half plane:
1 1
(5.83) → .
τ 2 − ξ2 (τ − |ξ| + i)(τ + |ξ| + i)
Here we note that the odd Green function would be non-causal, because of the
pole in the upper half plane, hence it should be given by poles:
1 1
(5.84) → ,
τ (τ 2 2
−ξ ) (τ + i)(τ − |ξ| + i)(τ + |ξ| + i)
yielding the second Green function:

0
1 e−ihξ,x−x i−iτ t
Z
(5.85) G0 (x − x0 , t) = dξdτ .
(2π)n (τ + i)(τ − |ξ| + i)(τ + |ξ| + i)
The complete causal solution is given by:

Z
(5.86) u(x, t) = dx0 G(x − x0 , t)g(x0 ) + G0 (x − x0 , t)(h(x0 ) − g(x0 )) ,
since:

1 1 1 1 1 1 1 1
(5.87) = + + − = + πiδ(τ ) .
τ + i 2 τ + i τ − i 2 τ + i τ − i τ
5.9. CONSTRAINED MINIMIZATION 91
5.8. Green function on a finite domain

Above we discussed the construction of a Green function or fundamental solu-
tion on an infinite domain. However, in most practical cases, the system will be on
a finite domain. Appropriate boundary conditions h(x, t) may not be met by the
Green function, similar to the existence argument in Section 5.1:
(5.88) G(x − x0 , t)|x0 ∈∂Ω 6= h(x0 , t) .
However, augmenting the Green function so that it satisfies the boundary condition
is an elliptic, stationary problem consisting of two parts: First construct a Green
function with vanishing boundary conditions, by adding a homogeneous solution
for all times with the same values at the boundary:
(5.89) ∆gx0 (x, t) = 0, gx0 (x, t) = G(x − x0 , t)|x0 ∈∂Ω .
Such that the augmented Green function G0 has vanishing boundary conditions:
(5.90) G0 (x, x0 , t) = G(x − x0 , t) − gx0 (x, t) .
Secondly, add to the total solution the quasi-stationary solution to the boundary
value problem:
(5.91) ∆f (x, t) = 0, f (x0 , t)|x0 ∈∂Ω = h(x0 , t) .
Given a problem with initial values u(x)t=0 and boundary values h(x ∈ ∂Ω), t), the
solution y(x, t) in Ω ⊗ [0, ∞) is given by:
Z
(5.92) y(x, t) = G0 (x, x0 , t)(u(x0 ) + ∂t2 f (x0 , t))dx0 + f (x, t) .
Ω
The time-derivative of the quasi-stationary solution f (x, t) serves as source for the
full solution.
5.9. Constrained minimization

In the theory of constrained partial differential equations one often resorts to
a method originated by Lagrange, where additional variables, Lagrange parame-
ters or multipliers, are introduced to solve the constraint simultaneously with the
constraints. In the modern approach it is called the KKT method. [Karush, 1939,
Kuhn and Tucker, 1951, Tröltzsch, 2005] However, in the literature older references
to similar approaches exist.[Carathéodory, 1935]
Considering the constrained problem for u:
(5.93) Au = f given Bu = g ,
can be rewritten, in the weak formulation, as a Lagrangian:
1
(5.94) L = hAu, ui − f u + p(Bu − g) ,
2
which satisfies the variational criteria:
(5.95) δ u L = δp L = 0 ,
where p is the Lagrange multiplier function. The corresponding equations are called
the KKT system or the saddle point problem, [Benzi et al., 2005] since it is expected
to have positive and negative eigenvalues:
A BT

u f
(5.96) = ,
B 0 p g
If the operator A is positive definite and, furthermore:

(5.97) λmin (A) > 4λmax (BA−1 B T ) ,
the system is positive definite. The condition above can be seen as the generalized
discriminant. In the case of the Stokes flow, the Schur complement BA−1 B T is the
identity:
1
(5.98) BA−1 B T = ∇ · ∇=1 .
∆
See also Section 8.11, where the Schur complement arises the Guyan projection, in
the matrix theory. [Golub and Van Loan, 1996]
5.10. Functional or variational derivative

In dealing with systems described by fields, or functions, the functional de-
rivative plays an important role. See also Section 2.7. However, for nonlinear
problems, and the finite approximation of such problems, the functional derivative
is not necessarily unique. It depends on the order of the approximations.
A functional J maps a function f to a number:
(5.99) J : f → J[f ] ∈ R .
For example, the Lebesgue norm L2[−1,1] is a functional:
Z 1
2
(5.100) J[f ] = kf k = dzf (z)2 .
−1
The functional derivative links the global functional and the variational equations,
such as the Euler-Lagrange or the Hamilton equations for fields. [Felsager, 1981] In
the case of stationary problems Sobolev norms are the functionals for elliptic partial
differential equations. [Dautray and Lions, 1990] However, careful mathematicians
tend to avoid an explicit definition of the variational equations related to the weak
problem under investigation. We, on the other hand, keep in mind that our final
goal is a consistent, but finite-dimensional approximation, for practical numerical
implementation. Consequently, we allow ourselves, in the mathematical description,
some latitude.
To define the derivative of a functional, we can proceed in different directions.
The directional, variational, or Gateaux, derivative arises from taking a small vari-
ation of the functional in the direction g:
J[f + g] − J[f ]
(5.101) δg J[f ] = lim .
→0
If such a limit exists, the variational, or functional, derivative can be written as a
linear operator on g:
(5.102) δg J[f ] = hJ 0 [f ], gi .
The stricter defined Fréchet derivative J 0 [f ], for an arbitrary g, only exists if the
norm of the linear operator associated with the directional derivative exists:
kδg J[f ] − J 0 [f ]gk
(5.103) lim =0 .
∀kgk→0 kgk
5.10. FUNCTIONAL OR VARIATIONAL DERIVATIVE 93
The directional derivative exists for many more functionals than the Fréchet de-
rivative will. The same situation applies to ordinary derivatives of multivariate
functions:
(5.104) lim f (x + δx, y + δy) 6= f (x, y) + fx0 (x, y)δx + fy0 (x, y)δy ,
δx,δy→0
where the vector F 0 (x, y) = (fx0 (x, y), fy0 (x, y)) would be the analogue of the Fréchet
derivative. Typical counterexamples are related the integrability condition dis-
cussed in Section 4.3.
Besides functional derivatives around particular points, given by the function
f , there is a special functional derivative, around the null-function, schematically
denoted by:
(5.105) δg J[0] = δg J .
This functional derivative is often encountered in quadratic functionals J2 [f ] =
hf, Qf i, where Q is a symmetric linear operator, such that the function is recovered
from the variational derivative:
(5.106) J20 [f ] = 2Qf ,
while in general
(5.107) δg J2 [f ] = 2gQf .
The corresponding Fréchet derivative is the linear operator: J20 = 2Q. It should be
noted that for nonlinear functionals such relations do not hold.
In the functional derivative we rely heavily on the underlying linear vector
space of functions. For example, for the functional which selects the value at a
single point z = z0 : J[f ] = f (z0 ), the functional derivative is the distribution:
(5.108) J 0 = δ(z − z0 ) ,
in the integral representation, where J (z) is the integrand corresponding to the
functional J.
In most numerical implementations of variational derivatives, the function space
is restricted to a finite dimensional space. Even more, for practical purposes a
Hilbert space is used, such that the function space and the functional-derivative
space are one and the same. As an example we will use the space of normalized
Legendre polynomials φi (z) with i ∈ {0, 1, , · · · , N } on the domain z ∈ [−1, 1].
They satisfy the orthonormality relation:
Z 1
(5.109) dzφi (z)φj (z) = δij .
−1
The symmetric unit operator in the space of Legendre polynomials to order N is
a projection operator P on the full space. For any measurable, i.e., integrable
function q(z), the projection g(z) on the finite subspace is given by:
XN Z 1
g(z) = P q(z) = φi (z) dzφi (z)q(z) .
i=0 −1
Hence, any function f and any functional derivative δg J[f ] is restricted to an ex-
pansion in terms of these polynomials:
N
X N
X N
X
(5.110) g(z) = ci φi (z) f (z) = ai φi (z) δg J[f ] = bi (a, c)φi (z) ,
i=0 i=0 i=0
where a, b, and c are the coefficient vectors, associated with f, J , and g. The
coefficient vector b will generally depend of the function f via its coefficient vector
a and the directional derivative direction g through c.
Hence, given a general nonlinear functional, and the functions g and f , the
results in the function space must be projected again onto the finite dimensional
space. There are two ways to proceed. First, determine the functional in terms
of the coefficient vector a, and project the gradient ∇a in terms of this finite
dimensional vector a onto the vector c:
(5.111) J[f ] → J(a), δg J[f ] → cT ∇a J(a) .
The second procedure to arrive at the finite-dimensional approximation of the
the functional derivative is to calculate the functional derivative first in the infinite-
dimensional space, and subsequently project onto the finite function space:
Z 1 Z 1
∂
(5.112) lim dz J [f (a) + q(z)] = dzδq(z) J [f (a)]q(z) ,
→0 −1 ∂ −1
where q(z) is an arbitrary measurable function, in the infinite-dimensional space.

The projected functional derivative is:
(5.113) P δq(z) J[f (a)] .
The function q(z) should also be projected onto the same function space, to yield
the coefficient vector c. The two approaches are identical in Hilbert space since the
domain of the variational function in the finite function space P g = g, therefore:
Z 1
(5.114) dzδg(z) J [f (a)]g(z) = hδg J [f (a)], P gi = hP δg J [f (a)], P gi .
−1
Clearly, this relation only holds on a self-dual space, such as the Hilbert space, since
then P 2 = P = P T .
From a pointwise comparison it is clear that the results are the same, for
sufficiently smooth functionals. In some cases it is easier to take the first route,
while in other cases the second route is faster.
For example, for the functional J [f ] = f 3 , the projected functional is easily
determined from the explicit integrals. The Fourier series einz are even better for
such functionals, since the non-zero terms are those with wave numbers summing
to zero:
Z π
(5.115) dzeinz eimz eikz = 2πδk+m+n .
−π
Similar addition theorems exist for other basis function on compact spaces. [Magnus
et al., 1966]
On the other hand, functionals like J [f ] = ef are more easily evaluated nu-
merically and pointwise:
Z 1
T X T
(5.116) δq J(a) = dzφ(z)ea φ(z) ≈ wi φ(zi )ea φ(zi ) ,
−1 i
where zi and wi are the Gauss-Legendre abscissas and weights. [Abramowitz and
Stegun, 1965]
Numerical implementation and the retention of properties of the physical sys-
tem dictate the approximation. It should be noted that for linear systems many
5.11. REMARKS ON PARTIAL DIFFERENTIAL EQUATIONS 95
approximations are equivalent. In the case of nonlinear systems the order in which
the different approximations are carried out will affectR the results.
1
Take a simple example of the functional J[f ] = −1 dzf 3 (z), in the subspace
f (z) = a0 + a1 z. We proceed to derive ∇a J along two routes. The functional
integral in the finite subspace is
Z 1
(5.117) J[f ] → J(a) = dz(a0 + a1 z)3 = 2a30 + 2a0 a21 .
−1
The variational derivative, along this route, is:
6a0 + 2a21

(5.118) ∇a J(a) = .
4a0 a1
The functional derivative, in the abstract function space, is:
(5.119) ∂q(z) J[f ] = 3f 2 (z) = 3(a0 + a1 z)2 .
It should be noted that this function lies outside the finite-dimensional subspace,
due to the term z 2 . The projection onto the two-dimensional function space {1, z}
is:
1
3a20 + a21
R
(5.120) P ∂q(z) J[f ] = R dz32 ∂q(z) J[f ] = 6= ∇a J(a) .
dz 2 z∂q(z) J[f ] 6a0 a1
Clearly, the two routes to the variational, or functional, derivative in the finite-
dimensional function space will lead to different results. In Chapter 11 we will
see that consistency conditions, in the general context of the equations of motion,
will yield the proper approximation the function space, given the conservation of
energy.
5.11. Remarks on partial differential equations

The link between the local differential equation, specifying the solution in a
small neighborhood of each point in space, and the global properties is through
variational principles. For physical systems these variational principles are in terms
of the Lagrangian or the Hamiltonian, or energy density of the system. Some
approaches start with field variables and conservation laws and try to construct
a second, or closure relation, for a consistent theory, which does not yet have an
energy density functional. It might even be difficult to construct such a functional,
for example, in the case of chemistry, where systems are often open in terms of the
energy flow. [Bird et al., 1960] In this chapter, the approach is followed which keeps
a close watch on the energy in the system, for consistency and stability.
Energy conservation, or energy density continuity, guides us to specify general
but appropriate boundary conditions, which keep track of the power flowing in and
out of the system.
Whether the energy principle restrict the choices of boundary conditions such
that inconsistent conditions yielding no solution or multiple solutions for the same
conditions is an open question. In many nonlinear cases with multiple local min-
ima in the energy, counter-examples exist for uniqueness. Existence is related to
the construction of evolution operators, which in the finite-order approximations
studied in the later chapters, seldom is a problem. First, the physical principles
and relevant properties of systems described by partial differential equations of this
chapter is studied in the next chapter.
CHAPTER 6
Models of infinite-dimensional systems
According to physics my desk is, for all its seeming fixity and
solidity, a swarm of vibrating molecules.
[W.V. Quine, Posits and reality]
Infinite-dimensional systems, as defined in Section 2.5, are systems described
by continuous functions, and, typically partial differential equations. None of the
models described here will be done great justice. They serve as advanced examples
of modelling infinite dimensional systems. For a more thorough analysis of each of
the models we refer to the respective literature in these fields.
Many models, whether they describe an elastic beam or the flow in a pipe, have
common features. They satisfy the same physical laws of conservation of energy
and momentum. [Fetter and Walecka, 1980, Landau and Lifschitz, 1987a, Truesdell
and Toupin, 1960] They have similar ways of damping of motion, and they exhibit
similar phenomena, such as wave motion, vibration, and balanced states under ex-
erted forces. The exhibition of these phenomena depends on the regime of external
circumstances. A system with a strong damping will show little vibration, unless
a large driving force at the right frequency is applied. In more complex models,
such as the Navier-Stokes equation, [Landau and Lifschitz, 1987b] different contri-
butions, such as the viscous or damping term, or the nonlinear inertial component,
of the model compete in the force and momentum balance. The domination of one
part, such as viscosity, over another, such as the fluid momentum, can be recovered
from dimensionless constants, in this case the Reynolds number, which combines
fluid properties with system parameters, such as system size.
In a system different effects, such as applied forces, friction, and inertia, are
combined into a force-balance equation:
(6.1) Finertia + Fdamping + Fpotential = Fapplied ,
where the forces on the left-hand side are determined by the system and the state
it is in. In discrete models these forces are numbers or vectors at worse. In infinite
systems, these forces are generally vector fields; vectors F(z) at each reference
position z. The force vector fields F are the dual of the field variables φ(z), which
describe the state of the system, their product is the work, which yields the local
energy, or the potential part of the Hamiltonian density through the line integral:
Z φ(z)
(6.2) H(φ(z)) = F(φ0 (z), z)dφ0 (z) .
φ0 (z)
In this case the reference position z is only an index. The choice of the field
variable φ determines the definition of the force, as the energy density is an invariant
quantity. Likewise, the kinetic part of the Hamiltonian defines the momentum field
97
98 6. MODELS OF INFINITE-DIMENSIONAL SYSTEMS
π(z):
Z φ̇(z)
(6.3) H(φ̇(z)) = π(φ0 (z), z)dφ̇0 (z) ,
φ̇0 (z)
where the mass density ρ(z) is the local relation, between π(z) and φ̇(z), at the
same position z:
(6.4) π(z) = ρ(z)φ̇(z) .
Through a variable transformation of field variables φ(z), the forces F(z) and the
momenta π(z) change as well. Their product, however, remains invariant, except
under coordinate transformations of the position coordinate z. In infinite systems
the coordinate z will also play a role in the differential operators in the Hamiltonian,
which intertwines the variable transformations of φ(z), and the coordinate trans-
formations of z. The invariant power-density products for the change of energy
F(φ(z), z)dφ(z) and π(φ(z), z)dφ̇(z) are retained.
6.1. Beam models

The beam models are important in structural mechanics. There are several
beam models in different degrees of sophistication, and for different purposes. The
most common beam model is the Euler-Bernoulli beam: [Love, 1944]
(6.5) ρA∂t2 y(x, t) = ∂x2 EI∂x2 y(x, t) ,
where ρ is the mass density, A is the cross-section area, E is the elasticity modulus,
and I the first moment in the bending direction under consideration. The vertical
deflection y is a function of the reference position along the beam x and the time t.
Beams are structural elements which are typically much longer than they are wide or
high. The length l of the beam remains constant, which means that the horizontal
position X(x) changes with the vertical position y(x). General deformations are
described by the position X(x) = (X, Y ), as function of a reference coordinate
x = (x, y), where the beam is described by x ∈ [0, l] and y = 0. See also Section 4.2.
In the years 1727 till 1776 Euler made extensive contributions to the theory
of beams with several members of the Bernoulli family. [Euler, 1744, Timoshenko,
1953, Truesdell and Noll, 1965] Johann Bernoulli introduced Euler and the rest of
the world to variational principles. Also his brother Jakob made significant contri-
butions, but the Bernoulli meant in Euler-Bernoulli equation is the son of Johann:
Daniel. Daniel Bernoulli and Euler worked out small displacement statics (the
Euler-Bernoulli equation), large-displacement solutions with bifurcations, [Antman
and Rosenfeld, 1978] and vibration dynamics with mode shapes.
If the horizontal position is fixed X(l) = l, the beam is more likely to act as a
string model than a beam model:
(6.6) ρA∂t2 y(x, t) = ∂x EA∂x y(x, t) ,
which is an approximation of the elastic energy in the length of the beam. In so-
called exact, or Cosserat, models, [Truesdell and Noll, 1965, Cohen and Muncaster,
1988] the beam is approximated by a line X = (X(x), Y (x)) for 0 < x < l, the
second reference coordinate y is absent, and possibly a perpendicular direction n on
that line, which is the orientation. See Figure 6.1. There are only a limited number
of invariants of a line, which are the, longitudinal strain, or stretch σ, curvature
6.1. BEAM MODELS 99
torsion s=1
γ beam
b
bending n
n
b r
r
body frame
s=0
Figure 6.1. The beam as a line, function of s, with local body

frame coordinates, the Frenet-Serret variables (n, b, r). The
change of the beam direction n = ∂s X(s)/|∂s X(s)| corresponds
to the bending, the body-frame rotation perpendicular to n yields
the torsion. See also Table 6.1.
k, and the torsion γ. There are many different ways to represent these quantities.
[Dubrovin et al., 1984] It is important to note that the rate of change of a unit
vector e(x) is perpendicular to that vector. Indeed, since eT e = 1,
(6.7) 0 = ∂x 1 = ∂x (eT e) = 2eT ∂x e .
One particular elegant way to represent these invariants, except for the stretch, is
to use the rotation group SO(3). The change of orientation of the beam, along the
beam, can be represented by the rotation of the body-fixed frame:
(6.8) ∂x R(x) = J(x)R(x) ,
where the anti-symmetric J(x) is the generator of rotations. It can be represented
by the angle vector ω:
(6.9) [J(x)]ij = ijk ωk ,
where ijk is the anti-symmetric Levi-Civita tensor. [Frankel, 2004] If the first
component of the vector represents the beam direction, ∂x X, and the second com-
ponent the beam orientation, n, then ω2 and ω3 are the two bending moments in,
and perpendicular to, the beam orientation, and ω1 is the beam torsion. [Lipschutz,
1969] The equations of the different beam models are given in Table 6.1.
One particularly unpleasant feature of the Euler beam model is the absence of a
constant or limiting wave velocity, or sound velocity. For most continuous, isotropic
media the velocity of sound will reach a limit as the frequency increases. For the
beam model, taking a high frequency ω, the wavelength k depends quadratically
on this frequency:
(6.10) ρAω 2 = EIk 4 ,
which means that the velocity v of a wave increases with frequency:
ρA √
(6.11) v= ω .
EI
name equations variables

Euler-Bernoulli ∂t2 y = ∂x4 y y: vertical displacement
Timoshenko ∂t2 W = −∂x ψ + ∂x2 W W : vertical displacement
∂t2 ψ = ∂x2 ψ − ψ + ∂x W ψ: rotational displacement
exact, Cosserat ns + f = rtt + btt {n, b, r}: orientation
Frenet-Serret ms + rs × n + k = b × rtt + b × btt s: reference coordinate
Table 6.1. Beam models
name condition 1 condition 2

clamped y = 0 (position) ∂x y = 0 (angle)
free ∂x2 y = 0 (torque) ∂x3 y = 0 (force)
supported y = 0 (position) ∂x2 y = 0 (torque)
Table 6.2. Typical boundary conditions for beams. In brackets,
the physical variable that is constrained, in the case of position
and angle, or has zero input, i.e., is free, in the case of force or
torque.
The Timoshenko beam (See also Table 6.1) cures part of this problem by introducing
a kinetic energy associated with the bending moment of the beam. [Hansen and
Snyder, 1997] However, this leads to two coupled equations, and still the stretch
and the shear are not accounted for. For vibrational analysis especially the shear
terms will be important if the wavelength approaches the cross-section size of the
beam. In that case the bending velocity will become equal to the shear velocity.
In the linear theory of elasticity a finite sound velocity arises from a fourth-order
time differential equation, leading to two characteristics, i.e., two sound velocities.
[Courant and Hilbert, 1961, John, 1982, Taylor and Yau, 2003]
For the modelling of beams the fact that the position of a point along the beam
X(x) depends on the integral over the beam from its fixed position X(0) = x0 leads
to most problems. In many approximations based on the bending and torsion of
the beam, the position will not be given accurately. Furthermore, energy might not
be conserved under motion, due to the finite numerical accuracy of the integration
over the beam coordinate x. If one has some idea of the possible positions of the
beam X(x), it would be advantageous to use that information and derive the elastic
energy and kinetic energy in some finite-dimensional configuration space, i.e., mode
shapes. The analysis of the function space X(x) ∈ D is central to the functional
system analysis.
The second aspect under investigation in this thesis are the wide variety of
boundary conditions. For the Euler-Bernoulli beam there are three well-known
boundary conditions. See Table 6.2. In terms of the power products, Eq. 6.2 and
Eq. 6.3, these boundary conditions correspond to the absence of two types of energy
transfer: the force-velocity and the torque-rotation, in the limited context of the
linearized beam model where y(x) ≈ 0. The beam is therefore isolated from its
surroundings, which is possible in a variety of manners.
6.2. FLUID MODELS 101
name equations variables

ideal, Euler ∂t u + (u · ∇)u = −∇p u: velocity
potential flow ∇(∂t φ + 12 |u|2 + w) = 0 φ: flow potential
R
irrotational u = ∇φ w = V dp: enthalpy
ideal gas ∂t ρ + ∇(ρu) = 0 ρ: density
∂t + (u · ∇)u = − Cρ ∇ργ γ = cp /cV
Stokes, ∂t u = ν∆u − ∇p ν = µρ: viscosity
viscous ∇·u=0 p: pressure
Navier-Stokes ∂t u + (u · ∇)u = ν∆u − ∇p
∇·u=0
generalized ∂t ρ + ∇ · (ρu) = 0 θ: temperature
Navier-Stokes, ∂t u + (u · ∇)u = µρ (∆u − 31 ∇∇·)u − ρ1 ∇p κ: heat conductivity
κ θ Ψ
thermal ∂t θ + u · ∇θ = cV ρ ∆θ + cV ρ (∇ · u)ρθ + cV ρ cV : specific heat
µ
Ψ= − 32 µ(∇
· u) + 2
2 (∂i uj + ∂j ui )
2
Ψ: dissipation function
Table 6.3. Fluid models
6.2. Fluid models

Fluid models are notoriously abstruse. [Mises and Friedrichs, 1971, Landau and
Lifschitz, 1987b, Sod, 1985] A wide combination of effects are taking into account
in the different models. It already starts with the co-moving coordinate system.
The time variation of a quantity at a particular point in the flow, say at a particle
that moves along with the stream, depends both on time and the velocity:
d ∂ ∂x ∂
(6.12) = + ·∇= + (u · ∇) .
dt ∂t ∂t ∂t
where u is the velocity.
Furthermore, fluid has a mass, inertia. Therefore it has the tendency to re-
tain its velocity and direction. Forces on the fluid are either external, through
boundaries and pressure, or internal, due to friction and gravity. If the friction is
neglected, and the flow dynamics is dominated by the inertia, it leads to the ideal,
or Euler, fluid model. This model can also be expressed in an energy relation,
which is called the potential model. If, on the other hand, the inertia is neglected,
and the forces in the flow only arise through friction, it yields the Stokes flow. If
both the friction and the inertia are taken into account, the common model is the
Navier-Stokes flow. However, due to the friction, the flow dissipates energy. The
mechanical energy is converted into heat. The energy conserving model, including
heat generation and flow, for fluid is the generalized, or thermal, Navier-Stokes
equation. Table 6.3 gives a list of well-known models for fluids.
Besides the microscopic fluid models, in terms of partial differential equations,
fluid systems can also be analyzed globally. The global analysis often leads more
quickly to a qualitative answer. One important number which arises in the global
analysis is the Reynolds number Re, which indicates if the solution will be laminar,
irrotational, or whether eddies and turbulence will arise. The number is a simple
variable dimensions typical symbols

length m l
time s t
mass kg m
density kg m−3 ρ
velocity m s−1 v, u
acceleration m s−2 a, g
force N = kg m s−2 F
pressure N m−2 = kg m−1 s−2 p
current A I, J
charge As Q, q
temperature K T, θ
energy J = N m = kg m2 s−2 E, H, W
power J s−1 = kg m2 s−3 P
Table 6.4. dimensional constants
combination of size, velocity and viscosity:

size |u|ρ size |u|
(6.13) Re = = .
µ ν
The size in the equation is a typical size of the problem, the diameter of a tube,
the width of a channel. Hence the Reynolds number may vary with the choice of
typical size. However, this order of magnitude is small compared to the several
orders of magnitude which occur in the variety of fluid problems.
In a sense, systems can be mapped to a simple one-dimensional problem of
the flow in a pipe, assuming the dimensions, such as length and velocity, can be
mapped from one problem to the next. The dynamical regimes of the archetypical
system, given by the value of the number, will apply to most systems. However,
sometimes care must be taken to choose the right value for the system parameters,
such as size or velocity. The flow in a pipe with the radius size = r and a mean
velocity u will have typically higher values Re for the regimes as a ball with radius
size = r, moving through a stationary fluid with the velocity u. However, these two
systems have the same regimes.
The dimensional constants of fluid dynamics, such as the density, viscosity, and
conductivity are combined with system parameters, such as size, velocity, and pres-
sure, to yield dimensionless numbers which will give a good indication of the flow
solution, not only laminar or turbulence, but also viscosity or inertia dominated,
and conduction or convection dominated.
6.3. Dimensional analysis

The construction of an invariant number, such as the Reynolds number, from
dimensional constants, as given in Table 6.4, material constants, as given in Ta-
ble 6.5, and system dimensions can be performed systematically using Bucking-
ham’s method. [Welty et al., 1984] Each dimensional constant can be expressed
in terms of length (m: meter), mass (kg: kilogram), time (s: seconds), current
(A: Ampère), and temperature (K: Kelvin). Expressing the different constants of
a problem in terms of the fundamental dimensions, gives a redundancy, which are
6.3. DIMENSIONAL ANALYSIS 103
variable dimensions typical symbols

density kg m−3 ρ
viscosity kg m−1 s−1 µ(η)
kinematical viscosity m2 s−1 ν
−2
elasticity modulus Nm = kg m−1 s−2 E
thermal diffusion W m K = kg m s−3 K−1
−1 −1
λ
heat capacity J kg−1 K−1 = m2 s−2 K−1 cp
Table 6.5. material properties
precisely the number of invariants. Mathematical physicists use this reduction ex-
tensively to remove most dimensional constants from the equations. Experimental
physicists have come up with a variety of experiments to determine the scaling be-
havior, i.e., the dimensionless invariants, for the same dimensionless constant, such
as Reynolds number. [Van Lammeren, 1944]
In physics and chemistry these dimensions play an important role. They imme-
diately disqualify certain models, due to the incorrect dimensions in the equations.
Every term in the force balance equation must have the dimensions of force, New-
tons: N = kg m s−2 . Likewise, in many cases one seeks the relevant dimension,
which causes a physical effect. For example, in deep water the relation between the
wave length and the wave velocity of a wave can only be determined by the few
physical constants in play: the gravitation and the density, which causes deep water
waves to make a fixed angle with a ship of approximately 23 degrees.[Lighthill, 1978]
In hadronic physics one still seeks the relevant physical variable of the dimension of
length, which causes the strongly interacting elementary particles to be of the size
they are. This, yet unknown, process is even referred to as dimensional reduction.
A systematic dimensional analysis of a system is Buckingham’s method. It
consists of a matrix, where horizontally all the terms of the equations of motion
are written, and vertically all the powers of the fundamental dimensions. The rank
of the corresponding matrix is the number of dimensional constants, the number
of columns minus the rank is the number of dimensionless constants. Thus equiv-
alently, the null-space of the powers matrix, For example, in fluid dynamics there
are two dimensionless constants, the Reynolds and the Euler numbers. The matrix,
with force, velocity, density, viscosity, and size, has the rank 3:
 
F v ρ µ l
 kg 1 0 1 1 0 
(6.14)   .
 m 1 1 −3 −1 1 
s −2 −1 0 −1 0
The null-space of the matrix gives rise to dimensionless constants. The two dimen-
sionless constants, which can be constructed from these five system dimensions are
the Euler number, corresponding to the null-vector (1, −2, −1, 0, −2):
F
(6.15) Eu = .
l2 ρv 2
and the Reynolds number, corresponding to the second independent null-vector
(0, 1, 1, −1, 1):
lρv lv
(6.16) Re = = ,
µ ν
name numerator force denominator force equation

2
Froude inertia force gravity force Fr = ugl
p
Euler pressure force inertia force Eu = ρu2
Reynolds inertia force viscous force Re = lu ν
νc
Prandtl momentum diffusion thermal diffusion Pr = κρp
Nusselt conductive resistance convective resistance Nu = λl κ
Peclet inertia force thermal diffusion Pe = Re · Pr
Stanton viscous force heat transfer St = cpλluρ
ρuc
Graetz heat transfer viscous force Gz = λl p
Knudsen container size mean free path Kn = Ll
l3
Grashof thermal expansion buoyancy viscous force Gr = αgT ν2
ν
Schmidt viscous force diffusion pressure Sc = D
2
Bond gravity force surface tension Bo = ρgL γ
Biot surface thermal conductivity material conductivity Bi = Rlλ
Fourier heat flow heat capacity Fo = ρcλT pl
2
Mach speed sound speed M = vc

v
Rossby inertial force Coriolis force Ro = Ωl
αcp ρ0 gT l3
Rayleigh thermal expansion buoyancy thermal diffusion Ra = γλ
Table 6.6. dimensionless constants, most of them are empirical
relations for particular complex thermofluidic phenomena. The
dimensionless constant width over length, d/l, for example for pipe
flows, will yield the same constant but for a different effect, in, e.g.,
St versus Gz.[Holman, 1981]
which makes the introduction of the kinematical viscosity ν natural, to simplify the
expression for the Reynolds number. Clearly many other choices of null-vectors are
possible, but they may be expressed in terms of the Reynolds and Euler number.
The Euler number therefore plays a smaller role in stationary and quasi-stationary
situations, which are characterized by the Reynolds number. A small list of well-
known dimensionless constants are given in Table 6.6.
The Euler number indicates the systems inertial resistance to motion due to
the force. The Reynolds number indicates the balance between viscous force and
momentum.
Another, less systematic, but more intuitive way to proceed in the recovering
the dimensionless characterizations of a system is by inspecting the equations of
motion. Total force in a system are the balance different forces, such as inertial
forces, applied forces, frictional forces, and potential forces. The equation of motion
is the direct sum of all these terms, and any ratio of two terms give dimensionless
number, which indicates the dominance of one term over the other. In Table 6.6
we give a number of common numerator and denominator forces. However, these
typical forces may depend on the interpretation. For example, an equation involving
pressures and an equation involving forces are fundamentally the same, although
in terms of dimensions they differ an area m2 , since pressure times area is force.
The interpretation may differ with these dimensions.
6.5. ENERGY-BASED DISCRETE FLUID MODEL 105
6.4. Laminar flow around a cylinder

One of the best known example of dimensional analysis is for the flow around a
cylinder. [Feynman et al., 1964] One is not yet able to fully model all the phenomena
[Landau and Lifschitz, 1987b] that occur in such flow. Especially the turbulent
flow and the transition between turbulent and turbulent boundary flow give rise to
many questions. However, the single dimensional constant; the Reynolds number
Re determines the different physical regimes, in the sense of typical flow behavior.
For the flow around a cylinder the Reynolds number Re is defined as:
[density][velocity][radius] ρvL
(6.17) Re = = .
[viscosity] µ
The Reynolds number regimes are the following:
< 20 Steady laminar flow, smoothly around the cylinder.
20 ∼ 150 Laminar eddy flow; behind the cylinder, steady back flow occurs.
150 ∼ 105 Turbulent flow; the eddies separate from the cylinder and move off irreg-
ularly.
< 105 Turbulent boundary layer flow; behind the cylinder the flow is completely
irregular like a trailing tail.
Hence, for many different situations, and many different fluids and gases, the
same situations occur for the same Reynolds number. Likewise, more complex
systems could be characterized by the same regimes. The flow in a pipe or around
a corner yields similar flow patterns as the cylinder problem.
6.5. Energy-based discrete fluid model

Different models of systems start with the choice of variables. A different
choice of variables may yield a solution where the original equations in the original
variables did not. Especially in the case of intrinsically nonlinear systems, such
as the full Navier-Stokes equations, the choice of variables may make the world of
difference. For example, entropy variables can be used to study the transition from
compressible to incompressible fluid dynamics.
In this section, we introduce a different type of discrete model for fluid, based
on the energy in the fluid. In many cases, the given variables, such as velocity
and pressure, and the given equations, such as the Navier-Stokes equations, are
taken for granted as the starting point of the modelling process. Not in all cases
these may be the appropriate variables. For example, the standard variables of
the Navier-Stokes equation are cumbersome for energy conservation. We show an
alternative choice of variables, for which energy conservation can be implemented
explicitly, as an example for modelling.
Instead of the pressure we use the energy as extra variable beside the velocity
vector u. The velocity field of fluid plays a dual role. First, it determines the ge-
ometry; the direction of the flow, and, second, it represents the momentum and the
kinetic energy of the flow. Different geometry, such as downstream, or Lagrangian,
coordinates are based on the geometric role of the velocity. It is also possible to
use a fixed grid with finite cells. Each cell has an associated energy Ei and a mean
velocity vi . The energy is passed on from one cell i to the next j through the mean
velocity 21 (vi + vj ). The amount is determined by the area of the common face aij ,
the normal nij , and the volume Vi and Vj :

aij nTij (vi + vj )
(6.18) ∆Eij = Ei ,
2Vi
where ∆Eij is the energy transferred from cell i to cell j if the velocity is in the
direction nTij (vi + vj ) > 0.
The energy of the fluid in a cell is mainly the kinetic energy, due to the incom-
pressible nature of fluid:
1
(6.19) Ei = ρi Vi viT vi .
2
In some cases the velocity is reduced in favor of a build up of pressure, or body
forces, such as gravitational forces may play a role. The differences between the
kinetic energy as given by the velocities vi and the energy Ei , which include pressure
and potential energy, give rise to forces which will change the momenta.
The viscosity affects the velocities and energies of neighboring cells perpendic-
ular to the direction of velocity. Furthermore, it depends on the differences of the
velocities in neighboring cells, instead of the mean, as velocity differences cause
shear forces. Here we will not investigate the particular dynamics which arises
from these effects. We have showed that energy and velocity, and the basic transfer
of energy, Eq. 6.19, due to the velocity, can be expressed in an energy-conserving
manner.
In practice separating the different energies, forces, and effects allows us to
write a numerical scheme to balance each of this relations. For incompressible
flows the compensating pressure should be balanced completely to assure that the
energy and velocity are consistent. At each time step an iteration to converge
to the incompressibility constraint ∇ · v = 0 may be required. In practice, the
incompressibility is never exact, as it would correspond to an infinite sound velocity.
However, for all practical purposes, the fluid may be perfectly stiff with respect to
the motion considered.
The numerical implementation will play an important role in further details of
the model and the integration scheme to be chosen. This goes beyond the scope of
this section. It is just to show a simple alternative in the study of fluid dynamics.
6.6. Control of the Voith-Schneider propeller

In the design of continuous components optimization plays an important role.
In Section 12.4 shape optimization is discussed as an example of structural optimiza-
tion. In this section we discuss an example which is closer to control optimization.
[Kirk, 2004] In that process, it is useful to characterize different components of a
larger system physically. In this case the blade characteristics of a propeller are
used to analyze and optimize the control of the Voith-Schneider propeller.
Ever since the mid-nineteenth century it is understood that lift, rather than
drag, is the preferred means of ship propulsion. The screw propeller is therefore
ubiquitous. However, it is usually designed for a constant linear velocity, with
larger, slower propellers, for big ocean ships, and proportionally smaller, faster
propellers for more versatile boats. However, the fixed screw propeller is not so
useful for boats with variable velocity and trust, and sharp turns, such as ferry
boats and tugs. In that case, variable pitch propellers and the Voith-Schneider
propeller are often used. The Voith-Schneider propeller consists of a set of rotating
6.6. CONTROL OF THE VOITH-SCHNEIDER PROPELLER 107
Vrotation
Vrotation
F
V
F F
θ
Vrotation
Vship
Vrotation
Figure 6.2. The Voith-Schneider propeller, viewed from the top.

The blades rotate around in the big circle, while the orientation
n(θ) changes along the circle.
vertical blades at the bottom of the hull. [Van Lammeren, 1944] The incident
angles of the blades with respect to the tangential velocity can be altered to create
positive or negative lift, depending on the position of the blade with respect to the
ship’s direction. See Figure 6.2.
The lift and drag of a foil, or blade, is a function of the velocity V and the
incidence angle α. The static drag γ acts opposite the direction of the velocity,
while the dynamic drag and the lift are due to the deflection of the momentum of
the flow. A column of water, in the direction of the flow will hit the blade, and
is deflected as light of a mirror, yielding a force. We represent the incidence angle
and the scalar velocity together as a vector velocity V, where the blade direction is
given by the unit normal vector n, perpendicular to the bottom, of which the flow
reflects. The incoming flow has direction V, the outgoing flow is reflected and has
direction V − 2n(VT n). The difference is the change in momentum and associated
with the force on the blade. The force on the blade is given by:
(6.20) F = −γV − 2A(VT n)n ,
where the static drag γ, the blade normal n, and effective cross section A are the
blade constants for water flow. See Figure 6.3. The flow hits the blade and the
component of the velocity perpendicular to the blade surface is reflected, causing
a change in momentum double to this component. These results are for stationary
laminar flow. [Breslin and Andersen, 1996] In the case of the Voith-Schneider
Force
Flow deflection
n bou
nda
V ry
A lay
er
effective cross section airfoil normal
Figure 6.3. The typical lift force of a blade in a flow can be

seen either in terms of the pressure, or as the momentum change
associated with the change of the flow direction. Subtleties at the
initial and final edge of the flow separating blade determine the
qualities, such as stalling angle and static drag, of the blade shape.
propeller the flow is neither stationary nor laminar, due to the rotation of the
multiple blades.
The Voith-Schneider propeller is notoriously difficult to model numerically, due
to its extent, and the complex flow around the moving and rotating blades. Instead,
we will study the propeller analytically, but in a systematic manner, such that we
are convinced that all the effects are named and estimated. At the moment we will
not use experimental input to avoid modelling bias, and to test the the strength
of the systematic approach, which could lead to new and different insight into old
problems.
The Voith-Schneider ship propeller is also a complex control problem. [Jürgens
et al., 2006] The control of the orientation of the blades n, the incidence angle,
during the revolution in the bigger circle, will increase or decrease the lift. To
optimize the thrust, due to the sideway lift, both the incidence angles and the
overall rotation velocity ω should be controlled. In many cases, the system is
optimized through model identification using experiments. In some cases long and
complicated numerical simulations are required to determine the optimal settings
for a number of particular cases, ship velocity, rotation velocity, thrust, and so on.
Blades have a static drag, a dynamical drag and a dynamical lift. [Kermode,
1987] The static drag is the minimal drag, γ, for a particular velocity. The dynam-
ical properties are due to the deflection of the water stream, causing both a lift and
a drag at the same time. It can be seen as a column of water hitting the blade and
being deflected. See Figure 6.3 The momentum rate-of-change of the water equals
the force on the blade. The amount of deflected water depends on the effective
area, A, of the blade and the angle of incidence. Changing the incidence angle n
during a cycle θ ∈ [0, 2π] the lift in different direction varies with the angle. The
mean lift, or thrust, is the integral over all angles:
Z 2π
1
(6.21) Fmean = dθF(θ, V, n(θ)) .
2π 0
6.6. CONTROL OF THE VOITH-SCHNEIDER PROPELLER 109
90
500
200
60 100
50
20
10
incidence angle
30 1
-30
-60
-90
0 60 120 180 240 300 360
blade position angle
Figure 6.4. The optimal incidence angles for the blades are plot-
ted for different values of γ/2A. The fast change at the angle 180
degrees is associated with the blade moving opposite the thrust
direction, where the static drag contributes to the thrust.
The velocity of the water near the blade is the sum of the rotation velocity of the
blade rωeθ and the ship velocity Vship . The laminar independent blade approxi-
mation of the thrust as function of the rotation velocity ω and the incidence angle
function n(θ) is:
Z 2π
1 T
(6.22) Fmean = − dθ(γVship + 2An(Vship n + rωeTθ n)) .
2π 0
The static drag averages to zero, but it will contribute to the required torque from
the ship motor:
Z 2π
1
(6.23) Taxis = − dθ(γr2 ω + 2AreTθ n(Vship
T
n + rωeTθ n)) .
2π 0
To first order, the velocity of the ship may be ignored: Vship = 0. The efficiency
can be optimized through the optimization of:
Z 2π Z 2π −1
r|Fmean | T T T 2
(6.24) max = dθ2A(ex n)(eθ n) dθ(γ + 2A(eθ n) ) ,
|Taxis | 0 0
where the force is set in the x-direction ex . The result is independent of the radius
r and the rotation velocity ω. It depends only on the ratio of the static and
dynamic drag γ/2A. The nonlinear optimization problem is easily solved with the
methods described in Chapter 10, after a small term (∂θ n)2 is added to remove the
solution jumping half a period ∆θ = π. The results of the optimization are shown
in Figure 6.4. The optimal blade orientations in a cycle are plotted for different
values of γ/2A.
The Voith-Schneider propeller is a nonideal gyrator. [Karnopp et al., 2000]

Given the axis velocity it yields the thrust of the propeller. The drag causes losses
in the propeller, independent of any ship velocity generated. Drag can be seen
as the internal friction of the propeller. The actual velocity will depend on other
factors, such as the ship hull shape and weight.
There are several improvements for this result. First the flow is not really
laminar, since the blade changes direction in the flow, which is even enhanced by
the changing incidence angle of the blade. Furthermore, the blades disturb the
flow, this disturbance increases with the increasing incidence angle. In the worst
case the blades cause a vortex which sucks the water up from below and swings
it out via the blades. Ventilation systems placed on roofs of buildings sometimes
work through this principle. Furthermore, changing the blade angles may cost a
considerable amount of force, which should be added to the operating torque Taxis .
The disturbances of the flow should be quantified in order to go beyond the
independent laminar results. We use the thrust of each blade as the basis of the
disturbance it causes. The system itself is a feedback loop.
We investigated only the simplest cases of interference between different aspects
of the fluid flow around the blades. However, this approach to modelling the Voith-
Schneider propeller allows us to experimentally investigate the actual magnitude
of different effects. A consistent model, resulting from experimental input, can be
used to investigate and implement more complex incidence angle dependencies.
In the laminar independent blade case we can already see the results of the
drag of the asymmetry of the solution. The drag in the direction of the thrust is
advantageous, while in the opposite direction it is unwanted. Hence, for the blade
angle opposite wanted thrust direction the drag should be minimized, which is at
a quarter circle angle of the maximal and minimal lift.
6.7. Elasticity
The standard mechanical engineering formulation of strain and stress starts
with the displacement u = (u, v, w) from which the elongations follow: [Love, 1944,
Truesdell and Noll, 1965, Washizu, 1968, Soutas-Little, 1973, Fetter and Walecka,
1980]

∂u ∂v ∂w
(6.25) (εx , εy , εz ) = , , .
∂x ∂y ∂z
The shear γ is defined as:

∂v ∂u ∂w ∂u ∂w ∂v
(6.26) (γxy , γxz , γyz ) = + , + , + .
∂x ∂y ∂x ∂z ∂y ∂z
The strain tensor arises from combining the stretches or elongations εi , with the
shear εij = 21 γxy such that is yield the symmetric strain tensor:
 
εx εxy εxz
(6.27) V =  εxy εy εyz  .
εxz εyz εz
The odd combinations of the mixed derivatives (i.e., ∂x v − ∂y u) correspond to
rotations, and do not cause any strain.
For the purpose of the tensor, the off-diagonal shear strain εij is introduced,
with a factor of a half, such that the sum of the squares of the coefficients yield the
6.7. ELASTICITY 111
linear shear energy. (In some literature, the factor of a half is in the off-diagonal
elements of the stress tensor, rather than in the definition of εij .)
The change elastic energy is the coefficient-wise, or direct product of the change
in strain dε, with the corresponding force or effort, which is called the stress tensor
S:
 
σx τxy τxz
(6.28) S =  τxy σy τyz  ,
τxz τyz σz
which is also denoted by σij , where σii = σi and σij = τij , such that the total
elastic energy is given be the integral over the change of strain:
 
Z X Z εi X Z γij
(6.29) W =  σi dεi + τij dγij  .
Volume i i≤j
Note that the summation over the shear terms is only over three terms: (i, j) =
{(1, 2), (1, 3), (2, 3)}. In the case of linear elasticity this reduces to:
1
Z X X
(6.30) W = σi ε i + τij γij .
2 Volume i
i≤j
The linear relation between stress and strain for isotropic material is:
0 1 0 10 1
Eεx 1 −ν −ν σx
B
B Eεy C B
C B −ν 1 −ν CB
CB σy C
C
B Eεz C B −ν −ν 1 CB σz C
(6.31) B C=B CB C ,
B
B Gγxy C B
C B 1 CB
CB τxy C
C
@ Gγxz A @ 1 A@ τxz A
Gγyz 1 τyz
where E is the Young modulus, ν = 1/m is the Poisson ratio, and G the shear
modulus. The shear modulus can be expressed in terms of the Young modulus and
the Poisson ratio:
(6.32) E = 2G(1 + ν) .
The matrix relation for the diagonal parts of the strain and stress tensors is
easily inverted:
 −1  
1 −ν −ν 1−ν ν ν
1
(6.33)  −ν 1 −ν  =  ν 1−ν ν  .
(1 + ν)(1 − 2ν)
−ν −ν 1 ν ν 1−ν
The inverse is singular for ν = 12 and ν = −1.
In itself elasticity is a well-defined problem. However, boundary conditions and
matching conditions are the complicating factors, since they are not expressed in
the stress and strain tensors, but in displacement u and force f :
X X
(6.34) fi |surface n = nj Sji = σi + τij ,
j j6=i
which requires one to separate rotations in the displacement u, and combine terms
to a force f . Furthermore, it might be possible to construct stresses and strain
which do not correspond to conservative force fields and displacements, respectively.
Therefore, there are a number of compatibility conditions.
In many cases not the full three-dimensional theory of elasticity is used. For
different types of two-dimensional theory, such as plates, and sections of a three-
dimensional structure, different specialized results exist. [Soutas-Little, 1973, Love,
1944] The lower-dimensional theory can be derived from the three-dimensional the-
ory by assuming homogeneity, and particular boundary conditions for the depen-
dent directions. The two typical cases are free boundaries, in the case of plates,
and fixed boundaries in the case of symmetry. The reduced order models depend
only on the remaining reference coordinates (x, y) or only x.
The reduction of the three-dimensional theory to two dimensions can be done
in two general manners. First, assuming the third direction to be free. Second,
assuming the third direction to be fixed. The first case occurs, for example, if one
considers a thin plate where the thickness of the plate can vary freely. The second
case occurs for cases with symmetry, the direction in which the system is symmetric
can be ignored, but the boundary conditions are fixed. Such conditions occur in
very long objects, where only the transverse deformation is considered, like railroad
tracks.
The simplest procedure to reduce the dimension is to use the form of the energy
relevant for the problem. For the fixed positions in one or two dimensions the
energy expressed in terms of the elongations εi is the best starting point as εz = 0
or εy = εz = 0 are simple reductions of the stiffness matrix:
 T   
ε E E⊥ E⊥ εx
1 x   k
(6.35) W = εy E⊥ Ek E⊥   εy  .
2
εz E⊥ E⊥ Ek εz
T
1 εx Ek E⊥ εx 1
(6.36) → W2d = → W1d = εx Ek , εx
2 εy E⊥ Ek εy 2
where Ek = (1 − ν)E/((1 − 2ν)(1 + ν)), E⊥ = νE/((1 − 2ν)(1 + ν)). Similarly, for

free motions, force in the irrelevant direction is zero σi = 0. Hence energy expressed
in terms of the stresses σx , σy , and σz is the easy starting point for reduction:
 T   
σx 1 −ν −ν σx
1
(6.37) W∗ =  σy   −ν 1 −ν   σy  ,
2E
σz −ν −ν 1 σz
yielding the two- and one-dimensional submatrix of the compliance matrix.

