Introduction to Systems Analysis

•
Dieter M. Imboden • Stefan Pfenninger

Introduction to Systems Analysis

Mathematically Modeling Natural Systems
Cartoons by Nikolas Stürchler

123
Dieter M. Imboden Stefan Pfenninger
Department of Environmental Systems International Institute for Applied Systems
Science Analysis (IIASA)
ETH Zurich Laxenburg
Zurich Austria
Switzerland

ISBN 978-3-642-30638-9 ISBN 978-3-642-30639-6 (eBook)

DOI 10.1007/978-3-642-30639-6
Springer Heidelberg New York Dordrecht London

Library of Congress Control Number: 2012948781

c Springer-Verlag Berlin Heidelberg 2013

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissim-
ilar methodology now known or hereafter developed. Exempted from this legal reservation are brief
excerpts in connection with reviews or scholarly analysis or material supplied specifically for the pur-
pose of being entered and executed on a computer system, for exclusive use by the purchaser of the
work. Duplication of this publication or parts thereof is permitted only under the provisions of the
Copyright Law of the Publisher’s location, in its current version, and permission for use must always
be obtained from Springer. Permissions for use may be obtained through RightsLink at the Copyright
Clearance Center. Violations are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publi-
cation does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of pub-
lication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Preface

Since the first publication of the German textbook Systemanalyse:

Einführung in die mathematische Modellierung natürlicher Systeme by
Dieter M. Imboden and Sabine Koch in 2003, many things have changed in
the academic world. Others have remained firm and constant; one of them
is the central role of the mathematical language as the common basis to
express the ever more complex findings of scientific research. In fact, an in-
troduction to the principles of mathematical modeling should still be part
of the foundation of every university science curriculum.
At the same time, teaching is becoming more international, especially
at the world’s leading research universities (to which ETH Zürich proudly
belongs). This in turn means that even in non-English-speaking countries,
English has become firmly established as the lingua franca in the classroom.
In fact, the first author (Dieter Imboden) was often asked for an English
version of the book. Among those who contacted him was one of his former
students and a young researcher, Stefan Pfenninger, at the time working at
the International Institute for Applied Systems Analysis (IIASA) in Laxen-
burg, Austria. Dieter responded with his regrets over the lack of an English
version of the book and spontaneously suggested to Stefan to help pre-
pare one. Sabine Koch, the junior author of the German book and by now
a successful ecological winemaker in Southern Germany, enthusiastically
supported the idea and was happy to yield the junior author position to
Stefan. The cartoonist Nikolas Stürchler, who had meanwhile joined the
Swiss Diplomatic Service (where he is rumored to use his skills livening up
drab international treaties), agreed to have his cartoons incorporated into
the English version as well. Thus, cheeky Dang, muddle-headed Professor
Dong, and clever dog Ding are still here to help the reader digest all the
equations with humor and (sometimes) deep philosophical insight.
Since the concept of the book has been proven in the lecture hall time
and again, we kept changes in the translated version to a minimum. With a
few exceptions, the numbering of equations, figures, tables, examples, and
problems correspond one-to-one to the German version. This will facilitate
the use of the book for lectures in which students are using either version.
The major exceptions to this are as follows. In Chap. 4, Example 4.13 has
been moved before the explanation of the two extreme cases (slow/rapid
fluctuation). The chapter ends with a new example on lake temperature
(Example 4.14), which was taken from Problem 4.6 of the German edition.

v
vi PREFACE

Consequently, Problem 4.6 now refers to a modified situation of that lake

temperature problem. The two other modifications are the addition of
Problem 6.9 (competition for living space) and the replacement of Prob-
lem 8.7 by a more realistic example (transport and reaction in a lake’s
water column).
The book is firmly rooted in the algebraic formulation of mathematical
models, their analytical solution, or—if solutions are too complex or do not
exist—in a thorough discussion of the anticipated model properties. The
reader will find neither mathematical proofs (just a summary of some useful
equations in the Appendix) nor any specific reference to the many existing
software tools. The latter make life easy for today’s modelers, but they
can also produce a false sense of security. We firmly believe that the user
of such software should always be able to interpret the result of computer
simulations, and that is only possible if she possesses at least a basic math-
ematical understanding. That being said, we also recognize the importance
of connecting directly and visually to one’s work and the ease with which
modern software allows us to formulate and solve modeling problems. We
therefore prepared interactive examples for several of the models presented
in the book on the web as an additional aid in understanding the underlying
principles (see the book website at www.systems-analysis.org).
Once again, the authors of the book were blessed with numerous favor-
able circumstances. After Dieter became President of the Research Council
of the Swiss National Science Foundation, the principal Swiss public re-
search funding agency, Nicolas Gruber succeeded him as environmental
physics professor at ETH Zürich. Nicolas aided the development of the
book in myriad ways: He graciously provided funding for travel and other
expenses and provided feedback as well as new material.

Zürich, Switzerland Dieter M. Imboden

Vienna, Austria Stefan Pfenninger
Contents

1 Introduction 1
1.1 Systems Analysis . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 What is a System? . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 What is a Model? . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 The Formation of Models in Natural Science . . . . . . . . . 9
1.5 Questions and Problems . . . . . . . . . . . . . . . . . . . . 13

2 Mathematical Models: A First Look 15

2.1 From System to Model . . . . . . . . . . . . . . . . . . . . . 15
2.2 Static Models . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Dynamic Models . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 Discrete Time Models . . . . . . . . . . . . . . . . . . . . . 26
2.5 Spatially Continuous Models . . . . . . . . . . . . . . . . . 27
2.6 Stochastic Models . . . . . . . . . . . . . . . . . . . . . . . 27
2.7 Questions and Problems . . . . . . . . . . . . . . . . . . . . 31

3 Static Models 33
3.1 Equilibrium Distribution Between Water and Air . . . . . . 34
3.2 Equilibrium Distribution Between Water and Sediment . . . 35
3.3 Multi-dimensional Static Models . . . . . . . . . . . . . . . 37
3.4 Questions and Problems . . . . . . . . . . . . . . . . . . . . 39

4 Linear Models with One Variable 43

4.1 The Linear One-Box Model as a Balance Equation . . . . . 44
4.2 Linear Models with Constant Coefficients . . . . . . . . . . 47
4.3 Time-Dependent Coefficients . . . . . . . . . . . . . . . . . 59
4.4 Questions and Problems . . . . . . . . . . . . . . . . . . . . 75

5 Linear Models with Several Variables 81

5.1 Linear Models with Two System Variables . . . . . . . . . . 81
5.2 Linear Models with Several System Variables . . . . . . . . 113
5.3 Questions and Problems . . . . . . . . . . . . . . . . . . . . 120

vii
viii CONTENTS

6 Nonlinear Models 125

6.1 Nonlinear Models with One System Variable . . . . . . . . . 126
6.2 Nonlinear Box Models with Several System Variables . . . . 143
6.3 Questions and Problems . . . . . . . . . . . . . . . . . . . . 160

7 Time-Discrete Models 165

7.1 Time-Discrete Models with One Variable . . . . . . . . . . . 165
7.2 Time-Discrete Models with Several Variables . . . . . . . . 180
7.3 Questions and Problems . . . . . . . . . . . . . . . . . . . . 186

8 Models in Time and Space 189

8.1 Mixing and Transformation . . . . . . . . . . . . . . . . . . 189
8.2 Advection, Diffusion and Exchange . . . . . . . . . . . . . . 192
8.3 Steady-State Transport/Transformation Models . . . . . . . 202
8.4 Time-Dependent Solutions of the Transport/
Transformation Equation . . . . . . . . . . . . . . . . . . . 213
8.5 Questions and Problems . . . . . . . . . . . . . . . . . . . . 221

A List of Symbols 227

B Dimensions and Units 231

B.1 Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
B.2 Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232

C Formulary 233
C.1 Linear Inhomogeneous First-Order Differential
Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
C.2 System of Two Linear First-Order Differential
Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
C.3 General Solution of the Linear Second-Order
Differential Equation with Constant Coefficients . . . . . . 235
C.4 Solution of Linear Differential Equations
with Imaginary Eigenvalues . . . . . . . . . . . . . . . . . . 237

D Eigenvalues 239
D.1 The n-Dimensional System . . . . . . . . . . . . . . . . . . 239
D.2 Explicit Solution for the Two-Dimensional System . . . . . 240

E Time-Dependent Diffusion Equation 243

E.1 The Normal or Gaussian Distribution . . . . . . . . . . . . 243
E.2 The Error Function . . . . . . . . . . . . . . . . . . . . . . . 244
E.3 The Principle of Linear Superposition . . . . . . . . . . . . 244

Bibliography 247

Index 249
Chapter 1

Introduction

Dang

I love the small

I always find surprises in maths.
authority
suspicious.

I like to look at things

from a different
point of view.

Some Remarks Before We Begin

This book was originally developed alongside the lecture Systems Analysis
at the Swiss Federal Institute of Technology (ETH) Zürich, on the basis
of lecture notes developed over 12 years. The lecture, together with others
on analysis, differential equations and linear algebra, belongs to the basic
mathematical knowledge imparted on students of environmental sciences
and other related areas at ETH Zürich.
The book was written to demonstrate practical applications of mathe-
matics without becoming tedious and overly complex, in other words, how
maths can be interesting and fun! Examples and exercises from various
environmental sciences like limnology, population ecology and environmen-
tal chemistry demonstrate what the mathematical tools can be used for.
A basic knowledge of environmental issues is beneficial for understanding
these examples in their context.
The book aims to be more than a mathematical treatise on the analysis
and modeling of natural systems, yet a certain set of basic mathematical
skills are still necessary. We will use linear differential equations, vector

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 1

DOI 10.1007/978-3-642-30639-6_1, © Springer-Verlag Berlin Heidelberg 2013
2 INTRODUCTION

and matrix calculus, linear algebra, and even take a glimpse at nonlinear
and partial differential equations. Most of the mathematical methods used
are covered in the appendices. Their treatment there is brief however, and
without proofs. Therefore it will not replace a good mathematics textbook
for someone who has not encountered this level of math before.
Although most of the examples used here are drawn from the environ-
mental sciences, this book is not an introduction to the theory of aquatic
or terrestrial environmental systems. Rather, a key goal of the book is to
demonstrate the virtually limitless practical potential of the methods pre-
sented. Readers are encouraged to use the tools provided to tackle problems
from their own areas of interest.
The first chapter is a short overview of the philosophy behind building
models in science and does not use mathematical terms yet.

Book website
Visit the book website at www.systems-analysis.org. You’ll find interactive versions
of several examples to help you understand the behavior of systems. You can also find
additional resources there, such as a list of further recommended books.

Ding & Dong

Ding: Studied at Oxford and MIT, since 1994
president of the Royal Society of Scientific Dong: Professor for networked knowledge at
Animals. 1997 honored as Dog of the Year , Ivory Tower University. Can give complicated
2000 became president of the School for answers to simple questions.
Animalic Systems Analysis (SASA).

1.1 Systems Analysis

Everything flows, nothing remains constant. Science sees our world as a
Heraclitus dynamic system. To describe and analyze this world, science has developed
(536–470 BC): the remarkable ability to divide systems into ever smaller pieces, into sub-
παντ α ρι systems which are again subdivided into sub-subsystems and so forth. The
panta rhei resulting entities are dynamic systems themselves, but they are brought
gr.: “everything flows” down to a level of complexity understandable for the human mind. In most
cases, the analysis of these systems is performed with the assistance of
 SYSTEMS ANALYSIS 3

mathematical methods.
The principle underlying all of this is called reductionism. The reduc-
tionistic principle is not without its critics, and sometimes it is given re-
sponsibility for spawning a sort of technological progress which is blind to
its negative consequences (environmental impacts such as biodiversity loss,
climate change, as well as other effects such as the growing division between
rich and poor).
We do not claim to be able to contribute to this important discussion.
We do argue, however, that to criticize, improve or even transcend an idea,
one must first understand it. On closer inspection, we discover that charac-
terizing a scientific approach as reductionistic or on the contrary as holistic
is always a question of one’s point of view. A cell biologist would probably
call a molecular biologist’s work reductionistic. Cell biology, however, will
appear as a holistic field to her. On the other hand, from a population
ecologist’s viewpoint, cell biology would itself appear reductionistic.
The division of science into different areas of expertise thus not only
takes place through by putting different disciplines into different drawers of
a chest (see illustration in the margin), but each drawer is again subdivided
into further drawers. The chest of drawers depicted is therefore only one
layer of reductionistic subdivision. In reality, many such chests would have
to be drawn and each of their drawers would be subdivided further and
further. Of course, this nested mass of drawers is not only a feature of
biology—the reductionistic approach is used in every scientific field.
Even though it is not our aim to highlight this debate in principle, it will
play a role indirectly throughout the book. The debate around reductionism Reductionism can be
forces us, time and again, to carefully examine the methods we develop thought of as a chest di-
here. As yet there are few concepts with which we could comprehend the vided into drawers that are
world in a holistic way (whatever that may mean), without falling back to themselves further subdi-
the method of subdividing it into smaller and more understandable parts. vided.
Even if we should develop such methods in the future, analyzing a system
by dissecting its individual parts will always remain an important part
of science. In any case, we want to think critically about our use of a
reductionistic approach, and not lose track of the bigger picture.
In this book we want to convey how to model and analyze real-world
systems. The underlying core principle is to construct simplified represen-
tations of such complex systems. These representations, called models,can
often be expressed mathematically (e.g. through a system of equations).
The construction of a mathematical model can be reduced to a few core
ideas, in particular:

• The balancing of mass, energy or the number of objects (e.g. individ-

uals of a biological species)
• The description of chemical or simple biological transformations thro-
ugh stoichiometric reaction equations
• The description of populations through equations describing growth,
death and interaction between individuals
4 INTRODUCTION

• The description of transport processes through exchange rates and

processes of diffusion or advection

• The application of statistical methods to describe systems with many

degrees of freedom

With this book you, the reader, should gain the ability to analyze real
systems and make quantitative statements about their behavior. This re-
quires the ability to translate empirically identified properties of these sys-
tems into a mathematical form, and to discuss the resulting equations (in
particular, differential equations), in some cases even to solve them an-
alytically. We will also delve into the behavior of more complex models,
without specifically discussing how to solve them. Finally, we also want to
demonstrate the limits of simulation and the emergence of mathematical
artifacts.
Before we look at different mathematical models in more detail, we first
have to clarify what exactly a system and a model are, and to understand
the significance of modeling in science.

1.2 What is a System?

 The ant and the dung beetle

A dung beetle watches an ant, desperately trying to lug a spruce needle
onto a vast anthill. “Why waste your strength?” the beetle asks, “one nee-
dle more or less won’t make a difference. What’s the point?” “My work is
part of building a system” the ant replies indignantly, and thinks to itself
“You’ll see that this is more than a pile of needles soon enough, when
we carry you inside the anthill’s vast tunnels after paralyzing you with our
poison!”

A system is a set of objects The word system originates from Greek and means configuration or compo-
between which relations sition. The New Oxford American Dictionary gives the following definition:
exist. “a set of connected things or parts forming a complex whole”. In a more
abstract way, one could also say that a system is a set of objects between
which relations exist.
A system is more than the What does this mean? A system is composed of different parts, the
sum of its parts. objects or system components. They are connected with each other through
mutual interactions which we call internal relations.
A pile of metal parts is therefore not yet a system. The parts have to be
connected in a meaningful manner. They might be connected to form the
system clockwork. If one cog is removed from that system, it will entirely
cease to function. This fact is reflected in the saying “a system is more than
the sum of its parts”.
 WHAT IS A SYSTEM? 5

Only through internal relations can the constituent parts become a sys-
tem. This is not always as straightforward as in the clockwork example:

Example 1.1: Sand heap

Is a sand heap a system? Many will say that it is not. One can easily
remove a grain of sand and what we call the sand heap remains nev-
ertheless. Furthermore, from a mineralogical point of view one learns
essentially the same whether looking at a couple of grains or at the
whole heap. Others, however, will contend that a sand heap is indeed
a system. Between the individual grains of sand, small gaps called pore
space exist. These pores provide space for various organisms, they can
store water, or in the case of sediments, provide information about the
sand heap’s formation through their structure.

Whether a number of objects constitute a system can therefore not be

decided in a general way without taking into account the particulars of the
specific question to be answered. For a construction firm a sand heap is
a source of raw material, whereas for a geologist it is a system containing
valuable information.

Fig. 1.1: Schematic de-

External relation
piction of a system with
three system variables
V1 , V2 and V3 . The sys-
tem boundary is the bor-
der between the system
and the environment. The
No feedback from environment acts on the
the system to the system (or rather, the sys-
System environment tem variables) through the
Environment
external relation. The sys-
tem variables are coupled
through the internal rela-
Part of a system’s definition is the system boundary, which forms the tions. The feedback from
(often virtual) border between the system and its environment. This does the system to the environ-
not mean, of course, that there are no interactions between a system and ment is neglected
the larger environment. On the contrary, it is these interactions across the
boundary that make a system interesting. We assume here, however, that
it is a one-way relationship: the system is influenced by the environment
through what we call the external relations, but the environment is not
influenced by the system. In other words, we assume that the environment
is infinitely large and infinitely robust in comparison to the system we are
looking at, therefore, any influence the system might have on the environ-
ment is so small that we can neglect it (Fig. 1.1).
6 INTRODUCTION

The next example illustrates the hierarchy of interactions described

above:

Example 1.2: The earth in the solar system

The earth is a subsystem of the solar system. It is influenced in par-
ticular by the energy reaching it from the sun, therefore, we can say it
is driven by that external relation. On the other hand, we can assume
that the earth has negligible influence on the sun. A sensible system
boundary between the earth and the surrounding solar system would
thus be drawn in the upper stratosphere (or maybe somewhat higher).

But setting the system boundary is not always as simple, as the next ex-
ample demonstrates:

Example 1.3: A lake as a system

Let’s look at a lake. Our goal is to describe the concentration of phos-
phorus in this lake. The lake’s body of water is confined by the water
surface and by the sediment. However, we will quickly realize that this
system boundary makes little sense: the amount of phosphorus in the
lake is also influenced by the dissolution of phosphorus from the sedi-
ment. As a result of prior phosphorus contaminations, a large amount
of phosphorus is stored in the sediment body. So we choose as the
system the body of water and the sediments combined. But then we
further note that predicting the phosphorus concentration in the lake is
impossible without knowing what happens in its upstream catchment
area (population growth, construction of sewage treatment plants, agri-
cultural policy, draining of wetlands, etc.). Therefore we choose as a
system: the body of water, the sediment, as well as the catchment
basin. By now, of course, our system is already a very complex entity
with countless influences. In order to create a concise model, we would
likely have to simplify or even neglect many of these influences.

Summing up, we can say that systems are theoretical constructs that help
us in understanding a part of the world. There is no absolutely valid way to
define a system’s boundary. In fact, the exact choice of boundary depends
on the question we are asking.
 WHAT IS A MODEL? 7

1.3 What is a Model?

Supermodel

We can trace the origin of our word model to ancient Rome, where Modulus is Latin for
small replicas of buildings were called modulus. Just as Roman architects “small-scale replica of a
needed replicas to assist them in designing large and complex buildings, if building”.
we want to analyze a natural system, we must first describe it in simplified
terms. Whether this is as a drawing, a functional diagram, a mathematical
equation or a verbal account, we can call such a description a model. So
we can say that a model is always a simplified image of a real system.
A model, therefore, does not necessarily have anything to do with maths
at the outset. It is simply a concept with which we can describe a com-
plex system in simplified terms. Models are used in architecture to plan
buildings, or in chemistry to describe atoms as spheres and the molecular
bonds as little rods between them (Fig. 1.2). Models are in use everywhere,
whether they are physical models or purely mental constructs.

a b Fig. 1.2: Examples of

models: (a) Architectural
model: Scaled-down execu-
tion of a planned or exist-
ing building. (b) Physical
model: Model of chemical
molecules. (c) Mathemati-
cal model: Kepler’s second
c d law of planetary motion.
Planet Yearly
deaths (d) Conceptional model:
Death rate
Excerpt from the World
Population
Sun model by Forrester and
Food per Desired food Meadows (see Meadows et
capita per capita al. 1972)
Amount of Capital in
food agriculture
8 INTRODUCTION

A model is the lens through which we look at the real system—it is not
an exact copy of that system. For the same system, there can be completely
different models. We could depict a landscape as a topographical map, or as
an impressionistic painting. In both cases, the simplification of the system
“landscape” is our explicit intention. It is impossible to say a priori which
model is better. One model is not superior to another because it is more
complex. Depending on the point of view that we want to emphasise and the
question we intend to answer, we include certain properties of the system
in the model while excluding others. You can think of it like a dog taken for
a walk in a forest. The dog’s model of the forest will be different than that
of its owner; it might predominantly consist of scents, for instance. But he
will be able to navigate the forest just as well or even better than his owner.
Good models allow us, through certain simplifications, to arrive at a
readily comprehensible representation of the real world that can be de-
scribed mathematically. Such models have a long tradition in science and
engineering. The only way to apprehend and analyze a real system is with
a model, that is, by framing it in terms that hold relevance for us. Often,
an important motivation to construct a model is to make predictions about
a system’s future behavior. As we will see, however, making predictions is
not the only role that models play.

Walking the dog

If a model is formulated with mathematical equations, it becomes a

mathematical model. Mathematical models were developed primarily in
the exact natural sciences and in engineering, above all in physics. They
allow us to develop consistent theories about our physical surroundings. Un-
der the influence of the phenomenal success these scientific disciplines had,
more and more additional disciplines started using mathematical models
throughout the twentieth century. This development did not stop short of
the social sciences and humanities, although it was hindered there until the
recent availability of immense amounts of data and with it the emergent
“digital” social science and humanities. To date, the field beyond natu-
ral science in which mathematical modeling has seen the most success is
economics.
 MODEL FORMATION 9

The development of systems-oriented branches of science (such as ecol-

ogy and environmental science) would not have been possible without mod-
els. More importantly, the models themselves would soon have been beyond
reach of mathematical analysis if not for the enormous progress in infor-
mation technology. The development of powerful computers now allows us
to simulate models numerically which even in the 1990s were deemed to
be unsolvable. On the other hand, some of the most influential ecological
models, such as the famous predator-prey equations by Lotka (1924) and
Volterra (1926) (see Sect. 6.2), were formulated more than 80 years ago,
when computers did not yet exist.
To summarize: A model is a concept for the simplified description of A model is a concept for
a complex system. It serves to characterize the important features of a the simplified description
system and to leave aside the peripheral ones. If the model is a mathemat- of a complex system.
ical one, the interactions between the system variables (internal relations)
and the ones between the system and its environment (external relations)
are formulated mathematically. This way, the model can be analyzed by
mathematical simulation.

1.4 The Formation of Models in Natural Science

To develop theories in science one needs models: without models there is
no generalizable or transferable insight. As we have seen above, “model”
does not necessarily mean a complex computer program. Even just the Ohm’s law:
assumption that in an electrical resistor current grows proportionally to I = U/R
voltage (Ohm’s law) is a model.
A classic example of how scientific understanding comes about illus-
trates the pivotal role of model formulation better than a long explanation:

Example 1.4: Our model of the universe

The process of formulating a scientific model can be divided into the
following phases:
1. Collecting Observations, measurements:
Tycho Brahe (1546–1601) collects data on the
movement of planets using his royally equipped
observatory Uranienborg near Copenhagen.
2. Arranging The search for an underlying principle of order
in the gathered data (model formation):
Johannes Kepler (1571–1630) arranges
Brahe’s orbital data in such a way that he
can describe it in terms of three mathematical
equations (Kepler’s laws).
10 INTRODUCTION

3. Understanding The search for an overarching principle, with

which the empirically found order can be
understood:
Isaac Newton (1643–1727) demonstrates that
Kepler’s laws can be explained through physical
principles that are valid far beyond astronomy.
4. Generalizing Can the discovered laws be applied in other
situations?
Albert Einstein (1879–1955) develops his
theory of general relativity based on the equiv-
alancy of inertial and gravitational mass
already postulated by Newton.
5. Predicting Can the (possibly generalized) laws be used to
predict observations or phenomena that have
not yet been observed?
Einstein’s theory of relativity remained a
mathematical construct until the astronomical
phenomena which he predicted were actually
observed.

Using this example, we can identify the two main tasks of systems analysis:
Model formation: The construction of a model begins with laying
down the system boundary, the definition of the system variables, and the
internal and external relations. The starting point is often empirical data,
for instance from well-designed experiments. These data are analyzed in
view of constructing a model. The model, usually a mathematical one,
should explain the data in the simplest possible form, in terms of the in-
ternal model structure and the external relations. In other words: systems
analysis looks for the (simplest) structure of the “black box” that can re-
produce the postulated connection between the state of the environment
and the system’s state (Fig. 1.3).

State of the State of

environment the system

Fig. 1.3: As a first step in mathematical model formation, the external relations
(state of the environment) and the system variables (Vi ) have to be defined and
quantified. The black box model attempts to describe the system variables from
the external state as simply and meaningfully as possible
 MODEL FORMATION 11

In the end, therefore, gathering data only makes sense if further insight
results from it, either as a generalizable statement, a prognosis, or as the
formulation of new questions, that is, the planning of new experiments. To
process observations in this manner, we need models.

Box model

I ve built a 3-dimensional box model. For the most dramatic demonstration, Wow! Wait a second, where is my money?
It conveys a deeper understanding of we need a modest input of money
how money flows in our society. similar to how we pay taxes.

Fascinating,
isn t it?

Model validation: After constructing a model compatible with the

data, and determining all model variables, we can use the model to predict
the system’s behavior under different conditions and compare these predic-
tions with further measurements. Through these comparisons, we might be
able to improve the model further. This process is called model validation.
In principle, the process of modeling is a process of infinite iteration,
which we abort as soon as we reach sufficient accuracy (Fig. 1.4).

Fig. 1.4: In model vali-

dation, observed data are
repeatedly compared with
adjusted model simulations
until a satisfactory corre-
spondence between model
and observational data is
achieved

Observations
Simulation Measurements
12 INTRODUCTION

With this, we’ve reached the end of the first chapter, in which we looked
at the importance of the modeling method in science. In the next chapter,
we will get to know the fundamental tools with which we can actually build
models.
At the end of each chapter, there is a set of questions and problems. In
the questions, we recap the chapter’s contents. They can usually be quickly
and easily answered. The problems however require more in-depth work,
but will greatly help with retaining and applying the key concepts and tools
introduced in the chapter.
 QUESTIONS AND PROBLEMS 13

1.5 Questions and Problems

Question 1.1: Why is there no simple and unambiguous definition to dif-
ferentiate between a reductionistic and a holistic description of a system?

Question 1.2: What is implied with the statement “The relation between
a system and its environment is asymmetrical”?

Question 1.3: Give a short definition for each of the following terms:
system, model, mathematical model.

Question 1.4: What distinguishes a model from reality?

Question 1.5: Give some of the most important decisions that have to be
made when constructing a model.

Question 1.6: Which functions do models have in science?

Question 1.7: What does model validation mean?

Problem 1.1: Model formation

Look at the following three systems and try to formulate a model. Consider
what system variables and external relations are important for the models.
What can be neglected? Where is the system boundary?
(a) Population growth (nationally, i.e. in a specific country)
(b) Population growth (globally, i.e. for the entire planet)
(c) Carbon dioxide in the atmosphere
(d) Lead (a heavy metal) in a lake’s sediment

Problem 1.2: Solar system

For a simplified description of the solar system, we choose the following
variables: the location and speed of the sun and the eight planets.
(a) What are the internal relations?
(b) What external relations does the system have?
(c) How does our system deviate from the real solar system as part of
the cosmos? Do our simplifications relate to the internal relations,
the external relations, or both of them?

Problem 1.3: From chemical elements to nuclear physics

Characterize, like we did in Example 1.4 for our model of the cosmos, the
path from the discovery of the chemical elements to the development of
nuclear physics.
Chapter 2

Mathematical Models: A First Look

In the second chapter, we want to look more closely at mathematical models

with which we can describe natural systems. Our aim is first and foremost
to familiarize ourselves with a model’s basic components, as well as with
different types of models. We will learn how to construct a model on the
basis of examples. For this, we will not yet need complex mathematics.

2.1 From System to Model

Before forming a model, we should think carefully about what we need the
model for. What questions do we want to examine with it? Our research
question helps us define the system, then select the system boundary and
the relevant system variables.
For this first step of model formation, we can use some very simple
tools. We draw our system as a box diagram. The boundary of the box
corresponds to the system boundary. Inside the box we draw the system
variables Vi . From outside the box, the external relations Ri influence the
system. In many cases, these external relations consist of mass flows from
the environment into the system. Therefore, the external relations are also
called input variables or mass input. But there could also be other kinds of
external relations, for instance, solar radiation, atmospheric temperature,
or a bank’s interest rate. We connect the system variables with arrows that
indicate the relations between them. These are the internal relations of
the system. They could, for instance, be mass flows between the system
variables.
One natural system that we will use as an example throughout this
book is a lake. Figure 2.1 shows a first simple model to describe a chemical
substance in a lake.

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 15

DOI 10.1007/978-3-642-30639-6_2, © Springer-Verlag Berlin Heidelberg 2013
16 MATHEMATICAL MODELS

Fig. 2.1: A lake as a box Mass

input
diagram with three vari-
ables

Example 2.1: A lake as a black box

I would
prefer a lake
The model has three variables: the amount or mass of a chemical sub-
to swim in! stance in the lake water Maq , in the lake’s sediment Msed and in it’s
living organisms Mbio (for instance in fish). The substance to be mod-
eled reaches the lake by a tributary river. This mass input represents
the system’s external relation. The loss of substance through the lake’s
outflow is symbolized by an arrow pointing out from the box. The ex-
change of substance between lake water, sediment and biomass is also
symbolized by arrows connecting the three variables. These are the
internal relations.

The dimension of a model The number of system variables is called the dimension of a model. The
equals the number of sys- example above represents a three-dimensional model with the three system
tem variables. variables Maq , Msed and Mbio . In order to formulate the model mathemat-
ically, we must describe the internal relations (the mass flows between the
three system variables) and the external relations. By doing so, we arrive at
the system or model equations. They usually contain additional quantities:
the model parameters. In the next section, we will illustrate the method to
create the system equations step by step with an example.

2.2 Static Models

We will begin with a simple example: a static model of a lake. It might
allow us to answer questions about the correlation between the influx of
a substance through the lake’s inlet and its average concentration in the
lake. The substance we are interested in is phosphorus, which often enters
surface waters from sewage and agricultural fertilisers.
Let’s first think about how the system looks schematically (Fig. 2.2).
Our model is one-dimensional, it has one system variable: the phosphorus
concentration in the lake Caq (unit: mg m−3 ). The external relation of
the system (the phosphorus input) is the influx of dissolved phosphorus
into the lake Jin (unit: kg year−1 ). It could either be constant or it could
 A FIRST LOOK 17

Phosphorus
Fig. 2.2: Static lake
input
model: Average phospho-
rus concentration in the
lake as a function of the
current phosphorus input

fluctuate with time. In a first attempt, we assume that there is an immediate Is there a correlation
relation between the phosphorus input at time t and the average phosphorus between Jin (t) and Caq (t)?
concentration in the lake measured at the same time. Mathematically, this
can be formulated as follows1 :

 
Caq (t) = Function of Jin = f Jin (t) (2.1)

We assume that the lake in question has been under surveillance for
several years and that a data set exists from which we can calculate Caq (t)
and Jin (t) for different times. To find the unknown function in Eq. (2.1), we
can sketch the values of Caq (t) and Jin (t) in a two-dimensional diagram.
Let’s look at a numerical example:

Example 2.2: Phosphorus in a lake — static model

In a lake, the nutrient load constantly increases over a period of many
years. This so-called eutrophication of the lake is established through
the following measurements. They demonstrate a correlation between
the yearly phosphorus input Jin (t) via inlets and wastewater treatment
plants, and the average total phosphorus concentration in the lake that
year Caq (t). The following data were measured in four not necessarily
consecutive years.
Jin (t) [kg year−1 ] 1,500 2,100 2,800 4,200
Caq (t) [mg m−3 ] 20 30 36 56

1 With this example, we by no means suggest that this model is reasonable. It

is merely a first attempt, the general validity of which we will later discuss in
detail.
18 MATHEMATICAL MODELS

Figure 2.3 shows the pairs of data in a two-dimensional diagram. All

data points lie roughly on a straight line. From that, we can conclude
that the function in question is a linear relation of the form f (x) = p·x.
Now, we can mathematically formulate the linear relation between
phosphorus input J(t) and average phosphorus concentration Caq (t):

Caq (t) = p · Jin (t) (2.2)

The coefficient p is is the model parameter. It can be calculated as the
ratio between the pairs of data. The resulting average value is:
 
Caq (t) mg m−3
p= = 1.33 × 10−2 (2.3)
Jin (t) kg year−1
where the bar atop the quotient signifies the average of the data pairs.

Fig. 2.3: The data pairs 60

Jin (t) and Caq (t) in a two- C aq (t )
Phosphorus concentration [ mg m−3]

dimensional space. The 50

data points lie approxi-
mately on a straight line
40

30

20

10

J in
0
0 1000 2000 3000 4000
Phosphorus input [ kg yr −1]

The above phosphorus model is called static, because for every value
of the external relation R (in this example, the input Jin (t)) there is a
precisely defined value of the system variable V. Mathematically speaking,
a static model with one system variable has the following form2 :

V = Function of {R} = f (R) (2.4)

2Not in all cases can we explicitly describe the system variable V (as we did in
Eq. 2.4). It can also be given through an implicit equation, such as eaV = bV.
 A FIRST LOOK 19

Equation (2.4) is called the system equation of the model. The function
f (R) can have an arbitrarily complicated form and can be defined by one
or several model parameters p. If the function f (R) is linear, as is the case
in Eq. (2.2), the model is called linear.
We have described the system “phosphorus in a lake” with a linear,
static model. By using Eq. (2.2), we can predict how the lake’s phospho-
rus concentration changes in response to a changed phosphorus input. We A static model cannot ad-
cannot, however, make any statement about how long it takes until the equately describe changes
phosphorus concentration reaches its new value after the input variable has through time.
been changed. In a static model, the change takes place instantaneously. A
dynamic model is needed in order to describe the transition from one state
to another.
Before we delve deeper into dynamic models in Sect. 2.3, we will briefly
discuss the importance of units and dimensions.

2.2.1 Handling Dimensions and Units

Mathematical models link model variables by mathematical relations. Model
variables generally have a dimension (see Appendix B). In most cases, the
combination of the three basic dimensions mass (M), time (T) and length
(L) is sufficient to express the dimension of every variable. Units are formed
by explicitly choosing a system of measurement for the dimensions; there-
fore, units are a specific representation of dimensions and should not be
confused with them. As long as we only work with algebraical expressions,
the choice of units is irrelevant. Only the dimensions have to match on both
sides of an equation. Yet, as soon we give specific values to variables in an
equation, a consistent set of units is needed.
Let’s take another look at Eq. (2.3) in the example above. The unit of
mass in the numerator (mg) is different from the one in the denominator
(kg). It can be adjusted through the conversion 1 mg = 10−6 kg. Then, the
following holds:

mg m−3
p = 1.33 × 10−2
kg year−1
mg m−3
= 1.33 × 10−8
mg year−1
= 1.33 × 10−8 year m−3

The parameter p therefore has the dimension of an inverse volume flow rate
(T L−3 ). We will get to know the physical significance of this parameter in
Sect. 4.2. For now, let’s look at another example:
20 MATHEMATICAL MODELS

Example 2.3: Dimensions and units

A wastewater treatment plant with a water volume of 1,000 m3 is ac-
cidentally contaminated with 10 kg of a toxic substance. How high is
the concentration of that substance in the treatment plant after the
accident?
The dimension of a concentration is (ML−3 ), or mass (M) per vol-
ume (L3 ). Therefore, we can calculate the average concentration C
in the plant with the data we have. To do so, we insert the mass of
the substance (in kg) and the plant’s volume (in m3 ). We then get the
concentration in kg m−3 :
10 kg
C= = 0.01 kg m−3 (2.5)
1, 000 m3
The units kg and m are so-called metric or SI units.a If possible, SI
units should be used in scientific work. Often, however, it makes sense
to convert the units into others that can more easily be visualized. The
units of the average concentration in the sewage treatment plant, for
instance, could be converted as follows:

103 g
0.01 kg m−3 = 0.01 × = 0.01 g L−1 = 10 mg L−1 (2.6)
103 L
Thus 10 mg of the toxic substance are dissolved in 1 L of water.

a SI units were defined as the “système international d’unités” at a confer-

ence in Paris in 1960. They form the international standard units.

Most common spreadsheet and statistics software allows the user to graphi-
cally display the empirical relation between data pairs (like in Fig. 2.3) and
to estimate the function by a regression. If we relate data pairs of two or
more measured variables by a linear or nonlinear regression, a static model
is produced where the regression equation becomes the model equation.
Since setting up model equations by regression analysis is very easy,
one will often forget to verify whether the equation makes any physical
sense. Even in literature, one occasionally finds model equations that have
different dimensions on both sides. Furthermore, equations are often set up
in a way that makes them valid only for a specific set of units.
In order to demonstrate this point, let’s look at experimental data on
the growth of an algae population under varying nitrate concentrations in
water. The measurements were analyzed by a nonlinear equation.
 A FIRST LOOK 21

Example 2.4: Nonlinear regression and the pitfall of units

A scientist determines the specific growth ratea G of an algae popula-
tion at different nitrate concentrations C in water. A nonlinear fit of
the data yields:
G = 0.1 C 0.6 (2.7)
G Specific growth (per hour)
C Nitrate concentration
Unfortunately, the author of the model forgot to specify in which units
the nitrate concentration should be inserted in Eq. (2.7). One user of
the model may express nitrate concentration in mol m−3 and receives
a value that is smaller by a factor of 4.9 than the result of another
user who fills in the concentration C in mgN L−1 . To make the formula
compatible for both users, they would both have to use the same units.
But which ones are the correct ones?
a Specific growth rate (G) is the relative increase in biomass per unit of
time. Independently of the dimension and unit in which biomass is given,
the dimension of G is always (T−1 ).

For a serious modeler there are three ways to prevent this kind of confusion:
(1) The easiest, but not always the best way: We simply write the correct
units in square brackets behind each variable:
G (h−1 ) = 0.1 (C (mgN L−1 ))0.6 (2.8)
This notation makes the equation confusing and difficult to read.
(2) A better way is to replace the numerical factor by the symbol p and
to separately give the value of p together with the correct unit:
G = p C 0.6 , with p = 0.1 h−1 (mgN L−1 )−0.6 (2.9)
G (h−1 ) Specific growth rate
C (mgN L−1 ) Nitrate concentration
If we would now like to insert the concentration in mol m−3 , we have
to keep in mind that 1 mgN L−1 = 1 gN m−3 = 14 1
mol m−3 . Thus,
the parameter can be converted as follows:
p = 0.1 h−1 (mgN L−1 )−0.6
1
= 0.1 h−1 ( mol m−3 )−0.6
14
= 0.1 h−1 (14)0.6 (mol m−3 )−0.6
= 0.49 h−1 (mol m−3 )−0.6
22 MATHEMATICAL MODELS

Since this procedure is rather complicated, it may be better to trans-

form the concentration values into the proper units and then apply
Eq. (2.7).
(3) The safest way, however, is to use non-dimensional variables wherever
fractional exponents could appear. Applied to Eq. (2.7), we should use
the following expression:
 0.6
0 C
G=G (2.10)
C0

G (h−1 ) Specific growth rate

C (mgN L−1 ) Nitrate concentration
where G0 is the “reference growth rate” at the “reference concentra-
tion” C 0 . The pair of reference values can be chosen arbitrarily, pref-
erentially somewhere in the middle of the range of available data. We
can satisfy ourselves that for C 0 = 1 mgN L−1 , G0 must be 0.1 h−1
for Eqs. (2.7) and (2.10) to be equivalent. Now, changing the units of
concentration is much easier. One only has to adjust the units of the
reference concentration accordingly.

 Tabloid units
Editor: “You want a front page headline because of a laughable 10
Curies in RadonCorp’s waste?”—Journalist: “How about 400 billion Bec-
querel?”a —“Much better. You’ve got journalism in your blood after all!”

a Curie and Becquerel are both units for the radioactivity of a radiation source.
1 Becquerel corresponds to one decay per second. 1 Curie = 37 billion Bec-
querel.

Summing up, verifying the dimensions and units when we set up and
solve system equations is one of the easiest and most efficient ways to
prevent mistakes. The following rules may help:
• The algebraic expressions to the left and right of the equals sign must
always have the same dimension.
• Addition and subtraction are only possible if all algebraic terms have
the same dimension.
• When inserting concrete numbers into an algebraic equation, their
units have to be compatible. In other words, they have to be chosen
so that the same combination of units results on both sides of the
equals sign.
• Exponents in equations are non-dimensional (and therefore also with-
out units). The same holds for arguments in transcendental functions
 A FIRST LOOK 23

such as sin(..), cos(..) or exp(..). Ensure that you only have fractional
powers or transcendental functions of non-dimensional variables in
your equations.

Appendix B gives an overview of dimensions and units.

2.3 Dynamic Models

Equilibrium
The only way forward now is a And if I trip out of the
dynamic move... equilibrium, I ll fall 600m into
the abyss.

11.06 seconds until a

new equilibrium is
reached.

In Sect. 2.2 we have described the system “phosphorus in a lake” with a

static model. As we discovered, with such a model we can calculate the
average concentration of phosphorus in the lake water as a function of the
input concentration. However, the model neglects the fact that it takes some
time for the system to adjust to the new input value. With a static model,
we can make no statement about how rapidly the phosphorus concentration
in the lake changes if there is a sudden change in the input concentration.
If we want to describe a system’s dynamics in time, static models are no
longer sufficient: we need a dynamic model.
Often, adjustment processes to a change in the external relation are
particularly important to understand natural systems, as is the case in the
following example:

Example 2.5: Carbon dioxide in the atmosphere

Burning fossil fuels releases carbon dioxide into the atmosphere. The
current atmospheric concentration of carbon dioxide (CO2 ) does not
directly depend on the current rate of fossil fuel consumption. Rather,
the gradual increase of CO2 mirrors the slow adjustment process of the
global carbon cycle to the increased input of CO2 into the atmosphere.
The CO2 concentration in the atmosphere will continue to rise even
if the growth of fossil fuel consumption were to be reduced or halted
completely. In Chap. 5, we will examine a dynamic model for the global
carbon cycle (in Example 5.12).
24 MATHEMATICAL MODELS

A dynamic model describes In general terms, a dynamic model can describe how rapidly the system
the system’s time- variable V changes when the external relation R varies. The dynamic model
dependent answer to an describes the time-dependent answer to the external change. Figure 2.4
external change. shows the adjustment of a linear system to a disturbance that occurs sud-
denly at time t = 0. The solid line represents the simulation with a dynamic
model. The system variable V(t) does not react immediately to the distur-
bance, but adjusts gradually to a new system state in equilibrium with
the changed external relation R. The static model, depicted as a dashed
line, disregards this adjustment behaviour; the system variable immediately
changes to the new system state.

Fig. 2.4: A dynamic

model can describe the
adjustment of the system
variable V(t) to a changed
external relation R (dis-
turbance). The state V0
corresponds to the equilib-
rium of the system variable
V with the external relation
R before the disturbance
occurred (t < 0)

The system equation of a dynamic model is significantly different from

Eq. (2.4). On the left-hand side of the equation we write the system vari-
able’s change with time:

Temporal change of V = f (R, V) (2.11)

If the temporal change of the system variable V is smooth, i.e. not made
up by sudden jumps,3 it can be described by the first derivative of the
system variable with respect to time. The mathematical model is therefore

3 Mathematicians call such a function differentiable.

 A FIRST LOOK 25

a first order differential equation4 with respect to time:

dV
= f (R, V) (2.12)
dt
The function f (R, V) is called the change or velocity function of the system
variable.
In order to calculate the adjustment of the model to a changed input
variable, we have to integrate the differential equation (2.12). To do so,
we need an initial value V0 , that is, the value of the system variable at
an arbitrary time t0 . The model calculation begins at time t0 . Usually,
we set t0 = 0. It is important to note that the solution of the differential
equation (2.12) depends on the initial value V0 . This is a crucial difference
to the static model (Eq. 2.4), where there is no such dependency.
In many cases a dynamic model is developed to draw up a balance of
a specific quantity, such as the mass of a chemical substance, the number
of individuals of a specific species, or even just a quantity of money. In the
resulting system equations we can characterize the temporal change of that
quantity in the system as the difference between input and output. Let’s
look at an example:

Example 2.6: Phosphorus in a lake — dynamic model

The dynamic phosphorus model of a lake can be described as the mass
balance of phosphorus in the lake water. First, let’s describe the mass
balance in words:
 
Temporal change of the
=
phosphorus mass in the lake
   
Inflow via inlet Outflow via outlet
−
per unit of time per unit of time
   
Net uptake through Net uptake through
− −
biota per unit of time sediment per unit of time

To integrate this equation, we need to specify the terms on the right-

hand side of the equals sign. We can either use explicit numerical values
or algebraic expressions that create a link to the system variable. In
addition, we need an initial value for the phosphorus mass in the lake.
We will discuss this in more detail in Chap. 4.

4 In a first order (n-th order) differential equation, the function that consti-

tutes its solution is no higher than in its first (n-th) derivative. The models
we deal with here are usually first order with respect to time. In physics, par-
ticularly in classical mechanics, they are often second order with respect to
time. For the solution of these differential equations this distinction is only
marginally relevant, because one can convert an n-order differential equation
into n first order equations.
26 MATHEMATICAL MODELS

2.4 Discrete Time Models

Leaps of nature
The Greek land tortoise moves Thus even in antiquity, the Kangaroo: Leaping
forward only slowly, correct conclusion was reached: Australian Marsupial with
but continuously. nature does not leap. elongated hind legs.

Ok, I basically
hate dogs.

Before the discovery of quantum mechanics scientists were convinced that

system variables such as temperature, velocity or concentration change con-
tinuously. It was postulated that nature does not leap. But at closer exam-
ination, this is also incorrect for many systems that have nothing to do
with quantum mechanics. For instance, let’s model an elephant herd in
a national park in Africa. The herd’s size changes with every birth and
every death by a whole number—that is, the variable “number of individ-
uals” leaps. There are also systems that only change at specific times. A
typical example is a savings account that does not have its interest added
continuously, but only once per year or once every quarter.
Not every variable changes If a system variable doesn’t change continuously, but—as shown in
continuously with time. Fig. 2.5—only at discrete points in time t0 , t1 , t2 , . . ., we can simulate the
system with a discrete time model of the form

V(k+1) = f (R(k) , V(k) ) (2.13)

This means that the state V(k+1) is calculated from V(k) according to
a specific “recipe”. The superscript index k in parentheses indicates the
system state at time tk . Unlike in a differential equation Eq. (2.12), the
system variable’s change in time is not described infinitesimally
 by the first
derivative, but by the change occurring during the time step tk+1 − tk .5
A more detailed description of these models follows in Chap. 7.

5 In order to show the connection of Eq. (2.13) to the differential equa-

tion (2.12) of the continuous model, the left-hand side of the former is often
written as the difference between two consecutive time steps, V(k+1) − V(k) . In
this way the differential equation (2.12) becomes a so-called difference equation.
See Chap. 7 for details.
 A FIRST LOOK 27

Fig. 2.5: With a discrete

model, the system variable
V(k) changes in “jumps” or
discrete time steps tk . It
may be driven by a exter-
nal relation R that changes
with time

2.5 Spatially Continuous Models

So far, we assumed that the system variable only depends on time. Yet,
natural systems also have a spatial dimension. As long as we are only
interested in the average phosphorus concentration and assume that it is
distributed homogeneously in the lake, we do not need to consider any
spatial variation of the system variable. Instead, we can think of the entire
lake volume as one large box. We call such a spatially homogeneous model
a box model.
Often, however, the spatial distribution of a substance in the system is
particularly important. For instance, we are interested in the oxygen distri-
bution in different depths z of a lake. Figure 2.6 shows the oxygen profile
in a lake at different times. To describe such a spatial distribution, we have
to treat the lake as a spatial continuum in one or all three spatial dimen-
sions. The oxygen distribution then appears in the system equation as a
function of both time t and depth z, which we write as C(z, t). The model
equation thus becomes a partial differential equation. We call these models
spatially continuous and distinguish them from box models. The analytic
solution of partial differential equations is usually difficult—they are solved
computationally. We will look at some simple examples in Chap. 8.

2.6 Stochastic Models

Solutions of differential equations have the property of being completely
determined once the initial conditions are given. In other words: A system
always develops in exactly the same way if it starts from the same initial
state. Such behavior is called deterministic.
Our experience with a multitude of natural systems teaches us some-
thing different: many phenomena, from the weather to the dynamics of
28 MATHEMATICAL MODELS

Fig. 2.6: The oxygen dis-

tribution in a lake can be
described as a function of
time t and depth z. The
figure shows the vertical
oxygen profile in a lake
with a maximum depth
of 30 m at three different
times t1 , t2 and t3

the stock exchange, appear to lie outside the possibility of deterministic

analysis and prediction. In most cases—except in the case of very simple
systems—we can at best predict the future within certain time limits. For
instance, we may be able to predict that the system will develop into the
state S1 with a probability of p1 , into the state S2 with the probability p2 ,
and so forth. This behavior is called stochastic.
There are many different reasons for the occurrence of stochastic pro-
cesses. The most obvious reason is connected to quantum mechanics, which
in several respects replaced our physical thinking over the course of the
twentieth century. In quantum mechanics, the deterministic equations of
classical mechanics are replaced with probability relations. For instance,
we cannot predict when exactly a specific radioactive isotope will decay,
that is, we cannot describe the isotope’s behavior with a classical (deter-
ministic) equation.
In many cases (but not in all), this kind of random process can be
eliminated by assuming a macroscopical perspective. For instance, if we
analyze a large number of identical isotopes, we can describe the behavior
of the overall system relatively accurately through the half-life of the iso-
tope. We can thus predict the time when half of the original isotopes will
have decayed. The stochastic events at the microscopic scale can (at least
approximately) be brought back to the deterministic scale by summing up
many individual processes. We find similar examples in thermodynamics:
The macroscopic variables pressure or temperature combine a large number
of microscopic (stochastic) molecular processes. Example 2.7 illustrates by
analogy the movement of atoms in an ideal gas: the movement is stochastic
on the level of the single molecules (balls), but their overall behavior results
in the macroscopic phenomenon of molecular diffusion (see Chap. 8).
 A FIRST LOOK 29

Example 2.7: The bed of nails, a stochastic system

Figure 2.7 depicts schematically a classic toy made of nails fixed on
an inclined wooden board—a bed of nails. A ball, starting from a
fixed position x = 0, rolls down the slope and periodically hits a nail
such that the ball’s path either continues to the left or right with the
same probability. The nails are depicted as squares, the numbers in the
squares give the probability that an individual ball arrives at this nail.
After n collisions (in the figure, n = 6) the ball lands with a specific
probability in one of the final positions (seven in our example). The
path of a specific ball cannot be predicted. But if a large number of
balls is sent through the system, and if the entire setup is built perfectly
symmetrically, we will find a characteristic distribution of balls in the
final boxes. These probabilities are called Bernoulli numbers.

Fig. 2.7: A schematic rep-

resentation of the path of
a ball through the bed of
nails. The path of an indi-
vidual ball cannot be pre-
dicted. If a large number of
balls are successively sent
through the system, we
find a characteristic prob-
ability distribution in the
final boxes. These proba-
bilities are called Bernoulli
numbers. Drawn in bold
are two individual paths
A and B. Path B leads to
the right six times, it has a
probability of 2−6 = 1/64.
The meaning of the vari-
ables labeled location (x)
and time (t) will be dis-
cussed in Sect. 8.4

In fact, we don’t need to go all the way to quantum mechanics to ex-

plain the stochastic behavior of some natural systems. Even the classical
equations of physics, using regular differential equations, contain the seed
of randomness. As we will see in Chap. 6, three coupled differential equa-
tions are sufficient to introduce stochastics into models. Such systems of
equations may have the special property that after a finite time, and start-
ing from almost identical initial states, they develop states which are very
different from each other. This happens even though the model equations
themselves are entirely deterministic. Since every observation is subjected
to some error, we are not able to determine the absolute equality of states.
30 MATHEMATICAL MODELS

With nonlinear differential Thus, apparently identical systems may develop differently. This is called
equations, almost identical deterministic chaos. The weather is a typical example for chaotic behavior.
initial conditions can later We are capable of predicting the weather over a period of a few days rela-
lead to completely different tively well. But our limited knowledge of the atmosphere’s state and other
system states. relevant factors means that a prognosis over 2 months will likely remain
impossible.
When modeling, we can use various techniques to give essentially
deterministic equations a whiff of stochastics. The simplest possibility is
to introduce a randomly varying external relation. In this book, however,
we will focus primarily on deterministic systems.
In this chapter, we got a general idea of some different model types and
thereby encountered some properties with which we can characterize mod-
els. These properties include the pairs static-dynamic, discrete-continuous
(either in space or in time) and deterministic-stochastic. In the following
chapters, we will discuss the different model types in more detail.

Deterministic Chaos
 QUESTIONS AND PROBLEMS 31

2.7 Questions and Problems

Question 2.1: Explain the difference between internal and external rela-
tions.

Question 2.2: Most dynamic models implicitly contain a static model.

The opposite is not true, however. Why?

Question 2.3: What is the difference between a dimension and a unit?

Which of the two is unambiguously defined for every variable?

Question 2.4: What does the dimension of a model mean? (Note: Com-
paring questions 2.3 and 2.4 will make clear that dimension has a different
meaning in mathematics than it does in physics.)

Question 2.5: Which of the following statements is correct?

• A box model is spatially discrete.

• A box model is always continuous in time.

Question 2.6: Look for some examples of systems that contain a stochastic
component.

Question 2.7: Grubs live in the ground for 3 years and hatch as May
beetles in spring of the fourth year. Sketch a simple regional model for
May beetles. In particular, you should think about the following issues:
choice of the system variables, type of model (dynamic or static, in space
or in time, discrete or continuous).

Problem 2.1: Mass balance

A small lake has an inlet and an outlet. Establish the mass balance of a
soluble substance in the lake, for the following two cases:

(a) A factory is built on the lake shore. From time t0 onwards, it continu-
ously discharges a substance into the lake. This substance is volatile,
i.e. it can evaporate into the atmosphere, but is not degraded in the
lake. It doesn’t deposit in the sediment.

(b) Through an accident, a large amount of a toxic substance enters the

lake at time t0 . This substance is degraded in the lake water and de-
posited in the sediment. After the accident, no more of the substance
is added to the lake.