Another approach would be to use Guyan reduction. See also section 8.11
on effective Hamiltonians. The boundary condition σz = 0 corresponds to the
minimal-energy value for εz given εx and εy :
 T   
ε E E⊥ E⊥ εx
∂W 1 x   k
(6.38) 0= = ∂ εz εy E⊥ Ek E⊥   εy  .
∂εz 2
εz E⊥ E⊥ Ek εz
This reduces to:

E⊥ (εx + εy )
(6.39) εz = − ,
Ek
6.8. VISCOELASTICITY 113
or as a constraint relation expressing all three elongations in terms of the two

independent εx , εy :
   
εx 1 0
 εy  =  0 εx
(6.40) 1  ,
−ν −ν εy
εz 1−ν 1−ν
with the elastic energy:

E 1 εx 1 ν εx
(6.41) W2d = .
2 (1 − ν 2 ) εy ν 1 εy
In the case of two unconstrained motions both εy and εz are expressed as a function
of the independent direction εx . The free motion corresponds to the minimal energy
for a given εx :
∂W ∂W
(6.42) = =0 ,
∂εy ∂εz
yielding the two constraints:
   
εx 1
(6.43)  εy  =  −ν  εx ,
εz −ν
with the elastic energy:
E 2
(6.44) W1d = ε ,
2 x
which is Young’s result for the elongation of a rod, which radius will decrease
appropriately, given by the constraint relation.
In short, inverting the energy function E = 12 xT Ax to a co-energy function
E = yT A−1 y where y = Ax and selecting the appropriate submatrix from the
∗
full matrix corresponds to inserting the solution of the variational minimum with
respect to the dual variable, since:
∂E
(6.45) x∗ = y = =0 ,
∂x
In some cases it may be convenient to work with strain variables, in other cases with
stress variables. In the linear theory there is no fundamental difference, as long as
the stiffness is non-singular. The singular stiffness matrix arises in the theory of
partial differential equation, for example, if we do not use the stress and the strain
but the force and displacement as our variables.
6.8. Viscoelasticity
Viscoelasticity [Christensen, 1971] is the dynamical theory of plasticity. Objects
under strain may deform, and part of these deformations can be permanent. Work
applied to the system is partly used to change the rest state. Such systems have
a hysteresis at best and usually orbits in force-position space which do not close.
The result varies with the trajectory in state space taken by the applied force.
Similar to heat transfer in fluids, the theory of viscoelasticity, especially thermo-
viscoelasticity, is a phenomenological theory due to the variety of effects. Two
simple and well-known characterizations of viscoelastic material are in terms of
springs and dampers, either in series in the Maxwell model, or in parallel in the
Kelvin, or Voigt, model. See Figure 6.5. The Maxwell model is a viscoelastic
Kelvin Maxwell
Figure 6.5. The Kelvin and the Maxwell models for viscoelastic-
ity. The respective generalized models consists of series and par-
allel combinations of the simple models with a single spring and
damper. The Kelvin, or Voigt, model corresponds to a viscoelastic
solid, while the Maxwell model is a viscoelastic fluid.
fluid, which can not sustain internal shear indefinitely, while the Kelvin model is
a viscoelastic solid, eventually regaining its original shape. In practice, both the
parallel and the series coupling are present, which, however, yield a complicated
network even for homogeneous materials. The generalized Kelvin model consists of
several Maxwell spring-damper combinations in parallel.
More complex models have more parameters to be fitted to experiment. Typ-
ical response results to be fitted to the generalized Kelvin or Maxwell models are
combinations of exponentials related to the damping and the elasticity. Another
typical experiment would be the propagation of waves in viscoelastic material. The
velocity and damping of the waves for different frequencies may be related the
parameters of the model at hand.
Viscoelasticity, especially in the combination with temperature effects, would
benefit from a microscopically based model, similar to the Vanderwaals model for
gas-liquid phases of materials. In applied mathematics the study of polymers and
percolation has led to considerable insight in the phases of polymers, glasses, and
liquid crystals, exhibiting the viscoelastic properties. The bulk properties, however,
have not been of much interest of the modern academic researchers. In the study
of friction dynamical and elastic effects are taken into account in dynamical models
depending on the history of the motion. [Dahl, 1976] Furthermore, the study of
macromolecules, such as DNA, in biochemistry, showed the complex static solu-
tions, corresponding to local minima in energy, slowly deforming through thermal
and mechanical excitations. Biological macromolecules, like enzymes, seem to en-
hance deformation processes through mechanical restrictions, combined with some
electrostatic forces from the polarity of the molecules.
6.9. Electrodynamics
The four Maxwell equations are the start of modern physics as we know it.
[Maxwell, 1891, Jackson, 1975] What were before separate phenomena, such as
magnetism, electricity, currents, and light finally came together in a single model.
Different pieces were already collected by great scientists such as Faraday, Coulomb,
6.9. ELECTRODYNAMICS 115
Ampère, and Ohm, but the crown was placed by Maxwell, who added the displace-
ment term to the relation between magnetic flux and current. [Taton, 1965] The
Maxwell equations were thereby consistent with the conservation of charge, satis-
fying the charge continuity relation.
The Maxwell equations themselves yielded a wave equation, which described
electromagnetic waves. The presence of the wave solution, and thereby electromag-
netic radiation leaving the system, makes the dynamical electromagnetic theory
complex, in the sense that it is not always useful to search for stationary or quasi-
stationary solutions to the Maxwell equations.
Despite the seeming simplicity of four Maxwell equations, the construction of an
appropriate model for a particular system, with the right independent, or states, and
dependent variables, or constitutive or closure relations, and the correct boundary
conditions, or input, can be a formidable task.
The four Maxwell equations are respectively, Biot-Savart law with the Maxwell
displacement term ∂t D, Faraday law, Poisson law and the Gauss law:
(6.46) ∇ × H = j + ∂t D ,
(6.47) ∇ × E = −∂t B ,
(6.48) ∇·D=ρ ,
(6.49) ∇·B=0 .
In many cases, materials have simple direct responses to an external field, such as
polarization and magnetization. The microscopic charge densities and eddy cur-
rents generated are not considered part of the macroscopic currents j and charges
ρ present in the Maxwell equations. The constants appearing in the Maxwell equa-
tions may vary from author to author. There are several units available to define
the electromagnetic fields. Here we use SI (MKSA) units, but Gauss and Heaviside
units are also very common. [Jackson, 1975]
The magnetic induction B is related to the magnetic field H through the in-
duced magnetization M:
1
(6.50) H= B−M .
µ0
while the electric displacement D is related to the the electric field E through the
induced polarization P:
(6.51) D = 0 E + P .
Modelling electrodynamical systems is complex for several reasons. First of all, the
Maxwell equations themselves are already a set of complicated equations, mainly
in the coupling of the equations, and the constraints ∇ · D = ρ and ∇ · B = 0,
which are difficult to implement on general fields D and B. Second, the choice of
independent and dependent variables, and the associated boundary conditions, is
not always clear from the system and its context. Usually, problems are set to be
either a static solution of the constraint, like electrostatic and magnetostatic, or a
electromagnetic wave trivially satisfying the constraints, like radiation and wave-
guide problems. The general electromagnetic problem, including light, or radiation,
of all frequencies, is often avoided. See also Table 6.7.
name variable boundary conditions sources

conductivity j V or E E
polarization D none (local relation) E
magnetization M none (local relation) H
electrostatics E or V E or V ρ
magnetostatics B B j
wave guides E × B(ω) E and B E×B
radiation E × B(ω) E×B ∂t j or ρ̈
electrodynamics E(t) and B(t) E and B j(t) or ρ(t)
Table 6.7. Different types of electrodynamical problems
Relativistic invariance of the Maxwell wave equation; it yielded a constant

wave velocity in vacuum, independent of the velocity of the observer, led to a
juxtaposition in modern physics. For the first time the equations dictated the
development rather than the experiments. Special and general relativity followed
from the analysis of the consistency of the equations of motion, i.e., the Maxwell
equations.
Nowadays, general electrodynamics is often treated in relativistic invariant ap-
proaches. This means that every space-time point is considered as separate. As an
initial boundary value problem (Section 5.3) such approaches will fail, since in that
case the initial surface t = 0, and the initial conditions, or kinematical variables,
are never defined. The kinematical variables, the electromagnetic fields and charge
and current densities, at a given time have to satisfy global conditions. See also
Section 11.11.
The energy density E of the electromagnetic field is a simple local relation:
1
DT E + B T H .

(6.52) E=
2
This energy density is deceivingly simple. The states D and B satisfy the Gauss
laws: ∇ · D = ρ and ∇ · B = 0, and they can therefore not be chosen freely.
Therefore, despite its local appearance, i.e., products of fields at the same point,
the energy density has a nonlocal nature. The change of energy reveals this partly.
The Maxwell equations determine the change of energy:
(6.53) Ė = −∇ · (E × H) − jT E ,
where the first term on the right-hand side is the divergence of the energy flux, given
by the vector of Poynting, E × H, and the second-term corresponds to the coupling
to mechanical energy due to the motion of charge, or current, j in an electric field
E; the power product of effort and flow. [Paynter, 1961] The direction of energy
flux, or radiation, is perpendicular to both the electric and magnetic fields E(D)
and H(B), which are also perpendicular with respect to each other. The Maxwell
equations, together with the energy density, Eq. 6.52, suggest that these are first-
order Hamilton equations of motion. This is not the case. Only the radiative field
has pairs of canonical variables. The remainder of the degrees of freedom are either
constraints, or longitudinal fields given by boundary conditions or by charge and
current sources. See also Section 11.11.
In Cartesian coordinates only six anti-symmetric combinations ∂i Ej Hk of the
27 contribute to the radiation in electrodynamics. Likewise, the material losses
6.10. HEAT TRANSFER 117
are easily described by constitutive relation expressing a current in terms of the

fields: j(D, B). On the other hand, the change in electromagnetic energy Ė cannot
be expressed in terms of applied currents, i.e., D(j) and B(j). Electromagnetic
radiation arises, from material sources, only due to the acceleration of char ges:
∂t j. [Jackson, 1975]
6.10. Heat transfer

The separation of different dynamical regimes for heat transfer is a complicated
problem, especially in the case of viscous fluids. The expansion of fluids and gases
with an increase in temperature yields an upward motion under gravitational forces.
The resulting convection occurs in many different types. Some well-known exam-
ples are the bubbles in boiling water and flames of a fire, which spell already the
complexity of the problem. In the absence of gravity, there is also no convection,
and a candle burns in space slow, blue, and spherical, due to the diffusion driven
burning process.
The convection processes depend on the material properties of the fluid such
as the thermal expansion of the fluid, the viscosity, heat diffusion, the density, and
the heat capacity. The circumstances can also cause variations, such as the column
height and width and other geometric parameters, the temperature difference, and
the presence of disturbances. [McAdams, 1951, Holman, 1981, Vekstein, 1992,
Landau and Lifschitz, 1987b]
Besides convection, there is also heat transfer due to diffusion, or conduction,
and due to radiation. Radiation plays an important role in the case of large surfaces
and large temperature differences. The Stefan law of heat transfer is:
ε 4 4
(6.54) j= σ(Tobject − Tsurrounding ) ,
2−ε
where ε is the emissivity correction for the deviation of the blackbody theory and
σ = 5.6696 10−8 Wm−2 K−4 the Stefan-Boltzmann constant. For example, in the
case of two flat surfaces of each a meter square relatively close together, the heat
transfer from the hotter surface at 370 degrees to the colder surface at 290 degrees
is 761 Watt, assuming the emissivity constant to be ε = 1. The emissivity constant
depends on the material, the wavelength, and the surface coating or roughness. The
wavelength in its turn depends on the temperature. Take for example glass. For
low temperatures it will radiate, and absorbs, very little heat, only close to melting
temperatures a glow appears. However, for most materials a typical emissivity is
ε ≈ 0.6 − 0.95.
In the case of conversion of sunlight into heat, for the purpose of generating
useful energy, by a Carnot cycle, [Callen, 1985] the absorption in the visible range
must be optimized, while the emissivity in the range of 300-500 Kelvin, the sub-
millimeter range of electromagnetic radiation, typical temperatures of sun-heated
objects should be minimized. Special coatings and surface treatments exist to
achieve this goal. Interestingly, the reciprocity theorem, stating that light will tra-
verse an optical path in both directions, in and out, with equal likelihood due to
the time-reversal invariance of the physical laws, is circumvented in this way.
At room temperature T = 290K, the conduction due to radiation for a ε = 1
object is approximately:
3
(6.55) j = 4σ(Tobject − Troom )Troom = 5.5(Tobject − Troom )Wm−2 K−1 .
In many case a value of 1. or 2. Wm−2 K−1 is used.

Normal conduction, due to diffusion, in solids and fluids is usually much larger.
Conduction is expressed in a heat conduction coefficient λ[Wm−1 K−1 ]. The length
scale m−1 refers to surface−1 length. The typical distance length scale, which would
yield the same conduction as radiation transfer at room temperature is ten to
thirty meters for metals, and couple of centimeters to a meter for minerals and
fluids. Aluminum, copper, gold and silver are typical good heat conductors.
Heat transfer is a typical case for dimensional analysis. [McAdams, 1951] The
relevance of different phenomena can be estimated by calculating the relevant num-
bers. However, in practice, experiments are the best test for such systems. In that
case, the dimensional analysis can still be used to interpolate between different
cases, with different dimensions or material constants. Although the underlying
mathematical theories of bulk processes of heat transfer are well known, detailed
surface characteristics, boundary layers, and other microscopic processes will yield
considerable variations in results. [Holman, 1981, Welty et al., 1984]
6.11. Multi-domain models

In Section 6.3 on dimensional analysis, the five fundamental dimensions: mass,
length, time, current, and temperature, of all physical systems were discussed. See
Table 6.4. In many fields of research, such as social sciences and economics, it is
often difficult to compare or combine one dynamical model with another. The vari-
ables may be the same, but the dynamical principles, such as Hamilton equations of
motion, if defined, are not of the same type, i.e., the effect one part has on another
is not defined. Within the physical domains it is possible to compare and combine
elements stemming from different domains, since the same variables and concepts
are used within all these domains. However, not all possible types of interactions
are defined yet. In particular the dynamical effects of the thermodynamic domain,
i.e., non-equilibrium thermodynamics, is mainly studied in linear approximations.
[De Groot and Mazur, 1962] These effects can play important roles at small scales
and in vacuum.
In particular, the energy, expressed in the Hamiltonian, can bring all these
different physical domains together. Each of the physical domains are independent
terms in the Hamiltonian. The multi-domain effects arise from interaction terms
in the Hamiltonian:
(6.56) H = Hmechanical + Helectromagnetic + Qthermal + · · · + Hinteraction .
The energy in the thermal domain, the heat Q, plays a slightly different role than
in the other domains, in the Hamiltonian language used in this thesis. [Breedveld,
1984] Typical interactions between the thermal and mechanical domains are the
thermal expansion, the temperature dependence of the stiffness, and the viscoelas-
ticity. [Kovalenko, 1969] Thermal effects are known for most materials. Thermal
expansion, changes in the stiffness, variations in magnetic properties and polar-
izability are only a few of the cross-over effects between the thermal and other
domains. In the modelling of modern high-precision machines these effects can no
longer be ignored.
The interaction terms in the Hamiltonian represent a consistent manner to in-
clude the cross-overs into other domains. In many cases, especially with the thermal
domain, the Hamiltonian is treated with a parametric temperature dependence. In
6.11. MULTI-DOMAIN MODELS 119
an open thermal environment, with a surrounding temperature, this might be a

good manner to proceed, at the loss of energy conservation. It is similar to the
electric resistors in electric networks, where energy leaves the system in the resis-
tor port as heat. However, at small scales and in vacuum, the heat capacitance
and heat generation will play a dynamical role, with temperatures throughout the
system which are never in balance with the surroundings.
The choice to restrict the attention and analysis to a typical physical subdo-
main, such as mechanical, electrodynamical, or thermal, is to simplify the model.
The other domains may yield only small effects in the problem under investigation.
The separation in different physical domains is mainly a conceptual one, to focus
the thoughts. However, one need to validate the choice to ignore the other physical
domains. The respective Hamiltonians Hdomain consist of all the types of energy
available in that domain. The interaction Hamiltonians form mainly the basis for
the ports or interfaces between different domains. Especially, the bilinear form of
the Hamiltonian, with a linear dependence on each of the variables from a particular
domain, has little energy-storage capabilities.
However, it is not always appropriate to express the interaction in terms of
an interaction Hamiltonian. The charged particle in a magnetic field will change
direction, without a change in energy. The interaction does not involve the ex-
change of energy, which is therefore not necessarily appropriate described by a
port-Hamiltonian, based on energy flux. However, the leading bilinear interaction
Hamiltonian between the electromagnetic and the mechanic domain is the potential
energy φ of charge density ρ:
Z
(6.57) Hem,mech = dzφ(z)ρ(z) ,
where Hem,mech is the energy in the mechanical domain, with the matter-charge
density ρ as variable, but a co-energy in the electromagnetic domain, with the
potential φ as effort variable. Matter has besides charge also mass. The power
continuity, and the appropriate port-variables, follows from the time-derivative of
the interaction Hamiltonian. The total energy does not change, but for each do-
main the energy change, and the power flow can be reconstructed. The change of
mechanical energy due to the interaction with the electromagnetic field is:
Z Z
(6.58) Ḣmech = dzφ(z)ρ̇(z) = dzET j ,
where we used the charge continuity equation. If the current j = qv is produced by

the velocity v of a particle charge q, the product qD can be seen as the force f on
the charged particle. Whether the electric field E or the electric displacement D is
used depends on whether the mechanical properties of the medium are taken into
account or not, i.e., whether the current is microscopic or macroscopic. In Eq. 6.58,
with E, the microscopic current is used. [Jackson, 1975, de Groot and Suttrop,
1972]
The Maxwell equations say very little about the forces on a mechanical charge q.
Lorentz introduced a fifth equation, nowadays known as the Lorentz force equation
of the force f on a charge q: [Lorentz, 1915, Jackson, 1975]
(6.59) f = qE + j × H .
The second term will not yield an energy transfer between the electromagnetic
and the mechanical domain, since the force is perpendicular to the motion in the
V F
piezo
V,i F,v
q x
Figure 6.6. The main function of a piezo is the couple two phys-
ical domains: the electric and the mechanical. It can be used in a
Carnot cycle to transform energy from one domain to another.
direction of the current j. However, this second term is the interaction term behind
electromotors and dynamos. Actually, in most cases the magnetic field H(j) is
generated by currents in coils itself, and a more appropriate description, without
the magnetic field H, is given by the Ampère force law: [Jackson, 1975]
µo djT1 (x1 )dj2 (x2 )(x1 − x2 )
I I
(6.60) f12 = − .
4π |x1 − x2 |3
The symmetric geometric form of the line integrals, over the two closed current
loops, j1 and j2 , is often approximated such that the orientation invariance of the
interaction is destroyed. [Breedveld, 2001]
The energy extends always across the different domains. Even equivalent de-
scriptions for the same energy may exist in different domains, such as, describing
the same electrostatic energy either in terms of the charges of of the fields. The
cross-over effects between can however also be small. In other cases there is little
point to separate the different domains, as the main function is to interlink the
domains. Piezo material is such a case. The electric and mechanical domain are
fundamentally coupled in a single energy function. Figure 6.6 shows the lumped
representation of a block piezoelectric material. The combination of electric and
mechanical capacitance or stiffness, for the same motion, will allow the transfer of
energy between the different domains.
It is important to notice that the interaction between domains in continuous
systems is often distributed. It is not the change in temperature in time, but
the spatial variation in temperature which causes the unwanted deformations in
mechanical structures. If the heat is trapped in the object, the temperature spatial
variations are smaller, although the temporal variations may be large, the uniform
expansion, at typically ∼ 10−5 /K, will not affect the overall shape.
Another peculiar aspect of the physical models is the alternative descriptions of
the same phenomena. In the world of applications different phenomena are seen as
separated. The underlying microscopic theory is however almost always based on
the electric features of the molecules and the atoms of which the material consists.
The microscopic peculiarities of these materials will lead to a range of features,
from interaction with light, to macroscopic electric properties, such as conductivity,
to thermal properties, and elasticity and compressibility. Crystals, or minerals,
are, for example generally, hard, brittle and incompressible, non-conducting, and
transparent, while metals have another range of features which span the different
physical domains. The most interesting aspect of fundamental physics is that these
different domains form together a coherent view of matter. The more microscopic
the model, the more these features arise from a common source. Unfortunately,
6.12. SOME REMARKS ON MODELLING 121
global features arise only with great difficulty from microscopic models, and, as
always, a balance between detail and feasibility must be found.
6.12. Some remarks on modelling

In this chapter we studied the models of physical systems within their own
right. This is the more traditional form of modelling. The distributed interactions
between different physical domains were discussed only in the last section. Our main
interest, the modular view of models, their interaction across domain boundaries,
is analyzed in the next chapter. The models of continuous systems are generally
much more complicated than lumped, low-dimensional models. They are generally
very rich in term of the phenomena that can occur. In modelling for engineering
purposes, the hope exists that the behavior is limited. Therefore we seek within the
full richness of continuous, or infinite-dimensional models, the degrees of freedom
relevant for the operations at hand. The typical operation, the typical input and
output, determines the relevant function. This is the central question posed and
answered in the remainder of the thesis.
CHAPTER 7
Port-based modelling of infinite systems
Global modelling, in terms of lumped parameter systems, can be insufficient

to recover all the relevant phenomena. Internal dynamics, such as vibrations, de-
formations, and delays, may spoil the global model. To improve upon such models
detailed local, internal, or distributed theories, of continuous systems, are required.
However, many local approaches neglect the global features and interconnections,
which have been the main interest.
In this chapter, the global, lumped and local, distributed approaches are linked
together, in more than one way. We show the advantages and disadvantages of
a collocated approach and the FEM approach, and show that the local theory of
the component can be augmented with a local theory of the boundary, in terms
of fluxes. These fluxes are the direct implementation of energy, force, or charge
balance, and form the natural link between the global, or global, modelling, and
the local description.
After a general motivation, the link between the global and local Newton force
law is established. Since this is based on general principles, no detailed local theory
is required. In the next section the features of local theories which lead to balance
laws are investigated, and applied to the inhomogeneous wave equation on a string.
In the next section, the interaction with the surrounding through boundary con-
ditions is investigated, first by domain splitting. In the next section by defining
appropriate fluxes, the boundary conditions, expressed in the energy flux, yield the
core model, which is the local equivalent of the lumped model. The core model
may vary with the type of boundary condition and resulting flux. Finally, the lead-
ing internal dynamics is discussed. The internal dynamics arises from the coupling
with the lumped, or core model, dynamics. We link back to the formal aspects of
Hamiltonian dynamics, and we end with some conclusions and an outlook.
7.1. Motivation
Modelling of systems is based on the natural view of the largest independent
units of a system. In a human body, for example, we recognize the members
separated by joints, although the digits of the hand are much smaller than the
upper leg. A medical specialist might, however, see different units, such as muscles,
organs, and bones. In principle, units do not need to be spatially distinct. Generally
only particular properties, such as electrical, or mechanical, are part of a unit of a
model.
Kirchhoff laws and rigid-body dynamics are successful descriptions for lumped
parameters of large units. However, for accurate measurement and design of the
model units more detailed understanding of the each unit is required. The body,
assumed to be rigid, turns out to be elastic, and electrical circuits may have delays,
or lags, and mutual interference between components, which are not present in the
123
124 7. PORT-BASED MODELLING OF INFINITE SYSTEMS
dynamics of the Kirchhoff laws. These are both examples of internal dynamics.
Instead of a couple of parameters to describe the model, an infinite number, or
continuous set, of parameters arise once the unit, or component, no longer acts
fully coherently or collectively.
The study and modelling of internal dynamics starts usually with a local theory,
expressed in terms of partial differential equations. Successful local theories are
consistent with the global, or lumped results. For example, any theory of elasticity
should propagate the force from one end of a free object to another, otherwise the
micro-balance law, the basis of the force balance, cannot be satisfied. Similarly,
local electromagnetic theories, at the basis of local modelling, should, in practice,
satisfy local current conservation in order to satisfy Kirchhoff laws at a global scale.
If a component is size-extensive, like a homogeneous and isotropic block of material,
one might split the block into small pieces, and local blocks will inherit the global
laws, yielding a direct link between the integral, of the block as a whole, and the
differential, of an infinitesimal part, version of the same conservation law.
Maxwell equations for the local description of electrodynamics were derived to
satisfy conservation laws, such as charge conservation. See Section 6.9. Maxwell
looked toward the theory of hydrodynamics, with divergence-free vector fields, for
inspiration of the final displacement current term in the Biot-Savart law, which
guaranteed the global conservation of charge. The local formulation in terms of
partial differential equations were only considered complete once the global physical
conservation laws were satisfied.
Hence, the local description is connected to the global description via the phys-
ical laws, expressed as conserved global quantities, such as charge and momentum,
on one hand, and, on the other hand, expressed as properties of the fields, like
charge, mass, and current densities. In establishing a close connection between the
local theory, expressed in partial differential equations or their FEM approxima-
tion, and the global theory, in terms of the lumped parameters, the conservation
laws should, and can, play a central role. This is not only for consistency, but
also to enable the modeler, or designer, to shift focus smoothly from global system
properties to details, which might be critical. Some results, in the more limited
context of linear FEM, appeared in [Ligterink et al., 2006].
7.2. Beyond rigid body dynamics

A simple and heuristic example of a bouncing ball should set the mind. In a
lumped description of the bounce of a ball, energy is lost at a bounce, to match
our observation. However, most balls have a life-span of hundreds of thousands
bounces, so it is unrealistic to expect the energy is directly converted through the
plasticity of the ball to heat and structure change, like with a lump of clay. In
a local description energy is converted from the global, or lumped, motion of the
ball, to the vibration of the ball, which is again converted into heat and sound. In
principle, the whole conversion process is interesting, however, depending on the
zeal of the investigator, he or she might wish to truncate the effort at a certain
detail. For example, the amount of conversion from kinetic energy to the domi-
nant elastic mode, and the other elastic energy in higher modes, might already be
enough information to determine the lost of global energy within the given accuracy.
The dominant elastic mode acts as an ideal spring for the ideal, energy-conserving
bounce, while other modes lead to decoherence and the lost of global motion.
7.2. BEYOND RIGID BODY DYNAMICS 125
Like in the example above, vibrational analysis lies at the core of most efforts
to bring to the surface the dominant effects of internal dynamics. This should not
be surprising. Vibrational analysis is the physical description of spectral decompo-
sition, which lies at the core of the most powerful mathematical tool for physical
systems, namely, functional analysis. However, our mental process might work
from the top down: to the global, or lumped, picture we add a couple of degrees
of freedom in the form of low-frequency vibrational modes, the mathematical local
description starts from the bottom up: from a model with an infinite number of
degrees of freedom we should distill a few relevant ones. Spectral decomposition is
a hard problem from the bottom up. See Section 5.2. It is only feasible for dis-
cretized FEM models, which truncate wave numbers at the inverse mesh spacing.
Therefore, they cannot get into the range of vibrational modes, which, for example,
lie at the basis of kinetic energy to heat conversion.
In an appropriate and powerful description of the local details of systems the
global description should meet the local description at some point. We should ask
ourselves the question: what are the local equivalents of the global variables we
use? Physical conservation laws should be the guiding principles in such questions.
For example, take the motion of an extended object. In the case of a rigid-body
description of the extended object the motion of the center-of-mass is a direct
consequence of the applied resultant force, also know as Newton force law:
(7.1) F = MA ,
where A is the acceleration of the center-of-mass and F the force on the center-of-
mass. The capital letters are used for the lumped variables, the small letter for the
distributed variables.
The corresponding local description is not immediately evident. Starting with
the sum, or lumped, mass M , it could easily be derived from the integration over
the mass density ρ:
Z
(7.2) M= ρ(x)dx .
Ω
Furthermore, the acceleration A can be considered the acceleration of the center-
of-mass coordinate X:
1
Z
(7.3) X= x(x)ρ(x)dx ,
M Ω
where the positions are normalized with the mass density, which yields:
(7.4) A = Ẍ ,
for a constant mass density ρ(x). The position X is with respect to an origin
x = 0, velocity and acceleration are affine and depend only on the orientation of
the coordinate system.
The displacement x(x) is a function of the reference coordinate x, such that the
density ρ(x) is invariant. All fields, such as the local forces f (x), are all functions
of the reference coordinates x.
For a rigid body the force law, Eq. 7.1, is the direct sum of all the local forces
f (x):
Z
(7.5) F= f (x)dx .
Ω
It is interesting to notice that, unlike the positions x, the forces are not weighted
by the mass density. The positions are therefore extensive, while the forces are
intensive. In that respect they pair together, which has consequences for the inner-
product, as we will see. In the case of a flexible object, internal forces f⊥ (x), which
sum to zero for the rigid-body force, still contributes to the internal dynamics.
The meaning of perpendicular is different from what one expect naively. Work and
energy are the invariant quantities in mechanics. The invariant inner-product and
the geometry should be based on the product of force and displacement, i.e., work.
If the force is applied to a particular point of the rigid body, the, as yet unknown,
internal force f⊥ (x) balance will make the object move through the parallel, or sum
force, fk , coherently; as a whole.
In the case of elasticity, the forces should be separated into two parts: the
forces that add up to the force causing the global motion, and the remainder. The
global forces should not cause any deformation hence they should be coherent with
the mass distribution in the system:
ρ(x)
(7.6) f (x) = fk (x) + f⊥ (x) = F + f⊥ (x) .
M
The global force F is defined through the local mass and force distributions. The
perpendicular part of the distributed force f⊥ is the remainder of the force, associ-
ated with internal force balance.
For the parallel forces, Newton law can be brought inside the integral, to yield
an exact local law by definition, for an object without internal stresses. The force
yields the corresponding change in momentum, for the collocated inertia ρ(x):
(7.7) fk (x) = ρ(x)A = Ṗ ,
while the remaining, perpendicular force does not cause any motion of the center-
of-mass, but only deformation. This result is most easily understood if the mass
integral is considered a mass-weighted inner product:
Z
(7.8) hf (x), ρ(x)g(x)i = f (x)ρ(x)g(x)dx .
Ω
The center-of-mass X and the parallel force F are both projections on the constant
and orthogonal unit vector fields e(x) = (ex , ey , ez ):
(7.9) X = he, ρxi ,
and:
(7.10) F = hf , ei .
Hence the perpendicular force is:
ρ(x)
(7.11) f⊥ (x) = f (x) − hf , ei ,
M
which inner-product hf , ei is not the standard self-dual inner-product; it represents
global work or energy and therefore is a product of force and displacement.
The associated decomposition of the displacement x(x) is
1
(7.12) x⊥ (x) = x(x) − Xe(x) = x(x) − he, ρxie(x) ,
M
such that:
(7.13) hfk , x⊥ i = hf⊥ , xk i = 0 ,
7.3. LOCAL THEORIES 127
where we used that hei , ρej i = M δij .

Hence it is possible to make a separation between internal dynamics and the
rigid dynamics for a local model. The bi-orthogonality is based on the work and
energy relations, since, dE = f dx is the change in energy density due to the applied
force. This work separates into two parts: the rigid-body energy and the internal
energy:
Z
(7.14) dE(x)dx = hf , dxi = hfk , dxk i + hf⊥ , dx⊥ i ,
Ω
since the cross terms vanish. Hence, the internal force f⊥ is perpendicular to the
global translations e associated with xk .
7.3. Local theories

The local, or distributed, theories in terms of partial differential equations,
which would for example yield the specific consequences of elasticity in the exam-
ple above, are more general than only the separation of rigid-body motion and
internal deformation. However, the partial differential equations have one particu-
lar property, which led to the recognition of rigid-body motion as a specific form of
global dynamics. The rigid-body motion has no associated elastic energy, whatever
model one chooses for the elastic, or internal dynamics. The fact that any isolated
system has rigid-body modes is the consequence of the translation invariance of
physical laws.
If we are guided by the local definition of the energy, as defined by the Hamil-
tonian density H, we find that the elastic energy vanished for any translation com-
bination T = (tx , ty , tz ) of the constant displacements ei (x):
(7.15) V (x) − V (x + T e) = 0 ,
where V is the potential part of the Hamiltonian. From the equation of motion it
follows that, without applied forces or position constraints the translations yield
conserved momenta (P = constant):
∂H
(7.16) Ṗ = − =0 .
∂T e
In many cases, such as Dirichlet boundary conditions, the boundary conditions
fix the values of P = 0. In other cases, such as Neumann boundary conditions,
the values of P = constant are only partly fixed. Even more general boundary
conditions exist, expressed in an input function u, such that P(u). The link between
boundary conditions and constants of motion hints at the possibility of separating
the direct consequences of the boundary conditions from the internal dynamics.
The Hamiltonian is positive, so is the elastic energy. By deforming an object
energy is added. Some restrictions apply to the form of the elastic energy. Typically
one can assume it is a function of a quadratic form, and similarly the differential
operators must appear to an even power for positivity. It can be written generically
as a weighted form:
(7.17) V ((Dx)2 , x) ,
where D is the differential, or moment operator. See, for example, Sections 2.8 and
5.3. In the case of elasticity the differential operator D could be the strain tensor:
1
(7.18) ij = Dx = (∂i xj (x) + ∂j xi (x)) .
2
In this case the solutions for which the elastic energy vanishes are besides the
translations e, also the rotations around an arbitrary axis n:
(7.19) ∂i (n × x)j + ∂j (n × x)i = 0 .
Hence not just the translations but also the rotations are undetermined in the
isolated system. The kernel of the differential operator D plays the central role.
By introducing the moment variable:
(7.20) q 0 = Dx ,
the potential energy of the system has a unique minimum q 0 = 0. The indeterminate
part, which is either fixed by the boundary conditions, or lead to a conserved
quantity, is now explicit in the equations of motion, and does no longer depend on
the precise details of the potential, now a function of q 0 .
In the theory of hyperbolic partial differential equations the existence and
uniqueness of a solution for given initial and boundary conditions is sought [Sakamoto,
1982] The proper initial conditions are determined by the Cauchy problem, on an
infinite domain or by local analysis. The boundary conditions might be inconsistent
with the initial condition. One way to avoid this situation is to choose the station-
ary, or pseudo-stationary solution for the boundary conditions as initial conditions.
A pseudo-stationary solution would, for example, be an accelerating solution for a
constant force. Pseudo-stationary solutions only have polynomial time-dependence
for the variables of the state for constant boundary values.
The stationary problem for the boundary conditions is called the associated
problem. The boundary conditions of the stationary problem are such that the
state is stationary. The solutions of the associated problem serve as source to the
internal dynamics. Since the solution to the associated problem already satisfies
the boundary conditions, the internal dynamics consists of modes with vanishing
boundary values. The associated problem, however, is not unique, but may vary
with the type of boundary conditions, but as we have seen for the rigid body, a
canonical pair of boundary conditions follows from one associated problem, namely,
the deformation due to the acceleration of the mass distribution. These pairs are
called the port variables.
For example, consider the Hamiltonian of an anisotropic wave equation in one
dimension, with a space dependent mass density ρ and stiffness κ:
p2 κ(x)q 02
(7.21) H0 = + ,
2ρ(x) 2
where p = ρ(x)q̇(x) is the canonical momentum, from Eq. 1.2, and q 0 = ∂x q(x) is
the moment variable. If a constant force is applied, the object will accelerate. The
internal deformation state, however, should be constant in time. The input force
is used to generate stationary equations. In these equations the momentum grows
linearly, i.e., the solution is polynomial in time. In order to arrive at an oscillatory
state, typical for the Hamilton equations of motion, the deformation itself must to
give rise to a change of momentum. This would be internal dynamics. States in
which internal dynamics is absent are called pseudo-stationary states. They follow
from the choice of input and the equations of motion.
The differential operator is D = −D ∗ = ∂x . Only constant functions vanish
for this differential operator. Hence the stationary state, corresponding to constant
7.3. LOCAL THEORIES 129
input, solution is:

F
(7.22) v= ,
u
which correspond to a constant moving and compressed, or stressed, state with a
velocity u and an internal stress F . These stationary states are, almost trivial,
solutions to the algebraic equations:
0
κ(x)−1 F

q (x)
(7.23) = .
p(x) ρ(x)u
The definition of the moment q 0 = Dy yields the displacement y:
Z x
1
(7.24) y(x) = y(0) + F dx0 0)
,
0 κ(x
where the integration constant y(0) = y0 + ut, since no restriction is placed on
the time-dependence or initial velocity. Hence, a stationary state can be defined
through the force F , the velocity u, and the initial displacement y0 at a point, e.g.,
x = 0 and t = 0.
The solution of constant acceleration, discussed above, would be a pseudo-
stationary solution:

0 0 ∂x 0 ∂x F (x)
(7.25) = v= .
pt ∂x 0 ∂x 0 u
κ(x)q 0 (x)

0 ∂x
= .
∂x 0 ρ(x)−1 (p(x) + pt (x)t)
The force F (x) on top row on the right-hand side is the input, therefore pt 6= 0. The
second equation must therefore have a zero left-hand side, in order not to generate
internal dynamics. The corresponding velocity u is the second port variable; the
output. Hence, from the top row it follows ρ(x)−1 pt (x) = a and ρ(x)−1 p(x) = u
must both be constant, and therefore:
Z x
(7.26) F (x) = F (0) + a ρ(x0 )dx0 ,
0
and the displacement y(x) is given by the same integral, where F (x) is no longer
constant:
Z x
1 2 F (x0 )
(7.27) y(x, t) = y0 + ut + at + dx0 .
2 0 κ(x0 )
A pseudo-stationary solution has a polynomial time-dependence of the states and,
in the linear case, also of the energy.
In the general case the equations of motion for the moment variables are the
equations of motion of the port-Hamiltonian H 0 : [Van der Schaft and Maschke,
2002]
0
δq H 0 (q 0 , p)

q̇ 0 D
(7.28) = ,
ṗ −D∗ 0 δp H 0 (q 0 , p)
where D∗ is the formal adjoint of the differential operator D, and p is the canonical
momentum for the displacement q(x). The difference between the port-Hamiltonian
and an ordinary Hamiltonian is the use of the moment variable q 0 instead of the
position q as variable. The moment variables describe only the internal dynamics.
For example, the translation q = constant will yield ∂z q = q 0 = 0. In the absence
of kinetic energy the translation is irrelevant, however, this is seldom the case. To
describe the stationary states, the configuration variables therefore may arise again.
The associated stationary problem would be the solution vs to:

0 D
(7.29) vs = 0 .
−D∗ 0
The pseudo-stationary problems vp would be of the type:

qt 0 D
(7.30) = vp ,
pt −D∗ 0
where polynomial solutions qt t and pt t are included, to allow for interlinked equa-
tions, where not the top and the bottom equation are satisfied independently, but
in combination, similar to the Jordan form in Section 3.1. The corresponding com-
ponents q 0 , p, qt , pt , which are constant in time, yield at most linear time-dependent
states, arise from solving the algebraic relations:
δq0 H 0 (q 0 , p, qt t, pt t)

(7.31) vp = ,
δp H 0 (q 0 , p, qt t, pt t)
where qt and pt are zero, for the stationary equation. The so-called zero modes qt
and pt play a special role in the Hamiltonian. The Hamiltonian depends parametric
on the zero modes. They are the direct consequence of boundary conditions, or
constants of motion in the absence of boundary conditions.
The displacement qt t may not contribute to the elastic energy, but it will con-
tribute to the kinetic energy; it is the constant kinetic energy associated with the
overall displacement velocity. Likewise, pt t will change the kinetic energy. The
Hamiltonian is time dependent due to the boundary conditions. If an object accel-
erates due to the applied force, the kinetic energy changes with the square of the
time.
7.4. Boundary conditions

The stationary and pseudo-stationary solutions are not fixed by the equations
of motion. They give rise to unknown parameters, such as the force, velocity, and
acceleration, which can be seen as the equivalents of constants of motion, in the
port-Hamiltonian system. These solutions yield values at the boundary such that
the parameters may be fixed by boundary conditions. The boundary-value problem
reduces to an algebraic problem.
Decomposition could be an approach to the boundary problems. The integrals
of the object domain Ω can be separated in parts Ω1 ∪ Ω2 :
Z Z Z
(7.32) → + .
Ω Ω1 Ω2
The definitions of the force F and the center-of-mass X hold just as well for each
part separately, as for the whole. The linearity of the integral yield the linearity
of the global, or global variables. This should not surprise us, as Newton has
sought extensive properties of mechanical systems. The work, or power, integral
would yield cross-terms between the different domains, related to the power transfer
between the domains:
Z
(7.33) dEdx = hf1 , dx1 iΩ1 + hf2 , dx2 iΩ2 + (f1 , dx2 )Ω1 ∩Ω2 + (f2 , dx1 )Ω2 ∩Ω1 . ,
Ω
7.4. BOUNDARY CONDITIONS 131
1 2
12
Figure 7.1. A larger linear problem, split in two, consists of two
internal parts with matrix representations A1αβ and A2αβ for the
energy consisting only of variables of one domain, and an overlap
term A12
αβ , with variables from both domains.
where the inner-product (·, ·) is defined on the common boundary of Ω1 and Ω2 .

The precise definition of the boundary, or trace, integrals (·, ·) depends on the local
theory. In the absence of spatial derivatives, the boundary terms vanish. They
can be seen as the end terms of partial integration, or the displacement flow along
force lines. However, without a local theory we can attach little meaning to these
boundary terms. Physical theories, however, must satisfy energy conservation and
force balance, which translates in this domain decomposition into a constraint on
the separation of rigid-body and internal dynamics:
(7.34) (fi , dxj ) = (f⊥i , dxj ) ,
as the result of the locality of the mass density. Discretization will generally destroy
this result, as we will see from the example of a string. If a homogeneous string of
length L is discretized in n segments with linear displacement, described by L + 1
end positions we find a stiffness matrix:
 
−1 1 0 ···
n 1 −2 1 0
Z 
(7.35) (∂x ·)2 → K =  0  .

L 1 −2 1 0
..

.. .. ..
. . . .
The mass matrix is the integral over the density of the linear displaced string:
 
2 1 0 ···
Z
nρ 
 1 4 1 0 
(7.36) ρ(·)2 → M =  .

6L  0 1 4 1 0

..

.. .. ..
. . . .
If the mass matrix would have been diagonal, the principles of the continuum
mechanics, for the separation of the rigid-body motion and the internal motion and
the corresponding forces, would have carried over to the discrete system. However,
for the tridiagonal mass matrix this is no longer the case. Moreover, for the time
simulation the mass matrix must be inverted, which yields a completely filled matrix
M −1 . As a consequence, the system can no longer be separated while yielding only
additional local boundary terms. Each node of the system is connected to each
other node, due to the nonlocal mass matrix inverse M −1 .
However, if we consider local, or band-diagonal matrices only, it is possible
to split the system in different manners. The larger matrix is split in two similar
forms, denoted by αβ. The remainder, consisting of the off-diagonal parts and
diagonal end terms are combined into a 2 × 2 overlap matrix. See Figure 7.1 If
these tridiagonal matrices are split into two equivalent matrices, the remainder is
part of the diagonal and the complete off-diagonal part across the split:
 
α β 0
 β α β 0 
 
 0 β α β 0 
(7.37) Aαβ =   ,
 0 β α β 0 

 0 β α β 
0 β α
which we denote by the Cartesian product in the two subspaces and the boundary
matrix:
0
1 2 α β
(7.38) Aαβ = Aαβ ⊗ Aαβ + ,
β α0 12
where typically the end point yields α0 = 21 α. This is a so-called force split, as the
two boundary points are not collocated, but adjacent, and free to move indepen-
dently. Their motion is limited by the kinetic and potential energy given by the
boundary matrices K12 and M12 . A film-layer model to connect two systems would
be an example of such a force split.
The other method of splitting systems to study boundary dynamics would
be the configuration split, which is more common. A configuration split would
correspond to a collocation of the boundary points. Rather than a dynamical
boundary connection, it is a constraint relation, with partly overlapping mass and
stiffness matrices. The boundary matrix A12 should be added one of the subspace
matrices A1αβ or A2αβ .
In the case of higher-order differential equations, such as the Euler-Bernoulli
beam, or higher dimensions, the splitting of systems, and the variety of boundary
terms and effects is even greater. In the continuum theory of elasticity it has led
to a number of variational principles, such as the three-field method and the four-
field method [Washizu, 1968] to retain the freedom to match boundaries in different
ways. Furthermore, the implementation of such splittings will only be feasible for
linear systems, where constant matrices are to be decomposed. The general theory
of decomposition of such systems, to solve iteratively, dates back to the work of
H.A. Schwarz.[Schwarz, 1870] Nowadays domain decomposition is a viable method
to generate static solutions of large FEM problems. [Toselli and Widlund, 2004]
7.5. Flux
Clearly, decomposing the linearized and discretized system is hardly unique, or
ideal. On the other hand, the rigid-body formulation, with the collocation at the
same x of density and displacement is limited. Instead, the boundary integral can
be coupled to the volume integrals, through the divergence theorem:
Z Z
(7.39) ∇ · Jdx = J · nda .
Ω ∂Ω
Therefore, an appropriate formulation should be expressed in terms of fluxes J.
The key flux is the energy flux S, since in the case of zero internal energy and zero
energy flux the solution, or internal state, is the unique static solution. In general,
energy is thus used to determine existence and uniqueness of a solution to a mixed
7.5. FLUX 133
boundary value problem of a hyperbolic equation. The energy flux is derived from
the canonical Hamiltonian H(q, p) of the system, through the differential operator D
encountered before in the definition of the elastic energy. However, the Hamiltonian
no longer needs to be derived from configuration space. The only link with an
underlying continuous reference space, with boundaries, is through the differential
operator D.
The energy of the system is given by the Hamiltonian density function: H(q, p),
which depends on the canonical position and momentum fields q(z) and p(z). The
energy is conserved, hence the Hamiltonian density satisfies the continuity equation,
Eq. 5.39:
(7.40) Ḣ + ∇ · S = 0 ,
where S is the energy flux. [Fetter and Walecka, 1980] The energy flux is the
key guiding principle in the analysis of the interaction between components. The
energy function is a positive operator of the state space, or phase space, and zero
energy corresponds to the lowest, rest, or ground state. Hence stability analysis
and dissipation uses energy and its in- and out-flow.
The energy flux is not uniquely defined. It depends on what quantities are
allowed to flow across boundaries. For example, an ideal balloon might change is
shape but not its content, while a porous paper bag can both change shape and
content. Constraints on the fields might complicate this matter further. The port-
Hamiltonian H 0 defines the energy-flux. The port-Hamiltonian [Van der Schaft and
Maschke, 2002] arises from the Hamiltonian through a variable substitution:
(7.41) H(q, p) = H 0 (Dq, p) ≡ H 0 (q 0 , p) ,
where D is an appropriate differential operator. The Hamilton equations, the equa-
tions of motion, are
0
δq 0 H 0

q̇ 0 D
(7.42) = ,
ṗ −D∗ 0 δp H 0
where D∗ is the formal adjoint of the differential operator D.
The energy flux follows from inserting the equations of motion into the time
derivative of the Hamiltonian density:
(7.43) Ḣ 0 = δq0 H 0 Dδp H 0 − δp H 0 D∗ δq0 H 0 ≡ −∇ · S .
The expression on the right-hand side is the divergence of the energy flux, which in
many cases can be written as the bilinear product using the boundary operator θ:
(7.44) δq0 H 0 θ(n)δp H 0 = S .
Clearly, appropriate boundary conditions are expressed in terms of δp H 0 and δq0 H 0
rather than the normal state variables q and p.
For the linear wave equation the Hamiltonian density is:
p(z)2 1
(7.45) H= + κ(z)(∇q(z))2 .
2ρ(z) 2
The associated differential operator is the gradient D = ∇, the adjoint is the
divergence D∗ = −∇·. The directional, or moment, variable is q 0 = ∇q. Therefore
the boundary operator θ is simply the surface normal:
(7.46) θ=n ,
and the energy flux follows from the Green identity:

p
(7.47) S = n · κ∇q .
ρ
More general differential operators appear if there are more that one type of energy
flux. For example, take the Euler-Bernoulli beam. The elastic energy is given by
the second derivative of the deflection y(z):
1
(7.48) Helastic = EI(∂z2 y(z))2 ,
2
which may seem like a single bending energy, however, it consists of two parts
in terms of boundary conditions. The boundary operator indicates, however, it
consists of two parts. The associated differential operator would be D = ∂z2 such
that q 0 = ∂z2 y, however, the boundary operator θ is:
(7.49) f θg = g∂z f − f ∂z g .
These two terms correspond to the two boundary conditions, which need to be set
for an Euler-Bernoulli beam, at each end, which lead to four possible solutions for
S = 0, three of those are familiar boundary conditions: free (δq0 H = ∂z δq0 H 0 = 0),
clamped (δp H 0 = ∂z δp H 0 = 0), and supported (δp H 0 = δq0 H 0 = 0), boundary
conditions, where in all cases the initial condition q = 0, at the boundaries, is
used. The fourth boundary condition corresponds to a fixed orientation, but a free
displacement (∂z δp H 0 = ∂z δq0 H 0 = 0), which is difficult to implement physically.
One could think of a clamped beam end moving freely along a rail.
The example of the Euler-Bernoulli beam showed that the physical boundary
conditions S = 0 are rather unlike the typical mathematical boundary conditions in
terms of the state variables q 0 and p at the boundary. They are, of course, related
through the constitutive relations, expressed in term of variational derivatives of
the port-Hamiltonian. In the case of isolated problems, associated with S = 0, such
distinction is of little relevance, especially in the linear case. The nonlinear, non-
isolated case S 6= 0 the distinction is very important. Not all boundary conditions
can be implemented straightforwardly.
The pseudo-stationary states, corresponding to the force F and the velocity u,
are, at one hand, related to physical boundary conditions, and, at the other hand,
to states q 0 and p. They form a bridge between the global forces and velocities and
the internal states. The model with only the degrees of freedom, such determined
uniquely by the boundary conditions, we call the core model.
7.6. Core model

The stationary states are not necessarily static. For example, a system moving
frictionless with constant velocity u is in a stationary state. No energy is transferred
in or out of the system, but can be transferred through the system as we will see in
the example below. The pseudo-stationary states have a linear time dependence,
which may correspond to a complicated energy dependence, if the Hamiltonian is
nonlinear. The equations for the quasi-stationary states do not depend on time.
The left-hand side, the time derivatives q̇ and ṗ are replaced by constant states q t
and pt of linear time-dependent solutions: q̇t (z)t = qt (z) and ṗt (z)t = pt (z). The
energy transfer, or power, can be calculated from the equations of motion:

qt 0 D F
(7.50) = ,
pt −D∗ 0 u
7.7. INTERNAL DYNAMICS 135
which yields a much simpler flux and boundary operator than in terms of the states:
(7.51) ∇ · S = uDF − FD ∗ u .
In the case of the differential operator D being the gradient D = ∇, or the diver-
gence D = ∇·, the boundary operators are:

(n · u)F, D = ∇·
(7.52) θ= .
(n · F)u, D = ∇
In the port-Hamiltonian the canonical variables q 0 and p are geometrically distinct.
As a consequence, depending on the choice of differential operator D, the one of
two port variables: u or F, is the boundary-extensive variable; the flux, while the
other variable is the boundary-intensive variable; the potential.
In many cases, the port flux is indeed a conserved flux, besides the energy
flux. For example, matter flow and charge flow are conserved fluxes, and so is
momentum flow in its peculiar way. It might therefore be important to formulate
the local theory such that the port flux corresponds to such conserved flux. In
many cases the procedure of recovering a Hamiltonian is therefore reversed. The
continuity equation is the first equation of motion, the second equation, the closure
relation is of less importance, and contains the material properties. Afterwards,
the closure relation is integrated to recover the Hamiltonian, or any other total
differential. For example, diffusion equations in physical chemistry arise from such
approach. Mass is almost mysteriously conserved for any constitutive relation,
relating density to pressure, due to the adjoint pair of the gradient operator and
the divergence operator.
7.7. Internal dynamics

The core model, rather than the boundary conditions, are the basis of internal
dynamics. Since the boundary conditions are satisfied by the states of the core
model, the internal dynamics has vanishing boundary conditions. Therefore, the
internal dynamics has gained some of independence of the core model state. In some
cases, therefore, the internal dynamics may be linearized, while the core model may
not.
Inserting the full model back into the equations of motion, taking into account
the time-dependence of the core states q0 = (q 0 +qt t), p0 = (p+pt t) which depend of
the port-variables F(t) and u(t), the internal dynamics appears on the right-hand
side of the equation:
δq 0 H 0 − F

∂F q0 Ḟ + ∂u q0 u̇ 0 D
(7.53) = .
∂F p0 Ḟ + ∂u p0 u̇ −D∗ 0 δp H − u
Hence, if the input is expanded in terms of polynomials in time, like in a spline
approximation, the internal dynamics yield an analogous expansion.
For example, a homogeneous truss, or beam segment, of length L, with elas-
ticity κ and mass density ρ, with the velocity u given as input, and the relative
distance determined by the external force F , yields a stationary state with a homo-
geneous compression q 0 = F/κ, and a homogeneous momentum p = ρu. Since the
stationary state is determined by two quantities F and u, the forces at the ends
are equal but opposite F1 = −F2 , and the velocities are equal u1 = u2 . In order to
affect a dynamical model in terms of the four port-variables F1 , F2 , u1 , and u2 , we
investigate the leading dynamical modes.
The total energy in the truss is:
Lρ 2 L
(7.54) E= (u + u1 u2 + u22 ) + (F1 − F2 )2 ,
6 1 8κ
which is similar to the linear finite-element approximation of a truss, expressed in

the applied boundary conditions. The power transfer through the truss is F1 u1 ,
into the truss at end point u1 and out again at u2 , with u1 = u2 and F2 = −F1 .
The four variables with the two constraints can be seen as the core model, fully
determined by the boundary conditions. The micro-balance determines the force
constraint, the power balance determines the velocity constraint. However, more
appropriately, the model can be extended to include dynamical relations, rather
than constraints between the end points, or port, variables (u1 , F1 ) and (u2 , F2 ).
The energy of the truss, in combination with the constraints, shows that any
dynamical model is incomplete. By a change of force or velocity, the power transfer
is nonzero, while for the core model, due to the constraints, the power transfer is
always zero. The model can be extended, by adding an approximation to the mean,
rigid-body inertial force: F1 +F2 to the overall, or rigid-body, motion approximated
by: u1 + u2 , and for the relative motion u1 − u2 an internal displacement. Instead
we proceed, more systematically, by investigating the dynamical effects of time
variation of the input.
If we now have a time-varying force difference F1 − F2 in the truss, the cor-
responding state q 0 will vary, yielding inertial effects. The state will vary with
time:
!
0 F1 (t)−F2 (t)
q 2κ
(7.55) = Ḟ1 (t)−Ḟ2 (t) x ,
p ρ
2 (u1 + u2 ) + ρ κ L
where x = [−L/2, L/2], the reference coordinate along the truss, which is the
solution to the differential equation:
p ∂ F1 (t) − F2 (t)
(7.56) ∂x = q̇ 0 = .
ρ ∂t 2κ
The time-variation of applied force will cause therefore variation in the relative
end-point velocity u1 − u2 .
Only at the next order, F̈1 − F̈2 the effect of the inertia will cause a change in the
forces at the end points. The boundary flux F u based analysis yield a polynomial
expansion in modes, due to the homogeneous and isotropic nature, rather than a
trigonometric expansion typical for Fi = 0 or ui = 0 boundary conditions:
2x
(7.57) q 0 (x) = d0 + d1 ,
L
where the constant deformation is d0 and space-dependent deformation is d1 . The

corresponding the canonical momenta are
2x
(7.58) p(x) = b0 + b1 .
L
7.8. PSEUDO-SYMPLECTIC OPERATORS 137
These four amplitudes d0 , d1 , b0 , and b1 are directly related to the four port vari-
ables:
   F1 −F2 
d0 2κ
 d1 F1 +F2
2κ
  
(7.59) 
 b0
=
  ρ(u1 +u2 )
 .

2
b1 ρ(u2 −u1 )
2
The energy due to the port-Hamiltonian, restricted to the space of these two modes
is given by (D = ∂x ):
Z L/2
Lb20 Lb2 κLd20 κLd21
(7.60) E= dxH 0 = + 1+ + .
−L/2 2ρ 6ρ 2 6
The equations of motion of the stationary system are:
 ˙   2b1 
d0 ρL
 d˙1   0 
(7.61)  ḃ0  =  − 2κd1  ,
   
L
ḃ1 0
which, once expressed in terms of F1 , F2 , u1 , and u2 , yield two conditions among
the four port variables, and two boundary conditions, either for F1 and u2 , or for
F2 and u1 . The other boundary conditions, like for F1 and F2 , would arise from
the pseudo-stationary state, with, in this case, pt 6= 0.
7.8. Pseudo-symplectic operators

From a formal perspective, the conserved energy of a Hamiltonian is not so
much the consequence of the Hamiltonian itself, but of the symplectic matrix J in
front:
(7.62) ẋ = J∇x H ,
where J T = −J, which in the canonical form reduces to J = (−I0 0I ). In the
port-Hamiltonian framework the symplectic form is replaced by an arbitrary skew-
adjoint differential operator, while the model reduction, based on the core model,
changes the differential operator D in shift operators, as shown in Section 2.11,
Sσ (x0 , x1 , x2 , · · · ) = (xσ , xσ+1 , xσ+2 , · · · ) acting on vectors (x0 , x1 , x2 , · · · ) in half-
infinite spaces:

0 I 0 D 0 Sσ
(7.63) J= → → .
−I 0 −D∗ 0 −Sσ0 0
The first step is the change of variable from q to q 0 = Dq. The second step is finding
the modes for which the differential operators D and D ∗ reduce to shift operators Sσ
and Sσ0 . The modes, with amplitudes (x0 , x1 , x2 , · · · ), are successively constructed
from the core model states.
Depending on the choice of boundary conditions, the values of σ and σ 0 change.
In the example of the wave equation above, Section 7.3, based on the stationary
state, σ = σ 0 = 1. The sum of the two null spaces σ +σ 0 is the dimension of the core
model; the states not determined by the equations of motion, but either constants
of motion, or determined by the port variables.
In the case of σ = 2 and σ 0 = 0 a constant nonzero qt exists. Likewise, in the

case of σ = 0 and σ 0 = 2 a constant nonzero pt exists. The first case corresponds
to u1 and u2 given, the second case corresponds to F1 and F2 given.
For the homogeneous string they are linear approximations of the states which
satisfy the particular boundary conditions. The index σ, σ 0 yields a classification of
boundary conditions.
In the case of the differential operator D and its adjoint D ∗ the signatures
are the same σ = σ 0 . Quite often this is not the case. See also Section 5.3. In
some cases the boundary conditions dictate a static core model, corresponding to
a simple spring model. The Euler-Bernoulli beam is a case where the dynamical
modes depend on the choice of boundary conditions. If force or torque is applied,
more dynamical degrees of freedom arise, than in the case of position and angle
boundary conditions.
The Euler-Bernoulli beam model, Eq. 7.48, will be discussed in detail in Sec-
tions 8.2 and 11.1. A segment of the Euler-Bernoulli beam can be modeled with a
third-order polynomial:
1 1
(7.64) y(x) = a0 + a1 x + a2 x2 + a3 x3 .
2 6
Acting on the coefficient vector a the differential operators D are represented by

shift operators:
(7.65)      
0 1 0 0 0 0 1 0 0 0 0 1
 0 0 1 0  0 0 0 1 0 0 0 0 
∂x2 =  3
  
∂x =  0 0 0 1 ,

 0 0 0 0  , ∂x =  0
   .
0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0
In the symmetric case with D = D ∗ = ∂x2 , the signature is σ = σ 0 = 2. This would

for example be the case of a beam position and direction controlled at one end, say
x = 0, with a0 and a1 given by the control, and a mass at the free end x = 1, such
that the forces and torque at x = 1 are associated with acceleration of the mass.
The static case usually studied for the Euler-Bernoulli beam, however, is asso-
ciated with the signature σ = 4 and σ 0 = 0. The positions and angles are given at
both end, and no force comes into play.
The case of σ = 0 and σ 0 = 4 is associated with forces and torques at both end.
This could correspond, for example, to a free-falling beam, with masses or springs
attached to both ends.
The typical study of the Euler-Bernoulli beam [Beitz et al., 1994] is the loaded
beam, where vertical forces are applied to the horizontal beam, clamped or sup-
ported. As a simple example, we consider a beam x ∈ [0, 2] with a vertical force F
applied in the middle x = 1. The situation is symmetric around x = 1, so we only
consider one half: x ∈ [0, 1]. The energy is given by:
1
1 2 1 2 2
Z
(7.66) E= dx ẏ + (∂x y) .
0 2 2
7.8. PSEUDO-SYMPLECTIC OPERATORS 139
F β F
x=0 x=2 x=0 x=2

x=1 x=1
F F
α α β
Figure 7.2. The two situations under consideration. Left the

clamped boundary conditions, right the supported boundary con-
ditions. A force F is applied in the middle. Below the correspond-
ing ideal models, with one and two degrees of freedom.
The mass matrix in terms of the coefficient vector a is determined by the integrals
over the product of modes:
(7.67)
1 1 1
  
1 2 6 24 1 −7.5 3 −52.5
 1 1 1 1 
30  M−1 = 16  −7.5 75 −337.5 630 

M= 2 3 8  .
 1 1 1 1   30 −337.5 1620 −3150 
6 8 20 72
1 1 1 1
24 30 72 252
−52.5 630 −3150 6300
The Hamiltonian for the finite approximation on [0, 1] is:

1 T −1
(7.68) H= π̇ M π̇a + 2a22 + 12a2 a3 + 18a23 ,
2 a
where πa = Mȧ is the momentum associated with the coefficient vector a. Two or
three degrees of freedom are given by position constraints, depending on supported
or clamped boundary conditions; one or two at x = 0, and one ∂x y = 0 at x = 1,
due to the symmetry. See Figure 7.2.
The absence of a0 and a1 in the elastic energy means the elastic energy has a
two-dimensional null-space. If the values of a0 and a1 are fixed by boundary con-
ditions, it means that their time-derivatives ȧ0 and ȧ1 are also fixed, as secondary
constraints, [Dirac, 1964] which has consequences for the momentum πa . These are
the four boundary conditions which make up σ + σ 0 .
In the case of clamped boundary conditions a0 = y0 and a1 = y00 there is
only a single mode φclamped with coefficient α, with parametric dependence on the
boundary conditions:
(7.69) φclamped(x) = y0 + y00 x(1 − x)2 + (3x2 − 2x3 )α .
The corresponding port-Hamiltonian, with D = ∂x3 , yield equations of motion with
a pseudo-symplectic operator of rank 4 and signature σ = 3 and σ 0 = 1:
0 ∂x3 δa H 0

ȧ
(7.70) = .
π̇a ∂x 0 δ πa H 0
The bottom row is the force equation, the top row is like a constraint equation.
Only the dynamics of a single degree of freedom is determined by the top equation,
due to the differential operator ∂x3 , with a null-space of dimension 3.
In the case of the supported beam, the boundary condition y00 is absent, and a
second mode with coefficient β arises:
(7.71) φsupported (x) = y0 + x(1 − x)2 β + (3x2 − 2x3 )α .
See Figure 7.2. We will not pursue the mapping between the coefficient vector
a and the reduced models variables α and β. In Sections 8.2 and 11.1 the prob-
lem is further investigated, especially concerned the matching of variables at the
boundaries, and the resulting constraints.
7.9. Conclusion
In this chapter we bring out different aspects of global versus local modelling.
Consistent, general, and versatile links between the two approaches are the fluxes,
which link micro-balance laws with their global equivalences. Among these fluxes,
the energy flux, or power, plays a central role, as the equations of motion are
determined by the energy density in canonical form; the Hamiltonian. The internal
dynamics, such as vibrations, are no longer a direct consequence of the interaction
with the surrounding, but due to the coupling between the core, or lumped, model,
and the remainder of the infinite number of degrees of freedom presence in the local
model.
Different physical problems and geometries are investigated in later chapters.
In some cases the core model might have a nontrivial solution, which has it conse-
quences for the internal dynamics. Furthermore, different control problems follow
from the core model. For example, for arbitrary dynamics, the pseudo-stationary
problem might be a better candidate of an optimal control problem, than the cor-
responding full dynamical problem.
CHAPTER 8
Linear finite element methods
In this chapter we will discuss the typical linear Finite Element Methods
(FEM), however, from the perspective of physical modelling, and linking the results
with typical lumped models. The success of FEM lies mainly in the efficiency that
a numerical results can be obtained for a large to very large system. Nowadays, a
typical desktop can treat a linear system with tens of thousands of variables eas-
ily. The typical simulation packages for lumped systems, which are generally much
more versatile and interactive, can handle at most a couple of hundreds of degrees
of freedom.
Hence it is not just extending the number of degrees of freedom in lumped
approaches, which would make them design and simulation tools for FEM. It is
necessary to devise computational methods more closely related to FEM, but with
the modularity and versatility of simulation software.
The groundwork to a marriage of FEM and dynamical modelling is the iden-
tification of typical concepts and principles of dynamical systems in FEM. For
example, the typical stiffness or conductance matrix of modular FEM is singular.
The null-space is associated with rigid-body modes and free flow. In a FEM ap-
proach the boundary conditions are set to eliminate the null-space from the solution
space. In modelling, however, the null-space plays an essential role in determining
how elements can be joined together in a complex system.
8.1. Introduction
The relation between the mathematical theory and the numerical solutions of
partial differential equations always has been a difficult one. Already is 1950, C.
Tompkins, in a study for the Office of Naval Research, mentions
the electronic digital calculating machines they [the engineers]
were developing would replace all mathematicians and, indeed,
all thinkers except the engineers building more machines.... my
task was frequently one of pointing out that the machines will
not run by themselves.
The mathematical “background”, produced by D. Bernstein [Bernstein, 1950] at the
order of C. Tompkins would, however, send most engineers running, and running
codes just hoping for the best. The mathematical theory of existence and unique-
ness of solutions of partial differential equations is complicated at best. Many a
mathematician has steered clear from partial differential equations. The treatment
of partial differential equations as operators on some normed space has been the
most prevailing and successful approach to uniqueness and existence. The bounds
on errors are, however, not always that useful, and certainly yield little feedback
on how to improve upon the result.
141
142 8. LINEAR FINITE ELEMENT METHODS
Fifty years later the uneasy distance between mathematics and numerical code
remains, except for the linear theory, which has developed nicely under strict and
easy bounds. Antman, noted for his work on nonlinear elasticity, remarks:
the danger with using sophisticated numerical procedures on
powerful computers is that they produce beautiful figures that
bear but a tenuous relation to the phenomena they purport to
describe, a danger more succinctly stated in Phil Colella’s para-
phrase of Mark Twain: ”There are lies, damned lies, and colored
graphics.”
For example, in fluid mechanics, the Navier-Stokes equation is a notoriously
hard problem. The Clay Institute will award anybody one million dollar who can
prove the existence of regular solutions for given initial and boundary conditions.
Despite this open problem millions of fluid dynamical problems are studied by the
means of FEM, and the engineers trust the numerical answers enough to fly to the
moon, with ship designs based on them.
In order to use FEM for other purposes than to generate large, sparse systems
of linear equations, we should keep the underlying continuous physical system in
mind. In FEM, typically only the state is retained, the flow is defined implicitly, by
the equations. For example, energy, and the energy density is useful to retain and
use in simulations. Physicists and mathematicians rely on the energy conservation
to test the correctness or stability of an approximation. In general, retaining the
energy, or any globally conserved quantity, such as charge or mass, is hard, as it
arises often as a global constraint on the systems. Local energy continuity, Eq. 5.39,
does not always come natural in the FEM formulation. A posteriori error checking,
on the generated results, is more easily done. In a later chapter we will turn our
eye to the open playing field of nonlinear FEM. Using an explicit energy-dependent
method in terms of Hamiltonians, and symplectic structures for simulation, we
assure ourselves of the numerical stability and correctness, even in the case of
nonlinearities which cause near singular behavior.
Solutions are bounded by energy. In the absence of energy, the solutions are
trivially zero. And the conservation of energy may cause the energy to localize in
a small region of the system, but even there it is bounded. Especially in FEM the
region in which the energy can focus, or localize, has a minimal size determined by
mesh, which yields an upper bound on the amplitude of the solution. However, if
all the energy focuses in a single element, it may signal the mesh is to coarse for
the problem at hand. Several methods of adaptive mesh refinement are based on
bounds similar to the energy bounds. Energy is like as set of parcels in a logistic
network. Based on inventory, parcels are redistributed. Diffusion equations are
only based on the state of the inventory, or state, while wave equations are based
on the direction the energy parcel arrived from. However, even with an explicit
energy, it remains complex to related the known partial differential equations to
such discrete network models, since the energy is typically a quadratic function of
the state, to guarantee the lower bound on the energy.
8.2. Spline as beam

Splines were originally a tool for the technical draftsman. It was a flexible
metal strip, which allowed the draftsman to draw a smooth curve to a set of points.
8.2. SPLINE AS BEAM 143
shift=1,1 shift=1,2
Figure 8.1. The successive basis functions for splines can be gen-
erated from the basic block function through shift, flip, end integra-
tion. The shift will determine the shape of the spline, such as in the
case of twice integrated block shapes below. The Euler-Bernoulli
beam is a three times integrated shift-and-flip block function. The
shift = 1,1 (first shift 1, then shift 1) generates a sharper peaked
spline than the shift = 1,2 (first shift 1, then shift 2) block function.
A connection which has almost been severed, since for FEM the spline approxi-
mation of a beam is considered rather cumbersome. Only recently, splines have
returned, from the domain of Computer-Aided Design, in the domain of structural
computations. [Hughes et al., 2005] In order to connect different numerical a phys-
ical aspects we discuss the simple example of cubic splines and the Euler-Bernoulli
beam. See also Figure 8.1.
The spline can serve as an introduction into a nontrivial and physical FEM
modelling. In particular we will discuss the effects of inputs, in terms of the applied
forces, and the elastic energy, which originally led to the Euler-Bernoulli beam
model. [Euler, 1744] A beam connected by both ends can be described by four
degrees of freedom, fixed by the position and angle at both end. Or likewise by the
force and torque at both end. Every point xi where a force Fi is applied requires
a cut of the beam in two separate segments. Four additional degrees of freedom
are introduced in the model. Three of those are constraint, or algebraic, relations
fixing the segments together. Only the fourth degree of freedom is the dual variable
associated with the applied force. We will see that this is the displacement y(xi ),
based on the elastic beam energy. The product of force and displacement is the
work applied.
The displacement y(x) as function of the horizontal reference position x in the
static case is the solution of the differential equation:
(8.1) EI∂x4 y(x) = F (x) .
The cubic spline solution y(x) is a piece-wise cubic function xi ≤ x ≤ xi+1 :
(8.2) y(x) = a0 + a1 x + a2 x2 + a3 x3 ,
where at each boundary xi the function and the first two derivatives from each
side matches, but the higher derivatives not necessarily. On each of the segments
(xi , xi+1 ) the spline is a solution of the force free Euler-Bernoulli equation:
(8.3) EI∂x4 y(x) = 0 x 6∈ {xi }ni=0
Only at the boundaries xi forces occur.
The left higher derivative and the right higher derivative yields a discontinuity,
related to a force at xi :
Fi
(8.4) lim ∂x3 y(xi + δ) − ∂x3 y(xi − δ) = .
δ→0 EI
Counting the degrees of freedom, each segment has four coefficients ai , three of
those are determined by the boundary conditions between the segments, the fourth
is the result of the force Fi+1 :
X
(8.5) F (x) = δ(x − xi )Fi .
i
Hence, there is a one-to-one map between the displacements y(xi ) and the
forces Fi , and endpoint conditions. For n segments, there are n − 1 forces Fi and
displacements y(xi ) and 4n coefficients aj , with 3(n − 1) relations. Four degrees
of freedom are not determined by either the forces or the displacements at the
intermediate boundaries between the segments. Between two endpoints these four
degrees of freedom are fixed. The boundary, or endpoint, conditions depend on the
choice of spline, and the forces used.
Traditionally, the Euler-Bernoulli beam boundary conditions are fixed by ei-
ther clamped or supported boundary conditions at the endpoints. See Table 6.2.
Clamped means both the displacement y and the derivative ∂x y are fixed, sup-
ported means the displacement and the second derivative of the displacement ∂x2 y
are fixed. Both cases yield a well-defined set of equations for all the coefficients,
using either the intermediate displacements y(xi ) or forces Fi . Hence, the spline
solution y(x) serves well as the exact solution of the approximated problem of a
distributed force F (x):
X X Z xi +x2 i+1
(8.6) F (x) → δ(x − xi )Fi = δ(x − xi ) x +x dx0 F (x0 ) .
i i−1
i i 2
The determination of the solution, even of the linear map between displace-
ments and forces, is a simple but tedious problem, people prefer to leave to the
computer. Each of the forces affects all of the displacements. One can think of a
large force in the middle, causing a deflection of the beam as a whole. One way to
proceed is the integrate the spline from one end to the other, with the unknown
integration constants, to be fixed afterwards by the endpoint boundary conditions.
In another approach, for each Fi the response of the displacements of all y(xi ) are
easily determined, and the total displacement is a linear combination of all dis-
placements, as long as the endpoint condition is zero, and thereby invariant for the
applied forces.
However, there is a different approach to the Euler-Bernoulli spline problem.
The pairs of displacements and forces (y(xi ), Fi ) are the associated with the power
supplied to the system. Each of the segments i has a bending moment, associated
with elastic energy:
Z xi+1
1
(8.7) Ei = EI dx(∂x2 y(x))2 .
2 xi
8.3. FINITE ELEMENT METHOD 145
For a finite energy y(x) and ∂x y(x) must be continuous across the boundary. The
static solution sought; the solution to the Euler-Bernoulli equation, is the solution
which minimizes the bending energy, for the given displacements of forces:
X
(8.8) min E ⇔ δE = δ Ei = 0 ,
which means the first variation of the energy, with respect to the independent
variables vanishes. Hence, if the forces are given, the solution is minimized for the
displacements:
xi+1
X Z xi+1 X X Z xi+1
2 2 2
(8.9) dx(∂x y(x)) = ∂x y∂x y − dx∂x y(x)∂x3 y(x) .
i xi i

ixi
xi
The first term on the right-hand side vanishes of each segment boundary, since the
spline is defined to have vanishing first and second derivatives across the boundaries.
The second term is easily integrated:
xi+1
X Z xi+1 X X Z xi+1
3 3
(8.10) − dx∂x y(x)∂x y(x) = − y∂x y + dxy(x)∂x4 y(x) .
i xi i

i xi
xi
The last term vanishes since ∂x4 y(x)

= 0 on each of the segments. It is not com-
pletely trivial. If the sum and the integral would have been interchanged, the
boundary effects would have been part of the last term.
Hence the elastic energy can be expressed as:
X1
(8.11) E =− y(xi )Fi ,
i
2
since the jump of ∂x3 y(x) across the segment boundary is the force, and the dis-
placement is continuous across the boundary.
For a given force Fi the displacements y(xi ) are minimized, and for a given
displacement the forces are minimized. Without any constraint, the minimization
problem would yield a zero solution. With the constraint, the minimization yield
the lowest attainable values for both displacement and force.
Later, when we discuss boundary polynomials, we will deal with numerical
implementations of spline-like approximations of continuous systems. Solving the
stationary situation can be achieved through several means.
8.3. Finite element method

Finite element method is a name for a collection of finite discretizations of
continuous systems. Some make the distinction like finite difference and finite
volume, and boundary element methods, but that is more a matter of interpretation
of the approximation, than fundamental differences.
The simplest interpretation consists of dividing the space in little cells and use
a very simple, lumped or discrete, submodel for every cell, such that the total model
is a lumped model of cells. The common picture is the approximation of a string
by a series of masses and springs, which is a unfortunate analogy, as we will see.
A series of masses and spring will always have a mass or a spring at the end,
suggesting some breaking of the duality between elastic and kinetic energy, inher-
ent in the infinite dimensional string model, which yields the wave equation. Much
more appropriate would be to turn the string into cells with piece-wise linear co-
ordinate dependence. Given the reference, or rest, coordinate of the string s, the
displacement x(s) is a function of that coordinate s. The appropriate finite element

approximation would arise from picking abscissas {si }ni=0 , and assigning finite el-
ement variables xi = x(si ) to these points, with linear dependence between the
points:
s − si+1 s − si
(8.12) x(s) = xi + xi+1 ,
si − si+1 si+1 − si
which is the function x(s) for cell i : s ∈ [si , si+1 ]. Whatever infinite dimensional
model we had, it simply translates into a finite, discrete model in terms of the
variables {xi }ni=0 . The kinetic energy of cell i is nothing but the integral over the
distributed kinetic energy:
1 si+1 ρ
Z
(8.13) Ekinetic = dsρẋ(s)2 = (si+1 − si )(ẋ2i + ẋi ẋi+1 + ẋ2i+1 ) ,
2 si 6
rather than a local mass, which would corresponds to a local model. Similar, the
potential, or elastic energy of cell i of a string would be the integral over the
distributed elastic energy, limited to the linear function of xi and xi+1 :
1 si+1 κ (xi+1 − xi )2
Z
(8.14) Eelastic = dsκ(∂s x(s))2 = ,
2 si 2 si+1 − si
where κ is the elasticity modulus.
The total co-energy is the system can be written as:
ρa T ρa T κ
(8.15) E ∗ = Ekinetic + Eelastic = ẋ ẋ − ẋ Kẋ + xT Kx .
2 12 2a
The kinetic term ẋT Kẋ is associated with the internal vibration of a segment, with
no associated center-of-mass motion. The mass matrix is non-diagonal: M = ρa(1−
1
6 K), which will lead to problem in time-integration, since time-integration requires
the inverse of the mass matrix, both in Lagrangian and Hamiltonian settings. See
Chapter 3.
In terms of matrices, only a single form arises, as here for five segments, of
length si+1 − si = a:
 
1 −1 0 0 0 0
 −1 2 −1 0 0 0 
 
 0 −1 2 −1 0 0 
(8.16) K=  .
 0
 0 −1 2 −1 0  
 0 0 0 −1 2 −1 
0 0 0 0 −1 1
The problems with the mass matrix are usually avoided by taking only the diag-
onal part of the mass matrix, ignoring the internal motion given by K. In the
approximation of localized masses at the nodes si , the energy would have been:
ρa T κ
(8.17) Emass−spring = ẋ ẋ + xT Kx .
2 2a
For the eigenvalues λi of the matrix K, the corresponding eigenfrequencies are:
6κλ
(8.18) ω2 = .
a2 ρ(6 + λ)
8.3. FINITE ELEMENT METHOD 147
For small frequencies, the result reduces to the typical stiffness over mass relation,
equivalent to the mass-spring result:
s
κλ
(8.19) ωsmall = .
a2 ρ
However, for large frequencies, each of the segments has a maximal frequency,
related to the internal mode of a single segment:
r
6κ
(8.20) ωmax = .
a2 ρ
If the wave length of the mode is shorter than the segment length a, the result is
no longer reliable in this finite approximation.
Interestingly, by examination of the matrix K it follows that the null-space
displacement x with Kx = 0 corresponds to the rigid-body motion. For an equal
displacement of all the modes xi = constant the elastic energy is zero:
 
1
 1 
(8.21) K .  = 0 .
 
 .. 
1
Furthermore, except for the endpoints, the linear function xi = si /a also yields a
zero result:
 
  −1
0
 1   0 
 
   0 
(8.22) K 2  =  .  .
   
 ..   .. 
 .   
 0 
n
1
These results follow easily from K being the discrete approximation of the second-
order differential operator, corresponding to the piece-wise linear approximation:
1
(8.23) ∂s2 → K .
a
Hence ∂s2 1 = 0 and ∂s2 s = 0. The boundary effects arise from the initial and final
node, x0 and xn , being part of only one segment. Boundary conditions should
be implemented to yield the appropriate solution ∂s2 s = 0. In practice, a linear
displacement function may arise as stationary solution, if a force is applied to the
boundaries.
For the string, this model using the piece-wise linear approximation is sufficient
to first order. Later, more complicated finite-order models will be discussed. For
example, for the beam, such a simple piece-wise linear model is insufficient. The
elastic energy in this approximation is either infinite or zero depending on the
perspective. On each of the segments:
1 si+1
Z
(8.24) Ebending = dsEI(∂s2 x(s))2 = 0 .
2 si
However, across the segment:

si +δ
1
Z
(8.25) Eboundary = lim dsEI(∂s2 x(s))2 = lim ∞ = ∞ ,
δ→0 2 si −δ δ→0
which shows two important aspects of any finite-order approximation of a con-

tinuous model. First, the approximation functions should be appropriate for the
differential operator. Second, the boundaries can contribute to the energy, depend-
ing on the formulation. In an appropriate function space neither aspect will occur,
however, special care must be taken to formulate the model in such a way that the
boundaries themselves do not contribute to the energy.
8.4. Stencils
Splines are an example of functions defined on a set of nodes, points, or abscis-
sas. Such a set of points are called a stencil. Stencils are used in all kind of discrete
methods, such as finite element methods. [Milne, 1970]
For example, to define a smooth function f (z) through a set of values fi = f (zi ),
one uses at least two points zi ≤ z ≤ zi+1 . The stencil is, in that case, the pair of
points {zi , zi+1 }:
zi+1 − z z − zi
(8.26) f (z) = f (zi ) + f (zi+1 ) .
zi+1 − zi zi+1 − zi
Such an approximation can serve again to an approximation for the derivative:
f (zi+1 ) − f (zi )
(8.27) ∂z f (z) = ,
zi+1 − zi
for zi ≤ z ≤ zi+1 .
The time-integration of an ordinary differential equation can be seen as a sten-
cil. For a n-th order differential equation at least n points {ti−n+1 , ti−n+2 , · · · ti } are
required. However, more advanced methods of time-integration use more nodes. A
smoother function f (t) is the result, which will improve the forward extrapolation.
For spatial functions the similar principles are used for other purposes, such as the
construction of a Laplacian ∆f (zi ) on a grid, using neighbor points.
In the case of hyperbolic differential equations, Godunov theorem states that
any higher-order approximation of the spatial derivatives will lead to oscillatory
results. It is the simulation equivalent of Gibbs phenomenon for Fourier series.
Nonlinear schemes of constructing derivatives have been developed to reduce the
oscillations. [Liu et al., 1994] This can be seen as the nonlinear weighting of several
stencils. Such stencils must be higher order, for example, for a point z : zi ≤ z ≤
zi+1 , the stencils {zi−1 , zi , zi+1 } {zi , zi+1 , zi+2 } can be used.
In the case of a linear approximation, the stencil function fi (z) is the depen-
dence of the interpolation f (z) on discrete value f (zi ):
∂f (z)
(8.28) fi (z) = .
∂f (zi )
The linear stencil function, or first-order spline, would be:
( z−z
zi −zi−1 , zi−1 ≤ z ≤ zi
i−1
(8.29) fi (z) = zi+1 −z .

zi+1 −zi , zi ≤ z ≤ zi+1
The stencil function with a smooth derivative can be multiple, but will always have
a larger stencil. Not only at the internal nodes zi the function must be smooth,
8.4. STENCILS 149
but also at the boundary nodes of the stencil. The cubic spline approximation can
be reformulated as stencil function on a stencil of five nodes:
 2
 − (z−zi−2 ) (zi−1 −z) , zi−2 ≤ z ≤ zi−1

 m−
(zi+1 −z)(z−zi−1 )
(8.30) fi (z) = (zi+1 −zi )(zi −zi−1 ) , zi−1 ≤ z ≤ zi+1 ,
 − (zi+2 −z)2 (z−zi+1 ) , z

≤z≤z

m+ i+1 i+2
where the coefficients m± are:

(zi±1 − zi±2 )2 (zi − zi−1 )(zi+1 − zi )
(8.31) m± = ,
(zi+1 − zi−1 )
such that the derivative at z = zi±1 is continuous. See also Figure 8.1, in Section 8.2
where this type of spline can be generated by a shift = 1, followed by a shift = 2.
Most stencil functions are a partition of unity. In the case of a constant function
sampled by f (zi ) = constant, the sum of all the stencil functions should reduce to
the constant unity function:
X
(8.32) fi (z) = 1 .
i
However, for example, in the case of positivity bounds fi (z) ≥ 0, one could consider
a simple stencil function, with a smooth derivative, which is not a partition of unity:
(z − zi−1 )2 (zi+1 − z)2
(8.33) fi (z) = , zi−1 ≤ z ≤ zi+1 .
(zi − zi−1 )2 (zi+1 − zi )2
However, typically, a maximal deviation of 12% occurs from the partition of unity.
See Figure 11.8.
Higher-order derivatives also require higher order stencils. The simplest ap-
proximation of the second derivative need a three-point stencil:
1 2 (zi+1 − zi )f (zi−1 ) − (zi+1 − zi−1 )f (zi ) + (zi − zi−1 )f (zi+1 )
∂ f (zi ) =
2 z (zi − zi−1 )(zi+1 − zi−1 )(zi+1 − zi )
f (zi − δ) − 2f (zi ) + f (zi + δ)

(8.34) → .
δ2
In higher dimensions the stencil can extend in different directions. The Lapla-
cian is the typical, isotropic differential operator. The simplest approximation in
two dimensions is on a cross stencil {zi−1,j , zi,j−1 , zi,j , zi+1,j , zi,j+1 }, which yields
the sum of the two second-order derivatives. In the case of the homogeneous square
grid the Laplacian is:
f (zi−1,j ) + f (zi,j+1 ) − 4f (zi,j ) + f (zi+1,j ) + f (zi,j+1 )
(8.35) ∆+ f (zi ) = .
δ2
However, one could also use a diagonal stencil {zi−1,j , zi,j−1 , zi,j , zi+1,j , zi,j+1 } on
a square grid:
(8.36)
f (zi−1,j−1 ) + f (zi−1,j+1 ) − 4f (zi,j ) + f (zi+1,j−1 ) + f (zi+1,j+1 )
∆× f (zi ) = ,
2δ 2
√
where 2δ is the diagonal grid size.
Combining the cross and the diagonal stencil in the right manner decrease the
order of the error. The error is typically defined as the order of the analytical
expansion which is no longer correctly represented by the discrete approximation
of the Laplacian.[Milne, 1970] The order-improved Laplace operator is the linear

combination:
1
(8.37) ∆ = (2∆+ + ∆× ) .
3
Besides square grids there are also triangular, honeycomb, and other grids. The
second-order analytical expansion can be used to define a proper approximation of
the Laplacian. In the case of general geometry, (rectangular, triangular, hexagonal,
etc.) the derivative should reproduce the continuous derivative as well as possible.
In order to check this we consider a symmetric system with one element center and
all his neighbors, of the same shape and size. The Laplacian should be:
X f (zj ) − f (zi )
(8.38) ∆f (zi ) = γ ,
kzi − zj k2
j:neighbors of i
where zi is the barycenter of center element i, with value f (zi ). The geometric
constant γ we are after. We consider a continuous function f (z) = 12 kz − zi k2 . The
integral can be calculated in close form:
f (zj ) − f (zi ) 1
Z X
(8.39) ∆f ddim z = Ai γ 2
= A i κi γ ,
cell i kzj − zi k 2
j:neighbors of i
where Ai is the area of element i and κi is the number of neighbors. Clearly the
continuous result should yield the total area times the dimension, since ∆ 12 kz −
zi k2 = dim. Hence we find:
2dim
(8.40) γ= .
κi
For any grid, this discrete approximation of the Laplacian leads to an error of the
order O(δ 3 ) where δ = kzj − zi k.
8.5. Energy stencils

The stencils of the differential operators are related to a particular discretization
of the energy functional, the energy stencil, which is a bilinear functional:
1
(8.41) E[f, g] = hDg, Df i ,
2
where f is the state variable and g is the associated dual, which in most cases
where a Hilbert-space inner product is used hx, yi = hy, xi, is from the same state
space. The energy functional associated with the Laplacian through the variation
principle is:
1
(8.42) δf E = δf h∇f, ∇f i ⇒ −∆f = 0 .
2
The stencil operator ∆+ corresponds to nearest-neighbor terms in the discretized
gradient only:
f (zi±1,j ) − f (zi,j )
(8.43) ∇x,± f = ,
|zi±1,j − zi,j |
where the direction of the difference vector δx,± = zi±1,j − zi,j determines the
direction (x, ±) of the gradient. The cross operator ∆× is related to diagonal
difference vectors:
(8.44) δ±,± = zi±1,j±1 − zi,j .
8.6. BOUNDARY CONDITIONS FOR STENCILS 151
domain point
boundary point
corner point
stencil
Figure 8.2. For each domain point there is a linear relation, defin-
ing the values on the domain points uniquely. Hence the set of
boundary points are such that all the stencils of ∆+ are covered.
In the case of the ∆× stencils four more points at the corners are
required.
The energy stencil function of the second-order improved Laplacian is the weighted
sum of both energies:
1 X 1 X
(8.45) E[f ] = (∇i,± f )2 + (∇±,± f )2 .
3 6
(i,±) (±,±)
The factors are determined from√the square of the relative distance between an edge
length a and a diagonal length 2a, which leads to the same result as Eq. 8.37.
8.6. Boundary conditions for stencils

A typical discretization of a partial differential equation on a grid leads to a
stencil as we have seen in the previous sections. A stencil is a linear relation among
the values on a number of nodes; the discretized version of the partial differential
equation, such as ∆f = 0. Hence, the successive application of stencils makes it
possible to reconstruct the solution in a domain given the values on the boundary.
See Figure 8.2. The stencils already show that the equations cannot be solved
successively from one corner to another. At the boundaries four domain points are
necessary, at the corner three points, intertwining all the equations.
Therefore not only the type of partial differential equation, but also the finite-
order approximation determines the type and number of boundary conditions, for
time-dependent systems as well. There is no real difference between the Laplace
equation and the wave equation in that respect. Different uses of the same stencil
lead to different boundary conditions. In the case that a system can be integrate
forwardly, we can denote a node in the stencil as the dependent variable. Using
the cross stencil ∆+ for the integration of the wave equation, the single node with
a larger t: f (zi , tj+1 ) is the normal dependent variable, expressed in terms of the
other values:
(8.46) f (zi , tj+1 ) = f (zi+1 , tj ) + f (zi−1 , tj ) − f (zi , tj−1 ) ,
which is the discretization of the homogeneous wave equation:
(8.47) (∂t2 − ∂z2 )f (z, t) = 0 .
The stencil is symmetric with respect to space. At each end of the domain a single
boundary condition is specified. Double boundary conditions only at one end of the
domain are also possible. This requires the x-stencil ∆× instead of the cross-stencil
∆+ . See Figure 8.3.
Figure 8.3. The boundary conditions corresponding to the sim-

plest wave stencils.
8.7. Finite volume versus finite difference

At lowest order, piece-wise linear approximations, of continuous systems, there
is no fundamental difference between finite-volume and finite-difference methods.
The equations are the same. At higher orders some differences arise. In general,
one could argue that finite volume methods have to match at common boundaries
between elements, and finite difference methods have overlapping stencils.
Spectral element methods are an extension to the finite volume method where a
set of basis functions on each element are matched at the boundary by neighboring
functions. For example, in fluid dynamics, the velocity fields approximations are
matched at the boundaries, using Gauss-Lobatto quadrature points. These are
associated with orthogonal polynomial solutions with nodes, or abscissas, at the
boundary and in the interior. The abscissas at the boundary are associated with the
weight function, the orthogonal expansion will be Gegenbauer or Jacobi polynomials
of a particular order. For example, on a domain z ∈ [−1, 1], a weight function
(1−x2 )j will introduce j-th order zeros at the boundary, associated with j boundary
values. The result has an error in the order of the square of the polynomial order,
since the inner-products vanish:
Z 1
1 (α) (α) π21−2α Γ(i + 2α)
(8.48) (1 − z 2 )α− 2 Ci (z)Cj (z)dz = ,
−1 i!(i + α)Γ(α)2
Pn (α)
where α = 52 . Any solution f (z) = 2
i=0 ai (1 − z )Ci (z) orthogonal to any
polynomial to the order n will yield a zero approximate integral, if the abscissas z i
(α)
are chosen as the zeros of Cn+1 (zi ) = 0, with the appropriate weights wi :
Z 1 X
(8.49) f (z)g(z)dz = wi f (zi )g(zi ) .
−1
There are the Gauss-Lobatto quadrature abscissas and weights. [Abramowitz and
Stegun, 1965, Hildebrand, 1987]
For example, in the case of fluid dynamics, the velocities have to be matched
at the element boundaries. The pressure, on the other hand is only needed at the
internal points. A different set of abscissas and weights are used for the pressure,
in the conjunction with the weights and abscissas for the velocities. The velocity
and the pressure are the intensive and the extensive variables of fluid dynamics.
The Navier-Stokes equation will tell us as much, under close examination of the
dimensions of the different variables. The energy, or the Hamiltonian, likewise
separates the state variables and the effort variables, or potentials.
The finite volume method has typically been associated [Saad, 1995] with con-
servation laws. The flow out of the volume equals the change of a conserved density
8.8. MODEL REDUCTION 153
such as mass or charge:

Z Z Z
(8.50) Q̇dn z = n · Jdn−1 z = ∇ · Jdn z .
Ω ∂Ω Ω
Hence it would be natural use the conserved density Qi on the finite volumes Ωi
to retain the conservation law exactly. The equations of motion would be the
instantaneous exchange of the conserved density:
(8.51) Q̇i = Jij , Q̇j = Jji ,
where Jij = −Jji . More complex function Qi (z) on the domain Ωi will yield a more
elaborate exchange term:
Z
(8.52) Jij = nij · J(z)dn−1 z .
∂Ωi ∩∂Ωj
The closure relation expresses the flow J(Q) in terms of the obvious state variable
Q(z). In practice, the boundary flow Jij , through the common boundary ∂Ωi ∩∂Ωj ,
depends on the conserved density on both whole domains Qi (z) on Ωi and Qj (z)
on Ωj .
Hence the finite-volume approach for the conserved density yields a finite-
difference approach for the conserved flow. This is a general result. In the case
that the closure relation J(Q) is implicitly used, the finite difference part is di-
rectly expressed in terms of the finite volume terms. The conservation law will still
be exact, but over the quality of the closure-relation approximation there remains
little control.
8.8. Model reduction

For the use of FEM in simulation and control model reduction is more or
less essential. The state vector x is too large to be handled and updated within
milliseconds, typical for interactive simulations and control on a personal computer.
Hence, from the large space x ∈ RN we should select a smaller space c ∈ Rn

of coefficients such that n N :
n
X
(8.53) x(c) = ci (t)xi ,
i=1
represents most of the dynamical behavior. The state vectors xi are the selected,
normalized modes. The time-dependence is given by the coefficient vector c.
In the process of model reduction two aspects play a role. First, the selection
of the modes xi . Secondly, the projection of the dynamics. The simplest approach
would be to project the dynamics on the same set of modes:
(8.54) xTj xi ċ = xTj Axi c + xTj Bu ,
where the matrix Eji = xTj xi is the mass matrix which multiplies the rate-of-change
of the coefficients vectors c, and Aji = xTj Axi is the reduced A matrix.
It is very important to notice that picking a set of modes does not yet introduce
an approximation error. The projection determines the loss of accuracy. The
dynamics of the modes may evolve the system outside the subspace spanned by the
modes:
X n
(8.55) x = x k + x⊥ = c i xi + x ⊥ ,
i=1
where the inner-product hx⊥ , xi i = 0. Hence the perpendicular part determines

the quality of the reduced order model.
Furthermore, the standard inner-product hy, xi = yT x may be highly unsuited
to determine the quality of the model. For example, modelling a heat-conductance
problem, with the temperature as state vector x, the air gaps in a system have a
low heat capacity and conductance. Hence, air gaps have little effect on the overall
temperature dynamics, but are slow to respond to changes. Hence, if the temper-
ature inner-product: hy, xi; the direct product of temperature with temperature
summed of all nodes, serves to determine the quality of the model, the air gap
temperature may play an important role in these inner-products. Hence a more
appropriate inner-product would be the one weighted with the heat capacitance
matrix (y, x) = yT Cx, however, it still has little physical meaning.
8.9. Proper orthogonal decomposition

Proper orthogonal decomposition [Holmes et al., 1996] is a form of model re-
duction which relies on the numerical experiment to select a number of modes.
The large FEM code simulates a long time sequence, with variable input, within
the appropriate range. From the time series a number of high-dimensional states
xi = x(ti ) are selected as modes from the model order reduction.
In order to select an appropriate set {xji }n0 from a larger set {xi }N
0 , one studies
the correlations in the larger set. If two vectors xi and xj are the same or parallel,
only one has to appear in the smaller set. Two vectors are parallel if:
(8.56) (xTi xj )2 = x2i x2j .
Constructing a matrix from the correlations, it corresponds to a vanishing deter-
minant:
x2 xTi xj

(8.57) detC = T i =0 .
xj xi x2j
From these two vectors only one is needed in the reduced set. Hence, selecting
appropriate modes for a proper orthogonal decomposition corresponds to principal
component analysis in statistical analysis. A single representative for vectors which
are nearly parallel is selected, and a number of directions with a large number of
vectors, in each of these directions is selected. This is done by determining the
largest eigenvalues, and the eigenvectors, of the correlation matrix.
In some cases the mean state:
N
1 X
(8.58) x= xi ,
N i=1
is subtracted from each state xi → xi − x, such that rate-of-change snapshots
(x(ti ) − x(ti + δ))/δ can be included among the state snapshots, to improve the
representation of the transient dynamics.
In the simple case that there are n blocks, with each ni identical unit-length
vectors, which are orthogonal to all other vectors, the eigenvalues of the large
8.10. MODEL DIMENSION 155
matrix arise from each block independently. A simple example is a state that
jumps between four different domains of the full state space:
 
0 |||||||| 0 0
 0 0 |||| 0 
(8.59) (x1 , x2 , x3 , · · · , xn ) = 
 ||||||
 ,
0 0 0 
0 0 0 ||
where ||||| represent a set of ni nearly identical column vectors. For a single block
i of size ni the eigenvalues are:
(8.60) λ i1 = n i , λij = 0 (j = 2, · · · , ni ) .
Hence a single dominant eigenvector represents a mean direction of the block, where
the eigenvalue counts the number of states from the large set appearing in the
direction of the eigenvector. In a realistic case the same principle holds. Eigenvalue
analysis selects dominant directions, while the smaller eigenvectors describe the
variations in the dominant directions, or principal components
Three remarks of caution are at place here. The covariance matrix C depends
on the sampling. If a particular state of the time-dependent FEM experiment is
sampled a lot, it will be dominant in the POD, while there might not be a physical
reason to include this state. For example, the initial state might be dominant if
one starts with sampling from the start. Hence often is started with sampling x(ti )
after a finite time. Moreover, time-sampling will be dominated by the stationary
behavior. Fast transient changes, usually of particular interest, might be swamped
by the slow variations, with large numbers of sampling points. This problem may
be cured in part by using difference vectors: x(ti+1 ) − x(ti ), rather than sampling
points x(ti ).
Furthermore, the inner-product used xTi xj in the covariance matrix might be
inappropriate. It might be dominated by the part of the states close to the fixed
boundary conditions, or irrelevant physics. If possible, the energy, or even better
the change of energy; the power, should lie at the basis of the inner-product used
in the covariance matrix C.
8.10. Model dimension

The dimension of a competent reduced-order model depends on the details
of the driven dynamics. For example, in many thermal cases the reduced-order
model which predicts the result with a percentile accuracy may be two or three-
dimensional. On the opposite end of the model spectrum are transmission lines,
where the dimension is given by the length L, signal velocity c and the input
sampling frequency f . The time the signal spent in a transmission line is τ = L/c.
If the sampling frequency is f , the amount of information in the transmission line
is f τ = f L/c, hence the dimension of the state space must be:
fL
(8.61) dim ≈ .
c
The time-dependence of the input signal is directly related to the wave length of
the transmission-line signal. The transmission line is like a queue. The number
of samples in the queue depends on the sampling frequency and the transmission
time.
However, in many cases the dimension may not be a priori clear. A large
FEM model lies at the basis and inputs leads to a certain dynamical range, which
determines the state space dimension, or model order. The Krylov method based on
the input vector B and the inverse of the evolution matrix A may break down due
to the lack of orthogonality between successive modes. The lack of orthogonality is
a physical manifestation of the fact that the dynamics is restricted to a subspace
of the full high-order FEM state space.
The angles between different state vectors can be used to determine the dimen-
sion of the subspace. If among three state vectors the angle of one pair equals the
sum of the other two pairs, the three vectors lie within the same two-dimensional
subspace. The two state vectors with the angle closest to π2 should remain while
the third state vector can be dropped from the reduced-order state space vectors.
8.11. Effective Hamiltonian

A simple way to reduce the number of variables and the dynamics of a FEM
model is the introduction of effective Hamiltonians, also known Guyan projection.
From the complete set of variables x a relevant set of configurations and state
representatives are selected through a projection operator P = PT . For example,
for an elastic structure, these could be the variables at the boundary. The remainder
is 1 − P = Q. The elastic energy of the remainder is minimized for arbitrary values
for Px = xport . Hence for each value of xport the values of xremainder are chosen
such that the energy xT Kx/2 is minimal:
(8.62) xremainder = −(QKQ)−1 QKPxport .
The matrix yielding the remainder from the retained part xport is known as the
Schur complement. [Golub and Van Loan, 1996] The state is described by xport ,
which determine the dimension of the subspace. The state however lies in the full
state space: x = xport + xremainder(xport ).
In terms of matrices we can order the port or boundary configurations such
that the first m rows and columns correspond to the retained variables:

PKP PKQ
(8.63) K= .
QKP QKQ
The elastic energy as function of the port configurations only is still a quadratic
function:
1 T 1
(8.64) Eelastic = x Kxport + xport KQ(QKQ)−1 QKxport .
2 port 2
The corresponding kinetic energy is:
1 T 1
(8.65) Ekinetic = ẋport Mẋport + ẋport KQ(QKQ)−1 M(QKQ)−1 QKẋport .
2 2
This would correspond to the first Krylov mode. The quality can be tested by the
overlap between the kinetic and the elastic mode. If these two modes are nearly
parallel, the main part of the dynamics is captured. Otherwise more modes or a
larger projection P may be used to yield a better approximation.
8.12. PLAQUETTE 157
8.12. Plaquette
The plaquette is the equivalent of the stencil in quantum field theories. It is
used in conjunction with the Maupertius, [Goldstein, 1980] or least action principle,
where the fields are integrated over a space-time domain, treating space and time
on equal footing. This makes the method especially suited for relativistic theories,
but also for time-dependent studies it has the advantage to incorporate the time-
integration into the finite-element method. [Creutz, 1983] The energy continuity
can be retained in the approximation.
The plaquette and link variables define quantum field theories by numerical
means. Most of these approaches are called Euclidean, which means that space
and time are treated on equal footing. Even more, the time axis is rotated in the
complex plane, which is called a Wick rotation, such that a wave equation in three
dimensions yields a Laplacian in four dimensions. Stable solutions of one can be
related to the other.
We will discuss the plaquette variable from the Hamiltonian perspective in a
single space dimension, without the Wick rotation. However, unlike the true Hamil-
tonian approach, time is discretized here as well. The energy is given by the Hamil-
tonian H, together with the corresponding energy flux S it yields a divergence-free
vector field in space-time x = (t, z), the two-dimensional version of the energy-
continuity relation, Eq. 5.39:
(8.66) ∂ t H + ∂z S = 0 ,
where the space-time vector field on (t, z) is given by:
00
H T
(8.67) J= = .
S T 01
In terms of the energy-momentum tensor T , the Hamiltonian and the flux are the
T 00 and the T 01 components respectively, if t = x0 and z = x1 . 1
The divergence theorem relates vector field to its integral over a plaquette
z ∈ [0, 1], t ∈ [0, 1], such that the values at the boundary satisfy the conservation
law:
Z Z Z
(8.68) dx∇ · J = dtS(t, 1) − S(t, 0) + dzH(1, z) − H(0, z) .
The integrals over each of the edges are called link variables. The single plaquette
is made up of four link variables.
Given a linear approximation for J, the four corners determine the complete
field, and given three corners, the fourth corner is determined, similar to the stencils
in Section 8.4. In more general approximations the duality between four corners
and four links still means a single condition arises from the energy conservation.
The vector field J is usually derived from the equations of motion, or their
generating functional, such as the Hamiltonian or the Lagrangian, for example via
the energy-momentum tensor:
δL jk
(8.69) T ij (q) = − g ∂k q + g ij L ,
δ∂i q
1Some details concerning signs and the Minskowski metric are not treated here, however,
different conventions will lead to different signs for both the fluxes and the derivatives. Here the
variables are treated as Euclidean. The sign convention for the fluxes is considered Euclidean as
well, which means the energy flow in space and time have the same sign.
where g ji = (10−10
) is the metric tensor, and L the Lagrangian. The description is
not unique, T 6= T ji . Hence, one energy flux might differ from another by a total
ij
divergence. The single requirement of energy conservation is that it can be written

as a rotation field, i.e., an exact 1-form:

∂z W
(8.70) J= .
−∂t W
However, starting from a general Hamiltonian, the link variables arise naturally
during discretization. The variable q(z, t) appears in the Hamiltonian as q̇ and ∂z q.
These two variables correspond to the link variables on a time-link and space-link
respectively:
Z
(8.71) qz (t) = dz∂z q(z, t) = q(1, t) − q(0, t) .
The links are constrained by the integrability condition: qz (1) − qz (0) + qt (0) −
qt (1) = 0.
Maupertius principle, or the Lagrangian variational principle, is the most ap-
propriate form to use the link variables:
Z 1 Z 1
(8.72) δ dtdzL(∂z q, q̇) → δ dtdzLp (qz (t), qt (z)) = 0 .
0 0
The wave equation, for example, reduces to an algebraic variational equation, with
the Lagrangian:
(8.73) L(qz (t), qt (z)) = qt2 − v −2 qz2 .
Under the assumption that the link variables are linear functions of the remaining
degree of freedom, the quadratic Lagrangian is easily integrated over each plaquette:
Z 1
1
(8.74) dz(qt (0) + (qt (1) − qt (0))z)2 = (qt (0)2 + qt (1)qt (0) + qt (1)2 ) .
0 3
The simulation can be done by filling the squares. The complete system for a line
consists of a chain of plaquettes. A single boundary time-link is required for a
unique solution to the hyperbolic variational problem. The constraints can be used
to re-express all the other time-links in terms of the space-links for t = 0 and t = 1:
(8.75) qt (1) = qz (1) − qz (0) + qt (0) .
The right time-link is determined by the left time-link and the space-links. For the
neighboring plaquette, the right time-link is the already determined left time-link
of the original plaquette. The time-links of the chain are determined recursively.
The variational problem allows one to reduce the degrees of freedom by half:
the space-links for t = 1 can be expressed in terms of the space-links for t = 0,
which is precisely the condition for a unique evolution.
↑
t z→
Clearly, the same procedures can be applied to non-linear systems, the adaption
lies in the single integrals:
Z 1
(8.76) dzT (qt (0) + (qt (1) − qt (0))z)2 = T (qt (0), qt (1)) ,
0
8.13. SIMPLEX 159
and, the potential term:

Z 1
(8.77) dtV(qz (0) + (qz (1) − qz (0))t)2 = V (qz (0), qz (1)) .
0
8.13. Simplex
In many modern FEM applications, the underlying mesh consists of triangles.
This is mainly the field of mathematicians, who appreciate the simplicity of the
simplex, and the field of software engineers, who base their methods on CAD/CAM
methods, which allow for a restricted set of smooth curves, which are more easily
meshed, or discretized, using triangulations than many other methods. There are
several ambitious research programs formulated to make triangular meshes the
industry standard, but, probably due to complications at the lowest levels, it has
never materialized. The complications for general modelling, with a direct access
and control over the mesh and the data, are large and various. Some of the details
of mesh refinements, described below, in three dimensions will make this clear.
In two dimensions, the mesh is clear, and some global properties, such a topo-
logical properties, are easily determined. For example, the Euler characteristic, χ E ,
is an invariant of a closed surface, which is related to the number of genuses, or
holes, or handles on a surface. For example a sphere has no handles, a donut one
handle, and a solid eight has two handles. The Euler characteristic χE is given by
the enumeration of all the elements of a surface:
(8.78) χE = vertices − edges + faces = 2 − handles .
For three-dimensional elements such relations do not exist. A characteristic is

defined:
(8.79) χ3 = vertices − edges + faces − volumes .
But it seems to have no useful topological meaning. One of the problems is the
number of ways to define a mesh, or mesh refinement, for a three-dimensional
simplex, such as the tetrahedron.
Starting from a cube, there is one easy division in identical simplices, based on
the symmetry principle. One can define one simplex by the corners: (0,0,0)-(1,0,0)-
(1,1,0)-(1,1,1) By reflections, all other simplices follow, yielding a filling with 6
simplices. One edge is the body diagonal, two edges are face diagonals, and three
edges are cube edges. [Maubach, 1994]
Given the positions of the four corners of the simplex xa , xb , xc , and xd , the
geometric properties follow simply:
1
(8.80) Areaface:abc =|xb × xc + xa × xb + xc × xa | ,
2
where the direction of the vector is the direction of the surface normal,
(8.81)
Volumesimplex:abcd = | det(xa −xb , xb −xc , xc −xd )| = |(xa −xb )·((xb −xc )×(xc −xd ))| .
The position of the center is:
xa + x b + x c + x d
(8.82) x0 = .
4
Figure 8.4. From left to right, the elongations εi , defined using

two nodes, the shear γxy using three nodes, and the masses cor-
responding the integrals over the patches for center nodes, edge
nodes, and corner nodes.
8.14. Transverse motion of a table

As a practical example for comparison we investigate the longitudinal motion
and vibration of a table. In several industrial setups large tables are used to position
a flat object. Fast motion will cause the table to deform and possibly vibrate.
We assume the table to be made of standard steel, with a Young’s modulus of
E = 20.0 109 Nm−2 , a Poisson ratio of ν = 0.30, and a specific density of ρ =
8.0 103 kgm−3 . The table is taken to be 2. × 2. meters. For relative numbers in the
linear theory, only the Poisson ratio ν is relevant; the scale can be set arbitrarily.
The three-dimensional elastic energy W of the table is expressed [Washizu,
1968, Landau and Lifschitz, 1959] in its elongations εx , εy , and εz , and the shears
γxy , γxz , and γyz :
 T   
ε E E⊥ E⊥ εx
1  x   k
Z
(8.83) W = εy E⊥ Ek E⊥   εy  + .
2 ε E E E ε
z ⊥ ⊥ k z
 T   
γxy G γxy
(8.84)  γxz   G   γxz  dV ,
γyz G γyz
where Ek = (1−ν)E/((1−2ν)(1+ν)), E⊥ = νE/((1−2ν)(1+ν)), and G = E/(2(1+
ν)). The elongations and shears can be expressed in terms of the displacements u,
v, and w:
∂u ∂v ∂u
(8.85) εx = , γxy = + , etc.
∂x ∂x ∂y
The two planar directions x and y are coupled, and the motion in the z direction
is free, with the associated inertia neglected:
δW δW δW
(8.86) = = =0 ,
δεz δγxz δγyz
such that the remaining two-dimensional elastic energy is:
T
E G 2
Z
εx 1 ν εx
(8.87) Wxy = + γ dA ,
2(1 − ν 2 ) εy ν 1 εy 2 xy
which correspond to the energy per unit thickness of the table. See also Section 6.7.
8.14. TRANSVERSE MOTION OF A TABLE 161
20
15
10
0
0 5 10 15 20
Figure 8.5. The static deformation, the black dots are the rest
positions of the grid points, and the gray dots are the displaced
positions. The scale is set for a good visual effect, the fixture is in
the center, indicated by the box.
The table has now three rigid-body modes in terms of the independent dis-
placements u and v, which are the two translations in the x and y direction, and
the rotation around an arbitrary axis in the z direction.
Each pair of nearest neighbors is used the calculate the elongations εi . The
shears require a set of three points: a corner with a point in the x-direction and the
y-direction. For the displacement a linear approximation between those points is
used. [Hughes, 1987] The mass matrix is taken diagonal, where edge have half and
corners have a quarter of the mass of a node away from the edge. See Figure 8.4.
The FEM model of the table consists of 20 points in each direction, which yields
2 × 20 × 20 = 800 displacement variables, and 19 × 19 = 361 cells. The elongations
for the diagonal part of the energy are the 2 × 20 × 19 = 760 elongations, one
for each nearest-neighbor pair of grid points. The off-diagonal energy uses the
elongations per cell, in order to be able to take the product of x and y elongations.
The shear γxy consists of 4 × 19 × 19 = 1449 terms, one for each corner of the cell.
The shear energy consists over the sum over cells with the average over the corner
products. If the average shear γxy would be calculated per cell first before taking
the product, it would yield additional zero modes, for ν = 0. The thickness of the
table is irrelevant for the longitudinal motion.
Unlike the wave equation for the longitudinal motions in a table, the x and
y motion are coupled in elastic theory. The eigenmodes have, therefore, fewer
degeneracies, but there are still some, associated with the modes with four-fold
symmetries. Furthermore, the shapes of the eigenmodes have more displacements
at the boundary, compared to the wave eigenmodes which is the trigonometric
series, due to more complex elastic energy. The eigenmodes are the result of the
diagonalization of the stiffness matrix:
(8.88) ωi2 Vi = KVi .