In both cases, you can assume that the substance’s concentration in the
inlet is zero.
32 MATHEMATICAL MODELS

Fig. 2.8: Phosphorus sed- 10

imentation, according to
Vollenweider (1976)

σp [a−1 ]
.1

σp = 10
z

.01
1 10 100 1000
average depth z [m]

Problem 2.2: Determining the dimension of parameters

The following differential equation describes the progression through time
of a substance’s concentration profile in a vertical well shaft:

∂C(z, t) ∂2C ∂C
= −k1 + k2 2 + k3 + k4 C(C ∗ − C)
∂t ∂z ∂z
Determine the dimensions (expressed by (M,L,T)) of the parameters k1 ,
k2 , k3 , k4 and C ∗ . C is the concentration of a substance, z a length and
t the time. Hint: How is the dimension of a first or second derivative of a
variable determined?

Problem 2.3: Phosphorus sedimentation

Empirically established correlations should also be correct in terms of their
dimensions. Figure 2.8 shows the relation between average lake depth z in
−1
meters and the specific phosphorus sedimentation rate σp = 10 z in year
for several lakes. What unit must the factor 10 have? What is the physical
meaning of that factor?

Problem 2.4: Bed of nails

If we extend the bed of nails shown in Fig. 2.7 by two steps (to n = 8),
the final boxes span from m = −8, −6, . . . , +6, +8. What is the probability
that a ball hits the box in the center (m = 0)? Calculate the Bernoulli
numbers at level n = 8.
Chapter 3

Static Models

 Static accounts
Anne to Bill: “Yesterday, I brought 10 dollars to the bank, and the teller
told me that I now had $100 in my account.”—Bill: “Hmm. . . So, if I bring
him 20 bucks tomorrow, my balance will be $200!”

In this chapter we want to take a closer look at static models. In Chap. 2 A static
we’ve seen that with a one-dimensional static model, the system variable V system?

has an exactly defined value for a given external relation R. A static model
therefore simply describes the equilibrium state between R and V and does
not give any indication about how rapidly V adjusts to changes in R. Thus,
static models are particularly useful when V quickly reacts to changes in R
(that is, in cases where we can neglect the time that V takes to change).

In Chap. 2 we learned that pairs of data (such as the phosphorus concen-

tration in a lake and the simultaneously measured phosphorus inflow) can
be used to construct a static model through linear or nonlinear regression.
Frequent applications of static models are generating calibration curves for
analytical devices, stoichiometrically examining chemical equilibrium reac-
tions, or describing the equilibrium distribution between two phases of a
chemical substance in a reactor or the environment.

In three examples, we now want to determine the distribution of benzene

between the three phases air, water and sediment. In doing so we will get
to know two frequently used static models: Henry’s law and the sorption
isotherm. Benzene is a common solvent and is also contained in gasoline.
It is a ring-shaped hydrocarbon with carcinogenic effects. Benzene

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 33

DOI 10.1007/978-3-642-30639-6_3, © Springer-Verlag Berlin Heidelberg 2013
34 STATIC MODELS

3.1 Equilibrium Distribution Between Water and Air

In the first example we want to describe the equilibrium distribution of
benzene between water and air. Let’s imagine an experiment:

Example 3.1: Henry coefficient of Benzene

A flask is half filled with water. Some benzene is dissolved in the water.
The flask is shaken for 5 min while kept closed with an airtight seal.
Afterwards, the benzene concentration is measured in the water (Caq )
and in the air (Cair ). If this experiment is repeated with a varying
amount of benzene, the following series of measurements might be the
result:

Caq [g L−1 ] 1 × 10−2 2 × 10−2 6 × 10−2 1 × 10−1

Cair [g L−1 ] 1.9 × 10−3 4.1 × 10−3 1.1 × 10−2 2 × 10−2

The data pairs are plotted in Fig. 3.1. They lie approximately on a
straight line, so we can formulate the following linear static model:

Cair = p · Caq (3.1)

Through linear regression we obtain the non-dimensional model pa-

rameter p = 0.196 ± 0.004.

Fig. 3.1: Two-dimensional 0.025

diagram for determining
the distribution coefficient
Benzene concentration Cair [g L−1]

Ka/w of benzene between 0.02

air and water. The linear
regression yields Ka/w =
0.196 ± 0.004
0.015

0.01

0.005

0
0 0.02 0.04 0.06 0.08 0.1 0.12
Benzene concentration Caq [g L−1]
 EQUILIBRIUM DISTRIBUTION 35

We can repeat similar experiments with other soluble substances. The

substance can initially exist in its gaseous form in air or dissolved in water.
Through time, it will always be distributed throughout both the air and
the water phases, and in describing this distribution we will obtain an
approximately linear model (at least as long as the added substance mass
is not too large or it is not a reactive substance). The model parameter p is
the air-water distribution coefficient, or rather, the non-dimensional Henry
coefficient Ka/w :
Cair
Ka/w = (3.2)
Caq
The distribution coefficient Ka/w is a substance-specific constant. Hand-
books such as the CRC Handbook of Chemistry and Physics usually give
the dimensional Henry coefficient KH (more about this in Problem 3.1).
Because the atmosphere is the most important transport medium for the
global dispersal of many substances, the Henry coefficient is of substantial
importance. In fact, many substances detected in surface waters originate
from the atmosphere, within which they are rapidly transported from their
point of emission and then dissolved into the water bodies where we mea-
sure them. As we will see later (Sect. 8.2), the Henry coefficient is also used
to dynamically model the phase transition between air and water.

William Henry (1775–1836)

In 1803, he formulated the law named after him: At equilibrium, the concentration of a
gas in water is proportional to its gas pressure over the liquid. William Henry was origi-
nally a physician. His own bad health eventually forced him to abandon that profession,
and he turned his interest to chemistry instead.

William Henry
(1775-1836)

3.2 Equilibrium Distribution Between Water and Sediment

A substance dissolved in water also interacts with surfaces of solid mate-
rials. In this second example, we will therefore look at the distribution of
benzene between water and suspended particles.

Example 3.2: Sorption isotherm of benzene

In this experiment we completely fill the flask with water and suspend
a defined quantity of dried lake sediment within. Again, we add various
amounts of benzene. The flask is then closed with an airtight seal and
shaken for 1 h. Afterwards, the concentration of benzene in the water
36 STATIC MODELS

and on the sediment particlesa is determined. We obtain the following

data:
Caq [g L−1 ] 1 × 10−2 2 × 10−2 6 × 10−2 1 × 10−1
Csed [g kg−1
sed ] 0.018 0.033 0.103 0.17

Once again we find an approximately linear relation between the data

pairs Caq and Csed . This is illustrated as a two-dimensional diagram in
Fig. 3.2. The average regression coefficient is p = 1.7 ± 0.008 L kg−1sed :

Csed = p · Caq (3.3)

aThe concentration on the sediment is given in the dimension substance
mass per sediment mass [ M M−1
sed ].

Fig. 3.2: Sorption 0.2

isotherm of benzene on
−1 ]

a lake sediment: the dis-

Benzenecon centration Csed [g kg sed

tribution coefficient is
Kd = 1.7 ± 0.008 L kg−1sed 0.15

0.1

0.05

0
0 0.02 0.04 0.06 0.08 1.1 0.12
Benzene concentration Caq [g L−1]

This experiment can also be repeated with other chemicals and particles.
The two-dimensional diagram representing the equilibrium distribution of
a chemical substance between water and a solid phase is called sorption
isotherm. In our example, the sorption isotherm is linear. The parameter
p is the slope of the sorption isotherm and is called distribution coefficient
Kd , with a dimension of [ L3 M−1 ]:
Csed
Kd = (3.4)
Caq
 MULTI-DIMENSIONAL STATIC MODELS 37

The distribution coefficient Kd depends on the properties of both the

chemical substance and the solid phase. The concept of a constant distri-
bution coefficient for the partition between a solution and a solid phase
is also used in dynamic transport models for chemicals. One use of such
models is to estimate the dispersion of a pollutant from a landfill site into
an aquifer.1

3.3 Multi-dimensional Static Models

In the third example, we will couple the two models in order to describe the
equilibrium concentration of benzene between water, air and sediment. In
order to minimize the system’s complexity, we look at a reactor (as shown
in the margin illustration) instead of a natural lake.

Example 3.3: Benzene in a reactor with air, water and sedi-

ment
The reactor is filled with a specified amount of water and dried sedi-
ment. Air flows through the reactor’s water volume, continuously mix-
ing it. The benzene concentration in that air is constant at Cair = 0.01
g L−1 . Which benzene concentration will we measure after about 1 h
(i.e., after equilibrium has been reached) in the water and on the sed-
iment particles?

The system “reactor” has an input quantity R (the benzene concentration

in the air Cair ). It further has two variables V1 and V2 : the benzene con-
centration in the water Caq and the benzene concentration on the sediment
particles Csed .

Cair → R
Caq → V1
Csed → V2

To draw up the system equations, we use the relations derived in the two
previous examples. This yields:

1
V1 = R (3.5)
Ka/w
V2 = Kd V1 (3.6)

1 An aquifer is a water-carrying layer in the ground consisting of sand, gravel

or permeable rock formations.
38 STATIC MODELS

This is a two-dimensional, linear, static model with the system parameters:

1
→ p1
Ka/w
K d → p2

Because we know R and the two system parameters, we can calculate the
benzene concentrations V1 and V2 directly with Eqs. (3.5) and (3.6). This
gives us:

1 1
Caq ≡ V1 = R= × 0.01 g L−1 = 0.05 g L−1
Ka/w 0.2
Csed ≡ V2 = Kd V1 = 1.7 L kgsed −1 × 0.05 g L−1 = 0.085 g kgsed −1

To calculate a model with two variables (a two-dimensional model), we

need to solve a system of two coupled equations (Eqs. 3.5 and 3.6) with two
unknown variables. Thus, in order to solve an n–dimensional model, an
n–dimensional system of equations with n unknowns has to be solved. The
system of equations for an n–dimensional static model has the following
general form:

Vi = fi (R1 , . . . , Rm , p1 , . . . , pq , V1 , . . . , Vn ) for i = 1, . . . , n (3.7)

Where:

Vi system variable, i = {1, . . . , n}

Rj external relation, j = {1, . . . , m}
pk model parameter, k = {1, . . . , q}
In our case, the system of equations can be solved analytically by hand.
In cases where this is not possible, software can solve the system compu-
tationally. One algorithm often used for this is the Gaussian elimination
method.
In this chapter, we have used three static models to describe the equi-
librium distribution of a chemical substance in the environment. We have
thereby encountered two important parameters, the air-water and the se-
diment-water distribution coefficients, both of which we will utilize again
for dynamic models.
However, all three examples have a weakness: we do not know how long
it takes for an equilibrium to be established between the different phases.
In the next chapter, dynamic models will let us investigate precisely that
question.
 QUESTIONS AND PROBLEMS 39

3.4 Questions and Problems

Question 3.1: State the types of equations with which static and dynamic
models are generally described.

Question 3.2: Which of the following laws and formulas from physics,
chemistry and biology correspond to a static model?
Don’t worry if you do not know some or all of these laws or can’t
clearly classify them. They are not necessary to understand the material.
In addition, some of them will be discussed in detail later in the book.

(a) Ohm’s law

(b) Kepler’s law of planetary motion, giving the relation between a

planet’s orbital period and the semi-major axis of its elliptical orbit

(c) Equation of pendular motion

(d) Ideal gas law

(e) Fourier’s law of heat convection

(f) Law of radioactive decay

(g) Lensmaker’s equation in optics

(h) The chemical equilibrium between dissolved carbonate and bicarbon-

ate in water

(i) The law for light passing through an optical prism

(j) Maxwell’s laws of electrodynamics

(k) The logistical growth law of a biological species

Problem 3.1: Dimensional Henry coefficient

Often, the concentration of a substance in the gas phase is given by the gas
pressure P of that substance. The Henry coefficient of a substance is then
defined as: KH = CPaq . Which dimension and unit does KH have? How is
it linked to the non-dimensional air-water distribution coefficient Ka/w , if
the gas phase of the substance in question can be approximated as an ideal
gas?

Problem 3.2: Henry coefficient of methyl bromide

Methyl bromide (CH3 Br) is a gas used in greenhouses to fight insect larvae,
for instance in salad cultivation. Using the following data pairs (measured
at a temperature of 20 ◦ C), estimate the Henry coefficient KH for methyl
bromide.
40 STATIC MODELS

P [atm] 0.2 0.3 0.4 0.6 0.7

Caq [mol L−1 ] 0.03 0.05 0.06 0.095 0.1

From KH , calculate the non-dimensional air-water distribution coefficient

Ka/w through the relation Ka/w = K H
RT . Hint: the solution of Problem 3.1
will be of use here.

Problem 3.3: Methyl bromide as ozone killer

Methyl bromide is one of the substances significantly contributing to ozone
destruction in the stratosphere. The concentration of methyl bromide in
a water sample taken from a water jar in a greenhouse is Caq = 0.001
mol L−1 . Calculate the total mass of methyl bromide in kg that was present
in the greenhouse at the time the sample was taken. The greenhouse is 5 m
wide, 20 m long and 2.5 m high. Assume that the system is at equilibrium
regarding the water/air distribution.

Problem 3.4: Nonlinear sorption isotherm

Nitroaromates are a class of substances that exhibit sorption on mineral
surfaces. The following table shows the result of sorption experiments with
1,4-Dinitrobenzene (1,4-DNB) on the clay mineral Kaolinite (data from
Haderlein et al. 1996):

Caq [μmol L−1 ] 0.17 0.51 1.8 3.6 7.6 19.5 26.5
1,4-Dinitrobenzene Cmin [μmol kg−1
min ] 241 633 1,640 2,850 4,240 6,100 7,060

Draw a two-dimensional diagram with the data pairs and try to find a
model with a maximum of two free parameters to describe the sorption.

Problem 3.5: Economic theory: supply and demand

A central element of economic theory is the assumption that, for a given
good, the price is determined by an equilibrium between supply and de-
mand (see e.g. Samuelson and Nordhaus 2009). The following empirical
observations serve as the basis for a graphical solution:
1. Demand for the good increases with decreasing price (Fig. 3.3a).

Fig. 3.3: Economic mod- a b

els: (a) Demand curve. (b)
Supply curve

Supply curve
Price per good

Demand curve

Sold goods Manufactured goods

 QUESTIONS AND PROBLEMS 41

2. Supply for the good increases with increasing price (Fig. 3.3b)
(a) Use the two curves to graphically determine the resulting price.
(b) What happens to the price of the good if, for a constant
supply curve (Fig. 3.3b), the demand grows, i.e. the demand
curve (Fig. 3.3a) moves to the right?
(c) What happens to the price and the number of goods sold if the
manufacturing costs decrease (for a given price, more goods can
be manufactured)?
Chapter 4

Linear Models with One Variable

In the remaining part of the book, we will exclusively deal with dynamic
models. As we already noted in the introduction, one of the principal char-
acteristics of natural systems is that they are not static.
Dynamic and continuous-time systems are usually described with differ-
ential equations. Although natural systems are hardly ever linear in reality,
we will first take a close look at linear models (and thus linear differen-
tial equations). The field of systems analysis is rooted in physics and the
technically-oriented sciences. In those fields, linear models play a particu-
larly important role. They have a key feature: they have analytical solutions
which can be linearly combined at will to construct new solutions. These
simple solutions are well-suited as building blocks for more complex models.
Real systems are usually nonlinear. Often, however, they can be approx- Nonlinear systems can of-
imated by linear models in a piecemeal way, that is, within limited ranges ten be approximated by lin-
of the system variable. As we will see in Chap. 6, the concepts developed for ear models piece by piece.
linear models serve as the point of departure for the analysis of nonlinear
models.
First, let us examine continuous-time systems with only one system
variable. The dynamic models resulting from this describe the change of a
single variable of the system. We do not spatially differentiate the system,
we treat it as one box, for instance as a thoroughly mixed volume. The
mathematical model for such a system is given by the following first-order
differential equation:
dV(t)
= R + pV(t) (4.1)
dt
Where:
V(t) System variable
R External relation
p Model parameter

Because such a model contains only a single system variable in a single

spatial box we also call it a one-box model. Equation (4.1) is an inho-
mogeneous, linear, first-order differential equation. The external relation R

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 43

DOI 10.1007/978-3-642-30639-6_4, © Springer-Verlag Berlin Heidelberg 2013
44 LINEAR MODELS WITH ONE VARIABLE

constitutes the inhomogeneous term of the differential equation. The

equation stays linear even if R and p are time-dependent.1 The impor-
tant condition is that R and p are not dependent on the system variable V
itself. If R = 0, the equation is called homogeneous.

4.1 The Linear One-Box Model as a Balance Equation

 Mathematician’s mass balance

Two mathematicians are standing outside a lecture hall, impatiently wait-
ing for all students to leave. The first says: “Now there’s only five left”.
Shortly afterwards eight people leave the room. The two mathematicians
stare at each other in astonishment for a moment. Then the second says:
“Now if three people go in, the room will finally be empty!”

Whatever the system variable of a one-dimensional system describes, we

can always characterize its dynamic behavior with a balance equation of
the following form:

dV
= {Production processes} − {Loss processes} (4.2)
dt

By “production” we mean all processes through which the value of the vari-
able V is increased. Apart from processes taking place within the system,
this also includes transport processes from the environment into the sys-
tem. “Loss” means all reaction and transport processes through which the
value of the variable V decreases.
If Eq. (4.2) describes a linear model, both process types consist of a sum
of zero-order and first-order processes:

{Production processes} = Jp + kp V
(4.3)
{Loss processes} = Jl + kl V

Jp and Jl are the zero-order rates independent of V, while (kp V) and (kl V)
are the first-order rates that are proportional to V. The parameters kp and
kl are called first-order specific (conversion) rates. They always have the
dimension [ T−1 ] independently of V.

1 If p is explicitly time-dependent, the term pV is a mixture of an internal and

external relation (see the explanations on page 45 following Example 4.1).
 BALANCE EQUATION 45

Balance equation

Production processes and loss processes dv/dt is not looking good at all. We don t like it when our students
are not balanced in other words, we are know less at the end of the semester
losing knowledge. than at the beginning.

Ok.
I see.

Oops...

If we plug Eq. (4.3) into Eq. (4.2), we get2 :

dV
= (Jp + kp V) − (Jl + kl V)
dt
= (Jp − Jl ) + (kp − kl ) V (4.4)
= J  + k V

The net coefficients J  = Jp − Jl and k  = kp − kl can be positive or

negative depending on the value of the individual parameters Jp , Jl , kp and
kl . What is important is that any one-dimensional linear model can be
reduced to the form of the last line of Eq. (4.4), independent of how diverse
the individual production and loss processes might be. In the most general
case, J  and k  are arbitrary time-dependent quantities but do not depend
on V.
Let’s look at an example:

Example 4.1: Fish growth in a pond

In a pond, the total amount of fish M (expressed as the total fish
biomass in kg) increases by egg deposition and fish growth after sub-
tracting mortality with the specific rate kp = 0.2 year−1 . Further-
more, the fish pond is restocked yearly with juvenile fish at Jp =
500 kg year−1 . On average, 40 % of the fish biomass is fished per year
(loss rate kl = 0.4 year−1). What is the dynamic equation for fish
biomass M ?

2 From now on, to make things easier to read, we will usually not write the
time dependence of the system variables. The notation V will stand for V(t).
46 LINEAR MODELS WITH ONE VARIABLE

The model consists of three production and loss terms. One of them is
zero-order:
J  = Jp = 500 kg year−1
The two others are first-order, they have the specific net rate:

k  = kp − kl = (0.2 − 0.4) year−1 = −0.2 year−1

So, overall:
dM
= 500 kg year−1 − (0.2 year−1 ) M
dt
We will return to this in Example 4.3.

In Sect. 1.2 we differentiated between internal and external relations. Where

do we find this distinction in Eq. (4.4) and Example 4.1? We will see that
to a certain degree the answer to this question depends on the context of
Eq. (4.4) and the underlying mathematical model.
In most cases, the first-order processes belong to the internal relations,
as they are controlled by the momentary value of variable V(t). If the spe-
cific rate k  is constant in time, the term (k  V) indeed only varies as a result
of the time-dependent variability of V(t). This implicit time dependence is
a typical attribute of internal relations.
Yet, k  may be time-dependent as well. This explicit time dependence
is then not a property of the model itself but an influence from outside. In
this case k  (t) is an external relation. Thus, the first-order process k  (t)V(t)
becomes a combination of internal and external relation. For instance, in
Example 4.1, at a given instant the fish breeder may decide to protect
the pond from fish-eating birds by laying out a net on its surface, thereby
making the parameter kl smaller. In the model this process would be an
external relation.
The zero-order term J  can also explicitly depend on time (but not
implicitly). If it does, then it certainly represents an external relation. If
it is constant, however, there is some room for interpretation: either we
interpret J  as a fixed (i.e. immutable) system parameter, or we see it
as an external relation. For that latter case, we can argue that although
the external relation is constant momentarily, it could be changed from
“outside the model” at any time.To stick to the fish pond example: the
rate of fish addition and the fishing rate could in principle be changed,
even if we choose to analyze the model only for the case of fixed values.
Summarizing, the important point is that an explicit time dependence of
system coefficients always represents external relations.
 CONSTANT COEFFICIENTS 47

4.2 Linear Models with Constant Coefficients

Linear decay process
In a linear decay process, the speed of
decay is proportional to the substance s
mass. In the case of our brain, this
means that each year, we lose 1/4 of
our knowledge.

First, we will look at the case of constant (time-independent) coefficients

J  and k  . We can interpret this model as free from external relations,
in other words, as autonomous. Starting from identical initial states, an
autonomous model always develops in the same way.3
We will again start from the general equation (4.4):

dV
= J  + k V (4.5)
dt
It has the initial condition V(0) = V0 and the following solution (as long
as k  is not zero):

J  k t J 
V(t) = (V0 + )e −  for k  = 0 (4.6)
k k
For the homogeneous differential equation
dV
= k V (4.7)
dt
the solution is accordingly:

V(t) = V0 ek t
(4.8)

By inserting Eq. (4.6) or (4.8) into the original differential equation, you
can easily validate the result given above. The initial condition V(t = 0) =
V0 is correctly expressed, as can be seen by setting t = 0 in the solutions.
Let’s first discuss the solution of the homogeneous Equation (4.8). As
can easily be seen, V(t) behaves entirely different for t −→ ∞ depending
on whether k  is negative or positive (we will not discuss the trivial case
of k  = 0). The two cases are shown in Fig. 4.1. They are the exponential

3 To be precise, this is only true for deterministic models (see Sect. 2.6).
48 LINEAR MODELS WITH ONE VARIABLE

growth and decay curves. For the case of a negative k  we introduce the
new specific rate k (which we will use again later):

k = −k  , k>0 (4.9)

To identify the exponential curve it helps to divide both sides of Eq. (4.8)
by V0 and then to calculate the natural logarithm on both sides:

V(t) 
ln( ) = ln V(t) − ln V0 = ln(ek t ) = k  t
V0

or:
ln V(t) = ln V0 + k  t (4.10)
Thus, if we plot ln V(t) against time t, the curve should be a straight line
with slope k  (see Fig. 4.1).

Fig. 4.1: Solution of the

linear homogeneous differ- (t) = 0 k t
e
ential equation for V0 = 1 ln (t)
with a positive (upper k >0
curve) and negative (lower 10
curve) coefficient k . The
half-logarithmic diagram
(on the right-hand side) re- 1
sults in a straight line with
slope k (see Eq. 4.10)
(t)

1
0

0 0.1
Relative system variable

0 −k t
(t) = e ln (t)
1
k = −k > 0
1

0.1

0 0.01
Time t →

Example 4.2: Exponential population growth

From 1850 to 1975 the world population grew from N 0 = 1 billion to
N (t) = 4 billion people. We assume that specific growth was constant
throughout that entire period (which of course is a gross oversimplifi-
cation). What is the specific growth rate kp ?
 CONSTANT COEFFICIENTS 49

Global population
The differential equation of the exponential growth model is analogous (billion people)
to Eq. (4.7):
dN
= kp · N (4.11)
dt
We arrive at the solution with Eq. (4.8):

N (t) = N 0 ekp t (4.12)

Solving Eq. (4.12) for kp and using t = 125 year, N (t)/N 0 = 4, we get:

1 N (t) 1
kp = ln = ln 4 ≈ 0.01 year−1 (4.13)
t N0 125 year

In many models, the specific total rate k  is negative. With the definition
Eq. (4.9) we can write the resulting differential equation in the following
form, also replacing J  with J to simplify further:

dV
=J −k V , k>0 (4.14)
dt
This equation has the solution:

J −kt J
V(t) = (V0 − )e + , k>0 (4.15)
k k

4.2.1 The Steady State

Buckingham Palace
Typical steady state:
not perturbed by
anything!

Cheeeeese!

After establishing the system equation for an autonomous linear model

and getting to know its solution, we now want to interpret it and use it to
draw conclusions about the model. The model is in a special state when
dt = 0. At that point V remains constant, and the model is in a so-called
dV

steady state. In mathematics, this state of a model is also called fixed point.
If the system variable is the average phosphorus concentration in a lake, the
concentration stays constant when all input and output processes exactly
balance each other. Since the size of some of these processes, e.g. the output,
50 LINEAR MODELS WITH ONE VARIABLE

depends on V, this occurs at a specific concentration which is called the

steady-state or equilibrium concentration.
Formally, we calculate the steady state value V∞4 by setting the left
side of Eq. (4.5) equal to zero and solving it for V = V∞ :

J J
V∞ = − = , k = 0 and k  = 0 (4.16)
k k

The homogeneous equation (4.7) has the steady state V∞ = 0.

In many cases, the meaning of the variable V precludes negative values
(e.g. when V represents the amount or concentration of a substance). In that
case Eq. (4.16) only makes sense if either J  or k  (but not both of them)
are negative. In Example 4.1 we encountered a case with J  > 0, k  < 0.
Let us now further analyze the dynamics of the fish pond.

Example 4.3: Steady-state fish biomass in the pond

In Example 4.1 we developed a model for a fish pond that is artificially
stocked with new fish while simultaneously being fished. We are now
interested in the steady state amount of fish biomass in the pond.
According to Eq. (4.16), the following holds:

J 500 kg year−1
M∞ = − = − = 2,500 kg
k −0.2 year−1
Note: Of course, this representation of the fish pond is very simplistic.
In nature, growth and death rates are never entirely constant, so that
M ∞ would always fluctuate around an average value.

Each dynamic model im- To summarize: By setting the left side of Eq. (4.5) equal to zero, the differen-
plicitly contains a static tial equation becomes the ordinary algebraic equation (4.16). This equation
model. can also be seen as a static model for V∞ as a function of J  and k  . In that
sense, dynamic models that have steady-state solutions implicitly contain
a static model as well.

4.2.2 The Linear Flow Reactor

An application of Eq. (4.14) is the so-called completely mixed linear flow
reactor. We can use it to describe a lake as a one-box model:

4 The symbol for infinity, which is used here as superscript, indicates that the
steady state is formally only reached after an infinite amount of time.
 CONSTANT COEFFICIENTS 51

Example 4.4: A lake as linear flow reactor

We consider a lake as a pool of water where a constant amount of wa-
ter flows through, that is, inflow and outflow are equal and the water
volume in the lake remains constant. Starting at time t0 , a constant
amount of a substance per time is added through the inflow. The sub-
stance is completely mixed in the lake and decays with a first-order
reaction. Therefore, the concentration is the same everywhere in the
lake, especially also in the water leaving the lake through the outflow.
Figure 4.2 shows this model as a box diagram.

Fig. 4.2: A lake as a com-

pletely mixed flow reactor
(one-box model) with a de-
cay reaction

To model this situation, we first establish a mass balance for the substance
with the help of Eq. (4.2):
       
Change of mass Supply Removal Reaction
= − −
in the lake per time per time per time per time
Mathematically, the mass balance can be formulated as a differential equa-
tion (M : total mass in the lake [ M ], R: reaction per time [ M T−1 ]):
dM
= Jin − Jout − R (4.17)
dt
For the mass input Jin and the mass removal Jout we can write:

Jin = QCin (4.18)

Jout = QC (4.19)

Q [ L3 T−1 ] Flow through the lake, inflow = outflow

Cin [ M L−3 ] Concentration of the substance in the inflow
C [ M L−3 ] Concentration of the substance in the lake and outflow

For the first-order reaction R of the substance in the lake we can write:

R = kr M (4.20)

where kr is the first-order reaction rate and has the dimension [ T−1 ]. Thus,
we end up with the following mass balance:
dM
= QCin − QC − kr M . (4.21)
dt
52 LINEAR MODELS WITH ONE VARIABLE

If we divide both sides of Eq. (4.21) by the lake’s constant volume V ,

we get the dynamic equation for the substance concentration in the lake
(C = MV ):
dC Q Q
= Cin − C − kr C
dt V V
= kw Cin − (kw + kr )C (4.22)
= kw Cin − ktot C

The quotient Q 1
V is called the specific flow rate kw . The inverse, kw , is
the renewal time τw of the lake water. The sum kw + kr is the total specific
1
rate ktot . The inverse, ktot , is the mean retention time of the substance in
the lake (influenced by both flow and decay).
Equation (4.22) can also be written in terms of the substance input Jin :

dC Jin
= − ktot C
dt V (4.23)
= jin − ktot C

The variable jin is the substance input per volume and time with the di-
mension [ ML−3 T−1 ]:

Jin Q
jin = = Cin = kw Cin (4.24)
V V
Comparing the dynamic equation for the flow reactor (4.23) with
Eq. (4.14) yields the following analogy: The system variable in the flow
reactor is C. The inhomogeneous term J corresponds to jin = kw Cin . The
coefficient k is given by the total specific rate ktot = kw + kr .
We obtain the equilibrium (steady state) concentration in the lake with
Eq. (4.16):

jin kw kw
C∞ = = Cin = Cin (4.25)
ktot ktot kw + kr

The concentration C ∞ is only reached if the substance input Jin , or jin ,

and the specific rate ktot remain constant.
Let us compare this result with the static lake model in Sect. 2.2. Ac-
cording to Eq. (2.2), the equilibrium concentration of phosphorus Caq ,
which we now call C ∞ , is:

Caq → C ∞ = p · Jin (4.26)

The equilibrium concentration C ∞ of the dynamic model can also be

expressed, by means of Eqs. (4.18) and (4.25), as a function of substance
input Jin :

kw 1 Q 1
C∞ = Cin = Cin = Jin (4.27)
ktot ktot V ktot · V
 CONSTANT COEFFICIENTS 53

The parameter p of the static model thus turns out to depend on the total
rate constant and the lake volume:
1
p= (4.28)
ktot · V
As noted before, p has the dimension [ TL−3 ].
In the following example we utilize the relation between model param-
eters and steady state:

Example 4.5: Algae growth in a pond with flowing water

In a pond with constant water through-flow and volume V = 104 m3 ,
favorable conditions let free-floating algae (phytoplankton) grow with
a specific growth rate kg = 0.5 d−1 . Through sedimentation, 20 % of
the plankton in the pond is lost daily. How large can the water flow Q
be at most, if the algae are not to disappear entirely from the pond?

First, we establish a mass balance for the system:

       
Change of algae Growth Outflow Sedimentation
= − −
in pond per time per time per time per time

We assume that the conditions for a linear flow reactor (one-box model)
are fulfilled. In that case, all three processes of the mass balance can be
described as linear functions:
dC
= kg C − kw C − ks C
dt (4.29)
= (kg − kw − ks )C

where:
kg = 0.5 d−1 specific growth rate of the algae
kw = Q/V flow rate through the pond
ks = 0.2 d−1 sedimentation rate of the algae in the pond

For the algae concentration not to decrease, the sum of all rates on the
right side of Eq. (4.29) must be positive or zero:

kg − kw − ks ≥ 0 (4.30)

Solving for kw = Q/V results in:

Q
kw = ≤ kg − ks (4.31)
V
For the flow Q we then get:

Q ≤ (kg − ks )V = 104 m3 (0.5 − 0.2) d−1 = 3,000 m3 d−1 (4.32)

54 LINEAR MODELS WITH ONE VARIABLE

4.2.3 Adjustment Behavior and Adjustment Time

Welcome to our weekly group Our first agenda item is One week later...
meeting on improving strategic team-conforming behavior, adaptation
research. Thank you all for coming. and identification. Dang?

Ties...

Up to this point, the dynamic treatment of a system has not brought

us anything beyond what we could have, in principle, derived from the
static model. However, the information contained within a dynamic model
reaches much beyond the steady state. For example, we are interested in
whether the system is indeed moving towards the steady state calculated
in Eq. (4.16) and if so, how long it will take to reach it.
Again, let’s consider a lake as a linear flow reactor. We now have to solve
the differential equation (4.22) by replacing the variables of the general
solution (4.15) with the variables of the flow reactor. To keep things simple,
we will again call the linear system’s total rate ktot simply k. It can, if not
explicitly stated otherwise, describe any process such as flow, decay (e.g.
radioactive), or sedimentation, but also (for k < 0 or k  > 0, see Eq. 4.9) a
linear growth process. We then get:

kw kw
C(t) = (C 0 − Cin )e−kt + Cin
k k (4.33)
= (C 0 − C ∞ )e−kt + C ∞

In the solution, the steady state concentration C ∞ = kw

k Cin appears. We
rewrite the equation in the form:

C(t) = C 0 e−kt + C ∞ (1 − e−kt ) (4.34)

and assume that k is positive. The first term on the right-hand side of
Eq. (4.34) describes the exponential decay of the initial concentration C 0
in the system. For t → ∞, this term becomes zero. The old system state
C 0 is “washed out” of the system; the term describes the so-called washing-
out or decay curve. In contrast, the second term describes how—starting
from concentration zero—the system develops towards the steady state
belonging to the (constant) input jin = kw Cin . We will call this term the
ingrowth curve. Figure 4.3 shows these two components of the solution as
well as their sum. In the language of mathematics one can also say that
the general solution of Eq. (4.34) is a linear superposition of two special
solutions, the net decay curve and the net ingrowth curve.
 CONSTANT COEFFICIENTS 55

C0
ln | C(t )−C ∞ |
C∞
Ingrowth curve

Decay curve
0
Time t Time t

Fig. 4.3: A linear system’s change from its initial state C 0 to the steady state C ∞
can be interpreted as the sum of two processes, the “decay” of the initial state C 0
(C 0 e−kt ) and the “ingrowth” of the end state C ∞ (C ∞ (1−e−kt )). The continuous
curve, the sum of the two dashed curves, shows the course of the system C(t).
Plotting on a logarithmic scale (on the right) gives a straight line with slope −k

Similar to Eq. (4.10) we can rewrite Eq. (4.33) in the following form:

C(t) − C ∞
ln( ) = ln(C(t) − C ∞ ) − ln(C 0 − C ∞ ) = −kt (4.35)
C0 − C∞

Like in Fig. 4.1, if ln(C(t) − C ∞ ) is plotted against t we get a straight line

with slope −k (Fig. 4.3).5
With that, we have used a dynamic model to calculate the development
of the concentration in the lake outside of the steady state. Furthermore,
we have shown that independent of the initial value C0 , the system always
moves towards the steady state if the rate k is positive.6
We now want to turn our attention to the second question: how long
does it take for the lake to reach its equilibrium concentration C ∞ ? In fact,
we already know the answer to this question: because the system adjusts
exponentially to a new steady state, the equilibrium concentration C ∞ is
only reached for t → ∞. It therefore makes more sense to calculate the time
after which the concentration falls below a specifically chosen deviation δC
from the equilibrium. We will call this the adjustment time. The adjustment time is the
A reasonable choice for δC is to define the residual deviation as a fixed time after which a model
fraction κ of the initial difference between steady state and initial state: has approached its steady
state up to a given dis-
∞ tance.
δC = κ |C − C | = κ δC ,
0 0
κ>0 (4.36)

5 The middle part of Eq. (4.35) only makes sense if (C 0 − C ∞ ) and thus also

(C 0 − C(t)) are positive. How to modify Fig. 4.3 for the case where C 0 < C ∞ is
left as an exercise to the reader.
6 As a reminder: In the original formulation of the linear differential equa-

tion (4.5) we introduced the first-order term with a positive sign (specific rate
k ). Thus, a finite steady state exists when k < 0.
56 LINEAR MODELS WITH ONE VARIABLE

By choosing the absolute value |C ∞ −C 0 |, δC is always defined as a positive

number, independent of whether the initial state is above or below the
steady state.
It depends on the user’s requirements or preferences what value is chosen
for the adjustment criterion κ. If, for example, the concentration C has a
measurement error of ±10 % anyway, it would make little sense to choose
κ smaller than 0.1. There are cases where the value of the system variable
is known with high precision, however, and adjustment is then understood
as a very small deviation from the steady state. In such a case we would
choose a small κ.
We define the adjustment time corresponding to κ as τκ . The concen-
tration at that time can be written with Eq. (4.34) as:

C(τκ ) = C 0 e−kτκ + C ∞ (1 − e−kτκ )

(4.37)
= C ∞ + e−kτκ (C 0 − C ∞ )

Conversely, the absolute value of the deviation between C ∞ and C(τκ ) shall
have the same value as κδC 0 :

|C ∞ − C(τκ )| = κ δC 0 = κ |C ∞ − C 0 | (4.38)

The combination of these two equations yields:

|C ∞ − C 0 | e−kτκ = κ |C ∞ − C 0 | (4.39)

And simplified further:

κ = e−kτκ (4.40)
Taking the natural logarithm on both sides gives the following result for
the adjustment time τκ :
ln κ
τκ = − (4.41)
k
Note: Because κ is a number between 0 and 1, ln κ is negative. Because of
the minus sign in Eq. (4.41) the adjustment time τκ is therefore positive.
If k < 0, the steady state would be at infinity and the definition of an
adjustment time would not make sense.
In many cases choosing κ = 0.05 (adjustment to 5 %) appears to be
sensible. According to Eq. (4.41) the following holds for the corresponding
adjustment time:
ln 0.05 2.9957 3
τ5 % = − = ≈ (4.42)
k k k
Note: Whichever way κ is chosen, the corresponding adjustment time will
always be proportional to k −1 . Only the factor in the numerator of
Eq. (4.41) changes.
Figure 4.4 shows how a linear system approaches the steady state. Due
to Eq. (4.39), the initial difference δC 0 disappears exponentially, that is,
as e−kt . After time t = k1 , δC has sunk to δC0 e−1 = 0.37 δC0 , after time
t = k2 to δC0 e−2 = 0.14 δC0 , etc. From Eq. (4.42) we already know that for
t = k3 the residual deviation is 0.05 or 5 %.
 CONSTANT COEFFICIENTS 57

C(t)
Fig. 4.4: Adjustment time
0.05 C 0 C for a linear system, drawn
0.14 C 0 for the case C 0 < C ∞ .
0.37 C 0
After elapsed time t >
τ5 % = k3 , the deviation
C0 from the steady state is
less than 5 %

C0
1/k 1/k 1/k
t

Let’s now look at some examples:

Example 4.6: Adjustment time

For a considerable time, substances have been discharged into a lake
with constant inflow and outflow. The lake’s water renewal time is
τw = 10 year. One substance is decomposed in the lake by a first-order
reaction with the reaction rate kr = 0.9 year−1 . The second substance
is conservative.a The inflow of both substances is suddenly stopped.
How long does it take for the respective concentrations in the lake to
drop to 5 % of their initial value?
For the conservative substance, the 5 % adjustment time is calculated
from just the specific flow rate kw = τ1w = 0.1 year−1:
3 3
τ5 % ≈ = year = 30 year
kw 0.1
For the reactive substance, the adjustment time is shorter, because in
addition to the washing-out process it is also decomposed in the lake:
3 3
τ5 % ≈ = year = 3 year
kw + kr (0.1 + 0.9)
a A conservative substance does not react with its environment, i.e. it
does not decompose and does not deposit. However, it is removed from
the lake via its outflow.

Example 4.7: Half-life

A well-known example of an adjustment time is the half-life τ1/2 of a
radioactive isotope. The half-life is defined as the time in which the
activity of a radioactive source A(t) decreases by half compared to the
initial activity A0 :
A(τ1/2 ) = 0.5A0 (4.43)
58 LINEAR MODELS WITH ONE VARIABLE

For the activity of a radioactive isotope with the decay constant kλ

the following holds:
A(t) = A0 e−kλ t (4.44)
With that we can calculate the half-life τ1/2 :

A(τ1/2 )
= 0.5 = e−kλ τ1/2
A0
ln 0.5 = −kλ τ1/2 (4.45)
ln 0.5 ln 2 0.693
τ1/2 = − = =
kλ kλ kλ
The half-life corresponds to the adjustment criterion κ = 0.5.

Until now, we calculated the adjustment time from a relative criterion (rel-
ative adjustment criterion κ): it only depends on the rate k, not on the
initial state of the system (see Eq. (4.41)). It can happen, however, that
we are interested in an absolute criterion, for instance, the condition that
the concentration should approach the steady state up to a deviation of no
more than 0.02 mg L−1 . As the following example shows, the corresponding
adjustment time then also depends on the initial concentration.

Example 4.8: Adjustment time to a concentration limit

A toxic substance created as a byproduct in a chemical plant is drained
with the effluent into a river. According to environmental regula-
tions, this substance’s concentration in the effluent must not exceed
0.1 µg L−1 . However, the values in the effluent fluctuate between 1 and
10 µg L−1 . Therefore it has to be treated in a reactor before being
discharged.
The reactor is not operated in through-flow mode. Instead, once filled
by the effluent it stays closed until the allowed concentration is reached.
The elimination rate for the substance in the reactor is kr = 0.1 h−1 .
How long must the substance stay in the reactor before we are certain
that it doesn’t exceed the legal concentration limit?
To calculate the adjustment to a concentration limit, we have to
take into account the initial concentration C 0 = 1 µg L−1 and C 0 =
10 µg L−1 . The initial concentration must be reduced to 10 % in the
first case and to 1 % in the second. This implies that the adjustment
criterion κ (see Eq. 4.36) is 0.1 for the first case and 0.01 for the second.
From Eq. (4.41) we calculate the corresponding adjustment times:

− ln 0.1 2.3
τ10 % = ≈ h = 23 h
kr 0.1
− ln 0.01 4.6
τ1 % = ≈ h = 46 h
kr 0.1
 TIME-DEPENDENT COEFFICIENTS 59

In order to meet the discharge concentration limit even for the worst
case (C 0 = 10 µg L−1 ), the effluent must be pretreated in the reactor
for at least 46 h, i.e. about 2 days. Note that even in the best case, i.e.
with an effluent concentration ten times smaller, we would still need
a retention time of about 1 day or just about half the time needed for
the worst-case effluent concentration.

4.3 Time-Dependent Coefficients

Until now we assumed that the coefficients J  and k  in the differential
equation (4.5) are constant over time. We have called such a model au-
tonomous. Natural systems, however, are changing all the time. For in-
stance, over the course of a year or even a day the temperature of a lake
can change markedly. This in turn may influence the reaction rate kr of a
substance or the growth rate kg of organisms in the lake. The material input
into a natural system is rarely constant. For instance, the concentrations in
the effluent of a sewage treatment plant typically vary with a daily rhythm.
Thus, natural systems must usually be seen as non-autonomous, which Non-autonomous mod-
means that the coefficients are time-dependent. Here we will limit our dis- els are explicitly time-
cussion to the case where only the inhomogeneous term J or J  , but not dependent.
the rate k or k  , is time-dependent7 :
dV
= J(t) − k · V (4.46)
dt
We continue to treat the specific rate k as constant. The case of a variable
k is discussed in Appendix C.1.
For the linear flow reactor with variable input we can write the following
dynamic equation:
dC Jin (t)
= jin (t) − kC, with jin (t) = (4.47)
dt V
Here, jin (t) is the time-dependent substance input per volume of the sys-
tem.
As before, we can calculate the steady state by setting the left side of
Eq. (4.47) equal to zero. We then get:
jin (t)
C ∞ (t) = (4.48)
k
Because jin (t) depends on time, so does the steady state. C ∞ (t) is that
concentration which the system would ultimately reach if the input at time
t (i.e. jin (t)) were to remain fixed at its current value for infinite time.

7 For the following discussion, we use Eq. (4.14) with the minus sign in front of

the first-order reaction, since this equation has a finite steady state for J > 0.
In principle one could also use Eq. (4.5) as the starting point.
60 LINEAR MODELS WITH ONE VARIABLE

The solution of Eq. (4.47) can be constructed with the aid of Eq. (C.8)
from Appendix C.1:
t =t

C(t) = C 0 e−kt + e−k(t−t ) jin (t )dt (4.49)
t =0

I would rather At first glance, this equation seems very complex. Yet, the first term on
have a bone than
these scrawny
the right-hand side should look familiar. In the solution of the case with
integrals! constant coefficients (Eq. 4.34) it describes the linear decay of the initial
state C 0 , that is, the decay curve (Fig. 4.3). In essence it describes the
behavior of all processes that occurred before time t = 0. In contrast,
the integral forming the second term of Eq. (4.49) constitutes the “book-
keeping” for the input from time t = 0 up to the present time t = t. This
also explains the role of the auxiliary variable t . Because we characterize
the present with the time t and look back from there, we need a second
variable t to keep track of which section of the past we are dealing with at
the moment.
In the integral, the input function jin (t) is summed up for the time
interval 0 to t. However, because the elimination process (−kC) is active
in the system as well, only a fraction of the introduced substance remains
at time t. The proportion still remaining in the system is larger the more

recent the substance input took place. The factor e−k(t−t ) describes exactly
this effect: for the current input t = t, the exponential function is 1. For
the input farthest in the past, i.e. for t = 0, the weighting is e−kt . The
influence of any input dating back even further is already contained in the
initial value C 0 .
Formally, we can therefore make the structure of the discussed integral
plausible by imagining that the continuous input is divided into many small
parts. Because the system is linear, the influence of all these small input
events can be summed up to get the overall temporal development of the
system. In mathematics this is called the linear superposition principle.
Figure 4.5 depicts this principle for two input events. The first input takes
place at time t1 with the magnitude j1 and duration Δt. Due to the input
event, the amount j1 Δt of substance per volume is added to the system.
For this first event, we can consider the system as a homogeneous system
with the initial condition C 0 = j1 Δt. The development of the concentration
for C1 (t) is then8 :
C1 (t) = j1 Δt e−k(t−t1 ) (4.50)
The second input event yields accordingly:
C2 (t) = j2 Δt e−k(t−t2 ) (4.51)
The expression (t−t2 ) in the exponential function indicates that the second
input took place at time t2 .

8 In the following expression, t is the starting time of the integration and (t −

1
t1 ) is the elapsed time since t1 .
 TIME-DEPENDENT COEFFICIENTS 61

jin(t ) j2 Fig. 4.5: In a linear sys-

tem the concentrations of
j1
two input events (curves
C1 and C2 ) sum up to
form the total concentra-
t t
tion (curve C)
t
t1 t2
Ci (t )

C1, C
C2

C1
t
t1 t2

The superposition of both solutions results in:

C(t) = C1 (t) + C2 (t) = j1 Δt e−k(t−t1 ) + j2 Δt e−k(t−t2 ) (4.52)

If the continuous input is partitioned in ever more and smaller input events,
the following sum results:
n
C(t) = ji Δt e−k(t−ti ) (4.53)
i=0

Finally, by transforming the sum into an integral (Δt → 0, n → ∞),

Eq. (4.53) becomes:
t

C(t) = jin (t )e−k(t−t ) dt (4.54)
0

Note that the input function jin (t ) can take an arbitrary form. In certain
cases, the resulting integral can be solved analytically, whereas in others
only a numerical integration is possible.

4.3.1 Exponentially Growing Input

Natural systems are often exposed to an external force which grows at an
increasing speed. Such an input can be described by an exponential function
(at least for a limited interval of time). We will explore this case with the
example of a lake facing increasing eutrophication.
62 LINEAR MODELS WITH ONE VARIABLE

Example 4.9: Increasing eutrophication of a lake

In a lake, the phosphorus load grows about 10 % per year. In the ref-
erence year (t = 0), the phosphorus input is jin (0) = 6 mg m−3 year−1
and the average phosphorus concentration in the lake is C 0 =
50 mg m−3 . We now want to calculate how large the phosphorus con-
centration in the lake will be 4 and 8 years later. The total elimination
rate of phosphorus in the lake is k = 0.12 year−1 .

We describe the increasing phosphorus input with an exponential function:

jin (t) = jin (0)eβt (4.55)

β is the rate at which the input grows each year, so for our example9
β = 0.1 year−1. The average phosphorus concentration in the lake, C(t),
can be calculated by inserting Eq. (4.55) into Eq. (4.49):
t
 
−kt
C(t) = C(0) e + e−k(t−t ) jin (0) eβt dt (4.56)
0

After performing the integrations, we get for β = −k:

jin (0) −kt jin (0) βt
C(t) = C(0)e−kt − e + e , β = −k (4.57)
k+β k+β
With this equation we can calculate the phosphorus concentration in the
lake. For t = 4 year we get:

6 mg m−3 year−1
C(t = 4) = 50 mg m−3 e−4×0.12 − e−4×0.12
(0.12 + 0.10) year−1
6 mg m−3 year−1
+ e4×0.1
(0.12 + 0.10) year−1
= (30.9 − 16.9 + 40.7) mg m−3
= 54.7 mg m−3

Correspondingly, for t = 8:

C(t = 8) = (19.1 − 10.4 + 60.7) mg m−3 = 69.4 mg m−3

Note the increasing speed at which C(t) is growing. In the first 4 years the
concentration increases by just under 5 %, while the next 4 years bring an
increase of 27 %!

9 To be precise, β would have to be slightly smaller than 0.1 year−1 , since

e0.1 = 1.105, which is slightly larger than 1.1. The “precise” β is calculated
through the relation β = ln 1.1 = 0.0953 year−1 .
 TIME-DEPENDENT COEFFICIENTS 63

Let us now take a closer look at Eq. (4.57). We observe that C(t)
depends on two specific rates. First, the development of the concentration
is driven by the already discussed total elimination rate k, which deter-
mines how quickly a substance introduced to the system is removed from
it. It is therefore a measure for the memory of the system. Second, it is
driven by β, which characterizes the speed with which the external relation
changes (in our example, the phosphorus input). Furthermore, we see that
because of the factor e−kt , the first two terms on the right-hand side of
Eq. (4.57) become smaller for increasing time and for very large times can
be neglected entirely (more precisely, they can be neglected for t k −1 ).
In contrast, the last term grows exponentially (because β > 0). For very
large times we can therefore approximate Eq. (4.57) by:
jin (0) βt jin (t)
C(t) ≈ e = , if t k −1 (4.58)
k+β k+β
jin (t) is the current input (see Eq. 4.55). This result reminds us of the
steady state (Eq. 4.48), except that it contains the additional factor β in
the denominator. The relative size of k and β therefore determines the
extent to which C(t) from Eq. (4.58) and C ∞ (t) from Eq. (4.48) differ from
each other. Let us now look at the two extreme cases:

Case A: Slowly Changing External Input, i.e. β k

In the denominator of Eq. (4.58), β can be neglected compared to k. Our
approximation is therefore:
jin (0) βt jin (t)
C(t) ≈ e = = C ∞ (t), for β k and t k −1 (4.59)
k k
In other words, if the rate of change of the input β is much smaller than the An adiabatic perturbation
total reaction rate of the system k, the system is almost in the state that is is so slow that the system
in equilibrium with the current input. In physics, an external change that can constantly adjust to
leaves the system approximately in its steady state is called an adiabatic the changing steady state.
perturbation.

Case B: Rapidly Changing External Input, i.e. β not k

In this situation, which we call the non-adiabiatic case, β cannot be ne-
glected in the denominator and we therefore get for t k −1 :
jin (t)
C(t) = , if t k −1 (4.60)
k+β
For an exponentially growing perturbation, i.e. for β > 0, the current sys-
tem state C(t) is therefore smaller than the steady state belonging to the
current input:
jin (t) jin (t)
C(t) = < = C ∞ (t) , if β > 0 (4.61)
k+β k
As long as the input keeps growing at the same rate, the system is unable
to reach the steady state. Instead, it keeps lagging behind a hypothetical
steady state.
64 LINEAR MODELS WITH ONE VARIABLE

After all this math, we want to go back to Example 4.9 and see what
happens when the input growth is stopped:

Example 4.10: Eutrophication of the lake is stopped

In the lake of the previous example, phosphorus input is stabilized af-
ter 20 years of increasing eutrophication. The environmental agency
in charge is optimistic but keeps monitoring the phosphorus concen-
tration’s further development. Surprisingly, it keeps rising for several
years. How can this be explained?

We need to examine the data from the lake more closely: The phospho-
rus load in the lake grew approximately exponentially with a rate of β =
0.1 year−1 for 20 years. The total specific elimination rate of phosphorus
in the lake is k = 0.12 year−1 . Obviously, the external perturbation of the
lake corresponds to a non-adiabatic situation.
Twenty years is long enough to apply the approximative solution
Eq. (4.58):
jin (t = 20 year)
C(t = 20 year) ≈
k+β
Because the two rates β and k have the same order of magnitude, after
20 years the lake is not in the steady state corresponding to its current
input. The hypothetical steady-state concentration for the lake would be:

jin (t = 20 year)
C ∞ (t = 20 year) =
k
The phosphorus concentration in the lake is smaller than the hypothetical
steady state concentration by:

C(t = 20 year) jin (t = 20 year) k k

= · =
C ∞ (t = 20 year) k+β jin (t = 20 year) k+β
0.12
= = 54 %
0.22
If the phosphorus input is stabilized after 20 years, the phosphorus concen-
tration in the lake keeps rising nevertheless—until the steady state concen-
tration C ∞ (t = 20a) corresponding to the input jin (t = 20a) is reached. In
our lake, the phosphorus concentration will still climb to 1.8 times its size.
To find out how long this increase will still take, we can calculate the 5 %
adjustment time:
3 3
τ5 % = = year = 25 year
k 0.12
The phosphorus concentration in the lake will keep rising for another
25 years, even though the phosphorus input is stabilized. Figure 4.6 shows
the development of the lake’s average phosphorus concentration.
 TIME-DEPENDENT COEFFICIENTS 65

400 Fig. 4.6: In a lake, the

phosphorus input grows ex-
ponentially for 20 years.
Then, the input is sta-
300 bilized. The black curve
is the calculated average
phosphorus concentration
]
3

in the lake, the grey curve

C(t) [mg m

200
is the steady state concen-
tration corresponding to
the current input. Despite
the stabilized input, the
average phosphorus con-
100 centration keeps rising for
years

Input is stabilized
0
0 10 20 30 40 50
Time [yr]

4.3.2 Exponentially Falling Input

The input into a natural system can also decrease, for instance, it could fall
exponentially. In that case, the rate β in Eq. (4.55) is negative. For a lake
this could mean that the phosphorus input decreases exponentially thanks
to environmental remediation efforts. The phosphorus concentration in the
lake can then be described analogously to Eq. (4.57) by replacing β with
β  = −β > 0:
jin (0) −kt jin (0) −β  t
C(t) = C(0) e−kt − e + e , k = β  (4.62)
k − β k − β
For k = β  the solution is composed of three terms. Because of the factors

e−kt and e−β t all of them become smaller and smaller with time until they
become virtually zero. The relative size of β  and k determines which term
survives longer when t → ∞. Once again we look at the two extreme cases
for large times t:

Case A: Slowly Falling External Input, i.e. β  k


In the denominator of Eq. (4.62), β can be neglected compared with k.
The third term becomes speed-determining, i.e. it tends towards zero the
slowest. We therefore get:
jin (0) −β  t jin (t)
C(t) ≈ e = = C ∞ (t), for β  k and t k −1
k k
(4.63)
Thus, it’s an adiabatic perturbation. The system remains approximatively
in equilibrium with the exponentially falling input. Let us look at an
example:
66 LINEAR MODELS WITH ONE VARIABLE

Example 4.11: Gradual lake remediation

For many years, a lake exhibits a constant average phosphorus concen-
tration of C 0 = 350 mg m−3 . Since this concentration is too high, the
phosphorus input, currently at jin (0) = 42 mg m−3 year−1 , is to be re-
duced through targeted remediation actions. The total elimination rate
of phosphorus in the lake is k = 0.12 year−1. Because the remediation
scheme can only gradually be put into place, the phosphorus input
is only reduced by 2 % per year. How does the average phosphorus
concentration C(t) in the lake react after remediation begins?