Table 8.1. The lowest eigenfrequencies ωi , with the correspond-

ing degeneracies. The important modes are indicated with an as-
terisk (*).
mode frequency (kHz) deg.

rigid 0.0 3
1* 1.6965 2
2 1.7989 1
3 1.9000 1
4* 2.0781 1
5* 2.6089 1
6 2.7307 1
7* 2.8066 2
8 3.3069 2
force/acceleration
inertia vibration
s
S
time deformation vibration
Figure 8.6. The two types of spectroscopic factors.
In the case of a degeneracy ωi1 = ωi2 · · · = ωij , the actual eigenvectors are irrele-
vant, only the space spanned by these eigenvectors is relevant:
(8.89) Ω{i1 ,··· ,ij } = span{Vi1 , · · · , Vij } .
The results are given in Table 8.1.
The static-deformation mode x0 is determined from solving the linear equation:
(8.90) Kx0 = F − M a ,
where x is taken perpendicular to the rigid-body modes. The iterative solver
SYMMLQ by Paige and Saunders [Paige and Saunders, 1975] can handle indefinite
K matrices, however, care must be taken that the right-hand side is perpendicular
to the null-space of K. Whether a vibrational mode is excited by the applied force
can be determine from the spectroscopic factors Si , which are the overlap between
the normalized eigenvectors and the static deformation mode:
(8.91) Si = kx0 kΩi ,
which is the length of x0 in the subspace of the eigenvectors of the i-th mode, which
reduces to
q
(8.92) Si = xT0 · Vi ,
for a non-degenerate eigenmode, where kVi k = 1.
8.14. TRANSVERSE MOTION OF A TABLE 163
Leading spectroscopic factors
spectroscopic factor S (arbr. scale)

3
mode 1
mode 4
mode 5
mode 7
2
0
−10 −5 0 5 10
shift of the fixture from center in grid points
Figure 8.7. The spectroscopic factors Si (lines), for the relax-

ation of deformation, of the four leading modes. The corresponding
calculations of Si = ωi2 si are indicated by the symbols.
A second type of spectroscopic constant is associated with the inertia of the

system:
(8.93) si = kFkΩi .
If we would measure the actual displacement at a certain point if a block-wave force
is applied, the oscillations in the rising edge would correspond to the vibrations
with the spectroscopic amplitudes si , while the vibrations in the falling edge would
correspond to the spectroscopic amplitudes Si . (See Figure 8.6.) In the rising edge
the inertia causes the vibrations:
(8.94) mẍ(t) + Kx(t) = F ,
with x(0) = 0, while in the falling edge it is the relaxation of the deformation:
(8.95) mẍ(t) + Kx(t) = 0 ,
with x(0) = x0 . The initial state is given by: Kx0 = F − M a. Hence it follows
that:
(8.96) si = ωi2 Si .
In practice it is easier to determine Si than si since x0 is much smoother than F.
(See Figure 8.7.) In the case of a linear model, the response to the rising edge and
the falling edge is the same. In practice, both numerical as nonlinear effects cause
differences between si and Si .
If the fixture is attached to the middle, the coupling between motion and vibra-
tion is mainly to mode 1, which has a twofold degeneracy. The symbols represent
the spectroscopic factors ωi−2 si , which quality is much poorer, due to the sharp
behavior of F, leading to numerical instabilities in the overlap. To determine the
eigensystem, the eigenvalues and eigenstates, the generic LAPACK [Anderson et al.,
1999] routine DSYEV was used.
We attach a fixture to the center of the table in the y-direction, to a square

of 4 × 4 = 16 grid points. The force is applied in the x direction, and the square
region is shifted along the x-axis. The spectroscopic factors depend on the position
of the fixture, and the leading factors are plotted (see Figure 8.7). However, the
spectroscopic factors do not diminish for higher modes, on average, they stay at
about 5% of the leading factor, throughout the whole spectrum. This is the result
the sharp edges of the fixture. Using only the lowest vibrational modes will only
be sufficient if the high modes damp out quickly, or if they are not excited due to
the slow changes of the applied force.
Another way (See Figure 8.8) to test the quality of the position of the fixture,
is to check the displacement, both the maximal and the mean values. The maximal
value is 3 to 5 times the mean value, and again the center position of the fixture
has the smallest values.
8.15. Simulation of dynamics

From the analysis of the spectroscopic factors it is clear that the lowest vi-
brational mode dominates the dynamics, if the fixture is placed in the center. We
compare the single mode result of the degenerate, vibrational mode, with the static-
deformation mode, and with the full theory. The input signal is a simple force bump,
with a time-scale which is comparable to the eigenfrequency of the table:
(8.97) F(t) = F0 t2 (T − t)2 ; 0<t<T ,
where T ≈ 3.3 periods.
We use a simple Euler stepping algorithm, Section 3.4, with step size δ:
„ « „ « „ «„ « „ «
xt+δ xt 0 1 xt 0
(8.98) = +δ +δ ,
vt+δ vt −K 0 vt Ft
where we have set the mass m = 1, yielding a unit mass matrix. No damping
is introduced. As output we follow a corner of the table, and its deviation from
the motion it would make if the table was a rigid body. The situation should be
in the clear advantage for vibrational modes, often the time-scale of an applied
force is ten times or more the typical period, rather than the 3.3 periods used here.
Furthermore, the absence of other low-lying vibrational modes means the dynamics
is dominated by this single mode. The design has already been optimized to reduced
vibrations.
The one-dimensional vibration reduced model:
(8.99) ẍv + ω12 xv = kF(t)kΩ1 ,
with the output:
CT V11 V1T1 F + CT V12 V1T2 F
(8.100) y= xv ,
kFkΩ1
where V11 and V12 are two orthonormal vectors in the first eigenmode.
Instead we can also use the static deformation under a constant acceleration
as a mode to study the dynamics. The static-deformation mode leads to a similar
reduced model:
xT0 F0 xT F(t)
(8.101) ẍm + 2 xm = 0 ,
x0 kx0 k
8.16. CONCLUSION 165
displacement (arbritrary scale)

maximal displacement
mean x−displacement
0.8 mean y−displacement
0.6
0.4
0.2
0
−10 −5 0 5 10
shift of the fixture
Figure 8.8. The maximal displacement, and the mean displace-

ments in x and y direction of the table for the static-deformation,
depending on the position of the fixture.
where Kx0 = F − M a, and with the output:

CT x 0
(8.102) y= xm .
kx0 k
The computational simplicity of the static-deformation approach, which requires
only the solution of a linear equation, is clear. For the vibrational mode, the large
eigenvalue equation had to be solved. The two-mode approximation is a straight-
forward implementation of the general results. The second mode was projected
perpendicular to the first mode to retain a unit mass matrix.
The results, in Figure 8.9, show the deviation of the lowest vibrational mode,
which is in practice a two-dimensional approximation, due to the degeneracy. The
x-motion is underestimated. The remaining vibration is tracked accurately. The
static-deformation mode alone gives accurate results, except for the remaining vi-
bration, after the force pulse. The two-dimensional input-based model yield the
most accurate results, both for the motion and the vibration. In the final vibration
there remains a small deviation. Including higher-order modes will not increase the
accuracy necessarily, because the higher modes have very large overlap with the
lower modes, and they will generate thereby numerical instabilities. Note, further-
more, in the absence of dynamics; if the mass of the static-deformation mode is
neglected, the shape of the force pulse already predicts 90% of the motion of the
table, the vibrational effects causes oscillations around that result.
8.16. Conclusion
The method of using the deformations of the table under a constant acceleration
as modes for a lumped dynamical model are further discussed in the Chapter 9. The
study of finite elements as invariant components, or port Hamiltonian elements, in
a network for the use in dynamics automatically leads to a more general, and often
nonlinear, formulation. This question is studied in Chapter 11.
force/displacement (arb. scale)

y−motion 800 dof full model
3000 2 dof vibration model
1 dof deformation model
2 dof deformation model
2000
1000
force
0
−1000
−2000 x−motion
0 10 20
time
Figure 8.9. The deviation of the rigid-body motion of a corner

of the table. The two-dimensional deformation model yields the
best results; its short-long-dashed line is almost indistinguishable
from the full results.
This chapter dealt with the more traditional approaches to dynamical FEM.
The specialized methods developed for fluid dynamics are not mentioned explicitly,
although being a large field of research. The connection between energy-based
port-Hamiltonian methods and general fluid dynamics is tenuous at best.
CHAPTER 9
Analytic input
The displacement of a string at a certain moment betrays some history of the

input. If the string is straight and at rest, there has not been any input for at
least as long as it takes a wave to travel from one end to another. Furthermore,
signals that traveled through the string before that time must have been canceled by
appropriate boundary conditions, at the opposite end as from where they emerged.
Hence, to some extent, for general continuous systems, it must be possible to

map the state of the system to the history of the input. This mapping is limited;
different histories may lead to the same state. Furthermore, the reverse may also
be true; the system could be described in terms of input variables. Rather than
to give a formal answer to this mapping problem, we seek a practical answer. In
many practical cases the detail of the input is limited to a discrete set. In our case
we use an analytic input, polynomial functions of time, as a discrete set.
State control has been about finding the state variables that described the
system uniquely. In the case of distributed or infinite systems, this is an impossible
task. The states of actual interest are those excited by the input and operational
circumstances. The input dictates the relevant state variables. Analytic input
space is a particular generic input for finding a corresponding state space. As
seen before, the stationary state, which nature depends on the type of input, is
an important state which characterizes the dominant quality, such as resistive,
inertial, or capacitance, in that setting. The stationary state has a natural place in
the analytic input based formulation, as the lowest order state. Instead of analytic
input, one could choose impulse response as basis for mapping out the state space.
It would have three drawbacks. First, it is dominated by the higher frequencies,
which is not typically the operational state of interest. It uses intrinsically the
linear approximation of the system, which may be a false assumption. Thirdly,
the dimension of the state space would in many cases be too high for a useful
characterization of the system in a small set of states.
For systems with an energy description, the magnitude of the states is bounded
by the integral of all the power supplied to the system. In a behavioral, finite
dimensional setting this has been formalized in the theory of dissipative systems.
[Willems, 1972a,b]
The use of real-time computing in control made elaborate smooth time-dependent
input possible. Usually such input is represented by splines, which are piece-wise
polynomial functions with arbitrary smooth boundaries between the pieces. The
spectral, or frequency, characteristics of such an input is not necessarily obvious.
However, it is clear that it is dominated by the low frequencies. The method based
on analytic input focuses therefore on the achievable control.
167
168 9. ANALYTIC INPUT
In this chapter the common results of control theory are established for analytic,
and piece-wise analytic input. Furthermore, the focus is shifted from the state to
the input. The general benefits of such a shift should be obvious, since the little
one knows about the state of a system, the more one knows about the input. In
many approaches a step input yields a state that converges to a constant state after
a finite time. Although the stationary state does not occur in a finite time window,
they are as limiting state relevant for the dynamics. The approach advocated in this
chapter associates a constant state with the constant input, and deviations from
constancy for input and state simultaneous. Such a framework naturally arrives at
analytic theory and polynomial expansion.
Finally, an input-centered approach is the right starting point to extend meth-
ods to nonlinear systems, where different state-space representations may be equiv-
alent but not equally practical, and the finite-dimensional input-centered approach
distinguishes the relevant subspaces, thereby reducing the complexity to the mini-
mum.
Some subtle differences between the standard control formulation and the ana-
lytic input control will arise. These differences are due to restrictions on the input.
However, these restrictions are quite natural; infinite slopes do not occur in most
physical system. Most singularities in the state arise from, for example, the neglect
of inertial forces. The focusing of waves, however, in nonlinear waves or position-
dependent dispersion may cause a singular shock waves, or caustic waves. [Courant
and Hilbert, 1961]
9.1. Stationary states

The stationary state that arises from constant input is in most cases the most
relevant nontrivial state. For a constant current or voltage, that leads to a constant
voltage or current as state, it corresponds to the state which determines the Ohmic
resistance. In the process of cooling, it is the analogous state of heat conductance,
or resistance. Likewise, for a constant force, it is an experimental verification of
the inertia, or total mass of the system, for a state with constant acceleration. A
state with constant acceleration is only a stationary state in a broad sense, also
referred to as a quasi-stationary state. Hence the stationary state for a given input
can be seen as the constitutive relation for the global or lumped response of the
system; resistance for flow, inertia or capacity for force. Furthermore, the stationary
state often gives a lower or upper bound for the dynamic response. For example, the
stationary flow is the maximal flow for a particular pressure or potential. From rest,
before that flow establishes, the flow is typically smaller because of the buffering
effects, and it reaches the stationary flow from below, in the case of resistance, or
more general, overcritically damped systems.
Altogether, a careful analysis of the stationary states, like in the linear case,
Eq. 3.4, for the particular input and output we choose, is the basis for a lumped or
modal view of distributed systems. A careful energy and power analysis determines
the appropriate matching of the lumped variables with modal shapes and mean
values. For example the total heat W in a system is the product of temperature
distribution T (z) as function of the position, and the capacity distribution, C(z),
also as function of the position. The total heat in the system is:
Z
(9.1) W = dzT (z)C(z) .
9.1. STATIONARY STATES 169
Tin R in T0 R out Tout
C0
Figure 9.1. The simplest network associated with a stationary

flow J and a distributed temperature T (z) and capacity. The mean
capacity C0 should be placed such that it gives the correct mean
temperature T0 .
The total capacity is given by the integral:

Z
(9.2) C0 = dzC(z) .
The mean temperature T0 for the total capacity C0 with the same total heat W
follows from the definition:
Z Z
C(z)
(9.3) W = T 0 C0 = dzT (z) dzC(z) .
C0
The mean temperature weighted with the relative capacity C(z)/C0 , and the total
capacity is the integral over the distributed capacity, such that if the temperature
is constant, the mean temperature is the constant temperature.
In the case of nonlinear constitutive relations, e.g., C(T, z), the product of the
mean values is no longer equal to the work W . With the modal approximation of
T (z) also a modal approximation of W (T0 ) is required for nonlinear systems. The
stationary state is a good starting point for such approximations. Even though
the original dynamics maybe highly nonlinear, small temporal variations around
the stationary state might be well-approximated by a linear model. For example,
consider a system which radiates of its heat. The heat transfer depends on the
fourth order of the temperature difference, but around a particular operational,
stationary state, with a temperature difference Tin −Tout , the variations in Tin +∆T
will lead to variations ∆J in the heat transfer J:
(9.4) ∆J ∼ (Tin + ∆T − Tout )4 − (Tin − Tout )4 ≈ 4(Tin − Tout )3 ∆T ,
where Tin and Tout are given by the stationary state T (z).
A simple network, Fig. 9.1, given the mean temperature T0 and the flow J,
and input and output temperatures Tin and Tout would yield a heat resistance
R = (Tin − Tout )/J. In a lumped network the total capacitance C0 should be
placed at the position where the mean temperature occurs, in between the incoming
resistance Rin and the outgoing resistance Rout :
Tin − T0 T0 − Tout
(9.5) Rin = R , Rout = R
Tin − Tout Tin − Tout
Hence, the heat flow and heat conservation dictates the transition from distributed
to lumped, depending on the situation at hand. The stationary flow is an appro-
priate situation, however, many other choices can be made. For example, instead
of the mean temperature T0 , one could choose the temperature at a sensor Tsensor .
The sensor temperature Tsensor together with the total capacitance C0 will not yield
the total heat in the system W 6= Tsensor C0 . The appropriate manner to do so is

the split the resistance in which the sensor resides Rin if T (zsensor ) > T0 or Rout
if T (zsensor ) < T0 , and rely on the energy-conserving network, but often another
capacitance Csensor is chosen to match Tsensor , i.e., W = Tsensor Csensor . Sensor out-
put dictates many input-output networks, where component consistency is lost. In
principle, for such simple networks, it makes no difference, but once it is placed in a
greater network, and possibly extended with other properties such as inductances,
these choices matter.
Typical input and output temperatures, like Tout and Tin , are not the mean,
or lumped, temperature T0 , used to determine the total heat W in the system. If
they are localized at, for example, zout and zin , they follow from the modal shape:
T (zi )/T0 . However, in many cases the temperature at the port itself is not constant,
but the temperature T (z) over an area of the boundary z ∈ ∂Ω.
Again the stationary state will yield a consistent definition of the heat flow in
and out of the system. The heat flow j(z) through the boundary sums to the total
heat flow in and out of the system:
Z
(9.6) J= dzj(z) .
∂Ω
The particular entropy generation in the system is associated with the temperature
difference between input and output:
Z Z
−1 j(z)
Z
−1
(9.7) S = dz(T (z)) j(z) = dz(T (z)) dzj(z) = T −1 J .
∂Ω ∂Ω J ∂Ω
In this case, like with C(z)/C0 in Eq. 9.3, the flow provides a measure j(z)/J, for
a consistent mean temperature at the boundary:
C(z) j(z)
Z Z
(9.8) dz = dz =1 .
Ω C0 ∂Ω J
The stationary state of the system gives us appropriate lumped parameters and
variables, which are consistent with the conserved quantities in the system, such as
energy and entropy. The state at rest, i.e., T = constant, would not have been able
to provide this linkage to lumped models.
Furthermore, the stationary state can be of a nonlinear model, it changes very
little in the definition of total and mean variables. The product, such as the power
product of effort and flow, in the integrand remains a product by definition. Non-
linear relations are constitutive relations, or closure relations, [Bird et al., 1960]
which link efforts and flows to the state variables.
To extend this notion of stationary states to lumped model beyond the simplest
approximation, described above, we view the system as the response to an analytical
input function of time. The stationary state is the response to a constant input, at
the next orders linear u(t) = a + bt and quadratic input u(t) = a + bt + ct2 are
considered. These functions may seem unnatural for large times t → ±∞, however,
the modal shapes, e.g., {ti (z)} for a response T (z, t) = t0 (z) + t1 (z)t + t2 (z)t2 + · · · ,
can be used to define a fast converging basis for time-variations of the input, which
are typically of the same order or slower than the longest time-scales of the response.
9.2. SERIAL NETWORK MODEL THROUGH TRIDIAGONALIZATION 171
9.2. Serial network model through tridiagonalization

Before considering time-dependence of a system as a response to an input,
we will examine the linkage between Lanczos tridiagonalization, described in Sec-
tion 2.12, serial networks, and extension of the simple network of the previous
section. The intermediate step between a distributed system, depending on a con-
tinuous position z, and a lumped model is often a large FEM model, which is sub-
sequently reduced to a lumped model with a small number of degrees of freedom.
[Davies, 1978] This last process is called model-order reduction and has received a
lot of attention in recent years. [Freund, 2003] The link between the reduced-order,
or lumped model, and the distributed model is tentative at best. Tridiagonalized
models, however, can be viewed as lumped, network models. In the matter of
truncating, or reducing, the tridiagonal model, the network representation helps to
retain charge conservation, from the continuous model to the lumped model.
Complete diagonalization is an iterative procedure which breaks to model into
isolated submodels, each corresponding to an eigenmode. However, the procedure
of diagonalization does not easily allow for the inclusion of damping or multiple
inputs and outputs. The algorithm for tridiagonalization is, on the other hand,
well-defined in terms of a finite number of steps. It also allows one to make a direct
link a network model.
Some details, such as conservation laws of the distributed system, are retained
in the proper network model, but are easily lost, if not the proper care is taken,
to consider the network model under reduction. [Freund, 2003] In this section we
investigate two cases of network models, which may correspond to a symmetric
matrix A.
A general symmetric matrix Ã can be brought to a tridiagonal form:
 
α 1 β1
 β1 α 2 β2 
 
 β 2 α 3 β 3 
(9.9) A=  .
 
 β 3 α 4 β 4 

 β 4 α 5 β 5 

.. ..
. .
Such a matrix can be mapped onto a serial network with capacitances and resis-
tances; a RC-network, or capacitances and inductances; a LC-network, depending
on the differential equation:
(9.10) ẋ = ARC x, ẍ = ALC x ,
where xi is a voltage. The resistors Ri and inductors Li are in between the voltage
points xi and xi+1 , connected to the ground via a capacitor Ci . See Figure 9.2. The
consequences of conservation laws, in this case charge conservation, are not a priori
clear from the coefficients αi and βi . In the RC-network and the LC-network, these
conditions are related to the capacitances Ci , as we will see.
In the case of a RC-network the variables are:
 −1 1 
R1 C 1 R1 C12
 1 −1 1 
R1 C12 R12 C2 R2 C23
(9.11) ARC = 1 −1 1  ,
 
 R2 C23 R23 C3 R3 C34 
.. ..
. .
V1 R1 V2 R2 V3 R3 V4
C1 C2 C3 C4
Figure 9.2. The serial RC-network which can represent a tridi-

agonal matrix ARC . In the case of a LC-network, the resistors R
are replaces by inductors L.
p
where Rij = Ri Rj /(Ri + Rj ) and Cij = Ci Cj . The replacement of the resistors
by inductors changes very little in the network, and the corresponding matrix has
the same structure:
 −1 1 
L1 C 1 L1 C12
 1 −1 1 
L1 C12 L12 C2 L2 C23
(9.12) ALC = 1 −1 1  ,
 
 L2 C23 L23 C3 L3 C34 
.. ..
. .
where analogous: Lij = Li Lj /(Li + Lj ).
It is important to notice that charge conservation means that:
 √ 
√C1
 C2 
 √  √ √ √ √
A  √C3  =

(9.13) C1 C2 C3 C4 · · · A = 0 ,
 
 C4 
..
 
.
which determines the relevant
√ null-space of A. The transformation
√ √ with the diag-
onal matrix consisting of Ci , such that A → diag( Ci )Adiag( Ci ) transforms
the voltage state xi = Vi into the charge state xi = Qi . The charge conservation
occurs both in the RC and the LC network, the same null-space arises.
Each of the voltages xi = Vi in the serial network is a coefficient of a mode
Vi (z). Hence a lumped model can arise from a numerical model order reduction of a
large A to ARC or ALC . However, in many cases the lumped parameters have little
meaning in the underlying model. The input is connected to all the nodes since,
typically, for every mode Vi (z), for every node zj , the overlap is nonzero: Vi (zj ) 6= 0.
Hence if an external connection is made at a single node zj , it is connected to all
the modes, and all the nodes of the lumped model, if in the tridiagonalization only
the internal dynamics and not the input and output are considered.
In the truncation of the tridiagonal matrix to a smaller matrix with the dom-
inant modes, the last diagonal coefficient (Cn R(n−1)n )−1 or (Cn L(n−1)n )−1 must
be adapted to respectively (Cn Rn−1 )−1 or (Cn Ln−1 )−1 , in order to retain charge
conservation. More generally, numerical errors in the tridiagonalization procedure
can be corrected to retain global charge conservation, lost in numerical errors. This
procedure can be based on the serial-network analogy.
Likewise, the complete diagonalization of A turns every state xi into the co-
efficient of a vibrational mode Vi (z). In both cases the input and output are
distributed, while the dynamics is, more or less, localized. The stationary state
connects input and output by a continuous flow, hence the stationary state lies the
9.3. ANALYTIC INPUT 173
basis for distributed dynamics and localized input and output. It is a fundamental
difference in perspective.
More appropriate is to rely on the input to determine the lowest mode, such
that the interconnection between the system and the environment is consistent for
every lumped, low-order approximation of the system. The stationary state is an
important aspect in this consistency.
9.3. Analytic input

The construction of modes, or relevant state vectors, from diagonalization, will
not include the stationary state for a static response. The dominant characteristic
of a system, being a resistor to heat flow, or an inertia under acceleration, is not
recognized, and given a place. Furthermore, the next-to-leading-order behavior,
such as the elasticity of the accelerating inertia is not singled out by a standard
modal analysis. Using a polynomial expansion in time of the input does, on the
other hand, separates and singles out the separate characteristics of the system
under the influence of particular input.
A state-space representation of the evolution (see Chapter 3) of a linear system
is given by:
(9.14) ẋ = Ax + Bu ,
and with the output:
(9.15) y = Cx + Du ,
where the input u ∈ R , the state x ∈ Rn , and the output y ∈ Rk are functions of
m
time only. If we insert an analytical expansion for the input:

(9.16) u(t) = u0 + u1 t + u2 t2 + · · · + uN tN ,
with similar expansions for the state x ∈ Rn×m :
(9.17) x(t) = x0 + x1 t + x2 t2 + · · · + xN tN ,
and the output y ∈ Rk×n . The typical dimension N , for a linear system does not
have to exceed the dimension of the state space. The projection onto the state
space determines the maximal dimension of both input richness and observability.
The input ui ti is mapped to the state space xj . If this mapping of a polynomial
in t to a subspace of the state space retains the dimension N of the polynomial,
we speak of a rich input. The full richness of the input are all the states possibly
excited by an arbitrary time-dependent input.
The evolution now yields a hierarchical set of linear equations for each order
tj :
(9.18) (j + 1)xj+1 = Axj + Buj ,
where the equation for j = 0 is the stationary equation.
The first question to address is the input richness. Analytic input yields a full
input richness if for N = n − 1 and ∃ui :
(9.19) rank{x0 , x1 , x2 , · · · , xn−1 } = n .
Such a definition is equivalent to the statement that there is a time t for which a
state x can be achieved, with the analytic control u.
The equation for xj is easily solved successively, starting from j = n − 1:
(9.20) xj = A−1 ((j + 1)xj+1 − Buj ) .
The states can be expressed in terms of the control u(t):

i
X (n − j)!
(9.21) xn−i = − Aj−i−1 Bun−j .
j=1
(n − i)!
The linear independence of the different ui and the upper-diagonal structure of the
coefficient matrix yield the input richness condition.
The input richness is equivalent to:
(9.22) rank{A−1 B, A−2 B, · · · , A−n B} = n .
This input richness is similar but not identical to the standard controllability condi-
tion, Eq. 3.29. Controllability is typically associated with low-dimensional systems.
Distributed systems have many more degrees of freedom than of interest for stan-
dard operation, hence is never controllable. The question is reversed: how many
states are participating in the response to typical, slow, i.e., analytic input.
The difference lies in the inverse of A. If A is singular, the stationary state,
associated with the right null-space of A, cannot be maintained with a single sta-
tionary input, as would be expected from the j = 0 equation.
9.4. Delay systems

Many distributed systems are, as input-output systems, characterized by a de-
lay τ . In terms of states it means that different states have time-shifted algebraic
relations: xj (t + τ ) = xi (t). Transmission cables are typical examples: the sig-
nal enters one end xi , and arrives τ time later at the other end xj . In terms of
differential equations, delay corresponds to a differential equation of infinite order:
(9.23)
τ2 τ3
eτ ∂t xj (t) = xj (t) + τ ∂t xj (t) + ∂t2 xj (t) + ∂t3 xj (t) + · · · = xj (t + τ ) = xi (t) ,
2! 3!
where ∂t is the differentiation with respect to time as operator, which relies on
an analytic expansion of xj (t). Such delay equations are truncated at a particular
order to analyze the effects of a feedback control loop. [Minorsky, 1942]
In terms of analytic input, delay equations are upper triangular matrices. The
shift of a polynomial u(t) = u0 + u1 t + u2 t2 + · · · un tn will be a polynomial of the
same order:
 
 2 3
 u0
1 τ τ τ ···
 0 1 2τ 3τ 2 4τ 3 · · ·   u1 
 
2 3 n 
(9.24) u(t + τ ) = (1, t, t , t , · · · , t )  0 0 1 3τ 6τ 2 · · ·  
  u 2  ,

  .. 
.

.. ..  . 
. ···
un
where the coefficients follow from the binomial expansion.

If the output is the delayed input, the upper-diagonal matrix, Eq. 9.24, will ap-
pear in the system matrix A. For large delays the order of the input tn , rather than
the number of sampling points, determines whether the system is well-conditioned
and stable.
9.6. FAILING INPUT RICHNESS 175
9.5. Finite time

A free system has a unique evolution given by its initial state x(0), which is
equivalent to x0 in the analytic case. The series {xi }∞ i=0 expands to infinity, how-
ever, only the first n terms are relevant to define A and thereby the free system
uniquely. In the case of input u the dimensionality of the input determines the di-
mensionality of the system that can be identified or controlled. In the standard case
an infinite number of derivatives of the input are assumed available to reconstruct
the state space. In practice, this is never the case.
In general, the input dimension is much lower than the dimension of any physi-
cal system, although many states correspond to irrelevant, damped high-frequency
eigenstates. The constructive analysis to optimize an input profile consisting of a
piece-wise polynomial function is a version of Heaviside calculus.
Piece-wise polynomial solutions can be constructed in many different ways. For
example, splines could lie at the basis of piece-wise polynomial solutions. Higher
order splines themselves can be constructed from successive convolutions of block
functions, which indicate the low frequency effect of splines. The k-th order spline
has a Fourier transform:
k
(k) sin T ω
(9.25) S (ω) = ,
ω
where T is the length of the block of unit height. Since a piece-wise polynomial
function can be written as the convolution of a Dirac comb with a spline, it is clear
that the k-power of the the spline yields a low-frequency dominance of the piece-
wise polynomial input data, depending both on the domain T , and on the order of
k.
For analytic input the state is a slave of the input, for the first n orders deter-
mined by the input. In the case of a piece-wise polynomial input this is no longer
the case. The state may be arbitrary.
The study of finite time is also part of the theory of input shaping. [Singer and
Seering, 1990] The typical step response is replaced by a smoother, typical spline,
function of time. The function is optimized by minimizing some cost function of
the defined output. There are many variations in how define the output and a cost
function to minimize. Furthermore, the shaping functions can have a large variety,
for example in fixed or free abscissas, order, etc.. Most of the approaches rely on
linearity of the system, and the problem shifts from input shaping toward system
identification. Here we start at completely the opposite end of the spectrum, we
know the system, but it is too large. The limitations of the input help us to define
a relevant subspace of the total state space.
9.6. Failing input richness

Unlike the controllability condition, which will simply fail for insufficiently rich
input, the richness condition will break down more progressively, in many cases.
The state vectors xi will still have components which span an additional subspace
with every new state vector xi+1 , associated with a part of the vector perpendicular
to all the previous vectors. This is due to the very large dimension of the state space,
based on FEM or some other high-order approximation. However, this component
will decrease. Within some tolerance a particular state vector xi+1 will no longer
be perpendicular to the space spanned by: {x0 , x1 , · · · , xi }:

kα0 x0 + α1 x1 + · · · + αi xi − xi+1 k
(9.26) min < .
α kxi+1 k
The input spans thereby a space of i dimensions; the full input richness is i for the
tolerance . In the theory of Krylov spaces a small input richness is an unwanted
feature; no longer the full state space is spanned by the input. For control purposes,
the small input richness is advantageous. Only a limited number of states are
excited, and, therefore, only a limited number of states are required to describe the
system dynamics. If, however, more states are included in the state-space model,
the dynamics will fail, with singular behavior associated with near-parallel state
vectors.
In the construction of the dynamics in the finite-dimensional subspace, Sec-
tion 9.8, we will consider the effects of failing input-richness, its causes and con-
sequences, further. The failing input richness can be seen as a lack of structure
in the system. For example, consider a force exerted on an elastic body in free
motion. The stationary response is the acceleration of the inertia as a whole. The
next term is the delay in acceleration, due to the elastic deformation of the body. If
the object is more or less homogeneous, it can be represented as a series of masses
and springs with a homogeneous spectrum. However, if the mass is located in one
place of the object and the elastic deformation has no associated mass, the system
is well-represented by a single mass-spring pair, and the input richness will fail. It
will fail, however, for physical reasons, the remainder of the spectrum, after the
dominant mode, occurs at much higher frequencies.
9.7. Smooth matching

The smoothness of the spline input u(t) determines the frequency band in which
the input occurs. Hence in terms of a piece-wise polynomial input, it is not the
order of the polynomial itself, but the smoothness of the transition from one part
to the next that determines the decay rate of the higher frequencies in the signal.
The matching means that at the end of each time slice t ∈ [0, T ] the state x(T )
is determined and used as initial state x(0) for the next time slice. Rather than
a state fully determined by the analytic input, the piece-wise analytic domain will
have the initial state as inhomogeneous term, and the equations of motion must be
integrated explicitly:
Z T
(9.27) x(T ) = eAT e−At Bu(t)dt + eAT x(0) .
0
The state x ∈ Rn is a long vector, which we expanded, in the analytic input
approach, in the {ti }N
i=0 polynomial dependence on the input. Hence, rather than
working with the very large dimension n, we will map the problem on the N -
dimensional space of analytic responses.
9.8. Projection on the input space

Previously, we used analytic input to construct a number of modes, or states,
which are relevant for the input we are considering. These modes are a particular
Krylov space. However, the modes themselves do not determine the dynamics, or
the reduced A, B, C, and D matrices. The dynamics of the full space does not
guarantee that the state remains in the subspace spanned by the Krylov vectors.
9.8. PROJECTION ON THE INPUT SPACE 177
The dynamics out of the Krylov space must be projected back into this subspace
under consideration. Many different choices for the constructed of the reduced
dynamics in the reduced subspace exists. We will give the general procedure to
construct reduced dynamics.
In numerical mathematics the controllability condition is similar to the span
of a Krylov space. The dimension of the Krylov space spanned in the rank of
controllability matrix. Hence, in the case of model order reduction, based on the
input vectors B, the controllabity tells us how well the system is approximated.
The difference between the standard controllability:
(9.28) X̃ = {B, AB, A2 B, · · · An−1 B} ,
and the analytic controllability:
(9.29) X = {B, A−1 B, A−2 B, · · · A−n+1 B} ,
has more significance in the numerical setting where the dimension of A is high,
than in the pure control setting.
In the original input dependence, states were constructed algebraically by col-
lecting the same orders of t and selecting the set of equations which would yield a
one-to-one map under general controllability conditions. In general the state can
be arbitrary. Furthermore, in physical control, the number of states is likely to be
infinite. However, from rest x = 0 or any constant state x = −A−1 Bu for constant
input, the time-dependent input determines the dominant subspace of the state
space. Eigenstates, not initially excited by the analytic input, are not very likely to
be excited at all. In order for such a situation to occur the diagonalization of the
matrix A should contain Jordan blocks larger than the input space dimension.
Hence we seek an approximated, or reduced-order, model determined by the
input u(t). The states xi are already determined from the analytic input-to-state
mapping, but a projection onto an n-dimensional dual x∗i is needed to close the set
of equations and yield a n-dimensional linear set:
(9.30) X∗ Ẋ = X∗ AX + X∗ Bu ,
where X is the matrix constructed from the columns x0 , x1 , · · · , and X∗ is likewise

constructed from the rows x∗0 , x∗1 , · · · . Four obvious choices for the dual X∗ are
easily given. The state dual X∗ = XT based on the l2 -norm of the state space,
the state inverse X∗ = (XT X)−1 XT which diagonalizes the left-hand side, the
weighted dual X∗ = XT A−1 , and the weighted inverse X∗ = (XT X)−1 XT A−1
which diagonalizes the subspace A matrix.
All of these dual spaces yield basic block-Hankel matrices for evaluation:
BT A−r B BT A−r−1 B BT A−r−2 B

 
 BT A−r−1 B BT A−r−2 B BT A−r−3 B 
(9.31) Hr =  BT A−r−2 B BT A−r−3 B
 ,
· 
T −r−3
B A B ·
which are known to lead to poor stability after the fast initial convergence. However,
this poor stability is related to higher frequencies in the system, which are, with
sufficiently smooth input, not necessary of interest, as pointed out in Section 9.6.
9.9. Laguerre projection

Analytic input has no generic time-scale, analytic functions scale proportionally
with the scaling of the time. This approach is very useful for hyperbolic systems,
which are not damped, but remain oscillating. Every input, however long ago, adds
to the state dynamics, depending on the phase, However, in many cases systems
are damped, and they have an inherent time-scale, given by the distance of the
poles in the transfer function to the real ω-axis, or imaginary s-axis. Hence, of
the time-dependent input only a window, the size of the minimal damping rate, is
relevant for the state of the system. The input richness is limited to the behavior in
that window. Analytic input cannot separate the behavior inside and outside the
window. More appropriately in such situations is to consider a Laguerre projection,
which is an orthogonal basis on the window function based on the exponential decay
in time of input.
Instead of using the state space, and the projection on the dual space, the time-
dependence of input and states can be used to determine to determine the state as
function of the input. In the case of the lowest mode having the characteristics of
a resistance, it yields a time constant λ, which can serve as the basic scale of an
orthogonal basis expansion in terms of Laguerre polynomials Li (t): [Weniger, 1985]
Z 0
(9.32) ci = λ eλt u(λt)Li (−λt)dt .
−∞
The Laguerre polynomials form an orthogonal basis: [Magnus et al., 1966]

Z 0
(9.33) et Lj (−t)Li (−t)dt = δij .
−∞
The normalization varies from author [Abramowitz and Stegun, 1965] to author, we
use [Magnus et al., 1966] as the standard. In a more general setting, with smooth
(0)
input, one can use the generalized Laguerre polynomials, where Li (t) = Li (t)
with the weight function (−t)α et , where α > −1 is to be optimized:
Z 0
(α) (α) i+α
(9.34) (−t)α et Lj (−t)Li (−t)dt = δij Γ(α + 1) .
−∞ i
The Laguerre polynomials satisfy the second-order differential equation:
(9.35) t∂t2 L(α) (α) (α)
n (t) + (α + 1 − t)∂t Ln (t) + nLn (t) = 0 .
In Section 9.3 the modes or states xi were constructed by solving algebraically

the evolution equation for every power of time, ti , necessitating a truncation of the
highest order in the analytic expansion of the state, xn tn . The Laguerre projection
is an alternative form of constructing similar, but not identical, states. The Fourier
transform of Laguerre functions e−t Ln (t) have interesting properties, with a single
pole and zero of (n + 1)-th and n-th order. [Wiener, 1949] The stationary state
arises from the lowest-order expansion in Laguerre polynomials with a weighted
time-average input u0 :
(9.36)
Z 0
dt ẋ(t)eλt − Ax(t)eλt − Bu0 eλt = x(0) + (1 − λ−1 A)x0 − λ−1 Bu0 .

0=
−∞
9.9. LAGUERRE PROJECTION 179
Hence there are four directions in state space: the state at t = 0: x(0), the average
state: x0 , the average change: Ax0 , and the control input average: Bu0 . The
system can be solved yielding the stationary state as average state: x 0 = −A−1 Bu0 ,
with x(0) = x0 . The equation can be balanced, given equal weight to the different
parts for λ = kAk.
In order to construct a set of modes in state space, similar to the analytic expan-
sion xi , we project the equation of motion onto the different Laguerre polynomials,
which define the weighted time averages xi and ui :
Z 0
dt ẋ(t)Li (−λt)eλt − Ax(t)Li (−λt)eλt − Bu(t)Li (−λt)eλt .

(9.37) 0 =
−∞
We use the following properties of Laguerre polynomials: [Magnus et al., 1966]

i−1
(α) (α)
X
(9.38) ∂t Li (t) = (−1)i−j Lj (t) ,
j=0
and

(α) n+α
(9.39) Li (0) = ,
n
which yields the j-th equation to solve xi and x(0) from the input Bu for α = 0:
j
X
(9.40) 0 = x(0) + (−1)j−i xi − λ−1 Axj − λ−1 Buj .
i=0
In matrix form the equations for the modes are:

(9.41)
1 0 0 0
0 1
0 1 ··· 0 1 0 −1
λ Ax0
1 0
λ−1 Bu0
1
x(0) x0
B x(0) C B
B −1 1 0 0 ··· C
C B x1 C B λ−1 Ax1 C B λ−1 Bu1 C
B . C+ B 1 −1 1 0 ··· CB . C=B C +B C .
.. ..
B C B CB C B C B C
@ .. A B .. .. C @ .. A @ . A @ . A
@ . . A
x(0) xn λ−1 Axn λ−1 Bun
··· ··· 1 −1 1
Setting x(0) = 0 for a moment, the modes xi can be solved as functions of Bui :
i
X
(9.42) xi = (λ − A)i−j−1 Buj .
j=0
In the general case the solution x(0) is sought such in the subspace spanned by
{A−i Bu}∞ i=0 , arising from the analytic expansion of (constant − A)
−1
. Unlike the
analytic case the vector x−1 = Bu plays a role in the Laguerre projection.
Depending on the particular application, different procedures can be followed.
For example, one might optimize the value of λ for a particular time-dependent in-
put, ui . It is important to notice the different linkages between the time-dependent,
and the restrictive input signal, and the subspaces of the large state space arising
from a FEM model. Input is central in restricting to the relevant dynamical sub-
space of the state space. This is a different approach from mapping a transfer
function between a significant, i.e., physical, input and a, usually rather arbitrary,
output, as function of the frequency. [Antoulas, 2005, Worden and Tomlinson, 2001]
input−dependent states
x=0 energy
restricted
space
E = constant
full state space
Figure 9.3. The energy and the input together restrict the full
state space to a small subspace of dynamical states, which are rel-
evant for the driven, or operational, dynamics. The input restricts
the dimension of the subspace, while the energy restricts the norm
E ≤ constant. These can be the operational settings of the partic-
ular system due to the restricted input.
9.10. Remarks on the limitations of linearity

In this chapter we restricted our attention mainly to linear models, typical
of large FEM applications. In this particular field physical knowledge is buried
deep in the large-matrix representations. However, time-dependent behavior may
still pry some of the relevant physics from these representations. For example, the
stationary state is particularly relevant for determining the main function of the
systems under investigation, e.g., being a resistive or capacitive element, and the
scales associated with both characteristics.
Extending this view we note that the operational setting of a system is usually
much more restrictive than the full state space of the FEM model. The time-
dependence of the input is used to determine the smaller, and more versatile, dy-
namical subspace of the full state space. The use of analytic input allows one to
include the stationary state in this subspace, a result commonly not achieved with
impulse response approaches, with a preset input signal, such as a step function or
oscillatory input. [Lee et al., 1999]
After the modes are selected, still some relevant information of the system
at hand is necessary to construct the dynamical equations in the subspace of the
selected modes. In this chapter we restricted ourselves to a noting different projec-
tions on the relevant dual subspaces. Physical principles, such as conservation laws
can be used to select the proper norm and projection. See Figure 9.3.
In the case of systems with nonlinearities the same physical principles can be
used to retain relevant modes from the full space. The input is to be expressed
terms of conserved quantities, such as energy, charge, or mass, such that the state
space dynamics is automatically restricted to the subspace limited by the same
conservation laws. In the case of nonlinear system, the conservation laws are still
linear. They are the guiding principles to steer clear from the instabilities and mul-
tiple, spurious solutions, typical on nonlinear equations. In the following chapters
these questions are investigated further.
CHAPTER 10
Real and virtual dynamics
For since the fabric of the universe is most perfect and the work
of a most wise creator, nothing at all takes place in the universe
in which some rule of maximum or minimum does not appear.
[L. Euler]
The term virtual work was coined by Lagrange to determine the internal balance
of forces, and the balance of forces between objects. He used the concept in the
introduction of generalized coordinates and the study of celestial mechanics in the
second half of the eighteenth century. [Gantmacher, 1970] D’Alembert consequently
popularized the concept.
Force is not an invariant concept, but the product of force and displacement is.
In the continuous setting force could be pressure, or some more complex distributed
source, expressed as a stress tensor. The product of any such representation of the
force and its dual displacement is the invariant work, independent of the represen-
tation.
Work allows for general coordinate transformations. The balance is the config-
uration, or state, of minimal energy, and a small, virtual, change of configuration
will always increase the energy. In other words: the virtual change of the state is
at the expense of virtual work. In this chapter we will take to analogy one step
further and assign to the iteration procedure a virtual time. The process of approx-
imating a solution successively is treated as a dynamical process. Once the analogy
is established, it is taken one step further by combining real dynamics with virtual
dynamics to describe dynamical systems with constraint relations.
Virtual work can be associated with many other problems. For example, con-
vergent iterative processes [Saad, 1995] are convergent in some norm, which can be
interpreted as the energy of the system E(x) = kxkE , and the Gateaux derivative
with respect to the state x is the associated virtual force:
δE(x)
(10.1) F=− .
δx
The iterative process can be identified as a damped dynamical process. Further-
more, constrained systems with dynamics can be treated as a whole: the real
dynamics can be combined with virtual dynamics of the constraint relations. The
linkage exists at multiple levels. First we investigate the time scale and damping
of a variational problem given by a sparse matrix. Second, we analyze the vir-
tual dynamics associated with multiple constraints. Finally, we combine the real
and virtual dynamics and reduce the influence of the virtual dynamics on the real
dynamics.
Linear variational problems with sparse matrices usually arise from FEM ap-
proach to partial differential equations. Such matrices have quite a lot of structure
181
182 10. REAL AND VIRTUAL DYNAMICS
from the underlying geometry. A paradigmatic example is the simplest approxi-

mation of the Laplacian ∆ in one dimension. The corresponding matrix operator
A ∈ RN ×N , associated with a piece-wise linear approximation is given by:
 
−1 1 0
 1 −2 1 0 
 
 . .. 
(10.2) A=  0 1 −2 1 0  .


 0 1 −2 1 

.. .. .. ..
. . . .
The particular Hilbert space l2 in which the solution x lies has a priori no signif-
icance. The nature of the problem is given by A and the inhomogeneous term b
for Ax = b. In general, the solution x = (x1 , x2 , · · · , xN ) will have all coefficients
xi 6= 0, in no particular order of size xi ∼ xj . A more natural space is the Krylov
space Kn (x0 , A) = {x0 , Ax0 , A2 x0 , · · · , An x0 ). The only candidate for the initial
vector x0 = b is the inhomogeneous term b, usually representing an applied force
of in flow.
An iterative method, for the solution x of the equation Ax = b, is based
on the successive multiplication of the matrix A with the vector b. If the vector
b = (1, 0, 0, 0, · · · , 0, 0) it takes N − 1 multiplications before the whole space is
filled, for example:
 
  −5
1
 0   9 
 
   −5 
(10.3) A 3 b = A3  0  =  1  .
   
 ..   
 .   0 
..
 
0
.
Only the first four coefficients are non-zero. As in most solutions to such variational
problems all the coefficients are nonzero, therefore, at least N − 1 multiplications
are required to find a solution. The matrix A (Eq. 10.2) is the linear discretization
of the Laplacian in one dimension. See Sect. 8.4 The walk through the space
from the input at one end, to the other end requires N steps, equivalent to N − 1
multiplications with A. The connectivity is N − 1, as it takes N − 1 steps to
walk through the whole state space. The conjugate gradient method, discussed in
Section 10.3, is a method which requires precisely N iterations to find the solution,
but the amount of work per iteration is more than the virtual dynamic approach
requires, as we will see. Furthermore, the dynamical interpretation allows us to
extend the method to nonlinear problems.
Another aspect of iterative methods is also immediately clear from inspecting
the vector A3 b. The coefficients are growing and they are alternating. The series
Aj b is not converging, since kAk = 2 > 1. If A is multiplied with an appropriate
factor (∆t)−1 , the series is convergent. The factor is the reciprocal of the discrete
time step ∆t, associated with evolution and dynamics, discussed in chapter 3.
Iterative procedures have a close link with integration methods for stable dynamics.
The number of steps n determine the subspace in which the approximate solu-
tion lies. This space is the Krylov space Kn (A, b). The connectivity indicates the
10. REAL AND VIRTUAL DYNAMICS 183
minimal size of the subspace. The larger the dimension of the underlying geometry
the smaller the connectivity compared with the dimension N of the matrix. In
treating the iterative solving of a numerical problem as a dynamical problem these
are all aspects that can be taken into account. Eventually more prior knowledge
will improve the convergence.
A last aspect, also clear from inspecting the vector A3 b, is that the sum of the
coefficients add up to zero. This is no accident, but the result of the underlying
PDE, the Laplace operator. Any constant function of position vanishes, as does
any linear function. These vectors lie in the left null-space of the Laplace operator,
Eq. 10.2. The FEM discretization should retain this property as this associated
with the force balance or mass conservation depending on the underlying problem.
Hence deviations from zero for the sum give some insight in the convergence, and
the position of the interacting boundaries and the forces at these boundaries. By
treating the iterative process as a numerical algorithm only, such aspects are often
forgotten.
Recently, the underlying physical system is been taken into account again, at a
limited scale, in methods referred to as structure preserving model-order reduction.
[Freund, 2003] For example, the positivity of the mass matrix and the stiffness
matrix are retained in the approximation, corresponding to a Krylov space K n
with its dimension much smaller than the dimension N of the full space.
To recapitulate, real and virtual dynamics refer to two aspects of a dynamical
simulation of a system. The real dynamics is the time evolution of a system as
determined by the equations of motion. Virtual dynamics refers to the iterative
solution of a numerical problem, [Saad, 1995, Antoulas, 2005] and treating it as a
dynamical problem. [Stuart and Humphries, 1998] These two problems are inti-
mately related, since they are both based on the series Aj b. First of all, in the
iterative solving of a linear equation an approximate solution converges to the true
solution. If each step is seen as a time step, the convergence process is similar to
that of a stable damping, usually associated with Lyapunov theory. [Lyapunov,
1992] Hence, turning the problem around, many numerical problems can be solved
once they are converted to a stable, virtual, dynamical problem.
On the other hand, constraints, or algebraic equations, may appear in dynami-
cal problems, which are simulated through the time integration of the equations of
motion. In each step the constraints need to be solved. The explicit solving yields
a corrector step; after each step the constraints are implemented, the implicit solv-
ing yields an adjusted step, which is already tangential to the constraint surface.
However, constraints may be introduced as virtual dynamics such that, even if the
initial state does not satisfy the constraint relations, the state converges to the con-
straint surface and remains there within some numerical bound. A simple generic
corrector step is the Newton iteration. If a function should satisfy H(q) = E, the
steepest descent yields the smallest linear correction q → q + δq, with min kδqk to
the solution H(q + δq) = E:
(E − H(q))∇q H
(10.4) δq = .
(∇q H)T ∇q H
See Figure 10.1. More elaborate schemes exists.
All of these different cases will be discussed in this chapter. The advantages of
using virtual dynamics are manifold. First of all, the methods described here can be
applied irrespective of the nature of the constraint, unlike most methods. Moreover,
H
H=E
H=E’
Figure 10.1. A corrector step in the direction of the gradient ∇H

is the smallest step to move from a trajectory with the incorrect
energy E 0 to the trajectory with the correct energy E.
both the constraints and the equations of motion may be nonlinear, which makes
the method versatile. Furthermore, the methods are generic, many problems can
be treated in the same manner, and the complexity will not change the general
dynamic approach.
10.1. Iterative methods

The relation between symmetric matrices and quadratic Hamiltonians is well-
established. The theory of matrices is commonly applied to dynamical systems to
derive results of stability and vibrational analysis. However, this relation works
both ways. The results above can be set into an explicit Hamiltonian dynamical
scheme. Furthermore, the connection with dynamical systems allows one to extend
iterative methods to nonlinear problems.
A simple example, related to the problem above, to set the mind, is the positive-
definite symmetric matrix A, and the linear equation Ax = b. If the matrix A is
large and sparse, a direct solver will take prohibitively long.
A corresponding Hamiltonian can be:
1 1
(10.5) H(q, p) = K + V = p2 + qT Aq .
2 2
The Hamiltonian describing the motion of a unit-mass particle is a multi-dimensional
ellipsoidal potential V , with a kinetic energy K. The equations of motion are:

q̇ 0 1 q
(10.6) = ,
ṗ −A 0 p
where the dots in q̇ and ṗ indicate the time derivatives. The solution q(t) and p(t)
will now oscillate around the minimum q = 0 and p = 0.
The right-hand side b be added to the equation can be seen as an applied force:

q̇ 0 1 q 0
(10.7) = + .
ṗ −A 0 p b
The solution (q(t), p(t)) will oscillate now around the new minimum, which is given
by (q, p) = (x, 0), where x being the solution of Ax = b. Adding a damping term
γ to the equations of motion will make the solution spiral toward the minimum:

q̇ 0 1 q 0
(10.8) = + .
ṗ −A −γ p b
The damping is optimal whenever it is neither too small, which keeps the solu-
tion oscillating, nor too large such that the changes are slow. Hence the optimal
damping, together with the proper step size ∆t, and integration routine will drive
the system the quickest to the solution (x, 0). Such method falls into the class
10.1. ITERATIVE METHODS 185
of direction set methods, like the conjugate gradient method, where p defines the
direction, or improved gradient.
The conditioning of the problem can be carried out by adding a mass term to
the momenta, or direction, p:
(diagA)−1

q̇ 0 q 0
(10.9) = + ,
ṗ −A −γ p b
where the unit matrix is replaced by the reciprocals of the diagonal elements of A.
If the problem is close to diagonal, all the eigenvalues will be of the same order,
which allows for a faster iteration scheme. In the study of vibrational problems the
equations of motion contain the mass M and the stiffness K:
M−1

q̇ 0 q
(10.10) = .
ṗ −K 0 p
The typical frequency is given by:
√
(10.11) ω= KM−1 .
Hence, due to the conditioning with the inverse of the diagonal part of A, the time
scale is independent of the problem with all frequencies ω ≈ 2π. If the step size is
too large, a simple Euler integration, associated with standard iteration, will spiral
outward in the absence of damping.
Damping and step size are concepts which give intuitive insight into the iterative
procedure. Furthermore, these concepts apply to linear systems and nonlinear
systems alike. The method can even be illustrated with different material blocks,
or balls, in a bowl. The block will move down to the bottom. If the damping, or
friction, is too large, it will move very slowly. If the friction is too small, it will not
come to rest, at the bottom. See Figure 10.2. Furthermore, the results of numerical
time integration can be taken to our advantage to improve the convergence and
optimize the memory usage. Beside the theory of time integration, the theory of
optimal control can define a fast and smooth path to be taken as ideal for the
damping.
In principle the path should straighten during the iteration. If one tries to
straighten it to quickly, it might miss flat directions in the potential defined by
A. On the other hand if it straightens to slowly, the step size has to remain small
to follow the bends in the path. An active, i.e., predictor-corrector, approach to
damping and step size may speed up the convergence of the iterative method.
One of the nicest parts of a dynamical interpretation is the strict bounds on the
convergence. Even though one might overshoot the solution with a large step, the
combination of the potential and the kinetic energy 12 p2 in the directive vector will
decrease, yielding less conservative convergence criteria for a decreasing directive
vector p, than the potential alone would be.
For example, take a pendulum with a small damping. In the first swing, the
pendulum swings beyond the minimum and the potential energy increases again.
The increase in kinetic energy corresponds to an increase in the directive vector,
which speeds up the convergence. Although the potential energy does not decrease
monotonically, the total energy decreases monotonically all along the path. An
adaptive damping will detect the minimum and increase the damping accordingly.
momentum
position
Figure 10.2. The trajectories in phase space for different damp-

ing. The straightest trajectory is not the fastest, as the damping
is high. The ellipse indicate the area of constant energy, in which
a dissipative system will remain. A particular horizontal cross sec-
tion indicates the position. In special cases, the motion might take
the system beyond that position, yielding a larger potential energy.
10.2. Adaptive damping and step size

Adaptive damping and step size can arise from tracking a succession of steps:

q(ti ) q(ti+1 ) q(ti+2 )
(10.12) , , ,··· .
p(ti ) p(ti+1 ) p(ti+2 )
Its basic properties are the Euclidean invariants, invariant under coordinate trans-
formations, such as the length dii+1 = |q(ti+1 ) − q(ti )| and the angle θi+1 i
=
T
q(ti+1 ) q(ti ).
We use a simple Euler step for the time integration:
(10.13) p(ti ) = p(ti−1 ) − (Aq(ti−1 ) + γp(ti−1 ) − b)∆ti−1 .
A similar update step is used for q:
(10.14) q(ti ) = q(ti−1 ) + p(ti )∆ti .
The damping of γ∆ti = 1, setting the momentum, or directive vector, p = 0, cor-
responds to the steepest descent, or gradient, method, which is prone to numerical
oscillations, [Booth, 1966, Press et al., 1992] between the sides of small valleys, so
to say. At each step the directional, or momentum, vector p is set to zero. A very
small damping, γ ≈ 0, will lead to dynamical damping, where the solution oscillates
around the minimum x. Somewhere in the middle lies the optimal damping.
The optimal time step is the maximal time step which does not lead to nu-
merical instabilities. For a well-conditioned problem, the norm of the matrix A is
kAk ≈ 1, such that a convergent series arises if ∆t < 1. With the presence of the
directive momentum vector the iteration matrix B is:

0 1
(10.15) B = ∆t .
−A −γ
The eigenvectors of B have a large imaginary part. Furthermore, in the absence
of damping γ = 0, the norm may be smaller than one, but the system will not
10.2. ADAPTIVE DAMPING AND STEP SIZE 187
converge to a solution, but instead oscillate around the minimum. The stability
analysis of the iterative series is therefore more complicated. One can also view
the directive vector p as the weighted summation of successive steepest descent
gradients:
X kAk
(10.16) kBk ≈ (1 − γ∆t)j kAk < .
j
γ∆t
Hence to avoid erratic behavior, a conservative estimate of the step size ∆t is based
on the weighted summation:
kAk
(10.17) ∆t < .
γ
Not all methods developed for large, sparse systems take into account the suc-
cessive nature of the matrix, arising mainly from FEM and network systems. In
order to connect to coefficients of the vector q a large number n of An are required,
typically of half the size of such systems, in the case of one-dimensional models. For
higher dimensions of the underlying model, like plates and blocks, the connectivity
is higher.
The criteria for tracking a solution in a FEM or network based model, yields a
rather trivial result. At each step, the solution progresses a little further through
the system, hence the main perpendicular part is the new component added first
to p, and then to q. In a dynamical interpretation, it is the wave front of the
hyperbolic equation propagating throughout the system. We find that for small
damping the convergence of the solution starts to oscillate with the period related
to the wave velocity and size of the system.
At every iteration An b → An+1 b, the solution extends only one or two mesh
points further away from b. Hence the larger the mesh, the more iterations are
required only to fill the vector x. Hence, typically, the number of iterations is
proportional to n. This is the period for an optimal time-step. Therefore, the
maximal step size ∆t ≈ kAk−1 . The time step determines the degree of exploration
of the solution space, bounded by the energy of the initial guess, while the damping
causes the convergence. In practice we find little variation in the optimal damping
of a large range of problems. Possibly, convergence can be improved by increasing
the damping as the convergence is reached.
Adaptive schemes work best if there is a detailed understanding of the system
at hand. The best possible scheme is a single iteration run, where the number of
iterations equals the connectivity of the system. This could be considered the dis-
crete equivalent of the direct integration of the inverse A−1 , the discrete equivalent
of the Green function. This would be the optimal convergence available for a given
system. However, in this case the Green function is not known, but implicit in the
matrix A. For example, if the equation Ax = b represents the displacement x as
the result of the force b, for the stiffness matrix A, the linear equation is the in-
ternal force balance. In terms of the force, the system is usually better understood
than in terms of the displacement, which depends on the constitutive relations.
Force equations can be integrated. Consider the iteration process as a long line
of masses connected by springs. If one mass is pulled with a force b, all the neigh-
boring masses will counteract the force. This counteraction is again counteracted
by the next-neighbor masses, and so on. The process described here corresponds to
the conjugate gradient algorithm described in the next section. At every stage the
exact solution for in a subspace Kn for that particular set of variables is sought. The
use of damping will yield a slower convergence for the same number of iterations,
but less computations per iteration. The dynamical approach makes it possible to
combine convergence in one particular subspace with motion in another. The stiff-
ness matrix of an isolated object is singular. Methods based on the positivity of the
stiffness matrix may fail for particular boundary conditions. An explicit reference
to the dynamics and the force can resolve numerical problems whose origins lie in
the problem definition. These principles are explored further in Chapter 12.
The stiffness matrix A corresponding to a string of masses and springs is singu-
lar. The same constant displacement of all the masses will not change the indenta-
tion of the springs and consequently will not change the force. The inhomogeneous
term, the right-hand side of the equation, representing the applied force, should
therefore be balanced; the sum force is zero. Hence the force is balanced locally
from source to sink, from the position the force acts, to the position where it is
counter-acted. In the case of a system response, the force may also be counter-acted
by resistive forces, or by inertial forces caused be acceleration.
Hence the convergence and quality of the iterative solution can be expressed
in terms of physical principles, involving local force balance underlying a static
solution. The constant displacement e = (1, 1, 1, 1, · · · , 1, 1, 1) plays an important
role in this analysis, since for the static solution eT x = 0. Hence deviations from
the static solution during iteration, which we will call force “absorption,” represent
themselves as eT xi . In the dynamical interpretation of the iteration process, the
absorption is related to the inertial and damping forces.
Part of the force is locally balanced, the other force lines run through the whole
structure, till fixed boundaries. By the lack of knowledge of the end of the force
lines the adaptive method should keep track of the force “transmission” versus the
force “absorption.” Given the solution xi+1 of a step in the iterative scheme:
(10.18) Axi = xi+1 .
A constant vector e = (1, 1, · · · , 1, 1) corresponds to a complete transmission:
(10.19) Ae = eT A = 0 .
Hence the force absorption can be expressed as a projection on e. The absorption
Fabsorption is given by the projection on the translation vector:
eT xi+1
(10.20) Fabsorption = √ ,
N kxi+1 k
where N is the dimension of the system. The force balance and localization at the
inhomogeneous term b, the force input, is an independent test of the correctness
and convergence of the iterative solution. In the cases the force is not balanced; the
sum force is not zero, the momentum grows due to the acceleration of the system,
and the force absorption is associated with the inertial force.
In the absence of knowledge of directions in the system, all the translations
are concatenated into one; e. Such knowledge, on the other hand, would make it
possible to separate e into ex , ey , etc..
The force transmission T is given by the change from one iteration to the next:
kxi+1 − Axi k
(10.21) T = .
kxi+1 k
10.3. CONJUGATE GRADIENT METHOD 189
Typically, T represents the size of the wave front. The typical instabilities for a large
time step occur if the action and counter-action remain local. Behind the leading
wave front, filling the state vector successively following the actual connectivity of
the underlying grid, the state converges locally, with an exponent determining the
quality of the iterative scheme.
Problems described by potential like V = qT Aq, where the matrix A is given
by the Laplacian and flow problems with pressure boundary conditions are examples
of systems with a high transmission.
10.3. Conjugate gradient method

The standard numerical method closely related to the dynamical approach is
the conjugate gradient method. Here we give only a short overview of the method,
which consists of two element: the residual vector r, which is a measure for the
error of the approximation so far, and its dual v, which introduces a new direction
perpendicular to all the old ones in a particular inner product. [Hestenes and
Stiefel, 1952, Stoer and Bulirsch, 1980, Golub and Van Loan, 1996, Press et al.,
1992, Saad, 1995, Antia, 2002] In this approach the positive definite matrix A is
used as an inner product:
(10.22) hv, xiA = vT Ax .
The iteration procedure resembles a Gram-Schmidt orthogonalization with respect
to the inner-product h·, ·iA . The approach is similar to the Lanczos tridiagonaliza-
tion, using the Krylov space, described in section 2.12.
The solution x is expanded in residuals ri , each the remainder of the equation
at that stage: ri = Axi−1 − b, while the next residual vector vi in the dual space
is orthogonalized with respect to the previous duals v0 , v1 , · · · , vi−1 .
For each iteration xi the smallest deviation of kAx − bk in the subspace
span{x0 , x1 , x2 , · · · , xi−1 } is sought. The residual part ri = Axi − b of the vector
xi is projected onto the residual dual vi :
hvi , ri iA
(10.23) vi+1 = ri − vi .
hvi , vi iA
Therefore, two successive vectors are perpendicular: hvi+1 , vi iA = 0. The algo-
rithm generates an orthogonal set of dual vectors, onto which the equation Ax − b
can be projected. During the generation of this set, the coefficients αi of the ex-
pansion:
X X hvi , ri iA
(10.24) x= α i vi = vi ,
i i
hvi , vi iA
can be generated simultaneously. It leads to a set of hierarchical equations, most

economically expressed, after some calculations, [Stoer and Bulirsch, 1980] as:
rTi ri
(10.25) ri+1 = ri − Avi+1 ,
hvi+1 , vi+1 iA
while the same part is added to the solution vector:
rTi ri
(10.26) xi+1 = xi + vi+1 .
hvi+1 , vi+1 iA
Furthermore, the iterative generation of the dual vectors can be expressed as:
rTi ri
(10.27) vi+1 = ri + T
vi .
ri−1 ri−1
The algorithm resembles the leap-frog integration (see Section 3.4), where r i and
vi are differences in steps in the position: ∆x = xi − xi−1 and momentum space:
∆p = pi − pi−1 , respectively. Except for the precise value of the step size and
damping, the pair of equations read:
(10.28) xi+1 = xi + αi pi ,
and
(10.29) pi+1 = pi − αi (Axi − b) .
Instead of relying for an exact computation of αi , which allows for a result in the
number of steps identical to the dimension of the matrix, we use some physical
principles to estimate two different constants, one for the position space x and one
for the momentum space p. The number of operation per step decreases, but the
convergence will decrease as well, since we did not choose the optimal αi for the
particular Krylov space K(A, b).
10.4. Dissipative models

The criterion for success of an iterative method can be expressed in terms of
the ever decreasing error norm. For a stationary method, using only the solution
vector x, the convergence to the solution is expressed as:
(10.30) kAxi − bk > kAxi+1 − bk .
For a symmetric positive definite matrix A another criterion would be:
1 T 1
(10.31) xi Axi − xTi b > xTi+1 Axi+1 − xTi+1 b ,
2 2
since the problem can be reformulated using b0 = A−1 b as the same statement in
a weighted norm kx − b0 kA . This norm is used in the conjugate gradient method.
In the dynamical case, the energy, or Hamiltonian, serves as norm for the
convergence:
1 T 1
(10.32) H= p p + xT Ax − xT b .
2 2
The Hamiltonian is bounded from below if A is a positive definite matrix. The
quadratic term will always dominate the inhomogeneous, or linear, term for large
kxk. The damping γ > 0 will always decrease the energy in the system. The
energy is dissipated. Lyapunov generalized this property of the dynamics of physical
systems to a wider class of systems. [Lyapunov, 1992] It became an anchor for
analyzing the stability of motion.
The choice of the step size should be such that the integration will not fail.
Extreme accurate time-integration is not required, as the solution converges to a
unique stationary value. However, the typical step size, limited by ∆t < kAk−1 ,
will change due to the presence of the momentum p.
10.5. DIFFUSIVE-DYNAMICS ITERATION USING SUM CONVERGENCE 191
10.5. Diffusive-dynamics iteration using sum convergence

Another norm, based on physical systems, could be used to design an iteration
scheme. In the case of diffusive processes, the dynamics converges to the stationary
flow, given the inflow and outflow in terms of the inhomogeneous vector b. The
matrix A has, as for most partial differential equations, the null-space solution:
(10.33) A[constant vector] = Ae = 0 .
For example, for a constant temperature, throughout the system, there will be no
heat flow. Likewise, for a constant electric potential there will be no electric current
flow. The presence of the sources b will change that situation, however, the greater
part of the matrix yield zero for a constant vector.
If the iteration scheme above is applied to a problem with a null-space e, care
must be taken to keep the solution x perpendicular to the null-space. By reformulat-
ing the problem, associating it with a parabolic system rather than a Hamiltonian
hyperbolic problem with damping, the numerical stability can be made implicit.
The same static force solution can be treated as a conserved charge diffusion, with
a distributed inhomogeneous term.
If the rows and columns are reordered such that the mesh points closer to the
inflow are the first rows and columns, and the mesh points closer to the outflow are
the last row and columns, one can expect the state vector x to be in an approximate
ascending order. This will improve the convergence as we will see.
The symmetric matrix A can be decomposed into an upper-diagonal part Au ,
a lower-diagonal part Al , and a diagonal part Ad , such that A = Au + Al + Ad . A
triangular matrix, consisting of an upper or lower part and the diagonal part can
be solved directly. This is the basis for some standard iteration schemes, such as
Gauss-Seidel, [Press et al., 1992, Antia, 2002] which decompose the matrix in two
parts, which can be solved directly:
(10.34) A = A 1 + A2 ,
to use in an iteration scheme like:
(10.35) xi+1 = A−1
2 (b − A1 xi ) ,
where A2 should be invertible. As long as kA1 kkA−12 k < 1 the iteration should
converge under general conditions. For Gauss-Seidel A1 = Au or A1 = Al , and
A2 = Ad + Al or A2 = Ad + Au respectively. For Jacobi iteration, A1 = Au + Al ,
and A2 = Ad .
The iteration scheme can be seen as the Euler integration of the dynamical
equation:
(A−12 A1 − 1)
(10.36) ẋ = x + A−1
2 b .
∆t
If we use for A1 = Au − Al and for A2 = Ad − 2Au the iteration matrix A1
is skew-symmetric. A skew-symmetric matrix corresponds to a gyroscopic process
in dynamics. It redistributes xi , but leaves the sum of all the coefficients invariant.
The values of the coefficients are like the charge density of the system, while the sum
is the total conserved charge. Hence only the matrix A2 affects the charge. The
matrix A2 has typically only negative coefficients, and determines the maximal step
size. It would correspond for a one-dimensional second-order differential equation
F F
Figure 10.3. The one-dimensional Hertzian problem. The verti-

cal springs, connected to the fixed world, have a stiffness α, while
the horizontal spring have a unit stiffness.
to rewriting it into an integro-differential equation:

Z source
(10.37) ∂k2 f (z) + Lf (z) = g(z) → ∂k f (z) = dzk (g(z) − Lf (z)) ,
sink
where the parallel direction zk is the direction from sink to source as defined by
the sign of the inhomogeneous term g. See also Section 12.3. By rewriting the
differential equation to an integro-differential equation the null-space of the stiffness
matrix A no longer poses a numerical problem.
10.6. The string model

To illustrate and test the dynamical principles of iterative methods, we examine
a string problem in detail. It is similar to an one-dimensional Hertzian contact
problem. See Figure 10.3. Such a problem is typical for most stable systems, based
on partial differential equations. Also Helmholtz problems yield similar results,
for example, for the penetration of electromagnetic fields in a conductor. The
parameter α determines the typical penetration depth. The parameter-dependent
matrix problem is mapped back to differential equations with an analytical solution.
The numerical problem can be studied without this analytic result, however, it yield
more insight in the physical aspects of system and solution. For example, the force
transmission will vary from complete transmission for α = 0 to no transmission for
large α. The choice of α, for the numerical example, is in the intermediate regime.
Given an elastic string stretched by an applied, opposite force b, at both ends.

The string is also attached to a frame, with a relative spring tension α. The stable
solution corresponds to the equation Ax = b, where x is the strain, or displacement.
The matrix equation is:
 
−1 1 0 
−1

 1 −2 1 0   0 
 
 . .  
. 

(10.38) Ax =  0
 1 −2 1 0 .  x − αx = 
  ..  = b .
0 1 −2 1
 



  0 
.. .. .. .. 1
. . . .
The minimum number of iterations required grows linearly with the size of the
system, since the effects of the force at the ends has to propagate through the
string one step at the time. Multigrid methods are developed to speed up this
basic convergence problem, but we will consider a fixed size of the system.
The matrix problem is the discretized version of the differential equation:
∂ 2 x(z)
(10.39) − αx(z) = 0 ,
∂z 2
10.6. THE STRING MODEL 193
penetration small α
depth
displacement
large
α
distance
Figure 10.4. The normalized analytical results of the Hertzian

problem for different values for α. For small values of α the force
is “transmitted,” from one end to the other, with a linear dis-
placement. For large values of α the force penetrates only a small
distance into the material.
with the boundary conditions corresponding to opposite forces at both ends (b =

(−1, 0, 0, · · · , 0, 0, 1)):

∂x(z)
(10.40) − αx(z) = ±1 .
∂z z=±N/2
The solution to the continuous problem is (see also Figure 10.4):

√
sinh αz
(10.41) x(z) = √ √ √ .
α cosh αN/2 − α sinh αN/2
Due to the boundary conditions, responsible for the inhomogeneous term b, the
value at the boundary grows with each iteration. The gradient term propagates
the effect inwards.
The step size is usually the largest step size possible for a stable iteration.
Since the string problem turns into a wave equation problem with dispersion, by
adding the directive vector p, the step size should satisfy criteria that the time
step should lie in the forward characteristic cone, determined by the characteristic
equation, and such, by the norm of A, the characteristic cone depends only slightly
on the considerable damping, therefore the step size may be chosen fixed. These are
typical results from the time integration of hyperbolic partial differential equations.
[Lax and Wendroff, 1960, Morton and Mayers, 1994]
The damping, on the other hand, should vary with the problem and the stage in
the iteration. At first, the damping should be small, such that the solution expands
throughout the domain with the maximal, or wave, velocity. Later the damping
should increase to damp out oscillations in the solution. A small damping is easily
recognized in the solution by an oscillation in the error norm:
(10.42) εi = kxexact − xiterative (ti )k .
The smaller the parameter α the larger the penetration depth, and typically
the slower the convergence. The optimal value of γ ≈ 0.07 − 0.11 varies only slowly
over the range of parameter α, only when the problem turns into a surface problem,
RMS error
1e-05
70 steps
140 steps
210 steps
280 steps
350 steps
1e-10 420 steps
0 0.1 0.2 0.3

damping
Figure 10.5. The convergence to the exact result after different

numbers of steps, for a variety of constant damping. The constant
α = 0.003 is in the intermediate regime and the dimension is N =
100.
with a small penetration depth, the optimal damping may be substantially larger
γ ≈ 0.25. See Figure 10.5. Even smaller than the range of damping γ is the
optimal variation of damping throughout the problem. Typically, the damping
should increase as the iteration proceeds by 20% to 30%, to reach the solution the
fastest. The size of the vector kpk is a good indication how far quick the convergence
is, or how large the system is. The size will grow at first. Once the solution starts
to converge, the size will decrease with the error. The typical relation between the
parameter α, and the optimal damping is, for α < 0.02:
(10.43) γoptimal ≈ 0.07 + 10α .
As the number of iterations is several times the size of the problem, the domain of
convergence is a large region around the optimal damping.
The damping is related to the convergence speed; the quicker the solution
convergences, normally for larger values of α, the larger the damping should be.
Very large values of α > 0.1 yield numerical instabilities, as the penetration depth
is in the order of the mesh size. Problems ill-defined by the coarseness of the grid
will be less stable for large damping.
10.7. Constraints
Constraints in dynamical systems are usually nonlinear, since most linear rela-
tions can be implemented without problems, yielding a lower dimensional dynamical
problem. Rigid-body dynamics is well-known for its constraints, which can either
by holonomic or non-holonomic, a term coined by Hertz. [Gantmacher, 1970] Even-
tually we are interested in very large sets of constraints for FEM problems, but for
the moment we will look at two simple problems to investigate the implementation
of constraints as virtual dynamics. The first case is in the absent of real dynamics.
The constraints determine the state almost uniquely for the Stewart platform, as
10.8. STEWART PLATFORM 195
x1
x3
x2
z3 z2
z1
Figure 10.6. The Stewart platform, actuated by six trusses, con-

nected to a fixed ground.
discrete local minima are found. In the mathematical pendulum (Section 10.10)
we investigate the decoupling of the virtual, or constraint, dynamics and the real
dynamics, and the influence of the virtual dynamics on the accuracy of the solution.
The generic methods described here are all local. Around the present position
in phase space we examine the constraint direction and the direction of the dynam-
ics. To first order, corresponding to local rectangular coordinates, we are able to
decouple the constraint direction, in which the virtual constraint forces act, and
the direction of dynamics, in which the system evolves. The nonlinear constraints
do not always allow for an analytical estimate of the accuracy and decoupling,
however, numerical results are very good.
Eventually these methods are meant to be applied to FEM systems with arbi-
trary boundary conditions. Many types of boundary conditions cannot be imple-
mented as input, but must be implemented as constraints, as, for example, fixed
positions at the boundary. It is possible to reduce the FEM model such that the
fixed boundary values are excluded from the model. However, that means the
resulting FEM model of the system can not be used again is a component of a
larger system in a general setting. For a general setting constrained motion and
the associated constraint force are the pair along which a system may interact with
neighboring components, or control.
10.8. Stewart platform

In analytical mechanics Cartesian coordinates are traded in for a smaller set of
generalized coordinates. For analytical calculations the smaller set of generalized
coordinates had clear advantages. In a number of cases it allows for an exact,
closed, solution of an isolated system. However, in many cases such a solution is
not sought. The complexity of a combined system, consisting of many subsystems,
already described in generalized coordinates, may be overwhelming.
In this chapter we advocated the approach of not performing any reduction,
using the Cartesian coordinates, and the Euclidean invariants. Any constraint can
be solved dynamically. Moreover, during simulation convergence to the constraint
surface may be implemented simultaneously with forward simulation.
As a simple example we discuss the Stewart platform, consisting of a plate
controlled by six extendable trusses, connected at each end with a ball joint. The
plate can be represented by the three joint positions, x1 , x2 , and x3 . See Figure 10.6.
The geometry and inertia of the plate depends solely of these coordinates. For
example, the kinetic energy of the plate can be calculated by:
1
Z X
(10.44) T = ρẋ2 (x1 , x2 , x3 ) = Mij ẋi ẋj .
2 ij
Two simple examples. First, consider the platform mass M located at the center:
x = 13 (x1 + x2 + x3 ), the mass matrix is therefore:
 
1 1 1
M
(10.45) M= 1 1 1  .
9
1 1 1
Secondly, consider the mass to be evenly distributed on a triangle spanned by the
three corners xi , where the triangle has three equal lengths. The center-of-mass
position x is given by the interpolation between the three points:
Z √3 Z u/√3 X
(10.46) x= du √ dv xi fi (u, v) ,
0 −u/ 3
where the weight function,

(1 − √u3 )
 
(10.47) f (u, v) =  12 ( √u3 − v)  ,

 
1 √ u
2 ( 3 + v)
the kinetic energy is the integral over the distributed mass:

1 2 X 1
Z Z
(10.48) ρẋ = ẋi ẋj ρfi (u, v)fj (u, v) .
2 ij
2
The weight functions fi (u, v) determine the mass-matrix coefficients, independent

of orientation or velocity. The weight functions satisfy:
(10.49) f1 (u, v) + f2 (u, v) + f3 (u, v) = 1 ,
such that a√collective translation is correctly
√ accounted for. The total area of the
triangle is 3. The mass matrix is ( 3ρ = M ):
 
2 1 1
M
(10.50) M= 1 2 1  .
12
1 1 2
Since the collective motion of all the nodes is associated with the total mass, the
elements of the mass matrix should always sum to the total mass M .
The platform itself can be represented by three stiff trusses V (xi − xj ) between
each pair of corners. The length at rest is normalized to 2, the same as in the case
of the weight functions fi (u, v) above. The trusses between the platform and the
floor have variable lengths lij , where the truss labelled ij is located at platform
corner xi and floor corner zj , such that i 6= j.
For all of these trusses we assume a simple elastic energy function:
1
(10.51) Vl (x − y) = k((x − y)2 − l2 )2 .
2
Hence, using the Hamiltonian consisting of all Vl ’s, the static configuration is the
minimal elastic energy. In the case the elastic energy is zero, the configuration is
10.9. VIRTUAL DYNAMICS 197
virtual dynamics
real
dynamics
constraint
surface
Figure 10.7. The virtual dynamics with the appropriate damping

converges to the real dynamics, which lies in the constraint surface.
allowed, i.e., all the constraints can be satisfied simultaneously. The mass can be
taken into account, even masses of each of the trusses can be added simply with
the mass matrix:

1 2 1
(10.52) Mtruss = Mtruss .
6 1 2
In order to converge to the static solution, the stiffness must be much larger than
the body forces present, for example, the gravity force. Furthermore, in a dynamical
context damping can be added which affects only the trusses. Using the standard
Rayleigh terms:
1
(10.53) R(ẋ − ẏ) = γ(ẋ − ẏ)2 ,
2
which yields only damping along the trusses. The motion of the truss as a whole is
not affected by the damping.
However, in cases with real dynamics and virtual dynamics from constraints,
the explicit configuration dependence should be put into the Rayleigh term to
guarantee the orthogonality of the damping to the real dynamics:
γ
(10.54) R(ẋ − ẏ, x − y) = 2 ((ẋ − ẏ)T (x − y))2 .
2l
Now the damping is explicitly in the direction of the truss. The damping no longer
relies on a small relative velocity ẋ − ẏ, which yield damping only in the direction
of the constraint due to the local orthogonality of virtual constraint dynamics and
real dynamics. This principle can be extended to more general constraints, as we
will see in a later section. See Figure 10.7.
The truss Hamiltonian with the truss Rayleigh term do not necessarily have
to be virtual terms, with very high stiffness, to yield iterative equations, which
converge to the constraint surface. In reality there are no pure rigid elements, all
elements have some elasticity, the truss elasticity Vl may be used to examine the
vibrational modes of the Stewart platform.
10.9. Virtual dynamics

The Cartesian Hamiltonian dynamics has only constant mass matrices. In the
symmetric case, under exchange of platform nodes, the mass matrices even have a
discrete steps
true trajectory
virtual step
time corrector step

step
Figure 10.8. The virtual dynamics combines the free dynamics

and the constraint in a single integration step. The corrector step
is formulated as damped, or dissipative, dynamics.
very simple form, which is easily inverted:

 −1
α+β β β
(10.55) M−1 =  β α+β β  .
β β α+β
α−1 − b
 
−b −b
(10.56) = −b α−1 − b −b  ,
−b −b α−1 − b
where b = β/(α(α + 3β)). The same matrix occurs for arbitrary dimensions Mα,β ,
such as in the case of a truss where the matrix is two-dimensional. The inverse
is has the same form, except the only the 3 is replaced by the dimension n of the
matrix in b = β/(α(α + nβ)). Hence in the case of the triangular homogeneous
platform the inverse of the mass matrix is:
 
3 −1 −1
3
(10.57) M−1 =  −1 3 1  .
M
−1 −1 3
The equations of motion for the position and momentum vector concatenated
from all positions and momenta of each point are given by:
M−1 p

ẋ
(10.58) = ,
ṗ −MRM−1 p − ∂x V
where the Rayleigh matrix R is constructed from the Rayleigh term R through:
1 T 1X
(10.59) ẋ Rẋ = Rij ẋi ẋj = R .
2 2 ij
The position-actuated Stewart platform has no real dynamics. The orientation of

the platform is fixed by the length of all six trusses lij , fixing the six degrees of
freedom of the platform. Hence there is only virtual dynamics; the convergence to
this solution, using the truss energies Vl and damping R.
In the case of force actuation, the truss lengths are the result of the system
response to the applied forces. There are at least two approaches to this problem.
First, use the configuration information to align the forces parallel to the trusses,
and add the corresponding applied resultant force to the Hamilton equations. Sec-
ond, use the truss length lij parameter as variable: Rather than assuming a rigid
10.10. PENDULUM EXAMPLE 199
truss with a high stiffness k, the truss can be a spring with a stiffness corresponding
to the length and actuation force:
Factuator
(10.60) k= ,
l
where the elastic energy of the truss is:
1
(10.61) Eelastic = kl2 .
2
The Hamiltonian can be updated in many different ways, to account for the actu-
ation.
In this study we will take the approach of updating the length, of the stiff truss,
according to the force. See Figure 10.8. Given x and y, the truss force is :
(10.62) F = −∂x Vl (x − y) = −2k((x − y)2 − l2 )|x − y| .
Hence, by adjusting l, the force F can be set to the actuator force:
F
(10.63) l2 = + (x − y)2 .
2k|x − y|
At each iteration step the truss lengths should be updated. Force actuation does
not guarantee stability, also not in the updating case. The power transfer is given
by the change of the elastic-truss energy. For stiff trusses the actual length change
is limited. However, the consequences for the dynamics of the platform may be
large, due to the large forces.
10.10. Pendulum example

In order to test the virtual dynamics, we consider the mathematical pendulum.
It consists of point mass in gravity rotating at a fixed length around the origin:

ẋ p
(10.64) = ,
ṗ −ey
with the constraint |x| = 1. The mass, gravity, and length are all set to unity.
The constraint can be eliminated to reduce the order of the system, retaining
a single angle θ:
(10.65) θ̈ = − sin θ ,
which is an anharmonic oscillator.
In the case of virtual dynamics, the truss (Eq. 10.51) connecting the mass to
the origin is taken to have a large stiffness k and a damping γ:

ẋ p
(10.66) = .
ṗ −γ(pT x)x − ey − k(x2 − 1)x
The elastic and damping forces due to the truss all act in the direction of the truss
x. In a decomposition of the coordinates in an angular and a radial coordinate
can be achieved by projecting the equations onto x and ez × x. The equation of
motion for the angular coordinate θ is the same as above, except √ for the presence
of the length x2 in the force term. The radial coordinate r = x2 yields a damped
nonlinear oscillator around r = 1:

ṙ p
(10.67) = ,
ṗ −pγr2 + cos θ − k(r 2 − 1)r + rθ̇2
-0.06
energy
-0.06005
-0.0601
0 10 20 30 40 50
time
Figure 10.9. The energy of the mathematical pendulum, as a

constrained two-dimensional system. The larger deviations are for
the smaller stiffness of k = 100, the larger stiffness is k = 500.
The damping is small, γ = 0.3, the slightly larger damping γ = 1
already damps out the oscillations. The larger oscillations are the
result of the swing. If the truss energy is taken into account, the
energy is constant, up to integration errors.
where cos θ is the gravity component in the direction of the truss, and r θ̇2 is the
centrifugal force.
In the static case p = θ̇ = 0, the truss is extended due to the gravitational
force:
1
(10.68) r =1+ .
2k
In the dynamic case, the truss dynamics is driven by the pendulum swing for small
amplitudes θ0 where: θ(t) = θ0 sin t. The force on the constrained, or virtual,
dynamics is now time-dependent, with a maximum at θ = 0:
(10.69) r̈ = −ṙγr 2 − k(r2 − 1)r + cos θ + rθ02 cos2 t .
The driving force will cause damping of the pendulum motion, due to the damping
forces in the constraint dynamics. However, this damping is small after the original
motion around the constraint is damped, and only the driving force remains. The
time-dependence of the driving force, the pendulum frequency, is much slower than
the frequency of the truss. Furthermore, the amplitude is small for large virtual
stiffness of the truss. Hence, the combination of the small amplitude and the slow
times reduces the virtual damping. Numerical investigations show that the energy
loss of the pendulum undergoes a sharp transition for stiffnesses of the order of
k ≈ 150. For a stiffness of k = 100 the pendulum looses about a fraction of
10.11. MIXING REAL AND VIRTUAL PARAMETERS 201
3.0 10−6 of its energy per swing. For k = 200 the losses are within the numerical
precision of the leap-frog integration method (10−10 ). The integration has been
carried out for tens of thousands of pendulum periods. Further improvements are
possible as we will see. See Figure 10.9.
10.11. Mixing real and virtual parameters

In the case of the Stewart platform, in the absence of real dynamics, the masses
have little meaning. The constraints determine the state fully. Inertial forces are
therefore irrelevant. However, in order to yield the fastest convergence either the
time-step, or the stiffness, should be adapted, in conjunction with the physical
mass. For the iterative methods, typically the time-step is adapted, to match the
inverse norm of the A-matrix. In physical systems, the mass and the time-step
are given by the phenomena under investigation. The stiffness and the damping
of the virtual dynamics should be made as large as possible to yield the fastest
convergence to the constraint surface.
The virtual stiffness alone does not determine the forces, the expected displace-
ments play also a role. For nonlinear virtual dynamics the distance to the origin
would be a reasonable measure for the size of the displacements. The degree of
connectivity, for example the two trusses at each joint of the platform, also yield a
factor in the bound on the forces. In the case of a discretized PDE, the dimension
plays the role of connectivity. A typical conservative estimate, based on numerical
experiments, for the maximal stiffness would be:
0.1m
(10.70) k≈ ,
(∆t)2
where ∆t is the step size, and m the total mass of the subsystem in which the
stiffness occurs. The damping may be a little larger:
0.15m
(10.71) γ≈ .
(∆t)2
One would like the state to take half to a quarter step toward the minimum in a
single time-step. The typical highest frequency of a harmonic oscillator would be:
r
2π nk
(10.72) ωmax = ≈ ,
Tmin m
where n is the connectivity; the number of springs attached to the mass m, and
Tmin the smallest period.
However, such a large stiffnesses rely on the stability of the integration method.
In many cases, the generated forces yield disproportionally large intermediate veloc-
ities, with respect to the relevant distances. The problem is therefore ill-conditioned,
with large stiffnesses and normal masses. For a generic integration method, the
stiffnesses are better approximated reciprocal to the step size:
0.1m
(10.73) k≈ ,
∆t
and a similar approximation for the damping. In the case of virtual dynamics
without real dynamics, the mass serves only the convergence. It can also be scaled
by the step size, to yield a well-conditioned problem.
10.12. General constraints

The example of the mathematical pendulum (Section 10.10) showed that the
constraint dynamics and the real dynamics is not fully decoupled. For particular
cases, with analytic results, a full decoupling might be feasible. We, however,
focus our attention on schemes that can be implemented numerically for general
constraints.
Given the constraint surfaces, the virtual dynamics, i.e., the dynamics perpen-
dicular to the constraint surface can be minimized. The potential dependence on
these perpendicular direction are removed by cross terms of real and virtual forces.
The effects of the curvature of the constraint surface, leading to typical virtual
inertial forces are more difficult to compensate. However, the curvature of the con-
straint surface may be used to adapt the perpendicular mass, associated to inertial
forces perpendicular to the constraint surface, in the system. The inertia associated
with the virtual dynamics can be used to minimize the perpendicular motion.
Given the Hamiltonian H0 (q, p) where q, p ∈ Rn and the constraints χj (q) = 0
for j ∈ {1, 2, 3, · · · , r}. The augmented Hamiltonian is:
r
X kj
(10.74) H0 (q, p) → H = H0 + Hc = H0 (q, p) + χj (q)2 .
j=1
2
Each constraint is added as a spring with the rest position χj = 0. The spring
stiffness kj is to be optimized for the simulation.
The augmented Rayleigh term is given by:
γX T γX
(10.75) R= (q̇ ∇χj )2 = (δp H T ∇χj )2 .
2 j 2 j
In the Hamiltonian formulation the Rayleigh term must be expressed in terms of

the momenta. The velocity is given as the variational derivative of the Hamiltonian.
The equations of motion are the Hamilton equations of motion with the damping
and the applied forces on the right-hand side:

q̇ δp H 0
(10.76) − = .
ṗ −δq H −δp R + Fapplied
The typical deviation from the constraint surface χj (q) = 0 is given by the vir-
tual force; the force perpendicular to the constraint surface. The direction perpen-
dicular to the constraint surface is given by the gradient of the constraint function
∇χj :
X ∇χj
(10.77) Fconstraint = −δq Hc = 2
∇χTj (δq H0 + ṗ − Fapplied ) .
j
|∇χ j |
The first term in between brackets can be removed by adding a gradient constraint
term Hg to the complete Hamiltonian H. This is the potential-energy dependence of
the original Hamiltonian H0 perpendicular to the constraint surface. The correction
term Hg to be added to the Hamiltonian compensates the virtual potential-energy
dependence:
X (∇χj )T δq H0
(10.78) Hg = − k j χj .
j
|∇χj |2
10.13. REMARKS ON LUMPED TO FINITE 203
constraint function χ
potential H0
2
constraint potential χ
gradient constraint
virtual Hamiltonian without gradient term
full virtual Hamiltonian H
constraint position
Figure 10.10. The potential perpendicular to the constraint di-

rection, with the different additions to the constraint Hamiltonian.
.
Furthermore, the applied force Fapplied is easily projected perpendicular to the

constraint surface:
X (∇χj )T Fapplied
(10.79) Fapplied → Fapplied − ∇χj .
j
|∇χj |2
See Figure 10.10.

Only the inertial constraint force (∇χj )T ṗ is not that easily compensated for,
standing in the way a complete decoupling of real and virtual dynamics. The inertial
constraint force is due to the curvature of the constraint surface. The curvature
causes the momenta to change direction. If the mass can be compensated for in
the direction of the constraint ∇χj , the inertial constraint force will also vanish.
However, such terms may give rise to instabilities, since massless objects tend to
oscillate fast.
10.13. Remarks on lumped to finite

In this chapter we treated lumped objects, such as trusses and point masses,
in nonlinear dynamical models. However, the methods developed here to study
dynamics can be applied to large nonlinear models. The construction of such finite
elements is investigated in the next chapter. The nonlinearity arises through the
consistency. The appropriate degrees of freedom are assigned to the internal dy-
namics and the general boundary conditions. Whether these boundary conditions
give rise to dynamics or constraint relations depends on the general setting. Even-
tually, we known that the simulation of such models can be carried out with the
methods described in this chapter.
CHAPTER 11
Nonlinear FEM
Don’t let them sell you a Turing machine without the tape.
[D. van Dalen, Filosofische grondslagen van de wiskunde]
Nonlinear FEM [Oden, 1972, Hughes, 1987] arises from two sources. First, the
constitutive relations might be nonlinear, as is often the case, e.g., in thermody-
namical and biological systems. In many other cases a linear constitutive relation is
used, valid or not, since it simplifies the methods to solve such problems explicitly.
In thermodynamical and biological systems these linear relations are known to fail,
For example, the pressure is approximately inversely proportional to the volume,
rather than linear. The large molecules in biology do not only have particular
chemical characteristics but also rheological and mechanical characteristics.
Secondly, the states considered may be far from the equilibrium state, or rest-
state, such that the linear theory, appropriate around the rest state, no longer
suffices. Beam models are well-known examples of the latter. The linear Euler-
Bernoulli equation (Eq. 6.5) yield correct results only for small deviations, such that
even the far end of the beam, with the largest deviation still points approximately
in the original direction.
A proper nonlinear model of a beam would take each segment or part indepen-
dently, and yield a result independent of the overall orientation. As a component, a
model can be part of a moving or rotating machine. These global motions will cause
inertial effects which must be described with the center-of-mass and moments-of-
inertia coordinates, which are not a priori given in a flexible body. The overall
orientation can be defined as the collective motion which will not affect the elastic,
or internal energy.
The orientation can be given by a rotation R and a translation a. A proper
invariant model of a beam with the position x(s) as function of the reference coor-
dinate s would have an elastic energy independent of the orientation, that is:
(11.1) Eelastic (x(s)) = Eelastic (Rx(s) + a) ,
for all rotations R and translations a.

The corresponding model would necessarily be nonlinear, if the orientation
needs to be accounted for too. The change of orientation (Ṙ, ȧ) is relevant for the
kinetic energy of the beam. A finite-dimensional, global part F of the distributed
force f (s); a finite number of components associated with the center-of-mass and the
moments of inertia, will only change the momentum and the angular momentum
of the beam.
The potential energy, such as elastic energy, is given by terms in the Hamilton-
ian which contain partial differential operators. Inherently in these equations are
the typical independence of translation, constant changes, and other invariances,
205
206 11. NONLINEAR FEM
kinetic potential
energy energy
element
network
Figure 11.1. Several element Hamiltonians make up one for one

element, which is combined in a network by variable collocation.
which give rise to constants of motion, and conserved quantities. The finite ele-
ment approximation should retain these qualities as part of the consistency of the
FEM approximation of the system. The most relevant conserved quantity is the
energy, as it will yield a stable dynamical FEM model, even for the most abstruse
nonlinearities.
In most cases the configuration state is the start of the analysis. In the case
of systems described by partial differential equations it is difficult to escape this
starting point as it closely ties the potential energy to the geometry of the rest
state, or reference state, i.e., the configuration. This is not necessarily an unwanted
starting point, as the kinetic energy is also closely tied to the configuration.
Combining the kinetic energy and the potential energy into a single Hamil-
tonian for an element gives rise to the port-Hamiltonian, due to the fact that the
kinetic energy depends on global orientation and the potential energy does not. See
Figure 11.1. The momentum and angular momentum are constants of motion for
an isolated model, but as a component for a larger system, the states associated
with the constants of motion are to be reassigned input via port variables. See
Section 7.8. The kinetic energy depends on the change of configuration. Hence the
typical variables involved in the potential energy and the kinetic energy are related
by a differential operator D. The null-space of this operator is to be augmented
with port variables.
For most physical systems similar properties arise, although the relation with
the configuration is less stringent. For example charge conservation in electromag-
netic systems is a conservation law, which leads to a total divergence as equation of
motion. The boundary conditions of the partial differential equation at the basis of
the potential energy are the missing variables. These variables are fixed by either
input-output considerations, or by connecting elements together in a larger grid
and thus reducing the number of variables to a well-defined set.
For rigid bodies the center-of-mass and the moments-of-inertia are static inertia
relations between the overall orientation and the associated kinetic energy. For a
deformable body, with elastic energy, this is no longer the case. The shape changes
with the motion, and causes interaction between the global motion and internal
11. NONLINEAR FEM 207
motion. The boundary conditions, causing both global motion and deformation
are therefore not easily decomposed.
From a heuristic viewpoint it is much easier to think of the boundary conditions
as global conditions, which restrict the changes in orientation, (Ṙ, ȧ) in Eq. 11.1,
or the changes of free motion associated with momentum and angular momen-
tum. However, such boundary conditions are distributed forces Fk (z), changing
the center-of-mass velocity ȧ and torques τ , with a change in rotation Ṙ as re-
sponse, which are singled out from the boundary forces F(z) through projections
as described in Section 7.2:
R 0
dz F(z 0 )
(11.2) Fk (z) = ρ(z) R 0 0 .
dz ρ(z )
The torques are also part of the distributed force, F(z). The torque τ which only
changes the overall orientation R around the rotation axis n, without internal
strain, is given by:
Z −1 Z
(11.3) τ = (ρz × n) dz”z” ⊗ z”ρ(z”) dz 0 z0 × F(z 0 ) ,
where n is the rotation axis at z = 0, and the inverse of the middle integral is the
variable inverse of the moment-of-inertia matrix J:
Z
(11.4) J = dzz ⊗ zρ(z) .
A strained object will deform and the density ρ(z) will vary with the force. Hence
the projections do not only depend on the applied force F, but on the object
response as well. Lumping together, or projecting, the distributed boundary force
F(z), to a single, but distributed force Fk (z) and torque τ (z) these forces are
associated with the inertia, from the mass distribution, ρ(z), of the object. Together
with the change in orientation, a combination of the translation velocity ȧ and the
rotation velocity Ṙ, the center-of-mass force Fk and the center-of-mass torque τ ,
form the typical power-bond pairs. [Paynter, 1961]
If the boundary force is separated in multiple lumped parts, additional lumped
variables, besides the center-of-mass variables, typically associated with inertia and
elasticity may arise. In the transmission line, and the beam models, one typically
thinks of end-point variables, where even combinations correspond to some mean, or
global, degree of freedom, similar to the center-of-mass, and the odd combination is
some internal degree of freedom, typically associated with its dominant aspect; the
elasticity or capacitance, although it also has inertial properties being a continuous
system.
This chapter builds upon earlier work. [Synge, 1957, Lanczos, 1961] In the early
days of numerical computing attempts to link finite-dimensional representations of
function spaces and the operators thereon to the infinite-dimensional physical spaces
and the differential operators played a larger role than they do nowadays. Repre-
sentations of operators, other than the ones that appear in the defining functional,
say the energy norm, are not considered essential. However, physical principles,
such as conservation laws, are written down as auxiliary differential operators, like
the finite-dimensional representation of flow or charge. Hence to construct together
with the finite-dimensional representation of the problem at hand, the underlying
differential operators, will give further insight in the physics, and may give one a
fighting chance to preserve physical principles.
Conservation laws should also be valid for the discretized models. For example,
take a diffusive problem and its discretization:
(11.5) f˙(z) = ∆f (z) ⇒ ẋ = Ax .
The formulation of the diffusive problem arises from some conservation law. Many
physical chemists [Bird et al., 1960, Sandler, 1977] dealing with such systems will
identify a conservation law, such as the continuity equation:
(11.6) f˙(z) = −∇ · J(z) ,
and a closure relation of the flow J in terms of the state f :
(11.7) J = −D(∇π(f )) ,
where π(f ) is the pressure function, and D(·) the diffusion function, related to
the free path length or a particle that diffuses. Only if the discretization A of the
Laplacian ∆ is constructed from the divergence ∇· and the gradient operator ∇, it is
easy to reconstruct the conservation law, eq. 11.6, with the diffusion equation 11.5,
also for the case of a nonlinear pressure function or a non-isotropic geometry with
a tensorial D.
Hence, we wish to construct consistent finite-dimensional representations of the
differential operators, and not just for the diffusive example above, that appear in
the models of physical systems and preserve from the start as much of their proper-
ties as possible, such that physical principles are retained in the finite-dimensional
systems.
In this chapter we will start with the spline approximation for each element of
a beam, showing that it suffers from two problems. First, the elastic energy of the
beam is not only given by the energy of the segment, but the boundary between
two segments may yield an infinite contribution to the energy, which can only
be solved by introducing constraints across the segments. Secondly, the standard
spline approximation corresponds to the Euler-Bernoulli beam model, which uses
explicitly a coordinate axis as rest position, with perpendicular displacements. This
is an approach which is not easily extended to an orientation-free description. In the
next section we show how these two problems can be cured, using invariant lines.
The boundary energy is zero and the segment energy is invariant under rotations
and translations.
In the following section we consider explicitly the partial differential operator,
the curl operator ∇×, and a polynomial approximation of the continuous states on
a three-dimensional cube. The counting of variables, boundary conditions, and the
null-space are the basis of constructing a consistent approximation of the continuous
model. The port-Hamiltonians elements of a triangular plate and electrodynamics
on the simplex are constructed.
11.1. The Euler-Bernoulli spline revisited

Before considering general element Hamiltonians, and the construction of the
port Hamiltonians we study the simple example of the Euler-Bernoulli beam and its
associated approximation in terms of piece-wise cubic splines: (z ∈ [zc − δ, zc + δ] ⊂
R)
(11.8) f (z) = c0 + c1 (z − zc ) + c2 (z − zc )2 + c3 (z − zc )3 = cT z i ,
11.1. THE EULER-BERNOULLI SPLINE REVISITED 209
Fi−1
Fi Fi+1
Fi+2 Fi+3
fi−1 fi
fi+1
f
i+2 fi+3
δ δ
z i−1 z i z z i+1 z i+2 z i+3

c
Figure 11.2. The spline consists of segments of length 2δ. The

center of each segment zc is used for the expansion of the polyno-
mial.
where zc is the center of the segment. See Figure 11.2, and also Sections 8.2 and
6.1.
The cubic spline depends on four variables, for which we can make a convenient
choice. These four variables can be paired, such that two variables appear on the
left-hand side of the domain and two on the right-hand side. The typical variables
one could use are the values and derivatives at the ends fi−1 = f (zi−1 ), fi = f (zi )
0
and fi−1 , fi0 . The map from these variables to the coefficients of the polynomial on
a domain z ∈ [zi−1 , zi ] with the center zc = (zi−1 + zi )/2 of length 2δ = zi − zi−1
is:
−1 
1 −δ δ 2 −δ 3
 
fi−1
 0 −1 2δ −3δ 2   fi−1 0 
(11.9) c= 1 δ δ2 3
   = Tf1 ,
δ   fi 
0 1 2δ 3δ 2 fi0
where f1 denotes the vector of variables of a single segment. It is comprises of
the left-end value and the derivative and the right-end value and derivative. Since
the elastic energy is given by the second derivative of the displacement, the elastic
energy is a quadratic function of the end point variables:
 
Z zi 0 0 0 0
κ κ  0 0 0 0 
(11.10) Eelastic = dz (∂z2 f (z))2 = f T TT   Tf .
zi−1 2 2  0 0 8δ 0 
0 0 0 24δ 3
Hence the elastic energy is given by a two-dimensional subspace of the configuration
state space fi . In the space of the four variables there is a two-dimensional subspace
which leaves the elastic energy unchanged. The two-dimensional null-space consists
of the piece-wise linear parts.
The concatenation of piece-wise linear segments will have kinks between them,
which yield infinite boundary energy. See Section 11.6. We will show this below.
The variables of the spline, the polynomial coefficients on each of the segments
define a smooth curve on each segment, but the problems occur at the boundaries.
In order to solve this, constraint relations among the coefficients on neighboring
segments arise. In the next section we will see that the common variables at the
boundaries, i.e., the port variables, are much more natural to define the shape of
the beam.
The spline approximation of the Euler-Bernoulli beam has two further con-
straints, in the second and third derivative across the successive segments. If the
second derivative does not match, it would yield a kink between to successive seg-
ments with an infinite bending energy. The value of the third derivative across the
segment boundary is associated with a local force Fi applied at zi . Hence from the
four variables c on each segment only one remains. It can be seen as the position
fi given by the local force Fi . Since these two relations are across the segment
boundaries, they link all the segments together, into a collective object. In some
cases it might be possible to find the coefficients successively, from one segment to
the next. But in many cases the problem has to be solved as a whole, in terms of
constraints.
Given the coefficients, these derivatives are easily constructed:
fi 1 ±δ δ 2 ±δ 3
   
 fi0   0 1 ±2δ 3δ 2 
(11.11)  (2)  =   c = Q± c ,
 fi   0 0 2 ±6δ 
(3) 0 0 0 6
fi
where the sign depends on the evaluation on the left-end or right-end of the segment.
The third and fourth row of the matrix Q± T are the basis of the constraint
relations. The second derivative vanishes at zi if:
(0, 0, 1, 0) Q+ Tf1i − Q− Tf1i+1 = 0 .