We use Eq. (4.62) and insert β  = 0.02 year−1 . Because β  is much smaller
than k, the third term of the equation is speed-determining. Figure 4.7
shows the calculated development of the average phosphorus concentration
C(t) and the influence of the three terms of Eq. (4.62). Terms 1 and 2
practically neutralize themselves and both tend rapidly towards zero, thus
C(t) is determined by the third term.

Fig. 4.7: Development of Average phosphorus concentration

average phosphorus con-
centration in a lake with Term 3
400
slow reduction of phospho-
rus input. The system’s
C(t)
behavior is determined by
the third term of Eq. (4.62) 200
3]

Term 1
C(t) [mg m

-200

Term 2

-400

0 10 20 30 40 50
Time t [yr]

Case B: Rapidly Falling External Input, i.e. β  k

In this case, k can be neglected in the denominator of Eq. (4.62), so the
first and second terms are speed-determining. For t β −1 we get:

jin (0) −kt

C(t) ≈ C 0 e−kt + e (4.64)
β
 TIME-DEPENDENT COEFFICIENTS 67

As an example, we look at the same lake again. This time, however, the
remediation measures are implemented rapidly:

Example 4.12: Rapid lake remediation

Due to drastic remediation measures the phosphorus input falls ex-
ponentially with the rate β  = 1 year−1. How does the phosphorus
concentration in the lake change?

Figure 4.8 shows the development calculated from Eq. (4.62) and the in-
fluence of the individual terms. As expected, term 3 rapidly tends to zero No
point in
and can be neglected. Term 2 declines more slowly, but because of the fac- hurrying...
tor jin (0)/β  it is much smaller than term 1 and can also be neglected.
Thus, the average phosphorus concentration in the lake is determined by
term 1. In contrast to Example 4.11, the remediation’s success is not de-
termined by the speed of input reduction but by the speed of the lake’s
response to the decreasing contamination. One could argue that an input
reduction rate β  much larger than the system’s innate elimination rate
k does not accelerate the recovery of the lake and is thus economically
inefficient.

Average phosphorus concentration Fig. 4.8: Development

400
of the average phospho-
rus concentration in a
lake with rapid phospho-
300 rus input reduction. The
C(t) system’s behavior is deter-
mined by the first term of
Eq. (4.62)
3]

200
C(t) [mg m

100
Term 1
Term 2
0
Term 3

-100
0 10 0 30 40 50
Time t [yr]
68 LINEAR MODELS WITH ONE VARIABLE

For the sake of completeness, we will briefly look at the special case
where k = β  ≡ −β. In this case, the integration of Eq. (4.56) does not lead
to Eq. (4.57), but is (with β  = −k):

t
 
−kt
C(t) = C(0) e + e−k(t−t ) jin (0) e−kt dt
0
t
(4.65)

= C(0) e−kt + e−kt jin (0)dt

0

In fact, the integrand no longer depends on the integration variable t and

Eq. (4.65) becomes:

C(t) = C(0) e−kt + t e−kt jin (0) (4.66)

In this case, the characteristic rates of the system and the external relation
are identical.

4.3.3 The Periodic Perturbation of a System

Oh, periodic
disturbance!

Many natural systems are subject to periodic influences, such as the

night/day cycle, the seasons, or the tides. The prototypical periodic fluc-
tuation is the sine function. By superimposing several sine functions with
different periods, time shifts and amplitudes, we can construct arbitrary
time-varying curves. Thus, if we analyze the behavior of a linear system
under the influence of a prototype function as depicted in Fig. 4.9, we have
a tool that can also be used to analyze other input functions. We use the
function:
jin (t) = j0 + j1 sin ωt (4.67)

j0 e.g. [ ML−3 T−1 ] Average Input

j1 e.g. [ ML−3 T−1 ] Amplitude of the input fluctuation
ω [ T−1 ] Angular frequency of the input fluctuation
ω = 2π
T ; T [ T ] Period of the input fluctuation
 TIME-DEPENDENT COEFFICIENTS 69

Fig. 4.9: Periodically fluc-

jin(t) tuating perturbation with
j1 average input j0 , amplitude
j1 and period T
j10

Time

Again, let’s look at the linear flow reactor as an example. The average
concentration in the reactor under the influence of a periodically fluctuating
input is described by the following differential equation:
dC
= jin (t) − kC = j0 + j1 sin ωt − kC (4.68)
dt
The solution is calculated with Eq. (4.49) by inserting Eq. (4.67) for jin (t)
and splitting the integral into two parts:
t t
 
C(t) = C 0 e−kt + j0 e−k(t−t ) dt + j1 sin ωt e−k(t−t ) dt (4.69)
0 0

The first integral yields:

j0
(1 − e−kt ) (4.70)
k
For the second integral we best consult a mathematical reference book
where we will find the following solution:
j1 j1 ω −kt ω
√ sin (ωt − η) + 2 e with η = arctan (4.71)
k2 + ω2 k + ω2 k
Thus, the overall concentration in the flow reactor with periodically fluc-
tuating input has the following form:
j0 j0 j1 j1 ω
C(t) = + (C0 − )e−kt + √ sin (ωt − η) + 2 2
e−kt
k k
  ω 2 + k 2 k + ω
   
1 2 3 4
(4.72)
It consists of four components: Terms 1 and 2 represent the dynamic behav-
ior of a linear system with constant input j0 (see Eq. 4.15). For large times,
i.e. for kt 1, only term 1 remains. Terms 3 and 4 describe the influence of
the fluctuating component of the input (j1 sin ωt) in an analogous way. For
kt 1, the last term disappears. This means that for kt 1, Eq. (4.72) is
reduced to the following expression:
j0 j1
C(t) ≈ +√ sin (ωt − η), for kt 1 (4.73)
k k + ω2
2
70 LINEAR MODELS WITH ONE VARIABLE

Let’s compare this expression with the variable steady state C ∞ (t) corre-
sponding to the momentary input j(t):

j(t) j0 j1
C ∞ (t) = = + sin ωt (4.74)
k k k
By comparing these two equations we notice the following:

1. Compared to the amplitude of Eq. (4.74), the amplitude

√ of the peri-
odic part of Eq. (4.73) is smaller by a factor of k/ k 2 + ω 2 .
2. The variation of Eq. (4.73) lags behind Eq. (4.74) by the phase η.

Let’s now look at the following example:

Example 4.13: Periodically fluctuating substance input with

an annual rhythm
In a small mountain lake the nutrient input fluctuates annually due to
the early summer snow melt. The angular frequency ω for the periodic
substance input therefore amounts to:
2π 2π
ω= = = 0.0172 d−1
T 365 d
We use the following equation to describe the variation of substance
input into the lake:

jin (t) = 6 × 10−3 µmol m−3 d−1 (1 + sin ωt)

Thus, j0 = j1 = 0.006 µmol m−3 d−1 . The total specific elimination

rate of the substance in the lake shall be

k = 0.01 d−1

Figure 4.10 shows the substance input jin (upper diagram) and the
average substance concentration C(t) (lower diagram, black curve).
The grey curve describes the adiabatic approximation, that is, the
concentration that would be in equilibrium with the current input.
Because k and ω are of similar magnitudes, the adiabatic approxima-
tion does not accurately represent the real concentration in the lake.
In fact, the amplitude is smaller by a factor of 0.5 and delayed by
η = arctan( 0.0172
0.01 ) = 1.04. To convert the delay into a time, we have
to compare η with the full period of the sine function (2π = 6.28) and
multiply it with the period of variation T = 365 d:
η 1.04
ΔT = T = 365 d = 60 d
2π 6.28
 TIME-DEPENDENT COEFFICIENTS 71

0.012

Substance input jin (t ) [mmol m−3 d−1] 0.01

0.008

0.006

0.004

0.002

jin (t )
0

T
1
Concentration C (t ) [mmol m−3]

0.8

0.6

0.4

0.2

C (t )
0
0 100 200 300 400 500 600 700
Time t [d]

Fig. 4.10: Development of the average concentration C(t) of a substance in a

lake whose substance input jin periodically fluctuates with a period of 1 year (black
curve in upper panel). The total elimination rate is k = 0.01 d−1 , the angular
frequency of the annual period is ω = 0.0172 d−1 . Lower panel, black curve: exact
solution; grey curve: adiabatic approximation. T is the delay of C(t) compared
to the input variation
72 LINEAR MODELS WITH ONE VARIABLE

Example 4.13 represents a situation where the two characteristic rate

constants, ω and k, are of similar magnitude. Remember that ω describes
the pace of the external (periodic) changes while k describes the speed of
the system’s response to these changes. Once again, we now want to look
at the two extreme situations, i.e. when the external perturbation is either
much slower or much faster than the system’s response.

Case A: Slowly Fluctuating Perturbation ω k

The variation of the perturbation proceeds slowly relative to the average
retention time k −1 of a substance in the system. The denominator of the
second term in Eq. (4.73) can be approximated by (k 2 + ω 2 )1/2 ≈ (k 2 )1/2 =
k. The phase difference η = arctan ωk becomes very small and can there-
fore be neglected in the sine function. Thus, from Eqs. (4.73) and (4.74) we
get:

j0 j1 sin (ωt)
C(t) ≈ + = C ∞ (t), for kt 1 and ω k (4.75)
k k

In this case, the average concentration in the reactor follows the steady
state that corresponds to the current input j(t) = j0 + j1 sin ωt. The slow
variation of the external relation thus represents an adiabatic perturba-
tion.

Case B: Rapidly Fluctuating Perturbance ω k

In this case, the variation of the external relation is swift compared to the
reaction time of the system. Because ω k, (k 2 + ω 2 )1/2 in Eq. (4.73)
can be approximated by ω. The phase difference η = arctan ωk tends to
2 for k → ∞. Thus, we get the following approximate solution for kt
π ω

1:

j0 j1 π
C(t) ≈ + sin (ωt − ), for kt 1 and ω k (4.76)
k ω 2

A linear system filters ex- The current system state C(t) lags behind the current perturbation jin
ternal fluctuations that are by a quarter period ( π2 ). The amplitude of the system fluctuation ( jω1 ) is
considerably faster than the reduced by the factor ωk 1 relative to the amplitude of the steady state
system’s reaction time. j1
( k , see Eq. 4.74). The system thus remains close to the average steady
state jk0 . In effect, a linear system filters external fluctuations out of the
system when they are substantially faster then the system’s reaction time
k −1 .
Let us demonstrate these extreme situations with the temperature
regime in a small lake.
 TIME-DEPENDENT COEFFICIENTS 73

Example 4.14: Water temperature in a small lake

The average water temperature Θ in a small, vertically well-mixed lake
can be approximated by the following linear differential equation:
dΘ
= kΘ (Θeq − Θ) (4.77)
dt
The specific thermal exchange rate kΘ (dimension T−1 ) is inversely
related to the lake’s mean depth. It also depends on meteorological pa-
rameters such as solar radiation, wind speed and humidity. For medium
latitudes and a lake mean depth of 5 m a typical value is kΘ = 0.1 d−1 .
Θeq is the equilibrium water temperature, that is, the water tempera-
ture which the lake would reach if all meteorological parameters were
to remain constant.a Θeq strongly varies with time of day, season and
with (irregular) weather events. By looking at Θeq statistically, we can
depict two dominant modes of variation, the annual and the diurnal,
and approximate each by a sinusoidal function:
2π
Θeq (t) = Θeq + Θamp sin t (4.78)
T
a) Annual temperature variation
2π
T = 365 d, thus ω = = 0.017 d−1 kΘ = 0.1 d−1
365 d

This case exhibits slow variation with η  0. From Eq. (4.75)

with j0 = kΘ Θeq , j1 = kΘ Θamp we get

2π
Θ(t)  Θeq + Θamp sin t = Θeq (t) (4.79)
T

The exact solution is plotted in the upper panel of Fig. 4.11.

b) Diurnal temperature variation

2π
T = 1 d, thus ω = = 6.28 d−1 kΘ = 0.1 d−1
1d

This case shows rapid variation with η  π

2, which corresponds
to a delay time of ΔT = 365
4 d = 91 d.
From Eq. (4.76) with j0 and j1 as above we get:

kΘ 2π π
Θ(t)  Θeq + ( )Θamp sin ( t − ) (4.80)
ω T 2
a Note that Θ
eq is not equal to air temperature since the direct exchange
of sensible heat between water and air is just one (minor) process of heat
exchange.
74 LINEAR MODELS WITH ONE VARIABLE

Since kωΘ = 0.1/6.28 = 0.016, the second term of Eq. (4.80) is fairly
small whatever the temperature fluctuation Θamp may be. Thus, water
temperature is virtually unaffected by the diurnal variability of the
relevant meteorological parameters (including the extreme day/night-
variation of short-wave solar radiation). The exact solution is plotted
in the lower panel of Fig. 4.11.

(t ) [ C]
and
eq (t )

Time t [d]
(t) [ C]
and
eq (t )

Time t [d]

Fig. 4.11: Average water temperature in a lake with annual variation (above)
and diurnal variation (below). kΘ = 0.1d−1 , Θeq = 10 ◦ C, Θamp = 10 ◦ C. The
black plot is the actual temperature Θ(t); the grey plot is the theoretical equilibrium
temperature Θeq (t). In the annual case, the actual temperature slightly lags behind
the theoretical equilibrium temperature, while in the diurnal case the fluctuation
is so rapid that the actual temperature cannot keep up and hardly deviates from
the average
 QUESTIONS AND PROBLEMS 75

4.4 Questions and Problems

Question 4.1: What characteristic properties do the solutions of linear
differential equations have?

Question 4.2: What is the difference between homogeneous and inhomo-

geneous linear differential equations?

Question 4.3: Write down the most general form of a linear, first-order
inhomogeneous differential equation.

Question 4.4: What requirement must a first-order differential equation

fulfill to possess a finite steady state?

Question 4.5: Which dimensions do the variables in Eqs. (4.17) and (4.22)
have?

Question 4.6: What is the adjustment time of a linear model? Why is this
time not unambiguously defined, but needs the specification of a conven-
tion? Name such a convention.

Question 4.7: Try to derive Eq. (4.57) by integration of Eq. (4.56).

Question 4.8: You measure the increase of molecular oxygen concentra-

tion in a constantly mixed, open-top glass beaker. The water within has
been completely de-gassed beforehand. How can you decide whether the
data can be described with a linear model?

Question 4.9: What is meant physically by calling the perturbation of a

linear model adiabatic? Give an example.

Question 4.10: The inhomogeneous term of a linear model fluctuates with

a period of 1 h around an average value. The system variable remains prac-
tically constant. What does this indicate about the system?

Problem 4.1: Adjustment behavior

What does Fig. 4.3 (ingrowth curve, decay curve, C(t)) look like for a lake
in which, starting from the steady state, the substance input is suddenly
cut in half?

Problem 4.2: Radioactive decay

The decay of a radioactive isotope is a first-order reaction often quantified
by half-life.
(a) Formulate a model for the radioactive isotope radon-222 (222 Rn), a
noble gas with a half-life of 3.8 days. Radon-222 is produced by the de-
cay of the radioactive element radium-226 (226 Ra, half-life 1,600 years)
76 LINEAR MODELS WITH ONE VARIABLE

through an α decay. The initial concentrations of 222 Rn and 226 Ra

are NRa = 1×108 atoms/liter and NRn = 5×104 atoms/liter, respec-
tively. You can neglect the reduction of 226 Ra concentration during
the observation period of several weeks. Why is that justifiable?
(b) Often, radioactive isotopes are not characterized by concentration,
but by their activity A. A is the number of decays that are registered
per time and volume. How do N and A depend on each other? Convert
the model developed for N into one for A.

Problem 4.3: Cars in a parking lot

In a downtown car park, 300 cars enter per hour. Fifteen percent of them
immediately leave the car park again, having found no free spot. The in-
dividual length of stay of parked cars varies. An observer notes that per
minute, 1 % of the parked cars leave the car park.
(a) How many parking spots does the car park have?
(b) How long does a car stay in the car park on average (not counting
the 15 % unsuccessful visitors)?
(c) We assume that the car park is opened at 6 a.m. and is completely
empty at that time. The parking behavior remains exactly as de-
scribed above: 300 cars enter per hour (of course, initially all of them
will find a spot), 1 % of parked cars leave every minute. When is the
car park full?

Problem 4.4: Phosphorus in a lake with sedimentation

A lake is loaded with 10 tons of phosphorus annually. On the one hand,
phosphorus is continuously added to the sediment at a rate which is propor-
tional to the concentration in the lake water. On the other hand, phosphorus
is washed out of the lake through the outflow. Calculate the steady-state
phosphorus concentration in the lake, if the annual phosphorus input is
reduced to 6 tons a year. How long does it take for the concentration to be
within 5 % of the new steady state?

The following is known about the lake:

Volume V = 0.2 km3
Through-flow of water Q = 0.1 km3 year−1
Sedimentation S = ks M [ mg year−1 ]
Sedimentation rate ks = 0.75 year−1
Average total phosphorus concentration
prior to the input reduction C 0 = 40 mg P m−3

Problem 4.5: Exponential population increase

In a city, new residents continuously arrive from the surrounding rural
area. From 1995 to 2010, the total immigration can be approximated by
 QUESTIONS AND PROBLEMS 77

an exponential function (growth rate β = 6.25 % per year). The birth rate
(b) in the city is 2.3 % and the death rate (d) is 0.5 % of the population
per year. Effects that are driven by the population’s age structure are not
taken into account. A census showed that in 1995, 10.9 million people lived
in the city, and 0.25 million migrated into it in the same year. How big
is the city’s population in the year 2020, if growth continues at the same
rate?

Problem 4.6: Lake water temperature

In Example 4.14 we introduced a simple linear model to describe the water
temperature Θ in a small, vertically well mixed lake. We neglected the
possible influence from the inlets and outflow.

(a) Modify Eq. (4.77) to account for the effect of the through-flow of water
(inlets, outlet). Assume that the temperature of the inflowing water
Θin is at equilibrium with the momentary meteorological conditions
(Θin (t) = Θeq (t)) while the temperature of the outflow is equal to
the mean lake temperature (Θout (t) = Θ(t)). The water through-flow
is expressed by the specific flow rate kw (Eq. 4.22).

(b) Apply the modified model to Lake A with kΘ = 0.03 d−1 , kw =

0.02 d−1 (corresponding to a mean water residence time of 50 d).
Calculate the lag phase ηA and the lag time ΔTA of the annual water
temperature variation relative to the external forcing Θeq (t). Note:
If we approximate the year by 360 days and express ηA in angular
degrees, then a 1◦ lag phase corresponds to a lag time of 1 day.

(c) Compare the result obtained in (b) with Lake B, which has the same
thermal properties but no through-flow. Calculate ηB and ΔTB .

(d) Assume that as in Eqs. (4.78) and (4.79), the annual temperature
variation in the two lakes can be described by a sinusoidal curve.
Calculate the ratio of the temperature amplitudes of the two lakes,
Θamp,A /Θamp,B . Explain the result qualitatively.

Problem 4.7: Dye in a well

A joker dumps 1 gram of a fluerescent dye (Fluorescein) into a water foun-
tain. The dye decays photochemically by a first-order reaction with the
reaction rate kr = 0.1 h−1 . The fountain has a volume of V = 2 m3 , and
water continuously flows through it with a rate of Q = 2 L/min.

(a) Draw a mass balance by using a box diagram and establish the dy-
namic equation for the dye concentration in the fountain.

(b) The dye remains visible in water down to a concentration of Ccrit =

10−5 g/L. How long does it take for the water to appear uncolored
again?
78 LINEAR MODELS WITH ONE VARIABLE

(c) Assume that the fountain is operated in a closed loop, i.e., the in-
flowing water is entirely made up of outflowing water. What is the
corresponding dynamic equation? Again, calculate the time tcrit for
this flow regime until the water appears uncolored.

Problem 4.8: Degradation process in a sewage plant

In the inlet of a sewage plant, the concentration Cin of a substance A is
measured. If the inlet concentration Cin is held constant over a long enough
time, one finds the concentration Cout in the plant’s outlet. The table shows
Cout for four different inlet concentrations of substance A:

Cin (mg/L) 20 50 80 130

Cout (mg/L) 1.0 2.5 4.0 6.5

Consider the sewage plant to be a completely mixed reactor with constant

flow Q. The average water retention time in the sewage plant is τw = 2 d.

(a) Draw a box diagram for the mass balance and establish a dynamic
equation for the concentration of substance A in the sewage plant.

(b) Give a mathematical relation for the degradation rate R [mg L−1
d−1 ] of substance A as a function of the concentration in the plant C.

(c) Calculate the specific degradation rate kr [d−1 ] of substance A in the

plant, under the assumption that the process is linear.

(d) Starting from the steady state Cin = 80 mg/L and Cout = 4.0 mg/L,
how long does it take for the concentration C in the plant to fall to
5 % of its initial value, if at time t = 0, Cin suddenly becomes zero?

(e) What is the answer to question (d) if you start from the steady state
with Cin = 20 mg/L and Caus = 1.0 mg/L? Explain your answer.

Problem 4.9: Copper accumulation on farm land

By way of atmospheric deposition and the use of fertilizers, 42 mg of copper
per square meter and year are added to agricultural land. Through washing-
out processes and the crop harvest, the soil loses 0.6 % of its copper content
per year.

(a) Establish the differential equation for the change in copper concen-
tration C in the soil.

(b) How large is the copper concentration in the soil at steady state?

(c) How long does it take until the concentration comes to within 5 % of
the steady state calculated in (b), if we assume an initial concentra-
tion of 0 mg copper per square meter?
 QUESTIONS AND PROBLEMS 79

(d) In fact, the average copper concentration measured in the soil is

Cmeasured = 6, 000 mg m−2 . How large would the total rate ktot of
a linear process that removes copper from the soil be, if Cmeasured
represents a steady state and the annual input is 42 mg m−2 year−1?
(e) Does it make sense to interpret the soil’s copper content with a steady-
state approach, if we know that the copper input grows by 1 % an-
nually and the copper concentration can only be measured with an
accuracy of 10 %? Justify your decision with a brief calculation.

Problem 4.10: Tritium in a lake

A lake with the constant volume V = 3 km3 has constant in- and out-
flows with Q = 300×106 m3 year−1 . For several decades, radioactive tritium
(decay constant kλ = 0.058 year−1 ) has been added to the inflow with a
constant concentration of Cin = 1 Bq/L. The average measured tritium
concentration in the lake is C = 0.44 Bq/L.
(a) Write down the mass balance for tritium in the lake, assuming that
the lake can be described as a completely mixed system with the three
processes inflow, outflow and radioactive decay.
(b) Using the tritium concentration measured in the lake, judge whether
there is another process removing tritium from the lake besides ra-
dioactive decay. If there is, calculate the reaction rate of this process
as a first-order constant kr .
(c) What happens if starting at time t0 the tritium input grows annually
by 2 %? How large is the tritium concentration in the lake at time
t0 + 10 years? (an approximation is sufficient)
(d) Answer question (c) for the case of an input growing by 20 % annually.
Chapter 5

Linear Models with Several Variables

So far we have only dealt with models described by a single system variable.
We called these models one-dimensional. To adequately describe a system,
however, we often need several variables that interact with each other. This
We dogs demand
is where multi-dimensional models come into play. NOT
There are two ways to move from a one-dimensional to a multi-dimen- to be thrown
sional model. First, it could be useful to divide the system variable V into into the same
box as
subcomponents. For instance, we want to split the total phosphorus con- humans!
centration in a lake into dissolved and particle-bound phosphorus. Or we
are interested in two substances A and B that can be converted into each
other by a chemical transformation. In both cases we need to add material
differentiation to our model.
Second, spatial differentiation could be necessary in some scenarios, for
instance, dividing the overall average phosphorus concentration in a lake
into surface and deep water concentration. In this case, we call our model
a two-box model, because it contains two spatially separate “boxes”. Of
course, this expression could also be used for the materially differentiated
model: the boxes would be material compartments rather than spatial ones.
Mathematically speaking it makes no difference whether we differentiate
spatially or materially. Both cases lead us to a system of two or more
coupled differential equations.

5.1 Linear Models with Two System Variables

5.1.1 The System of Linear Differential Equations and Its Eigenvalues

In this section, we want to limit our discussion to two-dimensional first-
order models with constant coefficients.1 Such models can be described by

1 Recall that in Chap. 4, we called models with constant coefficients au-

tonomous models.

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 81

DOI 10.1007/978-3-642-30639-6_5, © Springer-Verlag Berlin Heidelberg 2013
82 LINEAR MULTI-DIMENSIONAL MODELS

the following set of differential equations:

dV1
= R1 + p1,1 V1 + p1,2 V2
dt (5.1)
dV2
= R2 + p2,1 V1 + p2,2 V2
dt
The variables have the following meanings:
Vi i-th system variable
Ri inhomogeneous term (input) of the i-th system variable, constant
pi,j model parameter, constant

Eq. (5.1) can also be written as a matrix:

      
dV1 /dt R1 p1,1 p1,2 V1
= + (5.2)
dV2 /dt R2 p2,1 p2,2 V2
or
dV
=R+P·V (5.3)
dt
The system then consists of an input vector2 R, the coefficient matrix P
and the vector of the system variable V.

Example 5.1: Chemical reaction in a reactor (see Fig. 5.1)

We consider a fully mixed reactor with volume V and constant flow
rate Q. In the inflow, the substance A is continuously added with the
concentration Cin (JA = QCin ). Inside the reactor, a linear reaction
between substances A and B takes place. For example, substance A
could be hydrolyzed to substance B ≡ AH+ and back. The specific rates
of the forward and return reaction are kA and kB . Both substances
leave the reactor through the outflow. What are the dynamic equations
for CA and CB in the reactor?

As in Example 4.4 we begin by establishing the mass balance. This time

we need to do this for both substances, A and B, keeping in mind that MA
and MB are the masses of the substances in the reactor3:
dMA
= QCin − kA MA + kB MB − QCA
dt
dMB (5.4)
= kA MA − kB MB − QCB
dt

2 Variables printed in bold indicate vectors or matrices.

3 Strictly speaking, models that involve chemical conversion processes are atom
balance models. The concentrations should therefore be expressed in molar
units.
 TWO-DIMENSIONAL MODELS 83

Input Substance A Fig. 5.1: Linear reaction

JA = QCin in a completely mixed reac-
tor with through-flow

Outflow Substances A,B

QCA
QCB

If we divide both equations by the constant reactor volume V , we get

analogously to Eq. (4.22):
dCA
= kw Cin − kA CA + kB CB − kw CA
dt (5.5)
dCB
= kA CA − kB CB − kw CB
dt
where kw = Q/V , CA = MA /V and CB = MB /V .
We need to rewrite Eq. (5.5) so that we can compare them to Eq. (5.1)
and identify the parameters introduced there:
dCA
= kw Cin − (kA + kw )CA + kB CB
dt (5.6)
dCB
= kA CA − (kB + kw )CB
dt
Thus:
R1 → kw Cin , R2 → 0
p1,1 → −(kA + kw ), p1,2 → kB
p2,1 → kA , p2,2 → −(kB + kw )

As we will see, the coefficient matrix P plays a central role in the solution of
a system of linear differential equations (even more so in systems with more
than 2 dimensions). Therefore, we want to introduce four characteristic
quantities of the matrix P (see Appendix D):
1. The determinant of P:
det(P) = p1,1 p2,2 − p1,2 p2,1 (5.7)

2. The trace of P:
tr(P) = p1,1 + p2,2 (5.8)
84 LINEAR MULTI-DIMENSIONAL MODELS

3. The discriminant of P:

Δ(P) = tr(P)2 − 4 det(P)

= (p1,1 + p2,2 )2 − 4p1,1 p2,2 + 4p1,2 p2,1 (5.9)
= (p1,1 − p2,2 ) + 4p1,2 p2,1
2

4. The two eigenvalues4 of P:

1  
λi = tr(P) ± Δ(P) , i = 1, 2 (5.10)
2

The eigenvalues are the solutions of the so-called characteristic equation

λ2 − λ(p1,1 + p2,2 ) + p1,1 p2,2 − p1,2 p2,1 =

(5.11)
λ2 − λtr(P) + det(P) = 0

If the coefficient matrix P has two different eigenvalues λi (i = 1, 2), that

is, if Δ(P) = 0, the two-dimensional system of differential equations (5.1)
with constant coefficients Ri and pi,j has solutions of the form:

V1 (t) = a1,0 + a1,1 eλ1 t + a1,2 eλ2 t

(5.12)
V2 (t) = a2,0 + a2,1 eλ1 t + a2,2 eλ2 t

The two eigenvalues λi (i = 1, 2) have the dimension of a specific rate,

[ T ]−1 . The six coefficients ai,j (i = 1, 2; j = 0, 1, 2) depend on the coef-
ficients pi,j and the initial states V01 and V02 . In Appendix D, we discuss
in more detail how to get from the differential equation (5.1) to the solu-
tion (5.12).
If both eigenvalues are real and negative, all exponential functions in
Eq. (5.12) tend to zero for t → ∞. Table 5.1 shows what properties the
matrix P must have for this condition to be fulfilled. Since the exponential
functions tend to zero for t → ∞, the coefficients a1,0 and a2,0 are identical
with the system’s steady states V∞ ∞
1 and V2 . We find the steady states
by setting the left sides of Eq. (5.1) to zero, and solving the resulting two
coupled equations for the two unknowns V1 and V2 . We can then call them
V∞1 and V2 :
∞

p1,2 R2 − p2,2 R1 p1,2 R2 − p2,2 R1

a1,0 = V∞
1 = =
p1,1 p2,2 − p1,2 p2,1 det (P)
(5.13)
p2,1 R1 − p1,1 R2 p2,1 R1 − p1,1 R2
a2,0 = V∞ = =
2
p1,1 p2,2 − p1,2 p2,1 det (P)

4 “Eigen” is a German word that means “own”, used in the sense of the char-
acteristic (own) values belonging to a particular matrix.
 TWO-DIMENSIONAL MODELS 85

The four other constants ai,j in Eq. (5.12) are better legible if we don’t
express them by R1 and R2 , but rather by the constants a1,0 and a2,0 of
Eq. (5.13) (for λ1 , λ2 real and negative and λ1 = λ2 ):
 
a1,1 = q· (p1,1 − λ2 )(V01 − a1,0 ) + p1,2 (V02 − a2,0 )
 
a1,2 = −q· (p1,1 − λ1 )(V01 − a1,0 ) + p1,2 (V02 − a2,0 )
 
a2,1 = q· p2,1 (V01 − a1,0 ) + (p2,2 − λ2 )(V02 − a2,0 ) (5.14)
 
a2,2 = −q· p2,1 (V1 − a1,0 ) + (p2,2 − λ1 )(V2 − a2,0 )
0 0

 − 1
q = 1
λ1 −λ2 = √ 1
= (p1,1 − p2,2 )2 + 4 p1,2 p2,1 2
Δ(P)

Only in rare cases would we calculate the solution of Eq. (5.1) manually.
It makes more sense to use computer software for that. However, we can
get a qualitative view of the system’s behavior from the eigenvalues of the
coefficient matrix P. They are, as Eq. (5.10) shows, relatively easy to find,
at least for a two-dimensional system.

Example 5.2: Chemical reaction in a reactor (continued)

In Example 5.1 we analyzed the model of a chemical reactor. We
now want to numerically look at its steady state. Assuming we
are interested in substance B, we might want at least 90 % of the
introduced substance to be in the form B in the outflow. How big can
the flow rate Q be at most for this condition to be fulfilled?

We use the following data:

Reactor volume V = 10 m3
Reaction constants kA = 0.5 h−1
kB = 0.01 h−1

If the reactor is run with constant conditions for long enough, the concen-
trations CA and CB in the reactor (and therefore also in the outflow) take
∞ ∞
on the steady state values CA and CB . The conversion constraint of 90 %
can be written as:
∞
CB 1
0.9 ≤ ∞ ∞ = ∞ ∞
CA + CB CA /CB + 1
∞ ∞
Solved for CA /CB :
∞ ∞
CA /CB ≤ (1/9) = 0.11
86 LINEAR MULTI-DIMENSIONAL MODELS

∞ ∞
Before we use Eq. (5.13) to calculate CA and CB , we have to ensure
that the eigenvalues are real and negative and that the determinant is
det (P) = 0. According to Example 5.1, we have:
det (P) = (kA + kw )(kB + kw ) − kA kB
2
= kw + kw (kA + kB ) > 0
This is because all coefficients kw , kA , kB are larger than zero. Furthermore,
you can verify that all eigenvalues correspond to case 1 in Table 5.1. Thus
∞ ∞
we can now calculate the proportion CA /CB from Eq. (5.13), making use
of the fact that R2 = 0 (see Example 5.1):
∞
CA −p2,2 R1 kB + kw
∞ = = ≤ 0.11
CB p2,1 R1 kA
Solving for kw = Q/V gives:
Q
kw = ≤ 0.11kA − kB
V
= 0.11 × 0.5 h−1 − 0.01 h−1
= 0.045 h−1
or
Q ≤ 10 m3 × 0.045 h−1 = 0.45 m3 h−1
So the flow rate Q must not increase beyond 0.45 m3 h−1 .

5.1.2 Eigenvalues of a Linear System of Differential Equations: Their Meaning

If the two eigenvalues λi (i = 1, 2) are distinct, all solutions of a two-
dimensional system of linear differential equations look formally very simi-
lar (see Eq. 5.12). Vi (t) consists of three terms at most: a constant and two
exponential functions. In special cases some of these terms may be zero. It
doesn’t come as a surprise that the solutions of a three-dimensional system
contain an additional (third) exponential function. Correspondingly, there
is a third eigenvalue. An n-dimensional system has n eigenvalues.
The eigenvalues of a lin- The properties of the eigenvalues λi bring some variety to the solu-
ear system determine its tions in Eq. (5.12). Table 5.1 gives an overview of the possible eigenvalue
temporal behavior. types and conveys a qualitative impression of the corresponding solutions.
Note that the inhomogeneous terms R1 and R2 have no influence on the
eigenvalues.
The cases 4–6 in Table 5.1 require an additional explanation: If the dis-
criminant Δ(P) of the quadratic equation is smaller than zero, the eigen-
values are complex numbers. To interpret the solution Eq. (5.12), we must
know the meaning of an exponential function with a complex argument,
λi = a + i b5 :
eλi t = e(a+ib)t = eat eibt = eat (cos bt + i sin bt) (5.15)

5 See Appendix C.4.

 Table 5.1: Meaning of the eigenvalues λi

Eigenvaluesa λi = 12 [tr ± (Δ)1/2 ] Condition Properties Solution sketch

tr < 0, Δ > 0
(1) Real, both negative Solution with steady state
0 < Δ1/2 < −tr

tr < 0, Δ > 0
(2) Real, λ1 = 0, λ2 < 0 Δ1/2 = −tr Solution with steady state
(corresponds to det = 0)

Solution → ∞
(3) Real, at least one positive tr > 0, Δ > 0
(solution unbounded)b

(4) Both purely imaginary tr = 0, Δ < 0 Undamped oscillation

Conjugate-complex,
(5) tr < 0, Δ < 0 Damped oscillation
with negative real part

Conjugate-complex,
(6) tr > 0, Δ < 0 Oscillation → ±∞
with positive real part

a
To simplify notation, we don’t write the argument (P) in Δ, tr und det. In addition, the case Δ = 0(λ1 = λ2 ) is not
TWO-DIMENSIONAL MODELS

considered in this list. It is only relevant for special cases.

b
If only one λi > 0, the boundary conditions can still lead to finite solutions if in Eq. (5.12) the coefficient ai,j
87

belonging to the positive λi is zero.

88 LINEAR MULTI-DIMENSIONAL MODELS

A complex eigenvalue thus leads to an oscillating function. If the real part

of λi , Re(λi ) = a, is negative, the amplitude grows smaller. If however
Re(λi ) = a is positive, the amplitude grows to infinity. For a purely imag-
inary eigenvalue (a = 0) we get a periodic function with constant ampli-
tude, i.e. an undamped oscillation. We will discuss this case in more detail
in Sect. 5.1.6.
In the following sections we will consider examples for the different types
of eigenvalues. But first we want to introduce an important tool to analyze
multi-dimensional models: the phase space.

5.1.3 The Phase Space

One way to illustrate the dynamics of a model with several variables is the
phase diagram. In this diagram, the momentary state of the system is a
point in the n-dimensional phase space spanned by the n system variables
(see Fig. 5.2). The development of the system is then described by a curve,
also called trajectory. The system moves along this trajectory; if it reaches
a steady state, it stays at that point. Trajectories have two characteristic
properties: they do not cross and they have a direction that indicates the
course of development through time. In Chap. 6 on nonlinear models, the
phase diagram will play an important role in analyzing the properties of a
system.

Fig. 5.2: Phase diagram

for a two-dimensional
model. The trajectory is
the line along which the
model moves. It is deter-
mined by the initial state
V0 and the model equations

5.1.4 Linear Models with Real, Non-positive Eigenvalues

Real negative eigenvalues (case 1 in Table 5.1) frequently appear in linear
models that are based on the mass balance of a system. Figure 5.3 exempli-
fies the idea of a mass balance. First, we assume that the substance is not
produced by a first-order process, but only added from outside of the sys-
tem (inhomogeneous terms J1 and J2 ). Second, the elimination processes
are first order and either lead into the neighboring box (k1 M1 , k2 M2 ) or
out of the system (k3 M1 , k4 M2 ). All ki are positive.
 TWO-DIMENSIONAL MODELS 89

Fig. 5.3: A linear system

in which a substance can-
not be created by a first-
order process, but only
comes from outside the
system or a neighboring
box. This leads to a system
of differential equations
with real and non-positive
eigenvalues. All ki are pos-
itive
dM1
= J1 − (k1 + k3 )M1 + k2 M2
dt
dM2
= J2 + k1 M1 − (k2 + k4 )M2
dt

For the model in Fig. 5.3, the model parameters pi,j from Eq. (5.1) take the
following values:

p1,1 = −(k1 + k3 ); p1,2 = k2

p2,1 = k1 ; p2,2 = −(k2 + k4 )

One can confirm that under these conditions, the coefficients in matrix P
always have the following properties:

(a) The elements outside of the diagonal p1,2 and p2,1 are positive. The
discriminant Δ of the characteristic equation is therefore larger than
zero (Eq. 5.9) and the eigenvalues are real.
(b) The diagonal elements p1,1 and p2,2 are negative. Furthermore,
−p1,1 ≥ p2,1 and −p2,2 ≥ p1,2 . The eigenvalues are therefore neg-
ative (and real). The model tends to the steady state. Its adjustment
time τ is determined by the smaller of the eigenvalues. Analogously
to the one-dimensional system, the 5 % adjustment time is:
3 6
τ5 % ≈ =  (5.16)
min |λi | −tr(P) − Δ(P)

for λi = 0, Δ(P) > 0 and tr(P) < 0. However, Eq. (5.16) cannot be
interpreted with the same stringency as its one-dimensional counter-
part (Eq. 4.42). Individual variables of a multi-dimensional system of
equations can approach their steady state much more quickly if the
coefficients ai,j belonging to the smallest eigenvalue are small or even
zero for the respective variable. Therefore, Eq. (5.16) should always be
used merely as a first, rough estimate. A more detailed examination
would require knowing the explicit solution.
(c) If the system is conservative with respect to its mass balance, we
have a special case: there are neither internal sources nor sinks. This
90 LINEAR MULTI-DIMENSIONAL MODELS

means that the sum of both variables is constant and one of the
two eigenvalues is zero. For the coefficient matrix, −p1,1 = p2,1 and
−p2,2 = p1,2 . The first row of matrix P is identical with the second
row except for the algebraic sign. Such a matrix is called singular.
Its determinant is zero: det(P) = 0. Singular matrices have at least
one eigenvalue that is zero.

Example 5.3: Chemical reaction in a reactor (continued)

In Examples 5.1 and 5.2 we analyzed a reactor with a chemical reaction
between two substances, A and B. We now want to know how long it
takes for the reactor to reach a steady state.

To perform the calculation, we fix the through-flow Q at 0.4 m3 h−1 . Then,

kw = Q/V = 0.4 m3 h−1 /10 m3 = 0.04 h−1 . The system and its coefficient
matrix have the same structure as the model in Fig. 5.3. The coefficient
matrix is:

p1,1 = −(kA + kw ) = −(0.5 + 0.04) h−1 = −0.54 h−1

p1,2 = kB = 0.01 h−1
p2,1 = kA = 0.5 h−1
p2,2 = −(kB + kw ) = −(0.01 + 0.04) h−1 = −0.05 h−1

We therefore get, according to Eqs. (5.7)–(5.10):

det(P) = (0.54 × 0.05) h−2 − (0.01 × 0.5) h−2 = 2.2 × 10−2 h−2
tr(P) = −(0.54 + 0.05) h−1 = −0.59 h−1
Δ(P) = tr(P)2 − 4 det(P) = 0.260 h−2
1
λi = [−0.59 h−1 ± (0.260)1/2 h−1 ]
2
λ1 = −0.040 h−1 ; λ2 = −0.55 h−1

From any given initial state, the overall 5 % adjustment time according to
Eq. (5.16) is:
3 3
τ5 % = = = 75 h
|λ1 | 0.04 h−1
This time is about equal to the average retention time of the water in
the reactor (τw = V /Q = 25 h; τ5 % ∼ 3τw ). It depends only marginally
on the time for reaching an equilibrium between phases A and B. As
shown in Example 5.4, the latter depends on the sum of the two rates
(kA + kB )−1 ≈ 2 h.
In the next example, we look at a system with a singular system ma-
trix, i.e. det(P) = 0. Such systems only have a steady state if they are
homogeneous, that is, if R1 = R2 = 0.
 TWO-DIMENSIONAL MODELS 91

Example 5.4: Chemical reaction of two substances

We examine the forward and backward reaction between two chemical
substances with the masses M1 and M2 . This could be the acid-base
equilibrium of the acid AH and the corresponding base A− :

AH + H2 O  H3 O+ + A−

or shorter:

k1
M1  M2
k2

We assume that the speed of the forward and backward reactions is

proportional to the mass of the corresponding starting substance (i.e.,
it is a first-order reaction). If k1 and k2 are the specific rate constants
of the two reactions, the following mass balance results:
dM1
= −k1 M1 + k2 M2
dt (5.17)
dM2
= k1 M1 − k2 M2
dt

The coefficient matrix for this system is:

 
−k1 k2
P= (5.18)
k1 −k2
The two Eqs. (5.17) are linearly dependent. The matrix is therefore singular
(det(P) = 0). As per Table 5.1, case 2, one of the eigenvalues is then zero:
λ1 = 0. What does this mean for the modeled system?
To answer this question, we must look at the system’s mass balance.
The total mass in the system Mtot = M1 + M2 remains constant, and it is
0
determined by the total mass of the initial state Mtot = M10 + M20 :
dMtot dM1 dM2
= + =0 (5.19)
dt dt dt
In a linear system, an eigenvalue that is zero always indicates a con-
served quantity. In our case, the total mass Mtot is conserved. As follows
from Eq. (5.10), the second eigenvalue has the value (note that −tr(P) =
Δ(P)1/2 ):
λ2 = tr(P) = −(k1 + k2 ) (5.20)
The steady state of the model is calculated by setting dM i
dt = 0 in Eq. (5.17).
Since the equations are linearly dependent they both yield the same result:
k1 ∞
M2∞ = M = KM1∞ (5.21)
k2 1
92 LINEAR MULTI-DIMENSIONAL MODELS

Instead of one specific steady state or fixed point, we get a whole set of
potential steady states which all lie on a straight line, the so-called fixed-
point line. It corresponds to the thermodynamic equilibrium of the reaction
with the equilibrium constant K = kk12 .
We don’t need complex calculations to determine which point on the
fixed-point line the model is actually moving towards. We only need to
remember that the model obeys a mass conservation condition. Figure 5.4
shows the phase diagram of the homogeneous equilibrium reaction (5.17).
On the one hand, mass conservation means that the system can only move
0
along the line with slope (−1), which crosses the M1 or M2 axis at Mtot =
0 0
M1 + M2 . This is the system’s trajectory, labeled A in Fig. 5.4. On the
other hand, Eq. (5.21) says that all steady states must lie on a line through
the origin with slope ( kk12 ). This is the system’s fixed-point line, labeled B in
Fig. 5.4. The steady state in accordance to the initial condition (M10 , M20 )
lies at point S, the intersection between the two lines.

Fig. 5.4: Phase diagram

of the homogeneous chem-
ical equilibrium reaction
(Eq. 5.17). The arrows
on the trajectory (line A)
show the direction in which
the system moves in time.
Further explanations are in
the text

What remains is the confirmation that the model does indeed move towards
the fixed-point line. For this, we solve the system of equations (5.17) by
replacing M2 with Mtot0
− M1 in the first equation. We get the following
for M1 :
dM1 0
= k2 Mtot − (k1 + k2 )M1 (5.22)
dt
Eq. (5.22) is an inhomogeneous linear first-order differential equation with
the variable M1 . We already know its solution from Chap. 4 with Eqs. (4.6)
and (4.33):
M1 (t) = M1∞ + (M10 − M1∞ ) e−(k1 +k2 )t (5.23)

The steady state is:

k2
M1∞ = M0 (5.24)
k1 + k2 tot
Because the total rate (−[k1 +k2 ]) is negative, the system does in fact move
towards this steady state.
 TWO-DIMENSIONAL MODELS 93

0
The solution for M2 (t) results either from M2 = Mtot − M1 or analo-
gously to Eq. (5.23):

M2 (t) = M2∞ + (M20 − M2∞ ) e−(k1 +k2 )t (5.25)

with the steady state

k1
M2∞ = M0 (5.26)
k1 + k2 tot

As we already noted, one eigenvalue of the coefficient matrix is zero. The

second eigenvalue is λ2 = −(k1 + k2 ). This explains why, in contrast to
Eq. (5.12), the solutions (5.23) and (5.25) only consist of one exponential
function. Note that the eigenvalue λ2 , and thus the overall response time
of the system, only depends on the sum of the two reaction rates k1 and
k2 . It is therefore sufficient if just one of the reaction rates is large for the
response time to be small.
Let’s look at a second example leading to a homogeneous system of
equations. This time, however, the system is not conservative, that is,
det(P) = 0:

Example 5.5: Radioactive decay chain

Naturally occurring radioactive isotopes are often used to identify the
age of rocks. From the activity ratioa between mother and daughter
isotope, it is possible to draw conclusions about the rock’s age. The I got this decay
chain at a local
amount of mother isotopes N1 and daughter isotopes N2 can be de- thrift store.
scribed with the following system of differential equations: A real bargain!

dN1
= −kλ,1 N1
dt (5.27)
dN2
= kλ,1 N1 − kλ,2 N2
dt
kλ,i is the decay constant of the isotope i. It is linked to the half-life
τ1/2,i through the following relation: kλ,i = ln 2/τ1/2,i .

a The activity of a radioactive substance is defined as the product of the

amount of isotopes and the decay constant kλ . The activity is measured as
the number of decays per time with the unit Bq = 1 Becquerel = 1 decay
per second.

We will now examine this system of equations more closely. We notice that
N2 depends on N1 , but N1 doesn’t depend on N2 . Such a system is called
hierarchical. The hierarchy of the system is expressed in the so-called
94 LINEAR MULTI-DIMENSIONAL MODELS

triangular form of its coefficient matrix, in which all elements above the
diagonal are zero6 :
 
−kλ,1 0
P= (5.28)
kλ,1 −kλ,2
This makes determining the eigenvalues and therefore the solution of the
system of equations particularly easy. From the characteristic equation
(5.11) it follows that the eigenvalues are identical with the diagonal
elements:

λ2 + λ(kλ,1 + kλ,2 ) + kλ,1 kλ,2 = 0

(λ + kλ,1 )(λ + kλ,2 ) = 0
⇒ λ1 = −kλ,1
λ2 = −kλ,2

A hierarchical system of equations can be solved iteratively, i.e. from top

to bottom, by determining N1 (t) from the first equation and inserting the
result into the second one. The first differential equation has the solution
we already know from Eq. (4.8) in Chap. 4:

N1 (t) = N10 e−kλ,1 t (5.29)

If we now insert the solution for N1 into the second differential equation,
we get:
dN2
= kλ,1 N10 e−kλ,1 t − kλ,2 N2 (5.30)
dt
To solve this differential equation we can use the solution for the linear
one-box model with exponentially varying input. To do so we can view the
first term on the right-hand side of the equation as the input function of
the system variable N2 , which exponentially falls with kλ,1 . The initial value
of the input function is j0 = kλ,1 N10 . If we now apply Eq. (4.57), and with
the initial condition N20 = 0, we get:
kλ,1 kλ,1
N2 (t) = N 0 e−kλ,1 t − N 0 e−kλ,2 t
kλ,2 − kλ,1 1 kλ,2 − kλ,1 1
(5.31)
kλ,1
= N 0 · (e−kλ,1 t − e−kλ,2 t )
kλ,2 − kλ,1 1
We see that the solutions for N1 (t) and N2 (t) are sums of exponential
functions again. The exponential coefficients are the eigenvalues of the co-
efficient matrix. Because the system is hierarchical, the second exponential
function (with kλ,2 ) does not appear in the solution of N1 (t). Only the so-
lution of N2 (t) depends on both decay constants or eigenvalues. Finally, we

6 For another configuration of the system variables N , the triangular form

i
could also mean that all elements below the diagonal are zero.
 TWO-DIMENSIONAL MODELS 95

learn that for the homogeneous system Eq. (5.27) with negative eigenvalues,
the steady state is zero.
Let us summarize what we have learned from this example. A hierar-
chical system can be recognized by the triangular form of the coefficient
matrix. Therefore, the eigenvalues are identical with the diagonal elements
of this matrix and the system of equations can be solved iteratively. In
the next example, we will explore a hierarchical system consisting of two
spatially separate boxes:

Example 5.6: Transport of a radioactive isotope through two

lakes
A radioactive isotope with the decay constant kλ is inadvertently in-
troduced into two neighboring lakes. The outflow of one lake leads to
the other lake. Both lakes can be seen as completely mixed reactors.
Figure 5.5 shows the box diagram of the system. The isotope input Ji
takes place via tributaries or directly into the lake. The outflow of the
second lake (Q2 ) can be larger than the one of the first lake (Q1,2 ), if
inflows bring water directly into the second lake. This kind of two-box
model was used by Lerman (1972) to describe the concentration of the
radioactive isotope strontium-90 in the Great Lakes of North America.

Fig. 5.5: Transport of a

radioactive isotope through
a chain of lakes depicted as
a box diagram

First, we establish the mass balance for the two boxes: Mi is the total mass
of the isotope in lake i. The external isotope input is given by the input
function Ji [ MT−1 ], and the radioactive decay is described by −kλ Mi . The
transport processes we take into account are the flow from the first into the
second lake and the second lake’s outflow. We end up with the following
system of differential equations:
     
dM1 Input into Flow into Radioactive
= − −
dt lake 1 lake 2 decay
= J1 − Q1,2 C1 − kλ M1
96 LINEAR MULTI-DIMENSIONAL MODELS

   
dM2 Input into Outflow from
= + (5.32)
dt lake 2 lake 1
   
Outflow from Radioactive
− −
lake 2 decay
= J2 + Q1,2 C1 − Q2 C2 − kλ M2
Again, Ci = Mi /Vi means the concentration in lake i. For the concentra-
tion, the dynamic equations are:
dC1 J1 Q1,2
= − C1 − kλ C1
dt V1 V1
J1 Q1,2
= −( + kλ )C1
V1 V1
(5.33)
dC2 J2 Q1,2 Q2
= + C1 − C2 − kλ C2
dt V2 V2 V2
J2 Q1,2 Q2
= + C1 − ( + kλ )C2
V2 V2 V2
The coefficient matrix has the form:
⎛   ⎞
Q
− V1,2 + kλ 0
P=⎝ 1
Q1,2
  ⎠ (5.34)
V2 − Q2
V2 + kλ

The system is therefore hierarchical and has the following eigenvalues:

Q1,2 Q2
λ1 = −( + kλ ) λ2 = −( + kλ ) (5.35)
V1 V2
The system of equations can be solved with the same procedure as in Ex-
ample 5.5; we leave the individual steps as an exercise to the reader. The
solution of Eq. (5.33) for C1 (t) is:
C1 (t) = C1∞ + (C10 − C1∞ ) · eλ1 t (5.36)
with the steady state
J1 J1
C1∞ = = (5.37)
(−λ1 )V1 Q1,2 + V1 kλ
Because of the system’s hierarchy, only the exponential function with the
first eigenvalue (λ1 ) appears in Eq. (5.36)
Once again, the result (Eq. 5.36) can be inserted in Eq. (5.33). The re-
sulting differential equation with a time-dependent inhomogeneous term
can be solved according to the recipe from Eq. (4.57). In the end, the result
can be brought into the form consisting of three terms that we already
know from Eq. (5.12):
Q1,2 (C10 − C1∞ ) λ1 t
C2 (t) = C2∞ + e
Q2 − (V2 /V1 )Q1,2
  (5.38)
Q1,2 (C10 − C1∞ )
+ C20 − C2∞ − e λ2 t
Q2 − (V2 /V1 )Q1,2
 TWO-DIMENSIONAL MODELS 97

with
J2 + Q1,2 C1∞
C2∞ = (5.39)
Q2 + V2 kλ
To illustrate this rather complex-looking result, we can look at the concrete
case where the outflow of a relatively large lake leads into a much smaller
lake. Maybe you can already intuitively guess how a radioactive substance
with half-life τ1/2 of about 12 year (this could be tritium, a radioactive
isotope of hydrogen denoted 3 H) behaves in this system:

Lake 1: V1 = 100 × 109 m3

Q1,2 = 30 × 108 m3 year−1
J1 = 300 × 109 Bq year−1
⇒ Average retention time: τ1 = V1 /Q1,2 33 year
Lake 2: V2 = 5 × 109 m3
Q2 = 45 × 10 m3 year−1
8

J2 = 190 × 109 Bq year−1

⇒ Average retention time: τ2 = V2 /Q2 1.1 year
Decay constant of the isotope: kλ = 0.06 year−1
Half-life: τ1/2 = 11.6 year

We can now calculate the steady-state concentrations from Eqs. (5.37) and
(5.39):

C1∞ = 33 Bq m−3
C2∞ = 60 Bq m−3

Let’s assume that the isotope influx (J1 and J2 ) is stopped abruptly. Of
course, this assumption is unrealistic, because the lakes’ catchment area
also has a “memory”, for instance in the soil, where the isotope is washed
out only slowly. Nevertheless, we can hypothetically examine how C1 and
C2 (the concentration in the two lakes) would develop.
First, we can see that the only change in the system of equations (5.33)
concerns the two inhomogeneous terms (J1 = J2 = 0). The system matrix
P and therefore also the eigenvalues don’t change. They have the following
values:
λ1 = −0.09 year−1 ; λ2 = −0.96 year−1
As per Eq. (5.16), the (absolute) smaller eigenvalue λ1 determines the lake
chain’s reaction to the input change:
3
τ5 % = = 33 year
0.09 year−1
In fact, the two eigenvalues are each composed of the water flow rate and the
decay constant. In this example they can therefore explicitly be allocated
to the two subsystems. The eigenvalue λ1 describes the isotope in the first
lake, λ2 the one in the second lake. Figure 5.6 shows the reduction of
98 LINEAR MULTI-DIMENSIONAL MODELS

Fig. 5.6: After a quick 70

drop, the tritium concen-
tration in Lake 2 adjusts to 60

Tritium concentration [Bq m−3]

the much slower decrease
in Lake 1 upstream 50

40

30
Lake 1
20

10 Lake 2

0
0 5 10 15 20 25 30 35 40
Time t [yr]

tritium in the two lakes. Starting from the steady-state concentration, the
tritium concentration in the second lake declines very rapidly. The first lake,
however, reacts more slowly. If the second lake did not lie downstream of
another lake, its adjustment time would be τ5 %,Lake2 = 3/0.96 ≈ 3 year.
The memory of the downstream lake is thus located in the water of the
upstream lake, much in the same way as the memory of both these lakes
lies in the soil of their watersheds. In Problem 5.2 we will return to this
example.

5.1.5 Two-Box Model for Stratified Systems

In Chaps. 2 and 4 we explored several aspects of the dynamics of phospho-
rus in a lake. We now want to further extend our one-box model through
the following examples, while introducing an additional element for the
construction of models.

Example 5.7: Lake with stratification parameter

An investigation of a lake’s phosphorus balance shows that the lin-
ear flow reactor model is not always valid, because the phosphorus
concentration in the lake’s outflow is usually smaller than the aver-
age phosphorus concentration in the lake. Instead of being completely
mixed, it seems that a lake is sometimes stratified (or layered) in terms
of the phosphorus concentration. We could adjust the dynamic equa-
tion (4.22) for the concentration of a substance in the lake as follows:
 TWO-DIMENSIONAL MODELS 99

dC
= kw Cin − kw Cout − kr C (5.40)
dt
Cout [ ML−3 ] Average concentration in the outflow

The price for this extension is the appearance of a new system variable
Cout . This begs the question how it is related to C. As a first attempt,
we set:
Cout
= γ = const. (5.41)
C
γ [-] Stratification parameter for phosphorus

By doing so we can now eliminate the system variable Cout in Eq. (5.40)
and again get a one-box model. This model is different from Eq. (4.22):

dC
= kw Cin − (γkw + kr )C (5.42)
dt

The steady-state concentration in the lake is:

kw
C ∞ = Cin (5.43)
γkw + kr
In contrast, the steady-state concentration of the lake as completely mixed
flow reactor would be (see Eq. 4.25):
kw
C ∞ = Cin (5.44)
kw + kr
In summer, the phosphorus concentration at the surface of the water col-
umn usually decreases due to algal growth. This renders the stratification
parameter smaller than one: γ < 1. Thus, the average steady-state phos-
phorus concentration in the stratified lake is higher than in the completely
mixed lake. In contrast, the outflow concentration is reduced by the factor
γ. All in all, by keeping the water layers with above-average concentra-
tions away from the outlet, the stratification causes the retention time of
phosphorus in the lake to increase. This model has an important drawback,
however: we don’t know how to link the stratification parameter γ to other
factors, e.g. to the mean phosphorus concentration in the lake (which may
change under the influence of a varying external input). To correct for this
deficit of the one-box model, we need to construct a true two-box model.
Many freshwater lakes are thermally stratified in summer. Warm (and
thus light) water floats above the heavy and cold deeper water (Fig. 5.7).
Due to this temperature distribution, the vertical water circulation is sup-
pressed in the vertical layer with the largest density gradient. This zone
is called the thermocline. Above the thermocline lies the so-called (warm)
epilimnion, below it the (cold) hypolimnion.
In Chap. 8 we will encounter models that can describe the phosphorus
concentration in a stratified lake as a continuous function of depth z and
100 LINEAR MULTI-DIMENSIONAL MODELS

I love Temperature / Density

stratified
systems!

Epilimnion

Thermocline

Hypolimnion

Lake depth
Water Water
temperature density

Fig. 5.7: Vertical temperature and density profile of a lake during summer. The
thermocline separates the warm surface water, called epilimnion, from the cold
deep water, called hypolimnion. In the thermocline, vertical water and mass trans-
port is reduced in comparison to the layers above and below

time t, that is, as C(z, t). The resulting equations, however, are usually so
complex that they can only be solved numerically by computer.
Multi-box models stand It is not always justified to jump straight from using a one-box model
as an intermediate step to a continuous space-time-model. Often, an intermediate solution is most
between one-box models practical. Figure 5.7 suggests such a solution: the partition of the lake
and spatially continuous into two subsystems, the epilimnion and the hypolimnion, that can each
models. individually be treated as completely mixed. Similarly, in describing the
atmosphere, we could divide it into layers such as the troposphere and
stratosphere. This approach can deliver acceptable results for many types
of systems.
As Fig. 5.8 shows, the spatial segmentation of a system entails the ques-
tion of how to describe the mass flow between the different subsystems. We
will discuss this problem more extensively in Sect. 8.2. For now, and in the
following example, we will anticipate and use the outcome of that discussion
without further explanation.

Example 5.8: The stratified lake

We want to describe the behavior of a dissolved chemical substance in
a lake during the summer months with a two-box model (Fig. 5.8). We
choose the average concentrations in the epilimnion and hypolimnion
as system variables: CE = ME /VE and CH = MH /VH . The direct
substance input Jin = QCin and outflow Jout = QCE only influence
the epilimnion. We describe the water exchange with an exchange flow
Qex CE and Qex CH . For this, we assume that the dissolved substance
 TWO-DIMENSIONAL MODELS 101

moves passively with the water volume Qex that is continuously

exchanged between epilimnion and hypolimnion. In addition, a linear
decay process is taken into account in both boxes. Due to the difference
in water temperature and available light, the decay rates kr,E and kr,H
in the epilimnion and hypolimnion are not necessarily the same.

Fig. 5.8: Two-box model

for a substance with decay
reaction in a stratified lake

First, we establish the mass balance for the two boxes. For the epilimnion,
it is:
dME
= {Inflow} − {Outflow} − {Decay} ± {Exchange}
dt (5.45)
= QCin − QCE − kr,E ME − Qex CE + Qex CH

For the hypolimnion we get correspondingly:

dMH
= ± {Exchange} − {Decay}
dt (5.46)
= Qex CE − Qex CH − kr,H MH

After dividing Eq. (5.45) by the epilimnion volume VE and Eq. (5.46) by
the hypolimnion volume VH we get the dynamic equations for the concen-
trations:
dCE Q Q Qex Qex
= Cin − CE − kr,E CE − CE + CH
dt VE VE VE VE
= kw Cin − kw CE − kr,E CE − kex,E CE + kex,E CH
= kw Cin − (kw + kr,E + kex,E )CE + kex,E CH
(5.47)
dCH Qex Qex
= CE − CH − kr,H CH
dt VH VH
= kex,H CE − kex,H CH − kr,H CH
= kex,H CE − (kex,H + kr,H )CH
102 LINEAR MULTI-DIMENSIONAL MODELS

For Eqs. (5.47) we have defined the following specific rates:

Q Qex Qex
kw = , kex,E = , kex,H = (5.48)
VE VE VH
We now want to discuss the solutions of this model. First, we note that the
equations conform to the general form of Eq. (5.1). Thus, we are dealing
with a two-dimensional linear model. The inhomogeneous terms are:
R1 = kw Cin and R2 = 0 (5.49)
The coefficient matrix is composed as follows:
 
− (kw + kr,E + kex,E ) kex,E
P= (5.50)
kex,H − (kex,H + kr,H )
This is no longer a triangular matrix. To determine the eigenvalues, we
would have to explicitly solve the characteristic equation. Note that the
diagonal elements have a negative sign, whereas the coefficients outside of
the diagonal are positive. Furthermore, −p1,1 > p2,1 and −p2,2 > p1,2 . The
eigenvalues of such a matrix are, as we have seen in Sect. 5.1.4 (Case b),
always real and negative. This means that the steady states of the system
exist and will be reached. We can now use Eq. (5.13) to calculate the steady
state:
∞ (kex,H + kr,H )kw Cin
CE =
(kw + kr,E + kex,E ) (kex,H + kr,H ) − kex,E kex,H
(5.51)
∞ kex,H kw Cin
CH =
(kw + kr,E + kex,E ) (kex,H + kr,H ) − kex,E kex,H
An example with concrete numbers is shown in Table 5.2.
Note that at steady state the concentration in the epilimnion is larger
than in the hypolimnion. This seems to contradict our earlier discussion
about the stratification parameter and the vertical distribution of phospho-
rus during lake stratification. If such inconsistencies between model results
and real-world experience or data appear, it is always useful to check the
consistency of the model results. In other words, we want to verify quanti-
tatively whether the solution of the system of differential equations (5.47)
is formally correct.
Let’s look at the mass balance of the hypolimnion: At steady state,
the loss through decay at rate kr,H has to be compensated for by a net
transport through the thermocline. The net transport kex,H (CE − CH ) is
only directed from the epilimnion into the hypolimnion if CE > CH . Thus,
the model result is formally correct.
The apparent discrepancy between model and reality must therefore
have another reason. In fact, many substances (including phosphorus) are
influenced by another transport process: the (directed) transport by sedi-
menting (sinking) particles from the epilimnion into the hypolimnion. We
have failed to take this process into account in our model so far. In Prob-
lem 5.5 we will analyze its effects (for further information on this process,
also see Chap. 23 in Schwarzenbach et al. 2003).
 TWO-DIMENSIONAL MODELS 103

Table 5.2: Numeric example for the mass balance of a stratified lake

Volume: Vtot = 150 × 106 m3

VE = 50 × 106 m3
VH = 100 × 106 m3
Water flow: Q = 0.34 × 106 m3 d−1
Total substance input: Jin = 40 t year−1
Specific decay rates: kr,E = 0.02 d−1
kr,H = 0.002 d−1
Water exchange: Qex = 0.5 × 106 m3 d−1
Calculation of parameters:

Jin 1.1 × 108 mg d−1

Cin = = = 320 mg m−3
Q 0.34 × 106 m3 d−1
Q
kw = = 6.8 × 10−3 d−1
VE
Qex
kex,E = = 1 × 10−2 d−1
VE
Qex
kex,H = = 5 × 10−3 d−1
VH
Calculation of matrix elements pi,j and Ri

p1,1 = −(0.68 × 10−2 + 0.02 + 0.01) d−1 = −0.0368 d−1

p1,2 = 0.01 d−1
p2,1 = 0.005 d−1
p2,2 = −(0.005 d−1 + 0.002 d−1 ) = −0.007 d−1
R1 = kw Cin = 2.18 mg m−3 d−1
R2 = 0

Steady states:
∞
CE = 73 mg m−3 and ∞
CH = 52 mg m−3

Average value across the whole lake:

1 1 ∞ 2 ∞
C∞ = ∞
(VE CE ∞
+ VH CH ) = CE + CH = 59 mg m−3
Vtot 3 3
Stratification parameter in the steady state according to Example 5.7:
∞
Cout C∞ 73
γ= ∞
≈ E = = 1.24
C C∞ 59
104 LINEAR MULTI-DIMENSIONAL MODELS

We also have to think about another problem: how realistic is it to

assume that the lake is at steady state? We’ve mentioned before that lakes
in a temperate climate exhibit stratification primarily during summer. In
autumn, this stratification is broken up by two processes: the cooling of
the surface water and the influence of strong autumn winds. It is possible,
therefore, that during winter the lake is completely mixed and behaves like a
one-box model. We now want to calculate the adjustment time for the two-
box model and compare this with the duration of the lake’s stratification
period.
To estimate the adjustment time we need the eigenvalues of the coefficient
matrix P. We calculate them with the help of Eq. (5.10):

1 
λ1 = (tr(P) + Δ(P))
2
1 √
= (−0.0438 + 0.00109) d−1 ≈ −0.0054 d−1
2
1 
λ2 = (tr(P) − Δ(P))
2
1 √
= (−0.0438 − 0.00109) d−1 ≈ −0.038 d−1
2
The absolute smallest The adjustment time can be estimated with the (absolute) smallest eigen-
eigenvalue generally de- value (Eq. 5.16):
termines the adjustment
time of a linear system. 3 3
τ5 % ≈ = = 555 d
|λ1 | 0.0054 d−1

The lake would therefore have to remain stratified for more than a year for
the calculated equilibrium concentrations to be reached. As we have seen,
this is not what happens in reality.
Let us try to capture the complex mixing cycle of a lake in a temperate
European climate with a model. To do so, we assume that the lake is
completely mixed from the beginning of January until the end of April,
that is, for 120 days. Then, a stagnation period of 8 months begins, during
which the lake is stratified. The model of the lake thus changes back and
forth from a one-box model in winter to a two-box model in summer. The
initial concentrations at the beginning of the two-box phase are identical
0 0
(CE = CH ), and correspond to the final concentration of the preceding
one-box phase. At the transition from the two-box to the one-box phase,
the initial concentration is calculated from the volume-weighted final state
of the two-box phase:

CE (t) VE + CH (t) VH
C0 = (5.52)
VE + VH
Figure 5.9 shows the temporal variation of the dissolved substance de-
scribed in Example 5.8 in a lake undergoing an annual cycle of stratifica-
tion and destratification. For the stratification period, the numbers from
 TWO-DIMENSIONAL MODELS 105

90 Fig. 5.9: The mixing cycle

of a lake can be simulated
with a periodic change be-
80 tween a one-box and a two-
Concentration in mg m−3
box model. The example
shows the balance for a
70 dissolved substance with
the coefficients from Ta-
ble 5.2. The coefficients
60 for the one-box model are
calculated with Eqs. (5.52),
(5.53) and (5.54). The sub-
50 scripts E and H stand for
epilimnion (upper layer)
and hypolimnion (lower
40 layer), respectively
0 60 120 180 240 300 360
Days after onset of stratification

Table 5.2 were used. For the mixing period, the elimination rate kr and the
flow rate kw were adjusted as follows:
kr,E VE + kr,H VH
kr = (5.53)
VE + VH
Q
kw = (5.54)
VE + VH
The computation starts at time t = 0 with the onset of stratification. Con-
centrations in the epilimnion and the hypolimnion develop differently. In
the epilimnion, the concentration increases, while it decreases in the hy-
polimnion, yet the steady state is not reached. After 8 months (245 days),
the lake is instantaneously and completely mixed, resulting in an average
concentration in the entire lake which lies below the steady-state concen-
tration of the mixing period. Thus, the concentration increases further over
the next 120 days but is still well below the steady-state value when the
cycle starts over at the onset of stratification.
In reality, the transition between complete mixing in winter and com-
plete stratification in summer is not sudden, but gradual. Nevertheless,
the basic outcome of the mixing cycle is reproduced quite well with this
two-phase model.

5.1.6 Linear Models with Non-real Eigenvalues

The examples so far were all based on the principle of mass balance in
systems with linear transformation and transport processes. We have seen
that the coefficient matrix of such systems always leads to real, non-positive
eigenvalues and that the steady state can in fact be reached.
As already suggested in Sect. 5.1.2, linear two-dimensional systems
can also have non-real eigenvalues. Let’s look at the following linear
106 LINEAR MULTI-DIMENSIONAL MODELS

homogeneous system of differential equations, with the two functions y1 (t)

and y2 (t). For now, we will not think about what these variables could
stand for.

Example 5.9: Systems with purely imaginary eigenvalues

dy1
= k1 y2
dt (5.55)
dy2
= −k2 y1 with k1 , k2 > 0
dt

Note that the derivative of y1 with respect to time is proportional to y2 ,

the derivative of y2 proportional to y1 . This reciprocal dependence already
lets us suspect that the solutions must be trigonometric functions. But first
let’s look at the system’s coefficient matrix:
 
0 k1
P= (5.56)
−k2 0
Its characteristic equation (see Eq. 5.11) is:
λ2 + k1 k2 = 0
(5.57)
λ2 = −k1 k2
Because k1 and k2 are positive, we get two purely imaginary eigenvalues:
 
λ = ±i k1 k2 = ±iω with ω = k1 k2 (5.58)
The general solution of Eq. (5.12) applies to this model, too. Because the
system is homogeneous (Ri = 0), the steady state is yi∞ = 0 according to
Eq. (5.13). The solutions are therefore of the following form:
y1 (t) = a1,1 eiωt + a1,2 e−iωt
(5.59)
y2 (t) = a2,1 eiωt + a2,2 e−iωt
In other words, they can be expressed by a sum of two trigonometric func-
tions7 :
y1 (t) = b1,1 cos ωt + b1,2 sin ωt
(5.60)
y2 (t) = b2,1 cos ωt + b2,2 sin ωt
The constants bi,j are determined by the initial condition (y10 , y20 ). For ex-
ample, we can choose for t = 0:
y10 = 0 , y20 = 1 (5.61)

7 The relation between the descriptions in Eqs. (5.59) and (5.60) can be cal-
culated from Euler’s formula, eiωt = cos ωt + i sin ωt. If the variables yi (t) are
real functions, this also holds for the coefficients bi,j . The ai,j from Eq. (5.59)
however are generally complex numbers. For details see Appendix C.4.
 TWO-DIMENSIONAL MODELS 107

Then, with b1,1 = 0 and b2,1 = 1, we get from Eq. (5.60):

y1 (t) = b1,2 sin ωt

y2 (t) = cos ωt + b2,2 sin ωt .

To determine the two other constants bi,j , we insert Eq. (5.60) into the
differential equations (5.55) and get the identities:

dy1
= ωb1,2 cos ωt = k1 (cos ωt + b2,2 sin ωt)
dt
dy2
= −ω sin ωt + ωb2,2 cos ωt = −k2 b1,2 sin ωt
dt

which can only be fulfilled if b2,2 = 0 and b1,2 = ω/k2 = (k1 /k2 )1/2 .
We thus get this solution for our chosen initial condition:

k1
y1 (t) = sin ωt
k2 (5.62)
y2 (t) = cos ωt

In Fig. 5.10, the two functions are depicted as a time and a phase diagram
for the case kk12 = 4. The time diagram shows two undamped harmonic
oscillations that are shifted by a fourth of a period. The phase diagram
shows a closed elliptic trajectory. The system circulates the steady state
without ever reaching it.

Fig. 5.10: The solution

Eq. (5.60) with the ini-
tial condition y10 = 0 and
y20 = 1, above as a time
diagram, below as a phase
diagram. (k1 /k2 ) = 4
108 LINEAR MULTI-DIMENSIONAL MODELS

Undamped oscillations

Undamped
collision!

Undamped
oscillations.

We know from experience that undamped oscillations represent unusual

situations which do not really occur in natural systems. In fact, undamped
systems are unstable in a certain sense, since an arbitrarily small term
in one of the two equations radically changes the long-term behavior of
the system. As an illustration we modify Example 5.9 by introducing an
additional term into the second equation, the meaning of which will become
clear later.

Example 5.10: Oscillation with damping term

In the second equation of the model, Eq. (5.55), we introduce the ad-
ditional term −εy2 . For now, ε is an arbitrary positive parameter. The
modified model equations are:
dy1
= k1 y2
dt (5.63)
dy2
= −k2 y1 − εy2 , ε>0
dt

Note that mathematicians like to use the parameter ε whenever they want
to point out that this term can be arbitrarily small. As our analysis will
show, it is the mere existence of this additional term—and not its size—
which is responsible for a fundamental difference between Examples 5.9 and
5.10. The model has the coefficient matrix:
 
0 k1
P= (5.64)
−k2 −ε
Its characteristic equation is slightly modified compared to Eq. (5.57):
λ2 + λε + k1 k2 = 0 (5.65)
with the eigenvalues

1   ε ε2
λi = −ε ± ε2 − 4k1 k2 = − ± − k1 k2 (5.66)
2 2 4
 TWO-DIMENSIONAL MODELS 109

This result encompasses the entire behavior spectrum of a simple linear

oscillatory system as described by Eqs. (5.63). First, we note that Eq. (5.55)
is contained within Eq. (5.63) as a special case (ε = 0). That system has
undamped oscillations as its solutions. If we now let ε grow by an arbitrarily
small amount larger than zero, so that ε2 4k1 k2 remains valid, the
expression below the root of Eq. (5.66) remains negative and the eigenvalues
can be approximated as follows:

ε ε2
λi = − ± i k1 k2 −
2 4
ε 
≈ − ± i k1 k2 (5.67)
2
ε 
= − ± iω, ω = k1 k2
2
ω is identical with the definition in Eq. (5.58). Also note that ε can be
neglected in the root, but not outside of it, because the term outside trans-
forms the purely imaginary eigenvalues into complex numbers with a nega-
tive real part. From Table 5.1 (Case 5), we already know that the solution
of the corresponding differential equation is a damped equation.
Let’s take a closer look at these solutions. Analogously to Eq. (5.60),
they have the form:
ε ε
y1 (t) = a1,1 e(− 2 +iω)t + a1,2 e(− 2 −iω)t
ε  
= e− 2 t a1,1 eiωt + a1,2 e−iωt (5.68)
ε  
y2 (t) = e− 2 t a2,1 eiωt + a2,2 e−iωt
If these equations are expressed by sine and cosine functions, we get a
ε
modified form of Eq. (5.60) where the factor e− 2 t appears in front of the
right side of the equations. This additional term has a damping effect on
the oscillation and ensures that the model reaches the steady state zero for
t → ∞ (see Fig. 5.11a). The seemingly unimportant small term −εy2 in Linear models with purely
Eq. (5.63) thus radically changes the long-term behavior of the model. We imaginary eigenvalues are
call this kind of behavior a structural instability. structurally unstable.

Structural instability

Conducting
test sequence
for structural
stability...
110 LINEAR MULTI-DIMENSIONAL MODELS

Fig. 5.11: (a) Weakly a

damped oscillation, to the 1.5 1.0
left as time diagram, to 1.0
the right as phase diagram 0.5
0.5
for (k1 = 4, k2 = 1) and
ε = 1. (b) Strongly damped 0 0
oscillation (creeping pen-
-0.5
dulum) with ε = 6 and an -0.5
unchanged ki -1.0
-1.5 -1.0
0 2 4 6 8 10 -1.5 -1.0 -0.5 0 0.5 1.0 1.5

b
1.0 1.5
1.0
0.5
0.5
0 0
-0.5
-0.5
-1.0
-1.0 -1.5
0 0.5 1.0 1.5 2.0 2.5 3.0 -1.5 -1.0 -0.5 0 0.5 1.0 1.5

If ε increases, but the condition ε < 4k1 k2 still holds, the expression under
the root in Eq. (5.67) remains negative, i.e. the eigenvalues remain non-
real. However, the angular frequency ω (which, for weak damping, remains
practically the same as in the undamped case) now drops to lower values:

ε2
ω→ω = 
k1 k2 − (5.69)
4
A radical change of behavior takes place for

ε2
≥ k1 k2 (5.70)
4
According to Eq. (5.66), both eigenvalues are now real and negative and the
system “creeps” without oscillation towards the steady state y1 = y2 = 0
(see Fig. 5.11b). As we will see in Example 5.11, the equations introduced in
Examples 5.9 and 5.10 can be used to describe an oscillator or a pendulum,
among other things. Therefore, a system with strong damping is also called
a creeping oscillation or a creeping pendulum.
As a concrete example for a model with a structural instability, we use
a classic example from physics: the harmonic oscillator. In Chap. 6, we will
encounter another application from an entirely different field.
 TWO-DIMENSIONAL MODELS 111

Example 5.11: The harmonic oscillator

We consider the mass M , which is held at rest by a spring with constant
spring constant f . The spring constant describes the linear relation
between the spring force K and the spring displacement x from its
equilibrium position at x = 0:

K = −f x (5.71)

If M is displaced from its equilibrium position, the mass is subjected

to an acceleration according to Newton’s laws of motion:

d2 x
M = K = −f x (5.72)
dt2
x [L] Displacement from equilibrium position
f [ MT−2 ] Spring constant (force per displacement)
M [M] Mass
K [ MLT−2 ] Spring force

As we might remember from school, the solutions of Eq. (5.72) are har-
monic oscillations. To demonstrate the connection to the previous model
(Eq. 5.55), we choose the following new variables to describe the harmonic
oscillator:
y1 ≡ x
dx (5.73)
y2 ≡
dt
If we differentiate y1 with respect to time t, it follows that:
dy1 dx
= = y2 (5.74)
dt dt
On the other hand, if we convert Eq. (5.72) with the relation
d2 x dy2
= (5.75)
dt2 dt
then it follows that:
dy2
M = −f y1 (5.76)
dt
or: 
dy2 f
= −ω y1
2
with ω= (5.77)
dt M
The dynamic equation (5.72) for the harmonic oscillator is a second-order
differential equation. With the conversions above, we have transformed it
into a system of two first-order differential equations: (5.74) and (5.77).
This system has the following coefficient matrix:
 
0 1
P= (5.78)
−ω 2 0
112 LINEAR MULTI-DIMENSIONAL MODELS

This matrix has the same type of eigenvalues as Eq. (5.55). Therefore, the
Watch out!
Harmonic model of the harmonic oscillator must be structurally unstable as well. If
entities are
structurally
we introduce a very small damping term (γ: damping constant) that is
unstable! proportional to the speed of the moving mass M , dx
dt , we get:

d2 x dx
M = −f x − γ (5.79)
dt2 dt

Or, transformed to a first-order system of differential equations:

dy1
= y2
dt (5.80)
dy2 γ
= −ω 2 y1 − y2 = −ω 2 y1 − εy2
dt M

The damping constant per mass, γ/M , takes over the role of ε in Eq. (5.63).
The amplitude of the oscillation becomes smaller and smaller with time
t, until the oscillator eventually remains still at its equilibrium position
(x = 0). If the damping exceeds a certain value, the oscillator “creeps” into
this position. We will return to this type of differential equations when we
look at two-dimensional nonlinear systems in Chap. 6.
Let us now summarize our insight so far. To examine the behavior of a
model and its steady state, we need to calculate the eigenvalues λi of the
coefficient matrix. If both of them are real and negative or zero, the model
reaches a steady state. If non-real eigenvalues occur, the model still reaches
a steady state as long as their real parts are negative or zero. If purely
imaginary eigenvalues occur, the model undergoes an undamped oscillation.
Eigenvalues with positive real parts lead to solutions that diverge towards
infinity.

Multi-box model

Aw, damn it!

Now who clogged
the toilet?

A multi-box model at
its steady state: What
Everything flows... goes in, comes out.
It s that simple.
 MULTI-DIMENSIONAL MODELS 113

5.2 Linear Models with Several System Variables

Only a small step remains to move from linear models with two variables to
linear models with more variables. The coefficient matrix P now becomes
a (n × n) matrix, according to the number n of system variables:
⎛ ⎞
p1,1 p1,2 · · · p1,n
⎜ p2,1 p2,2 · · · p2,n ⎟
⎜ ⎟
P=⎜ . ⎟ (5.81)
⎝ .. ⎠
pn,1 pn,2 · · · pn,n
The solution of the resulting n-dimensional systems of differential equations
in turn consists of sums of exponential terms of the form eλi t , where λi
are the eigenvalues of the coefficient matrix.8 For a model with n system
variables we therefore get the following solution:

V1 (t) = a1,0 + a1,1 eλ1 t + a1,2 eλ2 t + . . . + a1,n eλn t

V2 (t) = a2,0 + a2,1 eλ1 t + a2,2 eλ2 t + . . . + a2,n eλn t
.... (5.82)
..
Vn (t) = an,0 + an,1 eλ1 t + an,2 eλ2 t + . . . + an,n eλn t
Or, shorter:
n
Vi (t) = ai,0 + ai,j eλj t , i = 1, . . . n (5.83)
j=1

As in the case of two-dimensional models, the eigenvalues of the coefficient

matrix are calculated with the characteristic equation. For larger matri-
ces this is easiest with an appropriate computer program. The behavior of
multi-dimensional models can then be analyzed by looking at the eigen-
values, just as we did for two-dimensional systems. If all eigenvalues are
real and negative, the coefficients ai,0 are identical with the corresponding
steady states V∞ i of the system.
Only in few cases does it make sense to construct very large linear
models. If circumstances require a complex model, it will often contain
nonlinear components as well.9 For this reason we will make do with a
single example here (a further one can be found in Problem 5.3).
The carbon cycle is one of the most important geochemical cycles in
our environment. We will try to describe it with a simplified linear multi-
dimensional model.
8 For simplification, we assume that the system does not have multiple
eigenvalues.
9 Certain optimization problems can be described by very large linear systems

of equations that are then solved with methods called linear programming.
114 LINEAR MULTI-DIMENSIONAL MODELS

Example 5.12: The global carbon cycle

Figure 5.12 shows the global carbon cycle in highly simplified form as
a box diagram. Carbon is stored in three geochemical reservoirs: the
atmosphere, the ocean surface and the deep sea. It is stored predomi-
nantly in the form of carbon dioxide (CO2 ) as well as carbonate (CO= 3
or HCO− 3 ). Of course, carbon is continuously exchanged between these
reservoirs. The fourth important reservoir is terrestrial biomass. Ocean
biomass is so small that we can neglect it. The size of the four boxes and
the width of the arrows roughly represent the system’s pre-industrial
steady state. By far the largest reservoir is the deep sea, while the at-
mosphere is the smallest. The fact that carbon exchange between the
deep sea and the other reservoirs is small will be important for the
model.
With the onset of rapid industrialization in the mid-nineteenth century,
the combustion of fossil fuels (initially mostly coal, then also oil and
gas) led to an increase of atmospheric CO2 concentration. The pre-
industrial CO2 concentration was around 280 ppmv.a The value in the
atmosphere box in Fig. 5.12 represents this concentration. Figure 5.13
shows the rapid increase of atmospheric CO2 concentration during the
past 150 years. By now, of course, the CO2 concentration has surpassed
390 ppmv and the carbon cycle is no longer in an equilibrium.

a The unit ppmv is a non-dimensional concentration measure. It means

parts per million per volume, that is, 1 ppmv means that a 10−6 volume
portion of the air consists of CO2 .

We want to construct a dynamic model from the pre-industrial steady state

(Fig. 5.12). Then, we will treat the disturbance of the system through fossil
fuel combustion as a singularity in time, i.e., we will introduce the total
amount of 300 × 1015 g C into the atmosphere at time t = 0 in one fell
swoop. This represents roughly the amount of carbon introduced to the
atmosphere by the burning of fossil fuels between 1,850 and 2,000, not
taking into account additional sources of carbon such as emissions from
land-use change and deforestation.
We want to answer three questions with our model:
1. How does the carbon inventory Mi (t) develop over time in the four
boxes?
2. Where will the additional carbon be stored in the long term?
3. How long will it take for the system to reach an equilibrium again, if
no further (anthropogenic or natural) disturbances occur?
How do we get from the static image in Fig. 5.12 to a dynamic model?
At this point we have to make an assumption about how the fluxes Fi,j
 MULTI-DIMENSIONAL MODELS 115

Atmosphere Fig. 5.12: Simplified

global carbon model with
the most important ex-
change fluxes. The num-
Land Ocean surface bers in the boxes give the
900
yearly carbon stock (unit
is 1015 g C), the numbers
Deep sea
next to the arrows give the
fluxes (unit 1015 g C a−1 ),
the percent values give the
relative proportion of the
entire carbon stock that
the box holds. The situa-
tion represents roughly the
pre-industrial steady state
(numbers are simplified
from Moore et al. 1994)

400 Fig. 5.13: Atmospheric

concentrations of CO2
over the last 2,000 years
(Source: Forster et al.
350 (2007))
CO2 (ppmv)

300

250
0 500 1000 1500 2000
Year

between the boxes react to changes of substance mass in the reservoirs. We

define them such that the first index (i) denominates the receiving box and
the second index (j) the originating box (see Table 5.3). Since we want to
construct a linear model, the simplest way is to describe the fluxes Fi,j as
linear functions of the content Mj of the originating box. To do so, we will
first look at the two boxes atmosphere and land (Fig. 5.14).
116 LINEAR MULTI-DIMENSIONAL MODELS

Table 5.3: Exchange rates and dynamic equations of the global linear carbon
model. Labeling of fluxes and rates: first index = number of the destination box,
second index = number of the originating box

Mi [ 1015 g C ] Reservoir size of the box i

Fi,j [ 1015 g C year−1 ] Flux from box j into box i
ki,j [ year−1 ] linear flux rate for the transport from box j
into box i
Calculation of the rates:
Because we assume Fi,j to be proportional to the originating reservoir j,
we get:
Fi,j
Fi,j = ki,j Mj thus ki,j =
Mj

Values of the specific transfer rates ki,j :

k1,2 = 0.061 a−1 (Land → atmosphere)

k2,1 = 0.183 a−1 (Atmosphere → land)
k1,3 = 0.111 a−1 (Ocean surface → atmosphere)
k3,1 = 0.167 a−1 (Atmosphere → ocean surface)
k3,4 = 1.14 × 10−3 a−1 (Deep sea → ocean surface)
k4,3 = 0.044 a−1 (Ocean surface → deep sea)

The dynamic mass balance equations for the reservoirs are:

dM1
= −(k2,1 + k3,1 )M1 + k1,2 M2 + k1,3 M3
dt
dM2
= k2,1 M1 − k1,2 M2
dt
dM3
= k3,1 M1 − (k1,3 + k4,3 )M3 + k3,4 M4
dt
dM4
= k4,3 M3 − k3,4 M4
dt

We calculate the specific exchange rates ki,j between land and atmo-
sphere as follows:
F1,2 110 × 1015 g a−1
k1,2 = = = 0.061 a−1
M2 1, 800 × 1015 g
F2,1 110 × 1015 g a−1
k2,1 = = = 0.183 a−1
M1 600 × 1015 g
Table 5.3 shows the calculations for the transfer rates and the transport
equations of the entire model. The smallest rates are, as we already noted,
the ones between the ocean surface and the deep sea.
 MULTI-DIMENSIONAL MODELS 117

Fig. 5.14: From the de-

scription of the steady
Atmosphere state we can calculate ex-
change rates for a linear
dynamic model by assum-
ing that the flux between
two boxes is a linear func-
tion of mass in the origi-
nating box (Excerpt from
Fig. 5.12)
Land

300 Fig. 5.15: Numerical so-

lution of the linear car-
Deviation from the steady state [1015 gC]

bon model. Starting from

250
the sudden addition of
Atmosphere 300 × 1015 g C to the at-
mosphere, the figure shows
200
the deviation of the four
Deep sea carbon reservoirs from
150
their pre-industrial steady
state as a function of
elapsed time. Note that af-
Land ter 50 years, the system is
100
still far from a new steady
state. For any point in
50 time the sum of all curves
Ocean surfacef is equal to the added mass
of 300 × 1015 g C
0
0 20 40 60 80 100
Time [yr]

How can this simple model be used to answer the three questions
posed above? The dynamic development of the four carbon reservoirs Mi (t)
can be calculated analytically or with the aid of a computer program.
In the first case, each variable Mi (t) is described by a constant and four
time-dependent exponential functions (see Eq. 5.83). The initial values of
Mi (i = 2, 3, 4) are those in Fig. 5.12, while the atmospheric reservoir (M1 )
is increased by the anthropogenic input of 300 × 1015 to 900 × 1015 g C.
Figure 5.15 shows the result of a numerical simulation of the model’s
development through time. As expected, the biggest reservoir, M4 (the
deep sea), reacts most slowly. That is because it’s only coupled to the
other reservoirs by a fairly small (relative to the reservoir size) mass flux.
The second question, regarding the new steady state, can easily be an-
swered from Table 5.3 without solving the system of equations. Because we
are dealing with a linear homogeneous model, the relative sizes of the Mi at
118 LINEAR MULTI-DIMENSIONAL MODELS

steady state remain unchanged, i.e. as given by the pre-industrial values of

Fig. 5.12. In other words: Each box absorbs an amount of the additional car-
bon (300 × 1015 g C) in proportion to its original size. Even though the deep
sea reacts slowly, it will—always according to our simple model—eventually
contain 91.4 % of the additional carbon. Overall, the 300 × 1015 g C only
make up 0.8 % of the sum of all reservoirs (38 300 × 1015 g C). In the new
steady state, the reservoirs will therefore have grown by less than 1 % of
their original size. In fact, the carbon drama of modern society is not pri-
marily one of total mass but results from the unfavorable fact that by burn-
ing fossil fuels we load a fairly small carbon reservoir which is dynamically
“far away” from the largest reservoir, the deep sea.10
Finally, to predict how long it will take for the new equilibrium to be
reached (thereby answering our last question), we need to calculate the
eigenvalues of the coefficient matrix P:
⎛ ⎞
−0.35 0.061 0.111 0
⎜ 0.183 −0.061 0 0 ⎟
P=⎜ ⎝ 0.167
⎟ (5.84)
0 −0.155 0.00114 ⎠
0 0 0.044 −0.00114

Because P is constructed from the steady-state solution of a homogeneous

system of equations, the matrix must be singular with one eigenvalue of
zero. The result of a numerical eigenvalue calculation confirms this predic-
tion.
We get:

λ1 = −0.443 year−1
λ2 = −0.114 year−1
λ3 = −0.00971 year−1
λ4 = 0

Thus, we can estimate the adjustment time with Eq. (5.16):

3 3
τ5 % ≈ = ≈ 310 year (5.85)
min (|λi | = 0) 0.00971 year−1

Of course, this carbon model is not very realistic. First, we treated the
anthropogenic disturbance as a single input event. Figure 5.13 however
shows that an exponentially growing input function comes much closer to
reality. A second simplification is the assumption that all fluxes are linear
functions of the reservoir sizes.
Nevertheless, the model illustrates important properties of the carbon
cycle that are confirmed by more complex models. As mentioned above,
the anthropogenic disturbance, although relatively small compared to the
natural fluxes, has a large impact on the atmosphere for two reasons: the
10 Of course, this leaves aside the increasing acidity of our oceans as a result of
their atmospheric carbon uptake, a process which has severe effects of its own.
 MULTI-DIMENSIONAL MODELS 119

disturbance is directed at the smallest reservoir, and that reservoir is po-

sitioned “far” from the main reservoir—the deep sea. Experienced systems
thinkers will be able to make qualitative statements like this just from
looking at Fig. 5.12 and the sizes of the different mass transfer rates ki,j .
120 LINEAR MULTI-DIMENSIONAL MODELS

5.3 Questions and Problems

Question 5.1: In many cases, one-dimensional models are insufficient to
describe important properties of a system. Name different possibilities to
further develop a one-dimensional model into a two-dimensional or multi-
dimensional model.

Question 5.2: To at least how many model parameters can a two-

dimensional linear homogeneous model be reduced?

Question 5.3: Which properties/characteristic quantities of the coefficient

matrix of a multi-dimensional model are important for the temporal be-
havior of the system variable?

Question 5.4: What is the characteristic equation of a matrix?

Question 5.5: Homogeneous linear systems of differential equations with

real eigenvalues only have a steady state if the coefficient matrix has a
certain property. Which one?

Question 5.6: What is meant by the structural instability of a model?

Question 5.7: What is a hierarchical linear model? Give a systemic and

a mathematical explanation/definition.

Question 5.8: Which property of a linear system is indicated by the oc-

currence of an eigenvalue of zero?

Question 5.9: For a car, the parts that behave like a creeping pendu-
lum play an important role. If they lose this property, they urgently need
replacement. Which parts are they?

Question 5.10: How can the overall reaction rate of a linear model be
determined? Are there special cases where some system variables move
towards a steady state much more rapidly than the overall rate would
indicate? Give a mathematical explanation and a concrete example.

Question 5.11: CO2 sequestration is the idea of capturing and pumping

anthropogenic CO2 into the deep sea. Why would this idea make sense from
a systems point of view, leaving aside the negative ecological consequences
and the technical difficulties? How would Fig. 5.15 look qualitatively, if the
entire anthropogenic carbon output were deposited in the deep sea right
from the outset?

Question 5.12: The net flux between two neighboring spatial boxes can
be described as a linear function of the concentration difference between
them. In Example 5.8 we applied this principle. What idea is it based on?
 QUESTIONS AND PROBLEMS 121

Problem 5.1: Reactor with two substances

In Examples 5.1–5.3 we analyzed a reactor with through-flow in which a
chemical conversion between the substances A and B takes place. Calculate
the development of the concentrations CA (t) and CB (t) for the parameters
in
given in Example 5.2, an input concentration of CA = 1 mol m−3 , and the
0 0
initial concentrations CA = CB = 0.

Problem 5.2: Chain of lakes

Let’s take another look at Example 5.6 with the radioactive isotopes in the
chain of lakes:
(a) The model equations for the total isotope mass in the two lakes
(Eq. 5.32) cannot be solved, because apart from Mi , they also contain
the “extrinsic” concentrations Ci . Transform the system of equations
so that a solvable two-dimensional linear system of differential equa-
tions for Mi results. Show that the eigenvalues are identical with those
of Eq. (5.33) for Ci .
(b) Combine the two lakes into a “super system” and formulate the dif-
ferential equation for the sum of the isotope mass in both lakes,
M = M1 + M2 . Why does this not result in a one-dimensional model?
Which additional assumption is needed for it to become one?
(c) Why doesn’t it make sense to apply the same procedure to the sum
of concentrations?

Problem 5.3: Radioactive decay chain with three isotopes

Consider the decay chain of three radioactive isotopes:
222
Rn → . . . 214
Pb → 214
Bi → . . . 210
Pb (→)

The dots indicate that we are neglecting certain intermediate products with
a very short half-life. Although 210 Pb is not stable either, it has a much
larger half-life than the other isotopes shown in the chain. Its decay plays
no role in the following problem.
To simplify, we choose the following notation (half-life in parentheses):

X≡ 222
Rn (3.8 d), Y ≡ 214
Pb (26.8 min), Z≡ 214
Bi (19.8 min)

Establish the system of differential equations for the activities Ai and cal-
culate the coefficient matrix. Then, calculate the activity of 214 Bi as a
function of time, AZ (t), and from that, derive a an approximate formula
for t > 10 h. The initial activities A0i are: A0X = 1, 000 Bq, A0Y , A0Z = 0 Bq.
(1 Bq = 1 decay per second).

Problem 5.4: Tritium in a sewage treatment plant

A simplified description of a sewage treatment plant is given in the box
diagram below (Fig. 5.16). At time t = 0, an accident causes a substantial
122 LINEAR MULTI-DIMENSIONAL MODELS

amount of the radioactive isotope tritium to enter the combined sedimenta-

tion and aeration tank. In the tank, a tritium activity of A1 (0) = 106 Bq/m3
(1 Bq = 1 Bequerel = 1 decay per second) is measured. After this isolated
input event, the tritium activity in the inflow is once again zero.

Fig. 5.16: Box diagram of Primary sedimentation

a sewage treatment plant Inflow Outflow
Aeration
Q Final sedimentation Q
V1 A1

QF QF

Digester
V2 A2

V1 = 104 m3 Q = 104 m3 d−1

V2 = 3, 000 m3 QF = 60 m3 d−1

(a) Establish the dynamic equations for the tritium activities in the com-
bined sedimentation/aeration tank (A1 (t)) and in the anaerobic di-
gestion tank (A2 (t)) for the time after the accident (t > 0).
(b) The half-life of tritium is t1/2 = 12 year. Compare the decay constant
of tritium with the different water transport rates of the system. How
large is the influence of radioactive decay on the tritium activity in
the plant’s outflow?
(c) Estimate the eigenvalues of the system. Hint: set the absolutely small-
est matrix element and the radioactive decay constant of tritium equal
to zero.
(d) With what rate does the activity in the outflow decline after a few
months?

Problem 5.5: Stratified lake with sedimentation

We consider a stratified lake with a total volume of Vtot = V1 + V2 =
3 × 108 m3 . The volume of the surface layer V1 is 1 × 108 m3 . Inflow and
outflow (rate Q = 1 × 106 m3 d−1 ) only occur via the surface layer. The
vertical exchange rate between the two layers is Qex = 5 × 106 m3 d−1 .
(a) How large is the average water retention time
– In the volume V1 with respect to inflow, outflow and exchange
rate?
– In the volume V2 with respect to the exchange rate?
 QUESTIONS AND PROBLEMS 123

(b) A conservative substance (inflow concentration Cin = 100 mg m−3 )

flows into the lake at time t = 0. For t < 0, C1 (t) = C2 (t) = 0.
Calculate the steady state concentrations C1∞ and C2∞ . How long
does it take for the system to approach these values up to 5 %?

(c) How do the answers to question (b) change if the substance is ra-
dioactive with a decay constant of kλ = 1 × 10−2 d−1 ?

(d) Answer question (b) for a non-radioactive (that is, conservative) sub-
stance which adsorbs on particles and then gets transported with
them from the surface into the deep water and finally to the sediment.
The relevant flows are linear functions of the masses M1 = C1 V1 and
M2 = C2 V2 :
Transport V1 → V2 : ks1 M1
Transport V2 → sediment : ks2 M2
Use ks1 = 0.02 d−1 , ks2 = 0.005 d−1.
Also, neglect the fact that the lake’s cross section decreases with
depth, i.e., assume that the entire sediment flux from V1 ends up in
the volume V2 . Remember that at the transition from equations for
Mi to those for Ci the volumes are different: not the same amount of
mass leaves V1 per volume as arrives in V2 per volume (Eqs. 5.45–5.48
may be helpful here).

(e) Compare the results from (b) to those from (d) and attempt to explain
qualitatively the respective ratios between C1∞ and C2∞ .

Problem 5.6: Conservative substance in a chain of lakes

We consider two successive, completely mixed lakes (volumes V1 and V2 ),
with a constant through-flow Q0 .
V1 = 1 × 105 m3 , V2 = 2 × 106 m3
Q0 = 1 × 105 m3 d−1

(a) Due to an accident, M = 200 kg of a conservative pollutant enters the

upper lake (Lake 1) at t = 0. Describe the concentration development
in both lakes, C1 (t) und C2 (t). How long does it take until C1 (t) and
C2 (t) have decreased to 10 µg/L?

(b) What is the maximum concentration reached in Lake 2, and when is

it reached?

(c) After 5 days, measurements show that the concentration in Lake 2’s
outflow is decreasing exponentially (e−αt ). How large is α?

(d) How would C1 (t) and C2 (t) change if, for storing electric energy, water
would be pumped at an average rate of Qp = 4 × 105 m3 d−1 from
Lake 2 to Lake 1, and the flow from Lake 1 to Lake 2 would increase
correspondingly by Qp ? When do C1 and C2 decrease below 10 µg/L?
124 LINEAR MULTI-DIMENSIONAL MODELS

(e) Explain the differences between the results from (a) and (d)
qualitatively.

Problem 5.7: Temperature control by a thermostat

The heating system of a building is controlled by a thermostat. The ther-
mostat opens and closes the valve of the system which feeds warm water
into the radiators mounted throughout the building. The system works as
follows:

1. Without heating, the building temperature y would decrease due to

heat loss through the walls and windows by the constant rate L:
dy
= −L
dt

2. The temperature increase per unit of time due to heating is propor-

tional to the position of the valve, x, where x = 0 means that the
valve is closed. For simplicity, it is assumed that x has no upper limit.
The proportionality factor is k1 .
3. The valve opens and closes with a speed which is linearly related to
the difference between the actual room temperature y and the target
room temperature y0 . The proportionality factor is k2 .

Answer the following questions:

(a) Formulate the set of differential equations for x and y and discuss the
properties of the solutions. Is a steady state reached?
(b) Discuss the limits of the model. Hint: in some cases the position
of the valve may become negative. What happens qualitatively to
the solution if the dynamic equations are modified such that dx/dt
becomes 0 as soon as x drops below 0?
(c) How should the coefficients k1 and k2 be chosen to make the oscilla-
tions of y small?
(d) Are there other means to dampen the temperature fluctuations in the
building?
Chapter 6

Nonlinear Models

Nonlinear
box model, Nice, keep
right? it up!

Natural systems are rarely linear. Even if they are, that linearity often
exists only within a limited range of the system variables Vi . Indeed, the
surprising diversity in the behavior of natural systems is mostly a result
nonlinear processes. Despite ever more powerful computers, the behavior
of many nonlinear systems can only be predicted over small periods: the
weather forecast is a prime example of this.
Continuous nonlinear systems (in time) are described by nonlinear dif-
ferential equations. If the system variables are continuous in time and space,
we are dealing with nonlinear partial differential equations, such as those
occurring in fluid dynamics. The field of fluid dynamics describes the dy-
namics of gaseous and fluid systems and forms the basis of atmospheric
science and oceanography. It can be used to model such phenomena as the
Gulf stream or a tropical cyclone. In this chapter, we will start with spa-
tially discrete models, i.e. our usual box model approach. In Chap. 8 we
will introduce spatially continuous models and get a first glimpse at the
colorful world of these types of models.
Nonlinear differential equations can only be solved analytically in spe-
cial cases. We usually have to apply numerical methods, which is not the
topic of this book. In some cases, however, nonlinear models can be ap-
proximated by linear equations so that their behavior near the steady state

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 125

DOI 10.1007/978-3-642-30639-6_6, © Springer-Verlag Berlin Heidelberg 2013
126 NONLINEAR MODELS

can be analyzed. But in this chapter we will also go beyond such approxi-
mations and discuss some behavioral traits of nonlinear models that have
no resemblance with the behavior of their linear siblings anymore.

6.1 Nonlinear Models with One System Variable

6.1.1 Autonomous Nonlinear Models

Let’s look at a first-order nonlinear system1 with one system variable V.
In the most general case, it can be described by the following differential
equation:

dV
= g(V(t), t) (6.1)
dt

With this seemingly cumbersome notation we want to point out that the
general function g can depend on time both explicitly and implicitly via V
(see Sect. 4.1). The explicit time dependence arises from the external and
the implicit time dependence from the internal relations.
In many cases (but not in all) we can functionally separate the influence
of external and internal relations, so that we can rewrite Eq. (6.1) in the
following form:

dV
= R(t) + f (V) (6.2)
dt

Some points to note: we want the function of the external relation R to be

independent of V. If it is constant in time, it can be integrated into the
function f . In that case the system is autonomous. We also don’t want f
to explicitly depend on time. Finally, f (V) determines whether the model
is linear: if f (V) is not of the form a + bV, the model is nonlinear.
Let’s examine an autonomous (i.e. R = 0) nonlinear model more closely.
A classic example is the logistic growth model. It was introduced by Ver-
hulst (1838) to describe population growth. Pearl and Reed (1920) used it
to model population dynamics in the United States from 1790 on. The logis-
tic growth model was the basis on which many later models in population
ecology were based (see e.g. Krebs 2001, May and McLean 2007).

1 As a reminder: First-order means that only the first derivative of V with re-
spect to time appears. Systems with higher derivatives, which are often found
in physics, can be transformed into multi-dimensional first-order systems (see
Example 5.11).
 ONE-BOX MODELS 127

Example 6.1: Logistic growth model

In Example 4.2 we discussed the exponential growth model. Exponen-
tial growth cannot go on forever, because the population would grow
to infinity. The logistic growth model “fixes” this behavior: for small
population sizes N , growth is exponential, but it levels off as popu-
lation increases, eventually becoming zero. In a similar fashion as in
Eq. (4.11), we can describe logistic growth with a specific growth rate
kp , which now itself depends on N :

dN
= kp (N ) · N (6.3)
dt
For kp (N ) we choose a function that decreases linearly from a max-
imum value kp◦ at N = 0, reaching zero at N = Nmax : kp =
kp◦ (1 − Nmax
N
). Inserting this in Eq. (6.3) yields:

dN N
= kp◦ (1 − )N ≡ f (N ) (6.4)
dt Nmax

The growth rate kp◦ has the dimension [ T ]−1 , Nmax has the same
dimension as N , e.g. a population number, or a population number
per surface area.

For the following considerations it will be helpful to plot the change function
f (N ). In our example, f (N ) is a parabola with a downward facing opening
that intersects the N -axis at N = 0 and N = Nmax (Fig. 6.1a). These two
N -intersects of f (N ) show where the function N (t) “stands still”, that is,
where dN dt = 0. Thus, they are called fixed points. The linear models we dealt
with so far had at most one fixed point, which we called the steady state.2 Ima
fixed point!
Now, however, we have two. This begs the question whether the system is
moving towards one of the fixed points, and if so, towards which one.
In the next section we will discuss a general recipe to answer that ques-
tion. For our current example, it is easy to see that the population N (t)
develops towards the value Nmax except in the case where the initial value
N 0 is zero, since from nothing, nothing can grow! In fact, the differen-
tial equation (6.4) can be analytically integrated, which gives the following
result3 :
N0
N (t) = Nmax ◦ (6.5)
(Nmax − N 0 ) e−kp t + N 0

2 From now on we will only speak of fixed points, since the concept of a
steady state makes less sense as we move towards higher-dimensional and more
abstract models.
3 Since discussing fundamental mathematical principles is not the aim of this

book, we will not go through the integration of Eq. (6.4) in detail. Just one
note: the solution can be derived by using the variable separation method.
128 NONLINEAR MODELS

Fig. 6.1: The logistic a b

0.3 1.2
growth model: (a) change
function dN
dt
= f (N ) for Nmax
1.0
kp◦ = 1 and Nmax = 1;
0.2

Change function f (N )
(b) behavior of the popula-

Population N (t)
0.8
tion size N (t) (see Eq. 6.5)
for kp◦ = 1, Nmax = 1 and 0.1 0.6
N 0 = 0.01
0.4
0
0.2
Nmax

-0.1 0
0 0.2 0.4 0.6 0.8 1.0 1.2 0 2 4 6 8 10
Population N Time t

Equation (6.5) is shown in Fig. 6.1b. If the initial state is N 0 Nmax ,

◦
N (t) initially (if t (kp◦ )−1 ) grows exponentially (N (t) = N 0 ekp t ) and, for
t (kp◦ )−1 , reaches the constant value Nmax .