(11.12)
The third derivative yields the force Fi :
Fi
(0, 0, 0, 1) Q+ Tf1i − Q− Tf1i+1 = −

(11.13) .
κ
In this case it is more convenient to have a two-segment vector, corresponding to
larger stencils as described in Section 8.4:
 
fi−1
 f0 
 i−1 
 fi 
(11.14) f2 = 
 fi0  ,

 
 fi+1 
0
fi+1
such that the constraints are expressed as vector products χTc f2i = 0 and χTf f2i = Fi .
The constraint vectors χc and χf are versatile. For example, starting from
0
the left, the successive coefficients fi+1 and fi+1 can be determined from the four
preceding coefficients f1 and the force Fi . The algorithm is equivalent to the inte-
gration scheme of the fourth-order differential equation, which requires five points
to determine the sixth:
F (z)
(11.15) ∂z4 f (z) = .
κ(z)
See Sections 8.4 and 3.6. If integration is possible, it is a fast method to deter-
mine the solution. Not in all cases constraints can be integrated, so to say, from
starting at one end and successively solving the equations to the other end. Sys-
tems integrable in such a manner we call hierarchical systems, as the equations can
11.2. INVARIANT LINES 211
be ordered hierarchically, like an upper or lower triangular matrix equation. In

the case of the Euler-Bernoulli beam, the boundary conditions are partially set at
both ends. Neither end suffices to solve the system completely. The corresponding
matrix equation is not triangular, but banded, where the dimension dimf2 = 6
determines the width of the band.
However, constraints can also be implemented dynamically, similar to iterative
methods in numerical analysis. See Chapter 10. A large, instead of infinite elastic
energy can be associated with the constraint. A Hamiltonian with such elastic
energy given by:
X kc X kf
(11.16) Econstraint = (χTc f2i )2 + (χTf f2i − Fi )2 ,
i
2 i
2
the first term is the constraint energy, the second term the applied force, also
implemented as constraint. With the appropriate damping, it will converge to the
constraint solution with Econstraint = 0, if possible. In many cases a large constraint
problem is converted to an optimization problem.
11.2. Invariant lines

Above we have seen that splines suffer from at least two problems. First, they
are not invariant under rotations, as it is tied to the z and f direction for the
reference coordinate and the displacement. A rotated beam will have a different
energy from the original beam. Secondly, the matching of the segments requires a
constraint equation, such that the elastic energy of the boundary remains finite, and
automatically zero since the energy is a quadratic function. An invariant function
depends only on invariant vector products. The beam is defined by positions of the
nodes xi and directional vectors at the same node vi . The position along the curve
is given by the reference coordinates s. The lowest order polynomial which gives
an initial direction vi−1 , a final direction vi , and a translation xi − xi−1 is a third
order polynomial. The position s ∈ [0, 1] along the segment is given by:
(11.17) xi (s) = xi−1 + (xi − xi−1 )(3s2 − 2s3 ) + vi−1 s(1 − s)2 + vi s2 (s − 1) .
Hence a smooth curve between xi−1 and xi depends on the tangential directions
vi−1 and vi at beginning and end. The size of the tangential vectors determine
the flatness at the endpoints. The coordinate s is not the length, i.e., |∂s x| 6= 1.
The beam may therefore be stretched or compressed along its length, unlike the
Euler-Bernoulli beam.
A beam put together from such segments will be invariant under rotations
and translations, since it depends only on vectors which all transform in the same
manner. Furthermore, the elastic energy is finite. The invariant curvature is given
[Guggenheimer, 1963] by:
(11.18) k(s) = ∂s x(s) × ∂s2 x(s) .
Inserting the explicit function, Eq. 11.17, in Eq. 11.18, for a segment yields a lengthy
expression:
(11.19) k(s) = 6(xi − xi−1 ) × vi−1 (1 − s)(12s2 − 9s + 1) .
+6(xi − xi−1 ) × vi s(12s2 − 15s + 4) + 2vi × vi−1 (1 − 2s)(9s2 − 9s + 1) .

γ = 0.5
γ = 1.0
γ = 1.5
Figure 11.3. A smooth curved defined by five points xi and two

end point directions v1 and v5 , of relatively small size compared to
the interposition distance |v| < |xi+1 −xi |. For increasing rigidities
the curvature shifts from the set points to the center of the lines.
However, examining the segment boundary, it yields only a finite jump in curvature
across the segments:
(11.20) lim k(±δ) = 6vi × (xi − xi±1 ) .
δ→0
Therefore, the elastic energy will have no infinite boundary contribution and no
constraints are required to remove them, as was in the case of the spline approxi-
mation. There is, however, a force acting on the segments, besides the local force
acting on the boundaries. The precise definition of the force depends on the choice
of variables. Was in the case of the spline approximated all the force located at
the segment boundary, with the awkward constraints, here the force is more evenly
distributed, without any constraints, except for possible forces at the boundary. It
should be noted that without force F (z) the beam shape is trivial; the purpose is
to describe the response to force in some distributed sense.
Eventually, one would use the positions and directions, xi and vi , of the nodes
in the discretized theory. A distributed force F(s) could be averaged over the
segments, like the forces Fi in the Euler-Bernoulli beam. The curvature k(s) and
the stiffness κ(s) determine the local, distributed force:
δE
(11.21) F(s) = = κk(s)∂s k(s) .
δs
Furthermore, the situation without jump and associated force Fi arises from:
(11.22) vi = γ(xi+1 − xi−1 ) ,
where γ is the rigidity. For small rigidity of the node, the force yielding the curved
beam is located mainly at the nodes xi ; for large rigidity the force acts mainly in be-
tween the nodes. A smooth distribution of the force corresponds to γ ≈ 1. Hence, a
11.3. POSTPONEMENT AND PRE-WORK 213
stationary spline like approximation through the points x0 , x1 , x2 , · · · arises simply

without constraints. See Figure 11.3. Only the first and last directional vectors, at
the ends of the line, are not fully determined: v0 = 0 corresponds to a supported-
like boundary, while a nonzero v0 is like a clamped boundary. The rigidity is the
consequence of the polynomial is s. If we would have chosen to construct the seg-
ments from piece-wise arc sections, the curvature is automatically homogeneous,
and would to depend on the scale introduced by the polynomials and set by the
rigidity. The Bezier curves and non-rational uniform B-Splines (NURBS) are only
a few of related smooth curves, independent of the orientation. [Foley and van
Dam, 1982, Farin, 2001] These curves are mainly used for computer graphics. Only
recently they are of interest of the FEM community as a means to bridge the gap
between computer aided design (CAD) and FEM analysis and simulation. [Hughes
et al., 2005]
The directions vi all transform in the same manner under the rotation R of
the affine transformation of Eq. 11.1. The positions xi also translate with a of the
full affine transformation. By construction the elastic energy remains the same.
Interestingly, the only contributions to the curvature are by the cross-product
of the relative positions xi+1 − xi and the directional vectors vi and vi+1 , and a
smaller contribution from vi × vi+1 . In fact, the elastic energy can be expressed as
weighted vector products:
(11.23)
 T  
Z 1 (xi − xi−1 ) × vi−1 (xi − xi−1 ) × vi−1
κk(s)2 κ vi−1   vi−1 
ds =   K  ,
0 2 2  (x i − x i−1 ) × v i   (x i − x i−1 ) × v i 
vi vi
where the stiffness matrix K is positive constant and symmetric. Furthermore, the
stretch energy is defined by:
Z 1
EA
(11.24) Estretch = ∂s xT (s)∂s x(s)ds ,
0 2
or in the case of a reference or rest length:
Z 1
EA
(11.25) Estretch = (∂s xT (s)∂s x(s) − 1)2 ds .
0 2
11.3. Postponement and pre-work

There are two main approaches to the numerical implementation of the con-
straints, which are the consequence of external and internal boundary conditions.
They both generate a model, which yield the displacement f (z) as direct result
of the boundary conditions and the applied forces Fi . The first approach we call
pre-work, and it works mainly in linear theory. For appropriate boundary condi-
tions the effect of the forces decouple from the effects of the boundary conditions.
If for each force Fi with clamped zero boundary conditions the displacement is
determined the total displacement is a linear combination of forces and boundary
values:
X
(11.26) f (z) = φ0 (z)f0 + ψ0 (z)f00 + φN (z)fN + ψN (z)fN
0
+ φi (z)Fi ,
where the modes ψ0 and ψN correspond to the second boundary condition at each
end, associated with the torque force. The four boundary conditions are given by
the boundary values f0 , f00 , fN , and fN
0
. Each force Fi at zi has its own modal
function:
(11.27) 
L(L−zi )2 zi z 2
2zi
z 3
 3 − 1 + , 0 ≤ z ≤ zi
φi (z) = 2
6κ L L
2
L L
3 .
Lz L−z L−z 2(L−z ) L−z

6κ
i
3 L i
L − 1+ L
i
L , zi ≤ z ≤ L
The four modal functions for the boundaries are given by the transformation matrix
T, with the distance δ = L, given in Eq. 11.9.
The disadvantage of such a modal decomposition is three-fold. First, for large
models, the number of modes grows out of control. Already the construction and
storage of such modes require a lot of computing power. Secondly, the bound-
ary conditions and input forces are set. Using, for example, force actuation at
the boundary, or the value of the displacement f (zi ) instead of the force Fi at a
particular node requires the partial inversion of the large modal matrix φi (zj ).
The approach described above will work very well in linear models. In nonlinear
models it is often prohibitively expensive. In that case every configuration changes
the forces and interactions, which need to be calculated at every stage. One seeks to
reduce and automate the calculations, and possibly adapt the model that it can be
done generically, sequential, or parallel. Such an approach we call postponement.
In this case the displacement is written as a general integral with four unknown
integration constants:
(11.28) ! ! !
Z z Z z0 Z z00 X Fi δ(z 000 − zi )
0 00 000
f (z) = dz dz dz + C3 + C2 + C1 +C0 ,
0 0 0 i
κ(z 000 )
a direct implementation of Eq. 11.15. The numerical implementation will be the
numerical scheme to calculate the respective integrals of the different unknown
constants.
The four constants Ci are given by the boundary conditions. They match with
the four boundary modes φ0 (z), ψ0 (z), φN (z), and ψN (z) above. However, in this
case the homogeneous displacement is constructed as we integrate along. Little
memory is required; the displacement f (z) is constructed on the fly. Any unknown
force Fi given by the constraint f (zi ) is added to the set of integration constants.
Nonlinearities will not affect the approach. However, different boundary condi-
tion, associated with different causalities, may require a different postponement
approach. For example, a beam with one free end, may be integrated from the
clamped end to the free end without constraints, while a beam clamped at both
end requires to constants Ci is integrated from one side to the other.
11.4. Consistency and conservation laws

Continuous theories are designed to satisfy conservation laws. A theory incon-
sistent with underlying physical laws, such as mass or energy conservation, is soon
rejected. The consistent theories remain. However, continuous theories have, in
principle, an infinite number of degrees of freedom and can only be analyzed nu-
merically in a finite order approximation. Within the process of approximating the
infinite number of degrees of freedom with a workable finite set, choices have to be
11.4. CONSISTENCY AND CONSERVATION LAWS 215
made. The simplest way to make consistent approximations is to resort to the vari-
ational principles. If the continuous function satisfies some minimization principle,
the appropriate finite-order approximation will also satisfy the same minimization
principle where the numerical value approximates the exact continuous result from
above. These principles have been the guiding principles in the development of
dynamical systems for centuries.
Conservation laws are continuity relations, in their generic form they relate the
flow J to the rate-of-change of a density ρ:
(11.29) ρ̇ + ∇ · J = 0 .
This equation immediately sums up all the trouble one would have in constructing
finite-order approximation of ρ and J. First, the divergence means that the con-
tinuous nature of the flow J, as function of position is related to a density ρ at a
point. Flow and density are therefore of a different geometric type. The same is
true for the time-derivative of the density, ρ̇, in the continuity equation. The flow
and the density are also of different dynamical types. Furthermore, if the continuity
equation holds, so must any derivative, spatial and temporal:
(11.30) ρ̈ + ∇ · J̇ = ∇ρ̇ − ∆J = 0 ,
where we used that typically the flow is defined rotation-free: ∇ × J = 0. Hence

the equations of motion, not resulting directly from conservation laws, must still
be consistent with the conservation laws.
A lot of effort is taken to satisfy as many as possible of the underlying physical
principles in the numerical approximation of the continuous theory. Conceptually, it
is difficult to determine which approach will retain the principles, and the resulting
symmetries and conservation laws. Our starting point is retaining as much of the
consistency of the equations of motion as possible. Dirac [Dirac, 1964] developed
a scheme to implement constraints based on the time dependence of constraints.
The constraints are inserted into the equations of motion, yielding possibly addi-
tional constraints, and eventually a consistent equations of motion based on the
commutator brackets. Our approach is to seek a consistent time dependence of
the finite-order approximation. Simply said, even though J(z) only has limited
variation as a function of the spatial coordinate z, the corresponding ρ(z) makes
the pair satisfy the continuity relation, Eq. 11.29, exactly. Any conservation law
based on the equations of motion will be satisfied as well. Both the states and the
Hamiltonian itself are approximated in the process.
The continuous Hamiltonian equations of motion depend on functions of the
continuous spatial variable z:

q̇(z) δp(z) H 0
(11.31) = + .
ṗ(z) −δq(z) H F(z)
The finite-dimensional set of equations of motion with q, p ∈ RN are the equations

of motion of the coefficients of some linear subspace of the continuous space:

q̇ ∇p H 0
(11.32) = + .
ṗ −∇q H f
The vectors q and p are the coefficient vectors, i.e., q = (q0 , q1 , q2 , · · · , qN ) of

corresponding modes or basis functions in the continuous space:
P
q(z) j φj (z)qj
(11.33) = P .
p(z) j ψj (z)pj
Like the divergence operator ∇· in the continuity relation, Eq. 11.29, the Hamil-
tonian of a continuous system will in general contain differential operators acting
on the continuous, or field variables. The differential operator D acting on the
functions q(z) and p(z) must have exact representations D in the coefficient space:
(11.34) Dq(z) = h(z) ⇔ Dq = h ,
P
such that: h(z) = φj (z)hj . For example, polynomial, exponential, and trigono-
metric functions will be mapped into the same functions under differentiation.
These are precisely the functions commonly used the modal approximations, of
linear, homogeneous systems.
Consistency of the equations of motion means the right-hand side and the left-
hand side of the continuous equation of motion, Eq. 11.31, using the modes φj (z)
and ψj (z) and the discrete equation of motion, Eq. 11.32 lie in the same subspace
of the infinite-dimensional continuous space. If that is the case, the conservation
laws derived from the equations of motion are also automatically satisfied. Hence,
rather than a metric approach, based on the norm as in the modal approach, we
use an algebraic approach, where we identify the operators.
Position-dependent masses and stiffnesses have to be restricted, for example,
to piece-wise linear mass densities ρ0 + ρ1 z and stiffnesses κ0 + κ1 z, such that the
symmetric formulation of the one-dimensional wave equation, in a polynomial space
of modes [Henrici, 1974] is indeed consistent:

q̇(z) ∂z (ρ0 + ρ1 z)p(z)
(11.35) = .
ṗ(z) −∂z (κ0 + κ1 z)q(z)
If q(z) and p(z) lie in a polynomial space of order N , the multiplication with the
linear material constant increases the order by one to N + 1, while the differentia-
tion lowers the dimension again to N . In the case of nonlinear Hamiltonians this
construction is more complicated, as we have seen in Section 5.10, however, luckily
the differential operators can be brought outside the Hamiltonian, yielding the port-
Hamiltonian formulation, discussed in Section 7.5. The projection ∂z [(ρ0 +ρ1 z)p(z)]
from an N -dimensional polynomial space back to the N -dimensional polynomial
space should have a nontrivial null-space, associated with a single boundary condi-
tion. In this case the one-dimensional null-space is spanned by:
1
(11.36) p0 (z) = ,
ρ0 + ρ 1 z
which does not lie in the polynomial subspace. As a result conservation laws will
fail for such approximations of position-dependent Hamiltonian with polynomial
function spaces. On the other hand, the truncation of (ρ0 +ρ1 z)p(z) back to the N -
dimensional polynomial space will yield a non-trivial null-space, an approximation
of p0 (z), and the requirement of a boundary condition. The particular null-space
p0 is in this case the consequence of the translation invariance, yielding solutions
with constant velocity:
(11.37) v = constant ⇒ (ρ0 + ρ1 z)p0 = constant .
11.5. ELEMENT HAMILTONIAN 217
In the general dynamical setting of evolution equations, discussed in Chapter 3,

the matrix A is nonsingular otherwise the dimension of the state could be reduced:
(11.38) ẋ = Ax + Bu .
A typical differential operator D at the basis of A, however, will be singular:
Dx0 = 0. The resulting A will also be singular. There are two solutions for con-
sistent dynamics: First, the matrix A inherits the null-space from the underlying
D operator, and the control vector B spans in the null-space of A:
(11.39) Ax0 = 0 ⇒ BT x0 6= 0 .
For example, consider a transmission line Axj (t) = xj+1 (t + δ), the signal moves to
the right, and an input with B = (1, 0, 0, · · · ) adds a new x0 (t) = u(t) at the input
end.
The second solution is a reformulation of the first one: The state x is restricted
to the image space of A, and B represents the coupling between the boundary
conditions, fixing the core model (Section 7.6) in the null-space of A, and the
internal states x.
In later sections consistency will appear in different forms. We rely on polyno-
mial spaces to have injective differential operators in the Hamiltonian, using only
piece-wise constant material properties. These differential operators are therefore
never of full rank. Differential operators may have complicated null-spaces. The
case of ∇× we analyze in some detail in Section 11.7. The boundary conditions
play an important role to yield consistency.
11.5. Element Hamiltonian

In the Hamiltonian nonlinear FEM, each element Hamiltonian is the energy
expressed in local variables q. The null-space of the Hamiltonian is the set of
variables q0 under which the energy is invariant:
(11.40) H(q) = H(q + q0 ) .
In the case of a quadratic Hamiltonian it corresponds to the eigenspace with eigen-
value 0, i.e., qT0 Kq0 = 0, if kKk ≥ 0.
This is not invariant under general coordinate transformations. For example
consider a Hamiltonian:
1
(11.41) H = (q12 + q22 ) .
2
In this case the null-space is the origin q1 = q2 = 0, however, under the introduction
of polar coordinates r, θ with r 2 = q12 + q22 and tan θ = y/x, the Hamiltonian
depends only on r. Such a coordinate transformation is similar to the action-angle
coordinates introduced in Section 2.2. The null-space is θ. In the case of discrete
approximations, these two sets of variables are distinct approximations, which will
lead to different subspaces of the full system. With the polar coordinates it would
be difficult to interconnect a grid of segments.
Elasticity Hamiltonians are invariant under translation and rotation of the ob-
ject as a whole. Hence the translation and rotation transformations should be
part of the null-space of the Hamiltonian. The kinetic energy, however, generally
depends on translation and rotation velocities.
order #variables #boundary eigenvalues (degeneracies)

0 3 3 03
10 6 6 2
1 24 24 0 4 8 12
2 81 78 029 46 86 96 122 1312 176 186 226 272
3 192 168 066 46 86 96 122 1312 166 176 186
20 22 246 2512 272 2912 326 346 366
12 6
Table 11.1. The dimensions and eigenvalues of the quadratic op-

erator A → (∇ × A)2 in a polynomial space of different low orders.
For example, given a block of elastic material. In the linear approximation it is

determined by eight three-dimensional positions: 24 variables. In the null-space of
the element Hamiltonian there are three dimensions associated with the translation
of the block, and another three with rotations, leaving an 18-dimensional elasticity
subspace. The strain tensor is normally only determined by four positions, 12
variables, of which 6 are associated with rotation and translation, leaving three
compression and three shear coefficients.
A rotational flow, or magnetic field, which invariant energy is given by:
k
(11.42) H(A) = (∇ × A)T (∇ × A) .
2
If the vector field A is approximated on a unit cube [0, 1] ⊗ [0, 1] ⊗ [0, 1] by linear
functions:
(11.43) A(x, y, z) = a000 + a100 x + a010 y + · · · + a111 xyz ,
there are again 24 variables aijk , almost identical to the elastic block. The 24
displacement positions can be mapped to 24 terms in a polynomial expansion. The
null-space, however, seems much more complicated. The constant function a000 ,
associated with laminar flow, or translation, in three directions is part of the null-
space. So are all the symmetric terms: ax100 , ay010 , and az001 . The three symmetric
permutation combinations:
(11.44) az010 = ay001 , ax001 = az100 , ay100 = ax010 ,
are also part of the null-space, and so is the combination:
(11.45) ax011 = ay101 = az110 .
Together they span the 10-dimensional null-space of the curl operator on polyno-
mials to order 1. For polynomials to order 2 the null-space is 29-dimensional, and
for order 3 the dimension is 66. See Table 11.1.
11.6. Boundary energy

In order to construct a proper FEM model, using element Hamiltonians, the
energy associated with the boundary should vanish. The separate Hamiltonians
interact due to their common variables, but any boundary energy is automatically
infinite, since it is the product of two delta functions. For example a discontinuous
step function f (z) = 0 for z < 0 and f (z) = 1 for z ≥ 0 will result in a delta
function under differentiation: ∂z f (z) = δ(z). The boundary energy is therefore
infinite:
1 1
Z Z
2
(11.46) E= dz(∂z f (z)) = dz(δ(z))2 = ∞ .
2 2
11.7. COUNTING DEGREES OF FREEDOM 219
The approximate solution will be discontinuous for a certain order of the derivative.
However, this order should be high enough that it does not yield a contribution to
the energy. The spline solution led to a boundary term for the second derivative,
while using the invariant-lines approach this contribution was absent.
In the case of first-order operators like ∇×, only the function itself has to
be continuous to yield a consistent result without boundary contributions. The
boundary energy for the common boundary ∂Ω = Ωi ∩ Ωj between domains Ωi and
Ωj is given by:
Z
(11.47) Eij = dz(nTij δ∇ D(fi − fj ))2 ,
Ω1 ∩Ω2
where the boundary operator nTij δ∇ H is defined as the the sum of terms where every
time a single differential operator is replaced by the component of the boundary
normal nij = −nji = (nxij , nyij , nzij ):
(11.48)
(nTij δ∇ ∂α1 ∂α2 · · · ∂αr f )2 = (nα α2 αr
ij ∂α2 · · · ∂αr f + ∂α1 nij · · · ∂αr f + ∂α1 ∂α2 · · · nij f )
1 2
.
The approximate function on the whole domain is the sum of smooth functions fi ,
extendable beyond the cell Ωi , times window functions wi (z) = 1 for z ∈ Ωi and
zero outside the cell domain Ωi :
X
(11.49) f (z) = wi (z)fi (z) .
Inserting f (z) in the Hamiltonian results in boundary terms due to the derivative
of the window functions:
Z Z Z
(11.50) dz∇wi (z) = dznT δ(z)|∂Ωi = dzn .
Ωi Ωi ∂Ωi
The boundary energy contribution only vanishes if nTij δ∇ D(fi

− fj ) = 0.
In the case of the energy given by (∇×A)2 the condition reduces to a continuity
condition
(11.51) n × Ai |Ωi ∩Ωj = n × Aj |Ωi ∩Ωj .
The solutions in the null-space of the Hamiltonian automatically satisfy the zero
boundary-energy condition.
11.7. Counting degrees of freedom

A system exists on a domain with a boundary. Furthermore, it consists of
elements each again with domains and boundaries. The common points of different
elements, i.e., the boundary points, are relations which reduce the number of degrees
of freedom. Furthermore, from the total set of degrees of freedom a subset will
be determined by the equations of motion and the initial state, the remainder
must be determined by the boundary conditions, or the input. A consistent set
of equations of motion are the result of assigning the right number of variables to
the boundary. However, since we make no distinction between internal boundaries,
between the elements, and external boundaries, for the boundary conditions, the
proper definition of variables on the elements should fix the system completely.
For triangulation, grid consisting of simplices, i.e., tetrahedrons, this extension
from elements to the full grid is not trivial. The number of elements with a common
vertex or edge is no longer fixed. If, therefore, the relations between degrees of
freedom is based on the identification of a common vertex or edge, it is not clear
how many relations will be given by a single identification. The result depends on
the type of triangulation of the domain.
In this section we will restrict ourselves to cubic cells. The number of neighbors
parts are fixed. The corner is shared by eight elements, an edge by four elements,
and a face by two.
Knowledge of the null-space of the element Hamiltonian allows us to address the
question of uniqueness and existence. For example, consider the magnetic problem
expressed in the vector potential A, with H(A) = (∇ × A)2 . The static solution
to some boundary conditions b is:
(11.52) min H(A), A|∂ = b
A
Uniqueness and existence requires a single minimum to exists. The Hamiltonian

does not have a single minimum, it has a null-space which yield a set of solutions
all with the same minimal energy. The boundary condition b should select from
this null-space a unique solution. In the continuous case typical boundary condi-
tions would still not yield a unique solution as the null space contains any function
Ak = ∇φ, given by an arbitrary φ, including functions that vanish at the bound-
ary. In the discrete approximation the typical basis functions do not vanish at the
boundary. In the case of a discrete approximation of a vector field A, it is generally
an unrestricted field A with Ak 6= 0. We would, however, like to separate the
longitudinal degrees of freedom in A, since they are not associated with dynamics.
Hence the null-space ∇× is separated in two parts: the degrees of freedom fixed
by boundary conditions, or port variables, and degrees of freedom associated with
static or constrained longitudinal fields.
In the continuous case, normally, to yield uniqueness a particular gauge choice,
for example the Coulomb gauge ∇ · A = 0 or the axial gauge Az = 0, is used
as auxiliary, or constraint equation. The gauge condition yields a different set of
boundary conditions than the original equations defined in terms of B = ∇ × A,
already for the fact that if A are polynomials of a particular order, the order of the
generated B is lower. In terms of the magnetic field B the boundary conditions
should yield ∇ · B = 0. For A these conditions are naturally fulfilled.
For the polynomial system, the gauge condition should yield a section of the
null-space, such that a unique projection null-space arises. The Coulomb gauge
condition yields the four equations:
(11.53) ax110 + az011 = ay110 + az101 = ax110 + az011 = ax100 + ay010 + az001 = 0 .
The remainder of the ten-dimensional null-space needs to be fixed by the boundary
conditions. The constant vector potential a000 can clearly only be fixed by bound-
ary conditions. Furthermore, the gauge condition restricts only one dimension in
the three-dimensional symmetric space {ax100 , ay010 , az001 }. Hence there are three
nontrivial boundary conditions, such that the intersection of the null-space with
the gauge-condition and the boundary condition ∂ : A|∂ = b is unique:
(11.54) rank{∇×, ∇·, ∂} = 24 ,
and consistent (see Section 11.4):
(11.55) rank{∇×} + rank{∇·} + rank{∂} = 24 ,
where ∇×, ∇·, and ∂ represent the Hamiltonian differential operator, the gauge
condition operator, and the boundary operator acting on the polynomial vector
11.7. COUNTING DEGREES OF FREEDOM 221
Figure 11.4. A third-order polynomial in each variable x, y, z on

each cell is determined by the values at 4dim points. The boundary
points are shared with neighboring cell, or the boundary of the
domain. The internal points are unaffected.
space A = aijk xi y j z k . The latter condition is often relaxed in weak formulations.

P
A minimum exists even for inconsistent equations. This means, the equations over
specify the field. A unique minimum, however, will still be found. The deviation
from the consistent result is due to the balancing of each of the conditions. Hence
the B = ∇ × A equation may not hold exactly, but to a particular order, depending
on the integration method in the weak approach, such as Galerkin method.
The uniqueness and existence conditions are the discretized version of the
Hodge-Helmholtz decomposition. [Hodge, 1941, Flanders, 1963] In the continuum
case the field A(x, y, z) is indeed readily decomposed into (Eq. 4.39):
(11.56) A(x, y, z) = Ak (x, y, z) + A⊥ (x, y, z) + Ah (x, y, z) ,
with ∇×Ak (x, y, z) = ∇×∇φ = 0, ∇·A⊥ (x, y, z) = 0, and ∆Ah (x, y, z) = 0, where
the parallel field Ak is fixed by the gauge condition, the rotational field A⊥ by the
original Hamiltonian, and the harmonic field Ah by the boundary conditions. For
a particular finite order polynomial approximation this decomposition no longer
holds.
We, however, have explicit information about the null-spaces of the element
Hamiltonians. The dimension of the null-space is given by dn , where n is the
polynomial degree. The gauge condition, typical of the continuum theory, is no
longer required. Assigning the correct number of variables to the boundary will
yield a consistent approximation with a unique result for a given boundary value
problem. See Figure 11.4.
A three-dimensional domain approximated by N 3 cells, each with a n-th order
polynomial basis functions based on equidistant nodes yields:
(11.57) #domain variables = cnN = 3(nN + 1)3 = 3n3 N 3 + 9n2 N 2 + 9nN + 3 .
The number of variables on the boundary of the domain:
(11.58) #domain boundary variables = bnN = 18(nN )2 + 6 .
The number of cells is:
(11.59) #cells = CN = N 3 .
The number of variables on a single cell:
(11.60) #cell variables = cn = 3(n + 1)3 = 3n3 + 9n2 + 9n + 3 .
The number of variables on the boundary of a cell:
(11.61) #cell boundary variables = bn = 18n2 + 6 .
For each cell, with n-th order polynomials in each direction lead to the condi-
tions:
(11.62) rank{∇×}cell + rank{∂}cell = cn − dn + bn > cn ,
and
(11.63) rank{∇×, ∂}cell = cn .
Since the null-space dimension dn is smaller than the number of boundary variables
bn , the first condition is not satisfied. For a single cell, with the typical boundary
conditions, the number of boundary conditions is larger than the number of unde-
termined variables, in the null-space of the Hamiltonian.
In the case of N cells, each with n-th order polynomial interpolations, there
are cN n variables, of which bN n on the boundary. The dimension of the rank{∇×}
is N 3 times the image space of a single cell operator:
(11.64) rank{∇×}domain = N 3 (cn − dn ) .
The rank of the domain boundary operator is:
(11.65) rank{∂}domain = bN n .
Hence for a large number of cells the number of degrees of freedom cnN minus the
number of relations from the single-cell Hamiltonians, rank{∇×}domain, grows like
N 3 n2 . If a single cell yields an inconsistent condition, so does a domain of cells:
(11.66) rank{∇×}domain + rank{∂}domain > rank{∇×, ∂}domain .
Besides the cumbersome counting to keep track of the variables and the rela-
tions, from the local Hamiltonian and the cell and domain boundaries, the incon-
sistent result is unwanted. Much more elegant is a treatment based on a single
cell, which can be connected consistently with other cells and boundaries. Such a
single cell, with port-variables common to adjacent cells, we will call an element
port Hamiltonian. The proper choice of port variables covers the null-space of the
Hamiltonian. The interconnection of these variables may lead to a network of alge-
braic relations, depending on the boundary conditions, which connect the system
to its surroundings.
11.8. The element Hamiltonian as port Hamiltonian

Constructing full finite-element Hamiltonians from cells, which fill a domain,
will lead to consistency problems, as we have seen in the previous sections. The
consistency problem manifests itself in an overdetermined set of equations, which
therefore are no longer a finite-dimensional subspace of the infinite-dimensional
space with exact representations of the differential operators, necessary to retain
conservation laws. The polynomial functions make sure the differential operators, in
the Hamiltonian and the conservation laws, have exact solutions. The mapping back
onto the finite-dimensional coefficient vector must be bijective. The overdetermined
set of equations leads to a projection.
Besides consistency problems, the typical formulation of the partial differential
equation is the dressed-down version of the partial differential equation that would
arise in interacting systems. The boundary conditions are typically in terms of the
fields, or state variables. Energy conservation, including the inflow and outflow
11.8. THE ELEMENT HAMILTONIAN AS PORT HAMILTONIAN 223
through the boundary, will be cumbersome integrals of these fields over the bound-
ary. [Sakamoto, 1982] Introducing physical boundary conditions would simplify the
expressions for the energy flow through the boundary. The appropriate, physical
boundary condition is some operator B, and the finite-dimensional representation
we seek reduces the problem:
(11.67) min H(A), given BA = b ⇒ min H(b, x) ,
A x
where the Hamiltonian is expressed in terms of the boundary condition b and

possibly additional internal variables x, which is a finite-dimensional representation
of the fields A(b, x).
Some advantages of the form H(b, x) are immediately clear. First of all, for
a given boundary condition b, the Hamiltonian is only a function of x, with a
parametric dependence on b. Furthermore, it is an unrestricted function of x;
in a time-dependent problem initial conditions are a value for the internal state
x at the initial time. For a domain, consisting of cells, the internal boundaries
are implemented by assigning Tji bi = Tij bj , where Tji and Tij identify the
common boundary for adjacent cells i and j. The interaction of the system with
its surroundings occurs in terms of the variables which lead to a simple bilinear
product, representing the power flow through the boundary.
The typical boundary values b are the force, the pressure, and the flow through
the boundary. These variables have more meaning in the global setting than the
state variables, such as density, flux, stress, in which the partial differential equation
and consequently the boundary conditions are normally expressed. In order to
arrive at these boundary variables, the equations of motion play a central role, both
in the derivation as in retaining consistency. The typical equations of motion of a
hyperbolic system are more general than the pseudo-symplectic port-Hamiltonian
equations (Eq. 7.28) of motion as the pair of operators: D and D c 6= D∗ do not have
to be the formal adjoint (Eq. 2.69) of each other. Different boundary conditions lead
to different operators D and D c , as we have seen in the section of initial boundary
value problems, Section 5.3.
The general equations of motion for a hyperbolic system are:

q̇(z) 0 D δDc∗ q(z) H
(11.68) = ,
ṗ(z) −Dc 0 δD∗ p(z) H
where D and Dc are positive and negative semi-definite operators associated with
the kinetic and the potential energy respectively. The discretized, finite dimensional
approximation of these equations is:

q̇ 0 D ∇Dc∗ q H
(11.69) = ,
ṗ −Dc 0 ∇D ∗ p H
where the consistent dimensions q ∈ Rr and p ∈ Rs are yet to be determined
from the properties of the matrices D and Dc . Although in the function space
q(z) and p(z) may lead to complicate forms of δq(z) H and δp(z) H for nonlinear
and position-dependent Hamiltonians, the local dimension, in configuration space,
of ∇q H and ∇p H are determined by the dimensions of r, s of q and p. The
mapping back the original function spaces q(z) and p(z) and their finite-dimensional
representations φi (z) and ψi (z) given in terms of the coefficients q and p must be
seen and treated as an approximation of the Hamiltonian or the closure relations.
Hence the conserved energy is no longer determined directly by q(z) and p(z), but
in the finite-dimensional representation.
The dimensions of the operators D and Dc determine the boundary variables
(∇q H, ∇p H) → b. The simplest procedure is to determine the time-dependence of
the Hamiltonian:
(11.70) Ḣ = ∇Tq HYD∇D∗ p H − ∇TDc∗ q HDcT Y∇p H ,
where Y is the matrix consisting of the integrals over products of basis functions:
Z
(11.71) Yij = dzφi (z)ψj (z) .
If the φi and ψi are limited respectively to N th order and M th order polynomials,

the integration must be exact of (N + M )th order polynomials to guarantee the
exact power conservation.
The finite-dimensional boundary operator is the matrix B associated with the
bilinear power flow h∇Dc∗ q H, ∇D∗ p HiB :
(11.72) B = PTDc DcT YD − DcT YDPD ,
where the projection operators PDc and PD project perpendicular to the null-space
of Dc and D respectively, which arises from the change of variation. This is the
difference between the normal variation and the differential variation, related to
the formal adjoint (Eq. 2.69):
(11.73) D∗ δDq H ⊂ δq H .
The two variations are identical for Dq 6= 0.
The special nature of the boundary operator is only clear from the actual
representations of the differential operators, D and D c , and the underlying finite-
dimensional subspaces, spanned by φi and ψi .
The image spaces of D and D c can be embedded in the polynomial domain
spaces, with a Cartesian vector space when appropriate, such as in the case of the
gradient operator:
N
X N
X
3
(11.74) ∇PN :∇ aijk xi y j z k = cijk xi y j z k ⊂ R3 ⊗ PN
3
,
i,j,k=0 i,j,k=0
with cijk being a vector in the direction of the gradient. Constant differential
operators acting on polynomials spaces P can always be embedded in the original
space:
(11.75) DP ⊂ E ⊗ P ,
where E is a vector space associated with the operator. Making, a priori a distinc-
tion between the domain space and the image space, as has been investigated in
the case of the ∇× operator, in Section 11.7, turns out to be rather cumbersome
with little gain, especially in the generic hyperbolic case D c 6= D∗ , investigated
here. In the case of the port-Hamiltonian D c = D∗ it would have been naturally
to restrict the function spaces accordingly to some bilinear product space, in the
port-Hamiltonian formulation called the power product. The two function spaces
φi and ψi are the natural dual of each other. Much work has been invested in
linking the underlying geometry with the function space using algebraic geometry.
[Hodge, 1941, Whitney, 1957, Morita, 2001, Bossavit, 1998] However, as we project
the infinite-dimensional function space (q(z), p(z)) onto a finite-dimensional space
11.8. THE ELEMENT HAMILTONIAN AS PORT HAMILTONIAN 225
(q, p) we will spare ourselves the trouble and deal with matrix-type operators only,
after the initial mapping of D and D c onto D and Dc .
However, for example, for the most common formulation of hyperbolic systems,
in this case the linear wave equation, the operators are not the adjoint of each other,
D = 1 and Dc = −∆:

q̇(z) 0 1 q(z)
(11.76) = ,
ṗ(z) ∆ 0 p(z)
noting that the Laplacian is a negative operator, without positive eigenvalues. If
the boundary operator B is to be restricted to the part of the equations which
are not determined by the equations of motion Y should not be symmetric, and
the subspace φi and ψi not equivalent. Only the projection operator PDc will be
nontrivial, while PD = 1.
The function q(z) needs to be augmented with boundary conditions in order to
yield a unique map onto ṗ(z). These are the boundary conditions for the elliptic
problem ∆q(z) = 0, such as the Dirichlet or Neumann conditions.
However, in a more general setting, instead of fixing the boundary conditions
for q(z), the equations can be made complete by adding an applied force to the
incomplete force, or Newton, equation, which is the second row:

q̇(z) 0 1 q(z) 0
(11.77) = + ,
ṗ(z) ∆ 0 p(z) F (z)
where the force F (z) lies in the left null-space of ∆. Although applied force F (z)
may appear reasonable, it will not convergence consistently, in some norm, to the
boundary integrals in any function space. The left null-space will include the
highest-order terms of the polynomial space, and consequently, a consistent F (z),
within the same function space, will be highly oscillatory, if the polynomial space
is sufficiently high.
The right null-space of ∆, on the other hand, contains only the constant and
linear functions. The null-space solution q0 (z) can be added to the state q(z)
without changing the dynamics (q̇(z), ṗ(z)). However, the time-variation of the
null-space solution q̇0 can be traced back to the momentum p0 (z). The core model,
Eq. 7.53, contains only q0 (z), the right null-space of ∆. The dimension of the null-
space determines the number of boundary, or port, variables. The mapping to the
discrete vectors q0 and p0 must be such that a network arises.
The geometry of port Hamiltonian elements is usually given by nodes, edges,
or faces. On its own is a single element just a blob where the structure has little
significance, but connecting elements together yields a geometry. A network arises
from the common, or connected, nodes, edges, and faces. If only two elements are
connected with another, the joint, or port, is a simple identification of variables.
Potentials and densities are the same, flows and forces are opposite if the element
normals n are opposite, which depends on the type of joint. The joint is the geomet-
ric aspect of the port. Connecting three or more elements in a single joint requires
some joint model. The identification of potentials and densities is straightforward,
but the extension of a normal on a boundary of two elements to the common edge
of three of more elements is ambiguous.
The right null-spaces of D and Dc yield the states of the core model. These
should be mapped to port variables, while the remainder of the variables vanish at
these joints. For example, in one spatial dimension z ∈ [0, 1], with D c = −∂z2 , the
null-space consists of the span{1, z}. The typical decomposition would be of a 4-th
order polynomial basis, on 5 equidistant nodes z = {0, 14 , 12 , 34 , 1} would be:
 1   
φ0 1−z
 φ20   z 
 φ1  =  (1 − z)z z − 1 z − 3  .
   
(11.78) 2 4 
 φ2   (1 − z)z z − 1 z − 3 
  
4 4
φ3 (1 − z)z z − 14 z − 12
The internal modes φ1 , φ2 , and φ3 vanish at the port nodes z = 0 and z = 1, and
the boundary, or port, modes φ10 and φ20 vanish for the Dc operator: Dc φ10 = 0
and Dc φ20 = 0. Furthermore, the null-space of D c is nicely separated in these two
modes each based at a single port node, φ10 at z = 0 and φ20 at z = 1. The Dc
matrix has two zero columns, associated with the right null-space:
0 0 − 13 − 19 − 47
 
4 8
 0 0 −7 − 19 − 13 
4 8 4
c
 
(11.79) D = − 0 0 96 96 96   .
 0 0 −192 −192 −192 
0 0 96 96 96
The boundary operator is:
(11.80) B = −(1 − PDc )Dc Y ,
since D = PD = 1.
11.9. From null-spaces to port variables

The null-space of the differential operator D is the null-space of the Hamiltonian
based on that operator. If the equations of motion are the Hamilton equations of
motion, the dynamics will not depend on degrees of freedom in the left null-space,
and degrees of freedom in the dual of right null-space are not determined by the
Hamilton equations of motion. The left and the right null space are connected if
the dual is based on the inner product of the exact integration in the polynomial
space.
In order to uniquely specify the dynamics of the full space, including the null-
spaces, the system needs to be augmented with boundary conditions or constants
of motion. In case of discrete models, at the basis of a FEM approach, it is useful to
assign the null-space to a boundary node, as we have seen in the previous section.
The typical approaches will lead to an identification of the left null-space of D,
based on the observability argument from Section 3.3. A control vector is added
which determines the rate of change of the state vector x outside the observability
range since it lies in the left null-space of A.
However, the left null-space will lead to highly oscillatory functions in a polyno-
mial space, as the highest-order polynomials vanish under differentiation. See also
Figure 5.1. The degree of polynomials are lowered under differentiation, hence, the
highest degrees lie in the image, or left, null-space. The norm, such as the Lebesgue
norm, determines the precise form of the null-space functions. The right null-space
is much more smooth and appropriate, and more closely related to lumped ap-
proaches which will dominate the reduced-order models. In order to be able to use
the right null-space, it is assigned to a boundary node, and the remainder of the
function space orthogonalized with respect to that projection.
11.9. FROM NULL-SPACES TO PORT VARIABLES 227
A node becomes a port by bi-orthogonalization of the remainder rather than

the mode itself, typical for multi-component structural analysis, such as Craig-
Bampton. [Craig and Chang, 1977, Craig, 1981] The modes associated with the
null-space are attached to particular nodes, the port nodes. The other, internal
modes have to vanish at the position of these port nodes, in order retain power
continuity.
The transformation matrix T0 from the coefficients of a polynomial expansion
φi (z) = z i of the modes to the values at the nodes is given by:
1 z1 z12 · · ·
 
2
 1 z 2 z2 · · · 
(11.81) T0 = zij =  1 z3 z 2 · · ·  ,
 
3
.. ..
 
. .
which is non-singular for zi 6= zj . The subscript zero in T0 indicates the basic

modes, not the port modes and the remainder.
The derivative in terms of the node values is given by:
 
0 1 0 ···
 0 0 2 0 
(11.82) D0 = T0  · · · 0 0 3  T−1 = T0 dT−1 .
 
 0 0
..

.
The null-space is the constant function, φ0 (z) = 1, which in the node vector is given
by q0 = (1, 1, 1, 1, · · · ); the first column of T0 , such that D0 q0 = 0. If the value at
z1 is the port variable, the other modes have to vanish at z1 . The transformation
matrix T from the modes to the nodal values is therefore of the following structure:
1 −z1 −z12 · · ·
   
1 0 0 ···
 1 z2 − z1 z22 − z12 · · ·   0 1 0 ··· 
..
 
(11.83) T =  1 z3 − z1 z 2 − z 2 · · ·  = T0   ,
 
3 1  . 0 1 ··· 
.. ..
   
. . ..
.
which is the matrix equivalent of subtraction of the Taylor expansion (Eq. 2.76)
around the port node z1 .
In the case of a set of ordered nodes z1 < z2 < z3 · · · < zN , the input port is
z1 , the output can be any port, however, typically the last one zN is picked. The
consistent theory of such elements requires two matrices. The first matrix is the dif-
ferential operator in the space of the particular modes, given by the transformation
matrix T:
(11.84) D = TdT−1 .
The other matrix is the exact integration of the modes on the space to define the
metric:
Z b ! i+j+1
− ai+j+1

−1 b
(11.85) Y=T i+j
z dz T = T T−1 ,
a i+j+1
j
i
Figure 11.5. A patch associated with the central point i is made

up of a polynomial of equations of lines constituting the boundary
of the domain. The order d of the polynomial depends on the
required smoothness.
where [a, b] is the domain. The differentiation and integration carries over from the
continuous space:
(11.86)
Z Z
∂z φi (z)φj (z)dz = (φi (z)∂z φj (z) + (∂z φi (z))φj (z))dz → [DT Y + YD]ij .
11.10. Plate port-Hamiltonian element

For small deflections of a plate from the flat surface z(x, y) = 0 the elastic
energy of a plate is given by the bi-harmonic form:
Z
(11.87) Eelastic = dxdy(∆z)2 ,
where the bending stiffness is set to unity, to simplify the equations. Hence the
values at the boundary should not only be equal, but also the derivatives across
the boundary should be equal. It is the extension of the Euler-Bernoulli beam to
two spatial dimensions. The differential operator D is the Laplacian.
The null-space up to second order, in each direction, is five dimensional:
(11.88) N = {1, x, y, xy, x2 − y 2 } .
The only nontrivial monomials are x2 + y 2 , x2 y, xy 2 , and x2 y 2 . See Section 11.7.
The smoothness condition for the boundary, such that the boundary energy
vanishes (see section 11.6), requires that the derivatives across the boundary are
continuous. Similar to the spline (Section 11.1) it is impossible to construct such
a solution, with second order polynomials alone. For example a solution could be:
(x, y) ∈ [0, 1] ⊗ [0, 1]
(11.89) φ
internal (x, y) = (x(1 − x)y(1 − y))
2
,
which requires at least fourth-order terms, x4 , y 4 , in the expansion. For a triangular
surface the situation is similar, for example:
(11.90) φ∆
internal (x, y) = (xy(1 − x − y))
2
,
for 0 < x < 1 and 0 < y < 1 − x, which only have fewer cross terms like xi y i , with
i + j > 4.
11.10. PLATE PORT-HAMILTONIAN ELEMENT 229
Figure 11.6. The triangulation of a sphere can be constructed

from matching a triangulated domain with its copy. The vertices
and edges at the boundary count for one, all other vertices, edges,
and faces occur twice in the sphere, with Euler characteristic χE =
2.
Furthermore, even is an element does not have any non-trivial internal modes,
which vanish at the boundary, a consistent port-Hamiltonian element may arise.
However, in the case of a second-order expansion, the dimension of the null-space is
more than half the total dimension. If the simply-connected surface is triangulated,
the Euler characteristic, Eq. 8.78, determines:
1 1
(11.91) 2 = 2 − Vb + E b = V i − E i + F ,
2 2
where V , E, and F are vertices, or nodes, edges, and faces respectively, and the
subscripts “i” and “b” stand for internal and boundary respectively. For a typical
triangulation of a surface a vertex is connected on average to six edges, hence
3Vi ≈ Ei which reduces the internal part of the degrees of freedom to:
(11.92) 2 × number of ports = 2Vi = F = number of elements .
See Figure 11.6.
Every port is a relation between an input and an output. If the state of an
element is given by a number of variables, the number of variables in the null-space
must be determined by the relations among the port variables. If z(x, y) is fully
specified by the values at the vertices, five of the nine degrees of freedom, corre-
sponding to the null-space, are to be determined by the ports, and the remaining
four should yield output at the ports, which cannot be.
Starting at degrees of freedom at each vertex (x, y), which we will use as ports,
Hence, a priori definitions of low-order modes will lead to inconsistencies in the
numbers of degrees of freedom we can assign to ports. The dimension of the null-
space is too large compared to the total number of degrees of freedom. For a
one-dimensional reference space, like a signal cable, half of the degrees of freedom
from the null-space are input, the other half is output. For higher dimensions, the
number of input is smaller, therefore the relative dimension of the null-space must
be smaller.
A locally smooth surface is defined by its value z(x, y) and a surface unit normal
n(x, y) with |n(x, y)| = 1. In total three degrees of freedom define a locally smooth
Figure 11.7. The white vertex in the center is associated with the
dark-grey patch. The stiffness matrix will have non-zero entries
due to overlap with the light-grey patches with the six black ver-
tices.
surface, which may act as port variables. The values at the common edge are to
be determined solely by the two vertices, at the two ends. However, this is usually
not possible.
One procedure to define such interpolations is to define a patch around the
vertex, on the domain of faces common to the vertex.
Y
(11.93) φpatch i (x, y) = (ci + ai x + bi y) (αhj,ki x + βhj,ki y + γhj,ki )d ,
hj,ki
2
where d = 2 for the plate and αhj,ki x + βhj,ki y + γhj,ki = 0, with βhj,ki + α2hj,ki = 1,
defines a line through the vertices j and k, on an edge hj, ki at the boundary of the
patch domain. See Figure 11.5. The constants ai , bi , and ci determine the value
z and the slope n at the central vertex i = (x, y). Outside the domain the the
function is zero. The patch is d-times differentiable.
Such a patch depends on all the neighboring faces. The port is not an intrinsic
quality of a triangular element; it depends on the number of common faces attached
to the element. If such elements are used to define a stiffness matrix, it is more
closely related to a finite-difference method than a finite-volume method. Further-
more, the order of the polynomial grows with the number of common faces. See
also Section 8.7. The stiffness matrix follows from the definition:
!2
∂2
Z X
(11.94) dxdy ∆ φpatch k =K .
∂(ai , bi , ci )∂(aj , bj , cj )
k
See Figure 11.7.

The patches used here will not sum to a constant function for constant coeffi-
cients, which is a common requirement. See Figure 11.8. The elastic energy will not
vanish if the patches do not sum to a constant. Normally, for higher-order elements
splines are used to construct a piece-wise lower-order polynomial which will sum
to a constant function. Such splines extend further over a mesh, generating a stiff-
ness matrix with more non-zero elements. Furthermore, for irregular triangulated
meshes the procedure to construct the patch is rather complex.
Using only the surface normal n(x, y) alone is a much simpler approach to
defining an element Hamiltonian. In Section 4.3 we have seen that a surface normal
determines the surface z(x, y) uniquely up to a constant displacement z0 . Further-
more, the linear interpolation of the surface normal between vertices automatically
yields a composite surface, consisting of surfaces on each element which has no
11.10. PLATE PORT-HAMILTONIAN ELEMENT 231
pacth functions
sum of patches
Figure 11.8. The patch functions may not sum to a constant

function, as is the case for the lower set with 16(z−2)2z 2 patches on
z ∈ [0, 2]. The upper set of patches consists of piece-wise 3z 2 − 2z 3
on z ∈ [0, 1] and 3(2−z)2 −2(2−z)3 ) on z ∈ [1, 2]. A spline similar
to these patches, with a continuous derivative, would consists of
three or four polynomial pieces.[Ahlberg et al., 1967]
boundary energy. We only have to define, or approximate, the integration around

the edge of the element which is conservative, such that z(x0 , y 0 ) − z(x, y) does
not depend on the path of integration between the points (x, y) and (x0 , y 0 ). This
is the integrability condition on the finite-dimensional functional subspace. See
Figure 11.9.
For the biharmonic approximation of the plate it is assumed that the deviation
from z = 0 is relatively small. Hence likewise we can approximate the normal
n(x, y) ≈ n = (1, nx (x, y), ny (x, y)), such that z is given by the integral:
Z x2 ,y2
(11.95) z(x2 , y2 ) − z(x1 , y1 ) = − nT dl(x, y) ,
x1 ,y1
where l(x, y) is a line element along the path from z(x1 , y1 ) to z(x2 , y2 ). The
integrability condition reduces to:
(11.96) ∂ x ny − ∂ y nx = 0 ,
such that all closed contours are zero.
If nx (x, y) and ny (x, y) are piecewise linear functions on a triangle, it means
there is a single condition among the six variables defining the normal:
(11.97) ∂x ny −∂y nx = ∂x (cy +cyx x+cyy y)−∂y (cx +cxy y +cxxx) = cyx −cxy = 0 .
Hence, the surface z(x, y) on the triangle is defined by six variables in total: the
integration constant z(x0 , y0 ) at a particular position x0 , y0 , and five coefficients
cy , cyx = cxy , cyy , cx , cxx for the normal.
The elastic energy in terms of the expansion of the normal n is given by:
1 1
Z
(11.98) Eelastic = dxdy(∇ · n)2 = VolΩ (cxx + cyy )2 .
2 Ω 2
surface
normal integrability
mismatch
z
y
x
Figure 11.9. The implicit definition of a surface, through the

surface normal n(x, y), may not always define a surface. The inte-
grability condition must be satisfied. Otherwise, the integral along
a closed contour will not end at the same height z as it started
from.
The configuration, the reconstruction of z(x, y) from n(x, y) and z(x0 , y0 ), is given
by: (x0 = 0, y0 = 0)
1 1
(11.99) z(x, y) = z(0, 0) + cx x + cy y + cxx x2 + cyy y 2 + cxy xy ,
2 2
which satisfies ∇z(x, y) = n. This is the small deviation approximation of the
mean curvature, Eq. 4.8. A single elastic energy term arises from the plate element,
consistent with the null-space, Eq. 11.88. The elastic energy is precisely the single
combination of monomials x2 + y 2 , which does not lie in the null-space to the
Laplace operator. Integrability and consistency are complementary views.
The difference lies in the associated port-variables, in terms of a set of sur-
face normals nx (x1 , y1 ), ny (x1 , y1 ), nx (x2 , y2 ), ny (x2 , y2 ), nx (x3 , y3 ), and ny (x3 , y3 )
at vertices 1, 2, and 3, where one component of one vertex normal is determined
by all the others, and a z-value at a single vertex is the sixth port-variable. Each
port is shared by several elements, yielding collocation relations; the same normal is
used for to define the surface z(x, y) for several elements. Since, the total number of
distinct ports, two per vertex, equals the total number of elements, from Eq. 11.92,
up to boundary effects, the one-dimensional state per element of elastic energy is
precisely covered by, on average, one distinct port variable per element.
The kinetic energy is associated with the velocity ż. The approximation ne-
glects the horizontal motion in the x − y-plane, therefore there is no kinetic energy
associated with it, and consequently no rotational energy. The kinetic energy is:
ρ 1X
Z
(11.100) Ekinetic = dxdy(ż(x, y))2 = ci Mij cj ,
2 Ω 2 ij
which depends explicitly on the shape of the triangle. The corresponding momenta
arise from the derivatives:
∂Ekinetic X
(11.101) pi = = Mij cj ,
∂ci j
11.11. ELECTROMAGNETISM ON A SIMPLEX 233
potential
E
q (charge) J (current)
Ω B
Figure 11.10. Static charges and currents outside the domain Ω

may cause source-dependent fields E and B inside the domain,
which are not dynamical, in the sense that the fields depend only
on the position and strength of the sources. The field lines are
closed in the sense that they enter and leave the domain Ω.
where ci is any of the coefficients z(0, 0), cx, cy , cxy , cxx − cyy , and cxx + cyy . The
element port Hamiltonian is:
1X 1
(11.102) H= pi Mij−1 pj + VolΩ (cxx + cyy )2 ,
2 ij 2
with the Hamilton equations of motion, and port variables z(0, 0), cx , cy , cxy , and
cxx − cyy in the null-space of the elastic energy. The mass matrix M is full rank.
11.11. Electromagnetism on a simplex

Electrodynamics, in terms of the electric and magnetic field D and B, is a rather
complex dynamical problem. The static fields are associated with the static charges
and the static currents respectively. These source-dependent fields are longitudinal,
i.e. ∇ × D = 0 and ∇ × B = 0, inside the simplex, and do not contribute to
the dynamics of the electromagnetic fields. See Figure 11.10. Appropriately, the
fields corresponding to given charges and currents should be separated from the
electrodynamics. They can be treated as ports localized at the position of the
charges and currents. This is discussed further in Chapter 12.
The null-space of the ∇× operator on the simplex, with a piece-wise linear ap-
proximation, gives an alternative view on the separation of dynamical and source-
dependent fields. The linear approximation of the fields are given by the twelve
values at the vertices; four three-dimensional vectors for D and likewise for B. The
dimension of the null-space of ∇× on this space is nine. Therefore only three vari-
ables correspond to dynamical, or internal energy on the simplex element. All the
other variables will also correspond to energy, but this energy may be determined
by just looking at the values on the boundary of the domain.
The fact that only three variables correspond to internal dynamics on a simplex
with four faces can be seen heuristically from the fact that the field lines must be
closed on the simplex, hence the flow E · n through each of the faces must sum to
zero:
X
(11.103) Areaf Ef · nf = 0 .
faces:f
Therefore one condition exists among the four face flows. In the finite-order approx-
imation, the condition changes in form but not in nature. For the simplicity we con-
sider a simplex with corners (0, 0, 0) − (1, 0, 0) − (0, 1, 0) − (0, 0, 1). The vector fields
at each of the corners are labelled respectively: E0 = E(0, 0, 0), Ex = E(1, 0, 0),
Ey = E(0, 1, 0), Ez = E(0, 0, 1). The linear interpolation yields a vector field inside
the simplex:
3
X
(11.104) E(x, y, z) = E0 +(Ex −E0 )x+(Ey −E0 )y +(Ez −E0 )z = E0 + E i xi ,
i=1
1 2 3
where x = x, x = y, and x = z. The surface normals n are −ex , −ey , −ez , and
√1 (ex + ey + ez ). The corresponding surface flows are for each of the faces:
3
(11.105)
1 T
e (2E0 + Ex + Ey ) z = 0 (cyclic{x, y, z})
Z
T
n E(x, y, z) = √ 6 z .
3 T
face 6 n (2E0 + Ex + Ey + Ez ) x+y+z =1
Clearly, the total flow in the linear approximation, through the boundary consti-
tuted by the four faces, is zero if, and only if, Eii = 0.
The curl operator ∇× maps this vector field onto the constant vector field, the
component i is given by:
(11.106) [∇ × E(x, y, z)]i = ijk Ejk .
The components E0i and Eii are in the null-space of ∇×. The latter, we have seen,
must always be zero, if there is no charge present and the flow is conserved.
The curl operators, part of the equations of motion of electrodynamics:
1
(11.107) ∂t D = ∇ × B .
µ
1
(11.108) ∂t B = − ∇ × D ,

are mapped onto R3×6 matrices for linear approximation on each simplex:
 2   2 
E3 E3
 E23     E23   1 
 3 
 E1  −1 1 0 0 0 0  3 
 E1  E0
(11.109) ∇ ×  1  =
   0 0 −1 1 0 0   1  =  E02  ,
 E31 
 E3 
 E2  0 0 0 0 −1 1  1 
 E2  E03
E12 E12
where the finite representation of Eij is each of the fields D and B.
The energy in the simplex is a quadratic function of the local vector field,
Eq. 11.104:
(11.110)
1 1 2 X 1 T 1 2 X 1 T
Z
E= E(x, y, z)2 dxdydz = E0 + E0 Ei + E + E Ej .
2 simplex 12 i
24 120 i i<j 120 i
11.11. ELECTROMAGNETISM ON A SIMPLEX 235
1 T
(11.111) =
E YE .
2
If the coefficient vector E is ordered like: E = (E01 , E02 , E03 , E32 , E23 , E13 , E31 , E21 , E12 ),
matrix Y ∈ R9×9 has the following structure:
 1 1 1

12 0 0 0 0 0 48 48 0
 0 1 1 1 
12 0 48 0 0 0 0 48 
1 1 1

 0 0 12 0 0 0 0 
48 48
1 1 1 
 
 0 0 120 0 0 0 0
48 240 
1 1 1

(11.112) Y=  0 0 48 0 120 240 0 0 0   .
 0 1 1 1
 1 0 48 0 240 120 0 0 0 
1 1


 48 0 0 0 0 0 120 240 0 
 1 1 1

48 0 0 0 0 0 240 120 0 
1 1 1
0 48 0 240 0 0 0 0 120
The discrete energy, Eq. 11.111, and the discrete equations of motion, Eq. 11.109,
yield the discrete electromagnetic power flow in and out of the simplex:
1 1 1 1
(11.113) Ė = DYḊ + BYḂ = DY(∇ × B) − BY(∇ × D) = DθB .
µ µ µ
The discrete curl matrix has only a non-zero block in the upper right corner. The
boundary operator θ has a complicated but symplectic structure, which follows
from the definition:
1 1
 
0 0 0 12 − 12 0 0 0 0
1 1
 0 0 0 0 0 12 − 12 0 0 
1 1 
 
 0 0 0 0 0 0 0 12 − 12 
 −1 1 1 1

0 0 0 0 − 24 − 48 0 
1  12
1
48
1 1 1 

(11.114) θ = 
12 0 0 0 0 0 48 24 − 48  .
µ   1 1 1 1 
0 − 12 0 − 48 0 0 0 48 − 24 
1 1 1 1

 0
12 0 24 − 48 0 0 0 48 

1 1 1 1

 0 0 − 12 48 − 24 − 48 0 0 0 
1 1 1 1
0 0 12 0 48 24 − 48 0 0
The rank of the boundary operator is rank[θ] = 6, compatible with the three-
dimensional null-space of ∇×.
Now we have all the ingredients to construct a port-Hamiltonian element. The
energy on the element is expressed in two diagonal blocks Y on the 18 dimensional
(D, B) space. The equations of motion is expressed in two off-diagonal ∇×, and
the boundary operator is θ. In the case of ordinary finite-element method only the
∇× operator is needed for the evolution of the system. In that case, for each node
value, e.g., Bx , the time-evolution is given by the mean of all the simplices common
to this node i:
δt X 1
(11.115) Bx,i (t + δt) − Bx,i (t) = ∇ × Ej ,
N j
simplicesj
where N is the number of simplices. The average due to the summation over
simplices destroys the power continuity and the divergence free form of the fields at
the level of the numerical finite difference. If the fields are smooth, and vary slowly
over the simplices these qualities are retained to a certain extent.
In the port-Hamiltonian element some of the variables Dij and Bji are dependent
variables and some are independent. The bilinear boundary operator tells us we can
construct six pairs of variables, one dependent and one independent, or one input
and one output. The remaining six variables, from the eighteen, are state variables.
Connecting the elements gives rise to a set of algebraic relations, which gives rise
to an interconnection network. Unlike in most FEM approaches, the manner in
which the elements are interconnected is not a priori fixed, and may depend on the
particular boundary conditions, electric or magnetic.
11.12. States from the null-space

In the case of a nonlinear Hamiltonian the state corresponding to a null-space
mode might be highly nontrivial. However, recovering these states will be important
to simplify the actual numerical implementation and simulations. They form the
linkage between the boundary control in terms of the null-space of the differential
operator D preyed from the Hamiltonian, and the internal dynamics, in terms
of local state variables. For example a null-space may correspond to an applied
voltage V , constant on a domain z ∈ [0, 1]. The charge density q0 (z) may be used
as state corresponding to the null-space. It is related implicitly to the applied
voltage through:
(11.116) V = δq H[q0 ] = constant .
The state q0 may be highly nontrivial to recover. However, it forms the linkage
between the input, the boundary conditions in terms of the physical variables, and
the mathematical model of the internal dynamics, in terms of the state variables.
11.13. Final remarks

The construction of energy-consistent elements and their ports for a general
nonlinear FEM has a number of requirements. Consistent equations of motion, to
study dynamical FEM, which is really the purpose of the whole exercise, are based
on non-singular operators with a unique zero state. Therefore the greater part
of constructing these elements is the counting of degrees of freedom, equations of
motion, and conditions. The linkage with the continuous theory and, for example,
continuity equations goes via integrability conditions, which guarantee that the
variables used correspond to existing configurations and unique energy states.
The construction of finite-elements to be used in a large network is very much a
local approach. The global conservation and balance laws rely on the proper imple-
mentation of algebraic relations and accurate time-integration. The next chapter
the approach is taken from the other side. The global properties are linked to a
coordinate system based on input and output. The internal dynamics and material
properties result only in local variations. In a sense, this approach is the simplest
geometric formulation of the core model, described in Section 7.6.
CHAPTER 12
Force lines and velocity fields
Geometry is the art of correct reasoning on incorrect figures.

[G. Pólya, How to solve it]
A global or system view of several components in conjunction necessarily has
less focus on the details. In a port-based approach the energy flow and the balance
relations throughout the system are the first concern. Continuous models, on the
other hand, are based on the local, small-scale balance and flow. The global prop-
erties in a continuous models follow only from the local analysis. In order retain the
global view within the local models an approximate global flow and balance can be
established first, where the local details will follow from more detailed analysis. An
experienced modeler may pick a dominant mode or static state. In this chapter we
use results from harmonic analysis and optimatization to establish global geometric
system views. See Figure 12.1.
The geometry of an extended object is usually considered independent from the
fields, such as displacements, currents, forces, and potentials, defined on the object.
They are, however, intimately related through the constitutive relations, implicit
in the Hamiltonian. In some cases, for special geometries, coordinate systems are
used which simplify the calculation of static solutions, such as polar coordinates for
spherical problems.
The input and output of a system; the sources of the fields, are instrumental
in determining the relevant field configurations. Hence a coordinate system on
an object, directly based on the fields, will be a geometric representation of the
Hamiltonian for the given interactions through the input in output. Simply said, for
a distributed system, with a single input and a single output, the parallel coordinate,
based on the field, maps the distance from input to output in a physical sense. The
surfaces perpendicular to the lines from input to output are the perpendicular
coordinates in the system. In electric systems they are the potential surfaces. The
lines are the streamlines, or integrals, of the vector field. In general, we make little
distinction between the vector fields and the lines.
Take for example a strained object due to two opposite forces applied at dif-
ferent positions. The force is conserved. On each closed surface enclosing only one
fixture position, the same amount of total force acts. In the language of Faraday
and Maxwell: the force lines are closed. The Poisson ratio determines the spread
of the forces lines between the two fixtures. The density of force lines is equal to
the local pressure on the surface.
12.1. Electrostatics
The energy of a static electric field can be calculated in different manners. The
total energy density of the field E will give the same result as the charge density ρ
237
238 12. FORCE LINES AND VELOCITY FIELDS
J
Figure 12.1. Given that for a conserved, incompressible flow, the
same amount J flows in as out, a global solution given by the solid
lines is first constructed. The details of the flow in each of the
components may change the position of the streamlines, but not
the total flow. These are the allowed degrees of freedom of the
internal dynamics, for example, those indicated by the dashed lines.
at the given potential V , or the Coulomb law between charges: [Jackson, 1975]
(12.1)
1 1 1 ρ(z)ρ(z 0 )
Z Z Z
2
Eelectrostatic = dz0 E (z) = dzρ(z)V (z) = dzdz 0 .
2 2 2 4π0 |z − z 0 |
One follows from the other through partial integration and Gauss or Poisson law:
1
(12.2) ∇·E= ρ .
4π0
(12.3) −∇V = E ,
and the Green function of the Laplacian:
f (z 0 )
Z
(12.4) f (z) = ∆z dz 0 .
4π|z − z 0 |
However, a constant can be added the potential V . This constant is the inte-
gration constant from the partial integration, and commonly set such that:
(12.5) lim V (z) = 0 .
|z|→∞
The surface of the domain is therefore set at infinity: |z| → ∞.

Hence, for a field E the energy density can be determined by looking only at
the charges ρ and the potential at those positions. In most electrostatic problems
either the charges ρ or the potential V is given. The energy of the system is the
amount of work to bring all the charges from infinity to the given configuration.
The relation between the charge and the potential is given by the capacity C
of the system, which is completely determined by the geometry. [Pólya and Szegö,
1951] The potential at the boundary is usually associated with conductors connected
12.2. FORCE LINES 239
to an external source, and therefore piece-wise constant. The electrostatic co-energy

can be given in terms of the potential as well:
1 2
(12.6) Eelectrostatic = V .
2C
Since it is a quadratic function the same capacitance C arises as in the energy,
expressed in terms of the charges. The capacity C can be determined from the
solution of the Laplace equation for piecewise constant boundaries, from the con-
stant potential at the boundary. The Laplace equation is an ideal starting point to
characterize the geometry in terms of the boundary.
An arbitrary scalar field f with the appropriate boundary conditions, i.e., port
values will give an upper and lower bound for the capacity. [Pólya and Szegö,
1951] The Dirichlet principle is an upper bound for the capacity with f = V at the
boundary ∂Ω:
1
Z
(12.7) C≤ d3 zk∇f k2 .
4πV 2 Ω
While, for an arbitrary divergent-free vector field f the lower bound is given by the
Thomson principle:
1 1
Z
(12.8) ≤ d3 zkf k2 ,
C 4πQ2 Ω
where the charge Q is given by the flow through the boundary ∂Ω:
1
Z
(12.9) d2 znT f = Q .
4π ∂Ω
Hence, by defining divergent free vectors fields f and scalar fields f the system
is already globally characterized, with the use of energy and variational principles.
In the next sections we will pursue these principles further. A divergence free
vector field, with only in flow and out flow at the boundaries or ports represents
any balanced quantity, such as mass, but also force, in the system. Using the
conservation law in the analysis will stabilize the numerical schemes. The reason
this is normally not pursued is because such conservation laws are global; the whole
domain is connected, a change at one place affects changes at other places. Through
the use of coordinates based on a particular choice of divergent free field, i.e., the
harmonic field, the coordinates can be used as level set for solutions to static,
dynamic, and optimization problems.
12.2. Force lines

Maxwell and Faraday introduced the concept of lines of forces, to treat the
electrostatic theory in analogy with fluid dynamics. [Maxwell, 1891] The lines of
forces are the lines along which a test charge will move in the electric field, perpen-
dicular to the surfaces of equal potential energy. In the more general case forces are
still balanced, but there might not be an underlying potential V responsible for the
force Fpotential = q∇V . Euler and Frobenius integrability conditions set rigorous
conditions on the fields in order to have an underlying potential. [Frankel, 2004] See
also Section 4.4. We will investigate this general case. Our purpose is to decompose
the interaction of an object with its environment in a global, transmission part, and
local, material effects. The global part is expressed as a harmonic field; solution of
the inhomogeneous Laplace equation, with its useful properties. The local parts of
the force field depend on the material.
Given an object with two fixtures, or ports, with equal but opposite force along
the line between the fixtures, such that the object is at rest. Through each surface
cutting the object in two, separating the two ports goes an equal amount of force
lines. The force of the port is distributed in the surface, and depending on the
geometry and material properties the force is more or less focused on the shortest
path between the two ports. The solution of the Laplacian, using the two ports as
sources of force, like charges, the force lines connect the two ports, like the electric
field. The conserved force lines are a divergent-free field. They correspond to the
stress tensor (Eq. 6.28):
3
X
(12.10) ∂i Sij = Fj ,
i=1
where Fj is the applied force, or pressure, at the fixtures.

Hence they can be constructed a force potential f driven by the input and
output forces Fin/out :
(12.11) −∆f = −∇ · (∇f ) = Fin/out .
The surfaces with f = constant define the parallel coordinates; the relative distance
between the ports. The force lines f associated with the Laplace equation spread
out. Near the ports the density of lines, the pressure, is higher, as there is a smaller
area of the sphere supporting the total applied force.
The stress tensor can be decomposed in terms of the force lines ∇f and the
internal balance tensor G, using the Hodge-Helmholtz decomposition (Eq. 4.39):
(12.12) S = ∇f + ∇ × G ,
or in terms of their components:
(12.13) Sij = ∂i fj + ikl ∂k Glj ,
where ijk is the fully anti-symmetric Levi-Civita tensor. The internal balance
tensor G are closed force lines, for each component of the force vector. See Fig. 12.2.
The internal balance fields can never contribute to the force at the ports, since
∇ · (∇ × ·) = 0. This makes the force lines ∇f with ∆f = Fin/out the natural choice
of the principal component of the coordinate system, independent of the material
or other constitutive relations. A similar decomposition was also proposed by Weyl
. [Weyl, 1940]
The force lines are associated with the harmonic approximation to a problem.
In the case of a fixed wave velocity, it would be the solution associated with this
wave velocity, propagating along the force lines, for a homogeneous mass density.
The internal balance may give rise to auxiliary modes. In the case of the wave
equation the force potential generates the static displacement φ, to an applied
force F:
eT f
(12.14) κ∆φ = F ⇒ φ = z .
κ
In the case of inhomogeneous elasticity modulus κ(x) it should already be taken
into account in the Laplacian:
(12.15) ∆φ = ∇ · (κ(x)∇φ) ,
12.3. MICROSCOPIC BALANCE 241
force lines coordinates internal balance force field
+ =
potential surfaces
force lines
port
Figure 12.2. A sketch of force lines ∇f , associated with the force

potential f , for a material with a small Poisson ratio ν, such that
the stress is localized along the axis connecting the ports. The
internal balance G compresses the force lines together. These are
only the longitudinal components of the force f = f ez .
since the energy is defined as κ(x)(∇φ)2 , which under partial integration leads to
the familiar form.
12.3. Microscopic balance

The force lines are not necessarily parallel to the force direction f , unlike the
force lines generated by a potential. The force direction is only the transmission of
the input force to the output force. Hence only the force line on the straight line
connecting input and output is a balanced force line, in the sense that force line
and force direction are parallel:
(12.16) rank[∇f ] = 1 .
The microscopic balance is automatically satisfied in the formulation in terms of
the vector fields ∇f and ∇ × G. The relation between the two is for a greater part
determined by the constitutive relations. However, some generic results exist for,
for example, elasticity. Such conditions find their basis in the invariant relation
between the stress and the strain. For example, results may not depend on the
overall rotation of the object, or its reflection. As a consequence, the strain tensor,
ij , is symmetric. If the work W is defined as the scalar product of stress and
strain, the stress tensor can be symmetric Sij = Sji too:
1X
Z
(12.17) W = dzij Sij .
2 ij
The asymmetric part will not contribute to the energy. However, this is not a
fundamental result, it holds only for isotropic materials. Materials exist with are
intrinsically chiral, i.e., left-handed or right-handed, one has only to think of screw-
like molecules, typical of biological molecules, and ferrites. For such systems few
microscopic theories of the continuous systems exist. [Maugin, 1988]
Force out
constant
potential
perpendicular force diffusion
Force in
Figure 12.3. The force from input to output generates a con-

served flow, indicated by the closed force lines. On any surface
intersecting the force lines, the material properties will cause a
conserved virtual dynamics, i.e., force diffusion, of the force lines.
On top of that will internal force balance G generate forces in
perpendicular directions.
On the other hand, energy must be an invariant property, independent of the

observer or its coordinate system. Hence the energy must depend on the invariants
made with f and G, where G is gauge invariant, meaning the energy may not
change with the addition of a conservative field ∇χ to G:
(12.18) E(f , G) = E(f , G + ∇χ) ,
where χ is an arbitrary non-singular scalar vector-valued function.
On a surface f = constant, the force line ∇f is perpendicular to the surface.
Furthermore, if the surface encloses the input or the output, the total transmission
(0)
force φf = nT ∇f is the same as the input and output force. The transmission force
is conserved. The microscopical material properties can only change the distribution
of force φf on the surface f = constant. See Figure 12.3. Hence, although both
(0)
the surface f = constant and the harmonic solution φf are global and possibly
complex to construct, the variation G due to material properties is a conserved
flow on each surface f = constant. The redistribution of the force over the surface
(0)
can be treated as a forced diffusion, where the harmonic force distribution φf and
the distribution on neighboring surfaces f ± δ are the source terms. The harmonic
solution already yields an appropriate coordinate system for this analysis.
As a simple two-dimensional example we study the force lines in a plane between
two point sources at x± = (±1, 0). In the theory of two-dimensional elasticity,
different forms of a generating potential, similar to f are proposed to bring the static
deformation equations to a biharmonic form. [Soutas-Little, 1973] In the literature
the input is generally not part of the consideration in constructing appropriate
fields, the separation in longitudinal and rotational fields is.
12.3. MICROSCOPIC BALANCE 243
3
η sink
source
ξ
2
-1
-2
-3
-4
-6 -4 -2 0 2 4 6
Figure 12.4. Bipolar coordinates for α = 1. The solid lines η =

constant are the force lines, integrated from the vector field ∇f .
The dashed lines ξ = constant are the lines with constant potential
f = constant. Both coordinates are circles in the plane.
The input is located at x− , the output at x+ . The harmonic solution f yield

the bipolar coordinate system. See Section 2.10. The harmonic function is the sum
of the two potentials, or Green functions, Eq. 4.45, in two dimensions: [Spiegel,
1964]

ei x − x+ ei |x − x+ |
(12.19) f = Re log = log ,
2π x − x− 2π |x − x− |
where the unit force is standing in an arbitrary direction ei . The imaginary part of
the complex function can be used as angular coordinates on the lines f = constant.
See Figure 12.4. The density of force lines on every surface is therefore unity. The
bipolar coordinates (ξ, η) can be seen as the parallel xk = ξ and perpendicular
coordinates x⊥ = η of the system. The parallel coordinate connects input with
output, while the perpendicular coordinates, depending on the dimension of the
space, is the span of all connections.
The simplest variation G of the force field is a mapping of the angular coor-
dinate η onto itself. The closed contours of G are therefore the same as the lines
η = constant running forward and backward the same amount. The most extreme
mapping −π ≥ η ≥ π to −π ≥ ηnew ≥ π is the step function:

1
(12.20) ηnew = 2π θ(η) − ,
2
where θ(x) = 1 for x ≥ 0 and θ(x) = 0 for x < 0. This will focus all the force along
the straight line connecting the source at x− and the sink at x+ . The symmetric
mappings ηnew (η) with ηnew (0) = 0 and ηnew (−η) = −ηnew (η) generate a whole
2π
η new
0
0
η old 2π
2
Figure 12.5. A sketch of a simple one parameter variation of force

lines in a material. The underlying harmonic field ensures force is
transmitted from source to sink. The variations are perpendicular
to the force lines.
class of force configurations, which focus more or less along the straight line. See
Figure 12.5. The interesting part is to derive efficient methods to find ηnew (η) for
different materials and input.
The bipolar coordinates may also serve for a near parallel FEM network rep-
resentation. Take for example the heat flow from x− to x+ , or in practice, a small
radius around the the source and sink. Since the harmonic field f suggests the
main flow in the direction of the gradient ∇f , the network may, in first instance,
be approximated by the parallel lines η = constant. In the case of isotropic ma-
terial, where the cells can be approximated by a uniform grid in (ξ, η)-space, the
capacitance of each cell ∆η × ∆ξ is given by:
ρ∆η∆ξ
(12.21) Ccell = 2
,
(cosh ξcell − cos ηcell
average average )
where ρ is the capacity density, ∆η∆ξ the surface element, and the denominator
follows from the Laplacian, Eq. 2.86. The resistance is in the same manner related
to the Laplace operator in the curved, bipolar coordinates. [Magnus et al., 1966]
A small cell is a wide as it is long. For a square cell, the resistance is independent
of the size. Hence all the resistances between each pair of neighboring cells i and j
are the same:
Tj − T i
(12.22) J = −λ∇T ⇒ Iij = yji λ = λ(Tj − Ti ) ,
xji
where J is the heat flux, λ the plane conductivity, xji the distance between cell
centers, and yji ≈ xji the length of the common edge. As an interesting conse-
quence, the stationary harmonic flow depends only at the temperatures of at the
source and sink, analogous to the energy of an electrostatic field in Eq. 12.1. The
time-dependent flow can be analyzed on a square grid (ξ, η), with a homogeneous
network of resistors λ−1 , and a variation in capacitances Ci , given by Eq. 12.21.
See Figure 12.6.
12.4. SHAPE OPTIMIZATION 245
source sink
FEM
bipolar
coordinates
source sink source sink
Figure 12.6. The near-parallel structure arises through a coor-

dinate transformation to bipolar coordinates. The network repre-
sentation of the FEM approximation has equal resistors, λ−1 , for
the square 5 × 5 geometry. The capacitors depend on the location
(ξ, η). The cross-connecting resistors can be ignored for small vari-
ations around the static, harmonic solution for heat diffusion or
electrostatics.
12.4. Shape optimization

The design of the shape of a domain Ω optimal for the transfer of force is a well-
known problem in engineering mathematics, which dates back to the days of Ha-
damard. [Hadamard, 1908, Courant and Hilbert, 1961, Pironneau, 1984, Cherkaev,
2000, Allaire et al., 2004] The optimization functional to be minimized can be the
maximal rigidity:
Z Z
(12.23) J[Ω] = dzuT f + dzuT F ,
Ω ∂Ω
where u is the displacement and f and F are the applied body force and surface
force.
The functional is the sum of the work done by the forces F applied to the
boundary, and the body forces f . Many other functionals exist for a variety of
criteria. In the case of vibrations, the structure defined by its outline Ω must also
be light to reduce the vibrational modes. This can, for example, be achieved by
adding a volume term to the functional
Z
(12.24) Jvolume [Ω] = α dz ,
Ω
where α is the relative contribution of the volume term. This is a common proce-
dure, [Allaire et al., 2004] however, it will not take into account critical structure,
or an eventual mode shape of the vibration.
The functional J[Ω] is minimized for the shape Ω of the object:
(12.25) min J[Ω] ,
Ω
which will not always lead to a well-defined problem. The functional derivative,
necessary for the optimization, with respect to the shape with the use of the sur-
face normal n. A level-set method may be the practical implementation of such
derivative.
The boundary ∂Ω may be defined through a level-set function ψ(z):


 ψ(z) = 0 z ∈ ∂Ω
(12.26) ψ(z) = ψ(z) < 0 z ∈ Ω .
ψ(z) > 0 z 6∈ Ω

The outward surface normal, necessary for variations with changes in shapes, is
given by the gradient of the level set ψ at ψ = 0:
∇ψ
(12.27) n= .
|∇ψ|
The optimization procedure using the level-set function is a rather abstract numeri-
cal method, which uses brute force computation of the functional and its variational,
or functional, derivative.
Using force lines in optimization could lead to a more heuristic approach, and,
since the force lines are conserved, it is intrinsically stable. The initial force lines in
the complete domain may be bundled to smaller areas, decreasing such the volume
term, Eq. 12.24, in the minimization functional, Eq. 12.23.
12.5. Following the force lines

Studying microscopically forces and the corresponding fields, which, together,
yield the energy density and power continuity, is not a common approach to the
study of continuous systems. Forces are not invariant concept in continuous sys-
tems. One person may talk about pressure, the other about force, discussing the
same phenomena. The energy is, on the contrary, an invariant concept. This is
only one of the reasons one has put aside force as central concept in the theory of
dynamics.
However, when dealing with actuated systems, force is an important concept.
The balancing of force is less trivial than for isolated systems. Most components
in a system act to transfer force, or effort, or potential, rather than anything else.
This transfer is described by the force lines. Components of systems are elements in
chains, and the geometrical concepts like flux lines, such as force lines and stream
lines bring this to light. The microscopical balance bring the material properties
into play.
CHAPTER 13
Conclusions
The depth of thought which goes into the formation of mathe-

matical concepts is later justified by the skill with which these
concepts are used.
[E. P. Wigner, The unreasonable effectiveness of math-
ematics in natural sciences]
Unfortunately, traditional modes of analysis have often placed
constraints upon the forms in which such [system-environment]
communication can occur.
[Henry M. Paynter, Analysis and design of engineering
systems]
This thesis addresses a single question:
What is an appropriate model of a continuous physical compo-
nent for analysis and simulation in larger systems?
This question was answered at many different levels, from conceptual, through
mathematical and physical, to numerical. Models are constructed and reconstructed
for the purposes at hand. For engineering purposes the function of a component
determines its model. As a consequence, the nature of the interaction with the
environment plays an important role, since a component’s function is established
by its interaction.
In this thesis we constructed finite-dimensional physical models of continuous,
or infinite-dimensional, components, for the use in larger systems with inputs and
outputs. An important part of such models is the interaction, through the common
boundary. The relation between the bulk and the boundary has many different
aspects which are discussed in this thesis.
First of all, the relation between bulk and boundary is also the relation between
the intrinsic properties and the interfaces, which is a system, or operational, rela-
tion. If the interface does not transmit a certain quantity or quality, this quality or
quantity does not have to appear in the component model, at least not explicitly.
Second, the relation between bulk and boundary is based on the geometric re-
lation between domain and boundary, which, with the help of differential geometry,
defines the relation between energy and energy flux, and, more generally, between
any conserved quantity and the corresponding conserved flow.
Third, the relation between bulk and boundary, is, in terms of the underlying
partial differential equations, the relation, or rather separation, in terms of ini-
tial conditions determining the states of the internal dynamics, and the boundary
conditions essential for connecting components together. The physical boundary
conditions follow from the energy flux through the boundary. The nature of the
247
248 13. CONCLUSIONS
boundary conditions in terms of the system description may vary with the circum-
stances. In some cases they will give rise to state variables associated with boundary
values, in other cases the boundary variables are constraints, or even parameters of
the model.
Finally, the relation between bulk and boundary is the relation between the
microscopic and the macroscopic properties, for example, between the charge den-
sity and the potential, and between the stress tensor and the force exerted on the
boundary. The internal and the external variables, used, respectively, in the partial
differential equations and the lumped models connected to the components, have
an uneasy relation to another. The consistency of the microscopic model in the
larger, system setting, follows from isolating the equivalents of the lumped model
parameter, in these models.
Modelling continuous systems remains difficult, mainly because the theories of

continuous systems, such as electrodynamics and elasticity, are difficult due to the
variety of phenomena they bring together. The internal dynamics draws all the
attention. Setting appropriate boundary conditions, for a start, and allowing a
component to interact through the boundary, in a more general system setting, are
complicated aspects, which receive little attention in the standard theory of con-
tinuous systems. For engineering purposes, only limited operations and variations
within these vast fields are relevant. Only the boundary conditions and input can
select or limit the dynamics to the intended function. The initial conditions play a
minor role.
The physical principles such as energy conservation and other conservation laws,
help to guide our attention toward the interface problem. The port-based modelling
is thereby be extended to such components. One nontrivial aspect of this extension
is the typical nonideal nature of continuous systems. Whether the appropriate
description of a component is more like a mass, a damper, or a capacitance, depends
more on the operational setting, i.e., the type and time-dependence of the boundary
conditions, than on the component itself.
The tendency exists to think in the terms of the methods and tools one uses.
Components described by partial differential equations are very much analyzed in
the language and with the methods developed for the intrinsic understanding of
these isolated components themselves. The relevant, modular features of a com-
ponent within greater schemes and systems are not the same as the features of a
component in an isolated setting.
The modular approach has brought electronics great advantages. Building a
large design in terms of functional blocks with well-defined properties seems to be
a successful route for rapid design, analysis, and prototyping. The improvement
and replacement of a part can be seen in the light of the whole, often with little
unknown interference between the components.
The intrinsic properties of a component in terms of its interactions does typi-
cally not exist. It would correspond to combining all the possible generic situations
and boundary conditions which may occur and select from them the largest set of
essential variables for which a core model is be constructed. In this thesis these
intrinsic properties and variables are inferred from the operators which generate
13. CONCLUSIONS 249
the dynamics. In some practical situations some the degrees of freedom may not
be set as initial conditions but arise as boundary conditions, or input.
In setting up a language for components in terms of their interactions we have
looked for inspiration towards the lumped models. Simple inputs will generate
responses that can be interpreted as the analogues of a lumped characteristic. The
stationary states, and some generalizations called quasi-stationary states, are the
result. They are non-trivial internal states, which are the direct consequence of
how a component interacts with its surroundings. They are not states in the sense
that they require some initial condition, to uniquely specify the evolution of the
system. Chapter 7 is central to this investigation.
Furthermore, we have tried to set up a network view based on the force balance
throughout the system. The conservation, or balance, laws are inherent to a global
physical view. Moreover, they are usually the complex physical laws which are
traded in for simple microscopical laws, which can be implemented easily in some
local finite-element scheme. For example, the whirls and tendrils of fluid motion will
draw away attention from the fact that the same amount of incompressible fluid that
goes into a container, or piping system, must come out as well. Setting up a quasi-
stationary, lumped-equivalent flow satisfying the conservation law as starting point
for the microscopical analysis, this analysis can be local, since the global laws are
given by the nontrivial starting state based on the interaction with the surroundings.
Only a preliminary analysis of a geometric view appears in Chapter 12, while the
aspects of these principles, generically referred to as transmission, are scattered
throughout the thesis, however, in particular, in Chapters 8 and 10.
The interface between different components is not naturally lumped, unlike the
interface between a continuous component and a lumped component. The local
product of effort and flow must be separated into a lumped, sum flow, satisfying
the balance law, and the consistently lumped, mean effort. The product should
still satisfy the power continuity. For FEM this restricts the numbers of internal
and boundary variables. In a given finite-dimensional approximation based on
states or configurations, the lumped power-continuity across the interface is only
satisfied approximately. Some principles exist, such as Saint-Venant principle for
elasticity, or the intermediate distance behavior of the Green function, or multi-
pole expansion, which justify these approximations. The stationary state forms
often the appropriate basis of the mean or lumped interface variable. Some partial
answers are provided in this thesis, in particular in Chapters 6, 9, and 10.
The language of most continuous systems is not just based on linear systems,
but on static linear systems, except for the vibrational analysis. For driven internal
dynamics a modelling language seldom exists, although in operational settings it
is the question of interest. Only a small number of well-known phenomena have
there own model, such as wind-induced waves, flow-induced vibrations, and thermal
convection. The practical use is in these cases obvious. However, also due to the
inherent nonlinearities in induced dynamics, such problems are often avoided in
modelling. From a functional design and optimization viewpoint, it is hard to
understand how they can be avoided. The analysis of FEM models of continuous
systems based on input characteristics is discussed in Chapter 9.
In this thesis we have shown that not the vibrational modes, but the core-model
modes are the dominant features of continuous components with input and output.
250 13. CONCLUSIONS
These modes can be linked to the stationary and quasi-stationary states for different
types of static input. The input as boundary conditions and boundary values are not
given in terms of the states, but in terms of the efforts, or variational derivatives of
the Hamiltonian. The consistency between initial conditions, in terms of states, and
physical boundary conditions, in terms of the Hamiltonian, is therefore strained.
In order to guarantee consistency the model is separated into two parts: the core
model for the boundary conditions, and the model for the internal dynamics.
The port-Hamiltonian follows from the ordinary Hamiltonian through a change
of variables, from configuration-like variables to energy-like variables, which opens
the, initially closed, system for interaction through the boundary. In Chapter 5 this
was done from a partial differential equations point of view, while the variational
viewpoint was taken in Chapter 7. The differential operator used in the variable
transformation determines the type of interface at the boundary. Three aspects
of port-Hamiltonian systems should survive in the numerical setting, for power
continuity. The exact action of the differential operator on the finite set of modes.
The exact integration of the model port-Hamiltonian density, associated with these
modes, over the domain. Finally, also the null-space of the differential operator,
yielding the degrees of freedom which are necessarily distinct interaction variables,
must be retained. This has been discussed in a general setting in Chapter 7, while
a number of practical cases were constructed in Chapter 11.
So far, the different integrability conditions, given in Chapter 4 and used in
Chapter 11, stood in the way of a uniform description of different port-Hamiltonian
models for open systems. The recipe to construct a port-Hamiltonian finite element
of arbitrary order and on an arbitrary domain, with a variety of boundary condi-
tions, has been implemented in a number of cases, in Chapter 11. The aspects
mentioned in the previous paragraph play an important role. The central place
the differential operator in the port-Hamiltonian has taken, in Chapter 7, is one of
most important contributions of this thesis to the study of the port-Hamiltonian
approach. This central role has been important for the construction, rather than
the reconstruction, of power-continuous continuous system components.
All is grace.
[Georges Bernanos, Diary of a country priest]
APPENDIX A
Literature
A.1. Functional analysis

Variational principles, partial differential equations, and numerical methods all
have been placed on a solid foundation by the use functional analysis. The shining
example of its power is still the classic work by Hilbert and Courant [Courant
and Hilbert, 1961]. More mathematical textbooks are by [Riesz and Nagy, 1955],
[Yosida, 1980], and [Zeidler, 1995]. In the latter some ground is covered to treat
nonlinear theory, and Banach spaces [Berger, 1977]. Applications can be found in
[Lebedev and Vorovich, 2000].
More specialized literature can be found on variational methods [Gel’fand and
Fomin, 1963, Mikhlin, 1964, Washizu, 1968, Carathéodory, 1935, Oden and Reddy,
1982], approximation theory [Kantorovich and Krylov, 1958, Funaro, 1992], and
optimal control [Lasiecka and Triggiani, 2000, Pironneau, 1984, Lions, 1970].
A.2. Partial differential equations

In the study of hyperbolic partial differential equations [Sommerfeld, 1992,
Hörmander, 1990, Gårding, 1998, Garabedian, 1964, John, 1982, Sakamoto, 1982]
there are many different approaches. We can distinguish a number of main direc-
tions. First, the study of two-dimensional problems (one space dimension and one
time dimension). This is a mainly mathematical field with many separate results,
but not really a field of general methods. Second, the study of linear equation with
constant coefficients. Here the theory of Fourier-Laplace transform proved very use-
ful, and many general methods exist. Third, the linear equations with coefficients
depending on space and on time. In this case the analysis falls apart in two classes:
local analysis and global analysis. The local analysis can prove the existence and
uniqueness of a solution in the neighborhood of the initial surface, while the global
analysis is restricted to problems with some symmetry. For neither case effective
numerical methods for the recovery of a solution exist.
In the numerical field one can separate approaches from strictly numerical which
require only little fore work, such as FEM, to almost completely analytical, where
only for the representation of the result functions need to be evaluated numerically.
Another separation in numerical approaches to PDE’s would be to distinguish vari-
ational approaches, such as Galerkin, which are inherently stable against direct
methods, with limited evaluations which aim at a direct simulation of the evolution
of the system.
The study of boundary condition,[Sakamoto, 1982] or the mixed problem, [Ha-
damard, 1923] or the initial-boundary value problem, of hyperbolic equation is
beset with difficulties. Several numerical techniques are applied to such problems,
[Lax and Wendroff, 1960, Morton and Mayers, 1994, Sod, 1985] and few results are
251
252 A. LITERATURE
scattered in the literature. The Lopatinski condition [Krantz, 1992, Gårding, 1998,
Hörmander, 1983] arises in many results.
A.3. Numerical methods

The most powerful and versatile of all numerical methods [Press et al., 1992,
Hildebrand, 1987, Stoer and Bulirsch, 1980] are those of linear algebra. Solutions
of sets of linear equations are easily found, and many problems, such as eigenvalue
problems [Wilkinson, 1965, Golub and Van Loan, 1996] fall into clear classes, which
specialized numerical methods. It is not surprising that most general problems are
approximated in such a way that a numerical linear problem arises.
Closely related to linear problems are Hilbert spaces, of which the space L2 ,
the Lebesgue space [Rudin, 1966] is the most common one. The norm is defined by
the integral, which is a linear functional of the space of functions. Hence, most of
the work before arriving at the eventual set of linear equations consists out of the
determination of appropriate integrals. There a number of methods for integration,
if the functions and the domain are sufficiently smooth and compact. If that is
not the case, specialized methods of integration are required, and the discrete,
numerical information of the integrand must be adapted to the problem at hand.
The smoothing of data is often achieved by interpolation by means of polynomial
fitting, such as spline fits. [Oden, 1972, Prenter, 1975, Funaro, 1992]
For nonlinear problems the situation is much less clear, different numerical
approaches are set to solve different problems, and are more or less tuned to a
particular class of problems. In many cases the problem, or solution, must be
augmented with additional information, such as a range of values for the solution,
of derivatives of functions, to be minimized or inverted.
For simulation of ODE’s and PDE’s a wide collection of methods exists. They
aim at the maximal accuracy for a particular step size. For PDE’s there are a num-
ber of bounds on the step size given the quality of initial conditions. A specialized
literature on conserved quantities, such as energy, retained during the simulation
exists. [Hairer et al., 2002, Serre, 2001] In the physics community the problem is
often reformulated to make such conservation laws explicit.
A.4. Geometry
A continuous problem is defined on a space with a notion of distance, given by
the metric. Euclid dominated geometry till the middle of the nineteenth century.
After the introduction of differentials through variational approaches to optimiza-
tion problems by Johann Bernoulli, Euler, Gauss, and others in the eighteenth
century the road lay open to reformulate geometry in the new language. [Mc-
Cleary, 1994, Spivak, 1970] The work of Riemann and Gauss are some highlights
of this programme. [Frankel, 2004] This was the birth of differential geometry, or
geometry in the small. The infinitesimal distance in differential geometry was and
is a concept beset with difficulties. Topology gave a more generic answer to the
notion of distance than geometry could. [Lefschetz, 1949] The two fields have up
to now still a uneasy relation to another. At the other end of differential geometry
the existence of shapes and invariants were questioned. Global solutions [Chern,
1989] of objects with negative curvature were sought, as were embeddings of man-
ifolds, or hypersurfaces, with metric in Euclidean space of higher dimension, such
as the surface of the sphere with constant positive curvature is embedded in three
A.6. DYNAMICS 253
dimensions. Manifolds were characterized by integrals of functions or operators

defined on the manifold. Invariant manifolds, group theory, varieties, and index
and characteristic numbers play an important role in algebraic geometry. [Hodge,
1941, Whitney, 1957] Algebraic geometry can be seen as the generic study of global
characteristic using tools as triangulations of surfaces and polynomial mappings.
A.5. Modelling
To arrive at a PDE describing a system, one requires understanding of the
underlying mechanical, physical, or chemical processes. Such information can be
found in specialized books and papers in the respective research field. There are,
however, attempts to collect and unify this type of modelling. Literature on mathe-
matical methods for physics stems from all ages, and range from very formal to very
practical. [Tayler, 2001, Temam and Miranville, 2000, Bossavit, 1998] Attempts to
detach the methods, such as operational calculus, [Churchill, 1944] from the actual
physical problems have led to abstract theories which survived mainly in exten-
sions to the standard approaches for linear systems. [Åström and Wittenmark,
1989, Ljung, 1999] In practice, such methods are mainly applied in cases where
the system is not known, which might manifest itself by a large noise component.
For hyperbolic PDE’s such black box approach is infeasible, as in many cases, the
problem itself will test the limits of computing power, and any overhead should be
reduced by the use all the information available on the system at hand.
The general modelling of complex, lumped systems in the field of system en-
gineering, which tries to reach a level of abstraction and generalization in order
to model and analyze complex interacting systems in with sufficient knowledge of
the parts of the components. [Paynter, 1961, Seireg, 1969, Karnopp et al., 2000,
Wellstead, 1979]
The extension to distributed systems and general invariances which associate
continuous variables to discrete systems, such as rotation angles, have been a trou-
blesome aspect of modelling. The unclear status of the work of Gabriel Kron is a
definite sign of this troubled use of elements of differential geometry in engineering.
[Grattan-Guinness, 1994, Paynter, 1961] From the Russian schools, Butkovskiy is
a known example of similar status, his work was more along the lines of opera-
tional calculus. [Butkovskiy, 1983] However, with robotics, and its limited number
of degrees of freedom, differential geometry is making a second entrance into the
modelling for the purpose of control. [Pappas et al., 1998, Nijmeijer and van der
Schaft, 1990, Slotine and Li, 1991]
A.6. Dynamics
In mechanics, [Arnold, 1989] or even more general dynamical descriptions, mod-
els can be formulated in equivalent forms, such as Hamiltonian of Lagrangian.
[Goldstein, 1980] However, not in all cases, all formulations exist or are of a feasible
form. In the mathematics there is a great interest in integrable systems, [Dubrovin
et al., 1984] for which a complete set of trajectories can be determined. Such solu-
tions usually have a underlying symmetry and are closely related to group theory,
in particular the theory of Lie groups.
The classical three-body problem is at the edge of the analytically solvable
problems. [Whittaker, 1917] More generic questions concerning dynamics involve
the periodicity and stability of motion. Chaotic attractors and patterns are the,
254 A. LITERATURE
mainly visual, manifestations of indeterministic stability, in which mankind likes to

classify aspects of nature. [Arnold, 1973, Hirsch and Smale, 1974]
A.7. Classical field theories

The mathematical formulation of multi-physics continuous systems received an
impetus with the work of Onsager in 1931, and Truesdell and Toupin in the Ency-
clopedia of Physics. [Truesdell and Toupin, 1960] Onsager derived relations between
flux dependencies on general principles. [Callen, 1985] On separate domains a lot of
literature is available, but interconnecting the different domains consistently, taking
care of energy conservation and entropy, has been a large and controversial research
field of the twentieth century. For example, from cosmology new questions arose
concerning the nature of classical field theories. [Soper, 1976] Probably the final
words are not yet said on the subject.
The thermodynamics of electromagnetism is discussed in, for example, [de Groot
and Suttrop, 1972, Hutter and van der Ven, 1978, Eringen and Maugin, 1998].
Truesdell and Noll give a overview of aspects of elasticity in [Truesdell and Noll,
1965]. Modern mechanical materials with electrodynamic properties are discussed
in [Maugin, 1988]. Heat transfer is another topic which interlinks many differ-
ent physical domains. [McAdams, 1951, Carslaw and Jaeger, 1959, Holman, 1981,
Vekstein, 1992]
However, the most diverse and active field of research in classical field theory
is the study of fluid dynamics.[Mises and Friedrichs, 1971, Landau and Lifschitz,
1987b, Lighthill, 1978] Turbulence, boundary layers, and free surfaces, such as
for droplets, are both from a fundamental and a practical perspective of great
interest.[De Gennes et al., 2003, Eggers, 1997]
A.8. Networks of dynamical systems

The bond-graph modelling [Paynter, 1961, Karnopp et al., 2000, Breedveld,
1984] of physical systems is one of the most successful methods to combine graph
theory [Mayeda, 1972] with physical modelling. A less successful attempt has been
put forward by Gabriel Kron. [Kron, 1965] The purpose of a graph or network
is to single out, or focus on, the global system, retaining the relevant parts of
the components. The network is the global view. Some more obscure attempts to
combine graphs with physical modelling can be found in, for example, [Mead, 1989],
extending the integrated circuit design to the physical domain and [Zemanian,
1991], using the results of electrical circuits to study infinite-dimensional networks
equivalent to continuous systems.
A.9. Control theory

Control theory is an integral part of the industrial revolution. [Bennett, 1979,
Maxwell, 1868] The complexity of the machines required some automated adjust-
ment of settings. The developments in control theory keep pace with the techno-
logical developments most of all. The classical system theory and control theory
predates the use of electronics and computers in control, both in design of the con-
trollers and in the controllers themselves. [Kalman et al., 1969, Brockett, 1970]
The work of H. S. Black on amplifiers in the telephone transmission lines first used
the electronic transistor in negative feedback. [Bode, 1960] Modern developments
A.10. HISTORY OF SCIENCE AND ENGINEERING 255
include all kind of optimization strategies, multivariable systems, and state mo-
del identification. [Skogestad and Postlethwaite, 1996, Åström and Wittenmark,
1989, Ogata, 1990, Polderman and Willems, 1998, Antoulas, 2005] Modern equip-
ment such as cd players and jet planes cannot function without a detailed plant
model in the controller. In these cases some like to speak of learning control or
adaptive control, if the plant model improves during operation. Well-known exam-
ples are learning control through neural networks and fuzzy logic. These methods
had their predecessor in the theory of dynamic programming, where a solution was
constructed recursively or algorithmically rather than explicitly. Dynamic program-
ming is again closely related to the application of variational methods in control
design, called optimal control. [Kirk, 2004]
A.10. History of science and engineering