6.1.2 Fixed Points of Nonlinear Models with One Variable

In Sect. 4.2.1 we saw that a linear one-dimensional differential equation has
at most one fixed point. If the inhomogeneous term is constant, the sys-
tem approaches this fixed point irrespective of its initial state. In contrast,
as Example 6.1 shows, a nonlinear model can have several fixed points.
This leads to the question whether there are any general rules describing
the system’s behavior around these fixed points. In this section, we will
first formulate an answer for one-dimensional models. The more complex
situation of multi-dimensional nonlinear models will follow in Sect. 6.2.
The following discussion assumes that the external relation R is constant
in time, since otherwise the system would not have any constant fixed
points. We therefore write Eq. (6.1) in the following form:

dV
= g(V), (6.6)
dt
We also allow for a possible constant external relation, incorporating it into
the change function g(V), and treat the system as autonomous. We assume
that g(V) has k roots V∞ i (i = 1, . . . , k), that is, the following holds for all
V∞
i :
g(V∞i ) =0 for i = 1, . . . , k (6.7)
The V∞i are then fixed points of the differential equation (6.6). If the system
variable V reaches a fixed point, the model stays put at that position,
because: 
dV 
=0 (6.8)
dt V∞
i
 ONE-BOX MODELS 129

V cannot move away from this point, except if the external relation R (and
with it, g(V)) were to change.
The fixed points of a one-dimensional systems are therefore the roots of
g(V). If g(V) is an n-degree polynomial, there is a maximum of n real roots. Intervals between
These, together with the virtual fixed points V = ∞ und V = −∞, divide fixed points of a one-
the V-axis into (n + 1) intervals (Fig. 6.2a). Because a one-dimensional dimensional nonlinear
system can only move along one variable axis, each fixed point represents a model are called invariant
barrier that the system cannot overcome. Sections between two neighboring areas.
fixed points are called invariant areas. They are a characteristic feature of
one-dimensional models. In two- and n-dimensional models, the system can
move in a two or n-dimensional space (see Fig. 5.2). Therefore, fixed points
can usually be “bypassed”.

a
g( )

Area I Area II Area III Area IV

−∞ +∞

Initial state

0 ∞ ∞ 0 ∞
1 2 3

dg
b Line with slope d ∞
G( ) 2

∞ ∞ ∞
1 2 3

Basin of attraction

Fig. 6.2: (a) A one-dimensional system with three fixed points partitions the
variable axis into four so-called invariant areas (I–IV). From the initial state
V0 , the system can only move to the fixed points V∞ ∞
2 and V3 . It cannot leave
the invariant area III, however. (b) Topographic illustration of the same stability
conditions as in (a), as what we call a stability landscape. The ball (which is
infinitely damped, i.e., does not behave like a ball in reality) stays still at V∞ ∞
2 . V1
∞
and V3 are unstable positions, at which a small nudge in either direction will set
the ball in motion. Relation between g(V) and G(V): g(V) = − dG(V) dV

Let’s imagine a model with an initial state V0 lying inside area III of
Fig. 6.2a. So far, we don’t know towards which of the neighboring fixed
130 NONLINEAR MODELS

points this model will move. But we can attempt to answer the question
intuitively: in area III, as we can see in Fig. 6.2a, g(V) is negative. According
to Eq. (6.6), dV/dt < 0, that is, V decreases and the model moves leftwards
on the V axis. This movement keeps up until the model reaches the fixed
point V∞2 , where it stops and remains.
To make these thoughts more formal, we can look at the change function
g(V) in the immediate vicinity of the fixed point V∞ ∞
i , e.g. at Vi + ε with
∞
a small ε. We can then describe g(V) at the point Vi as a Taylor series4
and abort after the linear term:

∞ ∞ dg(V) 
g(Vi + ε) ≈ g(Vi ) + ε (6.9)
dV V∞
i

However, since V∞
iis a fixed point, g(V∞
i )
= 0. Thus, it follows from
Eqs. (6.6) and (6.9):
 
dV  ∞ dg 
= g(V + ε) ≈ ε (6.10)
dt V∞ +ε i
dV V∞
i i

dV |Vi
dg
Because the derivative is independent of ε, our approximation in
∞

Eq. (6.10) replaces the function g(V) with a straight line in the vicinity of
V∞i . In other words, we have created a linear model out of a nonlinear one.
∞
Figure 6.2a shows the linearized velocity function in the proximity of V2 .
dg 
If we define k ≡ dV  ∞ 5 and describe ε as the deviation of the momentary
V2
state from the fixed point V∞ ∞
2 , ε = V − V2 then Eq. (6.10) becomes:

dV
= k(V − V∞ ∞
2 ) = −kV2 + kV (6.11)
dt
This is nothing else than the linear inhomogeneous differential equation
(4.5) with constant coefficients. If k < 0, the solution tends to the fixed
point V∞ 6
2 . For k > 0 the model moves away from it. For k = 0 the system
is indifferent: it just stands still where it is. These statements are valid
whether the initial state is to the left or right of the fixed point in question,
as long as the absolute value of the deviation |ε| is not too large. The fixed
point V∞2 in Eq. (6.2) “attracts” the system inside area II as well. We thus
call V∞
2 a stable fixed point or an attractor. Areas II and III together form
its basin of attraction.
In contrast, the model moves away from the fixed point V∞ 3 in both
directions as soon as it deviates to the left or right by the small value ε.

4 A function g(x) that is n times differentiable can be described as a power

series (Taylor series) in the vicinity of location x0 :
  (n)
g(x) = g(x0 ) + (x − x0 ) g (x
1!
0)
+ (x − x0 )2 g 2!
(x0 )
+ . . . + (x − x0 )n g n!(x0 ) + . . .
5 You can verify that k has the dimension [ T ]−1 .
6 If the system were indeed linear, V would tend to ∞. Since Eq. (6.11) is only

valid for the vicinity of V∞2 , this is of course not generally true for a nonlinear
model.
 ONE-BOX MODELS 131

Attractor
Stardate 476.3 - Log of the According to Houston we will fly past 26 years later...
shipwrecked astronauts Dang and Ding... the sun and land on Mars in about 30
years. No problem, we have a base Houston, it s getting hot here! You
there. In the meantime, we can relax... didn t miscalculate the attractor
by any chance!?! HOUSTON!...

Houston, Haha!
do you copy...

Formally, this can also be seen from Eq. (6.11) since here, k > 0. This
is called an unstable fixed point. It is like a needle standing on its tip:
theoretically, there is a point where it can stand still, but in reality even
just the disturbance caused by thermal movements prevents the needle from
remaining upright.
Figure 6.2b shows the stability relations by using a topographic function
G(V), where g(V) = − dG dV . Stable fixed points correspond to a minimum,
unstable ones to a maximum of the topographic function G(V).
The stable and unstable fixed point are shown in Fig. 6.3 (cases 1 and 2).
The figure further shows four functions (cases 3 through 6) whose slope
dg 
k ≡ dV  ∞ is zero at the fixed point. Cases 3 and 4 are one-sided attractors;
Vi
we call these fixed points one-sided stable. In case 5, within a finite area
around the fixed point, the model does not move. This fixed point is called
indifferent. Finally, case 6 shows a velocity function whose first nonzero
term of the Taylor series is the third differentiation. An example of such a
curve would be g(V∞ 3
6 + ε) = αε . If α < 0 (corresponding to case 6 in the
figure), the fixed point is stable, for α > 0 it is unstable.
Let’s summarize: the slope of the change function g(V) determines the
character of the fixed point, according to the following rules:

dg 
<0 Stable fixed point V∞
dV V∞ i

 i
dg 
>0 Unstable fixed point V∞ (6.12)
dV V∞ i

 i

dg  Higher-order differentiations of
=0
dV V∞ g(V) need to be considered
i

If a model only has one attractor towards which all systems initiated in Linear models have one
finite space move, that attractor is called asymptotically stable, or global. attractor at most. It is
Since the change function of a linear model is a straight line, a linear model asymptotically stable (or
has one root at most, and thus a single fixed point. This can be stable or global).
132 NONLINEAR MODELS

g( )
∞ ∞ ∞
1 2 3

±ε ±ε ±ε

0
1 2 3

g( )
∞ ∞
4 6

±ε ±ε

∞
5
0
4 5 6

Fig. 6.3: The various fixed points of a one-dimensional model dV dt

= g(V).
The fixed points are the roots of g(V). The slope of g at the fixed point (k ≡
dg 
dV f ixedpoint
) determines the fixed point’s character: (1) k < 0: stable, (2) k > 0:
unstable, (3) and (4) k = 0: one-sided stable, (5) k = 0: indifferent, (6) first and
second differentiation at V∞ 6 are zero (see text). The arrows give the model’s
direction of movement

unstable, depending on whether the specific rate k is positive or negative. If

k < 0, the only fixed point is stable and therefore an asymptotically stable
attractor. This is the reason why linear models behave so monotonously.

6.1.3 Non-autonomous Nonlinear Models

Let us now look at non-autonomous nonlinear models and assume that we
can explicitly separate the external relation from the internal relation, as
is the case in Eq. 6.2. The fixed points of the system thus fulfill the relation

f (V∞
i ) = −R(t) for i = 1, . . . , k (6.13)

Of course, it only makes sense to analyze the system’s behavior in the

vicinity of the fixed points V∞
i if R(t) has, at least for a certain time,
a constant value R(t) = Ro . The V∞ i are thus—in the sense of a static
model—functions of Ro . That is, a set of fixed points belongs to each Ro .
This set can also be empty (meaning that there are no finite fixed points).
Let’s look at an example:

Example 6.2: Fish pond with logistic growth

In a fish pond, the fish grow according to the logistic growth function
Eq. (6.4) (see Example 6.1). Furthermore, a certain amount of fish Jr
 ONE-BOX MODELS 133

is removed per unit of time by fishing. Therefore we can describe

the pond’s fish population with the following nonlinear differential
equation:
dN N
= kp◦ N (1 − ) − Jr = f (N ) − Jr ≡ g(N ) (6.14)
dt Nmax
where f (N ) is the logistic growth curve (Eq. 6.4).

We now want to graphically determine the stability properties of this model.

We find the fixed points by setting the change function g(V) equal to zero,
which is equivalent to f (N ) = Jr . Graphically, we find the fixed points by
moving the logistic growth curve from Fig. 6.1 downwards by the amount
Jr (Fig. 6.4a). From that, and with the help of Eq. (6.12), we immediately
see that N1∞ is an unstable and N2∞ a stable fixed point.

a b Fig. 6.4: Logistic growth

and fishing in a fish pond
f (N)
f (N) (Example 6.2). We can
either determine the
fixed points from the to-
Jcr it
Jr tal velocity function g(N )
(right-hand figure) or by
intersecting the logistic
growth function f (N ) with
N1∞ N2∞ N1∞ ∞ N2∞
Jr (left-hand figure). If
Ncr it
Jr > Jcrit , there are no
finite fixed points any more
g(N) = f (N) − Jr and the fish in the pond
eventually become extinct

For the following considerations another graphical representation can

be useful: in Fig. 6.4b we intersect the logistic growth curve f (N ) with the
fishing rate Jr and in so doing find the same fixed points whose stability
behavior we already know.
What happens to the system if we treat the fishing rate Jr as an external
relation? It seems that, if the initial state is N 0 < N1∞ , the fish population
grows less rapidly than it is being fished, so the fish die out. If N 0 > N1∞
however, an equilibrium between growth and fishing sets in. Furthermore,
we see that with increasing fishing rate Jr the fixed point N1∞ grows while
N2∞ decreases. Ultimately, the two fixed points meet at Ncrit ∞
= Nmax /2,
◦
which is when the fishing rate reaches Jcrit = kp (Nmax /4). For Jr > Jcrit
the fish always become extinct.
There is one final comment to be made: Eq. (6.14) has the drawback
that N (t) can become < 0, because the fishing rate Jr is constant whether
or not there are sufficient fish left in the pond. This problem could be
134 NONLINEAR MODELS

formally solved by setting dN/dt = 0 as soon as N = 0. In Problem 6.2 we

will discuss a form of Eq. (6.14) that eliminates this problem another way.

6.1.4 Hysteresis in Nonlinear Models

Hysteresis

Hysteresis = The lagging behind

of an effect in relation to the
Doesn t even
force that it is contingent upon.
seem to have any
effect to begin
with...

The dependence of a sys- We have seen that for a constant external relation Ro , nonlinear models
tem on its history is called can have several possible fixed points V∞ i . If Ro is changed adiabatically
hysteresis. (that is, slow enough), the V∞i change along with it, too. In the process, a
phenomenon called hysteresis can occur.7
We now want to show that hysteresis always occurs when the system
has several fixed points for a given constant value of the external relation
R. As a starting point, we choose Eq. (6.2) and replace R(t) by the constant
value J. According to Eq. (6.13) the fixed points are defined by the following
relation:
f (V∞
i ) = −J (6.15)
If f (V) is a continuous function and intersects the line −J at least twice,
then it must alternately intersect −J from above and below. In the first
case, the slope of f (V) is negative and, according to Eq. (6.12), the fixed
point is stable. In the second case, the fixed point is unstable. Thus, stable
and unstable fixed points alternate.8 Figure 6.5 shows a function f (V) that
has three roots, the first one at zero. One could describe f (V) by the third-
order polynomial

f (V) = −V(V − a)(V − b); a, b > 0 (6.16)

The polynomial has a local minimum at VA (value fA < 0) as well as a

local maximum at VB (value fB > 0).9

7 One of the most well-known hysteresis phenomena is the magnetization of a

ferromagnet.
8 As a simplification, we leave cases with df = 0 at the fixed point aside.
dV
9 We leave it to the reader to express these values explicitly as a function of

the two roots a and b.

 ONE-BOX MODELS 135

a b ∞

f( ) ∞
3
b

fB B

∞ ∞ ∞ J = − fA
1 A 2 3
a ∞ a
B b 2
−J
J = − fB
fA
A

∞
1
J

Fig. 6.5: (a) Velocity function of the dynamic model dV dt

= J + f (V). For a
constant J, the fixed points are found where f (V) intersects the horizontal line −J.
• = stable fixed point, ◦ = unstable fixed point. (b) The three fixed point branches
as a function of J: thick curve = stable fixed point, thin curve = unstable fixed
point. For a growing J the model jumps at J = −fA > 0 from V∞ ∞
1 to V3 . For
declining J, the jump takes place at J = −fB < 0

First, let’s look at the case J = 0. The three fixed points V∞ i (i = 1, 2, 3)

are then obviously identical with the roots of f (V). According to Eq. (6.12),
the fixed points at 0 and b are stable, and at a, unstable. If we now let J
grow, we will find the fixed points from the intercept points of f (V) with
the horizontal line −J. In the case of Fig. 6.5a, V∞ ∞
1 and V3 wander to the
right towards larger values of V, whereas V∞ 2 moves left. If −J reaches the
value fA , the two fixed points V∞ 1 and V∞
2 fuse, and only the (stable) fixed
point V∞3 remains.
Figure 6.5b plots the position of the fixed points as a function of the
external relation J. The thick curves represent stable fixed points, the thin
one the unstable fixed point. The figure shows that for J between −fB
and −fA , three “fixed point branches” exist. If J grows beyond −fA , the
system jumps from the V∞ ∞
1 branch directly to the V3 branch. In contrast,
if J declines from large values, the system jumps at J = −fB from the
V∞ ∞
3 branch to the V1 branch. These abrupt jumps take place at different
places depending on whether J is increasing or decreasing; this is what we
call hysteresis.
We can concretise the situation shown in Fig. 6.5 by looking at the
following model10 :

dV
= J − kV + ϕ(V) = J + f (V) (6.17)
dt
10 SeeScheffer et al. (2002). The authors give several examples of systems
whose behavior can be described by Eq. (6.17). One of them forms the basis
for Example 6.3, see below.
136 NONLINEAR MODELS

where ϕ(V) is a function that, at a critical value Vcrit , jumps relatively

steeply from a very small value (≈ 0) to the value ϕ◦ (see Fig. 6.6a below).
Such behavior can be described by the function
Vp
ϕ(V) = ϕ◦ (6.18)
Vp + Vpcrit
where the exponent p determines the steepness of the jump in the vicinity
of Vcrit . The bigger p is, the more abrupt the function ϕ(V) changes from
0 to ϕ◦ .
The compound function f (V) = −kV + ϕ(V) is shown in Fig. 6.6a on
the right. The structure of this function’s roots is the same as in the case
shown in Fig. 6.5a. Again, we can intersect the curve with the line (−J) and
observe the change of the three fixed points as a function of J (Fig. 6.6b). Be-
cause the middle (increasing) branch of f (V) rises approximately vertically,
the corresponding fixed point V∞ 2 ≈ Vcrit (as far as it still exists) is practi-
cally independent of J. The (catastrophic) jumps between the stable fixed
point branches happen at approximately J = kVcrit and J = kVcrit − ϕ◦ .

Fig. 6.6: Analysis of the a

behavior of Eq. (6.17) as
a function of the external
relation J.
(a) Velocity function
f (V) = −kV + ϕ(V ).
(b) Fixed points as a func-
tion of J and hysteresis of
the system.
(c) and (d) like (a) and
(b) but with a higher spe-
cific rate k, so that hys-
teresis takes place in the b
positive range of J.
• = stable fixed point,
◦ = unstable fixed point

c d
 ONE-BOX MODELS 137

In many cases, Eq. (6.17) describes a system for which only positive
values of J and k make sense (see Example 6.3 below). As Fig. 6.6b demon-
strates, this can imply that the jump to the upper fixed point branch is an
irreversible process, because the corresponding jump back would have to
occur under a negative value of J (area marked grey in Fig. 6.6b). If, how-
ever, kVcrit > ϕ◦ (Fig. 6.6c, d), both transitions occur at positive values of
J. In this case, for J = 0 the system only has one fixed point V = 0.

Example 6.3: Water transparency in a lakea

In shallow lakes there is a fragile interaction between nutrient concen-
tration and water transparency. High nutrient concentrations favor the
growth of phytoplankton and therefore increase the particle concen-
tration and turbidity (muddiness) of the water column: the higher the
turbidity, the lower the water transparency. Low transparency leads
to the sudden disappearance of plants growing underneath the wa-
ter surface (so-called macrophytes) because they no longer get enough
sunlight.
Up to a certain nutrient concentration, the macrophytes manage to
control the plankton concentration and thus improve their growth
conditions. Numerous mechanisms are involved in this process. The
nutrient concentration in the water is reduced by the macrophytes
themselves. Their growth improves conditions for zooplankton (e.g.
Daphnia—small crustaceans), which for their part reduce suspended
phytoplankton by feeding on them. Furthermore, the presence of
macrophytes reduces the re-suspension of sediment.
If the nutrient concentration exceeds a certain critical value, these
mechanisms no longer suffice to keep turbidity low, and the
macrophytes start to die off. The result is that the elimination of phy-
toplankton ceases. To return the system to its original state, the nutri-
ent concentration must (at least temporarily) be reduced to below the
value at which the water plants disappeared. Some sort of ecological
intervention may be necessary, such as the removal of certain fish that
decimate the algae-eating Daphnia.

a Adapted from Scheffer et al. (2002).

We can use the following simple model to describe the situation described
above. We choose the concentration of suspended phytoplankton B as our
system variable. B is simultaneously a measurement for the turbidity of
the water column:
dB
= JN + f (B) (6.19)
dt
with 
−fs (B) for B ≤ Bcrit
f (B) = (6.20)
−αfs (B) for B > Bcrit
138 NONLINEAR MODELS

Phytoplankton elimination
−fs (B) − fs (B)

Bcr it
Phytoplankton/turbidity B
b

s
te
hy
Phytoplankton/turbidity B

op
f1 (B)

r
ac
tm
ou
Critical
tes

ith
hy

W
p turbidity Bcr it
ro
ac
it hm
W

Nutrient concentration JN
c
Portion of the lake surface
with macrophytes

0.4

0.3

0.2

0.1

0.05 0.10 0.15 0.20 0.25

Total phosphorus concentration [mg/L]

Fig. 6.7: (a) Elimination function of the turbidity with and without macrophytes.
Note that qualitatively, the function resembles the one in Fig. 6.6c. (b) Steady-
state turbidity as a function of nutrient concentration. (c) Measurements in Dutch
lakes show that hysteresis in the system is indeed related to the distribution of
macrophytes. The macrophytes disappear at a higher phosphorus concentration
than the one where they reappear as the concentration decreases (From Scheffer
et al. (2002), data in (c) from Meijer (2000))

The inhomogeneous term JN describes the nutrient input. In lakes, the

critical nutrient is usually phosphorus. The function fs (B) describes the
elimination of phytoplankton, with fs (B) monotonously increasing with B
(Fig. 6.7a). The constant α is the “reduction coefficient”; it describes by
 ONE-BOX MODELS 139

which factor the function fs (B) is reduced if the critical nutrient concen-
tration Bcrit is overshot and the macrophytes die off.
Figure 6.7b shows the fixed-point phytoplankton concentration (and
accordingly the turbidity) B ∞ as a function of nutrient input JN . We get
B ∞ from Eq. (6.19) by setting dB/dt = 0:

f (B ∞ ) = −JN (6.21)

The model shows hysteresis. Measurements in Dutch lakes show that

the phosphorus concentration at which macrophytes disappear is ins fact
higher than the one at which they reappear as the nutrient concentration
decreases (Fig. 6.7c).
Let’s now take another look at hysteresis from a somewhat different
angle:

Example 6.4: Nonlinear phosphorus model

A weak point of the lake models we’ve discussed so far (Examples 2.2,
4.9, 4.10–4.12) is the inadequate description of phosphorus deposition
into the sediment. So far, we described it as a linear function of the
average phosphorus concentration (ks C). Yet observations show that
in most lakes, the specific phosphorus sedimentation rate ks decreases
with growing concentration and can even become negative. Such a neg-
ative rate would indicate re-dissolution of phosphorus from the sedi-
ment into the water.
Once again, we describe the lake as a linear flow reactor (see Exam-
ple 4.4), but introduce a concentration-dependent specific phosphorus
sedimentation rate ks (C). Compared to Eq. (4.22), the dynamic equa-
tion for the average phosphorus concentration is slightly changed:
dC
= kw Cin − kw C − ks (C)C (6.22)
dt
This equation is no longer linear, because ks is a function of C. To
model the behavior of ks (C) as simply as possible, we choose the fol-
lowing function:

ks for C < Ccrit
ks (C) = (6.23)
0 for C ≥ Ccrit

The function ks (C) is plotted in Fig. 6.8. In reality, ks (C) would de-
crease continuously, rather than abruptly as in our model. Eq. (6.23)
has the advantage that the model can be analyzed relatively easily
while retaining all important properties of a realistic model.
140 NONLINEAR MODELS

Fig. 6.8: Specific sedimen- ks (C)

tation rate ks as a function
of the average phospho- ks
rus concentration C in
the lake: if the phospho-
rus concentration is higher
than Ccrit , ks becomes zero C
and no more phosphorus is Ccr it
added to the sediment

If we insert Eq. (6.23) into Eq. (6.22), the following model equation re-
sults:

dC ks C for C < Ccrit
= kw Cin − kw C − (6.24)
dt 0 for C ≥ Ccrit

This is actually a superposition of two linear models. The current value

of C determines which of the two linear equations is valid. If, in a given
case, the concentration C never were to exceed the critical value Ccrit , the
model would not “know” about its double life, and only behave according
to Eq. (4.22).11
Remember that Eq. (6.24) is non-autonomous due to its input term
(kw Cin ). In general, that is, if Cin is not constant in time, the model will
not reach a fixed point. Yet, as in previous examples, it is useful to analyze
the hypothetical fixed point as a function of an assumed constant input
concentration Cin . In other words, we are looking for the static model,
C ∞ = f (Cin ), which the dynamic model (Eq. 6.24) implicitly contains.
Much like the dynamic model is a superposition of two linear models, the
static model is a superposition of two fixed-point lines:
⎧
∞
⎨ Cin kwk+k
w
s
≡ CA for C ∞ < Ccrit
∞
C (Cin ) = (6.25)
⎩ ∞
Cin ≡ CB for C ∞ ≥ Ccrit

The fixed-point lines are plotted in Fig. 6.9. Line A (C ∞ < Ccrit ) has a
slope of kwk+k
w
s
, line B a slope of 1. The horizontal line at Ccrit delimits the
scope of the two fixed-point lines. We note that for an input concentration
Cin between Ccrit and C  = kwk+k w
s
Ccrit > Ccrit , both fixed-point lines are
technically valid. How does the model really behave?
Let’s imagine that initially, Cin is so small that the corresponding fixed-
∞
point concentration is CA < Ccrit . If Cin now grows so slowly that the

11 Webrought up the phenomenon of linearity within a limited area in the in-

troduction to Chap. 6.
 ONE-BOX MODELS 141

C∞ B Fig. 6.9: Relation between

the input concentration
CB∞ = Cin Cin and the two fixed-point
lines (Eq. 6.25). If Cin
A lies between Ccrit and C 
(shaded area in the graph),
two fixed points exist. The
CA∞ = kw
C
kw +ks in model’s previous history
Ccr it then decides which fixed
point is actually reached:
again, we see hysteresis

Cin
kw +ks
Ccr it C = kw
Ccr it

model adiabatically moves along in its fixed point,12 then, at an input

∞
concentration of C  , CA reaches the critical value Ccrit :

∞ kw kw kw + ks
CA (C  ) = C = Ccrit = Ccrit (6.26)
kw + ks kw + ks kw

At Ccrit , the model now jumps to the upper fixed-point line B. If, on the
other hand, the model moves downwards on B coming from large values of
Cin , it won’t jump back to A at Cin = C  but only at Cin = Ccrit .
To summarize: despite its simplicity, the model (6.24) has allowed us to
see two phenomena that are typical for nonlinear models (and do not occur
in linear models):

1. The model has critical states in which even the smallest of changes
in the input (external relation) lead to large changes in the behavior
of the system variable.

2. There is an interval of the input variable Cin in which the behavior

of the system depends on its prior history (hysteresis).

6.1.5 Synergism in Nonlinear Models

Apart from the occurrence of hysteresis, nonlinear models have another
property which fundamentally differentiates them from linear models: syn-
ergistic effects. They are the phenomena by which several forces influencing
a system from the outside can influence each other either by reinforcement
or compensation.

12 Question to the reader: How should we specify the attribute slow here?
142 NONLINEAR MODELS

Synergism
...action!

1, 2, and...

Grr.. a nice
synergism!

Synergism: interaction be- To understand more clearly what we mean by “influence each other”
tween different external we remember those systems where a mutual influence doesn’t exist: linear
forces, so that the over- systems. In Sect. 4.3 we looked at a linear model with time-dependent input
all influence is smaller or and saw that the input can be divided into many individual input chunks—
larger than the sum of the the system “processes” each input as if the others didn’t exist (see Fig. 4.5).
individual influences. The combination of all input events is obtained from the sum (superposi-
tion) of all individual events. In general, this method cannot be applied to
nonlinear models.
To illustrate this, we can think of the following question: what is the in-
fluence of water polluted with zinc and mercury on the growth of a specific
algae species? We can assume that each of the two heavy metal concentra-
tions on its own would reduce the algae growth rate by 10 %. If the effects
on algae growth of the two heavy metals were independent of each other,
the overall effect would be a growth rate reduction of 20 %. This is the
case without synergism. In reality, however, we will often observe that the
effects of two pollutants mutually reinforce each other, so the reduction
of growth rate might be more than 20 %. Some substances may also com-
pensate each other’s effects (think poison and antidote). In both cases we
speak of synergism.

Example 6.5: Synergism in the nonlinear phosphorus model

In Fig. 6.10, we look at the influence of two input events on the behav-
ior of phosphorus concentration in a lake with nonlinear sedimentation
(Eq. 6.24). If the sum of the two events leads to a phosphorus concen-
tration greater than the critical value Ccrit , the phosphorus concen-
tration’s behavior is no longer identical with the sum of the individual
events.
 MULTIDIMENSIONAL MODELS 143

C(t) Fig. 6.10: In the phospho-

rus model from Eq. (6.24),
e −k 2 t the actual concentration
development (black contin-
Ccr it
e −k 1 t uous line) is not identical
with the sum of all individ-
e −k 1 t
e −k 1 t ual events (dashed line)

Time t
t1 t2

Remember: synergistic effects can only be described by nonlinear mod-

els. We thus begin to see how in many ways, the possibilities of linear
models are too limited to adequately describe real-world phenomena.

6.2 Nonlinear Box Models with Several System Variables

6.2.1 The Jacobian Matrix
Multi-dimensional nonlinear models greatly increase the potential of math-
ematical modeling. The diversity gained, however, comes at the cost of only
being able to analytically solve the systems of differential equations for a
limited number of special cases. Most nonlinear models have to be solved
numerically. In contrast to analytical solutions, computer simulations have
the drawback that certain characteristic properties of the model cannot
be analyzed (or only with a large number of simulations). One of these
properties is a system’s behavior in the vicinity of a fixed point.
We will now look at that behavior, and in the process learn some tricks
that let us, to an extent, skirt around numerical solutions. Let’s consider
an n-dimensional model that is described by the following system of n
first-order differential equations:
dVi
= gi (V1 , . . . , Vn ), i = {1, . . . , n} (6.27)
dt
As in the one-dimensional case (see Eq. 6.6), we assume that all exist-
ing external relations are constant in time and thus integrated into the
functions gi .
The fixed points of the system of differential equations (6.27) are the q
solutions of the n-dimensional system of regular but nonlinear equations13 :

0 = gi Vk1 , Vk2 , . . . , Vkn , k = {1, 2, . . . , q}, i = {1, 2, . . . , n} (6.28)

13 Note that here we slightly modify the notation introduced in Eq. (6.7). We
leave out the symbol ∞, the numbering of the fixed point k becomes the upper
index, and the lower index indicates the variable.
144 NONLINEAR MODELS

If the system were linear, it would have at most q = n solutions. For sys-
tems of nonlinear equations, the number of fixed points q can be arbitrarily
large.
To examine the system’s behavior around a fixed point k, we can build
on the method that we used for one-dimensional models in Sect. 6.1.2. Be-
cause it is difficult to imagine surfaces in more than three dimensions, we
first look at a two-dimensional system with the variables V1 and V2 . We
can represent each of the two velocity functions gi (i = 1, 2) as a surface
in three-dimensional space. Two of the dimensions are spanned by the two
variables themselves. This is the phase space as we have already encoun-
tered it, or more precisely, the phase plane. Along the third axis, standing
perpendicular to this plane, we plot the value of g1 or g2 . This way we get
two topographical maps with mountains and valleys. Where both of these
topographies intersect the phase plane there is a fixed point. Around this
fixed point, we can approximate the development of g1 and g2 with tan-
gents in the direction of V1 or V2 , just as we did in Eq. (6.9) (or Fig. 6.2).
∂g1 ∂g1 ∂g2 ∂g2
In total we get four tangents, with the slopes ∂V 1
, ∂V2
, ∂V1
and ∂V2
. The
symbol for the partial derivative (∂) indicates that the slope is being deter-
mined along the chosen direction (V1 or V2 ) and that the other coordinate
is held constant. Applying this principle to more than two dimensions, we
get n2 such slopes. Below, we will explicitly plot these velocity surfaces for
Example 6.6 (Fig. 6.13).
Somewhat more formally, we assume that the system is arbitrarily close
to the fixed point k, so:

Vi = Vki + εi , i = {1, . . . , n} (6.29)

or, written as vectors:

V = Vk + ε (6.30)

Because by definition, at the fixed point itself all change functions gi are
zero, we can approximate them in the vicinity of Vk with the following
Taylor series, aborting after the first-order derivatives:
 
n
∂gi 
gi (V) = gi (Vk + ε) =  · εj i = {1, . . . , n} (6.31)
∂Vj 
j=1 Vk

The vertical line with the argument Vk tells us to calculate the partial
derivatives at the fixed point Vk . Because Vk is constant in time, dVi /dt =
dεi /dt. Thus, as an approximation, Eq. (6.27) becomes an n-dimensional
linear system of the deviation variables εi :

n
dεi
= Bi,j (Vk )εj i = {1, . . . , n} (6.32)
dt j=1
 MULTIDIMENSIONAL MODELS 145

with
 
k ∂gi 
Bi,j (V ) = (6.33)
∂Vj Vk

The Bi,j (Vk ) can be written as a matrix:

⎛ ⎞
∂g1
∂V1
∂g1
∂V2 . . . 
⎜ ∂g2 ∂g2
... ⎟ 
B(Vk ) = ⎝ ∂V1 ∂V2 ⎠ (6.34)
.. .. 
. 
..
. . Vk

where the argument Vk indicates that the derivatives of the velocity func-
tions gi in the matrix are to be calculated at the selected fixed point k. Finally something
useful to come
B is the so-called Jacobian matrix. It describes the system of nonlinear out of maths,
equations in an approximated linearized manner in the vicinity of the fixed this Jacobian
mattress!
points. Written in matrix form, Eq. 6.32 is:

dε
= B(Vk )ε (6.35)
dt

The expressions (6.32) and (6.35) are thus linear approximations of the
nonlinear system (6.27). By way of the eigenvalues of the Jacobian matrix
we can now analyze the behavior of the model around the fixed point. We
must keep in mind however that under some circumstances such a stability
analysis may only be valid in very small ε surroundings of Vk . Furthermore,
there are nonlinear models whose velocity functions gi are not differentiable,
so that no Jacobian matrix can be calculated. Finally, there are cases where
the linearized model cannot make any statement about the real model (see
Sect. 6.2.4).

6.2.2 Characterizing the Fixed Points of Multi-dimensional Systems

Again, for simplification, we consider a two-dimensional model. In contrast
to the one-dimensional case the fixed points Vk don’t divide the phase space
into invariant areas. It is also possible that the model does not even reach
a fixed point, but rather oscillates around it, as we have already seen in the
example of the harmonic oscillator (Example 5.11).
In Sect. 5.1.1 we analyzed the properties of the eigenvalues of a two-
dimensional linear model and summarized the results (Table 5.1). In the
vicinity of their fixed points, nonlinear two-dimensional models can be ap-
proximated by linear models. The Jacobian matrix thereby takes over the
role of the coefficient matrix P which we introduced in Eq. (5.3). At this
point we want to summarize the classification of the fixed points (as de-
tailed in Table 5.1) once again; at the same time, in Fig. 6.11, we want to
visualize the trajectories in the two-dimensional phase space near a fixed
point. Each pattern has a specific pair of eigenvalues λ1 and λ2 of the
Jacobian matrix B associated with it.
146 NONLINEAR MODELS

a) b)

c) d)

e) f)

Fig. 6.11: Characterization of the behavior of two-dimensional models in their

phase space near a fixed point Vk . V1 und V2 are the system variables. (a) Stable
star, (b) unstable star, (c) saddle point (unstable), (d) stable fixed point with
oscillation, (e) unstable fixed point with oscillation, (f) undamped oscillation:
center. See text for further explanations
 MULTIDIMENSIONAL MODELS 147

Real eigenvalues

(a) Both eigenvalues are real and negative (λ1 , λ2 < 0): stable fixed point,
also called stable star (Fig. 6.11a).

(b) Both eigenvalues are real and positive (λ1 , λ2 > 0): unstable fixed
point, also called unstable star (Fig. 6.11b).

(c) Eigenvalues are real and have different signs, e.g. (λ1 > 0, λ2 < 0):
Unstable fixed point, also called saddle point (Fig. 6.11c).

It is easily possible to extend the definitions above to one-sided stable and

indifferent situations, that is, to include cases with an eigenvalue of zero.

Non-real eigenvalues
If the eigenvalues are non-real, λ1 and λ2 only differ in the sign of their
imaginary parts.14 The system variables oscillate around the fixed point in
three possible ways:

(d) Re(λ1 , λ2 ) < 0: The model moves on a spiral towards the fixed point
(Fig. 6.11d); the fixed point is stable.

(e) Re(λ1 , λ2 ) > 0: The model moves away from the fixed point on a
spiral (Fig. 6.11e); the fixed point is unstable.

(f) Re(λ1 , λ2 ) = 0, i.e. purely imaginary eigenvalues. The model moves

on a closed curve around the fixed point without ever reaching it
(Fig. 6.11f). The fixed point is called a center.

It can be shown that, except for case f (center), examining the fixed points
by linearization does indeed yield an accurate description of the nonlinear
system.15 This holds at least in a (possibly very small) area around the
fixed point. In the case of a center (case f), we cannot decide whether If the fixed point is a
the behavior of the real (nonlinear) model follows categories d, e or f. We center, the system’s be-
will illustrate this with Example 6.9. Furthermore, there is still another havior in its vicinity
possibility: the convergence of the trajectory towards a closed curve around cannot be determined by
the fixed point. This case is called a limit cycle and we will illustrate it in linearization.
Example 6.8 below.
For nonlinear models with more than two dimensions, the eigenval-
ues of the Jacobian matrix are still useful to analyze the behavior near
the fixed point. If n = 4, for example, the eigenvalues can be calculated
from the solution of a fourth-order equation. The more dimensions a model
has, the more diverse the possible combinations of real and non-real parts
of the eigenvalues become. Thus the diversity of the behavior of fixed
points also rises. In certain cases, e.g. if all solutions are real, we can still
14 Because they are the solutions of a system of quadratic equations, non-real
eigenvalues only occur as conjugate-complex pairs (see Eq. 5.10).
15 As a reminder: the calculation of the Jacobian matrix presupposes that the

velocity functions gi are differentiable at the fixed point.

148 NONLINEAR MODELS

make qualitative statements relatively easily. A more general discussion of

n-dimensional models goes beyond the scope of this book, however.
In the following, we will concentrate on two-dimensional nonlinear mod-
els and, to conclude the chapter, introduce a special three-dimensional
model.

6.2.3 Predator-Prey Model

 Wisdom of the hare
An old hare is lecturing a young one. “It’s time you learnt about the
predator-prey model.” — “Why is that important?” — “It’s an ancient wis-
dom that gives hope to us hares. It foretells that there are lynxes; they eat
hares. But as more and more hares are eaten, the lynxes start starving
and their numbers dwindle. Thus the population of hares increases again,
and the circle of life begins anew. . . it means that we hares can never die
out completely!” — “But. . . I thought that there have been no lynxes here
for a long time. The humans eradicated them all!” — “Well, yes, but one
day there could be lynxes again. So mark my words!” The little hare thinks
for a while and then smiles: “Ok, I think I understand. The people hunt us,
so there are less of us, and in turn, less people too, so then there are
more of us again. . . ” — “Each model has its limits. . . ” the old hare wants
to continue his lecture, but the young one has already run off to spread
the news to his friends.

We begin our discussion of two-dimensional models with the predator-prey

model by Lotka and Volterra.

Example 6.6: Predator-prey model by Lotka and Volterra

One of the most well-known models from mathematical ecology is
the predator-prey model, developed almost simultaneously by Albert
Lotka (1924) and Vito Volterra (1926). In its original form, it describes
a theory of competition between two species. Applied to an interac-
tion between a predator and its prey, we can reduce the model to the
following assumptions:

1. A population of prey animals X increases with the specific net

rate k1 .
2. A population of predators Y dies with the specific net rate k2 .
 MULTIDIMENSIONAL MODELS 149

3. The prey is eaten by the predators, which results in an increase

of predators and a corresponding decrease of prey by k3 XY .a

Figure 6.12 graphically shows the model. Mathematically, it is de-

scribed by two coupled differential equations:
dX
= k1 X − k3 XY
dt (6.36)
dY
= −k2 Y + k3 XY
dt
a Of course, the assumption that one eaten prey animal results in the
“creation” of one additional predator, as suggested by Eq. (6.36), is not
particularly realistic. The model could be improved by replacing k3 with
k3∗ = αk3 (α < 1) in the second Eq. (6.36). Although this changes
Eqs. (6.37), (6.40) and (6.41), it does not change the eigenvalues of the
linearized system or the fundamental structure of the system of equa-
tions (6.36).

k1 X k3 XY
Fig. 6.12: Model by Lotka
k2 Y
Prey X Predator Y and Volterra for the inter-
action between predator Y
and prey X

Because of the two terms ±k3 XY , the model is nonlinear. It has two
fixed points, first, the trivial fixed point X1∞ = 0, Y1∞ = 0, and second:
k2 k1
X2∞ = , Y2∞ = (6.37)
k3 k3
We get this result by setting the left sides of Eqs. (6.36) equal to zero.
To determine the stability of the fixed points, we will analyze the model’s
behavior in the vicinity of the second fixed point by introducing new vari-
ables. This is somewhat more laborious than the method with the Jacobian
matrix, but also easier to understand. We define δX and δY as the deviation
of the population sizes from the fixed point:
δX = X − X2∞ , δY = Y − Y2∞ (6.38)
Since X2∞ , Y2∞ are constant, it follows that:
dδX dX dδY dY
= and = (6.39)
dt dt dt dt
In Eqs. (6.36), X and Y can be replaced by the new variables δX and δY .
After some algebraic conversions, we get:
dδX
= −k2 δY − k3 δY δX
dt (6.40)
dδY
= k1 δX + k3 δY δX
dt
150 NONLINEAR MODELS

Fig. 6.13: Velocity func-

tions gX (X, Y ) and
gY (X, Y ) of the Lotka-
Volterra model (6.36).
The nontrivial fixed point
(X2∞ , Y2∞ ) is located where
the two plains intersect
the zero plain. The slopes
of the plains at this point
along the X and Y axes
correspond to the ele-
ments of the Jacobian
matrix (Eq. 6.34). The
model parameters are as
in Fig. 6.14

The method of linearization implies that we allow only small deviations

from the fixed point and disregard products of small quantities, δX δY .
This results in:
dδX dδY
= −k2 δY , = k1 δX (6.41)
dt dt
Of course, the resulting system of linear differential equations (6.41) could
also have been produced by the method of the Jacobian matrix (see
Problem 6.3). Figure 6.13 shows the two velocity functions gX (X, Y ) and
gY (X, Y ). The fixed point is located where the two planes intersect the
zero plain.
Maybe you noticed that we are already acquainted with Eq. 6.41 from
Example 5.9: we only have to replace δY with y1 and δX with y2 . This way
we know without further calculations that the eigenvalues are purely imag-
inary (λ1 = ±i(k1 k2 )1/2 , see Eq. 5.58). The nonlinear fixed point (6.37)
is therefore a center. In Chap. 5 we showed that the undamped harmonic
oscillator (Example 5.11) can also be described by this type of equation.
The solutions are undamped harmonic oscillations of the pendulum’s dis-
placement and speed, with the speed lagging behind the displacement by
a quarter period. Likewise, Fig. 5.10 could be carried over to the predator-
prey model: y1 could be interpreted as the number of predators and y2 as
the number of prey.
Before we reach any premature conclusions, we need to remember that
in the case of the harmonic oscillator the system of equations itself had
the form of Eq. (5.55) (or 6.41), whereas in the case of the Lotka-Volterra
model, we got Eq. (6.41) only as a linearized approximation of the original
system Eq. (6.36). Furthermore, the fixed point is a center. In Sect. 6.2.2
we noted that for a center, the method of linearizing (the Jacobian matrix
method) says nothing about how the nonlinear model actually behaves at
the fixed point. In Example 6.8 we will show that for a center, the real (i.e.
nonlinear) model can behave according to one of three behavioral patterns:
it either moves towards the fixed point (the fixed point is an attractor),
 MULTIDIMENSIONAL MODELS 151

250 250
Y∞
2
Prey X, Predator Y
200
X(t) 200

Prey X
150 Y(t) 150
X∞
100 X∞
2
2 100

Y∞
50 2 50

0 0
0 10 20 30 40 0 50 100 150 200
Time t [yr] Predator Y

Fig. 6.14: Numerical solutions of the Lotka-Volterra predator-prey model. On

the left, temporal behavior of the two system variables X, Y . On the right, the
phase space with fixed point at (X2∞ , Y2∞ ). Parameters used: k1 = 0.5 year−1 ,
k2 = 0.8 year−1 , k3 = 0.008 year−1 . Initial condition: X 0 = 50, Y 0 = 30. Non-
trivial fixed point at X2∞ = 100, Y2∞ = 62.5

moves away from it, or it circles it on a closed trajectory like the undamped
harmonic oscillator (see Fig. 5.10).
It can be shown that the Lotka-Volterra model belongs to the third
category (see e.g. Arrowsmith and Place 1992). In Fig. 6.14, X and Y are
shown both as functions of time and in the two-dimensional phase space.
The plot shows that the trajectory describes a closed curve, but not an
ellipse as in Fig. 5.10. Correspondingly, X(t) and Y (t) are also not pure sine
or cosine functions. This non-harmonic (yet undamped) behavior reminds
us that the real system of Eqs. (6.36) is nonlinear and that Eq. (6.41) is
merely a linear approximation.
The kinship between the Lotka-Volterra model and the model of the
harmonic oscillator suggests that they also share the feature of structural
instability. The following example demonstrates that this is indeed the case.

Predator
Example 6.7: Predator-prey model with self-interaction with self-
interaction!
We will now make a minor change to the predator-prey model from
Example 6.6 by assuming that for large prey populations, the prey
animals themselves impede and potentially even kill or eat one another.
We can describe the resulting negative effect on the prey population
dynamics in Eq. (6.36) with an additional term of the form (−k4 X 2 ). k4
is an arbitrarily small coefficient of self-interaction. Note that because
the variable X appears squared, the new term is only relevant for large
values of X. The modified system equations are:
dX
= k1 X − k3 XY − k4 X 2
dt (6.42)
dY
= −k2 Y + k3 XY
dt
152 NONLINEAR MODELS

The system still has two fixed points. The trivial fixed point (X1∞ =
Y1∞ = 0) remains unchanged. The nontrivial fixed point however is slightly
changed compared to Eq. (6.37):
k2 k1  k2
X2∞ = , Y2∞ = − ; with  = k4 (6.43)
k3 k3 k3 k3
Problem 6.4 asks you to calculate the Jacobian matrix of Eq. (6.42) and
determine the eigenvalues at the nontrivial fixed point. From the results
of that calculation, the deviation of X and Y from the nontrivial fixed
point is described by the following approximative system of linear equa-
tions:
dδX
= −k2 δY − δX
dt (6.44)
dδY
= (k1 − )δX
dt
We assume that the prey’s self-interaction is arbitrarily small but nonzero.
In particular,  k1 holds. The system (6.44) thus essentially differs from
the original equations (6.41) only by the additional term −δX in the first
equation (just like the damped pendulum, Eq. 5.63). We already know what
this means for the eigenvalues (see Eq. 5.67)—they receive a (very small)
negative real part:
 
λi ≈ − ± iω, ω = k1 k2 (6.45)
2
With that, the fixed point (X2∞ , Y2∞ ) becomes stable, that is, an at-
tractor (Fig. 6.11d). The oscillation of the predator and prey populations
is damped around the fixed point. The system eventually assumes the con-
stant values X2∞ and Y2∞ . Furthermore, because the fixed point is no longer
a center (as it was in our first predator-prey model), we know now that the
nonlinear system qualitatively behaves like its linearized counterpart. Fig-
ure 6.15 shows the dynamic behavior of the modified predator-prey model.
The Lotka-Volterra model The Lotka-Volterra model profoundly influenced the thinking of the first
profoundly influenced the ecological modelers. The Hudson Bay Company’s 100-year catch statistics
beginning of ecological sys- of the Canada lynx and its prey, the snowshoe hare, were often cited as
tems modeling. a textbook example of the periodic oscillations of a predator-prey pair as
predicted by the model (Fig. 6.16). The catch statistics show the population
density of both species. It exhibits a regular fluctuation with a period of
about 10 years. With some goodwill, we can see that the maximum hare
population does indeed occur just before the maximum of the lynx—as
predicted by the Lotka-Volterra model.
This apparent confirmation of the Lotka-Volterra model does not stand
up to critical examination, however. First, we know from our mathematical
considerations that the model is structurally unstable. Thus there are many
reasons why the fluctuations should disappear. Second, the snowshoe hare
population also fluctuates in areas where there are no lynxes. And finally,
the lynx does not subsist exclusively on a snowshoe hare diet.
 MULTIDIMENSIONAL MODELS 153

200 200

Prey X, Predator Y X(t)

150 150
¥

Prey X
Y(t) X2
¥
100 X2 100

¥
50 Y2 50

Y2
0 0
0 10 20 30 40 0 50 100 150 200
Time t [yr] Predator Y

Fig. 6.15: Modified predator-prey model: compared to Fig. 6.14, the dynamic prey
equation has an additional prey-prey self-interaction term (Eq. 6.42). Therefore
the oscillations of the predator and prey populations Y and X are damped and
the system moves towards the nontrivial fixed point. The position of the fixed
point itself and the period of the oscillations change only very slightly for a small
self-interaction. Parameters and initial condition are like in Fig. 6.14, and k4 =
5 × 10−4 a−1

160

140 Hare
Lynx
120
Number in thousand

100

80

60

40

20

1845 1855 1865 1875 1885 1895 1905 1915 1925 1935

Fig. 6.16: Population fluctuations of snowshoe hare and lynx, from the catch
statistics of the Hudson Bay Company. This is a classic example of cyclical os-
cillation of population densities (Source: MacLulich 1937 cited in Odum 1971)

In short, this example demonstrates that building a model for measured

data can stimulate the development of an interesting new theory even if it
turns out not to be that relevant for the original data. Indeed, the Lotka-
Volterra model could be credited as the origin of mathematical ecology.
Thus it doesn’t surprise that many further models emerged based on it,
attempting to remedy the deficiencies of their predecessor. One of these is
the Holling-Tanner model (May 1974, p. 84).
154 NONLINEAR MODELS

Example 6.8: Holling-Tanner predator-prey model

In the Holling-Tanner model, the interaction between prey X and
predator Y is expressed by the following system of nonlinear differ-
ential equations:
dX X XY
= r(1 − )X − w (6.46)
dt Xk X + KX
dY Y
= s(1 − J )Y (6.47)
dt X
with r, s, w, J, Xk , KX > 0

In Eq. (6.46), the first term describes the logistic growth of the prey
(see Example 6.1). Xk is the hypothetical equilibrium population of X
if there were no predators. The second term of Eq. (6.46) simulates the
decimation of prey by the predators. In contrast to the Lotka-Volterra
model, the predators have a limited appetite: for X KX , it reaches
the value −wY which is independent of X. Eq. (6.47) describes the
growth of the predator population. It is only positive if JY /X < 1, i.e.
if Y < X/J. In other words: the predator population can only grow if
at least J prey animals are available per predator.

For a detailed discussion of the properties of this model, we refer to Arrow-

smith and Place (1992). For us, the important point is that the solutions
of the Holling-Tanner model also show periodic oscillations of X and Y
around an equilibrium value. However, in contrast to the Lotka-Volterra
model, it is structurally stable. Apart from the trivial (Y = X = 0) and
semi-trivial fixed point (Y = 0, X = Xk ), the model has the nontrivial
fixed point (X ∞ , Y ∞ ) that results from the intersection of a parabola and
a line (Fig. 6.18).16 The parabola results from Eq. (6.46) with dX/dt = 0:
r X∞
Y∞ = (1 − )(X ∞ + KX ) (6.48)
w Xk
The line results from Eq. (6.47) with dY /dt = 0:
X∞
Y∞ = (6.49)
J
In the phase space, the system’s trajectory eventually moves to a closed
curve around the fixed point (X ∞ , Y ∞ ). In contrast to the Lotka-Volterra
model, this closed cycle is not fixed by the initial state once and for all.
Rather, the trajectories move towards an “attraction curve” and then follow
it (Fig. 6.17). Such a curve is called a limit cycle. A limit cycle is thus a
“curve attractor” to which the system is attracted. Figure 6.18 shows the
tuning of the periodic oscillations of X(t) and Y (t) to the limit cycle.

16 Thestability properties of the semi-trivial fixed point will be discussed in

Problem 6.5.
 MULTIDIMENSIONAL MODELS 155

250 250
Fig. 6.17: Numerical
Y¥
Prey X, Predator Y
200 200
solution of the Holling-
X(t) Tanner model. The sys-
150 150 tem moves towards the

Prey X
limit cycle. Parameters
X¥
100 X¥ 100 used: r = 2.5 year−1 , s =
Y(t) 0.225 year−1 , w = 5 year−1 ,
50 Y¥ 50 Xk = 300, KX = 50,
J=2. Initial conditions:
0
0 10 20 30 40 50 20 30 40 50 60
0 X 0 = 50, Y 0 = 60
Time t [yr] Predator Y

Fig. 6.18: Determina-

Y = X/J
tion of the nontrivial fixed
60
Limit cycle point (X ∞ , T ∞ ) from two
curves (Eqs. 6.48 and
6.49) and calculation of
two trajectories. The first
Predator Y

50 (X 0 = 50, Y 0 = 50) ap-

(X¥,Y¥) proaches the limit cycle
from its outside, the second
(X 0 = 105, Y 0 = 45) from
the inside. Model parame-
40 ters are as in Fig. 6.17

Y = r/w (1−X/XK) (X+KX)

30
40 80 120 160 200
Prey X

6.2.4 The Behavior of Models in the Vicinity of Centers

As we already mentioned earlier, in the vicinity of a center, the behavior

of a nonlinear system cannot be determined with the linearization method.
The following example (from Arrowsmith and Place 1992) demonstrates
this.
156 NONLINEAR MODELS

Example 6.9: Two nonlinear models with center

We examine the following two linear models (x, y are the system
variables).