The history of science is relevant for a better understanding of the scientific
concepts. Some concepts may seem natural today, but have a troubled journey of
existence, we have come to accept. [Dijksterhuis, 1950, Kronig, 1958, Taton, 1965,
Bennett, 1979, Grattan-Guinness, 1994] On the theory of classical field theory the
contributions of Clifford Truesdell and collaborators are well-known. Furthermore,
there are specific contributions of fluid mechanics, [Tokaty, 1971] elasticity, [Timo-
shenko, 1953].
Bibliography
M. J. Ablowitz and P. A. Clarkson. Solitons, nonlinear evolution equations and inverse

scattering. Cambridge University Press, Cambridge, 1991.
R. Abraham, J. E. Marsden, and T. Ratiu. Manifolds, tensor analysis, and applications,
volume 75 of Applied Mathematical Sciences. Springer, New York, second edition, 1991.
M. Abramowitz and I. A. Stegun. Handbook of mathematical functions. Dover, New York,
1965.
J. H. Ahlberg, E. N. Nilson, and J. L. Walsh. The theory of splines and their applications.
Academic Press, New York, 1967.
N. I. Akhiezer. The classical moment problem. Oliver and Boyd, Edinburgh, 1965.
G. Allaire, F. Jouve, and A.-M. Toader. Structural optimization using sensitivity analysis
and a level-set method. Journal of Computational Physics, 194:363–393, 2004.
E. Anderson, Z. Bai, C. H. Bishop, S. Blackford, J. W. Demmel, J. J. Dongarra,
J. Du Croz, A. Greenbaum, S. Hammarling, A. Mckenney, and D. C. Sorensen. Lapack
users’ guide. SIAM, Philadelphia, 1999.
H. M. Antia. Numerical methods for scientists and engineers. Birkhäuser, Basel, second
edition, 2002.
S. S. Antman and G. Rosenfeld. Global behavior of buckled states of nonlinearly elastic
rods. SIAM Review, 20:513–566, 1978.
A. C. Antoulas. Approximation of large-scale dynamical systems. SIAM, Philadephia,
2005.
V. I. Arnold. Ordinary differential equations. MIT Press, Cambridge (Mass), 1973.
V. I. Arnold. Mathematical methods of classical mechanics. Springer, New York, second
edition, 1989.
V. I. Arnold. Lectures on partial differential equations. Springer, Berlin, 2004.
K. J. Åström and B. Wittenmark. Adaptive control. Addison-Wesley, Reading (MA),
1989.
A. O. Barut. Electrodynamics and Classical Theory of Fields and Particles. Dover, New
York, 1981.
W. Beitz, K.-H. Küttner, and B.J. Davies. Dubbel Handbook of mechanical engineering.
Springer, London, 1994.
S. Bennett. A history of control engineering 1800-1930. The Institution of Electrical
Engineers (Peter Peregrinus Ltd), London, 1979.
M. Benzi, G. H. Golub, and J. Liesen. Numerical solution of saddle point problems. Acta
Numerica, 14:1–137, 2005.
M. S. Berger. Nonlinearity and functional analysis. Academic Press, New York, 1977.
D. L. Bernstein. Existence theorems in partial differential equations. Princeton University
Press, Princeton (NJ), 1950.
L. Bers, F. John, and M. Schechter. Partial differential equations, volume III of Lectures
in applied mathematics. Interscience, New York, 1964.
C. B. Biezeno and R. Grammel. Technische dynamik. Springer, Berlin, 1953.
R. B. Bird, W. E. Stewart, and E. N. Lightfoot. Transport phenomena. Wiley, New York,
1960.
257
258 Bibliography
A. V. Bitsadze. Boundary value problems for second order elliptic equations. North-
Holland, Amsterdam, 1968.
H. Bode. Feedback: the history of an idea. In Proceedings of the symposium on active
networks and feedback systems, Brooklyn, 1960.
A. D. Booth. Numerical methods. Butterworths, London, third edition, 1966.
M. Born, E. Wolf, and A. B. Bhatia. Principles of optics. Pergamon, Oxford, sixth edition,
1980.
A. Bossavit. Computational electromagnetism. Academic Press, San Diego, 1998.
P. C. Breedveld. Physical system theory in terms of bond graphs. PhD thesis, University
Twente, Enschede, 1984.
P. C. Breedveld. A generic dynamic model of multiphase electromechanical transduction
in rotating machinery. In WESIC 2001, pages 381–394, Enschede, 2001. University
Twente.
J. P. Breslin and P. Andersen. Hydrodynamics of ship propellers, volume 3 of Cambridge
ocean technology. Cambridge, Cambridge, 1996.
R. W. Brockett. Finite dimensional linear systems. Wiley, New York, 1970.
A. G. Butkovskiy. Structural theory of distributed systems. Ellis-Horwood, Chichester
(UK), 1983.
H. B. Callen. Thermodynamics and an introduction to thermostatistics. Wiley, New York,
1985.
C. Carathéodory. Variationsrechnung und partielle differentialgleichungen erster ordnung.
Teubner, Leipzig, 1935.
H. S. Carslaw and J. C. Jaeger. Conduction of heat in solids. Clarendon Press, Oxford,
second edition, 1959.
A. Cherkaev. Variational methods for structural optimization. Springer, New York, 2000.
S. S. Chern. Global differential geometry, volume 27 of Studies in Mathematics. The
Mathematical Association of America, 1989.
B. H. Chirgwin and C. Plumpton. A course of mathematics for engineers and scientists,
volume 4. Pergamon Press, Oxford, 1964.
R. M. Christensen. Theory of Viscoelasticity. Academic Press, New York, 1971.
R. V. Churchill. Modern operational mathematics in engineering. McGraw-Hill, New York,
1944.
H. Cohen and R. G. Muncaster. The theory of pseudo-rigid bodies, volume 33 of Springer
tracts in natural philosophy. Springer, New York, 1988.
J. B. Conway. Functions of one complex variable, volume 11 of Graduate Texts in Math-
ematics. Springer, New York, second edition, 1978.
R. Courant and D. Hilbert. Methods of mathematical physics: Partial differential equa-
tions, volume II. Interscience, New York, 1961.
R. R. Craig. Structural dynamics, an introduction to computer methods. Wiley, New York,
1981.
R. R. Jr. Craig and C.-J. Chang. On the use of attachment modes in substructure coupling
for dynamic analysis. In Proc. AIAA/ASME 18th Structures, Structural Dynamics and
Materials Conference, volume B, 1977.
M. Creutz. Quarks, gluons, and lattices. Cambridge University Press, Cambridge, 1983.
R. F. Curtain and H. Zwart. An introduction to infinite-dimensional linear systems theory,
volume 21 of Text in Applied Mathematics. Springer, New York, 1995.
C. M. Dafermos. Hyperbolic conservation laws in continuum physics. Springer, Berlin,
1999.
P. R. Dahl. Solid friction damping of mechanical vibrations. AIAA Journal, 14:1675–1683,
1976.
R. Dautray and J.-L. Lions. Physical origins and classical methods, volume 1 of Math-
ematical analysis and numerical methods for science and technology. Springer, New
York, 1990.
Bibliography 259
B. Davies. Integral transforms and their applications, volume 25 of Applied Mathematical

Sciences. Springer, New York, 1978.
P.-G. de Gennes, F. Brochard-Wyard, D. Quéré, and A. Reisinger. Capillarity and wetting
phenomena : drops, bubbles, pearls, waves. Springer, New York, 2003.
S. R. de Groot and P. Mazur. Non-equilibrium thermodynamics. North-Holland, Amster-
dam, 1962.
S. R. de Groot and L. G. Suttrop. Foundations of electrodynamics. North-Holland, Ams-
terdam, 1972.
E. J. Dijksterhuis. De mechanisering van het wereldbeeld. Meulenhoff, Amsterdam, 1950.
P. A. M. Dirac. Lectures on quantum mechanics. Belfer Graduate School, Yeshiva Uni-
versity, New York, 1964.
B. A. Dubrovin, A. T. Fomenko, and S. P. Novikov. Modern geometry -methods and appli-
cations, Part I. The geometry of surfaces, transformation groups, and fields. Springer,
New York, 1984.
J. Eggers. Nonlinear dynamics and breakup of free-surface flows. Reviews of Modern
Physics, 69:865–929, 1997.
A. C. Eringen and G. A. Maugin. Electrodynamics of continua, volume I and II. Springer,
Berlin, 1998.
L. Euler. Methodus inveniendi lineas curvas maximi minimive proprietate gaudentes, sive
solutio problematis isoperimetrici lattissimo sensu accepti (A method for finding curved
lines enjoying properties of maximum or minimum, or solution of isoperimetric problems
in the broadest accepted sense). Bousquet, Lausanne, 1744.
D. Farin. Shape. In B. Engquist and W. Schmid, editors, Mathematics unlimited –2001
and beyond, pages 463–477, New York, 2001. Springer.
B. Felsager. Geometry, particles and fields. Odense University Press, Odense, 1981.
A. L. Fetter and J. D. Walecka. Theoretical mechanics of particles and continua. McGraw-
Hill, New York, 1980.
R. P. Feynman, R. B. Leighton, and M. Sands. The Feynman lectures of physics, volume II.
Addison-Wesley, Palo Alto, 1964.
H. Flanders. Differential forms: with applications to the physical sciences. Academic
Press, New York, 1963.
J. D. Foley and A. van Dam. Fundamentals of interactive computer graphics. Addison-
Wesley, Reading (Mass), 1982.
O. Föllinger. Lineare abtastsysteme. Oldenbourg, München, 1974.
T. Frankel. The geometry of physics. Cambridge University Press, Cambridge, second
edition, 2004.
R. W. Freund. Model reduction methods based on krylov subspaces. Acta Numerica, 12:
267–319, 2003.
K. O. Friedrichs. Energy principle and energy method. lecture notes, New York University,
Institute of Mathematical Sciences, New York, 1954.
K. O. Friedrichs. Spectral theory of operators in Hilbert space. Springer, New York, 1973.
D. Funaro. Polynomial approximation of differential equations. Springer, Berlin, 1992.
F. Gantmacher. Lectures in analytical mechanics. MIR, Moscow, 1970.
P. R. Garabedian. Partial differential equations. Wiley, New York, 1964.
L. Gårding. Hyperbolic equations in the twentieth century. Seminaires et Congres, 3:
37–68, 1998.
B. R. Gelbaum and J. M. H. Olmsted. Counterexamples in analysis. Holden-Day, San
Fransisco, 1964.
I. M. Gel’fand and S. V. Fomin. Calculus of variations. Prentice-Hall, Englewood Cliffs
(NJ), 1963.
I. M. Gel’fand and G. E. Shilov. Generalized Functions, volume 1 Properties and opera-
tions. Academic Press, Boston, 1964.
E.-D. Gilles. Systeme met verteilten parametern. R. Oldenbourg, München, 1973.
260 Bibliography
I. Gohberg and S. Goldberg. Basic operator theory. Birkhäuser, Boston, 1981.

H. Goldstein. Classical mechanics. Addison-Wesley, Reading (Mass), second edition, 1980.
G. H. Golub and C. F Van Loan. Matrix Computations. John Hopkins, Baltimore, third
edition, 1996.
B. R. Gossick. Hamilton’s principle and physical systems. Academic Press, New York,
1967.
I. Grattan-Guinness. History and philosophy of the mathematical sciences, volume 2. John
Hopkins University Press, Baltimore, 1994.
G. Grosche, V. Ziegler, D. Ziegler, and E. Zeidler. Tuebner -Taschenbuch der Mathematik,
volume II. Tuebner, Stuttgart, 1995.
H. W. Guggenheimer. Differential Geometry. McGraw-Hill, New York, 1963.
J. Hadamard. Mémoire sur le problème d’analyse relatif à l’équilibre de plaques élastique
encastrés. Mémoires présentés par divers savants á l’Académie des Sciences, 33:1–128,
1908.
J. S. Hadamard. Lectures on Cauchy’s problem in linear partial differential equations.
Dover, New York, 1923.
E. Hairer, C. Lubich, and G. Wanner. Geometric numerical integration, volume 31 of
Springer series in computational mathematics. Springer, Berlin, 2002.
C. H. Hansen and S. D. Snyder. Active control of noise and vibration. Chapman and Hall,
London, 1997.
O. Heaviside. On the forces, stresses, and fluxes of energy in the electromagnetic field.
Proceedings of the Royal Society of London, 50:126–129, 1891.
O. Heaviside. Heaviside centenary volume. Institute of Electrical Engineers, London, 1950.
P. Henrici. Applied and computational complex analysis: Power series, integration, con-
formal mapping, location of zeros, volume 1. Wiley, New York, 1974.
M. R. Hestenes and E. Stiefel. Methods of conjugate gradients for solving linear systems.
Journal of Research of the National Bureau of Standards, 49:409–436, 1952.
R. L. Highdon. Initial-boundary value problems for linear hyperbolic problems. SIAM
Review, 28:177–217, 1986.
B. F. Hildebrand. Introduction to numerical analysis. Dover, New York, second edition,
1987.
M. W. Hirsch and S. Smale. Differential equations, dynamical systems and linear algebra.
Acedemic Press, San Diego, 1974.
W. V. D. Hodge. The theory and applications of harmonic integrals. Cambridge University
Press, Cambridge, 1941.
N. Hogan. Impedance control: an approach to manipulation i-iii. ASME Journal od
Dynamic Systems Measurement and Control, 107:1–24, 1985.
J. P. Holman. Heat transfer. McGraw-Hill, New York, fifth edition, 1981.
P. Holmes, J. L. Lumley, and G. Berkooz. Turbulence, coherent structures, dynamical
systems and symmetry. Cambridge, Cambridge, 1996.
L. Hörmander. The analysis of linear partial differential operators, volume 2 Differential
operators with constant coefficients. Springer, Berlin, 1983.
L. Hörmander. The analysis of linear partial differential operators, volume 1 Distribution
theory and Fourier analysis. Springer, Berlin, 1990.
T. J. Hughes, J. A. Cottrell, and Y. Bazilevs. Isogeometric analysis: Cad, finite element,
nurbs, exact geometry, and mesh refinement. Computer Methods in Applied Mechanics
and Engineering, 194:435–458, 2005.
T. J. R. Hughes. The finite element method. Prentice-Hall, Englewood Cliff (NJ), 1987.
K. Hutter and A. A.F. van der Ven. Field matter interactions in thermoelastic solids,
volume 88 of Lecture Notes in Physics. Springer, Berlin, 1978.
C. Huygens. Horologium oscillatorium. Société Hollandaise des Sciences, Den Haag, 1673.
A. Isodori. Nonlinear control systems, volume 72 of Lecture Notes in Control and Infor-
mation Sciences. Springer, Berlin, 1985.
Bibliography 261
T. A. Ivey and J. M. Landsberg. Cartan for beginners: differential geometry via moving
frames and exterior differential systems, volume 61 of Graduate studies in mathematics.
AMS, Providence, Rhode Island, 2000.
J. D. Jackson. Classical electrodynamics. Wiley, New York, second edition, 1975.
F. John. Partial differential equations, volume 1 of Applied Mathematical Sciences.
Springer, New York, fourth edition, 1982.
D. Jürgens, M. Palm, S. Singer, and K. Urban. Numerical optimization of the voith-
schneider propeller. Ulm University preprint, 2006.
R. E. Kalman, P. L. Falb, and M. A. Arbib. Topics in mathematical system theory.
McGraw-Hill, New York, 1969.
E. Kamke. Differentialgleichungen, volume II. Geest und Portig, Leipzig, fourth edition,
1962.
L. V. Kantorovich and V. I. Krylov. Approximate methods of higher analysis. Noordhoff,
Groningen, 1958.
D. C. Karnopp, D. L. Margolis, and R. C. Rosenberg. System dynamics: modeling and
simulation of mechatronic system. Wiley, New York, 2000.
W. Karush. Minima of functions of several variables. PhD thesis, University of Chicago,
Chicago (IL), 1939.
T. Kato. Perturbation theory for linear operators. Springer, Berlin, 1966.
A. C. Kermode. Mechanics of flight. Longman, London, 1987.
D. E. Kirk. Optimal control theory. Dover, New York, 2004.
G. A. Korn and T. M. Korn. Mathematical handbook for scientists and engineers. McGraw-
Hill, New-York, second edition, 1968.
A. D. Kovalenko. Thermoelasticity. Wolters-Noordhoff, Groningen, 1969.
S. G. Krantz. Partial differential equations and complex analysis. CRC Press, Boca Raton
(FL), 1992.
H. O. Kreiss. Initial boundary value problems for hyperbolic systems. Communications
in Pure and Applied Mathematics, 23:277–298, 1970.
S. Krenk. State-space time integration with energy control and fourth-order accuracy for
linear dynamic systems. Int. J. Numer. Meth. Engng, 65:595–619, 2006.
G. Kron. Non-riemannian dynamics of rotating electrical machinery. Journal of Mathe-
matics and Physics, 13:103–194, 1934.
G. Kron. Invisible dual (n-1)-networks induced by electric 1-networks. IEEE Transactions
on Circuit Theory, CT12:464–476, 1965.
R. Kronig. Leerboek der natuurkunde. Scheltema en Holkema, Amsterdam, 1958.
H. W. Kuhn and A. W. Tucker. Nonlinear programming. In Proceedings of 2nd Berkeley
Symposium, pages 481–492, Berkeley, 1951. University of California Press.
O. A. Ladyzhenskaya. The boundary value problems of mathematical physics, volume 49
of Applied Mathematical Sciences. Springer, New York, 1985.
C. Lanczos. Linear differential operators. Van Nostrand, London, 1961.
C. Lanczos. The variational principles of mechanics. University of Toronto Press, Toronto,
1966.
L. D. Landau and E. M. Lifschitz. Theory of elasticity. Pergamon, Oxford, 1959.
L. D. Landau and E. M. Lifschitz. Mechanics, volume 1 of Course in theoretical physics.
Pergamon, Oxford, third edition, 1976.
L. D. Landau and E. M. Lifschitz. The classical theory of fields. Pergamon, Oxford, fourth
edition, 1987a.
L. D. Landau and E. M. Lifschitz. Fluid Mechanics. Pergamon Press, Oxford, second
edition, 1987b.
I. Lasiecka and R. Triggiani. Control theory for partial differential equations: continu-
ous and approximation theories: II abstract hyperbolic-like systems over a finite time
horizon, volume 75/2 of Encyclopedia of mathematics and its applications. Cambridge,
Cambridge, 2000.
262 Bibliography
P. D. Lax and B. Wendroff. System of conservation laws. Communications on Pure and

Applied Mathematics, 13:217–237, 1960.
G. Lebeau. Control for hyperbolic equations. Journées Équations aux Dérivées partielles,
21:1–24, 1992.
L. P. Lebedev and I. I. Vorovich. Functional analysis in mechanics. Springer, New York,
2000.
A. Y. Lee, W. S. Tsuha, G. K. Man, and G. A. Macala. Model reduction methodologies for
articulated multi-flexible body systems, volume 15 (Nonlinear Techniques) of Structural
Dynamic Systems Computational Techniques and Optimization, pages 195–271. Gordon
Breach, 1999.
S. Lefschetz. Introduction to topology. Princeton University Press, Princeton (NJ), 1949.
M. J. Lighthill. Waves in fluids. Cambridge University Press, Cambridge, 1978.
N. E. Ligterink, P. C. Breedveld, and A. J. Van der Schaft. Internal dynamics around
operational fem states. preprint, submitted for publication, 2006.
J. L. Lions. Optimal control of systems governed by partial differential equations. Springer,
Berlin, 1970.
J. Liouville. Démonstration d’un théoreme du a m. sturm et relatif a une classe de fonctions
transcendantes. Journal de mathématiques pures et appliquées, 2:269–277, 1836.
M. M. Lipschutz. Differential geometry. Schaum’s outline series. McGraw-Hill, New York,
1969.
X.-D. Liu, S. Osher, and T. Chan. Weighted essentially non-oscillatory schemes. Journal
of Computational Physics, 115:200–212, 1994.
L. A. Liusternik and V. J. Sobolev. Elements of functional analysis. Ungar, New York,
1961.
L. Ljung. System identification: theory for the user. Prentice Hall, Upper Saddle River
(NJ), 1999.
H. A. Lorentz. The theory of electrons. Dover, New York, second edition, 1915.
A. E. H. Love. A treatise on the mathematical theory of elasticity. Dover, New York, 1944.
A. M. Lyapunov. The general problem of the stability of motion. Taylor & Francis, London,
annotated translation of the 1892 classic edition, 1992.
W. Magnus, F. Oberhettinger, and R. P. Soni. Formulas and theorems for the special
functions of mathematical physics. Springer, New York, third edition, 1966.
A. Majda and S. Osher. Initial-boundary value problems for hyperbolic equations with
uniformly characteristic boundary. Communications on Pure and Applied Mathematics,
28:607–675, 1975.
J. M. Maubach. Local bisection refinement of n-simplicial grids generated by reflection.
SIAM Journal for Scientific Computing, 16:210–227, 1994.
G. A. Maugin. Continuum mechanics of electromagnetic solids, volume 33 of Applied
mathematics and mechanics. North-Holland, Amsterdam, 1988.
J. C. Maxwell. On governors. Proceedings of the Royal Society of London, 16:270–283,
1868.
J. C. Maxwell. A treatise on electricity and magnetism. Clarendon Press, Oxford, third
edition, 1891.
W. Mayeda. Graph theory. Wiley, New York, 1972.
W. A. McAdams. Heat Transmission. McGraw-Hill, New York, second edition, 1951.
J. McCleary. Geometry from a differentiable viewpoint. Cambridge University Press,
Cambridge, 1994.
C. Mead. Analog VLSI and neural systems. Addison-Wesley, Reading (Mass), 1989.
S. G. Mikhlin. Variational methods in mathematical physics. Pergamon, Oxford, 1964.
W. A. Milne. Numerical solution of differential equations. Dover, New York, second
edition, 1970.
N. Minorsky. Self-excited oscillations in dynamical systems possessing retarded action.
Journal of Applied Mechanics, 9:65–71, 1942.
Bibliography 263
R. von Mises and K. O. Friedrichs. Fluid dynamics. Springer, New York, 1971.
S. Morita. Geometry of differential forms, volume 201 of Translations of mathematical
monographs. American Mathematical Society, Providence, Rhode Island, 2001.
K. W. Morton and D. F. Mayers. Numerical solution of partial differential equations.
Cambridge University Press, Cambridge, 1994.
H. Nijmeijer and A.J. van der Schaft. Nonlinear dynamical control theory. Springer, New
York, 1990.
E. Noether. Invariante variationsprobleme. Nachr. Ges. Wiss. Göttingen, 171, 1918.
J. T. Oden. Finite elements of nonlinear continua. McGraw-Hill, New York, 1972.
J. T. Oden and J. N. Reddy. An Introduction to the mathematical theory of finite elements.
Wileys, New York, 1982.
K. Ogata. Modern control engineering. Prentice-Hall, Englewood Cliffs (NJ), 1990.
C. C. Paige and M. A. Saunders. Solution of sparse indefinite systems of linear equations.
SIAM J. Numer. Anal., 12:617–629, 1975.
G. J. Pappas, J. Lygeros, D. Tilbury, and S. Sastry. Exterior differential systems in
control and robotics. In J. Baillieul, Sastry S. S., and Sussman H. J., editors, Essays
on mathematical robotics, volume 104 of The IMA volumes in mathematics and its
applications, pages 271–372, New York, 1998. Springer.
H. M. Paynter. Analysis and design of engineering systems. MIT Press, Cambridge (MA),
1961.
O. Pironneau. Optimal shape design for elliptic systems. Springer, New York, 1984.
J. W. Polderman and J. C. Willems. Introduction to mathematical systems theory, vol-
ume 26 of Texts in Applied Mathematics. Springer, New York, 1998.
G. Pólya and G. Szegö. Isoperimetric inequalities in mathematical physics, volume 27 of
Annals of mathematical studies. Princeton University Press, Princeton (NJ), 1951.
P. M. Prenter. Splines and variational methods. Wiley, New York, 1975.
W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery. Numerical Recipes
in C: The Art of Scientific Computing. Cambridge, New York, 1992.
F. Riesz and B. S. Nagy. Functional analysis. Frederick Ungar, New York, 1955.
L. Rosenfeld. Questions of irreversibility and ergodicity. In Teorie ergodiche, Scuola
Internazionale di Fisica –Enrico Fermi– (Societa Italiana di Fisica), pages 1–20, New
York, 1961. Academic Press.
W. Rudin. Real and complex analysis. McGraw-Hill, New York, second edition, 1966.
Y. Saad. Iterative methods for sparse linear systems. PWS Publishing, Boston, 1995.
R. Sakamoto. hyperbolic boundary value problems. Cambridge, Cambridge, 1982.
S. I. Sandler. Chemical and engineering thermodynamics. Wiley, New York, 1977.
L. Schwartz. Mathematics for the physical sciences. Hermann, Paris, 1966.
H. A. Schwarz. Über einen Grenzübergang durch alternierendes Verfahren. Vierteljahrss-
chrift der Naturforschenden Gesellschaft in Zürich, 15:272–286, 1870.
A. Seireg. Mechanical system analysis. International Textbook Company, Scranton (PA),
1969.
D. Serre. Systems of conservation laws. In B. Engquist and W. Schmid, editors, Mathe-
matics unlimited –2001 and beyond, pages 1061–1080, New York, 2001. Springer.
G. E. Shilov. Generalized functions and partial differential equations. Gordon Breach,
New York, 1968.
N. C. Singer and W. P. Seering. Preshaping command inputs to reduce system vibration.
Journal of Dynamical Systems, Measurement and Control, 112:76–82, 1990.
S. Skogestad and I. Postlethwaite. Multivariable feedback control. Wiley, Chichester, 1996.
J.-J. E. Slotine and W. Li. Applied nonlinear control. Prentice Hall, Englewood Cliffs
(NJ), 1991.
G. A. Sod. Numerical methods in fluid mechanics. Cambridge, Cambridge, 1985.
A. Sommerfeld. Partieller differentalgleichungen der physik. Deutsch, Thun, sixth edition,
1992.
264 Bibliography
D. E. Soper. Classical field theory. Wiley, New York, 1976.

R. W. Soutas-Little. Elasticity. Prentice-Hall, Englewood Cliffs (NJ), 1973.
M. R. Spiegel. Complex variables. Schaum’s Outlines. McGraw-Hill, New York, 1964.
M. Spivak. A comprehensive introduction to differential geometry. Publish or Perish,
Houston, 1970.
T. Stieltjes. Recherches sur les fractions continues. Anns. Fac. Sci. Univ. Toulouse, 8,9:
J1–11, A5–47, 1894-1895.
J. Stoer and R. Bulirsch. Introduction to Numerical Analysis. Springer, New York, 1980.
R. F. Streater and A. S. Wightman. CPT, spin and statistics, and all that. Benjamin,
New York, 1964.
A. M. Stuart and A. R. Humphries. Dynamical systems and numerical analysis, volume 2
of Cambridge monographs on applied and computational mathematics. Cambridge Uni-
versity Press, Cambridge, 1998.
J. C. F. Sturm. Mémoire sur les équations diffrentielles linéaires du second ordre. Journal
de mathématiques pures et appliquées, 1:106–186,278–289,290–308,, 1836.
J. L. Synge. The hypercircle in mathematical physics: a method for the approximate
solution of boundary value problems. Cambridge Unversity Press, Cambridge, 1957.
R. Taton. Science in the nineteenth century, volume 3 of General History of Sciences.
Thames and Hudson, London, 1965.
A. B. Tayler. Mathematical models in applied mechanics. Oxford University Press, Oxford,
2001.
S. W. Taylor and S. C. B. Yau. Boundary control of a rotating timoshenko beam. Aus-
tralian and New Zealand Industrial and Applied Mathematics Journal, 44:E143–E184,
2003.
R. Temam and A. Miranville. Mathematical modeling in continuum mechanics. Cambridge
Press, Cambridge, 2000.
D. W. Theobald. The concept of energy. Spon, London, 1966.
W. Thirring. Classical mathematical physics. Springer, New York, third edition, 1997.
S. P. Timoshenko. History of strength of materials. McGraw-Hill, New York, 1953.
G. A. Tokaty. A history and philosophy of fluidmechanics. G. T. Foulis and Co, Henley-
on-Thames (Oxfordshire), 1971.
A. Toselli and O. Widlund. Domain decomposition methods - algorithms and theory,
volume 34 of Springer Series in Computational Mathematics. Springer, Berlin, 2004.
F. Tröltzsch. Optimale Steuerung partieller Differentialgleichungen. Vieweg, Wiesbaden,
2005.
J. L. Troutman. Variational calculus with elementary convexity. Undergraduate Texts in
Mathematics. Springer, New York, 1983.
C. Truesdell and W. Noll. Non-linear field theories of mechanics, volume III/3 of Handbuch
für Physik. Springer, Berlin, 1965.
C. Truesdell and R. A. Toupin. The classical field theories, volume III/1 of Encyclopedia
of Physics. Springer, Berlin, 1960.
B. van der Pol and H. Bremmer. Operational calculus. Cambridge Univesity Press, Com-
bridge, 1959.
A. J. van der Schaft and B. M. Maschke. Hamiltonian formulation of distributed-parameter
systems with boundary energy flow. Journal of Geometry and Physics, 42:166–194, 2002.
W. P. A. van Lammeren. Weerstand en voortstuwing van schepen. Stam, Amsterdam,
1944.
G. E. Vekstein. Physics of continuous media: A collection of problems with solutions for
physics students. Adam-Hilger, Bristol, 1992.
R. R. Waiboer, R. G. K. M. Aarts, and J. B. Jonker. Application of a perturbation method
for realistic dynamic simulation of industrial robots. Multibody Systems Dynamics, 13:
323–338, 2005.
Bibliography 265
K. Washizu. Variational methods in elasticity and plasticity. Pergamon Press, Oxford,

1968.
P. E. Wellstead. Introduction to physical system modelling. Academic Press, New York,
1979.
J. R. Welty, C. E. Wicks, and R. E. Wilson. Fundamentals of momentum, heat, and mass
transfer. Wiley, New York, 1984.
W. L. Wendland. Elliptic systems in the plane. Pitman, London, 1979.
E. J. Weniger. Weakly convergent expansions of a plane wave and their use in fourier
integrals. Journal of Mathematical Physics, 26:276–291, 1985.
H. Weyl. The method of orthogonal projection in potential theory. Duke Mathematical
Journal, 7:411–444, 1940.
H. Weyl. Space time matter. Dover, New York, 1952.
H. Whitney. Geometric integration theory. Princeton Press, Princeton, 1957.
E. T. Whittaker. A treatise on the analytical dynamics of particles and rigid bodies.
Cambridge Press, Cambridge, 1917.
D. V. Widder. The laplace transform. Princeton Unversity Press, Princeton, 1941.
N. Wiener. Extrapolation, interpolation and smoothing of stationary time series. Wiley,
New York, 1949.
J. A. Wilkinson. The algebraic eigenvalue problem. Orford University Press, Oxford, 1965.
J. C. Willems. Dissipative dynamical systems, part i: general theory. Archive for Rational
Mechanics and Analysis, 45:321–351, 1972a.
J. C. Willems. Dissipative dynamical systems, part ii: linear systems with quadratic
supply rates. Archive for Rational Mechanics and Analysis, 45:352–393, 1972b.
K. Worden and G. R. Tomlinson. Nonlinearity in structural dynamics. IoP Publishing,
Bristol, 2001.
K. Yosida. Functional analysis, volume 123 of Grundlehren der mathematische wis-
senschaften. Springer, Berlin, sixth edition, 1980.
E. Zeidler. Applied functional analysis: Applications to mathematical physics. Springer,
New York, 1995.
A. H. Zemanian. Infinite electrical networks. Cambridge University Press, Cambridge,
1991.
Index
A matrix, 21, 46, 50, 51, 53, 55, 60, 153, 100, 115, 127, 128, 134, 139, 144, 193,
156, 171, 173, 177, 184, 187, 208, 217 195, 208, 213, 219, 222
boundary energy, 208, 209, 218, 228
abscissas, 94, 145, 148, 152, 175 boundary operator, 40, 78, 88, 133, 135, 221,
absorption, 117, 188 222, 224–226, 235
acoustic, 7 boundary value, x, 55, 73, 77, 78, 84, 85,
action, 37, 39, 75, 80, 157 128, 152, 214, 223
action angle, 23, 217 bridge model, 29
action reaction, x, 3, 14, 29 Buckingham method, 102
actuator, 8, 199 buoyancy, 104
adaptive, 61, 185, 188, 255
cable, 174, 229
adjoint, 40, 45, 75, 83, 134, 135, 138, 225
CAD/CAM, 159, 213
affine, 125, 205, 213
canonical pair, 12, 13, 23, 26, 27, 57, 116,
Ampère, 103, 115, 120
128, 133, 135
analytic, 34, 59, 63, 85, 87, 167, 179
capacitance, 80, 119, 120, 154, 167, 169, 207,
analyticity, 34, 54, 74, 86
239, 244
anharmonic oscillator, 199
capacitive, 3, 180
anisotropic, 128
capacitor, 49, 171, 244
annulus, 87
Carnot cycle, 117, 120
Arnoldi tridiagonalization, 47
Cartan, 5, 66
atom, 121
Cartesian coordinates, 43, 116, 195, 197, 224
attractor, 32, 254 Cauchy, 43, 54, 77, 86, 89, 128
audio compression, 61 Cauchy-Riemann, 34, 42, 85
causal, 90
Banach space, 32, 33, 35, 251 causality, 4, 55, 78, 90
band diagonal matrix, 131 center-of-mass, 125, 126, 130, 146, 196, 205,
barycenter, 150 207
beam, 9, 98, 100, 132, 134, 135, 138, 142, characteristic, 39, 88, 90, 100, 193
148, 149, 205, 207, 208, 211, 214, 228 charge, 3, 102, 115, 116, 119, 152, 171, 172,
behavioral theory, 3, 167 192, 233, 238, 240
Bezier curve, 213 chiral, 241
bi-orthogonal, 127, 226 Christoffel symbol, 38
bias, 108 circuit, 1, 3, 124, 254
bifurcation, 98 clamped, 100, 134, 138, 139, 144, 213
biharmonic, 72, 231, 242 closure relation, 95, 115, 135, 153, 170, 208,
Biot number, 104 224
Biot-Savart law, 115, 124 co-energy, 113, 119, 146, 238
bipolar coordinates, 43, 242, 244 co-moving, 101
blackbody radiation, 117 coherent, 2, 124, 126
bond graph, ix collocation, 29, 132, 206, 232
Bond number, 104 commutator, 67, 215
boundary condition, x, 5, 6, 10, 38, 40, 41, compact, 32, 34, 47, 86, 94, 252
54, 71, 73, 74, 77, 78, 80, 83, 85, 91, complex coordinates, 42, 43
267
268 INDEX
compliance, 112 divergence free, 68, 157, 235, 239

compressibility, 106 droplet, 254
conditional stability, 6, 84 dual, 41, 74, 97, 113, 143, 151, 181, 189, 225,
conductance, 6, 141, 154, 168 226
conductor, 118, 192, 239
conformal, 42, 43 eccentric cylinders, 43
conjugate gradient method, 47, 182, 185, 187, eigenfrequency, 164
189, 190 eigenfunction, 74, 75
constitutive relation, 115, 117, 134, 135, 168, eigenmode, 32, 59, 161, 162, 164, 171
170, 187, 205, 237, 240, 241 eigenspace, 217
constraint, 14, 27, 43, 51, 67, 86, 91, 106, eigenspectrum, 46
112, 115, 116, 127, 132, 136, 139, 143, eigenstate, 83, 175
145, 181, 183, 194, 199, 202, 210, 213– eigensystem, 163
215, 220 eigenvalue, 53, 64, 74, 92, 147, 154, 163, 165,
controllability, 51, 55, 174, 175, 177 185, 217, 218, 225, 252
elastic energy, 9, 10, 30, 66, 98, 111, 124,
convection, 102, 117, 249
127, 133, 134, 139, 143, 145, 146, 156,
convolution, 175
160, 196, 208, 209, 212, 228, 232
Cosserat, 98, 100
elasticity, 9, 72, 87, 98, 100, 103, 111, 113,
Coulomb, 115, 220, 237
121, 126, 127, 132, 135, 141, 146, 173,
countable, 32, 35, 46
218, 241, 242, 249
Courant-Friedrichs-Lewy, 36
electrodynamics, 7, 39, 74, 115, 117, 124,
covariance matrix, 155
233, 234
Craig-Bampton method, 227
electromotor, 119
curvature, 58, 64, 202, 211, 213, 232
electrostatics, 44, 67, 72, 114, 115, 120, 237,
239, 244
D’Alembert, 181
element Hamiltonian, 206, 208, 218, 220–
D’Alembertian, 72, 88
222, 231
Darboux theorem, 27, 43
elongation, 110, 112, 113, 160
degeneracy, 162, 163
emissivity, 117
delay, 2, 7, 174
energy conservation, x, 3, 7, 19, 26, 36, 56,
density, 5, 10, 82, 95, 101–103, 125, 128, 152,
63, 68, 77, 105, 119, 130, 142, 157, 214
160, 192, 215
energy continuity, 63, 71, 82, 84, 142, 157
design, 1, 2, 5, 8, 106, 142, 164, 213, 245
energy flux, 63, 68, 71, 80, 82, 84, 116, 119,
diagonalization, 47, 53, 59, 162, 171, 173,
123, 133, 134, 140, 157, 247
177, 189
energy-momentum tensor, 20, 81, 157
differentiability, 33
enthalpy, 101
differential operator, 5, 12, 34, 40, 46, 68,
entropy, 105, 170
70, 72, 78, 80, 83, 87, 88, 98, 127, 128,
Euclidean space, 43, 157, 186, 195, 253
130, 133, 135, 137, 139, 147, 149, 150,
Euler characteristic, 159, 229
206, 208, 216, 219, 222, 224, 226–228,
Euler-Bernoulli beam model, 98, 100, 134,
236, 250
138, 142, 205, 208, 211, 212, 228
diffusion, 103, 104, 117, 118, 135, 191, 208,
Euler-Lagrange equations, 51, 82
242
evolution, 20–22, 27, 49–51, 53, 54, 59, 73,
dimensionality, 175 77, 78, 80, 156, 158, 173, 175, 178, 182,
Dirichlet condition, 6, 43, 55, 73, 78, 83, 86, 217
127, 225 existence, x, 5, 35, 66, 71–73, 77, 80, 88, 128,
Dirichlet principle, 239 141, 220, 221
discontinuity, 144 exponentiation, 53, 54
discretization, 145, 150, 151 exterior product, 43, 64
dispersion, 168, 193
displacement, 7, 11–13, 30, 98, 100, 110, 111, Faraday, 37, 115, 237, 240
113, 125–127, 129, 131, 143, 144, 164, feed forward, 62
167, 181, 187, 192, 201, 208, 209, 211, feedback, 1, 61, 62, 110, 174, 255
213, 218, 230, 237, 240, 245 FEM, 5, 6, 60, 62, 125, 132, 141–143, 154,
dissipation, 49, 66, 101, 133 156, 159, 161, 171, 180, 181, 205, 206,
distribution theory, 88, 93 213, 217, 218, 244
divergence, 40, 68, 72, 78, 80, 81, 83, 116, ferrite, 241
132, 133, 135, 216 film layer, 132
INDEX 269
finite difference, 145, 152, 153 Hertzian problem, 192

finite volume, 145, 152 Hessian, 64
flame, 117 Hilbert space, 33, 35, 93, 94, 182
flow, 6, 68, 82, 84, 101, 104–106, 108, 168, Hodge-Helmholtz decomposition, 69, 221, 240
191, 234, 237, 242, 244 holonomic constraint, 194
fluid, 68, 101, 104–106, 113, 117, 152 Huygens, 19
flux, 63, 68, 71, 72, 80, 82, 84, 85, 115, 123, hyperbolic, 39, 71, 72, 74, 76, 77, 80, 84, 87,
223, 244 128, 133, 148, 158, 187, 193, 223, 225
force balance, 69, 103, 124, 126, 131, 183, hysteresis, 113
187, 188, 242
formal adjoint, 40, 74, 130, 133, 223 identification, 175, 255
Fourier number, 104 impedance, 6, 77, 80, 81, 84
Fourier transform, 72, 88, 94, 148, 175, 178 incompressibility, 106
Fourier-Laplace transform, 13, 36, 88, 251 inductance, 170, 171
Fréchet, 92 infinite-dimensional, 27, 208, 222, 225
frame, 98 initial boundary value problem, 77, 83, 84,
Frenet-Serret coordinates, 65, 98, 100 88, 116, 223
frequency, 8, 20, 23, 25, 28, 46, 76, 97, 100, input richness, 173–175, 178
125, 147, 155, 162, 167, 175, 176, 185 input-state-output model, 49, 54
friction, 24, 43, 101, 110, 114, 185 instability, 163, 165, 180, 186, 189, 194, 203
Frobenius integrability, 67, 240 integrability, 48, 63, 66, 67, 86, 158, 231,
Froude number, 104 236, 240, 250
functional, 11, 36, 38, 40, 52, 72, 73, 75, 82, interconnection, 6, 8, 123, 222, 236
92, 94, 95, 150, 157, 245 interpolation, 148, 222, 230, 231, 252
functional derivative, 92–95, 245 invariance, 20, 36, 50, 74, 81, 115, 117, 120,
127, 217, 253
Galerkin, 35, 221, 251 involutive, 67
Gateaux derivative, 92, 181 irreversibility, 32
Gauss curvature, 65 irrotational, 101
Gauss law, 238 isotropic, 72, 99, 111, 124, 149, 241, 244
Gauss-Legendre integration, 94
Gauss-Lobatto integration, 152 Jacobian, 29
Gauss-Seidel method, 191 Jordan block, 53, 55, 130, 177
Gegenbauer polynomials, 152 Joukowsky (Zhukovsky), 43
Generalized functions, 36
genus, 159 Kelvin (temperature), 103
geodesic, 39 Kelvin viscoelasticity, 113
geometric integration, 56 kinematical variables, 19, 20, 23, 77, 78, 80,
Godunov theorem, 148 116
Gossick dissipation, 25 kinematical viscosity, 103
gradient, 21, 26, 69, 72, 94, 133, 135, 186, Kirchhoff laws, 124
193, 208, 224, 244, 246 Knudsen number, 104
Graetz number, 104 Krylov space, 46, 156, 177, 182, 183, 189,
Gram-Schmidt orthogonalization, 189 190
Gramian, 51, 54
graph, 9, 35, 254 Lagrangian, 12, 39, 82, 91, 105, 146, 157,
Grashof number, 104 158, 253
Green function, 43, 70, 74, 76, 88, 90, 91, Laguerre projection, 178, 179
187, 238, 243 Lancaster scheme, 52
Green identity, 5, 40, 133 Lanczos tridiagonalization, 47, 48, 171, 189
Guyan projection, 92, 112, 156 Laplace-Beltrami operator, 72
Laplacian, 44, 70, 72, 78, 82, 86, 148–151,
Hadamard, 245 157, 181, 208, 225, 228, 238, 244
Hankel matrix, 177 leap-frog integration, 57, 201
harmonic function, 69, 86, 242 leapfrog integration, 56
heat, 32, 72, 101, 104, 117, 120, 124, 154, Lebesgue integration, 33, 36, 47, 79, 92, 226,
169, 170, 244 252
Heaviside, 74, 89, 115, 175 Legendre polynomials, 79, 93
Hermitian conjugate, 53 Levi-Civita tensor, 99, 240
270 INDEX
Lie derivative, 67 piezo, 120

linearity, 73, 130, 175, 180 plaquette, 157, 158
Liouville theorem, 22 plasticity, 113, 124
liquid crystals, 114 plate model, 112, 228, 230–232
Lopatinski condition, 87, 88, 252 Poincaré, 32, 83
Lorentz force, 119 Poisson law, 115, 238
Lyapunov, x, 183, 190 Poisson ratio, 111, 160, 237, 240
polarization, 115
Mach number, 104 port Hamiltonian, 12, 14, 78, 129, 133, 135,
magnetism, 115 206, 222, 223, 225, 228, 233, 235, 250
manifold, 253 postponement, 214
mass matrix, 13, 31, 33, 38, 131, 139, 146, power continuity, 63, 84, 119, 227
153, 183, 196, 197, 233 Poynting vector, 40, 116
mass-spring system, 146 Prandlt number, 104
mass-spring-damper model, 61 proper orthogonal decomposition (POD), 154
Maupertius principle, 158 pseudo stationary, 128–130, 134, 137
Maxwell, 116, 119, 237, 239 pseudo symplectic, 223
Maxwell equations, 115
Maxwell viscoelasticity, 113 quarternions, 43
Maxwell-Pfaff-Cartan, 66 quasi stationary, 91, 168
mechanics, 23, 52, 98, 181, 195, 253
mesh, 36, 125, 142, 159, 187, 194, 230 radiation, 32, 115–117
metric tensor, 38, 72, 157 Rayleigh damping, 197, 198, 202
model order reduction, 154, 155, 172, 177 Rayleigh dissipation, 25
modular, 3, 6, 9, 29, 121, 141 Rayleigh number, 104
moment problem, 47 relativity, 49, 116
moment variable, 14, 82, 128, 129 relaxation, 163
monomial, 45, 228, 232 residue theorem, 43
Reynolds number, 97, 101–104
Navier-Stokes equations, 97, 101, 105, 142, rigidity, 213, 245
152 Robin boundary condition, 83
network model, 3, 8, 142, 169, 171, 172, 187, Rossby number, 104
222, 236, 254 rotation matrix, 58, 66
Neumann condition, 6, 55, 73, 78, 83, 86, Runge-Kutta integration, 56
127, 225
Newton, 3, 29, 52, 103, 123, 125, 126, 130, saddle point, 24, 91
225 Saint-Venant principle, 9, 249
Noether, 20 Sakamoto, 88
nonlinearity, 24, 204 scaling, 103, 178
norm, 33, 43, 46, 56, 71, 72, 77, 81, 83, 84, scattering, 7, 39, 66, 72
92, 177, 180, 186, 190, 193, 201, 252 Schauder basis, 32
null-space, 13, 27, 45, 51, 78, 80, 86, 103, Schmidt number, 104
137, 139, 141, 147, 172, 174, 191, 217– Schwartz space, 36
221, 225, 226, 228, 229, 232, 236, 250 self-adjoint, 74, 75
Nusselt number, 104 self-dual, 94
semi-group, 36
observability, 55 separable space, 36
Ohm, 115, 168 serial network, 171, 172
optimization, 36, 74, 106, 109, 245, 246 shear modulus, 111
orthogonal, 28, 47, 76, 126, 152, 154, 156, signal, 7, 9, 61, 80, 90, 155, 164, 174
178, 189, 226 simplex, 159, 208, 233, 234
output, 3, 49–51, 53, 55, 61, 129, 170, 173, simply connected, 34, 86
179, 206, 227, 229, 237, 240 simulation, 4, 29, 46, 49, 60, 62, 131, 141,
142, 148, 153, 195, 252
Peclet number, 104 skew symmetric, 43, 137, 192
pendulum, 23, 25, 185, 199, 201 soap film, 65
penetration depth, 192, 193 Sobolev space, 35, 46, 92
periodicity, 254 space-time, 20, 82, 116, 157
PID, 62 sparse matrix, 181
INDEX 271
spectral, 53, 54, 76 Voith-Schneider propeller, 106

spectroscopic, 162 voltage, 3, 171, 172, 236
splines, 143, 175, 211, 230
spring-damper model, 114 Wanderwaals model, 114
stability, 177, 199, 201 wave equation, 72, 87, 115, 128, 151
Stanton number, 104 Weierstrass, 32, 34, 35
state space, 48, 51, 54, 55, 63, 84, 113, 133, Young modulus, 111, 113, 160
150, 154, 173, 175, 177, 179
static deformation, 8, 162, 164, 242 zero energy, x, 133
Stefan law, 117
stencil, 148–152
step size, 36, 56, 58, 164, 185–187, 190, 192,
193, 201
Stewart platform, 195, 197, 198, 201
Stieltjes, 36, 47
stiffness matrix, 13, 53, 58, 60, 112, 113, 131,
132, 162, 183, 188, 192, 213, 230
Stokes theorem, 43
streamline, 43, 67, 69, 237
stress tensor, 68, 111–113, 181, 240, 241
Sturm-Liouville, 74–76
SYMMLQ algorithm, 162
symplectic, 12, 22, 26, 27, 29, 42, 53, 57–59,
137, 235
temperature, 9, 102, 117, 119, 120, 154, 169,

191
tetrahedron (see simplex), 159
thermodynamics, 9, 32, 67, 118, 254
Thomson principle, 239
time integration, 186, 193
time reversal, 74
time scale, 181, 185
Timoshenko beam model, 100
topology, 5, 33, 63, 87, 159
torque, 9, 100, 138, 207
torsion, 9, 66, 98
total divergence, 39, 68, 82, 158
trajectory, 62, 113, 183
transfer matrix, 51
triangulation, 159, 219, 229, 253
tridiagonal matrix, 47, 131, 172, 189
trigonometric series, 13, 32, 45, 76, 216
truncation, 32, 46, 172, 178, 217
turbulence, 6, 84, 101
uniqueness, 35, 72, 73, 77, 85, 88, 128

unitary matrix, 28
unitary operator, 36
vacuum, 118
variational derivative, 13, 40, 58, 60, 85, 92,
93, 95, 134, 202
vertex, 219, 229, 230, 232
vibration, 7, 76, 80, 100, 124, 160, 164, 165,
173, 185, 245, 249
viscoelasticity, 113
viscosity, 43, 97, 101, 103, 106
Voigt model (Kelvin model), 113

Thesis Ligterink

Uploaded by

Document Informationclick to expand document informationFunctional Sysem Dynamics sebagai materi mata kuliah di S2 Teknik Mesin Universitas Sebelas Maret (Dinamika Sistem) bagian pertama, bagian yang kedua "iterative learning control" dan bagian ketiga "einstein relativistic based control".

Document Informationclick to expand document information

Copyright:

Available Formats

Thesis Ligterink

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Thesis Ligterink

Uploaded by

Copyright:

Available Formats

FUNCTIONAL SYSTEM DYNAMICS

This research was financially supported by

Abstract. Functional system dynamics is the analysis, modelling, and simu-

Proefschrift Universiteit Twente

university homepage: http://www.utwente.nl/

FUNCTIONAL SYSTEM DYNAMICS

ter verkrijging van

Norbert Emiel Ligterink

geboren op 26 augustus 1968

promotoren: Prof.dr. A. J. van der Schaft

assistent-promotor: Dr.ir. P. C. Breedveld

Chapter 2. Mathematical models of dynamical systems 19

Chapter 4. Differential geometry 63

4.2. Frenet-Serret coordinates 65

Chapter 5. Partial differential equations 71

Chapter 6. Models of infinite-dimensional systems 97

Chapter 7. Port-based modelling of infinite systems 123

Chapter 8. Linear finite element methods 141

8.6. Boundary conditions for stencils 151

Chapter 9. Analytic input 167

Chapter 10. Real and virtual dynamics 181

Chapter 11. Nonlinear FEM 205

11.10. Plate port-Hamiltonian element 228

... untutored mind

To deceive yourself, and me with you, you speak too much.

I was in a printing house in Hell and saw the method in which

1.1. The sad history of operational calculus

1.3. Port-based modelling

1.4. Stages of existence

the manner in which components are interconnected. Furthermore, these intercon-

From the continuous, or distributed, model a, as large as possible, discrete,

1.5. Boundary impedance

1.6. Transmission versus vibration

The function of a component is generally not to vibrate or oscillate. The

Figure 1.1. Most local variations in the boundary conditions are

of lumped energy-based bond-graph network modelling. [Paynter, 1961, Karnopp

1.7. Preliminary example

Preferably, an instantaneous approximation is sought, for example, the force

Force is not an arbitrary input to the system. Systems described by partial

Often, we make no distinction between the Hamiltonian as the energy density

• Chapter 9 Analytic input

1.9. Nederlandse samenvatting

randwaarden, en het model voor de interne dynamica, waarvoor begin-

Mathematical models of dynamical systems

We perceive motion and time together. It sometimes happens

2.2. One-dimensional dynamical system

Figure 2.1. Some trajectories of the harmonic oscillator in

gradient of the energy:

incompressible flows, which is equivalent to conformal mappings of the complex

Figure 2.3. Some trajectories of the mathematical pendulum in

equations of motion. While investigating dynamical systems in higher dimensions,

given by the equations of motion:

(2.13) Ė = ∇x E T ẋ = −γ ẋ2 + ẋF .

At this point it makes no difference whether we constrain the motion xi (t), or

2.3. Finite-dimensional dynamical systems

(2.23) = J12 ω12 + J13 ω13 + J23 ω23 = J · ω .

The rotation dynamics of the rigid body is a well-known example of an odd-

Figure 2.4. On the energy surface in three-dimensional space, the