Model 1:
dx
= −y + x(x2 + y 2 )
dt (6.50)
dy
= x + y(x2 + y 2 )
dt
Model 2:
dx
= −y − x(x2 + y 2 )
dt (6.51)
dy
= x − y(x2 + y 2 )
dt

The only difference between the two models is the sign of the second term
on the right-hand side of the equations. Both models have the fixed point
x = y = 0. The Jacobian matrices at the fixed point are identical:
 
0 −1
B= (6.52)
1 0

We already know that the eigenvalues are purely imaginary (λi = ±i).
The fixed points of the linearized systems are centers. But how do the
nonlinearized systems really behave in the proximity of the fixed point? To
answer this question, we introduce two new variables r and θ to replace x
and y:  y
r = x2 + y 2 , θ = arctan (6.53)
x
Of course, we didn’t choose these particular notations for r and θ by ac-
cident: they are the polar coordinates in the original phase space (x, y).
Now we have to formulate the new dynamic equations, which requires
Center
Test series #1 complete:
Center is an attractor.
 MULTIDIMENSIONAL MODELS 157

some effort, but nothing more than the chain rule of differentiation. We
demonstrate the procedure with model 117 :

dr dr dx dr dy
= +
dt dx dt dy dt
x dx y dy
= +
x2 + y 2 dt x2 + y 2 dt
−xy  xy 
= + x2 x2 + y 2 +  + y 2 x2 + y 2
2
x +y 2 2
x +y 2
3
= (x2 + y 2 ) 2 = r3
(6.54)

dθ dθ dx dθ dy
= +
dt dx dt dy dt
−y/x2 1/x
= 2
[−y + x(x2 + y 2 ] + [x + y(x2 + y 2 ]
1 + (x/y) 1 + (x/y)2
y2 x2
= 2 − xy + + xy = 1
x + y2 x2 + y 2

For model 2, the result is almost identical, except that dr/dt has an inverted
sign:
dr dθ
= −r3 ; =1 (6.55)
dt dt
The new variables r and θ have decoupled the two systems of differential
equations (6.50) and (6.51). The arc function θ(t) has the trivial form:

θ(t) = θ0 + t (6.56)

This means that in both cases, the trajectory turns with constant angu-
lar velocity in a (mathematically) positive direction of rotation (which is
counter-clockwise) around the coordinate origin. The difference between
the two models is that in the case of Eq. (6.50) (model 1) the radius grows
with time (outward spiral), whereas in the second case, the radius decreases
(inward spiral) until the system comes to a rest in the fixed point x = y = 0.
The two solutions are shown in Fig. 6.19 together with the trajectory of the
undamped linear oscillator (r = const). Note that the third graph, the lin-
ear approximation of both nonlinear models, is not capable of making a
distinction between them.

6.2.5 Nonlinear Models with Three and More Variables

In nonlinear systems of differential equations with more than two vari-
ables, a new phenomenon can occur: deterministic chaos. A system becomes

17 Remember that d
dz
(arctan z) = 1
1+z 2
.
158 NONLINEAR MODELS

a) Model 1 b) Model 2 c) Linear approximation

Fig. 6.19: The behavior
of a system near a cen-
ter cannot be predicted
from the Jacobian ma-
trix. Model 1 (Eq. 6.50)
moves away from the cen-
ter at x = y = 0, model 2
(Eq. 6.51) moves towards
it. The linear approxima-
tion is a circular trajectory
around the center

chaotic if two trajectories starting from arbitrarily close initial points sud-
denly develop completely differently. Often, chaotic systems remain close
to an attractor for a certain timespan, only to suddenly jump into a com-
pletely different course that unfolds around another attractor. Such at-
tractors are called strange attractors. The chaos is called deterministic (as
opposed to stochastic) because the underlying equations are entirely de-
terministic, i.e., they contain no random quantities. Because in reality, a
system’s state can never be measured to perfect precision (no measurement
is without measurement error), there are always tiny differences between
apparently identical states. These differences can be responsible for the
diverging development from two seemingly equivalent initial states.
A well-founded discussion of deterministically chaotic systems would go
beyond the scope of this book. We merely want to discuss one example:
the Lorenz system. In order for deterministic chaos to occur, the system of
differential equations has to be at least three-dimensional.18

Example 6.10: Lorenz model and deterministic chaos

In 1963, the atmospheric physicist Lorenz established a simple model
to describe the coupling between heat convection and heat transport
in a fluid (water, air, etc). He used three variables (x, y, z) with the
following meaning:
x Circulation speed of the convection source
y Temperature difference between the upwards and downwards
flowing fluid
z Deviation of the vertical temperature gradient from
the equilibrium

18 For difference models, one dimension is sufficient (see Chap. 7).

 MULTIDIMENSIONAL MODELS 159

40 40

35 35

30 30

z 25 25 z

20 20

15 15

10 10
−20 −10 0 10 20 −20 −10 0 10 20
x y

Fig. 6.20: Solution of the Lorenz model (Example 6.10) with the parameters
a = 10, b = 28 and c = 8/3. Initial conditions (x0 , y 0 , z 0 ) = (−11.3, −13.3, 28.0).
Attractor at (±8.4853, ±8.4853, 27.0000). On the left, trajectories in three-
dimensional phase space are projected onto the (x, z) plane, on the right, onto
the (y, z) plane

The system of differential equations has the following form:

dx
= a(y − x)
dt
dy
= bx − y − xz (6.57)
dt
dz
= xy − cz
dt
a, b, c are positive constants with b > 1. The system has three fixed
points:

(1) x = y = z = 0

(2) x = y = (b − 1)c, z = b − 1

(3) x = y = − (b − 1)c, z = b − 1

For certain parameter values a, b, c, the system seems to jump ran-

domly between trajectories around the second and third fixed point.
The resulting form has been called the “Lorenz butterfly” (Fig. 6.20).
More about the Lorenz equations and chaos more generally can be
found in Strogatz (1994).
160 NONLINEAR MODELS

6.3 Questions and Problems

Question 6.1: Explain the difference between implicit and explicit time
dependence of a model.

Question 6.2: Define the term autonomous system.

Question 6.3: Give examples of systems where the implicit and explicit
time dependence cannot be broken up as we did in Eq. (6.2). You can modify
systems we introduced previously to get such examples.

Question 6.4: Give some properties that distinguish nonlinear from linear
systems.

Question 6.5: What do we mean by the area of attraction of a fixed point?

Question 6.6: Characterize a linear system by its fixed points.

Question 6.7: What is an asymptotically stable fixed point?

Question 6.8: What do we mean with the term hysteresis in a model?

Under which circumstances does it occur in a model with one variable?

Question 6.9: Why don’t linear models exhibit hysteresis?

Question 6.10: Nonlinear models can describe synergistic effects. What

do we mean by that?

Question 6.11: Which properties of a linear model are we speaking about

if we say that linear models show no synergism?

Question 6.12: Which mathematical idea is behind the Jacobian matrix?

What can we use this matrix for?

Question 6.13: What is the Jacobian matrix of a one-dimensional model?

Question 6.14: What do we mean by a center? What is special about it?

Question 6.15: What advantage does the Holling-Tanner predator-prey

model have over the Lotka-Volterra model?

Question 6.16: What do we mean by deterministic chaos? When can it

occur in systems of differential equations?

Problem 6.1: Mountain lake with time-dependent elimination rate

In Example 4.13 we analyzed a mountain lake with a yearly fluctuating nu-
trient input as a linear model. The total substance elimination rate of the
lake was given as k = 0.01 d−1 . In fact, this rate also fluctuates over the
year, because the substance-specific rate kr depends on the season, as does
the outflow rate kw . To keep things simple, we assume that the total elim-
 QUESTIONS AND PROBLEMS 161

ination rate is k = 0.005 d−1 for the first 6 months of the year, while for
the rest of the year it is k = 0.015 d−1 . Analyze the resulting model, in
particular by answering the following questions:

(a) Does this change make the model nonlinear?

(b) Can the model be described in the form of Eq. (6.2)?

(c) Identify the external relations.

(d) Sketch out a way to solve the model.

Problem 6.2: Fish in a pond

In Example 6.2 we discussed the model of a fish pond. We assumed that
the fish grow logistically and that the fishing rate is constant. The model
has the disadvantage that it can formally lead to negative fish populations.
Modify the model so that the fishing rate is a linear function of the fish
population N . The specific fishing rate is kf .

(a) Establish the model equation and determine the fixed point(s) and
its/their stability.

(b) Draw the model space graphically with the aid of the individual com-
ponents of the change function (as we did in Fig. 6.4).

(c) How must we choose kf so that the fixed-point fishing yield is

maximized?

(d) Calculate the maximum allowable specific rate kf so that the fish
population does not go extinct.

Problem 6.3: Jacobian matrix of the Lotka-Volterra model

Calculate the Jacobian matrix of the Lotka-Volterra model (Eq. 6.36) as
a function of the two system variables X and Y . Evaluate the matrix at
the two fixed points of the model, determine the corresponding eigenvalue
pairs and characterize the two fixed points according to their stability.

Problem 6.4: Predator-prey model with self-interaction

Calculate the Jacobian matrix for the predator-prey model with
self-interaction (Eq. 6.42). Use it to determine the eigenvalues at the fixed
point and characterize the fixed point’s stability.

Problem 6.5: Semi-trivial fixed point of the Holling-Tanner model

One of the fixed points of the Holling-Tanner model (Example 6.8) is X =
Xk , Y = 0. Examine the stability of the system at this fixed point with
the Jacobian matrix and draw a qualitative sketch of the trajectories in the
area surrounding the fixed point.
162 NONLINEAR MODELS

Problem 6.6: Nonlinear biomass growth in a pond

In a completely mixed pond with an average water retention time of τw =
V /Q (V : pond volume, Q: flow rate), the growth of biomass concentration
B depends on the concentration N of a limiting nutrient. Biomass B and
nutrient N are expressed with the same units, e.g. N = phosphorus per vol-
ume as dissolved nutrient and B= phosphorus per volume incorporated into
the biomass. The growth function G (transfer of phosphorus per volume
and time into the biomass) has the following form (N0 > 0, kw > 0):

fw N (N0 − N ) B for 0 ≤ N ≤ N0
G(N , B) =
0 for N > N0
The decay function (recirculation of biomass phosphorus into dissolved nu-
trient phosphorus) has the form:
R(N , B) = kA B, kA > 0
The system has a constant nutrient input JN (mass per time), the biomass
input is zero.
(a) Draw a box diagram for the N/B system (don’t forget the through-
flow!).
(b) Graphically discuss the development of the growth function through
time.
(c) Determine the fixed points and their stability.
(d) What is the condition (expressed by the coefficients) for a finite
biomass to establish itself in the lake?

Problem 6.7: Lotka-Volterra with two prey animals

We consider a Lotka-Volterra model consisting of three species (two prey
animals X1 , X2 , one predator Y ). The following processes are taken into
account:
Linear growth of the prey X1 , k1 X1
Linear growth of the prey X2 , k2 X2
Linear death rate of the predator, k3 Y
Feeding rate of the predator for prey X1 : k4 X1 Y
Feeding rate of the predator for prey X2 : k5 X2 Y
(a) Draw a box diagram and formulate the dynamic equations.
(b) Identify the fixed points and their stability.
(c) Which species survive in the long run?
Numerical values: k1 = 0.05 year−1
k2 = 0.03 year−1
k3 = 0.02 year−1
k4 = 0.002 year−1
k5 = 0.001 year−1
(The variables X1 , X2 , Y are non-dimensional.)
 QUESTIONS AND PROBLEMS 163

k1X1 k3X1Y k2Y Fig. 6.21: Lotka-Volterra

X1 Y
model with a prey niche.
k4X1 k5X2 X1 : prey in unprotected
habitat, X2 : prey in niche
k1X2
X2 (protected from predators),
Y : predator

Problem 6.8: Lotka-Volterra with prey niche

In the following modified Lotka-Volterra model, the prey animal can retreat
into a protected niche (Fig. 6.21).

(a) Establish the dynamic equations.

(b) What are the conditions for the coefficients ki so that a nontrivial19
fixed point exists?

(c) What does the condition found in (b) mean biologically? How could
we meaningfully modify the model (i.e. by introducing an improve-
ment which makes sense from a biological point of view) so that a
nontrivial fixed point always exists?

Problem 6.9: Competition for living space

Two species A and B are in direct competition for the same living space.
The dynamics of the two populations, whose size are denoted by A and B,
are influenced by the following processes:

1. Without the competing species, population A grows linearly with the

constant rate k1 .

2. Similarly, without competition population B grows according to the

following logistic model:

dB
= (k3 − k4 B)B if A=0
dt

3. The competition between populations A and B leads to a reduction

of growth in both populations. The reduction is proportional to the
product AB, where for A the proportionality factor is k2 while for B
it is k5 .

(a) Write down the dynamic equations for both populations.

(b) Determine all fixed points A∞ ∞

i , Bi (the subscript i refers to the
different fixed points).
19 The trivial fixed point is X1 = X2 = Y = 0.
164 NONLINEAR MODELS

(c) Determine the stability properties of those fixed points where both
populations are different from zero. Hint: use the Jacobian matrix.

Use the following values for question (c) and assume that A and B are
non-dimensional:

k1 = 4 a−1 ; k2 = 1 a−1 ; k3 = 6 a−1 ; k4 = 1 a−1 ; k5 = 1 a−1

Chapter 7

Time-Discrete Models

7.1 Time-Discrete Models with One Variable

In Sect. 2.4 we briefly introduced systems that change stepwise at certain
times rather than continuously. Such systems and the corresponding models
are called discrete in time or simply time-discrete. Starting from the initial
time t0 with n = 0, the times at which the stepwise changes take place are
numbered sequentially: t1 , t2 , t3 , . . .. The system variable V is only defined
for these discrete times. We indicate them with the superscript1 (n): V(n) .
A time-discrete model is like a recipe that tells us how, using the variable
at a previous time, we can calculate its value at the next discrete time
step (n + 1). It is not even that important whether the index (n) actually
describes consecutive time steps, or simply a sequence of events. In the
stochastic bed of nails (Fig. 2.7), we let a ball roll down an inclined plane
along a random path. Although we interpreted the left vertical axis in
Fig. 2.7 as time axis, we could just as well have seen it as a series of impacts.
At each impact (“time step”), the ball hits a nail and, according to a given
probability distribution, rolls down to the next (“later in time”) level.
As an illustration of how discrete models function, we choose a sim-
ple example from number theory. It states that the series of all quadratic
numbers Q(n) ≡ n2 can be obtained with the following recipe:

Q(n+1) = Q(n) + 2 Q(n) + 1 (7.1)

As a starting point we choose Q(0) = 0. The relation in Eq. (7.1) can either
be proven mathematically or simply checked for its validity by explicit
calculation of Q(1) , Q(2) and so on. We thus have a recipe to determine
Q(n+1) from Q(n) , that is, how to calculate (26)2 if we already know that
(25)2 = 625: (26)2 = 625 + 2 × 25 + 1 = 676. If we interpret Q(n) as a time

1 The index is put into brackets to distinguish it from a power exponent.

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 165

DOI 10.1007/978-3-642-30639-6_7, © Springer-Verlag Berlin Heidelberg 2013
166 TIME-DISCRETE MODELS

series, this simple recipe describes how a discrete-time system develops.

The instructions in Eq. (7.1) are then similar to a differential equation in
time.

7.1.1 Difference Equations and the Numerical Solution

of Differential Equations
At first sight, we could expect that analyzing and solving an iterative equa-
tion such as Eq. (7.1) is much easier than solving a differential equation such
as Eq. (4.7):
dV
= k · V (4.7)
dt
The algebraic solution of Eq. (4.7) cannot be found with a simple computer
program,2 whereas Eq. (7.1) seems an obvious candidate to be tackled with
a simple iterative program. In fact, the most primitive integration algorithm
for ordinary differential equations is based on the idea of approximating the
continuous progression of time with finite intervals of size Δt. We define:

V(tn ) ≡ V(n) with tn = t0 + n · Δt (7.2)

In discrete models, differ- We then write Eq. (4.7) for a finite time step with a so-called difference
ence equations take the equation:
place of differential equa- dV ΔV V(n+1) − V(n)
tions.
−→ = = k  · V(n) (7.3)
dt Δt Δt
Solving for V(n+1) gives us:

V(n+1) = (1 + Δt · k  )V(n) (7.4)

If we begin with the initial state V(0) and apply the equation m consecutive
times, we get:
V(m) = (1 + Δt · k  )m V(0) (7.5)
This expression does not seem to have much in common with the exact
solution of the differential equation (4.7), which, as we learned in Eq. (4.8),
is an exponential function. Of course, the numerical approximation of the
solution of (4.7) should eventually transform into the exact solution if only
we choose the time step Δt small enough. If the integration should take
place over the entire time period T , a given time step results in m = T /Δt
iterations. In the limit case, we have Δt −→ 0 and m −→ ∞. Inserted in
Eq. (7.5), V(m) ≡ V(T ) thus becomes a limit of the form:

k  T m (0)
V(T ) = lim [(1 + ) ]V (7.6)
m→∞ m

2By “simple” we exclude advanced software such as Mathematica with which

we could algebraically integrate.
 TIME-DISCRETE MODELS WITH ONE VARIABLE 167

A m
As can be learnt from mathematical handbooks, limm→∞ (1 + m) =
eA , so with A ≡ k  T :

V(T ) = ek T V0 (7.7)
This result is in full accordance with Eq. (4.8).
Let us exemplify the relation between discrete and continuous temporal
change with an example which is familiar to all of us: the calculation of
interest on a bank account.

Example 7.1: Bank interest

At the start of the year, you deposit 1,000e in your bank account.
The bank offers a generous yearly interest rate of 5 %. Calculate the
account balance at the end of the year, if
(a) The interest is added to the account once per year;
(b) The interest is added quarterly;
(b) The interest is added monthly;
(d) The interest is added continuously.

We choose the variables N0 = 1,000e and p = 0.05. For the different

scenarios, our account balance is then:

(a) N (t = 1) = N0 (1 + p) = 1.05× 1,000e =1,050e

(b) N (t = 14 ) = N0 (1 + p4 ), thus
N (t = 24 ) = N (t = 14 )(1 + p4 ) = N0 (1 + p4 )2
...
⇒ N (t = 1) = N0 (1 + p4 )4 = (1 + 0.05 4
4 ) N0 = 1050.95e

(c) Analogously to (b) we get

p 12
N (t = 1) = N0 (1 + 12 ) = 1051.16e

(d) To calculate the continuous interest payment, we use Eq. (7.7):

N (t = 1) = N0 ep = N0 e0.05 = 1051.27e

Fortunately for us depositors, the fact that interest is only payed once a
year does not make such a big difference for small interest rates. For larger
rates, it would make a difference however. For instance, for a rate of 20 %
per year the balance would grow by 22.14 % over a year for continuous
interest payment, compared to the “mere” 20 % if the interest is added
once per year.
In many (but not in all) cases, the numerical solution of a differen-
tial equation approaches the exact solution if the time step is chosen small
168 TIME-DISCRETE MODELS

enough.3 One could think, therefore, that a time-discrete difference

equation is nothing more than a course-grained variant of a differential
equation. One would then further assume that no fundamentally new sys-
tem behavior is to be expected. These premature conclusions would be
wrong, however. Let’s take it step by step and look at how these models
really behave.

7.1.2 First-Order Linear Models

We first examine the following inhomogeneous, linear, discrete, first-order
equation with constant coefficients a0 and I:
V(n+1) = I + a0 V(n) (7.8)
Eq. (7.8) is of first order because V(n+1) only depends on the immediate pre-
decessor value V(n) . We already encountered the case I = 0 with Eq. (7.4)
and in Example 7.1. As we will see, terms of the form a1 V(n−1) , a2 V(n−2)
etc. can occur on the right-hand side of Eq. (7.8). In such a case, the equa-
tion would become second, third or higher order.
Let’s think of the linear differential equation that is related to the time-
discrete linear Eq. (7.8). What does it look like? In other words: for which
differential equation does Eq. (7.8) form a numerical approximation?
To answer this question, we need to bring Eq. (7.8) into the difference
form (Eq. 7.3) by subtracting V(n) on both sides:
V(n+1) − V(n) = I + (a0 − 1)V(n)
V(n+1) − V(n)
or = J + k  V(n) (7.9)
Δt
I a0 − 1
with J = , k =
Δt Δt
As we already know, the differential equation corresponding to Eq. (7.9),
∞
dt = J + k V, has the fixed point V = − kJ for k  < 0. If our anal-
dV 

ogy between differential and difference equation is correct, with an infinite

number of iterations and if (a0 − 1) < 0 or a0 < 1, Eq. (7.9) would have to
tend to:
I/Δt I
V∞ = − =− (7.10)
(a0 − 1)/Δt a0 − 1
We can check whether our conjecture is true by applying Eq. (7.8) several
times in a row, starting with V(0) ≡ V0 :
V(1) = I + a0 V0
V(2) = I + a0 V(1) = I + a0 (I + a0 V0 ) = I(1 + a0 ) + a20 V0
V(3) = I + a0 V(2) = . . . = I(1 + a0 + a20 ) + a30 V0

3 We don’t want to suggest that the “forward integration” we used in Eq. (7.3)
is a particularly good strategy for numerical integration. There are in fact more
efficient and reliable methods.
 TIME-DISCRETE MODELS WITH ONE VARIABLE 169

From this we can already see the general rule:

n−1
V(n) = I · (a0 )i + an0 V0 (7.11)
i=0

With the formula for the calculation of geometric sums,

n−1
an0 − 1
(a0 )i = (a0 = 1) (7.12)
i=0
a0 − 1

we finally get (for a0 = 1)4 :

an0 − 1 I I
V(n) = I + an0 V0 = − + an0 ( + V0 ) (7.13)
a0 − 1 a0 − 1 a0 − 1
If |a0 | < 1, that is if −1 < a0 < 1, the second term in Eq. (7.13) disappears
for n → ∞, so that:
I
V∞ = lim V(n) = − , −1 < a0 < 1 (7.14)
n→∞ a0 − 1
The assumption formulated in Eq. (7.10) is therefore only partially correct,
because there is not only an upper bound (a0 < 1). Equation (7.14) shows
a lower bound, too: a0 must also not become too negative (in fact, not more
negative than −1). Otherwise the system oscillates back and forth between
increasing positive and negative values. Figure 7.1 summarizes the different
cases. Finally, we would like to note that similar to the case of an inhomo-
geneous linear differential equation, Eq. (7.8) could have been transformed
into a homogeneous difference equation by introducing a new variable. It
could then have been solved according to Eq. (7.5) (see Problem 7.1).

Example 7.2: Fish population in a pond

In a fish pond, fishing and natural mortality mean that only about
40 % of the individuals of a given fish species survive each year. Each
spring, 1,200 individuals are set free in the pond. At the beginning of
the year 2000, there were 1,000 fish in the pond.

(a) Calculate the fish population N (n) at the beginning of 2001 and
2002.
(b) In the long term, what is the fish population at the beginning of
the year?
(c) If, at the beginning of 2000, there had been no fish in the
pond, how long would it take for the difference to the first case
(N 0 =1,000 in the year 2000) to be less than ten fish?

4 For a0 = 1, the result follows directly from Eq. (7.11): V(n) = nI + V0

170 TIME-DISCRETE MODELS

Fig. 7.1: Different so- a b

lutions of the discrete 2 1
a 0 = −0.5
linear first-order model
0.8
(Eq. 7.8) with I = 1. (a) 1.5 a 0 = 0.5
0 < a0 < 1: direct conver- 0.6
gence, (b) −1 < a0 < 0:

n
1
oscillating convergence, 0.4
(c) a0 > 1: direct diver-
0.5
gence, (d) a0 < −1: oscil- 0.2
lating divergence
0 0
c d
300 50

250

200 0
n

150

100 50

50 a0 = 2 a 0 = −2

0 100
0 2 4 6 8 0 2 4 6 8
n n

(a) We begin our analysis in the year 2000 (N 0 =1,000) and apply
Eq. (7.8) with a0 = 0.4, I =1,200:
In the year 2001: N (1) =1,200+0.4×1,000 = 1,600
In the year 2002: N (2) =1,200+0.4×1,600 = 1,840

(b) Because |a0 | < 1, in the long run, at the end of each winter (i.e.
just before the new fish are set free) the fish population reaches the
following value (Eq. 7.14):

1, 200
N∞ = − = 2, 000
0.4 − 1

(c) If we apply Eq. (7.13) to our example,

I
N (n) = (1 − an0 ) + an0 N 0 ,
(1 − a0 )

we can see that the influence of the initial state N 0 is limited to the
last term. In the first case, this term is (0.4)n × 1,000, in the second
case it is zero. Thus we must ask ourselves for which n the following
relation is valid:
(0.4)n × 1, 000 = 10
 TIME-DISCRETE MODELS WITH ONE VARIABLE 171

If we transform the expression and take the logarithm on both sides,

it follows that5 :

log(10/1, 000) log(0.01) −2

n= = = =5
log(0.4) log(0.4) −0.40

After 5 years at most, it no longer matters whether initially there

were 1,000 or no fish in the pond.

7.1.3 Higher-Order Linear Models

Time-discrete models of higher (e.g. q) order are expressed by an iterative
equation in which q + 1 consecutive parameter values (V (n) , V (n+1) , . . . ,
V (n+q) ) are put in a relation to each other. If the system is linear, the
equation can be brought into the following form:

I + a0 V(n+q) + a1 V(n+q−1) + . . . + aq V(n) = 0 (7.15)

In fact, the inhomogeneous Eq. (7.15) can be transformed into a homoge-

neous one by choosing a new variable which lacks the term I. More precisely,
we choose the variable:

V̂(j) = V(j) + b, j = 1, 2, . . . (7.16)

"
q
If ai = 0, and if we choose
i=0

I
b= (7.17)
"
q
ai
i=0

then an iterative equation for V̂(n) emerges which, apart from the miss-
ing term I, has exactly the form of Eq. (7.15). Note that the solution of
Eq. (7.15) requires the specification of q initial conditions (e.g. V(0) , V(1) ,
. . ., V(q−1) ).

Example 7.3: Bottle recycling

A beverage manufacturer launches a new product in a special return-
able bottle. Due to capacity limitations, the company can produce a
maximum of 10, 000 bottles per month. Experience shows that 70 % of

5 Either the natural logarithm or the common (base-10) logarithm work. The
following numbers correspond to the common logarithm.
172 TIME-DISCRETE MODELS

bottles are returned in the following month, and another 10 % a month

later. The rest of the bottles is lost. The beverage company would like
to know:

(a) How the monthly available number of bottles develops through

time, and
(b) How many bottles it can refill per month in the long run.

The sale of the product begins in month 1 with the first 10, 000 bottles.

The iterative equation for the problem has the form

N (n+2) = 0.7N (n+1) + 0.1N (n) + 10, 000 (7.18)

where N (n) is the number of available bottles in month n after launching

the product. The initial conditions are:

N (0) = 0
N (1) = 10, 000

Without having to calculate much, we can easily answer the question about
the long-term availability of bottles (that is, the fixed point N ∞ ). Twenty
percent of the bottles never come back, which means they need to be
replaced by new bottles (10,000 per month). This implies that, at the fixed
point, 50, 000 bottles are filled each month, since 20 % of that are the 10,000
new ones. The bottle inventory develops as follows:

2nd month: N (2) = 0.7N (1) + 0.1N (0) + 10, 000

= 7, 000 + 0 + 10, 000 = 17, 000
(3)
3rd month: N = 0.7N (2) + 0.1N (1) + 10, 000
= 11, 900 + 1, 000 + 10, 000 = 22, 900

At the fixed point, the following must hold: N (n+2) = N (n+1) = N (n) =
N ∞ . Therefore:

N ∞ = 0.7N ∞ + 0.1N ∞ + 10, 000

10, 000
→ N∞ = = 50, 000
0.2

Calculating the monthly values N (n) with the “recipe” in Eq. (7.18) is easily
done on a computer; this would show that the value indeed asymptotically
approaches N ∞ = 50, 000.
 TIME-DISCRETE MODELS WITH ONE VARIABLE 173

Recycling bottles

An alcoholic buys N bottles of After one week, there are

Bottle
beer each day and returns half 84 empty bottles in the apartment.
Return
of them to the store How large is N?
the next day.
How large is N?

Boohooo, stop! I
Answer me,
don t know!
professor!

Sigh..

In Sect. 7.1.2 we saw that a time-discrete model can also lead to divergent
solutions (see Fig. 7.1). It would therefore be useful to have a mathematical
tool to analyze equations of the form (7.15) more generally.
To that end, we look at the homogeneous form of Eq. (7.15), which we get
via the transformation in Eq. (7.16):

a0 V̂(n+q) + a1 V̂(n+q−1) + . . . + aq V̂(n) = 0 (7.19)

We now claim that the following function6 is a solution of Eq. (7.19):

V̂(n) = λn (7.20)

Inserting into Eq. (7.19) yields:

a0 λn+q + a1 λn+q−1 + . . . + aq λn = 0

After dividing by λn we get an equation that is independent of the iteration

step n:
a0 λq + a1 λq−1 + . . . + aq = 0 (7.21)
This is the characteristic equation of Eq. (7.19). In general, it has q (real
or complex) solutions λj (j = 1, . . . , q). The general solution of Eq. (7.19)
therefore results from a linear combination of all λj :
q
V̂(n) = Aj (λj )n (7.22)
j=1

The Aj are determined by the initial conditions. We want to limit our

discussion of Eq. (7.22) to those situations in which the roots of the char-
acteristic Eq. (7.21), λj , are real, and differentiate between two cases:

1. The absolute value of all λj is smaller than 1: |λj | < 1. Then, V̂(n)
converges to zero for n → ∞.
6 Equation (7.20) makes it clear why we prudently write V(n) : The n in paren-

theses gives us the n-th iteration step, whereas λn actually means the n-th
power.
174 TIME-DISCRETE MODELS

2. At least one root is ≥ 1 in absolute value. Then, it follows that V̂(n) →

±∞ (excluding special initial conditions).
As a reminder: If Eq. (7.19) indeed developed from an inhomogeneous equa-
tion, then, because of Eq. (7.16), the following holds for the first case
(V̂(n) → 0):
I
lim V(n) = lim V̂(n) − b = −b = − q (7.23)
n→∞ n→∞ "
ai
i=0
Let’s go back to Example 7.3 and analyze it again with the general theory.
To do so, we write Eq. (7.18) in the form of Eq. (7.15):
10, 000 − N (n+2) + 0.7N (n+1) + 0.1N (n) = 0 (7.24)
The characteristic equation is:
−λ2 + 0.7λ + 0.1 = 0
1
→ λ = − [−0.7 ± {(0.7)2 + 0.4}1/2 ]
2
1
= − [−0.7 ± 0.943]
2
→ λ1 = +0.822; λ2 = −0.122
We are dealing with case 1 (convergence), therefore, as per Eq. (7.23):
10, 000 10, 000
N ∞ = −b = − = = 50, 000
(−1 + 0.7 + 0.1) 0.2
This conforms to our earlier result. Finally, the general solution of Eq. (7.24)
with the initial values of N (0) = 0, N (1) = 10, 000 is (also see Problem 7.4):
N (n) = 50, 000 − 48, 850 (0.822)n − 1, 150 (−0.122)n (7.25)

7.1.4 Nonlinear Models

Even though the question of whether linear difference models tend to a fixed
point is not as easily answered as for the case of linear differential equations,
simple recipes (as in Fig. 7.1) to analyze their long-term behavior do exist.
In the case of nonlinear systems, however, the matter becomes quite a bit
more complex. In this section we will only look at first-order nonlinear
models and merely give one example of the multifaceted world of such
models.
We consider a (time-discrete) difference model of the form:
V(n+1) − V(n) = g(V(n) ) (7.26)
g(V(n) ) is an arbitrary function of V(n) , for instance the logistic growth
function (see Eq. 6.4):
g(V(n) ) = k(b − V(n) )V(n) , k, b > 0 (7.27)
 TIME-DISCRETE MODELS WITH ONE VARIABLE 175

Line with slope

Fig. 7.2: As in the case of one-dimensional nonlinear differential equations, the

behavior of a nonlinear difference model near a fixed point can be analyzed with
local linearization. To do so, the function g(V) is replaced by a straight line at
the fixed point V∞ j . The line has the slope dV |V∞
dg
j
. X(n) indicates the deviation of
V(n) from the fixed point.

Obviously, the roots of the function g(V(n) ) are the fixed points of the
model, since there, according to (7.26), V(n+1) = V(n) . As in the case of
nonlinear differential equations we analyze the behavior of the system at
the fixed point V∞
j by introducing a new discrete variable:

X(n) = V(n) − V∞
j (7.28)

It then follows from Eq. (7.26):

X(n+1) − X(n) = g(V(n) ) = g(V∞

j +X
(n)
) (7.29)

If the deviation from the fixed point (X(n) ) is small, g can be developed
into a Taylor series and broken off after the linear term (see Fig. 7.2):

∞ ∞ dg 
g(Vj + X ) = g(Vj ) +
(n)
X(n) + . . . (7.30)
dV V∞
j

Since V∞ ∞
j is a fixed point, by definition, g(Vj ) = 0. We get a linear differ-
ence equation for X (n)7
:

dg 
X (n+1)
−X =
(n)
X(n) (7.31)
dV V∞
j

A simple transformation results in:

  
dg 
X(n+1) = 1 + X(n) = A X(n) (7.32)
dV V∞
j

7 Obviously, the procedure is completely analogous to the one we used for non-
linear differential equations in Chap. 6.
176 TIME-DISCRETE MODELS

From our earlier discussion (e.g. in Fig. 7.1), we already know that Eq. (7.32)
tends to the fixed point if the following holds:
  
 dg  
|A| = 1 + <1 (7.33)
dV  ∞ 
Vj


dg 
Expressed by dV  this means, in accordance with the continuous case
V∞
j
(Eq. 6.12), that the fixed point V∞
j is unstable if the following holds:

dg 
≥ 0 ⇔ V∞ unstable (7.34)
dV V∞ j
j

The inverse is not true however: in fact, because of Eq. (7.33) the fixed
point can also be unstable due to a strongly negative slope of the function
g. Altogether, the conditions for stability are:

dg 
0> > −2 ⇔ V∞ stable (7.35)
dV  ∞ Vj
j

Example 7.4: Discrete logistic growth

The time-discrete logistic growth model is well-suited to have a brief
exemplary look at the surprising world of nonlinear difference models.
Eq. (7.27) describes discrete logistic growth. The model has two fixed
points: V∞ ∞
1 = 0 and V2 = b. Combining the logistic growth equation
with the model function (7.26) yields the following model equation:

Discrete V(n+1) − V(n) = k(b − V(n) )V(n) ≡ g(V) , k, b > 0

growth is PRIVATE BANK
the name
of our g(V) has the derivative:
game!

dg
= −2kV(n) + kb , (7.36)
dV
For the two fixed points this results in:
 
dg  dg 
= +kb , = −kb , k, b > 0 (7.37)
dV V∞ dV V∞
1 2

As long as kb < 2, the second fixed point V∞

2 is stable. The first fixed
point, V∞
1 = 0, is always unstable.

For the following analysis of the logistic growth model we closely follow
Arrowsmith and Place (1992). With the new variable x(n) = 1+kbk
V(n) the
model equation takes the form
x(n+1) = μ(1 − x(n) )x(n) ≡ Fμ (x(n) ) , μ ≡ 1 + kb (7.38)
 TIME-DISCRETE MODELS WITH ONE VARIABLE 177

The function Fμ (x) is a downward-facing parabola with its apex at x = 12 ,

intersecting the x-axis at x = 0 and x = 1 (Fig. 7.3a). Note that Fμ (x)
only depends on the parameter μ: we add it as an index to the function F
because it will play an important role.
The model has reached an equilibrium x∞ if the following holds:

x∞ = Fμ (x∞ ) = μx∞ (1 − x∞ ) (7.39)

Solving for x∞ gives (next to the trivial solution x∞ = 0):

μ−1
x∞ = (7.40)
μ

Graphically speaking, we get x∞ as the intersection of the parabola y =

Fμ (x) with the line y = x (Fig. 7.3a). The evolution of the system from
an initial point x(0) to the fixed point x∞ can also be shown graphically
(Fig. 7.3b).

a b y
y

Fixed point:
Fµ (x)≡µx (1−x)=x
B B’
y=µx(1−x)
Fµ (x (0)) A A’ C
C’

y=x

x x
x∞ 1 x (0) x (1) x
∞ (2)
x 1

Fig. 7.3: (a) Graphical determination of the fixed point of the discrete model
x(n+1) = μx(n) (1 − x(n) ) ≡ Fμ (x(n) ). See text for further explanations. (b) The
movement of the model from the initial state x(0) to the fixed point x∞ can be
determined graphically: After the first iteration, Fμ (x(0) ) takes on the new x value
x(1) , at point A. That x value can be found by mirroring x(1) = Fμ (x(0) ) on the
line x = y (point A’). The value Fμ (x(1) ), belonging to x(1) , lies at point B. By
mirroring B on the line x = y (point B’) we get x(2) or point C, and so forth, all
the way to the fixed point x∞ (From Arrowsmith and Place (1992))

As shown above, the nontrivial fixed point is stable for 0 < kb < 2
which translates into the stability condition 1 < μ < 3. Yet, for μ ≥ 3
there are fixed points as well, although of a different kind. For instance, for
μ = 3.2, the system ends on a closed path that oscillates between the two
178 TIME-DISCRETE MODELS

Fig. 7.4: Behavior of the y

discrete model x(n+1) =
μx(n) (1 − x(n) ) for μ = 3.2.
In contrast to Fig. 7.3b,
the trajectory is not mov-
ing towards the fixed point
x∞ = μ−1 μ
= 0.6875, but
rather oscillates between
the two points x∞ ∞
1 and x2 .
More about this case is
found in Fig. 7.5c
x
x 1∞ x ∞ x 2∞

values x∞ ∞
1 and x2 lying on either side of x
∞
(Fig. 7.4).8 In other words: if
(n)
we only consider every other value x , i.e. only either the even or odd n,
these sequences again reach a fixed point. We thus have to formulate a new
specification from the iterative equation Eq. (7.38) that directly takes us
from x(n) to x(n+2) , skipping the intermediate step x(n+1) . We name this
(2) (2)
“double-jump function” Fμ . To get Fμ out of Fμ , we insert y = μx(1−x)
in the function Fμ (y) = μx(1 − y):

Fμ(2) (x) = μ2 x(1 − x)[1 − μx(1 − x)] (7.41)

(2)
Fμ is a fourth-order polynomial and must not be confused with the square
of Fμ . Its fixed points can be calculated—as in the case of Fμ —from the
following equation:
x∞ = Fμ(2) (x∞ ) (7.42)
(2)
Graphically, they can be found from the intersection of y = Fμ (x)
with the line y = x.
(2)
In Fig. 7.5 we plotted Fμ (x) for three different values of μ. We are
interested in the nontrivial (x = 0) intersections of the curve with the line
x = y. For μ = 2.8 (Fig. 7.5a) there is a single intersection point x∞ . Of
course, it is identical with the fixed point of the “single-jump function”
Fμ (as per Eq. 7.40, x∞ = 1.82.8 = 0.643), since if the system doesn’t move
anymore after a single jump, it will also remain motionless after two, three,
or q jumps.
(2)
For μ = 3, the curve y = Fμ (x) has a slope of exactly 1 at the
intersection—as does the function y = x (Fig. 7.5b). For even larger values
of μ (e.g. for μ = 3.2, see Fig. 7.5c), the original intersection point x∞ is
divided into two new intersection points x∞ ∞
1 and x2 . These two drift apart

8 Strictly speaking this is only true if the initial point does not lie at x∞ =
μ−1
μ
= (3.2−1)
3.2
= 0.6875.
 TIME-DISCRETE MODELS WITH ONE VARIABLE 179

a b c
y=x

y = Fµ(2)(x)

µ = 2.8 µ=3 µ = 3.2

x x x
x∞ x∞ x 1∞ x ∞ x 2∞

Fig. 7.5: (a) The fixed points of the “double-jump function” (Eq. 7.32) result
(2)
from the intersection of the curves y = Fμ (x) and y = x. (b) For μ = 3,
the system shows a bifurcation: The intersect x∞ splits into the two points x∞
1
and x∞2 (period doubling). (c) “Two-jumps function” for the situation shown in
Fig. 7.4 (From Arrowsmith and Place (1992))

further for an increasing μ. The “double-jump system” then has three fixed
points in total (the middle one, x∞ , being unstable). The “single-jump
system” in contrast jumps back and forth between x∞ ∞
1 and x2 . In fact,
Fig. 7.5c represents the case shown as “single-jump system” in Fig. 7.4.
The behavior of the system at μ = 3 is called phase doubling. The points
at which this doubling occurs are called bifurcation points. In order to get
further bifurcation points, we have to examine the “four-jumps function”
(4) (8) (16)
Fμ , or Fμ , Fμ and so forth. In fact, the model experiences further
phase doublings with growing μ. Yet the story does not end here: Above
the critical value μ = 3.8284 . . ., completely new situations for which no
periodic cycles exist appear on the scene. For μ > 4 these aperiodic fluctu-
ations play the predominant role if the system did not start on a periodic
solution by chance. This strange behavior is another example of determin-
istic chaos, which we already encountered in Sect. 6.2.5.
The emergence of deterministic, but unpredictable (that is, apparently Even one-dimensional non-
chaotic) behavior is not actually that surprising—we know it very well from linear difference models
our own experience. We can take the 104 cards from a Solitaire card game can exhibit chaotic (not
as an example. They can be mixed in 104! ways (factorial of 104, a number easily predictable) behavior.

Bifurcation
A father sends forth his two sons to Nothing but chance brought luck
try their luck in the world. They arrive to one of them while setting the
at a fork in the road and throw a coin to other on the road to ruin.
decide who goes left and who right. The
first son finds a princess and marries
her, the second encounters a dragon
and is eaten... What did
Einstein say? God
does not play dice.
180 TIME-DISCRETE MODELS

with more than 160 digits). If we now deal a certain set of cards from our
deck, and make our decisions strictly according to certain rules, we will be
able to win the game for certain sets of cards but not for others. Switching
two cards could be enough to turn a set without a solution into one with
a solution. Even though all steps are completely determined, we will be
unable to deduce any regularity that could let us predict whether the game
will work out or not.

7.2 Time-Discrete Models with Several Variables

7.2.1 Linear Models
After this short glance at the complex world of chaos, we’ll try to get
to firmer ground again and to conclude this chapter by having a look at
multi-dimensional discrete models. In essence, we will concentrate on the
discussion of first-order multi-dimensional linear models and only explicitly
solve the two-dimensional case.
A first-order q-dimensional linear difference model can be expressed by
the following algebraic system of equations:
q
(n+1) (n)
Vi = Ii + pij Vj , i = 1...q (7.43)
j=1

The pij are the elements of the coefficient matrix P = (pij ) which we
already introduced for multi-dimensional linear differential equations (see
Eq. 5.81). So we can also write Eq. (7.43) in matrix form:

V(n+1) = I + PV(n) (7.44)

P is a (q × q) matrix; V(n+1) ,V(n) and I are q-dimensional vectors.

Outwardly, the difference equation (7.44) now looks like Eq. (7.8), except
that the coefficient a0 from there has become the matrix P while V(n)
and I have become vectors. Indeed, formally we can write the solution for
Eq. (7.44) in the same way as Eq. (7.11):
n−1
V(n) = (P)n · V0 + (P)i I (7.45)
i=0

Calculating the power of matrices is quite complicated in general, so the

formal solution Eq. (7.45) is of little use for concrete calculations. There is
one important exception, however: if the matrix P were to have a diagonal
form by chance, its powers would be diagonal too, so that we could simply
take the power of the diagonal elements of the original matrix. If P is
diagonal, this also means that the original algebraic system of equations
(7.43) can be broken down into q disjoint equations which can all be solved
independently of each other.
And with that, we are (as in Sect. 5.2), once again faced with eigen-
values. If we should succeed to diagonalize the coefficient matrix through
 TIME-DISCRETE MODELS WITH SEVERAL VARIABLES 181

the introduction of new variables, we could use all the tools we learnt for
one-dimensional systems in Sect. 7.1.2 to solve Eq. (7.43).
Instead of more theoretical considerations and the risk of loosing track
of the essential issues, we now want to turn to the concrete example of a
two-dimensional system and use it to illustrate the important concepts step
by step.

7.2.2 Linear Models with Two Variables

Example 7.5: Who wins against the bank?

Two gamblers (X and Y) agree on the following rules of a game with
money, to be applied successively through each round of the game:
1. Player X receives from or pays to the bank 2/3 of the amount
which she currently posesses, minus 1/3 of the amount that player
Y has.
2. At the same time player Y receives from or pays to the bank 1/3
of the amount that he currently has, minus 2/3 of the amount
that player X holds.
When we write “receives from or pays to”, that is because the balances
X (k) and Y (k) can also become negative, in which case the term that
described money received from the bank now represents money lost to
it.
The initial balance is X 0 = 600 and Y 0 = 900. How do the players’
balances develop? Determine a general formula for the balance after k
rounds, X (k) and Y (k) . Does the bank lose money by performing these
transactions?

First, we establish the iterative equation by translating the rules into math-
ematical expressions step by step. The change of balance in the k-th step
is:
2 1
X (k) − X (k−1) = X (k−1) − Y (k−1)
3 3 (7.46)
2 1
Y (k) − Y (k−1) = − X (k−1) + Y (k−1)
3 3
Eq. (7.46) can be transformed to:

5 (k−1) 1 (k−1)
X (k) = X − Y
3 3 (7.47)
2 (k−1) 4 (k−1)
Y (k) =− X + Y
3 3
182 TIME-DISCRETE MODELS

Or, in matrix form (and calling the (2 × 2) matrix P):

      
X (k) 5/3 −1/3 X (k−1) X (k−1)
= =P (7.48)
Y (k) −2/3 4/3 Y (k−1) Y (k−1)

Whether the bank goes bust or not can be decided by looking at the de-
velopment of the sum S (k) = X (k) + Y (k) throughout the game. If we add
up the two equations from Eq. (7.47), we get:

S (k) ≡ X (k) + Y (k) = X (k−1) + Y (k−1) ≡ S (k−1) (7.49)

So there is no danger for the bank; the sum of balances stays constant at
S 0 = X 0 + Y 0 =1,500. On the other hand, however, the bank also doesn’t
make any profit. But what about the distribution of the total amount
amongst the two players?

Table 7.1: Example 7.5: State of the game after the first four rounds

k X (k) Y (k) S (k)

0 600 900 1,500
1 700 800 1,500
2 900 600 1,500
3 1,300 200 1,500
4 2,100 −600 1,500

Table 7.1 shows the state of the game after the first four rounds, calculated
by applying Eq. (7.47) successively. Not a good game for Y! Even though
he began with a head start, after only four rounds he’s already in debt.
How does the game go on?
It would be elegant if we could formulate the solution of the system
of equations similarly as the one for the one-dimensional system (Eq. 7.8).
Let’s recall Eq. (7.49): the difference equation of the sum S (k) ≡ X (k) +Y (k)
is solvable by itself, without calculating the individual solutions of X (k)
and Y (k) . We will apply a trick (the benefit of which will become apparent
soon): we’ll look for a second linear combination9 of X (k) and Y (k) whose
iterative equation is solvable by itself. Indeed, such a form exists:

T (k) = 2X (k) − Y (k) (7.50)

9 A linear combination of two variables X and Y is an arbitrary linear expres-

sion of the form aX + bY . Of course the sum of X and Y is a linear combina-

tion with a = b = 1.
 TIME-DISCRETE MODELS WITH SEVERAL VARIABLES 183

If, with the aid of the original iterative equations (7.47), we replace the
variables X (k) and Y (k) in Eq. (7.50) with their “predecessors” X (k−1) and
Y (k−1) , it follows that:
   
5 (k−1) 1 (k−1) 2 (k−1) 4 (k−1)
T (k)
=2 X − Y − − X + Y
3 3 3 3
   
10 2 2 4
= + X (k−1) − + Y (k−1)
3 3 3 3
= 4X (k−1) − 2Y (k−1) = 2 (2X (k−1) − Y (k−1) )
So:
T (k) = 2T (k−1) (7.51)
The solution of the iterative equation now becomes very simple:
T (k) = 2k T 0 = 2k (2X 0 − Y 0 ) = 2k × 300 (7.52)
What remains is to replace the original variables X (k) and Y (k) with the
new variables S (k) and T (k) , whose solutions we know (see Eqs. 7.49 and
7.52). After a bit of algebra we find:
1
X (k) = (S (k) + T (k) )
3 (7.53)
1
Y (k) = (2S (k) − T (k) )
3
By using the solutions for S (k) and T (k) we can thus also reconstruct those
for X (k) und Y (k) :
1
X (k) = (S (0) + 2k T (0) )
3
1
= (1, 500 + 2k × 300)
3 (7.54)
1
Y (k)
= (2S (0) − 2k T (0) )
3
1
= (3, 000 − 2k × 300)
3
It now immediately becomes clear that player Y must loose. In Problem 7.6
we ask how large the initial sums X (0) and Y (0) have to be so that (with
the same rules) player Y wins or the amounts of the two players remain
constant. Of course, the latter situation represents a fixed point.
How could one have guessed the trick we applied to solve Eq. (7.46)?
In the following section, we will discuss the deeper meaning of the solution
and the method we used to get to it.

7.2.3 The Role of Eigenvalues and Eigenfunctions

As elegant as the solution of Example 7.5 may be, it does not explain which
intuition we followed when we introduced the new variables S (k) and T (k)
184 TIME-DISCRETE MODELS

(Eqs. 7.49 and 7.50). Although it was possible, with some experience, to
see from the system of equations (7.46) or (7.47) that the sum of the two
variables X (k) and Y (k) remains constant, the choice of Eq. (7.50) appears
quite random.
Of course, behind this method lurks a mathematical concept that we
frequently mentioned: the determination of the eigenvalues (and associated
eigenfunctions) of a linear system. In Appendix D we explain the procedure
for a n-dimensional system. Without using computational methods, only
two-dimensional systems can be solved by hand with reasonable effort.
Table 7.2 on the next page summarizes the approach for Example 7.5.
There would be a lot more to say about time-discrete models and their
difference equations (we haven’t even mentioned multi-dimensional non-
linear models in passing!). Nevertheless, we want to close the discussion
here. There is other literature available to delve deeper into the topic (e.g.
Luenberger 1979).
 TIME-DISCRETE MODELS WITH SEVERAL VARIABLES 185

Table 7.2: The solution of two-dimensional difference models with eigenfunc-

tions, explained with Example 7.5. Appendix D gives an overview of all equations
used here

The matrix P of the system (7.47) is:

 
5/3 −1/3
P= (7.55)
−2/3 4/3

with
5 4
tr(P) = + =3
3 3
1
det(P) = (20 − 2) = 2
9
Eigenvalues (D.13):
1  1
λi = [3 ± 32 − 4 × 2] = [3 ± 1]
2 2
λ1 = 2, λ2 = 1
Eigenfunctions (D.17):
2 2 1
z1 = − X + (2 − 5/3)Y = − X + Y
3 3 3
1
= − (2X − Y )
3
1 1
z2 = (1 − 4/3)X − Y = − (X + Y )
3 3

Note: Except for the constant factors, z1 and z2 are consistent with the
variables S und T , that is, z1 = − T3 and z2 = − S3 . The eigenfunctions are
only determined to a constant factor.
Determinant (D.21):
4 5 1 1
D = −2 × 2 + 2 × + 1 × = [−12 + 8 + 5] =
3 3 3 3
So, according to (D.20) the reverse transformation is:
 
1 5
X = 3 − z1 − (2 − )z2 = −(z1 + z2 )
3 3
 
4 2
Y = 3 −(1 − )z1 − z2 = z1 − 2z2
3 3

The result conforms with Eq. (7.53), if—as mentioned before—we take into
account that z1 = −T /3 and z2 = −S/3.
186 TIME-DISCRETE MODELS

7.3 Questions and Problems

Question 7.1: Explain the difference between a time-discrete model and
a model that is continuous in time.

Question 7.2: Explain the relationship between the terms difference model
and time-discrete model.

Question 7.3: Why is it important whether the interest is paid out yearly
or quarterly for interest rates above 10 % per annum, but much less so for
2 %?

Question 7.4: Explain the difference between order and dimension of a

difference model.

Question 7.5: Give examples of algebraic formulas for:

(a) A one-dimensional, linear first-order difference model.

(b) Like (a), but nonlinear.
(c) Like (a), but second-order.
(d) Like (a), but two-dimensional.
(e) Like (d), but non-linear.

Question 7.6: What is the difference between one-dimensional nonlinear

difference models from the corresponding differential models, in terms of
the stability of a fixed point?
(2)
Question 7.7: Explain the difference between (Fμ )2 (Eq. 7.38) and Fμ
(Eq. 7.41).

Question 7.8: What do we mean by deterministic chaos in difference

models?

Question 7.9:

(a) Which dimension must a nonlinear first-order difference model have

at least, so that in principle, deterministic chaotic behavior could
occur?

(b) What is the answer to that question for nonlinear systems of differ-
ential equations?

Question 7.10: What is the characteristic equation of a qth-order linear

difference model? Is there a characteristic equation for nonlinear models?

Problem 7.1: Elimination of the inhomogeneous term

Calculate the fixed point V∞ of the time-discrete model
V(n+1) = I + a0 V(n) by introducing a new variable V̂(n) which converts the
model equation into the homogeneous relation V̂(n+1) = bV̂(n) . Derive a
corresponding recipe for a linear qth-order iteration model.
 QUESTIONS AND PROBLEMS 187

Problem 7.2: Consumer loan

A credit institute charges its credit card holders a monthly interest of 2.5 %
on the amount they owe. What yearly interest rate does this equate to?

Problem 7.3: Reactor with chloride

In a chemical reactor, 10 kg of chloride are added weekly. All chloride is
completely mixed with the water in the reactor. Once per week, 20 % of
the solution in the tank is emptied.
(a) How much chloride is there in the tank immediately before the weekly
discharge, if the reactor has already been in operation for a very long
time?
(b) If the reactor contains no chloride at time t = 0, how many weeks
does it take until the amount of chloride in the tank has reached the
value calculated in (a) to within a 5 % deviation?
Note: chloride is conservative in the reactor, i.e. it does not react. The
reactor is completely mixed before the weekly discharge.

Problem 7.4: Returnable bottles

Let’s take another look at the beverage manufacturer from Example 7.3.
(a) Prove Eq. (7.25).
(b) Deduce a corresponding relation, if in the first month initially only
5,000 bottles are produced, but 10,000 in every following month.

Problem 7.5: Fibonacci numbers

The Fibonacci numbers are a sequence of numbers that can be defined by
the iterative equation y (n+2) = y (n+1) + y (n) . With y (1) = y (2) = 1, they
are therefore 1, 1, 2, 3, 5, 8,. . . . With the characteristic equation, develop a
relation of the following form that describes the Fibonacci numbers:

y (n) = A1 λn1 + A2 λn2 . . .

Note: the result will deliver irrational

√ expressions for both the λ values and
the coefficients Ai (they contain 5). Nevertheless, the resulting y (n) will
be rational, even integer numbers!

Problem 7.6: Gambling

In Example 7.5, for unchanged rules, how do we have to choose the initial
balances for the two players (X 0 and Y 0 ), so that
(a) player Y wins?
(b) both players keep a constant balance?
188 TIME-DISCRETE MODELS

Problem 7.7: Students in a study program

In a given subject, 120 students start studying per year. On average, five
students per year drop out prematurely. The study duration is 5 years.
How many students are enrolled in the program in total (i.e. in all years
combined) once the system has reached a fixed point?

Problem 7.8: Fish in a pond

In a pond, each winter J = 10, 000 fish are introduced. In summer, together
with the offspring from natural reproduction, these fish form age cohort 1,
(i)
N1 , of the i-th year. A fraction of a = 0.1 of this age group survives the
winter (the rest is either caught by fishermen or dies), it becomes next year’s
(i+1)
age cohort 2, N2 . Higher-order age cohorts are not taken into account
because due to the high death rate, they are very small. The contribution of
(i)
natural reproduction in N1 is calculated as follows: each fish of age cohort
(i−1)
1 of the previous year produces b1 = 0.5 offspring (that is, b1 N1 ), each
(i−1)
fish of age cohort 2 of the previous year produces b2 = 2 (i.e. b2 N2 ).
(i) (i)
(a) Draw a box diagram to describe N1 and N2 .
(i) (i)
(b) Establish the iterative equations for N1 and N2 .
(c) Convert the two coupled first-order difference equations into a single
(i)
second-order difference equation for N1 .
(i) (i)
(d) Calculate N1 and N2 by iterating through the first years. Use the
(0) (0)
initial values N1 = 0, N2 = 0.
(e) To solve this equation, convert the inhomogeneous into a homoge-
neous equation by using a new variable (see Eq. 7.16). Look for a
(i) (i)
general solution for N1 (and subsequently also for N2 ) by using
the characteristic equation.
(f) How large is the fish population at the fixed point (if there is one)?
(∞) (∞)
In other words, calculate N1 + N2 .
Chapter 8

Models in Time and Space

Transport and transformation

After 1 week of marathon training. After 2 weeks of marathon training. After 3 weeks of marathon training.

8.1 Mixing and Transformation

Natural systems have a spatial structure, but until now, we either com-
pletely ignored it (as in the case of the one-box model) or we described it
in a highly simplified manner (as in the two-box model of a stratified lake,
Example 5.8).
In this chapter, we will discuss situations where continuity in time as
well as in space is important. In the most general case, this implies mathe-
matically describing the system variable Vi as a continuous function of the
spatial coordinates (x, y, z)1 and of time t:
Vi → Vi (x, y, z, t) (8.1)
Vi can represent any scalar quantity, for instance the three-dimensional
temperature distribution in the Atlantic Ocean as a function of time, or the

1 x, y, z are the three axes of a Cartesian coordinate system. Of course we

could also use different coordinates, such as the spherical coordinates r, θ, ϕ.

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 189

DOI 10.1007/978-3-642-30639-6_8, © Springer-Verlag Berlin Heidelberg 2013
190 MODELS IN TIME AND SPACE

north/south component of wind velocity in the troposphere. In the latter

case, the variable could be combined with the wind velocities along the
other two Cartesian coordinate axes (east/west and upwards/downwards)
to form the three-dimensional vector field of wind speed, which plays an
important role in weather and climate models.
In our examples, we will primarily deal with scalar fields (temperature,
concentrations, etc) rather than vector fields. In addition, we will concen-
trate on one-dimensional spatial models. For instance, we want to model
the dynamics of the vertical oxygen distribution in a lake C(z, t) (as shown
in Chap. 2, Fig. 2.6). Before we start building our model, however, we have
to make sure that it is justified to consider the concentration variations
along the vertical axis only (i.e. to neglect the concentration changes along
the horizontal axes x and y).
Usually, measurements show relatively similar O2 concentrations across
a lake at a given depth, even in lakes spanning several kilometers, whereas
the vertical changes vary significantly over depths as little as a few meters.
This suggests that it is justified to approximate the three-dimensional O2
distribution C(x, y, z, t) with a single2 vertical profile C(z, t), just as in
Fig. 2.6. Limnologists would explain the one-dimensional behavior of lakes
as follows. O2 production by photosynthesizing plankton and O2 consump-
tion due to respiration and decay of organic matter primarily depend on
the available light intensity, i.e. vary primarily with depth. Physics adds
another observation that completes the explanation: in lakes (as in the sea),
mixing processes along the horizontal axes are much quicker and more in-
tensive than in the vertical. Potential concentration differences along the
horizontal dimension are therefore evened out quickly, whereas they can
remain for a long time along the vertical dimension.
What makes sense for the lake example can be used as the basis for a
more general theory of the spatial structure of scalar fields. At each location
in a natural system, transport and transformation processes occur simul-
taneously.3 The relative speed of the two processes determines whether
spatial structures can form and remain, or whether transport processes
constantly counterbalance them and keep the system in a state of homo-
geneous mixing. In Fig. 8.1 we compare typical mixing times τmix with
transformation times τr . The former indicate, for instance, how long it
takes for a substance introduced to the atmosphere to vertically propagate
in the troposphere (vertical tropospherical mixing time). One could also
look at horizontal tropospherical mixing in the southern or northern hemi-
sphere, global tropospherical mixing, or the mixing between troposphere
and stratosphere. These mixing processes (ordered above by increasing du-
ration) are system-specific, i.e. they depend on the properties of the system
in which the mixing processes occur. In the examples above, they depend
on the mixing dynamics of the atmosphere.

2 Most appropriately an average profile, or one measured at the lake’s center.

3 Transport processes occur not only in fluid systems (atmosphere and hydro-
sphere), but also in solid materials, only much slower and therefore much less
obvious (e.g. molecular diffusion in solid matter).
 MIXING AND TRANSFORMATION 191

r mix Fig. 8.1: Schematic dis-

Transformation processes play of the relative mag-
at local equilibrium nitude of transformation
(spatial time τr and mixing time
inhomogeneities)
τmix . For τr  τmix , the

t
an
mix

v
system is spatially com-

le
re
pletely mixed (homoge-

e
Mixing time

ar
neous). For τr  τmix ,

n
io
at
all system components are

rm
r mix

fo
Rapid mixing locally at equilibrium but

ns
tra
d
(spatial inhomogeneities spatial inhomogeneities
in r
an
disappear) prevail. At the transition
ix x ∼
g
i

area (τr ∼ τmix ) both mix-

m
M

ing and transformation

determine the system’s
Transformation time r spatial structure

While the mixing time is system-specific, the transformation time4 τr

is a substance-specific quantity. For instance, τr could be defined by the
half-life τ1/2 of a radioactive isotope (as in Example 4.7) or the adjustment
time of a reactive chemical substance (as in Example 4.8). In a system
characterized by a given mixing time τmix , there can be substances with a
transformation time of τr either much larger, much smaller or of a similar
magnitude as τmix . In Fig. 8.1, these substances would be, respectively,
in the area of rapid mixing, the area of transformation processes at local
equilibrium, or in the transition area (shaded diagonal). Examples for the
diverse relations between transformation and transport are qualitatively
shown in Fig. 8.2.

Input Input Input

Substance A Substance B Substance C

Fig. 8.2: Spatial distribution of three substances A, B and C in a system with

rapid horizontal (h) and slow vertical (v) mixing (τmix,h  τmix,v ). The trans-
formation time of substance A is largest (low reactivity), that of substance C
smallest (high reactivity). Substance A is distributed almost equally across the
system, substance C very unequally. Substance B exhibits vertical inhomogeneity,
but is horizontally well mixed

4 The time needed to reach locally complete (or almost complete) reactive
equilibrium.
192 MODELS IN TIME AND SPACE

8.2 Advection, Diffusion and Exchange

In the following subsections, we will discuss the processes responsible for
transport and mixing and show how the mixing times (τmix ) introduced
above can be quantified.

8.2.1 Advection (Directed Transport)

There are essentially two types of transport processes: directed and undi-
Transport processes can rected (random) ones. The directed process is called advection. In a fluid
be directed or the result medium (e.g. water, air), advection results from flow. Flow can be quanti-
of many small random fied with the flow vector v = (vx , vy , vz ). All substances that are contained
movements. in a fluid element moved by flow are transported in the same direction,
that of the flow vector v (Fig. 8.3a). Therefore, we call advective transport
directed.

a b

Fad = C v

C Fdiff

Fig. 8.3: (a) The transport in a flow field (Fad , advective transport) is directed,
i.e. it has the same direction for all substances contained in a given volume of
fluid, and an absolute value that is proportional to the various substance con-
centrations C. (b) Diffusive transport results from a large number of random
individual movements. The net flux, Fdiff , is directed from the higher to the lower
concentration (Fick’s first law)

The advective mass flow Fad (transported mass per time and per area
perpendicular to the flow) is:

Fad = C v [M L−2 T −1 ] (8.2)

where C [ M L−3 ] is the concentration of the substance. This vector equa-

tion can also be written component-wise:

Fad,x = Cvx ; Fad,y = Cvy ; Fad,z = Cvz (8.3)

x, y, z are the Cartesian coordinates and Fad,x , vx , . . . the respective com-

ponents of the vectors Fad,x and v.
 ADVECTION, DIFFUSION AND EXCHANGE 193

The distance covered in time t by a fluid element and the substances

it contains is xad = vx t (analogous for the other coordinates). Strictly
speaking, however, this is only true if t is chosen small enough so that the
flow can be considered to be straight during this time. The mixing time by
advection in a system with the extent L can be written as:
L
τmix,ad = (8.4)
v
where for simplicity we have omitted the coordinate subscripts x, y or z.

Directed and undirected movement

First attempt Second attempt After the 10th attempt

8.2.2 Diffusion: Undirected Transport

Diffusive transport is a result of individual random movement. The best
illustration is the thermal movement of gas molecules. As an example, let’s
look at the behavior of benzene molecules which are inhomogeneously (i.e.
irregularly) distributed in a volume of air (Fig. 8.3b). Even though the
velocity vectors of the individual benzene molecules don’t point in the same
direction, for this particular situation they nonetheless cause a net benzene
transport from the left to the right side. This is simply because the benzene
concentration on the left side is higher, and therefore, the probability that
a molecule crosses the “virtual boundary” from left to right is higher than
the other way around.
Mathematically, this behavior is described by Fick’s first law. Written
for one spatial dimension (x-axis), it reads5 :
∂C
Fdiff , x = −Dx [M L−2 T −1 ] (8.5)
∂x
Dx [ L2 T−1 ] Diffusion coefficient in x direction
∂C −4
∂x [M L ] Partial derivative of substance
concentration in x direction

5 Because C depends on several coordinates (t, x, . . .), we use the symbol for
the partial derivative (∂) in the following discussion, to remind you that when
calculating the derivative the other coordinates are held constant.
194 MODELS IN TIME AND SPACE

Analogous equations are valid for y and z directions. If the cause of the
transport process is thermal movement of the molecules, Dx is called the
molecular diffusion coefficient. Usually, it is the same for all three dimen-
sions (Dx = Dy = Dz ≡ D). Thus the three components of Eq. (8.5) can
be summarized in one vector equation:
Fdiff = −D ∇C (8.6)
∇ is the gradient operator:
⎛ ⎞
∂C/∂x
∇C = ⎝ ∂C/∂y ⎠ (8.7)
∂C/∂z
Eq. (8.5) is called a flux-gradient model (as is Eq. 8.6, of course). The minus
sign shows that transport is always directed against the gradient, i.e. from
the higher to the lower concentration. Wherever undirected random pro-
cesses are involved, flux-gradient models appear, as for instance in heat
conduction (Fourier’s law).
Determining the mixing time τmix,diff for diffusion is somewhat more
difficult than for advection. We could ask ourselves, for instance, how a
concentration jump (a front) in a fluid behaves under the influence of diffu-
sion, or how quickly a substance initially concentrated in one spot spreads
through space (in one, two, or three dimensions). In all these cases we can
express the diffusive mixing distance with the following equation:
Ldiff = a(Dt)1/2 (8.8)
where the numerical factor depends on whether diffusion occurs along one,
two √
or three spatial dimensions. In the case of one-dimensional diffusion,
a = 2. This yields the famous Einstein-Smoluchowski relation6 :
Ldiff = (2Dt)1/2 (8.9)
Note that in contrast to advection, in which the distance
√ L grows linearly
with time t, diffusion only advances proportionally to t since it is undi-
rected and not “goal-oriented”. Solving Eq. (8.8) for t yields the diffusive
mixing time τmix,diff :
1 L2 L2
τmix,diff = = (8.10)
a2 D 2D
where the expression on the far right corresponds to the Einstein-Smolu-
chowski relation for one dimension.

8.2.3 Exchange at Boundary Layers

In Fig. 8.3b, we imply that mass flow takes place through an imaginary
boundary layer. In many situations, however, the boundary is physically

6 In Chap. 8.4.1 we will derive this relation for a special case (see Eq. 8.54)
 ADVECTION, DIFFUSION AND EXCHANGE 195

real, for instance between a liquid and a gas. Transport across this boundary
is called gas exchange. At the interface between the gaseous and liquid
phase, the concentration (expressed as mass per volume) exhibits a sudden
jump, i.e. a mathematical discontinuity, so that the concentration gradient
needed for the application of Eq. (8.5) becomes infinitely large.
We thus need a different approach. In Sect. 3.1 we showed that the
equilibrium between two chemical phases A and B can be described by a
static equilibrium relation of the following form (eq = equilibrium):
 
CA
KA/B = (8.11)
CB eq
eq
Thus, for a given CB there is a value CA = KA/B CB which is in equilibrium
with CB . If A signifies air (A = a) and B water (B = w), then KA/B is
the non-dimensional Henry coefficient Ka/w (Eq. 3.2).
The mass flow across a real boundary (such as the water surface) is
described by the following expression:
eq
FA/B = vA/B (CA − CA ) [M L−2 T −1 ] (8.12)

FA/B : Mass flux per area and time between

A and B (FA/B > 0, if net transport
from A to B)
vA/B [ L T−1 ]: Exchange velocity at the boundary
layer
eq
CA = CB KA/B [ M L−3 ]: Concentration in A that is at equi-
librium with the concentration in B,
CB (Eq. 8.11).

Equation (8.12) is the result of undirected transport as well. The molecules

(or other objects) cross the boundary layer in both directions. At equilib-
eq
rium (CA = CA ), the two fluxes compensate each other (FA/B = 0).
The exchange model Eq. (8.12) can also be applied to the flux bet-
ween two identical phases. In Example 5.8 we described the mass exchange
between the epilimnion and the hypolimnion of a stratified lake, with Qex
describing the water volume exchanged per time. If Qex is normalized by
the area A through which the exchange takes place, we get a quantity
with the dimension of a velocity which we call the exchange velocity: vex =
Qex /A [ L T−1 ]. In the following example, we will encounter yet another
application of the exchange model Eq. (8.12).

Example 8.1: Museum with a Roman gold treasure

In a museum, a separate room A showcases a famous Roman treasure of
gold coins. The room is only reachable from the neighboring room B
(see Fig. 8.4). On weekends, when visitor numbers peak, room A is
196 MODELS IN TIME AND SPACE

often hopelessly overcrowded. The museum’s director therefore decides

to analyze visitor behavior so that museum staff in room B can foresee
precarious conditions and take measures in time. The study shows the
following result:
(1) Each minute, 30 % of visitors in room B decide to enter the
treasure chamber A.
(2) Each minute, 10 % of visitors inside the treasure chamber (A)
decide to leave.
The goal is to prevent visitor numbers in A from exceeding the critical
value of NAcrit = 30.

Fig. 8.4: The flow of vis-

itors between exhibition
room A containing the Room B Roman Treasure
roman treasure, and room
B through which visitors
must pass, can be described
by an exchange model. NA ,
NB NA
NB = number of visitors
in the rooms A and B

Room A

We use kB NB to denote the transition rate of visitors from B to A, and

kA NA for A to B. NA and NB are the number of visitors in the respective
rooms. From observations (1) and (2) it follows that kB = 0.3 min−1 and
kA = 0.1 min−1 . The net flow of visitors from B to A is:

FB→A = kB NB − kA NA (8.13)

Room A is in equilibrium with room B if FB→A = 0, that is, for:

kB
NAeq = NB (8.14)
kA

We can replace NB in Eq. (8.13) with NAeq . It then follows that:

 
kA eq
FB→A = kB N − kA NA = kA (NAeq − NA ) (8.15)
kB A
 ADVECTION, DIFFUSION AND EXCHANGE 197

This relation has the same form as the exchange flow in Eq. (8.12). For
NA ≤ NAcrit to remain true, NAeq < NAcrit must hold, therefore:
kB
NAeq = NB ≤ NAcrit (8.16)
kA
This condition is fulfilled as long as the following is true:
kA crit 0.1
NB ≤ N = 30 = 10 (8.17)
kB A 0.3
If NB suddenly exceeds this value, the museum staff has time of about
τ ∼ 1/kA = 10 min to bar entry to room A.
It’s important to note that the formalism developed in Eqs. (8.13)–(8.17)
implies a precision and controllability which in reality does not exist. The
visitor behavior which we used to determine the model parameters kA and
kB is only meaningful in a statistical sense. The real values fluctuate around
their mean. We could describe this fluctuation by the standard deviations
σA and σB . When dealing with statistical fluctuations, absolute require-
ments such as “NA must not exceed NAcrit ” often cannot be fulfilled except
through absolute conditions (such as simply keeping room A or B locked).
For real (statistical) systems it is often more appropriate to work with
probabilistic requirements such as “NA shall not exceed NAcrit with a prob-
ability larger than p (e.g. p = 5 % or 0.05)”. The situation then becomes
a typical problem of probability theory. And finally, in order to ease the
demand on museum staff to manage visitor flows, we could of course also in-
stall light barriers at the treasure chamber’s entrance and exit. This would
allow us to constantly monitor NA and trigger a signal as soon as NAcrit is
exceeded.
Boundary layers

Typical case of an
exchange across the
boundary...

8.2.4 Gas Exchange

Fick’s first law (Eq. 8.5) and the exchange model (Eq. 8.12) are two differ-
ent mathematical approaches to the same physical phenomenon. In both
cases, the net flux results from the summation of many random individual
198 MODELS IN TIME AND SPACE

Fig. 8.5: Film model to

describe gas exchange of a z
Ca
volatile substance. See text
for explanations Air eq Ca
Cw = K a/ w
Molecular
boundary layer δ

Water
Cw

C (z)

processes. With the example of gas exchange we will now show that the
two models are not just physically but also mathematically related.
The so-called film model for the exchange of a volatile substance (sub-
stance with a large Henry coefficient) assumes that on the water side of
the water surface, a thin film exists (the molecular boundary layer with
thickness δ) through which dissolved substances can only travel by molec-
ular diffusion (Fig. 8.5). In the air and in the water beneath the boundary
layer, mixing is so rapid that we can assume the concentrations Ca and
Cw as spatially constant. At the very boundary between air and water, the
eq
two phases are at equilibrium (Cw = Ca /Ka/w , Ka/w = non-dimensional
Henry coefficient). Inside the boundary layer we observe a linear concen-
eq
tration profile from Cw to Cw . The mass flow, according to Fick’s first law,
is then as follows (see Eq. 8.5):

∂C C eq − Cw
Fw→a = −D = −D w
∂x δ (8.18)
D
= vw/a (Cw − Cw
eq
), vw/a =
δ
Thus, the exchange velocity vw/a can also be interpreted by the Fickian
diffusion model. A similar case will be discussed in Example 8.2. More about
modeling the exchange at boundary layers can be found in Schwarzenbach
et al. (2003).

8.2.5 Turbulent Diffusion

In fact, making the distinction between advective and diffusive transport
is not as easy as suggested so far. Flow processes in natural systems (such
as the atmosphere or hydrosphere) are almost always turbulent. Turbu-
lent flow—as opposed to laminar flow—consists of a superposition of flow
patterns with a range of different spatial and temporal structures. If, for
instance, wind velocity along a certain direction is continuously measured
 ADVECTION, DIFFUSION AND EXCHANGE 199

Table 8.1: Typical diffusion coefficients in the environment (From Schwarzen-

bach et al. 2003)

System Diffusion coefficient

(cm2 s−1 )a

Molecular
In water 10−6 –10−5
In air 10−1
Turbulent, in the ocean
Vertical, surface layerb 10−1 –104
Vertical, deep layer 1–10
Horizontal 10 –108
2

Turbulent, in lakes
Vertical, epilimnionb 10−1 –104
Vertical, hypolimnion 10−3 –10−1
Horizontalc 101 –107
Turbulent, in the atmosphere
Verticald 104 –105
In running water
Turbulent, vertical 1–10
Turbulent, lateral 10 –103
1

Longitudinal dispersion 10−5 –106

a
1 cm2 s−1 = 8.64 m2 d−1
b
Maximal values
c
Horizontal diffusion depends on the considered spatial extent
d
In the atmosphere, horizontal transport primarily occurs by advection
(wind)

at steady state as a function of time, the average wind velocity v̄ can be

determined for different averaging intervals (e.g., 1 min, 5 min, 1 h, etc.). At
a given point in time, the actual wind velocity v(t) deviates by v  (t) from
v̄: v(t) = v̄ + v  (t).7 The turbulent velocity component v  (t) fluctuates in an
apparently random manner around an average value of 0. Note that the sep-
aration between average velocity and fluctuation depends on the duration
of the averaging interval and is therefore to a certain extent arbitrary.
Due to the stochastic (that is, unpredictable) nature of v  , the advective
transport equation (8.2) can only be evaluated if we limit ourselves to the
7 v̄ is, of course, time-dependent too, but due the averaging, its variation is

smoother than that of v  .

200 MODELS IN TIME AND SPACE

average flow. It can be shown that the effect of the fluctuation v  on the
mass flow results in an additional undirected transport. Therefore it often
makes sense to describe this so-called turbulent mass flow in the same form
as Eq. (8.5):
∂C
Fturb,x = −Kx (8.19)
∂x
The molecular diffusion coefficient is replaced by the much larger turbulent
diffusion coefficient K. Because the effect of turbulence is usually more
pronounced along the horizontal than the vertical axes (see Sect. 8.1), it
makes sense to add a subscript to K to indicate along which axis turbulent
transport is being considered (i.e. Kx , Ky or Kz ).
The molecular diffusion coefficient is a property of the diffusing sub-
stance and the fluid in which diffusion takes place. It’s usually listed in
physical or chemical manuals. In contrast, the turbulent diffusion coeffi-
cient depends on the medium (lake, ocean, atmosphere, etc.) and on time.
For instance, it makes a significant difference whether we look at turbulent
diffusion in the atmosphere during a calm or windy period, a thunderstorm
or a tornado. Because of the arbitrary distinction between average flow and
turbulence resulting from the averaging interval, the horizontal diffusion in
ocean and atmosphere depends on the spatial extent of the diffusion pro-
cess considered. The figures given in Table 8.1 are typical values; the actual
quantitative determination of turbulent diffusion coefficients can be a very
time-consuming procedure. More about this can be found in Schwarzenbach
et al. (2003), as well as a detailed explanation of the related phenomenon of
dispersion, which appears along the direction of advective flow (for instance
in flowing waters).

Example 8.2: Turbulent exchange through the thermocline

Turbulent
boundary layer!
Figure 8.6 shows the vertical concentration profile of dissolved molecu-
lar oxygen (O2 ) measured in June in a small lake with a depth of 25 m.
Because of measured biomass sedimentation, we estimate that below
10 m the O2 consumption rate per area and time is R = 1.5 g m−2 d−1 .

From temperature measurements, scientists studying the lake have de-

termined a vertical turbulent diffusion coefficient of Kz = 0.6 m2 d−1
for the thermocline (the layer between 5 and 10 m depth). In greater
depths, Kz is about ten times as large, in the layer above 5 m even a
hundredfold larger.

Estimate whether the vertical O2 transport can compensate the O2

loss and if not, how long it will take for the deep water to become
completely anoxic (O2 -free). The current stock of O2 in the deep water
is M  = 30 g m−2 .
 ADVECTION, DIFFUSION AND EXCHANGE 201

O 2 concentration [ mg/L ] Fig. 8.6: Vertical profile

6 12
of dissolved molecular
oxygen (O2 ) concentration
A
in a lake (see Example 8.2
5m and Problem 8.2)

10 m
Depth z B

Oxygen consumption rate R

O2

25 m
Lake bed

The bottleneck for vertical O2 transport is the thermocline; transport is

significantly faster above and below. The flux of O2 across the thermocline
per area and time can be calculated from Eq. (8.18):

Kz
Fthermocline = (CA − CB ) = vex (CA − CB )
δ

With:
Kz = 0.6 m2 d−1
δ = 5m
vex = Kz /δS = 0.12 m d−1
CA = 12 mg/L = 12 g m−3
CB = 6 g m−3
(CA , CB : O2 concentration at A and B)
Thus: Fthermocline = 0.12 m d−1 × 6 g m−3 = 0.72 g m−2 d−1

In the deep water, a net O2 loss of (R − Fthermocline ) ∼ 0.8 g m−2 d−1

occurs. The O2 stock M  = 30 g m−2 only lasts for about another 5 weeks.
Note that if the O2 reserves in the deep water decline, the concentration
at the lower boundary of the thermocline changes too, and with it, the flow
through the thermocline. If CA remains constant, Fthermocline takes on the
maximum value (for CB = 0) of:

Fthermocline (max.) = vex CA = 0.12 m d−1 × 12 g m−3 = 1.44 g m−2 d−1

Only after all oxygen in the deep water has completely disappeared will
Fthermocline be of similar magnitude as the loss rate R. In Problem 8.2 we
will deal with the question whether CA does indeed remain constant.
202 MODELS IN TIME AND SPACE

8.3 Steady-State Transport/Transformation Models

So far, we looked at transport with equations describing mass flow and
mixing time. In this section, we want to combine transport processes with
transformation processes in order to construct integrated transport and
transformation models.

8.3.1 The Divergence Theorem: From Flux to Local Change

To get from mass fluxes to the local change of a system property (such
as concentration or temperature), we consider the test volume shown in
Fig. 8.7. It has a length of Δx along the x-axis; its frontal area A is perpen-
dicular to the x-axis. We assume that mass transport in y and z direction is
zero. Furthermore, the property under consideration in the test volume is
conservative, i.e. its transformation rate is zero. Then, the balance equation
for the test volume V = AΔx is:
∂M
= AFx (x) − AFx (x + Δx) (8.20)
∂t
Where:
M : Total value of the property (e.g., mass, energy, etc.) in V
Fx (x), Fx (x + Δx): Property flux (e.g., mass or energy flux)
per area and time in x direction at location x and x + Δx, respectively.

Fig. 8.7: The divergence Area A

theorem establishes a link z Fx (x + Δ x)
between flux and local
change. See text for fur-
ther explanations
Area A
x
x + Δx

Fx (x) Δx

y
x

The average property density (e.g., concentration) in V is C = M/V . For

a constant V , the following holds:
∂C ∂C
V = AΔx = A(Fx (x) − Fx (x + Δx))
∂t ∂t
After dividing by AΔx:
∂C Fx (x) − Fx (x + Δx) Fx (x + Δx) − Fx (x)
= =−
∂t Δx Δx
 STEADY-STATE MODELS 203

If we now take the limit Δx → 0, the right-hand side will be the negative
derivative of Fx with respect to x:
∂C ∂Fx
=− (8.21)
∂t ∂x
In the same way we can calculate the effect on C of the other flow compo-
nents (Fy , Fz ). This ultimately gives us the following result:
 
∂C ∂Fx ∂Fy ∂Fz
=− + + = −∇F (8.22)
∂t ∂x ∂y ∂z
On the right-hand side is the so-called divergence of the transport
vector F .8 Equation (8.22) is the divergence theorem. It states that the
change of a scalar quantity (concentration C) inside an infinitely small test
volume is equal to the divergence of the vector F describing the flux of
this quantity. In other words: if we know all the fluxes at the surface of
a volume, we know what happens inside that volume (of course assuming
that the volume contains no sources or sinks).
It is remarkable that, to derive the divergence theorem in Eq. (8.22),
we didn’t need any information about how the flux came to be and how
it is mathematically formulated. The theorem holds for any conservative
vector field. We now want to apply it to the advective and diffusive flow
equation9 :
For advection (Eq. 8.3):
 
∂C ∂ ∂ ∂
= − (Cvx ) − (Cvy ) − (Cvz )
∂t ad ∂x ∂y ∂z
  (8.23)
∂vx ∂vy ∂vz ∂C ∂C ∂C
= −C + + − vx − vy − vz
∂x ∂y ∂z ∂x ∂y ∂z
It can be shown that for a non-compressible fluid, the divergence of the
flow field is zero, so the expression in parentheses in Eq. (8.23) disappears.
What remains is:
 
∂C ∂C ∂C ∂C
= −vx − vy − vz (8.24)
∂t ad ∂x ∂y ∂z
For diffusion (Eq. 8.6):
       
∂C ∂ ∂C ∂ ∂C ∂ ∂C
=− −Dx − −Dy − −Dz
∂t diff ∂x ∂x ∂y ∂y ∂z ∂z
∂2C ∂2C ∂2C
= Dx + D y + D z
∂x2 ∂y 2 ∂z 2
∂Dx ∂C ∂Dy ∂C ∂Dz ∂C
+ + +
∂x ∂x ∂y ∂y ∂z ∂z
(8.25)

8 Note the different meaning of the symbol ∇ in Eqs. (8.22) and (8.6). In the
latter case, ∇ creates a vector (the gradient vector) from a scalar (C), in the
former case, a scalar (the divergence) from a vector(F ).
9 To do so, we need the differentiation rule for a product of two functions.
204 MODELS IN TIME AND SPACE

If the diffusion coefficients are spatially constant, the last three terms of
Fick’s second law describes Eq. (8.25) disappear. The first three terms then form Fick’s second law of
the diffusion-induced diffusion. Let’s look at the case of isotropic diffusion (Dx = Dy = Dz ≡ D,
change of concentration i.e. diffusion that is equal in all three spatial directions, which is usually
over time. the case for molecular but not turbulent diffusion). In this case, Fick’s
second law takes on the following form, in which it usually appears in the
literature:  2 
∂C ∂ C ∂2C ∂2C
=D + + (8.26)
∂t ∂x2 ∂y 2 ∂z 2
Table 8.2 gives an overview of the effects of diffusion and advection on the
flow and local change of a scalar (such as concentration or temperature).

Table 8.2: The effect of diffusion and advection on the flow or the local change of
a property. The equations are formulated for the case of a concentration, C. Note
that when moving through the table each step to the right or downwards results
in a change of sign and an increase of the derivative by one order

Flow F Local concentration change

[ M L−2 T−1 ] ∂C
∂t [ ML
−3 −1 a
T ]

Advection +vx C −vx ∂C

∂x

2
Diffusion −Dx ∂C
∂x +Dx ∂∂xC2

a
For vx , Dx = constant. Analogous equations exist for the other spatial
coordinates y and z

8.3.2 Transport and Transformation: Steady State

We consider a one-dimensional system with diffusion, advection and trans-
formation. The local concentration change is (if Dx , vx = constant):
     
∂C ∂C ∂C ∂C
= + +
∂t ∂t diff ∂t ad ∂t transformation
2
  (8.27)
∂ C ∂C ∂C
= Dx 2 − vx +
∂x ∂x ∂t transformation

If we limit ourselves to transformations of zero and first order we get10 :

∂C ∂2C ∂C
= Dx 2 − vx − kr C + J (8.28)
∂t ∂x ∂x

10 It
makes sense to write the zero-order reaction as a positive and the first-
order reaction as a negative function.
 STEADY-STATE MODELS 205

This is a linear partial second-order differential equation. Its solutions de-

pend on the initial and boundary conditions. We will discuss examples in
Sect. 8.4 and in Appendix E. For the moment, we want to look at the steady
state. We determine it by setting the left side of Eq. (8.28) equal to zero.
Because only the derivatives with respect to the spatial coordinate x re-
main, the equation becomes a regular differential equation (except that we
are now looking at derivatives with respect to space rather than time).
The steady-state solution of Eq. (8.28) is described by:

d2 C dC
Dx 2
− vx − kr C + J = 0 (8.29)
dx dx

Note that we replaced the partial with the regular derivative.

Equation (8.29) is a (regular) linear, inhomogeneous second-order dif-
ferential equation. We can transform it into a system of coupled linear first-
order differential equations. In Sect. 5.1.1, we learned that, if the coefficients
(Dx , vx , kr , J) are constant, the solutions generally consist of two exponen-
tial functions and a constant term. You can find the result in Appendix C.3.
Applied to Eq. (8.29), we get:

J
C(x) = A1 eλ1 x + A2 eλ2 x + (8.30)
kr

The λi (i = 1, 2) are again the eigenvalues of the system. As in Chaps. 5 and

7 (Eqs. 5.11 and 7.21), we determine them from the characteristic equation.
In the case of a second-order linear system, the characteristic equation is a
second-order equation, in other words a quadratic one:

Dx λ2i − vx λi − kr = 0 (8.31)

Its solutions are:

1  
λi = vx ± (vx2 + 4Dx kr )1/2
2Dx
#  1/2 $ (8.32)
vx 4Dx kr
= 1± 1+ [ L−1 ]
2Dx vx2

The eigenvalues λi determine the spatial structure of the solution, for in-
stance the behavior of C(x) for x → ± ∞. Their dimension is [ L−1 ]. Note
that the eigenvalues don’t depend on the inhomogeneous term J.
The coefficients Ai depend on the boundary conditions of the system.
A second-order equation needs two boundary conditions for its solution to
be unique. If we are looking for the solution of Eq. (8.29) for a finite interval,
e.g. on the x-axis between x = 0 and x = xL , we have six possibilities to
206 MODELS IN TIME AND SPACE

specify the boundary conditions. We can either directly specify the two
boundary values:
C0 ≡ C(x = 0) ; CL ≡ C(x = xL ) (8.33)
Or we can specify the first derivatives at the boundaries:
 
 dC   dC 
C0 ≡ ; CL ≡ (8.34)
dx x=0 dx x=xL
Finally, we can take any combination of value and derivative. If the system
is unbounded (xL → ∞), the coefficient Ai in Eq. (8.30) which is associated
with an eigenvalue λi with a positive real part must be zero. Otherwise,
the solution C(x) becomes infinite for x → ∞. Explicit expressions for Ai
for all these possibilities are compiled in Table 8.3.

Table 8.3: Coefficients Ai of the solution of the steady-state linear trans-

port/transformation equation (Eqs. 8.29 and 8.30) for different boundary con-
ditions at x = 0 and x = xL .

Boundary conditionsa A1 A2

(CL − kJr )−(C0 − kJr )eλ2 xL −(CL − kJr )+(C0 − kJr )eλ1 xL
(1) C0 , CL eλ1 xL −eλ2 xL eλ1 xL −eλ2 xL


−C0 eλ2 xL −CL 
+C0 eλ1 xL
(2) C0 , CL
CL
λ1 (eλ1 xL −eλ2 xL ) λ2 (eλ1 xL −eλ2 xL )

C0 −λ2 (C0 − kJr ) −C0 +λ1 (C0 − kJr )

(3) C0 , C0 λ1 −λ2 λ1 −λ2


CL −(C0 − kJr )λ2 eλ2 xL 
−CL +(C0 − kJr )λ1 eλ1 xL
(4) C0 , CL λ1 eλ1 xL −λ2 eλ2 xL λ1 eλ1 xL −λ2 eλ2 xL

If the system is unbounded on one side (e.g. xL → ∞), the solution

is determined by a single boundary condition, either C0 or C0 :

(5) C0 0 C0 − J/kr

(6) C0 0 C0 /λ2

a
See definitions in Eqs. (8.33) and (8.34). Note that the two missing com-
binations (CL , CL and C0 , CL ) follow from the other four simply by ex-
changing the boundaries (x = 0 becomes x = L and vice versa)
 STEADY-STATE MODELS 207

Example 8.3: Radon-222 in the ocean

The decay of radium-226 (226 Ra) in deep sea sediments produces the
radioactive noble gas radon-222 (222 Rn, half-life τ1/2 = 3.8 d). Radon
diffuses from the sediment surface into the water with a flux F (con-
stant per time and area). Once in the water, it is distributed by turbu-
lent diffusion. We are interested in the steady-state radon profile as a
function of the vertical distance h from the sediment C(h). JRa is the
production of 222 Rn in the water column per volume and time from
the decay of dissolved 226 Ra. Because as a first approximation we can
describe the sediment surface as an infinitely large horizontal plain, the
horizontal concentration gradients and therefore the horizontal diffu-
sive fluxes are zero. The average vertical advective flow vz is also zero,
since the sediment is neither a source nor a sink of water.

With vz = 0, and according to Eq. (8.29), the steady-state vertical trans-

port/transformation equation for the concentration C of 222 Rn is:
∂C ∂2C
= Kz 2 − λRn C + JRa = 0 (8.35)
∂t ∂h

h(m): Height above sediment

Kz (m2 d−1 ): Vertical turbulent diffusion coefficient (instead of Dz )
kRn (d−1 ): Decay constant of 222 Rn (see Eq. 4.45)
kRn = ln 2/τ1/2 = 0.693/3.8 d = 0.18 d−1
C(Bq m−3 ): Activity of 222 Rn in the water
(1 Bq = 1 Becquerel = 1 decay per second)
JRa (Bq m−3 s−1 ): In-situ production of 222 Rn through decay of 226 Ra

The eigenvalues of Eq. (8.35) follow from Eq. (8.32):

λi = ±(kRn /Kz )1/2
Since 222 Rn decays rather rapidly compared to the vertical mixing time in
the deep ocean, the water column can be seen as infinitely deep. Thus A1
(corresponding to the positive eigenvalue λ1 ) in Eq. (8.30) must be zero. To
determine the remaining coefficient A2 , we express the vertical 222 Rn flux
at the sediment surface (h=0), FRn , in terms of Fick’s first law (Eq. 8.5):

∂C  FRn
=− (8.36)
∂h h=0 Kz
According to Table 8.3 (case 6) and Eq. (8.30), the solution is11 :
JRa FRn −(kRn /Kz )1/2 h
C(h) = − e
kRn Kz λ2
(8.37)
JRa FRn 1/2
= − e−(kRn /Kz ) h
kRn (Kz kRn )1/2

11 Note that λ2 = −(λRn /Kz )1/2 .

208 MODELS IN TIME AND SPACE

The 222 Rn activity above the sediment falls exponentially towards the
constant value C∞ = JRa /kRn . In Problem 8.4 we will look into how 222 Rn
measurements can be used to determine the vertical turbulent diffusion
coefficient in the ocean.

8.3.3 Spatial Structures: Péclet und Damköhler Numbers

We want to bring out the influence of the eigenvalues λi on the spatial
structure of the solution of Eq. (8.29) more clearly. To do so, we intro-
duce the non-dimensional spatial coordinate ξ = x/xL which normalizes
the distance between the two boundaries x = {0, xL } onto the inter-
val ξ = {0, 1}.12 To write the solution (8.30) in terms of the new spa-
tial coordinate, we only have to multiply the eigenvalues with xL , since
λi x = λi ξ:
#  1/2 $
xL vx 4Dx kr
λi = λi xL = 1± 1+ [-] (8.38)
2Dx vx2

The λi are the non-dimensional eigenvalues. The solution of Eq. (8.30) is
now:
  J
C(ξ = x/xL ) = A1 eλ1 ξ + A2 eλ2 ξ + (8.39)
kr

Since ξ only varies between 0 and 1, the size of λi determines the influence
of the spatial functions exp(λi ξ) on the shape of C(ξ). For instance, if
|λi | 1, then exp(λi ξ) barely varies with ξ and is about 1 everywhere. If,
on the other hand, (−λi ) 1, then exp(λi ξ) drops off from 1 (at ξ = 0)
to 0 over a very short distance. In other words, the spatial variation of this
term is limited to a small zone close to the boundary while over the rest of
the interval the term is zero.
To systematically discuss the different patterns of spatial behavior of
Eqs. (8.30) and (8.39), we rewrite the non-dimensional eigenvalues in terms
of two non-dimensional numbers, Pe and Da:

Pe
λi = [sgn(vx ) ± (1 + 4 Da)1/2 ], i = 1, 2 (8.40)
2

Here, sgn(vx ) = is the sign of vx . Pe and Da are defined as follows:

|vx |xL
Pe ≡ ≥0 Péclet number (8.41)
Dx

12 Note that the following considerations are only valid if the extent of the sys-
tem is finite, i.e. if it is enclosed by finite boundaries.
 STEADY-STATE MODELS 209

Dx kr
Da ≡ ≥0 Damköhler number13 (8.42)
vx2

In Problem 8.5 we will show that the Péclet number Pe can be seen as a The Péclet and Damköhler
measure for the relative velocity of advection versus diffusion in the interval numbers measure the rel-
xL (for Pe 1, advection is faster than diffusion and vice versa). The ative influence of directed
Damköhler number Da measures the ratio between diffusive and advective transport, diffusive trans-
transport during the average life span of the reactive substance, kr−1 . port, and transformation.
Figure 8.8 gives an overview of the influence of Pe and Da on λi and
schematically shows the shape of the steady-state concentration profile
C(x) between x = 0 and x = xL . We can divide the extreme cases into four
groups (explained in more detail on the next page, following Fig. 8.8).

CL
slow A: purely diffusive
2 reaction
k r <<D/L (with little reaction)
C0
diffusion
0 xL
& reaction
CL
vx2 << Dkr otherwise
fast
(Da>>1) B: reactive-diffusive
reaction
C0
transport 0 xL
& reaction
CL
slow C: purely diffusive
vx2 >>
Dkr (with little advection)
|vx |<< D/L advection
(Da<<1) C0
(Pe<<1)
diffusion 0 xL
& advection
CL
otherwise vx<0 D: advective-diffusive
fast
advection vx >0 (diffusion at boundaries)
C0
0 xL

Fig. 8.8: One-dimensional diffusion/advection/transformation profiles at steady-

state: the non-dimensional Péclet and Damköhler numbers determine both the
form of the profile and the influence of diffusion (coefficient Dx ), advection
(velocity vx ) and the specific reaction (rate kr ). See text for further explanations

If Da 1 (i.e. vx2 Dx kr ), according to Eq. (8.40) the non-

1/2
dimensional eigenvalues are virtually independent of vx : λi ∼ Pe(Da) =
1/2
(kr /Dx ) xL . Therefore we call this the diffusion and reaction regime. It
can be further subdivided:

13 Inthe literature, we find different types of Damköhler numbers, see e.g.

Domenico and Schwartz (1998). For our purposes, the definition in Eq. (8.42)
is the adequate form.
210 MODELS IN TIME AND SPACE

Case A: Purely Diffusive Regime: Da 1 and kr Dx /x2L

The reactivity is too slow to significantly alter the linear diffusion profile
(which therefore looks similar to e.g. Fig. 8.5).

Case B: Reactive-Diffusive Regime: Da 1, kr is not Dx /x2L

The time needed for diffusion over the distance xL is large enough to allow
for a significant reaction during diffusive transfer (Example 8.3 belongs to
this category).
If Da 1 (i.e. vx2 Dx kr ), the substance behaves approximately
conservative (kr ∼ 0). In this case, which we call the diffusion and ad-
vection regime, the size of the Péclet number decides whether diffusion or
advection is the dominant transport process.

Case C: Purely Diffusive Regime: Da 1, Pe 1

The substance is quasi-conservative throughout its transport (Da 1); the
transport process is determined by diffusion (Pe 1).

Case D: Advective-Diffusive Regime: Da 1, Pe not 1

As in case C, but advection dominates except near the boundaries.
To summarize, we can say that the structure of a concentration pro-
file is determined by the eigenvalues of the linear second-order differential
equation (8.29). The eigenvalues themselves are determined by Eq. (8.32),
which depends on the three coefficients Dx , vx und kr . kr is a substance-
dependent property, vx a property of the system, and Dx generally a prop-
erty of both. This means that in a given system with its distinct transport
regime, different substances experience different transport/transformation
ratios. In a system with (substance-independent) turbulent diffusion, the
Péclet number is only a property of the system. The following example
illustrates this.

Example 8.4: Vertical transport in the ocean

The quicker I
decay, the less In the Atlantic and Pacific oceans, there is a depth zone (typically
deeply I sink!
between 1 and 5 km depth, see Fig. 8.9), in which water flows upward
extremely slowly (called upwelling). This happens as a compensation
for deep water formation in the North Atlantic and in the Antarc-
tic Circumpolar Current. This depth zone is also affected by vertical
turbulent diffusion, compared to which molecular diffusion can be ne-
glected.
 STEADY-STATE MODELS 211

Water surface Fig. 8.9: A simple model

0m of the vertical water col-
Surface water (C SW) umn in the Atlantic and
1 km
Pacific Oceans. The deep
water is bounded by the
Depth bottom water and the sur-
face water. Water slowly
Deep water moves upwards through
the deep water layer
(upwelling), while turbu-
5 km lent mixing takes place
simultaneously
Bottom water (C BW)

Sediment

We consider the steady-state vertical profiles of two dissolved radioac-

tive isotopes, radium-226 (226 Ra, τ1/2 =1,600 year) and the heavy
hydrogen isotope tritium (3 H, τ1/2 = 12 year). We assume that the
concentrations at the lower deep water boundary are controlled by the
bottom water, and at the upper deep water boundary by the surface
water. These concentrations are constant. Radium diffuses out of the
sediment, its concentration in the surface water is zero (CSW = 0).
Tritium enters the surface water via the atmosphere and then diffuses
downwards. In the bottom water there is no tritium (CBW = 0).a
a Of course, these assumptions strongly simplify the real picture. In par-
ticular, atmospheric tritium concentration, primarily determined by hu-
man activities, has changed significantly over the recent past. Further-
more, we don’t take into account the production of radium from the decay
of uranium in the water column (i.e. J = 0).

Numerical values used:

Kz = 1 cm2 s−1 = 3.2 × 103 m2 year−1 Vertical turbulent diffusivity

vz = 2 m year−1 Speed of upwelling
kRa = 0.693/1, 600 year = 4.3 × 10−4 year−1 Decay constant of 226 Ra
kT = 0.693/12 year = 0.058 year−1 Decay constant of tritium

Coordinate system: positive upwards, ξ = 0 at the lower boundary, ξ = 1

at the upper boundary, xL = 4 km = 4 × 103 m.

Radium-226
We normalize the concentrations with the lower boundary concentration
CBW . In other words, we set C0 = C(ξ = 0) = 1, CL = 0. We then
212 MODELS IN TIME AND SPACE

Fig. 8.10: Relative ver- Relative concentration

0 1
tical concentration pro- ξ =1 1km
file of two radionuclides Tritium
in the deep water of
the Pacific Ocean (see
Fig. 8.9). The short-lived

Standardized depth ξ
tritium (τ1/2 =12 year) 226
Radium

Absolute depth
is barely influenced by
the upwelling. For the
long-lived radium-226
(τ1/2 =1,600 year)
however, there is a clear
difference between the
profiles with and without 226
Radium
upwelling without upwelling

ξ =0 5km
0 1
Bottom water

calculate the non-dimensional parameters:

vx xL 2 m year−1 × 4 × 103 m
Pe = = = 2.5
Kz 3.2 × 103 m2 year−1
Kz kRa 3.2 × 103 m2 year−1 × 4.3 × 10−4 year−1
Da = = = 0.344
vz2 (2 m year−1 )2
Eigenvalues : λ1 = 3.18, λ2 = −0.68

From Table 8.3, case (1) with C0 = 1, CL = 0, J = 0:

A1 = −0.0217, A2 = 1.0217

The normalized radium profile thus has the form:

CRa (ξ) = −0.0217e3.18ξ + 1.0217e−0.68ξ (8.43)

The profile is shown in Fig. 8.10. To illustrate the effect of upwelling, the
corresponding profile without upwelling (vz = 0) is included in the figure.
It becomes clear that, due to upwelling, at a given depth the concentration
is shifted to higher values.

Tritium (3 H)
Boundary conditions: C0 = 0, CL = 1

Pe = 2.5 (as for 226

Ra)
3.2 × 10 m year−1 × 0.058 year−1
3 2
Da = = 46.4
(2 m year−1 )2
 TIME-DEPENDENT SOLUTIONS 213

Eigenvalues14: λ1 = 18.3, λ2 = −15.8

1 −λ  
A1 =  ∼ e (note: eλ1 eλ2 )
1

−e
eλ1 λ2

A2 = −A1 ∼ −e−λ1
The standardized profile therefore has the form:
      
CT (ξ) = e−λ1 eλ1 ξ − e−λ1 eλ2 ξ ∼ e−λ1 eλ1 ξ = e−λ1 (1−ξ)
(8.44)
= e−18.3(1−ξ)
The profile corresponds to an exponential curve that rapidly declines in
negative ξ direction, with the boundary value CT (ξ = 1) = 1. The influence
of upwelling on the tritium profile is weak.
Sustaining the radium and tritium profiles at a steady state implies that
there is a continuous flux of isotopes from and to the deep water across its
upper and lower boundaries. In Problem 8.8 we will discuss the relative size
of these boundary fluxes and compare them to the loss of isotopes due to
radioactive decay.

8.4 Time-Dependent Solutions of the Transport/Transformation

Equation
As we already noted at the outset of Chap. 8, spatially continuous models
lead to partial differential equations. These equations can only be solved
algebraically in special cases. The dynamic equations for fluids,15 the so-
called Navier-Stokes equations, which are particularly important for natural
systems, unfortunately don’t belong to the category of generally solvable
partial differential equations. In fact, the numerical analysis of the coupled
ocean-atmosphere system is one of the major challenges for computational
science and supercomputing. Thus, neither the weather nor the climate are
easily predictable.
Dealing with such models lies beyond the scope of this book. However,
we don’t want to conclude this chapter without at least throwing a glance
at some specific time-dependent solutions of the transport equation. We
will begin with the one-dimensional case.

8.4.1 The One-Dimensional Diffusion Equation

We consider the one-dimensional diffusion equation with constant diffu-
sion coefficient Dx . This equation results from Eq. (8.26) by dropping the
derivatives with respect to y and z (Fick’s second law):
∂C ∂2C
= Dx 2 (8.45)
∂t ∂x
14 Since the Damköhler number Da  1, the eigenvalues would approximately
be λi = ± xL (kr /Kz )1/2 = ±17.
15 “Fluid” is used here as a general expression for liquid and gaseous systems.

The most important fluid systems on Earth are the ocean and the atmosphere.
214 MODELS IN TIME AND SPACE

The solutions of Eq. (8.45) depend on the initial and boundary conditions.
We assume that the system is infinite in both directions along the x-axis,
i.e. the substance described by the concentration C can spread along the
x-axis from −∞ to +∞. We’ll now look at two different initial conditions.

(1) The Delta Function

At time t = 0, the substance mass m is concentrated entirely at point
x = 0.16 To concentrate a finite mass onto a point with (by definition) no
spatial extent, mathematicians created the delta function δ. It is defined as
follows:
 ∞
0 for x = 0
δ(x) = where δ(x) dx = 1 (8.46)
∞ for x=0
−∞

The initial condition of Eq. (8.45) is thus:

C(x, t = 0) = m δ(x) (8.47)

We now claim that the solution of Eq. (8.45) with the initial condition
Eq. (8.47) has the following form:
 
m x2
C(x, t) = exp − (8.48)
2(πDx t)1/2 4Dx t

You can verify that this is indeed a solution of Fick’s second law by inserting
Eq. (8.48) into Eq. (8.45) (see Problem 8.9). C(x, t) is a normal distribution
with the variance:
σx = (2Dx t)1/2 (8.49)

The normal distribution (also called Gaussian distribution), a proba-

bility distribution often used in statistics for random variables distributed
around an average value, is defined in Appendix E.1. The total substance
mass (calculated from the integral of C(x, t) between x = −∞ and +∞)
is m, i.e. constant with respect to time. This result is not entirely surprising;
we already saw in Example 2.7 (the bed of nails) that random processes
lead to a normal distribution whose standard deviation increases with time.
For t = 0, σx and the denominator of Eq. (8.48) become zero, i.e.
C(x, t = 0) describes a normal distribution with variance σx = 0. The
delta function introduced in Eq. (8.46) can be defined as the limit of a nor-
mal distribution with σx → 0, thus, Eq. (8.48) does indeed fulfill the initial
condition of Eq. (8.46).

16 Note that for the one-dimensional case, m has the dimension [M L−2 ] (mass
per area perpendicular to the x-axis).
 TIME-DEPENDENT SOLUTIONS 215

Example 8.5: Substance spreading in a canal

Due to an accident, a certain mass M of the herbicide atrazine enters
an abandoned canal with stagnant water. It rapidly expands across
the canal’s cross section (width w, average depth h). Mixing along
the canal axis takes place by turbulent diffusion (turbulent diffusion
coefficient Kx ). How long will it take for the atrazine concentration
C(x, t) to drop below the tolerance value for drinking water Ctol along
the entire length of the canal?
Numerical values: M = 1 kg, Ctol = 0.1 µg/L = 10−4 g m−3
w = 50 m, h = 8 m, Kx = 0.1 m2 s−1

We assume that the atrazine enters the canal at x = 0 and neglect the
time needed to distribute it homogeneously across the canal’s cross-section.
Its further spread can be described by Eq. (8.45) and the initial condition
Eq. (8.47), where the initial mass per cross-sectional area is:

M 103 g
m= = = 2.5 g m−2
wh 50m × 8m

Note that the concentration maximum remains at the location of inflow

x = 0, where its time-dependent value is:
m
C(x = 0, t) =
2(πKx t)1/2

We are looking for the time t0 at which C(x = 0, t0 ) = Ctol :

m
Ctol =
2(πKx t0 )1/2

Solving for t0 yields:

m2 (2.5 g m−2 )2
t0 = =
2
4πKx Ctol 4π × 0.1 m2 s−1 (10−4 g m−3 )2

= 5.0 × 108 s ∼ 16year (!)

This unrealistically long time indicates that the herbicide’s disappearance is

probably not determined by diffusion but rather by other processes (chem-
ical conversion, sorption on suspended particles and the sediment surface).
In Example 8.7 we will show that diffusive dispersion in two and three
dimensions radically reduces the diffusion time t0 .
We can easily generalize the solution of this problem to the case where
in addition to horizontal diffusion, a longitudinal flow with velocity vx
exists. The maximum of the concentration distribution shifts according to
216 MODELS IN TIME AND SPACE

Fig. 8.11: Effects of C C

advection and diffusion
t0
on concentration: advection
moves the center of the t1
curve without changing its t2
t0 t1 t2
form (left), while diffusion
broadens the distribution x x
without moving its cen-
ter. Because the underlying
equations are linear, the
two processes can be super-
imposed

the equation xm = vx t, while the distribution relative to xm keeps the form

of Eq. (8.48). Written out, the solution is (see Problem 8.9):
 
m (x − vx t)2
C(x, t) = exp − (8.50)
2(πDx t)1/2 4Dx t
Note that the flow has no influence on the time calculated in Example 8.5,
but only moves the location of the concentration maximum along the x-axis.
The effects of advection and diffusion are compared in Fig. 8.11.

(2) Diffusion with Constant Boundary Concentration

A second type of solution for Eq. (8.45) results if the initial concentra-
tion is a step function (see Fig. 8.12):

C0 for x ≤ 0
C(x, t = 0) = (8.51)
0 for x > 0

It describes an initial situation with two regions separated by a boundary:

a region with constant concentration C0 on the left-hand side and a region
with zero concentration on the right-hand side. At time t = 0, transport
across the boundary at x = 0 begins. The substance can now diffuse from
the left (x < 0) to the right (x > 0). It is further assumed that the stock of
substance on the left side of the boundary is large and transport towards
the boundary fast enough so that the boundary concentration at x = 0
remains constant at C0 : C(x < 0, t) = C0 .
The problem can be solved by way of the superposition principle, which we
know to be valid for linear equations (see Appendix E.3). The result is:
∞  
C0 ξ2
C(x, t) = exp − dξ
(πDx t)1/2 4Dx t
x (8.52)
x
= C0 erfc( ); x≥0
(2Dx t)1/2
 TIME-DEPENDENT SOLUTIONS 217

Fig. 8.12: Diffusion from

a boundary with constant
C0 concentration: the concen-
tration profile constantly
advances into the (ini-
tially empty) right section
(Eq. 8.52). The point x1/2
C0 where the concentration
2 reaches half of the bound-
ary value moves to the
t1 t2 t3 right with a velocity pro-
portional to t1/2 (Eq. 8.54)

x =0 x1/2(t 3)

erfc is the complementary error function (see Appendix E.2), which has
the value 1 for x = 0 and 0 for x → ∞.17 As shown in Fig. 8.12, the solu-
tion (Eq. 8.52) consists of concentration profiles which—starting from the
boundary value C0 at x = 0—intrude further and further into the positive
section. One way to quantify the speed of this process is by following the
migration of the distance x1/2 (t) at which the concentration is half the
value at the edge, i.e., C0 /2. We can calculate x1/2 (t) from the following
equation:
x1/2 (t)
erfc( ) = 0.5 (8.53)
2(Dx t)1/2

The erf and erfc functions are given in tables in most mathematical hand-
books. In such a table we can find that erfc(y) has the value 0.5 if y = 0.48.
Thus:
x1/2 (t)
0.48 =
2(Dx t)1/2

Solving for x1/2 (t), we get:

x1/2 (t) = 2 × 0.48(Dxt)1/2 ∼ (Dx t)1/2 (8.54)

Except for a numerical factor, the movement of the “half-concentration

front” with time thus follows the Einstein-Smoluchowski relation (Eq. 8.9).

17 Note that in the definition of erfc the spatial variable x appears as the lower

boundary of the integral while ξ is just an auxiliary integration variable.

218 MODELS IN TIME AND SPACE

Example 8.6: Oxygen diffusion in a pipe

An abandoned, water-filled vertical groundwater pipe (length: 8 m) is
opened at time t = 0 and brought into contact with the atmosphere.
Molecular oxygen (O2 ) diffuses into the initially completely oxygen-
free water inside the pipe (molecular diffusivity of O2 at T=10 ◦ C:
D = 1.5 × 10−5 cm2 s−1 , O2 equilibrium concentration in water in
contact with the atmosphere at T=10 ◦ C: Cweq
= 11.3 mg/L).

We want to answer two questions:

(a) How long will it take for the O2 concentration 1 m below the
water surface to reach the value 0.1 Cxeq ?
(b) How long does it take to reach 0.5 Cxeq ?

Since the atmospheric O2 concentration and thus the O2 concentration in

the water at the upper end of the pipe (x = 0) stays constant, the oxygen
begins to propagate as per Eq. (8.52) as soon as the pipe is opened to air.
For the moment, let us assume that the pipe is infinitely deep. We will later
check whether the finite pipe length (8 m) substantially changes the result.
 
(a) 0.1Cweq
is reached when erfc is 2(Dxxt)1/2 = 0.1. According to an erfc
table, this is equivalent to 2(Dxxt)1/2 = 1.15. Solving for t = t0.1 with x =
1 m, Dx = 1.5 × 10−5 cm2 s−1 = 1.5 × 10−9 m2 s−1 results in:
x2 1 m2
t0.1 = = = 1.3 × 108 s ∼ 4 a (!)
(2 × 1.15)2 Dx (2.3)2 1.5 × 10−9 m2 s−1
(b) The answer of the second question follows directly from Eq. (8.54):
x2 1 m2
t0.5 = = = 6.7 × 108 s ∼ 21 a (!)
Dx 1.5 × 10−9 m2 s−1
Note that at that time (after 21 years) the O2 concentration at x = 8 m
would be (Eq. 8.54):
C(x = 8 m, t0.5 ) 8m
= erfc( )
C0 2(1.5 × 10−9 m2 s−1 × 6.7 × 108 s)1/2
= erfc(4) < 10−5
In other words, after 21 years the concentration at the lower end of the pipe
would still be so small that the diffusion process doesn’t really “feel” the
finite length of the pipe yet. Therefore, it is justified to assume an infinite
pipe length and to use Eq. (8.52).
Note that the calculated times (t0.1 , t0.5 ) are so large that we should
really question the validity of our assumptions. Any kind of tiny convective
movement (caused e.g. due to small temperature and thus density gradients
along the pipe) would reduce the calculated times by several orders of
magnitude.
 TIME-DEPENDENT SOLUTIONS 219

8.4.2 Multi-dimensional Diffusion

To conclude, we will briefly touch on multi-dimensional diffusion. For a
constant diffusion coefficient Dx = Dy = Dz = D (isotropic diffusion),
we have already derived the three-dimensional version of Fick’s second
law (Eq. 8.26). By dropping one of the three terms, we also get a two-
dimensional variant.
For the case of a (two- or three-dimensional) point-shaped initial condi-
tion18 the solutions are (not surprisingly) two- or three-dimensional normal
distributions of the following form:

2-dimensional:
 2 
m(2) x + y2
C(x, y, t) = exp − (8.55)
4(πDt) 4Dt

3-dimensional:
 2 
m(3) x + y2 + z 2
C(x, y, z, t) = exp − (8.56)
8(πDt)3/2 4Dt

Note that m(2) has dimension [ ML−1 ], while m(3) has dimension [ M ]. The
concentration distributions are circularly or spherically symmetric.
In Example 8.5 we’ve seen that in the one-dimensional case the max-
imum concentration of a point source drops with t−1/2 , i.e. rather slowly.
As we can see from Eqs. (8.55) and (8.56), the coefficient declines with t−1
in the two-dimensional case and with t−3/2 in the three-dimensional case.
Let’s therefore have a last look at the behavior of a pollutant, but this time
for the case of two-dimensional diffusion.

Example 8.7: Pollutant dispersal in a lake

We want to examine how quickly diffusive horizontal isotropic mixing
lets the concentration of a pollutant fall below a critical value. To do
so, we will modify Example 8.5 (atrazine in a canal) as follows: At
time t0 , the amount M of the herbicide atrazine enters the surface
layer of a lake, is rapidly distributed vertically within the epilimnion
(height h) and then further spreads through isotropic horizontal
diffusion (Kx = Ky = K). How long will it take for the atrazine
concentration to fall below the tolerance value for drinking water Ctol
throughout the lake?
Numbers: M = 1kg, Ctol = 0.1 µg/L = 10−4 g m−3
h = 10m, Kx = 0.1 m2 s−1

18 We get a three-dimensional point from the product of three delta functions:

C(x, y, z, t = 0) =m(3) δ(x)δ(y)δ(z), whereby m(3) has the dimension of a mass.
220 MODELS IN TIME AND SPACE

We assume that the horizontal concentration distribution initially corre-

sponds to a two-dimensional delta function with the total mass per depth
m(2) = M/h = 103 g/10 m = 100 g m−3 . The horizontal concentration dis-
tribution is then described by Eq. (8.55). The maximum concentration at
x = y = 0 as a function of time is:

m(2)
Cmax (t) = = Ctol
4πKt
Solving for t ≡ t0 :

m(2) 100 g m−1

t0 = =
4πKCtol 4π × 0.1 m2 s−1 × 10−4 g m−3 (8.57)
= 8.0 × 10 s = 9 d
5

Compared to one-dimensional mixing, this time is much smaller. Of course,

the atrazine could be moved by advection in addition to diffusion. As we
showed in Eq. (8.50) for the one-dimensional case, this does not change the
value of Cmax , but only the location where it occurs.
 QUESTIONS AND PROBLEMS 221

8.5 Questions and Problems

Question 8.1: Why do models that are continuous in time and space lead
to partial differential equations?

Question 8.2: What is the difference between diffusive and advective

transport?

Question 8.3: Define typical times for diffusion, advection and transfor-
mation.

Question 8.4: The ratio between mixing time τmix and transformation
time τr determines whether spatial differentiation is appropriate when mod-
eling a system. What is the decision criteria?

Question 8.5: Give examples for processes that can be described by an

exchange model.

Question 8.6: What are some of the commonalities and differences

between molecular and turbulent diffusion?

Question 8.7: Explain the meaning of the divergence theorem.

Question 8.8: Explain the meaning of the Péclet and Damköhler numbers.

Question 8.9: What is the difference between the diffusive-advective and

the diffusive-reactive regime of a space/time model?

Question 8.10: The Einstein-Smoluchowski relation states that diffusion

in one dimension spatially advances at the speed of t1/2 . What does this
mean for the maximum concentration of a substance which is initially con-
centrated in one point? In two and three dimensions, this effect is in play
across multiple dimensions simultaneously. What then is the corresponding
rule for the maximum concentration in two- or three-dimensional diffusion?

Question 8.11: How must we modify Eqs. (8.55) and (8.56) in order to
also include the influence of advection?

Problem 8.1: Tetrachloroethene in a pond

We consider a pond with little through-flow (volume V = 2 × 10 4 m3 ,
surface A = 5 × 103 m2 ). Due to an accident, a small amount M = 2 kg of
tetrachloroethene (a substance used to clean electronic components) is in-
troduced. Tetrachloroethene is a volatile substance; its atmospheric con-
centration is small and can be neglected in the calculation of gas exchange.

(a) How long will it take until the average concentration in the pond falls
below the value Ccrit = 0.1 µg/L, if the average gas exchange velocity
of tetrachloroethene at the pond’s surface is vW/L = 0.4 m d−1 ? No
other elimination processes of tetrachloroethene play a role in the
pond. Furthermore, you can assume that tetrachloroethene rapidly
becomes completely mixed in the water body.
222 MODELS IN TIME AND SPACE

(b) By how much is this time shortened if we also take a through-flow of

Q =2,000 m3 per day into account?

Problem 8.2: Dissolved oxygen in a lake

In Example 8.2 we assumed that the O2 concentration in the well-mixed sur-
face layer of the lake remains roughly constant (CA = 12 mg/L), while the
deep water remained anoxic (oxygen-free), i.e. (CB → 0). Assess whether
this assumption makes sense, using the following information:
3.0 g m−2 d−1 : O2 production via photosynthesis in the surface layer
9 mg/L: O2 equilibrium concentration with the atmosphere (at 20 ◦ C)
vW/L = 0.6 m d−1 : Water-air exchange velocity at the lake surface

As a reminder: The O2 flux through the thermocline into the deep water is
initially 0.72 g m−2 d−1 and then rises up to about 1.4 g m−2 d−1 when CB
drops to zero.

Problem 8.3: Geothermal heat flux at the bottom of a lake

To determine the geothermal heat flux through a lake’s sediments, a mea-
suring rod with attached thermistors is pushed into the sediment. After
waiting long enough, a steady-state temperature profile will be reached
along the rod (that is, along the vertical axis). From this temperature pro-
file we can calculate the geothermal heat flux Fth (unit: W m−2 ), if we
assume that the sediment predominantly consists of water. This is an ac-
ceptable assumption, since fresh sediment consists of 85–95 % water. The
heat transfer can be described by Fourier’s law, which has the same form
as Fick’s first law:
−2 −1
Fth = −γth dT
dx , γth : Thermal conductivity [Wm K ]
γth (Water, 4 ◦ C) = 0.58 Wm−1 K−1

(a) Determine the heat flux Fth .

(b) Calculate the vertical temperature gradient in the water column above
the sediment which is in equilibrium with Fth . Assume that we are
dealing with a hypothetical lake with vertical walls. The vertical tur-
bulent diffusion coefficient in the water column is Kz = 0.05 cm2 s−1 =
5×10−6 m2 s−1 .
Temperature data from the sediment
Depth in the T [ ◦C ]
sediment [ cm ]
0 4.250
10 4.269
20 4.285
30 4.304
50 4.342
100 4.425
 QUESTIONS AND PROBLEMS 223

Problem 8.4: Vertical turbulent diffusivity

In a 50 m deep lake with a flat bottom, the following profile of the radioac-
tive noble gas radon-222 is measured at the deepest spot:
222
Height above ground Rn activity
[ m] [ Bq/Liter]
0.5 53
1.0 47
2.0 37
3.0 30
4.0 25
The in-situ production of 222
Rn from 226
Ra decay is JRa = 1.8 BqL−1d−1 .
(a) Calculate the vertical turbulent diffusion coefficient Kz above the lake
floor by using a one-dimensional diffusion/reaction model.
(b) How large is the flux FRn of 222 Rn from the sediment (expressed in
Bq m−2 d−1 )? The decay constant of radon-222 is 0.181 d−1.

Note: As per Eq. (8.37), the “excess activity” of radon, Cexc ≡ C −JRa /λRn ,
is an exponential function. Therefore ln(Cexc ) should be a linear function
1
of height h with slope −(λRn /Kz ) 2 . With λRn known we can therefore
determine Kz . The flux follows from Eq. (8.36).

Problem 8.5: Péclet and Damköhler numbers

Use the concepts we developed in Sects. 8.2.1 and 8.2.2 for mixing times
and transport distances to show that
(a) The Péclet number (Eq. 8.41) is the ratio of diffusive to advective
mixing time over the distance xL ;

(b) The Damköhler number (Eq. 8.42) is the square of the quotient of
diffusive and advective transport distance in time τ = kr−1 .

Problem 8.6: Volatile substance in the groundwater

A leaking tank at an electronics plant causes the volatile organic solvent
trichloroethene (TCE) to enter the groundwater beneath the plant, where
a concentration of 25 mg m−3 is measured. The water table is at 4 m below
the surface.
Via gas exchange, TCE enters the air-filled pore space of the so-called
vadose zone (also called unsaturated zone), and from there, diffuses into the
free atmosphere. The atmospheric TCE concentration is very small and can Trichloroethene
be neglected. The non-dimensional Henry coefficient between air and water
(see Eq. 3.2) for TCE at a water temperature of 10 ◦ C is Ka/w = 0.22.
The effective diffusion coefficient of TCE through the air filled pores of
the vadose zone is Dground ∼ 7 × 10−3 m2 d−1 . Note that for the diffusion
through the gas phase only 15 % of the total cross section is available.
224 MODELS IN TIME AND SPACE

(a) Calculate and draw the steady-state TCE profile from the ground-
water up to the surface assuming that TCE is not degraded in the
soil.
(b) Sketch the steady-state concentration profile of TCE, assuming that
it degraded while migrating upwards through the soil. Compare the
two profiles (a) and (b).
(c) Calculate the steady-state flux of TCE from the groundwater into the
atmosphere per surface area (m2 ) and day for case (a) (no decompo-
sition). Remember that only 15 % of the cross section is available for
the diffusion of TCE through air.
(d) Using Eq. (8.10), estimate the time τstat needed for a steady-state
TCE profile to develop.
(e) The answer to question (d) will show that the time to reach a diffusive
steady state is several years. We should therefore additionally consider
the influence from vertical air flow, as can result from air pressure
differences. This might significantly alter the flux calculated in (c)
and the time calculated in (d). Investigate this question for the case
of a small air current of 0.1 m d−1 . Calculate the modified values for
F und τstat . Calculating the Péclet number may be helpful for this.

Note: This problem is explored in more detail in Schwarzenbach et al. (2003,

Illustrative Example 19.2).

Problem 8.7: Transport and reaction in a lake’s water column

The measurement of four different volatile chemicals in a lake results in dis-
tinct vertical steady-state concentration profiles (see Fig. 8.13). Note that
in order to interpret and understand these profiles, the scale and units of
the coordinates are not relevant and thus omitted. The following simplified
assumptions can be made:

1. Vertical transport of the chemicals is by turbulent diffusion only, with

a diffusion coefficient which is constant across the whole water col-
umn (it simplifies the mathematics, although physically it’s not very
realistic!).
2. The lake consists of an oxic upper and an anoxic lower layer. Anoxic
means that the concentration of dissolved molecular oxygen (O2 ) is
zero.
3. All substances enter the lake by air-water exchange at the water sur-
face. The exchange rate is assumed to be fast, so that the surface
concentration in the water is in equilibrium with the atmosphere and
constant with time. (Note that in Fig. 8.13 the surface concentra-
tions of all chemicals are drawn at the same value for convenience,
although in reality they may be different. We can do that because the
concentration axis has no absolute scale.)
 QUESTIONS AND PROBLEMS 225

C0 Fig. 8.13: Concentration

z=0
profiles of four different
volatile chemicals in a lake

Aerobic zone
Anaerobic zone

1 2 3 4
zb

4. The lake bottom acts like an impermeable boundary. The chemicals

can neither enter the sediments nor is there a flux out of the sedi-
ments.

5. The four chemicals A, B, C, D have the following specific properties:

A is inert (no reaction in the water). B is degraded in the anoxic
part by a first-order reaction, but it is inert in the oxic layer. C is
degraded by a first-order reaction in the oxic layer, but inert in the
anoxic layer. D is degraded in both layers; the first-order reaction rate
is independent of the O2 concentration, i.e. it is equal in the whole
water column.

Answer the following questions:

(a) Identify the profiles shown in Fig. 8.13 by the four chemicals A, B, C,
and D.

(b) Derive the analytical solutions for the curves of chemicals A and D.
Use Kz and kr as vertical diffusivity and first-order reaction rate
constant.

(c) How do the profiles for chemicals B, C, and D qualitatively change if

kr is increased? Go to the limits of an extremely fast reaction.

Problem 8.8: Radium and tritium in the ocean deep water

In Example 8.4 we discussed the steady-state concentration profiles of 226 Ra
and tritium in the deep water of the ocean. The calculated profiles (Eq. 8.43
for 226 Ra, Eq. 8.44 for tritium) correspond to an equilibrium between mass
flows at the upper and lower boundaries of the deep water and the radioac-
tive decay within the deep water. Establish a mass balance of the deep
water column consisting of the following components for each of the two
radioactive isotopes:
226 MODELS IN TIME AND SPACE

1. Flux between surface and deep water, divided into the contribution
of diffusion and advection;

2. The same, but for the flux between bottom and deep water;
3. Radioactive decay.

Note: To calculate the flux per surface, use the standardized concentration
profiles. Choose the sign so that all fluxes into the deep water are counted
positively in the balance while all fluxes out of the deep water appear as
negative numbers. The radioactive decay is also a negative contribution to
the balance.

Problem 8.9: Time-dependent diffusion/advection equation

(a) By forming the corresponding spatial and temporal derivatives, show
that Eq. (8.48) is indeed a solution of Fick’s second law in the one-
dimensional case (Eq. 8.45).
(b) Show that the pure advection equation ∂C/∂t = −vx ∂C ∂x has solutions
of the form C(x − vx t), whereby C(x) is the initial concentration.
(c) Confirm that the combination of (a) and (b) equates to Eq. (8.50).

Problem 8.10: Symmetrical diffusion at a boundary surface

Look for solutions of the differential equation (8.48) for the initial condition
Eq. (8.51), which represents a concentration jump from C0 to 0 at x = 0.
Unlike in the solution Eq. (8.52), the concentration in the left section does
not stay constant.

Note: Use the method from Appendix E.3. The diffusion of the mass element
C0 dξ now proceeds in both directions. Using qualitative considerations,
first make a sketch of the distribution of C(x) through the boundary surface
at x = 0.
Appendix A

List of Symbols

Only symbols that appear more than once are listed here.

A, Ai Area [ L2 ]
B(Vk ) Jacobian matrix, calculated at the fixed point Vk
C Concentration [ ML−3 ]
C0 Initial concentration
C∞ Steady state concentration
CX Concentration of substance X [ MX L−3 ]
Caq Concentration in water [ ML−3 ]
Cin Concentration in the inflow [ ML−3 ]
Cair Concentration in air [ ML−3 ]
Csed Concentration in sediment [ MM−3 sed ]
D Diffusion coefficient
Da Damköhler number (Eq. 8.42)
Fi , F Flux
Fad Advective flux
Fdiff Diffusive flux
Fi,j Mass flux from box j to box i in a multi-dimensional model
[ MT−1 ]
I Inhomogeneous term in discrete models
Jp Zero-order production process
Jl Zero-order loss process
J Zero-order net process
Jin Substance input [ MT−1 ]
Jout Substance output [ MT−1 ]
KA/B Chemical distribution (or partition) coefficient between A
and B
Kd Distribution coefficient between particles and water
[ L3 M−1
sed ]
Ka/w Non-dimensional air/water distribution coefficient (or
non-dimensional Henry coefficient)
Kx Turbulent diffusion coefficient in direction x
Ldiff Characteristic diffusion distance

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 227

DOI 10.1007/978-3-642-30639-6, © Springer-Verlag Berlin Heidelberg 2013
228 APPENDIX A

M Mass [ M ]
N Quantity, e.g. of individuals or atoms
P Coefficient matrix, P = (pi,j )
Pe Péclet number (Eq. 8.41)
Q Flow rate [ L3 T−1 ]
Qex Exchange rate [ L3 T−1 ]
R, Ri External relation
T Period of a periodically fluctuating input [ T ]
V Volume [ L3 ]
VE Volume of the epilimnion [ L3 ]
VH Volume of the hypolimnion [ L3 ]
V Generic system variable
V Vector of the system variable V = (Vi )
Vi i-th system variable in a multi-dimensional model
V0 Initial value of the system variable V at time t = 0
V∞ Steady state value of the system variable V

ai Coefficient in discrete models

ai,j Coefficients of the solutions of linear differential equations
f (V) Change function for an autonomous system
g(V, t) Change function for a nonlinear differential equation
jin Substance input set in relation to the system’s volume
[ MT−1 L−3 ]
j0 Mean value of a periodically fluctuating input [ MT−1 L−3 ]
j1 Amplitude of a periodically fluctuating input [ MT−1 L−3 ]
k k = −k  [ T−1 ]
k Net first-order coefficient describing substance conversion
[ T−1 ]
kex,E Exchange rate from epilimnion into hypolimnion [ T−1 ]
kex,H Exchange rate from hypolimnion into epilimnion [ T−1 ]
ki,j Specific exchange rate from box j into box i
kp First-order production or growth rate [ T−1 ]
kr Linear reaction rate [ T−1 ]
ks Sedimentation rate [ T−1 ]
ktot Specific total rate [ T−1 ]
kl First-order coefficient describing a loss process [ T−1 ]
kw Specific flow rate kw = VQ [ T−1 ]
m Mass per area
kλ Radioactive decay rate [ T−1 ]
p Model parameter
pi,j General model parameter in a multi-dimensional model
t Time [ T ]
tn Time steps in discrete models
t0 Starting time from which the system is modeled
vA/B Exchange velocity between A and B
vw/a Water/air exchange velocity [ L−1 ]
 LIST OF SYMBOLS 229

vx Advective velocity in direction x

x Spatial coordinate (usually horizontal)
x1/2 Distance at which concentration has half of the boundary
value
y Spatial coordinate (usually vertical)
yi Generic variable in a differential equation
z Vertical spatial coordinate, e.g. depth in a lake [ L ]

β Growth rate of an exponentially growing input [ T−1 ]

β  = −β Reduction rate of an exponentially falling input [ T−1 ]
γ Stratification parameter
ε Positive parameter (usually small)
κ Relative convergence towards an equilibrium concentration
λi Eigenvalues of the coefficient matrix
ω Angular frequency of a periodically fluctuating input
−1
(ω = 2πT ) [T ]
τκ , τ5 % Adjustment time to κ or to 5 % [ T ]
τ1/2 Half-life of a radioactive isotope [ T ]

det(..) Determinant of a matrix

tr(..) Trace of a matrix
Δ(..) Discriminant of a matrix

Indices
i, j,. . . Subscript: variable or box number
(n), (k),. . . Superscript: iteration step of a discrete variable
◦, ∞ Superscript: initial value or steady-state/fixed-point value
Appendix B

Dimensions and Units

B.1 Dimensions

Mathematical models tie together various measured quantities. These quan-

tities usually have a dimension. The basic dimensions in physics are (indi-
cated with capital letters):

[L] Length [M] Mass

[T] Time [Q] Electric charge

Other common dimensions include:

[K] Temperature
[N] Number (of individuals, atoms, etc.)

All further dimensions can be composed from these basic dimensions, as

the following examples demonstrate:

[ LT−1 ] for velocity

[ MLT−2 ] for force
[ ML2 T−2 ] for energy or work
[ ML−1 T−2 ] for pressure

Occasionally it can make sense to add an index to the dimensions, such as:

[ MA L−3 ] concentration of substance A

[ L3r ] volume of the reactor
[ L3w L−3
tot ] porosity, i.e. water volume per total volume

If dealing with indexed dimensions, only quantities with the same index
can be reduced.

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 231

DOI 10.1007/978-3-642-30639-6, © Springer-Verlag Berlin Heidelberg 2013
232 APPENDIX B

B.2 Units
From the dimensions we arrive at units by choosing a specific system of
measurement, for instance the metric system. The metric system leads us
to the SI units, which are the international standard and commonly used
in most scientific fields:

m Meter for length [ L ]

kg Kilogram for mass [ M ]
s Second for time [ T ]
C Coulomb for electric charge [ Q ]

The units derived from the basic dimensions usually have their own names:

Newton N for kg m s−2 (force)

Joule J for N m or kg m2 s−2 (energy, work)
Watt W for J s−1 or kg m2 s−3 (power)
Pascal Pa for N m−2 or kg m−1 s−2 (pressure)
Liter L for dm3 (volume)
Becquerel Bq for s−1 (radioactive activity)
Curie Cu 1 Cu = 3.7 × 1010 Bq (radioactive activity)

Temperature (in degrees Celsius ◦ C or Kelvin K) is a bit of a special case.

It is an intensive quantity, that is, energy per volume. In practice, K is
often treated as another basic dimension (and basic unit).
Appendix C

Formulary

C.1 Linear Inhomogeneous First-Order Differential Equation

dV
= J(t) − k(t) · V (C.1)
dt
In the most general case, the coefficients J and k can be time-dependent.
J is the inhomogeneous term. The solution is
t

0 −ϕ(t) −ϕ(t)
V(t) = V e +e eϕ(t ) J(t )dt (C.2)
0

with the exponent ϕ(t) defined as

t

ϕ(t) = k(t )dt (C.3)

0

and the initial condition

V0 = V(t = 0) (C.4)

C.1.1 Special Cases

Homogeneous Equation: J = 0
The solution for a constant k is

V(t) = V0 e−k·t (C.5)

For a time-dependent k, it is

V(t) = V0 e−ϕ(t) (C.6)

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 233

DOI 10.1007/978-3-642-30639-6, © Springer-Verlag Berlin Heidelberg 2013
234 APPENDIX C

Inhomogeneous Equation with Constant Coefficients J and k

 J
V(t) = V∞ + V0 − V∞ e−k·t , with steady state V∞ = (C.7)
k

Inhomogeneous Equation with Constant k but Time-Dependent

External Parameter J(t)
The general solution is

t

0 −k·t
V(t) = V e + e−k·(t−t ) J(t )dt (C.8)
0

With expontentially growing input J(t) = J(0)eβt it is

J(0) −kt J(0) βt

V(t) = V0 e−kt − e + e (C.9)
k+β k+β

And finally, with periodically fluctuating input J(t) = J0 + J1 sin ωt we get

J0 J0
V(t) = + (V0 − )e−kt +
k k
(C.10)
J1 ω J1 ω −kt
+√ sin(ωt − arctan ) + 2 e
2
ω +k 2 k k + ω2

C.2 System of Two Linear First-Order Differential Equations

By writing

. dVi
Vi = (C.11)
dt
the following system of equations results:
.
V1 = J1 + p1,1 V1 + p1,2 V2
. (C.12)
V2 = J2 + p2,1 V1 + p2,2 V2

The solution for constant coefficients Ji and pij , if det(P) = 0 is

V1 (t) = a1,0 + a1,1 eλ1 t + a1,2 eλ2 t

(C.13)
V2 (t) = a2,0 + a2,1 eλ1 t + a2,2 eλ2 t

The eigenvalues λ (characteristic rates) are thereby

%
1 
λ1,2 = (p1,1 + p2,2 ) ± (p1,1 − p2,2 )2 + 4p1,2 p2,1 (C.14)
2
 FORMULARY 235

If both eigenvalues are real and negative, the model has the following steady
state:

p1,2 J2 − p2,2 J1
a1,0 = V∞
1 =
p1,1 p2,2 − p1,2 p2,1
(C.15)
p2,1 J1 − p1,1 J2
a2,0 = V∞ =
2
p1,1 p2,2 − p1,2 p2,1
The elements of the matrix Ai,j are calculated as follows:

 
a1,1 = q · (p1,1 − λ2 )(V01 − a1,0 ) + p1,2 (V02 − a2,0 )
 
a1,2 = −q · (p1,1 − λ1 )(V01 − a1,0 ) + p1,2 (V02 − a2,0 )
  (C.16)
a2,1 = q · p2,1 (V01 − a1,0 ) + (p2,2 − λ2 )(V02 − a2,0 )
 
a2,2 = −q · p2,1 (V01 − a1,0 ) + (p2,2 − λ1 )(V02 − a2,0 )

with
1 1 1
q= =  = √ (C.17)
λ1 − λ2 (p1,1 − p2,2 ) + 4p1,2 p2,1
2 Δ

C.3 General Solution of the Linear Second-Order Differential

Equation with Constant Coefficients
The inhomogeneous differential equation of the function f (x)

d2 f df
a +b + cf + d = 0, a, c = 0 (C.18)
dx2 dx
d
can be transformed by using g = f + c into the homogeneous equation

d2 g dg
a +b + cg = 0 (C.19)
dx2 dx
It has the solution

g(x) = A1 eλ1 x + A2 eλ2 x (C.20)

where λ1 and λ2 are

1 1 
λ1 = −b + (b2 − 4ac) and λ2 = (−b − b2 − 4ac)
2a 2a
The solution for f (x) is therefore

d
f (x) = A1 eλ1 x + A2 eλ2 x − (C.21)
c
236 APPENDIX C

The parameters A1 and A2 are determined by the boundary conditions,

such as
d
f (0) = f0 , thus g(0) = f0 +
c
d
f (xL ) = fL , thus g(xL ) = fL +
c
With these boundary conditions, the following results for the parameters
A1 and A2 :

(fL + dc ) − (f0 + dc )eλ2 xL

A1 =
eλ1 xL − eλ2 xL
(C.22)
(f0 + dc eλ2 xL ) − (fL + dc )
A2 =
eλ1 xL − eλ2 xL

C.3.1 Special Case: b = 0

The differential equation (C.18) then has the form

d2 f
a + cf + d = 0 (C.23)
dx2
The approach to the solution is the same, except that λ1 and λ2 are calcu-
lated as follows:
 
c c
λ1 = − and λ2 = − − (C.24)
a a

C.3.2 Non-dimensional Description of the Solution

With the following definitions we can describe the solution of Eq. (C.18)
within the interval [0, xL ] in a non-dimensional form:
x xL b ac
ξ= , Pe = − , Da = − (C.25)
xL a b2
The new non-dimensional quantities Pe and Da are the Péclet and Damköhler
numbers, respectively. The non-dimensional spatial coordinate ξ varies be-
tween 0 and 1 if x remains limited to the interval [0, xL ]. The following now
holds:
dg dg dξ 1 dg
= · =
dx dξ dx xL dξ
(C.26)
d2 g 1 d2 g
= 2 · 2
dx2 xL dξ
With this, we can rewrite Eq. (C.19) with ξ as

d2 g dg
− Pe − Pe2 Dag = 0 (C.27)
dξ 2 dξ
 FORMULARY 237

The solution then becomes

∗ ∗
g(ξ) = A1 eλ1 ξ + A2 eλ2 ξ
Pe √ (C.28)
with λ∗i = (1 ± 1 + 4Da)
2
For the Damköhler number Da we can now distinguish between the following
two extreme situations:
λ∗1 ≈ Pe √
• |Da| 1:
λ∗2 ≈ −DaPe, |λ∗2 | |λ∗1 | (use: 1 + ε ≈ 1 + 2ε )
∗
If simultaneously, |Pe| 1, then the exponential functions eλi ξ barely
vary in the interval ξ ∈ [0, 1], i.e. the function g(ξ) is practically
constant.
√
• |Da| 1 : λ∗i = ±Pe Da
If simultaneously, |Pe| ≈ 1 or larger, then the large sizes of the eigen-
values λ∗i entail strong variations in the solution function g(ξ) in the
interval ξ ∈ [0, 1].

C.4 Solution of Linear Differential Equations with

Imaginary Eigenvalues
The solution of a two-dimensional system of linear differential equations
with the imaginary eigenvalues λ = ±iω has the form (see Eq. 5.55)

yj (t) = aj1 eiωt + aj2 e−iωt , j = 1, 2 (C.29)

With Euler’s formula

eiωt = cos ωt + i sin ωt (C.30)

we can bring Eq. (C.29) into the form

yj (t) = bj1 cos ωt + bj2 sin ωt, j = 1, 2 (C.31)

The bi,j are defined as

b11 = (a11 + a12 ), b12 = i(a11 − a12 )

(C.32)
b21 = (a21 + a22 ), b22 = i(a21 − a22 )
Finally, we can also write Eq. (C.31) as a single trigonometric function with
phase shift ϕj , e.g. as

yj (t) = cj sin(ωt + ϕj ), j = 1, 2 (C.33)

where cj and ϕj are

 
bj1
cj = (b2j1 + b2j2 )1/2 ) ; ϕj = arctg , j = 1, 2 (C.34)
bj2
238 APPENDIX C

Usually, the system variables yj (t) are real functions, which means that the
coefficients bj1 , bj2 and cj are real and the pairs (aj1 , aj2 ) are conjugate-
complex. All these coefficients are calculated from the initial conditions
y1◦ and y2◦ .

Let’s consider an example. We will use the equations of the harmonic

oscillator (Example 5.11, Eqs. 5.73 and 5.77):
dy1 dy2
= y2 ; = −ω 2 y1 (C.35)
dt dt
The initial values are y1◦ = A, y2◦ = B. The eigenvalues are λ = ± iω.
If we insert the general solution (Eq. C.31) into the differential equations
(Eq. C.35) and compare the coefficients of cos ωt and sin ωt on both sides
of the equals sign, we get

b11 = A ; b12 = B
ω (C.36)
b21 = B ; b22 = −ω A
This means that y1 (t) and y2 (t) are

y1 (t) = A cos ωt + B
ω sin ωt (C.37)
y2 (t) = B cos ωt − ω A sin ωt
The solutions, as per Eq. (C.33), are then

B 2 1/2
 Aω
y1 (t) = (A2 + ω2 ) sin(ωt + ϕ1 ), ϕ1 = arctg B
 (C.38)
2 2
y2 (t) = (ω A + B ) 2 1/2
sin(ωt + ϕ2 ), ϕ2 = −arctg B
Aω

Finally, we can also write the solutions in the form of Eq. (C.29) as
 −iωt
y1 (t) = 1/2 (A − i B
ω )e
iωt
+ 1/2 A + i B
ω e
(C.39)
y2 (t) = 1/2 (B + iωA)eiωt + 1/2(B − iωt)e−iωt
The coefficients are, as mentioned, complex numbers.
Appendix D

Eigenvalues

D.1 The n-Dimensional System

The system consists of the n differential equations

n
dVi
= Ri + pij Vj , i = 1, · · · , n (D.1)
dt j=1

.
In vector and matrix form, and with Vi = dVi
dt , we can write it as
.
V = R + PV
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
R1 V1 p11 ... p1n
⎜ ⎟ ⎜ ⎟ ⎜ .. ⎟
R = ⎝ ... ⎠ V = ⎝ ... ⎠ P=⎝ . ⎠ (D.2)
Rn Vn pn1 . . . pnn

With the matrix U we introduce n new variables Zi :

n
Z = U−1 V, or Zi = (U −1 )ij Vj (D.3)
j=1

If we multiply Eq. (D.2) on both sides from the left with U−1 and add the
identity matrix E = UU−1 in between P and V on the right-hand side,
then it follows that
.
U−1 V = U−1 R + U−1 P(UU−1 )V (D.4)

or, in terms of Z,
.
Z = U−1 R + ΛZ with Λ = U−1 PU (D.5)

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 239

DOI 10.1007/978-3-642-30639-6, © Springer-Verlag Berlin Heidelberg 2013
240 APPENDIX D

With a well-chosen transformation matrix U, the new system matrix Λ

becomes diagonal1:
⎛ ⎞
λ1 0
⎜ .. ⎟
Λ=⎝ . ⎠ (D.6)
0 λn
The λi are the eigenvalues of the matrix P. The corresponding eigenfunc-
tions are the column vectors of the matrix U.
As per Appendix C.1, the diagonalized system (Eq. D.5) can now be
solved row-wise. For instance, for
.
Zk = Tk + λk Zk , Tk = k-th component of the vector U−1 R (D.7)

we get the solution

Tk
Zk (t) = Zk∞ + (Zk0 − Zk∞ )eλk t , Zk∞ = − (D.8)
λk
Through the inverse transformation V = UZ we can calculate the time-
dependent solutions Vi (t) from the solutions in Eq. (D.8).

D.2 Explicit Solution for the Two-Dimensional System

D.2.1 The Eigenvalues
The eigenvalues λi are calculated with the so-called characteristic equation
of P,
det(Eλ − P) = 0 (D.9)
In the above equation, det(..) is the determinant and E the two-dimensional
identity matrix, which is
 
1 0
E= (D.10)
0 1
Written out, Eq. (D.9) thus has the following form:
 
λ − p1,1 −p1,2
det =0 (D.11)
−p2,1 λ − p2,2
To determine the two eigenvalues λ1 und λ2 , Eq. (D.11) leads to the
quadratic equation

λ2 − λ(p1,1 + p2,2 ) + p1,1 p2,2 − p1,2 p2,1 = 0 (D.12)

1 If one or more of the eigenvalues λ are identical, the diagonalization no

i
longer succeeds completely, but only up to the diagonally arranged sub-
matrices. This case is not discussed here.
 EIGENVALUES 241

The solution of the quadratic equation can be written as

1  
λi = tr(P) ± Δ(P) , i = 1, 2 (D.13)
2
Here tr(P) is the so-called trace of the quadratic matrix, i.e. the sum of its
diagonal elements:
tr(P) = p1,1 + p2,2 (D.14)
Δ(P) is the discriminant of the quadratic equation:
2
Δ(P) = (p1,1 + p2,2 ) − 4p1,1 p2,2 + 4p1,2 p2,1
2
= (p1,1 − p2,2 ) + 4p1,2 p2,1 (D.15)
= tr(P) − 4 det(P)
2

D.2.2 Properties of the Eigenvalues λi

Mass balance models lead to distinct systems of differential equations, i.e. to
certain properties of the matrix P and its eigenvalues λi .

Conservative Equations (Homogeneous)

The system of equations is homogeneous if R = 0. The matrix P then has
the following properties:

• The diagonal elements pii are negative. All elements outside of the
diagonal, pij with i = j, are positive or zero.

• The sum of all elements in one column of P is zero (as should be

expected for a conservative mass balance). From this it follows that
the matrix is singular, i.e. its determinant is zero.

These properties are responsible for the fact that P has one eigenvalue
λ = 0. The corresponding transformed variable U−1 y is the sum of all
yi , that is, the total mass in the system (which is constant). All other
eigenvalues λi are negative, i.e. there is a steady state.

Non-conservative Equations (Inhomogeneous)

Not all xi = 0. In addition, at least one sum of column elements is negative.
Then, the matrix P is not singular, i.e. the inverse matrix P−1 exists. It
can be used to calculate the steady state with the relation

y = −P−1 x (D.16)
Because the pij are still constrained,

< 0 for i = j
pij
> 0 for i = j
Therefore all eigenvalues λi are negative, so that a steady state exists in
this case, too.
242 APPENDIX D

D.2.3 Eigenfunctions
The eigenfunctions of the matrix P are only determined up to a constant
(positive or negative) factor. For instance, they can be written as

z1 = p2,1 V1 + (λ1 − p1,1 )V2

(D.17)
z2 = (λ2 − p2,2 )V1 + p1,2 V2
Written in the new variables zi , the system of homogeneous differential
equations becomes
dz1
= λ1 z1
dt (D.18)
dz2
= λ2 z2
dt
with the solutions

z1 (t) = z1o eλ1 ,t z2 (t) = z2o eλ2 ,t (D.19)

The initial values zio
can be calculated from the transformation equations
(D.17) by inserting Vo1 and Vo2 into them. The inverse transformation to
Eq. (D.17) is
1
V1 = [p1,2 z1 − (λ1 − p1,1 )z2 ]
D (D.20)
1
V2 = [−(λ2 − p2,2 )z1 + p2,1 z2 ]
D
with

D = p1,2 p2,1 − (λ1 − p1,1 )(λ2 − p2,2 )

(D.21)
= −2 det(P) + λ1 p2,2 + λ2 p1,1
Note that we used the following relation: λ1 λ2 = det(P).
Appendix E

Time-Dependent Diffusion Equation

We will look at some mathematical functions that are useful for the explicit
time-dependent solution of the diffusion equation. More can be found in
Crank (1975) and in Schwarzenbach et al. (2003, Chaps. 18 and 22).

E.1 The Normal or Gaussian Distribution

Figure E.1 shows the Gaussian distribution G(x) (often also called the
normal distribution).

G(x 0) = G 0 Fig. E.1: Gauss or nor-

mal distribution around
the mean value x0 with
variance σ
G(x 0 ± σ ) = e−0.5 G 0 = 0.607 G 0

x0 − σ x0 x0 + σ

The mathematical formulation is

1 (x − x0 )2
G(x) = √ exp (− ) (E.1)
2π σ 2σ 2
where x0 is the mean value of the distribution and σ is the variance. The
Gaussian distribution G(x) is standardized, so
∞

G(x)dx = 1 (E.2)
−∞

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 243

DOI 10.1007/978-3-642-30639-6, © Springer-Verlag Berlin Heidelberg 2013
244 APPENDIX E

The values the distribution takes on for x0 ± 2σ and x0 ± 3σ are

G(x0 ± 2σ) = e−2 G0 = 0.135 G0

G(x0 ± 3σ) = e−9/2 G0 = 0.011 G0

The area between the edges x = x0 − σ and x = x0 + σ is 68.3 % of the

total integral of G(x) standardized to 1. The area between x = x0 ± 2σ
encompasses 95.4 %. The special solution of the diffusion equation is thus
a Gaussian distribution with the time-dependent variance
√
σ(t) = 2Dt (E.3)
From it, we can determine the diffusion coefficient D as

1 dσ 2
D= (E.4)
2 dt

E.2 The Error Function

The error function erf(x) is defined as
x
2
erf(x) = √ exp (−ξ 2 )dξ (E.5)
π
0

The complement of the error function is

∞
2
erfc(x) = 1 − erf(x) = √ exp (−ξ 2 )dξ (E.6)
π
x

The error function has the following properties:

erf(x = 1) = 1, erfc(x = 1) = 0
erf(x = 0) = 0, erfc(x = 0) = 1

Its values can be looked up in tables in the literature, for instance in

Abramowitz and Stegun (1972).

E.3 The Principle of Linear Superposition

The influence of the layer dξ on the concentration at point x is calculated
from Eq. (8.48) with m = C0 dξ. Because the concentration only moves in
positive x direction (for x < 0, C = C0 =constant), the factor 2 disappears
in the denominator of Eq. (8.48). Thus the influence of the layer dξ can be
described by
C 0 dξ ξ2
√ exp (− )
Dπt 4Dt
 TIME-DEPENDENT DIFFUSION EQUATION 245

C0

dx x =0 x

The superposition of all dξ sections results in

∞
C0 ξ2
C(x, t) = √ exp (− )dξ (E.7)
Dπt 4Dt
x

Note: the integration boundaries in (E.7) result because the smallest dis-
tance of an element dξ to the point x is given by x, the largest distance by
∞.
Equation (E.7) can be transformed
√ by the substitution of the variable
ξ with η = 2√ξDt and thus dξ = 2 Dt dη:

∞
2C 0 x
C(x, t) = √ exp (−η 2 )dη = C 0 erfc( √ ) (E.8)
π 2 Dt
√x
2 Dt

The function erfc is the complement of the error function (see Sect. E.2):
x x
erfc( √ ) = 1 − erf( √ ) (E.9)
2 Dt 2 Dt
Bibliography

Abramowitz M, Stegun IA (1972) Handbook of mathematical functions.

Wiley, New York
Arrowsmith DK, Place CM (1992) Dynamical systems: differential equa-
tions, maps and chaotic behaviour. Chapman and Hall, London
Crank J (1975) The Mathematics of Diffusion. Clarendon Press, Oxford
Domenico PA, Schwartz FW (1998) Physical and Chemical Hydrology. 2.
Ed., Wiley, New York
Forster, P., V. Ramaswamy, P. Artaxo, T. Berntsen, R. Betts, D.W. Fa-
hey, J. Haywood, J. Lean, D.C. Lowe, G. Myhre, J. Nganga, R. Prinn,
G. Raga, M. Schulz and R. Van Dorland (2007) Changes in Atmo-
spheric Constituents and in Radiative Forcing. In: Climate Change 2007:
The Physical Science Basis. Contribution of Working Group I to the
Fourth Assessment Report of the Intergovernmental Panel on Climate
Change [Solomon, S., D. Qin, M. Manning, Z. Chen, M. Marquis, K.B.
Averyt, M.Tignor and H.L. Miller (eds.)]. Cambridge University Press,
Cambridge, United Kingdom and New York, NY, USA.
Haderlein SB, Weissmahr K, Schwarzenbach RP (1996) Specific adsorption
of nitroaromatic explosives and pesticides to clay minerals. Environ. Sci.
Technol. 30:612–622
Krebs CJ (2001) Ecology: The Experimental Analysis of Distribution and
Abundance. Benjamin Cummings, San Francisco, 5. Ed.
Lerman A (1972) Strontium-90 in the great lakes: Concentration-time
model. J. Geophys. Res. 77:3256–3264
Lotka AJ (1924) Elements of Physical Biology. Williams and Wilkins Co,
Inc. Republished by Dover Publications, New York 1956, under the title
Elements of Mathematical Biology
Luenberger DG (1979) Introduction to Dynamic Systems. Wiley, New York
May RM (1974) Stability and complexity in model ecosystems. Princeton
University Press, 2nd Ed.

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 247

DOI 10.1007/978-3-642-30639-6, © Springer-Verlag Berlin Heidelberg 2013
248 BIBLIOGRAPHY

May RM and McLean AR (eds) (2007) Theoretical Ecology: Principles and

Applications. Oxford University Press, New York, 3rd Ed.

MacLulich DA (1937) Fluctuations in the numbers of the varying hare

(Lepus americanus). Univ. Toronto Studies, Biol. Ser. 43
Meadows DH, Meadows DL, Randers J, Behrens WW (1972) The Limits
to Growth. Universe Books, New York
Meijer ML (2000) Biomanipulation in the Netherlands—15 Years of Expe-
rience. Wageningen Univ, Wageningen

Moore WS, Berrien III, Brasswell BH (1994) The lifetime of excess atmo-
spheric carbon dioxide. Glob. Biogeochem. Cycl. 8:23–38
Odum EP (1971) Fundamentals of Ecology. Saunders, Philadelphia
Pearl R, Reed LJ (1920) On the rate of growth of the population of the
United States since 1790 and its mathematical representation. Proc. Nat.
Acad. Csi. U.S. 6:275–288

Samuelson PA, Nordhaus W (2009) Economics. McGraw-Hill, New York,

19th Ed.
Schwarzenbach RP, Gschwend PM, Imboden DM (2003) Environmental
Organic Chemistry, 2nd edition. Wiley, New York
Scheffer M, Carpenter S, Foley JA, Folke C, Walker B (2002) Catastrophic
shifts in ecosystems. Nature 413:591–596
Strogatz SH (1994) Nonlinear dynamics and chaos: with applications to
physics, biology, chemistry, and engineering. Westview Press, Boulder
Verhulst PF (1838) Notice sur la loi que la population suit dans son
accroissement. Corresp. Math. Phys. 10:113–121
Vollenweider RA (1976) Advances in defining critical loading levels for
phosphorus in lake eutrophication. Mem. Ist. Ital. Idrobiol. 33:53–83
Volterra V (1926) Fluctuations in the abundance of a species considered
mathematically. Nature 118:558–560
Index

Adiabatic perturbation, 63, 65, 72, nonlinear, 174

141 Differential equation, 25
Adjustment time, 54, 55, 57, 58, 89, first-order, 25, 43
118 inhomogeneous, 43
relative factor, 55 initial value, 25
Advection, 192, 209 linear, 43
Asymptotically stable, 131 nonlinear, 125
Attractor, 130, 150 numerical solution, 166
asymptotically stable, 131 partial, 27
Autonomous, 47, 126 Diffusion, 192, 209
at a boundary area, 217
Balance equation, 44 molecular, 28, 194
Basic dimensions, 231 multi-dimensional, 219
Basin of attraction, 130 turbulent, 198, 199
Bernoulli numbers, 29 Diffusion equation
Bifurcation point, 179 one-dimensional, 213
Black box, 10 time-dependent, 243
Boundary conditions, 205 Dimension, 231
Box model, 11, 27 of a model, 16
of a variable, 19, 20
Carbon cycle, 114, 117
Discrete model, 165
Carbon dioxide, in the atmosphere,
first-order linear, 168, 170
23, 114, 115
Center, 147, 150, 156 higher-order linear, 171
Change function, 25, 127, 128 nonlinear, 174
Characteristic equation, 84, 113, 173, with several variables, 180
205, 240 Discriminant of a matrix, 84, 241
Coefficient matrix, 82 Distribution coefficient, 36
singular, 90 Divergence theorem, 202
triangular, 94
Creeping pendulum, 110 Economics, supply and demand, 40
Economic theory, 40
Damköhler number, 208 Eigenfunction, 183, 242
Decay curve, 55, 60 Eigenvalue, 84, 87, 147, 183, 205, 239
Determinant, 83 purely imaginary, 106
Deterministic chaos, 30, 157, 179 of a triangular matrix, 94
Difference equation, 26, 166 Einstein-Smoluchowski relation, 194,
first-order linear, 168 217
fixed point, 177 Epilimnion of a lake, 99

D.M. Imboden and S. Pfenninger, Introduction to Systems Analysis, 249

DOI 10.1007/978-3-642-30639-6, © Springer-Verlag Berlin Heidelberg 2013
250 INDEX

Equilibrium concentration, 50 Ingrowth curve, 55

Equilibrium distribution Inhomogeneous term, 44
sorption, 35 Initial condition, 47
between water and air, 34 Input
between water and sediment, 35 exponential growth, 61
Error function, 217, 244 variable, 15
Eutrophication, 17 Interest rate, 167
Exchange, 192 Internal relation, 4, 9, 15, 46
at boundary, 100, 194 Invariant area, 129
turbulent, 200
velocity, 195, 198 Jacobian matrix, 143, 145, 161
External relation, 5, 10, 15, 46
Lake
Fibonacci numbers, 187 as black box, 16
Fick’s first law, 193 eutrophication, 62, 64
Fick’s second law, 204, 219 flow reactor, 51, 54
Film model, gas exchange, 198 one-box model, 50
Fish pond, 45, 50, 132, 161, 169 oxygen, 28
Fixed point, 49, 127 periodic input, 70
asymptotically stable, 131 phosphorus, 16, 17, 25, 65–67,
classification, 146 76, 139, 142
as a function of the external radioactive isotope, 95
relation, 135 static model, 17
indifferent, 131 stratified, 98, 100, 122
line, 92, 140
as system, 6
one-sided stable, 131, 132
water renewal time, 57
stable, 130, 131
water temperature, 73
unstable, 131
Limit cycle, 147, 154
Flow rate, 51
Linear model
specific, 52
eigenvalues, 86
Flow reactor, linear, 50
with constant coefficients, 47
Flow vector, 192
with one variable, 43
Flux-gradient model, 194
periodic perturbation, 68
Gas exchange, 195, 197 with several variables, 81, 113
Gaussian distribution, 214, 243 with time-dependent
Gradient operator, 194 coefficients, 59
Growth model Linear superposition, 54, 60, 244
algae, 53 Linearization, 144, 150, 175
fish, 45, 132, 169 Logistic growth, 126, 127, 132
discrete, 176
Half-life, 57, 191 Lorenz system, 158
Harmonic oscillator, 111 Lotka, 9
Henry coefficient, 34, 35, 195 Lotka-Volterra. See Predator-prey
Hierarchical system, 93 model148
Holling-Tanner model, 154
Hypolimnion of a lake, 99 Mass balance, 25, 51, 82
Hysteresis, 134, 141 Mass flow, 192
 INDEX 251

Matrix, 82 Prediction, 8, 10
singular, 91 Probability, relations, 28
Mixing, 189
Mixing time, 191, 194 Radioactive decay, 93, 121
Model, 3, 7, 43 Radioactivity, 58
autonomous, 47 Rate
continuous in space, 27 first-order, 44
continuous in time, 26 specific, 44
deterministic, 27 total specific, 52
discrete time, 26, 165 zero-order, 44
dynamic, 23 Reaction
equation, 17 first-order, 91
formation, 10 homogeneous, 92
linear, 19 Reaction rate, first-order, 51
mathematical, 3, 7, 8, 15 Reactor, 50, 82, 85, 90
parameter, 16 Reductionism, 3
scientific theory, 9 Renewal time, 52
static, 16, 33, 50 Retention time, 52
stochastic, 27
Saddle point, 147
Singular matrix, 90
Non-autonomous, 59
Specific rate, 44
Nonlinear model
Stable star, 147
autonomous, 126
Steady state, 49
non-autonomous, 132
Steady-state concentration, 50
with one variable, 126
Strange attractor, 158
with several variables, 143 Structural instability, 109, 151
with three variables, 157 Superposition, linear, 54
Normal distribution, 214, 243 Synergism, 141, 142
System, 4
One-box model, linear, 44 boundary, 5, 10, 15
Oscillation stochastic, 29
creeping, 110 System equation, 19
damped, 108 of a dynamic model, 24
undamped, 88 Systems analysis, task of, 10
System variable, 10, 15
Péclet number, 208
Partial derivative, 144, 193 Thermocline, 99
Phase diagram, 88, 107, 110 Time dependence
Phase doubling, 179 explicit, 46, 59
Phase space, 88, 145 implicit, 46
Population model, 48 Time-discrete model, 165
Predator-prey model, 9, 148 Trace of a matrix, 83, 241
Holling-Tanner, 154, 161 Trajectory, 88, 145
Lotka-Volterra, 148, 151, Transformation, 189, 209
161–163 Transformation time, 191
velocity functions, 150 Transport
with self-interaction, 151 directed, 192
252 INDEX

by particles, 102 Unstable star, 147

undirected, 193
Transport/transformation equation,
Validation, 11
213
Vector, 82
Turbulent diffusion, 198
Vector field, 190
Two-box model, 81
Velocity function, 25, 135
for a stratified lake, 98, 105
Volterra, 9
Units, 19, 231
in the model equation, 21 World model, 7